WSL2-Linux-Kernel/fs/btrfs/tree-log.c

7021 строка
193 KiB
C
Исходник Обычный вид История

// SPDX-License-Identifier: GPL-2.0
/*
* Copyright (C) 2008 Oracle. All rights reserved.
*/
#include <linux/sched.h>
include cleanup: Update gfp.h and slab.h includes to prepare for breaking implicit slab.h inclusion from percpu.h percpu.h is included by sched.h and module.h and thus ends up being included when building most .c files. percpu.h includes slab.h which in turn includes gfp.h making everything defined by the two files universally available and complicating inclusion dependencies. percpu.h -> slab.h dependency is about to be removed. Prepare for this change by updating users of gfp and slab facilities include those headers directly instead of assuming availability. As this conversion needs to touch large number of source files, the following script is used as the basis of conversion. http://userweb.kernel.org/~tj/misc/slabh-sweep.py The script does the followings. * Scan files for gfp and slab usages and update includes such that only the necessary includes are there. ie. if only gfp is used, gfp.h, if slab is used, slab.h. * When the script inserts a new include, it looks at the include blocks and try to put the new include such that its order conforms to its surrounding. It's put in the include block which contains core kernel includes, in the same order that the rest are ordered - alphabetical, Christmas tree, rev-Xmas-tree or at the end if there doesn't seem to be any matching order. * If the script can't find a place to put a new include (mostly because the file doesn't have fitting include block), it prints out an error message indicating which .h file needs to be added to the file. The conversion was done in the following steps. 1. The initial automatic conversion of all .c files updated slightly over 4000 files, deleting around 700 includes and adding ~480 gfp.h and ~3000 slab.h inclusions. The script emitted errors for ~400 files. 2. Each error was manually checked. Some didn't need the inclusion, some needed manual addition while adding it to implementation .h or embedding .c file was more appropriate for others. This step added inclusions to around 150 files. 3. The script was run again and the output was compared to the edits from #2 to make sure no file was left behind. 4. Several build tests were done and a couple of problems were fixed. e.g. lib/decompress_*.c used malloc/free() wrappers around slab APIs requiring slab.h to be added manually. 5. The script was run on all .h files but without automatically editing them as sprinkling gfp.h and slab.h inclusions around .h files could easily lead to inclusion dependency hell. Most gfp.h inclusion directives were ignored as stuff from gfp.h was usually wildly available and often used in preprocessor macros. Each slab.h inclusion directive was examined and added manually as necessary. 6. percpu.h was updated not to include slab.h. 7. Build test were done on the following configurations and failures were fixed. CONFIG_GCOV_KERNEL was turned off for all tests (as my distributed build env didn't work with gcov compiles) and a few more options had to be turned off depending on archs to make things build (like ipr on powerpc/64 which failed due to missing writeq). * x86 and x86_64 UP and SMP allmodconfig and a custom test config. * powerpc and powerpc64 SMP allmodconfig * sparc and sparc64 SMP allmodconfig * ia64 SMP allmodconfig * s390 SMP allmodconfig * alpha SMP allmodconfig * um on x86_64 SMP allmodconfig 8. percpu.h modifications were reverted so that it could be applied as a separate patch and serve as bisection point. Given the fact that I had only a couple of failures from tests on step 6, I'm fairly confident about the coverage of this conversion patch. If there is a breakage, it's likely to be something in one of the arch headers which should be easily discoverable easily on most builds of the specific arch. Signed-off-by: Tejun Heo <tj@kernel.org> Guess-its-ok-by: Christoph Lameter <cl@linux-foundation.org> Cc: Ingo Molnar <mingo@redhat.com> Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
2010-03-24 11:04:11 +03:00
#include <linux/slab.h>
#include <linux/blkdev.h>
Btrfs: turbo charge fsync At least for the vm workload. Currently on fsync we will 1) Truncate all items in the log tree for the given inode if they exist and 2) Copy all items for a given inode into the log The problem with this is that for things like VMs you can have lots of extents from the fragmented writing behavior, and worst yet you may have only modified a few extents, not the entire thing. This patch fixes this problem by tracking which transid modified our extent, and then when we do the tree logging we find all of the extents we've modified in our current transaction, sort them and commit them. We also only truncate up to the xattrs of the inode and copy that stuff in normally, and then just drop any extents in the range we have that exist in the log already. Here are some numbers of a 50 meg fio job that does random writes and fsync()s after every write Original Patched SATA drive 82KB/s 140KB/s Fusion drive 431KB/s 2532KB/s So around 2-6 times faster depending on your hardware. There are a few corner cases, for example if you truncate at all we have to do it the old way since there is no way to be sure what is in the log is ok. This probably could be done smarter, but if you write-fsync-truncate-write-fsync you deserve what you get. All this work is in RAM of course so if your inode gets evicted from cache and you read it in and fsync it we'll do it the slow way if we are still in the same transaction that we last modified the inode in. The biggest cool part of this is that it requires no changes to the recovery code, so if you fsync with this patch and crash and load an old kernel, it will run the recovery and be a-ok. I have tested this pretty thoroughly with an fsync tester and everything comes back fine, as well as xfstests. Thanks, Signed-off-by: Josef Bacik <jbacik@fusionio.com>
2012-08-17 21:14:17 +04:00
#include <linux/list_sort.h>
#include <linux/iversion.h>
#include "misc.h"
btrfs: Remove custom crc32c init code The custom crc32 init code was introduced in 14a958e678cd ("Btrfs: fix btrfs boot when compiled as built-in") to enable using btrfs as a built-in. However, later as pointed out by 60efa5eb2e88 ("Btrfs: use late_initcall instead of module_init") this wasn't enough and finally btrfs was switched to late_initcall which comes after the generic crc32c implementation is initiliased. The latter commit superseeded the former. Now that we don't have to maintain our own code let's just remove it and switch to using the generic implementation. Despite touching a lot of files the patch is really simple. Here is the gist of the changes: 1. Select LIBCRC32C rather than the low-level modules. 2. s/btrfs_crc32c/crc32c/g 3. replace hash.h with linux/crc32c.h 4. Move the btrfs namehash funcs to ctree.h and change the tree accordingly. I've tested this with btrfs being both a module and a built-in and xfstest doesn't complain. Does seem to fix the longstanding problem of not automatically selectiong the crc32c module when btrfs is used. Possibly there is a workaround in dracut. The modinfo confirms that now all the module dependencies are there: before: depends: zstd_compress,zstd_decompress,raid6_pq,xor,zlib_deflate after: depends: libcrc32c,zstd_compress,zstd_decompress,raid6_pq,xor,zlib_deflate Signed-off-by: Nikolay Borisov <nborisov@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> [ add more info to changelog from mails ] Signed-off-by: David Sterba <dsterba@suse.com>
2018-01-08 12:45:05 +03:00
#include "ctree.h"
#include "tree-log.h"
#include "disk-io.h"
#include "locking.h"
#include "print-tree.h"
#include "backref.h"
#include "compression.h"
#include "qgroup.h"
#include "block-group.h"
#include "space-info.h"
#include "zoned.h"
#include "inode-item.h"
/* magic values for the inode_only field in btrfs_log_inode:
*
* LOG_INODE_ALL means to log everything
* LOG_INODE_EXISTS means to log just enough to recreate the inode
* during log replay
*/
enum {
LOG_INODE_ALL,
LOG_INODE_EXISTS,
LOG_OTHER_INODE,
LOG_OTHER_INODE_ALL,
};
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 17:24:20 +03:00
/*
* directory trouble cases
*
* 1) on rename or unlink, if the inode being unlinked isn't in the fsync
* log, we must force a full commit before doing an fsync of the directory
* where the unlink was done.
* ---> record transid of last unlink/rename per directory
*
* mkdir foo/some_dir
* normal commit
* rename foo/some_dir foo2/some_dir
* mkdir foo/some_dir
* fsync foo/some_dir/some_file
*
* The fsync above will unlink the original some_dir without recording
* it in its new location (foo2). After a crash, some_dir will be gone
* unless the fsync of some_file forces a full commit
*
* 2) we must log any new names for any file or dir that is in the fsync
* log. ---> check inode while renaming/linking.
*
* 2a) we must log any new names for any file or dir during rename
* when the directory they are being removed from was logged.
* ---> check inode and old parent dir during rename
*
* 2a is actually the more important variant. With the extra logging
* a crash might unlink the old name without recreating the new one
*
* 3) after a crash, we must go through any directories with a link count
* of zero and redo the rm -rf
*
* mkdir f1/foo
* normal commit
* rm -rf f1/foo
* fsync(f1)
*
* The directory f1 was fully removed from the FS, but fsync was never
* called on f1, only its parent dir. After a crash the rm -rf must
* be replayed. This must be able to recurse down the entire
* directory tree. The inode link count fixup code takes care of the
* ugly details.
*/
/*
* stages for the tree walking. The first
* stage (0) is to only pin down the blocks we find
* the second stage (1) is to make sure that all the inodes
* we find in the log are created in the subvolume.
*
* The last stage is to deal with directories and links and extents
* and all the other fun semantics
*/
enum {
LOG_WALK_PIN_ONLY,
LOG_WALK_REPLAY_INODES,
LOG_WALK_REPLAY_DIR_INDEX,
LOG_WALK_REPLAY_ALL,
};
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 17:24:20 +03:00
static int btrfs_log_inode(struct btrfs_trans_handle *trans,
struct btrfs_inode *inode,
int inode_only,
Btrfs: fix data corruption after fast fsync and writeback error When we do a fast fsync, we start all ordered operations and then while they're running in parallel we visit the list of modified extent maps and construct their matching file extent items and write them to the log btree. After that, in btrfs_sync_log() we wait for all the ordered operations to finish (via btrfs_wait_logged_extents). The problem with this is that we were completely ignoring errors that can happen in the extent write path, such as -ENOSPC, a temporary -ENOMEM or -EIO errors for example. When such error happens, it means we have parts of the on disk extent that weren't written to, and so we end up logging file extent items that point to these extents that contain garbage/random data - so after a crash/reboot plus log replay, we get our inode's metadata pointing to those extents. This worked in contrast with the full (non-fast) fsync path, where we start all ordered operations, wait for them to finish and then write to the log btree. In this path, after each ordered operation completes we check if it's flagged with an error (BTRFS_ORDERED_IOERR) and return -EIO if so (via btrfs_wait_ordered_range). So if an error happens with any ordered operation, just return a -EIO error to userspace, so that it knows that not all of its previous writes were durably persisted and the application can take proper action (like redo the writes for e.g.) - and definitely not leave any file extent items in the log refer to non fully written extents. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2014-09-05 18:14:39 +04:00
struct btrfs_log_ctx *ctx);
static int link_to_fixup_dir(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_path *path, u64 objectid);
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 17:24:20 +03:00
static noinline int replay_dir_deletes(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_root *log,
struct btrfs_path *path,
u64 dirid, int del_all);
static void wait_log_commit(struct btrfs_root *root, int transid);
/*
* tree logging is a special write ahead log used to make sure that
* fsyncs and O_SYNCs can happen without doing full tree commits.
*
* Full tree commits are expensive because they require commonly
* modified blocks to be recowed, creating many dirty pages in the
* extent tree an 4x-6x higher write load than ext3.
*
* Instead of doing a tree commit on every fsync, we use the
* key ranges and transaction ids to find items for a given file or directory
* that have changed in this transaction. Those items are copied into
* a special tree (one per subvolume root), that tree is written to disk
* and then the fsync is considered complete.
*
* After a crash, items are copied out of the log-tree back into the
* subvolume tree. Any file data extents found are recorded in the extent
* allocation tree, and the log-tree freed.
*
* The log tree is read three times, once to pin down all the extents it is
* using in ram and once, once to create all the inodes logged in the tree
* and once to do all the other items.
*/
/*
* start a sub transaction and setup the log tree
* this increments the log tree writer count to make the people
* syncing the tree wait for us to finish
*/
static int start_log_trans(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_log_ctx *ctx)
{
struct btrfs_fs_info *fs_info = root->fs_info;
btrfs: do not block inode logging for so long during transaction commit Early on during a transaction commit we acquire the tree_log_mutex and hold it until after we write the super blocks. But before writing the extent buffers dirtied by the transaction and the super blocks we unblock the transaction by setting its state to TRANS_STATE_UNBLOCKED and setting fs_info->running_transaction to NULL. This means that after that and before writing the super blocks, new transactions can start. However if any transaction wants to log an inode, it will block waiting for the transaction commit to write its dirty extent buffers and the super blocks because the tree_log_mutex is only released after those operations are complete, and starting a new log transaction blocks on that mutex (at start_log_trans()). Writing the dirty extent buffers and the super blocks can take a very significant amount of time to complete, but we could allow the tasks wanting to log an inode to proceed with most of their steps: 1) create the log trees 2) log metadata in the trees 3) write their dirty extent buffers They only need to wait for the previous transaction commit to complete (write its super blocks) before they attempt to write their super blocks, otherwise we could end up with a corrupt filesystem after a crash. So change start_log_trans() to use the root tree's log_mutex to serialize for the creation of the log root tree instead of using the tree_log_mutex, and make btrfs_sync_log() acquire the tree_log_mutex before writing the super blocks. This allows for inode logging to wait much less time when there is a previous transaction that is still committing, often not having to wait at all, as by the time when we try to sync the log the previous transaction already wrote its super blocks. This patch belongs to a patch set that is comprised of the following patches: btrfs: fix race causing unnecessary inode logging during link and rename btrfs: fix race that results in logging old extents during a fast fsync btrfs: fix race that causes unnecessary logging of ancestor inodes btrfs: fix race that makes inode logging fallback to transaction commit btrfs: fix race leading to unnecessary transaction commit when logging inode btrfs: do not block inode logging for so long during transaction commit The following script that uses dbench was used to measure the impact of the whole patchset: $ cat test-dbench.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/btrfs MOUNT_OPTIONS="-o ssd" echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f -m single -d single $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -D $MNT -t 300 64 umount $MNT The test was run on a machine with 12 cores, 64G of ram, using a NVMe device and a non-debug kernel configuration (Debian's default). Before patch set: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 11277211 0.250 85.340 Close 8283172 0.002 6.479 Rename 477515 1.935 86.026 Unlink 2277936 0.770 87.071 Deltree 256 15.732 81.379 Mkdir 128 0.003 0.009 Qpathinfo 10221180 0.056 44.404 Qfileinfo 1789967 0.002 4.066 Qfsinfo 1874399 0.003 9.176 Sfileinfo 918589 0.061 10.247 Find 3951758 0.341 54.040 WriteX 5616547 0.047 85.079 ReadX 17676028 0.005 9.704 LockX 36704 0.003 1.800 UnlockX 36704 0.002 0.687 Flush 790541 14.115 676.236 Throughput 1179.19 MB/sec 64 clients 64 procs max_latency=676.240 ms After patch set: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 12687926 0.171 86.526 Close 9320780 0.002 8.063 Rename 537253 1.444 78.576 Unlink 2561827 0.559 87.228 Deltree 374 11.499 73.549 Mkdir 187 0.003 0.005 Qpathinfo 11500300 0.061 36.801 Qfileinfo 2017118 0.002 7.189 Qfsinfo 2108641 0.003 4.825 Sfileinfo 1033574 0.008 8.065 Find 4446553 0.408 47.835 WriteX 6335667 0.045 84.388 ReadX 19887312 0.003 9.215 LockX 41312 0.003 1.394 UnlockX 41312 0.002 1.425 Flush 889233 13.014 623.259 Throughput 1339.32 MB/sec 64 clients 64 procs max_latency=623.265 ms +12.7% throughput, -8.2% max latency Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-11-25 15:19:28 +03:00
struct btrfs_root *tree_root = fs_info->tree_root;
const bool zoned = btrfs_is_zoned(fs_info);
int ret = 0;
bool created = false;
btrfs: do not block inode logging for so long during transaction commit Early on during a transaction commit we acquire the tree_log_mutex and hold it until after we write the super blocks. But before writing the extent buffers dirtied by the transaction and the super blocks we unblock the transaction by setting its state to TRANS_STATE_UNBLOCKED and setting fs_info->running_transaction to NULL. This means that after that and before writing the super blocks, new transactions can start. However if any transaction wants to log an inode, it will block waiting for the transaction commit to write its dirty extent buffers and the super blocks because the tree_log_mutex is only released after those operations are complete, and starting a new log transaction blocks on that mutex (at start_log_trans()). Writing the dirty extent buffers and the super blocks can take a very significant amount of time to complete, but we could allow the tasks wanting to log an inode to proceed with most of their steps: 1) create the log trees 2) log metadata in the trees 3) write their dirty extent buffers They only need to wait for the previous transaction commit to complete (write its super blocks) before they attempt to write their super blocks, otherwise we could end up with a corrupt filesystem after a crash. So change start_log_trans() to use the root tree's log_mutex to serialize for the creation of the log root tree instead of using the tree_log_mutex, and make btrfs_sync_log() acquire the tree_log_mutex before writing the super blocks. This allows for inode logging to wait much less time when there is a previous transaction that is still committing, often not having to wait at all, as by the time when we try to sync the log the previous transaction already wrote its super blocks. This patch belongs to a patch set that is comprised of the following patches: btrfs: fix race causing unnecessary inode logging during link and rename btrfs: fix race that results in logging old extents during a fast fsync btrfs: fix race that causes unnecessary logging of ancestor inodes btrfs: fix race that makes inode logging fallback to transaction commit btrfs: fix race leading to unnecessary transaction commit when logging inode btrfs: do not block inode logging for so long during transaction commit The following script that uses dbench was used to measure the impact of the whole patchset: $ cat test-dbench.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/btrfs MOUNT_OPTIONS="-o ssd" echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f -m single -d single $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -D $MNT -t 300 64 umount $MNT The test was run on a machine with 12 cores, 64G of ram, using a NVMe device and a non-debug kernel configuration (Debian's default). Before patch set: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 11277211 0.250 85.340 Close 8283172 0.002 6.479 Rename 477515 1.935 86.026 Unlink 2277936 0.770 87.071 Deltree 256 15.732 81.379 Mkdir 128 0.003 0.009 Qpathinfo 10221180 0.056 44.404 Qfileinfo 1789967 0.002 4.066 Qfsinfo 1874399 0.003 9.176 Sfileinfo 918589 0.061 10.247 Find 3951758 0.341 54.040 WriteX 5616547 0.047 85.079 ReadX 17676028 0.005 9.704 LockX 36704 0.003 1.800 UnlockX 36704 0.002 0.687 Flush 790541 14.115 676.236 Throughput 1179.19 MB/sec 64 clients 64 procs max_latency=676.240 ms After patch set: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 12687926 0.171 86.526 Close 9320780 0.002 8.063 Rename 537253 1.444 78.576 Unlink 2561827 0.559 87.228 Deltree 374 11.499 73.549 Mkdir 187 0.003 0.005 Qpathinfo 11500300 0.061 36.801 Qfileinfo 2017118 0.002 7.189 Qfsinfo 2108641 0.003 4.825 Sfileinfo 1033574 0.008 8.065 Find 4446553 0.408 47.835 WriteX 6335667 0.045 84.388 ReadX 19887312 0.003 9.215 LockX 41312 0.003 1.394 UnlockX 41312 0.002 1.425 Flush 889233 13.014 623.259 Throughput 1339.32 MB/sec 64 clients 64 procs max_latency=623.265 ms +12.7% throughput, -8.2% max latency Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-11-25 15:19:28 +03:00
/*
* First check if the log root tree was already created. If not, create
* it before locking the root's log_mutex, just to keep lockdep happy.
*/
if (!test_bit(BTRFS_ROOT_HAS_LOG_TREE, &tree_root->state)) {
mutex_lock(&tree_root->log_mutex);
if (!fs_info->log_root_tree) {
ret = btrfs_init_log_root_tree(trans, fs_info);
if (!ret) {
btrfs: do not block inode logging for so long during transaction commit Early on during a transaction commit we acquire the tree_log_mutex and hold it until after we write the super blocks. But before writing the extent buffers dirtied by the transaction and the super blocks we unblock the transaction by setting its state to TRANS_STATE_UNBLOCKED and setting fs_info->running_transaction to NULL. This means that after that and before writing the super blocks, new transactions can start. However if any transaction wants to log an inode, it will block waiting for the transaction commit to write its dirty extent buffers and the super blocks because the tree_log_mutex is only released after those operations are complete, and starting a new log transaction blocks on that mutex (at start_log_trans()). Writing the dirty extent buffers and the super blocks can take a very significant amount of time to complete, but we could allow the tasks wanting to log an inode to proceed with most of their steps: 1) create the log trees 2) log metadata in the trees 3) write their dirty extent buffers They only need to wait for the previous transaction commit to complete (write its super blocks) before they attempt to write their super blocks, otherwise we could end up with a corrupt filesystem after a crash. So change start_log_trans() to use the root tree's log_mutex to serialize for the creation of the log root tree instead of using the tree_log_mutex, and make btrfs_sync_log() acquire the tree_log_mutex before writing the super blocks. This allows for inode logging to wait much less time when there is a previous transaction that is still committing, often not having to wait at all, as by the time when we try to sync the log the previous transaction already wrote its super blocks. This patch belongs to a patch set that is comprised of the following patches: btrfs: fix race causing unnecessary inode logging during link and rename btrfs: fix race that results in logging old extents during a fast fsync btrfs: fix race that causes unnecessary logging of ancestor inodes btrfs: fix race that makes inode logging fallback to transaction commit btrfs: fix race leading to unnecessary transaction commit when logging inode btrfs: do not block inode logging for so long during transaction commit The following script that uses dbench was used to measure the impact of the whole patchset: $ cat test-dbench.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/btrfs MOUNT_OPTIONS="-o ssd" echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f -m single -d single $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -D $MNT -t 300 64 umount $MNT The test was run on a machine with 12 cores, 64G of ram, using a NVMe device and a non-debug kernel configuration (Debian's default). Before patch set: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 11277211 0.250 85.340 Close 8283172 0.002 6.479 Rename 477515 1.935 86.026 Unlink 2277936 0.770 87.071 Deltree 256 15.732 81.379 Mkdir 128 0.003 0.009 Qpathinfo 10221180 0.056 44.404 Qfileinfo 1789967 0.002 4.066 Qfsinfo 1874399 0.003 9.176 Sfileinfo 918589 0.061 10.247 Find 3951758 0.341 54.040 WriteX 5616547 0.047 85.079 ReadX 17676028 0.005 9.704 LockX 36704 0.003 1.800 UnlockX 36704 0.002 0.687 Flush 790541 14.115 676.236 Throughput 1179.19 MB/sec 64 clients 64 procs max_latency=676.240 ms After patch set: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 12687926 0.171 86.526 Close 9320780 0.002 8.063 Rename 537253 1.444 78.576 Unlink 2561827 0.559 87.228 Deltree 374 11.499 73.549 Mkdir 187 0.003 0.005 Qpathinfo 11500300 0.061 36.801 Qfileinfo 2017118 0.002 7.189 Qfsinfo 2108641 0.003 4.825 Sfileinfo 1033574 0.008 8.065 Find 4446553 0.408 47.835 WriteX 6335667 0.045 84.388 ReadX 19887312 0.003 9.215 LockX 41312 0.003 1.394 UnlockX 41312 0.002 1.425 Flush 889233 13.014 623.259 Throughput 1339.32 MB/sec 64 clients 64 procs max_latency=623.265 ms +12.7% throughput, -8.2% max latency Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-11-25 15:19:28 +03:00
set_bit(BTRFS_ROOT_HAS_LOG_TREE, &tree_root->state);
created = true;
}
btrfs: do not block inode logging for so long during transaction commit Early on during a transaction commit we acquire the tree_log_mutex and hold it until after we write the super blocks. But before writing the extent buffers dirtied by the transaction and the super blocks we unblock the transaction by setting its state to TRANS_STATE_UNBLOCKED and setting fs_info->running_transaction to NULL. This means that after that and before writing the super blocks, new transactions can start. However if any transaction wants to log an inode, it will block waiting for the transaction commit to write its dirty extent buffers and the super blocks because the tree_log_mutex is only released after those operations are complete, and starting a new log transaction blocks on that mutex (at start_log_trans()). Writing the dirty extent buffers and the super blocks can take a very significant amount of time to complete, but we could allow the tasks wanting to log an inode to proceed with most of their steps: 1) create the log trees 2) log metadata in the trees 3) write their dirty extent buffers They only need to wait for the previous transaction commit to complete (write its super blocks) before they attempt to write their super blocks, otherwise we could end up with a corrupt filesystem after a crash. So change start_log_trans() to use the root tree's log_mutex to serialize for the creation of the log root tree instead of using the tree_log_mutex, and make btrfs_sync_log() acquire the tree_log_mutex before writing the super blocks. This allows for inode logging to wait much less time when there is a previous transaction that is still committing, often not having to wait at all, as by the time when we try to sync the log the previous transaction already wrote its super blocks. This patch belongs to a patch set that is comprised of the following patches: btrfs: fix race causing unnecessary inode logging during link and rename btrfs: fix race that results in logging old extents during a fast fsync btrfs: fix race that causes unnecessary logging of ancestor inodes btrfs: fix race that makes inode logging fallback to transaction commit btrfs: fix race leading to unnecessary transaction commit when logging inode btrfs: do not block inode logging for so long during transaction commit The following script that uses dbench was used to measure the impact of the whole patchset: $ cat test-dbench.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/btrfs MOUNT_OPTIONS="-o ssd" echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f -m single -d single $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -D $MNT -t 300 64 umount $MNT The test was run on a machine with 12 cores, 64G of ram, using a NVMe device and a non-debug kernel configuration (Debian's default). Before patch set: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 11277211 0.250 85.340 Close 8283172 0.002 6.479 Rename 477515 1.935 86.026 Unlink 2277936 0.770 87.071 Deltree 256 15.732 81.379 Mkdir 128 0.003 0.009 Qpathinfo 10221180 0.056 44.404 Qfileinfo 1789967 0.002 4.066 Qfsinfo 1874399 0.003 9.176 Sfileinfo 918589 0.061 10.247 Find 3951758 0.341 54.040 WriteX 5616547 0.047 85.079 ReadX 17676028 0.005 9.704 LockX 36704 0.003 1.800 UnlockX 36704 0.002 0.687 Flush 790541 14.115 676.236 Throughput 1179.19 MB/sec 64 clients 64 procs max_latency=676.240 ms After patch set: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 12687926 0.171 86.526 Close 9320780 0.002 8.063 Rename 537253 1.444 78.576 Unlink 2561827 0.559 87.228 Deltree 374 11.499 73.549 Mkdir 187 0.003 0.005 Qpathinfo 11500300 0.061 36.801 Qfileinfo 2017118 0.002 7.189 Qfsinfo 2108641 0.003 4.825 Sfileinfo 1033574 0.008 8.065 Find 4446553 0.408 47.835 WriteX 6335667 0.045 84.388 ReadX 19887312 0.003 9.215 LockX 41312 0.003 1.394 UnlockX 41312 0.002 1.425 Flush 889233 13.014 623.259 Throughput 1339.32 MB/sec 64 clients 64 procs max_latency=623.265 ms +12.7% throughput, -8.2% max latency Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-11-25 15:19:28 +03:00
}
mutex_unlock(&tree_root->log_mutex);
if (ret)
return ret;
}
mutex_lock(&root->log_mutex);
again:
if (root->log_root) {
int index = (root->log_transid + 1) % 2;
if (btrfs_need_log_full_commit(trans)) {
ret = -EAGAIN;
goto out;
}
if (zoned && atomic_read(&root->log_commit[index])) {
wait_log_commit(root, root->log_transid - 1);
goto again;
}
if (!root->log_start_pid) {
clear_bit(BTRFS_ROOT_MULTI_LOG_TASKS, &root->state);
root->log_start_pid = current->pid;
} else if (root->log_start_pid != current->pid) {
set_bit(BTRFS_ROOT_MULTI_LOG_TASKS, &root->state);
}
} else {
/*
* This means fs_info->log_root_tree was already created
* for some other FS trees. Do the full commit not to mix
* nodes from multiple log transactions to do sequential
* writing.
*/
if (zoned && !created) {
ret = -EAGAIN;
goto out;
}
ret = btrfs_add_log_tree(trans, root);
if (ret)
goto out;
btrfs: check if a log root exists before locking the log_mutex on unlink This brings back an optimization that commit e678934cbe5f02 ("btrfs: Remove unnecessary check from join_running_log_trans") removed, but in a different form. So it's almost equivalent to a revert. That commit removed an optimization where we avoid locking a root's log_mutex when there is no log tree created in the current transaction. The affected code path is triggered through unlink operations. That commit was based on the assumption that the optimization was not necessary because we used to have the following checks when the patch was authored: int btrfs_del_dir_entries_in_log(...) { (...) if (dir->logged_trans < trans->transid) return 0; ret = join_running_log_trans(root); (...) } int btrfs_del_inode_ref_in_log(...) { (...) if (inode->logged_trans < trans->transid) return 0; ret = join_running_log_trans(root); (...) } However before that patch was merged, another patch was merged first which replaced those checks because they were buggy. That other patch corresponds to commit 803f0f64d17769 ("Btrfs: fix fsync not persisting dentry deletions due to inode evictions"). The assumption that if the logged_trans field of an inode had a smaller value then the current transaction's generation (transid) meant that the inode was not logged in the current transaction was only correct if the inode was not evicted and reloaded in the current transaction. So the corresponding bug fix changed those checks and replaced them with the following helper function: static bool inode_logged(struct btrfs_trans_handle *trans, struct btrfs_inode *inode) { if (inode->logged_trans == trans->transid) return true; if (inode->last_trans == trans->transid && test_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags) && !test_bit(BTRFS_FS_LOG_RECOVERING, &trans->fs_info->flags)) return true; return false; } So if we have a subvolume without a log tree in the current transaction (because we had no fsyncs), every time we unlink an inode we can end up trying to lock the log_mutex of the root through join_running_log_trans() twice, once for the inode being unlinked (by btrfs_del_inode_ref_in_log()) and once for the parent directory (with btrfs_del_dir_entries_in_log()). This means if we have several unlink operations happening in parallel for inodes in the same subvolume, and the those inodes and/or their parent inode were changed in the current transaction, we end up having a lot of contention on the log_mutex. The test robots from intel reported a -30.7% performance regression for a REAIM test after commit e678934cbe5f02 ("btrfs: Remove unnecessary check from join_running_log_trans"). So just bring back the optimization to join_running_log_trans() where we check first if a log root exists before trying to lock the log_mutex. This is done by checking for a bit that is set on the root when a log tree is created and removed when a log tree is freed (at transaction commit time). Commit e678934cbe5f02 ("btrfs: Remove unnecessary check from join_running_log_trans") was merged in the 5.4 merge window while commit 803f0f64d17769 ("Btrfs: fix fsync not persisting dentry deletions due to inode evictions") was merged in the 5.3 merge window. But the first commit was actually authored before the second commit (May 23 2019 vs June 19 2019). Reported-by: kernel test robot <rong.a.chen@intel.com> Link: https://lore.kernel.org/lkml/20200611090233.GL12456@shao2-debian/ Fixes: e678934cbe5f02 ("btrfs: Remove unnecessary check from join_running_log_trans") CC: stable@vger.kernel.org # 5.4+ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-06-15 12:38:44 +03:00
set_bit(BTRFS_ROOT_HAS_LOG_TREE, &root->state);
clear_bit(BTRFS_ROOT_MULTI_LOG_TASKS, &root->state);
root->log_start_pid = current->pid;
}
atomic_inc(&root->log_writers);
if (!ctx->logging_new_name) {
int index = root->log_transid % 2;
list_add_tail(&ctx->list, &root->log_ctxs[index]);
ctx->log_transid = root->log_transid;
}
out:
mutex_unlock(&root->log_mutex);
return ret;
}
/*
* returns 0 if there was a log transaction running and we were able
* to join, or returns -ENOENT if there were not transactions
* in progress
*/
static int join_running_log_trans(struct btrfs_root *root)
{
const bool zoned = btrfs_is_zoned(root->fs_info);
int ret = -ENOENT;
btrfs: check if a log root exists before locking the log_mutex on unlink This brings back an optimization that commit e678934cbe5f02 ("btrfs: Remove unnecessary check from join_running_log_trans") removed, but in a different form. So it's almost equivalent to a revert. That commit removed an optimization where we avoid locking a root's log_mutex when there is no log tree created in the current transaction. The affected code path is triggered through unlink operations. That commit was based on the assumption that the optimization was not necessary because we used to have the following checks when the patch was authored: int btrfs_del_dir_entries_in_log(...) { (...) if (dir->logged_trans < trans->transid) return 0; ret = join_running_log_trans(root); (...) } int btrfs_del_inode_ref_in_log(...) { (...) if (inode->logged_trans < trans->transid) return 0; ret = join_running_log_trans(root); (...) } However before that patch was merged, another patch was merged first which replaced those checks because they were buggy. That other patch corresponds to commit 803f0f64d17769 ("Btrfs: fix fsync not persisting dentry deletions due to inode evictions"). The assumption that if the logged_trans field of an inode had a smaller value then the current transaction's generation (transid) meant that the inode was not logged in the current transaction was only correct if the inode was not evicted and reloaded in the current transaction. So the corresponding bug fix changed those checks and replaced them with the following helper function: static bool inode_logged(struct btrfs_trans_handle *trans, struct btrfs_inode *inode) { if (inode->logged_trans == trans->transid) return true; if (inode->last_trans == trans->transid && test_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags) && !test_bit(BTRFS_FS_LOG_RECOVERING, &trans->fs_info->flags)) return true; return false; } So if we have a subvolume without a log tree in the current transaction (because we had no fsyncs), every time we unlink an inode we can end up trying to lock the log_mutex of the root through join_running_log_trans() twice, once for the inode being unlinked (by btrfs_del_inode_ref_in_log()) and once for the parent directory (with btrfs_del_dir_entries_in_log()). This means if we have several unlink operations happening in parallel for inodes in the same subvolume, and the those inodes and/or their parent inode were changed in the current transaction, we end up having a lot of contention on the log_mutex. The test robots from intel reported a -30.7% performance regression for a REAIM test after commit e678934cbe5f02 ("btrfs: Remove unnecessary check from join_running_log_trans"). So just bring back the optimization to join_running_log_trans() where we check first if a log root exists before trying to lock the log_mutex. This is done by checking for a bit that is set on the root when a log tree is created and removed when a log tree is freed (at transaction commit time). Commit e678934cbe5f02 ("btrfs: Remove unnecessary check from join_running_log_trans") was merged in the 5.4 merge window while commit 803f0f64d17769 ("Btrfs: fix fsync not persisting dentry deletions due to inode evictions") was merged in the 5.3 merge window. But the first commit was actually authored before the second commit (May 23 2019 vs June 19 2019). Reported-by: kernel test robot <rong.a.chen@intel.com> Link: https://lore.kernel.org/lkml/20200611090233.GL12456@shao2-debian/ Fixes: e678934cbe5f02 ("btrfs: Remove unnecessary check from join_running_log_trans") CC: stable@vger.kernel.org # 5.4+ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-06-15 12:38:44 +03:00
if (!test_bit(BTRFS_ROOT_HAS_LOG_TREE, &root->state))
return ret;
mutex_lock(&root->log_mutex);
again:
if (root->log_root) {
int index = (root->log_transid + 1) % 2;
ret = 0;
if (zoned && atomic_read(&root->log_commit[index])) {
wait_log_commit(root, root->log_transid - 1);
goto again;
}
atomic_inc(&root->log_writers);
}
mutex_unlock(&root->log_mutex);
return ret;
}
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 17:24:20 +03:00
/*
* This either makes the current running log transaction wait
* until you call btrfs_end_log_trans() or it makes any future
* log transactions wait until you call btrfs_end_log_trans()
*/
void btrfs_pin_log_trans(struct btrfs_root *root)
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 17:24:20 +03:00
{
atomic_inc(&root->log_writers);
}
/*
* indicate we're done making changes to the log tree
* and wake up anyone waiting to do a sync
*/
void btrfs_end_log_trans(struct btrfs_root *root)
{
if (atomic_dec_and_test(&root->log_writers)) {
/* atomic_dec_and_test implies a barrier */
cond_wake_up_nomb(&root->log_writer_wait);
}
}
static void btrfs_wait_tree_block_writeback(struct extent_buffer *buf)
{
filemap_fdatawait_range(buf->pages[0]->mapping,
buf->start, buf->start + buf->len - 1);
}
/*
* the walk control struct is used to pass state down the chain when
* processing the log tree. The stage field tells us which part
* of the log tree processing we are currently doing. The others
* are state fields used for that specific part
*/
struct walk_control {
/* should we free the extent on disk when done? This is used
* at transaction commit time while freeing a log tree
*/
int free;
/* pin only walk, we record which extents on disk belong to the
* log trees
*/
int pin;
/* what stage of the replay code we're currently in */
int stage;
Btrfs: fix warning when replaying log after fsync of a tmpfile When replaying a log which contains a tmpfile (which necessarily has a link count of 0) we end up calling inc_nlink(), at fs/btrfs/tree-log.c:replay_one_buffer(), which produces a warning like the following: [195191.943673] WARNING: CPU: 0 PID: 6924 at fs/inode.c:342 inc_nlink+0x33/0x40 [195191.943723] CPU: 0 PID: 6924 Comm: mount Not tainted 4.19.0-rc6-btrfs-next-38 #1 [195191.943724] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.11.2-0-gf9626ccb91-prebuilt.qemu-project.org 04/01/2014 [195191.943726] RIP: 0010:inc_nlink+0x33/0x40 [195191.943728] RSP: 0018:ffffb96e425e3870 EFLAGS: 00010246 [195191.943730] RAX: 0000000000000000 RBX: ffff8c0d1e6af4f0 RCX: 0000000000000006 [195191.943731] RDX: 0000000000000000 RSI: 0000000000000000 RDI: ffff8c0d1e6af4f0 [195191.943731] RBP: 0000000000000097 R08: 0000000000000001 R09: 0000000000000000 [195191.943732] R10: 0000000000000000 R11: 0000000000000000 R12: ffffb96e425e3a60 [195191.943733] R13: ffff8c0d10cff0c8 R14: ffff8c0d0d515348 R15: ffff8c0d78a1b3f8 [195191.943735] FS: 00007f570ee24480(0000) GS:ffff8c0dfb200000(0000) knlGS:0000000000000000 [195191.943736] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [195191.943737] CR2: 00005593286277c8 CR3: 00000000bb8f2006 CR4: 00000000003606f0 [195191.943739] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [195191.943740] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [195191.943741] Call Trace: [195191.943778] replay_one_buffer+0x797/0x7d0 [btrfs] [195191.943802] walk_up_log_tree+0x1c1/0x250 [btrfs] [195191.943809] ? rcu_read_lock_sched_held+0x3f/0x70 [195191.943825] walk_log_tree+0xae/0x1d0 [btrfs] [195191.943840] btrfs_recover_log_trees+0x1d7/0x4d0 [btrfs] [195191.943856] ? replay_dir_deletes+0x280/0x280 [btrfs] [195191.943870] open_ctree+0x1c3b/0x22a0 [btrfs] [195191.943887] btrfs_mount_root+0x6b4/0x800 [btrfs] [195191.943894] ? rcu_read_lock_sched_held+0x3f/0x70 [195191.943899] ? pcpu_alloc+0x55b/0x7c0 [195191.943906] ? mount_fs+0x3b/0x140 [195191.943908] mount_fs+0x3b/0x140 [195191.943912] ? __init_waitqueue_head+0x36/0x50 [195191.943916] vfs_kern_mount+0x62/0x160 [195191.943927] btrfs_mount+0x134/0x890 [btrfs] [195191.943936] ? rcu_read_lock_sched_held+0x3f/0x70 [195191.943938] ? pcpu_alloc+0x55b/0x7c0 [195191.943943] ? mount_fs+0x3b/0x140 [195191.943952] ? btrfs_remount+0x570/0x570 [btrfs] [195191.943954] mount_fs+0x3b/0x140 [195191.943956] ? __init_waitqueue_head+0x36/0x50 [195191.943960] vfs_kern_mount+0x62/0x160 [195191.943963] do_mount+0x1f9/0xd40 [195191.943967] ? memdup_user+0x4b/0x70 [195191.943971] ksys_mount+0x7e/0xd0 [195191.943974] __x64_sys_mount+0x21/0x30 [195191.943977] do_syscall_64+0x60/0x1b0 [195191.943980] entry_SYSCALL_64_after_hwframe+0x49/0xbe [195191.943983] RIP: 0033:0x7f570e4e524a [195191.943986] RSP: 002b:00007ffd83589478 EFLAGS: 00000206 ORIG_RAX: 00000000000000a5 [195191.943989] RAX: ffffffffffffffda RBX: 0000563f335b2060 RCX: 00007f570e4e524a [195191.943990] RDX: 0000563f335b2240 RSI: 0000563f335b2280 RDI: 0000563f335b2260 [195191.943992] RBP: 0000000000000000 R08: 0000000000000000 R09: 0000000000000020 [195191.943993] R10: 00000000c0ed0000 R11: 0000000000000206 R12: 0000563f335b2260 [195191.943994] R13: 0000563f335b2240 R14: 0000000000000000 R15: 00000000ffffffff [195191.944002] irq event stamp: 8688 [195191.944010] hardirqs last enabled at (8687): [<ffffffff9cb004c3>] console_unlock+0x503/0x640 [195191.944012] hardirqs last disabled at (8688): [<ffffffff9ca037dd>] trace_hardirqs_off_thunk+0x1a/0x1c [195191.944018] softirqs last enabled at (8638): [<ffffffff9cc0a5d1>] __set_page_dirty_nobuffers+0x101/0x150 [195191.944020] softirqs last disabled at (8634): [<ffffffff9cc26bbe>] wb_wakeup_delayed+0x2e/0x60 [195191.944022] ---[ end trace 5d6e873a9a0b811a ]--- This happens because the inode does not have the flag I_LINKABLE set, which is a runtime only flag, not meant to be persisted, set when the inode is created through open(2) if the flag O_EXCL is not passed to it. Except for the warning, there are no other consequences (like corruptions or metadata inconsistencies). Since it's pointless to replay a tmpfile as it would be deleted in a later phase of the log replay procedure (it has a link count of 0), fix this by not logging tmpfiles and if a tmpfile is found in a log (created by a kernel without this change), skip the replay of the inode. A test case for fstests follows soon. Fixes: 471d557afed1 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay") CC: stable@vger.kernel.org # 4.18+ Reported-by: Martin Steigerwald <martin@lichtvoll.de> Link: https://lore.kernel.org/linux-btrfs/3666619.NTnn27ZJZE@merkaba/ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2018-10-08 13:12:55 +03:00
/*
* Ignore any items from the inode currently being processed. Needs
* to be set every time we find a BTRFS_INODE_ITEM_KEY and we are in
* the LOG_WALK_REPLAY_INODES stage.
*/
bool ignore_cur_inode;
/* the root we are currently replaying */
struct btrfs_root *replay_dest;
/* the trans handle for the current replay */
struct btrfs_trans_handle *trans;
/* the function that gets used to process blocks we find in the
* tree. Note the extent_buffer might not be up to date when it is
* passed in, and it must be checked or read if you need the data
* inside it
*/
int (*process_func)(struct btrfs_root *log, struct extent_buffer *eb,
struct walk_control *wc, u64 gen, int level);
};
/*
* process_func used to pin down extents, write them or wait on them
*/
static int process_one_buffer(struct btrfs_root *log,
struct extent_buffer *eb,
struct walk_control *wc, u64 gen, int level)
{
struct btrfs_fs_info *fs_info = log->fs_info;
int ret = 0;
/*
* If this fs is mixed then we need to be able to process the leaves to
* pin down any logged extents, so we have to read the block.
*/
if (btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
ret = btrfs_read_buffer(eb, gen, level, NULL);
if (ret)
return ret;
}
if (wc->pin) {
ret = btrfs_pin_extent_for_log_replay(wc->trans, eb->start,
eb->len);
if (ret)
return ret;
if (btrfs_buffer_uptodate(eb, gen, 0) &&
btrfs_header_level(eb) == 0)
ret = btrfs_exclude_logged_extents(eb);
}
return ret;
}
btrfs: insert items in batches when logging a directory when possible When logging a directory, we scan its directory items from the subvolume tree and then copy one by one into the log tree. This is not efficient since we generally are able to insert several items in a batch, using a single btree operation for adding several items at once. The reason we copy items one by one is that we must check if each item was previously logged in the current transaction, and if it was we either overwrite it or skip it in case its content did not change in the subvolume tree (this can happen only for dir item keys, but not for dir index keys), and doing such check makes it a bit cumbersome to attempt batch insertions. However the chances for doing batch insertions are very frequent and always happen when: 1) Logging the directory for the first time in the current transaction, as none of the items exist in the log tree yet; 2) Logging new dir index keys, because the offset for new dir index keys comes from a monotonically increasing counter. This means if we keep adding dentries to a directory, through creation of new files and sub-directories or by adding new links or renaming from some other directory into the one we are logging, all the new dir index keys have a new offset that is greater than the offset of any previously logged index keys, so we can insert them in batches into the log tree. For dir item keys, since their offset depends on the result of an hash function against the dentry's name, unless the directory is being logged for the first time in the current transaction, the chances being able to insert the items in the log using batches is pretty much random and not predictable, as it depends on the names of the dentries, but still happens often enough. So change directory logging to keep track of consecutive directory items that don't exist yet in the log and batch insert them. This patch is part of a patchset comprised of the following 5 patches: btrfs: remove root argument from btrfs_log_inode() and its callees btrfs: remove redundant log root assignment from log_dir_items() btrfs: factor out the copying loop of dir items from log_dir_items() btrfs: insert items in batches when logging a directory when possible btrfs: keep track of the last logged keys when logging a directory This is patch 4/5. The change log of the last patch (5/5) has performance results. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-16 13:32:13 +03:00
static int do_overwrite_item(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_path *path,
struct extent_buffer *eb, int slot,
struct btrfs_key *key)
{
int ret;
u32 item_size;
u64 saved_i_size = 0;
int save_old_i_size = 0;
unsigned long src_ptr;
unsigned long dst_ptr;
int overwrite_root = 0;
bool inode_item = key->type == BTRFS_INODE_ITEM_KEY;
if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID)
overwrite_root = 1;
item_size = btrfs_item_size(eb, slot);
src_ptr = btrfs_item_ptr_offset(eb, slot);
btrfs: insert items in batches when logging a directory when possible When logging a directory, we scan its directory items from the subvolume tree and then copy one by one into the log tree. This is not efficient since we generally are able to insert several items in a batch, using a single btree operation for adding several items at once. The reason we copy items one by one is that we must check if each item was previously logged in the current transaction, and if it was we either overwrite it or skip it in case its content did not change in the subvolume tree (this can happen only for dir item keys, but not for dir index keys), and doing such check makes it a bit cumbersome to attempt batch insertions. However the chances for doing batch insertions are very frequent and always happen when: 1) Logging the directory for the first time in the current transaction, as none of the items exist in the log tree yet; 2) Logging new dir index keys, because the offset for new dir index keys comes from a monotonically increasing counter. This means if we keep adding dentries to a directory, through creation of new files and sub-directories or by adding new links or renaming from some other directory into the one we are logging, all the new dir index keys have a new offset that is greater than the offset of any previously logged index keys, so we can insert them in batches into the log tree. For dir item keys, since their offset depends on the result of an hash function against the dentry's name, unless the directory is being logged for the first time in the current transaction, the chances being able to insert the items in the log using batches is pretty much random and not predictable, as it depends on the names of the dentries, but still happens often enough. So change directory logging to keep track of consecutive directory items that don't exist yet in the log and batch insert them. This patch is part of a patchset comprised of the following 5 patches: btrfs: remove root argument from btrfs_log_inode() and its callees btrfs: remove redundant log root assignment from log_dir_items() btrfs: factor out the copying loop of dir items from log_dir_items() btrfs: insert items in batches when logging a directory when possible btrfs: keep track of the last logged keys when logging a directory This is patch 4/5. The change log of the last patch (5/5) has performance results. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-16 13:32:13 +03:00
/* Our caller must have done a search for the key for us. */
ASSERT(path->nodes[0] != NULL);
/*
* And the slot must point to the exact key or the slot where the key
* should be at (the first item with a key greater than 'key')
*/
if (path->slots[0] < btrfs_header_nritems(path->nodes[0])) {
struct btrfs_key found_key;
btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
ret = btrfs_comp_cpu_keys(&found_key, key);
ASSERT(ret >= 0);
} else {
ret = 1;
}
if (ret == 0) {
char *src_copy;
char *dst_copy;
u32 dst_size = btrfs_item_size(path->nodes[0],
path->slots[0]);
if (dst_size != item_size)
goto insert;
if (item_size == 0) {
btrfs_release_path(path);
return 0;
}
dst_copy = kmalloc(item_size, GFP_NOFS);
src_copy = kmalloc(item_size, GFP_NOFS);
if (!dst_copy || !src_copy) {
btrfs_release_path(path);
kfree(dst_copy);
kfree(src_copy);
return -ENOMEM;
}
read_extent_buffer(eb, src_copy, src_ptr, item_size);
dst_ptr = btrfs_item_ptr_offset(path->nodes[0], path->slots[0]);
read_extent_buffer(path->nodes[0], dst_copy, dst_ptr,
item_size);
ret = memcmp(dst_copy, src_copy, item_size);
kfree(dst_copy);
kfree(src_copy);
/*
* they have the same contents, just return, this saves
* us from cowing blocks in the destination tree and doing
* extra writes that may not have been done by a previous
* sync
*/
if (ret == 0) {
btrfs_release_path(path);
return 0;
}
/*
* We need to load the old nbytes into the inode so when we
* replay the extents we've logged we get the right nbytes.
*/
if (inode_item) {
struct btrfs_inode_item *item;
u64 nbytes;
u32 mode;
item = btrfs_item_ptr(path->nodes[0], path->slots[0],
struct btrfs_inode_item);
nbytes = btrfs_inode_nbytes(path->nodes[0], item);
item = btrfs_item_ptr(eb, slot,
struct btrfs_inode_item);
btrfs_set_inode_nbytes(eb, item, nbytes);
/*
* If this is a directory we need to reset the i_size to
* 0 so that we can set it up properly when replaying
* the rest of the items in this log.
*/
mode = btrfs_inode_mode(eb, item);
if (S_ISDIR(mode))
btrfs_set_inode_size(eb, item, 0);
}
} else if (inode_item) {
struct btrfs_inode_item *item;
u32 mode;
/*
* New inode, set nbytes to 0 so that the nbytes comes out
* properly when we replay the extents.
*/
item = btrfs_item_ptr(eb, slot, struct btrfs_inode_item);
btrfs_set_inode_nbytes(eb, item, 0);
/*
* If this is a directory we need to reset the i_size to 0 so
* that we can set it up properly when replaying the rest of
* the items in this log.
*/
mode = btrfs_inode_mode(eb, item);
if (S_ISDIR(mode))
btrfs_set_inode_size(eb, item, 0);
}
insert:
btrfs_release_path(path);
/* try to insert the key into the destination tree */
Btrfs: fix fsync log replay for inodes with a mix of regular refs and extrefs If we have an inode with a large number of hard links, some of which may be extrefs, turn a regular ref into an extref, fsync the inode and then replay the fsync log (after a crash/reboot), we can endup with an fsync log that makes the replay code always fail with -EOVERFLOW when processing the inode's references. This is easy to reproduce with the test case I made for xfstests. Its steps are the following: _scratch_mkfs "-O extref" >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create a test file with 3001 hard links. This number is large enough to # make btrfs start using extrefs at some point even if the fs has the maximum # possible leaf/node size (64Kb). echo "hello world" > $SCRATCH_MNT/foo for i in `seq 1 3000`; do ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link_`printf "%04d" $i` done # Make sure all metadata and data are durably persisted. sync # Now remove one link, add a new one with a new name, add another new one with # the same name as the one we just removed and fsync the inode. rm -f $SCRATCH_MNT/foo_link_0001 ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link_3001 ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link_0001 rm -f $SCRATCH_MNT/foo_link_0002 ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link_3002 ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link_3003 $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foo # Simulate a crash/power loss. This makes sure the next mount # will see an fsync log and will replay that log. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # Check that the number of hard links is correct, we are able to remove all # the hard links and read the file's data. This is just to verify we don't # get stale file handle errors (due to dangling directory index entries that # point to inodes that no longer exist). echo "Link count: $(stat --format=%h $SCRATCH_MNT/foo)" [ -f $SCRATCH_MNT/foo ] || echo "Link foo is missing" for ((i = 1; i <= 3003; i++)); do name=foo_link_`printf "%04d" $i` if [ $i -eq 2 ]; then [ -f $SCRATCH_MNT/$name ] && echo "Link $name found" else [ -f $SCRATCH_MNT/$name ] || echo "Link $name is missing" fi done rm -f $SCRATCH_MNT/foo_link_* cat $SCRATCH_MNT/foo rm -f $SCRATCH_MNT/foo status=0 exit The fix is simply to correct the overflow condition when overwriting a reference item because it was wrong, trying to increase the item in the fs/subvol tree by an impossible amount. Also ensure that we don't insert one normal ref and one ext ref for the same dentry - this happened because processing a dir index entry from the parent in the log happened when the normal ref item was full, which made the logic insert an extref and later when the normal ref had enough room, it would be inserted again when processing the ref item from the child inode in the log. This issue has been present since the introduction of the extrefs feature (2012). A test case for xfstests follows soon. This test only passes if the previous patch titled "Btrfs: fix fsync when extend references are added to an inode" is applied too. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-01-14 04:52:25 +03:00
path->skip_release_on_error = 1;
ret = btrfs_insert_empty_item(trans, root, path,
key, item_size);
Btrfs: fix fsync log replay for inodes with a mix of regular refs and extrefs If we have an inode with a large number of hard links, some of which may be extrefs, turn a regular ref into an extref, fsync the inode and then replay the fsync log (after a crash/reboot), we can endup with an fsync log that makes the replay code always fail with -EOVERFLOW when processing the inode's references. This is easy to reproduce with the test case I made for xfstests. Its steps are the following: _scratch_mkfs "-O extref" >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create a test file with 3001 hard links. This number is large enough to # make btrfs start using extrefs at some point even if the fs has the maximum # possible leaf/node size (64Kb). echo "hello world" > $SCRATCH_MNT/foo for i in `seq 1 3000`; do ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link_`printf "%04d" $i` done # Make sure all metadata and data are durably persisted. sync # Now remove one link, add a new one with a new name, add another new one with # the same name as the one we just removed and fsync the inode. rm -f $SCRATCH_MNT/foo_link_0001 ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link_3001 ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link_0001 rm -f $SCRATCH_MNT/foo_link_0002 ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link_3002 ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link_3003 $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foo # Simulate a crash/power loss. This makes sure the next mount # will see an fsync log and will replay that log. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # Check that the number of hard links is correct, we are able to remove all # the hard links and read the file's data. This is just to verify we don't # get stale file handle errors (due to dangling directory index entries that # point to inodes that no longer exist). echo "Link count: $(stat --format=%h $SCRATCH_MNT/foo)" [ -f $SCRATCH_MNT/foo ] || echo "Link foo is missing" for ((i = 1; i <= 3003; i++)); do name=foo_link_`printf "%04d" $i` if [ $i -eq 2 ]; then [ -f $SCRATCH_MNT/$name ] && echo "Link $name found" else [ -f $SCRATCH_MNT/$name ] || echo "Link $name is missing" fi done rm -f $SCRATCH_MNT/foo_link_* cat $SCRATCH_MNT/foo rm -f $SCRATCH_MNT/foo status=0 exit The fix is simply to correct the overflow condition when overwriting a reference item because it was wrong, trying to increase the item in the fs/subvol tree by an impossible amount. Also ensure that we don't insert one normal ref and one ext ref for the same dentry - this happened because processing a dir index entry from the parent in the log happened when the normal ref item was full, which made the logic insert an extref and later when the normal ref had enough room, it would be inserted again when processing the ref item from the child inode in the log. This issue has been present since the introduction of the extrefs feature (2012). A test case for xfstests follows soon. This test only passes if the previous patch titled "Btrfs: fix fsync when extend references are added to an inode" is applied too. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-01-14 04:52:25 +03:00
path->skip_release_on_error = 0;
/* make sure any existing item is the correct size */
Btrfs: fix fsync log replay for inodes with a mix of regular refs and extrefs If we have an inode with a large number of hard links, some of which may be extrefs, turn a regular ref into an extref, fsync the inode and then replay the fsync log (after a crash/reboot), we can endup with an fsync log that makes the replay code always fail with -EOVERFLOW when processing the inode's references. This is easy to reproduce with the test case I made for xfstests. Its steps are the following: _scratch_mkfs "-O extref" >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create a test file with 3001 hard links. This number is large enough to # make btrfs start using extrefs at some point even if the fs has the maximum # possible leaf/node size (64Kb). echo "hello world" > $SCRATCH_MNT/foo for i in `seq 1 3000`; do ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link_`printf "%04d" $i` done # Make sure all metadata and data are durably persisted. sync # Now remove one link, add a new one with a new name, add another new one with # the same name as the one we just removed and fsync the inode. rm -f $SCRATCH_MNT/foo_link_0001 ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link_3001 ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link_0001 rm -f $SCRATCH_MNT/foo_link_0002 ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link_3002 ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link_3003 $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foo # Simulate a crash/power loss. This makes sure the next mount # will see an fsync log and will replay that log. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # Check that the number of hard links is correct, we are able to remove all # the hard links and read the file's data. This is just to verify we don't # get stale file handle errors (due to dangling directory index entries that # point to inodes that no longer exist). echo "Link count: $(stat --format=%h $SCRATCH_MNT/foo)" [ -f $SCRATCH_MNT/foo ] || echo "Link foo is missing" for ((i = 1; i <= 3003; i++)); do name=foo_link_`printf "%04d" $i` if [ $i -eq 2 ]; then [ -f $SCRATCH_MNT/$name ] && echo "Link $name found" else [ -f $SCRATCH_MNT/$name ] || echo "Link $name is missing" fi done rm -f $SCRATCH_MNT/foo_link_* cat $SCRATCH_MNT/foo rm -f $SCRATCH_MNT/foo status=0 exit The fix is simply to correct the overflow condition when overwriting a reference item because it was wrong, trying to increase the item in the fs/subvol tree by an impossible amount. Also ensure that we don't insert one normal ref and one ext ref for the same dentry - this happened because processing a dir index entry from the parent in the log happened when the normal ref item was full, which made the logic insert an extref and later when the normal ref had enough room, it would be inserted again when processing the ref item from the child inode in the log. This issue has been present since the introduction of the extrefs feature (2012). A test case for xfstests follows soon. This test only passes if the previous patch titled "Btrfs: fix fsync when extend references are added to an inode" is applied too. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-01-14 04:52:25 +03:00
if (ret == -EEXIST || ret == -EOVERFLOW) {
u32 found_size;
found_size = btrfs_item_size(path->nodes[0],
path->slots[0]);
if (found_size > item_size)
btrfs_truncate_item(path, item_size, 1);
else if (found_size < item_size)
btrfs_extend_item(path, item_size - found_size);
} else if (ret) {
return ret;
}
dst_ptr = btrfs_item_ptr_offset(path->nodes[0],
path->slots[0]);
/* don't overwrite an existing inode if the generation number
* was logged as zero. This is done when the tree logging code
* is just logging an inode to make sure it exists after recovery.
*
* Also, don't overwrite i_size on directories during replay.
* log replay inserts and removes directory items based on the
* state of the tree found in the subvolume, and i_size is modified
* as it goes
*/
if (key->type == BTRFS_INODE_ITEM_KEY && ret == -EEXIST) {
struct btrfs_inode_item *src_item;
struct btrfs_inode_item *dst_item;
src_item = (struct btrfs_inode_item *)src_ptr;
dst_item = (struct btrfs_inode_item *)dst_ptr;
Btrfs: fix fsync data loss after adding hard link to inode We have a scenario where after the fsync log replay we can lose file data that had been previously fsync'ed if we added an hard link for our inode and after that we sync'ed the fsync log (for example by fsync'ing some other file or directory). This is because when adding an hard link we updated the inode item in the log tree with an i_size value of 0. At that point the new inode item was in memory only and a subsequent fsync log replay would not make us lose the file data. However if after adding the hard link we sync the log tree to disk, by fsync'ing some other file or directory for example, we ended up losing the file data after log replay, because the inode item in the persisted log tree had an an i_size of zero. This is easy to reproduce, and the following excerpt from my test for xfstests shows this: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create one file with data and fsync it. # This made the btrfs fsync log persist the data and the inode metadata with # a correct inode->i_size (4096 bytes). $XFS_IO_PROG -f -c "pwrite -S 0xaa -b 4K 0 4K" -c "fsync" \ $SCRATCH_MNT/foo | _filter_xfs_io # Now add one hard link to our file. This made the btrfs code update the fsync # log, in memory only, with an inode metadata having a size of 0. ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link # Now force persistence of the fsync log to disk, for example, by fsyncing some # other file. touch $SCRATCH_MNT/bar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/bar # Before a power loss or crash, we could read the 4Kb of data from our file as # expected. echo "File content before:" od -t x1 $SCRATCH_MNT/foo # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # After the fsync log replay, because the fsync log had a value of 0 for our # inode's i_size, we couldn't read anymore the 4Kb of data that we previously # wrote and fsync'ed. The size of the file became 0 after the fsync log replay. echo "File content after:" od -t x1 $SCRATCH_MNT/foo Another alternative test, that doesn't need to fsync an inode in the same transaction it was created, is: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create our test file with some data. $XFS_IO_PROG -f -c "pwrite -S 0xaa -b 8K 0 8K" \ $SCRATCH_MNT/foo | _filter_xfs_io # Make sure the file is durably persisted. sync # Append some data to our file, to increase its size. $XFS_IO_PROG -f -c "pwrite -S 0xcc -b 4K 8K 4K" \ $SCRATCH_MNT/foo | _filter_xfs_io # Fsync the file, so from this point on if a crash/power failure happens, our # new data is guaranteed to be there next time the fs is mounted. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foo # Add one hard link to our file. This made btrfs write into the in memory fsync # log a special inode with generation 0 and an i_size of 0 too. Note that this # didn't update the inode in the fsync log on disk. ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link # Now make sure the in memory fsync log is durably persisted. # Creating and fsync'ing another file will do it. touch $SCRATCH_MNT/bar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/bar # As expected, before the crash/power failure, we should be able to read the # 12Kb of file data. echo "File content before:" od -t x1 $SCRATCH_MNT/foo # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # After mounting the fs again, the fsync log was replayed. # The btrfs fsync log replay code didn't update the i_size of the persisted # inode because the inode item in the log had a special generation with a # value of 0 (and it couldn't know the correct i_size, since that inode item # had a 0 i_size too). This made the last 4Kb of file data inaccessible and # effectively lost. echo "File content after:" od -t x1 $SCRATCH_MNT/foo This isn't a new issue/regression. This problem has been around since the log tree code was added in 2008: Btrfs: Add a write ahead tree log to optimize synchronous operations (commit e02119d5a7b4396c5a872582fddc8bd6d305a70a) Test cases for xfstests follow soon. CC: <stable@vger.kernel.org> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-02-13 15:30:56 +03:00
if (btrfs_inode_generation(eb, src_item) == 0) {
struct extent_buffer *dst_eb = path->nodes[0];
Btrfs: fix metadata inconsistencies after directory fsync We can get into inconsistency between inodes and directory entries after fsyncing a directory. The issue is that while a directory gets the new dentries persisted in the fsync log and replayed at mount time, the link count of the inode that directory entries point to doesn't get updated, staying with an incorrect link count (smaller then the correct value). This later leads to stale file handle errors when accessing (including attempt to delete) some of the links if all the other ones are removed, which also implies impossibility to delete the parent directories, since the dentries can not be removed. Another issue is that (unlike ext3/4, xfs, f2fs, reiserfs, nilfs2), when fsyncing a directory, new files aren't logged (their metadata and dentries) nor any child directories. So this patch fixes this issue too, since it has the same resolution as the incorrect inode link count issue mentioned before. This is very easy to reproduce, and the following excerpt from my test case for xfstests shows how: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create our main test file and directory. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 8K" $SCRATCH_MNT/foo | _filter_xfs_io mkdir $SCRATCH_MNT/mydir # Make sure all metadata and data are durably persisted. sync # Add a hard link to 'foo' inside our test directory and fsync only the # directory. The btrfs fsync implementation had a bug that caused the new # directory entry to be visible after the fsync log replay but, the inode # of our file remained with a link count of 1. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_2 # Add a few more links and new files. # This is just to verify nothing breaks or gives incorrect results after the # fsync log is replayed. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_3 $XFS_IO_PROG -f -c "pwrite -S 0xff 0 64K" $SCRATCH_MNT/hello | _filter_xfs_io ln $SCRATCH_MNT/hello $SCRATCH_MNT/mydir/hello_2 # Add some subdirectories and new files and links to them. This is to verify # that after fsyncing our top level directory 'mydir', all the subdirectories # and their files/links are registered in the fsync log and exist after the # fsync log is replayed. mkdir -p $SCRATCH_MNT/mydir/x/y/z ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/foo_y_link ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/z/foo_z_link touch $SCRATCH_MNT/mydir/x/y/z/qwerty # Now fsync only our top directory. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/mydir # And fsync now our new file named 'hello', just to verify later that it has # the expected content and that the previous fsync on the directory 'mydir' had # no bad influence on this fsync. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/hello # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # Verify the content of our file 'foo' remains the same as before, 8192 bytes, # all with the value 0xaa. echo "File 'foo' content after log replay:" od -t x1 $SCRATCH_MNT/foo # Remove the first name of our inode. Because of the directory fsync bug, the # inode's link count was 1 instead of 5, so removing the 'foo' name ended up # deleting the inode and the other names became stale directory entries (still # visible to applications). Attempting to remove or access the remaining # dentries pointing to that inode resulted in stale file handle errors and # made it impossible to remove the parent directories since it was impossible # for them to become empty. echo "file 'foo' link count after log replay: $(stat -c %h $SCRATCH_MNT/foo)" rm -f $SCRATCH_MNT/foo # Now verify that all files, links and directories created before fsyncing our # directory exist after the fsync log was replayed. [ -f $SCRATCH_MNT/mydir/foo_2 ] || echo "Link mydir/foo_2 is missing" [ -f $SCRATCH_MNT/mydir/foo_3 ] || echo "Link mydir/foo_3 is missing" [ -f $SCRATCH_MNT/hello ] || echo "File hello is missing" [ -f $SCRATCH_MNT/mydir/hello_2 ] || echo "Link mydir/hello_2 is missing" [ -f $SCRATCH_MNT/mydir/x/y/foo_y_link ] || \ echo "Link mydir/x/y/foo_y_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/foo_z_link ] || \ echo "Link mydir/x/y/z/foo_z_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/qwerty ] || \ echo "File mydir/x/y/z/qwerty is missing" # We expect our file here to have a size of 64Kb and all the bytes having the # value 0xff. echo "file 'hello' content after log replay:" od -t x1 $SCRATCH_MNT/hello # Now remove all files/links, under our test directory 'mydir', and verify we # can remove all the directories. rm -f $SCRATCH_MNT/mydir/x/y/z/* rmdir $SCRATCH_MNT/mydir/x/y/z rm -f $SCRATCH_MNT/mydir/x/y/* rmdir $SCRATCH_MNT/mydir/x/y rmdir $SCRATCH_MNT/mydir/x rm -f $SCRATCH_MNT/mydir/* rmdir $SCRATCH_MNT/mydir # An fsck, run by the fstests framework everytime a test finishes, also detected # the inconsistency and printed the following error message: # # root 5 inode 257 errors 2001, no inode item, link count wrong # unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref # unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref status=0 exit The expected golden output for the test is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 5 file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 Which is the output after this patch and when running the test against ext3/4, xfs, f2fs, reiserfs or nilfs2. Without this patch, the test's output is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 1 Link mydir/foo_2 is missing Link mydir/foo_3 is missing Link mydir/x/y/foo_y_link is missing Link mydir/x/y/z/foo_z_link is missing File mydir/x/y/z/qwerty is missing file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y/z': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x': No such file or directory rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_2': Stale file handle rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_3': Stale file handle rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir': Directory not empty Fsck, without this fix, also complains about the wrong link count: root 5 inode 257 errors 2001, no inode item, link count wrong unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref So fix this by logging the inodes that the dentries point to when fsyncing a directory. A test case for xfstests follows. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-03-20 20:19:46 +03:00
const u64 ino_size = btrfs_inode_size(eb, src_item);
Btrfs: fix fsync data loss after adding hard link to inode We have a scenario where after the fsync log replay we can lose file data that had been previously fsync'ed if we added an hard link for our inode and after that we sync'ed the fsync log (for example by fsync'ing some other file or directory). This is because when adding an hard link we updated the inode item in the log tree with an i_size value of 0. At that point the new inode item was in memory only and a subsequent fsync log replay would not make us lose the file data. However if after adding the hard link we sync the log tree to disk, by fsync'ing some other file or directory for example, we ended up losing the file data after log replay, because the inode item in the persisted log tree had an an i_size of zero. This is easy to reproduce, and the following excerpt from my test for xfstests shows this: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create one file with data and fsync it. # This made the btrfs fsync log persist the data and the inode metadata with # a correct inode->i_size (4096 bytes). $XFS_IO_PROG -f -c "pwrite -S 0xaa -b 4K 0 4K" -c "fsync" \ $SCRATCH_MNT/foo | _filter_xfs_io # Now add one hard link to our file. This made the btrfs code update the fsync # log, in memory only, with an inode metadata having a size of 0. ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link # Now force persistence of the fsync log to disk, for example, by fsyncing some # other file. touch $SCRATCH_MNT/bar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/bar # Before a power loss or crash, we could read the 4Kb of data from our file as # expected. echo "File content before:" od -t x1 $SCRATCH_MNT/foo # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # After the fsync log replay, because the fsync log had a value of 0 for our # inode's i_size, we couldn't read anymore the 4Kb of data that we previously # wrote and fsync'ed. The size of the file became 0 after the fsync log replay. echo "File content after:" od -t x1 $SCRATCH_MNT/foo Another alternative test, that doesn't need to fsync an inode in the same transaction it was created, is: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create our test file with some data. $XFS_IO_PROG -f -c "pwrite -S 0xaa -b 8K 0 8K" \ $SCRATCH_MNT/foo | _filter_xfs_io # Make sure the file is durably persisted. sync # Append some data to our file, to increase its size. $XFS_IO_PROG -f -c "pwrite -S 0xcc -b 4K 8K 4K" \ $SCRATCH_MNT/foo | _filter_xfs_io # Fsync the file, so from this point on if a crash/power failure happens, our # new data is guaranteed to be there next time the fs is mounted. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foo # Add one hard link to our file. This made btrfs write into the in memory fsync # log a special inode with generation 0 and an i_size of 0 too. Note that this # didn't update the inode in the fsync log on disk. ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link # Now make sure the in memory fsync log is durably persisted. # Creating and fsync'ing another file will do it. touch $SCRATCH_MNT/bar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/bar # As expected, before the crash/power failure, we should be able to read the # 12Kb of file data. echo "File content before:" od -t x1 $SCRATCH_MNT/foo # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # After mounting the fs again, the fsync log was replayed. # The btrfs fsync log replay code didn't update the i_size of the persisted # inode because the inode item in the log had a special generation with a # value of 0 (and it couldn't know the correct i_size, since that inode item # had a 0 i_size too). This made the last 4Kb of file data inaccessible and # effectively lost. echo "File content after:" od -t x1 $SCRATCH_MNT/foo This isn't a new issue/regression. This problem has been around since the log tree code was added in 2008: Btrfs: Add a write ahead tree log to optimize synchronous operations (commit e02119d5a7b4396c5a872582fddc8bd6d305a70a) Test cases for xfstests follow soon. CC: <stable@vger.kernel.org> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-02-13 15:30:56 +03:00
Btrfs: fix metadata inconsistencies after directory fsync We can get into inconsistency between inodes and directory entries after fsyncing a directory. The issue is that while a directory gets the new dentries persisted in the fsync log and replayed at mount time, the link count of the inode that directory entries point to doesn't get updated, staying with an incorrect link count (smaller then the correct value). This later leads to stale file handle errors when accessing (including attempt to delete) some of the links if all the other ones are removed, which also implies impossibility to delete the parent directories, since the dentries can not be removed. Another issue is that (unlike ext3/4, xfs, f2fs, reiserfs, nilfs2), when fsyncing a directory, new files aren't logged (their metadata and dentries) nor any child directories. So this patch fixes this issue too, since it has the same resolution as the incorrect inode link count issue mentioned before. This is very easy to reproduce, and the following excerpt from my test case for xfstests shows how: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create our main test file and directory. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 8K" $SCRATCH_MNT/foo | _filter_xfs_io mkdir $SCRATCH_MNT/mydir # Make sure all metadata and data are durably persisted. sync # Add a hard link to 'foo' inside our test directory and fsync only the # directory. The btrfs fsync implementation had a bug that caused the new # directory entry to be visible after the fsync log replay but, the inode # of our file remained with a link count of 1. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_2 # Add a few more links and new files. # This is just to verify nothing breaks or gives incorrect results after the # fsync log is replayed. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_3 $XFS_IO_PROG -f -c "pwrite -S 0xff 0 64K" $SCRATCH_MNT/hello | _filter_xfs_io ln $SCRATCH_MNT/hello $SCRATCH_MNT/mydir/hello_2 # Add some subdirectories and new files and links to them. This is to verify # that after fsyncing our top level directory 'mydir', all the subdirectories # and their files/links are registered in the fsync log and exist after the # fsync log is replayed. mkdir -p $SCRATCH_MNT/mydir/x/y/z ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/foo_y_link ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/z/foo_z_link touch $SCRATCH_MNT/mydir/x/y/z/qwerty # Now fsync only our top directory. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/mydir # And fsync now our new file named 'hello', just to verify later that it has # the expected content and that the previous fsync on the directory 'mydir' had # no bad influence on this fsync. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/hello # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # Verify the content of our file 'foo' remains the same as before, 8192 bytes, # all with the value 0xaa. echo "File 'foo' content after log replay:" od -t x1 $SCRATCH_MNT/foo # Remove the first name of our inode. Because of the directory fsync bug, the # inode's link count was 1 instead of 5, so removing the 'foo' name ended up # deleting the inode and the other names became stale directory entries (still # visible to applications). Attempting to remove or access the remaining # dentries pointing to that inode resulted in stale file handle errors and # made it impossible to remove the parent directories since it was impossible # for them to become empty. echo "file 'foo' link count after log replay: $(stat -c %h $SCRATCH_MNT/foo)" rm -f $SCRATCH_MNT/foo # Now verify that all files, links and directories created before fsyncing our # directory exist after the fsync log was replayed. [ -f $SCRATCH_MNT/mydir/foo_2 ] || echo "Link mydir/foo_2 is missing" [ -f $SCRATCH_MNT/mydir/foo_3 ] || echo "Link mydir/foo_3 is missing" [ -f $SCRATCH_MNT/hello ] || echo "File hello is missing" [ -f $SCRATCH_MNT/mydir/hello_2 ] || echo "Link mydir/hello_2 is missing" [ -f $SCRATCH_MNT/mydir/x/y/foo_y_link ] || \ echo "Link mydir/x/y/foo_y_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/foo_z_link ] || \ echo "Link mydir/x/y/z/foo_z_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/qwerty ] || \ echo "File mydir/x/y/z/qwerty is missing" # We expect our file here to have a size of 64Kb and all the bytes having the # value 0xff. echo "file 'hello' content after log replay:" od -t x1 $SCRATCH_MNT/hello # Now remove all files/links, under our test directory 'mydir', and verify we # can remove all the directories. rm -f $SCRATCH_MNT/mydir/x/y/z/* rmdir $SCRATCH_MNT/mydir/x/y/z rm -f $SCRATCH_MNT/mydir/x/y/* rmdir $SCRATCH_MNT/mydir/x/y rmdir $SCRATCH_MNT/mydir/x rm -f $SCRATCH_MNT/mydir/* rmdir $SCRATCH_MNT/mydir # An fsck, run by the fstests framework everytime a test finishes, also detected # the inconsistency and printed the following error message: # # root 5 inode 257 errors 2001, no inode item, link count wrong # unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref # unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref status=0 exit The expected golden output for the test is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 5 file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 Which is the output after this patch and when running the test against ext3/4, xfs, f2fs, reiserfs or nilfs2. Without this patch, the test's output is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 1 Link mydir/foo_2 is missing Link mydir/foo_3 is missing Link mydir/x/y/foo_y_link is missing Link mydir/x/y/z/foo_z_link is missing File mydir/x/y/z/qwerty is missing file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y/z': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x': No such file or directory rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_2': Stale file handle rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_3': Stale file handle rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir': Directory not empty Fsck, without this fix, also complains about the wrong link count: root 5 inode 257 errors 2001, no inode item, link count wrong unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref So fix this by logging the inodes that the dentries point to when fsyncing a directory. A test case for xfstests follows. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-03-20 20:19:46 +03:00
/*
* For regular files an ino_size == 0 is used only when
* logging that an inode exists, as part of a directory
* fsync, and the inode wasn't fsynced before. In this
* case don't set the size of the inode in the fs/subvol
* tree, otherwise we would be throwing valid data away.
*/
Btrfs: fix fsync data loss after adding hard link to inode We have a scenario where after the fsync log replay we can lose file data that had been previously fsync'ed if we added an hard link for our inode and after that we sync'ed the fsync log (for example by fsync'ing some other file or directory). This is because when adding an hard link we updated the inode item in the log tree with an i_size value of 0. At that point the new inode item was in memory only and a subsequent fsync log replay would not make us lose the file data. However if after adding the hard link we sync the log tree to disk, by fsync'ing some other file or directory for example, we ended up losing the file data after log replay, because the inode item in the persisted log tree had an an i_size of zero. This is easy to reproduce, and the following excerpt from my test for xfstests shows this: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create one file with data and fsync it. # This made the btrfs fsync log persist the data and the inode metadata with # a correct inode->i_size (4096 bytes). $XFS_IO_PROG -f -c "pwrite -S 0xaa -b 4K 0 4K" -c "fsync" \ $SCRATCH_MNT/foo | _filter_xfs_io # Now add one hard link to our file. This made the btrfs code update the fsync # log, in memory only, with an inode metadata having a size of 0. ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link # Now force persistence of the fsync log to disk, for example, by fsyncing some # other file. touch $SCRATCH_MNT/bar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/bar # Before a power loss or crash, we could read the 4Kb of data from our file as # expected. echo "File content before:" od -t x1 $SCRATCH_MNT/foo # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # After the fsync log replay, because the fsync log had a value of 0 for our # inode's i_size, we couldn't read anymore the 4Kb of data that we previously # wrote and fsync'ed. The size of the file became 0 after the fsync log replay. echo "File content after:" od -t x1 $SCRATCH_MNT/foo Another alternative test, that doesn't need to fsync an inode in the same transaction it was created, is: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create our test file with some data. $XFS_IO_PROG -f -c "pwrite -S 0xaa -b 8K 0 8K" \ $SCRATCH_MNT/foo | _filter_xfs_io # Make sure the file is durably persisted. sync # Append some data to our file, to increase its size. $XFS_IO_PROG -f -c "pwrite -S 0xcc -b 4K 8K 4K" \ $SCRATCH_MNT/foo | _filter_xfs_io # Fsync the file, so from this point on if a crash/power failure happens, our # new data is guaranteed to be there next time the fs is mounted. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foo # Add one hard link to our file. This made btrfs write into the in memory fsync # log a special inode with generation 0 and an i_size of 0 too. Note that this # didn't update the inode in the fsync log on disk. ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link # Now make sure the in memory fsync log is durably persisted. # Creating and fsync'ing another file will do it. touch $SCRATCH_MNT/bar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/bar # As expected, before the crash/power failure, we should be able to read the # 12Kb of file data. echo "File content before:" od -t x1 $SCRATCH_MNT/foo # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # After mounting the fs again, the fsync log was replayed. # The btrfs fsync log replay code didn't update the i_size of the persisted # inode because the inode item in the log had a special generation with a # value of 0 (and it couldn't know the correct i_size, since that inode item # had a 0 i_size too). This made the last 4Kb of file data inaccessible and # effectively lost. echo "File content after:" od -t x1 $SCRATCH_MNT/foo This isn't a new issue/regression. This problem has been around since the log tree code was added in 2008: Btrfs: Add a write ahead tree log to optimize synchronous operations (commit e02119d5a7b4396c5a872582fddc8bd6d305a70a) Test cases for xfstests follow soon. CC: <stable@vger.kernel.org> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-02-13 15:30:56 +03:00
if (S_ISREG(btrfs_inode_mode(eb, src_item)) &&
Btrfs: fix metadata inconsistencies after directory fsync We can get into inconsistency between inodes and directory entries after fsyncing a directory. The issue is that while a directory gets the new dentries persisted in the fsync log and replayed at mount time, the link count of the inode that directory entries point to doesn't get updated, staying with an incorrect link count (smaller then the correct value). This later leads to stale file handle errors when accessing (including attempt to delete) some of the links if all the other ones are removed, which also implies impossibility to delete the parent directories, since the dentries can not be removed. Another issue is that (unlike ext3/4, xfs, f2fs, reiserfs, nilfs2), when fsyncing a directory, new files aren't logged (their metadata and dentries) nor any child directories. So this patch fixes this issue too, since it has the same resolution as the incorrect inode link count issue mentioned before. This is very easy to reproduce, and the following excerpt from my test case for xfstests shows how: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create our main test file and directory. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 8K" $SCRATCH_MNT/foo | _filter_xfs_io mkdir $SCRATCH_MNT/mydir # Make sure all metadata and data are durably persisted. sync # Add a hard link to 'foo' inside our test directory and fsync only the # directory. The btrfs fsync implementation had a bug that caused the new # directory entry to be visible after the fsync log replay but, the inode # of our file remained with a link count of 1. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_2 # Add a few more links and new files. # This is just to verify nothing breaks or gives incorrect results after the # fsync log is replayed. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_3 $XFS_IO_PROG -f -c "pwrite -S 0xff 0 64K" $SCRATCH_MNT/hello | _filter_xfs_io ln $SCRATCH_MNT/hello $SCRATCH_MNT/mydir/hello_2 # Add some subdirectories and new files and links to them. This is to verify # that after fsyncing our top level directory 'mydir', all the subdirectories # and their files/links are registered in the fsync log and exist after the # fsync log is replayed. mkdir -p $SCRATCH_MNT/mydir/x/y/z ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/foo_y_link ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/z/foo_z_link touch $SCRATCH_MNT/mydir/x/y/z/qwerty # Now fsync only our top directory. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/mydir # And fsync now our new file named 'hello', just to verify later that it has # the expected content and that the previous fsync on the directory 'mydir' had # no bad influence on this fsync. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/hello # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # Verify the content of our file 'foo' remains the same as before, 8192 bytes, # all with the value 0xaa. echo "File 'foo' content after log replay:" od -t x1 $SCRATCH_MNT/foo # Remove the first name of our inode. Because of the directory fsync bug, the # inode's link count was 1 instead of 5, so removing the 'foo' name ended up # deleting the inode and the other names became stale directory entries (still # visible to applications). Attempting to remove or access the remaining # dentries pointing to that inode resulted in stale file handle errors and # made it impossible to remove the parent directories since it was impossible # for them to become empty. echo "file 'foo' link count after log replay: $(stat -c %h $SCRATCH_MNT/foo)" rm -f $SCRATCH_MNT/foo # Now verify that all files, links and directories created before fsyncing our # directory exist after the fsync log was replayed. [ -f $SCRATCH_MNT/mydir/foo_2 ] || echo "Link mydir/foo_2 is missing" [ -f $SCRATCH_MNT/mydir/foo_3 ] || echo "Link mydir/foo_3 is missing" [ -f $SCRATCH_MNT/hello ] || echo "File hello is missing" [ -f $SCRATCH_MNT/mydir/hello_2 ] || echo "Link mydir/hello_2 is missing" [ -f $SCRATCH_MNT/mydir/x/y/foo_y_link ] || \ echo "Link mydir/x/y/foo_y_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/foo_z_link ] || \ echo "Link mydir/x/y/z/foo_z_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/qwerty ] || \ echo "File mydir/x/y/z/qwerty is missing" # We expect our file here to have a size of 64Kb and all the bytes having the # value 0xff. echo "file 'hello' content after log replay:" od -t x1 $SCRATCH_MNT/hello # Now remove all files/links, under our test directory 'mydir', and verify we # can remove all the directories. rm -f $SCRATCH_MNT/mydir/x/y/z/* rmdir $SCRATCH_MNT/mydir/x/y/z rm -f $SCRATCH_MNT/mydir/x/y/* rmdir $SCRATCH_MNT/mydir/x/y rmdir $SCRATCH_MNT/mydir/x rm -f $SCRATCH_MNT/mydir/* rmdir $SCRATCH_MNT/mydir # An fsck, run by the fstests framework everytime a test finishes, also detected # the inconsistency and printed the following error message: # # root 5 inode 257 errors 2001, no inode item, link count wrong # unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref # unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref status=0 exit The expected golden output for the test is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 5 file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 Which is the output after this patch and when running the test against ext3/4, xfs, f2fs, reiserfs or nilfs2. Without this patch, the test's output is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 1 Link mydir/foo_2 is missing Link mydir/foo_3 is missing Link mydir/x/y/foo_y_link is missing Link mydir/x/y/z/foo_z_link is missing File mydir/x/y/z/qwerty is missing file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y/z': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x': No such file or directory rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_2': Stale file handle rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_3': Stale file handle rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir': Directory not empty Fsck, without this fix, also complains about the wrong link count: root 5 inode 257 errors 2001, no inode item, link count wrong unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref So fix this by logging the inodes that the dentries point to when fsyncing a directory. A test case for xfstests follows. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-03-20 20:19:46 +03:00
S_ISREG(btrfs_inode_mode(dst_eb, dst_item)) &&
ino_size != 0)
btrfs_set_inode_size(dst_eb, dst_item, ino_size);
goto no_copy;
Btrfs: fix fsync data loss after adding hard link to inode We have a scenario where after the fsync log replay we can lose file data that had been previously fsync'ed if we added an hard link for our inode and after that we sync'ed the fsync log (for example by fsync'ing some other file or directory). This is because when adding an hard link we updated the inode item in the log tree with an i_size value of 0. At that point the new inode item was in memory only and a subsequent fsync log replay would not make us lose the file data. However if after adding the hard link we sync the log tree to disk, by fsync'ing some other file or directory for example, we ended up losing the file data after log replay, because the inode item in the persisted log tree had an an i_size of zero. This is easy to reproduce, and the following excerpt from my test for xfstests shows this: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create one file with data and fsync it. # This made the btrfs fsync log persist the data and the inode metadata with # a correct inode->i_size (4096 bytes). $XFS_IO_PROG -f -c "pwrite -S 0xaa -b 4K 0 4K" -c "fsync" \ $SCRATCH_MNT/foo | _filter_xfs_io # Now add one hard link to our file. This made the btrfs code update the fsync # log, in memory only, with an inode metadata having a size of 0. ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link # Now force persistence of the fsync log to disk, for example, by fsyncing some # other file. touch $SCRATCH_MNT/bar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/bar # Before a power loss or crash, we could read the 4Kb of data from our file as # expected. echo "File content before:" od -t x1 $SCRATCH_MNT/foo # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # After the fsync log replay, because the fsync log had a value of 0 for our # inode's i_size, we couldn't read anymore the 4Kb of data that we previously # wrote and fsync'ed. The size of the file became 0 after the fsync log replay. echo "File content after:" od -t x1 $SCRATCH_MNT/foo Another alternative test, that doesn't need to fsync an inode in the same transaction it was created, is: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create our test file with some data. $XFS_IO_PROG -f -c "pwrite -S 0xaa -b 8K 0 8K" \ $SCRATCH_MNT/foo | _filter_xfs_io # Make sure the file is durably persisted. sync # Append some data to our file, to increase its size. $XFS_IO_PROG -f -c "pwrite -S 0xcc -b 4K 8K 4K" \ $SCRATCH_MNT/foo | _filter_xfs_io # Fsync the file, so from this point on if a crash/power failure happens, our # new data is guaranteed to be there next time the fs is mounted. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foo # Add one hard link to our file. This made btrfs write into the in memory fsync # log a special inode with generation 0 and an i_size of 0 too. Note that this # didn't update the inode in the fsync log on disk. ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link # Now make sure the in memory fsync log is durably persisted. # Creating and fsync'ing another file will do it. touch $SCRATCH_MNT/bar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/bar # As expected, before the crash/power failure, we should be able to read the # 12Kb of file data. echo "File content before:" od -t x1 $SCRATCH_MNT/foo # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # After mounting the fs again, the fsync log was replayed. # The btrfs fsync log replay code didn't update the i_size of the persisted # inode because the inode item in the log had a special generation with a # value of 0 (and it couldn't know the correct i_size, since that inode item # had a 0 i_size too). This made the last 4Kb of file data inaccessible and # effectively lost. echo "File content after:" od -t x1 $SCRATCH_MNT/foo This isn't a new issue/regression. This problem has been around since the log tree code was added in 2008: Btrfs: Add a write ahead tree log to optimize synchronous operations (commit e02119d5a7b4396c5a872582fddc8bd6d305a70a) Test cases for xfstests follow soon. CC: <stable@vger.kernel.org> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-02-13 15:30:56 +03:00
}
if (overwrite_root &&
S_ISDIR(btrfs_inode_mode(eb, src_item)) &&
S_ISDIR(btrfs_inode_mode(path->nodes[0], dst_item))) {
save_old_i_size = 1;
saved_i_size = btrfs_inode_size(path->nodes[0],
dst_item);
}
}
copy_extent_buffer(path->nodes[0], eb, dst_ptr,
src_ptr, item_size);
if (save_old_i_size) {
struct btrfs_inode_item *dst_item;
dst_item = (struct btrfs_inode_item *)dst_ptr;
btrfs_set_inode_size(path->nodes[0], dst_item, saved_i_size);
}
/* make sure the generation is filled in */
if (key->type == BTRFS_INODE_ITEM_KEY) {
struct btrfs_inode_item *dst_item;
dst_item = (struct btrfs_inode_item *)dst_ptr;
if (btrfs_inode_generation(path->nodes[0], dst_item) == 0) {
btrfs_set_inode_generation(path->nodes[0], dst_item,
trans->transid);
}
}
no_copy:
btrfs_mark_buffer_dirty(path->nodes[0]);
btrfs_release_path(path);
return 0;
}
btrfs: insert items in batches when logging a directory when possible When logging a directory, we scan its directory items from the subvolume tree and then copy one by one into the log tree. This is not efficient since we generally are able to insert several items in a batch, using a single btree operation for adding several items at once. The reason we copy items one by one is that we must check if each item was previously logged in the current transaction, and if it was we either overwrite it or skip it in case its content did not change in the subvolume tree (this can happen only for dir item keys, but not for dir index keys), and doing such check makes it a bit cumbersome to attempt batch insertions. However the chances for doing batch insertions are very frequent and always happen when: 1) Logging the directory for the first time in the current transaction, as none of the items exist in the log tree yet; 2) Logging new dir index keys, because the offset for new dir index keys comes from a monotonically increasing counter. This means if we keep adding dentries to a directory, through creation of new files and sub-directories or by adding new links or renaming from some other directory into the one we are logging, all the new dir index keys have a new offset that is greater than the offset of any previously logged index keys, so we can insert them in batches into the log tree. For dir item keys, since their offset depends on the result of an hash function against the dentry's name, unless the directory is being logged for the first time in the current transaction, the chances being able to insert the items in the log using batches is pretty much random and not predictable, as it depends on the names of the dentries, but still happens often enough. So change directory logging to keep track of consecutive directory items that don't exist yet in the log and batch insert them. This patch is part of a patchset comprised of the following 5 patches: btrfs: remove root argument from btrfs_log_inode() and its callees btrfs: remove redundant log root assignment from log_dir_items() btrfs: factor out the copying loop of dir items from log_dir_items() btrfs: insert items in batches when logging a directory when possible btrfs: keep track of the last logged keys when logging a directory This is patch 4/5. The change log of the last patch (5/5) has performance results. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-16 13:32:13 +03:00
/*
* Item overwrite used by replay and tree logging. eb, slot and key all refer
* to the src data we are copying out.
*
* root is the tree we are copying into, and path is a scratch
* path for use in this function (it should be released on entry and
* will be released on exit).
*
* If the key is already in the destination tree the existing item is
* overwritten. If the existing item isn't big enough, it is extended.
* If it is too large, it is truncated.
*
* If the key isn't in the destination yet, a new item is inserted.
*/
static int overwrite_item(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_path *path,
struct extent_buffer *eb, int slot,
struct btrfs_key *key)
{
int ret;
/* Look for the key in the destination tree. */
ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
if (ret < 0)
return ret;
return do_overwrite_item(trans, root, path, eb, slot, key);
}
/*
* simple helper to read an inode off the disk from a given root
* This can only be called for subvolume roots and not for the log
*/
static noinline struct inode *read_one_inode(struct btrfs_root *root,
u64 objectid)
{
struct inode *inode;
inode = btrfs_iget(root->fs_info->sb, objectid, root);
if (IS_ERR(inode))
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 18:45:14 +04:00
inode = NULL;
return inode;
}
/* replays a single extent in 'eb' at 'slot' with 'key' into the
* subvolume 'root'. path is released on entry and should be released
* on exit.
*
* extents in the log tree have not been allocated out of the extent
* tree yet. So, this completes the allocation, taking a reference
* as required if the extent already exists or creating a new extent
* if it isn't in the extent allocation tree yet.
*
* The extent is inserted into the file, dropping any existing extents
* from the file that overlap the new one.
*/
static noinline int replay_one_extent(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_path *path,
struct extent_buffer *eb, int slot,
struct btrfs_key *key)
{
struct btrfs_drop_extents_args drop_args = { 0 };
struct btrfs_fs_info *fs_info = root->fs_info;
int found_type;
u64 extent_end;
u64 start = key->offset;
u64 nbytes = 0;
struct btrfs_file_extent_item *item;
struct inode *inode = NULL;
unsigned long size;
int ret = 0;
item = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
found_type = btrfs_file_extent_type(eb, item);
if (found_type == BTRFS_FILE_EXTENT_REG ||
found_type == BTRFS_FILE_EXTENT_PREALLOC) {
nbytes = btrfs_file_extent_num_bytes(eb, item);
extent_end = start + nbytes;
/*
* We don't add to the inodes nbytes if we are prealloc or a
* hole.
*/
if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
nbytes = 0;
} else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
size = btrfs_file_extent_ram_bytes(eb, item);
nbytes = btrfs_file_extent_ram_bytes(eb, item);
extent_end = ALIGN(start + size,
fs_info->sectorsize);
} else {
ret = 0;
goto out;
}
inode = read_one_inode(root, key->objectid);
if (!inode) {
ret = -EIO;
goto out;
}
/*
* first check to see if we already have this extent in the
* file. This must be done before the btrfs_drop_extents run
* so we don't try to drop this extent.
*/
ret = btrfs_lookup_file_extent(trans, root, path,
btrfs_ino(BTRFS_I(inode)), start, 0);
if (ret == 0 &&
(found_type == BTRFS_FILE_EXTENT_REG ||
found_type == BTRFS_FILE_EXTENT_PREALLOC)) {
struct btrfs_file_extent_item cmp1;
struct btrfs_file_extent_item cmp2;
struct btrfs_file_extent_item *existing;
struct extent_buffer *leaf;
leaf = path->nodes[0];
existing = btrfs_item_ptr(leaf, path->slots[0],
struct btrfs_file_extent_item);
read_extent_buffer(eb, &cmp1, (unsigned long)item,
sizeof(cmp1));
read_extent_buffer(leaf, &cmp2, (unsigned long)existing,
sizeof(cmp2));
/*
* we already have a pointer to this exact extent,
* we don't have to do anything
*/
if (memcmp(&cmp1, &cmp2, sizeof(cmp1)) == 0) {
btrfs_release_path(path);
goto out;
}
}
btrfs_release_path(path);
/* drop any overlapping extents */
drop_args.start = start;
drop_args.end = extent_end;
drop_args.drop_cache = true;
ret = btrfs_drop_extents(trans, root, BTRFS_I(inode), &drop_args);
if (ret)
goto out;
if (found_type == BTRFS_FILE_EXTENT_REG ||
found_type == BTRFS_FILE_EXTENT_PREALLOC) {
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 18:45:14 +04:00
u64 offset;
unsigned long dest_offset;
struct btrfs_key ins;
if (btrfs_file_extent_disk_bytenr(eb, item) == 0 &&
btrfs_fs_incompat(fs_info, NO_HOLES))
goto update_inode;
ret = btrfs_insert_empty_item(trans, root, path, key,
sizeof(*item));
if (ret)
goto out;
dest_offset = btrfs_item_ptr_offset(path->nodes[0],
path->slots[0]);
copy_extent_buffer(path->nodes[0], eb, dest_offset,
(unsigned long)item, sizeof(*item));
ins.objectid = btrfs_file_extent_disk_bytenr(eb, item);
ins.offset = btrfs_file_extent_disk_num_bytes(eb, item);
ins.type = BTRFS_EXTENT_ITEM_KEY;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 18:45:14 +04:00
offset = key->offset - btrfs_file_extent_offset(eb, item);
/*
* Manually record dirty extent, as here we did a shallow
* file extent item copy and skip normal backref update,
* but modifying extent tree all by ourselves.
* So need to manually record dirty extent for qgroup,
* as the owner of the file extent changed from log tree
* (doesn't affect qgroup) to fs/file tree(affects qgroup)
*/
ret = btrfs_qgroup_trace_extent(trans,
btrfs_file_extent_disk_bytenr(eb, item),
btrfs_file_extent_disk_num_bytes(eb, item),
GFP_NOFS);
if (ret < 0)
goto out;
if (ins.objectid > 0) {
struct btrfs_ref ref = { 0 };
u64 csum_start;
u64 csum_end;
LIST_HEAD(ordered_sums);
/*
* is this extent already allocated in the extent
* allocation tree? If so, just add a reference
*/
ret = btrfs_lookup_data_extent(fs_info, ins.objectid,
ins.offset);
if (ret < 0) {
goto out;
} else if (ret == 0) {
btrfs_init_generic_ref(&ref,
BTRFS_ADD_DELAYED_REF,
ins.objectid, ins.offset, 0);
btrfs_init_data_ref(&ref,
root->root_key.objectid,
key->objectid, offset, 0, false);
ret = btrfs_inc_extent_ref(trans, &ref);
if (ret)
goto out;
} else {
/*
* insert the extent pointer in the extent
* allocation tree
*/
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 18:45:14 +04:00
ret = btrfs_alloc_logged_file_extent(trans,
root->root_key.objectid,
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 18:45:14 +04:00
key->objectid, offset, &ins);
if (ret)
goto out;
}
btrfs_release_path(path);
if (btrfs_file_extent_compression(eb, item)) {
csum_start = ins.objectid;
csum_end = csum_start + ins.offset;
} else {
csum_start = ins.objectid +
btrfs_file_extent_offset(eb, item);
csum_end = csum_start +
btrfs_file_extent_num_bytes(eb, item);
}
ret = btrfs_lookup_csums_range(root->log_root,
csum_start, csum_end - 1,
&ordered_sums, 0);
if (ret)
goto out;
Btrfs: fix file read corruption after extent cloning and fsync If we partially clone one extent of a file into a lower offset of the file, fsync the file, power fail and then mount the fs to trigger log replay, we can get multiple checksum items in the csum tree that overlap each other and result in checksum lookup failures later. Those failures can make file data read requests assume a checksum value of 0, but they will not return an error (-EIO for example) to userspace exactly because the expected checksum value 0 is a special value that makes the read bio endio callback return success and set all the bytes of the corresponding page with the value 0x01 (at fs/btrfs/inode.c:__readpage_endio_check()). From a userspace perspective this is equivalent to file corruption because we are not returning what was written to the file. Details about how this can happen, and why, are included inline in the following reproducer test case for fstests and the comment added to tree-log.c. seq=`basename $0` seqres=$RESULT_DIR/$seq echo "QA output created by $seq" tmp=/tmp/$$ status=1 # failure is the default! trap "_cleanup; exit \$status" 0 1 2 3 15 _cleanup() { _cleanup_flakey rm -f $tmp.* } # get standard environment, filters and checks . ./common/rc . ./common/filter . ./common/dmflakey # real QA test starts here _need_to_be_root _supported_fs btrfs _supported_os Linux _require_scratch _require_dm_flakey _require_cloner _require_metadata_journaling $SCRATCH_DEV rm -f $seqres.full _scratch_mkfs >>$seqres.full 2>&1 _init_flakey _mount_flakey # Create our test file with a single 100K extent starting at file # offset 800K. We fsync the file here to make the fsync log tree gets # a single csum item that covers the whole 100K extent, which causes # the second fsync, done after the cloning operation below, to not # leave in the log tree two csum items covering two sub-ranges # ([0, 20K[ and [20K, 100K[)) of our extent. $XFS_IO_PROG -f -c "pwrite -S 0xaa 800K 100K" \ -c "fsync" \ $SCRATCH_MNT/foo | _filter_xfs_io # Now clone part of our extent into file offset 400K. This adds a file # extent item to our inode's metadata that points to the 100K extent # we created before, using a data offset of 20K and a data length of # 20K, so that it refers to the sub-range [20K, 40K[ of our original # extent. $CLONER_PROG -s $((800 * 1024 + 20 * 1024)) -d $((400 * 1024)) \ -l $((20 * 1024)) $SCRATCH_MNT/foo $SCRATCH_MNT/foo # Now fsync our file to make sure the extent cloning is durably # persisted. This fsync will not add a second csum item to the log # tree containing the checksums for the blocks in the sub-range # [20K, 40K[ of our extent, because there was already a csum item in # the log tree covering the whole extent, added by the first fsync # we did before. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foo echo "File digest before power failure:" md5sum $SCRATCH_MNT/foo | _filter_scratch # Silently drop all writes and ummount to simulate a crash/power # failure. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again, mount to trigger log replay and validate file # contents. # The fsync log replay first processes the file extent item # corresponding to the file offset 400K (the one which refers to the # [20K, 40K[ sub-range of our 100K extent) and then processes the file # extent item for file offset 800K. It used to happen that when # processing the later, it erroneously left in the csum tree 2 csum # items that overlapped each other, 1 for the sub-range [20K, 40K[ and # 1 for the whole range of our extent. This introduced a problem where # subsequent lookups for the checksums of blocks within the range # [40K, 100K[ of our extent would not find anything because lookups in # the csum tree ended up looking only at the smaller csum item, the # one covering the subrange [20K, 40K[. This made read requests assume # an expected checksum with a value of 0 for those blocks, which caused # checksum verification failure when the read operations finished. # However those checksum failure did not result in read requests # returning an error to user space (like -EIO for e.g.) because the # expected checksum value had the special value 0, and in that case # btrfs set all bytes of the corresponding pages with the value 0x01 # and produce the following warning in dmesg/syslog: # # "BTRFS warning (device dm-0): csum failed ino 257 off 917504 csum\ # 1322675045 expected csum 0" # _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey echo "File digest after log replay:" # Must match the same digest he had after cloning the extent and # before the power failure happened. md5sum $SCRATCH_MNT/foo | _filter_scratch _unmount_flakey status=0 exit Signed-off-by: Filipe Manana <fdmanana@suse.com> Reviewed-by: Liu Bo <bo.li.liu@oracle.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-08-19 13:09:40 +03:00
/*
* Now delete all existing cums in the csum root that
* cover our range. We do this because we can have an
* extent that is completely referenced by one file
* extent item and partially referenced by another
* file extent item (like after using the clone or
* extent_same ioctls). In this case if we end up doing
* the replay of the one that partially references the
* extent first, and we do not do the csum deletion
* below, we can get 2 csum items in the csum tree that
* overlap each other. For example, imagine our log has
* the two following file extent items:
*
* key (257 EXTENT_DATA 409600)
* extent data disk byte 12845056 nr 102400
* extent data offset 20480 nr 20480 ram 102400
*
* key (257 EXTENT_DATA 819200)
* extent data disk byte 12845056 nr 102400
* extent data offset 0 nr 102400 ram 102400
*
* Where the second one fully references the 100K extent
* that starts at disk byte 12845056, and the log tree
* has a single csum item that covers the entire range
* of the extent:
*
* key (EXTENT_CSUM EXTENT_CSUM 12845056) itemsize 100
*
* After the first file extent item is replayed, the
* csum tree gets the following csum item:
*
* key (EXTENT_CSUM EXTENT_CSUM 12865536) itemsize 20
*
* Which covers the 20K sub-range starting at offset 20K
* of our extent. Now when we replay the second file
* extent item, if we do not delete existing csum items
* that cover any of its blocks, we end up getting two
* csum items in our csum tree that overlap each other:
*
* key (EXTENT_CSUM EXTENT_CSUM 12845056) itemsize 100
* key (EXTENT_CSUM EXTENT_CSUM 12865536) itemsize 20
*
* Which is a problem, because after this anyone trying
* to lookup up for the checksum of any block of our
* extent starting at an offset of 40K or higher, will
* end up looking at the second csum item only, which
* does not contain the checksum for any block starting
* at offset 40K or higher of our extent.
*/
while (!list_empty(&ordered_sums)) {
struct btrfs_ordered_sum *sums;
struct btrfs_root *csum_root;
sums = list_entry(ordered_sums.next,
struct btrfs_ordered_sum,
list);
csum_root = btrfs_csum_root(fs_info,
sums->bytenr);
Btrfs: fix file read corruption after extent cloning and fsync If we partially clone one extent of a file into a lower offset of the file, fsync the file, power fail and then mount the fs to trigger log replay, we can get multiple checksum items in the csum tree that overlap each other and result in checksum lookup failures later. Those failures can make file data read requests assume a checksum value of 0, but they will not return an error (-EIO for example) to userspace exactly because the expected checksum value 0 is a special value that makes the read bio endio callback return success and set all the bytes of the corresponding page with the value 0x01 (at fs/btrfs/inode.c:__readpage_endio_check()). From a userspace perspective this is equivalent to file corruption because we are not returning what was written to the file. Details about how this can happen, and why, are included inline in the following reproducer test case for fstests and the comment added to tree-log.c. seq=`basename $0` seqres=$RESULT_DIR/$seq echo "QA output created by $seq" tmp=/tmp/$$ status=1 # failure is the default! trap "_cleanup; exit \$status" 0 1 2 3 15 _cleanup() { _cleanup_flakey rm -f $tmp.* } # get standard environment, filters and checks . ./common/rc . ./common/filter . ./common/dmflakey # real QA test starts here _need_to_be_root _supported_fs btrfs _supported_os Linux _require_scratch _require_dm_flakey _require_cloner _require_metadata_journaling $SCRATCH_DEV rm -f $seqres.full _scratch_mkfs >>$seqres.full 2>&1 _init_flakey _mount_flakey # Create our test file with a single 100K extent starting at file # offset 800K. We fsync the file here to make the fsync log tree gets # a single csum item that covers the whole 100K extent, which causes # the second fsync, done after the cloning operation below, to not # leave in the log tree two csum items covering two sub-ranges # ([0, 20K[ and [20K, 100K[)) of our extent. $XFS_IO_PROG -f -c "pwrite -S 0xaa 800K 100K" \ -c "fsync" \ $SCRATCH_MNT/foo | _filter_xfs_io # Now clone part of our extent into file offset 400K. This adds a file # extent item to our inode's metadata that points to the 100K extent # we created before, using a data offset of 20K and a data length of # 20K, so that it refers to the sub-range [20K, 40K[ of our original # extent. $CLONER_PROG -s $((800 * 1024 + 20 * 1024)) -d $((400 * 1024)) \ -l $((20 * 1024)) $SCRATCH_MNT/foo $SCRATCH_MNT/foo # Now fsync our file to make sure the extent cloning is durably # persisted. This fsync will not add a second csum item to the log # tree containing the checksums for the blocks in the sub-range # [20K, 40K[ of our extent, because there was already a csum item in # the log tree covering the whole extent, added by the first fsync # we did before. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foo echo "File digest before power failure:" md5sum $SCRATCH_MNT/foo | _filter_scratch # Silently drop all writes and ummount to simulate a crash/power # failure. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again, mount to trigger log replay and validate file # contents. # The fsync log replay first processes the file extent item # corresponding to the file offset 400K (the one which refers to the # [20K, 40K[ sub-range of our 100K extent) and then processes the file # extent item for file offset 800K. It used to happen that when # processing the later, it erroneously left in the csum tree 2 csum # items that overlapped each other, 1 for the sub-range [20K, 40K[ and # 1 for the whole range of our extent. This introduced a problem where # subsequent lookups for the checksums of blocks within the range # [40K, 100K[ of our extent would not find anything because lookups in # the csum tree ended up looking only at the smaller csum item, the # one covering the subrange [20K, 40K[. This made read requests assume # an expected checksum with a value of 0 for those blocks, which caused # checksum verification failure when the read operations finished. # However those checksum failure did not result in read requests # returning an error to user space (like -EIO for e.g.) because the # expected checksum value had the special value 0, and in that case # btrfs set all bytes of the corresponding pages with the value 0x01 # and produce the following warning in dmesg/syslog: # # "BTRFS warning (device dm-0): csum failed ino 257 off 917504 csum\ # 1322675045 expected csum 0" # _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey echo "File digest after log replay:" # Must match the same digest he had after cloning the extent and # before the power failure happened. md5sum $SCRATCH_MNT/foo | _filter_scratch _unmount_flakey status=0 exit Signed-off-by: Filipe Manana <fdmanana@suse.com> Reviewed-by: Liu Bo <bo.li.liu@oracle.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-08-19 13:09:40 +03:00
if (!ret)
ret = btrfs_del_csums(trans, csum_root,
sums->bytenr,
sums->len);
if (!ret)
ret = btrfs_csum_file_blocks(trans,
csum_root,
sums);
list_del(&sums->list);
kfree(sums);
}
if (ret)
goto out;
} else {
btrfs_release_path(path);
}
} else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
/* inline extents are easy, we just overwrite them */
ret = overwrite_item(trans, root, path, eb, slot, key);
if (ret)
goto out;
}
btrfs: use the file extent tree infrastructure We want to use this everywhere we modify the file extent items permanently. These include: 1) Inserting new file extents for writes and prealloc extents. 2) Truncating inode items. 3) btrfs_cont_expand(). 4) Insert inline extents. 5) Insert new extents from log replay. 6) Insert a new extent for clone, as it could be past i_size. 7) Hole punching For hole punching in particular it might seem it's not necessary because anybody extending would use btrfs_cont_expand, however there is a corner that still can give us trouble. Start with an empty file and fallocate KEEP_SIZE 1M-2M We now have a 0 length file, and a hole file extent from 0-1M, and a prealloc extent from 1M-2M. Now punch 1M-1.5M Because this is past i_size we have [HOLE EXTENT][ NOTHING ][PREALLOC] [0 1M][1M 1.5M][1.5M 2M] with an i_size of 0. Now if we pwrite 0-1.5M we'll increas our i_size to 1.5M, but our disk_i_size is still 0 until the ordered extent completes. However if we now immediately truncate 2M on the file we'll just call btrfs_cont_expand(inode, 1.5M, 2M), since our old i_size is 1.5M. If we commit the transaction here and crash we'll expose the gap. To fix this we need to clear the file extent mapping for the range that we punched but didn't insert a corresponding file extent for. This will mean the truncate will only get an disk_i_size set to 1M if we crash before the finish ordered io happens. I've written an xfstest to reproduce the problem and validate this fix. Reviewed-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-01-17 17:02:22 +03:00
ret = btrfs_inode_set_file_extent_range(BTRFS_I(inode), start,
extent_end - start);
if (ret)
goto out;
update_inode:
btrfs: update the number of bytes used by an inode atomically There are several occasions where we do not update the inode's number of used bytes atomically, resulting in a concurrent stat(2) syscall to report a value of used blocks that does not correspond to a valid value, that is, a value that does not match neither what we had before the operation nor what we get after the operation completes. In extreme cases it can result in stat(2) reporting zero used blocks, which can cause problems for some userspace tools where they can consider a file with a non-zero size and zero used blocks as completely sparse and skip reading data, as reported/discussed a long time ago in some threads like the following: https://lists.gnu.org/archive/html/bug-tar/2016-07/msg00001.html The cases where this can happen are the following: -> Case 1 If we do a write (buffered or direct IO) against a file region for which there is already an allocated extent (or multiple extents), then we have a short time window where we can report a number of used blocks to stat(2) that does not take into account the file region being overwritten. This short time window happens when completing the ordered extent(s). This happens because when we drop the extents in the write range we decrement the inode's number of bytes and later on when we insert the new extent(s) we increment the number of bytes in the inode, resulting in a short time window where a stat(2) syscall can get an incorrect number of used blocks. If we do writes that overwrite an entire file, then we have a short time window where we report 0 used blocks to stat(2). Example reproducer: $ cat reproducer-1.sh #!/bin/bash MNT=/mnt/sdi DEV=/dev/sdi stat_loop() { trap "wait; exit" SIGTERM local filepath=$1 local expected=$2 local got while :; do got=$(stat -c %b $filepath) if [ $got -ne $expected ]; then echo -n "ERROR: unexpected used blocks" echo " (got: $got expected: $expected)" fi done } mkfs.btrfs -f $DEV > /dev/null # mkfs.xfs -f $DEV > /dev/null # mkfs.ext4 -F $DEV > /dev/null # mkfs.f2fs -f $DEV > /dev/null # mkfs.reiserfs -f $DEV > /dev/null mount $DEV $MNT xfs_io -f -s -c "pwrite -b 64K 0 64K" $MNT/foobar >/dev/null expected=$(stat -c %b $MNT/foobar) # Create a process to keep calling stat(2) on the file and see if the # reported number of blocks used (disk space used) changes, it should # not because we are not increasing the file size nor punching holes. stat_loop $MNT/foobar $expected & loop_pid=$! for ((i = 0; i < 50000; i++)); do xfs_io -s -c "pwrite -b 64K 0 64K" $MNT/foobar >/dev/null done kill $loop_pid &> /dev/null wait umount $DEV $ ./reproducer-1.sh ERROR: unexpected used blocks (got: 0 expected: 128) ERROR: unexpected used blocks (got: 0 expected: 128) (...) Note that since this is a short time window where the race can happen, the reproducer may not be able to always trigger the bug in one run, or it may trigger it multiple times. -> Case 2 If we do a buffered write against a file region that does not have any allocated extents, like a hole or beyond EOF, then during ordered extent completion we have a short time window where a concurrent stat(2) syscall can report a number of used blocks that does not correspond to the value before or after the write operation, a value that is actually larger than the value after the write completes. This happens because once we start a buffered write into an unallocated file range we increment the inode's 'new_delalloc_bytes', to make sure any stat(2) call gets a correct used blocks value before delalloc is flushed and completes. However at ordered extent completion, after we inserted the new extent, we increment the inode's number of bytes used with the size of the new extent, and only later, when clearing the range in the inode's iotree, we decrement the inode's 'new_delalloc_bytes' counter with the size of the extent. So this results in a short time window where a concurrent stat(2) syscall can report a number of used blocks that accounts for the new extent twice. Example reproducer: $ cat reproducer-2.sh #!/bin/bash MNT=/mnt/sdi DEV=/dev/sdi stat_loop() { trap "wait; exit" SIGTERM local filepath=$1 local expected=$2 local got while :; do got=$(stat -c %b $filepath) if [ $got -ne $expected ]; then echo -n "ERROR: unexpected used blocks" echo " (got: $got expected: $expected)" fi done } mkfs.btrfs -f $DEV > /dev/null # mkfs.xfs -f $DEV > /dev/null # mkfs.ext4 -F $DEV > /dev/null # mkfs.f2fs -f $DEV > /dev/null # mkfs.reiserfs -f $DEV > /dev/null mount $DEV $MNT touch $MNT/foobar write_size=$((64 * 1024)) for ((i = 0; i < 16384; i++)); do offset=$(($i * $write_size)) xfs_io -c "pwrite -S 0xab $offset $write_size" $MNT/foobar >/dev/null blocks_used=$(stat -c %b $MNT/foobar) # Fsync the file to trigger writeback and keep calling stat(2) on it # to see if the number of blocks used changes. stat_loop $MNT/foobar $blocks_used & loop_pid=$! xfs_io -c "fsync" $MNT/foobar kill $loop_pid &> /dev/null wait $loop_pid done umount $DEV $ ./reproducer-2.sh ERROR: unexpected used blocks (got: 265472 expected: 265344) ERROR: unexpected used blocks (got: 284032 expected: 283904) (...) Note that since this is a short time window where the race can happen, the reproducer may not be able to always trigger the bug in one run, or it may trigger it multiple times. -> Case 3 Another case where such problems happen is during other operations that replace extents in a file range with other extents. Those operations are extent cloning, deduplication and fallocate's zero range operation. The cause of the problem is similar to the first case. When we drop the extents from a range, we decrement the inode's number of bytes, and later on, after inserting the new extents we increment it. Since this is not done atomically, a concurrent stat(2) call can see and return a number of used blocks that is smaller than it should be, does not match the number of used blocks before or after the clone/deduplication/zero operation. Like for the first case, when doing a clone, deduplication or zero range operation against an entire file, we end up having a time window where we can report 0 used blocks to a stat(2) call. Example reproducer: $ cat reproducer-3.sh #!/bin/bash MNT=/mnt/sdi DEV=/dev/sdi mkfs.btrfs -f $DEV > /dev/null # mkfs.xfs -f -m reflink=1 $DEV > /dev/null mount $DEV $MNT extent_size=$((64 * 1024)) num_extents=16384 file_size=$(($extent_size * $num_extents)) # File foo has many small extents. xfs_io -f -s -c "pwrite -S 0xab -b $extent_size 0 $file_size" $MNT/foo \ > /dev/null # File bar has much less extents and has exactly the same data as foo. xfs_io -f -c "pwrite -S 0xab 0 $file_size" $MNT/bar > /dev/null expected=$(stat -c %b $MNT/foo) # Now deduplicate bar into foo. While the deduplication is in progres, # the number of used blocks/file size reported by stat should not change xfs_io -c "dedupe $MNT/bar 0 0 $file_size" $MNT/foo > /dev/null & dedupe_pid=$! while [ -n "$(ps -p $dedupe_pid -o pid=)" ]; do used=$(stat -c %b $MNT/foo) if [ $used -ne $expected ]; then echo "Unexpected blocks used: $used (expected: $expected)" fi done umount $DEV $ ./reproducer-3.sh Unexpected blocks used: 2076800 (expected: 2097152) Unexpected blocks used: 2097024 (expected: 2097152) Unexpected blocks used: 2079872 (expected: 2097152) (...) Note that since this is a short time window where the race can happen, the reproducer may not be able to always trigger the bug in one run, or it may trigger it multiple times. So fix this by: 1) Making btrfs_drop_extents() not decrement the VFS inode's number of bytes, and instead return the number of bytes; 2) Making any code that drops extents and adds new extents update the inode's number of bytes atomically, while holding the btrfs inode's spinlock, which is also used by the stat(2) callback to get the inode's number of bytes; 3) For ranges in the inode's iotree that are marked as 'delalloc new', corresponding to previously unallocated ranges, increment the inode's number of bytes when clearing the 'delalloc new' bit from the range, in the same critical section that decrements the inode's 'new_delalloc_bytes' counter, delimited by the btrfs inode's spinlock. An alternative would be to have btrfs_getattr() wait for any IO (ordered extents in progress) and locking the whole range (0 to (u64)-1) while it it computes the number of blocks used. But that would mean blocking stat(2), which is a very used syscall and expected to be fast, waiting for writes, clone/dedupe, fallocate, page reads, fiemap, etc. CC: stable@vger.kernel.org # 5.4+ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-11-04 14:07:34 +03:00
btrfs_update_inode_bytes(BTRFS_I(inode), nbytes, drop_args.bytes_found);
ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
out:
if (inode)
iput(inode);
return ret;
}
/*
* when cleaning up conflicts between the directory names in the
* subvolume, directory names in the log and directory names in the
* inode back references, we may have to unlink inodes from directories.
*
* This is a helper function to do the unlink of a specific directory
* item
*/
static noinline int drop_one_dir_item(struct btrfs_trans_handle *trans,
struct btrfs_path *path,
struct btrfs_inode *dir,
struct btrfs_dir_item *di)
{
struct btrfs_root *root = dir->root;
struct inode *inode;
char *name;
int name_len;
struct extent_buffer *leaf;
struct btrfs_key location;
int ret;
leaf = path->nodes[0];
btrfs_dir_item_key_to_cpu(leaf, di, &location);
name_len = btrfs_dir_name_len(leaf, di);
name = kmalloc(name_len, GFP_NOFS);
if (!name)
return -ENOMEM;
read_extent_buffer(leaf, name, (unsigned long)(di + 1), name_len);
btrfs_release_path(path);
inode = read_one_inode(root, location.objectid);
if (!inode) {
ret = -EIO;
goto out;
}
ret = link_to_fixup_dir(trans, root, path, location.objectid);
if (ret)
goto out;
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 17:24:20 +03:00
ret = btrfs_unlink_inode(trans, dir, BTRFS_I(inode), name,
name_len);
if (ret)
goto out;
else
ret = btrfs_run_delayed_items(trans);
out:
kfree(name);
iput(inode);
return ret;
}
/*
* See if a given name and sequence number found in an inode back reference are
* already in a directory and correctly point to this inode.
*
* Returns: < 0 on error, 0 if the directory entry does not exists and 1 if it
* exists.
*/
static noinline int inode_in_dir(struct btrfs_root *root,
struct btrfs_path *path,
u64 dirid, u64 objectid, u64 index,
const char *name, int name_len)
{
struct btrfs_dir_item *di;
struct btrfs_key location;
int ret = 0;
di = btrfs_lookup_dir_index_item(NULL, root, path, dirid,
index, name, name_len, 0);
if (IS_ERR(di)) {
ret = PTR_ERR(di);
goto out;
} else if (di) {
btrfs_dir_item_key_to_cpu(path->nodes[0], di, &location);
if (location.objectid != objectid)
goto out;
} else {
goto out;
}
btrfs_release_path(path);
di = btrfs_lookup_dir_item(NULL, root, path, dirid, name, name_len, 0);
if (IS_ERR(di)) {
ret = PTR_ERR(di);
goto out;
} else if (di) {
btrfs_dir_item_key_to_cpu(path->nodes[0], di, &location);
if (location.objectid == objectid)
ret = 1;
}
out:
btrfs_release_path(path);
return ret;
}
/*
* helper function to check a log tree for a named back reference in
* an inode. This is used to decide if a back reference that is
* found in the subvolume conflicts with what we find in the log.
*
* inode backreferences may have multiple refs in a single item,
* during replay we process one reference at a time, and we don't
* want to delete valid links to a file from the subvolume if that
* link is also in the log.
*/
static noinline int backref_in_log(struct btrfs_root *log,
struct btrfs_key *key,
u64 ref_objectid,
Btrfs: fix fsync log replay for inodes with a mix of regular refs and extrefs If we have an inode with a large number of hard links, some of which may be extrefs, turn a regular ref into an extref, fsync the inode and then replay the fsync log (after a crash/reboot), we can endup with an fsync log that makes the replay code always fail with -EOVERFLOW when processing the inode's references. This is easy to reproduce with the test case I made for xfstests. Its steps are the following: _scratch_mkfs "-O extref" >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create a test file with 3001 hard links. This number is large enough to # make btrfs start using extrefs at some point even if the fs has the maximum # possible leaf/node size (64Kb). echo "hello world" > $SCRATCH_MNT/foo for i in `seq 1 3000`; do ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link_`printf "%04d" $i` done # Make sure all metadata and data are durably persisted. sync # Now remove one link, add a new one with a new name, add another new one with # the same name as the one we just removed and fsync the inode. rm -f $SCRATCH_MNT/foo_link_0001 ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link_3001 ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link_0001 rm -f $SCRATCH_MNT/foo_link_0002 ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link_3002 ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link_3003 $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foo # Simulate a crash/power loss. This makes sure the next mount # will see an fsync log and will replay that log. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # Check that the number of hard links is correct, we are able to remove all # the hard links and read the file's data. This is just to verify we don't # get stale file handle errors (due to dangling directory index entries that # point to inodes that no longer exist). echo "Link count: $(stat --format=%h $SCRATCH_MNT/foo)" [ -f $SCRATCH_MNT/foo ] || echo "Link foo is missing" for ((i = 1; i <= 3003; i++)); do name=foo_link_`printf "%04d" $i` if [ $i -eq 2 ]; then [ -f $SCRATCH_MNT/$name ] && echo "Link $name found" else [ -f $SCRATCH_MNT/$name ] || echo "Link $name is missing" fi done rm -f $SCRATCH_MNT/foo_link_* cat $SCRATCH_MNT/foo rm -f $SCRATCH_MNT/foo status=0 exit The fix is simply to correct the overflow condition when overwriting a reference item because it was wrong, trying to increase the item in the fs/subvol tree by an impossible amount. Also ensure that we don't insert one normal ref and one ext ref for the same dentry - this happened because processing a dir index entry from the parent in the log happened when the normal ref item was full, which made the logic insert an extref and later when the normal ref had enough room, it would be inserted again when processing the ref item from the child inode in the log. This issue has been present since the introduction of the extrefs feature (2012). A test case for xfstests follows soon. This test only passes if the previous patch titled "Btrfs: fix fsync when extend references are added to an inode" is applied too. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-01-14 04:52:25 +03:00
const char *name, int namelen)
{
struct btrfs_path *path;
int ret;
path = btrfs_alloc_path();
if (!path)
return -ENOMEM;
ret = btrfs_search_slot(NULL, log, key, path, 0, 0);
if (ret < 0) {
goto out;
} else if (ret == 1) {
ret = 0;
goto out;
}
if (key->type == BTRFS_INODE_EXTREF_KEY)
ret = !!btrfs_find_name_in_ext_backref(path->nodes[0],
path->slots[0],
ref_objectid,
name, namelen);
else
ret = !!btrfs_find_name_in_backref(path->nodes[0],
path->slots[0],
name, namelen);
out:
btrfs_free_path(path);
return ret;
}
static inline int __add_inode_ref(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_path *path,
struct btrfs_root *log_root,
struct btrfs_inode *dir,
struct btrfs_inode *inode,
u64 inode_objectid, u64 parent_objectid,
u64 ref_index, char *name, int namelen,
int *search_done)
{
int ret;
char *victim_name;
int victim_name_len;
struct extent_buffer *leaf;
struct btrfs_dir_item *di;
struct btrfs_key search_key;
struct btrfs_inode_extref *extref;
again:
/* Search old style refs */
search_key.objectid = inode_objectid;
search_key.type = BTRFS_INODE_REF_KEY;
search_key.offset = parent_objectid;
ret = btrfs_search_slot(NULL, root, &search_key, path, 0, 0);
if (ret == 0) {
struct btrfs_inode_ref *victim_ref;
unsigned long ptr;
unsigned long ptr_end;
leaf = path->nodes[0];
/* are we trying to overwrite a back ref for the root directory
* if so, just jump out, we're done
*/
if (search_key.objectid == search_key.offset)
return 1;
/* check all the names in this back reference to see
* if they are in the log. if so, we allow them to stay
* otherwise they must be unlinked as a conflict
*/
ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
ptr_end = ptr + btrfs_item_size(leaf, path->slots[0]);
while (ptr < ptr_end) {
victim_ref = (struct btrfs_inode_ref *)ptr;
victim_name_len = btrfs_inode_ref_name_len(leaf,
victim_ref);
victim_name = kmalloc(victim_name_len, GFP_NOFS);
if (!victim_name)
return -ENOMEM;
read_extent_buffer(leaf, victim_name,
(unsigned long)(victim_ref + 1),
victim_name_len);
ret = backref_in_log(log_root, &search_key,
parent_objectid, victim_name,
victim_name_len);
if (ret < 0) {
kfree(victim_name);
return ret;
} else if (!ret) {
inc_nlink(&inode->vfs_inode);
btrfs_release_path(path);
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 17:24:20 +03:00
ret = btrfs_unlink_inode(trans, dir, inode,
victim_name, victim_name_len);
kfree(victim_name);
if (ret)
return ret;
ret = btrfs_run_delayed_items(trans);
if (ret)
return ret;
*search_done = 1;
goto again;
}
kfree(victim_name);
ptr = (unsigned long)(victim_ref + 1) + victim_name_len;
}
/*
* NOTE: we have searched root tree and checked the
* corresponding ref, it does not need to check again.
*/
*search_done = 1;
}
btrfs_release_path(path);
/* Same search but for extended refs */
extref = btrfs_lookup_inode_extref(NULL, root, path, name, namelen,
inode_objectid, parent_objectid, 0,
0);
if (!IS_ERR_OR_NULL(extref)) {
u32 item_size;
u32 cur_offset = 0;
unsigned long base;
struct inode *victim_parent;
leaf = path->nodes[0];
item_size = btrfs_item_size(leaf, path->slots[0]);
base = btrfs_item_ptr_offset(leaf, path->slots[0]);
while (cur_offset < item_size) {
extref = (struct btrfs_inode_extref *)(base + cur_offset);
victim_name_len = btrfs_inode_extref_name_len(leaf, extref);
if (btrfs_inode_extref_parent(leaf, extref) != parent_objectid)
goto next;
victim_name = kmalloc(victim_name_len, GFP_NOFS);
if (!victim_name)
return -ENOMEM;
read_extent_buffer(leaf, victim_name, (unsigned long)&extref->name,
victim_name_len);
search_key.objectid = inode_objectid;
search_key.type = BTRFS_INODE_EXTREF_KEY;
search_key.offset = btrfs_extref_hash(parent_objectid,
victim_name,
victim_name_len);
ret = backref_in_log(log_root, &search_key,
parent_objectid, victim_name,
victim_name_len);
if (ret < 0) {
btrfs: fix memory leak in __add_inode_ref() Line 1169 (#3) allocates a memory chunk for victim_name by kmalloc(), but when the function returns in line 1184 (#4) victim_name allocated by line 1169 (#3) is not freed, which will lead to a memory leak. There is a similar snippet of code in this function as allocating a memory chunk for victim_name in line 1104 (#1) as well as releasing the memory in line 1116 (#2). We should kfree() victim_name when the return value of backref_in_log() is less than zero and before the function returns in line 1184 (#4). 1057 static inline int __add_inode_ref(struct btrfs_trans_handle *trans, 1058 struct btrfs_root *root, 1059 struct btrfs_path *path, 1060 struct btrfs_root *log_root, 1061 struct btrfs_inode *dir, 1062 struct btrfs_inode *inode, 1063 u64 inode_objectid, u64 parent_objectid, 1064 u64 ref_index, char *name, int namelen, 1065 int *search_done) 1066 { 1104 victim_name = kmalloc(victim_name_len, GFP_NOFS); // #1: kmalloc (victim_name-1) 1105 if (!victim_name) 1106 return -ENOMEM; 1112 ret = backref_in_log(log_root, &search_key, 1113 parent_objectid, victim_name, 1114 victim_name_len); 1115 if (ret < 0) { 1116 kfree(victim_name); // #2: kfree (victim_name-1) 1117 return ret; 1118 } else if (!ret) { 1169 victim_name = kmalloc(victim_name_len, GFP_NOFS); // #3: kmalloc (victim_name-2) 1170 if (!victim_name) 1171 return -ENOMEM; 1180 ret = backref_in_log(log_root, &search_key, 1181 parent_objectid, victim_name, 1182 victim_name_len); 1183 if (ret < 0) { 1184 return ret; // #4: missing kfree (victim_name-2) 1185 } else if (!ret) { 1241 return 0; 1242 } Fixes: d3316c8233bb ("btrfs: Properly handle backref_in_log retval") CC: stable@vger.kernel.org # 5.10+ Reviewed-by: Qu Wenruo <wqu@suse.com> Reviewed-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Jianglei Nie <niejianglei2021@163.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-12-09 09:56:31 +03:00
kfree(victim_name);
return ret;
} else if (!ret) {
ret = -ENOENT;
victim_parent = read_one_inode(root,
parent_objectid);
if (victim_parent) {
inc_nlink(&inode->vfs_inode);
btrfs_release_path(path);
ret = btrfs_unlink_inode(trans,
BTRFS_I(victim_parent),
inode,
victim_name,
victim_name_len);
if (!ret)
ret = btrfs_run_delayed_items(
trans);
}
iput(victim_parent);
kfree(victim_name);
if (ret)
return ret;
*search_done = 1;
goto again;
}
kfree(victim_name);
next:
cur_offset += victim_name_len + sizeof(*extref);
}
*search_done = 1;
}
btrfs_release_path(path);
/* look for a conflicting sequence number */
di = btrfs_lookup_dir_index_item(trans, root, path, btrfs_ino(dir),
ref_index, name, namelen, 0);
if (IS_ERR(di)) {
return PTR_ERR(di);
} else if (di) {
ret = drop_one_dir_item(trans, path, dir, di);
if (ret)
return ret;
}
btrfs_release_path(path);
/* look for a conflicting name */
di = btrfs_lookup_dir_item(trans, root, path, btrfs_ino(dir),
name, namelen, 0);
if (IS_ERR(di)) {
return PTR_ERR(di);
} else if (di) {
ret = drop_one_dir_item(trans, path, dir, di);
if (ret)
return ret;
}
btrfs_release_path(path);
return 0;
}
static int extref_get_fields(struct extent_buffer *eb, unsigned long ref_ptr,
u32 *namelen, char **name, u64 *index,
u64 *parent_objectid)
{
struct btrfs_inode_extref *extref;
extref = (struct btrfs_inode_extref *)ref_ptr;
*namelen = btrfs_inode_extref_name_len(eb, extref);
*name = kmalloc(*namelen, GFP_NOFS);
if (*name == NULL)
return -ENOMEM;
read_extent_buffer(eb, *name, (unsigned long)&extref->name,
*namelen);
Btrfs: fix log replay failure after unlink and link combination If we have a file with 2 (or more) hard links in the same directory, remove one of the hard links, create a new file (or link an existing file) in the same directory with the name of the removed hard link, and then finally fsync the new file, we end up with a log that fails to replay, causing a mount failure. Example: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/foo $ ln /mnt/testdir/foo /mnt/testdir/bar $ sync $ unlink /mnt/testdir/bar $ touch /mnt/testdir/bar $ xfs_io -c "fsync" /mnt/testdir/bar <power failure> $ mount /dev/sdb /mnt mount: mount(2) failed: /mnt: No such file or directory When replaying the log, for that example, we also see the following in dmesg/syslog: [71813.671307] BTRFS info (device dm-0): failed to delete reference to bar, inode 258 parent 257 [71813.674204] ------------[ cut here ]------------ [71813.675694] BTRFS: Transaction aborted (error -2) [71813.677236] WARNING: CPU: 1 PID: 13231 at fs/btrfs/inode.c:4128 __btrfs_unlink_inode+0x17b/0x355 [btrfs] [71813.679669] Modules linked in: btrfs xfs f2fs dm_flakey dm_mod dax ghash_clmulni_intel ppdev pcbc aesni_intel aes_x86_64 crypto_simd cryptd glue_helper evdev psmouse i2c_piix4 parport_pc i2c_core pcspkr sg serio_raw parport button sunrpc loop autofs4 ext4 crc16 mbcache jbd2 zstd_decompress zstd_compress xxhash raid10 raid456 async_raid6_recov async_memcpy async_pq async_xor async_tx xor raid6_pq libcrc32c crc32c_generic raid1 raid0 multipath linear md_mod ata_generic sd_mod virtio_scsi ata_piix libata virtio_pci virtio_ring crc32c_intel floppy virtio e1000 scsi_mod [last unloaded: btrfs] [71813.679669] CPU: 1 PID: 13231 Comm: mount Tainted: G W 4.15.0-rc9-btrfs-next-56+ #1 [71813.679669] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.10.2-0-g5f4c7b1-prebuilt.qemu-project.org 04/01/2014 [71813.679669] RIP: 0010:__btrfs_unlink_inode+0x17b/0x355 [btrfs] [71813.679669] RSP: 0018:ffffc90001cef738 EFLAGS: 00010286 [71813.679669] RAX: 0000000000000025 RBX: ffff880217ce4708 RCX: 0000000000000001 [71813.679669] RDX: 0000000000000000 RSI: ffffffff81c14bae RDI: 00000000ffffffff [71813.679669] RBP: ffffc90001cef7c0 R08: 0000000000000001 R09: 0000000000000001 [71813.679669] R10: ffffc90001cef5e0 R11: ffffffff8343f007 R12: ffff880217d474c8 [71813.679669] R13: 00000000fffffffe R14: ffff88021ccf1548 R15: 0000000000000101 [71813.679669] FS: 00007f7cee84c480(0000) GS:ffff88023fc80000(0000) knlGS:0000000000000000 [71813.679669] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [71813.679669] CR2: 00007f7cedc1abf9 CR3: 00000002354b4003 CR4: 00000000001606e0 [71813.679669] Call Trace: [71813.679669] btrfs_unlink_inode+0x17/0x41 [btrfs] [71813.679669] drop_one_dir_item+0xfa/0x131 [btrfs] [71813.679669] add_inode_ref+0x71e/0x851 [btrfs] [71813.679669] ? __lock_is_held+0x39/0x71 [71813.679669] ? replay_one_buffer+0x53/0x53a [btrfs] [71813.679669] replay_one_buffer+0x4a4/0x53a [btrfs] [71813.679669] ? rcu_read_unlock+0x3a/0x57 [71813.679669] ? __lock_is_held+0x39/0x71 [71813.679669] walk_up_log_tree+0x101/0x1d2 [btrfs] [71813.679669] walk_log_tree+0xad/0x188 [btrfs] [71813.679669] btrfs_recover_log_trees+0x1fa/0x31e [btrfs] [71813.679669] ? replay_one_extent+0x544/0x544 [btrfs] [71813.679669] open_ctree+0x1cf6/0x2209 [btrfs] [71813.679669] btrfs_mount_root+0x368/0x482 [btrfs] [71813.679669] ? trace_hardirqs_on_caller+0x14c/0x1a6 [71813.679669] ? __lockdep_init_map+0x176/0x1c2 [71813.679669] ? mount_fs+0x64/0x10b [71813.679669] mount_fs+0x64/0x10b [71813.679669] vfs_kern_mount+0x68/0xce [71813.679669] btrfs_mount+0x13e/0x772 [btrfs] [71813.679669] ? trace_hardirqs_on_caller+0x14c/0x1a6 [71813.679669] ? __lockdep_init_map+0x176/0x1c2 [71813.679669] ? mount_fs+0x64/0x10b [71813.679669] mount_fs+0x64/0x10b [71813.679669] vfs_kern_mount+0x68/0xce [71813.679669] do_mount+0x6e5/0x973 [71813.679669] ? memdup_user+0x3e/0x5c [71813.679669] SyS_mount+0x72/0x98 [71813.679669] entry_SYSCALL_64_fastpath+0x1e/0x8b [71813.679669] RIP: 0033:0x7f7cedf150ba [71813.679669] RSP: 002b:00007ffca71da688 EFLAGS: 00000206 [71813.679669] Code: 7f a0 e8 51 0c fd ff 48 8b 43 50 f0 0f ba a8 30 2c 00 00 02 72 17 41 83 fd fb 74 11 44 89 ee 48 c7 c7 7d 11 7f a0 e8 38 f5 8d e0 <0f> ff 44 89 e9 ba 20 10 00 00 eb 4d 48 8b 4d b0 48 8b 75 88 4c [71813.679669] ---[ end trace 83bd473fc5b4663b ]--- [71813.854764] BTRFS: error (device dm-0) in __btrfs_unlink_inode:4128: errno=-2 No such entry [71813.886994] BTRFS: error (device dm-0) in btrfs_replay_log:2307: errno=-2 No such entry (Failed to recover log tree) [71813.903357] BTRFS error (device dm-0): cleaner transaction attach returned -30 [71814.128078] BTRFS error (device dm-0): open_ctree failed This happens because the log has inode reference items for both inode 258 (the first file we created) and inode 259 (the second file created), and when processing the reference item for inode 258, we replace the corresponding item in the subvolume tree (which has two names, "foo" and "bar") witht he one in the log (which only has one name, "foo") without removing the corresponding dir index keys from the parent directory. Later, when processing the inode reference item for inode 259, which has a name of "bar" associated to it, we notice that dir index entries exist for that name and for a different inode, so we attempt to unlink that name, which fails because the inode reference item for inode 258 no longer has the name "bar" associated to it, making a call to btrfs_unlink_inode() fail with a -ENOENT error. Fix this by unlinking all the names in an inode reference item from a subvolume tree that are not present in the inode reference item found in the log tree, before overwriting it with the item from the log tree. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2018-02-28 18:56:10 +03:00
if (index)
*index = btrfs_inode_extref_index(eb, extref);
if (parent_objectid)
*parent_objectid = btrfs_inode_extref_parent(eb, extref);
return 0;
}
static int ref_get_fields(struct extent_buffer *eb, unsigned long ref_ptr,
u32 *namelen, char **name, u64 *index)
{
struct btrfs_inode_ref *ref;
ref = (struct btrfs_inode_ref *)ref_ptr;
*namelen = btrfs_inode_ref_name_len(eb, ref);
*name = kmalloc(*namelen, GFP_NOFS);
if (*name == NULL)
return -ENOMEM;
read_extent_buffer(eb, *name, (unsigned long)(ref + 1), *namelen);
Btrfs: fix log replay failure after unlink and link combination If we have a file with 2 (or more) hard links in the same directory, remove one of the hard links, create a new file (or link an existing file) in the same directory with the name of the removed hard link, and then finally fsync the new file, we end up with a log that fails to replay, causing a mount failure. Example: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/foo $ ln /mnt/testdir/foo /mnt/testdir/bar $ sync $ unlink /mnt/testdir/bar $ touch /mnt/testdir/bar $ xfs_io -c "fsync" /mnt/testdir/bar <power failure> $ mount /dev/sdb /mnt mount: mount(2) failed: /mnt: No such file or directory When replaying the log, for that example, we also see the following in dmesg/syslog: [71813.671307] BTRFS info (device dm-0): failed to delete reference to bar, inode 258 parent 257 [71813.674204] ------------[ cut here ]------------ [71813.675694] BTRFS: Transaction aborted (error -2) [71813.677236] WARNING: CPU: 1 PID: 13231 at fs/btrfs/inode.c:4128 __btrfs_unlink_inode+0x17b/0x355 [btrfs] [71813.679669] Modules linked in: btrfs xfs f2fs dm_flakey dm_mod dax ghash_clmulni_intel ppdev pcbc aesni_intel aes_x86_64 crypto_simd cryptd glue_helper evdev psmouse i2c_piix4 parport_pc i2c_core pcspkr sg serio_raw parport button sunrpc loop autofs4 ext4 crc16 mbcache jbd2 zstd_decompress zstd_compress xxhash raid10 raid456 async_raid6_recov async_memcpy async_pq async_xor async_tx xor raid6_pq libcrc32c crc32c_generic raid1 raid0 multipath linear md_mod ata_generic sd_mod virtio_scsi ata_piix libata virtio_pci virtio_ring crc32c_intel floppy virtio e1000 scsi_mod [last unloaded: btrfs] [71813.679669] CPU: 1 PID: 13231 Comm: mount Tainted: G W 4.15.0-rc9-btrfs-next-56+ #1 [71813.679669] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.10.2-0-g5f4c7b1-prebuilt.qemu-project.org 04/01/2014 [71813.679669] RIP: 0010:__btrfs_unlink_inode+0x17b/0x355 [btrfs] [71813.679669] RSP: 0018:ffffc90001cef738 EFLAGS: 00010286 [71813.679669] RAX: 0000000000000025 RBX: ffff880217ce4708 RCX: 0000000000000001 [71813.679669] RDX: 0000000000000000 RSI: ffffffff81c14bae RDI: 00000000ffffffff [71813.679669] RBP: ffffc90001cef7c0 R08: 0000000000000001 R09: 0000000000000001 [71813.679669] R10: ffffc90001cef5e0 R11: ffffffff8343f007 R12: ffff880217d474c8 [71813.679669] R13: 00000000fffffffe R14: ffff88021ccf1548 R15: 0000000000000101 [71813.679669] FS: 00007f7cee84c480(0000) GS:ffff88023fc80000(0000) knlGS:0000000000000000 [71813.679669] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [71813.679669] CR2: 00007f7cedc1abf9 CR3: 00000002354b4003 CR4: 00000000001606e0 [71813.679669] Call Trace: [71813.679669] btrfs_unlink_inode+0x17/0x41 [btrfs] [71813.679669] drop_one_dir_item+0xfa/0x131 [btrfs] [71813.679669] add_inode_ref+0x71e/0x851 [btrfs] [71813.679669] ? __lock_is_held+0x39/0x71 [71813.679669] ? replay_one_buffer+0x53/0x53a [btrfs] [71813.679669] replay_one_buffer+0x4a4/0x53a [btrfs] [71813.679669] ? rcu_read_unlock+0x3a/0x57 [71813.679669] ? __lock_is_held+0x39/0x71 [71813.679669] walk_up_log_tree+0x101/0x1d2 [btrfs] [71813.679669] walk_log_tree+0xad/0x188 [btrfs] [71813.679669] btrfs_recover_log_trees+0x1fa/0x31e [btrfs] [71813.679669] ? replay_one_extent+0x544/0x544 [btrfs] [71813.679669] open_ctree+0x1cf6/0x2209 [btrfs] [71813.679669] btrfs_mount_root+0x368/0x482 [btrfs] [71813.679669] ? trace_hardirqs_on_caller+0x14c/0x1a6 [71813.679669] ? __lockdep_init_map+0x176/0x1c2 [71813.679669] ? mount_fs+0x64/0x10b [71813.679669] mount_fs+0x64/0x10b [71813.679669] vfs_kern_mount+0x68/0xce [71813.679669] btrfs_mount+0x13e/0x772 [btrfs] [71813.679669] ? trace_hardirqs_on_caller+0x14c/0x1a6 [71813.679669] ? __lockdep_init_map+0x176/0x1c2 [71813.679669] ? mount_fs+0x64/0x10b [71813.679669] mount_fs+0x64/0x10b [71813.679669] vfs_kern_mount+0x68/0xce [71813.679669] do_mount+0x6e5/0x973 [71813.679669] ? memdup_user+0x3e/0x5c [71813.679669] SyS_mount+0x72/0x98 [71813.679669] entry_SYSCALL_64_fastpath+0x1e/0x8b [71813.679669] RIP: 0033:0x7f7cedf150ba [71813.679669] RSP: 002b:00007ffca71da688 EFLAGS: 00000206 [71813.679669] Code: 7f a0 e8 51 0c fd ff 48 8b 43 50 f0 0f ba a8 30 2c 00 00 02 72 17 41 83 fd fb 74 11 44 89 ee 48 c7 c7 7d 11 7f a0 e8 38 f5 8d e0 <0f> ff 44 89 e9 ba 20 10 00 00 eb 4d 48 8b 4d b0 48 8b 75 88 4c [71813.679669] ---[ end trace 83bd473fc5b4663b ]--- [71813.854764] BTRFS: error (device dm-0) in __btrfs_unlink_inode:4128: errno=-2 No such entry [71813.886994] BTRFS: error (device dm-0) in btrfs_replay_log:2307: errno=-2 No such entry (Failed to recover log tree) [71813.903357] BTRFS error (device dm-0): cleaner transaction attach returned -30 [71814.128078] BTRFS error (device dm-0): open_ctree failed This happens because the log has inode reference items for both inode 258 (the first file we created) and inode 259 (the second file created), and when processing the reference item for inode 258, we replace the corresponding item in the subvolume tree (which has two names, "foo" and "bar") witht he one in the log (which only has one name, "foo") without removing the corresponding dir index keys from the parent directory. Later, when processing the inode reference item for inode 259, which has a name of "bar" associated to it, we notice that dir index entries exist for that name and for a different inode, so we attempt to unlink that name, which fails because the inode reference item for inode 258 no longer has the name "bar" associated to it, making a call to btrfs_unlink_inode() fail with a -ENOENT error. Fix this by unlinking all the names in an inode reference item from a subvolume tree that are not present in the inode reference item found in the log tree, before overwriting it with the item from the log tree. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2018-02-28 18:56:10 +03:00
if (index)
*index = btrfs_inode_ref_index(eb, ref);
return 0;
}
Btrfs: fix log replay failure after unlink and link combination If we have a file with 2 (or more) hard links in the same directory, remove one of the hard links, create a new file (or link an existing file) in the same directory with the name of the removed hard link, and then finally fsync the new file, we end up with a log that fails to replay, causing a mount failure. Example: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/foo $ ln /mnt/testdir/foo /mnt/testdir/bar $ sync $ unlink /mnt/testdir/bar $ touch /mnt/testdir/bar $ xfs_io -c "fsync" /mnt/testdir/bar <power failure> $ mount /dev/sdb /mnt mount: mount(2) failed: /mnt: No such file or directory When replaying the log, for that example, we also see the following in dmesg/syslog: [71813.671307] BTRFS info (device dm-0): failed to delete reference to bar, inode 258 parent 257 [71813.674204] ------------[ cut here ]------------ [71813.675694] BTRFS: Transaction aborted (error -2) [71813.677236] WARNING: CPU: 1 PID: 13231 at fs/btrfs/inode.c:4128 __btrfs_unlink_inode+0x17b/0x355 [btrfs] [71813.679669] Modules linked in: btrfs xfs f2fs dm_flakey dm_mod dax ghash_clmulni_intel ppdev pcbc aesni_intel aes_x86_64 crypto_simd cryptd glue_helper evdev psmouse i2c_piix4 parport_pc i2c_core pcspkr sg serio_raw parport button sunrpc loop autofs4 ext4 crc16 mbcache jbd2 zstd_decompress zstd_compress xxhash raid10 raid456 async_raid6_recov async_memcpy async_pq async_xor async_tx xor raid6_pq libcrc32c crc32c_generic raid1 raid0 multipath linear md_mod ata_generic sd_mod virtio_scsi ata_piix libata virtio_pci virtio_ring crc32c_intel floppy virtio e1000 scsi_mod [last unloaded: btrfs] [71813.679669] CPU: 1 PID: 13231 Comm: mount Tainted: G W 4.15.0-rc9-btrfs-next-56+ #1 [71813.679669] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.10.2-0-g5f4c7b1-prebuilt.qemu-project.org 04/01/2014 [71813.679669] RIP: 0010:__btrfs_unlink_inode+0x17b/0x355 [btrfs] [71813.679669] RSP: 0018:ffffc90001cef738 EFLAGS: 00010286 [71813.679669] RAX: 0000000000000025 RBX: ffff880217ce4708 RCX: 0000000000000001 [71813.679669] RDX: 0000000000000000 RSI: ffffffff81c14bae RDI: 00000000ffffffff [71813.679669] RBP: ffffc90001cef7c0 R08: 0000000000000001 R09: 0000000000000001 [71813.679669] R10: ffffc90001cef5e0 R11: ffffffff8343f007 R12: ffff880217d474c8 [71813.679669] R13: 00000000fffffffe R14: ffff88021ccf1548 R15: 0000000000000101 [71813.679669] FS: 00007f7cee84c480(0000) GS:ffff88023fc80000(0000) knlGS:0000000000000000 [71813.679669] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [71813.679669] CR2: 00007f7cedc1abf9 CR3: 00000002354b4003 CR4: 00000000001606e0 [71813.679669] Call Trace: [71813.679669] btrfs_unlink_inode+0x17/0x41 [btrfs] [71813.679669] drop_one_dir_item+0xfa/0x131 [btrfs] [71813.679669] add_inode_ref+0x71e/0x851 [btrfs] [71813.679669] ? __lock_is_held+0x39/0x71 [71813.679669] ? replay_one_buffer+0x53/0x53a [btrfs] [71813.679669] replay_one_buffer+0x4a4/0x53a [btrfs] [71813.679669] ? rcu_read_unlock+0x3a/0x57 [71813.679669] ? __lock_is_held+0x39/0x71 [71813.679669] walk_up_log_tree+0x101/0x1d2 [btrfs] [71813.679669] walk_log_tree+0xad/0x188 [btrfs] [71813.679669] btrfs_recover_log_trees+0x1fa/0x31e [btrfs] [71813.679669] ? replay_one_extent+0x544/0x544 [btrfs] [71813.679669] open_ctree+0x1cf6/0x2209 [btrfs] [71813.679669] btrfs_mount_root+0x368/0x482 [btrfs] [71813.679669] ? trace_hardirqs_on_caller+0x14c/0x1a6 [71813.679669] ? __lockdep_init_map+0x176/0x1c2 [71813.679669] ? mount_fs+0x64/0x10b [71813.679669] mount_fs+0x64/0x10b [71813.679669] vfs_kern_mount+0x68/0xce [71813.679669] btrfs_mount+0x13e/0x772 [btrfs] [71813.679669] ? trace_hardirqs_on_caller+0x14c/0x1a6 [71813.679669] ? __lockdep_init_map+0x176/0x1c2 [71813.679669] ? mount_fs+0x64/0x10b [71813.679669] mount_fs+0x64/0x10b [71813.679669] vfs_kern_mount+0x68/0xce [71813.679669] do_mount+0x6e5/0x973 [71813.679669] ? memdup_user+0x3e/0x5c [71813.679669] SyS_mount+0x72/0x98 [71813.679669] entry_SYSCALL_64_fastpath+0x1e/0x8b [71813.679669] RIP: 0033:0x7f7cedf150ba [71813.679669] RSP: 002b:00007ffca71da688 EFLAGS: 00000206 [71813.679669] Code: 7f a0 e8 51 0c fd ff 48 8b 43 50 f0 0f ba a8 30 2c 00 00 02 72 17 41 83 fd fb 74 11 44 89 ee 48 c7 c7 7d 11 7f a0 e8 38 f5 8d e0 <0f> ff 44 89 e9 ba 20 10 00 00 eb 4d 48 8b 4d b0 48 8b 75 88 4c [71813.679669] ---[ end trace 83bd473fc5b4663b ]--- [71813.854764] BTRFS: error (device dm-0) in __btrfs_unlink_inode:4128: errno=-2 No such entry [71813.886994] BTRFS: error (device dm-0) in btrfs_replay_log:2307: errno=-2 No such entry (Failed to recover log tree) [71813.903357] BTRFS error (device dm-0): cleaner transaction attach returned -30 [71814.128078] BTRFS error (device dm-0): open_ctree failed This happens because the log has inode reference items for both inode 258 (the first file we created) and inode 259 (the second file created), and when processing the reference item for inode 258, we replace the corresponding item in the subvolume tree (which has two names, "foo" and "bar") witht he one in the log (which only has one name, "foo") without removing the corresponding dir index keys from the parent directory. Later, when processing the inode reference item for inode 259, which has a name of "bar" associated to it, we notice that dir index entries exist for that name and for a different inode, so we attempt to unlink that name, which fails because the inode reference item for inode 258 no longer has the name "bar" associated to it, making a call to btrfs_unlink_inode() fail with a -ENOENT error. Fix this by unlinking all the names in an inode reference item from a subvolume tree that are not present in the inode reference item found in the log tree, before overwriting it with the item from the log tree. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2018-02-28 18:56:10 +03:00
/*
* Take an inode reference item from the log tree and iterate all names from the
* inode reference item in the subvolume tree with the same key (if it exists).
* For any name that is not in the inode reference item from the log tree, do a
* proper unlink of that name (that is, remove its entry from the inode
* reference item and both dir index keys).
*/
static int unlink_old_inode_refs(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_path *path,
struct btrfs_inode *inode,
struct extent_buffer *log_eb,
int log_slot,
struct btrfs_key *key)
{
int ret;
unsigned long ref_ptr;
unsigned long ref_end;
struct extent_buffer *eb;
again:
btrfs_release_path(path);
ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
if (ret > 0) {
ret = 0;
goto out;
}
if (ret < 0)
goto out;
eb = path->nodes[0];
ref_ptr = btrfs_item_ptr_offset(eb, path->slots[0]);
ref_end = ref_ptr + btrfs_item_size(eb, path->slots[0]);
Btrfs: fix log replay failure after unlink and link combination If we have a file with 2 (or more) hard links in the same directory, remove one of the hard links, create a new file (or link an existing file) in the same directory with the name of the removed hard link, and then finally fsync the new file, we end up with a log that fails to replay, causing a mount failure. Example: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/foo $ ln /mnt/testdir/foo /mnt/testdir/bar $ sync $ unlink /mnt/testdir/bar $ touch /mnt/testdir/bar $ xfs_io -c "fsync" /mnt/testdir/bar <power failure> $ mount /dev/sdb /mnt mount: mount(2) failed: /mnt: No such file or directory When replaying the log, for that example, we also see the following in dmesg/syslog: [71813.671307] BTRFS info (device dm-0): failed to delete reference to bar, inode 258 parent 257 [71813.674204] ------------[ cut here ]------------ [71813.675694] BTRFS: Transaction aborted (error -2) [71813.677236] WARNING: CPU: 1 PID: 13231 at fs/btrfs/inode.c:4128 __btrfs_unlink_inode+0x17b/0x355 [btrfs] [71813.679669] Modules linked in: btrfs xfs f2fs dm_flakey dm_mod dax ghash_clmulni_intel ppdev pcbc aesni_intel aes_x86_64 crypto_simd cryptd glue_helper evdev psmouse i2c_piix4 parport_pc i2c_core pcspkr sg serio_raw parport button sunrpc loop autofs4 ext4 crc16 mbcache jbd2 zstd_decompress zstd_compress xxhash raid10 raid456 async_raid6_recov async_memcpy async_pq async_xor async_tx xor raid6_pq libcrc32c crc32c_generic raid1 raid0 multipath linear md_mod ata_generic sd_mod virtio_scsi ata_piix libata virtio_pci virtio_ring crc32c_intel floppy virtio e1000 scsi_mod [last unloaded: btrfs] [71813.679669] CPU: 1 PID: 13231 Comm: mount Tainted: G W 4.15.0-rc9-btrfs-next-56+ #1 [71813.679669] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.10.2-0-g5f4c7b1-prebuilt.qemu-project.org 04/01/2014 [71813.679669] RIP: 0010:__btrfs_unlink_inode+0x17b/0x355 [btrfs] [71813.679669] RSP: 0018:ffffc90001cef738 EFLAGS: 00010286 [71813.679669] RAX: 0000000000000025 RBX: ffff880217ce4708 RCX: 0000000000000001 [71813.679669] RDX: 0000000000000000 RSI: ffffffff81c14bae RDI: 00000000ffffffff [71813.679669] RBP: ffffc90001cef7c0 R08: 0000000000000001 R09: 0000000000000001 [71813.679669] R10: ffffc90001cef5e0 R11: ffffffff8343f007 R12: ffff880217d474c8 [71813.679669] R13: 00000000fffffffe R14: ffff88021ccf1548 R15: 0000000000000101 [71813.679669] FS: 00007f7cee84c480(0000) GS:ffff88023fc80000(0000) knlGS:0000000000000000 [71813.679669] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [71813.679669] CR2: 00007f7cedc1abf9 CR3: 00000002354b4003 CR4: 00000000001606e0 [71813.679669] Call Trace: [71813.679669] btrfs_unlink_inode+0x17/0x41 [btrfs] [71813.679669] drop_one_dir_item+0xfa/0x131 [btrfs] [71813.679669] add_inode_ref+0x71e/0x851 [btrfs] [71813.679669] ? __lock_is_held+0x39/0x71 [71813.679669] ? replay_one_buffer+0x53/0x53a [btrfs] [71813.679669] replay_one_buffer+0x4a4/0x53a [btrfs] [71813.679669] ? rcu_read_unlock+0x3a/0x57 [71813.679669] ? __lock_is_held+0x39/0x71 [71813.679669] walk_up_log_tree+0x101/0x1d2 [btrfs] [71813.679669] walk_log_tree+0xad/0x188 [btrfs] [71813.679669] btrfs_recover_log_trees+0x1fa/0x31e [btrfs] [71813.679669] ? replay_one_extent+0x544/0x544 [btrfs] [71813.679669] open_ctree+0x1cf6/0x2209 [btrfs] [71813.679669] btrfs_mount_root+0x368/0x482 [btrfs] [71813.679669] ? trace_hardirqs_on_caller+0x14c/0x1a6 [71813.679669] ? __lockdep_init_map+0x176/0x1c2 [71813.679669] ? mount_fs+0x64/0x10b [71813.679669] mount_fs+0x64/0x10b [71813.679669] vfs_kern_mount+0x68/0xce [71813.679669] btrfs_mount+0x13e/0x772 [btrfs] [71813.679669] ? trace_hardirqs_on_caller+0x14c/0x1a6 [71813.679669] ? __lockdep_init_map+0x176/0x1c2 [71813.679669] ? mount_fs+0x64/0x10b [71813.679669] mount_fs+0x64/0x10b [71813.679669] vfs_kern_mount+0x68/0xce [71813.679669] do_mount+0x6e5/0x973 [71813.679669] ? memdup_user+0x3e/0x5c [71813.679669] SyS_mount+0x72/0x98 [71813.679669] entry_SYSCALL_64_fastpath+0x1e/0x8b [71813.679669] RIP: 0033:0x7f7cedf150ba [71813.679669] RSP: 002b:00007ffca71da688 EFLAGS: 00000206 [71813.679669] Code: 7f a0 e8 51 0c fd ff 48 8b 43 50 f0 0f ba a8 30 2c 00 00 02 72 17 41 83 fd fb 74 11 44 89 ee 48 c7 c7 7d 11 7f a0 e8 38 f5 8d e0 <0f> ff 44 89 e9 ba 20 10 00 00 eb 4d 48 8b 4d b0 48 8b 75 88 4c [71813.679669] ---[ end trace 83bd473fc5b4663b ]--- [71813.854764] BTRFS: error (device dm-0) in __btrfs_unlink_inode:4128: errno=-2 No such entry [71813.886994] BTRFS: error (device dm-0) in btrfs_replay_log:2307: errno=-2 No such entry (Failed to recover log tree) [71813.903357] BTRFS error (device dm-0): cleaner transaction attach returned -30 [71814.128078] BTRFS error (device dm-0): open_ctree failed This happens because the log has inode reference items for both inode 258 (the first file we created) and inode 259 (the second file created), and when processing the reference item for inode 258, we replace the corresponding item in the subvolume tree (which has two names, "foo" and "bar") witht he one in the log (which only has one name, "foo") without removing the corresponding dir index keys from the parent directory. Later, when processing the inode reference item for inode 259, which has a name of "bar" associated to it, we notice that dir index entries exist for that name and for a different inode, so we attempt to unlink that name, which fails because the inode reference item for inode 258 no longer has the name "bar" associated to it, making a call to btrfs_unlink_inode() fail with a -ENOENT error. Fix this by unlinking all the names in an inode reference item from a subvolume tree that are not present in the inode reference item found in the log tree, before overwriting it with the item from the log tree. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2018-02-28 18:56:10 +03:00
while (ref_ptr < ref_end) {
char *name = NULL;
int namelen;
u64 parent_id;
if (key->type == BTRFS_INODE_EXTREF_KEY) {
ret = extref_get_fields(eb, ref_ptr, &namelen, &name,
NULL, &parent_id);
} else {
parent_id = key->offset;
ret = ref_get_fields(eb, ref_ptr, &namelen, &name,
NULL);
}
if (ret)
goto out;
if (key->type == BTRFS_INODE_EXTREF_KEY)
ret = !!btrfs_find_name_in_ext_backref(log_eb, log_slot,
parent_id, name,
namelen);
Btrfs: fix log replay failure after unlink and link combination If we have a file with 2 (or more) hard links in the same directory, remove one of the hard links, create a new file (or link an existing file) in the same directory with the name of the removed hard link, and then finally fsync the new file, we end up with a log that fails to replay, causing a mount failure. Example: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/foo $ ln /mnt/testdir/foo /mnt/testdir/bar $ sync $ unlink /mnt/testdir/bar $ touch /mnt/testdir/bar $ xfs_io -c "fsync" /mnt/testdir/bar <power failure> $ mount /dev/sdb /mnt mount: mount(2) failed: /mnt: No such file or directory When replaying the log, for that example, we also see the following in dmesg/syslog: [71813.671307] BTRFS info (device dm-0): failed to delete reference to bar, inode 258 parent 257 [71813.674204] ------------[ cut here ]------------ [71813.675694] BTRFS: Transaction aborted (error -2) [71813.677236] WARNING: CPU: 1 PID: 13231 at fs/btrfs/inode.c:4128 __btrfs_unlink_inode+0x17b/0x355 [btrfs] [71813.679669] Modules linked in: btrfs xfs f2fs dm_flakey dm_mod dax ghash_clmulni_intel ppdev pcbc aesni_intel aes_x86_64 crypto_simd cryptd glue_helper evdev psmouse i2c_piix4 parport_pc i2c_core pcspkr sg serio_raw parport button sunrpc loop autofs4 ext4 crc16 mbcache jbd2 zstd_decompress zstd_compress xxhash raid10 raid456 async_raid6_recov async_memcpy async_pq async_xor async_tx xor raid6_pq libcrc32c crc32c_generic raid1 raid0 multipath linear md_mod ata_generic sd_mod virtio_scsi ata_piix libata virtio_pci virtio_ring crc32c_intel floppy virtio e1000 scsi_mod [last unloaded: btrfs] [71813.679669] CPU: 1 PID: 13231 Comm: mount Tainted: G W 4.15.0-rc9-btrfs-next-56+ #1 [71813.679669] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.10.2-0-g5f4c7b1-prebuilt.qemu-project.org 04/01/2014 [71813.679669] RIP: 0010:__btrfs_unlink_inode+0x17b/0x355 [btrfs] [71813.679669] RSP: 0018:ffffc90001cef738 EFLAGS: 00010286 [71813.679669] RAX: 0000000000000025 RBX: ffff880217ce4708 RCX: 0000000000000001 [71813.679669] RDX: 0000000000000000 RSI: ffffffff81c14bae RDI: 00000000ffffffff [71813.679669] RBP: ffffc90001cef7c0 R08: 0000000000000001 R09: 0000000000000001 [71813.679669] R10: ffffc90001cef5e0 R11: ffffffff8343f007 R12: ffff880217d474c8 [71813.679669] R13: 00000000fffffffe R14: ffff88021ccf1548 R15: 0000000000000101 [71813.679669] FS: 00007f7cee84c480(0000) GS:ffff88023fc80000(0000) knlGS:0000000000000000 [71813.679669] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [71813.679669] CR2: 00007f7cedc1abf9 CR3: 00000002354b4003 CR4: 00000000001606e0 [71813.679669] Call Trace: [71813.679669] btrfs_unlink_inode+0x17/0x41 [btrfs] [71813.679669] drop_one_dir_item+0xfa/0x131 [btrfs] [71813.679669] add_inode_ref+0x71e/0x851 [btrfs] [71813.679669] ? __lock_is_held+0x39/0x71 [71813.679669] ? replay_one_buffer+0x53/0x53a [btrfs] [71813.679669] replay_one_buffer+0x4a4/0x53a [btrfs] [71813.679669] ? rcu_read_unlock+0x3a/0x57 [71813.679669] ? __lock_is_held+0x39/0x71 [71813.679669] walk_up_log_tree+0x101/0x1d2 [btrfs] [71813.679669] walk_log_tree+0xad/0x188 [btrfs] [71813.679669] btrfs_recover_log_trees+0x1fa/0x31e [btrfs] [71813.679669] ? replay_one_extent+0x544/0x544 [btrfs] [71813.679669] open_ctree+0x1cf6/0x2209 [btrfs] [71813.679669] btrfs_mount_root+0x368/0x482 [btrfs] [71813.679669] ? trace_hardirqs_on_caller+0x14c/0x1a6 [71813.679669] ? __lockdep_init_map+0x176/0x1c2 [71813.679669] ? mount_fs+0x64/0x10b [71813.679669] mount_fs+0x64/0x10b [71813.679669] vfs_kern_mount+0x68/0xce [71813.679669] btrfs_mount+0x13e/0x772 [btrfs] [71813.679669] ? trace_hardirqs_on_caller+0x14c/0x1a6 [71813.679669] ? __lockdep_init_map+0x176/0x1c2 [71813.679669] ? mount_fs+0x64/0x10b [71813.679669] mount_fs+0x64/0x10b [71813.679669] vfs_kern_mount+0x68/0xce [71813.679669] do_mount+0x6e5/0x973 [71813.679669] ? memdup_user+0x3e/0x5c [71813.679669] SyS_mount+0x72/0x98 [71813.679669] entry_SYSCALL_64_fastpath+0x1e/0x8b [71813.679669] RIP: 0033:0x7f7cedf150ba [71813.679669] RSP: 002b:00007ffca71da688 EFLAGS: 00000206 [71813.679669] Code: 7f a0 e8 51 0c fd ff 48 8b 43 50 f0 0f ba a8 30 2c 00 00 02 72 17 41 83 fd fb 74 11 44 89 ee 48 c7 c7 7d 11 7f a0 e8 38 f5 8d e0 <0f> ff 44 89 e9 ba 20 10 00 00 eb 4d 48 8b 4d b0 48 8b 75 88 4c [71813.679669] ---[ end trace 83bd473fc5b4663b ]--- [71813.854764] BTRFS: error (device dm-0) in __btrfs_unlink_inode:4128: errno=-2 No such entry [71813.886994] BTRFS: error (device dm-0) in btrfs_replay_log:2307: errno=-2 No such entry (Failed to recover log tree) [71813.903357] BTRFS error (device dm-0): cleaner transaction attach returned -30 [71814.128078] BTRFS error (device dm-0): open_ctree failed This happens because the log has inode reference items for both inode 258 (the first file we created) and inode 259 (the second file created), and when processing the reference item for inode 258, we replace the corresponding item in the subvolume tree (which has two names, "foo" and "bar") witht he one in the log (which only has one name, "foo") without removing the corresponding dir index keys from the parent directory. Later, when processing the inode reference item for inode 259, which has a name of "bar" associated to it, we notice that dir index entries exist for that name and for a different inode, so we attempt to unlink that name, which fails because the inode reference item for inode 258 no longer has the name "bar" associated to it, making a call to btrfs_unlink_inode() fail with a -ENOENT error. Fix this by unlinking all the names in an inode reference item from a subvolume tree that are not present in the inode reference item found in the log tree, before overwriting it with the item from the log tree. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2018-02-28 18:56:10 +03:00
else
ret = !!btrfs_find_name_in_backref(log_eb, log_slot,
name, namelen);
Btrfs: fix log replay failure after unlink and link combination If we have a file with 2 (or more) hard links in the same directory, remove one of the hard links, create a new file (or link an existing file) in the same directory with the name of the removed hard link, and then finally fsync the new file, we end up with a log that fails to replay, causing a mount failure. Example: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/foo $ ln /mnt/testdir/foo /mnt/testdir/bar $ sync $ unlink /mnt/testdir/bar $ touch /mnt/testdir/bar $ xfs_io -c "fsync" /mnt/testdir/bar <power failure> $ mount /dev/sdb /mnt mount: mount(2) failed: /mnt: No such file or directory When replaying the log, for that example, we also see the following in dmesg/syslog: [71813.671307] BTRFS info (device dm-0): failed to delete reference to bar, inode 258 parent 257 [71813.674204] ------------[ cut here ]------------ [71813.675694] BTRFS: Transaction aborted (error -2) [71813.677236] WARNING: CPU: 1 PID: 13231 at fs/btrfs/inode.c:4128 __btrfs_unlink_inode+0x17b/0x355 [btrfs] [71813.679669] Modules linked in: btrfs xfs f2fs dm_flakey dm_mod dax ghash_clmulni_intel ppdev pcbc aesni_intel aes_x86_64 crypto_simd cryptd glue_helper evdev psmouse i2c_piix4 parport_pc i2c_core pcspkr sg serio_raw parport button sunrpc loop autofs4 ext4 crc16 mbcache jbd2 zstd_decompress zstd_compress xxhash raid10 raid456 async_raid6_recov async_memcpy async_pq async_xor async_tx xor raid6_pq libcrc32c crc32c_generic raid1 raid0 multipath linear md_mod ata_generic sd_mod virtio_scsi ata_piix libata virtio_pci virtio_ring crc32c_intel floppy virtio e1000 scsi_mod [last unloaded: btrfs] [71813.679669] CPU: 1 PID: 13231 Comm: mount Tainted: G W 4.15.0-rc9-btrfs-next-56+ #1 [71813.679669] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.10.2-0-g5f4c7b1-prebuilt.qemu-project.org 04/01/2014 [71813.679669] RIP: 0010:__btrfs_unlink_inode+0x17b/0x355 [btrfs] [71813.679669] RSP: 0018:ffffc90001cef738 EFLAGS: 00010286 [71813.679669] RAX: 0000000000000025 RBX: ffff880217ce4708 RCX: 0000000000000001 [71813.679669] RDX: 0000000000000000 RSI: ffffffff81c14bae RDI: 00000000ffffffff [71813.679669] RBP: ffffc90001cef7c0 R08: 0000000000000001 R09: 0000000000000001 [71813.679669] R10: ffffc90001cef5e0 R11: ffffffff8343f007 R12: ffff880217d474c8 [71813.679669] R13: 00000000fffffffe R14: ffff88021ccf1548 R15: 0000000000000101 [71813.679669] FS: 00007f7cee84c480(0000) GS:ffff88023fc80000(0000) knlGS:0000000000000000 [71813.679669] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [71813.679669] CR2: 00007f7cedc1abf9 CR3: 00000002354b4003 CR4: 00000000001606e0 [71813.679669] Call Trace: [71813.679669] btrfs_unlink_inode+0x17/0x41 [btrfs] [71813.679669] drop_one_dir_item+0xfa/0x131 [btrfs] [71813.679669] add_inode_ref+0x71e/0x851 [btrfs] [71813.679669] ? __lock_is_held+0x39/0x71 [71813.679669] ? replay_one_buffer+0x53/0x53a [btrfs] [71813.679669] replay_one_buffer+0x4a4/0x53a [btrfs] [71813.679669] ? rcu_read_unlock+0x3a/0x57 [71813.679669] ? __lock_is_held+0x39/0x71 [71813.679669] walk_up_log_tree+0x101/0x1d2 [btrfs] [71813.679669] walk_log_tree+0xad/0x188 [btrfs] [71813.679669] btrfs_recover_log_trees+0x1fa/0x31e [btrfs] [71813.679669] ? replay_one_extent+0x544/0x544 [btrfs] [71813.679669] open_ctree+0x1cf6/0x2209 [btrfs] [71813.679669] btrfs_mount_root+0x368/0x482 [btrfs] [71813.679669] ? trace_hardirqs_on_caller+0x14c/0x1a6 [71813.679669] ? __lockdep_init_map+0x176/0x1c2 [71813.679669] ? mount_fs+0x64/0x10b [71813.679669] mount_fs+0x64/0x10b [71813.679669] vfs_kern_mount+0x68/0xce [71813.679669] btrfs_mount+0x13e/0x772 [btrfs] [71813.679669] ? trace_hardirqs_on_caller+0x14c/0x1a6 [71813.679669] ? __lockdep_init_map+0x176/0x1c2 [71813.679669] ? mount_fs+0x64/0x10b [71813.679669] mount_fs+0x64/0x10b [71813.679669] vfs_kern_mount+0x68/0xce [71813.679669] do_mount+0x6e5/0x973 [71813.679669] ? memdup_user+0x3e/0x5c [71813.679669] SyS_mount+0x72/0x98 [71813.679669] entry_SYSCALL_64_fastpath+0x1e/0x8b [71813.679669] RIP: 0033:0x7f7cedf150ba [71813.679669] RSP: 002b:00007ffca71da688 EFLAGS: 00000206 [71813.679669] Code: 7f a0 e8 51 0c fd ff 48 8b 43 50 f0 0f ba a8 30 2c 00 00 02 72 17 41 83 fd fb 74 11 44 89 ee 48 c7 c7 7d 11 7f a0 e8 38 f5 8d e0 <0f> ff 44 89 e9 ba 20 10 00 00 eb 4d 48 8b 4d b0 48 8b 75 88 4c [71813.679669] ---[ end trace 83bd473fc5b4663b ]--- [71813.854764] BTRFS: error (device dm-0) in __btrfs_unlink_inode:4128: errno=-2 No such entry [71813.886994] BTRFS: error (device dm-0) in btrfs_replay_log:2307: errno=-2 No such entry (Failed to recover log tree) [71813.903357] BTRFS error (device dm-0): cleaner transaction attach returned -30 [71814.128078] BTRFS error (device dm-0): open_ctree failed This happens because the log has inode reference items for both inode 258 (the first file we created) and inode 259 (the second file created), and when processing the reference item for inode 258, we replace the corresponding item in the subvolume tree (which has two names, "foo" and "bar") witht he one in the log (which only has one name, "foo") without removing the corresponding dir index keys from the parent directory. Later, when processing the inode reference item for inode 259, which has a name of "bar" associated to it, we notice that dir index entries exist for that name and for a different inode, so we attempt to unlink that name, which fails because the inode reference item for inode 258 no longer has the name "bar" associated to it, making a call to btrfs_unlink_inode() fail with a -ENOENT error. Fix this by unlinking all the names in an inode reference item from a subvolume tree that are not present in the inode reference item found in the log tree, before overwriting it with the item from the log tree. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2018-02-28 18:56:10 +03:00
if (!ret) {
struct inode *dir;
btrfs_release_path(path);
dir = read_one_inode(root, parent_id);
if (!dir) {
ret = -ENOENT;
kfree(name);
goto out;
}
ret = btrfs_unlink_inode(trans, BTRFS_I(dir),
Btrfs: fix log replay failure after unlink and link combination If we have a file with 2 (or more) hard links in the same directory, remove one of the hard links, create a new file (or link an existing file) in the same directory with the name of the removed hard link, and then finally fsync the new file, we end up with a log that fails to replay, causing a mount failure. Example: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/foo $ ln /mnt/testdir/foo /mnt/testdir/bar $ sync $ unlink /mnt/testdir/bar $ touch /mnt/testdir/bar $ xfs_io -c "fsync" /mnt/testdir/bar <power failure> $ mount /dev/sdb /mnt mount: mount(2) failed: /mnt: No such file or directory When replaying the log, for that example, we also see the following in dmesg/syslog: [71813.671307] BTRFS info (device dm-0): failed to delete reference to bar, inode 258 parent 257 [71813.674204] ------------[ cut here ]------------ [71813.675694] BTRFS: Transaction aborted (error -2) [71813.677236] WARNING: CPU: 1 PID: 13231 at fs/btrfs/inode.c:4128 __btrfs_unlink_inode+0x17b/0x355 [btrfs] [71813.679669] Modules linked in: btrfs xfs f2fs dm_flakey dm_mod dax ghash_clmulni_intel ppdev pcbc aesni_intel aes_x86_64 crypto_simd cryptd glue_helper evdev psmouse i2c_piix4 parport_pc i2c_core pcspkr sg serio_raw parport button sunrpc loop autofs4 ext4 crc16 mbcache jbd2 zstd_decompress zstd_compress xxhash raid10 raid456 async_raid6_recov async_memcpy async_pq async_xor async_tx xor raid6_pq libcrc32c crc32c_generic raid1 raid0 multipath linear md_mod ata_generic sd_mod virtio_scsi ata_piix libata virtio_pci virtio_ring crc32c_intel floppy virtio e1000 scsi_mod [last unloaded: btrfs] [71813.679669] CPU: 1 PID: 13231 Comm: mount Tainted: G W 4.15.0-rc9-btrfs-next-56+ #1 [71813.679669] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.10.2-0-g5f4c7b1-prebuilt.qemu-project.org 04/01/2014 [71813.679669] RIP: 0010:__btrfs_unlink_inode+0x17b/0x355 [btrfs] [71813.679669] RSP: 0018:ffffc90001cef738 EFLAGS: 00010286 [71813.679669] RAX: 0000000000000025 RBX: ffff880217ce4708 RCX: 0000000000000001 [71813.679669] RDX: 0000000000000000 RSI: ffffffff81c14bae RDI: 00000000ffffffff [71813.679669] RBP: ffffc90001cef7c0 R08: 0000000000000001 R09: 0000000000000001 [71813.679669] R10: ffffc90001cef5e0 R11: ffffffff8343f007 R12: ffff880217d474c8 [71813.679669] R13: 00000000fffffffe R14: ffff88021ccf1548 R15: 0000000000000101 [71813.679669] FS: 00007f7cee84c480(0000) GS:ffff88023fc80000(0000) knlGS:0000000000000000 [71813.679669] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [71813.679669] CR2: 00007f7cedc1abf9 CR3: 00000002354b4003 CR4: 00000000001606e0 [71813.679669] Call Trace: [71813.679669] btrfs_unlink_inode+0x17/0x41 [btrfs] [71813.679669] drop_one_dir_item+0xfa/0x131 [btrfs] [71813.679669] add_inode_ref+0x71e/0x851 [btrfs] [71813.679669] ? __lock_is_held+0x39/0x71 [71813.679669] ? replay_one_buffer+0x53/0x53a [btrfs] [71813.679669] replay_one_buffer+0x4a4/0x53a [btrfs] [71813.679669] ? rcu_read_unlock+0x3a/0x57 [71813.679669] ? __lock_is_held+0x39/0x71 [71813.679669] walk_up_log_tree+0x101/0x1d2 [btrfs] [71813.679669] walk_log_tree+0xad/0x188 [btrfs] [71813.679669] btrfs_recover_log_trees+0x1fa/0x31e [btrfs] [71813.679669] ? replay_one_extent+0x544/0x544 [btrfs] [71813.679669] open_ctree+0x1cf6/0x2209 [btrfs] [71813.679669] btrfs_mount_root+0x368/0x482 [btrfs] [71813.679669] ? trace_hardirqs_on_caller+0x14c/0x1a6 [71813.679669] ? __lockdep_init_map+0x176/0x1c2 [71813.679669] ? mount_fs+0x64/0x10b [71813.679669] mount_fs+0x64/0x10b [71813.679669] vfs_kern_mount+0x68/0xce [71813.679669] btrfs_mount+0x13e/0x772 [btrfs] [71813.679669] ? trace_hardirqs_on_caller+0x14c/0x1a6 [71813.679669] ? __lockdep_init_map+0x176/0x1c2 [71813.679669] ? mount_fs+0x64/0x10b [71813.679669] mount_fs+0x64/0x10b [71813.679669] vfs_kern_mount+0x68/0xce [71813.679669] do_mount+0x6e5/0x973 [71813.679669] ? memdup_user+0x3e/0x5c [71813.679669] SyS_mount+0x72/0x98 [71813.679669] entry_SYSCALL_64_fastpath+0x1e/0x8b [71813.679669] RIP: 0033:0x7f7cedf150ba [71813.679669] RSP: 002b:00007ffca71da688 EFLAGS: 00000206 [71813.679669] Code: 7f a0 e8 51 0c fd ff 48 8b 43 50 f0 0f ba a8 30 2c 00 00 02 72 17 41 83 fd fb 74 11 44 89 ee 48 c7 c7 7d 11 7f a0 e8 38 f5 8d e0 <0f> ff 44 89 e9 ba 20 10 00 00 eb 4d 48 8b 4d b0 48 8b 75 88 4c [71813.679669] ---[ end trace 83bd473fc5b4663b ]--- [71813.854764] BTRFS: error (device dm-0) in __btrfs_unlink_inode:4128: errno=-2 No such entry [71813.886994] BTRFS: error (device dm-0) in btrfs_replay_log:2307: errno=-2 No such entry (Failed to recover log tree) [71813.903357] BTRFS error (device dm-0): cleaner transaction attach returned -30 [71814.128078] BTRFS error (device dm-0): open_ctree failed This happens because the log has inode reference items for both inode 258 (the first file we created) and inode 259 (the second file created), and when processing the reference item for inode 258, we replace the corresponding item in the subvolume tree (which has two names, "foo" and "bar") witht he one in the log (which only has one name, "foo") without removing the corresponding dir index keys from the parent directory. Later, when processing the inode reference item for inode 259, which has a name of "bar" associated to it, we notice that dir index entries exist for that name and for a different inode, so we attempt to unlink that name, which fails because the inode reference item for inode 258 no longer has the name "bar" associated to it, making a call to btrfs_unlink_inode() fail with a -ENOENT error. Fix this by unlinking all the names in an inode reference item from a subvolume tree that are not present in the inode reference item found in the log tree, before overwriting it with the item from the log tree. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2018-02-28 18:56:10 +03:00
inode, name, namelen);
kfree(name);
iput(dir);
/*
* Whenever we need to check if a name exists or not, we
* check the subvolume tree. So after an unlink we must
* run delayed items, so that future checks for a name
* during log replay see that the name does not exists
* anymore.
*/
if (!ret)
ret = btrfs_run_delayed_items(trans);
Btrfs: fix log replay failure after unlink and link combination If we have a file with 2 (or more) hard links in the same directory, remove one of the hard links, create a new file (or link an existing file) in the same directory with the name of the removed hard link, and then finally fsync the new file, we end up with a log that fails to replay, causing a mount failure. Example: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/foo $ ln /mnt/testdir/foo /mnt/testdir/bar $ sync $ unlink /mnt/testdir/bar $ touch /mnt/testdir/bar $ xfs_io -c "fsync" /mnt/testdir/bar <power failure> $ mount /dev/sdb /mnt mount: mount(2) failed: /mnt: No such file or directory When replaying the log, for that example, we also see the following in dmesg/syslog: [71813.671307] BTRFS info (device dm-0): failed to delete reference to bar, inode 258 parent 257 [71813.674204] ------------[ cut here ]------------ [71813.675694] BTRFS: Transaction aborted (error -2) [71813.677236] WARNING: CPU: 1 PID: 13231 at fs/btrfs/inode.c:4128 __btrfs_unlink_inode+0x17b/0x355 [btrfs] [71813.679669] Modules linked in: btrfs xfs f2fs dm_flakey dm_mod dax ghash_clmulni_intel ppdev pcbc aesni_intel aes_x86_64 crypto_simd cryptd glue_helper evdev psmouse i2c_piix4 parport_pc i2c_core pcspkr sg serio_raw parport button sunrpc loop autofs4 ext4 crc16 mbcache jbd2 zstd_decompress zstd_compress xxhash raid10 raid456 async_raid6_recov async_memcpy async_pq async_xor async_tx xor raid6_pq libcrc32c crc32c_generic raid1 raid0 multipath linear md_mod ata_generic sd_mod virtio_scsi ata_piix libata virtio_pci virtio_ring crc32c_intel floppy virtio e1000 scsi_mod [last unloaded: btrfs] [71813.679669] CPU: 1 PID: 13231 Comm: mount Tainted: G W 4.15.0-rc9-btrfs-next-56+ #1 [71813.679669] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.10.2-0-g5f4c7b1-prebuilt.qemu-project.org 04/01/2014 [71813.679669] RIP: 0010:__btrfs_unlink_inode+0x17b/0x355 [btrfs] [71813.679669] RSP: 0018:ffffc90001cef738 EFLAGS: 00010286 [71813.679669] RAX: 0000000000000025 RBX: ffff880217ce4708 RCX: 0000000000000001 [71813.679669] RDX: 0000000000000000 RSI: ffffffff81c14bae RDI: 00000000ffffffff [71813.679669] RBP: ffffc90001cef7c0 R08: 0000000000000001 R09: 0000000000000001 [71813.679669] R10: ffffc90001cef5e0 R11: ffffffff8343f007 R12: ffff880217d474c8 [71813.679669] R13: 00000000fffffffe R14: ffff88021ccf1548 R15: 0000000000000101 [71813.679669] FS: 00007f7cee84c480(0000) GS:ffff88023fc80000(0000) knlGS:0000000000000000 [71813.679669] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [71813.679669] CR2: 00007f7cedc1abf9 CR3: 00000002354b4003 CR4: 00000000001606e0 [71813.679669] Call Trace: [71813.679669] btrfs_unlink_inode+0x17/0x41 [btrfs] [71813.679669] drop_one_dir_item+0xfa/0x131 [btrfs] [71813.679669] add_inode_ref+0x71e/0x851 [btrfs] [71813.679669] ? __lock_is_held+0x39/0x71 [71813.679669] ? replay_one_buffer+0x53/0x53a [btrfs] [71813.679669] replay_one_buffer+0x4a4/0x53a [btrfs] [71813.679669] ? rcu_read_unlock+0x3a/0x57 [71813.679669] ? __lock_is_held+0x39/0x71 [71813.679669] walk_up_log_tree+0x101/0x1d2 [btrfs] [71813.679669] walk_log_tree+0xad/0x188 [btrfs] [71813.679669] btrfs_recover_log_trees+0x1fa/0x31e [btrfs] [71813.679669] ? replay_one_extent+0x544/0x544 [btrfs] [71813.679669] open_ctree+0x1cf6/0x2209 [btrfs] [71813.679669] btrfs_mount_root+0x368/0x482 [btrfs] [71813.679669] ? trace_hardirqs_on_caller+0x14c/0x1a6 [71813.679669] ? __lockdep_init_map+0x176/0x1c2 [71813.679669] ? mount_fs+0x64/0x10b [71813.679669] mount_fs+0x64/0x10b [71813.679669] vfs_kern_mount+0x68/0xce [71813.679669] btrfs_mount+0x13e/0x772 [btrfs] [71813.679669] ? trace_hardirqs_on_caller+0x14c/0x1a6 [71813.679669] ? __lockdep_init_map+0x176/0x1c2 [71813.679669] ? mount_fs+0x64/0x10b [71813.679669] mount_fs+0x64/0x10b [71813.679669] vfs_kern_mount+0x68/0xce [71813.679669] do_mount+0x6e5/0x973 [71813.679669] ? memdup_user+0x3e/0x5c [71813.679669] SyS_mount+0x72/0x98 [71813.679669] entry_SYSCALL_64_fastpath+0x1e/0x8b [71813.679669] RIP: 0033:0x7f7cedf150ba [71813.679669] RSP: 002b:00007ffca71da688 EFLAGS: 00000206 [71813.679669] Code: 7f a0 e8 51 0c fd ff 48 8b 43 50 f0 0f ba a8 30 2c 00 00 02 72 17 41 83 fd fb 74 11 44 89 ee 48 c7 c7 7d 11 7f a0 e8 38 f5 8d e0 <0f> ff 44 89 e9 ba 20 10 00 00 eb 4d 48 8b 4d b0 48 8b 75 88 4c [71813.679669] ---[ end trace 83bd473fc5b4663b ]--- [71813.854764] BTRFS: error (device dm-0) in __btrfs_unlink_inode:4128: errno=-2 No such entry [71813.886994] BTRFS: error (device dm-0) in btrfs_replay_log:2307: errno=-2 No such entry (Failed to recover log tree) [71813.903357] BTRFS error (device dm-0): cleaner transaction attach returned -30 [71814.128078] BTRFS error (device dm-0): open_ctree failed This happens because the log has inode reference items for both inode 258 (the first file we created) and inode 259 (the second file created), and when processing the reference item for inode 258, we replace the corresponding item in the subvolume tree (which has two names, "foo" and "bar") witht he one in the log (which only has one name, "foo") without removing the corresponding dir index keys from the parent directory. Later, when processing the inode reference item for inode 259, which has a name of "bar" associated to it, we notice that dir index entries exist for that name and for a different inode, so we attempt to unlink that name, which fails because the inode reference item for inode 258 no longer has the name "bar" associated to it, making a call to btrfs_unlink_inode() fail with a -ENOENT error. Fix this by unlinking all the names in an inode reference item from a subvolume tree that are not present in the inode reference item found in the log tree, before overwriting it with the item from the log tree. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2018-02-28 18:56:10 +03:00
if (ret)
goto out;
goto again;
}
kfree(name);
ref_ptr += namelen;
if (key->type == BTRFS_INODE_EXTREF_KEY)
ref_ptr += sizeof(struct btrfs_inode_extref);
else
ref_ptr += sizeof(struct btrfs_inode_ref);
}
ret = 0;
out:
btrfs_release_path(path);
return ret;
}
Btrfs: fix mount failure after fsync due to hard link recreation If we end up with logging an inode reference item which has the same name but different index from the one we have persisted, we end up failing when replaying the log with an errno value of -EEXIST. The error comes from btrfs_add_link(), which is called from add_inode_ref(), when we are replaying an inode reference item. Example scenario where this happens: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ touch /mnt/foo $ ln /mnt/foo /mnt/bar $ sync # Rename the first hard link (foo) to a new name and rename the second # hard link (bar) to the old name of the first hard link (foo). $ mv /mnt/foo /mnt/qwerty $ mv /mnt/bar /mnt/foo # Create a new file, in the same parent directory, with the old name of # the second hard link (bar) and fsync this new file. # We do this instead of calling fsync on foo/qwerty because if we did # that the fsync resulted in a full transaction commit, not triggering # the problem. $ touch /mnt/bar $ xfs_io -c "fsync" /mnt/bar <power fail> $ mount /dev/sdb /mnt mount: mount /dev/sdb on /mnt failed: File exists So fix this by checking if a conflicting inode reference exists (same name, same parent but different index), removing it (and the associated dir index entries from the parent inode) if it exists, before attempting to add the new reference. A test case for fstests follows soon. CC: stable@vger.kernel.org # 4.4+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2018-07-20 12:59:06 +03:00
static int btrfs_inode_ref_exists(struct inode *inode, struct inode *dir,
const u8 ref_type, const char *name,
const int namelen)
{
struct btrfs_key key;
struct btrfs_path *path;
const u64 parent_id = btrfs_ino(BTRFS_I(dir));
int ret;
path = btrfs_alloc_path();
if (!path)
return -ENOMEM;
key.objectid = btrfs_ino(BTRFS_I(inode));
key.type = ref_type;
if (key.type == BTRFS_INODE_REF_KEY)
key.offset = parent_id;
else
key.offset = btrfs_extref_hash(parent_id, name, namelen);
ret = btrfs_search_slot(NULL, BTRFS_I(inode)->root, &key, path, 0, 0);
if (ret < 0)
goto out;
if (ret > 0) {
ret = 0;
goto out;
}
if (key.type == BTRFS_INODE_EXTREF_KEY)
ret = !!btrfs_find_name_in_ext_backref(path->nodes[0],
path->slots[0], parent_id, name, namelen);
Btrfs: fix mount failure after fsync due to hard link recreation If we end up with logging an inode reference item which has the same name but different index from the one we have persisted, we end up failing when replaying the log with an errno value of -EEXIST. The error comes from btrfs_add_link(), which is called from add_inode_ref(), when we are replaying an inode reference item. Example scenario where this happens: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ touch /mnt/foo $ ln /mnt/foo /mnt/bar $ sync # Rename the first hard link (foo) to a new name and rename the second # hard link (bar) to the old name of the first hard link (foo). $ mv /mnt/foo /mnt/qwerty $ mv /mnt/bar /mnt/foo # Create a new file, in the same parent directory, with the old name of # the second hard link (bar) and fsync this new file. # We do this instead of calling fsync on foo/qwerty because if we did # that the fsync resulted in a full transaction commit, not triggering # the problem. $ touch /mnt/bar $ xfs_io -c "fsync" /mnt/bar <power fail> $ mount /dev/sdb /mnt mount: mount /dev/sdb on /mnt failed: File exists So fix this by checking if a conflicting inode reference exists (same name, same parent but different index), removing it (and the associated dir index entries from the parent inode) if it exists, before attempting to add the new reference. A test case for fstests follows soon. CC: stable@vger.kernel.org # 4.4+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2018-07-20 12:59:06 +03:00
else
ret = !!btrfs_find_name_in_backref(path->nodes[0], path->slots[0],
name, namelen);
Btrfs: fix mount failure after fsync due to hard link recreation If we end up with logging an inode reference item which has the same name but different index from the one we have persisted, we end up failing when replaying the log with an errno value of -EEXIST. The error comes from btrfs_add_link(), which is called from add_inode_ref(), when we are replaying an inode reference item. Example scenario where this happens: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ touch /mnt/foo $ ln /mnt/foo /mnt/bar $ sync # Rename the first hard link (foo) to a new name and rename the second # hard link (bar) to the old name of the first hard link (foo). $ mv /mnt/foo /mnt/qwerty $ mv /mnt/bar /mnt/foo # Create a new file, in the same parent directory, with the old name of # the second hard link (bar) and fsync this new file. # We do this instead of calling fsync on foo/qwerty because if we did # that the fsync resulted in a full transaction commit, not triggering # the problem. $ touch /mnt/bar $ xfs_io -c "fsync" /mnt/bar <power fail> $ mount /dev/sdb /mnt mount: mount /dev/sdb on /mnt failed: File exists So fix this by checking if a conflicting inode reference exists (same name, same parent but different index), removing it (and the associated dir index entries from the parent inode) if it exists, before attempting to add the new reference. A test case for fstests follows soon. CC: stable@vger.kernel.org # 4.4+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2018-07-20 12:59:06 +03:00
out:
btrfs_free_path(path);
return ret;
}
static int add_link(struct btrfs_trans_handle *trans,
Btrfs: fix fsync after succession of renames of different files After a succession of rename operations of different files and fsyncing one of them, such that each file gets a new name that corresponds to an old name of another file, we can end up with a log that will cause a failure when attempted to replay at mount time (an EEXIST error). We currently have correct behaviour when such succession of renames involves only two files, but if there are more files involved, we end up not logging all the inodes that are needed, therefore resulting in a failure when attempting to replay the log. Example: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/fname1 $ touch /mnt/testdir/fname2 $ sync $ mv /mnt/testdir/fname1 /mnt/testdir/fname3 $ mv /mnt/testdir/fname2 /mnt/testdir/fname4 $ ln /mnt/testdir/fname3 /mnt/testdir/fname2 $ touch /mnt/testdir/fname1 $ xfs_io -c "fsync" /mnt/testdir/fname1 <power failure> $ mount /dev/sdb /mnt mount: mount /dev/sdb on /mnt failed: File exists So fix this by checking all inode dependencies when logging an inode. That is, if one logged inode A has a new name that matches the old name of some other inode B, check if inode B has a new name that matches the old name of some other inode C, and so on. This fix is implemented not by doing any recursive function calls but by using an iterative method using a linked list that is used in a first-in-first-out fashion. A test case for fstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-02-13 15:14:03 +03:00
struct inode *dir, struct inode *inode, const char *name,
int namelen, u64 ref_index)
{
struct btrfs_root *root = BTRFS_I(dir)->root;
Btrfs: fix fsync after succession of renames of different files After a succession of rename operations of different files and fsyncing one of them, such that each file gets a new name that corresponds to an old name of another file, we can end up with a log that will cause a failure when attempted to replay at mount time (an EEXIST error). We currently have correct behaviour when such succession of renames involves only two files, but if there are more files involved, we end up not logging all the inodes that are needed, therefore resulting in a failure when attempting to replay the log. Example: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/fname1 $ touch /mnt/testdir/fname2 $ sync $ mv /mnt/testdir/fname1 /mnt/testdir/fname3 $ mv /mnt/testdir/fname2 /mnt/testdir/fname4 $ ln /mnt/testdir/fname3 /mnt/testdir/fname2 $ touch /mnt/testdir/fname1 $ xfs_io -c "fsync" /mnt/testdir/fname1 <power failure> $ mount /dev/sdb /mnt mount: mount /dev/sdb on /mnt failed: File exists So fix this by checking all inode dependencies when logging an inode. That is, if one logged inode A has a new name that matches the old name of some other inode B, check if inode B has a new name that matches the old name of some other inode C, and so on. This fix is implemented not by doing any recursive function calls but by using an iterative method using a linked list that is used in a first-in-first-out fashion. A test case for fstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-02-13 15:14:03 +03:00
struct btrfs_dir_item *dir_item;
struct btrfs_key key;
struct btrfs_path *path;
struct inode *other_inode = NULL;
int ret;
path = btrfs_alloc_path();
if (!path)
return -ENOMEM;
dir_item = btrfs_lookup_dir_item(NULL, root, path,
btrfs_ino(BTRFS_I(dir)),
name, namelen, 0);
if (!dir_item) {
btrfs_release_path(path);
goto add_link;
} else if (IS_ERR(dir_item)) {
ret = PTR_ERR(dir_item);
goto out;
}
/*
* Our inode's dentry collides with the dentry of another inode which is
* in the log but not yet processed since it has a higher inode number.
* So delete that other dentry.
*/
btrfs_dir_item_key_to_cpu(path->nodes[0], dir_item, &key);
btrfs_release_path(path);
other_inode = read_one_inode(root, key.objectid);
if (!other_inode) {
ret = -ENOENT;
goto out;
}
ret = btrfs_unlink_inode(trans, BTRFS_I(dir), BTRFS_I(other_inode),
Btrfs: fix fsync after succession of renames of different files After a succession of rename operations of different files and fsyncing one of them, such that each file gets a new name that corresponds to an old name of another file, we can end up with a log that will cause a failure when attempted to replay at mount time (an EEXIST error). We currently have correct behaviour when such succession of renames involves only two files, but if there are more files involved, we end up not logging all the inodes that are needed, therefore resulting in a failure when attempting to replay the log. Example: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/fname1 $ touch /mnt/testdir/fname2 $ sync $ mv /mnt/testdir/fname1 /mnt/testdir/fname3 $ mv /mnt/testdir/fname2 /mnt/testdir/fname4 $ ln /mnt/testdir/fname3 /mnt/testdir/fname2 $ touch /mnt/testdir/fname1 $ xfs_io -c "fsync" /mnt/testdir/fname1 <power failure> $ mount /dev/sdb /mnt mount: mount /dev/sdb on /mnt failed: File exists So fix this by checking all inode dependencies when logging an inode. That is, if one logged inode A has a new name that matches the old name of some other inode B, check if inode B has a new name that matches the old name of some other inode C, and so on. This fix is implemented not by doing any recursive function calls but by using an iterative method using a linked list that is used in a first-in-first-out fashion. A test case for fstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-02-13 15:14:03 +03:00
name, namelen);
if (ret)
goto out;
/*
* If we dropped the link count to 0, bump it so that later the iput()
* on the inode will not free it. We will fixup the link count later.
*/
if (other_inode->i_nlink == 0)
inc_nlink(other_inode);
ret = btrfs_run_delayed_items(trans);
if (ret)
goto out;
add_link:
ret = btrfs_add_link(trans, BTRFS_I(dir), BTRFS_I(inode),
name, namelen, 0, ref_index);
out:
iput(other_inode);
btrfs_free_path(path);
return ret;
}
/*
* replay one inode back reference item found in the log tree.
* eb, slot and key refer to the buffer and key found in the log tree.
* root is the destination we are replaying into, and path is for temp
* use by this function. (it should be released on return).
*/
static noinline int add_inode_ref(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_root *log,
struct btrfs_path *path,
struct extent_buffer *eb, int slot,
struct btrfs_key *key)
{
struct inode *dir = NULL;
struct inode *inode = NULL;
unsigned long ref_ptr;
unsigned long ref_end;
char *name = NULL;
int namelen;
int ret;
int search_done = 0;
int log_ref_ver = 0;
u64 parent_objectid;
u64 inode_objectid;
u64 ref_index = 0;
int ref_struct_size;
ref_ptr = btrfs_item_ptr_offset(eb, slot);
ref_end = ref_ptr + btrfs_item_size(eb, slot);
if (key->type == BTRFS_INODE_EXTREF_KEY) {
struct btrfs_inode_extref *r;
ref_struct_size = sizeof(struct btrfs_inode_extref);
log_ref_ver = 1;
r = (struct btrfs_inode_extref *)ref_ptr;
parent_objectid = btrfs_inode_extref_parent(eb, r);
} else {
ref_struct_size = sizeof(struct btrfs_inode_ref);
parent_objectid = key->offset;
}
inode_objectid = key->objectid;
/*
* it is possible that we didn't log all the parent directories
* for a given inode. If we don't find the dir, just don't
* copy the back ref in. The link count fixup code will take
* care of the rest
*/
dir = read_one_inode(root, parent_objectid);
if (!dir) {
ret = -ENOENT;
goto out;
}
inode = read_one_inode(root, inode_objectid);
if (!inode) {
ret = -EIO;
goto out;
}
while (ref_ptr < ref_end) {
if (log_ref_ver) {
ret = extref_get_fields(eb, ref_ptr, &namelen, &name,
&ref_index, &parent_objectid);
/*
* parent object can change from one array
* item to another.
*/
if (!dir)
dir = read_one_inode(root, parent_objectid);
if (!dir) {
ret = -ENOENT;
goto out;
}
} else {
ret = ref_get_fields(eb, ref_ptr, &namelen, &name,
&ref_index);
}
if (ret)
goto out;
ret = inode_in_dir(root, path, btrfs_ino(BTRFS_I(dir)),
btrfs_ino(BTRFS_I(inode)), ref_index,
name, namelen);
if (ret < 0) {
goto out;
} else if (ret == 0) {
/*
* look for a conflicting back reference in the
* metadata. if we find one we have to unlink that name
* of the file before we add our new link. Later on, we
* overwrite any existing back reference, and we don't
* want to create dangling pointers in the directory.
*/
if (!search_done) {
ret = __add_inode_ref(trans, root, path, log,
BTRFS_I(dir),
BTRFS_I(inode),
inode_objectid,
parent_objectid,
ref_index, name, namelen,
&search_done);
if (ret) {
if (ret == 1)
ret = 0;
goto out;
}
}
Btrfs: fix mount failure after fsync due to hard link recreation If we end up with logging an inode reference item which has the same name but different index from the one we have persisted, we end up failing when replaying the log with an errno value of -EEXIST. The error comes from btrfs_add_link(), which is called from add_inode_ref(), when we are replaying an inode reference item. Example scenario where this happens: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ touch /mnt/foo $ ln /mnt/foo /mnt/bar $ sync # Rename the first hard link (foo) to a new name and rename the second # hard link (bar) to the old name of the first hard link (foo). $ mv /mnt/foo /mnt/qwerty $ mv /mnt/bar /mnt/foo # Create a new file, in the same parent directory, with the old name of # the second hard link (bar) and fsync this new file. # We do this instead of calling fsync on foo/qwerty because if we did # that the fsync resulted in a full transaction commit, not triggering # the problem. $ touch /mnt/bar $ xfs_io -c "fsync" /mnt/bar <power fail> $ mount /dev/sdb /mnt mount: mount /dev/sdb on /mnt failed: File exists So fix this by checking if a conflicting inode reference exists (same name, same parent but different index), removing it (and the associated dir index entries from the parent inode) if it exists, before attempting to add the new reference. A test case for fstests follows soon. CC: stable@vger.kernel.org # 4.4+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2018-07-20 12:59:06 +03:00
/*
* If a reference item already exists for this inode
* with the same parent and name, but different index,
* drop it and the corresponding directory index entries
* from the parent before adding the new reference item
* and dir index entries, otherwise we would fail with
* -EEXIST returned from btrfs_add_link() below.
*/
ret = btrfs_inode_ref_exists(inode, dir, key->type,
name, namelen);
if (ret > 0) {
ret = btrfs_unlink_inode(trans,
Btrfs: fix mount failure after fsync due to hard link recreation If we end up with logging an inode reference item which has the same name but different index from the one we have persisted, we end up failing when replaying the log with an errno value of -EEXIST. The error comes from btrfs_add_link(), which is called from add_inode_ref(), when we are replaying an inode reference item. Example scenario where this happens: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ touch /mnt/foo $ ln /mnt/foo /mnt/bar $ sync # Rename the first hard link (foo) to a new name and rename the second # hard link (bar) to the old name of the first hard link (foo). $ mv /mnt/foo /mnt/qwerty $ mv /mnt/bar /mnt/foo # Create a new file, in the same parent directory, with the old name of # the second hard link (bar) and fsync this new file. # We do this instead of calling fsync on foo/qwerty because if we did # that the fsync resulted in a full transaction commit, not triggering # the problem. $ touch /mnt/bar $ xfs_io -c "fsync" /mnt/bar <power fail> $ mount /dev/sdb /mnt mount: mount /dev/sdb on /mnt failed: File exists So fix this by checking if a conflicting inode reference exists (same name, same parent but different index), removing it (and the associated dir index entries from the parent inode) if it exists, before attempting to add the new reference. A test case for fstests follows soon. CC: stable@vger.kernel.org # 4.4+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2018-07-20 12:59:06 +03:00
BTRFS_I(dir),
BTRFS_I(inode),
name, namelen);
/*
* If we dropped the link count to 0, bump it so
* that later the iput() on the inode will not
* free it. We will fixup the link count later.
*/
if (!ret && inode->i_nlink == 0)
inc_nlink(inode);
/*
* Whenever we need to check if a name exists or
* not, we check the subvolume tree. So after an
* unlink we must run delayed items, so that future
* checks for a name during log replay see that the
* name does not exists anymore.
*/
if (!ret)
ret = btrfs_run_delayed_items(trans);
Btrfs: fix mount failure after fsync due to hard link recreation If we end up with logging an inode reference item which has the same name but different index from the one we have persisted, we end up failing when replaying the log with an errno value of -EEXIST. The error comes from btrfs_add_link(), which is called from add_inode_ref(), when we are replaying an inode reference item. Example scenario where this happens: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ touch /mnt/foo $ ln /mnt/foo /mnt/bar $ sync # Rename the first hard link (foo) to a new name and rename the second # hard link (bar) to the old name of the first hard link (foo). $ mv /mnt/foo /mnt/qwerty $ mv /mnt/bar /mnt/foo # Create a new file, in the same parent directory, with the old name of # the second hard link (bar) and fsync this new file. # We do this instead of calling fsync on foo/qwerty because if we did # that the fsync resulted in a full transaction commit, not triggering # the problem. $ touch /mnt/bar $ xfs_io -c "fsync" /mnt/bar <power fail> $ mount /dev/sdb /mnt mount: mount /dev/sdb on /mnt failed: File exists So fix this by checking if a conflicting inode reference exists (same name, same parent but different index), removing it (and the associated dir index entries from the parent inode) if it exists, before attempting to add the new reference. A test case for fstests follows soon. CC: stable@vger.kernel.org # 4.4+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2018-07-20 12:59:06 +03:00
}
if (ret < 0)
goto out;
/* insert our name */
ret = add_link(trans, dir, inode, name, namelen,
Btrfs: fix fsync after succession of renames of different files After a succession of rename operations of different files and fsyncing one of them, such that each file gets a new name that corresponds to an old name of another file, we can end up with a log that will cause a failure when attempted to replay at mount time (an EEXIST error). We currently have correct behaviour when such succession of renames involves only two files, but if there are more files involved, we end up not logging all the inodes that are needed, therefore resulting in a failure when attempting to replay the log. Example: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/fname1 $ touch /mnt/testdir/fname2 $ sync $ mv /mnt/testdir/fname1 /mnt/testdir/fname3 $ mv /mnt/testdir/fname2 /mnt/testdir/fname4 $ ln /mnt/testdir/fname3 /mnt/testdir/fname2 $ touch /mnt/testdir/fname1 $ xfs_io -c "fsync" /mnt/testdir/fname1 <power failure> $ mount /dev/sdb /mnt mount: mount /dev/sdb on /mnt failed: File exists So fix this by checking all inode dependencies when logging an inode. That is, if one logged inode A has a new name that matches the old name of some other inode B, check if inode B has a new name that matches the old name of some other inode C, and so on. This fix is implemented not by doing any recursive function calls but by using an iterative method using a linked list that is used in a first-in-first-out fashion. A test case for fstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-02-13 15:14:03 +03:00
ref_index);
if (ret)
goto out;
ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
if (ret)
goto out;
}
/* Else, ret == 1, we already have a perfect match, we're done. */
ref_ptr = (unsigned long)(ref_ptr + ref_struct_size) + namelen;
kfree(name);
name = NULL;
if (log_ref_ver) {
iput(dir);
dir = NULL;
}
}
Btrfs: fix log replay failure after unlink and link combination If we have a file with 2 (or more) hard links in the same directory, remove one of the hard links, create a new file (or link an existing file) in the same directory with the name of the removed hard link, and then finally fsync the new file, we end up with a log that fails to replay, causing a mount failure. Example: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/foo $ ln /mnt/testdir/foo /mnt/testdir/bar $ sync $ unlink /mnt/testdir/bar $ touch /mnt/testdir/bar $ xfs_io -c "fsync" /mnt/testdir/bar <power failure> $ mount /dev/sdb /mnt mount: mount(2) failed: /mnt: No such file or directory When replaying the log, for that example, we also see the following in dmesg/syslog: [71813.671307] BTRFS info (device dm-0): failed to delete reference to bar, inode 258 parent 257 [71813.674204] ------------[ cut here ]------------ [71813.675694] BTRFS: Transaction aborted (error -2) [71813.677236] WARNING: CPU: 1 PID: 13231 at fs/btrfs/inode.c:4128 __btrfs_unlink_inode+0x17b/0x355 [btrfs] [71813.679669] Modules linked in: btrfs xfs f2fs dm_flakey dm_mod dax ghash_clmulni_intel ppdev pcbc aesni_intel aes_x86_64 crypto_simd cryptd glue_helper evdev psmouse i2c_piix4 parport_pc i2c_core pcspkr sg serio_raw parport button sunrpc loop autofs4 ext4 crc16 mbcache jbd2 zstd_decompress zstd_compress xxhash raid10 raid456 async_raid6_recov async_memcpy async_pq async_xor async_tx xor raid6_pq libcrc32c crc32c_generic raid1 raid0 multipath linear md_mod ata_generic sd_mod virtio_scsi ata_piix libata virtio_pci virtio_ring crc32c_intel floppy virtio e1000 scsi_mod [last unloaded: btrfs] [71813.679669] CPU: 1 PID: 13231 Comm: mount Tainted: G W 4.15.0-rc9-btrfs-next-56+ #1 [71813.679669] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.10.2-0-g5f4c7b1-prebuilt.qemu-project.org 04/01/2014 [71813.679669] RIP: 0010:__btrfs_unlink_inode+0x17b/0x355 [btrfs] [71813.679669] RSP: 0018:ffffc90001cef738 EFLAGS: 00010286 [71813.679669] RAX: 0000000000000025 RBX: ffff880217ce4708 RCX: 0000000000000001 [71813.679669] RDX: 0000000000000000 RSI: ffffffff81c14bae RDI: 00000000ffffffff [71813.679669] RBP: ffffc90001cef7c0 R08: 0000000000000001 R09: 0000000000000001 [71813.679669] R10: ffffc90001cef5e0 R11: ffffffff8343f007 R12: ffff880217d474c8 [71813.679669] R13: 00000000fffffffe R14: ffff88021ccf1548 R15: 0000000000000101 [71813.679669] FS: 00007f7cee84c480(0000) GS:ffff88023fc80000(0000) knlGS:0000000000000000 [71813.679669] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [71813.679669] CR2: 00007f7cedc1abf9 CR3: 00000002354b4003 CR4: 00000000001606e0 [71813.679669] Call Trace: [71813.679669] btrfs_unlink_inode+0x17/0x41 [btrfs] [71813.679669] drop_one_dir_item+0xfa/0x131 [btrfs] [71813.679669] add_inode_ref+0x71e/0x851 [btrfs] [71813.679669] ? __lock_is_held+0x39/0x71 [71813.679669] ? replay_one_buffer+0x53/0x53a [btrfs] [71813.679669] replay_one_buffer+0x4a4/0x53a [btrfs] [71813.679669] ? rcu_read_unlock+0x3a/0x57 [71813.679669] ? __lock_is_held+0x39/0x71 [71813.679669] walk_up_log_tree+0x101/0x1d2 [btrfs] [71813.679669] walk_log_tree+0xad/0x188 [btrfs] [71813.679669] btrfs_recover_log_trees+0x1fa/0x31e [btrfs] [71813.679669] ? replay_one_extent+0x544/0x544 [btrfs] [71813.679669] open_ctree+0x1cf6/0x2209 [btrfs] [71813.679669] btrfs_mount_root+0x368/0x482 [btrfs] [71813.679669] ? trace_hardirqs_on_caller+0x14c/0x1a6 [71813.679669] ? __lockdep_init_map+0x176/0x1c2 [71813.679669] ? mount_fs+0x64/0x10b [71813.679669] mount_fs+0x64/0x10b [71813.679669] vfs_kern_mount+0x68/0xce [71813.679669] btrfs_mount+0x13e/0x772 [btrfs] [71813.679669] ? trace_hardirqs_on_caller+0x14c/0x1a6 [71813.679669] ? __lockdep_init_map+0x176/0x1c2 [71813.679669] ? mount_fs+0x64/0x10b [71813.679669] mount_fs+0x64/0x10b [71813.679669] vfs_kern_mount+0x68/0xce [71813.679669] do_mount+0x6e5/0x973 [71813.679669] ? memdup_user+0x3e/0x5c [71813.679669] SyS_mount+0x72/0x98 [71813.679669] entry_SYSCALL_64_fastpath+0x1e/0x8b [71813.679669] RIP: 0033:0x7f7cedf150ba [71813.679669] RSP: 002b:00007ffca71da688 EFLAGS: 00000206 [71813.679669] Code: 7f a0 e8 51 0c fd ff 48 8b 43 50 f0 0f ba a8 30 2c 00 00 02 72 17 41 83 fd fb 74 11 44 89 ee 48 c7 c7 7d 11 7f a0 e8 38 f5 8d e0 <0f> ff 44 89 e9 ba 20 10 00 00 eb 4d 48 8b 4d b0 48 8b 75 88 4c [71813.679669] ---[ end trace 83bd473fc5b4663b ]--- [71813.854764] BTRFS: error (device dm-0) in __btrfs_unlink_inode:4128: errno=-2 No such entry [71813.886994] BTRFS: error (device dm-0) in btrfs_replay_log:2307: errno=-2 No such entry (Failed to recover log tree) [71813.903357] BTRFS error (device dm-0): cleaner transaction attach returned -30 [71814.128078] BTRFS error (device dm-0): open_ctree failed This happens because the log has inode reference items for both inode 258 (the first file we created) and inode 259 (the second file created), and when processing the reference item for inode 258, we replace the corresponding item in the subvolume tree (which has two names, "foo" and "bar") witht he one in the log (which only has one name, "foo") without removing the corresponding dir index keys from the parent directory. Later, when processing the inode reference item for inode 259, which has a name of "bar" associated to it, we notice that dir index entries exist for that name and for a different inode, so we attempt to unlink that name, which fails because the inode reference item for inode 258 no longer has the name "bar" associated to it, making a call to btrfs_unlink_inode() fail with a -ENOENT error. Fix this by unlinking all the names in an inode reference item from a subvolume tree that are not present in the inode reference item found in the log tree, before overwriting it with the item from the log tree. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2018-02-28 18:56:10 +03:00
/*
* Before we overwrite the inode reference item in the subvolume tree
* with the item from the log tree, we must unlink all names from the
* parent directory that are in the subvolume's tree inode reference
* item, otherwise we end up with an inconsistent subvolume tree where
* dir index entries exist for a name but there is no inode reference
* item with the same name.
*/
ret = unlink_old_inode_refs(trans, root, path, BTRFS_I(inode), eb, slot,
key);
if (ret)
goto out;
/* finally write the back reference in the inode */
ret = overwrite_item(trans, root, path, eb, slot, key);
out:
btrfs_release_path(path);
kfree(name);
iput(dir);
iput(inode);
return ret;
}
static int count_inode_extrefs(struct btrfs_root *root,
struct btrfs_inode *inode, struct btrfs_path *path)
{
int ret = 0;
int name_len;
unsigned int nlink = 0;
u32 item_size;
u32 cur_offset = 0;
u64 inode_objectid = btrfs_ino(inode);
u64 offset = 0;
unsigned long ptr;
struct btrfs_inode_extref *extref;
struct extent_buffer *leaf;
while (1) {
ret = btrfs_find_one_extref(root, inode_objectid, offset, path,
&extref, &offset);
if (ret)
break;
leaf = path->nodes[0];
item_size = btrfs_item_size(leaf, path->slots[0]);
ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
Btrfs: fix fsync when extend references are added to an inode If we added an extended reference to an inode and fsync'ed it, the log replay code would make our inode have an incorrect link count, which was lower then the expected/correct count. This resulted in stale directory index entries after deleting some of the hard links, and any access to the dangling directory entries resulted in -ESTALE errors because the entries pointed to inode items that don't exist anymore. This is easy to reproduce with the test case I made for xfstests, and the bulk of that test is: _scratch_mkfs "-O extref" >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create a test file with 3001 hard links. This number is large enough to # make btrfs start using extrefs at some point even if the fs has the maximum # possible leaf/node size (64Kb). echo "hello world" > $SCRATCH_MNT/foo for i in `seq 1 3000`; do ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link_`printf "%04d" $i` done # Make sure all metadata and data are durably persisted. sync # Add one more link to the inode that ends up being a btrfs extref and fsync # the inode. ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link_3001 $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foo # Simulate a crash/power loss. This makes sure the next mount # will see an fsync log and will replay that log. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # Now after the fsync log replay btrfs left our inode with a wrong link count N, # which was smaller than the correct link count M (N < M). # So after removing N hard links, the remaining M - N directory entries were # still visible to user space but it was impossible to do anything with them # because they pointed to an inode that didn't exist anymore. This resulted in # stale file handle errors (-ESTALE) when accessing those dentries for example. # # So remove all hard links except the first one and then attempt to read the # file, to verify we don't get an -ESTALE error when accessing the inodel # # The btrfs fsck tool also detected the incorrect inode link count and it # reported an error message like the following: # # root 5 inode 257 errors 2001, no inode item, link count wrong # unresolved ref dir 256 index 2978 namelen 13 name foo_link_2976 filetype 1 errors 4, no inode ref # # The fstests framework automatically calls fsck after a test is run, so we # don't need to call fsck explicitly here. rm -f $SCRATCH_MNT/foo_link_* cat $SCRATCH_MNT/foo status=0 exit So make sure an fsync always flushes the delayed inode item, so that the fsync log contains it (needed in order to trigger the link count fixup code) and fix the extref counting function, which always return -ENOENT to its caller (and made it assume there were always 0 extrefs). This issue has been present since the introduction of the extrefs feature (2012). A test case for xfstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-01-13 19:40:04 +03:00
cur_offset = 0;
while (cur_offset < item_size) {
extref = (struct btrfs_inode_extref *) (ptr + cur_offset);
name_len = btrfs_inode_extref_name_len(leaf, extref);
nlink++;
cur_offset += name_len + sizeof(*extref);
}
offset++;
btrfs_release_path(path);
}
btrfs_release_path(path);
Btrfs: fix fsync when extend references are added to an inode If we added an extended reference to an inode and fsync'ed it, the log replay code would make our inode have an incorrect link count, which was lower then the expected/correct count. This resulted in stale directory index entries after deleting some of the hard links, and any access to the dangling directory entries resulted in -ESTALE errors because the entries pointed to inode items that don't exist anymore. This is easy to reproduce with the test case I made for xfstests, and the bulk of that test is: _scratch_mkfs "-O extref" >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create a test file with 3001 hard links. This number is large enough to # make btrfs start using extrefs at some point even if the fs has the maximum # possible leaf/node size (64Kb). echo "hello world" > $SCRATCH_MNT/foo for i in `seq 1 3000`; do ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link_`printf "%04d" $i` done # Make sure all metadata and data are durably persisted. sync # Add one more link to the inode that ends up being a btrfs extref and fsync # the inode. ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link_3001 $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foo # Simulate a crash/power loss. This makes sure the next mount # will see an fsync log and will replay that log. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # Now after the fsync log replay btrfs left our inode with a wrong link count N, # which was smaller than the correct link count M (N < M). # So after removing N hard links, the remaining M - N directory entries were # still visible to user space but it was impossible to do anything with them # because they pointed to an inode that didn't exist anymore. This resulted in # stale file handle errors (-ESTALE) when accessing those dentries for example. # # So remove all hard links except the first one and then attempt to read the # file, to verify we don't get an -ESTALE error when accessing the inodel # # The btrfs fsck tool also detected the incorrect inode link count and it # reported an error message like the following: # # root 5 inode 257 errors 2001, no inode item, link count wrong # unresolved ref dir 256 index 2978 namelen 13 name foo_link_2976 filetype 1 errors 4, no inode ref # # The fstests framework automatically calls fsck after a test is run, so we # don't need to call fsck explicitly here. rm -f $SCRATCH_MNT/foo_link_* cat $SCRATCH_MNT/foo status=0 exit So make sure an fsync always flushes the delayed inode item, so that the fsync log contains it (needed in order to trigger the link count fixup code) and fix the extref counting function, which always return -ENOENT to its caller (and made it assume there were always 0 extrefs). This issue has been present since the introduction of the extrefs feature (2012). A test case for xfstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-01-13 19:40:04 +03:00
if (ret < 0 && ret != -ENOENT)
return ret;
return nlink;
}
static int count_inode_refs(struct btrfs_root *root,
struct btrfs_inode *inode, struct btrfs_path *path)
{
int ret;
struct btrfs_key key;
unsigned int nlink = 0;
unsigned long ptr;
unsigned long ptr_end;
int name_len;
u64 ino = btrfs_ino(inode);
key.objectid = ino;
key.type = BTRFS_INODE_REF_KEY;
key.offset = (u64)-1;
while (1) {
ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
if (ret < 0)
break;
if (ret > 0) {
if (path->slots[0] == 0)
break;
path->slots[0]--;
}
process_slot:
btrfs_item_key_to_cpu(path->nodes[0], &key,
path->slots[0]);
if (key.objectid != ino ||
key.type != BTRFS_INODE_REF_KEY)
break;
ptr = btrfs_item_ptr_offset(path->nodes[0], path->slots[0]);
ptr_end = ptr + btrfs_item_size(path->nodes[0],
path->slots[0]);
while (ptr < ptr_end) {
struct btrfs_inode_ref *ref;
ref = (struct btrfs_inode_ref *)ptr;
name_len = btrfs_inode_ref_name_len(path->nodes[0],
ref);
ptr = (unsigned long)(ref + 1) + name_len;
nlink++;
}
if (key.offset == 0)
break;
if (path->slots[0] > 0) {
path->slots[0]--;
goto process_slot;
}
key.offset--;
btrfs_release_path(path);
}
btrfs_release_path(path);
return nlink;
}
/*
* There are a few corners where the link count of the file can't
* be properly maintained during replay. So, instead of adding
* lots of complexity to the log code, we just scan the backrefs
* for any file that has been through replay.
*
* The scan will update the link count on the inode to reflect the
* number of back refs found. If it goes down to zero, the iput
* will free the inode.
*/
static noinline int fixup_inode_link_count(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct inode *inode)
{
struct btrfs_path *path;
int ret;
u64 nlink = 0;
u64 ino = btrfs_ino(BTRFS_I(inode));
path = btrfs_alloc_path();
if (!path)
return -ENOMEM;
ret = count_inode_refs(root, BTRFS_I(inode), path);
if (ret < 0)
goto out;
nlink = ret;
ret = count_inode_extrefs(root, BTRFS_I(inode), path);
if (ret < 0)
goto out;
nlink += ret;
ret = 0;
if (nlink != inode->i_nlink) {
set_nlink(inode, nlink);
ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
if (ret)
goto out;
}
BTRFS_I(inode)->index_cnt = (u64)-1;
if (inode->i_nlink == 0) {
if (S_ISDIR(inode->i_mode)) {
ret = replay_dir_deletes(trans, root, NULL, path,
ino, 1);
if (ret)
goto out;
}
ret = btrfs_insert_orphan_item(trans, root, ino);
if (ret == -EEXIST)
ret = 0;
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 17:24:20 +03:00
}
out:
btrfs_free_path(path);
return ret;
}
static noinline int fixup_inode_link_counts(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_path *path)
{
int ret;
struct btrfs_key key;
struct inode *inode;
key.objectid = BTRFS_TREE_LOG_FIXUP_OBJECTID;
key.type = BTRFS_ORPHAN_ITEM_KEY;
key.offset = (u64)-1;
while (1) {
ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
if (ret < 0)
break;
if (ret == 1) {
ret = 0;
if (path->slots[0] == 0)
break;
path->slots[0]--;
}
btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
if (key.objectid != BTRFS_TREE_LOG_FIXUP_OBJECTID ||
key.type != BTRFS_ORPHAN_ITEM_KEY)
break;
ret = btrfs_del_item(trans, root, path);
if (ret)
break;
btrfs_release_path(path);
inode = read_one_inode(root, key.offset);
if (!inode) {
ret = -EIO;
break;
}
ret = fixup_inode_link_count(trans, root, inode);
iput(inode);
if (ret)
break;
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 17:24:20 +03:00
/*
* fixup on a directory may create new entries,
* make sure we always look for the highset possible
* offset
*/
key.offset = (u64)-1;
}
btrfs_release_path(path);
return ret;
}
/*
* record a given inode in the fixup dir so we can check its link
* count when replay is done. The link count is incremented here
* so the inode won't go away until we check it
*/
static noinline int link_to_fixup_dir(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_path *path,
u64 objectid)
{
struct btrfs_key key;
int ret = 0;
struct inode *inode;
inode = read_one_inode(root, objectid);
if (!inode)
return -EIO;
key.objectid = BTRFS_TREE_LOG_FIXUP_OBJECTID;
key.type = BTRFS_ORPHAN_ITEM_KEY;
key.offset = objectid;
ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
btrfs_release_path(path);
if (ret == 0) {
if (!inode->i_nlink)
set_nlink(inode, 1);
else
inc_nlink(inode);
ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
} else if (ret == -EEXIST) {
ret = 0;
}
iput(inode);
return ret;
}
/*
* when replaying the log for a directory, we only insert names
* for inodes that actually exist. This means an fsync on a directory
* does not implicitly fsync all the new files in it
*/
static noinline int insert_one_name(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
u64 dirid, u64 index,
char *name, int name_len,
struct btrfs_key *location)
{
struct inode *inode;
struct inode *dir;
int ret;
inode = read_one_inode(root, location->objectid);
if (!inode)
return -ENOENT;
dir = read_one_inode(root, dirid);
if (!dir) {
iput(inode);
return -EIO;
}
ret = btrfs_add_link(trans, BTRFS_I(dir), BTRFS_I(inode), name,
name_len, 1, index);
/* FIXME, put inode into FIXUP list */
iput(inode);
iput(dir);
return ret;
}
btrfs: only copy dir index keys when logging a directory Currently, when logging a directory, we copy both dir items and dir index items from the fs/subvolume tree to the log tree. Both items have exactly the same data (same struct btrfs_dir_item), the difference lies in the key values, where a dir index key contains the index number of a directory entry while the dir item key does not, as it's used for doing fast lookups of an entry by name, while the former is used for sorting entries when listing a directory. We can exploit that and log only the dir index items, since they contain all the information needed to correctly add, replace and delete directory entries when replaying a log tree. Logging only the dir index items is also backward and forward compatible: an unpatched kernel (without this change) can correctly replay a log tree generated by a patched kernel (with this patch), and a patched kernel can correctly replay a log tree generated by an unpatched kernel. The backward compatibility is ensured because: 1) For inserting a new dentry: a dentry is only inserted when we find a new dir index key - we can only insert if we know the dir index offset, which is encoded in the dir index key's offset; 2) For deleting dentries: during log replay, before adding or replacing dentries, we first replay dentry deletions. Whenever we find a dir item key or a dir index key in the subvolume/fs tree that is not logged in a range for which the log tree is authoritative, we do the unlink of the dentry, which removes both the existing dir item key and the dir index key. Therefore logging just dir index keys is enough to ensure dentry deletions are correctly replayed; 3) For dentry replacements: they work when we log only dir index keys and this is mostly due to a combination of 1) and 2). If we replace a dentry with name "foobar" to point from inode A to inode B, then we know the dir index key for the new dentry is different from the old one, as it has an index number (key offset) larger than the old one. This results in replaying a deletion, through replay_dir_deletes(), that causes the old dentry to be removed, both the dir item key and the dir index key, as mentioned at 2). Then when processing the new dir index key, we add the new dentry, adding both a new dir item key and a new index key pointing to inode B, as stated in 1). The forward compatibility, the ability for a patched kernel to replay a log created by an older, unpatched kernel, comes from the changes required for making sure we are able to replay a log that only contains dir index keys - we simply ignore every dir item key we find. So modify directory logging to log only dir index items, and modify the log replay process to ignore dir item keys, from log trees created by an unpatched kernel, and process only with dir index keys. This reduces the amount of logged metadata by about half, and therefore the time spent logging or fsyncing large directories (less CPU time and less IO). The following test script was used to measure this change: #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=1000000 NUM_FILE_DELETES=10000 mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The tests were run on a physical machine, with a non-debug kernel (Debian's default kernel config), for different values of $NUM_NEW_FILES and $NUM_FILE_DELETES, and the results were the following: ** Before patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 8412 ms after adding 1000000 files dir fsync took 500 ms after deleting 10000 files ** After patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 4252 ms after adding 1000000 files (-49.5%) dir fsync took 269 ms after deleting 10000 files (-46.2%) ** Before patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 745 ms after adding 100000 files dir fsync took 59 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 404 ms after adding 100000 files (-45.8%) dir fsync took 31 ms after deleting 1000 files (-47.5%) ** Before patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 67 ms after adding 10000 files dir fsync took 9 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 36 ms after adding 10000 files (-46.3%) dir fsync took 5 ms after deleting 1000 files (-44.4%) ** Before patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 9 ms after adding 1000 files dir fsync took 4 ms after deleting 100 files ** After patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 7 ms after adding 1000 files (-22.2%) dir fsync took 3 ms after deleting 100 files (-25.0%) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-10-25 19:31:53 +03:00
static int delete_conflicting_dir_entry(struct btrfs_trans_handle *trans,
struct btrfs_inode *dir,
struct btrfs_path *path,
struct btrfs_dir_item *dst_di,
const struct btrfs_key *log_key,
u8 log_type,
bool exists)
{
struct btrfs_key found_key;
btrfs_dir_item_key_to_cpu(path->nodes[0], dst_di, &found_key);
/* The existing dentry points to the same inode, don't delete it. */
if (found_key.objectid == log_key->objectid &&
found_key.type == log_key->type &&
found_key.offset == log_key->offset &&
btrfs_dir_type(path->nodes[0], dst_di) == log_type)
return 1;
/*
* Don't drop the conflicting directory entry if the inode for the new
* entry doesn't exist.
*/
if (!exists)
return 0;
return drop_one_dir_item(trans, path, dir, dst_di);
}
/*
* take a single entry in a log directory item and replay it into
* the subvolume.
*
* if a conflicting item exists in the subdirectory already,
* the inode it points to is unlinked and put into the link count
* fix up tree.
*
* If a name from the log points to a file or directory that does
* not exist in the FS, it is skipped. fsyncs on directories
* do not force down inodes inside that directory, just changes to the
* names or unlinks in a directory.
Btrfs: fix stale directory entries after fsync log replay We have another case where after an fsync log replay we get an inode with a wrong link count (smaller than it should be) and a number of directory entries greater than its link count. This happens when we add a new link hard link to our inode A and then we fsync some other inode B that has the side effect of logging the parent directory inode too. In this case at log replay time we add the new hard link to our inode (the item with key BTRFS_INODE_REF_KEY) when processing the parent directory but we never adjust the link count of our inode A. As a result we get stale dir entries for our inode A that can never be deleted and therefore it makes it impossible to remove the parent directory (as its i_size can never decrease back to 0). A simple reproducer for fstests that triggers this issue: seq=`basename $0` seqres=$RESULT_DIR/$seq echo "QA output created by $seq" tmp=/tmp/$$ status=1 # failure is the default! trap "_cleanup; exit \$status" 0 1 2 3 15 _cleanup() { _cleanup_flakey rm -f $tmp.* } # get standard environment, filters and checks . ./common/rc . ./common/filter . ./common/dmflakey # real QA test starts here _need_to_be_root _supported_fs generic _supported_os Linux _require_scratch _require_dm_flakey _require_metadata_journaling $SCRATCH_DEV rm -f $seqres.full _scratch_mkfs >>$seqres.full 2>&1 _init_flakey _mount_flakey # Create our test directory and files. mkdir $SCRATCH_MNT/testdir touch $SCRATCH_MNT/testdir/foo touch $SCRATCH_MNT/testdir/bar # Make sure everything done so far is durably persisted. sync # Create one hard link for file foo and another one for file bar. After # that fsync only the file bar. ln $SCRATCH_MNT/testdir/bar $SCRATCH_MNT/testdir/bar_link ln $SCRATCH_MNT/testdir/foo $SCRATCH_MNT/testdir/foo_link $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/testdir/bar # Silently drop all writes on scratch device to simulate power failure. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again and mount the fs to trigger log/journal replay. _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # Now verify both our files have a link count of 2. echo "Link count for file foo: $(stat --format=%h $SCRATCH_MNT/testdir/foo)" echo "Link count for file bar: $(stat --format=%h $SCRATCH_MNT/testdir/bar)" # We should be able to remove all the links of our files in testdir, and # after that the parent directory should become empty and therefore # possible to remove it. rm -f $SCRATCH_MNT/testdir/* rmdir $SCRATCH_MNT/testdir _unmount_flakey # The fstests framework will call fsck against our filesystem which will verify # that all metadata is in a consistent state. status=0 exit The test fails with: -Link count for file foo: 2 +Link count for file foo: 1 Link count for file bar: 2 +rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/testdir/foo_link': Stale file handle +rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/testdir': Directory not empty (...) _check_btrfs_filesystem: filesystem on /dev/sdc is inconsistent And fsck's output: (...) checking fs roots root 5 inode 258 errors 2001, no inode item, link count wrong unresolved ref dir 257 index 5 namelen 8 name foo_link filetype 1 errors 4, no inode ref Checking filesystem on /dev/sdc (...) So fix this by marking inodes for link count fixup at log replay time whenever a directory entry is replayed if the entry was created in the transaction where the fsync was made and if it points to a non-directory inode. This isn't a new problem/regression, the issue exists for a long time, possibly since the log tree feature was added (2008). Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-07-16 01:26:43 +03:00
*
* Returns < 0 on error, 0 if the name wasn't replayed (dentry points to a
* non-existing inode) and 1 if the name was replayed.
*/
static noinline int replay_one_name(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_path *path,
struct extent_buffer *eb,
struct btrfs_dir_item *di,
struct btrfs_key *key)
{
char *name;
int name_len;
btrfs: only copy dir index keys when logging a directory Currently, when logging a directory, we copy both dir items and dir index items from the fs/subvolume tree to the log tree. Both items have exactly the same data (same struct btrfs_dir_item), the difference lies in the key values, where a dir index key contains the index number of a directory entry while the dir item key does not, as it's used for doing fast lookups of an entry by name, while the former is used for sorting entries when listing a directory. We can exploit that and log only the dir index items, since they contain all the information needed to correctly add, replace and delete directory entries when replaying a log tree. Logging only the dir index items is also backward and forward compatible: an unpatched kernel (without this change) can correctly replay a log tree generated by a patched kernel (with this patch), and a patched kernel can correctly replay a log tree generated by an unpatched kernel. The backward compatibility is ensured because: 1) For inserting a new dentry: a dentry is only inserted when we find a new dir index key - we can only insert if we know the dir index offset, which is encoded in the dir index key's offset; 2) For deleting dentries: during log replay, before adding or replacing dentries, we first replay dentry deletions. Whenever we find a dir item key or a dir index key in the subvolume/fs tree that is not logged in a range for which the log tree is authoritative, we do the unlink of the dentry, which removes both the existing dir item key and the dir index key. Therefore logging just dir index keys is enough to ensure dentry deletions are correctly replayed; 3) For dentry replacements: they work when we log only dir index keys and this is mostly due to a combination of 1) and 2). If we replace a dentry with name "foobar" to point from inode A to inode B, then we know the dir index key for the new dentry is different from the old one, as it has an index number (key offset) larger than the old one. This results in replaying a deletion, through replay_dir_deletes(), that causes the old dentry to be removed, both the dir item key and the dir index key, as mentioned at 2). Then when processing the new dir index key, we add the new dentry, adding both a new dir item key and a new index key pointing to inode B, as stated in 1). The forward compatibility, the ability for a patched kernel to replay a log created by an older, unpatched kernel, comes from the changes required for making sure we are able to replay a log that only contains dir index keys - we simply ignore every dir item key we find. So modify directory logging to log only dir index items, and modify the log replay process to ignore dir item keys, from log trees created by an unpatched kernel, and process only with dir index keys. This reduces the amount of logged metadata by about half, and therefore the time spent logging or fsyncing large directories (less CPU time and less IO). The following test script was used to measure this change: #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=1000000 NUM_FILE_DELETES=10000 mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The tests were run on a physical machine, with a non-debug kernel (Debian's default kernel config), for different values of $NUM_NEW_FILES and $NUM_FILE_DELETES, and the results were the following: ** Before patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 8412 ms after adding 1000000 files dir fsync took 500 ms after deleting 10000 files ** After patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 4252 ms after adding 1000000 files (-49.5%) dir fsync took 269 ms after deleting 10000 files (-46.2%) ** Before patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 745 ms after adding 100000 files dir fsync took 59 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 404 ms after adding 100000 files (-45.8%) dir fsync took 31 ms after deleting 1000 files (-47.5%) ** Before patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 67 ms after adding 10000 files dir fsync took 9 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 36 ms after adding 10000 files (-46.3%) dir fsync took 5 ms after deleting 1000 files (-44.4%) ** Before patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 9 ms after adding 1000 files dir fsync took 4 ms after deleting 100 files ** After patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 7 ms after adding 1000 files (-22.2%) dir fsync took 3 ms after deleting 100 files (-25.0%) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-10-25 19:31:53 +03:00
struct btrfs_dir_item *dir_dst_di;
struct btrfs_dir_item *index_dst_di;
bool dir_dst_matches = false;
bool index_dst_matches = false;
struct btrfs_key log_key;
btrfs: only copy dir index keys when logging a directory Currently, when logging a directory, we copy both dir items and dir index items from the fs/subvolume tree to the log tree. Both items have exactly the same data (same struct btrfs_dir_item), the difference lies in the key values, where a dir index key contains the index number of a directory entry while the dir item key does not, as it's used for doing fast lookups of an entry by name, while the former is used for sorting entries when listing a directory. We can exploit that and log only the dir index items, since they contain all the information needed to correctly add, replace and delete directory entries when replaying a log tree. Logging only the dir index items is also backward and forward compatible: an unpatched kernel (without this change) can correctly replay a log tree generated by a patched kernel (with this patch), and a patched kernel can correctly replay a log tree generated by an unpatched kernel. The backward compatibility is ensured because: 1) For inserting a new dentry: a dentry is only inserted when we find a new dir index key - we can only insert if we know the dir index offset, which is encoded in the dir index key's offset; 2) For deleting dentries: during log replay, before adding or replacing dentries, we first replay dentry deletions. Whenever we find a dir item key or a dir index key in the subvolume/fs tree that is not logged in a range for which the log tree is authoritative, we do the unlink of the dentry, which removes both the existing dir item key and the dir index key. Therefore logging just dir index keys is enough to ensure dentry deletions are correctly replayed; 3) For dentry replacements: they work when we log only dir index keys and this is mostly due to a combination of 1) and 2). If we replace a dentry with name "foobar" to point from inode A to inode B, then we know the dir index key for the new dentry is different from the old one, as it has an index number (key offset) larger than the old one. This results in replaying a deletion, through replay_dir_deletes(), that causes the old dentry to be removed, both the dir item key and the dir index key, as mentioned at 2). Then when processing the new dir index key, we add the new dentry, adding both a new dir item key and a new index key pointing to inode B, as stated in 1). The forward compatibility, the ability for a patched kernel to replay a log created by an older, unpatched kernel, comes from the changes required for making sure we are able to replay a log that only contains dir index keys - we simply ignore every dir item key we find. So modify directory logging to log only dir index items, and modify the log replay process to ignore dir item keys, from log trees created by an unpatched kernel, and process only with dir index keys. This reduces the amount of logged metadata by about half, and therefore the time spent logging or fsyncing large directories (less CPU time and less IO). The following test script was used to measure this change: #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=1000000 NUM_FILE_DELETES=10000 mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The tests were run on a physical machine, with a non-debug kernel (Debian's default kernel config), for different values of $NUM_NEW_FILES and $NUM_FILE_DELETES, and the results were the following: ** Before patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 8412 ms after adding 1000000 files dir fsync took 500 ms after deleting 10000 files ** After patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 4252 ms after adding 1000000 files (-49.5%) dir fsync took 269 ms after deleting 10000 files (-46.2%) ** Before patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 745 ms after adding 100000 files dir fsync took 59 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 404 ms after adding 100000 files (-45.8%) dir fsync took 31 ms after deleting 1000 files (-47.5%) ** Before patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 67 ms after adding 10000 files dir fsync took 9 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 36 ms after adding 10000 files (-46.3%) dir fsync took 5 ms after deleting 1000 files (-44.4%) ** Before patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 9 ms after adding 1000 files dir fsync took 4 ms after deleting 100 files ** After patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 7 ms after adding 1000 files (-22.2%) dir fsync took 3 ms after deleting 100 files (-25.0%) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-10-25 19:31:53 +03:00
struct btrfs_key search_key;
struct inode *dir;
u8 log_type;
bool exists;
int ret;
btrfs: only copy dir index keys when logging a directory Currently, when logging a directory, we copy both dir items and dir index items from the fs/subvolume tree to the log tree. Both items have exactly the same data (same struct btrfs_dir_item), the difference lies in the key values, where a dir index key contains the index number of a directory entry while the dir item key does not, as it's used for doing fast lookups of an entry by name, while the former is used for sorting entries when listing a directory. We can exploit that and log only the dir index items, since they contain all the information needed to correctly add, replace and delete directory entries when replaying a log tree. Logging only the dir index items is also backward and forward compatible: an unpatched kernel (without this change) can correctly replay a log tree generated by a patched kernel (with this patch), and a patched kernel can correctly replay a log tree generated by an unpatched kernel. The backward compatibility is ensured because: 1) For inserting a new dentry: a dentry is only inserted when we find a new dir index key - we can only insert if we know the dir index offset, which is encoded in the dir index key's offset; 2) For deleting dentries: during log replay, before adding or replacing dentries, we first replay dentry deletions. Whenever we find a dir item key or a dir index key in the subvolume/fs tree that is not logged in a range for which the log tree is authoritative, we do the unlink of the dentry, which removes both the existing dir item key and the dir index key. Therefore logging just dir index keys is enough to ensure dentry deletions are correctly replayed; 3) For dentry replacements: they work when we log only dir index keys and this is mostly due to a combination of 1) and 2). If we replace a dentry with name "foobar" to point from inode A to inode B, then we know the dir index key for the new dentry is different from the old one, as it has an index number (key offset) larger than the old one. This results in replaying a deletion, through replay_dir_deletes(), that causes the old dentry to be removed, both the dir item key and the dir index key, as mentioned at 2). Then when processing the new dir index key, we add the new dentry, adding both a new dir item key and a new index key pointing to inode B, as stated in 1). The forward compatibility, the ability for a patched kernel to replay a log created by an older, unpatched kernel, comes from the changes required for making sure we are able to replay a log that only contains dir index keys - we simply ignore every dir item key we find. So modify directory logging to log only dir index items, and modify the log replay process to ignore dir item keys, from log trees created by an unpatched kernel, and process only with dir index keys. This reduces the amount of logged metadata by about half, and therefore the time spent logging or fsyncing large directories (less CPU time and less IO). The following test script was used to measure this change: #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=1000000 NUM_FILE_DELETES=10000 mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The tests were run on a physical machine, with a non-debug kernel (Debian's default kernel config), for different values of $NUM_NEW_FILES and $NUM_FILE_DELETES, and the results were the following: ** Before patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 8412 ms after adding 1000000 files dir fsync took 500 ms after deleting 10000 files ** After patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 4252 ms after adding 1000000 files (-49.5%) dir fsync took 269 ms after deleting 10000 files (-46.2%) ** Before patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 745 ms after adding 100000 files dir fsync took 59 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 404 ms after adding 100000 files (-45.8%) dir fsync took 31 ms after deleting 1000 files (-47.5%) ** Before patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 67 ms after adding 10000 files dir fsync took 9 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 36 ms after adding 10000 files (-46.3%) dir fsync took 5 ms after deleting 1000 files (-44.4%) ** Before patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 9 ms after adding 1000 files dir fsync took 4 ms after deleting 100 files ** After patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 7 ms after adding 1000 files (-22.2%) dir fsync took 3 ms after deleting 100 files (-25.0%) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-10-25 19:31:53 +03:00
bool update_size = true;
Btrfs: fix stale directory entries after fsync log replay We have another case where after an fsync log replay we get an inode with a wrong link count (smaller than it should be) and a number of directory entries greater than its link count. This happens when we add a new link hard link to our inode A and then we fsync some other inode B that has the side effect of logging the parent directory inode too. In this case at log replay time we add the new hard link to our inode (the item with key BTRFS_INODE_REF_KEY) when processing the parent directory but we never adjust the link count of our inode A. As a result we get stale dir entries for our inode A that can never be deleted and therefore it makes it impossible to remove the parent directory (as its i_size can never decrease back to 0). A simple reproducer for fstests that triggers this issue: seq=`basename $0` seqres=$RESULT_DIR/$seq echo "QA output created by $seq" tmp=/tmp/$$ status=1 # failure is the default! trap "_cleanup; exit \$status" 0 1 2 3 15 _cleanup() { _cleanup_flakey rm -f $tmp.* } # get standard environment, filters and checks . ./common/rc . ./common/filter . ./common/dmflakey # real QA test starts here _need_to_be_root _supported_fs generic _supported_os Linux _require_scratch _require_dm_flakey _require_metadata_journaling $SCRATCH_DEV rm -f $seqres.full _scratch_mkfs >>$seqres.full 2>&1 _init_flakey _mount_flakey # Create our test directory and files. mkdir $SCRATCH_MNT/testdir touch $SCRATCH_MNT/testdir/foo touch $SCRATCH_MNT/testdir/bar # Make sure everything done so far is durably persisted. sync # Create one hard link for file foo and another one for file bar. After # that fsync only the file bar. ln $SCRATCH_MNT/testdir/bar $SCRATCH_MNT/testdir/bar_link ln $SCRATCH_MNT/testdir/foo $SCRATCH_MNT/testdir/foo_link $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/testdir/bar # Silently drop all writes on scratch device to simulate power failure. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again and mount the fs to trigger log/journal replay. _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # Now verify both our files have a link count of 2. echo "Link count for file foo: $(stat --format=%h $SCRATCH_MNT/testdir/foo)" echo "Link count for file bar: $(stat --format=%h $SCRATCH_MNT/testdir/bar)" # We should be able to remove all the links of our files in testdir, and # after that the parent directory should become empty and therefore # possible to remove it. rm -f $SCRATCH_MNT/testdir/* rmdir $SCRATCH_MNT/testdir _unmount_flakey # The fstests framework will call fsck against our filesystem which will verify # that all metadata is in a consistent state. status=0 exit The test fails with: -Link count for file foo: 2 +Link count for file foo: 1 Link count for file bar: 2 +rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/testdir/foo_link': Stale file handle +rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/testdir': Directory not empty (...) _check_btrfs_filesystem: filesystem on /dev/sdc is inconsistent And fsck's output: (...) checking fs roots root 5 inode 258 errors 2001, no inode item, link count wrong unresolved ref dir 257 index 5 namelen 8 name foo_link filetype 1 errors 4, no inode ref Checking filesystem on /dev/sdc (...) So fix this by marking inodes for link count fixup at log replay time whenever a directory entry is replayed if the entry was created in the transaction where the fsync was made and if it points to a non-directory inode. This isn't a new problem/regression, the issue exists for a long time, possibly since the log tree feature was added (2008). Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-07-16 01:26:43 +03:00
bool name_added = false;
dir = read_one_inode(root, key->objectid);
if (!dir)
return -EIO;
name_len = btrfs_dir_name_len(eb, di);
name = kmalloc(name_len, GFP_NOFS);
if (!name) {
ret = -ENOMEM;
goto out;
}
log_type = btrfs_dir_type(eb, di);
read_extent_buffer(eb, name, (unsigned long)(di + 1),
name_len);
btrfs_dir_item_key_to_cpu(eb, di, &log_key);
ret = btrfs_lookup_inode(trans, root, path, &log_key, 0);
btrfs_release_path(path);
if (ret < 0)
goto out;
exists = (ret == 0);
ret = 0;
btrfs: only copy dir index keys when logging a directory Currently, when logging a directory, we copy both dir items and dir index items from the fs/subvolume tree to the log tree. Both items have exactly the same data (same struct btrfs_dir_item), the difference lies in the key values, where a dir index key contains the index number of a directory entry while the dir item key does not, as it's used for doing fast lookups of an entry by name, while the former is used for sorting entries when listing a directory. We can exploit that and log only the dir index items, since they contain all the information needed to correctly add, replace and delete directory entries when replaying a log tree. Logging only the dir index items is also backward and forward compatible: an unpatched kernel (without this change) can correctly replay a log tree generated by a patched kernel (with this patch), and a patched kernel can correctly replay a log tree generated by an unpatched kernel. The backward compatibility is ensured because: 1) For inserting a new dentry: a dentry is only inserted when we find a new dir index key - we can only insert if we know the dir index offset, which is encoded in the dir index key's offset; 2) For deleting dentries: during log replay, before adding or replacing dentries, we first replay dentry deletions. Whenever we find a dir item key or a dir index key in the subvolume/fs tree that is not logged in a range for which the log tree is authoritative, we do the unlink of the dentry, which removes both the existing dir item key and the dir index key. Therefore logging just dir index keys is enough to ensure dentry deletions are correctly replayed; 3) For dentry replacements: they work when we log only dir index keys and this is mostly due to a combination of 1) and 2). If we replace a dentry with name "foobar" to point from inode A to inode B, then we know the dir index key for the new dentry is different from the old one, as it has an index number (key offset) larger than the old one. This results in replaying a deletion, through replay_dir_deletes(), that causes the old dentry to be removed, both the dir item key and the dir index key, as mentioned at 2). Then when processing the new dir index key, we add the new dentry, adding both a new dir item key and a new index key pointing to inode B, as stated in 1). The forward compatibility, the ability for a patched kernel to replay a log created by an older, unpatched kernel, comes from the changes required for making sure we are able to replay a log that only contains dir index keys - we simply ignore every dir item key we find. So modify directory logging to log only dir index items, and modify the log replay process to ignore dir item keys, from log trees created by an unpatched kernel, and process only with dir index keys. This reduces the amount of logged metadata by about half, and therefore the time spent logging or fsyncing large directories (less CPU time and less IO). The following test script was used to measure this change: #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=1000000 NUM_FILE_DELETES=10000 mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The tests were run on a physical machine, with a non-debug kernel (Debian's default kernel config), for different values of $NUM_NEW_FILES and $NUM_FILE_DELETES, and the results were the following: ** Before patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 8412 ms after adding 1000000 files dir fsync took 500 ms after deleting 10000 files ** After patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 4252 ms after adding 1000000 files (-49.5%) dir fsync took 269 ms after deleting 10000 files (-46.2%) ** Before patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 745 ms after adding 100000 files dir fsync took 59 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 404 ms after adding 100000 files (-45.8%) dir fsync took 31 ms after deleting 1000 files (-47.5%) ** Before patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 67 ms after adding 10000 files dir fsync took 9 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 36 ms after adding 10000 files (-46.3%) dir fsync took 5 ms after deleting 1000 files (-44.4%) ** Before patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 9 ms after adding 1000 files dir fsync took 4 ms after deleting 100 files ** After patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 7 ms after adding 1000 files (-22.2%) dir fsync took 3 ms after deleting 100 files (-25.0%) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-10-25 19:31:53 +03:00
dir_dst_di = btrfs_lookup_dir_item(trans, root, path, key->objectid,
name, name_len, 1);
if (IS_ERR(dir_dst_di)) {
ret = PTR_ERR(dir_dst_di);
goto out;
btrfs: only copy dir index keys when logging a directory Currently, when logging a directory, we copy both dir items and dir index items from the fs/subvolume tree to the log tree. Both items have exactly the same data (same struct btrfs_dir_item), the difference lies in the key values, where a dir index key contains the index number of a directory entry while the dir item key does not, as it's used for doing fast lookups of an entry by name, while the former is used for sorting entries when listing a directory. We can exploit that and log only the dir index items, since they contain all the information needed to correctly add, replace and delete directory entries when replaying a log tree. Logging only the dir index items is also backward and forward compatible: an unpatched kernel (without this change) can correctly replay a log tree generated by a patched kernel (with this patch), and a patched kernel can correctly replay a log tree generated by an unpatched kernel. The backward compatibility is ensured because: 1) For inserting a new dentry: a dentry is only inserted when we find a new dir index key - we can only insert if we know the dir index offset, which is encoded in the dir index key's offset; 2) For deleting dentries: during log replay, before adding or replacing dentries, we first replay dentry deletions. Whenever we find a dir item key or a dir index key in the subvolume/fs tree that is not logged in a range for which the log tree is authoritative, we do the unlink of the dentry, which removes both the existing dir item key and the dir index key. Therefore logging just dir index keys is enough to ensure dentry deletions are correctly replayed; 3) For dentry replacements: they work when we log only dir index keys and this is mostly due to a combination of 1) and 2). If we replace a dentry with name "foobar" to point from inode A to inode B, then we know the dir index key for the new dentry is different from the old one, as it has an index number (key offset) larger than the old one. This results in replaying a deletion, through replay_dir_deletes(), that causes the old dentry to be removed, both the dir item key and the dir index key, as mentioned at 2). Then when processing the new dir index key, we add the new dentry, adding both a new dir item key and a new index key pointing to inode B, as stated in 1). The forward compatibility, the ability for a patched kernel to replay a log created by an older, unpatched kernel, comes from the changes required for making sure we are able to replay a log that only contains dir index keys - we simply ignore every dir item key we find. So modify directory logging to log only dir index items, and modify the log replay process to ignore dir item keys, from log trees created by an unpatched kernel, and process only with dir index keys. This reduces the amount of logged metadata by about half, and therefore the time spent logging or fsyncing large directories (less CPU time and less IO). The following test script was used to measure this change: #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=1000000 NUM_FILE_DELETES=10000 mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The tests were run on a physical machine, with a non-debug kernel (Debian's default kernel config), for different values of $NUM_NEW_FILES and $NUM_FILE_DELETES, and the results were the following: ** Before patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 8412 ms after adding 1000000 files dir fsync took 500 ms after deleting 10000 files ** After patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 4252 ms after adding 1000000 files (-49.5%) dir fsync took 269 ms after deleting 10000 files (-46.2%) ** Before patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 745 ms after adding 100000 files dir fsync took 59 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 404 ms after adding 100000 files (-45.8%) dir fsync took 31 ms after deleting 1000 files (-47.5%) ** Before patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 67 ms after adding 10000 files dir fsync took 9 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 36 ms after adding 10000 files (-46.3%) dir fsync took 5 ms after deleting 1000 files (-44.4%) ** Before patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 9 ms after adding 1000 files dir fsync took 4 ms after deleting 100 files ** After patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 7 ms after adding 1000 files (-22.2%) dir fsync took 3 ms after deleting 100 files (-25.0%) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-10-25 19:31:53 +03:00
} else if (dir_dst_di) {
ret = delete_conflicting_dir_entry(trans, BTRFS_I(dir), path,
dir_dst_di, &log_key, log_type,
exists);
if (ret < 0)
goto out;
dir_dst_matches = (ret == 1);
}
btrfs: only copy dir index keys when logging a directory Currently, when logging a directory, we copy both dir items and dir index items from the fs/subvolume tree to the log tree. Both items have exactly the same data (same struct btrfs_dir_item), the difference lies in the key values, where a dir index key contains the index number of a directory entry while the dir item key does not, as it's used for doing fast lookups of an entry by name, while the former is used for sorting entries when listing a directory. We can exploit that and log only the dir index items, since they contain all the information needed to correctly add, replace and delete directory entries when replaying a log tree. Logging only the dir index items is also backward and forward compatible: an unpatched kernel (without this change) can correctly replay a log tree generated by a patched kernel (with this patch), and a patched kernel can correctly replay a log tree generated by an unpatched kernel. The backward compatibility is ensured because: 1) For inserting a new dentry: a dentry is only inserted when we find a new dir index key - we can only insert if we know the dir index offset, which is encoded in the dir index key's offset; 2) For deleting dentries: during log replay, before adding or replacing dentries, we first replay dentry deletions. Whenever we find a dir item key or a dir index key in the subvolume/fs tree that is not logged in a range for which the log tree is authoritative, we do the unlink of the dentry, which removes both the existing dir item key and the dir index key. Therefore logging just dir index keys is enough to ensure dentry deletions are correctly replayed; 3) For dentry replacements: they work when we log only dir index keys and this is mostly due to a combination of 1) and 2). If we replace a dentry with name "foobar" to point from inode A to inode B, then we know the dir index key for the new dentry is different from the old one, as it has an index number (key offset) larger than the old one. This results in replaying a deletion, through replay_dir_deletes(), that causes the old dentry to be removed, both the dir item key and the dir index key, as mentioned at 2). Then when processing the new dir index key, we add the new dentry, adding both a new dir item key and a new index key pointing to inode B, as stated in 1). The forward compatibility, the ability for a patched kernel to replay a log created by an older, unpatched kernel, comes from the changes required for making sure we are able to replay a log that only contains dir index keys - we simply ignore every dir item key we find. So modify directory logging to log only dir index items, and modify the log replay process to ignore dir item keys, from log trees created by an unpatched kernel, and process only with dir index keys. This reduces the amount of logged metadata by about half, and therefore the time spent logging or fsyncing large directories (less CPU time and less IO). The following test script was used to measure this change: #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=1000000 NUM_FILE_DELETES=10000 mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The tests were run on a physical machine, with a non-debug kernel (Debian's default kernel config), for different values of $NUM_NEW_FILES and $NUM_FILE_DELETES, and the results were the following: ** Before patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 8412 ms after adding 1000000 files dir fsync took 500 ms after deleting 10000 files ** After patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 4252 ms after adding 1000000 files (-49.5%) dir fsync took 269 ms after deleting 10000 files (-46.2%) ** Before patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 745 ms after adding 100000 files dir fsync took 59 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 404 ms after adding 100000 files (-45.8%) dir fsync took 31 ms after deleting 1000 files (-47.5%) ** Before patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 67 ms after adding 10000 files dir fsync took 9 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 36 ms after adding 10000 files (-46.3%) dir fsync took 5 ms after deleting 1000 files (-44.4%) ** Before patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 9 ms after adding 1000 files dir fsync took 4 ms after deleting 100 files ** After patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 7 ms after adding 1000 files (-22.2%) dir fsync took 3 ms after deleting 100 files (-25.0%) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-10-25 19:31:53 +03:00
btrfs_release_path(path);
index_dst_di = btrfs_lookup_dir_index_item(trans, root, path,
key->objectid, key->offset,
name, name_len, 1);
if (IS_ERR(index_dst_di)) {
ret = PTR_ERR(index_dst_di);
goto out;
btrfs: only copy dir index keys when logging a directory Currently, when logging a directory, we copy both dir items and dir index items from the fs/subvolume tree to the log tree. Both items have exactly the same data (same struct btrfs_dir_item), the difference lies in the key values, where a dir index key contains the index number of a directory entry while the dir item key does not, as it's used for doing fast lookups of an entry by name, while the former is used for sorting entries when listing a directory. We can exploit that and log only the dir index items, since they contain all the information needed to correctly add, replace and delete directory entries when replaying a log tree. Logging only the dir index items is also backward and forward compatible: an unpatched kernel (without this change) can correctly replay a log tree generated by a patched kernel (with this patch), and a patched kernel can correctly replay a log tree generated by an unpatched kernel. The backward compatibility is ensured because: 1) For inserting a new dentry: a dentry is only inserted when we find a new dir index key - we can only insert if we know the dir index offset, which is encoded in the dir index key's offset; 2) For deleting dentries: during log replay, before adding or replacing dentries, we first replay dentry deletions. Whenever we find a dir item key or a dir index key in the subvolume/fs tree that is not logged in a range for which the log tree is authoritative, we do the unlink of the dentry, which removes both the existing dir item key and the dir index key. Therefore logging just dir index keys is enough to ensure dentry deletions are correctly replayed; 3) For dentry replacements: they work when we log only dir index keys and this is mostly due to a combination of 1) and 2). If we replace a dentry with name "foobar" to point from inode A to inode B, then we know the dir index key for the new dentry is different from the old one, as it has an index number (key offset) larger than the old one. This results in replaying a deletion, through replay_dir_deletes(), that causes the old dentry to be removed, both the dir item key and the dir index key, as mentioned at 2). Then when processing the new dir index key, we add the new dentry, adding both a new dir item key and a new index key pointing to inode B, as stated in 1). The forward compatibility, the ability for a patched kernel to replay a log created by an older, unpatched kernel, comes from the changes required for making sure we are able to replay a log that only contains dir index keys - we simply ignore every dir item key we find. So modify directory logging to log only dir index items, and modify the log replay process to ignore dir item keys, from log trees created by an unpatched kernel, and process only with dir index keys. This reduces the amount of logged metadata by about half, and therefore the time spent logging or fsyncing large directories (less CPU time and less IO). The following test script was used to measure this change: #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=1000000 NUM_FILE_DELETES=10000 mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The tests were run on a physical machine, with a non-debug kernel (Debian's default kernel config), for different values of $NUM_NEW_FILES and $NUM_FILE_DELETES, and the results were the following: ** Before patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 8412 ms after adding 1000000 files dir fsync took 500 ms after deleting 10000 files ** After patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 4252 ms after adding 1000000 files (-49.5%) dir fsync took 269 ms after deleting 10000 files (-46.2%) ** Before patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 745 ms after adding 100000 files dir fsync took 59 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 404 ms after adding 100000 files (-45.8%) dir fsync took 31 ms after deleting 1000 files (-47.5%) ** Before patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 67 ms after adding 10000 files dir fsync took 9 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 36 ms after adding 10000 files (-46.3%) dir fsync took 5 ms after deleting 1000 files (-44.4%) ** Before patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 9 ms after adding 1000 files dir fsync took 4 ms after deleting 100 files ** After patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 7 ms after adding 1000 files (-22.2%) dir fsync took 3 ms after deleting 100 files (-25.0%) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-10-25 19:31:53 +03:00
} else if (index_dst_di) {
ret = delete_conflicting_dir_entry(trans, BTRFS_I(dir), path,
index_dst_di, &log_key,
log_type, exists);
if (ret < 0)
goto out;
btrfs: only copy dir index keys when logging a directory Currently, when logging a directory, we copy both dir items and dir index items from the fs/subvolume tree to the log tree. Both items have exactly the same data (same struct btrfs_dir_item), the difference lies in the key values, where a dir index key contains the index number of a directory entry while the dir item key does not, as it's used for doing fast lookups of an entry by name, while the former is used for sorting entries when listing a directory. We can exploit that and log only the dir index items, since they contain all the information needed to correctly add, replace and delete directory entries when replaying a log tree. Logging only the dir index items is also backward and forward compatible: an unpatched kernel (without this change) can correctly replay a log tree generated by a patched kernel (with this patch), and a patched kernel can correctly replay a log tree generated by an unpatched kernel. The backward compatibility is ensured because: 1) For inserting a new dentry: a dentry is only inserted when we find a new dir index key - we can only insert if we know the dir index offset, which is encoded in the dir index key's offset; 2) For deleting dentries: during log replay, before adding or replacing dentries, we first replay dentry deletions. Whenever we find a dir item key or a dir index key in the subvolume/fs tree that is not logged in a range for which the log tree is authoritative, we do the unlink of the dentry, which removes both the existing dir item key and the dir index key. Therefore logging just dir index keys is enough to ensure dentry deletions are correctly replayed; 3) For dentry replacements: they work when we log only dir index keys and this is mostly due to a combination of 1) and 2). If we replace a dentry with name "foobar" to point from inode A to inode B, then we know the dir index key for the new dentry is different from the old one, as it has an index number (key offset) larger than the old one. This results in replaying a deletion, through replay_dir_deletes(), that causes the old dentry to be removed, both the dir item key and the dir index key, as mentioned at 2). Then when processing the new dir index key, we add the new dentry, adding both a new dir item key and a new index key pointing to inode B, as stated in 1). The forward compatibility, the ability for a patched kernel to replay a log created by an older, unpatched kernel, comes from the changes required for making sure we are able to replay a log that only contains dir index keys - we simply ignore every dir item key we find. So modify directory logging to log only dir index items, and modify the log replay process to ignore dir item keys, from log trees created by an unpatched kernel, and process only with dir index keys. This reduces the amount of logged metadata by about half, and therefore the time spent logging or fsyncing large directories (less CPU time and less IO). The following test script was used to measure this change: #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=1000000 NUM_FILE_DELETES=10000 mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The tests were run on a physical machine, with a non-debug kernel (Debian's default kernel config), for different values of $NUM_NEW_FILES and $NUM_FILE_DELETES, and the results were the following: ** Before patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 8412 ms after adding 1000000 files dir fsync took 500 ms after deleting 10000 files ** After patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 4252 ms after adding 1000000 files (-49.5%) dir fsync took 269 ms after deleting 10000 files (-46.2%) ** Before patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 745 ms after adding 100000 files dir fsync took 59 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 404 ms after adding 100000 files (-45.8%) dir fsync took 31 ms after deleting 1000 files (-47.5%) ** Before patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 67 ms after adding 10000 files dir fsync took 9 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 36 ms after adding 10000 files (-46.3%) dir fsync took 5 ms after deleting 1000 files (-44.4%) ** Before patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 9 ms after adding 1000 files dir fsync took 4 ms after deleting 100 files ** After patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 7 ms after adding 1000 files (-22.2%) dir fsync took 3 ms after deleting 100 files (-25.0%) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-10-25 19:31:53 +03:00
index_dst_matches = (ret == 1);
}
btrfs: only copy dir index keys when logging a directory Currently, when logging a directory, we copy both dir items and dir index items from the fs/subvolume tree to the log tree. Both items have exactly the same data (same struct btrfs_dir_item), the difference lies in the key values, where a dir index key contains the index number of a directory entry while the dir item key does not, as it's used for doing fast lookups of an entry by name, while the former is used for sorting entries when listing a directory. We can exploit that and log only the dir index items, since they contain all the information needed to correctly add, replace and delete directory entries when replaying a log tree. Logging only the dir index items is also backward and forward compatible: an unpatched kernel (without this change) can correctly replay a log tree generated by a patched kernel (with this patch), and a patched kernel can correctly replay a log tree generated by an unpatched kernel. The backward compatibility is ensured because: 1) For inserting a new dentry: a dentry is only inserted when we find a new dir index key - we can only insert if we know the dir index offset, which is encoded in the dir index key's offset; 2) For deleting dentries: during log replay, before adding or replacing dentries, we first replay dentry deletions. Whenever we find a dir item key or a dir index key in the subvolume/fs tree that is not logged in a range for which the log tree is authoritative, we do the unlink of the dentry, which removes both the existing dir item key and the dir index key. Therefore logging just dir index keys is enough to ensure dentry deletions are correctly replayed; 3) For dentry replacements: they work when we log only dir index keys and this is mostly due to a combination of 1) and 2). If we replace a dentry with name "foobar" to point from inode A to inode B, then we know the dir index key for the new dentry is different from the old one, as it has an index number (key offset) larger than the old one. This results in replaying a deletion, through replay_dir_deletes(), that causes the old dentry to be removed, both the dir item key and the dir index key, as mentioned at 2). Then when processing the new dir index key, we add the new dentry, adding both a new dir item key and a new index key pointing to inode B, as stated in 1). The forward compatibility, the ability for a patched kernel to replay a log created by an older, unpatched kernel, comes from the changes required for making sure we are able to replay a log that only contains dir index keys - we simply ignore every dir item key we find. So modify directory logging to log only dir index items, and modify the log replay process to ignore dir item keys, from log trees created by an unpatched kernel, and process only with dir index keys. This reduces the amount of logged metadata by about half, and therefore the time spent logging or fsyncing large directories (less CPU time and less IO). The following test script was used to measure this change: #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=1000000 NUM_FILE_DELETES=10000 mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The tests were run on a physical machine, with a non-debug kernel (Debian's default kernel config), for different values of $NUM_NEW_FILES and $NUM_FILE_DELETES, and the results were the following: ** Before patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 8412 ms after adding 1000000 files dir fsync took 500 ms after deleting 10000 files ** After patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 4252 ms after adding 1000000 files (-49.5%) dir fsync took 269 ms after deleting 10000 files (-46.2%) ** Before patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 745 ms after adding 100000 files dir fsync took 59 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 404 ms after adding 100000 files (-45.8%) dir fsync took 31 ms after deleting 1000 files (-47.5%) ** Before patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 67 ms after adding 10000 files dir fsync took 9 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 36 ms after adding 10000 files (-46.3%) dir fsync took 5 ms after deleting 1000 files (-44.4%) ** Before patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 9 ms after adding 1000 files dir fsync took 4 ms after deleting 100 files ** After patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 7 ms after adding 1000 files (-22.2%) dir fsync took 3 ms after deleting 100 files (-25.0%) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-10-25 19:31:53 +03:00
btrfs_release_path(path);
if (dir_dst_matches && index_dst_matches) {
ret = 0;
update_size = false;
goto out;
}
/*
* Check if the inode reference exists in the log for the given name,
* inode and parent inode
*/
btrfs: only copy dir index keys when logging a directory Currently, when logging a directory, we copy both dir items and dir index items from the fs/subvolume tree to the log tree. Both items have exactly the same data (same struct btrfs_dir_item), the difference lies in the key values, where a dir index key contains the index number of a directory entry while the dir item key does not, as it's used for doing fast lookups of an entry by name, while the former is used for sorting entries when listing a directory. We can exploit that and log only the dir index items, since they contain all the information needed to correctly add, replace and delete directory entries when replaying a log tree. Logging only the dir index items is also backward and forward compatible: an unpatched kernel (without this change) can correctly replay a log tree generated by a patched kernel (with this patch), and a patched kernel can correctly replay a log tree generated by an unpatched kernel. The backward compatibility is ensured because: 1) For inserting a new dentry: a dentry is only inserted when we find a new dir index key - we can only insert if we know the dir index offset, which is encoded in the dir index key's offset; 2) For deleting dentries: during log replay, before adding or replacing dentries, we first replay dentry deletions. Whenever we find a dir item key or a dir index key in the subvolume/fs tree that is not logged in a range for which the log tree is authoritative, we do the unlink of the dentry, which removes both the existing dir item key and the dir index key. Therefore logging just dir index keys is enough to ensure dentry deletions are correctly replayed; 3) For dentry replacements: they work when we log only dir index keys and this is mostly due to a combination of 1) and 2). If we replace a dentry with name "foobar" to point from inode A to inode B, then we know the dir index key for the new dentry is different from the old one, as it has an index number (key offset) larger than the old one. This results in replaying a deletion, through replay_dir_deletes(), that causes the old dentry to be removed, both the dir item key and the dir index key, as mentioned at 2). Then when processing the new dir index key, we add the new dentry, adding both a new dir item key and a new index key pointing to inode B, as stated in 1). The forward compatibility, the ability for a patched kernel to replay a log created by an older, unpatched kernel, comes from the changes required for making sure we are able to replay a log that only contains dir index keys - we simply ignore every dir item key we find. So modify directory logging to log only dir index items, and modify the log replay process to ignore dir item keys, from log trees created by an unpatched kernel, and process only with dir index keys. This reduces the amount of logged metadata by about half, and therefore the time spent logging or fsyncing large directories (less CPU time and less IO). The following test script was used to measure this change: #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=1000000 NUM_FILE_DELETES=10000 mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The tests were run on a physical machine, with a non-debug kernel (Debian's default kernel config), for different values of $NUM_NEW_FILES and $NUM_FILE_DELETES, and the results were the following: ** Before patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 8412 ms after adding 1000000 files dir fsync took 500 ms after deleting 10000 files ** After patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 4252 ms after adding 1000000 files (-49.5%) dir fsync took 269 ms after deleting 10000 files (-46.2%) ** Before patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 745 ms after adding 100000 files dir fsync took 59 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 404 ms after adding 100000 files (-45.8%) dir fsync took 31 ms after deleting 1000 files (-47.5%) ** Before patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 67 ms after adding 10000 files dir fsync took 9 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 36 ms after adding 10000 files (-46.3%) dir fsync took 5 ms after deleting 1000 files (-44.4%) ** Before patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 9 ms after adding 1000 files dir fsync took 4 ms after deleting 100 files ** After patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 7 ms after adding 1000 files (-22.2%) dir fsync took 3 ms after deleting 100 files (-25.0%) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-10-25 19:31:53 +03:00
search_key.objectid = log_key.objectid;
search_key.type = BTRFS_INODE_REF_KEY;
search_key.offset = key->objectid;
ret = backref_in_log(root->log_root, &search_key, 0, name, name_len);
if (ret < 0) {
goto out;
} else if (ret) {
/* The dentry will be added later. */
ret = 0;
update_size = false;
goto out;
}
btrfs: only copy dir index keys when logging a directory Currently, when logging a directory, we copy both dir items and dir index items from the fs/subvolume tree to the log tree. Both items have exactly the same data (same struct btrfs_dir_item), the difference lies in the key values, where a dir index key contains the index number of a directory entry while the dir item key does not, as it's used for doing fast lookups of an entry by name, while the former is used for sorting entries when listing a directory. We can exploit that and log only the dir index items, since they contain all the information needed to correctly add, replace and delete directory entries when replaying a log tree. Logging only the dir index items is also backward and forward compatible: an unpatched kernel (without this change) can correctly replay a log tree generated by a patched kernel (with this patch), and a patched kernel can correctly replay a log tree generated by an unpatched kernel. The backward compatibility is ensured because: 1) For inserting a new dentry: a dentry is only inserted when we find a new dir index key - we can only insert if we know the dir index offset, which is encoded in the dir index key's offset; 2) For deleting dentries: during log replay, before adding or replacing dentries, we first replay dentry deletions. Whenever we find a dir item key or a dir index key in the subvolume/fs tree that is not logged in a range for which the log tree is authoritative, we do the unlink of the dentry, which removes both the existing dir item key and the dir index key. Therefore logging just dir index keys is enough to ensure dentry deletions are correctly replayed; 3) For dentry replacements: they work when we log only dir index keys and this is mostly due to a combination of 1) and 2). If we replace a dentry with name "foobar" to point from inode A to inode B, then we know the dir index key for the new dentry is different from the old one, as it has an index number (key offset) larger than the old one. This results in replaying a deletion, through replay_dir_deletes(), that causes the old dentry to be removed, both the dir item key and the dir index key, as mentioned at 2). Then when processing the new dir index key, we add the new dentry, adding both a new dir item key and a new index key pointing to inode B, as stated in 1). The forward compatibility, the ability for a patched kernel to replay a log created by an older, unpatched kernel, comes from the changes required for making sure we are able to replay a log that only contains dir index keys - we simply ignore every dir item key we find. So modify directory logging to log only dir index items, and modify the log replay process to ignore dir item keys, from log trees created by an unpatched kernel, and process only with dir index keys. This reduces the amount of logged metadata by about half, and therefore the time spent logging or fsyncing large directories (less CPU time and less IO). The following test script was used to measure this change: #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=1000000 NUM_FILE_DELETES=10000 mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The tests were run on a physical machine, with a non-debug kernel (Debian's default kernel config), for different values of $NUM_NEW_FILES and $NUM_FILE_DELETES, and the results were the following: ** Before patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 8412 ms after adding 1000000 files dir fsync took 500 ms after deleting 10000 files ** After patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 4252 ms after adding 1000000 files (-49.5%) dir fsync took 269 ms after deleting 10000 files (-46.2%) ** Before patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 745 ms after adding 100000 files dir fsync took 59 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 404 ms after adding 100000 files (-45.8%) dir fsync took 31 ms after deleting 1000 files (-47.5%) ** Before patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 67 ms after adding 10000 files dir fsync took 9 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 36 ms after adding 10000 files (-46.3%) dir fsync took 5 ms after deleting 1000 files (-44.4%) ** Before patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 9 ms after adding 1000 files dir fsync took 4 ms after deleting 100 files ** After patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 7 ms after adding 1000 files (-22.2%) dir fsync took 3 ms after deleting 100 files (-25.0%) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-10-25 19:31:53 +03:00
search_key.objectid = log_key.objectid;
search_key.type = BTRFS_INODE_EXTREF_KEY;
search_key.offset = key->objectid;
ret = backref_in_log(root->log_root, &search_key, key->objectid, name,
name_len);
if (ret < 0) {
goto out;
} else if (ret) {
Btrfs: fix fsync log replay for inodes with a mix of regular refs and extrefs If we have an inode with a large number of hard links, some of which may be extrefs, turn a regular ref into an extref, fsync the inode and then replay the fsync log (after a crash/reboot), we can endup with an fsync log that makes the replay code always fail with -EOVERFLOW when processing the inode's references. This is easy to reproduce with the test case I made for xfstests. Its steps are the following: _scratch_mkfs "-O extref" >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create a test file with 3001 hard links. This number is large enough to # make btrfs start using extrefs at some point even if the fs has the maximum # possible leaf/node size (64Kb). echo "hello world" > $SCRATCH_MNT/foo for i in `seq 1 3000`; do ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link_`printf "%04d" $i` done # Make sure all metadata and data are durably persisted. sync # Now remove one link, add a new one with a new name, add another new one with # the same name as the one we just removed and fsync the inode. rm -f $SCRATCH_MNT/foo_link_0001 ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link_3001 ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link_0001 rm -f $SCRATCH_MNT/foo_link_0002 ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link_3002 ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link_3003 $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foo # Simulate a crash/power loss. This makes sure the next mount # will see an fsync log and will replay that log. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # Check that the number of hard links is correct, we are able to remove all # the hard links and read the file's data. This is just to verify we don't # get stale file handle errors (due to dangling directory index entries that # point to inodes that no longer exist). echo "Link count: $(stat --format=%h $SCRATCH_MNT/foo)" [ -f $SCRATCH_MNT/foo ] || echo "Link foo is missing" for ((i = 1; i <= 3003; i++)); do name=foo_link_`printf "%04d" $i` if [ $i -eq 2 ]; then [ -f $SCRATCH_MNT/$name ] && echo "Link $name found" else [ -f $SCRATCH_MNT/$name ] || echo "Link $name is missing" fi done rm -f $SCRATCH_MNT/foo_link_* cat $SCRATCH_MNT/foo rm -f $SCRATCH_MNT/foo status=0 exit The fix is simply to correct the overflow condition when overwriting a reference item because it was wrong, trying to increase the item in the fs/subvol tree by an impossible amount. Also ensure that we don't insert one normal ref and one ext ref for the same dentry - this happened because processing a dir index entry from the parent in the log happened when the normal ref item was full, which made the logic insert an extref and later when the normal ref had enough room, it would be inserted again when processing the ref item from the child inode in the log. This issue has been present since the introduction of the extrefs feature (2012). A test case for xfstests follows soon. This test only passes if the previous patch titled "Btrfs: fix fsync when extend references are added to an inode" is applied too. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-01-14 04:52:25 +03:00
/* The dentry will be added later. */
ret = 0;
update_size = false;
goto out;
}
btrfs_release_path(path);
ret = insert_one_name(trans, root, key->objectid, key->offset,
name, name_len, &log_key);
Btrfs: fix fsync log replay for inodes with a mix of regular refs and extrefs If we have an inode with a large number of hard links, some of which may be extrefs, turn a regular ref into an extref, fsync the inode and then replay the fsync log (after a crash/reboot), we can endup with an fsync log that makes the replay code always fail with -EOVERFLOW when processing the inode's references. This is easy to reproduce with the test case I made for xfstests. Its steps are the following: _scratch_mkfs "-O extref" >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create a test file with 3001 hard links. This number is large enough to # make btrfs start using extrefs at some point even if the fs has the maximum # possible leaf/node size (64Kb). echo "hello world" > $SCRATCH_MNT/foo for i in `seq 1 3000`; do ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link_`printf "%04d" $i` done # Make sure all metadata and data are durably persisted. sync # Now remove one link, add a new one with a new name, add another new one with # the same name as the one we just removed and fsync the inode. rm -f $SCRATCH_MNT/foo_link_0001 ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link_3001 ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link_0001 rm -f $SCRATCH_MNT/foo_link_0002 ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link_3002 ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link_3003 $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foo # Simulate a crash/power loss. This makes sure the next mount # will see an fsync log and will replay that log. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # Check that the number of hard links is correct, we are able to remove all # the hard links and read the file's data. This is just to verify we don't # get stale file handle errors (due to dangling directory index entries that # point to inodes that no longer exist). echo "Link count: $(stat --format=%h $SCRATCH_MNT/foo)" [ -f $SCRATCH_MNT/foo ] || echo "Link foo is missing" for ((i = 1; i <= 3003; i++)); do name=foo_link_`printf "%04d" $i` if [ $i -eq 2 ]; then [ -f $SCRATCH_MNT/$name ] && echo "Link $name found" else [ -f $SCRATCH_MNT/$name ] || echo "Link $name is missing" fi done rm -f $SCRATCH_MNT/foo_link_* cat $SCRATCH_MNT/foo rm -f $SCRATCH_MNT/foo status=0 exit The fix is simply to correct the overflow condition when overwriting a reference item because it was wrong, trying to increase the item in the fs/subvol tree by an impossible amount. Also ensure that we don't insert one normal ref and one ext ref for the same dentry - this happened because processing a dir index entry from the parent in the log happened when the normal ref item was full, which made the logic insert an extref and later when the normal ref had enough room, it would be inserted again when processing the ref item from the child inode in the log. This issue has been present since the introduction of the extrefs feature (2012). A test case for xfstests follows soon. This test only passes if the previous patch titled "Btrfs: fix fsync when extend references are added to an inode" is applied too. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-01-14 04:52:25 +03:00
if (ret && ret != -ENOENT && ret != -EEXIST)
goto out;
Btrfs: fix stale directory entries after fsync log replay We have another case where after an fsync log replay we get an inode with a wrong link count (smaller than it should be) and a number of directory entries greater than its link count. This happens when we add a new link hard link to our inode A and then we fsync some other inode B that has the side effect of logging the parent directory inode too. In this case at log replay time we add the new hard link to our inode (the item with key BTRFS_INODE_REF_KEY) when processing the parent directory but we never adjust the link count of our inode A. As a result we get stale dir entries for our inode A that can never be deleted and therefore it makes it impossible to remove the parent directory (as its i_size can never decrease back to 0). A simple reproducer for fstests that triggers this issue: seq=`basename $0` seqres=$RESULT_DIR/$seq echo "QA output created by $seq" tmp=/tmp/$$ status=1 # failure is the default! trap "_cleanup; exit \$status" 0 1 2 3 15 _cleanup() { _cleanup_flakey rm -f $tmp.* } # get standard environment, filters and checks . ./common/rc . ./common/filter . ./common/dmflakey # real QA test starts here _need_to_be_root _supported_fs generic _supported_os Linux _require_scratch _require_dm_flakey _require_metadata_journaling $SCRATCH_DEV rm -f $seqres.full _scratch_mkfs >>$seqres.full 2>&1 _init_flakey _mount_flakey # Create our test directory and files. mkdir $SCRATCH_MNT/testdir touch $SCRATCH_MNT/testdir/foo touch $SCRATCH_MNT/testdir/bar # Make sure everything done so far is durably persisted. sync # Create one hard link for file foo and another one for file bar. After # that fsync only the file bar. ln $SCRATCH_MNT/testdir/bar $SCRATCH_MNT/testdir/bar_link ln $SCRATCH_MNT/testdir/foo $SCRATCH_MNT/testdir/foo_link $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/testdir/bar # Silently drop all writes on scratch device to simulate power failure. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again and mount the fs to trigger log/journal replay. _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # Now verify both our files have a link count of 2. echo "Link count for file foo: $(stat --format=%h $SCRATCH_MNT/testdir/foo)" echo "Link count for file bar: $(stat --format=%h $SCRATCH_MNT/testdir/bar)" # We should be able to remove all the links of our files in testdir, and # after that the parent directory should become empty and therefore # possible to remove it. rm -f $SCRATCH_MNT/testdir/* rmdir $SCRATCH_MNT/testdir _unmount_flakey # The fstests framework will call fsck against our filesystem which will verify # that all metadata is in a consistent state. status=0 exit The test fails with: -Link count for file foo: 2 +Link count for file foo: 1 Link count for file bar: 2 +rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/testdir/foo_link': Stale file handle +rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/testdir': Directory not empty (...) _check_btrfs_filesystem: filesystem on /dev/sdc is inconsistent And fsck's output: (...) checking fs roots root 5 inode 258 errors 2001, no inode item, link count wrong unresolved ref dir 257 index 5 namelen 8 name foo_link filetype 1 errors 4, no inode ref Checking filesystem on /dev/sdc (...) So fix this by marking inodes for link count fixup at log replay time whenever a directory entry is replayed if the entry was created in the transaction where the fsync was made and if it points to a non-directory inode. This isn't a new problem/regression, the issue exists for a long time, possibly since the log tree feature was added (2008). Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-07-16 01:26:43 +03:00
if (!ret)
name_added = true;
update_size = false;
ret = 0;
btrfs: only copy dir index keys when logging a directory Currently, when logging a directory, we copy both dir items and dir index items from the fs/subvolume tree to the log tree. Both items have exactly the same data (same struct btrfs_dir_item), the difference lies in the key values, where a dir index key contains the index number of a directory entry while the dir item key does not, as it's used for doing fast lookups of an entry by name, while the former is used for sorting entries when listing a directory. We can exploit that and log only the dir index items, since they contain all the information needed to correctly add, replace and delete directory entries when replaying a log tree. Logging only the dir index items is also backward and forward compatible: an unpatched kernel (without this change) can correctly replay a log tree generated by a patched kernel (with this patch), and a patched kernel can correctly replay a log tree generated by an unpatched kernel. The backward compatibility is ensured because: 1) For inserting a new dentry: a dentry is only inserted when we find a new dir index key - we can only insert if we know the dir index offset, which is encoded in the dir index key's offset; 2) For deleting dentries: during log replay, before adding or replacing dentries, we first replay dentry deletions. Whenever we find a dir item key or a dir index key in the subvolume/fs tree that is not logged in a range for which the log tree is authoritative, we do the unlink of the dentry, which removes both the existing dir item key and the dir index key. Therefore logging just dir index keys is enough to ensure dentry deletions are correctly replayed; 3) For dentry replacements: they work when we log only dir index keys and this is mostly due to a combination of 1) and 2). If we replace a dentry with name "foobar" to point from inode A to inode B, then we know the dir index key for the new dentry is different from the old one, as it has an index number (key offset) larger than the old one. This results in replaying a deletion, through replay_dir_deletes(), that causes the old dentry to be removed, both the dir item key and the dir index key, as mentioned at 2). Then when processing the new dir index key, we add the new dentry, adding both a new dir item key and a new index key pointing to inode B, as stated in 1). The forward compatibility, the ability for a patched kernel to replay a log created by an older, unpatched kernel, comes from the changes required for making sure we are able to replay a log that only contains dir index keys - we simply ignore every dir item key we find. So modify directory logging to log only dir index items, and modify the log replay process to ignore dir item keys, from log trees created by an unpatched kernel, and process only with dir index keys. This reduces the amount of logged metadata by about half, and therefore the time spent logging or fsyncing large directories (less CPU time and less IO). The following test script was used to measure this change: #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=1000000 NUM_FILE_DELETES=10000 mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The tests were run on a physical machine, with a non-debug kernel (Debian's default kernel config), for different values of $NUM_NEW_FILES and $NUM_FILE_DELETES, and the results were the following: ** Before patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 8412 ms after adding 1000000 files dir fsync took 500 ms after deleting 10000 files ** After patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 4252 ms after adding 1000000 files (-49.5%) dir fsync took 269 ms after deleting 10000 files (-46.2%) ** Before patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 745 ms after adding 100000 files dir fsync took 59 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 404 ms after adding 100000 files (-45.8%) dir fsync took 31 ms after deleting 1000 files (-47.5%) ** Before patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 67 ms after adding 10000 files dir fsync took 9 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 36 ms after adding 10000 files (-46.3%) dir fsync took 5 ms after deleting 1000 files (-44.4%) ** Before patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 9 ms after adding 1000 files dir fsync took 4 ms after deleting 100 files ** After patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 7 ms after adding 1000 files (-22.2%) dir fsync took 3 ms after deleting 100 files (-25.0%) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-10-25 19:31:53 +03:00
out:
if (!ret && update_size) {
btrfs_i_size_write(BTRFS_I(dir), dir->i_size + name_len * 2);
ret = btrfs_update_inode(trans, root, BTRFS_I(dir));
}
kfree(name);
iput(dir);
if (!ret && name_added)
ret = 1;
return ret;
}
btrfs: only copy dir index keys when logging a directory Currently, when logging a directory, we copy both dir items and dir index items from the fs/subvolume tree to the log tree. Both items have exactly the same data (same struct btrfs_dir_item), the difference lies in the key values, where a dir index key contains the index number of a directory entry while the dir item key does not, as it's used for doing fast lookups of an entry by name, while the former is used for sorting entries when listing a directory. We can exploit that and log only the dir index items, since they contain all the information needed to correctly add, replace and delete directory entries when replaying a log tree. Logging only the dir index items is also backward and forward compatible: an unpatched kernel (without this change) can correctly replay a log tree generated by a patched kernel (with this patch), and a patched kernel can correctly replay a log tree generated by an unpatched kernel. The backward compatibility is ensured because: 1) For inserting a new dentry: a dentry is only inserted when we find a new dir index key - we can only insert if we know the dir index offset, which is encoded in the dir index key's offset; 2) For deleting dentries: during log replay, before adding or replacing dentries, we first replay dentry deletions. Whenever we find a dir item key or a dir index key in the subvolume/fs tree that is not logged in a range for which the log tree is authoritative, we do the unlink of the dentry, which removes both the existing dir item key and the dir index key. Therefore logging just dir index keys is enough to ensure dentry deletions are correctly replayed; 3) For dentry replacements: they work when we log only dir index keys and this is mostly due to a combination of 1) and 2). If we replace a dentry with name "foobar" to point from inode A to inode B, then we know the dir index key for the new dentry is different from the old one, as it has an index number (key offset) larger than the old one. This results in replaying a deletion, through replay_dir_deletes(), that causes the old dentry to be removed, both the dir item key and the dir index key, as mentioned at 2). Then when processing the new dir index key, we add the new dentry, adding both a new dir item key and a new index key pointing to inode B, as stated in 1). The forward compatibility, the ability for a patched kernel to replay a log created by an older, unpatched kernel, comes from the changes required for making sure we are able to replay a log that only contains dir index keys - we simply ignore every dir item key we find. So modify directory logging to log only dir index items, and modify the log replay process to ignore dir item keys, from log trees created by an unpatched kernel, and process only with dir index keys. This reduces the amount of logged metadata by about half, and therefore the time spent logging or fsyncing large directories (less CPU time and less IO). The following test script was used to measure this change: #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=1000000 NUM_FILE_DELETES=10000 mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The tests were run on a physical machine, with a non-debug kernel (Debian's default kernel config), for different values of $NUM_NEW_FILES and $NUM_FILE_DELETES, and the results were the following: ** Before patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 8412 ms after adding 1000000 files dir fsync took 500 ms after deleting 10000 files ** After patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 4252 ms after adding 1000000 files (-49.5%) dir fsync took 269 ms after deleting 10000 files (-46.2%) ** Before patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 745 ms after adding 100000 files dir fsync took 59 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 404 ms after adding 100000 files (-45.8%) dir fsync took 31 ms after deleting 1000 files (-47.5%) ** Before patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 67 ms after adding 10000 files dir fsync took 9 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 36 ms after adding 10000 files (-46.3%) dir fsync took 5 ms after deleting 1000 files (-44.4%) ** Before patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 9 ms after adding 1000 files dir fsync took 4 ms after deleting 100 files ** After patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 7 ms after adding 1000 files (-22.2%) dir fsync took 3 ms after deleting 100 files (-25.0%) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-10-25 19:31:53 +03:00
/* Replay one dir item from a BTRFS_DIR_INDEX_KEY key. */
static noinline int replay_one_dir_item(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_path *path,
struct extent_buffer *eb, int slot,
struct btrfs_key *key)
{
btrfs: only copy dir index keys when logging a directory Currently, when logging a directory, we copy both dir items and dir index items from the fs/subvolume tree to the log tree. Both items have exactly the same data (same struct btrfs_dir_item), the difference lies in the key values, where a dir index key contains the index number of a directory entry while the dir item key does not, as it's used for doing fast lookups of an entry by name, while the former is used for sorting entries when listing a directory. We can exploit that and log only the dir index items, since they contain all the information needed to correctly add, replace and delete directory entries when replaying a log tree. Logging only the dir index items is also backward and forward compatible: an unpatched kernel (without this change) can correctly replay a log tree generated by a patched kernel (with this patch), and a patched kernel can correctly replay a log tree generated by an unpatched kernel. The backward compatibility is ensured because: 1) For inserting a new dentry: a dentry is only inserted when we find a new dir index key - we can only insert if we know the dir index offset, which is encoded in the dir index key's offset; 2) For deleting dentries: during log replay, before adding or replacing dentries, we first replay dentry deletions. Whenever we find a dir item key or a dir index key in the subvolume/fs tree that is not logged in a range for which the log tree is authoritative, we do the unlink of the dentry, which removes both the existing dir item key and the dir index key. Therefore logging just dir index keys is enough to ensure dentry deletions are correctly replayed; 3) For dentry replacements: they work when we log only dir index keys and this is mostly due to a combination of 1) and 2). If we replace a dentry with name "foobar" to point from inode A to inode B, then we know the dir index key for the new dentry is different from the old one, as it has an index number (key offset) larger than the old one. This results in replaying a deletion, through replay_dir_deletes(), that causes the old dentry to be removed, both the dir item key and the dir index key, as mentioned at 2). Then when processing the new dir index key, we add the new dentry, adding both a new dir item key and a new index key pointing to inode B, as stated in 1). The forward compatibility, the ability for a patched kernel to replay a log created by an older, unpatched kernel, comes from the changes required for making sure we are able to replay a log that only contains dir index keys - we simply ignore every dir item key we find. So modify directory logging to log only dir index items, and modify the log replay process to ignore dir item keys, from log trees created by an unpatched kernel, and process only with dir index keys. This reduces the amount of logged metadata by about half, and therefore the time spent logging or fsyncing large directories (less CPU time and less IO). The following test script was used to measure this change: #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=1000000 NUM_FILE_DELETES=10000 mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The tests were run on a physical machine, with a non-debug kernel (Debian's default kernel config), for different values of $NUM_NEW_FILES and $NUM_FILE_DELETES, and the results were the following: ** Before patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 8412 ms after adding 1000000 files dir fsync took 500 ms after deleting 10000 files ** After patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 4252 ms after adding 1000000 files (-49.5%) dir fsync took 269 ms after deleting 10000 files (-46.2%) ** Before patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 745 ms after adding 100000 files dir fsync took 59 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 404 ms after adding 100000 files (-45.8%) dir fsync took 31 ms after deleting 1000 files (-47.5%) ** Before patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 67 ms after adding 10000 files dir fsync took 9 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 36 ms after adding 10000 files (-46.3%) dir fsync took 5 ms after deleting 1000 files (-44.4%) ** Before patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 9 ms after adding 1000 files dir fsync took 4 ms after deleting 100 files ** After patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 7 ms after adding 1000 files (-22.2%) dir fsync took 3 ms after deleting 100 files (-25.0%) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-10-25 19:31:53 +03:00
int ret;
struct btrfs_dir_item *di;
btrfs: only copy dir index keys when logging a directory Currently, when logging a directory, we copy both dir items and dir index items from the fs/subvolume tree to the log tree. Both items have exactly the same data (same struct btrfs_dir_item), the difference lies in the key values, where a dir index key contains the index number of a directory entry while the dir item key does not, as it's used for doing fast lookups of an entry by name, while the former is used for sorting entries when listing a directory. We can exploit that and log only the dir index items, since they contain all the information needed to correctly add, replace and delete directory entries when replaying a log tree. Logging only the dir index items is also backward and forward compatible: an unpatched kernel (without this change) can correctly replay a log tree generated by a patched kernel (with this patch), and a patched kernel can correctly replay a log tree generated by an unpatched kernel. The backward compatibility is ensured because: 1) For inserting a new dentry: a dentry is only inserted when we find a new dir index key - we can only insert if we know the dir index offset, which is encoded in the dir index key's offset; 2) For deleting dentries: during log replay, before adding or replacing dentries, we first replay dentry deletions. Whenever we find a dir item key or a dir index key in the subvolume/fs tree that is not logged in a range for which the log tree is authoritative, we do the unlink of the dentry, which removes both the existing dir item key and the dir index key. Therefore logging just dir index keys is enough to ensure dentry deletions are correctly replayed; 3) For dentry replacements: they work when we log only dir index keys and this is mostly due to a combination of 1) and 2). If we replace a dentry with name "foobar" to point from inode A to inode B, then we know the dir index key for the new dentry is different from the old one, as it has an index number (key offset) larger than the old one. This results in replaying a deletion, through replay_dir_deletes(), that causes the old dentry to be removed, both the dir item key and the dir index key, as mentioned at 2). Then when processing the new dir index key, we add the new dentry, adding both a new dir item key and a new index key pointing to inode B, as stated in 1). The forward compatibility, the ability for a patched kernel to replay a log created by an older, unpatched kernel, comes from the changes required for making sure we are able to replay a log that only contains dir index keys - we simply ignore every dir item key we find. So modify directory logging to log only dir index items, and modify the log replay process to ignore dir item keys, from log trees created by an unpatched kernel, and process only with dir index keys. This reduces the amount of logged metadata by about half, and therefore the time spent logging or fsyncing large directories (less CPU time and less IO). The following test script was used to measure this change: #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=1000000 NUM_FILE_DELETES=10000 mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The tests were run on a physical machine, with a non-debug kernel (Debian's default kernel config), for different values of $NUM_NEW_FILES and $NUM_FILE_DELETES, and the results were the following: ** Before patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 8412 ms after adding 1000000 files dir fsync took 500 ms after deleting 10000 files ** After patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 4252 ms after adding 1000000 files (-49.5%) dir fsync took 269 ms after deleting 10000 files (-46.2%) ** Before patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 745 ms after adding 100000 files dir fsync took 59 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 404 ms after adding 100000 files (-45.8%) dir fsync took 31 ms after deleting 1000 files (-47.5%) ** Before patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 67 ms after adding 10000 files dir fsync took 9 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 36 ms after adding 10000 files (-46.3%) dir fsync took 5 ms after deleting 1000 files (-44.4%) ** Before patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 9 ms after adding 1000 files dir fsync took 4 ms after deleting 100 files ** After patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 7 ms after adding 1000 files (-22.2%) dir fsync took 3 ms after deleting 100 files (-25.0%) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-10-25 19:31:53 +03:00
/* We only log dir index keys, which only contain a single dir item. */
ASSERT(key->type == BTRFS_DIR_INDEX_KEY);
Btrfs: fix stale directory entries after fsync log replay We have another case where after an fsync log replay we get an inode with a wrong link count (smaller than it should be) and a number of directory entries greater than its link count. This happens when we add a new link hard link to our inode A and then we fsync some other inode B that has the side effect of logging the parent directory inode too. In this case at log replay time we add the new hard link to our inode (the item with key BTRFS_INODE_REF_KEY) when processing the parent directory but we never adjust the link count of our inode A. As a result we get stale dir entries for our inode A that can never be deleted and therefore it makes it impossible to remove the parent directory (as its i_size can never decrease back to 0). A simple reproducer for fstests that triggers this issue: seq=`basename $0` seqres=$RESULT_DIR/$seq echo "QA output created by $seq" tmp=/tmp/$$ status=1 # failure is the default! trap "_cleanup; exit \$status" 0 1 2 3 15 _cleanup() { _cleanup_flakey rm -f $tmp.* } # get standard environment, filters and checks . ./common/rc . ./common/filter . ./common/dmflakey # real QA test starts here _need_to_be_root _supported_fs generic _supported_os Linux _require_scratch _require_dm_flakey _require_metadata_journaling $SCRATCH_DEV rm -f $seqres.full _scratch_mkfs >>$seqres.full 2>&1 _init_flakey _mount_flakey # Create our test directory and files. mkdir $SCRATCH_MNT/testdir touch $SCRATCH_MNT/testdir/foo touch $SCRATCH_MNT/testdir/bar # Make sure everything done so far is durably persisted. sync # Create one hard link for file foo and another one for file bar. After # that fsync only the file bar. ln $SCRATCH_MNT/testdir/bar $SCRATCH_MNT/testdir/bar_link ln $SCRATCH_MNT/testdir/foo $SCRATCH_MNT/testdir/foo_link $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/testdir/bar # Silently drop all writes on scratch device to simulate power failure. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again and mount the fs to trigger log/journal replay. _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # Now verify both our files have a link count of 2. echo "Link count for file foo: $(stat --format=%h $SCRATCH_MNT/testdir/foo)" echo "Link count for file bar: $(stat --format=%h $SCRATCH_MNT/testdir/bar)" # We should be able to remove all the links of our files in testdir, and # after that the parent directory should become empty and therefore # possible to remove it. rm -f $SCRATCH_MNT/testdir/* rmdir $SCRATCH_MNT/testdir _unmount_flakey # The fstests framework will call fsck against our filesystem which will verify # that all metadata is in a consistent state. status=0 exit The test fails with: -Link count for file foo: 2 +Link count for file foo: 1 Link count for file bar: 2 +rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/testdir/foo_link': Stale file handle +rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/testdir': Directory not empty (...) _check_btrfs_filesystem: filesystem on /dev/sdc is inconsistent And fsck's output: (...) checking fs roots root 5 inode 258 errors 2001, no inode item, link count wrong unresolved ref dir 257 index 5 namelen 8 name foo_link filetype 1 errors 4, no inode ref Checking filesystem on /dev/sdc (...) So fix this by marking inodes for link count fixup at log replay time whenever a directory entry is replayed if the entry was created in the transaction where the fsync was made and if it points to a non-directory inode. This isn't a new problem/regression, the issue exists for a long time, possibly since the log tree feature was added (2008). Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-07-16 01:26:43 +03:00
btrfs: only copy dir index keys when logging a directory Currently, when logging a directory, we copy both dir items and dir index items from the fs/subvolume tree to the log tree. Both items have exactly the same data (same struct btrfs_dir_item), the difference lies in the key values, where a dir index key contains the index number of a directory entry while the dir item key does not, as it's used for doing fast lookups of an entry by name, while the former is used for sorting entries when listing a directory. We can exploit that and log only the dir index items, since they contain all the information needed to correctly add, replace and delete directory entries when replaying a log tree. Logging only the dir index items is also backward and forward compatible: an unpatched kernel (without this change) can correctly replay a log tree generated by a patched kernel (with this patch), and a patched kernel can correctly replay a log tree generated by an unpatched kernel. The backward compatibility is ensured because: 1) For inserting a new dentry: a dentry is only inserted when we find a new dir index key - we can only insert if we know the dir index offset, which is encoded in the dir index key's offset; 2) For deleting dentries: during log replay, before adding or replacing dentries, we first replay dentry deletions. Whenever we find a dir item key or a dir index key in the subvolume/fs tree that is not logged in a range for which the log tree is authoritative, we do the unlink of the dentry, which removes both the existing dir item key and the dir index key. Therefore logging just dir index keys is enough to ensure dentry deletions are correctly replayed; 3) For dentry replacements: they work when we log only dir index keys and this is mostly due to a combination of 1) and 2). If we replace a dentry with name "foobar" to point from inode A to inode B, then we know the dir index key for the new dentry is different from the old one, as it has an index number (key offset) larger than the old one. This results in replaying a deletion, through replay_dir_deletes(), that causes the old dentry to be removed, both the dir item key and the dir index key, as mentioned at 2). Then when processing the new dir index key, we add the new dentry, adding both a new dir item key and a new index key pointing to inode B, as stated in 1). The forward compatibility, the ability for a patched kernel to replay a log created by an older, unpatched kernel, comes from the changes required for making sure we are able to replay a log that only contains dir index keys - we simply ignore every dir item key we find. So modify directory logging to log only dir index items, and modify the log replay process to ignore dir item keys, from log trees created by an unpatched kernel, and process only with dir index keys. This reduces the amount of logged metadata by about half, and therefore the time spent logging or fsyncing large directories (less CPU time and less IO). The following test script was used to measure this change: #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=1000000 NUM_FILE_DELETES=10000 mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The tests were run on a physical machine, with a non-debug kernel (Debian's default kernel config), for different values of $NUM_NEW_FILES and $NUM_FILE_DELETES, and the results were the following: ** Before patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 8412 ms after adding 1000000 files dir fsync took 500 ms after deleting 10000 files ** After patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 4252 ms after adding 1000000 files (-49.5%) dir fsync took 269 ms after deleting 10000 files (-46.2%) ** Before patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 745 ms after adding 100000 files dir fsync took 59 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 404 ms after adding 100000 files (-45.8%) dir fsync took 31 ms after deleting 1000 files (-47.5%) ** Before patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 67 ms after adding 10000 files dir fsync took 9 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 36 ms after adding 10000 files (-46.3%) dir fsync took 5 ms after deleting 1000 files (-44.4%) ** Before patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 9 ms after adding 1000 files dir fsync took 4 ms after deleting 100 files ** After patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 7 ms after adding 1000 files (-22.2%) dir fsync took 3 ms after deleting 100 files (-25.0%) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-10-25 19:31:53 +03:00
di = btrfs_item_ptr(eb, slot, struct btrfs_dir_item);
ret = replay_one_name(trans, root, path, eb, di, key);
if (ret < 0)
return ret;
Btrfs: fix stale directory entries after fsync log replay We have another case where after an fsync log replay we get an inode with a wrong link count (smaller than it should be) and a number of directory entries greater than its link count. This happens when we add a new link hard link to our inode A and then we fsync some other inode B that has the side effect of logging the parent directory inode too. In this case at log replay time we add the new hard link to our inode (the item with key BTRFS_INODE_REF_KEY) when processing the parent directory but we never adjust the link count of our inode A. As a result we get stale dir entries for our inode A that can never be deleted and therefore it makes it impossible to remove the parent directory (as its i_size can never decrease back to 0). A simple reproducer for fstests that triggers this issue: seq=`basename $0` seqres=$RESULT_DIR/$seq echo "QA output created by $seq" tmp=/tmp/$$ status=1 # failure is the default! trap "_cleanup; exit \$status" 0 1 2 3 15 _cleanup() { _cleanup_flakey rm -f $tmp.* } # get standard environment, filters and checks . ./common/rc . ./common/filter . ./common/dmflakey # real QA test starts here _need_to_be_root _supported_fs generic _supported_os Linux _require_scratch _require_dm_flakey _require_metadata_journaling $SCRATCH_DEV rm -f $seqres.full _scratch_mkfs >>$seqres.full 2>&1 _init_flakey _mount_flakey # Create our test directory and files. mkdir $SCRATCH_MNT/testdir touch $SCRATCH_MNT/testdir/foo touch $SCRATCH_MNT/testdir/bar # Make sure everything done so far is durably persisted. sync # Create one hard link for file foo and another one for file bar. After # that fsync only the file bar. ln $SCRATCH_MNT/testdir/bar $SCRATCH_MNT/testdir/bar_link ln $SCRATCH_MNT/testdir/foo $SCRATCH_MNT/testdir/foo_link $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/testdir/bar # Silently drop all writes on scratch device to simulate power failure. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again and mount the fs to trigger log/journal replay. _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # Now verify both our files have a link count of 2. echo "Link count for file foo: $(stat --format=%h $SCRATCH_MNT/testdir/foo)" echo "Link count for file bar: $(stat --format=%h $SCRATCH_MNT/testdir/bar)" # We should be able to remove all the links of our files in testdir, and # after that the parent directory should become empty and therefore # possible to remove it. rm -f $SCRATCH_MNT/testdir/* rmdir $SCRATCH_MNT/testdir _unmount_flakey # The fstests framework will call fsck against our filesystem which will verify # that all metadata is in a consistent state. status=0 exit The test fails with: -Link count for file foo: 2 +Link count for file foo: 1 Link count for file bar: 2 +rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/testdir/foo_link': Stale file handle +rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/testdir': Directory not empty (...) _check_btrfs_filesystem: filesystem on /dev/sdc is inconsistent And fsck's output: (...) checking fs roots root 5 inode 258 errors 2001, no inode item, link count wrong unresolved ref dir 257 index 5 namelen 8 name foo_link filetype 1 errors 4, no inode ref Checking filesystem on /dev/sdc (...) So fix this by marking inodes for link count fixup at log replay time whenever a directory entry is replayed if the entry was created in the transaction where the fsync was made and if it points to a non-directory inode. This isn't a new problem/regression, the issue exists for a long time, possibly since the log tree feature was added (2008). Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-07-16 01:26:43 +03:00
btrfs: only copy dir index keys when logging a directory Currently, when logging a directory, we copy both dir items and dir index items from the fs/subvolume tree to the log tree. Both items have exactly the same data (same struct btrfs_dir_item), the difference lies in the key values, where a dir index key contains the index number of a directory entry while the dir item key does not, as it's used for doing fast lookups of an entry by name, while the former is used for sorting entries when listing a directory. We can exploit that and log only the dir index items, since they contain all the information needed to correctly add, replace and delete directory entries when replaying a log tree. Logging only the dir index items is also backward and forward compatible: an unpatched kernel (without this change) can correctly replay a log tree generated by a patched kernel (with this patch), and a patched kernel can correctly replay a log tree generated by an unpatched kernel. The backward compatibility is ensured because: 1) For inserting a new dentry: a dentry is only inserted when we find a new dir index key - we can only insert if we know the dir index offset, which is encoded in the dir index key's offset; 2) For deleting dentries: during log replay, before adding or replacing dentries, we first replay dentry deletions. Whenever we find a dir item key or a dir index key in the subvolume/fs tree that is not logged in a range for which the log tree is authoritative, we do the unlink of the dentry, which removes both the existing dir item key and the dir index key. Therefore logging just dir index keys is enough to ensure dentry deletions are correctly replayed; 3) For dentry replacements: they work when we log only dir index keys and this is mostly due to a combination of 1) and 2). If we replace a dentry with name "foobar" to point from inode A to inode B, then we know the dir index key for the new dentry is different from the old one, as it has an index number (key offset) larger than the old one. This results in replaying a deletion, through replay_dir_deletes(), that causes the old dentry to be removed, both the dir item key and the dir index key, as mentioned at 2). Then when processing the new dir index key, we add the new dentry, adding both a new dir item key and a new index key pointing to inode B, as stated in 1). The forward compatibility, the ability for a patched kernel to replay a log created by an older, unpatched kernel, comes from the changes required for making sure we are able to replay a log that only contains dir index keys - we simply ignore every dir item key we find. So modify directory logging to log only dir index items, and modify the log replay process to ignore dir item keys, from log trees created by an unpatched kernel, and process only with dir index keys. This reduces the amount of logged metadata by about half, and therefore the time spent logging or fsyncing large directories (less CPU time and less IO). The following test script was used to measure this change: #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=1000000 NUM_FILE_DELETES=10000 mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The tests were run on a physical machine, with a non-debug kernel (Debian's default kernel config), for different values of $NUM_NEW_FILES and $NUM_FILE_DELETES, and the results were the following: ** Before patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 8412 ms after adding 1000000 files dir fsync took 500 ms after deleting 10000 files ** After patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 4252 ms after adding 1000000 files (-49.5%) dir fsync took 269 ms after deleting 10000 files (-46.2%) ** Before patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 745 ms after adding 100000 files dir fsync took 59 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 404 ms after adding 100000 files (-45.8%) dir fsync took 31 ms after deleting 1000 files (-47.5%) ** Before patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 67 ms after adding 10000 files dir fsync took 9 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 36 ms after adding 10000 files (-46.3%) dir fsync took 5 ms after deleting 1000 files (-44.4%) ** Before patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 9 ms after adding 1000 files dir fsync took 4 ms after deleting 100 files ** After patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 7 ms after adding 1000 files (-22.2%) dir fsync took 3 ms after deleting 100 files (-25.0%) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-10-25 19:31:53 +03:00
/*
* If this entry refers to a non-directory (directories can not have a
* link count > 1) and it was added in the transaction that was not
* committed, make sure we fixup the link count of the inode the entry
* points to. Otherwise something like the following would result in a
* directory pointing to an inode with a wrong link that does not account
* for this dir entry:
*
* mkdir testdir
* touch testdir/foo
* touch testdir/bar
* sync
*
* ln testdir/bar testdir/bar_link
* ln testdir/foo testdir/foo_link
* xfs_io -c "fsync" testdir/bar
*
* <power failure>
*
* mount fs, log replay happens
*
* File foo would remain with a link count of 1 when it has two entries
* pointing to it in the directory testdir. This would make it impossible
* to ever delete the parent directory has it would result in stale
* dentries that can never be deleted.
*/
if (ret == 1 && btrfs_dir_type(eb, di) != BTRFS_FT_DIR) {
struct btrfs_path *fixup_path;
struct btrfs_key di_key;
Btrfs: fix stale directory entries after fsync log replay We have another case where after an fsync log replay we get an inode with a wrong link count (smaller than it should be) and a number of directory entries greater than its link count. This happens when we add a new link hard link to our inode A and then we fsync some other inode B that has the side effect of logging the parent directory inode too. In this case at log replay time we add the new hard link to our inode (the item with key BTRFS_INODE_REF_KEY) when processing the parent directory but we never adjust the link count of our inode A. As a result we get stale dir entries for our inode A that can never be deleted and therefore it makes it impossible to remove the parent directory (as its i_size can never decrease back to 0). A simple reproducer for fstests that triggers this issue: seq=`basename $0` seqres=$RESULT_DIR/$seq echo "QA output created by $seq" tmp=/tmp/$$ status=1 # failure is the default! trap "_cleanup; exit \$status" 0 1 2 3 15 _cleanup() { _cleanup_flakey rm -f $tmp.* } # get standard environment, filters and checks . ./common/rc . ./common/filter . ./common/dmflakey # real QA test starts here _need_to_be_root _supported_fs generic _supported_os Linux _require_scratch _require_dm_flakey _require_metadata_journaling $SCRATCH_DEV rm -f $seqres.full _scratch_mkfs >>$seqres.full 2>&1 _init_flakey _mount_flakey # Create our test directory and files. mkdir $SCRATCH_MNT/testdir touch $SCRATCH_MNT/testdir/foo touch $SCRATCH_MNT/testdir/bar # Make sure everything done so far is durably persisted. sync # Create one hard link for file foo and another one for file bar. After # that fsync only the file bar. ln $SCRATCH_MNT/testdir/bar $SCRATCH_MNT/testdir/bar_link ln $SCRATCH_MNT/testdir/foo $SCRATCH_MNT/testdir/foo_link $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/testdir/bar # Silently drop all writes on scratch device to simulate power failure. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again and mount the fs to trigger log/journal replay. _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # Now verify both our files have a link count of 2. echo "Link count for file foo: $(stat --format=%h $SCRATCH_MNT/testdir/foo)" echo "Link count for file bar: $(stat --format=%h $SCRATCH_MNT/testdir/bar)" # We should be able to remove all the links of our files in testdir, and # after that the parent directory should become empty and therefore # possible to remove it. rm -f $SCRATCH_MNT/testdir/* rmdir $SCRATCH_MNT/testdir _unmount_flakey # The fstests framework will call fsck against our filesystem which will verify # that all metadata is in a consistent state. status=0 exit The test fails with: -Link count for file foo: 2 +Link count for file foo: 1 Link count for file bar: 2 +rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/testdir/foo_link': Stale file handle +rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/testdir': Directory not empty (...) _check_btrfs_filesystem: filesystem on /dev/sdc is inconsistent And fsck's output: (...) checking fs roots root 5 inode 258 errors 2001, no inode item, link count wrong unresolved ref dir 257 index 5 namelen 8 name foo_link filetype 1 errors 4, no inode ref Checking filesystem on /dev/sdc (...) So fix this by marking inodes for link count fixup at log replay time whenever a directory entry is replayed if the entry was created in the transaction where the fsync was made and if it points to a non-directory inode. This isn't a new problem/regression, the issue exists for a long time, possibly since the log tree feature was added (2008). Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-07-16 01:26:43 +03:00
btrfs: only copy dir index keys when logging a directory Currently, when logging a directory, we copy both dir items and dir index items from the fs/subvolume tree to the log tree. Both items have exactly the same data (same struct btrfs_dir_item), the difference lies in the key values, where a dir index key contains the index number of a directory entry while the dir item key does not, as it's used for doing fast lookups of an entry by name, while the former is used for sorting entries when listing a directory. We can exploit that and log only the dir index items, since they contain all the information needed to correctly add, replace and delete directory entries when replaying a log tree. Logging only the dir index items is also backward and forward compatible: an unpatched kernel (without this change) can correctly replay a log tree generated by a patched kernel (with this patch), and a patched kernel can correctly replay a log tree generated by an unpatched kernel. The backward compatibility is ensured because: 1) For inserting a new dentry: a dentry is only inserted when we find a new dir index key - we can only insert if we know the dir index offset, which is encoded in the dir index key's offset; 2) For deleting dentries: during log replay, before adding or replacing dentries, we first replay dentry deletions. Whenever we find a dir item key or a dir index key in the subvolume/fs tree that is not logged in a range for which the log tree is authoritative, we do the unlink of the dentry, which removes both the existing dir item key and the dir index key. Therefore logging just dir index keys is enough to ensure dentry deletions are correctly replayed; 3) For dentry replacements: they work when we log only dir index keys and this is mostly due to a combination of 1) and 2). If we replace a dentry with name "foobar" to point from inode A to inode B, then we know the dir index key for the new dentry is different from the old one, as it has an index number (key offset) larger than the old one. This results in replaying a deletion, through replay_dir_deletes(), that causes the old dentry to be removed, both the dir item key and the dir index key, as mentioned at 2). Then when processing the new dir index key, we add the new dentry, adding both a new dir item key and a new index key pointing to inode B, as stated in 1). The forward compatibility, the ability for a patched kernel to replay a log created by an older, unpatched kernel, comes from the changes required for making sure we are able to replay a log that only contains dir index keys - we simply ignore every dir item key we find. So modify directory logging to log only dir index items, and modify the log replay process to ignore dir item keys, from log trees created by an unpatched kernel, and process only with dir index keys. This reduces the amount of logged metadata by about half, and therefore the time spent logging or fsyncing large directories (less CPU time and less IO). The following test script was used to measure this change: #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=1000000 NUM_FILE_DELETES=10000 mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The tests were run on a physical machine, with a non-debug kernel (Debian's default kernel config), for different values of $NUM_NEW_FILES and $NUM_FILE_DELETES, and the results were the following: ** Before patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 8412 ms after adding 1000000 files dir fsync took 500 ms after deleting 10000 files ** After patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 4252 ms after adding 1000000 files (-49.5%) dir fsync took 269 ms after deleting 10000 files (-46.2%) ** Before patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 745 ms after adding 100000 files dir fsync took 59 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 404 ms after adding 100000 files (-45.8%) dir fsync took 31 ms after deleting 1000 files (-47.5%) ** Before patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 67 ms after adding 10000 files dir fsync took 9 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 36 ms after adding 10000 files (-46.3%) dir fsync took 5 ms after deleting 1000 files (-44.4%) ** Before patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 9 ms after adding 1000 files dir fsync took 4 ms after deleting 100 files ** After patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 7 ms after adding 1000 files (-22.2%) dir fsync took 3 ms after deleting 100 files (-25.0%) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-10-25 19:31:53 +03:00
fixup_path = btrfs_alloc_path();
if (!fixup_path)
return -ENOMEM;
btrfs_dir_item_key_to_cpu(eb, di, &di_key);
ret = link_to_fixup_dir(trans, root, fixup_path, di_key.objectid);
btrfs_free_path(fixup_path);
}
btrfs: only copy dir index keys when logging a directory Currently, when logging a directory, we copy both dir items and dir index items from the fs/subvolume tree to the log tree. Both items have exactly the same data (same struct btrfs_dir_item), the difference lies in the key values, where a dir index key contains the index number of a directory entry while the dir item key does not, as it's used for doing fast lookups of an entry by name, while the former is used for sorting entries when listing a directory. We can exploit that and log only the dir index items, since they contain all the information needed to correctly add, replace and delete directory entries when replaying a log tree. Logging only the dir index items is also backward and forward compatible: an unpatched kernel (without this change) can correctly replay a log tree generated by a patched kernel (with this patch), and a patched kernel can correctly replay a log tree generated by an unpatched kernel. The backward compatibility is ensured because: 1) For inserting a new dentry: a dentry is only inserted when we find a new dir index key - we can only insert if we know the dir index offset, which is encoded in the dir index key's offset; 2) For deleting dentries: during log replay, before adding or replacing dentries, we first replay dentry deletions. Whenever we find a dir item key or a dir index key in the subvolume/fs tree that is not logged in a range for which the log tree is authoritative, we do the unlink of the dentry, which removes both the existing dir item key and the dir index key. Therefore logging just dir index keys is enough to ensure dentry deletions are correctly replayed; 3) For dentry replacements: they work when we log only dir index keys and this is mostly due to a combination of 1) and 2). If we replace a dentry with name "foobar" to point from inode A to inode B, then we know the dir index key for the new dentry is different from the old one, as it has an index number (key offset) larger than the old one. This results in replaying a deletion, through replay_dir_deletes(), that causes the old dentry to be removed, both the dir item key and the dir index key, as mentioned at 2). Then when processing the new dir index key, we add the new dentry, adding both a new dir item key and a new index key pointing to inode B, as stated in 1). The forward compatibility, the ability for a patched kernel to replay a log created by an older, unpatched kernel, comes from the changes required for making sure we are able to replay a log that only contains dir index keys - we simply ignore every dir item key we find. So modify directory logging to log only dir index items, and modify the log replay process to ignore dir item keys, from log trees created by an unpatched kernel, and process only with dir index keys. This reduces the amount of logged metadata by about half, and therefore the time spent logging or fsyncing large directories (less CPU time and less IO). The following test script was used to measure this change: #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=1000000 NUM_FILE_DELETES=10000 mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The tests were run on a physical machine, with a non-debug kernel (Debian's default kernel config), for different values of $NUM_NEW_FILES and $NUM_FILE_DELETES, and the results were the following: ** Before patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 8412 ms after adding 1000000 files dir fsync took 500 ms after deleting 10000 files ** After patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 4252 ms after adding 1000000 files (-49.5%) dir fsync took 269 ms after deleting 10000 files (-46.2%) ** Before patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 745 ms after adding 100000 files dir fsync took 59 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 404 ms after adding 100000 files (-45.8%) dir fsync took 31 ms after deleting 1000 files (-47.5%) ** Before patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 67 ms after adding 10000 files dir fsync took 9 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 36 ms after adding 10000 files (-46.3%) dir fsync took 5 ms after deleting 1000 files (-44.4%) ** Before patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 9 ms after adding 1000 files dir fsync took 4 ms after deleting 100 files ** After patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 7 ms after adding 1000 files (-22.2%) dir fsync took 3 ms after deleting 100 files (-25.0%) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-10-25 19:31:53 +03:00
Btrfs: fix stale directory entries after fsync log replay We have another case where after an fsync log replay we get an inode with a wrong link count (smaller than it should be) and a number of directory entries greater than its link count. This happens when we add a new link hard link to our inode A and then we fsync some other inode B that has the side effect of logging the parent directory inode too. In this case at log replay time we add the new hard link to our inode (the item with key BTRFS_INODE_REF_KEY) when processing the parent directory but we never adjust the link count of our inode A. As a result we get stale dir entries for our inode A that can never be deleted and therefore it makes it impossible to remove the parent directory (as its i_size can never decrease back to 0). A simple reproducer for fstests that triggers this issue: seq=`basename $0` seqres=$RESULT_DIR/$seq echo "QA output created by $seq" tmp=/tmp/$$ status=1 # failure is the default! trap "_cleanup; exit \$status" 0 1 2 3 15 _cleanup() { _cleanup_flakey rm -f $tmp.* } # get standard environment, filters and checks . ./common/rc . ./common/filter . ./common/dmflakey # real QA test starts here _need_to_be_root _supported_fs generic _supported_os Linux _require_scratch _require_dm_flakey _require_metadata_journaling $SCRATCH_DEV rm -f $seqres.full _scratch_mkfs >>$seqres.full 2>&1 _init_flakey _mount_flakey # Create our test directory and files. mkdir $SCRATCH_MNT/testdir touch $SCRATCH_MNT/testdir/foo touch $SCRATCH_MNT/testdir/bar # Make sure everything done so far is durably persisted. sync # Create one hard link for file foo and another one for file bar. After # that fsync only the file bar. ln $SCRATCH_MNT/testdir/bar $SCRATCH_MNT/testdir/bar_link ln $SCRATCH_MNT/testdir/foo $SCRATCH_MNT/testdir/foo_link $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/testdir/bar # Silently drop all writes on scratch device to simulate power failure. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again and mount the fs to trigger log/journal replay. _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # Now verify both our files have a link count of 2. echo "Link count for file foo: $(stat --format=%h $SCRATCH_MNT/testdir/foo)" echo "Link count for file bar: $(stat --format=%h $SCRATCH_MNT/testdir/bar)" # We should be able to remove all the links of our files in testdir, and # after that the parent directory should become empty and therefore # possible to remove it. rm -f $SCRATCH_MNT/testdir/* rmdir $SCRATCH_MNT/testdir _unmount_flakey # The fstests framework will call fsck against our filesystem which will verify # that all metadata is in a consistent state. status=0 exit The test fails with: -Link count for file foo: 2 +Link count for file foo: 1 Link count for file bar: 2 +rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/testdir/foo_link': Stale file handle +rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/testdir': Directory not empty (...) _check_btrfs_filesystem: filesystem on /dev/sdc is inconsistent And fsck's output: (...) checking fs roots root 5 inode 258 errors 2001, no inode item, link count wrong unresolved ref dir 257 index 5 namelen 8 name foo_link filetype 1 errors 4, no inode ref Checking filesystem on /dev/sdc (...) So fix this by marking inodes for link count fixup at log replay time whenever a directory entry is replayed if the entry was created in the transaction where the fsync was made and if it points to a non-directory inode. This isn't a new problem/regression, the issue exists for a long time, possibly since the log tree feature was added (2008). Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-07-16 01:26:43 +03:00
return ret;
}
/*
* directory replay has two parts. There are the standard directory
* items in the log copied from the subvolume, and range items
* created in the log while the subvolume was logged.
*
* The range items tell us which parts of the key space the log
* is authoritative for. During replay, if a key in the subvolume
* directory is in a logged range item, but not actually in the log
* that means it was deleted from the directory before the fsync
* and should be removed.
*/
static noinline int find_dir_range(struct btrfs_root *root,
struct btrfs_path *path,
u64 dirid,
u64 *start_ret, u64 *end_ret)
{
struct btrfs_key key;
u64 found_end;
struct btrfs_dir_log_item *item;
int ret;
int nritems;
if (*start_ret == (u64)-1)
return 1;
key.objectid = dirid;
key.type = BTRFS_DIR_LOG_INDEX_KEY;
key.offset = *start_ret;
ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
if (ret < 0)
goto out;
if (ret > 0) {
if (path->slots[0] == 0)
goto out;
path->slots[0]--;
}
if (ret != 0)
btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
if (key.type != BTRFS_DIR_LOG_INDEX_KEY || key.objectid != dirid) {
ret = 1;
goto next;
}
item = btrfs_item_ptr(path->nodes[0], path->slots[0],
struct btrfs_dir_log_item);
found_end = btrfs_dir_log_end(path->nodes[0], item);
if (*start_ret >= key.offset && *start_ret <= found_end) {
ret = 0;
*start_ret = key.offset;
*end_ret = found_end;
goto out;
}
ret = 1;
next:
/* check the next slot in the tree to see if it is a valid item */
nritems = btrfs_header_nritems(path->nodes[0]);
Btrfs: fix tree search logic when replaying directory entry deletes If a log tree has a layout like the following: leaf N: ... item 240 key (282 DIR_LOG_ITEM 0) itemoff 8189 itemsize 8 dir log end 1275809046 leaf N + 1: item 0 key (282 DIR_LOG_ITEM 3936149215) itemoff 16275 itemsize 8 dir log end 18446744073709551615 ... When we pass the value 1275809046 + 1 as the parameter start_ret to the function tree-log.c:find_dir_range() (done by replay_dir_deletes()), we end up with path->slots[0] having the value 239 (points to the last item of leaf N, item 240). Because the dir log item in that position has an offset value smaller than *start_ret (1275809046 + 1) we need to move on to the next leaf, however the logic for that is wrong since it compares the current slot to the number of items in the leaf, which is smaller and therefore we don't lookup for the next leaf but instead we set the slot to point to an item that does not exist, at slot 240, and we later operate on that slot which has unexpected content or in the worst case can result in an invalid memory access (accessing beyond the last page of leaf N's extent buffer). So fix the logic that checks when we need to lookup at the next leaf by first incrementing the slot and only after to check if that slot is beyond the last item of the current leaf. Signed-off-by: Robbie Ko <robbieko@synology.com> Reviewed-by: Filipe Manana <fdmanana@suse.com> Fixes: e02119d5a7b4 (Btrfs: Add a write ahead tree log to optimize synchronous operations) Cc: stable@vger.kernel.org # 2.6.29+ Signed-off-by: Filipe Manana <fdmanana@suse.com> [Modified changelog for clarity and correctness]
2016-10-07 12:30:47 +03:00
path->slots[0]++;
if (path->slots[0] >= nritems) {
ret = btrfs_next_leaf(root, path);
if (ret)
goto out;
}
btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
if (key.type != BTRFS_DIR_LOG_INDEX_KEY || key.objectid != dirid) {
ret = 1;
goto out;
}
item = btrfs_item_ptr(path->nodes[0], path->slots[0],
struct btrfs_dir_log_item);
found_end = btrfs_dir_log_end(path->nodes[0], item);
*start_ret = key.offset;
*end_ret = found_end;
ret = 0;
out:
btrfs_release_path(path);
return ret;
}
/*
* this looks for a given directory item in the log. If the directory
* item is not in the log, the item is removed and the inode it points
* to is unlinked
*/
static noinline int check_item_in_log(struct btrfs_trans_handle *trans,
struct btrfs_root *log,
struct btrfs_path *path,
struct btrfs_path *log_path,
struct inode *dir,
struct btrfs_key *dir_key)
{
struct btrfs_root *root = BTRFS_I(dir)->root;
int ret;
struct extent_buffer *eb;
int slot;
struct btrfs_dir_item *di;
int name_len;
char *name;
struct inode *inode = NULL;
struct btrfs_key location;
/*
* Currenly we only log dir index keys. Even if we replay a log created
* by an older kernel that logged both dir index and dir item keys, all
* we need to do is process the dir index keys, we (and our caller) can
* safely ignore dir item keys (key type BTRFS_DIR_ITEM_KEY).
*/
ASSERT(dir_key->type == BTRFS_DIR_INDEX_KEY);
eb = path->nodes[0];
slot = path->slots[0];
di = btrfs_item_ptr(eb, slot, struct btrfs_dir_item);
name_len = btrfs_dir_name_len(eb, di);
name = kmalloc(name_len, GFP_NOFS);
if (!name) {
ret = -ENOMEM;
goto out;
}
read_extent_buffer(eb, name, (unsigned long)(di + 1), name_len);
if (log) {
struct btrfs_dir_item *log_di;
log_di = btrfs_lookup_dir_index_item(trans, log, log_path,
dir_key->objectid,
dir_key->offset,
name, name_len, 0);
if (IS_ERR(log_di)) {
ret = PTR_ERR(log_di);
goto out;
} else if (log_di) {
/* The dentry exists in the log, we have nothing to do. */
ret = 0;
goto out;
}
}
btrfs_dir_item_key_to_cpu(eb, di, &location);
btrfs_release_path(path);
btrfs_release_path(log_path);
inode = read_one_inode(root, location.objectid);
if (!inode) {
ret = -EIO;
goto out;
}
ret = link_to_fixup_dir(trans, root, path, location.objectid);
if (ret)
goto out;
inc_nlink(inode);
ret = btrfs_unlink_inode(trans, BTRFS_I(dir), BTRFS_I(inode), name,
name_len);
if (ret)
goto out;
ret = btrfs_run_delayed_items(trans);
if (ret)
goto out;
/*
* Unlike dir item keys, dir index keys can only have one name (entry) in
* them, as there are no key collisions since each key has a unique offset
* (an index number), so we're done.
*/
out:
btrfs_release_path(path);
btrfs_release_path(log_path);
kfree(name);
iput(inode);
return ret;
}
Btrfs: remove deleted xattrs on fsync log replay If we deleted xattrs from a file and fsynced the file, after a log replay the xattrs would remain associated to the file. This was an unexpected behaviour and differs from what other filesystems do, such as for example xfs and ext3/4. Fix this by, on fsync log replay, check if every xattr in the fs/subvol tree (that belongs to a logged inode) has a matching xattr in the log, and if it does not, delete it from the fs/subvol tree. This is a similar approach to what we do for dentries when we replay a directory from the fsync log. This issue is trivial to reproduce, and the following excerpt from my test for xfstests triggers the issue: _crash_and_mount() { # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey } rm -f $seqres.full _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create out test file and add 3 xattrs to it. touch $SCRATCH_MNT/foobar $SETFATTR_PROG -n user.attr1 -v val1 $SCRATCH_MNT/foobar $SETFATTR_PROG -n user.attr2 -v val2 $SCRATCH_MNT/foobar $SETFATTR_PROG -n user.attr3 -v val3 $SCRATCH_MNT/foobar # Make sure everything is durably persisted. sync # Now delete the second xattr and fsync the inode. $SETFATTR_PROG -x user.attr2 $SCRATCH_MNT/foobar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foobar _crash_and_mount # After the fsync log is replayed, the file should have only 2 xattrs, the ones # named user.attr1 and user.attr3. The btrfs fsync log replay bug left the file # with the 3 xattrs that we had before deleting the second one and fsyncing the # file. echo "xattr names and values after first fsync log replay:" $GETFATTR_PROG --absolute-names --dump $SCRATCH_MNT/foobar | _filter_scratch # Now write some data to our file, fsync it, remove the first xattr, add a new # hard link to our file and commit the fsync log by fsyncing some other new # file. This is to verify that after log replay our first xattr does not exist # anymore. echo "hello world!" >> $SCRATCH_MNT/foobar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foobar $SETFATTR_PROG -x user.attr1 $SCRATCH_MNT/foobar ln $SCRATCH_MNT/foobar $SCRATCH_MNT/foobar_link touch $SCRATCH_MNT/qwerty $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/qwerty _crash_and_mount # Now only the xattr with name user.attr3 should be set in our file. echo "xattr names and values after second fsync log replay:" $GETFATTR_PROG --absolute-names --dump $SCRATCH_MNT/foobar | _filter_scratch status=0 exit The expected golden output, which is produced with this patch applied or when testing against xfs or ext3/4, is: xattr names and values after first fsync log replay: # file: SCRATCH_MNT/foobar user.attr1="val1" user.attr3="val3" xattr names and values after second fsync log replay: # file: SCRATCH_MNT/foobar user.attr3="val3" Without this patch applied, the output is: xattr names and values after first fsync log replay: # file: SCRATCH_MNT/foobar user.attr1="val1" user.attr2="val2" user.attr3="val3" xattr names and values after second fsync log replay: # file: SCRATCH_MNT/foobar user.attr1="val1" user.attr2="val2" user.attr3="val3" A patch with a test case for xfstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-02-23 22:53:35 +03:00
static int replay_xattr_deletes(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_root *log,
struct btrfs_path *path,
const u64 ino)
{
struct btrfs_key search_key;
struct btrfs_path *log_path;
int i;
int nritems;
int ret;
log_path = btrfs_alloc_path();
if (!log_path)
return -ENOMEM;
search_key.objectid = ino;
search_key.type = BTRFS_XATTR_ITEM_KEY;
search_key.offset = 0;
again:
ret = btrfs_search_slot(NULL, root, &search_key, path, 0, 0);
if (ret < 0)
goto out;
process_leaf:
nritems = btrfs_header_nritems(path->nodes[0]);
for (i = path->slots[0]; i < nritems; i++) {
struct btrfs_key key;
struct btrfs_dir_item *di;
struct btrfs_dir_item *log_di;
u32 total_size;
u32 cur;
btrfs_item_key_to_cpu(path->nodes[0], &key, i);
if (key.objectid != ino || key.type != BTRFS_XATTR_ITEM_KEY) {
ret = 0;
goto out;
}
di = btrfs_item_ptr(path->nodes[0], i, struct btrfs_dir_item);
total_size = btrfs_item_size(path->nodes[0], i);
Btrfs: remove deleted xattrs on fsync log replay If we deleted xattrs from a file and fsynced the file, after a log replay the xattrs would remain associated to the file. This was an unexpected behaviour and differs from what other filesystems do, such as for example xfs and ext3/4. Fix this by, on fsync log replay, check if every xattr in the fs/subvol tree (that belongs to a logged inode) has a matching xattr in the log, and if it does not, delete it from the fs/subvol tree. This is a similar approach to what we do for dentries when we replay a directory from the fsync log. This issue is trivial to reproduce, and the following excerpt from my test for xfstests triggers the issue: _crash_and_mount() { # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey } rm -f $seqres.full _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create out test file and add 3 xattrs to it. touch $SCRATCH_MNT/foobar $SETFATTR_PROG -n user.attr1 -v val1 $SCRATCH_MNT/foobar $SETFATTR_PROG -n user.attr2 -v val2 $SCRATCH_MNT/foobar $SETFATTR_PROG -n user.attr3 -v val3 $SCRATCH_MNT/foobar # Make sure everything is durably persisted. sync # Now delete the second xattr and fsync the inode. $SETFATTR_PROG -x user.attr2 $SCRATCH_MNT/foobar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foobar _crash_and_mount # After the fsync log is replayed, the file should have only 2 xattrs, the ones # named user.attr1 and user.attr3. The btrfs fsync log replay bug left the file # with the 3 xattrs that we had before deleting the second one and fsyncing the # file. echo "xattr names and values after first fsync log replay:" $GETFATTR_PROG --absolute-names --dump $SCRATCH_MNT/foobar | _filter_scratch # Now write some data to our file, fsync it, remove the first xattr, add a new # hard link to our file and commit the fsync log by fsyncing some other new # file. This is to verify that after log replay our first xattr does not exist # anymore. echo "hello world!" >> $SCRATCH_MNT/foobar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foobar $SETFATTR_PROG -x user.attr1 $SCRATCH_MNT/foobar ln $SCRATCH_MNT/foobar $SCRATCH_MNT/foobar_link touch $SCRATCH_MNT/qwerty $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/qwerty _crash_and_mount # Now only the xattr with name user.attr3 should be set in our file. echo "xattr names and values after second fsync log replay:" $GETFATTR_PROG --absolute-names --dump $SCRATCH_MNT/foobar | _filter_scratch status=0 exit The expected golden output, which is produced with this patch applied or when testing against xfs or ext3/4, is: xattr names and values after first fsync log replay: # file: SCRATCH_MNT/foobar user.attr1="val1" user.attr3="val3" xattr names and values after second fsync log replay: # file: SCRATCH_MNT/foobar user.attr3="val3" Without this patch applied, the output is: xattr names and values after first fsync log replay: # file: SCRATCH_MNT/foobar user.attr1="val1" user.attr2="val2" user.attr3="val3" xattr names and values after second fsync log replay: # file: SCRATCH_MNT/foobar user.attr1="val1" user.attr2="val2" user.attr3="val3" A patch with a test case for xfstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-02-23 22:53:35 +03:00
cur = 0;
while (cur < total_size) {
u16 name_len = btrfs_dir_name_len(path->nodes[0], di);
u16 data_len = btrfs_dir_data_len(path->nodes[0], di);
u32 this_len = sizeof(*di) + name_len + data_len;
char *name;
name = kmalloc(name_len, GFP_NOFS);
if (!name) {
ret = -ENOMEM;
goto out;
}
read_extent_buffer(path->nodes[0], name,
(unsigned long)(di + 1), name_len);
log_di = btrfs_lookup_xattr(NULL, log, log_path, ino,
name, name_len, 0);
btrfs_release_path(log_path);
if (!log_di) {
/* Doesn't exist in log tree, so delete it. */
btrfs_release_path(path);
di = btrfs_lookup_xattr(trans, root, path, ino,
name, name_len, -1);
kfree(name);
if (IS_ERR(di)) {
ret = PTR_ERR(di);
goto out;
}
ASSERT(di);
ret = btrfs_delete_one_dir_name(trans, root,
path, di);
if (ret)
goto out;
btrfs_release_path(path);
search_key = key;
goto again;
}
kfree(name);
if (IS_ERR(log_di)) {
ret = PTR_ERR(log_di);
goto out;
}
cur += this_len;
di = (struct btrfs_dir_item *)((char *)di + this_len);
}
}
ret = btrfs_next_leaf(root, path);
if (ret > 0)
ret = 0;
else if (ret == 0)
goto process_leaf;
out:
btrfs_free_path(log_path);
btrfs_release_path(path);
return ret;
}
/*
* deletion replay happens before we copy any new directory items
* out of the log or out of backreferences from inodes. It
* scans the log to find ranges of keys that log is authoritative for,
* and then scans the directory to find items in those ranges that are
* not present in the log.
*
* Anything we don't find in the log is unlinked and removed from the
* directory.
*/
static noinline int replay_dir_deletes(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_root *log,
struct btrfs_path *path,
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 17:24:20 +03:00
u64 dirid, int del_all)
{
u64 range_start;
u64 range_end;
int ret = 0;
struct btrfs_key dir_key;
struct btrfs_key found_key;
struct btrfs_path *log_path;
struct inode *dir;
dir_key.objectid = dirid;
dir_key.type = BTRFS_DIR_INDEX_KEY;
log_path = btrfs_alloc_path();
if (!log_path)
return -ENOMEM;
dir = read_one_inode(root, dirid);
/* it isn't an error if the inode isn't there, that can happen
* because we replay the deletes before we copy in the inode item
* from the log
*/
if (!dir) {
btrfs_free_path(log_path);
return 0;
}
range_start = 0;
range_end = 0;
while (1) {
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 17:24:20 +03:00
if (del_all)
range_end = (u64)-1;
else {
ret = find_dir_range(log, path, dirid,
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 17:24:20 +03:00
&range_start, &range_end);
if (ret < 0)
goto out;
else if (ret > 0)
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 17:24:20 +03:00
break;
}
dir_key.offset = range_start;
while (1) {
int nritems;
ret = btrfs_search_slot(NULL, root, &dir_key, path,
0, 0);
if (ret < 0)
goto out;
nritems = btrfs_header_nritems(path->nodes[0]);
if (path->slots[0] >= nritems) {
ret = btrfs_next_leaf(root, path);
if (ret == 1)
break;
else if (ret < 0)
goto out;
}
btrfs_item_key_to_cpu(path->nodes[0], &found_key,
path->slots[0]);
if (found_key.objectid != dirid ||
found_key.type != dir_key.type) {
ret = 0;
goto out;
}
if (found_key.offset > range_end)
break;
ret = check_item_in_log(trans, log, path,
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 17:24:20 +03:00
log_path, dir,
&found_key);
if (ret)
goto out;
if (found_key.offset == (u64)-1)
break;
dir_key.offset = found_key.offset + 1;
}
btrfs_release_path(path);
if (range_end == (u64)-1)
break;
range_start = range_end + 1;
}
ret = 0;
out:
btrfs_release_path(path);
btrfs_free_path(log_path);
iput(dir);
return ret;
}
/*
* the process_func used to replay items from the log tree. This
* gets called in two different stages. The first stage just looks
* for inodes and makes sure they are all copied into the subvolume.
*
* The second stage copies all the other item types from the log into
* the subvolume. The two stage approach is slower, but gets rid of
* lots of complexity around inodes referencing other inodes that exist
* only in the log (references come from either directory items or inode
* back refs).
*/
static int replay_one_buffer(struct btrfs_root *log, struct extent_buffer *eb,
struct walk_control *wc, u64 gen, int level)
{
int nritems;
struct btrfs_path *path;
struct btrfs_root *root = wc->replay_dest;
struct btrfs_key key;
int i;
int ret;
ret = btrfs_read_buffer(eb, gen, level, NULL);
if (ret)
return ret;
level = btrfs_header_level(eb);
if (level != 0)
return 0;
path = btrfs_alloc_path();
if (!path)
return -ENOMEM;
nritems = btrfs_header_nritems(eb);
for (i = 0; i < nritems; i++) {
btrfs_item_key_to_cpu(eb, &key, i);
/* inode keys are done during the first stage */
if (key.type == BTRFS_INODE_ITEM_KEY &&
wc->stage == LOG_WALK_REPLAY_INODES) {
struct btrfs_inode_item *inode_item;
u32 mode;
inode_item = btrfs_item_ptr(eb, i,
struct btrfs_inode_item);
Btrfs: fix warning when replaying log after fsync of a tmpfile When replaying a log which contains a tmpfile (which necessarily has a link count of 0) we end up calling inc_nlink(), at fs/btrfs/tree-log.c:replay_one_buffer(), which produces a warning like the following: [195191.943673] WARNING: CPU: 0 PID: 6924 at fs/inode.c:342 inc_nlink+0x33/0x40 [195191.943723] CPU: 0 PID: 6924 Comm: mount Not tainted 4.19.0-rc6-btrfs-next-38 #1 [195191.943724] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.11.2-0-gf9626ccb91-prebuilt.qemu-project.org 04/01/2014 [195191.943726] RIP: 0010:inc_nlink+0x33/0x40 [195191.943728] RSP: 0018:ffffb96e425e3870 EFLAGS: 00010246 [195191.943730] RAX: 0000000000000000 RBX: ffff8c0d1e6af4f0 RCX: 0000000000000006 [195191.943731] RDX: 0000000000000000 RSI: 0000000000000000 RDI: ffff8c0d1e6af4f0 [195191.943731] RBP: 0000000000000097 R08: 0000000000000001 R09: 0000000000000000 [195191.943732] R10: 0000000000000000 R11: 0000000000000000 R12: ffffb96e425e3a60 [195191.943733] R13: ffff8c0d10cff0c8 R14: ffff8c0d0d515348 R15: ffff8c0d78a1b3f8 [195191.943735] FS: 00007f570ee24480(0000) GS:ffff8c0dfb200000(0000) knlGS:0000000000000000 [195191.943736] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [195191.943737] CR2: 00005593286277c8 CR3: 00000000bb8f2006 CR4: 00000000003606f0 [195191.943739] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [195191.943740] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [195191.943741] Call Trace: [195191.943778] replay_one_buffer+0x797/0x7d0 [btrfs] [195191.943802] walk_up_log_tree+0x1c1/0x250 [btrfs] [195191.943809] ? rcu_read_lock_sched_held+0x3f/0x70 [195191.943825] walk_log_tree+0xae/0x1d0 [btrfs] [195191.943840] btrfs_recover_log_trees+0x1d7/0x4d0 [btrfs] [195191.943856] ? replay_dir_deletes+0x280/0x280 [btrfs] [195191.943870] open_ctree+0x1c3b/0x22a0 [btrfs] [195191.943887] btrfs_mount_root+0x6b4/0x800 [btrfs] [195191.943894] ? rcu_read_lock_sched_held+0x3f/0x70 [195191.943899] ? pcpu_alloc+0x55b/0x7c0 [195191.943906] ? mount_fs+0x3b/0x140 [195191.943908] mount_fs+0x3b/0x140 [195191.943912] ? __init_waitqueue_head+0x36/0x50 [195191.943916] vfs_kern_mount+0x62/0x160 [195191.943927] btrfs_mount+0x134/0x890 [btrfs] [195191.943936] ? rcu_read_lock_sched_held+0x3f/0x70 [195191.943938] ? pcpu_alloc+0x55b/0x7c0 [195191.943943] ? mount_fs+0x3b/0x140 [195191.943952] ? btrfs_remount+0x570/0x570 [btrfs] [195191.943954] mount_fs+0x3b/0x140 [195191.943956] ? __init_waitqueue_head+0x36/0x50 [195191.943960] vfs_kern_mount+0x62/0x160 [195191.943963] do_mount+0x1f9/0xd40 [195191.943967] ? memdup_user+0x4b/0x70 [195191.943971] ksys_mount+0x7e/0xd0 [195191.943974] __x64_sys_mount+0x21/0x30 [195191.943977] do_syscall_64+0x60/0x1b0 [195191.943980] entry_SYSCALL_64_after_hwframe+0x49/0xbe [195191.943983] RIP: 0033:0x7f570e4e524a [195191.943986] RSP: 002b:00007ffd83589478 EFLAGS: 00000206 ORIG_RAX: 00000000000000a5 [195191.943989] RAX: ffffffffffffffda RBX: 0000563f335b2060 RCX: 00007f570e4e524a [195191.943990] RDX: 0000563f335b2240 RSI: 0000563f335b2280 RDI: 0000563f335b2260 [195191.943992] RBP: 0000000000000000 R08: 0000000000000000 R09: 0000000000000020 [195191.943993] R10: 00000000c0ed0000 R11: 0000000000000206 R12: 0000563f335b2260 [195191.943994] R13: 0000563f335b2240 R14: 0000000000000000 R15: 00000000ffffffff [195191.944002] irq event stamp: 8688 [195191.944010] hardirqs last enabled at (8687): [<ffffffff9cb004c3>] console_unlock+0x503/0x640 [195191.944012] hardirqs last disabled at (8688): [<ffffffff9ca037dd>] trace_hardirqs_off_thunk+0x1a/0x1c [195191.944018] softirqs last enabled at (8638): [<ffffffff9cc0a5d1>] __set_page_dirty_nobuffers+0x101/0x150 [195191.944020] softirqs last disabled at (8634): [<ffffffff9cc26bbe>] wb_wakeup_delayed+0x2e/0x60 [195191.944022] ---[ end trace 5d6e873a9a0b811a ]--- This happens because the inode does not have the flag I_LINKABLE set, which is a runtime only flag, not meant to be persisted, set when the inode is created through open(2) if the flag O_EXCL is not passed to it. Except for the warning, there are no other consequences (like corruptions or metadata inconsistencies). Since it's pointless to replay a tmpfile as it would be deleted in a later phase of the log replay procedure (it has a link count of 0), fix this by not logging tmpfiles and if a tmpfile is found in a log (created by a kernel without this change), skip the replay of the inode. A test case for fstests follows soon. Fixes: 471d557afed1 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay") CC: stable@vger.kernel.org # 4.18+ Reported-by: Martin Steigerwald <martin@lichtvoll.de> Link: https://lore.kernel.org/linux-btrfs/3666619.NTnn27ZJZE@merkaba/ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2018-10-08 13:12:55 +03:00
/*
* If we have a tmpfile (O_TMPFILE) that got fsync'ed
* and never got linked before the fsync, skip it, as
* replaying it is pointless since it would be deleted
* later. We skip logging tmpfiles, but it's always
* possible we are replaying a log created with a kernel
* that used to log tmpfiles.
*/
if (btrfs_inode_nlink(eb, inode_item) == 0) {
wc->ignore_cur_inode = true;
continue;
} else {
wc->ignore_cur_inode = false;
}
Btrfs: remove deleted xattrs on fsync log replay If we deleted xattrs from a file and fsynced the file, after a log replay the xattrs would remain associated to the file. This was an unexpected behaviour and differs from what other filesystems do, such as for example xfs and ext3/4. Fix this by, on fsync log replay, check if every xattr in the fs/subvol tree (that belongs to a logged inode) has a matching xattr in the log, and if it does not, delete it from the fs/subvol tree. This is a similar approach to what we do for dentries when we replay a directory from the fsync log. This issue is trivial to reproduce, and the following excerpt from my test for xfstests triggers the issue: _crash_and_mount() { # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey } rm -f $seqres.full _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create out test file and add 3 xattrs to it. touch $SCRATCH_MNT/foobar $SETFATTR_PROG -n user.attr1 -v val1 $SCRATCH_MNT/foobar $SETFATTR_PROG -n user.attr2 -v val2 $SCRATCH_MNT/foobar $SETFATTR_PROG -n user.attr3 -v val3 $SCRATCH_MNT/foobar # Make sure everything is durably persisted. sync # Now delete the second xattr and fsync the inode. $SETFATTR_PROG -x user.attr2 $SCRATCH_MNT/foobar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foobar _crash_and_mount # After the fsync log is replayed, the file should have only 2 xattrs, the ones # named user.attr1 and user.attr3. The btrfs fsync log replay bug left the file # with the 3 xattrs that we had before deleting the second one and fsyncing the # file. echo "xattr names and values after first fsync log replay:" $GETFATTR_PROG --absolute-names --dump $SCRATCH_MNT/foobar | _filter_scratch # Now write some data to our file, fsync it, remove the first xattr, add a new # hard link to our file and commit the fsync log by fsyncing some other new # file. This is to verify that after log replay our first xattr does not exist # anymore. echo "hello world!" >> $SCRATCH_MNT/foobar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foobar $SETFATTR_PROG -x user.attr1 $SCRATCH_MNT/foobar ln $SCRATCH_MNT/foobar $SCRATCH_MNT/foobar_link touch $SCRATCH_MNT/qwerty $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/qwerty _crash_and_mount # Now only the xattr with name user.attr3 should be set in our file. echo "xattr names and values after second fsync log replay:" $GETFATTR_PROG --absolute-names --dump $SCRATCH_MNT/foobar | _filter_scratch status=0 exit The expected golden output, which is produced with this patch applied or when testing against xfs or ext3/4, is: xattr names and values after first fsync log replay: # file: SCRATCH_MNT/foobar user.attr1="val1" user.attr3="val3" xattr names and values after second fsync log replay: # file: SCRATCH_MNT/foobar user.attr3="val3" Without this patch applied, the output is: xattr names and values after first fsync log replay: # file: SCRATCH_MNT/foobar user.attr1="val1" user.attr2="val2" user.attr3="val3" xattr names and values after second fsync log replay: # file: SCRATCH_MNT/foobar user.attr1="val1" user.attr2="val2" user.attr3="val3" A patch with a test case for xfstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-02-23 22:53:35 +03:00
ret = replay_xattr_deletes(wc->trans, root, log,
path, key.objectid);
if (ret)
break;
mode = btrfs_inode_mode(eb, inode_item);
if (S_ISDIR(mode)) {
ret = replay_dir_deletes(wc->trans,
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 17:24:20 +03:00
root, log, path, key.objectid, 0);
if (ret)
break;
}
ret = overwrite_item(wc->trans, root, path,
eb, i, &key);
if (ret)
break;
Btrfs: fix loss of prealloc extents past i_size after fsync log replay Currently if we allocate extents beyond an inode's i_size (through the fallocate system call) and then fsync the file, we log the extents but after a power failure we replay them and then immediately drop them. This behaviour happens since about 2009, commit c71bf099abdd ("Btrfs: Avoid orphan inodes cleanup while replaying log"), because it marks the inode as an orphan instead of dropping any extents beyond i_size before replaying logged extents, so after the log replay, and while the mount operation is still ongoing, we find the inode marked as an orphan and then perform a truncation (drop extents beyond the inode's i_size). Because the processing of orphan inodes is still done right after replaying the log and before the mount operation finishes, the intention of that commit does not make any sense (at least as of today). However reverting that behaviour is not enough, because we can not simply discard all extents beyond i_size and then replay logged extents, because we risk dropping extents beyond i_size created in past transactions, for example: add prealloc extent beyond i_size fsync - clears the flag BTRFS_INODE_NEEDS_FULL_SYNC from the inode transaction commit add another prealloc extent beyond i_size fsync - triggers the fast fsync path power failure In that scenario, we would drop the first extent and then replay the second one. To fix this just make sure that all prealloc extents beyond i_size are logged, and if we find too many (which is far from a common case), fallback to a full transaction commit (like we do when logging regular extents in the fast fsync path). Trivial reproducer: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ xfs_io -f -c "pwrite -S 0xab 0 256K" /mnt/foo $ sync $ xfs_io -c "falloc -k 256K 1M" /mnt/foo $ xfs_io -c "fsync" /mnt/foo <power failure> # mount to replay log $ mount /dev/sdb /mnt # at this point the file only has one extent, at offset 0, size 256K A test case for fstests follows soon, covering multiple scenarios that involve adding prealloc extents with previous shrinking truncates and without such truncates. Fixes: c71bf099abdd ("Btrfs: Avoid orphan inodes cleanup while replaying log") Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2018-04-06 00:55:12 +03:00
/*
* Before replaying extents, truncate the inode to its
* size. We need to do it now and not after log replay
* because before an fsync we can have prealloc extents
* added beyond the inode's i_size. If we did it after,
* through orphan cleanup for example, we would drop
* those prealloc extents just after replaying them.
*/
if (S_ISREG(mode)) {
struct btrfs_drop_extents_args drop_args = { 0 };
Btrfs: fix loss of prealloc extents past i_size after fsync log replay Currently if we allocate extents beyond an inode's i_size (through the fallocate system call) and then fsync the file, we log the extents but after a power failure we replay them and then immediately drop them. This behaviour happens since about 2009, commit c71bf099abdd ("Btrfs: Avoid orphan inodes cleanup while replaying log"), because it marks the inode as an orphan instead of dropping any extents beyond i_size before replaying logged extents, so after the log replay, and while the mount operation is still ongoing, we find the inode marked as an orphan and then perform a truncation (drop extents beyond the inode's i_size). Because the processing of orphan inodes is still done right after replaying the log and before the mount operation finishes, the intention of that commit does not make any sense (at least as of today). However reverting that behaviour is not enough, because we can not simply discard all extents beyond i_size and then replay logged extents, because we risk dropping extents beyond i_size created in past transactions, for example: add prealloc extent beyond i_size fsync - clears the flag BTRFS_INODE_NEEDS_FULL_SYNC from the inode transaction commit add another prealloc extent beyond i_size fsync - triggers the fast fsync path power failure In that scenario, we would drop the first extent and then replay the second one. To fix this just make sure that all prealloc extents beyond i_size are logged, and if we find too many (which is far from a common case), fallback to a full transaction commit (like we do when logging regular extents in the fast fsync path). Trivial reproducer: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ xfs_io -f -c "pwrite -S 0xab 0 256K" /mnt/foo $ sync $ xfs_io -c "falloc -k 256K 1M" /mnt/foo $ xfs_io -c "fsync" /mnt/foo <power failure> # mount to replay log $ mount /dev/sdb /mnt # at this point the file only has one extent, at offset 0, size 256K A test case for fstests follows soon, covering multiple scenarios that involve adding prealloc extents with previous shrinking truncates and without such truncates. Fixes: c71bf099abdd ("Btrfs: Avoid orphan inodes cleanup while replaying log") Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2018-04-06 00:55:12 +03:00
struct inode *inode;
u64 from;
inode = read_one_inode(root, key.objectid);
if (!inode) {
ret = -EIO;
break;
}
from = ALIGN(i_size_read(inode),
root->fs_info->sectorsize);
drop_args.start = from;
drop_args.end = (u64)-1;
drop_args.drop_cache = true;
ret = btrfs_drop_extents(wc->trans, root,
BTRFS_I(inode),
&drop_args);
Btrfs: fix loss of prealloc extents past i_size after fsync log replay Currently if we allocate extents beyond an inode's i_size (through the fallocate system call) and then fsync the file, we log the extents but after a power failure we replay them and then immediately drop them. This behaviour happens since about 2009, commit c71bf099abdd ("Btrfs: Avoid orphan inodes cleanup while replaying log"), because it marks the inode as an orphan instead of dropping any extents beyond i_size before replaying logged extents, so after the log replay, and while the mount operation is still ongoing, we find the inode marked as an orphan and then perform a truncation (drop extents beyond the inode's i_size). Because the processing of orphan inodes is still done right after replaying the log and before the mount operation finishes, the intention of that commit does not make any sense (at least as of today). However reverting that behaviour is not enough, because we can not simply discard all extents beyond i_size and then replay logged extents, because we risk dropping extents beyond i_size created in past transactions, for example: add prealloc extent beyond i_size fsync - clears the flag BTRFS_INODE_NEEDS_FULL_SYNC from the inode transaction commit add another prealloc extent beyond i_size fsync - triggers the fast fsync path power failure In that scenario, we would drop the first extent and then replay the second one. To fix this just make sure that all prealloc extents beyond i_size are logged, and if we find too many (which is far from a common case), fallback to a full transaction commit (like we do when logging regular extents in the fast fsync path). Trivial reproducer: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ xfs_io -f -c "pwrite -S 0xab 0 256K" /mnt/foo $ sync $ xfs_io -c "falloc -k 256K 1M" /mnt/foo $ xfs_io -c "fsync" /mnt/foo <power failure> # mount to replay log $ mount /dev/sdb /mnt # at this point the file only has one extent, at offset 0, size 256K A test case for fstests follows soon, covering multiple scenarios that involve adding prealloc extents with previous shrinking truncates and without such truncates. Fixes: c71bf099abdd ("Btrfs: Avoid orphan inodes cleanup while replaying log") Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2018-04-06 00:55:12 +03:00
if (!ret) {
btrfs: update the number of bytes used by an inode atomically There are several occasions where we do not update the inode's number of used bytes atomically, resulting in a concurrent stat(2) syscall to report a value of used blocks that does not correspond to a valid value, that is, a value that does not match neither what we had before the operation nor what we get after the operation completes. In extreme cases it can result in stat(2) reporting zero used blocks, which can cause problems for some userspace tools where they can consider a file with a non-zero size and zero used blocks as completely sparse and skip reading data, as reported/discussed a long time ago in some threads like the following: https://lists.gnu.org/archive/html/bug-tar/2016-07/msg00001.html The cases where this can happen are the following: -> Case 1 If we do a write (buffered or direct IO) against a file region for which there is already an allocated extent (or multiple extents), then we have a short time window where we can report a number of used blocks to stat(2) that does not take into account the file region being overwritten. This short time window happens when completing the ordered extent(s). This happens because when we drop the extents in the write range we decrement the inode's number of bytes and later on when we insert the new extent(s) we increment the number of bytes in the inode, resulting in a short time window where a stat(2) syscall can get an incorrect number of used blocks. If we do writes that overwrite an entire file, then we have a short time window where we report 0 used blocks to stat(2). Example reproducer: $ cat reproducer-1.sh #!/bin/bash MNT=/mnt/sdi DEV=/dev/sdi stat_loop() { trap "wait; exit" SIGTERM local filepath=$1 local expected=$2 local got while :; do got=$(stat -c %b $filepath) if [ $got -ne $expected ]; then echo -n "ERROR: unexpected used blocks" echo " (got: $got expected: $expected)" fi done } mkfs.btrfs -f $DEV > /dev/null # mkfs.xfs -f $DEV > /dev/null # mkfs.ext4 -F $DEV > /dev/null # mkfs.f2fs -f $DEV > /dev/null # mkfs.reiserfs -f $DEV > /dev/null mount $DEV $MNT xfs_io -f -s -c "pwrite -b 64K 0 64K" $MNT/foobar >/dev/null expected=$(stat -c %b $MNT/foobar) # Create a process to keep calling stat(2) on the file and see if the # reported number of blocks used (disk space used) changes, it should # not because we are not increasing the file size nor punching holes. stat_loop $MNT/foobar $expected & loop_pid=$! for ((i = 0; i < 50000; i++)); do xfs_io -s -c "pwrite -b 64K 0 64K" $MNT/foobar >/dev/null done kill $loop_pid &> /dev/null wait umount $DEV $ ./reproducer-1.sh ERROR: unexpected used blocks (got: 0 expected: 128) ERROR: unexpected used blocks (got: 0 expected: 128) (...) Note that since this is a short time window where the race can happen, the reproducer may not be able to always trigger the bug in one run, or it may trigger it multiple times. -> Case 2 If we do a buffered write against a file region that does not have any allocated extents, like a hole or beyond EOF, then during ordered extent completion we have a short time window where a concurrent stat(2) syscall can report a number of used blocks that does not correspond to the value before or after the write operation, a value that is actually larger than the value after the write completes. This happens because once we start a buffered write into an unallocated file range we increment the inode's 'new_delalloc_bytes', to make sure any stat(2) call gets a correct used blocks value before delalloc is flushed and completes. However at ordered extent completion, after we inserted the new extent, we increment the inode's number of bytes used with the size of the new extent, and only later, when clearing the range in the inode's iotree, we decrement the inode's 'new_delalloc_bytes' counter with the size of the extent. So this results in a short time window where a concurrent stat(2) syscall can report a number of used blocks that accounts for the new extent twice. Example reproducer: $ cat reproducer-2.sh #!/bin/bash MNT=/mnt/sdi DEV=/dev/sdi stat_loop() { trap "wait; exit" SIGTERM local filepath=$1 local expected=$2 local got while :; do got=$(stat -c %b $filepath) if [ $got -ne $expected ]; then echo -n "ERROR: unexpected used blocks" echo " (got: $got expected: $expected)" fi done } mkfs.btrfs -f $DEV > /dev/null # mkfs.xfs -f $DEV > /dev/null # mkfs.ext4 -F $DEV > /dev/null # mkfs.f2fs -f $DEV > /dev/null # mkfs.reiserfs -f $DEV > /dev/null mount $DEV $MNT touch $MNT/foobar write_size=$((64 * 1024)) for ((i = 0; i < 16384; i++)); do offset=$(($i * $write_size)) xfs_io -c "pwrite -S 0xab $offset $write_size" $MNT/foobar >/dev/null blocks_used=$(stat -c %b $MNT/foobar) # Fsync the file to trigger writeback and keep calling stat(2) on it # to see if the number of blocks used changes. stat_loop $MNT/foobar $blocks_used & loop_pid=$! xfs_io -c "fsync" $MNT/foobar kill $loop_pid &> /dev/null wait $loop_pid done umount $DEV $ ./reproducer-2.sh ERROR: unexpected used blocks (got: 265472 expected: 265344) ERROR: unexpected used blocks (got: 284032 expected: 283904) (...) Note that since this is a short time window where the race can happen, the reproducer may not be able to always trigger the bug in one run, or it may trigger it multiple times. -> Case 3 Another case where such problems happen is during other operations that replace extents in a file range with other extents. Those operations are extent cloning, deduplication and fallocate's zero range operation. The cause of the problem is similar to the first case. When we drop the extents from a range, we decrement the inode's number of bytes, and later on, after inserting the new extents we increment it. Since this is not done atomically, a concurrent stat(2) call can see and return a number of used blocks that is smaller than it should be, does not match the number of used blocks before or after the clone/deduplication/zero operation. Like for the first case, when doing a clone, deduplication or zero range operation against an entire file, we end up having a time window where we can report 0 used blocks to a stat(2) call. Example reproducer: $ cat reproducer-3.sh #!/bin/bash MNT=/mnt/sdi DEV=/dev/sdi mkfs.btrfs -f $DEV > /dev/null # mkfs.xfs -f -m reflink=1 $DEV > /dev/null mount $DEV $MNT extent_size=$((64 * 1024)) num_extents=16384 file_size=$(($extent_size * $num_extents)) # File foo has many small extents. xfs_io -f -s -c "pwrite -S 0xab -b $extent_size 0 $file_size" $MNT/foo \ > /dev/null # File bar has much less extents and has exactly the same data as foo. xfs_io -f -c "pwrite -S 0xab 0 $file_size" $MNT/bar > /dev/null expected=$(stat -c %b $MNT/foo) # Now deduplicate bar into foo. While the deduplication is in progres, # the number of used blocks/file size reported by stat should not change xfs_io -c "dedupe $MNT/bar 0 0 $file_size" $MNT/foo > /dev/null & dedupe_pid=$! while [ -n "$(ps -p $dedupe_pid -o pid=)" ]; do used=$(stat -c %b $MNT/foo) if [ $used -ne $expected ]; then echo "Unexpected blocks used: $used (expected: $expected)" fi done umount $DEV $ ./reproducer-3.sh Unexpected blocks used: 2076800 (expected: 2097152) Unexpected blocks used: 2097024 (expected: 2097152) Unexpected blocks used: 2079872 (expected: 2097152) (...) Note that since this is a short time window where the race can happen, the reproducer may not be able to always trigger the bug in one run, or it may trigger it multiple times. So fix this by: 1) Making btrfs_drop_extents() not decrement the VFS inode's number of bytes, and instead return the number of bytes; 2) Making any code that drops extents and adds new extents update the inode's number of bytes atomically, while holding the btrfs inode's spinlock, which is also used by the stat(2) callback to get the inode's number of bytes; 3) For ranges in the inode's iotree that are marked as 'delalloc new', corresponding to previously unallocated ranges, increment the inode's number of bytes when clearing the 'delalloc new' bit from the range, in the same critical section that decrements the inode's 'new_delalloc_bytes' counter, delimited by the btrfs inode's spinlock. An alternative would be to have btrfs_getattr() wait for any IO (ordered extents in progress) and locking the whole range (0 to (u64)-1) while it it computes the number of blocks used. But that would mean blocking stat(2), which is a very used syscall and expected to be fast, waiting for writes, clone/dedupe, fallocate, page reads, fiemap, etc. CC: stable@vger.kernel.org # 5.4+ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-11-04 14:07:34 +03:00
inode_sub_bytes(inode,
drop_args.bytes_found);
Btrfs: fix warning when replaying log after fsync of a tmpfile When replaying a log which contains a tmpfile (which necessarily has a link count of 0) we end up calling inc_nlink(), at fs/btrfs/tree-log.c:replay_one_buffer(), which produces a warning like the following: [195191.943673] WARNING: CPU: 0 PID: 6924 at fs/inode.c:342 inc_nlink+0x33/0x40 [195191.943723] CPU: 0 PID: 6924 Comm: mount Not tainted 4.19.0-rc6-btrfs-next-38 #1 [195191.943724] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.11.2-0-gf9626ccb91-prebuilt.qemu-project.org 04/01/2014 [195191.943726] RIP: 0010:inc_nlink+0x33/0x40 [195191.943728] RSP: 0018:ffffb96e425e3870 EFLAGS: 00010246 [195191.943730] RAX: 0000000000000000 RBX: ffff8c0d1e6af4f0 RCX: 0000000000000006 [195191.943731] RDX: 0000000000000000 RSI: 0000000000000000 RDI: ffff8c0d1e6af4f0 [195191.943731] RBP: 0000000000000097 R08: 0000000000000001 R09: 0000000000000000 [195191.943732] R10: 0000000000000000 R11: 0000000000000000 R12: ffffb96e425e3a60 [195191.943733] R13: ffff8c0d10cff0c8 R14: ffff8c0d0d515348 R15: ffff8c0d78a1b3f8 [195191.943735] FS: 00007f570ee24480(0000) GS:ffff8c0dfb200000(0000) knlGS:0000000000000000 [195191.943736] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [195191.943737] CR2: 00005593286277c8 CR3: 00000000bb8f2006 CR4: 00000000003606f0 [195191.943739] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [195191.943740] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [195191.943741] Call Trace: [195191.943778] replay_one_buffer+0x797/0x7d0 [btrfs] [195191.943802] walk_up_log_tree+0x1c1/0x250 [btrfs] [195191.943809] ? rcu_read_lock_sched_held+0x3f/0x70 [195191.943825] walk_log_tree+0xae/0x1d0 [btrfs] [195191.943840] btrfs_recover_log_trees+0x1d7/0x4d0 [btrfs] [195191.943856] ? replay_dir_deletes+0x280/0x280 [btrfs] [195191.943870] open_ctree+0x1c3b/0x22a0 [btrfs] [195191.943887] btrfs_mount_root+0x6b4/0x800 [btrfs] [195191.943894] ? rcu_read_lock_sched_held+0x3f/0x70 [195191.943899] ? pcpu_alloc+0x55b/0x7c0 [195191.943906] ? mount_fs+0x3b/0x140 [195191.943908] mount_fs+0x3b/0x140 [195191.943912] ? __init_waitqueue_head+0x36/0x50 [195191.943916] vfs_kern_mount+0x62/0x160 [195191.943927] btrfs_mount+0x134/0x890 [btrfs] [195191.943936] ? rcu_read_lock_sched_held+0x3f/0x70 [195191.943938] ? pcpu_alloc+0x55b/0x7c0 [195191.943943] ? mount_fs+0x3b/0x140 [195191.943952] ? btrfs_remount+0x570/0x570 [btrfs] [195191.943954] mount_fs+0x3b/0x140 [195191.943956] ? __init_waitqueue_head+0x36/0x50 [195191.943960] vfs_kern_mount+0x62/0x160 [195191.943963] do_mount+0x1f9/0xd40 [195191.943967] ? memdup_user+0x4b/0x70 [195191.943971] ksys_mount+0x7e/0xd0 [195191.943974] __x64_sys_mount+0x21/0x30 [195191.943977] do_syscall_64+0x60/0x1b0 [195191.943980] entry_SYSCALL_64_after_hwframe+0x49/0xbe [195191.943983] RIP: 0033:0x7f570e4e524a [195191.943986] RSP: 002b:00007ffd83589478 EFLAGS: 00000206 ORIG_RAX: 00000000000000a5 [195191.943989] RAX: ffffffffffffffda RBX: 0000563f335b2060 RCX: 00007f570e4e524a [195191.943990] RDX: 0000563f335b2240 RSI: 0000563f335b2280 RDI: 0000563f335b2260 [195191.943992] RBP: 0000000000000000 R08: 0000000000000000 R09: 0000000000000020 [195191.943993] R10: 00000000c0ed0000 R11: 0000000000000206 R12: 0000563f335b2260 [195191.943994] R13: 0000563f335b2240 R14: 0000000000000000 R15: 00000000ffffffff [195191.944002] irq event stamp: 8688 [195191.944010] hardirqs last enabled at (8687): [<ffffffff9cb004c3>] console_unlock+0x503/0x640 [195191.944012] hardirqs last disabled at (8688): [<ffffffff9ca037dd>] trace_hardirqs_off_thunk+0x1a/0x1c [195191.944018] softirqs last enabled at (8638): [<ffffffff9cc0a5d1>] __set_page_dirty_nobuffers+0x101/0x150 [195191.944020] softirqs last disabled at (8634): [<ffffffff9cc26bbe>] wb_wakeup_delayed+0x2e/0x60 [195191.944022] ---[ end trace 5d6e873a9a0b811a ]--- This happens because the inode does not have the flag I_LINKABLE set, which is a runtime only flag, not meant to be persisted, set when the inode is created through open(2) if the flag O_EXCL is not passed to it. Except for the warning, there are no other consequences (like corruptions or metadata inconsistencies). Since it's pointless to replay a tmpfile as it would be deleted in a later phase of the log replay procedure (it has a link count of 0), fix this by not logging tmpfiles and if a tmpfile is found in a log (created by a kernel without this change), skip the replay of the inode. A test case for fstests follows soon. Fixes: 471d557afed1 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay") CC: stable@vger.kernel.org # 4.18+ Reported-by: Martin Steigerwald <martin@lichtvoll.de> Link: https://lore.kernel.org/linux-btrfs/3666619.NTnn27ZJZE@merkaba/ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2018-10-08 13:12:55 +03:00
/* Update the inode's nbytes. */
Btrfs: fix loss of prealloc extents past i_size after fsync log replay Currently if we allocate extents beyond an inode's i_size (through the fallocate system call) and then fsync the file, we log the extents but after a power failure we replay them and then immediately drop them. This behaviour happens since about 2009, commit c71bf099abdd ("Btrfs: Avoid orphan inodes cleanup while replaying log"), because it marks the inode as an orphan instead of dropping any extents beyond i_size before replaying logged extents, so after the log replay, and while the mount operation is still ongoing, we find the inode marked as an orphan and then perform a truncation (drop extents beyond the inode's i_size). Because the processing of orphan inodes is still done right after replaying the log and before the mount operation finishes, the intention of that commit does not make any sense (at least as of today). However reverting that behaviour is not enough, because we can not simply discard all extents beyond i_size and then replay logged extents, because we risk dropping extents beyond i_size created in past transactions, for example: add prealloc extent beyond i_size fsync - clears the flag BTRFS_INODE_NEEDS_FULL_SYNC from the inode transaction commit add another prealloc extent beyond i_size fsync - triggers the fast fsync path power failure In that scenario, we would drop the first extent and then replay the second one. To fix this just make sure that all prealloc extents beyond i_size are logged, and if we find too many (which is far from a common case), fallback to a full transaction commit (like we do when logging regular extents in the fast fsync path). Trivial reproducer: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ xfs_io -f -c "pwrite -S 0xab 0 256K" /mnt/foo $ sync $ xfs_io -c "falloc -k 256K 1M" /mnt/foo $ xfs_io -c "fsync" /mnt/foo <power failure> # mount to replay log $ mount /dev/sdb /mnt # at this point the file only has one extent, at offset 0, size 256K A test case for fstests follows soon, covering multiple scenarios that involve adding prealloc extents with previous shrinking truncates and without such truncates. Fixes: c71bf099abdd ("Btrfs: Avoid orphan inodes cleanup while replaying log") Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2018-04-06 00:55:12 +03:00
ret = btrfs_update_inode(wc->trans,
root, BTRFS_I(inode));
Btrfs: fix loss of prealloc extents past i_size after fsync log replay Currently if we allocate extents beyond an inode's i_size (through the fallocate system call) and then fsync the file, we log the extents but after a power failure we replay them and then immediately drop them. This behaviour happens since about 2009, commit c71bf099abdd ("Btrfs: Avoid orphan inodes cleanup while replaying log"), because it marks the inode as an orphan instead of dropping any extents beyond i_size before replaying logged extents, so after the log replay, and while the mount operation is still ongoing, we find the inode marked as an orphan and then perform a truncation (drop extents beyond the inode's i_size). Because the processing of orphan inodes is still done right after replaying the log and before the mount operation finishes, the intention of that commit does not make any sense (at least as of today). However reverting that behaviour is not enough, because we can not simply discard all extents beyond i_size and then replay logged extents, because we risk dropping extents beyond i_size created in past transactions, for example: add prealloc extent beyond i_size fsync - clears the flag BTRFS_INODE_NEEDS_FULL_SYNC from the inode transaction commit add another prealloc extent beyond i_size fsync - triggers the fast fsync path power failure In that scenario, we would drop the first extent and then replay the second one. To fix this just make sure that all prealloc extents beyond i_size are logged, and if we find too many (which is far from a common case), fallback to a full transaction commit (like we do when logging regular extents in the fast fsync path). Trivial reproducer: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ xfs_io -f -c "pwrite -S 0xab 0 256K" /mnt/foo $ sync $ xfs_io -c "falloc -k 256K 1M" /mnt/foo $ xfs_io -c "fsync" /mnt/foo <power failure> # mount to replay log $ mount /dev/sdb /mnt # at this point the file only has one extent, at offset 0, size 256K A test case for fstests follows soon, covering multiple scenarios that involve adding prealloc extents with previous shrinking truncates and without such truncates. Fixes: c71bf099abdd ("Btrfs: Avoid orphan inodes cleanup while replaying log") Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2018-04-06 00:55:12 +03:00
}
iput(inode);
if (ret)
break;
}
ret = link_to_fixup_dir(wc->trans, root,
path, key.objectid);
if (ret)
break;
}
Btrfs: fix warning when replaying log after fsync of a tmpfile When replaying a log which contains a tmpfile (which necessarily has a link count of 0) we end up calling inc_nlink(), at fs/btrfs/tree-log.c:replay_one_buffer(), which produces a warning like the following: [195191.943673] WARNING: CPU: 0 PID: 6924 at fs/inode.c:342 inc_nlink+0x33/0x40 [195191.943723] CPU: 0 PID: 6924 Comm: mount Not tainted 4.19.0-rc6-btrfs-next-38 #1 [195191.943724] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.11.2-0-gf9626ccb91-prebuilt.qemu-project.org 04/01/2014 [195191.943726] RIP: 0010:inc_nlink+0x33/0x40 [195191.943728] RSP: 0018:ffffb96e425e3870 EFLAGS: 00010246 [195191.943730] RAX: 0000000000000000 RBX: ffff8c0d1e6af4f0 RCX: 0000000000000006 [195191.943731] RDX: 0000000000000000 RSI: 0000000000000000 RDI: ffff8c0d1e6af4f0 [195191.943731] RBP: 0000000000000097 R08: 0000000000000001 R09: 0000000000000000 [195191.943732] R10: 0000000000000000 R11: 0000000000000000 R12: ffffb96e425e3a60 [195191.943733] R13: ffff8c0d10cff0c8 R14: ffff8c0d0d515348 R15: ffff8c0d78a1b3f8 [195191.943735] FS: 00007f570ee24480(0000) GS:ffff8c0dfb200000(0000) knlGS:0000000000000000 [195191.943736] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [195191.943737] CR2: 00005593286277c8 CR3: 00000000bb8f2006 CR4: 00000000003606f0 [195191.943739] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [195191.943740] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [195191.943741] Call Trace: [195191.943778] replay_one_buffer+0x797/0x7d0 [btrfs] [195191.943802] walk_up_log_tree+0x1c1/0x250 [btrfs] [195191.943809] ? rcu_read_lock_sched_held+0x3f/0x70 [195191.943825] walk_log_tree+0xae/0x1d0 [btrfs] [195191.943840] btrfs_recover_log_trees+0x1d7/0x4d0 [btrfs] [195191.943856] ? replay_dir_deletes+0x280/0x280 [btrfs] [195191.943870] open_ctree+0x1c3b/0x22a0 [btrfs] [195191.943887] btrfs_mount_root+0x6b4/0x800 [btrfs] [195191.943894] ? rcu_read_lock_sched_held+0x3f/0x70 [195191.943899] ? pcpu_alloc+0x55b/0x7c0 [195191.943906] ? mount_fs+0x3b/0x140 [195191.943908] mount_fs+0x3b/0x140 [195191.943912] ? __init_waitqueue_head+0x36/0x50 [195191.943916] vfs_kern_mount+0x62/0x160 [195191.943927] btrfs_mount+0x134/0x890 [btrfs] [195191.943936] ? rcu_read_lock_sched_held+0x3f/0x70 [195191.943938] ? pcpu_alloc+0x55b/0x7c0 [195191.943943] ? mount_fs+0x3b/0x140 [195191.943952] ? btrfs_remount+0x570/0x570 [btrfs] [195191.943954] mount_fs+0x3b/0x140 [195191.943956] ? __init_waitqueue_head+0x36/0x50 [195191.943960] vfs_kern_mount+0x62/0x160 [195191.943963] do_mount+0x1f9/0xd40 [195191.943967] ? memdup_user+0x4b/0x70 [195191.943971] ksys_mount+0x7e/0xd0 [195191.943974] __x64_sys_mount+0x21/0x30 [195191.943977] do_syscall_64+0x60/0x1b0 [195191.943980] entry_SYSCALL_64_after_hwframe+0x49/0xbe [195191.943983] RIP: 0033:0x7f570e4e524a [195191.943986] RSP: 002b:00007ffd83589478 EFLAGS: 00000206 ORIG_RAX: 00000000000000a5 [195191.943989] RAX: ffffffffffffffda RBX: 0000563f335b2060 RCX: 00007f570e4e524a [195191.943990] RDX: 0000563f335b2240 RSI: 0000563f335b2280 RDI: 0000563f335b2260 [195191.943992] RBP: 0000000000000000 R08: 0000000000000000 R09: 0000000000000020 [195191.943993] R10: 00000000c0ed0000 R11: 0000000000000206 R12: 0000563f335b2260 [195191.943994] R13: 0000563f335b2240 R14: 0000000000000000 R15: 00000000ffffffff [195191.944002] irq event stamp: 8688 [195191.944010] hardirqs last enabled at (8687): [<ffffffff9cb004c3>] console_unlock+0x503/0x640 [195191.944012] hardirqs last disabled at (8688): [<ffffffff9ca037dd>] trace_hardirqs_off_thunk+0x1a/0x1c [195191.944018] softirqs last enabled at (8638): [<ffffffff9cc0a5d1>] __set_page_dirty_nobuffers+0x101/0x150 [195191.944020] softirqs last disabled at (8634): [<ffffffff9cc26bbe>] wb_wakeup_delayed+0x2e/0x60 [195191.944022] ---[ end trace 5d6e873a9a0b811a ]--- This happens because the inode does not have the flag I_LINKABLE set, which is a runtime only flag, not meant to be persisted, set when the inode is created through open(2) if the flag O_EXCL is not passed to it. Except for the warning, there are no other consequences (like corruptions or metadata inconsistencies). Since it's pointless to replay a tmpfile as it would be deleted in a later phase of the log replay procedure (it has a link count of 0), fix this by not logging tmpfiles and if a tmpfile is found in a log (created by a kernel without this change), skip the replay of the inode. A test case for fstests follows soon. Fixes: 471d557afed1 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay") CC: stable@vger.kernel.org # 4.18+ Reported-by: Martin Steigerwald <martin@lichtvoll.de> Link: https://lore.kernel.org/linux-btrfs/3666619.NTnn27ZJZE@merkaba/ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2018-10-08 13:12:55 +03:00
if (wc->ignore_cur_inode)
continue;
if (key.type == BTRFS_DIR_INDEX_KEY &&
wc->stage == LOG_WALK_REPLAY_DIR_INDEX) {
ret = replay_one_dir_item(wc->trans, root, path,
eb, i, &key);
if (ret)
break;
}
if (wc->stage < LOG_WALK_REPLAY_ALL)
continue;
/* these keys are simply copied */
if (key.type == BTRFS_XATTR_ITEM_KEY) {
ret = overwrite_item(wc->trans, root, path,
eb, i, &key);
if (ret)
break;
} else if (key.type == BTRFS_INODE_REF_KEY ||
key.type == BTRFS_INODE_EXTREF_KEY) {
ret = add_inode_ref(wc->trans, root, log, path,
eb, i, &key);
if (ret && ret != -ENOENT)
break;
ret = 0;
} else if (key.type == BTRFS_EXTENT_DATA_KEY) {
ret = replay_one_extent(wc->trans, root, path,
eb, i, &key);
if (ret)
break;
}
btrfs: only copy dir index keys when logging a directory Currently, when logging a directory, we copy both dir items and dir index items from the fs/subvolume tree to the log tree. Both items have exactly the same data (same struct btrfs_dir_item), the difference lies in the key values, where a dir index key contains the index number of a directory entry while the dir item key does not, as it's used for doing fast lookups of an entry by name, while the former is used for sorting entries when listing a directory. We can exploit that and log only the dir index items, since they contain all the information needed to correctly add, replace and delete directory entries when replaying a log tree. Logging only the dir index items is also backward and forward compatible: an unpatched kernel (without this change) can correctly replay a log tree generated by a patched kernel (with this patch), and a patched kernel can correctly replay a log tree generated by an unpatched kernel. The backward compatibility is ensured because: 1) For inserting a new dentry: a dentry is only inserted when we find a new dir index key - we can only insert if we know the dir index offset, which is encoded in the dir index key's offset; 2) For deleting dentries: during log replay, before adding or replacing dentries, we first replay dentry deletions. Whenever we find a dir item key or a dir index key in the subvolume/fs tree that is not logged in a range for which the log tree is authoritative, we do the unlink of the dentry, which removes both the existing dir item key and the dir index key. Therefore logging just dir index keys is enough to ensure dentry deletions are correctly replayed; 3) For dentry replacements: they work when we log only dir index keys and this is mostly due to a combination of 1) and 2). If we replace a dentry with name "foobar" to point from inode A to inode B, then we know the dir index key for the new dentry is different from the old one, as it has an index number (key offset) larger than the old one. This results in replaying a deletion, through replay_dir_deletes(), that causes the old dentry to be removed, both the dir item key and the dir index key, as mentioned at 2). Then when processing the new dir index key, we add the new dentry, adding both a new dir item key and a new index key pointing to inode B, as stated in 1). The forward compatibility, the ability for a patched kernel to replay a log created by an older, unpatched kernel, comes from the changes required for making sure we are able to replay a log that only contains dir index keys - we simply ignore every dir item key we find. So modify directory logging to log only dir index items, and modify the log replay process to ignore dir item keys, from log trees created by an unpatched kernel, and process only with dir index keys. This reduces the amount of logged metadata by about half, and therefore the time spent logging or fsyncing large directories (less CPU time and less IO). The following test script was used to measure this change: #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=1000000 NUM_FILE_DELETES=10000 mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The tests were run on a physical machine, with a non-debug kernel (Debian's default kernel config), for different values of $NUM_NEW_FILES and $NUM_FILE_DELETES, and the results were the following: ** Before patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 8412 ms after adding 1000000 files dir fsync took 500 ms after deleting 10000 files ** After patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 4252 ms after adding 1000000 files (-49.5%) dir fsync took 269 ms after deleting 10000 files (-46.2%) ** Before patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 745 ms after adding 100000 files dir fsync took 59 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 404 ms after adding 100000 files (-45.8%) dir fsync took 31 ms after deleting 1000 files (-47.5%) ** Before patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 67 ms after adding 10000 files dir fsync took 9 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 36 ms after adding 10000 files (-46.3%) dir fsync took 5 ms after deleting 1000 files (-44.4%) ** Before patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 9 ms after adding 1000 files dir fsync took 4 ms after deleting 100 files ** After patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 7 ms after adding 1000 files (-22.2%) dir fsync took 3 ms after deleting 100 files (-25.0%) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-10-25 19:31:53 +03:00
/*
* We don't log BTRFS_DIR_ITEM_KEY keys anymore, only the
* BTRFS_DIR_INDEX_KEY items which we use to derive the
* BTRFS_DIR_ITEM_KEY items. If we are replaying a log from an
* older kernel with such keys, ignore them.
*/
}
btrfs_free_path(path);
return ret;
}
/*
* Correctly adjust the reserved bytes occupied by a log tree extent buffer
*/
static void unaccount_log_buffer(struct btrfs_fs_info *fs_info, u64 start)
{
struct btrfs_block_group *cache;
cache = btrfs_lookup_block_group(fs_info, start);
if (!cache) {
btrfs_err(fs_info, "unable to find block group for %llu", start);
return;
}
spin_lock(&cache->space_info->lock);
spin_lock(&cache->lock);
cache->reserved -= fs_info->nodesize;
cache->space_info->bytes_reserved -= fs_info->nodesize;
spin_unlock(&cache->lock);
spin_unlock(&cache->space_info->lock);
btrfs_put_block_group(cache);
}
static noinline int walk_down_log_tree(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_path *path, int *level,
struct walk_control *wc)
{
struct btrfs_fs_info *fs_info = root->fs_info;
u64 bytenr;
u64 ptr_gen;
struct extent_buffer *next;
struct extent_buffer *cur;
u32 blocksize;
int ret = 0;
while (*level > 0) {
struct btrfs_key first_key;
cur = path->nodes[*level];
WARN_ON(btrfs_header_level(cur) != *level);
if (path->slots[*level] >=
btrfs_header_nritems(cur))
break;
bytenr = btrfs_node_blockptr(cur, path->slots[*level]);
ptr_gen = btrfs_node_ptr_generation(cur, path->slots[*level]);
btrfs_node_key_to_cpu(cur, &first_key, path->slots[*level]);
blocksize = fs_info->nodesize;
next = btrfs_find_create_tree_block(fs_info, bytenr,
btrfs_header_owner(cur),
*level - 1);
if (IS_ERR(next))
return PTR_ERR(next);
if (*level == 1) {
ret = wc->process_func(root, next, wc, ptr_gen,
*level - 1);
if (ret) {
free_extent_buffer(next);
return ret;
}
path->slots[*level]++;
if (wc->free) {
ret = btrfs_read_buffer(next, ptr_gen,
*level - 1, &first_key);
if (ret) {
free_extent_buffer(next);
return ret;
}
if (trans) {
btrfs_tree_lock(next);
btrfs_clean_tree_block(next);
btrfs_wait_tree_block_writeback(next);
btrfs_tree_unlock(next);
ret = btrfs_pin_reserved_extent(trans,
bytenr, blocksize);
if (ret) {
free_extent_buffer(next);
return ret;
}
btrfs_redirty_list_add(
trans->transaction, next);
} else {
if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &next->bflags))
clear_extent_buffer_dirty(next);
unaccount_log_buffer(fs_info, bytenr);
}
}
free_extent_buffer(next);
continue;
}
ret = btrfs_read_buffer(next, ptr_gen, *level - 1, &first_key);
if (ret) {
free_extent_buffer(next);
return ret;
}
if (path->nodes[*level-1])
free_extent_buffer(path->nodes[*level-1]);
path->nodes[*level-1] = next;
*level = btrfs_header_level(next);
path->slots[*level] = 0;
cond_resched();
}
path->slots[*level] = btrfs_header_nritems(path->nodes[*level]);
cond_resched();
return 0;
}
static noinline int walk_up_log_tree(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_path *path, int *level,
struct walk_control *wc)
{
struct btrfs_fs_info *fs_info = root->fs_info;
int i;
int slot;
int ret;
for (i = *level; i < BTRFS_MAX_LEVEL - 1 && path->nodes[i]; i++) {
slot = path->slots[i];
if (slot + 1 < btrfs_header_nritems(path->nodes[i])) {
path->slots[i]++;
*level = i;
WARN_ON(*level == 0);
return 0;
} else {
ret = wc->process_func(root, path->nodes[*level], wc,
btrfs_header_generation(path->nodes[*level]),
*level);
if (ret)
return ret;
if (wc->free) {
struct extent_buffer *next;
next = path->nodes[*level];
if (trans) {
btrfs_tree_lock(next);
btrfs_clean_tree_block(next);
btrfs_wait_tree_block_writeback(next);
btrfs_tree_unlock(next);
ret = btrfs_pin_reserved_extent(trans,
path->nodes[*level]->start,
path->nodes[*level]->len);
if (ret)
return ret;
btrfs: fix re-dirty process of tree-log nodes There is a report of a transaction abort of -EAGAIN with the following script. #!/bin/sh for d in sda sdb; do mkfs.btrfs -d single -m single -f /dev/\${d} done mount /dev/sda /mnt/test mount /dev/sdb /mnt/scratch for dir in test scratch; do echo 3 >/proc/sys/vm/drop_caches fio --directory=/mnt/\${dir} --name=fio.\${dir} --rw=read --size=50G --bs=64m \ --numjobs=$(nproc) --time_based --ramp_time=5 --runtime=480 \ --group_reporting |& tee /dev/shm/fio.\${dir} echo 3 >/proc/sys/vm/drop_caches done for d in sda sdb; do umount /dev/\${d} done The stack trace is shown in below. [3310.967991] BTRFS: error (device sda) in btrfs_commit_transaction:2341: errno=-11 unknown (Error while writing out transaction) [3310.968060] BTRFS info (device sda): forced readonly [3310.968064] BTRFS warning (device sda): Skipping commit of aborted transaction. [3310.968065] ------------[ cut here ]------------ [3310.968066] BTRFS: Transaction aborted (error -11) [3310.968074] WARNING: CPU: 14 PID: 1684 at fs/btrfs/transaction.c:1946 btrfs_commit_transaction.cold+0x209/0x2c8 [3310.968131] CPU: 14 PID: 1684 Comm: fio Not tainted 5.14.10-300.fc35.x86_64 #1 [3310.968135] Hardware name: DIAWAY Tartu/Tartu, BIOS V2.01.B10 04/08/2021 [3310.968137] RIP: 0010:btrfs_commit_transaction.cold+0x209/0x2c8 [3310.968144] RSP: 0018:ffffb284ce393e10 EFLAGS: 00010282 [3310.968147] RAX: 0000000000000026 RBX: ffff973f147b0f60 RCX: 0000000000000027 [3310.968149] RDX: ffff974ecf098a08 RSI: 0000000000000001 RDI: ffff974ecf098a00 [3310.968150] RBP: ffff973f147b0f08 R08: 0000000000000000 R09: ffffb284ce393c48 [3310.968151] R10: ffffb284ce393c40 R11: ffffffff84f47468 R12: ffff973f101bfc00 [3310.968153] R13: ffff971f20cf2000 R14: 00000000fffffff5 R15: ffff973f147b0e58 [3310.968154] FS: 00007efe65468740(0000) GS:ffff974ecf080000(0000) knlGS:0000000000000000 [3310.968157] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [3310.968158] CR2: 000055691bcbe260 CR3: 000000105cfa4001 CR4: 0000000000770ee0 [3310.968160] PKRU: 55555554 [3310.968161] Call Trace: [3310.968167] ? dput+0xd4/0x300 [3310.968174] btrfs_sync_file+0x3f1/0x490 [3310.968180] __x64_sys_fsync+0x33/0x60 [3310.968185] do_syscall_64+0x3b/0x90 [3310.968190] entry_SYSCALL_64_after_hwframe+0x44/0xae [3310.968194] RIP: 0033:0x7efe6557329b [3310.968200] RSP: 002b:00007ffe0236ebc0 EFLAGS: 00000293 ORIG_RAX: 000000000000004a [3310.968203] RAX: ffffffffffffffda RBX: 0000000000000000 RCX: 00007efe6557329b [3310.968204] RDX: 0000000000000000 RSI: 00007efe58d77010 RDI: 0000000000000006 [3310.968205] RBP: 0000000004000000 R08: 0000000000000000 R09: 00007efe58d77010 [3310.968207] R10: 0000000016cacc0c R11: 0000000000000293 R12: 00007efe5ce95980 [3310.968208] R13: 0000000000000000 R14: 00007efe6447c790 R15: 0000000c80000000 [3310.968212] ---[ end trace 1a346f4d3c0d96ba ]--- [3310.968214] BTRFS: error (device sda) in cleanup_transaction:1946: errno=-11 unknown The abort occurs because of a write hole while writing out freeing tree nodes of a tree-log tree. For zoned btrfs, we re-dirty a freed tree node to ensure btrfs can write the region and does not leave a hole on write on a zoned device. The current code fails to re-dirty a node when the tree-log tree's depth is greater or equal to 2. That leads to a transaction abort with -EAGAIN. Fix the issue by properly re-dirtying a node on walking up the tree. Fixes: d3575156f662 ("btrfs: zoned: redirty released extent buffers") CC: stable@vger.kernel.org # 5.12+ Link: https://github.com/kdave/btrfs-progs/issues/415 Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com> Signed-off-by: Naohiro Aota <naohiro.aota@wdc.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-11-30 06:40:21 +03:00
btrfs_redirty_list_add(trans->transaction,
next);
} else {
if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &next->bflags))
clear_extent_buffer_dirty(next);
unaccount_log_buffer(fs_info,
path->nodes[*level]->start);
}
}
free_extent_buffer(path->nodes[*level]);
path->nodes[*level] = NULL;
*level = i + 1;
}
}
return 1;
}
/*
* drop the reference count on the tree rooted at 'snap'. This traverses
* the tree freeing any blocks that have a ref count of zero after being
* decremented.
*/
static int walk_log_tree(struct btrfs_trans_handle *trans,
struct btrfs_root *log, struct walk_control *wc)
{
struct btrfs_fs_info *fs_info = log->fs_info;
int ret = 0;
int wret;
int level;
struct btrfs_path *path;
int orig_level;
path = btrfs_alloc_path();
if (!path)
return -ENOMEM;
level = btrfs_header_level(log->node);
orig_level = level;
path->nodes[level] = log->node;
atomic_inc(&log->node->refs);
path->slots[level] = 0;
while (1) {
wret = walk_down_log_tree(trans, log, path, &level, wc);
if (wret > 0)
break;
if (wret < 0) {
ret = wret;
goto out;
}
wret = walk_up_log_tree(trans, log, path, &level, wc);
if (wret > 0)
break;
if (wret < 0) {
ret = wret;
goto out;
}
}
/* was the root node processed? if not, catch it here */
if (path->nodes[orig_level]) {
ret = wc->process_func(log, path->nodes[orig_level], wc,
btrfs_header_generation(path->nodes[orig_level]),
orig_level);
if (ret)
goto out;
if (wc->free) {
struct extent_buffer *next;
next = path->nodes[orig_level];
if (trans) {
btrfs_tree_lock(next);
btrfs_clean_tree_block(next);
btrfs_wait_tree_block_writeback(next);
btrfs_tree_unlock(next);
ret = btrfs_pin_reserved_extent(trans,
next->start, next->len);
if (ret)
goto out;
btrfs: fix re-dirty process of tree-log nodes There is a report of a transaction abort of -EAGAIN with the following script. #!/bin/sh for d in sda sdb; do mkfs.btrfs -d single -m single -f /dev/\${d} done mount /dev/sda /mnt/test mount /dev/sdb /mnt/scratch for dir in test scratch; do echo 3 >/proc/sys/vm/drop_caches fio --directory=/mnt/\${dir} --name=fio.\${dir} --rw=read --size=50G --bs=64m \ --numjobs=$(nproc) --time_based --ramp_time=5 --runtime=480 \ --group_reporting |& tee /dev/shm/fio.\${dir} echo 3 >/proc/sys/vm/drop_caches done for d in sda sdb; do umount /dev/\${d} done The stack trace is shown in below. [3310.967991] BTRFS: error (device sda) in btrfs_commit_transaction:2341: errno=-11 unknown (Error while writing out transaction) [3310.968060] BTRFS info (device sda): forced readonly [3310.968064] BTRFS warning (device sda): Skipping commit of aborted transaction. [3310.968065] ------------[ cut here ]------------ [3310.968066] BTRFS: Transaction aborted (error -11) [3310.968074] WARNING: CPU: 14 PID: 1684 at fs/btrfs/transaction.c:1946 btrfs_commit_transaction.cold+0x209/0x2c8 [3310.968131] CPU: 14 PID: 1684 Comm: fio Not tainted 5.14.10-300.fc35.x86_64 #1 [3310.968135] Hardware name: DIAWAY Tartu/Tartu, BIOS V2.01.B10 04/08/2021 [3310.968137] RIP: 0010:btrfs_commit_transaction.cold+0x209/0x2c8 [3310.968144] RSP: 0018:ffffb284ce393e10 EFLAGS: 00010282 [3310.968147] RAX: 0000000000000026 RBX: ffff973f147b0f60 RCX: 0000000000000027 [3310.968149] RDX: ffff974ecf098a08 RSI: 0000000000000001 RDI: ffff974ecf098a00 [3310.968150] RBP: ffff973f147b0f08 R08: 0000000000000000 R09: ffffb284ce393c48 [3310.968151] R10: ffffb284ce393c40 R11: ffffffff84f47468 R12: ffff973f101bfc00 [3310.968153] R13: ffff971f20cf2000 R14: 00000000fffffff5 R15: ffff973f147b0e58 [3310.968154] FS: 00007efe65468740(0000) GS:ffff974ecf080000(0000) knlGS:0000000000000000 [3310.968157] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [3310.968158] CR2: 000055691bcbe260 CR3: 000000105cfa4001 CR4: 0000000000770ee0 [3310.968160] PKRU: 55555554 [3310.968161] Call Trace: [3310.968167] ? dput+0xd4/0x300 [3310.968174] btrfs_sync_file+0x3f1/0x490 [3310.968180] __x64_sys_fsync+0x33/0x60 [3310.968185] do_syscall_64+0x3b/0x90 [3310.968190] entry_SYSCALL_64_after_hwframe+0x44/0xae [3310.968194] RIP: 0033:0x7efe6557329b [3310.968200] RSP: 002b:00007ffe0236ebc0 EFLAGS: 00000293 ORIG_RAX: 000000000000004a [3310.968203] RAX: ffffffffffffffda RBX: 0000000000000000 RCX: 00007efe6557329b [3310.968204] RDX: 0000000000000000 RSI: 00007efe58d77010 RDI: 0000000000000006 [3310.968205] RBP: 0000000004000000 R08: 0000000000000000 R09: 00007efe58d77010 [3310.968207] R10: 0000000016cacc0c R11: 0000000000000293 R12: 00007efe5ce95980 [3310.968208] R13: 0000000000000000 R14: 00007efe6447c790 R15: 0000000c80000000 [3310.968212] ---[ end trace 1a346f4d3c0d96ba ]--- [3310.968214] BTRFS: error (device sda) in cleanup_transaction:1946: errno=-11 unknown The abort occurs because of a write hole while writing out freeing tree nodes of a tree-log tree. For zoned btrfs, we re-dirty a freed tree node to ensure btrfs can write the region and does not leave a hole on write on a zoned device. The current code fails to re-dirty a node when the tree-log tree's depth is greater or equal to 2. That leads to a transaction abort with -EAGAIN. Fix the issue by properly re-dirtying a node on walking up the tree. Fixes: d3575156f662 ("btrfs: zoned: redirty released extent buffers") CC: stable@vger.kernel.org # 5.12+ Link: https://github.com/kdave/btrfs-progs/issues/415 Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com> Signed-off-by: Naohiro Aota <naohiro.aota@wdc.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-11-30 06:40:21 +03:00
btrfs_redirty_list_add(trans->transaction, next);
} else {
if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &next->bflags))
clear_extent_buffer_dirty(next);
unaccount_log_buffer(fs_info, next->start);
}
}
}
out:
btrfs_free_path(path);
return ret;
}
/*
* helper function to update the item for a given subvolumes log root
* in the tree of log roots
*/
static int update_log_root(struct btrfs_trans_handle *trans,
btrfs: fix incorrect updating of log root tree We've historically had reports of being unable to mount file systems because the tree log root couldn't be read. Usually this is the "parent transid failure", but could be any of the related errors, including "fsid mismatch" or "bad tree block", depending on which block got allocated. The modification of the individual log root items are serialized on the per-log root root_mutex. This means that any modification to the per-subvol log root_item is completely protected. However we update the root item in the log root tree outside of the log root tree log_mutex. We do this in order to allow multiple subvolumes to be updated in each log transaction. This is problematic however because when we are writing the log root tree out we update the super block with the _current_ log root node information. Since these two operations happen independently of each other, you can end up updating the log root tree in between writing out the dirty blocks and setting the super block to point at the current root. This means we'll point at the new root node that hasn't been written out, instead of the one we should be pointing at. Thus whatever garbage or old block we end up pointing at complains when we mount the file system later and try to replay the log. Fix this by copying the log's root item into a local root item copy. Then once we're safely under the log_root_tree->log_mutex we update the root item in the log_root_tree. This way we do not modify the log_root_tree while we're committing it, fixing the problem. CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Chris Mason <clm@fb.com> Reviewed-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-09-30 23:27:25 +03:00
struct btrfs_root *log,
struct btrfs_root_item *root_item)
{
struct btrfs_fs_info *fs_info = log->fs_info;
int ret;
if (log->log_transid == 1) {
/* insert root item on the first sync */
ret = btrfs_insert_root(trans, fs_info->log_root_tree,
btrfs: fix incorrect updating of log root tree We've historically had reports of being unable to mount file systems because the tree log root couldn't be read. Usually this is the "parent transid failure", but could be any of the related errors, including "fsid mismatch" or "bad tree block", depending on which block got allocated. The modification of the individual log root items are serialized on the per-log root root_mutex. This means that any modification to the per-subvol log root_item is completely protected. However we update the root item in the log root tree outside of the log root tree log_mutex. We do this in order to allow multiple subvolumes to be updated in each log transaction. This is problematic however because when we are writing the log root tree out we update the super block with the _current_ log root node information. Since these two operations happen independently of each other, you can end up updating the log root tree in between writing out the dirty blocks and setting the super block to point at the current root. This means we'll point at the new root node that hasn't been written out, instead of the one we should be pointing at. Thus whatever garbage or old block we end up pointing at complains when we mount the file system later and try to replay the log. Fix this by copying the log's root item into a local root item copy. Then once we're safely under the log_root_tree->log_mutex we update the root item in the log_root_tree. This way we do not modify the log_root_tree while we're committing it, fixing the problem. CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Chris Mason <clm@fb.com> Reviewed-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-09-30 23:27:25 +03:00
&log->root_key, root_item);
} else {
ret = btrfs_update_root(trans, fs_info->log_root_tree,
btrfs: fix incorrect updating of log root tree We've historically had reports of being unable to mount file systems because the tree log root couldn't be read. Usually this is the "parent transid failure", but could be any of the related errors, including "fsid mismatch" or "bad tree block", depending on which block got allocated. The modification of the individual log root items are serialized on the per-log root root_mutex. This means that any modification to the per-subvol log root_item is completely protected. However we update the root item in the log root tree outside of the log root tree log_mutex. We do this in order to allow multiple subvolumes to be updated in each log transaction. This is problematic however because when we are writing the log root tree out we update the super block with the _current_ log root node information. Since these two operations happen independently of each other, you can end up updating the log root tree in between writing out the dirty blocks and setting the super block to point at the current root. This means we'll point at the new root node that hasn't been written out, instead of the one we should be pointing at. Thus whatever garbage or old block we end up pointing at complains when we mount the file system later and try to replay the log. Fix this by copying the log's root item into a local root item copy. Then once we're safely under the log_root_tree->log_mutex we update the root item in the log_root_tree. This way we do not modify the log_root_tree while we're committing it, fixing the problem. CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Chris Mason <clm@fb.com> Reviewed-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-09-30 23:27:25 +03:00
&log->root_key, root_item);
}
return ret;
}
static void wait_log_commit(struct btrfs_root *root, int transid)
{
DEFINE_WAIT(wait);
int index = transid % 2;
/*
* we only allow two pending log transactions at a time,
* so we know that if ours is more than 2 older than the
* current transaction, we're done
*/
for (;;) {
prepare_to_wait(&root->log_commit_wait[index],
&wait, TASK_UNINTERRUPTIBLE);
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 17:24:20 +03:00
if (!(root->log_transid_committed < transid &&
atomic_read(&root->log_commit[index])))
break;
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 17:24:20 +03:00
mutex_unlock(&root->log_mutex);
schedule();
mutex_lock(&root->log_mutex);
}
finish_wait(&root->log_commit_wait[index], &wait);
}
static void wait_for_writer(struct btrfs_root *root)
{
DEFINE_WAIT(wait);
for (;;) {
prepare_to_wait(&root->log_writer_wait, &wait,
TASK_UNINTERRUPTIBLE);
if (!atomic_read(&root->log_writers))
break;
mutex_unlock(&root->log_mutex);
schedule();
Btrfs: fix scheduler warning when syncing log We try to lock a mutex while the current task state is not TASK_RUNNING, which results in the following warning when CONFIG_DEBUG_LOCK_ALLOC=y: [30736.772501] ------------[ cut here ]------------ [30736.774545] WARNING: CPU: 9 PID: 19972 at kernel/sched/core.c:7300 __might_sleep+0x8b/0xa8() [30736.783453] do not call blocking ops when !TASK_RUNNING; state=2 set at [<ffffffff8107499b>] prepare_to_wait+0x43/0x89 [30736.786261] Modules linked in: dm_flakey dm_mod crc32c_generic btrfs xor raid6_pq nfsd auth_rpcgss oid_registry nfs_acl nfs lockd grace fscache sunrpc loop parport_pc psmouse parport pcspkr microcode serio_raw evdev processor thermal_sys i2c_piix4 i2c_core button ext4 crc16 jbd2 mbcache sg sr_mod cdrom sd_mod ata_generic virtio_scsi floppy ata_piix libata virtio_pci virtio_ring e1000 virtio scsi_mod [30736.794323] CPU: 9 PID: 19972 Comm: fsstress Not tainted 3.19.0-rc7-btrfs-next-5+ #1 [30736.795821] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.7.5-0-ge51488c-20140602_164612-nilsson.home.kraxel.org 04/01/2014 [30736.798788] 0000000000000009 ffff88042743fbd8 ffffffff814248ed ffff88043d32f2d8 [30736.800504] ffff88042743fc28 ffff88042743fc18 ffffffff81045338 0000000000000001 [30736.802131] ffffffff81064514 ffffffff817c52d1 000000000000026d 0000000000000000 [30736.803676] Call Trace: [30736.804256] [<ffffffff814248ed>] dump_stack+0x4c/0x65 [30736.805245] [<ffffffff81045338>] warn_slowpath_common+0xa1/0xbb [30736.806360] [<ffffffff81064514>] ? __might_sleep+0x8b/0xa8 [30736.807391] [<ffffffff81045398>] warn_slowpath_fmt+0x46/0x48 [30736.808511] [<ffffffff8107499b>] ? prepare_to_wait+0x43/0x89 [30736.809620] [<ffffffff8107499b>] ? prepare_to_wait+0x43/0x89 [30736.810691] [<ffffffff81064514>] __might_sleep+0x8b/0xa8 [30736.811703] [<ffffffff81426eaf>] mutex_lock_nested+0x2f/0x3a0 [30736.812889] [<ffffffff8107bfa1>] ? trace_hardirqs_on_caller+0x18f/0x1ab [30736.814138] [<ffffffff8107bfca>] ? trace_hardirqs_on+0xd/0xf [30736.819878] [<ffffffffa038cfff>] wait_for_writer.isra.12+0x91/0xaa [btrfs] [30736.821260] [<ffffffff810748bd>] ? signal_pending_state+0x31/0x31 [30736.822410] [<ffffffffa0391f0a>] btrfs_sync_log+0x160/0x947 [btrfs] [30736.823574] [<ffffffff8107bfa1>] ? trace_hardirqs_on_caller+0x18f/0x1ab [30736.824847] [<ffffffff8107bfca>] ? trace_hardirqs_on+0xd/0xf [30736.825972] [<ffffffffa036e555>] btrfs_sync_file+0x2b0/0x319 [btrfs] [30736.827684] [<ffffffff8117901a>] vfs_fsync_range+0x21/0x23 [30736.828932] [<ffffffff81179038>] vfs_fsync+0x1c/0x1e [30736.829917] [<ffffffff8117928b>] do_fsync+0x34/0x4e [30736.830862] [<ffffffff811794b3>] SyS_fsync+0x10/0x14 [30736.831819] [<ffffffff8142a512>] system_call_fastpath+0x12/0x17 [30736.832982] ---[ end trace c0b57df60d32ae5c ]--- Fix this my acquiring the mutex after calling finish_wait(), which sets the task's state to TASK_RUNNING. Signed-off-by: Filipe Manana <fdmanana@suse.com> Reviewed-by: Liu Bo <bo.li.liu@oracle.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-02-11 14:12:39 +03:00
mutex_lock(&root->log_mutex);
}
finish_wait(&root->log_writer_wait, &wait);
}
static inline void btrfs_remove_log_ctx(struct btrfs_root *root,
struct btrfs_log_ctx *ctx)
{
mutex_lock(&root->log_mutex);
list_del_init(&ctx->list);
mutex_unlock(&root->log_mutex);
}
/*
* Invoked in log mutex context, or be sure there is no other task which
* can access the list.
*/
static inline void btrfs_remove_all_log_ctxs(struct btrfs_root *root,
int index, int error)
{
struct btrfs_log_ctx *ctx;
struct btrfs_log_ctx *safe;
list_for_each_entry_safe(ctx, safe, &root->log_ctxs[index], list) {
list_del_init(&ctx->list);
ctx->log_ret = error;
}
}
/*
* btrfs_sync_log does sends a given tree log down to the disk and
* updates the super blocks to record it. When this call is done,
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 17:24:20 +03:00
* you know that any inodes previously logged are safely on disk only
* if it returns 0.
*
* Any other return value means you need to call btrfs_commit_transaction.
* Some of the edge cases for fsyncing directories that have had unlinks
* or renames done in the past mean that sometimes the only safe
* fsync is to commit the whole FS. When btrfs_sync_log returns -EAGAIN,
* that has happened.
*/
int btrfs_sync_log(struct btrfs_trans_handle *trans,
struct btrfs_root *root, struct btrfs_log_ctx *ctx)
{
int index1;
int index2;
int mark;
int ret;
struct btrfs_fs_info *fs_info = root->fs_info;
struct btrfs_root *log = root->log_root;
struct btrfs_root *log_root_tree = fs_info->log_root_tree;
btrfs: fix incorrect updating of log root tree We've historically had reports of being unable to mount file systems because the tree log root couldn't be read. Usually this is the "parent transid failure", but could be any of the related errors, including "fsid mismatch" or "bad tree block", depending on which block got allocated. The modification of the individual log root items are serialized on the per-log root root_mutex. This means that any modification to the per-subvol log root_item is completely protected. However we update the root item in the log root tree outside of the log root tree log_mutex. We do this in order to allow multiple subvolumes to be updated in each log transaction. This is problematic however because when we are writing the log root tree out we update the super block with the _current_ log root node information. Since these two operations happen independently of each other, you can end up updating the log root tree in between writing out the dirty blocks and setting the super block to point at the current root. This means we'll point at the new root node that hasn't been written out, instead of the one we should be pointing at. Thus whatever garbage or old block we end up pointing at complains when we mount the file system later and try to replay the log. Fix this by copying the log's root item into a local root item copy. Then once we're safely under the log_root_tree->log_mutex we update the root item in the log_root_tree. This way we do not modify the log_root_tree while we're committing it, fixing the problem. CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Chris Mason <clm@fb.com> Reviewed-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-09-30 23:27:25 +03:00
struct btrfs_root_item new_root_item;
int log_transid = 0;
struct btrfs_log_ctx root_log_ctx;
struct blk_plug plug;
btrfs: do not block inode logging for so long during transaction commit Early on during a transaction commit we acquire the tree_log_mutex and hold it until after we write the super blocks. But before writing the extent buffers dirtied by the transaction and the super blocks we unblock the transaction by setting its state to TRANS_STATE_UNBLOCKED and setting fs_info->running_transaction to NULL. This means that after that and before writing the super blocks, new transactions can start. However if any transaction wants to log an inode, it will block waiting for the transaction commit to write its dirty extent buffers and the super blocks because the tree_log_mutex is only released after those operations are complete, and starting a new log transaction blocks on that mutex (at start_log_trans()). Writing the dirty extent buffers and the super blocks can take a very significant amount of time to complete, but we could allow the tasks wanting to log an inode to proceed with most of their steps: 1) create the log trees 2) log metadata in the trees 3) write their dirty extent buffers They only need to wait for the previous transaction commit to complete (write its super blocks) before they attempt to write their super blocks, otherwise we could end up with a corrupt filesystem after a crash. So change start_log_trans() to use the root tree's log_mutex to serialize for the creation of the log root tree instead of using the tree_log_mutex, and make btrfs_sync_log() acquire the tree_log_mutex before writing the super blocks. This allows for inode logging to wait much less time when there is a previous transaction that is still committing, often not having to wait at all, as by the time when we try to sync the log the previous transaction already wrote its super blocks. This patch belongs to a patch set that is comprised of the following patches: btrfs: fix race causing unnecessary inode logging during link and rename btrfs: fix race that results in logging old extents during a fast fsync btrfs: fix race that causes unnecessary logging of ancestor inodes btrfs: fix race that makes inode logging fallback to transaction commit btrfs: fix race leading to unnecessary transaction commit when logging inode btrfs: do not block inode logging for so long during transaction commit The following script that uses dbench was used to measure the impact of the whole patchset: $ cat test-dbench.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/btrfs MOUNT_OPTIONS="-o ssd" echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f -m single -d single $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -D $MNT -t 300 64 umount $MNT The test was run on a machine with 12 cores, 64G of ram, using a NVMe device and a non-debug kernel configuration (Debian's default). Before patch set: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 11277211 0.250 85.340 Close 8283172 0.002 6.479 Rename 477515 1.935 86.026 Unlink 2277936 0.770 87.071 Deltree 256 15.732 81.379 Mkdir 128 0.003 0.009 Qpathinfo 10221180 0.056 44.404 Qfileinfo 1789967 0.002 4.066 Qfsinfo 1874399 0.003 9.176 Sfileinfo 918589 0.061 10.247 Find 3951758 0.341 54.040 WriteX 5616547 0.047 85.079 ReadX 17676028 0.005 9.704 LockX 36704 0.003 1.800 UnlockX 36704 0.002 0.687 Flush 790541 14.115 676.236 Throughput 1179.19 MB/sec 64 clients 64 procs max_latency=676.240 ms After patch set: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 12687926 0.171 86.526 Close 9320780 0.002 8.063 Rename 537253 1.444 78.576 Unlink 2561827 0.559 87.228 Deltree 374 11.499 73.549 Mkdir 187 0.003 0.005 Qpathinfo 11500300 0.061 36.801 Qfileinfo 2017118 0.002 7.189 Qfsinfo 2108641 0.003 4.825 Sfileinfo 1033574 0.008 8.065 Find 4446553 0.408 47.835 WriteX 6335667 0.045 84.388 ReadX 19887312 0.003 9.215 LockX 41312 0.003 1.394 UnlockX 41312 0.002 1.425 Flush 889233 13.014 623.259 Throughput 1339.32 MB/sec 64 clients 64 procs max_latency=623.265 ms +12.7% throughput, -8.2% max latency Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-11-25 15:19:28 +03:00
u64 log_root_start;
u64 log_root_level;
mutex_lock(&root->log_mutex);
log_transid = ctx->log_transid;
if (root->log_transid_committed >= log_transid) {
mutex_unlock(&root->log_mutex);
return ctx->log_ret;
}
index1 = log_transid % 2;
if (atomic_read(&root->log_commit[index1])) {
wait_log_commit(root, log_transid);
mutex_unlock(&root->log_mutex);
return ctx->log_ret;
}
ASSERT(log_transid == root->log_transid);
atomic_set(&root->log_commit[index1], 1);
/* wait for previous tree log sync to complete */
if (atomic_read(&root->log_commit[(index1 + 1) % 2]))
wait_log_commit(root, log_transid - 1);
while (1) {
int batch = atomic_read(&root->log_batch);
/* when we're on an ssd, just kick the log commit out */
if (!btrfs_test_opt(fs_info, SSD) &&
test_bit(BTRFS_ROOT_MULTI_LOG_TASKS, &root->state)) {
mutex_unlock(&root->log_mutex);
schedule_timeout_uninterruptible(1);
mutex_lock(&root->log_mutex);
}
wait_for_writer(root);
if (batch == atomic_read(&root->log_batch))
break;
}
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 17:24:20 +03:00
/* bail out if we need to do a full commit */
if (btrfs_need_log_full_commit(trans)) {
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 17:24:20 +03:00
ret = -EAGAIN;
mutex_unlock(&root->log_mutex);
goto out;
}
if (log_transid % 2 == 0)
mark = EXTENT_DIRTY;
else
mark = EXTENT_NEW;
/* we start IO on all the marked extents here, but we don't actually
* wait for them until later.
*/
blk_start_plug(&plug);
ret = btrfs_write_marked_extents(fs_info, &log->dirty_log_pages, mark);
/*
* -EAGAIN happens when someone, e.g., a concurrent transaction
* commit, writes a dirty extent in this tree-log commit. This
* concurrent write will create a hole writing out the extents,
* and we cannot proceed on a zoned filesystem, requiring
* sequential writing. While we can bail out to a full commit
* here, but we can continue hoping the concurrent writing fills
* the hole.
*/
if (ret == -EAGAIN && btrfs_is_zoned(fs_info))
ret = 0;
if (ret) {
blk_finish_plug(&plug);
btrfs_abort_transaction(trans, ret);
btrfs_set_log_full_commit(trans);
mutex_unlock(&root->log_mutex);
goto out;
}
btrfs: fix incorrect updating of log root tree We've historically had reports of being unable to mount file systems because the tree log root couldn't be read. Usually this is the "parent transid failure", but could be any of the related errors, including "fsid mismatch" or "bad tree block", depending on which block got allocated. The modification of the individual log root items are serialized on the per-log root root_mutex. This means that any modification to the per-subvol log root_item is completely protected. However we update the root item in the log root tree outside of the log root tree log_mutex. We do this in order to allow multiple subvolumes to be updated in each log transaction. This is problematic however because when we are writing the log root tree out we update the super block with the _current_ log root node information. Since these two operations happen independently of each other, you can end up updating the log root tree in between writing out the dirty blocks and setting the super block to point at the current root. This means we'll point at the new root node that hasn't been written out, instead of the one we should be pointing at. Thus whatever garbage or old block we end up pointing at complains when we mount the file system later and try to replay the log. Fix this by copying the log's root item into a local root item copy. Then once we're safely under the log_root_tree->log_mutex we update the root item in the log_root_tree. This way we do not modify the log_root_tree while we're committing it, fixing the problem. CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Chris Mason <clm@fb.com> Reviewed-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-09-30 23:27:25 +03:00
/*
* We _must_ update under the root->log_mutex in order to make sure we
* have a consistent view of the log root we are trying to commit at
* this moment.
*
* We _must_ copy this into a local copy, because we are not holding the
* log_root_tree->log_mutex yet. This is important because when we
* commit the log_root_tree we must have a consistent view of the
* log_root_tree when we update the super block to point at the
* log_root_tree bytenr. If we update the log_root_tree here we'll race
* with the commit and possibly point at the new block which we may not
* have written out.
*/
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 18:45:14 +04:00
btrfs_set_root_node(&log->root_item, log->node);
btrfs: fix incorrect updating of log root tree We've historically had reports of being unable to mount file systems because the tree log root couldn't be read. Usually this is the "parent transid failure", but could be any of the related errors, including "fsid mismatch" or "bad tree block", depending on which block got allocated. The modification of the individual log root items are serialized on the per-log root root_mutex. This means that any modification to the per-subvol log root_item is completely protected. However we update the root item in the log root tree outside of the log root tree log_mutex. We do this in order to allow multiple subvolumes to be updated in each log transaction. This is problematic however because when we are writing the log root tree out we update the super block with the _current_ log root node information. Since these two operations happen independently of each other, you can end up updating the log root tree in between writing out the dirty blocks and setting the super block to point at the current root. This means we'll point at the new root node that hasn't been written out, instead of the one we should be pointing at. Thus whatever garbage or old block we end up pointing at complains when we mount the file system later and try to replay the log. Fix this by copying the log's root item into a local root item copy. Then once we're safely under the log_root_tree->log_mutex we update the root item in the log_root_tree. This way we do not modify the log_root_tree while we're committing it, fixing the problem. CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Chris Mason <clm@fb.com> Reviewed-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-09-30 23:27:25 +03:00
memcpy(&new_root_item, &log->root_item, sizeof(new_root_item));
root->log_transid++;
log->log_transid = root->log_transid;
root->log_start_pid = 0;
/*
* IO has been started, blocks of the log tree have WRITTEN flag set
* in their headers. new modifications of the log will be written to
* new positions. so it's safe to allow log writers to go in.
*/
mutex_unlock(&root->log_mutex);
if (btrfs_is_zoned(fs_info)) {
mutex_lock(&fs_info->tree_root->log_mutex);
if (!log_root_tree->node) {
ret = btrfs_alloc_log_tree_node(trans, log_root_tree);
if (ret) {
mutex_unlock(&fs_info->tree_root->log_mutex);
goto out;
}
}
mutex_unlock(&fs_info->tree_root->log_mutex);
}
btrfs_init_log_ctx(&root_log_ctx, NULL);
mutex_lock(&log_root_tree->log_mutex);
index2 = log_root_tree->log_transid % 2;
list_add_tail(&root_log_ctx.list, &log_root_tree->log_ctxs[index2]);
root_log_ctx.log_transid = log_root_tree->log_transid;
btrfs: fix incorrect updating of log root tree We've historically had reports of being unable to mount file systems because the tree log root couldn't be read. Usually this is the "parent transid failure", but could be any of the related errors, including "fsid mismatch" or "bad tree block", depending on which block got allocated. The modification of the individual log root items are serialized on the per-log root root_mutex. This means that any modification to the per-subvol log root_item is completely protected. However we update the root item in the log root tree outside of the log root tree log_mutex. We do this in order to allow multiple subvolumes to be updated in each log transaction. This is problematic however because when we are writing the log root tree out we update the super block with the _current_ log root node information. Since these two operations happen independently of each other, you can end up updating the log root tree in between writing out the dirty blocks and setting the super block to point at the current root. This means we'll point at the new root node that hasn't been written out, instead of the one we should be pointing at. Thus whatever garbage or old block we end up pointing at complains when we mount the file system later and try to replay the log. Fix this by copying the log's root item into a local root item copy. Then once we're safely under the log_root_tree->log_mutex we update the root item in the log_root_tree. This way we do not modify the log_root_tree while we're committing it, fixing the problem. CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Chris Mason <clm@fb.com> Reviewed-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-09-30 23:27:25 +03:00
/*
* Now we are safe to update the log_root_tree because we're under the
* log_mutex, and we're a current writer so we're holding the commit
* open until we drop the log_mutex.
*/
ret = update_log_root(trans, log, &new_root_item);
if (ret) {
if (!list_empty(&root_log_ctx.list))
list_del_init(&root_log_ctx.list);
blk_finish_plug(&plug);
btrfs_set_log_full_commit(trans);
if (ret != -ENOSPC) {
btrfs_abort_transaction(trans, ret);
mutex_unlock(&log_root_tree->log_mutex);
goto out;
}
btrfs_wait_tree_log_extents(log, mark);
mutex_unlock(&log_root_tree->log_mutex);
ret = -EAGAIN;
goto out;
}
if (log_root_tree->log_transid_committed >= root_log_ctx.log_transid) {
blk_finish_plug(&plug);
list_del_init(&root_log_ctx.list);
mutex_unlock(&log_root_tree->log_mutex);
ret = root_log_ctx.log_ret;
goto out;
}
index2 = root_log_ctx.log_transid % 2;
if (atomic_read(&log_root_tree->log_commit[index2])) {
blk_finish_plug(&plug);
ret = btrfs_wait_tree_log_extents(log, mark);
wait_log_commit(log_root_tree,
root_log_ctx.log_transid);
mutex_unlock(&log_root_tree->log_mutex);
if (!ret)
ret = root_log_ctx.log_ret;
goto out;
}
ASSERT(root_log_ctx.log_transid == log_root_tree->log_transid);
atomic_set(&log_root_tree->log_commit[index2], 1);
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 17:24:20 +03:00
if (atomic_read(&log_root_tree->log_commit[(index2 + 1) % 2])) {
wait_log_commit(log_root_tree,
root_log_ctx.log_transid - 1);
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 17:24:20 +03:00
}
/*
* now that we've moved on to the tree of log tree roots,
* check the full commit flag again
*/
if (btrfs_need_log_full_commit(trans)) {
blk_finish_plug(&plug);
btrfs_wait_tree_log_extents(log, mark);
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 17:24:20 +03:00
mutex_unlock(&log_root_tree->log_mutex);
ret = -EAGAIN;
goto out_wake_log_root;
}
ret = btrfs_write_marked_extents(fs_info,
&log_root_tree->dirty_log_pages,
EXTENT_DIRTY | EXTENT_NEW);
blk_finish_plug(&plug);
/*
* As described above, -EAGAIN indicates a hole in the extents. We
* cannot wait for these write outs since the waiting cause a
* deadlock. Bail out to the full commit instead.
*/
if (ret == -EAGAIN && btrfs_is_zoned(fs_info)) {
btrfs_set_log_full_commit(trans);
btrfs_wait_tree_log_extents(log, mark);
mutex_unlock(&log_root_tree->log_mutex);
goto out_wake_log_root;
} else if (ret) {
btrfs_set_log_full_commit(trans);
btrfs_abort_transaction(trans, ret);
mutex_unlock(&log_root_tree->log_mutex);
goto out_wake_log_root;
}
ret = btrfs_wait_tree_log_extents(log, mark);
if (!ret)
ret = btrfs_wait_tree_log_extents(log_root_tree,
EXTENT_NEW | EXTENT_DIRTY);
if (ret) {
btrfs_set_log_full_commit(trans);
mutex_unlock(&log_root_tree->log_mutex);
goto out_wake_log_root;
}
btrfs: do not block inode logging for so long during transaction commit Early on during a transaction commit we acquire the tree_log_mutex and hold it until after we write the super blocks. But before writing the extent buffers dirtied by the transaction and the super blocks we unblock the transaction by setting its state to TRANS_STATE_UNBLOCKED and setting fs_info->running_transaction to NULL. This means that after that and before writing the super blocks, new transactions can start. However if any transaction wants to log an inode, it will block waiting for the transaction commit to write its dirty extent buffers and the super blocks because the tree_log_mutex is only released after those operations are complete, and starting a new log transaction blocks on that mutex (at start_log_trans()). Writing the dirty extent buffers and the super blocks can take a very significant amount of time to complete, but we could allow the tasks wanting to log an inode to proceed with most of their steps: 1) create the log trees 2) log metadata in the trees 3) write their dirty extent buffers They only need to wait for the previous transaction commit to complete (write its super blocks) before they attempt to write their super blocks, otherwise we could end up with a corrupt filesystem after a crash. So change start_log_trans() to use the root tree's log_mutex to serialize for the creation of the log root tree instead of using the tree_log_mutex, and make btrfs_sync_log() acquire the tree_log_mutex before writing the super blocks. This allows for inode logging to wait much less time when there is a previous transaction that is still committing, often not having to wait at all, as by the time when we try to sync the log the previous transaction already wrote its super blocks. This patch belongs to a patch set that is comprised of the following patches: btrfs: fix race causing unnecessary inode logging during link and rename btrfs: fix race that results in logging old extents during a fast fsync btrfs: fix race that causes unnecessary logging of ancestor inodes btrfs: fix race that makes inode logging fallback to transaction commit btrfs: fix race leading to unnecessary transaction commit when logging inode btrfs: do not block inode logging for so long during transaction commit The following script that uses dbench was used to measure the impact of the whole patchset: $ cat test-dbench.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/btrfs MOUNT_OPTIONS="-o ssd" echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f -m single -d single $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -D $MNT -t 300 64 umount $MNT The test was run on a machine with 12 cores, 64G of ram, using a NVMe device and a non-debug kernel configuration (Debian's default). Before patch set: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 11277211 0.250 85.340 Close 8283172 0.002 6.479 Rename 477515 1.935 86.026 Unlink 2277936 0.770 87.071 Deltree 256 15.732 81.379 Mkdir 128 0.003 0.009 Qpathinfo 10221180 0.056 44.404 Qfileinfo 1789967 0.002 4.066 Qfsinfo 1874399 0.003 9.176 Sfileinfo 918589 0.061 10.247 Find 3951758 0.341 54.040 WriteX 5616547 0.047 85.079 ReadX 17676028 0.005 9.704 LockX 36704 0.003 1.800 UnlockX 36704 0.002 0.687 Flush 790541 14.115 676.236 Throughput 1179.19 MB/sec 64 clients 64 procs max_latency=676.240 ms After patch set: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 12687926 0.171 86.526 Close 9320780 0.002 8.063 Rename 537253 1.444 78.576 Unlink 2561827 0.559 87.228 Deltree 374 11.499 73.549 Mkdir 187 0.003 0.005 Qpathinfo 11500300 0.061 36.801 Qfileinfo 2017118 0.002 7.189 Qfsinfo 2108641 0.003 4.825 Sfileinfo 1033574 0.008 8.065 Find 4446553 0.408 47.835 WriteX 6335667 0.045 84.388 ReadX 19887312 0.003 9.215 LockX 41312 0.003 1.394 UnlockX 41312 0.002 1.425 Flush 889233 13.014 623.259 Throughput 1339.32 MB/sec 64 clients 64 procs max_latency=623.265 ms +12.7% throughput, -8.2% max latency Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-11-25 15:19:28 +03:00
log_root_start = log_root_tree->node->start;
log_root_level = btrfs_header_level(log_root_tree->node);
log_root_tree->log_transid++;
mutex_unlock(&log_root_tree->log_mutex);
/*
btrfs: do not block inode logging for so long during transaction commit Early on during a transaction commit we acquire the tree_log_mutex and hold it until after we write the super blocks. But before writing the extent buffers dirtied by the transaction and the super blocks we unblock the transaction by setting its state to TRANS_STATE_UNBLOCKED and setting fs_info->running_transaction to NULL. This means that after that and before writing the super blocks, new transactions can start. However if any transaction wants to log an inode, it will block waiting for the transaction commit to write its dirty extent buffers and the super blocks because the tree_log_mutex is only released after those operations are complete, and starting a new log transaction blocks on that mutex (at start_log_trans()). Writing the dirty extent buffers and the super blocks can take a very significant amount of time to complete, but we could allow the tasks wanting to log an inode to proceed with most of their steps: 1) create the log trees 2) log metadata in the trees 3) write their dirty extent buffers They only need to wait for the previous transaction commit to complete (write its super blocks) before they attempt to write their super blocks, otherwise we could end up with a corrupt filesystem after a crash. So change start_log_trans() to use the root tree's log_mutex to serialize for the creation of the log root tree instead of using the tree_log_mutex, and make btrfs_sync_log() acquire the tree_log_mutex before writing the super blocks. This allows for inode logging to wait much less time when there is a previous transaction that is still committing, often not having to wait at all, as by the time when we try to sync the log the previous transaction already wrote its super blocks. This patch belongs to a patch set that is comprised of the following patches: btrfs: fix race causing unnecessary inode logging during link and rename btrfs: fix race that results in logging old extents during a fast fsync btrfs: fix race that causes unnecessary logging of ancestor inodes btrfs: fix race that makes inode logging fallback to transaction commit btrfs: fix race leading to unnecessary transaction commit when logging inode btrfs: do not block inode logging for so long during transaction commit The following script that uses dbench was used to measure the impact of the whole patchset: $ cat test-dbench.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/btrfs MOUNT_OPTIONS="-o ssd" echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f -m single -d single $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -D $MNT -t 300 64 umount $MNT The test was run on a machine with 12 cores, 64G of ram, using a NVMe device and a non-debug kernel configuration (Debian's default). Before patch set: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 11277211 0.250 85.340 Close 8283172 0.002 6.479 Rename 477515 1.935 86.026 Unlink 2277936 0.770 87.071 Deltree 256 15.732 81.379 Mkdir 128 0.003 0.009 Qpathinfo 10221180 0.056 44.404 Qfileinfo 1789967 0.002 4.066 Qfsinfo 1874399 0.003 9.176 Sfileinfo 918589 0.061 10.247 Find 3951758 0.341 54.040 WriteX 5616547 0.047 85.079 ReadX 17676028 0.005 9.704 LockX 36704 0.003 1.800 UnlockX 36704 0.002 0.687 Flush 790541 14.115 676.236 Throughput 1179.19 MB/sec 64 clients 64 procs max_latency=676.240 ms After patch set: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 12687926 0.171 86.526 Close 9320780 0.002 8.063 Rename 537253 1.444 78.576 Unlink 2561827 0.559 87.228 Deltree 374 11.499 73.549 Mkdir 187 0.003 0.005 Qpathinfo 11500300 0.061 36.801 Qfileinfo 2017118 0.002 7.189 Qfsinfo 2108641 0.003 4.825 Sfileinfo 1033574 0.008 8.065 Find 4446553 0.408 47.835 WriteX 6335667 0.045 84.388 ReadX 19887312 0.003 9.215 LockX 41312 0.003 1.394 UnlockX 41312 0.002 1.425 Flush 889233 13.014 623.259 Throughput 1339.32 MB/sec 64 clients 64 procs max_latency=623.265 ms +12.7% throughput, -8.2% max latency Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-11-25 15:19:28 +03:00
* Here we are guaranteed that nobody is going to write the superblock
* for the current transaction before us and that neither we do write
* our superblock before the previous transaction finishes its commit
* and writes its superblock, because:
*
* 1) We are holding a handle on the current transaction, so no body
* can commit it until we release the handle;
*
* 2) Before writing our superblock we acquire the tree_log_mutex, so
* if the previous transaction is still committing, and hasn't yet
* written its superblock, we wait for it to do it, because a
* transaction commit acquires the tree_log_mutex when the commit
* begins and releases it only after writing its superblock.
*/
btrfs: do not block inode logging for so long during transaction commit Early on during a transaction commit we acquire the tree_log_mutex and hold it until after we write the super blocks. But before writing the extent buffers dirtied by the transaction and the super blocks we unblock the transaction by setting its state to TRANS_STATE_UNBLOCKED and setting fs_info->running_transaction to NULL. This means that after that and before writing the super blocks, new transactions can start. However if any transaction wants to log an inode, it will block waiting for the transaction commit to write its dirty extent buffers and the super blocks because the tree_log_mutex is only released after those operations are complete, and starting a new log transaction blocks on that mutex (at start_log_trans()). Writing the dirty extent buffers and the super blocks can take a very significant amount of time to complete, but we could allow the tasks wanting to log an inode to proceed with most of their steps: 1) create the log trees 2) log metadata in the trees 3) write their dirty extent buffers They only need to wait for the previous transaction commit to complete (write its super blocks) before they attempt to write their super blocks, otherwise we could end up with a corrupt filesystem after a crash. So change start_log_trans() to use the root tree's log_mutex to serialize for the creation of the log root tree instead of using the tree_log_mutex, and make btrfs_sync_log() acquire the tree_log_mutex before writing the super blocks. This allows for inode logging to wait much less time when there is a previous transaction that is still committing, often not having to wait at all, as by the time when we try to sync the log the previous transaction already wrote its super blocks. This patch belongs to a patch set that is comprised of the following patches: btrfs: fix race causing unnecessary inode logging during link and rename btrfs: fix race that results in logging old extents during a fast fsync btrfs: fix race that causes unnecessary logging of ancestor inodes btrfs: fix race that makes inode logging fallback to transaction commit btrfs: fix race leading to unnecessary transaction commit when logging inode btrfs: do not block inode logging for so long during transaction commit The following script that uses dbench was used to measure the impact of the whole patchset: $ cat test-dbench.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/btrfs MOUNT_OPTIONS="-o ssd" echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f -m single -d single $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -D $MNT -t 300 64 umount $MNT The test was run on a machine with 12 cores, 64G of ram, using a NVMe device and a non-debug kernel configuration (Debian's default). Before patch set: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 11277211 0.250 85.340 Close 8283172 0.002 6.479 Rename 477515 1.935 86.026 Unlink 2277936 0.770 87.071 Deltree 256 15.732 81.379 Mkdir 128 0.003 0.009 Qpathinfo 10221180 0.056 44.404 Qfileinfo 1789967 0.002 4.066 Qfsinfo 1874399 0.003 9.176 Sfileinfo 918589 0.061 10.247 Find 3951758 0.341 54.040 WriteX 5616547 0.047 85.079 ReadX 17676028 0.005 9.704 LockX 36704 0.003 1.800 UnlockX 36704 0.002 0.687 Flush 790541 14.115 676.236 Throughput 1179.19 MB/sec 64 clients 64 procs max_latency=676.240 ms After patch set: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 12687926 0.171 86.526 Close 9320780 0.002 8.063 Rename 537253 1.444 78.576 Unlink 2561827 0.559 87.228 Deltree 374 11.499 73.549 Mkdir 187 0.003 0.005 Qpathinfo 11500300 0.061 36.801 Qfileinfo 2017118 0.002 7.189 Qfsinfo 2108641 0.003 4.825 Sfileinfo 1033574 0.008 8.065 Find 4446553 0.408 47.835 WriteX 6335667 0.045 84.388 ReadX 19887312 0.003 9.215 LockX 41312 0.003 1.394 UnlockX 41312 0.002 1.425 Flush 889233 13.014 623.259 Throughput 1339.32 MB/sec 64 clients 64 procs max_latency=623.265 ms +12.7% throughput, -8.2% max latency Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-11-25 15:19:28 +03:00
mutex_lock(&fs_info->tree_log_mutex);
/*
* The previous transaction writeout phase could have failed, and thus
* marked the fs in an error state. We must not commit here, as we
* could have updated our generation in the super_for_commit and
* writing the super here would result in transid mismatches. If there
* is an error here just bail.
*/
if (BTRFS_FS_ERROR(fs_info)) {
ret = -EIO;
btrfs_set_log_full_commit(trans);
btrfs_abort_transaction(trans, ret);
mutex_unlock(&fs_info->tree_log_mutex);
goto out_wake_log_root;
}
btrfs: do not block inode logging for so long during transaction commit Early on during a transaction commit we acquire the tree_log_mutex and hold it until after we write the super blocks. But before writing the extent buffers dirtied by the transaction and the super blocks we unblock the transaction by setting its state to TRANS_STATE_UNBLOCKED and setting fs_info->running_transaction to NULL. This means that after that and before writing the super blocks, new transactions can start. However if any transaction wants to log an inode, it will block waiting for the transaction commit to write its dirty extent buffers and the super blocks because the tree_log_mutex is only released after those operations are complete, and starting a new log transaction blocks on that mutex (at start_log_trans()). Writing the dirty extent buffers and the super blocks can take a very significant amount of time to complete, but we could allow the tasks wanting to log an inode to proceed with most of their steps: 1) create the log trees 2) log metadata in the trees 3) write their dirty extent buffers They only need to wait for the previous transaction commit to complete (write its super blocks) before they attempt to write their super blocks, otherwise we could end up with a corrupt filesystem after a crash. So change start_log_trans() to use the root tree's log_mutex to serialize for the creation of the log root tree instead of using the tree_log_mutex, and make btrfs_sync_log() acquire the tree_log_mutex before writing the super blocks. This allows for inode logging to wait much less time when there is a previous transaction that is still committing, often not having to wait at all, as by the time when we try to sync the log the previous transaction already wrote its super blocks. This patch belongs to a patch set that is comprised of the following patches: btrfs: fix race causing unnecessary inode logging during link and rename btrfs: fix race that results in logging old extents during a fast fsync btrfs: fix race that causes unnecessary logging of ancestor inodes btrfs: fix race that makes inode logging fallback to transaction commit btrfs: fix race leading to unnecessary transaction commit when logging inode btrfs: do not block inode logging for so long during transaction commit The following script that uses dbench was used to measure the impact of the whole patchset: $ cat test-dbench.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/btrfs MOUNT_OPTIONS="-o ssd" echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f -m single -d single $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -D $MNT -t 300 64 umount $MNT The test was run on a machine with 12 cores, 64G of ram, using a NVMe device and a non-debug kernel configuration (Debian's default). Before patch set: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 11277211 0.250 85.340 Close 8283172 0.002 6.479 Rename 477515 1.935 86.026 Unlink 2277936 0.770 87.071 Deltree 256 15.732 81.379 Mkdir 128 0.003 0.009 Qpathinfo 10221180 0.056 44.404 Qfileinfo 1789967 0.002 4.066 Qfsinfo 1874399 0.003 9.176 Sfileinfo 918589 0.061 10.247 Find 3951758 0.341 54.040 WriteX 5616547 0.047 85.079 ReadX 17676028 0.005 9.704 LockX 36704 0.003 1.800 UnlockX 36704 0.002 0.687 Flush 790541 14.115 676.236 Throughput 1179.19 MB/sec 64 clients 64 procs max_latency=676.240 ms After patch set: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 12687926 0.171 86.526 Close 9320780 0.002 8.063 Rename 537253 1.444 78.576 Unlink 2561827 0.559 87.228 Deltree 374 11.499 73.549 Mkdir 187 0.003 0.005 Qpathinfo 11500300 0.061 36.801 Qfileinfo 2017118 0.002 7.189 Qfsinfo 2108641 0.003 4.825 Sfileinfo 1033574 0.008 8.065 Find 4446553 0.408 47.835 WriteX 6335667 0.045 84.388 ReadX 19887312 0.003 9.215 LockX 41312 0.003 1.394 UnlockX 41312 0.002 1.425 Flush 889233 13.014 623.259 Throughput 1339.32 MB/sec 64 clients 64 procs max_latency=623.265 ms +12.7% throughput, -8.2% max latency Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-11-25 15:19:28 +03:00
btrfs_set_super_log_root(fs_info->super_for_commit, log_root_start);
btrfs_set_super_log_root_level(fs_info->super_for_commit, log_root_level);
ret = write_all_supers(fs_info, 1);
btrfs: do not block inode logging for so long during transaction commit Early on during a transaction commit we acquire the tree_log_mutex and hold it until after we write the super blocks. But before writing the extent buffers dirtied by the transaction and the super blocks we unblock the transaction by setting its state to TRANS_STATE_UNBLOCKED and setting fs_info->running_transaction to NULL. This means that after that and before writing the super blocks, new transactions can start. However if any transaction wants to log an inode, it will block waiting for the transaction commit to write its dirty extent buffers and the super blocks because the tree_log_mutex is only released after those operations are complete, and starting a new log transaction blocks on that mutex (at start_log_trans()). Writing the dirty extent buffers and the super blocks can take a very significant amount of time to complete, but we could allow the tasks wanting to log an inode to proceed with most of their steps: 1) create the log trees 2) log metadata in the trees 3) write their dirty extent buffers They only need to wait for the previous transaction commit to complete (write its super blocks) before they attempt to write their super blocks, otherwise we could end up with a corrupt filesystem after a crash. So change start_log_trans() to use the root tree's log_mutex to serialize for the creation of the log root tree instead of using the tree_log_mutex, and make btrfs_sync_log() acquire the tree_log_mutex before writing the super blocks. This allows for inode logging to wait much less time when there is a previous transaction that is still committing, often not having to wait at all, as by the time when we try to sync the log the previous transaction already wrote its super blocks. This patch belongs to a patch set that is comprised of the following patches: btrfs: fix race causing unnecessary inode logging during link and rename btrfs: fix race that results in logging old extents during a fast fsync btrfs: fix race that causes unnecessary logging of ancestor inodes btrfs: fix race that makes inode logging fallback to transaction commit btrfs: fix race leading to unnecessary transaction commit when logging inode btrfs: do not block inode logging for so long during transaction commit The following script that uses dbench was used to measure the impact of the whole patchset: $ cat test-dbench.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/btrfs MOUNT_OPTIONS="-o ssd" echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f -m single -d single $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -D $MNT -t 300 64 umount $MNT The test was run on a machine with 12 cores, 64G of ram, using a NVMe device and a non-debug kernel configuration (Debian's default). Before patch set: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 11277211 0.250 85.340 Close 8283172 0.002 6.479 Rename 477515 1.935 86.026 Unlink 2277936 0.770 87.071 Deltree 256 15.732 81.379 Mkdir 128 0.003 0.009 Qpathinfo 10221180 0.056 44.404 Qfileinfo 1789967 0.002 4.066 Qfsinfo 1874399 0.003 9.176 Sfileinfo 918589 0.061 10.247 Find 3951758 0.341 54.040 WriteX 5616547 0.047 85.079 ReadX 17676028 0.005 9.704 LockX 36704 0.003 1.800 UnlockX 36704 0.002 0.687 Flush 790541 14.115 676.236 Throughput 1179.19 MB/sec 64 clients 64 procs max_latency=676.240 ms After patch set: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 12687926 0.171 86.526 Close 9320780 0.002 8.063 Rename 537253 1.444 78.576 Unlink 2561827 0.559 87.228 Deltree 374 11.499 73.549 Mkdir 187 0.003 0.005 Qpathinfo 11500300 0.061 36.801 Qfileinfo 2017118 0.002 7.189 Qfsinfo 2108641 0.003 4.825 Sfileinfo 1033574 0.008 8.065 Find 4446553 0.408 47.835 WriteX 6335667 0.045 84.388 ReadX 19887312 0.003 9.215 LockX 41312 0.003 1.394 UnlockX 41312 0.002 1.425 Flush 889233 13.014 623.259 Throughput 1339.32 MB/sec 64 clients 64 procs max_latency=623.265 ms +12.7% throughput, -8.2% max latency Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-11-25 15:19:28 +03:00
mutex_unlock(&fs_info->tree_log_mutex);
if (ret) {
btrfs_set_log_full_commit(trans);
btrfs_abort_transaction(trans, ret);
goto out_wake_log_root;
}
/*
* We know there can only be one task here, since we have not yet set
* root->log_commit[index1] to 0 and any task attempting to sync the
* log must wait for the previous log transaction to commit if it's
* still in progress or wait for the current log transaction commit if
* someone else already started it. We use <= and not < because the
* first log transaction has an ID of 0.
*/
ASSERT(root->last_log_commit <= log_transid);
root->last_log_commit = log_transid;
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 17:24:20 +03:00
out_wake_log_root:
mutex_lock(&log_root_tree->log_mutex);
btrfs_remove_all_log_ctxs(log_root_tree, index2, ret);
log_root_tree->log_transid_committed++;
atomic_set(&log_root_tree->log_commit[index2], 0);
mutex_unlock(&log_root_tree->log_mutex);
/*
* The barrier before waitqueue_active (in cond_wake_up) is needed so
* all the updates above are seen by the woken threads. It might not be
* necessary, but proving that seems to be hard.
*/
cond_wake_up(&log_root_tree->log_commit_wait[index2]);
out:
mutex_lock(&root->log_mutex);
btrfs_remove_all_log_ctxs(root, index1, ret);
root->log_transid_committed++;
atomic_set(&root->log_commit[index1], 0);
mutex_unlock(&root->log_mutex);
/*
* The barrier before waitqueue_active (in cond_wake_up) is needed so
* all the updates above are seen by the woken threads. It might not be
* necessary, but proving that seems to be hard.
*/
cond_wake_up(&root->log_commit_wait[index1]);
return ret;
}
static void free_log_tree(struct btrfs_trans_handle *trans,
struct btrfs_root *log)
{
int ret;
struct walk_control wc = {
.free = 1,
.process_func = process_one_buffer
};
if (log->node) {
ret = walk_log_tree(trans, log, &wc);
if (ret) {
btrfs: skip reserved bytes warning on unmount after log cleanup failure After the recent changes made by commit c2e39305299f01 ("btrfs: clear extent buffer uptodate when we fail to write it") and its followup fix, commit 651740a5024117 ("btrfs: check WRITE_ERR when trying to read an extent buffer"), we can now end up not cleaning up space reservations of log tree extent buffers after a transaction abort happens, as well as not cleaning up still dirty extent buffers. This happens because if writeback for a log tree extent buffer failed, then we have cleared the bit EXTENT_BUFFER_UPTODATE from the extent buffer and we have also set the bit EXTENT_BUFFER_WRITE_ERR on it. Later on, when trying to free the log tree with free_log_tree(), which iterates over the tree, we can end up getting an -EIO error when trying to read a node or a leaf, since read_extent_buffer_pages() returns -EIO if an extent buffer does not have EXTENT_BUFFER_UPTODATE set and has the EXTENT_BUFFER_WRITE_ERR bit set. Getting that -EIO means that we return immediately as we can not iterate over the entire tree. In that case we never update the reserved space for an extent buffer in the respective block group and space_info object. When this happens we get the following traces when unmounting the fs: [174957.284509] BTRFS: error (device dm-0) in cleanup_transaction:1913: errno=-5 IO failure [174957.286497] BTRFS: error (device dm-0) in free_log_tree:3420: errno=-5 IO failure [174957.399379] ------------[ cut here ]------------ [174957.402497] WARNING: CPU: 2 PID: 3206883 at fs/btrfs/block-group.c:127 btrfs_put_block_group+0x77/0xb0 [btrfs] [174957.407523] Modules linked in: btrfs overlay dm_zero (...) [174957.424917] CPU: 2 PID: 3206883 Comm: umount Tainted: G W 5.16.0-rc5-btrfs-next-109 #1 [174957.426689] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.14.0-0-g155821a1990b-prebuilt.qemu.org 04/01/2014 [174957.428716] RIP: 0010:btrfs_put_block_group+0x77/0xb0 [btrfs] [174957.429717] Code: 21 48 8b bd (...) [174957.432867] RSP: 0018:ffffb70d41cffdd0 EFLAGS: 00010206 [174957.433632] RAX: 0000000000000001 RBX: ffff8b09c3848000 RCX: ffff8b0758edd1c8 [174957.434689] RDX: 0000000000000001 RSI: ffffffffc0b467e7 RDI: ffff8b0758edd000 [174957.436068] RBP: ffff8b0758edd000 R08: 0000000000000000 R09: 0000000000000000 [174957.437114] R10: 0000000000000246 R11: 0000000000000000 R12: ffff8b09c3848148 [174957.438140] R13: ffff8b09c3848198 R14: ffff8b0758edd188 R15: dead000000000100 [174957.439317] FS: 00007f328fb82800(0000) GS:ffff8b0a2d200000(0000) knlGS:0000000000000000 [174957.440402] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [174957.441164] CR2: 00007fff13563e98 CR3: 0000000404f4e005 CR4: 0000000000370ee0 [174957.442117] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [174957.443076] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [174957.443948] Call Trace: [174957.444264] <TASK> [174957.444538] btrfs_free_block_groups+0x255/0x3c0 [btrfs] [174957.445238] close_ctree+0x301/0x357 [btrfs] [174957.445803] ? call_rcu+0x16c/0x290 [174957.446250] generic_shutdown_super+0x74/0x120 [174957.446832] kill_anon_super+0x14/0x30 [174957.447305] btrfs_kill_super+0x12/0x20 [btrfs] [174957.447890] deactivate_locked_super+0x31/0xa0 [174957.448440] cleanup_mnt+0x147/0x1c0 [174957.448888] task_work_run+0x5c/0xa0 [174957.449336] exit_to_user_mode_prepare+0x1e5/0x1f0 [174957.449934] syscall_exit_to_user_mode+0x16/0x40 [174957.450512] do_syscall_64+0x48/0xc0 [174957.450980] entry_SYSCALL_64_after_hwframe+0x44/0xae [174957.451605] RIP: 0033:0x7f328fdc4a97 [174957.452059] Code: 03 0c 00 f7 (...) [174957.454320] RSP: 002b:00007fff13564ec8 EFLAGS: 00000246 ORIG_RAX: 00000000000000a6 [174957.455262] RAX: 0000000000000000 RBX: 00007f328feea264 RCX: 00007f328fdc4a97 [174957.456131] RDX: 0000000000000000 RSI: 0000000000000000 RDI: 0000560b8ae51dd0 [174957.457118] RBP: 0000560b8ae51ba0 R08: 0000000000000000 R09: 00007fff13563c40 [174957.458005] R10: 00007f328fe49fc0 R11: 0000000000000246 R12: 0000000000000000 [174957.459113] R13: 0000560b8ae51dd0 R14: 0000560b8ae51cb0 R15: 0000000000000000 [174957.460193] </TASK> [174957.460534] irq event stamp: 0 [174957.461003] hardirqs last enabled at (0): [<0000000000000000>] 0x0 [174957.461947] hardirqs last disabled at (0): [<ffffffffb0e94214>] copy_process+0x934/0x2040 [174957.463147] softirqs last enabled at (0): [<ffffffffb0e94214>] copy_process+0x934/0x2040 [174957.465116] softirqs last disabled at (0): [<0000000000000000>] 0x0 [174957.466323] ---[ end trace bc7ee0c490bce3af ]--- [174957.467282] ------------[ cut here ]------------ [174957.468184] WARNING: CPU: 2 PID: 3206883 at fs/btrfs/block-group.c:3976 btrfs_free_block_groups+0x330/0x3c0 [btrfs] [174957.470066] Modules linked in: btrfs overlay dm_zero (...) [174957.483137] CPU: 2 PID: 3206883 Comm: umount Tainted: G W 5.16.0-rc5-btrfs-next-109 #1 [174957.484691] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.14.0-0-g155821a1990b-prebuilt.qemu.org 04/01/2014 [174957.486853] RIP: 0010:btrfs_free_block_groups+0x330/0x3c0 [btrfs] [174957.488050] Code: 00 00 00 ad de (...) [174957.491479] RSP: 0018:ffffb70d41cffde0 EFLAGS: 00010206 [174957.492520] RAX: ffff8b08d79310b0 RBX: ffff8b09c3848000 RCX: 0000000000000000 [174957.493868] RDX: 0000000000000001 RSI: fffff443055ee600 RDI: ffffffffb1131846 [174957.495183] RBP: ffff8b08d79310b0 R08: 0000000000000000 R09: 0000000000000000 [174957.496580] R10: 0000000000000001 R11: 0000000000000000 R12: ffff8b08d7931000 [174957.498027] R13: ffff8b09c38492b0 R14: dead000000000122 R15: dead000000000100 [174957.499438] FS: 00007f328fb82800(0000) GS:ffff8b0a2d200000(0000) knlGS:0000000000000000 [174957.500990] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [174957.502117] CR2: 00007fff13563e98 CR3: 0000000404f4e005 CR4: 0000000000370ee0 [174957.503513] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [174957.504864] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [174957.506167] Call Trace: [174957.506654] <TASK> [174957.507047] close_ctree+0x301/0x357 [btrfs] [174957.507867] ? call_rcu+0x16c/0x290 [174957.508567] generic_shutdown_super+0x74/0x120 [174957.509447] kill_anon_super+0x14/0x30 [174957.510194] btrfs_kill_super+0x12/0x20 [btrfs] [174957.511123] deactivate_locked_super+0x31/0xa0 [174957.511976] cleanup_mnt+0x147/0x1c0 [174957.512610] task_work_run+0x5c/0xa0 [174957.513309] exit_to_user_mode_prepare+0x1e5/0x1f0 [174957.514231] syscall_exit_to_user_mode+0x16/0x40 [174957.515069] do_syscall_64+0x48/0xc0 [174957.515718] entry_SYSCALL_64_after_hwframe+0x44/0xae [174957.516688] RIP: 0033:0x7f328fdc4a97 [174957.517413] Code: 03 0c 00 f7 d8 (...) [174957.521052] RSP: 002b:00007fff13564ec8 EFLAGS: 00000246 ORIG_RAX: 00000000000000a6 [174957.522514] RAX: 0000000000000000 RBX: 00007f328feea264 RCX: 00007f328fdc4a97 [174957.523950] RDX: 0000000000000000 RSI: 0000000000000000 RDI: 0000560b8ae51dd0 [174957.525375] RBP: 0000560b8ae51ba0 R08: 0000000000000000 R09: 00007fff13563c40 [174957.526763] R10: 00007f328fe49fc0 R11: 0000000000000246 R12: 0000000000000000 [174957.528058] R13: 0000560b8ae51dd0 R14: 0000560b8ae51cb0 R15: 0000000000000000 [174957.529404] </TASK> [174957.529843] irq event stamp: 0 [174957.530256] hardirqs last enabled at (0): [<0000000000000000>] 0x0 [174957.531061] hardirqs last disabled at (0): [<ffffffffb0e94214>] copy_process+0x934/0x2040 [174957.532075] softirqs last enabled at (0): [<ffffffffb0e94214>] copy_process+0x934/0x2040 [174957.533083] softirqs last disabled at (0): [<0000000000000000>] 0x0 [174957.533865] ---[ end trace bc7ee0c490bce3b0 ]--- [174957.534452] BTRFS info (device dm-0): space_info 4 has 1070841856 free, is not full [174957.535404] BTRFS info (device dm-0): space_info total=1073741824, used=2785280, pinned=0, reserved=49152, may_use=0, readonly=65536 zone_unusable=0 [174957.537029] BTRFS info (device dm-0): global_block_rsv: size 0 reserved 0 [174957.537859] BTRFS info (device dm-0): trans_block_rsv: size 0 reserved 0 [174957.538697] BTRFS info (device dm-0): chunk_block_rsv: size 0 reserved 0 [174957.539552] BTRFS info (device dm-0): delayed_block_rsv: size 0 reserved 0 [174957.540403] BTRFS info (device dm-0): delayed_refs_rsv: size 0 reserved 0 This also means that in case we have log tree extent buffers that are still dirty, we can end up not cleaning them up in case we find an extent buffer with EXTENT_BUFFER_WRITE_ERR set on it, as in that case we have no way for iterating over the rest of the tree. This issue is very often triggered with test cases generic/475 and generic/648 from fstests. The issue could almost be fixed by iterating over the io tree attached to each log root which keeps tracks of the range of allocated extent buffers, log_root->dirty_log_pages, however that does not work and has some inconveniences: 1) After we sync the log, we clear the range of the extent buffers from the io tree, so we can't find them after writeback. We could keep the ranges in the io tree, with a separate bit to signal they represent extent buffers already written, but that means we need to hold into more memory until the transaction commits. How much more memory is used depends a lot on whether we are able to allocate contiguous extent buffers on disk (and how often) for a log tree - if we are able to, then a single extent state record can represent multiple extent buffers, otherwise we need multiple extent state record structures to track each extent buffer. In fact, my earlier approach did that: https://lore.kernel.org/linux-btrfs/3aae7c6728257c7ce2279d6660ee2797e5e34bbd.1641300250.git.fdmanana@suse.com/ However that can cause a very significant negative impact on performance, not only due to the extra memory usage but also because we get a larger and deeper dirty_log_pages io tree. We got a report that, on beefy machines at least, we can get such performance drop with fsmark for example: https://lore.kernel.org/linux-btrfs/20220117082426.GE32491@xsang-OptiPlex-9020/ 2) We would be doing it only to deal with an unexpected and exceptional case, which is basically failure to read an extent buffer from disk due to IO failures. On a healthy system we don't expect transaction aborts to happen after all; 3) Instead of relying on iterating the log tree or tracking the ranges of extent buffers in the dirty_log_pages io tree, using the radix tree that tracks extent buffers (fs_info->buffer_radix) to find all log tree extent buffers is not reliable either, because after writeback of an extent buffer it can be evicted from memory by the release page callback of the btree inode (btree_releasepage()). Since there's no way to be able to properly cleanup a log tree without being able to read its extent buffers from disk and without using more memory to track the logical ranges of the allocated extent buffers do the following: 1) When we fail to cleanup a log tree, setup a flag that indicates that failure; 2) Trigger writeback of all log tree extent buffers that are still dirty, and wait for the writeback to complete. This is just to cleanup their state, page states, page leaks, etc; 3) When unmounting the fs, ignore if the number of bytes reserved in a block group and in a space_info is not 0 if, and only if, we failed to cleanup a log tree. Also ignore only for metadata block groups and the metadata space_info object. This is far from a perfect solution, but it serves to silence test failures such as those from generic/475 and generic/648. However having a non-zero value for the reserved bytes counters on unmount after a transaction abort, is not such a terrible thing and it's completely harmless, it does not affect the filesystem integrity in any way. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-01-18 16:39:34 +03:00
/*
* We weren't able to traverse the entire log tree, the
* typical scenario is getting an -EIO when reading an
* extent buffer of the tree, due to a previous writeback
* failure of it.
*/
set_bit(BTRFS_FS_STATE_LOG_CLEANUP_ERROR,
&log->fs_info->fs_state);
/*
* Some extent buffers of the log tree may still be dirty
* and not yet written back to storage, because we may
* have updates to a log tree without syncing a log tree,
* such as during rename and link operations. So flush
* them out and wait for their writeback to complete, so
* that we properly cleanup their state and pages.
*/
btrfs_write_marked_extents(log->fs_info,
&log->dirty_log_pages,
EXTENT_DIRTY | EXTENT_NEW);
btrfs_wait_tree_log_extents(log,
EXTENT_DIRTY | EXTENT_NEW);
if (trans)
btrfs_abort_transaction(trans, ret);
else
btrfs_handle_fs_error(log->fs_info, ret, NULL);
}
btrfs: fix error handling in free_log_tree When we hit an I/O error in free_log_tree->walk_log_tree during file system shutdown we can crash due to there not being a valid transaction handle. Use btrfs_handle_fs_error when there's no transaction handle to use. BUG: unable to handle kernel NULL pointer dereference at 0000000000000060 IP: free_log_tree+0xd2/0x140 [btrfs] PGD 0 P4D 0 Oops: 0000 [#1] SMP DEBUG_PAGEALLOC PTI Modules linked in: <modules> CPU: 2 PID: 23544 Comm: umount Tainted: G W 4.12.14-kvmsmall #9 SLE15 (unreleased) Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.0.0-prebuilt.qemu-project.org 04/01/2014 task: ffff96bfd3478880 task.stack: ffffa7cf40d78000 RIP: 0010:free_log_tree+0xd2/0x140 [btrfs] RSP: 0018:ffffa7cf40d7bd10 EFLAGS: 00010282 RAX: 00000000fffffffb RBX: 00000000fffffffb RCX: 0000000000000002 RDX: 0000000000000000 RSI: ffff96c02f07d4c8 RDI: 0000000000000282 RBP: ffff96c013cf1000 R08: ffff96c02f07d4c8 R09: ffff96c02f07d4d0 R10: 0000000000000000 R11: 0000000000000002 R12: 0000000000000000 R13: ffff96c005e800c0 R14: ffffa7cf40d7bdb8 R15: 0000000000000000 FS: 00007f17856bcfc0(0000) GS:ffff96c03f600000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 0000000000000060 CR3: 0000000045ed6002 CR4: 00000000003606e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: ? wait_for_writer+0xb0/0xb0 [btrfs] btrfs_free_log+0x17/0x30 [btrfs] btrfs_drop_and_free_fs_root+0x9a/0xe0 [btrfs] btrfs_free_fs_roots+0xc0/0x130 [btrfs] ? wait_for_completion+0xf2/0x100 close_ctree+0xea/0x2e0 [btrfs] ? kthread_stop+0x161/0x260 generic_shutdown_super+0x6c/0x120 kill_anon_super+0xe/0x20 btrfs_kill_super+0x13/0x100 [btrfs] deactivate_locked_super+0x3f/0x70 cleanup_mnt+0x3b/0x70 task_work_run+0x78/0x90 exit_to_usermode_loop+0x77/0xa6 do_syscall_64+0x1c5/0x1e0 entry_SYSCALL_64_after_hwframe+0x42/0xb7 RIP: 0033:0x7f1784f90827 RSP: 002b:00007ffdeeb03118 EFLAGS: 00000246 ORIG_RAX: 00000000000000a6 RAX: 0000000000000000 RBX: 0000556a60c62970 RCX: 00007f1784f90827 RDX: 0000000000000001 RSI: 0000000000000000 RDI: 0000556a60c62b50 RBP: 0000000000000000 R08: 0000000000000005 R09: 00000000ffffffff R10: 0000556a60c63900 R11: 0000000000000246 R12: 0000556a60c62b50 R13: 00007f17854a81c4 R14: 0000000000000000 R15: 0000000000000000 RIP: free_log_tree+0xd2/0x140 [btrfs] RSP: ffffa7cf40d7bd10 CR2: 0000000000000060 Fixes: 681ae50917df9 ("Btrfs: cleanup reserved space when freeing tree log on error") CC: <stable@vger.kernel.org> # v3.13 Signed-off-by: Jeff Mahoney <jeffm@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2018-09-06 23:59:33 +03:00
}
clear_extent_bits(&log->dirty_log_pages, 0, (u64)-1,
EXTENT_DIRTY | EXTENT_NEW | EXTENT_NEED_WAIT);
btrfs: fix corrupt log due to concurrent fsync of inodes with shared extents When we have extents shared amongst different inodes in the same subvolume, if we fsync them in parallel we can end up with checksum items in the log tree that represent ranges which overlap. For example, consider we have inodes A and B, both sharing an extent that covers the logical range from X to X + 64KiB: 1) Task A starts an fsync on inode A; 2) Task B starts an fsync on inode B; 3) Task A calls btrfs_csum_file_blocks(), and the first search in the log tree, through btrfs_lookup_csum(), returns -EFBIG because it finds an existing checksum item that covers the range from X - 64KiB to X; 4) Task A checks that the checksum item has not reached the maximum possible size (MAX_CSUM_ITEMS) and then releases the search path before it does another path search for insertion (through a direct call to btrfs_search_slot()); 5) As soon as task A releases the path and before it does the search for insertion, task B calls btrfs_csum_file_blocks() and gets -EFBIG too, because there is an existing checksum item that has an end offset that matches the start offset (X) of the checksum range we want to log; 6) Task B releases the path; 7) Task A does the path search for insertion (through btrfs_search_slot()) and then verifies that the checksum item that ends at offset X still exists and extends its size to insert the checksums for the range from X to X + 64KiB; 8) Task A releases the path and returns from btrfs_csum_file_blocks(), having inserted the checksums into an existing checksum item that got its size extended. At this point we have one checksum item in the log tree that covers the logical range from X - 64KiB to X + 64KiB; 9) Task B now does a search for insertion using btrfs_search_slot() too, but it finds that the previous checksum item no longer ends at the offset X, it now ends at an of offset X + 64KiB, so it leaves that item untouched. Then it releases the path and calls btrfs_insert_empty_item() that inserts a checksum item with a key offset corresponding to X and a size for inserting a single checksum (4 bytes in case of crc32c). Subsequent iterations end up extending this new checksum item so that it contains the checksums for the range from X to X + 64KiB. So after task B returns from btrfs_csum_file_blocks() we end up with two checksum items in the log tree that have overlapping ranges, one for the range from X - 64KiB to X + 64KiB, and another for the range from X to X + 64KiB. Having checksum items that represent ranges which overlap, regardless of being in the log tree or in the chekcsums tree, can lead to problems where checksums for a file range end up not being found. This type of problem has happened a few times in the past and the following commits fixed them and explain in detail why having checksum items with overlapping ranges is problematic: 27b9a8122ff71a "Btrfs: fix csum tree corruption, duplicate and outdated checksums" b84b8390d6009c "Btrfs: fix file read corruption after extent cloning and fsync" 40e046acbd2f36 "Btrfs: fix missing data checksums after replaying a log tree" Since this specific instance of the problem can only happen when logging inodes, because it is the only case where concurrent attempts to insert checksums for the same range can happen, fix the issue by using an extent io tree as a range lock to serialize checksum insertion during inode logging. This issue could often be reproduced by the test case generic/457 from fstests. When it happens it produces the following trace: BTRFS critical (device dm-0): corrupt leaf: root=18446744073709551610 block=30625792 slot=42, csum end range (15020032) goes beyond the start range (15015936) of the next csum item BTRFS info (device dm-0): leaf 30625792 gen 7 total ptrs 49 free space 2402 owner 18446744073709551610 BTRFS info (device dm-0): refs 1 lock (w:0 r:0 bw:0 br:0 sw:0 sr:0) lock_owner 0 current 15884 item 0 key (18446744073709551606 128 13979648) itemoff 3991 itemsize 4 item 1 key (18446744073709551606 128 13983744) itemoff 3987 itemsize 4 item 2 key (18446744073709551606 128 13987840) itemoff 3983 itemsize 4 item 3 key (18446744073709551606 128 13991936) itemoff 3979 itemsize 4 item 4 key (18446744073709551606 128 13996032) itemoff 3975 itemsize 4 item 5 key (18446744073709551606 128 14000128) itemoff 3971 itemsize 4 (...) BTRFS error (device dm-0): block=30625792 write time tree block corruption detected ------------[ cut here ]------------ WARNING: CPU: 1 PID: 15884 at fs/btrfs/disk-io.c:539 btree_csum_one_bio+0x268/0x2d0 [btrfs] Modules linked in: btrfs dm_thin_pool ... CPU: 1 PID: 15884 Comm: fsx Tainted: G W 5.6.0-rc7-btrfs-next-58 #1 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.12.0-59-gc9ba5276e321-prebuilt.qemu.org 04/01/2014 RIP: 0010:btree_csum_one_bio+0x268/0x2d0 [btrfs] Code: c7 c7 ... RSP: 0018:ffffbb0109e6f8e0 EFLAGS: 00010296 RAX: 0000000000000000 RBX: ffffe1c0847b6080 RCX: 0000000000000000 RDX: 0000000000000000 RSI: ffffffffaa963988 RDI: 0000000000000001 RBP: ffff956a4f4d2000 R08: 0000000000000000 R09: 0000000000000001 R10: 0000000000000526 R11: 0000000000000000 R12: ffff956a5cd28bb0 R13: 0000000000000000 R14: ffff956a649c9388 R15: 000000011ed82000 FS: 00007fb419959e80(0000) GS:ffff956a7aa00000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 0000000000fe6d54 CR3: 0000000138696005 CR4: 00000000003606e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: btree_submit_bio_hook+0x67/0xc0 [btrfs] submit_one_bio+0x31/0x50 [btrfs] btree_write_cache_pages+0x2db/0x4b0 [btrfs] ? __filemap_fdatawrite_range+0xb1/0x110 do_writepages+0x23/0x80 __filemap_fdatawrite_range+0xd2/0x110 btrfs_write_marked_extents+0x15e/0x180 [btrfs] btrfs_sync_log+0x206/0x10a0 [btrfs] ? kmem_cache_free+0x315/0x3b0 ? btrfs_log_inode+0x1e8/0xf90 [btrfs] ? __mutex_unlock_slowpath+0x45/0x2a0 ? lockref_put_or_lock+0x9/0x30 ? dput+0x2d/0x580 ? dput+0xb5/0x580 ? btrfs_sync_file+0x464/0x4d0 [btrfs] btrfs_sync_file+0x464/0x4d0 [btrfs] do_fsync+0x38/0x60 __x64_sys_fsync+0x10/0x20 do_syscall_64+0x5c/0x280 entry_SYSCALL_64_after_hwframe+0x49/0xbe RIP: 0033:0x7fb41953a6d0 Code: 48 3d ... RSP: 002b:00007ffcc86bd218 EFLAGS: 00000246 ORIG_RAX: 000000000000004a RAX: ffffffffffffffda RBX: 000000000000000d RCX: 00007fb41953a6d0 RDX: 0000000000000009 RSI: 0000000000040000 RDI: 0000000000000003 RBP: 0000000000040000 R08: 0000000000000001 R09: 0000000000000009 R10: 0000000000000064 R11: 0000000000000246 R12: 0000556cf4b2c060 R13: 0000000000000100 R14: 0000000000000000 R15: 0000556cf322b420 irq event stamp: 0 hardirqs last enabled at (0): [<0000000000000000>] 0x0 hardirqs last disabled at (0): [<ffffffffa96bdedf>] copy_process+0x74f/0x2020 softirqs last enabled at (0): [<ffffffffa96bdedf>] copy_process+0x74f/0x2020 softirqs last disabled at (0): [<0000000000000000>] 0x0 ---[ end trace d543fc76f5ad7fd8 ]--- In that trace the tree checker detected the overlapping checksum items at the time when we triggered writeback for the log tree when syncing the log. Another trace that can happen is due to BUG_ON() when deleting checksum items while logging an inode: BTRFS critical (device dm-0): slot 81 key (18446744073709551606 128 13635584) new key (18446744073709551606 128 13635584) BTRFS info (device dm-0): leaf 30949376 gen 7 total ptrs 98 free space 8527 owner 18446744073709551610 BTRFS info (device dm-0): refs 4 lock (w:1 r:0 bw:0 br:0 sw:1 sr:0) lock_owner 13473 current 13473 item 0 key (257 1 0) itemoff 16123 itemsize 160 inode generation 7 size 262144 mode 100600 item 1 key (257 12 256) itemoff 16103 itemsize 20 item 2 key (257 108 0) itemoff 16050 itemsize 53 extent data disk bytenr 13631488 nr 4096 extent data offset 0 nr 131072 ram 131072 (...) ------------[ cut here ]------------ kernel BUG at fs/btrfs/ctree.c:3153! invalid opcode: 0000 [#1] PREEMPT SMP DEBUG_PAGEALLOC PTI CPU: 1 PID: 13473 Comm: fsx Not tainted 5.6.0-rc7-btrfs-next-58 #1 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.12.0-59-gc9ba5276e321-prebuilt.qemu.org 04/01/2014 RIP: 0010:btrfs_set_item_key_safe+0x1ea/0x270 [btrfs] Code: 0f b6 ... RSP: 0018:ffff95e3889179d0 EFLAGS: 00010282 RAX: 0000000000000000 RBX: 0000000000000051 RCX: 0000000000000000 RDX: 0000000000000000 RSI: ffffffffb7763988 RDI: 0000000000000001 RBP: fffffffffffffff6 R08: 0000000000000000 R09: 0000000000000001 R10: 00000000000009ef R11: 0000000000000000 R12: ffff8912a8ba5a08 R13: ffff95e388917a06 R14: ffff89138dcf68c8 R15: ffff95e388917ace FS: 00007fe587084e80(0000) GS:ffff8913baa00000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 00007fe587091000 CR3: 0000000126dac005 CR4: 00000000003606e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: btrfs_del_csums+0x2f4/0x540 [btrfs] copy_items+0x4b5/0x560 [btrfs] btrfs_log_inode+0x910/0xf90 [btrfs] btrfs_log_inode_parent+0x2a0/0xe40 [btrfs] ? dget_parent+0x5/0x370 btrfs_log_dentry_safe+0x4a/0x70 [btrfs] btrfs_sync_file+0x42b/0x4d0 [btrfs] __x64_sys_msync+0x199/0x200 do_syscall_64+0x5c/0x280 entry_SYSCALL_64_after_hwframe+0x49/0xbe RIP: 0033:0x7fe586c65760 Code: 00 f7 ... RSP: 002b:00007ffe250f98b8 EFLAGS: 00000246 ORIG_RAX: 000000000000001a RAX: ffffffffffffffda RBX: 00000000000040e1 RCX: 00007fe586c65760 RDX: 0000000000000004 RSI: 0000000000006b51 RDI: 00007fe58708b000 RBP: 0000000000006a70 R08: 0000000000000003 R09: 00007fe58700cb61 R10: 0000000000000100 R11: 0000000000000246 R12: 00000000000000e1 R13: 00007fe58708b000 R14: 0000000000006b51 R15: 0000558de021a420 Modules linked in: dm_log_writes ... ---[ end trace c92a7f447a8515f5 ]--- CC: stable@vger.kernel.org # 4.4+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-05-18 14:14:50 +03:00
extent_io_tree_release(&log->log_csum_range);
btrfs_put_root(log);
}
/*
* free all the extents used by the tree log. This should be called
* at commit time of the full transaction
*/
int btrfs_free_log(struct btrfs_trans_handle *trans, struct btrfs_root *root)
{
if (root->log_root) {
free_log_tree(trans, root->log_root);
root->log_root = NULL;
btrfs: check if a log root exists before locking the log_mutex on unlink This brings back an optimization that commit e678934cbe5f02 ("btrfs: Remove unnecessary check from join_running_log_trans") removed, but in a different form. So it's almost equivalent to a revert. That commit removed an optimization where we avoid locking a root's log_mutex when there is no log tree created in the current transaction. The affected code path is triggered through unlink operations. That commit was based on the assumption that the optimization was not necessary because we used to have the following checks when the patch was authored: int btrfs_del_dir_entries_in_log(...) { (...) if (dir->logged_trans < trans->transid) return 0; ret = join_running_log_trans(root); (...) } int btrfs_del_inode_ref_in_log(...) { (...) if (inode->logged_trans < trans->transid) return 0; ret = join_running_log_trans(root); (...) } However before that patch was merged, another patch was merged first which replaced those checks because they were buggy. That other patch corresponds to commit 803f0f64d17769 ("Btrfs: fix fsync not persisting dentry deletions due to inode evictions"). The assumption that if the logged_trans field of an inode had a smaller value then the current transaction's generation (transid) meant that the inode was not logged in the current transaction was only correct if the inode was not evicted and reloaded in the current transaction. So the corresponding bug fix changed those checks and replaced them with the following helper function: static bool inode_logged(struct btrfs_trans_handle *trans, struct btrfs_inode *inode) { if (inode->logged_trans == trans->transid) return true; if (inode->last_trans == trans->transid && test_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags) && !test_bit(BTRFS_FS_LOG_RECOVERING, &trans->fs_info->flags)) return true; return false; } So if we have a subvolume without a log tree in the current transaction (because we had no fsyncs), every time we unlink an inode we can end up trying to lock the log_mutex of the root through join_running_log_trans() twice, once for the inode being unlinked (by btrfs_del_inode_ref_in_log()) and once for the parent directory (with btrfs_del_dir_entries_in_log()). This means if we have several unlink operations happening in parallel for inodes in the same subvolume, and the those inodes and/or their parent inode were changed in the current transaction, we end up having a lot of contention on the log_mutex. The test robots from intel reported a -30.7% performance regression for a REAIM test after commit e678934cbe5f02 ("btrfs: Remove unnecessary check from join_running_log_trans"). So just bring back the optimization to join_running_log_trans() where we check first if a log root exists before trying to lock the log_mutex. This is done by checking for a bit that is set on the root when a log tree is created and removed when a log tree is freed (at transaction commit time). Commit e678934cbe5f02 ("btrfs: Remove unnecessary check from join_running_log_trans") was merged in the 5.4 merge window while commit 803f0f64d17769 ("Btrfs: fix fsync not persisting dentry deletions due to inode evictions") was merged in the 5.3 merge window. But the first commit was actually authored before the second commit (May 23 2019 vs June 19 2019). Reported-by: kernel test robot <rong.a.chen@intel.com> Link: https://lore.kernel.org/lkml/20200611090233.GL12456@shao2-debian/ Fixes: e678934cbe5f02 ("btrfs: Remove unnecessary check from join_running_log_trans") CC: stable@vger.kernel.org # 5.4+ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-06-15 12:38:44 +03:00
clear_bit(BTRFS_ROOT_HAS_LOG_TREE, &root->state);
}
return 0;
}
int btrfs_free_log_root_tree(struct btrfs_trans_handle *trans,
struct btrfs_fs_info *fs_info)
{
if (fs_info->log_root_tree) {
free_log_tree(trans, fs_info->log_root_tree);
fs_info->log_root_tree = NULL;
btrfs: do not block inode logging for so long during transaction commit Early on during a transaction commit we acquire the tree_log_mutex and hold it until after we write the super blocks. But before writing the extent buffers dirtied by the transaction and the super blocks we unblock the transaction by setting its state to TRANS_STATE_UNBLOCKED and setting fs_info->running_transaction to NULL. This means that after that and before writing the super blocks, new transactions can start. However if any transaction wants to log an inode, it will block waiting for the transaction commit to write its dirty extent buffers and the super blocks because the tree_log_mutex is only released after those operations are complete, and starting a new log transaction blocks on that mutex (at start_log_trans()). Writing the dirty extent buffers and the super blocks can take a very significant amount of time to complete, but we could allow the tasks wanting to log an inode to proceed with most of their steps: 1) create the log trees 2) log metadata in the trees 3) write their dirty extent buffers They only need to wait for the previous transaction commit to complete (write its super blocks) before they attempt to write their super blocks, otherwise we could end up with a corrupt filesystem after a crash. So change start_log_trans() to use the root tree's log_mutex to serialize for the creation of the log root tree instead of using the tree_log_mutex, and make btrfs_sync_log() acquire the tree_log_mutex before writing the super blocks. This allows for inode logging to wait much less time when there is a previous transaction that is still committing, often not having to wait at all, as by the time when we try to sync the log the previous transaction already wrote its super blocks. This patch belongs to a patch set that is comprised of the following patches: btrfs: fix race causing unnecessary inode logging during link and rename btrfs: fix race that results in logging old extents during a fast fsync btrfs: fix race that causes unnecessary logging of ancestor inodes btrfs: fix race that makes inode logging fallback to transaction commit btrfs: fix race leading to unnecessary transaction commit when logging inode btrfs: do not block inode logging for so long during transaction commit The following script that uses dbench was used to measure the impact of the whole patchset: $ cat test-dbench.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/btrfs MOUNT_OPTIONS="-o ssd" echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f -m single -d single $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -D $MNT -t 300 64 umount $MNT The test was run on a machine with 12 cores, 64G of ram, using a NVMe device and a non-debug kernel configuration (Debian's default). Before patch set: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 11277211 0.250 85.340 Close 8283172 0.002 6.479 Rename 477515 1.935 86.026 Unlink 2277936 0.770 87.071 Deltree 256 15.732 81.379 Mkdir 128 0.003 0.009 Qpathinfo 10221180 0.056 44.404 Qfileinfo 1789967 0.002 4.066 Qfsinfo 1874399 0.003 9.176 Sfileinfo 918589 0.061 10.247 Find 3951758 0.341 54.040 WriteX 5616547 0.047 85.079 ReadX 17676028 0.005 9.704 LockX 36704 0.003 1.800 UnlockX 36704 0.002 0.687 Flush 790541 14.115 676.236 Throughput 1179.19 MB/sec 64 clients 64 procs max_latency=676.240 ms After patch set: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 12687926 0.171 86.526 Close 9320780 0.002 8.063 Rename 537253 1.444 78.576 Unlink 2561827 0.559 87.228 Deltree 374 11.499 73.549 Mkdir 187 0.003 0.005 Qpathinfo 11500300 0.061 36.801 Qfileinfo 2017118 0.002 7.189 Qfsinfo 2108641 0.003 4.825 Sfileinfo 1033574 0.008 8.065 Find 4446553 0.408 47.835 WriteX 6335667 0.045 84.388 ReadX 19887312 0.003 9.215 LockX 41312 0.003 1.394 UnlockX 41312 0.002 1.425 Flush 889233 13.014 623.259 Throughput 1339.32 MB/sec 64 clients 64 procs max_latency=623.265 ms +12.7% throughput, -8.2% max latency Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-11-25 15:19:28 +03:00
clear_bit(BTRFS_ROOT_HAS_LOG_TREE, &fs_info->tree_root->state);
}
return 0;
}
Btrfs: fix fsync not persisting dentry deletions due to inode evictions In order to avoid searches on a log tree when unlinking an inode, we check if the inode being unlinked was logged in the current transaction, as well as the inode of its parent directory. When any of the inodes are logged, we proceed to delete directory items and inode reference items from the log, to ensure that if a subsequent fsync of only the inode being unlinked or only of the parent directory when the other is not fsync'ed as well, does not result in the entry still existing after a power failure. That check however is not reliable when one of the inodes involved (the one being unlinked or its parent directory's inode) is evicted, since the logged_trans field is transient, that is, it is not stored on disk, so it is lost when the inode is evicted and loaded into memory again (which is set to zero on load). As a consequence the checks currently being done by btrfs_del_dir_entries_in_log() and btrfs_del_inode_ref_in_log() always return true if the inode was evicted before, regardless of the inode having been logged or not before (and in the current transaction), this results in the dentry being unlinked still existing after a log replay if after the unlink operation only one of the inodes involved is fsync'ed. Example: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/dir $ touch /mnt/dir/foo $ xfs_io -c fsync /mnt/dir/foo # Keep an open file descriptor on our directory while we evict inodes. # We just want to evict the file's inode, the directory's inode must not # be evicted. $ ( cd /mnt/dir; while true; do :; done ) & $ pid=$! # Wait a bit to give time to background process to chdir to our test # directory. $ sleep 0.5 # Trigger eviction of the file's inode. $ echo 2 > /proc/sys/vm/drop_caches # Unlink our file and fsync the parent directory. After a power failure # we don't expect to see the file anymore, since we fsync'ed the parent # directory. $ rm -f $SCRATCH_MNT/dir/foo $ xfs_io -c fsync /mnt/dir <power failure> $ mount /dev/sdb /mnt $ ls /mnt/dir foo $ --> file still there, unlink not persisted despite explicit fsync on dir Fix this by checking if the inode has the full_sync bit set in its runtime flags as well, since that bit is set everytime an inode is loaded from disk, or for other less common cases such as after a shrinking truncate or failure to allocate extent maps for holes, and gets cleared after the first fsync. Also consider the inode as possibly logged only if it was last modified in the current transaction (besides having the full_fsync flag set). Fixes: 3a5f1d458ad161 ("Btrfs: Optimize btree walking while logging inodes") CC: stable@vger.kernel.org # 4.4+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-06-19 15:05:39 +03:00
/*
btrfs: avoid inode logging during rename and link when possible During a rename or link operation, we need to determine if an inode was previously logged or not, and if it was, do some update to the logged inode. We used to rely exclusively on the logged_trans field of struct btrfs_inode to determine that, but that was not reliable because the value of that field is not persisted in the inode item, so it's lost when an inode is evicted and loaded back again. That led to several issues in the past, such as not persisting deletions (such as the case fixed by commit 803f0f64d17769 ("Btrfs: fix fsync not persisting dentry deletions due to inode evictions")), or resulting in losing a file after an inode eviction followed by a rename (commit ecc64fab7d49c6 ("btrfs: fix lost inode on log replay after mix of fsync, rename and inode eviction")), besides other issues. So the inode_logged() helper was introduced and used to determine if an inode was possibly logged before in the current transaction, with the caveat that it could return false positives, in the sense that even if an inode was not logged before in the current transaction, it could still return true, but never to return false in case the inode was logged. >From a functional point of view that is fine, but from a performance perspective it can introduce significant latencies to rename and link operations, as they will end up doing inode logging even when it is not necessary. Recently on a 5.15 kernel, an openSUSE Tumbleweed user reported package installations and upgrades, with the zypper tool, were often taking a long time to complete. With strace it could be observed that zypper was spending about 99% of its time on rename operations, and then with further analysis we checked that directory logging was happening too frequently. Taking into account that installation/upgrade of some of the packages needed a few thousand file renames, the slowdown was very noticeable for the user. The issue was caused indirectly due to an excessive number of inode evictions on a 5.15 kernel, about 100x more compared to a 5.13, 5.14 or a 5.16-rc8 kernel. While triggering the inode evictions if something outside btrfs' control, btrfs could still behave better by eliminating the false positives from the inode_logged() helper. So change inode_logged() to actually eliminate such false positives caused by inode eviction and when an inode was never logged since the filesystem was mounted, as both cases relate to when the logged_trans field of struct btrfs_inode has a value of zero. When it can not determine if the inode was logged based only on the logged_trans value, lookup for the existence of the inode item in the log tree - if it's there then we known the inode was logged, if it's not there then it can not have been logged in the current transaction. Once we determine if the inode was logged, update the logged_trans value to avoid future calls to have to search in the log tree again. Alternatively, we could start storing logged_trans in the on disk inode item structure (struct btrfs_inode_item) in the unused space it still has, but that would be a bit odd because: 1) We only care about logged_trans since the filesystem was mounted, we don't care about its value from a previous mount. Having it persisted in the inode item structure would not make the best use of the precious unused space; 2) In order to get logged_trans persisted before inode eviction, we would have to update the delayed inode when we finish logging the inode and update its logged_trans in struct btrfs_inode, which makes it a bit cumbersome since we need to check if the delayed inode exists, if not create it and populate it and deal with any errors (-ENOMEM mostly). This change is part of a patchset comprised of the following patches: 1/5 btrfs: add helper to delete a dir entry from a log tree 2/5 btrfs: pass the dentry to btrfs_log_new_name() instead of the inode 3/5 btrfs: avoid logging all directory changes during renames 4/5 btrfs: stop doing unnecessary log updates during a rename 5/5 btrfs: avoid inode logging during rename and link when possible The following test script mimics part of what the zypper tool does during package installations/upgrades. It does not triggers inode evictions, but it's similar because it triggers false positives from the inode_logged() helper, because the inodes have a logged_trans of 0, there's a log tree due to a fsync of an unrelated file and the directory inode has its last_trans field set to the current transaction: $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=10000 mkfs.btrfs -f $DEV mount $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync # Now do some change to an unrelated file and fsync it. # This is just to create a log tree to make sure that inode_logged() # does not return false when called against "testdir". xfs_io -f -c "pwrite 0 4K" -c "fsync" $MNT/foo # Do some change to testdir. This is to make sure inode_logged() # will return true when called against "testdir", because its # logged_trans is 0, it was changed in the current transaction # and there's a log tree. echo -n > $MNT/testdir/file_$((NUM_FILES + 1)) echo "Renaming $NUM_FILES files..." start=$(date +%s%N) for ((i = 1; i <= $NUM_FILES; i++)); do mv $MNT/testdir/file_$i $MNT/testdir/file_$i-RPMDELETE done end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "Renames took $dur milliseconds" umount $MNT Testing this change on a box using a non-debug kernel (Debian's default kernel config) gave the following results: NUM_FILES=10000, before patchset: 27837 ms NUM_FILES=10000, after patches 1/5 to 4/5 applied: 9236 ms (-66.8%) NUM_FILES=10000, after whole patchset applied: 8902 ms (-68.0%) NUM_FILES=5000, before patchset: 9127 ms NUM_FILES=5000, after patches 1/5 to 4/5 applied: 4640 ms (-49.2%) NUM_FILES=5000, after whole patchset applied: 4441 ms (-51.3%) NUM_FILES=2000, before patchset: 2528 ms NUM_FILES=2000, after patches 1/5 to 4/5 applied: 1983 ms (-21.6%) NUM_FILES=2000, after whole patchset applied: 1747 ms (-30.9%) NUM_FILES=1000, before patchset: 1085 ms NUM_FILES=1000, after patches 1/5 to 4/5 applied: 893 ms (-17.7%) NUM_FILES=1000, after whole patchset applied: 867 ms (-20.1%) Running dbench on the same physical machine with the following script: $ cat run-dbench.sh #!/bin/bash NUM_JOBS=$(nproc --all) DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 MOUNT_OPTIONS="-o ssd" MKFS_OPTIONS="-O no-holes -R free-space-tree" echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -D $MNT -t 120 $NUM_JOBS umount $MNT Before patchset: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 3761352 0.032 143.843 Close 2762770 0.002 2.273 Rename 159304 0.291 67.037 Unlink 759784 0.207 143.998 Deltree 72 4.028 15.977 Mkdir 36 0.003 0.006 Qpathinfo 3409780 0.013 9.678 Qfileinfo 596772 0.001 0.878 Qfsinfo 625189 0.003 1.245 Sfileinfo 306443 0.006 1.840 Find 1318106 0.063 19.798 WriteX 1871137 0.021 8.532 ReadX 5897325 0.003 3.567 LockX 12252 0.003 0.258 UnlockX 12252 0.002 0.100 Flush 263666 3.327 155.632 Throughput 980.047 MB/sec 12 clients 12 procs max_latency=155.636 ms After whole patchset applied: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 4195584 0.033 107.742 Close 3081932 0.002 1.935 Rename 177641 0.218 14.905 Unlink 847333 0.166 107.822 Deltree 118 5.315 15.247 Mkdir 59 0.004 0.048 Qpathinfo 3802612 0.014 10.302 Qfileinfo 666748 0.001 1.034 Qfsinfo 697329 0.003 0.944 Sfileinfo 341712 0.006 2.099 Find 1470365 0.065 9.359 WriteX 2093921 0.021 8.087 ReadX 6576234 0.003 3.407 LockX 13660 0.003 0.308 UnlockX 13660 0.002 0.114 Flush 294090 2.906 115.539 Throughput 1093.11 MB/sec 12 clients 12 procs max_latency=115.544 ms +11.5% throughput -25.8% max latency rename max latency -77.8% Link: https://bugzilla.opensuse.org/show_bug.cgi?id=1193549 Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-01-20 14:00:10 +03:00
* Check if an inode was logged in the current transaction. This correctly deals
* with the case where the inode was logged but has a logged_trans of 0, which
* happens if the inode is evicted and loaded again, as logged_trans is an in
* memory only field (not persisted).
*
* Returns 1 if the inode was logged before in the transaction, 0 if it was not,
* and < 0 on error.
Btrfs: fix fsync not persisting dentry deletions due to inode evictions In order to avoid searches on a log tree when unlinking an inode, we check if the inode being unlinked was logged in the current transaction, as well as the inode of its parent directory. When any of the inodes are logged, we proceed to delete directory items and inode reference items from the log, to ensure that if a subsequent fsync of only the inode being unlinked or only of the parent directory when the other is not fsync'ed as well, does not result in the entry still existing after a power failure. That check however is not reliable when one of the inodes involved (the one being unlinked or its parent directory's inode) is evicted, since the logged_trans field is transient, that is, it is not stored on disk, so it is lost when the inode is evicted and loaded into memory again (which is set to zero on load). As a consequence the checks currently being done by btrfs_del_dir_entries_in_log() and btrfs_del_inode_ref_in_log() always return true if the inode was evicted before, regardless of the inode having been logged or not before (and in the current transaction), this results in the dentry being unlinked still existing after a log replay if after the unlink operation only one of the inodes involved is fsync'ed. Example: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/dir $ touch /mnt/dir/foo $ xfs_io -c fsync /mnt/dir/foo # Keep an open file descriptor on our directory while we evict inodes. # We just want to evict the file's inode, the directory's inode must not # be evicted. $ ( cd /mnt/dir; while true; do :; done ) & $ pid=$! # Wait a bit to give time to background process to chdir to our test # directory. $ sleep 0.5 # Trigger eviction of the file's inode. $ echo 2 > /proc/sys/vm/drop_caches # Unlink our file and fsync the parent directory. After a power failure # we don't expect to see the file anymore, since we fsync'ed the parent # directory. $ rm -f $SCRATCH_MNT/dir/foo $ xfs_io -c fsync /mnt/dir <power failure> $ mount /dev/sdb /mnt $ ls /mnt/dir foo $ --> file still there, unlink not persisted despite explicit fsync on dir Fix this by checking if the inode has the full_sync bit set in its runtime flags as well, since that bit is set everytime an inode is loaded from disk, or for other less common cases such as after a shrinking truncate or failure to allocate extent maps for holes, and gets cleared after the first fsync. Also consider the inode as possibly logged only if it was last modified in the current transaction (besides having the full_fsync flag set). Fixes: 3a5f1d458ad161 ("Btrfs: Optimize btree walking while logging inodes") CC: stable@vger.kernel.org # 4.4+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-06-19 15:05:39 +03:00
*/
btrfs: avoid inode logging during rename and link when possible During a rename or link operation, we need to determine if an inode was previously logged or not, and if it was, do some update to the logged inode. We used to rely exclusively on the logged_trans field of struct btrfs_inode to determine that, but that was not reliable because the value of that field is not persisted in the inode item, so it's lost when an inode is evicted and loaded back again. That led to several issues in the past, such as not persisting deletions (such as the case fixed by commit 803f0f64d17769 ("Btrfs: fix fsync not persisting dentry deletions due to inode evictions")), or resulting in losing a file after an inode eviction followed by a rename (commit ecc64fab7d49c6 ("btrfs: fix lost inode on log replay after mix of fsync, rename and inode eviction")), besides other issues. So the inode_logged() helper was introduced and used to determine if an inode was possibly logged before in the current transaction, with the caveat that it could return false positives, in the sense that even if an inode was not logged before in the current transaction, it could still return true, but never to return false in case the inode was logged. >From a functional point of view that is fine, but from a performance perspective it can introduce significant latencies to rename and link operations, as they will end up doing inode logging even when it is not necessary. Recently on a 5.15 kernel, an openSUSE Tumbleweed user reported package installations and upgrades, with the zypper tool, were often taking a long time to complete. With strace it could be observed that zypper was spending about 99% of its time on rename operations, and then with further analysis we checked that directory logging was happening too frequently. Taking into account that installation/upgrade of some of the packages needed a few thousand file renames, the slowdown was very noticeable for the user. The issue was caused indirectly due to an excessive number of inode evictions on a 5.15 kernel, about 100x more compared to a 5.13, 5.14 or a 5.16-rc8 kernel. While triggering the inode evictions if something outside btrfs' control, btrfs could still behave better by eliminating the false positives from the inode_logged() helper. So change inode_logged() to actually eliminate such false positives caused by inode eviction and when an inode was never logged since the filesystem was mounted, as both cases relate to when the logged_trans field of struct btrfs_inode has a value of zero. When it can not determine if the inode was logged based only on the logged_trans value, lookup for the existence of the inode item in the log tree - if it's there then we known the inode was logged, if it's not there then it can not have been logged in the current transaction. Once we determine if the inode was logged, update the logged_trans value to avoid future calls to have to search in the log tree again. Alternatively, we could start storing logged_trans in the on disk inode item structure (struct btrfs_inode_item) in the unused space it still has, but that would be a bit odd because: 1) We only care about logged_trans since the filesystem was mounted, we don't care about its value from a previous mount. Having it persisted in the inode item structure would not make the best use of the precious unused space; 2) In order to get logged_trans persisted before inode eviction, we would have to update the delayed inode when we finish logging the inode and update its logged_trans in struct btrfs_inode, which makes it a bit cumbersome since we need to check if the delayed inode exists, if not create it and populate it and deal with any errors (-ENOMEM mostly). This change is part of a patchset comprised of the following patches: 1/5 btrfs: add helper to delete a dir entry from a log tree 2/5 btrfs: pass the dentry to btrfs_log_new_name() instead of the inode 3/5 btrfs: avoid logging all directory changes during renames 4/5 btrfs: stop doing unnecessary log updates during a rename 5/5 btrfs: avoid inode logging during rename and link when possible The following test script mimics part of what the zypper tool does during package installations/upgrades. It does not triggers inode evictions, but it's similar because it triggers false positives from the inode_logged() helper, because the inodes have a logged_trans of 0, there's a log tree due to a fsync of an unrelated file and the directory inode has its last_trans field set to the current transaction: $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=10000 mkfs.btrfs -f $DEV mount $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync # Now do some change to an unrelated file and fsync it. # This is just to create a log tree to make sure that inode_logged() # does not return false when called against "testdir". xfs_io -f -c "pwrite 0 4K" -c "fsync" $MNT/foo # Do some change to testdir. This is to make sure inode_logged() # will return true when called against "testdir", because its # logged_trans is 0, it was changed in the current transaction # and there's a log tree. echo -n > $MNT/testdir/file_$((NUM_FILES + 1)) echo "Renaming $NUM_FILES files..." start=$(date +%s%N) for ((i = 1; i <= $NUM_FILES; i++)); do mv $MNT/testdir/file_$i $MNT/testdir/file_$i-RPMDELETE done end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "Renames took $dur milliseconds" umount $MNT Testing this change on a box using a non-debug kernel (Debian's default kernel config) gave the following results: NUM_FILES=10000, before patchset: 27837 ms NUM_FILES=10000, after patches 1/5 to 4/5 applied: 9236 ms (-66.8%) NUM_FILES=10000, after whole patchset applied: 8902 ms (-68.0%) NUM_FILES=5000, before patchset: 9127 ms NUM_FILES=5000, after patches 1/5 to 4/5 applied: 4640 ms (-49.2%) NUM_FILES=5000, after whole patchset applied: 4441 ms (-51.3%) NUM_FILES=2000, before patchset: 2528 ms NUM_FILES=2000, after patches 1/5 to 4/5 applied: 1983 ms (-21.6%) NUM_FILES=2000, after whole patchset applied: 1747 ms (-30.9%) NUM_FILES=1000, before patchset: 1085 ms NUM_FILES=1000, after patches 1/5 to 4/5 applied: 893 ms (-17.7%) NUM_FILES=1000, after whole patchset applied: 867 ms (-20.1%) Running dbench on the same physical machine with the following script: $ cat run-dbench.sh #!/bin/bash NUM_JOBS=$(nproc --all) DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 MOUNT_OPTIONS="-o ssd" MKFS_OPTIONS="-O no-holes -R free-space-tree" echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -D $MNT -t 120 $NUM_JOBS umount $MNT Before patchset: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 3761352 0.032 143.843 Close 2762770 0.002 2.273 Rename 159304 0.291 67.037 Unlink 759784 0.207 143.998 Deltree 72 4.028 15.977 Mkdir 36 0.003 0.006 Qpathinfo 3409780 0.013 9.678 Qfileinfo 596772 0.001 0.878 Qfsinfo 625189 0.003 1.245 Sfileinfo 306443 0.006 1.840 Find 1318106 0.063 19.798 WriteX 1871137 0.021 8.532 ReadX 5897325 0.003 3.567 LockX 12252 0.003 0.258 UnlockX 12252 0.002 0.100 Flush 263666 3.327 155.632 Throughput 980.047 MB/sec 12 clients 12 procs max_latency=155.636 ms After whole patchset applied: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 4195584 0.033 107.742 Close 3081932 0.002 1.935 Rename 177641 0.218 14.905 Unlink 847333 0.166 107.822 Deltree 118 5.315 15.247 Mkdir 59 0.004 0.048 Qpathinfo 3802612 0.014 10.302 Qfileinfo 666748 0.001 1.034 Qfsinfo 697329 0.003 0.944 Sfileinfo 341712 0.006 2.099 Find 1470365 0.065 9.359 WriteX 2093921 0.021 8.087 ReadX 6576234 0.003 3.407 LockX 13660 0.003 0.308 UnlockX 13660 0.002 0.114 Flush 294090 2.906 115.539 Throughput 1093.11 MB/sec 12 clients 12 procs max_latency=115.544 ms +11.5% throughput -25.8% max latency rename max latency -77.8% Link: https://bugzilla.opensuse.org/show_bug.cgi?id=1193549 Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-01-20 14:00:10 +03:00
static int inode_logged(struct btrfs_trans_handle *trans,
struct btrfs_inode *inode,
struct btrfs_path *path_in)
Btrfs: fix fsync not persisting dentry deletions due to inode evictions In order to avoid searches on a log tree when unlinking an inode, we check if the inode being unlinked was logged in the current transaction, as well as the inode of its parent directory. When any of the inodes are logged, we proceed to delete directory items and inode reference items from the log, to ensure that if a subsequent fsync of only the inode being unlinked or only of the parent directory when the other is not fsync'ed as well, does not result in the entry still existing after a power failure. That check however is not reliable when one of the inodes involved (the one being unlinked or its parent directory's inode) is evicted, since the logged_trans field is transient, that is, it is not stored on disk, so it is lost when the inode is evicted and loaded into memory again (which is set to zero on load). As a consequence the checks currently being done by btrfs_del_dir_entries_in_log() and btrfs_del_inode_ref_in_log() always return true if the inode was evicted before, regardless of the inode having been logged or not before (and in the current transaction), this results in the dentry being unlinked still existing after a log replay if after the unlink operation only one of the inodes involved is fsync'ed. Example: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/dir $ touch /mnt/dir/foo $ xfs_io -c fsync /mnt/dir/foo # Keep an open file descriptor on our directory while we evict inodes. # We just want to evict the file's inode, the directory's inode must not # be evicted. $ ( cd /mnt/dir; while true; do :; done ) & $ pid=$! # Wait a bit to give time to background process to chdir to our test # directory. $ sleep 0.5 # Trigger eviction of the file's inode. $ echo 2 > /proc/sys/vm/drop_caches # Unlink our file and fsync the parent directory. After a power failure # we don't expect to see the file anymore, since we fsync'ed the parent # directory. $ rm -f $SCRATCH_MNT/dir/foo $ xfs_io -c fsync /mnt/dir <power failure> $ mount /dev/sdb /mnt $ ls /mnt/dir foo $ --> file still there, unlink not persisted despite explicit fsync on dir Fix this by checking if the inode has the full_sync bit set in its runtime flags as well, since that bit is set everytime an inode is loaded from disk, or for other less common cases such as after a shrinking truncate or failure to allocate extent maps for holes, and gets cleared after the first fsync. Also consider the inode as possibly logged only if it was last modified in the current transaction (besides having the full_fsync flag set). Fixes: 3a5f1d458ad161 ("Btrfs: Optimize btree walking while logging inodes") CC: stable@vger.kernel.org # 4.4+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-06-19 15:05:39 +03:00
{
btrfs: avoid inode logging during rename and link when possible During a rename or link operation, we need to determine if an inode was previously logged or not, and if it was, do some update to the logged inode. We used to rely exclusively on the logged_trans field of struct btrfs_inode to determine that, but that was not reliable because the value of that field is not persisted in the inode item, so it's lost when an inode is evicted and loaded back again. That led to several issues in the past, such as not persisting deletions (such as the case fixed by commit 803f0f64d17769 ("Btrfs: fix fsync not persisting dentry deletions due to inode evictions")), or resulting in losing a file after an inode eviction followed by a rename (commit ecc64fab7d49c6 ("btrfs: fix lost inode on log replay after mix of fsync, rename and inode eviction")), besides other issues. So the inode_logged() helper was introduced and used to determine if an inode was possibly logged before in the current transaction, with the caveat that it could return false positives, in the sense that even if an inode was not logged before in the current transaction, it could still return true, but never to return false in case the inode was logged. >From a functional point of view that is fine, but from a performance perspective it can introduce significant latencies to rename and link operations, as they will end up doing inode logging even when it is not necessary. Recently on a 5.15 kernel, an openSUSE Tumbleweed user reported package installations and upgrades, with the zypper tool, were often taking a long time to complete. With strace it could be observed that zypper was spending about 99% of its time on rename operations, and then with further analysis we checked that directory logging was happening too frequently. Taking into account that installation/upgrade of some of the packages needed a few thousand file renames, the slowdown was very noticeable for the user. The issue was caused indirectly due to an excessive number of inode evictions on a 5.15 kernel, about 100x more compared to a 5.13, 5.14 or a 5.16-rc8 kernel. While triggering the inode evictions if something outside btrfs' control, btrfs could still behave better by eliminating the false positives from the inode_logged() helper. So change inode_logged() to actually eliminate such false positives caused by inode eviction and when an inode was never logged since the filesystem was mounted, as both cases relate to when the logged_trans field of struct btrfs_inode has a value of zero. When it can not determine if the inode was logged based only on the logged_trans value, lookup for the existence of the inode item in the log tree - if it's there then we known the inode was logged, if it's not there then it can not have been logged in the current transaction. Once we determine if the inode was logged, update the logged_trans value to avoid future calls to have to search in the log tree again. Alternatively, we could start storing logged_trans in the on disk inode item structure (struct btrfs_inode_item) in the unused space it still has, but that would be a bit odd because: 1) We only care about logged_trans since the filesystem was mounted, we don't care about its value from a previous mount. Having it persisted in the inode item structure would not make the best use of the precious unused space; 2) In order to get logged_trans persisted before inode eviction, we would have to update the delayed inode when we finish logging the inode and update its logged_trans in struct btrfs_inode, which makes it a bit cumbersome since we need to check if the delayed inode exists, if not create it and populate it and deal with any errors (-ENOMEM mostly). This change is part of a patchset comprised of the following patches: 1/5 btrfs: add helper to delete a dir entry from a log tree 2/5 btrfs: pass the dentry to btrfs_log_new_name() instead of the inode 3/5 btrfs: avoid logging all directory changes during renames 4/5 btrfs: stop doing unnecessary log updates during a rename 5/5 btrfs: avoid inode logging during rename and link when possible The following test script mimics part of what the zypper tool does during package installations/upgrades. It does not triggers inode evictions, but it's similar because it triggers false positives from the inode_logged() helper, because the inodes have a logged_trans of 0, there's a log tree due to a fsync of an unrelated file and the directory inode has its last_trans field set to the current transaction: $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=10000 mkfs.btrfs -f $DEV mount $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync # Now do some change to an unrelated file and fsync it. # This is just to create a log tree to make sure that inode_logged() # does not return false when called against "testdir". xfs_io -f -c "pwrite 0 4K" -c "fsync" $MNT/foo # Do some change to testdir. This is to make sure inode_logged() # will return true when called against "testdir", because its # logged_trans is 0, it was changed in the current transaction # and there's a log tree. echo -n > $MNT/testdir/file_$((NUM_FILES + 1)) echo "Renaming $NUM_FILES files..." start=$(date +%s%N) for ((i = 1; i <= $NUM_FILES; i++)); do mv $MNT/testdir/file_$i $MNT/testdir/file_$i-RPMDELETE done end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "Renames took $dur milliseconds" umount $MNT Testing this change on a box using a non-debug kernel (Debian's default kernel config) gave the following results: NUM_FILES=10000, before patchset: 27837 ms NUM_FILES=10000, after patches 1/5 to 4/5 applied: 9236 ms (-66.8%) NUM_FILES=10000, after whole patchset applied: 8902 ms (-68.0%) NUM_FILES=5000, before patchset: 9127 ms NUM_FILES=5000, after patches 1/5 to 4/5 applied: 4640 ms (-49.2%) NUM_FILES=5000, after whole patchset applied: 4441 ms (-51.3%) NUM_FILES=2000, before patchset: 2528 ms NUM_FILES=2000, after patches 1/5 to 4/5 applied: 1983 ms (-21.6%) NUM_FILES=2000, after whole patchset applied: 1747 ms (-30.9%) NUM_FILES=1000, before patchset: 1085 ms NUM_FILES=1000, after patches 1/5 to 4/5 applied: 893 ms (-17.7%) NUM_FILES=1000, after whole patchset applied: 867 ms (-20.1%) Running dbench on the same physical machine with the following script: $ cat run-dbench.sh #!/bin/bash NUM_JOBS=$(nproc --all) DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 MOUNT_OPTIONS="-o ssd" MKFS_OPTIONS="-O no-holes -R free-space-tree" echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -D $MNT -t 120 $NUM_JOBS umount $MNT Before patchset: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 3761352 0.032 143.843 Close 2762770 0.002 2.273 Rename 159304 0.291 67.037 Unlink 759784 0.207 143.998 Deltree 72 4.028 15.977 Mkdir 36 0.003 0.006 Qpathinfo 3409780 0.013 9.678 Qfileinfo 596772 0.001 0.878 Qfsinfo 625189 0.003 1.245 Sfileinfo 306443 0.006 1.840 Find 1318106 0.063 19.798 WriteX 1871137 0.021 8.532 ReadX 5897325 0.003 3.567 LockX 12252 0.003 0.258 UnlockX 12252 0.002 0.100 Flush 263666 3.327 155.632 Throughput 980.047 MB/sec 12 clients 12 procs max_latency=155.636 ms After whole patchset applied: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 4195584 0.033 107.742 Close 3081932 0.002 1.935 Rename 177641 0.218 14.905 Unlink 847333 0.166 107.822 Deltree 118 5.315 15.247 Mkdir 59 0.004 0.048 Qpathinfo 3802612 0.014 10.302 Qfileinfo 666748 0.001 1.034 Qfsinfo 697329 0.003 0.944 Sfileinfo 341712 0.006 2.099 Find 1470365 0.065 9.359 WriteX 2093921 0.021 8.087 ReadX 6576234 0.003 3.407 LockX 13660 0.003 0.308 UnlockX 13660 0.002 0.114 Flush 294090 2.906 115.539 Throughput 1093.11 MB/sec 12 clients 12 procs max_latency=115.544 ms +11.5% throughput -25.8% max latency rename max latency -77.8% Link: https://bugzilla.opensuse.org/show_bug.cgi?id=1193549 Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-01-20 14:00:10 +03:00
struct btrfs_path *path = path_in;
struct btrfs_key key;
int ret;
Btrfs: fix fsync not persisting dentry deletions due to inode evictions In order to avoid searches on a log tree when unlinking an inode, we check if the inode being unlinked was logged in the current transaction, as well as the inode of its parent directory. When any of the inodes are logged, we proceed to delete directory items and inode reference items from the log, to ensure that if a subsequent fsync of only the inode being unlinked or only of the parent directory when the other is not fsync'ed as well, does not result in the entry still existing after a power failure. That check however is not reliable when one of the inodes involved (the one being unlinked or its parent directory's inode) is evicted, since the logged_trans field is transient, that is, it is not stored on disk, so it is lost when the inode is evicted and loaded into memory again (which is set to zero on load). As a consequence the checks currently being done by btrfs_del_dir_entries_in_log() and btrfs_del_inode_ref_in_log() always return true if the inode was evicted before, regardless of the inode having been logged or not before (and in the current transaction), this results in the dentry being unlinked still existing after a log replay if after the unlink operation only one of the inodes involved is fsync'ed. Example: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/dir $ touch /mnt/dir/foo $ xfs_io -c fsync /mnt/dir/foo # Keep an open file descriptor on our directory while we evict inodes. # We just want to evict the file's inode, the directory's inode must not # be evicted. $ ( cd /mnt/dir; while true; do :; done ) & $ pid=$! # Wait a bit to give time to background process to chdir to our test # directory. $ sleep 0.5 # Trigger eviction of the file's inode. $ echo 2 > /proc/sys/vm/drop_caches # Unlink our file and fsync the parent directory. After a power failure # we don't expect to see the file anymore, since we fsync'ed the parent # directory. $ rm -f $SCRATCH_MNT/dir/foo $ xfs_io -c fsync /mnt/dir <power failure> $ mount /dev/sdb /mnt $ ls /mnt/dir foo $ --> file still there, unlink not persisted despite explicit fsync on dir Fix this by checking if the inode has the full_sync bit set in its runtime flags as well, since that bit is set everytime an inode is loaded from disk, or for other less common cases such as after a shrinking truncate or failure to allocate extent maps for holes, and gets cleared after the first fsync. Also consider the inode as possibly logged only if it was last modified in the current transaction (besides having the full_fsync flag set). Fixes: 3a5f1d458ad161 ("Btrfs: Optimize btree walking while logging inodes") CC: stable@vger.kernel.org # 4.4+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-06-19 15:05:39 +03:00
if (inode->logged_trans == trans->transid)
btrfs: avoid inode logging during rename and link when possible During a rename or link operation, we need to determine if an inode was previously logged or not, and if it was, do some update to the logged inode. We used to rely exclusively on the logged_trans field of struct btrfs_inode to determine that, but that was not reliable because the value of that field is not persisted in the inode item, so it's lost when an inode is evicted and loaded back again. That led to several issues in the past, such as not persisting deletions (such as the case fixed by commit 803f0f64d17769 ("Btrfs: fix fsync not persisting dentry deletions due to inode evictions")), or resulting in losing a file after an inode eviction followed by a rename (commit ecc64fab7d49c6 ("btrfs: fix lost inode on log replay after mix of fsync, rename and inode eviction")), besides other issues. So the inode_logged() helper was introduced and used to determine if an inode was possibly logged before in the current transaction, with the caveat that it could return false positives, in the sense that even if an inode was not logged before in the current transaction, it could still return true, but never to return false in case the inode was logged. >From a functional point of view that is fine, but from a performance perspective it can introduce significant latencies to rename and link operations, as they will end up doing inode logging even when it is not necessary. Recently on a 5.15 kernel, an openSUSE Tumbleweed user reported package installations and upgrades, with the zypper tool, were often taking a long time to complete. With strace it could be observed that zypper was spending about 99% of its time on rename operations, and then with further analysis we checked that directory logging was happening too frequently. Taking into account that installation/upgrade of some of the packages needed a few thousand file renames, the slowdown was very noticeable for the user. The issue was caused indirectly due to an excessive number of inode evictions on a 5.15 kernel, about 100x more compared to a 5.13, 5.14 or a 5.16-rc8 kernel. While triggering the inode evictions if something outside btrfs' control, btrfs could still behave better by eliminating the false positives from the inode_logged() helper. So change inode_logged() to actually eliminate such false positives caused by inode eviction and when an inode was never logged since the filesystem was mounted, as both cases relate to when the logged_trans field of struct btrfs_inode has a value of zero. When it can not determine if the inode was logged based only on the logged_trans value, lookup for the existence of the inode item in the log tree - if it's there then we known the inode was logged, if it's not there then it can not have been logged in the current transaction. Once we determine if the inode was logged, update the logged_trans value to avoid future calls to have to search in the log tree again. Alternatively, we could start storing logged_trans in the on disk inode item structure (struct btrfs_inode_item) in the unused space it still has, but that would be a bit odd because: 1) We only care about logged_trans since the filesystem was mounted, we don't care about its value from a previous mount. Having it persisted in the inode item structure would not make the best use of the precious unused space; 2) In order to get logged_trans persisted before inode eviction, we would have to update the delayed inode when we finish logging the inode and update its logged_trans in struct btrfs_inode, which makes it a bit cumbersome since we need to check if the delayed inode exists, if not create it and populate it and deal with any errors (-ENOMEM mostly). This change is part of a patchset comprised of the following patches: 1/5 btrfs: add helper to delete a dir entry from a log tree 2/5 btrfs: pass the dentry to btrfs_log_new_name() instead of the inode 3/5 btrfs: avoid logging all directory changes during renames 4/5 btrfs: stop doing unnecessary log updates during a rename 5/5 btrfs: avoid inode logging during rename and link when possible The following test script mimics part of what the zypper tool does during package installations/upgrades. It does not triggers inode evictions, but it's similar because it triggers false positives from the inode_logged() helper, because the inodes have a logged_trans of 0, there's a log tree due to a fsync of an unrelated file and the directory inode has its last_trans field set to the current transaction: $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=10000 mkfs.btrfs -f $DEV mount $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync # Now do some change to an unrelated file and fsync it. # This is just to create a log tree to make sure that inode_logged() # does not return false when called against "testdir". xfs_io -f -c "pwrite 0 4K" -c "fsync" $MNT/foo # Do some change to testdir. This is to make sure inode_logged() # will return true when called against "testdir", because its # logged_trans is 0, it was changed in the current transaction # and there's a log tree. echo -n > $MNT/testdir/file_$((NUM_FILES + 1)) echo "Renaming $NUM_FILES files..." start=$(date +%s%N) for ((i = 1; i <= $NUM_FILES; i++)); do mv $MNT/testdir/file_$i $MNT/testdir/file_$i-RPMDELETE done end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "Renames took $dur milliseconds" umount $MNT Testing this change on a box using a non-debug kernel (Debian's default kernel config) gave the following results: NUM_FILES=10000, before patchset: 27837 ms NUM_FILES=10000, after patches 1/5 to 4/5 applied: 9236 ms (-66.8%) NUM_FILES=10000, after whole patchset applied: 8902 ms (-68.0%) NUM_FILES=5000, before patchset: 9127 ms NUM_FILES=5000, after patches 1/5 to 4/5 applied: 4640 ms (-49.2%) NUM_FILES=5000, after whole patchset applied: 4441 ms (-51.3%) NUM_FILES=2000, before patchset: 2528 ms NUM_FILES=2000, after patches 1/5 to 4/5 applied: 1983 ms (-21.6%) NUM_FILES=2000, after whole patchset applied: 1747 ms (-30.9%) NUM_FILES=1000, before patchset: 1085 ms NUM_FILES=1000, after patches 1/5 to 4/5 applied: 893 ms (-17.7%) NUM_FILES=1000, after whole patchset applied: 867 ms (-20.1%) Running dbench on the same physical machine with the following script: $ cat run-dbench.sh #!/bin/bash NUM_JOBS=$(nproc --all) DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 MOUNT_OPTIONS="-o ssd" MKFS_OPTIONS="-O no-holes -R free-space-tree" echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -D $MNT -t 120 $NUM_JOBS umount $MNT Before patchset: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 3761352 0.032 143.843 Close 2762770 0.002 2.273 Rename 159304 0.291 67.037 Unlink 759784 0.207 143.998 Deltree 72 4.028 15.977 Mkdir 36 0.003 0.006 Qpathinfo 3409780 0.013 9.678 Qfileinfo 596772 0.001 0.878 Qfsinfo 625189 0.003 1.245 Sfileinfo 306443 0.006 1.840 Find 1318106 0.063 19.798 WriteX 1871137 0.021 8.532 ReadX 5897325 0.003 3.567 LockX 12252 0.003 0.258 UnlockX 12252 0.002 0.100 Flush 263666 3.327 155.632 Throughput 980.047 MB/sec 12 clients 12 procs max_latency=155.636 ms After whole patchset applied: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 4195584 0.033 107.742 Close 3081932 0.002 1.935 Rename 177641 0.218 14.905 Unlink 847333 0.166 107.822 Deltree 118 5.315 15.247 Mkdir 59 0.004 0.048 Qpathinfo 3802612 0.014 10.302 Qfileinfo 666748 0.001 1.034 Qfsinfo 697329 0.003 0.944 Sfileinfo 341712 0.006 2.099 Find 1470365 0.065 9.359 WriteX 2093921 0.021 8.087 ReadX 6576234 0.003 3.407 LockX 13660 0.003 0.308 UnlockX 13660 0.002 0.114 Flush 294090 2.906 115.539 Throughput 1093.11 MB/sec 12 clients 12 procs max_latency=115.544 ms +11.5% throughput -25.8% max latency rename max latency -77.8% Link: https://bugzilla.opensuse.org/show_bug.cgi?id=1193549 Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-01-20 14:00:10 +03:00
return 1;
Btrfs: fix fsync not persisting dentry deletions due to inode evictions In order to avoid searches on a log tree when unlinking an inode, we check if the inode being unlinked was logged in the current transaction, as well as the inode of its parent directory. When any of the inodes are logged, we proceed to delete directory items and inode reference items from the log, to ensure that if a subsequent fsync of only the inode being unlinked or only of the parent directory when the other is not fsync'ed as well, does not result in the entry still existing after a power failure. That check however is not reliable when one of the inodes involved (the one being unlinked or its parent directory's inode) is evicted, since the logged_trans field is transient, that is, it is not stored on disk, so it is lost when the inode is evicted and loaded into memory again (which is set to zero on load). As a consequence the checks currently being done by btrfs_del_dir_entries_in_log() and btrfs_del_inode_ref_in_log() always return true if the inode was evicted before, regardless of the inode having been logged or not before (and in the current transaction), this results in the dentry being unlinked still existing after a log replay if after the unlink operation only one of the inodes involved is fsync'ed. Example: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/dir $ touch /mnt/dir/foo $ xfs_io -c fsync /mnt/dir/foo # Keep an open file descriptor on our directory while we evict inodes. # We just want to evict the file's inode, the directory's inode must not # be evicted. $ ( cd /mnt/dir; while true; do :; done ) & $ pid=$! # Wait a bit to give time to background process to chdir to our test # directory. $ sleep 0.5 # Trigger eviction of the file's inode. $ echo 2 > /proc/sys/vm/drop_caches # Unlink our file and fsync the parent directory. After a power failure # we don't expect to see the file anymore, since we fsync'ed the parent # directory. $ rm -f $SCRATCH_MNT/dir/foo $ xfs_io -c fsync /mnt/dir <power failure> $ mount /dev/sdb /mnt $ ls /mnt/dir foo $ --> file still there, unlink not persisted despite explicit fsync on dir Fix this by checking if the inode has the full_sync bit set in its runtime flags as well, since that bit is set everytime an inode is loaded from disk, or for other less common cases such as after a shrinking truncate or failure to allocate extent maps for holes, and gets cleared after the first fsync. Also consider the inode as possibly logged only if it was last modified in the current transaction (besides having the full_fsync flag set). Fixes: 3a5f1d458ad161 ("Btrfs: Optimize btree walking while logging inodes") CC: stable@vger.kernel.org # 4.4+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-06-19 15:05:39 +03:00
btrfs: avoid inode logging during rename and link when possible During a rename or link operation, we need to determine if an inode was previously logged or not, and if it was, do some update to the logged inode. We used to rely exclusively on the logged_trans field of struct btrfs_inode to determine that, but that was not reliable because the value of that field is not persisted in the inode item, so it's lost when an inode is evicted and loaded back again. That led to several issues in the past, such as not persisting deletions (such as the case fixed by commit 803f0f64d17769 ("Btrfs: fix fsync not persisting dentry deletions due to inode evictions")), or resulting in losing a file after an inode eviction followed by a rename (commit ecc64fab7d49c6 ("btrfs: fix lost inode on log replay after mix of fsync, rename and inode eviction")), besides other issues. So the inode_logged() helper was introduced and used to determine if an inode was possibly logged before in the current transaction, with the caveat that it could return false positives, in the sense that even if an inode was not logged before in the current transaction, it could still return true, but never to return false in case the inode was logged. >From a functional point of view that is fine, but from a performance perspective it can introduce significant latencies to rename and link operations, as they will end up doing inode logging even when it is not necessary. Recently on a 5.15 kernel, an openSUSE Tumbleweed user reported package installations and upgrades, with the zypper tool, were often taking a long time to complete. With strace it could be observed that zypper was spending about 99% of its time on rename operations, and then with further analysis we checked that directory logging was happening too frequently. Taking into account that installation/upgrade of some of the packages needed a few thousand file renames, the slowdown was very noticeable for the user. The issue was caused indirectly due to an excessive number of inode evictions on a 5.15 kernel, about 100x more compared to a 5.13, 5.14 or a 5.16-rc8 kernel. While triggering the inode evictions if something outside btrfs' control, btrfs could still behave better by eliminating the false positives from the inode_logged() helper. So change inode_logged() to actually eliminate such false positives caused by inode eviction and when an inode was never logged since the filesystem was mounted, as both cases relate to when the logged_trans field of struct btrfs_inode has a value of zero. When it can not determine if the inode was logged based only on the logged_trans value, lookup for the existence of the inode item in the log tree - if it's there then we known the inode was logged, if it's not there then it can not have been logged in the current transaction. Once we determine if the inode was logged, update the logged_trans value to avoid future calls to have to search in the log tree again. Alternatively, we could start storing logged_trans in the on disk inode item structure (struct btrfs_inode_item) in the unused space it still has, but that would be a bit odd because: 1) We only care about logged_trans since the filesystem was mounted, we don't care about its value from a previous mount. Having it persisted in the inode item structure would not make the best use of the precious unused space; 2) In order to get logged_trans persisted before inode eviction, we would have to update the delayed inode when we finish logging the inode and update its logged_trans in struct btrfs_inode, which makes it a bit cumbersome since we need to check if the delayed inode exists, if not create it and populate it and deal with any errors (-ENOMEM mostly). This change is part of a patchset comprised of the following patches: 1/5 btrfs: add helper to delete a dir entry from a log tree 2/5 btrfs: pass the dentry to btrfs_log_new_name() instead of the inode 3/5 btrfs: avoid logging all directory changes during renames 4/5 btrfs: stop doing unnecessary log updates during a rename 5/5 btrfs: avoid inode logging during rename and link when possible The following test script mimics part of what the zypper tool does during package installations/upgrades. It does not triggers inode evictions, but it's similar because it triggers false positives from the inode_logged() helper, because the inodes have a logged_trans of 0, there's a log tree due to a fsync of an unrelated file and the directory inode has its last_trans field set to the current transaction: $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=10000 mkfs.btrfs -f $DEV mount $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync # Now do some change to an unrelated file and fsync it. # This is just to create a log tree to make sure that inode_logged() # does not return false when called against "testdir". xfs_io -f -c "pwrite 0 4K" -c "fsync" $MNT/foo # Do some change to testdir. This is to make sure inode_logged() # will return true when called against "testdir", because its # logged_trans is 0, it was changed in the current transaction # and there's a log tree. echo -n > $MNT/testdir/file_$((NUM_FILES + 1)) echo "Renaming $NUM_FILES files..." start=$(date +%s%N) for ((i = 1; i <= $NUM_FILES; i++)); do mv $MNT/testdir/file_$i $MNT/testdir/file_$i-RPMDELETE done end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "Renames took $dur milliseconds" umount $MNT Testing this change on a box using a non-debug kernel (Debian's default kernel config) gave the following results: NUM_FILES=10000, before patchset: 27837 ms NUM_FILES=10000, after patches 1/5 to 4/5 applied: 9236 ms (-66.8%) NUM_FILES=10000, after whole patchset applied: 8902 ms (-68.0%) NUM_FILES=5000, before patchset: 9127 ms NUM_FILES=5000, after patches 1/5 to 4/5 applied: 4640 ms (-49.2%) NUM_FILES=5000, after whole patchset applied: 4441 ms (-51.3%) NUM_FILES=2000, before patchset: 2528 ms NUM_FILES=2000, after patches 1/5 to 4/5 applied: 1983 ms (-21.6%) NUM_FILES=2000, after whole patchset applied: 1747 ms (-30.9%) NUM_FILES=1000, before patchset: 1085 ms NUM_FILES=1000, after patches 1/5 to 4/5 applied: 893 ms (-17.7%) NUM_FILES=1000, after whole patchset applied: 867 ms (-20.1%) Running dbench on the same physical machine with the following script: $ cat run-dbench.sh #!/bin/bash NUM_JOBS=$(nproc --all) DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 MOUNT_OPTIONS="-o ssd" MKFS_OPTIONS="-O no-holes -R free-space-tree" echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -D $MNT -t 120 $NUM_JOBS umount $MNT Before patchset: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 3761352 0.032 143.843 Close 2762770 0.002 2.273 Rename 159304 0.291 67.037 Unlink 759784 0.207 143.998 Deltree 72 4.028 15.977 Mkdir 36 0.003 0.006 Qpathinfo 3409780 0.013 9.678 Qfileinfo 596772 0.001 0.878 Qfsinfo 625189 0.003 1.245 Sfileinfo 306443 0.006 1.840 Find 1318106 0.063 19.798 WriteX 1871137 0.021 8.532 ReadX 5897325 0.003 3.567 LockX 12252 0.003 0.258 UnlockX 12252 0.002 0.100 Flush 263666 3.327 155.632 Throughput 980.047 MB/sec 12 clients 12 procs max_latency=155.636 ms After whole patchset applied: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 4195584 0.033 107.742 Close 3081932 0.002 1.935 Rename 177641 0.218 14.905 Unlink 847333 0.166 107.822 Deltree 118 5.315 15.247 Mkdir 59 0.004 0.048 Qpathinfo 3802612 0.014 10.302 Qfileinfo 666748 0.001 1.034 Qfsinfo 697329 0.003 0.944 Sfileinfo 341712 0.006 2.099 Find 1470365 0.065 9.359 WriteX 2093921 0.021 8.087 ReadX 6576234 0.003 3.407 LockX 13660 0.003 0.308 UnlockX 13660 0.002 0.114 Flush 294090 2.906 115.539 Throughput 1093.11 MB/sec 12 clients 12 procs max_latency=115.544 ms +11.5% throughput -25.8% max latency rename max latency -77.8% Link: https://bugzilla.opensuse.org/show_bug.cgi?id=1193549 Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-01-20 14:00:10 +03:00
/*
* If logged_trans is not 0, then we know the inode logged was not logged
* in this transaction, so we can return false right away.
*/
if (inode->logged_trans > 0)
return 0;
/*
* If no log tree was created for this root in this transaction, then
* the inode can not have been logged in this transaction. In that case
* set logged_trans to anything greater than 0 and less than the current
* transaction's ID, to avoid the search below in a future call in case
* a log tree gets created after this.
*/
if (!test_bit(BTRFS_ROOT_HAS_LOG_TREE, &inode->root->state)) {
inode->logged_trans = trans->transid - 1;
return 0;
}
/*
* We have a log tree and the inode's logged_trans is 0. We can't tell
* for sure if the inode was logged before in this transaction by looking
* only at logged_trans. We could be pessimistic and assume it was, but
* that can lead to unnecessarily logging an inode during rename and link
* operations, and then further updating the log in followup rename and
* link operations, specially if it's a directory, which adds latency
* visible to applications doing a series of rename or link operations.
*
* A logged_trans of 0 here can mean several things:
*
* 1) The inode was never logged since the filesystem was mounted, and may
* or may have not been evicted and loaded again;
*
* 2) The inode was logged in a previous transaction, then evicted and
* then loaded again;
*
* 3) The inode was logged in the current transaction, then evicted and
* then loaded again.
*
* For cases 1) and 2) we don't want to return true, but we need to detect
* case 3) and return true. So we do a search in the log root for the inode
* item.
*/
key.objectid = btrfs_ino(inode);
key.type = BTRFS_INODE_ITEM_KEY;
key.offset = 0;
if (!path) {
path = btrfs_alloc_path();
if (!path)
return -ENOMEM;
}
ret = btrfs_search_slot(NULL, inode->root->log_root, &key, path, 0, 0);
if (path_in)
btrfs_release_path(path);
else
btrfs_free_path(path);
btrfs: check if a log tree exists at inode_logged() In case an inode was never logged since it was loaded from disk and was modified in the current transaction (its ->last_trans matches the ID of the current transaction), inode_logged() returns true even if there's no existing log tree. In this case we can simply check if a log tree exists and return false if it does not. This avoids a caller of inode_logged() doing some unnecessary, but harmless, work. For btrfs_log_new_name() it avoids it logging an inode in case it was never logged since it was loaded from disk and there is currently no log tree for the inode's root. For the remaining callers of inode_logged(), btrfs_del_dir_entries_in_log() and btrfs_del_inode_ref_in_log(), it has no effect since they already check if a log tree exists through their calls to join_running_log_trans(). So just add a check to inode_logged() to verify if a log tree exists, and return false if it does not. This patch is part of a patch set comprised of the following patches: btrfs: check if a log tree exists at inode_logged() btrfs: remove no longer needed checks for NULL log context btrfs: do not log new dentries when logging that a new name exists btrfs: always update the logged transaction when logging new names btrfs: avoid expensive search when dropping inode items from log btrfs: add helper to truncate inode items when logging inode btrfs: avoid expensive search when truncating inode items from the log btrfs: avoid search for logged i_size when logging inode if possible btrfs: avoid attempt to drop extents when logging inode for the first time btrfs: do not commit delayed inode when logging a file in full sync mode This is patch 1/10 and test results are listed in the change log of the last patch in the set. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-08-31 17:30:31 +03:00
/*
btrfs: avoid inode logging during rename and link when possible During a rename or link operation, we need to determine if an inode was previously logged or not, and if it was, do some update to the logged inode. We used to rely exclusively on the logged_trans field of struct btrfs_inode to determine that, but that was not reliable because the value of that field is not persisted in the inode item, so it's lost when an inode is evicted and loaded back again. That led to several issues in the past, such as not persisting deletions (such as the case fixed by commit 803f0f64d17769 ("Btrfs: fix fsync not persisting dentry deletions due to inode evictions")), or resulting in losing a file after an inode eviction followed by a rename (commit ecc64fab7d49c6 ("btrfs: fix lost inode on log replay after mix of fsync, rename and inode eviction")), besides other issues. So the inode_logged() helper was introduced and used to determine if an inode was possibly logged before in the current transaction, with the caveat that it could return false positives, in the sense that even if an inode was not logged before in the current transaction, it could still return true, but never to return false in case the inode was logged. >From a functional point of view that is fine, but from a performance perspective it can introduce significant latencies to rename and link operations, as they will end up doing inode logging even when it is not necessary. Recently on a 5.15 kernel, an openSUSE Tumbleweed user reported package installations and upgrades, with the zypper tool, were often taking a long time to complete. With strace it could be observed that zypper was spending about 99% of its time on rename operations, and then with further analysis we checked that directory logging was happening too frequently. Taking into account that installation/upgrade of some of the packages needed a few thousand file renames, the slowdown was very noticeable for the user. The issue was caused indirectly due to an excessive number of inode evictions on a 5.15 kernel, about 100x more compared to a 5.13, 5.14 or a 5.16-rc8 kernel. While triggering the inode evictions if something outside btrfs' control, btrfs could still behave better by eliminating the false positives from the inode_logged() helper. So change inode_logged() to actually eliminate such false positives caused by inode eviction and when an inode was never logged since the filesystem was mounted, as both cases relate to when the logged_trans field of struct btrfs_inode has a value of zero. When it can not determine if the inode was logged based only on the logged_trans value, lookup for the existence of the inode item in the log tree - if it's there then we known the inode was logged, if it's not there then it can not have been logged in the current transaction. Once we determine if the inode was logged, update the logged_trans value to avoid future calls to have to search in the log tree again. Alternatively, we could start storing logged_trans in the on disk inode item structure (struct btrfs_inode_item) in the unused space it still has, but that would be a bit odd because: 1) We only care about logged_trans since the filesystem was mounted, we don't care about its value from a previous mount. Having it persisted in the inode item structure would not make the best use of the precious unused space; 2) In order to get logged_trans persisted before inode eviction, we would have to update the delayed inode when we finish logging the inode and update its logged_trans in struct btrfs_inode, which makes it a bit cumbersome since we need to check if the delayed inode exists, if not create it and populate it and deal with any errors (-ENOMEM mostly). This change is part of a patchset comprised of the following patches: 1/5 btrfs: add helper to delete a dir entry from a log tree 2/5 btrfs: pass the dentry to btrfs_log_new_name() instead of the inode 3/5 btrfs: avoid logging all directory changes during renames 4/5 btrfs: stop doing unnecessary log updates during a rename 5/5 btrfs: avoid inode logging during rename and link when possible The following test script mimics part of what the zypper tool does during package installations/upgrades. It does not triggers inode evictions, but it's similar because it triggers false positives from the inode_logged() helper, because the inodes have a logged_trans of 0, there's a log tree due to a fsync of an unrelated file and the directory inode has its last_trans field set to the current transaction: $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=10000 mkfs.btrfs -f $DEV mount $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync # Now do some change to an unrelated file and fsync it. # This is just to create a log tree to make sure that inode_logged() # does not return false when called against "testdir". xfs_io -f -c "pwrite 0 4K" -c "fsync" $MNT/foo # Do some change to testdir. This is to make sure inode_logged() # will return true when called against "testdir", because its # logged_trans is 0, it was changed in the current transaction # and there's a log tree. echo -n > $MNT/testdir/file_$((NUM_FILES + 1)) echo "Renaming $NUM_FILES files..." start=$(date +%s%N) for ((i = 1; i <= $NUM_FILES; i++)); do mv $MNT/testdir/file_$i $MNT/testdir/file_$i-RPMDELETE done end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "Renames took $dur milliseconds" umount $MNT Testing this change on a box using a non-debug kernel (Debian's default kernel config) gave the following results: NUM_FILES=10000, before patchset: 27837 ms NUM_FILES=10000, after patches 1/5 to 4/5 applied: 9236 ms (-66.8%) NUM_FILES=10000, after whole patchset applied: 8902 ms (-68.0%) NUM_FILES=5000, before patchset: 9127 ms NUM_FILES=5000, after patches 1/5 to 4/5 applied: 4640 ms (-49.2%) NUM_FILES=5000, after whole patchset applied: 4441 ms (-51.3%) NUM_FILES=2000, before patchset: 2528 ms NUM_FILES=2000, after patches 1/5 to 4/5 applied: 1983 ms (-21.6%) NUM_FILES=2000, after whole patchset applied: 1747 ms (-30.9%) NUM_FILES=1000, before patchset: 1085 ms NUM_FILES=1000, after patches 1/5 to 4/5 applied: 893 ms (-17.7%) NUM_FILES=1000, after whole patchset applied: 867 ms (-20.1%) Running dbench on the same physical machine with the following script: $ cat run-dbench.sh #!/bin/bash NUM_JOBS=$(nproc --all) DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 MOUNT_OPTIONS="-o ssd" MKFS_OPTIONS="-O no-holes -R free-space-tree" echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -D $MNT -t 120 $NUM_JOBS umount $MNT Before patchset: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 3761352 0.032 143.843 Close 2762770 0.002 2.273 Rename 159304 0.291 67.037 Unlink 759784 0.207 143.998 Deltree 72 4.028 15.977 Mkdir 36 0.003 0.006 Qpathinfo 3409780 0.013 9.678 Qfileinfo 596772 0.001 0.878 Qfsinfo 625189 0.003 1.245 Sfileinfo 306443 0.006 1.840 Find 1318106 0.063 19.798 WriteX 1871137 0.021 8.532 ReadX 5897325 0.003 3.567 LockX 12252 0.003 0.258 UnlockX 12252 0.002 0.100 Flush 263666 3.327 155.632 Throughput 980.047 MB/sec 12 clients 12 procs max_latency=155.636 ms After whole patchset applied: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 4195584 0.033 107.742 Close 3081932 0.002 1.935 Rename 177641 0.218 14.905 Unlink 847333 0.166 107.822 Deltree 118 5.315 15.247 Mkdir 59 0.004 0.048 Qpathinfo 3802612 0.014 10.302 Qfileinfo 666748 0.001 1.034 Qfsinfo 697329 0.003 0.944 Sfileinfo 341712 0.006 2.099 Find 1470365 0.065 9.359 WriteX 2093921 0.021 8.087 ReadX 6576234 0.003 3.407 LockX 13660 0.003 0.308 UnlockX 13660 0.002 0.114 Flush 294090 2.906 115.539 Throughput 1093.11 MB/sec 12 clients 12 procs max_latency=115.544 ms +11.5% throughput -25.8% max latency rename max latency -77.8% Link: https://bugzilla.opensuse.org/show_bug.cgi?id=1193549 Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-01-20 14:00:10 +03:00
* Logging an inode always results in logging its inode item. So if we
* did not find the item we know the inode was not logged for sure.
*/
btrfs: avoid inode logging during rename and link when possible During a rename or link operation, we need to determine if an inode was previously logged or not, and if it was, do some update to the logged inode. We used to rely exclusively on the logged_trans field of struct btrfs_inode to determine that, but that was not reliable because the value of that field is not persisted in the inode item, so it's lost when an inode is evicted and loaded back again. That led to several issues in the past, such as not persisting deletions (such as the case fixed by commit 803f0f64d17769 ("Btrfs: fix fsync not persisting dentry deletions due to inode evictions")), or resulting in losing a file after an inode eviction followed by a rename (commit ecc64fab7d49c6 ("btrfs: fix lost inode on log replay after mix of fsync, rename and inode eviction")), besides other issues. So the inode_logged() helper was introduced and used to determine if an inode was possibly logged before in the current transaction, with the caveat that it could return false positives, in the sense that even if an inode was not logged before in the current transaction, it could still return true, but never to return false in case the inode was logged. >From a functional point of view that is fine, but from a performance perspective it can introduce significant latencies to rename and link operations, as they will end up doing inode logging even when it is not necessary. Recently on a 5.15 kernel, an openSUSE Tumbleweed user reported package installations and upgrades, with the zypper tool, were often taking a long time to complete. With strace it could be observed that zypper was spending about 99% of its time on rename operations, and then with further analysis we checked that directory logging was happening too frequently. Taking into account that installation/upgrade of some of the packages needed a few thousand file renames, the slowdown was very noticeable for the user. The issue was caused indirectly due to an excessive number of inode evictions on a 5.15 kernel, about 100x more compared to a 5.13, 5.14 or a 5.16-rc8 kernel. While triggering the inode evictions if something outside btrfs' control, btrfs could still behave better by eliminating the false positives from the inode_logged() helper. So change inode_logged() to actually eliminate such false positives caused by inode eviction and when an inode was never logged since the filesystem was mounted, as both cases relate to when the logged_trans field of struct btrfs_inode has a value of zero. When it can not determine if the inode was logged based only on the logged_trans value, lookup for the existence of the inode item in the log tree - if it's there then we known the inode was logged, if it's not there then it can not have been logged in the current transaction. Once we determine if the inode was logged, update the logged_trans value to avoid future calls to have to search in the log tree again. Alternatively, we could start storing logged_trans in the on disk inode item structure (struct btrfs_inode_item) in the unused space it still has, but that would be a bit odd because: 1) We only care about logged_trans since the filesystem was mounted, we don't care about its value from a previous mount. Having it persisted in the inode item structure would not make the best use of the precious unused space; 2) In order to get logged_trans persisted before inode eviction, we would have to update the delayed inode when we finish logging the inode and update its logged_trans in struct btrfs_inode, which makes it a bit cumbersome since we need to check if the delayed inode exists, if not create it and populate it and deal with any errors (-ENOMEM mostly). This change is part of a patchset comprised of the following patches: 1/5 btrfs: add helper to delete a dir entry from a log tree 2/5 btrfs: pass the dentry to btrfs_log_new_name() instead of the inode 3/5 btrfs: avoid logging all directory changes during renames 4/5 btrfs: stop doing unnecessary log updates during a rename 5/5 btrfs: avoid inode logging during rename and link when possible The following test script mimics part of what the zypper tool does during package installations/upgrades. It does not triggers inode evictions, but it's similar because it triggers false positives from the inode_logged() helper, because the inodes have a logged_trans of 0, there's a log tree due to a fsync of an unrelated file and the directory inode has its last_trans field set to the current transaction: $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=10000 mkfs.btrfs -f $DEV mount $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync # Now do some change to an unrelated file and fsync it. # This is just to create a log tree to make sure that inode_logged() # does not return false when called against "testdir". xfs_io -f -c "pwrite 0 4K" -c "fsync" $MNT/foo # Do some change to testdir. This is to make sure inode_logged() # will return true when called against "testdir", because its # logged_trans is 0, it was changed in the current transaction # and there's a log tree. echo -n > $MNT/testdir/file_$((NUM_FILES + 1)) echo "Renaming $NUM_FILES files..." start=$(date +%s%N) for ((i = 1; i <= $NUM_FILES; i++)); do mv $MNT/testdir/file_$i $MNT/testdir/file_$i-RPMDELETE done end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "Renames took $dur milliseconds" umount $MNT Testing this change on a box using a non-debug kernel (Debian's default kernel config) gave the following results: NUM_FILES=10000, before patchset: 27837 ms NUM_FILES=10000, after patches 1/5 to 4/5 applied: 9236 ms (-66.8%) NUM_FILES=10000, after whole patchset applied: 8902 ms (-68.0%) NUM_FILES=5000, before patchset: 9127 ms NUM_FILES=5000, after patches 1/5 to 4/5 applied: 4640 ms (-49.2%) NUM_FILES=5000, after whole patchset applied: 4441 ms (-51.3%) NUM_FILES=2000, before patchset: 2528 ms NUM_FILES=2000, after patches 1/5 to 4/5 applied: 1983 ms (-21.6%) NUM_FILES=2000, after whole patchset applied: 1747 ms (-30.9%) NUM_FILES=1000, before patchset: 1085 ms NUM_FILES=1000, after patches 1/5 to 4/5 applied: 893 ms (-17.7%) NUM_FILES=1000, after whole patchset applied: 867 ms (-20.1%) Running dbench on the same physical machine with the following script: $ cat run-dbench.sh #!/bin/bash NUM_JOBS=$(nproc --all) DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 MOUNT_OPTIONS="-o ssd" MKFS_OPTIONS="-O no-holes -R free-space-tree" echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -D $MNT -t 120 $NUM_JOBS umount $MNT Before patchset: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 3761352 0.032 143.843 Close 2762770 0.002 2.273 Rename 159304 0.291 67.037 Unlink 759784 0.207 143.998 Deltree 72 4.028 15.977 Mkdir 36 0.003 0.006 Qpathinfo 3409780 0.013 9.678 Qfileinfo 596772 0.001 0.878 Qfsinfo 625189 0.003 1.245 Sfileinfo 306443 0.006 1.840 Find 1318106 0.063 19.798 WriteX 1871137 0.021 8.532 ReadX 5897325 0.003 3.567 LockX 12252 0.003 0.258 UnlockX 12252 0.002 0.100 Flush 263666 3.327 155.632 Throughput 980.047 MB/sec 12 clients 12 procs max_latency=155.636 ms After whole patchset applied: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 4195584 0.033 107.742 Close 3081932 0.002 1.935 Rename 177641 0.218 14.905 Unlink 847333 0.166 107.822 Deltree 118 5.315 15.247 Mkdir 59 0.004 0.048 Qpathinfo 3802612 0.014 10.302 Qfileinfo 666748 0.001 1.034 Qfsinfo 697329 0.003 0.944 Sfileinfo 341712 0.006 2.099 Find 1470365 0.065 9.359 WriteX 2093921 0.021 8.087 ReadX 6576234 0.003 3.407 LockX 13660 0.003 0.308 UnlockX 13660 0.002 0.114 Flush 294090 2.906 115.539 Throughput 1093.11 MB/sec 12 clients 12 procs max_latency=115.544 ms +11.5% throughput -25.8% max latency rename max latency -77.8% Link: https://bugzilla.opensuse.org/show_bug.cgi?id=1193549 Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-01-20 14:00:10 +03:00
if (ret < 0) {
return ret;
} else if (ret > 0) {
/*
* Set logged_trans to a value greater than 0 and less then the
* current transaction to avoid doing the search in future calls.
*/
inode->logged_trans = trans->transid - 1;
return 0;
}
/*
* The inode was previously logged and then evicted, set logged_trans to
* the current transacion's ID, to avoid future tree searches as long as
* the inode is not evicted again.
*/
inode->logged_trans = trans->transid;
/*
* If it's a directory, then we must set last_dir_index_offset to the
* maximum possible value, so that the next attempt to log the inode does
* not skip checking if dir index keys found in modified subvolume tree
* leaves have been logged before, otherwise it would result in attempts
* to insert duplicate dir index keys in the log tree. This must be done
* because last_dir_index_offset is an in-memory only field, not persisted
* in the inode item or any other on-disk structure, so its value is lost
* once the inode is evicted.
*/
if (S_ISDIR(inode->vfs_inode.i_mode))
inode->last_dir_index_offset = (u64)-1;
Btrfs: fix fsync not persisting dentry deletions due to inode evictions In order to avoid searches on a log tree when unlinking an inode, we check if the inode being unlinked was logged in the current transaction, as well as the inode of its parent directory. When any of the inodes are logged, we proceed to delete directory items and inode reference items from the log, to ensure that if a subsequent fsync of only the inode being unlinked or only of the parent directory when the other is not fsync'ed as well, does not result in the entry still existing after a power failure. That check however is not reliable when one of the inodes involved (the one being unlinked or its parent directory's inode) is evicted, since the logged_trans field is transient, that is, it is not stored on disk, so it is lost when the inode is evicted and loaded into memory again (which is set to zero on load). As a consequence the checks currently being done by btrfs_del_dir_entries_in_log() and btrfs_del_inode_ref_in_log() always return true if the inode was evicted before, regardless of the inode having been logged or not before (and in the current transaction), this results in the dentry being unlinked still existing after a log replay if after the unlink operation only one of the inodes involved is fsync'ed. Example: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/dir $ touch /mnt/dir/foo $ xfs_io -c fsync /mnt/dir/foo # Keep an open file descriptor on our directory while we evict inodes. # We just want to evict the file's inode, the directory's inode must not # be evicted. $ ( cd /mnt/dir; while true; do :; done ) & $ pid=$! # Wait a bit to give time to background process to chdir to our test # directory. $ sleep 0.5 # Trigger eviction of the file's inode. $ echo 2 > /proc/sys/vm/drop_caches # Unlink our file and fsync the parent directory. After a power failure # we don't expect to see the file anymore, since we fsync'ed the parent # directory. $ rm -f $SCRATCH_MNT/dir/foo $ xfs_io -c fsync /mnt/dir <power failure> $ mount /dev/sdb /mnt $ ls /mnt/dir foo $ --> file still there, unlink not persisted despite explicit fsync on dir Fix this by checking if the inode has the full_sync bit set in its runtime flags as well, since that bit is set everytime an inode is loaded from disk, or for other less common cases such as after a shrinking truncate or failure to allocate extent maps for holes, and gets cleared after the first fsync. Also consider the inode as possibly logged only if it was last modified in the current transaction (besides having the full_fsync flag set). Fixes: 3a5f1d458ad161 ("Btrfs: Optimize btree walking while logging inodes") CC: stable@vger.kernel.org # 4.4+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-06-19 15:05:39 +03:00
btrfs: avoid inode logging during rename and link when possible During a rename or link operation, we need to determine if an inode was previously logged or not, and if it was, do some update to the logged inode. We used to rely exclusively on the logged_trans field of struct btrfs_inode to determine that, but that was not reliable because the value of that field is not persisted in the inode item, so it's lost when an inode is evicted and loaded back again. That led to several issues in the past, such as not persisting deletions (such as the case fixed by commit 803f0f64d17769 ("Btrfs: fix fsync not persisting dentry deletions due to inode evictions")), or resulting in losing a file after an inode eviction followed by a rename (commit ecc64fab7d49c6 ("btrfs: fix lost inode on log replay after mix of fsync, rename and inode eviction")), besides other issues. So the inode_logged() helper was introduced and used to determine if an inode was possibly logged before in the current transaction, with the caveat that it could return false positives, in the sense that even if an inode was not logged before in the current transaction, it could still return true, but never to return false in case the inode was logged. >From a functional point of view that is fine, but from a performance perspective it can introduce significant latencies to rename and link operations, as they will end up doing inode logging even when it is not necessary. Recently on a 5.15 kernel, an openSUSE Tumbleweed user reported package installations and upgrades, with the zypper tool, were often taking a long time to complete. With strace it could be observed that zypper was spending about 99% of its time on rename operations, and then with further analysis we checked that directory logging was happening too frequently. Taking into account that installation/upgrade of some of the packages needed a few thousand file renames, the slowdown was very noticeable for the user. The issue was caused indirectly due to an excessive number of inode evictions on a 5.15 kernel, about 100x more compared to a 5.13, 5.14 or a 5.16-rc8 kernel. While triggering the inode evictions if something outside btrfs' control, btrfs could still behave better by eliminating the false positives from the inode_logged() helper. So change inode_logged() to actually eliminate such false positives caused by inode eviction and when an inode was never logged since the filesystem was mounted, as both cases relate to when the logged_trans field of struct btrfs_inode has a value of zero. When it can not determine if the inode was logged based only on the logged_trans value, lookup for the existence of the inode item in the log tree - if it's there then we known the inode was logged, if it's not there then it can not have been logged in the current transaction. Once we determine if the inode was logged, update the logged_trans value to avoid future calls to have to search in the log tree again. Alternatively, we could start storing logged_trans in the on disk inode item structure (struct btrfs_inode_item) in the unused space it still has, but that would be a bit odd because: 1) We only care about logged_trans since the filesystem was mounted, we don't care about its value from a previous mount. Having it persisted in the inode item structure would not make the best use of the precious unused space; 2) In order to get logged_trans persisted before inode eviction, we would have to update the delayed inode when we finish logging the inode and update its logged_trans in struct btrfs_inode, which makes it a bit cumbersome since we need to check if the delayed inode exists, if not create it and populate it and deal with any errors (-ENOMEM mostly). This change is part of a patchset comprised of the following patches: 1/5 btrfs: add helper to delete a dir entry from a log tree 2/5 btrfs: pass the dentry to btrfs_log_new_name() instead of the inode 3/5 btrfs: avoid logging all directory changes during renames 4/5 btrfs: stop doing unnecessary log updates during a rename 5/5 btrfs: avoid inode logging during rename and link when possible The following test script mimics part of what the zypper tool does during package installations/upgrades. It does not triggers inode evictions, but it's similar because it triggers false positives from the inode_logged() helper, because the inodes have a logged_trans of 0, there's a log tree due to a fsync of an unrelated file and the directory inode has its last_trans field set to the current transaction: $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=10000 mkfs.btrfs -f $DEV mount $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync # Now do some change to an unrelated file and fsync it. # This is just to create a log tree to make sure that inode_logged() # does not return false when called against "testdir". xfs_io -f -c "pwrite 0 4K" -c "fsync" $MNT/foo # Do some change to testdir. This is to make sure inode_logged() # will return true when called against "testdir", because its # logged_trans is 0, it was changed in the current transaction # and there's a log tree. echo -n > $MNT/testdir/file_$((NUM_FILES + 1)) echo "Renaming $NUM_FILES files..." start=$(date +%s%N) for ((i = 1; i <= $NUM_FILES; i++)); do mv $MNT/testdir/file_$i $MNT/testdir/file_$i-RPMDELETE done end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "Renames took $dur milliseconds" umount $MNT Testing this change on a box using a non-debug kernel (Debian's default kernel config) gave the following results: NUM_FILES=10000, before patchset: 27837 ms NUM_FILES=10000, after patches 1/5 to 4/5 applied: 9236 ms (-66.8%) NUM_FILES=10000, after whole patchset applied: 8902 ms (-68.0%) NUM_FILES=5000, before patchset: 9127 ms NUM_FILES=5000, after patches 1/5 to 4/5 applied: 4640 ms (-49.2%) NUM_FILES=5000, after whole patchset applied: 4441 ms (-51.3%) NUM_FILES=2000, before patchset: 2528 ms NUM_FILES=2000, after patches 1/5 to 4/5 applied: 1983 ms (-21.6%) NUM_FILES=2000, after whole patchset applied: 1747 ms (-30.9%) NUM_FILES=1000, before patchset: 1085 ms NUM_FILES=1000, after patches 1/5 to 4/5 applied: 893 ms (-17.7%) NUM_FILES=1000, after whole patchset applied: 867 ms (-20.1%) Running dbench on the same physical machine with the following script: $ cat run-dbench.sh #!/bin/bash NUM_JOBS=$(nproc --all) DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 MOUNT_OPTIONS="-o ssd" MKFS_OPTIONS="-O no-holes -R free-space-tree" echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -D $MNT -t 120 $NUM_JOBS umount $MNT Before patchset: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 3761352 0.032 143.843 Close 2762770 0.002 2.273 Rename 159304 0.291 67.037 Unlink 759784 0.207 143.998 Deltree 72 4.028 15.977 Mkdir 36 0.003 0.006 Qpathinfo 3409780 0.013 9.678 Qfileinfo 596772 0.001 0.878 Qfsinfo 625189 0.003 1.245 Sfileinfo 306443 0.006 1.840 Find 1318106 0.063 19.798 WriteX 1871137 0.021 8.532 ReadX 5897325 0.003 3.567 LockX 12252 0.003 0.258 UnlockX 12252 0.002 0.100 Flush 263666 3.327 155.632 Throughput 980.047 MB/sec 12 clients 12 procs max_latency=155.636 ms After whole patchset applied: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 4195584 0.033 107.742 Close 3081932 0.002 1.935 Rename 177641 0.218 14.905 Unlink 847333 0.166 107.822 Deltree 118 5.315 15.247 Mkdir 59 0.004 0.048 Qpathinfo 3802612 0.014 10.302 Qfileinfo 666748 0.001 1.034 Qfsinfo 697329 0.003 0.944 Sfileinfo 341712 0.006 2.099 Find 1470365 0.065 9.359 WriteX 2093921 0.021 8.087 ReadX 6576234 0.003 3.407 LockX 13660 0.003 0.308 UnlockX 13660 0.002 0.114 Flush 294090 2.906 115.539 Throughput 1093.11 MB/sec 12 clients 12 procs max_latency=115.544 ms +11.5% throughput -25.8% max latency rename max latency -77.8% Link: https://bugzilla.opensuse.org/show_bug.cgi?id=1193549 Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-01-20 14:00:10 +03:00
return 1;
Btrfs: fix fsync not persisting dentry deletions due to inode evictions In order to avoid searches on a log tree when unlinking an inode, we check if the inode being unlinked was logged in the current transaction, as well as the inode of its parent directory. When any of the inodes are logged, we proceed to delete directory items and inode reference items from the log, to ensure that if a subsequent fsync of only the inode being unlinked or only of the parent directory when the other is not fsync'ed as well, does not result in the entry still existing after a power failure. That check however is not reliable when one of the inodes involved (the one being unlinked or its parent directory's inode) is evicted, since the logged_trans field is transient, that is, it is not stored on disk, so it is lost when the inode is evicted and loaded into memory again (which is set to zero on load). As a consequence the checks currently being done by btrfs_del_dir_entries_in_log() and btrfs_del_inode_ref_in_log() always return true if the inode was evicted before, regardless of the inode having been logged or not before (and in the current transaction), this results in the dentry being unlinked still existing after a log replay if after the unlink operation only one of the inodes involved is fsync'ed. Example: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/dir $ touch /mnt/dir/foo $ xfs_io -c fsync /mnt/dir/foo # Keep an open file descriptor on our directory while we evict inodes. # We just want to evict the file's inode, the directory's inode must not # be evicted. $ ( cd /mnt/dir; while true; do :; done ) & $ pid=$! # Wait a bit to give time to background process to chdir to our test # directory. $ sleep 0.5 # Trigger eviction of the file's inode. $ echo 2 > /proc/sys/vm/drop_caches # Unlink our file and fsync the parent directory. After a power failure # we don't expect to see the file anymore, since we fsync'ed the parent # directory. $ rm -f $SCRATCH_MNT/dir/foo $ xfs_io -c fsync /mnt/dir <power failure> $ mount /dev/sdb /mnt $ ls /mnt/dir foo $ --> file still there, unlink not persisted despite explicit fsync on dir Fix this by checking if the inode has the full_sync bit set in its runtime flags as well, since that bit is set everytime an inode is loaded from disk, or for other less common cases such as after a shrinking truncate or failure to allocate extent maps for holes, and gets cleared after the first fsync. Also consider the inode as possibly logged only if it was last modified in the current transaction (besides having the full_fsync flag set). Fixes: 3a5f1d458ad161 ("Btrfs: Optimize btree walking while logging inodes") CC: stable@vger.kernel.org # 4.4+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-06-19 15:05:39 +03:00
}
/*
* Delete a directory entry from the log if it exists.
*
* Returns < 0 on error
* 1 if the entry does not exists
* 0 if the entry existed and was successfully deleted
*/
static int del_logged_dentry(struct btrfs_trans_handle *trans,
struct btrfs_root *log,
struct btrfs_path *path,
u64 dir_ino,
const char *name, int name_len,
u64 index)
{
struct btrfs_dir_item *di;
/*
* We only log dir index items of a directory, so we don't need to look
* for dir item keys.
*/
di = btrfs_lookup_dir_index_item(trans, log, path, dir_ino,
index, name, name_len, -1);
if (IS_ERR(di))
return PTR_ERR(di);
else if (!di)
return 1;
/*
* We do not need to update the size field of the directory's
* inode item because on log replay we update the field to reflect
* all existing entries in the directory (see overwrite_item()).
*/
return btrfs_delete_one_dir_name(trans, log, path, di);
}
/*
* If both a file and directory are logged, and unlinks or renames are
* mixed in, we have a few interesting corners:
*
* create file X in dir Y
* link file X to X.link in dir Y
* fsync file X
* unlink file X but leave X.link
* fsync dir Y
*
* After a crash we would expect only X.link to exist. But file X
* didn't get fsync'd again so the log has back refs for X and X.link.
*
* We solve this by removing directory entries and inode backrefs from the
* log when a file that was logged in the current transaction is
* unlinked. Any later fsync will include the updated log entries, and
* we'll be able to reconstruct the proper directory items from backrefs.
*
* This optimizations allows us to avoid relogging the entire inode
* or the entire directory.
*/
void btrfs_del_dir_entries_in_log(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
const char *name, int name_len,
struct btrfs_inode *dir, u64 index)
{
struct btrfs_path *path;
int ret;
btrfs: avoid inode logging during rename and link when possible During a rename or link operation, we need to determine if an inode was previously logged or not, and if it was, do some update to the logged inode. We used to rely exclusively on the logged_trans field of struct btrfs_inode to determine that, but that was not reliable because the value of that field is not persisted in the inode item, so it's lost when an inode is evicted and loaded back again. That led to several issues in the past, such as not persisting deletions (such as the case fixed by commit 803f0f64d17769 ("Btrfs: fix fsync not persisting dentry deletions due to inode evictions")), or resulting in losing a file after an inode eviction followed by a rename (commit ecc64fab7d49c6 ("btrfs: fix lost inode on log replay after mix of fsync, rename and inode eviction")), besides other issues. So the inode_logged() helper was introduced and used to determine if an inode was possibly logged before in the current transaction, with the caveat that it could return false positives, in the sense that even if an inode was not logged before in the current transaction, it could still return true, but never to return false in case the inode was logged. >From a functional point of view that is fine, but from a performance perspective it can introduce significant latencies to rename and link operations, as they will end up doing inode logging even when it is not necessary. Recently on a 5.15 kernel, an openSUSE Tumbleweed user reported package installations and upgrades, with the zypper tool, were often taking a long time to complete. With strace it could be observed that zypper was spending about 99% of its time on rename operations, and then with further analysis we checked that directory logging was happening too frequently. Taking into account that installation/upgrade of some of the packages needed a few thousand file renames, the slowdown was very noticeable for the user. The issue was caused indirectly due to an excessive number of inode evictions on a 5.15 kernel, about 100x more compared to a 5.13, 5.14 or a 5.16-rc8 kernel. While triggering the inode evictions if something outside btrfs' control, btrfs could still behave better by eliminating the false positives from the inode_logged() helper. So change inode_logged() to actually eliminate such false positives caused by inode eviction and when an inode was never logged since the filesystem was mounted, as both cases relate to when the logged_trans field of struct btrfs_inode has a value of zero. When it can not determine if the inode was logged based only on the logged_trans value, lookup for the existence of the inode item in the log tree - if it's there then we known the inode was logged, if it's not there then it can not have been logged in the current transaction. Once we determine if the inode was logged, update the logged_trans value to avoid future calls to have to search in the log tree again. Alternatively, we could start storing logged_trans in the on disk inode item structure (struct btrfs_inode_item) in the unused space it still has, but that would be a bit odd because: 1) We only care about logged_trans since the filesystem was mounted, we don't care about its value from a previous mount. Having it persisted in the inode item structure would not make the best use of the precious unused space; 2) In order to get logged_trans persisted before inode eviction, we would have to update the delayed inode when we finish logging the inode and update its logged_trans in struct btrfs_inode, which makes it a bit cumbersome since we need to check if the delayed inode exists, if not create it and populate it and deal with any errors (-ENOMEM mostly). This change is part of a patchset comprised of the following patches: 1/5 btrfs: add helper to delete a dir entry from a log tree 2/5 btrfs: pass the dentry to btrfs_log_new_name() instead of the inode 3/5 btrfs: avoid logging all directory changes during renames 4/5 btrfs: stop doing unnecessary log updates during a rename 5/5 btrfs: avoid inode logging during rename and link when possible The following test script mimics part of what the zypper tool does during package installations/upgrades. It does not triggers inode evictions, but it's similar because it triggers false positives from the inode_logged() helper, because the inodes have a logged_trans of 0, there's a log tree due to a fsync of an unrelated file and the directory inode has its last_trans field set to the current transaction: $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=10000 mkfs.btrfs -f $DEV mount $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync # Now do some change to an unrelated file and fsync it. # This is just to create a log tree to make sure that inode_logged() # does not return false when called against "testdir". xfs_io -f -c "pwrite 0 4K" -c "fsync" $MNT/foo # Do some change to testdir. This is to make sure inode_logged() # will return true when called against "testdir", because its # logged_trans is 0, it was changed in the current transaction # and there's a log tree. echo -n > $MNT/testdir/file_$((NUM_FILES + 1)) echo "Renaming $NUM_FILES files..." start=$(date +%s%N) for ((i = 1; i <= $NUM_FILES; i++)); do mv $MNT/testdir/file_$i $MNT/testdir/file_$i-RPMDELETE done end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "Renames took $dur milliseconds" umount $MNT Testing this change on a box using a non-debug kernel (Debian's default kernel config) gave the following results: NUM_FILES=10000, before patchset: 27837 ms NUM_FILES=10000, after patches 1/5 to 4/5 applied: 9236 ms (-66.8%) NUM_FILES=10000, after whole patchset applied: 8902 ms (-68.0%) NUM_FILES=5000, before patchset: 9127 ms NUM_FILES=5000, after patches 1/5 to 4/5 applied: 4640 ms (-49.2%) NUM_FILES=5000, after whole patchset applied: 4441 ms (-51.3%) NUM_FILES=2000, before patchset: 2528 ms NUM_FILES=2000, after patches 1/5 to 4/5 applied: 1983 ms (-21.6%) NUM_FILES=2000, after whole patchset applied: 1747 ms (-30.9%) NUM_FILES=1000, before patchset: 1085 ms NUM_FILES=1000, after patches 1/5 to 4/5 applied: 893 ms (-17.7%) NUM_FILES=1000, after whole patchset applied: 867 ms (-20.1%) Running dbench on the same physical machine with the following script: $ cat run-dbench.sh #!/bin/bash NUM_JOBS=$(nproc --all) DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 MOUNT_OPTIONS="-o ssd" MKFS_OPTIONS="-O no-holes -R free-space-tree" echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -D $MNT -t 120 $NUM_JOBS umount $MNT Before patchset: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 3761352 0.032 143.843 Close 2762770 0.002 2.273 Rename 159304 0.291 67.037 Unlink 759784 0.207 143.998 Deltree 72 4.028 15.977 Mkdir 36 0.003 0.006 Qpathinfo 3409780 0.013 9.678 Qfileinfo 596772 0.001 0.878 Qfsinfo 625189 0.003 1.245 Sfileinfo 306443 0.006 1.840 Find 1318106 0.063 19.798 WriteX 1871137 0.021 8.532 ReadX 5897325 0.003 3.567 LockX 12252 0.003 0.258 UnlockX 12252 0.002 0.100 Flush 263666 3.327 155.632 Throughput 980.047 MB/sec 12 clients 12 procs max_latency=155.636 ms After whole patchset applied: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 4195584 0.033 107.742 Close 3081932 0.002 1.935 Rename 177641 0.218 14.905 Unlink 847333 0.166 107.822 Deltree 118 5.315 15.247 Mkdir 59 0.004 0.048 Qpathinfo 3802612 0.014 10.302 Qfileinfo 666748 0.001 1.034 Qfsinfo 697329 0.003 0.944 Sfileinfo 341712 0.006 2.099 Find 1470365 0.065 9.359 WriteX 2093921 0.021 8.087 ReadX 6576234 0.003 3.407 LockX 13660 0.003 0.308 UnlockX 13660 0.002 0.114 Flush 294090 2.906 115.539 Throughput 1093.11 MB/sec 12 clients 12 procs max_latency=115.544 ms +11.5% throughput -25.8% max latency rename max latency -77.8% Link: https://bugzilla.opensuse.org/show_bug.cgi?id=1193549 Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-01-20 14:00:10 +03:00
ret = inode_logged(trans, dir, NULL);
if (ret == 0)
return;
else if (ret < 0) {
btrfs_set_log_full_commit(trans);
return;
btrfs: avoid inode logging during rename and link when possible During a rename or link operation, we need to determine if an inode was previously logged or not, and if it was, do some update to the logged inode. We used to rely exclusively on the logged_trans field of struct btrfs_inode to determine that, but that was not reliable because the value of that field is not persisted in the inode item, so it's lost when an inode is evicted and loaded back again. That led to several issues in the past, such as not persisting deletions (such as the case fixed by commit 803f0f64d17769 ("Btrfs: fix fsync not persisting dentry deletions due to inode evictions")), or resulting in losing a file after an inode eviction followed by a rename (commit ecc64fab7d49c6 ("btrfs: fix lost inode on log replay after mix of fsync, rename and inode eviction")), besides other issues. So the inode_logged() helper was introduced and used to determine if an inode was possibly logged before in the current transaction, with the caveat that it could return false positives, in the sense that even if an inode was not logged before in the current transaction, it could still return true, but never to return false in case the inode was logged. >From a functional point of view that is fine, but from a performance perspective it can introduce significant latencies to rename and link operations, as they will end up doing inode logging even when it is not necessary. Recently on a 5.15 kernel, an openSUSE Tumbleweed user reported package installations and upgrades, with the zypper tool, were often taking a long time to complete. With strace it could be observed that zypper was spending about 99% of its time on rename operations, and then with further analysis we checked that directory logging was happening too frequently. Taking into account that installation/upgrade of some of the packages needed a few thousand file renames, the slowdown was very noticeable for the user. The issue was caused indirectly due to an excessive number of inode evictions on a 5.15 kernel, about 100x more compared to a 5.13, 5.14 or a 5.16-rc8 kernel. While triggering the inode evictions if something outside btrfs' control, btrfs could still behave better by eliminating the false positives from the inode_logged() helper. So change inode_logged() to actually eliminate such false positives caused by inode eviction and when an inode was never logged since the filesystem was mounted, as both cases relate to when the logged_trans field of struct btrfs_inode has a value of zero. When it can not determine if the inode was logged based only on the logged_trans value, lookup for the existence of the inode item in the log tree - if it's there then we known the inode was logged, if it's not there then it can not have been logged in the current transaction. Once we determine if the inode was logged, update the logged_trans value to avoid future calls to have to search in the log tree again. Alternatively, we could start storing logged_trans in the on disk inode item structure (struct btrfs_inode_item) in the unused space it still has, but that would be a bit odd because: 1) We only care about logged_trans since the filesystem was mounted, we don't care about its value from a previous mount. Having it persisted in the inode item structure would not make the best use of the precious unused space; 2) In order to get logged_trans persisted before inode eviction, we would have to update the delayed inode when we finish logging the inode and update its logged_trans in struct btrfs_inode, which makes it a bit cumbersome since we need to check if the delayed inode exists, if not create it and populate it and deal with any errors (-ENOMEM mostly). This change is part of a patchset comprised of the following patches: 1/5 btrfs: add helper to delete a dir entry from a log tree 2/5 btrfs: pass the dentry to btrfs_log_new_name() instead of the inode 3/5 btrfs: avoid logging all directory changes during renames 4/5 btrfs: stop doing unnecessary log updates during a rename 5/5 btrfs: avoid inode logging during rename and link when possible The following test script mimics part of what the zypper tool does during package installations/upgrades. It does not triggers inode evictions, but it's similar because it triggers false positives from the inode_logged() helper, because the inodes have a logged_trans of 0, there's a log tree due to a fsync of an unrelated file and the directory inode has its last_trans field set to the current transaction: $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=10000 mkfs.btrfs -f $DEV mount $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync # Now do some change to an unrelated file and fsync it. # This is just to create a log tree to make sure that inode_logged() # does not return false when called against "testdir". xfs_io -f -c "pwrite 0 4K" -c "fsync" $MNT/foo # Do some change to testdir. This is to make sure inode_logged() # will return true when called against "testdir", because its # logged_trans is 0, it was changed in the current transaction # and there's a log tree. echo -n > $MNT/testdir/file_$((NUM_FILES + 1)) echo "Renaming $NUM_FILES files..." start=$(date +%s%N) for ((i = 1; i <= $NUM_FILES; i++)); do mv $MNT/testdir/file_$i $MNT/testdir/file_$i-RPMDELETE done end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "Renames took $dur milliseconds" umount $MNT Testing this change on a box using a non-debug kernel (Debian's default kernel config) gave the following results: NUM_FILES=10000, before patchset: 27837 ms NUM_FILES=10000, after patches 1/5 to 4/5 applied: 9236 ms (-66.8%) NUM_FILES=10000, after whole patchset applied: 8902 ms (-68.0%) NUM_FILES=5000, before patchset: 9127 ms NUM_FILES=5000, after patches 1/5 to 4/5 applied: 4640 ms (-49.2%) NUM_FILES=5000, after whole patchset applied: 4441 ms (-51.3%) NUM_FILES=2000, before patchset: 2528 ms NUM_FILES=2000, after patches 1/5 to 4/5 applied: 1983 ms (-21.6%) NUM_FILES=2000, after whole patchset applied: 1747 ms (-30.9%) NUM_FILES=1000, before patchset: 1085 ms NUM_FILES=1000, after patches 1/5 to 4/5 applied: 893 ms (-17.7%) NUM_FILES=1000, after whole patchset applied: 867 ms (-20.1%) Running dbench on the same physical machine with the following script: $ cat run-dbench.sh #!/bin/bash NUM_JOBS=$(nproc --all) DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 MOUNT_OPTIONS="-o ssd" MKFS_OPTIONS="-O no-holes -R free-space-tree" echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -D $MNT -t 120 $NUM_JOBS umount $MNT Before patchset: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 3761352 0.032 143.843 Close 2762770 0.002 2.273 Rename 159304 0.291 67.037 Unlink 759784 0.207 143.998 Deltree 72 4.028 15.977 Mkdir 36 0.003 0.006 Qpathinfo 3409780 0.013 9.678 Qfileinfo 596772 0.001 0.878 Qfsinfo 625189 0.003 1.245 Sfileinfo 306443 0.006 1.840 Find 1318106 0.063 19.798 WriteX 1871137 0.021 8.532 ReadX 5897325 0.003 3.567 LockX 12252 0.003 0.258 UnlockX 12252 0.002 0.100 Flush 263666 3.327 155.632 Throughput 980.047 MB/sec 12 clients 12 procs max_latency=155.636 ms After whole patchset applied: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 4195584 0.033 107.742 Close 3081932 0.002 1.935 Rename 177641 0.218 14.905 Unlink 847333 0.166 107.822 Deltree 118 5.315 15.247 Mkdir 59 0.004 0.048 Qpathinfo 3802612 0.014 10.302 Qfileinfo 666748 0.001 1.034 Qfsinfo 697329 0.003 0.944 Sfileinfo 341712 0.006 2.099 Find 1470365 0.065 9.359 WriteX 2093921 0.021 8.087 ReadX 6576234 0.003 3.407 LockX 13660 0.003 0.308 UnlockX 13660 0.002 0.114 Flush 294090 2.906 115.539 Throughput 1093.11 MB/sec 12 clients 12 procs max_latency=115.544 ms +11.5% throughput -25.8% max latency rename max latency -77.8% Link: https://bugzilla.opensuse.org/show_bug.cgi?id=1193549 Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-01-20 14:00:10 +03:00
}
ret = join_running_log_trans(root);
if (ret)
return;
mutex_lock(&dir->log_mutex);
path = btrfs_alloc_path();
if (!path) {
ret = -ENOMEM;
goto out_unlock;
}
ret = del_logged_dentry(trans, root->log_root, path, btrfs_ino(dir),
name, name_len, index);
btrfs_free_path(path);
out_unlock:
mutex_unlock(&dir->log_mutex);
if (ret < 0)
btrfs_set_log_full_commit(trans);
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 17:24:20 +03:00
btrfs_end_log_trans(root);
}
/* see comments for btrfs_del_dir_entries_in_log */
void btrfs_del_inode_ref_in_log(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
const char *name, int name_len,
struct btrfs_inode *inode, u64 dirid)
{
struct btrfs_root *log;
u64 index;
int ret;
btrfs: avoid inode logging during rename and link when possible During a rename or link operation, we need to determine if an inode was previously logged or not, and if it was, do some update to the logged inode. We used to rely exclusively on the logged_trans field of struct btrfs_inode to determine that, but that was not reliable because the value of that field is not persisted in the inode item, so it's lost when an inode is evicted and loaded back again. That led to several issues in the past, such as not persisting deletions (such as the case fixed by commit 803f0f64d17769 ("Btrfs: fix fsync not persisting dentry deletions due to inode evictions")), or resulting in losing a file after an inode eviction followed by a rename (commit ecc64fab7d49c6 ("btrfs: fix lost inode on log replay after mix of fsync, rename and inode eviction")), besides other issues. So the inode_logged() helper was introduced and used to determine if an inode was possibly logged before in the current transaction, with the caveat that it could return false positives, in the sense that even if an inode was not logged before in the current transaction, it could still return true, but never to return false in case the inode was logged. >From a functional point of view that is fine, but from a performance perspective it can introduce significant latencies to rename and link operations, as they will end up doing inode logging even when it is not necessary. Recently on a 5.15 kernel, an openSUSE Tumbleweed user reported package installations and upgrades, with the zypper tool, were often taking a long time to complete. With strace it could be observed that zypper was spending about 99% of its time on rename operations, and then with further analysis we checked that directory logging was happening too frequently. Taking into account that installation/upgrade of some of the packages needed a few thousand file renames, the slowdown was very noticeable for the user. The issue was caused indirectly due to an excessive number of inode evictions on a 5.15 kernel, about 100x more compared to a 5.13, 5.14 or a 5.16-rc8 kernel. While triggering the inode evictions if something outside btrfs' control, btrfs could still behave better by eliminating the false positives from the inode_logged() helper. So change inode_logged() to actually eliminate such false positives caused by inode eviction and when an inode was never logged since the filesystem was mounted, as both cases relate to when the logged_trans field of struct btrfs_inode has a value of zero. When it can not determine if the inode was logged based only on the logged_trans value, lookup for the existence of the inode item in the log tree - if it's there then we known the inode was logged, if it's not there then it can not have been logged in the current transaction. Once we determine if the inode was logged, update the logged_trans value to avoid future calls to have to search in the log tree again. Alternatively, we could start storing logged_trans in the on disk inode item structure (struct btrfs_inode_item) in the unused space it still has, but that would be a bit odd because: 1) We only care about logged_trans since the filesystem was mounted, we don't care about its value from a previous mount. Having it persisted in the inode item structure would not make the best use of the precious unused space; 2) In order to get logged_trans persisted before inode eviction, we would have to update the delayed inode when we finish logging the inode and update its logged_trans in struct btrfs_inode, which makes it a bit cumbersome since we need to check if the delayed inode exists, if not create it and populate it and deal with any errors (-ENOMEM mostly). This change is part of a patchset comprised of the following patches: 1/5 btrfs: add helper to delete a dir entry from a log tree 2/5 btrfs: pass the dentry to btrfs_log_new_name() instead of the inode 3/5 btrfs: avoid logging all directory changes during renames 4/5 btrfs: stop doing unnecessary log updates during a rename 5/5 btrfs: avoid inode logging during rename and link when possible The following test script mimics part of what the zypper tool does during package installations/upgrades. It does not triggers inode evictions, but it's similar because it triggers false positives from the inode_logged() helper, because the inodes have a logged_trans of 0, there's a log tree due to a fsync of an unrelated file and the directory inode has its last_trans field set to the current transaction: $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=10000 mkfs.btrfs -f $DEV mount $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync # Now do some change to an unrelated file and fsync it. # This is just to create a log tree to make sure that inode_logged() # does not return false when called against "testdir". xfs_io -f -c "pwrite 0 4K" -c "fsync" $MNT/foo # Do some change to testdir. This is to make sure inode_logged() # will return true when called against "testdir", because its # logged_trans is 0, it was changed in the current transaction # and there's a log tree. echo -n > $MNT/testdir/file_$((NUM_FILES + 1)) echo "Renaming $NUM_FILES files..." start=$(date +%s%N) for ((i = 1; i <= $NUM_FILES; i++)); do mv $MNT/testdir/file_$i $MNT/testdir/file_$i-RPMDELETE done end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "Renames took $dur milliseconds" umount $MNT Testing this change on a box using a non-debug kernel (Debian's default kernel config) gave the following results: NUM_FILES=10000, before patchset: 27837 ms NUM_FILES=10000, after patches 1/5 to 4/5 applied: 9236 ms (-66.8%) NUM_FILES=10000, after whole patchset applied: 8902 ms (-68.0%) NUM_FILES=5000, before patchset: 9127 ms NUM_FILES=5000, after patches 1/5 to 4/5 applied: 4640 ms (-49.2%) NUM_FILES=5000, after whole patchset applied: 4441 ms (-51.3%) NUM_FILES=2000, before patchset: 2528 ms NUM_FILES=2000, after patches 1/5 to 4/5 applied: 1983 ms (-21.6%) NUM_FILES=2000, after whole patchset applied: 1747 ms (-30.9%) NUM_FILES=1000, before patchset: 1085 ms NUM_FILES=1000, after patches 1/5 to 4/5 applied: 893 ms (-17.7%) NUM_FILES=1000, after whole patchset applied: 867 ms (-20.1%) Running dbench on the same physical machine with the following script: $ cat run-dbench.sh #!/bin/bash NUM_JOBS=$(nproc --all) DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 MOUNT_OPTIONS="-o ssd" MKFS_OPTIONS="-O no-holes -R free-space-tree" echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -D $MNT -t 120 $NUM_JOBS umount $MNT Before patchset: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 3761352 0.032 143.843 Close 2762770 0.002 2.273 Rename 159304 0.291 67.037 Unlink 759784 0.207 143.998 Deltree 72 4.028 15.977 Mkdir 36 0.003 0.006 Qpathinfo 3409780 0.013 9.678 Qfileinfo 596772 0.001 0.878 Qfsinfo 625189 0.003 1.245 Sfileinfo 306443 0.006 1.840 Find 1318106 0.063 19.798 WriteX 1871137 0.021 8.532 ReadX 5897325 0.003 3.567 LockX 12252 0.003 0.258 UnlockX 12252 0.002 0.100 Flush 263666 3.327 155.632 Throughput 980.047 MB/sec 12 clients 12 procs max_latency=155.636 ms After whole patchset applied: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 4195584 0.033 107.742 Close 3081932 0.002 1.935 Rename 177641 0.218 14.905 Unlink 847333 0.166 107.822 Deltree 118 5.315 15.247 Mkdir 59 0.004 0.048 Qpathinfo 3802612 0.014 10.302 Qfileinfo 666748 0.001 1.034 Qfsinfo 697329 0.003 0.944 Sfileinfo 341712 0.006 2.099 Find 1470365 0.065 9.359 WriteX 2093921 0.021 8.087 ReadX 6576234 0.003 3.407 LockX 13660 0.003 0.308 UnlockX 13660 0.002 0.114 Flush 294090 2.906 115.539 Throughput 1093.11 MB/sec 12 clients 12 procs max_latency=115.544 ms +11.5% throughput -25.8% max latency rename max latency -77.8% Link: https://bugzilla.opensuse.org/show_bug.cgi?id=1193549 Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-01-20 14:00:10 +03:00
ret = inode_logged(trans, inode, NULL);
if (ret == 0)
return;
btrfs: avoid inode logging during rename and link when possible During a rename or link operation, we need to determine if an inode was previously logged or not, and if it was, do some update to the logged inode. We used to rely exclusively on the logged_trans field of struct btrfs_inode to determine that, but that was not reliable because the value of that field is not persisted in the inode item, so it's lost when an inode is evicted and loaded back again. That led to several issues in the past, such as not persisting deletions (such as the case fixed by commit 803f0f64d17769 ("Btrfs: fix fsync not persisting dentry deletions due to inode evictions")), or resulting in losing a file after an inode eviction followed by a rename (commit ecc64fab7d49c6 ("btrfs: fix lost inode on log replay after mix of fsync, rename and inode eviction")), besides other issues. So the inode_logged() helper was introduced and used to determine if an inode was possibly logged before in the current transaction, with the caveat that it could return false positives, in the sense that even if an inode was not logged before in the current transaction, it could still return true, but never to return false in case the inode was logged. >From a functional point of view that is fine, but from a performance perspective it can introduce significant latencies to rename and link operations, as they will end up doing inode logging even when it is not necessary. Recently on a 5.15 kernel, an openSUSE Tumbleweed user reported package installations and upgrades, with the zypper tool, were often taking a long time to complete. With strace it could be observed that zypper was spending about 99% of its time on rename operations, and then with further analysis we checked that directory logging was happening too frequently. Taking into account that installation/upgrade of some of the packages needed a few thousand file renames, the slowdown was very noticeable for the user. The issue was caused indirectly due to an excessive number of inode evictions on a 5.15 kernel, about 100x more compared to a 5.13, 5.14 or a 5.16-rc8 kernel. While triggering the inode evictions if something outside btrfs' control, btrfs could still behave better by eliminating the false positives from the inode_logged() helper. So change inode_logged() to actually eliminate such false positives caused by inode eviction and when an inode was never logged since the filesystem was mounted, as both cases relate to when the logged_trans field of struct btrfs_inode has a value of zero. When it can not determine if the inode was logged based only on the logged_trans value, lookup for the existence of the inode item in the log tree - if it's there then we known the inode was logged, if it's not there then it can not have been logged in the current transaction. Once we determine if the inode was logged, update the logged_trans value to avoid future calls to have to search in the log tree again. Alternatively, we could start storing logged_trans in the on disk inode item structure (struct btrfs_inode_item) in the unused space it still has, but that would be a bit odd because: 1) We only care about logged_trans since the filesystem was mounted, we don't care about its value from a previous mount. Having it persisted in the inode item structure would not make the best use of the precious unused space; 2) In order to get logged_trans persisted before inode eviction, we would have to update the delayed inode when we finish logging the inode and update its logged_trans in struct btrfs_inode, which makes it a bit cumbersome since we need to check if the delayed inode exists, if not create it and populate it and deal with any errors (-ENOMEM mostly). This change is part of a patchset comprised of the following patches: 1/5 btrfs: add helper to delete a dir entry from a log tree 2/5 btrfs: pass the dentry to btrfs_log_new_name() instead of the inode 3/5 btrfs: avoid logging all directory changes during renames 4/5 btrfs: stop doing unnecessary log updates during a rename 5/5 btrfs: avoid inode logging during rename and link when possible The following test script mimics part of what the zypper tool does during package installations/upgrades. It does not triggers inode evictions, but it's similar because it triggers false positives from the inode_logged() helper, because the inodes have a logged_trans of 0, there's a log tree due to a fsync of an unrelated file and the directory inode has its last_trans field set to the current transaction: $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=10000 mkfs.btrfs -f $DEV mount $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync # Now do some change to an unrelated file and fsync it. # This is just to create a log tree to make sure that inode_logged() # does not return false when called against "testdir". xfs_io -f -c "pwrite 0 4K" -c "fsync" $MNT/foo # Do some change to testdir. This is to make sure inode_logged() # will return true when called against "testdir", because its # logged_trans is 0, it was changed in the current transaction # and there's a log tree. echo -n > $MNT/testdir/file_$((NUM_FILES + 1)) echo "Renaming $NUM_FILES files..." start=$(date +%s%N) for ((i = 1; i <= $NUM_FILES; i++)); do mv $MNT/testdir/file_$i $MNT/testdir/file_$i-RPMDELETE done end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "Renames took $dur milliseconds" umount $MNT Testing this change on a box using a non-debug kernel (Debian's default kernel config) gave the following results: NUM_FILES=10000, before patchset: 27837 ms NUM_FILES=10000, after patches 1/5 to 4/5 applied: 9236 ms (-66.8%) NUM_FILES=10000, after whole patchset applied: 8902 ms (-68.0%) NUM_FILES=5000, before patchset: 9127 ms NUM_FILES=5000, after patches 1/5 to 4/5 applied: 4640 ms (-49.2%) NUM_FILES=5000, after whole patchset applied: 4441 ms (-51.3%) NUM_FILES=2000, before patchset: 2528 ms NUM_FILES=2000, after patches 1/5 to 4/5 applied: 1983 ms (-21.6%) NUM_FILES=2000, after whole patchset applied: 1747 ms (-30.9%) NUM_FILES=1000, before patchset: 1085 ms NUM_FILES=1000, after patches 1/5 to 4/5 applied: 893 ms (-17.7%) NUM_FILES=1000, after whole patchset applied: 867 ms (-20.1%) Running dbench on the same physical machine with the following script: $ cat run-dbench.sh #!/bin/bash NUM_JOBS=$(nproc --all) DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 MOUNT_OPTIONS="-o ssd" MKFS_OPTIONS="-O no-holes -R free-space-tree" echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -D $MNT -t 120 $NUM_JOBS umount $MNT Before patchset: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 3761352 0.032 143.843 Close 2762770 0.002 2.273 Rename 159304 0.291 67.037 Unlink 759784 0.207 143.998 Deltree 72 4.028 15.977 Mkdir 36 0.003 0.006 Qpathinfo 3409780 0.013 9.678 Qfileinfo 596772 0.001 0.878 Qfsinfo 625189 0.003 1.245 Sfileinfo 306443 0.006 1.840 Find 1318106 0.063 19.798 WriteX 1871137 0.021 8.532 ReadX 5897325 0.003 3.567 LockX 12252 0.003 0.258 UnlockX 12252 0.002 0.100 Flush 263666 3.327 155.632 Throughput 980.047 MB/sec 12 clients 12 procs max_latency=155.636 ms After whole patchset applied: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 4195584 0.033 107.742 Close 3081932 0.002 1.935 Rename 177641 0.218 14.905 Unlink 847333 0.166 107.822 Deltree 118 5.315 15.247 Mkdir 59 0.004 0.048 Qpathinfo 3802612 0.014 10.302 Qfileinfo 666748 0.001 1.034 Qfsinfo 697329 0.003 0.944 Sfileinfo 341712 0.006 2.099 Find 1470365 0.065 9.359 WriteX 2093921 0.021 8.087 ReadX 6576234 0.003 3.407 LockX 13660 0.003 0.308 UnlockX 13660 0.002 0.114 Flush 294090 2.906 115.539 Throughput 1093.11 MB/sec 12 clients 12 procs max_latency=115.544 ms +11.5% throughput -25.8% max latency rename max latency -77.8% Link: https://bugzilla.opensuse.org/show_bug.cgi?id=1193549 Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-01-20 14:00:10 +03:00
else if (ret < 0) {
btrfs_set_log_full_commit(trans);
return;
}
ret = join_running_log_trans(root);
if (ret)
return;
log = root->log_root;
mutex_lock(&inode->log_mutex);
ret = btrfs_del_inode_ref(trans, log, name, name_len, btrfs_ino(inode),
dirid, &index);
mutex_unlock(&inode->log_mutex);
if (ret < 0 && ret != -ENOENT)
btrfs_set_log_full_commit(trans);
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 17:24:20 +03:00
btrfs_end_log_trans(root);
}
/*
* creates a range item in the log for 'dirid'. first_offset and
* last_offset tell us which parts of the key space the log should
* be considered authoritative for.
*/
static noinline int insert_dir_log_key(struct btrfs_trans_handle *trans,
struct btrfs_root *log,
struct btrfs_path *path,
btrfs: only copy dir index keys when logging a directory Currently, when logging a directory, we copy both dir items and dir index items from the fs/subvolume tree to the log tree. Both items have exactly the same data (same struct btrfs_dir_item), the difference lies in the key values, where a dir index key contains the index number of a directory entry while the dir item key does not, as it's used for doing fast lookups of an entry by name, while the former is used for sorting entries when listing a directory. We can exploit that and log only the dir index items, since they contain all the information needed to correctly add, replace and delete directory entries when replaying a log tree. Logging only the dir index items is also backward and forward compatible: an unpatched kernel (without this change) can correctly replay a log tree generated by a patched kernel (with this patch), and a patched kernel can correctly replay a log tree generated by an unpatched kernel. The backward compatibility is ensured because: 1) For inserting a new dentry: a dentry is only inserted when we find a new dir index key - we can only insert if we know the dir index offset, which is encoded in the dir index key's offset; 2) For deleting dentries: during log replay, before adding or replacing dentries, we first replay dentry deletions. Whenever we find a dir item key or a dir index key in the subvolume/fs tree that is not logged in a range for which the log tree is authoritative, we do the unlink of the dentry, which removes both the existing dir item key and the dir index key. Therefore logging just dir index keys is enough to ensure dentry deletions are correctly replayed; 3) For dentry replacements: they work when we log only dir index keys and this is mostly due to a combination of 1) and 2). If we replace a dentry with name "foobar" to point from inode A to inode B, then we know the dir index key for the new dentry is different from the old one, as it has an index number (key offset) larger than the old one. This results in replaying a deletion, through replay_dir_deletes(), that causes the old dentry to be removed, both the dir item key and the dir index key, as mentioned at 2). Then when processing the new dir index key, we add the new dentry, adding both a new dir item key and a new index key pointing to inode B, as stated in 1). The forward compatibility, the ability for a patched kernel to replay a log created by an older, unpatched kernel, comes from the changes required for making sure we are able to replay a log that only contains dir index keys - we simply ignore every dir item key we find. So modify directory logging to log only dir index items, and modify the log replay process to ignore dir item keys, from log trees created by an unpatched kernel, and process only with dir index keys. This reduces the amount of logged metadata by about half, and therefore the time spent logging or fsyncing large directories (less CPU time and less IO). The following test script was used to measure this change: #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=1000000 NUM_FILE_DELETES=10000 mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The tests were run on a physical machine, with a non-debug kernel (Debian's default kernel config), for different values of $NUM_NEW_FILES and $NUM_FILE_DELETES, and the results were the following: ** Before patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 8412 ms after adding 1000000 files dir fsync took 500 ms after deleting 10000 files ** After patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 4252 ms after adding 1000000 files (-49.5%) dir fsync took 269 ms after deleting 10000 files (-46.2%) ** Before patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 745 ms after adding 100000 files dir fsync took 59 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 404 ms after adding 100000 files (-45.8%) dir fsync took 31 ms after deleting 1000 files (-47.5%) ** Before patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 67 ms after adding 10000 files dir fsync took 9 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 36 ms after adding 10000 files (-46.3%) dir fsync took 5 ms after deleting 1000 files (-44.4%) ** Before patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 9 ms after adding 1000 files dir fsync took 4 ms after deleting 100 files ** After patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 7 ms after adding 1000 files (-22.2%) dir fsync took 3 ms after deleting 100 files (-25.0%) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-10-25 19:31:53 +03:00
u64 dirid,
u64 first_offset, u64 last_offset)
{
int ret;
struct btrfs_key key;
struct btrfs_dir_log_item *item;
key.objectid = dirid;
key.offset = first_offset;
btrfs: only copy dir index keys when logging a directory Currently, when logging a directory, we copy both dir items and dir index items from the fs/subvolume tree to the log tree. Both items have exactly the same data (same struct btrfs_dir_item), the difference lies in the key values, where a dir index key contains the index number of a directory entry while the dir item key does not, as it's used for doing fast lookups of an entry by name, while the former is used for sorting entries when listing a directory. We can exploit that and log only the dir index items, since they contain all the information needed to correctly add, replace and delete directory entries when replaying a log tree. Logging only the dir index items is also backward and forward compatible: an unpatched kernel (without this change) can correctly replay a log tree generated by a patched kernel (with this patch), and a patched kernel can correctly replay a log tree generated by an unpatched kernel. The backward compatibility is ensured because: 1) For inserting a new dentry: a dentry is only inserted when we find a new dir index key - we can only insert if we know the dir index offset, which is encoded in the dir index key's offset; 2) For deleting dentries: during log replay, before adding or replacing dentries, we first replay dentry deletions. Whenever we find a dir item key or a dir index key in the subvolume/fs tree that is not logged in a range for which the log tree is authoritative, we do the unlink of the dentry, which removes both the existing dir item key and the dir index key. Therefore logging just dir index keys is enough to ensure dentry deletions are correctly replayed; 3) For dentry replacements: they work when we log only dir index keys and this is mostly due to a combination of 1) and 2). If we replace a dentry with name "foobar" to point from inode A to inode B, then we know the dir index key for the new dentry is different from the old one, as it has an index number (key offset) larger than the old one. This results in replaying a deletion, through replay_dir_deletes(), that causes the old dentry to be removed, both the dir item key and the dir index key, as mentioned at 2). Then when processing the new dir index key, we add the new dentry, adding both a new dir item key and a new index key pointing to inode B, as stated in 1). The forward compatibility, the ability for a patched kernel to replay a log created by an older, unpatched kernel, comes from the changes required for making sure we are able to replay a log that only contains dir index keys - we simply ignore every dir item key we find. So modify directory logging to log only dir index items, and modify the log replay process to ignore dir item keys, from log trees created by an unpatched kernel, and process only with dir index keys. This reduces the amount of logged metadata by about half, and therefore the time spent logging or fsyncing large directories (less CPU time and less IO). The following test script was used to measure this change: #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=1000000 NUM_FILE_DELETES=10000 mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The tests were run on a physical machine, with a non-debug kernel (Debian's default kernel config), for different values of $NUM_NEW_FILES and $NUM_FILE_DELETES, and the results were the following: ** Before patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 8412 ms after adding 1000000 files dir fsync took 500 ms after deleting 10000 files ** After patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 4252 ms after adding 1000000 files (-49.5%) dir fsync took 269 ms after deleting 10000 files (-46.2%) ** Before patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 745 ms after adding 100000 files dir fsync took 59 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 404 ms after adding 100000 files (-45.8%) dir fsync took 31 ms after deleting 1000 files (-47.5%) ** Before patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 67 ms after adding 10000 files dir fsync took 9 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 36 ms after adding 10000 files (-46.3%) dir fsync took 5 ms after deleting 1000 files (-44.4%) ** Before patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 9 ms after adding 1000 files dir fsync took 4 ms after deleting 100 files ** After patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 7 ms after adding 1000 files (-22.2%) dir fsync took 3 ms after deleting 100 files (-25.0%) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-10-25 19:31:53 +03:00
key.type = BTRFS_DIR_LOG_INDEX_KEY;
ret = btrfs_insert_empty_item(trans, log, path, &key, sizeof(*item));
if (ret)
return ret;
item = btrfs_item_ptr(path->nodes[0], path->slots[0],
struct btrfs_dir_log_item);
btrfs_set_dir_log_end(path->nodes[0], item, last_offset);
btrfs_mark_buffer_dirty(path->nodes[0]);
btrfs_release_path(path);
return 0;
}
btrfs: insert items in batches when logging a directory when possible When logging a directory, we scan its directory items from the subvolume tree and then copy one by one into the log tree. This is not efficient since we generally are able to insert several items in a batch, using a single btree operation for adding several items at once. The reason we copy items one by one is that we must check if each item was previously logged in the current transaction, and if it was we either overwrite it or skip it in case its content did not change in the subvolume tree (this can happen only for dir item keys, but not for dir index keys), and doing such check makes it a bit cumbersome to attempt batch insertions. However the chances for doing batch insertions are very frequent and always happen when: 1) Logging the directory for the first time in the current transaction, as none of the items exist in the log tree yet; 2) Logging new dir index keys, because the offset for new dir index keys comes from a monotonically increasing counter. This means if we keep adding dentries to a directory, through creation of new files and sub-directories or by adding new links or renaming from some other directory into the one we are logging, all the new dir index keys have a new offset that is greater than the offset of any previously logged index keys, so we can insert them in batches into the log tree. For dir item keys, since their offset depends on the result of an hash function against the dentry's name, unless the directory is being logged for the first time in the current transaction, the chances being able to insert the items in the log using batches is pretty much random and not predictable, as it depends on the names of the dentries, but still happens often enough. So change directory logging to keep track of consecutive directory items that don't exist yet in the log and batch insert them. This patch is part of a patchset comprised of the following 5 patches: btrfs: remove root argument from btrfs_log_inode() and its callees btrfs: remove redundant log root assignment from log_dir_items() btrfs: factor out the copying loop of dir items from log_dir_items() btrfs: insert items in batches when logging a directory when possible btrfs: keep track of the last logged keys when logging a directory This is patch 4/5. The change log of the last patch (5/5) has performance results. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-16 13:32:13 +03:00
static int flush_dir_items_batch(struct btrfs_trans_handle *trans,
struct btrfs_root *log,
struct extent_buffer *src,
struct btrfs_path *dst_path,
int start_slot,
int count)
{
char *ins_data = NULL;
btrfs: loop only once over data sizes array when inserting an item batch When inserting a batch of items into a btree, we end up looping over the data sizes array 3 times: 1) Once in the caller of btrfs_insert_empty_items(), when it populates the array with the data sizes for each item; 2) Once at btrfs_insert_empty_items() to sum the elements of the data sizes array and compute the total data size; 3) And then once again at setup_items_for_insert(), where we do exactly the same as what we do at btrfs_insert_empty_items(), to compute the total data size. That is not bad for small arrays, but when the arrays have hundreds of elements, the time spent on looping is not negligible. For example when doing batch inserts of delayed items for dir index items or when logging a directory, it's common to have 200 to 260 dir index items in a single batch when using a leaf size of 16K and using file names between 8 and 12 characters. For a 64K leaf size, multiply that by 4. Taking into account that during directory logging or when flushing delayed dir index items we can have many of those large batches, the time spent on the looping adds up quickly. It's also more important to avoid it at setup_items_for_insert(), since we are holding a write lock on a leaf and, in some cases, on upper nodes of the btree, which causes us to block other tasks that want to access the leaf and nodes for longer than necessary. So change the code so that setup_items_for_insert() and btrfs_insert_empty_items() no longer compute the total data size, and instead rely on the caller to supply it. This makes us loop over the array only once, where we can both populate the data size array and compute the total data size, taking advantage of spatial and temporal locality. To make this more manageable, use a structure to contain all the relevant details for a batch of items (keys array, data sizes array, total data size, number of items), and use it as an argument for btrfs_insert_empty_items() and setup_items_for_insert(). This patch is part of a small patchset that is comprised of the following patches: btrfs: loop only once over data sizes array when inserting an item batch btrfs: unexport setup_items_for_insert() btrfs: use single bulk copy operations when logging directories This is patch 1/3 and performance results, and the specific tests, are included in the changelog of patch 3/3. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-24 14:28:13 +03:00
struct btrfs_item_batch batch;
btrfs: insert items in batches when logging a directory when possible When logging a directory, we scan its directory items from the subvolume tree and then copy one by one into the log tree. This is not efficient since we generally are able to insert several items in a batch, using a single btree operation for adding several items at once. The reason we copy items one by one is that we must check if each item was previously logged in the current transaction, and if it was we either overwrite it or skip it in case its content did not change in the subvolume tree (this can happen only for dir item keys, but not for dir index keys), and doing such check makes it a bit cumbersome to attempt batch insertions. However the chances for doing batch insertions are very frequent and always happen when: 1) Logging the directory for the first time in the current transaction, as none of the items exist in the log tree yet; 2) Logging new dir index keys, because the offset for new dir index keys comes from a monotonically increasing counter. This means if we keep adding dentries to a directory, through creation of new files and sub-directories or by adding new links or renaming from some other directory into the one we are logging, all the new dir index keys have a new offset that is greater than the offset of any previously logged index keys, so we can insert them in batches into the log tree. For dir item keys, since their offset depends on the result of an hash function against the dentry's name, unless the directory is being logged for the first time in the current transaction, the chances being able to insert the items in the log using batches is pretty much random and not predictable, as it depends on the names of the dentries, but still happens often enough. So change directory logging to keep track of consecutive directory items that don't exist yet in the log and batch insert them. This patch is part of a patchset comprised of the following 5 patches: btrfs: remove root argument from btrfs_log_inode() and its callees btrfs: remove redundant log root assignment from log_dir_items() btrfs: factor out the copying loop of dir items from log_dir_items() btrfs: insert items in batches when logging a directory when possible btrfs: keep track of the last logged keys when logging a directory This is patch 4/5. The change log of the last patch (5/5) has performance results. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-16 13:32:13 +03:00
struct extent_buffer *dst;
btrfs: use single bulk copy operations when logging directories When logging a directory and inserting a batch of directory items, we are copying the data of each item from a leaf in the fs/subvolume tree to a leaf in a log tree, separately. This is not really needed, since we are copying from a contiguous memory area into another one, so we can use a single copy operation to copy all items at once. This patch is part of a small patchset that is comprised of the following patches: btrfs: loop only once over data sizes array when inserting an item batch btrfs: unexport setup_items_for_insert() btrfs: use single bulk copy operations when logging directories This is patch 3/3. The following test was used to compare performance of a branch without the patchset versus one branch that has the whole patchset applied: $ cat dir-fsync-test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=1000000 NUM_FILE_DELETES=1000 LEAF_SIZE=16K mkfs.btrfs -f -n $LEAF_SIZE $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done # Fsync the directory, this will log the new dir items and the inodes # they point to, because these are new inodes. start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done # Fsync the directory, this will only log dir items, there are no # dentries pointing to new inodes. start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" umount $MNT The tests were run on a non-debug kernel (Debian's default kernel config) and were the following: *** with a leaf size of 16K, before patchset *** dir fsync took 8482 ms after adding 1000000 files dir fsync took 166 ms after deleting 1000 files *** with a leaf size of 16K, after patchset *** dir fsync took 8196 ms after adding 1000000 files (-3.4%) dir fsync took 143 ms after deleting 1000 files (-14.9%) *** with a leaf size of 64K, before patchset *** dir fsync took 12851 ms after adding 1000000 files dir fsync took 466 ms after deleting 1000 files *** with a leaf size of 64K, after patchset *** dir fsync took 12287 ms after adding 1000000 files (-4.5%) dir fsync took 414 ms after deleting 1000 files (-11.8%) Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-24 14:28:15 +03:00
unsigned long src_offset;
unsigned long dst_offset;
btrfs: insert items in batches when logging a directory when possible When logging a directory, we scan its directory items from the subvolume tree and then copy one by one into the log tree. This is not efficient since we generally are able to insert several items in a batch, using a single btree operation for adding several items at once. The reason we copy items one by one is that we must check if each item was previously logged in the current transaction, and if it was we either overwrite it or skip it in case its content did not change in the subvolume tree (this can happen only for dir item keys, but not for dir index keys), and doing such check makes it a bit cumbersome to attempt batch insertions. However the chances for doing batch insertions are very frequent and always happen when: 1) Logging the directory for the first time in the current transaction, as none of the items exist in the log tree yet; 2) Logging new dir index keys, because the offset for new dir index keys comes from a monotonically increasing counter. This means if we keep adding dentries to a directory, through creation of new files and sub-directories or by adding new links or renaming from some other directory into the one we are logging, all the new dir index keys have a new offset that is greater than the offset of any previously logged index keys, so we can insert them in batches into the log tree. For dir item keys, since their offset depends on the result of an hash function against the dentry's name, unless the directory is being logged for the first time in the current transaction, the chances being able to insert the items in the log using batches is pretty much random and not predictable, as it depends on the names of the dentries, but still happens often enough. So change directory logging to keep track of consecutive directory items that don't exist yet in the log and batch insert them. This patch is part of a patchset comprised of the following 5 patches: btrfs: remove root argument from btrfs_log_inode() and its callees btrfs: remove redundant log root assignment from log_dir_items() btrfs: factor out the copying loop of dir items from log_dir_items() btrfs: insert items in batches when logging a directory when possible btrfs: keep track of the last logged keys when logging a directory This is patch 4/5. The change log of the last patch (5/5) has performance results. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-16 13:32:13 +03:00
struct btrfs_key key;
u32 item_size;
int ret;
int i;
ASSERT(count > 0);
btrfs: loop only once over data sizes array when inserting an item batch When inserting a batch of items into a btree, we end up looping over the data sizes array 3 times: 1) Once in the caller of btrfs_insert_empty_items(), when it populates the array with the data sizes for each item; 2) Once at btrfs_insert_empty_items() to sum the elements of the data sizes array and compute the total data size; 3) And then once again at setup_items_for_insert(), where we do exactly the same as what we do at btrfs_insert_empty_items(), to compute the total data size. That is not bad for small arrays, but when the arrays have hundreds of elements, the time spent on looping is not negligible. For example when doing batch inserts of delayed items for dir index items or when logging a directory, it's common to have 200 to 260 dir index items in a single batch when using a leaf size of 16K and using file names between 8 and 12 characters. For a 64K leaf size, multiply that by 4. Taking into account that during directory logging or when flushing delayed dir index items we can have many of those large batches, the time spent on the looping adds up quickly. It's also more important to avoid it at setup_items_for_insert(), since we are holding a write lock on a leaf and, in some cases, on upper nodes of the btree, which causes us to block other tasks that want to access the leaf and nodes for longer than necessary. So change the code so that setup_items_for_insert() and btrfs_insert_empty_items() no longer compute the total data size, and instead rely on the caller to supply it. This makes us loop over the array only once, where we can both populate the data size array and compute the total data size, taking advantage of spatial and temporal locality. To make this more manageable, use a structure to contain all the relevant details for a batch of items (keys array, data sizes array, total data size, number of items), and use it as an argument for btrfs_insert_empty_items() and setup_items_for_insert(). This patch is part of a small patchset that is comprised of the following patches: btrfs: loop only once over data sizes array when inserting an item batch btrfs: unexport setup_items_for_insert() btrfs: use single bulk copy operations when logging directories This is patch 1/3 and performance results, and the specific tests, are included in the changelog of patch 3/3. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-24 14:28:13 +03:00
batch.nr = count;
btrfs: insert items in batches when logging a directory when possible When logging a directory, we scan its directory items from the subvolume tree and then copy one by one into the log tree. This is not efficient since we generally are able to insert several items in a batch, using a single btree operation for adding several items at once. The reason we copy items one by one is that we must check if each item was previously logged in the current transaction, and if it was we either overwrite it or skip it in case its content did not change in the subvolume tree (this can happen only for dir item keys, but not for dir index keys), and doing such check makes it a bit cumbersome to attempt batch insertions. However the chances for doing batch insertions are very frequent and always happen when: 1) Logging the directory for the first time in the current transaction, as none of the items exist in the log tree yet; 2) Logging new dir index keys, because the offset for new dir index keys comes from a monotonically increasing counter. This means if we keep adding dentries to a directory, through creation of new files and sub-directories or by adding new links or renaming from some other directory into the one we are logging, all the new dir index keys have a new offset that is greater than the offset of any previously logged index keys, so we can insert them in batches into the log tree. For dir item keys, since their offset depends on the result of an hash function against the dentry's name, unless the directory is being logged for the first time in the current transaction, the chances being able to insert the items in the log using batches is pretty much random and not predictable, as it depends on the names of the dentries, but still happens often enough. So change directory logging to keep track of consecutive directory items that don't exist yet in the log and batch insert them. This patch is part of a patchset comprised of the following 5 patches: btrfs: remove root argument from btrfs_log_inode() and its callees btrfs: remove redundant log root assignment from log_dir_items() btrfs: factor out the copying loop of dir items from log_dir_items() btrfs: insert items in batches when logging a directory when possible btrfs: keep track of the last logged keys when logging a directory This is patch 4/5. The change log of the last patch (5/5) has performance results. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-16 13:32:13 +03:00
if (count == 1) {
btrfs_item_key_to_cpu(src, &key, start_slot);
item_size = btrfs_item_size(src, start_slot);
btrfs: loop only once over data sizes array when inserting an item batch When inserting a batch of items into a btree, we end up looping over the data sizes array 3 times: 1) Once in the caller of btrfs_insert_empty_items(), when it populates the array with the data sizes for each item; 2) Once at btrfs_insert_empty_items() to sum the elements of the data sizes array and compute the total data size; 3) And then once again at setup_items_for_insert(), where we do exactly the same as what we do at btrfs_insert_empty_items(), to compute the total data size. That is not bad for small arrays, but when the arrays have hundreds of elements, the time spent on looping is not negligible. For example when doing batch inserts of delayed items for dir index items or when logging a directory, it's common to have 200 to 260 dir index items in a single batch when using a leaf size of 16K and using file names between 8 and 12 characters. For a 64K leaf size, multiply that by 4. Taking into account that during directory logging or when flushing delayed dir index items we can have many of those large batches, the time spent on the looping adds up quickly. It's also more important to avoid it at setup_items_for_insert(), since we are holding a write lock on a leaf and, in some cases, on upper nodes of the btree, which causes us to block other tasks that want to access the leaf and nodes for longer than necessary. So change the code so that setup_items_for_insert() and btrfs_insert_empty_items() no longer compute the total data size, and instead rely on the caller to supply it. This makes us loop over the array only once, where we can both populate the data size array and compute the total data size, taking advantage of spatial and temporal locality. To make this more manageable, use a structure to contain all the relevant details for a batch of items (keys array, data sizes array, total data size, number of items), and use it as an argument for btrfs_insert_empty_items() and setup_items_for_insert(). This patch is part of a small patchset that is comprised of the following patches: btrfs: loop only once over data sizes array when inserting an item batch btrfs: unexport setup_items_for_insert() btrfs: use single bulk copy operations when logging directories This is patch 1/3 and performance results, and the specific tests, are included in the changelog of patch 3/3. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-24 14:28:13 +03:00
batch.keys = &key;
batch.data_sizes = &item_size;
batch.total_data_size = item_size;
btrfs: insert items in batches when logging a directory when possible When logging a directory, we scan its directory items from the subvolume tree and then copy one by one into the log tree. This is not efficient since we generally are able to insert several items in a batch, using a single btree operation for adding several items at once. The reason we copy items one by one is that we must check if each item was previously logged in the current transaction, and if it was we either overwrite it or skip it in case its content did not change in the subvolume tree (this can happen only for dir item keys, but not for dir index keys), and doing such check makes it a bit cumbersome to attempt batch insertions. However the chances for doing batch insertions are very frequent and always happen when: 1) Logging the directory for the first time in the current transaction, as none of the items exist in the log tree yet; 2) Logging new dir index keys, because the offset for new dir index keys comes from a monotonically increasing counter. This means if we keep adding dentries to a directory, through creation of new files and sub-directories or by adding new links or renaming from some other directory into the one we are logging, all the new dir index keys have a new offset that is greater than the offset of any previously logged index keys, so we can insert them in batches into the log tree. For dir item keys, since their offset depends on the result of an hash function against the dentry's name, unless the directory is being logged for the first time in the current transaction, the chances being able to insert the items in the log using batches is pretty much random and not predictable, as it depends on the names of the dentries, but still happens often enough. So change directory logging to keep track of consecutive directory items that don't exist yet in the log and batch insert them. This patch is part of a patchset comprised of the following 5 patches: btrfs: remove root argument from btrfs_log_inode() and its callees btrfs: remove redundant log root assignment from log_dir_items() btrfs: factor out the copying loop of dir items from log_dir_items() btrfs: insert items in batches when logging a directory when possible btrfs: keep track of the last logged keys when logging a directory This is patch 4/5. The change log of the last patch (5/5) has performance results. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-16 13:32:13 +03:00
} else {
btrfs: loop only once over data sizes array when inserting an item batch When inserting a batch of items into a btree, we end up looping over the data sizes array 3 times: 1) Once in the caller of btrfs_insert_empty_items(), when it populates the array with the data sizes for each item; 2) Once at btrfs_insert_empty_items() to sum the elements of the data sizes array and compute the total data size; 3) And then once again at setup_items_for_insert(), where we do exactly the same as what we do at btrfs_insert_empty_items(), to compute the total data size. That is not bad for small arrays, but when the arrays have hundreds of elements, the time spent on looping is not negligible. For example when doing batch inserts of delayed items for dir index items or when logging a directory, it's common to have 200 to 260 dir index items in a single batch when using a leaf size of 16K and using file names between 8 and 12 characters. For a 64K leaf size, multiply that by 4. Taking into account that during directory logging or when flushing delayed dir index items we can have many of those large batches, the time spent on the looping adds up quickly. It's also more important to avoid it at setup_items_for_insert(), since we are holding a write lock on a leaf and, in some cases, on upper nodes of the btree, which causes us to block other tasks that want to access the leaf and nodes for longer than necessary. So change the code so that setup_items_for_insert() and btrfs_insert_empty_items() no longer compute the total data size, and instead rely on the caller to supply it. This makes us loop over the array only once, where we can both populate the data size array and compute the total data size, taking advantage of spatial and temporal locality. To make this more manageable, use a structure to contain all the relevant details for a batch of items (keys array, data sizes array, total data size, number of items), and use it as an argument for btrfs_insert_empty_items() and setup_items_for_insert(). This patch is part of a small patchset that is comprised of the following patches: btrfs: loop only once over data sizes array when inserting an item batch btrfs: unexport setup_items_for_insert() btrfs: use single bulk copy operations when logging directories This is patch 1/3 and performance results, and the specific tests, are included in the changelog of patch 3/3. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-24 14:28:13 +03:00
struct btrfs_key *ins_keys;
u32 *ins_sizes;
btrfs: insert items in batches when logging a directory when possible When logging a directory, we scan its directory items from the subvolume tree and then copy one by one into the log tree. This is not efficient since we generally are able to insert several items in a batch, using a single btree operation for adding several items at once. The reason we copy items one by one is that we must check if each item was previously logged in the current transaction, and if it was we either overwrite it or skip it in case its content did not change in the subvolume tree (this can happen only for dir item keys, but not for dir index keys), and doing such check makes it a bit cumbersome to attempt batch insertions. However the chances for doing batch insertions are very frequent and always happen when: 1) Logging the directory for the first time in the current transaction, as none of the items exist in the log tree yet; 2) Logging new dir index keys, because the offset for new dir index keys comes from a monotonically increasing counter. This means if we keep adding dentries to a directory, through creation of new files and sub-directories or by adding new links or renaming from some other directory into the one we are logging, all the new dir index keys have a new offset that is greater than the offset of any previously logged index keys, so we can insert them in batches into the log tree. For dir item keys, since their offset depends on the result of an hash function against the dentry's name, unless the directory is being logged for the first time in the current transaction, the chances being able to insert the items in the log using batches is pretty much random and not predictable, as it depends on the names of the dentries, but still happens often enough. So change directory logging to keep track of consecutive directory items that don't exist yet in the log and batch insert them. This patch is part of a patchset comprised of the following 5 patches: btrfs: remove root argument from btrfs_log_inode() and its callees btrfs: remove redundant log root assignment from log_dir_items() btrfs: factor out the copying loop of dir items from log_dir_items() btrfs: insert items in batches when logging a directory when possible btrfs: keep track of the last logged keys when logging a directory This is patch 4/5. The change log of the last patch (5/5) has performance results. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-16 13:32:13 +03:00
ins_data = kmalloc(count * sizeof(u32) +
count * sizeof(struct btrfs_key), GFP_NOFS);
if (!ins_data)
return -ENOMEM;
ins_sizes = (u32 *)ins_data;
ins_keys = (struct btrfs_key *)(ins_data + count * sizeof(u32));
btrfs: loop only once over data sizes array when inserting an item batch When inserting a batch of items into a btree, we end up looping over the data sizes array 3 times: 1) Once in the caller of btrfs_insert_empty_items(), when it populates the array with the data sizes for each item; 2) Once at btrfs_insert_empty_items() to sum the elements of the data sizes array and compute the total data size; 3) And then once again at setup_items_for_insert(), where we do exactly the same as what we do at btrfs_insert_empty_items(), to compute the total data size. That is not bad for small arrays, but when the arrays have hundreds of elements, the time spent on looping is not negligible. For example when doing batch inserts of delayed items for dir index items or when logging a directory, it's common to have 200 to 260 dir index items in a single batch when using a leaf size of 16K and using file names between 8 and 12 characters. For a 64K leaf size, multiply that by 4. Taking into account that during directory logging or when flushing delayed dir index items we can have many of those large batches, the time spent on the looping adds up quickly. It's also more important to avoid it at setup_items_for_insert(), since we are holding a write lock on a leaf and, in some cases, on upper nodes of the btree, which causes us to block other tasks that want to access the leaf and nodes for longer than necessary. So change the code so that setup_items_for_insert() and btrfs_insert_empty_items() no longer compute the total data size, and instead rely on the caller to supply it. This makes us loop over the array only once, where we can both populate the data size array and compute the total data size, taking advantage of spatial and temporal locality. To make this more manageable, use a structure to contain all the relevant details for a batch of items (keys array, data sizes array, total data size, number of items), and use it as an argument for btrfs_insert_empty_items() and setup_items_for_insert(). This patch is part of a small patchset that is comprised of the following patches: btrfs: loop only once over data sizes array when inserting an item batch btrfs: unexport setup_items_for_insert() btrfs: use single bulk copy operations when logging directories This is patch 1/3 and performance results, and the specific tests, are included in the changelog of patch 3/3. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-24 14:28:13 +03:00
batch.keys = ins_keys;
batch.data_sizes = ins_sizes;
batch.total_data_size = 0;
btrfs: insert items in batches when logging a directory when possible When logging a directory, we scan its directory items from the subvolume tree and then copy one by one into the log tree. This is not efficient since we generally are able to insert several items in a batch, using a single btree operation for adding several items at once. The reason we copy items one by one is that we must check if each item was previously logged in the current transaction, and if it was we either overwrite it or skip it in case its content did not change in the subvolume tree (this can happen only for dir item keys, but not for dir index keys), and doing such check makes it a bit cumbersome to attempt batch insertions. However the chances for doing batch insertions are very frequent and always happen when: 1) Logging the directory for the first time in the current transaction, as none of the items exist in the log tree yet; 2) Logging new dir index keys, because the offset for new dir index keys comes from a monotonically increasing counter. This means if we keep adding dentries to a directory, through creation of new files and sub-directories or by adding new links or renaming from some other directory into the one we are logging, all the new dir index keys have a new offset that is greater than the offset of any previously logged index keys, so we can insert them in batches into the log tree. For dir item keys, since their offset depends on the result of an hash function against the dentry's name, unless the directory is being logged for the first time in the current transaction, the chances being able to insert the items in the log using batches is pretty much random and not predictable, as it depends on the names of the dentries, but still happens often enough. So change directory logging to keep track of consecutive directory items that don't exist yet in the log and batch insert them. This patch is part of a patchset comprised of the following 5 patches: btrfs: remove root argument from btrfs_log_inode() and its callees btrfs: remove redundant log root assignment from log_dir_items() btrfs: factor out the copying loop of dir items from log_dir_items() btrfs: insert items in batches when logging a directory when possible btrfs: keep track of the last logged keys when logging a directory This is patch 4/5. The change log of the last patch (5/5) has performance results. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-16 13:32:13 +03:00
for (i = 0; i < count; i++) {
const int slot = start_slot + i;
btrfs_item_key_to_cpu(src, &ins_keys[i], slot);
ins_sizes[i] = btrfs_item_size(src, slot);
btrfs: loop only once over data sizes array when inserting an item batch When inserting a batch of items into a btree, we end up looping over the data sizes array 3 times: 1) Once in the caller of btrfs_insert_empty_items(), when it populates the array with the data sizes for each item; 2) Once at btrfs_insert_empty_items() to sum the elements of the data sizes array and compute the total data size; 3) And then once again at setup_items_for_insert(), where we do exactly the same as what we do at btrfs_insert_empty_items(), to compute the total data size. That is not bad for small arrays, but when the arrays have hundreds of elements, the time spent on looping is not negligible. For example when doing batch inserts of delayed items for dir index items or when logging a directory, it's common to have 200 to 260 dir index items in a single batch when using a leaf size of 16K and using file names between 8 and 12 characters. For a 64K leaf size, multiply that by 4. Taking into account that during directory logging or when flushing delayed dir index items we can have many of those large batches, the time spent on the looping adds up quickly. It's also more important to avoid it at setup_items_for_insert(), since we are holding a write lock on a leaf and, in some cases, on upper nodes of the btree, which causes us to block other tasks that want to access the leaf and nodes for longer than necessary. So change the code so that setup_items_for_insert() and btrfs_insert_empty_items() no longer compute the total data size, and instead rely on the caller to supply it. This makes us loop over the array only once, where we can both populate the data size array and compute the total data size, taking advantage of spatial and temporal locality. To make this more manageable, use a structure to contain all the relevant details for a batch of items (keys array, data sizes array, total data size, number of items), and use it as an argument for btrfs_insert_empty_items() and setup_items_for_insert(). This patch is part of a small patchset that is comprised of the following patches: btrfs: loop only once over data sizes array when inserting an item batch btrfs: unexport setup_items_for_insert() btrfs: use single bulk copy operations when logging directories This is patch 1/3 and performance results, and the specific tests, are included in the changelog of patch 3/3. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-24 14:28:13 +03:00
batch.total_data_size += ins_sizes[i];
btrfs: insert items in batches when logging a directory when possible When logging a directory, we scan its directory items from the subvolume tree and then copy one by one into the log tree. This is not efficient since we generally are able to insert several items in a batch, using a single btree operation for adding several items at once. The reason we copy items one by one is that we must check if each item was previously logged in the current transaction, and if it was we either overwrite it or skip it in case its content did not change in the subvolume tree (this can happen only for dir item keys, but not for dir index keys), and doing such check makes it a bit cumbersome to attempt batch insertions. However the chances for doing batch insertions are very frequent and always happen when: 1) Logging the directory for the first time in the current transaction, as none of the items exist in the log tree yet; 2) Logging new dir index keys, because the offset for new dir index keys comes from a monotonically increasing counter. This means if we keep adding dentries to a directory, through creation of new files and sub-directories or by adding new links or renaming from some other directory into the one we are logging, all the new dir index keys have a new offset that is greater than the offset of any previously logged index keys, so we can insert them in batches into the log tree. For dir item keys, since their offset depends on the result of an hash function against the dentry's name, unless the directory is being logged for the first time in the current transaction, the chances being able to insert the items in the log using batches is pretty much random and not predictable, as it depends on the names of the dentries, but still happens often enough. So change directory logging to keep track of consecutive directory items that don't exist yet in the log and batch insert them. This patch is part of a patchset comprised of the following 5 patches: btrfs: remove root argument from btrfs_log_inode() and its callees btrfs: remove redundant log root assignment from log_dir_items() btrfs: factor out the copying loop of dir items from log_dir_items() btrfs: insert items in batches when logging a directory when possible btrfs: keep track of the last logged keys when logging a directory This is patch 4/5. The change log of the last patch (5/5) has performance results. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-16 13:32:13 +03:00
}
}
btrfs: loop only once over data sizes array when inserting an item batch When inserting a batch of items into a btree, we end up looping over the data sizes array 3 times: 1) Once in the caller of btrfs_insert_empty_items(), when it populates the array with the data sizes for each item; 2) Once at btrfs_insert_empty_items() to sum the elements of the data sizes array and compute the total data size; 3) And then once again at setup_items_for_insert(), where we do exactly the same as what we do at btrfs_insert_empty_items(), to compute the total data size. That is not bad for small arrays, but when the arrays have hundreds of elements, the time spent on looping is not negligible. For example when doing batch inserts of delayed items for dir index items or when logging a directory, it's common to have 200 to 260 dir index items in a single batch when using a leaf size of 16K and using file names between 8 and 12 characters. For a 64K leaf size, multiply that by 4. Taking into account that during directory logging or when flushing delayed dir index items we can have many of those large batches, the time spent on the looping adds up quickly. It's also more important to avoid it at setup_items_for_insert(), since we are holding a write lock on a leaf and, in some cases, on upper nodes of the btree, which causes us to block other tasks that want to access the leaf and nodes for longer than necessary. So change the code so that setup_items_for_insert() and btrfs_insert_empty_items() no longer compute the total data size, and instead rely on the caller to supply it. This makes us loop over the array only once, where we can both populate the data size array and compute the total data size, taking advantage of spatial and temporal locality. To make this more manageable, use a structure to contain all the relevant details for a batch of items (keys array, data sizes array, total data size, number of items), and use it as an argument for btrfs_insert_empty_items() and setup_items_for_insert(). This patch is part of a small patchset that is comprised of the following patches: btrfs: loop only once over data sizes array when inserting an item batch btrfs: unexport setup_items_for_insert() btrfs: use single bulk copy operations when logging directories This is patch 1/3 and performance results, and the specific tests, are included in the changelog of patch 3/3. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-24 14:28:13 +03:00
ret = btrfs_insert_empty_items(trans, log, dst_path, &batch);
btrfs: insert items in batches when logging a directory when possible When logging a directory, we scan its directory items from the subvolume tree and then copy one by one into the log tree. This is not efficient since we generally are able to insert several items in a batch, using a single btree operation for adding several items at once. The reason we copy items one by one is that we must check if each item was previously logged in the current transaction, and if it was we either overwrite it or skip it in case its content did not change in the subvolume tree (this can happen only for dir item keys, but not for dir index keys), and doing such check makes it a bit cumbersome to attempt batch insertions. However the chances for doing batch insertions are very frequent and always happen when: 1) Logging the directory for the first time in the current transaction, as none of the items exist in the log tree yet; 2) Logging new dir index keys, because the offset for new dir index keys comes from a monotonically increasing counter. This means if we keep adding dentries to a directory, through creation of new files and sub-directories or by adding new links or renaming from some other directory into the one we are logging, all the new dir index keys have a new offset that is greater than the offset of any previously logged index keys, so we can insert them in batches into the log tree. For dir item keys, since their offset depends on the result of an hash function against the dentry's name, unless the directory is being logged for the first time in the current transaction, the chances being able to insert the items in the log using batches is pretty much random and not predictable, as it depends on the names of the dentries, but still happens often enough. So change directory logging to keep track of consecutive directory items that don't exist yet in the log and batch insert them. This patch is part of a patchset comprised of the following 5 patches: btrfs: remove root argument from btrfs_log_inode() and its callees btrfs: remove redundant log root assignment from log_dir_items() btrfs: factor out the copying loop of dir items from log_dir_items() btrfs: insert items in batches when logging a directory when possible btrfs: keep track of the last logged keys when logging a directory This is patch 4/5. The change log of the last patch (5/5) has performance results. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-16 13:32:13 +03:00
if (ret)
goto out;
dst = dst_path->nodes[0];
btrfs: use single bulk copy operations when logging directories When logging a directory and inserting a batch of directory items, we are copying the data of each item from a leaf in the fs/subvolume tree to a leaf in a log tree, separately. This is not really needed, since we are copying from a contiguous memory area into another one, so we can use a single copy operation to copy all items at once. This patch is part of a small patchset that is comprised of the following patches: btrfs: loop only once over data sizes array when inserting an item batch btrfs: unexport setup_items_for_insert() btrfs: use single bulk copy operations when logging directories This is patch 3/3. The following test was used to compare performance of a branch without the patchset versus one branch that has the whole patchset applied: $ cat dir-fsync-test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=1000000 NUM_FILE_DELETES=1000 LEAF_SIZE=16K mkfs.btrfs -f -n $LEAF_SIZE $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done # Fsync the directory, this will log the new dir items and the inodes # they point to, because these are new inodes. start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done # Fsync the directory, this will only log dir items, there are no # dentries pointing to new inodes. start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" umount $MNT The tests were run on a non-debug kernel (Debian's default kernel config) and were the following: *** with a leaf size of 16K, before patchset *** dir fsync took 8482 ms after adding 1000000 files dir fsync took 166 ms after deleting 1000 files *** with a leaf size of 16K, after patchset *** dir fsync took 8196 ms after adding 1000000 files (-3.4%) dir fsync took 143 ms after deleting 1000 files (-14.9%) *** with a leaf size of 64K, before patchset *** dir fsync took 12851 ms after adding 1000000 files dir fsync took 466 ms after deleting 1000 files *** with a leaf size of 64K, after patchset *** dir fsync took 12287 ms after adding 1000000 files (-4.5%) dir fsync took 414 ms after deleting 1000 files (-11.8%) Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-24 14:28:15 +03:00
/*
* Copy all the items in bulk, in a single copy operation. Item data is
* organized such that it's placed at the end of a leaf and from right
* to left. For example, the data for the second item ends at an offset
* that matches the offset where the data for the first item starts, the
* data for the third item ends at an offset that matches the offset
* where the data of the second items starts, and so on.
* Therefore our source and destination start offsets for copy match the
* offsets of the last items (highest slots).
*/
dst_offset = btrfs_item_ptr_offset(dst, dst_path->slots[0] + count - 1);
src_offset = btrfs_item_ptr_offset(src, start_slot + count - 1);
copy_extent_buffer(dst, src, dst_offset, src_offset, batch.total_data_size);
btrfs: insert items in batches when logging a directory when possible When logging a directory, we scan its directory items from the subvolume tree and then copy one by one into the log tree. This is not efficient since we generally are able to insert several items in a batch, using a single btree operation for adding several items at once. The reason we copy items one by one is that we must check if each item was previously logged in the current transaction, and if it was we either overwrite it or skip it in case its content did not change in the subvolume tree (this can happen only for dir item keys, but not for dir index keys), and doing such check makes it a bit cumbersome to attempt batch insertions. However the chances for doing batch insertions are very frequent and always happen when: 1) Logging the directory for the first time in the current transaction, as none of the items exist in the log tree yet; 2) Logging new dir index keys, because the offset for new dir index keys comes from a monotonically increasing counter. This means if we keep adding dentries to a directory, through creation of new files and sub-directories or by adding new links or renaming from some other directory into the one we are logging, all the new dir index keys have a new offset that is greater than the offset of any previously logged index keys, so we can insert them in batches into the log tree. For dir item keys, since their offset depends on the result of an hash function against the dentry's name, unless the directory is being logged for the first time in the current transaction, the chances being able to insert the items in the log using batches is pretty much random and not predictable, as it depends on the names of the dentries, but still happens often enough. So change directory logging to keep track of consecutive directory items that don't exist yet in the log and batch insert them. This patch is part of a patchset comprised of the following 5 patches: btrfs: remove root argument from btrfs_log_inode() and its callees btrfs: remove redundant log root assignment from log_dir_items() btrfs: factor out the copying loop of dir items from log_dir_items() btrfs: insert items in batches when logging a directory when possible btrfs: keep track of the last logged keys when logging a directory This is patch 4/5. The change log of the last patch (5/5) has performance results. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-16 13:32:13 +03:00
btrfs_release_path(dst_path);
out:
kfree(ins_data);
return ret;
}
static int process_dir_items_leaf(struct btrfs_trans_handle *trans,
struct btrfs_inode *inode,
struct btrfs_path *path,
struct btrfs_path *dst_path,
btrfs: stop copying old dir items when logging a directory When logging a directory, we go over every leaf of the subvolume tree that was changed in the current transaction and copy all its dir index keys to the log tree. That includes copying dir index keys created in past transactions. This is done mostly for simplicity, as after logging the keys we log an item that specifies the start and end ranges of the keys we logged. That item is then used during log replay to figure out which keys need to be deleted - every key in that range that we find in the subvolume tree and is not in the log tree, needs to be deleted. Now that we log only dir index keys, and not dir item keys anymore, when we remove dentries from a directory (due to unlink and rename operations), we can get entire leaves that we changed only for deleting old dir index keys, or that have few dir index keys that are new - this is due to the fact that the offset for new index keys comes from a monotonically increasing counter. We can avoid logging dir index keys from past transactions, and in order to track the deletions, only log range items (BTRFS_DIR_LOG_INDEX_KEY key type) when we find gaps between consecutive index keys. This massively reduces the amount of logged metadata when we have deleted directory entries, even if it's a small percentage of the total number of entries. The reduction comes from both less items that are logged and instead of logging many dir index items (struct btrfs_dir_item), which have a size of 30 bytes plus a file name, we typically log just a few range items (struct btrfs_dir_log_item), which take only 8 bytes each. Even if no entries were deleted from a directory and only new entries were added, we typically still get a reduction on the amount of logged metadata, because it's very likely the first leaf that got the new dir index entries also has several old dir index entries. So change the logging logic to not log dir index keys created in past transactions and log a range item for every gap it finds between each pair of consecutive index keys, to ensure deletions are tracked and replayed on log replay. This patch is part of a patchset comprised of the following patches: 1/4 btrfs: don't log unnecessary boundary keys when logging directory 2/4 btrfs: put initial index value of a directory in a constant 3/4 btrfs: stop copying old dir items when logging a directory 4/4 btrfs: stop trying to log subdirectories created in past transactions The following test was run on a branch without this patchset and on a branch with the first three patches applied: $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=1000000 NUM_FILE_DELETES=10000 MKFS_OPTIONS="-O no-holes -R free-space-tree" MOUNT_OPTIONS="-o ssd" mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync del_inc=$(( $NUM_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The test was run on a non-debug kernel (Debian's default kernel config), and the results were the following for various values of NUM_FILES and NUM_FILE_DELETES: ** before, NUM_FILES = 1 000 000, NUM_FILE_DELETES = 10 000 ** dir fsync took 585 ms after deleting 10000 files ** after, NUM_FILES = 1 000 000, NUM_FILE_DELETES = 10 000 ** dir fsync took 34 ms after deleting 10000 files (-94.2%) ** before, NUM_FILES = 100 000, NUM_FILE_DELETES = 1 000 ** dir fsync took 50 ms after deleting 1000 files ** after, NUM_FILES = 100 000, NUM_FILE_DELETES = 1 000 ** dir fsync took 7 ms after deleting 1000 files (-86.0%) ** before, NUM_FILES = 10 000, NUM_FILE_DELETES = 100 ** dir fsync took 9 ms after deleting 100 files ** after, NUM_FILES = 10 000, NUM_FILE_DELETES = 100 ** dir fsync took 5 ms after deleting 100 files (-44.4%) Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-12-15 15:20:00 +03:00
struct btrfs_log_ctx *ctx,
u64 *last_old_dentry_offset)
{
struct btrfs_root *log = inode->root->log_root;
struct extent_buffer *src = path->nodes[0];
const int nritems = btrfs_header_nritems(src);
const u64 ino = btrfs_ino(inode);
btrfs: insert items in batches when logging a directory when possible When logging a directory, we scan its directory items from the subvolume tree and then copy one by one into the log tree. This is not efficient since we generally are able to insert several items in a batch, using a single btree operation for adding several items at once. The reason we copy items one by one is that we must check if each item was previously logged in the current transaction, and if it was we either overwrite it or skip it in case its content did not change in the subvolume tree (this can happen only for dir item keys, but not for dir index keys), and doing such check makes it a bit cumbersome to attempt batch insertions. However the chances for doing batch insertions are very frequent and always happen when: 1) Logging the directory for the first time in the current transaction, as none of the items exist in the log tree yet; 2) Logging new dir index keys, because the offset for new dir index keys comes from a monotonically increasing counter. This means if we keep adding dentries to a directory, through creation of new files and sub-directories or by adding new links or renaming from some other directory into the one we are logging, all the new dir index keys have a new offset that is greater than the offset of any previously logged index keys, so we can insert them in batches into the log tree. For dir item keys, since their offset depends on the result of an hash function against the dentry's name, unless the directory is being logged for the first time in the current transaction, the chances being able to insert the items in the log using batches is pretty much random and not predictable, as it depends on the names of the dentries, but still happens often enough. So change directory logging to keep track of consecutive directory items that don't exist yet in the log and batch insert them. This patch is part of a patchset comprised of the following 5 patches: btrfs: remove root argument from btrfs_log_inode() and its callees btrfs: remove redundant log root assignment from log_dir_items() btrfs: factor out the copying loop of dir items from log_dir_items() btrfs: insert items in batches when logging a directory when possible btrfs: keep track of the last logged keys when logging a directory This is patch 4/5. The change log of the last patch (5/5) has performance results. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-16 13:32:13 +03:00
bool last_found = false;
int batch_start = 0;
int batch_size = 0;
int i;
for (i = path->slots[0]; i < nritems; i++) {
btrfs: stop copying old dir items when logging a directory When logging a directory, we go over every leaf of the subvolume tree that was changed in the current transaction and copy all its dir index keys to the log tree. That includes copying dir index keys created in past transactions. This is done mostly for simplicity, as after logging the keys we log an item that specifies the start and end ranges of the keys we logged. That item is then used during log replay to figure out which keys need to be deleted - every key in that range that we find in the subvolume tree and is not in the log tree, needs to be deleted. Now that we log only dir index keys, and not dir item keys anymore, when we remove dentries from a directory (due to unlink and rename operations), we can get entire leaves that we changed only for deleting old dir index keys, or that have few dir index keys that are new - this is due to the fact that the offset for new index keys comes from a monotonically increasing counter. We can avoid logging dir index keys from past transactions, and in order to track the deletions, only log range items (BTRFS_DIR_LOG_INDEX_KEY key type) when we find gaps between consecutive index keys. This massively reduces the amount of logged metadata when we have deleted directory entries, even if it's a small percentage of the total number of entries. The reduction comes from both less items that are logged and instead of logging many dir index items (struct btrfs_dir_item), which have a size of 30 bytes plus a file name, we typically log just a few range items (struct btrfs_dir_log_item), which take only 8 bytes each. Even if no entries were deleted from a directory and only new entries were added, we typically still get a reduction on the amount of logged metadata, because it's very likely the first leaf that got the new dir index entries also has several old dir index entries. So change the logging logic to not log dir index keys created in past transactions and log a range item for every gap it finds between each pair of consecutive index keys, to ensure deletions are tracked and replayed on log replay. This patch is part of a patchset comprised of the following patches: 1/4 btrfs: don't log unnecessary boundary keys when logging directory 2/4 btrfs: put initial index value of a directory in a constant 3/4 btrfs: stop copying old dir items when logging a directory 4/4 btrfs: stop trying to log subdirectories created in past transactions The following test was run on a branch without this patchset and on a branch with the first three patches applied: $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=1000000 NUM_FILE_DELETES=10000 MKFS_OPTIONS="-O no-holes -R free-space-tree" MOUNT_OPTIONS="-o ssd" mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync del_inc=$(( $NUM_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The test was run on a non-debug kernel (Debian's default kernel config), and the results were the following for various values of NUM_FILES and NUM_FILE_DELETES: ** before, NUM_FILES = 1 000 000, NUM_FILE_DELETES = 10 000 ** dir fsync took 585 ms after deleting 10000 files ** after, NUM_FILES = 1 000 000, NUM_FILE_DELETES = 10 000 ** dir fsync took 34 ms after deleting 10000 files (-94.2%) ** before, NUM_FILES = 100 000, NUM_FILE_DELETES = 1 000 ** dir fsync took 50 ms after deleting 1000 files ** after, NUM_FILES = 100 000, NUM_FILE_DELETES = 1 000 ** dir fsync took 7 ms after deleting 1000 files (-86.0%) ** before, NUM_FILES = 10 000, NUM_FILE_DELETES = 100 ** dir fsync took 9 ms after deleting 100 files ** after, NUM_FILES = 10 000, NUM_FILE_DELETES = 100 ** dir fsync took 5 ms after deleting 100 files (-44.4%) Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-12-15 15:20:00 +03:00
struct btrfs_dir_item *di;
struct btrfs_key key;
int ret;
btrfs_item_key_to_cpu(src, &key, i);
btrfs: only copy dir index keys when logging a directory Currently, when logging a directory, we copy both dir items and dir index items from the fs/subvolume tree to the log tree. Both items have exactly the same data (same struct btrfs_dir_item), the difference lies in the key values, where a dir index key contains the index number of a directory entry while the dir item key does not, as it's used for doing fast lookups of an entry by name, while the former is used for sorting entries when listing a directory. We can exploit that and log only the dir index items, since they contain all the information needed to correctly add, replace and delete directory entries when replaying a log tree. Logging only the dir index items is also backward and forward compatible: an unpatched kernel (without this change) can correctly replay a log tree generated by a patched kernel (with this patch), and a patched kernel can correctly replay a log tree generated by an unpatched kernel. The backward compatibility is ensured because: 1) For inserting a new dentry: a dentry is only inserted when we find a new dir index key - we can only insert if we know the dir index offset, which is encoded in the dir index key's offset; 2) For deleting dentries: during log replay, before adding or replacing dentries, we first replay dentry deletions. Whenever we find a dir item key or a dir index key in the subvolume/fs tree that is not logged in a range for which the log tree is authoritative, we do the unlink of the dentry, which removes both the existing dir item key and the dir index key. Therefore logging just dir index keys is enough to ensure dentry deletions are correctly replayed; 3) For dentry replacements: they work when we log only dir index keys and this is mostly due to a combination of 1) and 2). If we replace a dentry with name "foobar" to point from inode A to inode B, then we know the dir index key for the new dentry is different from the old one, as it has an index number (key offset) larger than the old one. This results in replaying a deletion, through replay_dir_deletes(), that causes the old dentry to be removed, both the dir item key and the dir index key, as mentioned at 2). Then when processing the new dir index key, we add the new dentry, adding both a new dir item key and a new index key pointing to inode B, as stated in 1). The forward compatibility, the ability for a patched kernel to replay a log created by an older, unpatched kernel, comes from the changes required for making sure we are able to replay a log that only contains dir index keys - we simply ignore every dir item key we find. So modify directory logging to log only dir index items, and modify the log replay process to ignore dir item keys, from log trees created by an unpatched kernel, and process only with dir index keys. This reduces the amount of logged metadata by about half, and therefore the time spent logging or fsyncing large directories (less CPU time and less IO). The following test script was used to measure this change: #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=1000000 NUM_FILE_DELETES=10000 mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The tests were run on a physical machine, with a non-debug kernel (Debian's default kernel config), for different values of $NUM_NEW_FILES and $NUM_FILE_DELETES, and the results were the following: ** Before patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 8412 ms after adding 1000000 files dir fsync took 500 ms after deleting 10000 files ** After patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 4252 ms after adding 1000000 files (-49.5%) dir fsync took 269 ms after deleting 10000 files (-46.2%) ** Before patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 745 ms after adding 100000 files dir fsync took 59 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 404 ms after adding 100000 files (-45.8%) dir fsync took 31 ms after deleting 1000 files (-47.5%) ** Before patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 67 ms after adding 10000 files dir fsync took 9 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 36 ms after adding 10000 files (-46.3%) dir fsync took 5 ms after deleting 1000 files (-44.4%) ** Before patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 9 ms after adding 1000 files dir fsync took 4 ms after deleting 100 files ** After patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 7 ms after adding 1000 files (-22.2%) dir fsync took 3 ms after deleting 100 files (-25.0%) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-10-25 19:31:53 +03:00
if (key.objectid != ino || key.type != BTRFS_DIR_INDEX_KEY) {
btrfs: insert items in batches when logging a directory when possible When logging a directory, we scan its directory items from the subvolume tree and then copy one by one into the log tree. This is not efficient since we generally are able to insert several items in a batch, using a single btree operation for adding several items at once. The reason we copy items one by one is that we must check if each item was previously logged in the current transaction, and if it was we either overwrite it or skip it in case its content did not change in the subvolume tree (this can happen only for dir item keys, but not for dir index keys), and doing such check makes it a bit cumbersome to attempt batch insertions. However the chances for doing batch insertions are very frequent and always happen when: 1) Logging the directory for the first time in the current transaction, as none of the items exist in the log tree yet; 2) Logging new dir index keys, because the offset for new dir index keys comes from a monotonically increasing counter. This means if we keep adding dentries to a directory, through creation of new files and sub-directories or by adding new links or renaming from some other directory into the one we are logging, all the new dir index keys have a new offset that is greater than the offset of any previously logged index keys, so we can insert them in batches into the log tree. For dir item keys, since their offset depends on the result of an hash function against the dentry's name, unless the directory is being logged for the first time in the current transaction, the chances being able to insert the items in the log using batches is pretty much random and not predictable, as it depends on the names of the dentries, but still happens often enough. So change directory logging to keep track of consecutive directory items that don't exist yet in the log and batch insert them. This patch is part of a patchset comprised of the following 5 patches: btrfs: remove root argument from btrfs_log_inode() and its callees btrfs: remove redundant log root assignment from log_dir_items() btrfs: factor out the copying loop of dir items from log_dir_items() btrfs: insert items in batches when logging a directory when possible btrfs: keep track of the last logged keys when logging a directory This is patch 4/5. The change log of the last patch (5/5) has performance results. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-16 13:32:13 +03:00
last_found = true;
break;
}
btrfs: stop copying old dir items when logging a directory When logging a directory, we go over every leaf of the subvolume tree that was changed in the current transaction and copy all its dir index keys to the log tree. That includes copying dir index keys created in past transactions. This is done mostly for simplicity, as after logging the keys we log an item that specifies the start and end ranges of the keys we logged. That item is then used during log replay to figure out which keys need to be deleted - every key in that range that we find in the subvolume tree and is not in the log tree, needs to be deleted. Now that we log only dir index keys, and not dir item keys anymore, when we remove dentries from a directory (due to unlink and rename operations), we can get entire leaves that we changed only for deleting old dir index keys, or that have few dir index keys that are new - this is due to the fact that the offset for new index keys comes from a monotonically increasing counter. We can avoid logging dir index keys from past transactions, and in order to track the deletions, only log range items (BTRFS_DIR_LOG_INDEX_KEY key type) when we find gaps between consecutive index keys. This massively reduces the amount of logged metadata when we have deleted directory entries, even if it's a small percentage of the total number of entries. The reduction comes from both less items that are logged and instead of logging many dir index items (struct btrfs_dir_item), which have a size of 30 bytes plus a file name, we typically log just a few range items (struct btrfs_dir_log_item), which take only 8 bytes each. Even if no entries were deleted from a directory and only new entries were added, we typically still get a reduction on the amount of logged metadata, because it's very likely the first leaf that got the new dir index entries also has several old dir index entries. So change the logging logic to not log dir index keys created in past transactions and log a range item for every gap it finds between each pair of consecutive index keys, to ensure deletions are tracked and replayed on log replay. This patch is part of a patchset comprised of the following patches: 1/4 btrfs: don't log unnecessary boundary keys when logging directory 2/4 btrfs: put initial index value of a directory in a constant 3/4 btrfs: stop copying old dir items when logging a directory 4/4 btrfs: stop trying to log subdirectories created in past transactions The following test was run on a branch without this patchset and on a branch with the first three patches applied: $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=1000000 NUM_FILE_DELETES=10000 MKFS_OPTIONS="-O no-holes -R free-space-tree" MOUNT_OPTIONS="-o ssd" mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync del_inc=$(( $NUM_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The test was run on a non-debug kernel (Debian's default kernel config), and the results were the following for various values of NUM_FILES and NUM_FILE_DELETES: ** before, NUM_FILES = 1 000 000, NUM_FILE_DELETES = 10 000 ** dir fsync took 585 ms after deleting 10000 files ** after, NUM_FILES = 1 000 000, NUM_FILE_DELETES = 10 000 ** dir fsync took 34 ms after deleting 10000 files (-94.2%) ** before, NUM_FILES = 100 000, NUM_FILE_DELETES = 1 000 ** dir fsync took 50 ms after deleting 1000 files ** after, NUM_FILES = 100 000, NUM_FILE_DELETES = 1 000 ** dir fsync took 7 ms after deleting 1000 files (-86.0%) ** before, NUM_FILES = 10 000, NUM_FILE_DELETES = 100 ** dir fsync took 9 ms after deleting 100 files ** after, NUM_FILES = 10 000, NUM_FILE_DELETES = 100 ** dir fsync took 5 ms after deleting 100 files (-44.4%) Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-12-15 15:20:00 +03:00
di = btrfs_item_ptr(src, i, struct btrfs_dir_item);
btrfs: keep track of the last logged keys when logging a directory After the first time we log a directory in the current transaction, for each directory item in a changed leaf of the subvolume tree, we have to check if we previously logged the item, in order to overwrite it in case its data changed or skip it in case its data hasn't changed. Checking if we have logged each item before not only wastes times, but it also adds lock contention on the log tree. So in order to minimize the number of times we do such checks, keep track of the offset of the last key we logged for a directory and, on the next time we log the directory, skip the checks for any new keys that have an offset greater than the offset we have previously saved. This is specially effective for index keys, because the offset for these keys comes from a monotonically increasing counter. This patch is part of a patchset comprised of the following 5 patches: btrfs: remove root argument from btrfs_log_inode() and its callees btrfs: remove redundant log root assignment from log_dir_items() btrfs: factor out the copying loop of dir items from log_dir_items() btrfs: insert items in batches when logging a directory when possible btrfs: keep track of the last logged keys when logging a directory This is patch 5/5. The following test was used on a non-debug kernel to measure the impact it has on a directory fsync: $ cat test-dir-fsync.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=100000 NUM_FILE_DELETES=1000 mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done # fsync the directory, this will log the new dir items and the inodes # they point to, because these are new inodes. start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done # fsync the directory, this will only log dir items, there are no # dentries pointing to new inodes. start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" umount $MNT Test results with NUM_NEW_FILES set to 100 000 and 1 000 000: **** before patchset, 100 000 files, 1000 deletes **** dir fsync took 848 ms after adding 100000 files dir fsync took 175 ms after deleting 1000 files **** after patchset, 100 000 files, 1000 deletes **** dir fsync took 758 ms after adding 100000 files (-11.2%) dir fsync took 63 ms after deleting 1000 files (-94.1%) **** before patchset, 1 000 000 files, 1000 deletes **** dir fsync took 9945 ms after adding 1000000 files dir fsync took 473 ms after deleting 1000 files **** after patchset, 1 000 000 files, 1000 deletes **** dir fsync took 8677 ms after adding 1000000 files (-13.6%) dir fsync took 146 ms after deleting 1000 files (-105.6%) Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-16 13:32:14 +03:00
ctx->last_dir_item_offset = key.offset;
btrfs: stop copying old dir items when logging a directory When logging a directory, we go over every leaf of the subvolume tree that was changed in the current transaction and copy all its dir index keys to the log tree. That includes copying dir index keys created in past transactions. This is done mostly for simplicity, as after logging the keys we log an item that specifies the start and end ranges of the keys we logged. That item is then used during log replay to figure out which keys need to be deleted - every key in that range that we find in the subvolume tree and is not in the log tree, needs to be deleted. Now that we log only dir index keys, and not dir item keys anymore, when we remove dentries from a directory (due to unlink and rename operations), we can get entire leaves that we changed only for deleting old dir index keys, or that have few dir index keys that are new - this is due to the fact that the offset for new index keys comes from a monotonically increasing counter. We can avoid logging dir index keys from past transactions, and in order to track the deletions, only log range items (BTRFS_DIR_LOG_INDEX_KEY key type) when we find gaps between consecutive index keys. This massively reduces the amount of logged metadata when we have deleted directory entries, even if it's a small percentage of the total number of entries. The reduction comes from both less items that are logged and instead of logging many dir index items (struct btrfs_dir_item), which have a size of 30 bytes plus a file name, we typically log just a few range items (struct btrfs_dir_log_item), which take only 8 bytes each. Even if no entries were deleted from a directory and only new entries were added, we typically still get a reduction on the amount of logged metadata, because it's very likely the first leaf that got the new dir index entries also has several old dir index entries. So change the logging logic to not log dir index keys created in past transactions and log a range item for every gap it finds between each pair of consecutive index keys, to ensure deletions are tracked and replayed on log replay. This patch is part of a patchset comprised of the following patches: 1/4 btrfs: don't log unnecessary boundary keys when logging directory 2/4 btrfs: put initial index value of a directory in a constant 3/4 btrfs: stop copying old dir items when logging a directory 4/4 btrfs: stop trying to log subdirectories created in past transactions The following test was run on a branch without this patchset and on a branch with the first three patches applied: $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=1000000 NUM_FILE_DELETES=10000 MKFS_OPTIONS="-O no-holes -R free-space-tree" MOUNT_OPTIONS="-o ssd" mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync del_inc=$(( $NUM_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The test was run on a non-debug kernel (Debian's default kernel config), and the results were the following for various values of NUM_FILES and NUM_FILE_DELETES: ** before, NUM_FILES = 1 000 000, NUM_FILE_DELETES = 10 000 ** dir fsync took 585 ms after deleting 10000 files ** after, NUM_FILES = 1 000 000, NUM_FILE_DELETES = 10 000 ** dir fsync took 34 ms after deleting 10000 files (-94.2%) ** before, NUM_FILES = 100 000, NUM_FILE_DELETES = 1 000 ** dir fsync took 50 ms after deleting 1000 files ** after, NUM_FILES = 100 000, NUM_FILE_DELETES = 1 000 ** dir fsync took 7 ms after deleting 1000 files (-86.0%) ** before, NUM_FILES = 10 000, NUM_FILE_DELETES = 100 ** dir fsync took 9 ms after deleting 100 files ** after, NUM_FILES = 10 000, NUM_FILE_DELETES = 100 ** dir fsync took 5 ms after deleting 100 files (-44.4%) Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-12-15 15:20:00 +03:00
/*
* Skip ranges of items that consist only of dir item keys created
* in past transactions. However if we find a gap, we must log a
* dir index range item for that gap, so that index keys in that
* gap are deleted during log replay.
*/
if (btrfs_dir_transid(src, di) < trans->transid) {
if (key.offset > *last_old_dentry_offset + 1) {
ret = insert_dir_log_key(trans, log, dst_path,
ino, *last_old_dentry_offset + 1,
key.offset - 1);
/*
* -EEXIST should never happen because when we
* log a directory in full mode (LOG_INODE_ALL)
* we drop all BTRFS_DIR_LOG_INDEX_KEY keys from
* the log tree.
*/
ASSERT(ret != -EEXIST);
if (ret < 0)
return ret;
}
*last_old_dentry_offset = key.offset;
continue;
}
/*
* We must make sure that when we log a directory entry, the
* corresponding inode, after log replay, has a matching link
* count. For example:
*
* touch foo
* mkdir mydir
* sync
* ln foo mydir/bar
* xfs_io -c "fsync" mydir
* <crash>
* <mount fs and log replay>
*
* Would result in a fsync log that when replayed, our file inode
* would have a link count of 1, but we get two directory entries
* pointing to the same inode. After removing one of the names,
* it would not be possible to remove the other name, which
* resulted always in stale file handle errors, and would not be
* possible to rmdir the parent directory, since its i_size could
* never be decremented to the value BTRFS_EMPTY_DIR_SIZE,
* resulting in -ENOTEMPTY errors.
*/
btrfs: insert items in batches when logging a directory when possible When logging a directory, we scan its directory items from the subvolume tree and then copy one by one into the log tree. This is not efficient since we generally are able to insert several items in a batch, using a single btree operation for adding several items at once. The reason we copy items one by one is that we must check if each item was previously logged in the current transaction, and if it was we either overwrite it or skip it in case its content did not change in the subvolume tree (this can happen only for dir item keys, but not for dir index keys), and doing such check makes it a bit cumbersome to attempt batch insertions. However the chances for doing batch insertions are very frequent and always happen when: 1) Logging the directory for the first time in the current transaction, as none of the items exist in the log tree yet; 2) Logging new dir index keys, because the offset for new dir index keys comes from a monotonically increasing counter. This means if we keep adding dentries to a directory, through creation of new files and sub-directories or by adding new links or renaming from some other directory into the one we are logging, all the new dir index keys have a new offset that is greater than the offset of any previously logged index keys, so we can insert them in batches into the log tree. For dir item keys, since their offset depends on the result of an hash function against the dentry's name, unless the directory is being logged for the first time in the current transaction, the chances being able to insert the items in the log using batches is pretty much random and not predictable, as it depends on the names of the dentries, but still happens often enough. So change directory logging to keep track of consecutive directory items that don't exist yet in the log and batch insert them. This patch is part of a patchset comprised of the following 5 patches: btrfs: remove root argument from btrfs_log_inode() and its callees btrfs: remove redundant log root assignment from log_dir_items() btrfs: factor out the copying loop of dir items from log_dir_items() btrfs: insert items in batches when logging a directory when possible btrfs: keep track of the last logged keys when logging a directory This is patch 4/5. The change log of the last patch (5/5) has performance results. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-16 13:32:13 +03:00
if (!ctx->log_new_dentries) {
struct btrfs_key di_key;
btrfs_dir_item_key_to_cpu(src, di, &di_key);
btrfs: stop copying old dir items when logging a directory When logging a directory, we go over every leaf of the subvolume tree that was changed in the current transaction and copy all its dir index keys to the log tree. That includes copying dir index keys created in past transactions. This is done mostly for simplicity, as after logging the keys we log an item that specifies the start and end ranges of the keys we logged. That item is then used during log replay to figure out which keys need to be deleted - every key in that range that we find in the subvolume tree and is not in the log tree, needs to be deleted. Now that we log only dir index keys, and not dir item keys anymore, when we remove dentries from a directory (due to unlink and rename operations), we can get entire leaves that we changed only for deleting old dir index keys, or that have few dir index keys that are new - this is due to the fact that the offset for new index keys comes from a monotonically increasing counter. We can avoid logging dir index keys from past transactions, and in order to track the deletions, only log range items (BTRFS_DIR_LOG_INDEX_KEY key type) when we find gaps between consecutive index keys. This massively reduces the amount of logged metadata when we have deleted directory entries, even if it's a small percentage of the total number of entries. The reduction comes from both less items that are logged and instead of logging many dir index items (struct btrfs_dir_item), which have a size of 30 bytes plus a file name, we typically log just a few range items (struct btrfs_dir_log_item), which take only 8 bytes each. Even if no entries were deleted from a directory and only new entries were added, we typically still get a reduction on the amount of logged metadata, because it's very likely the first leaf that got the new dir index entries also has several old dir index entries. So change the logging logic to not log dir index keys created in past transactions and log a range item for every gap it finds between each pair of consecutive index keys, to ensure deletions are tracked and replayed on log replay. This patch is part of a patchset comprised of the following patches: 1/4 btrfs: don't log unnecessary boundary keys when logging directory 2/4 btrfs: put initial index value of a directory in a constant 3/4 btrfs: stop copying old dir items when logging a directory 4/4 btrfs: stop trying to log subdirectories created in past transactions The following test was run on a branch without this patchset and on a branch with the first three patches applied: $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=1000000 NUM_FILE_DELETES=10000 MKFS_OPTIONS="-O no-holes -R free-space-tree" MOUNT_OPTIONS="-o ssd" mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync del_inc=$(( $NUM_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The test was run on a non-debug kernel (Debian's default kernel config), and the results were the following for various values of NUM_FILES and NUM_FILE_DELETES: ** before, NUM_FILES = 1 000 000, NUM_FILE_DELETES = 10 000 ** dir fsync took 585 ms after deleting 10000 files ** after, NUM_FILES = 1 000 000, NUM_FILE_DELETES = 10 000 ** dir fsync took 34 ms after deleting 10000 files (-94.2%) ** before, NUM_FILES = 100 000, NUM_FILE_DELETES = 1 000 ** dir fsync took 50 ms after deleting 1000 files ** after, NUM_FILES = 100 000, NUM_FILE_DELETES = 1 000 ** dir fsync took 7 ms after deleting 1000 files (-86.0%) ** before, NUM_FILES = 10 000, NUM_FILE_DELETES = 100 ** dir fsync took 9 ms after deleting 100 files ** after, NUM_FILES = 10 000, NUM_FILE_DELETES = 100 ** dir fsync took 5 ms after deleting 100 files (-44.4%) Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-12-15 15:20:00 +03:00
if (di_key.type != BTRFS_ROOT_ITEM_KEY)
btrfs: insert items in batches when logging a directory when possible When logging a directory, we scan its directory items from the subvolume tree and then copy one by one into the log tree. This is not efficient since we generally are able to insert several items in a batch, using a single btree operation for adding several items at once. The reason we copy items one by one is that we must check if each item was previously logged in the current transaction, and if it was we either overwrite it or skip it in case its content did not change in the subvolume tree (this can happen only for dir item keys, but not for dir index keys), and doing such check makes it a bit cumbersome to attempt batch insertions. However the chances for doing batch insertions are very frequent and always happen when: 1) Logging the directory for the first time in the current transaction, as none of the items exist in the log tree yet; 2) Logging new dir index keys, because the offset for new dir index keys comes from a monotonically increasing counter. This means if we keep adding dentries to a directory, through creation of new files and sub-directories or by adding new links or renaming from some other directory into the one we are logging, all the new dir index keys have a new offset that is greater than the offset of any previously logged index keys, so we can insert them in batches into the log tree. For dir item keys, since their offset depends on the result of an hash function against the dentry's name, unless the directory is being logged for the first time in the current transaction, the chances being able to insert the items in the log using batches is pretty much random and not predictable, as it depends on the names of the dentries, but still happens often enough. So change directory logging to keep track of consecutive directory items that don't exist yet in the log and batch insert them. This patch is part of a patchset comprised of the following 5 patches: btrfs: remove root argument from btrfs_log_inode() and its callees btrfs: remove redundant log root assignment from log_dir_items() btrfs: factor out the copying loop of dir items from log_dir_items() btrfs: insert items in batches when logging a directory when possible btrfs: keep track of the last logged keys when logging a directory This is patch 4/5. The change log of the last patch (5/5) has performance results. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-16 13:32:13 +03:00
ctx->log_new_dentries = true;
}
btrfs: avoid inode logging during rename and link when possible During a rename or link operation, we need to determine if an inode was previously logged or not, and if it was, do some update to the logged inode. We used to rely exclusively on the logged_trans field of struct btrfs_inode to determine that, but that was not reliable because the value of that field is not persisted in the inode item, so it's lost when an inode is evicted and loaded back again. That led to several issues in the past, such as not persisting deletions (such as the case fixed by commit 803f0f64d17769 ("Btrfs: fix fsync not persisting dentry deletions due to inode evictions")), or resulting in losing a file after an inode eviction followed by a rename (commit ecc64fab7d49c6 ("btrfs: fix lost inode on log replay after mix of fsync, rename and inode eviction")), besides other issues. So the inode_logged() helper was introduced and used to determine if an inode was possibly logged before in the current transaction, with the caveat that it could return false positives, in the sense that even if an inode was not logged before in the current transaction, it could still return true, but never to return false in case the inode was logged. >From a functional point of view that is fine, but from a performance perspective it can introduce significant latencies to rename and link operations, as they will end up doing inode logging even when it is not necessary. Recently on a 5.15 kernel, an openSUSE Tumbleweed user reported package installations and upgrades, with the zypper tool, were often taking a long time to complete. With strace it could be observed that zypper was spending about 99% of its time on rename operations, and then with further analysis we checked that directory logging was happening too frequently. Taking into account that installation/upgrade of some of the packages needed a few thousand file renames, the slowdown was very noticeable for the user. The issue was caused indirectly due to an excessive number of inode evictions on a 5.15 kernel, about 100x more compared to a 5.13, 5.14 or a 5.16-rc8 kernel. While triggering the inode evictions if something outside btrfs' control, btrfs could still behave better by eliminating the false positives from the inode_logged() helper. So change inode_logged() to actually eliminate such false positives caused by inode eviction and when an inode was never logged since the filesystem was mounted, as both cases relate to when the logged_trans field of struct btrfs_inode has a value of zero. When it can not determine if the inode was logged based only on the logged_trans value, lookup for the existence of the inode item in the log tree - if it's there then we known the inode was logged, if it's not there then it can not have been logged in the current transaction. Once we determine if the inode was logged, update the logged_trans value to avoid future calls to have to search in the log tree again. Alternatively, we could start storing logged_trans in the on disk inode item structure (struct btrfs_inode_item) in the unused space it still has, but that would be a bit odd because: 1) We only care about logged_trans since the filesystem was mounted, we don't care about its value from a previous mount. Having it persisted in the inode item structure would not make the best use of the precious unused space; 2) In order to get logged_trans persisted before inode eviction, we would have to update the delayed inode when we finish logging the inode and update its logged_trans in struct btrfs_inode, which makes it a bit cumbersome since we need to check if the delayed inode exists, if not create it and populate it and deal with any errors (-ENOMEM mostly). This change is part of a patchset comprised of the following patches: 1/5 btrfs: add helper to delete a dir entry from a log tree 2/5 btrfs: pass the dentry to btrfs_log_new_name() instead of the inode 3/5 btrfs: avoid logging all directory changes during renames 4/5 btrfs: stop doing unnecessary log updates during a rename 5/5 btrfs: avoid inode logging during rename and link when possible The following test script mimics part of what the zypper tool does during package installations/upgrades. It does not triggers inode evictions, but it's similar because it triggers false positives from the inode_logged() helper, because the inodes have a logged_trans of 0, there's a log tree due to a fsync of an unrelated file and the directory inode has its last_trans field set to the current transaction: $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=10000 mkfs.btrfs -f $DEV mount $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync # Now do some change to an unrelated file and fsync it. # This is just to create a log tree to make sure that inode_logged() # does not return false when called against "testdir". xfs_io -f -c "pwrite 0 4K" -c "fsync" $MNT/foo # Do some change to testdir. This is to make sure inode_logged() # will return true when called against "testdir", because its # logged_trans is 0, it was changed in the current transaction # and there's a log tree. echo -n > $MNT/testdir/file_$((NUM_FILES + 1)) echo "Renaming $NUM_FILES files..." start=$(date +%s%N) for ((i = 1; i <= $NUM_FILES; i++)); do mv $MNT/testdir/file_$i $MNT/testdir/file_$i-RPMDELETE done end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "Renames took $dur milliseconds" umount $MNT Testing this change on a box using a non-debug kernel (Debian's default kernel config) gave the following results: NUM_FILES=10000, before patchset: 27837 ms NUM_FILES=10000, after patches 1/5 to 4/5 applied: 9236 ms (-66.8%) NUM_FILES=10000, after whole patchset applied: 8902 ms (-68.0%) NUM_FILES=5000, before patchset: 9127 ms NUM_FILES=5000, after patches 1/5 to 4/5 applied: 4640 ms (-49.2%) NUM_FILES=5000, after whole patchset applied: 4441 ms (-51.3%) NUM_FILES=2000, before patchset: 2528 ms NUM_FILES=2000, after patches 1/5 to 4/5 applied: 1983 ms (-21.6%) NUM_FILES=2000, after whole patchset applied: 1747 ms (-30.9%) NUM_FILES=1000, before patchset: 1085 ms NUM_FILES=1000, after patches 1/5 to 4/5 applied: 893 ms (-17.7%) NUM_FILES=1000, after whole patchset applied: 867 ms (-20.1%) Running dbench on the same physical machine with the following script: $ cat run-dbench.sh #!/bin/bash NUM_JOBS=$(nproc --all) DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 MOUNT_OPTIONS="-o ssd" MKFS_OPTIONS="-O no-holes -R free-space-tree" echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -D $MNT -t 120 $NUM_JOBS umount $MNT Before patchset: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 3761352 0.032 143.843 Close 2762770 0.002 2.273 Rename 159304 0.291 67.037 Unlink 759784 0.207 143.998 Deltree 72 4.028 15.977 Mkdir 36 0.003 0.006 Qpathinfo 3409780 0.013 9.678 Qfileinfo 596772 0.001 0.878 Qfsinfo 625189 0.003 1.245 Sfileinfo 306443 0.006 1.840 Find 1318106 0.063 19.798 WriteX 1871137 0.021 8.532 ReadX 5897325 0.003 3.567 LockX 12252 0.003 0.258 UnlockX 12252 0.002 0.100 Flush 263666 3.327 155.632 Throughput 980.047 MB/sec 12 clients 12 procs max_latency=155.636 ms After whole patchset applied: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 4195584 0.033 107.742 Close 3081932 0.002 1.935 Rename 177641 0.218 14.905 Unlink 847333 0.166 107.822 Deltree 118 5.315 15.247 Mkdir 59 0.004 0.048 Qpathinfo 3802612 0.014 10.302 Qfileinfo 666748 0.001 1.034 Qfsinfo 697329 0.003 0.944 Sfileinfo 341712 0.006 2.099 Find 1470365 0.065 9.359 WriteX 2093921 0.021 8.087 ReadX 6576234 0.003 3.407 LockX 13660 0.003 0.308 UnlockX 13660 0.002 0.114 Flush 294090 2.906 115.539 Throughput 1093.11 MB/sec 12 clients 12 procs max_latency=115.544 ms +11.5% throughput -25.8% max latency rename max latency -77.8% Link: https://bugzilla.opensuse.org/show_bug.cgi?id=1193549 Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-01-20 14:00:10 +03:00
if (!ctx->logged_before)
btrfs: insert items in batches when logging a directory when possible When logging a directory, we scan its directory items from the subvolume tree and then copy one by one into the log tree. This is not efficient since we generally are able to insert several items in a batch, using a single btree operation for adding several items at once. The reason we copy items one by one is that we must check if each item was previously logged in the current transaction, and if it was we either overwrite it or skip it in case its content did not change in the subvolume tree (this can happen only for dir item keys, but not for dir index keys), and doing such check makes it a bit cumbersome to attempt batch insertions. However the chances for doing batch insertions are very frequent and always happen when: 1) Logging the directory for the first time in the current transaction, as none of the items exist in the log tree yet; 2) Logging new dir index keys, because the offset for new dir index keys comes from a monotonically increasing counter. This means if we keep adding dentries to a directory, through creation of new files and sub-directories or by adding new links or renaming from some other directory into the one we are logging, all the new dir index keys have a new offset that is greater than the offset of any previously logged index keys, so we can insert them in batches into the log tree. For dir item keys, since their offset depends on the result of an hash function against the dentry's name, unless the directory is being logged for the first time in the current transaction, the chances being able to insert the items in the log using batches is pretty much random and not predictable, as it depends on the names of the dentries, but still happens often enough. So change directory logging to keep track of consecutive directory items that don't exist yet in the log and batch insert them. This patch is part of a patchset comprised of the following 5 patches: btrfs: remove root argument from btrfs_log_inode() and its callees btrfs: remove redundant log root assignment from log_dir_items() btrfs: factor out the copying loop of dir items from log_dir_items() btrfs: insert items in batches when logging a directory when possible btrfs: keep track of the last logged keys when logging a directory This is patch 4/5. The change log of the last patch (5/5) has performance results. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-16 13:32:13 +03:00
goto add_to_batch;
btrfs: keep track of the last logged keys when logging a directory After the first time we log a directory in the current transaction, for each directory item in a changed leaf of the subvolume tree, we have to check if we previously logged the item, in order to overwrite it in case its data changed or skip it in case its data hasn't changed. Checking if we have logged each item before not only wastes times, but it also adds lock contention on the log tree. So in order to minimize the number of times we do such checks, keep track of the offset of the last key we logged for a directory and, on the next time we log the directory, skip the checks for any new keys that have an offset greater than the offset we have previously saved. This is specially effective for index keys, because the offset for these keys comes from a monotonically increasing counter. This patch is part of a patchset comprised of the following 5 patches: btrfs: remove root argument from btrfs_log_inode() and its callees btrfs: remove redundant log root assignment from log_dir_items() btrfs: factor out the copying loop of dir items from log_dir_items() btrfs: insert items in batches when logging a directory when possible btrfs: keep track of the last logged keys when logging a directory This is patch 5/5. The following test was used on a non-debug kernel to measure the impact it has on a directory fsync: $ cat test-dir-fsync.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=100000 NUM_FILE_DELETES=1000 mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done # fsync the directory, this will log the new dir items and the inodes # they point to, because these are new inodes. start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done # fsync the directory, this will only log dir items, there are no # dentries pointing to new inodes. start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" umount $MNT Test results with NUM_NEW_FILES set to 100 000 and 1 000 000: **** before patchset, 100 000 files, 1000 deletes **** dir fsync took 848 ms after adding 100000 files dir fsync took 175 ms after deleting 1000 files **** after patchset, 100 000 files, 1000 deletes **** dir fsync took 758 ms after adding 100000 files (-11.2%) dir fsync took 63 ms after deleting 1000 files (-94.1%) **** before patchset, 1 000 000 files, 1000 deletes **** dir fsync took 9945 ms after adding 1000000 files dir fsync took 473 ms after deleting 1000 files **** after patchset, 1 000 000 files, 1000 deletes **** dir fsync took 8677 ms after adding 1000000 files (-13.6%) dir fsync took 146 ms after deleting 1000 files (-105.6%) Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-16 13:32:14 +03:00
/*
* If we were logged before and have logged dir items, we can skip
* checking if any item with a key offset larger than the last one
* we logged is in the log tree, saving time and avoiding adding
btrfs: avoid inode logging during rename and link when possible During a rename or link operation, we need to determine if an inode was previously logged or not, and if it was, do some update to the logged inode. We used to rely exclusively on the logged_trans field of struct btrfs_inode to determine that, but that was not reliable because the value of that field is not persisted in the inode item, so it's lost when an inode is evicted and loaded back again. That led to several issues in the past, such as not persisting deletions (such as the case fixed by commit 803f0f64d17769 ("Btrfs: fix fsync not persisting dentry deletions due to inode evictions")), or resulting in losing a file after an inode eviction followed by a rename (commit ecc64fab7d49c6 ("btrfs: fix lost inode on log replay after mix of fsync, rename and inode eviction")), besides other issues. So the inode_logged() helper was introduced and used to determine if an inode was possibly logged before in the current transaction, with the caveat that it could return false positives, in the sense that even if an inode was not logged before in the current transaction, it could still return true, but never to return false in case the inode was logged. >From a functional point of view that is fine, but from a performance perspective it can introduce significant latencies to rename and link operations, as they will end up doing inode logging even when it is not necessary. Recently on a 5.15 kernel, an openSUSE Tumbleweed user reported package installations and upgrades, with the zypper tool, were often taking a long time to complete. With strace it could be observed that zypper was spending about 99% of its time on rename operations, and then with further analysis we checked that directory logging was happening too frequently. Taking into account that installation/upgrade of some of the packages needed a few thousand file renames, the slowdown was very noticeable for the user. The issue was caused indirectly due to an excessive number of inode evictions on a 5.15 kernel, about 100x more compared to a 5.13, 5.14 or a 5.16-rc8 kernel. While triggering the inode evictions if something outside btrfs' control, btrfs could still behave better by eliminating the false positives from the inode_logged() helper. So change inode_logged() to actually eliminate such false positives caused by inode eviction and when an inode was never logged since the filesystem was mounted, as both cases relate to when the logged_trans field of struct btrfs_inode has a value of zero. When it can not determine if the inode was logged based only on the logged_trans value, lookup for the existence of the inode item in the log tree - if it's there then we known the inode was logged, if it's not there then it can not have been logged in the current transaction. Once we determine if the inode was logged, update the logged_trans value to avoid future calls to have to search in the log tree again. Alternatively, we could start storing logged_trans in the on disk inode item structure (struct btrfs_inode_item) in the unused space it still has, but that would be a bit odd because: 1) We only care about logged_trans since the filesystem was mounted, we don't care about its value from a previous mount. Having it persisted in the inode item structure would not make the best use of the precious unused space; 2) In order to get logged_trans persisted before inode eviction, we would have to update the delayed inode when we finish logging the inode and update its logged_trans in struct btrfs_inode, which makes it a bit cumbersome since we need to check if the delayed inode exists, if not create it and populate it and deal with any errors (-ENOMEM mostly). This change is part of a patchset comprised of the following patches: 1/5 btrfs: add helper to delete a dir entry from a log tree 2/5 btrfs: pass the dentry to btrfs_log_new_name() instead of the inode 3/5 btrfs: avoid logging all directory changes during renames 4/5 btrfs: stop doing unnecessary log updates during a rename 5/5 btrfs: avoid inode logging during rename and link when possible The following test script mimics part of what the zypper tool does during package installations/upgrades. It does not triggers inode evictions, but it's similar because it triggers false positives from the inode_logged() helper, because the inodes have a logged_trans of 0, there's a log tree due to a fsync of an unrelated file and the directory inode has its last_trans field set to the current transaction: $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=10000 mkfs.btrfs -f $DEV mount $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync # Now do some change to an unrelated file and fsync it. # This is just to create a log tree to make sure that inode_logged() # does not return false when called against "testdir". xfs_io -f -c "pwrite 0 4K" -c "fsync" $MNT/foo # Do some change to testdir. This is to make sure inode_logged() # will return true when called against "testdir", because its # logged_trans is 0, it was changed in the current transaction # and there's a log tree. echo -n > $MNT/testdir/file_$((NUM_FILES + 1)) echo "Renaming $NUM_FILES files..." start=$(date +%s%N) for ((i = 1; i <= $NUM_FILES; i++)); do mv $MNT/testdir/file_$i $MNT/testdir/file_$i-RPMDELETE done end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "Renames took $dur milliseconds" umount $MNT Testing this change on a box using a non-debug kernel (Debian's default kernel config) gave the following results: NUM_FILES=10000, before patchset: 27837 ms NUM_FILES=10000, after patches 1/5 to 4/5 applied: 9236 ms (-66.8%) NUM_FILES=10000, after whole patchset applied: 8902 ms (-68.0%) NUM_FILES=5000, before patchset: 9127 ms NUM_FILES=5000, after patches 1/5 to 4/5 applied: 4640 ms (-49.2%) NUM_FILES=5000, after whole patchset applied: 4441 ms (-51.3%) NUM_FILES=2000, before patchset: 2528 ms NUM_FILES=2000, after patches 1/5 to 4/5 applied: 1983 ms (-21.6%) NUM_FILES=2000, after whole patchset applied: 1747 ms (-30.9%) NUM_FILES=1000, before patchset: 1085 ms NUM_FILES=1000, after patches 1/5 to 4/5 applied: 893 ms (-17.7%) NUM_FILES=1000, after whole patchset applied: 867 ms (-20.1%) Running dbench on the same physical machine with the following script: $ cat run-dbench.sh #!/bin/bash NUM_JOBS=$(nproc --all) DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 MOUNT_OPTIONS="-o ssd" MKFS_OPTIONS="-O no-holes -R free-space-tree" echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -D $MNT -t 120 $NUM_JOBS umount $MNT Before patchset: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 3761352 0.032 143.843 Close 2762770 0.002 2.273 Rename 159304 0.291 67.037 Unlink 759784 0.207 143.998 Deltree 72 4.028 15.977 Mkdir 36 0.003 0.006 Qpathinfo 3409780 0.013 9.678 Qfileinfo 596772 0.001 0.878 Qfsinfo 625189 0.003 1.245 Sfileinfo 306443 0.006 1.840 Find 1318106 0.063 19.798 WriteX 1871137 0.021 8.532 ReadX 5897325 0.003 3.567 LockX 12252 0.003 0.258 UnlockX 12252 0.002 0.100 Flush 263666 3.327 155.632 Throughput 980.047 MB/sec 12 clients 12 procs max_latency=155.636 ms After whole patchset applied: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 4195584 0.033 107.742 Close 3081932 0.002 1.935 Rename 177641 0.218 14.905 Unlink 847333 0.166 107.822 Deltree 118 5.315 15.247 Mkdir 59 0.004 0.048 Qpathinfo 3802612 0.014 10.302 Qfileinfo 666748 0.001 1.034 Qfsinfo 697329 0.003 0.944 Sfileinfo 341712 0.006 2.099 Find 1470365 0.065 9.359 WriteX 2093921 0.021 8.087 ReadX 6576234 0.003 3.407 LockX 13660 0.003 0.308 UnlockX 13660 0.002 0.114 Flush 294090 2.906 115.539 Throughput 1093.11 MB/sec 12 clients 12 procs max_latency=115.544 ms +11.5% throughput -25.8% max latency rename max latency -77.8% Link: https://bugzilla.opensuse.org/show_bug.cgi?id=1193549 Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-01-20 14:00:10 +03:00
* contention on the log tree. We can only rely on the value of
* last_dir_index_offset when we know for sure that the inode was
* previously logged in the current transaction.
btrfs: keep track of the last logged keys when logging a directory After the first time we log a directory in the current transaction, for each directory item in a changed leaf of the subvolume tree, we have to check if we previously logged the item, in order to overwrite it in case its data changed or skip it in case its data hasn't changed. Checking if we have logged each item before not only wastes times, but it also adds lock contention on the log tree. So in order to minimize the number of times we do such checks, keep track of the offset of the last key we logged for a directory and, on the next time we log the directory, skip the checks for any new keys that have an offset greater than the offset we have previously saved. This is specially effective for index keys, because the offset for these keys comes from a monotonically increasing counter. This patch is part of a patchset comprised of the following 5 patches: btrfs: remove root argument from btrfs_log_inode() and its callees btrfs: remove redundant log root assignment from log_dir_items() btrfs: factor out the copying loop of dir items from log_dir_items() btrfs: insert items in batches when logging a directory when possible btrfs: keep track of the last logged keys when logging a directory This is patch 5/5. The following test was used on a non-debug kernel to measure the impact it has on a directory fsync: $ cat test-dir-fsync.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=100000 NUM_FILE_DELETES=1000 mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done # fsync the directory, this will log the new dir items and the inodes # they point to, because these are new inodes. start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done # fsync the directory, this will only log dir items, there are no # dentries pointing to new inodes. start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" umount $MNT Test results with NUM_NEW_FILES set to 100 000 and 1 000 000: **** before patchset, 100 000 files, 1000 deletes **** dir fsync took 848 ms after adding 100000 files dir fsync took 175 ms after deleting 1000 files **** after patchset, 100 000 files, 1000 deletes **** dir fsync took 758 ms after adding 100000 files (-11.2%) dir fsync took 63 ms after deleting 1000 files (-94.1%) **** before patchset, 1 000 000 files, 1000 deletes **** dir fsync took 9945 ms after adding 1000000 files dir fsync took 473 ms after deleting 1000 files **** after patchset, 1 000 000 files, 1000 deletes **** dir fsync took 8677 ms after adding 1000000 files (-13.6%) dir fsync took 146 ms after deleting 1000 files (-105.6%) Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-16 13:32:14 +03:00
*/
btrfs: only copy dir index keys when logging a directory Currently, when logging a directory, we copy both dir items and dir index items from the fs/subvolume tree to the log tree. Both items have exactly the same data (same struct btrfs_dir_item), the difference lies in the key values, where a dir index key contains the index number of a directory entry while the dir item key does not, as it's used for doing fast lookups of an entry by name, while the former is used for sorting entries when listing a directory. We can exploit that and log only the dir index items, since they contain all the information needed to correctly add, replace and delete directory entries when replaying a log tree. Logging only the dir index items is also backward and forward compatible: an unpatched kernel (without this change) can correctly replay a log tree generated by a patched kernel (with this patch), and a patched kernel can correctly replay a log tree generated by an unpatched kernel. The backward compatibility is ensured because: 1) For inserting a new dentry: a dentry is only inserted when we find a new dir index key - we can only insert if we know the dir index offset, which is encoded in the dir index key's offset; 2) For deleting dentries: during log replay, before adding or replacing dentries, we first replay dentry deletions. Whenever we find a dir item key or a dir index key in the subvolume/fs tree that is not logged in a range for which the log tree is authoritative, we do the unlink of the dentry, which removes both the existing dir item key and the dir index key. Therefore logging just dir index keys is enough to ensure dentry deletions are correctly replayed; 3) For dentry replacements: they work when we log only dir index keys and this is mostly due to a combination of 1) and 2). If we replace a dentry with name "foobar" to point from inode A to inode B, then we know the dir index key for the new dentry is different from the old one, as it has an index number (key offset) larger than the old one. This results in replaying a deletion, through replay_dir_deletes(), that causes the old dentry to be removed, both the dir item key and the dir index key, as mentioned at 2). Then when processing the new dir index key, we add the new dentry, adding both a new dir item key and a new index key pointing to inode B, as stated in 1). The forward compatibility, the ability for a patched kernel to replay a log created by an older, unpatched kernel, comes from the changes required for making sure we are able to replay a log that only contains dir index keys - we simply ignore every dir item key we find. So modify directory logging to log only dir index items, and modify the log replay process to ignore dir item keys, from log trees created by an unpatched kernel, and process only with dir index keys. This reduces the amount of logged metadata by about half, and therefore the time spent logging or fsyncing large directories (less CPU time and less IO). The following test script was used to measure this change: #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=1000000 NUM_FILE_DELETES=10000 mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The tests were run on a physical machine, with a non-debug kernel (Debian's default kernel config), for different values of $NUM_NEW_FILES and $NUM_FILE_DELETES, and the results were the following: ** Before patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 8412 ms after adding 1000000 files dir fsync took 500 ms after deleting 10000 files ** After patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 4252 ms after adding 1000000 files (-49.5%) dir fsync took 269 ms after deleting 10000 files (-46.2%) ** Before patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 745 ms after adding 100000 files dir fsync took 59 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 404 ms after adding 100000 files (-45.8%) dir fsync took 31 ms after deleting 1000 files (-47.5%) ** Before patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 67 ms after adding 10000 files dir fsync took 9 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 36 ms after adding 10000 files (-46.3%) dir fsync took 5 ms after deleting 1000 files (-44.4%) ** Before patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 9 ms after adding 1000 files dir fsync took 4 ms after deleting 100 files ** After patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 7 ms after adding 1000 files (-22.2%) dir fsync took 3 ms after deleting 100 files (-25.0%) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-10-25 19:31:53 +03:00
if (key.offset > inode->last_dir_index_offset)
btrfs: keep track of the last logged keys when logging a directory After the first time we log a directory in the current transaction, for each directory item in a changed leaf of the subvolume tree, we have to check if we previously logged the item, in order to overwrite it in case its data changed or skip it in case its data hasn't changed. Checking if we have logged each item before not only wastes times, but it also adds lock contention on the log tree. So in order to minimize the number of times we do such checks, keep track of the offset of the last key we logged for a directory and, on the next time we log the directory, skip the checks for any new keys that have an offset greater than the offset we have previously saved. This is specially effective for index keys, because the offset for these keys comes from a monotonically increasing counter. This patch is part of a patchset comprised of the following 5 patches: btrfs: remove root argument from btrfs_log_inode() and its callees btrfs: remove redundant log root assignment from log_dir_items() btrfs: factor out the copying loop of dir items from log_dir_items() btrfs: insert items in batches when logging a directory when possible btrfs: keep track of the last logged keys when logging a directory This is patch 5/5. The following test was used on a non-debug kernel to measure the impact it has on a directory fsync: $ cat test-dir-fsync.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=100000 NUM_FILE_DELETES=1000 mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done # fsync the directory, this will log the new dir items and the inodes # they point to, because these are new inodes. start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done # fsync the directory, this will only log dir items, there are no # dentries pointing to new inodes. start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" umount $MNT Test results with NUM_NEW_FILES set to 100 000 and 1 000 000: **** before patchset, 100 000 files, 1000 deletes **** dir fsync took 848 ms after adding 100000 files dir fsync took 175 ms after deleting 1000 files **** after patchset, 100 000 files, 1000 deletes **** dir fsync took 758 ms after adding 100000 files (-11.2%) dir fsync took 63 ms after deleting 1000 files (-94.1%) **** before patchset, 1 000 000 files, 1000 deletes **** dir fsync took 9945 ms after adding 1000000 files dir fsync took 473 ms after deleting 1000 files **** after patchset, 1 000 000 files, 1000 deletes **** dir fsync took 8677 ms after adding 1000000 files (-13.6%) dir fsync took 146 ms after deleting 1000 files (-105.6%) Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-16 13:32:14 +03:00
goto add_to_batch;
btrfs: insert items in batches when logging a directory when possible When logging a directory, we scan its directory items from the subvolume tree and then copy one by one into the log tree. This is not efficient since we generally are able to insert several items in a batch, using a single btree operation for adding several items at once. The reason we copy items one by one is that we must check if each item was previously logged in the current transaction, and if it was we either overwrite it or skip it in case its content did not change in the subvolume tree (this can happen only for dir item keys, but not for dir index keys), and doing such check makes it a bit cumbersome to attempt batch insertions. However the chances for doing batch insertions are very frequent and always happen when: 1) Logging the directory for the first time in the current transaction, as none of the items exist in the log tree yet; 2) Logging new dir index keys, because the offset for new dir index keys comes from a monotonically increasing counter. This means if we keep adding dentries to a directory, through creation of new files and sub-directories or by adding new links or renaming from some other directory into the one we are logging, all the new dir index keys have a new offset that is greater than the offset of any previously logged index keys, so we can insert them in batches into the log tree. For dir item keys, since their offset depends on the result of an hash function against the dentry's name, unless the directory is being logged for the first time in the current transaction, the chances being able to insert the items in the log using batches is pretty much random and not predictable, as it depends on the names of the dentries, but still happens often enough. So change directory logging to keep track of consecutive directory items that don't exist yet in the log and batch insert them. This patch is part of a patchset comprised of the following 5 patches: btrfs: remove root argument from btrfs_log_inode() and its callees btrfs: remove redundant log root assignment from log_dir_items() btrfs: factor out the copying loop of dir items from log_dir_items() btrfs: insert items in batches when logging a directory when possible btrfs: keep track of the last logged keys when logging a directory This is patch 4/5. The change log of the last patch (5/5) has performance results. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-16 13:32:13 +03:00
/*
* Check if the key was already logged before. If not we can add
* it to a batch for bulk insertion.
*/
ret = btrfs_search_slot(NULL, log, &key, dst_path, 0, 0);
if (ret < 0) {
return ret;
} else if (ret > 0) {
btrfs_release_path(dst_path);
goto add_to_batch;
}
/*
* Item exists in the log. Overwrite the item in the log if it
* has different content or do nothing if it has exactly the same
* content. And then flush the current batch if any - do it after
* overwriting the current item, or we would deadlock otherwise,
* since we are holding a path for the existing item.
*/
ret = do_overwrite_item(trans, log, dst_path, src, i, &key);
if (ret < 0)
return ret;
if (batch_size > 0) {
ret = flush_dir_items_batch(trans, log, src, dst_path,
batch_start, batch_size);
if (ret < 0)
return ret;
batch_size = 0;
}
continue;
add_to_batch:
if (batch_size == 0)
batch_start = i;
batch_size++;
}
btrfs: insert items in batches when logging a directory when possible When logging a directory, we scan its directory items from the subvolume tree and then copy one by one into the log tree. This is not efficient since we generally are able to insert several items in a batch, using a single btree operation for adding several items at once. The reason we copy items one by one is that we must check if each item was previously logged in the current transaction, and if it was we either overwrite it or skip it in case its content did not change in the subvolume tree (this can happen only for dir item keys, but not for dir index keys), and doing such check makes it a bit cumbersome to attempt batch insertions. However the chances for doing batch insertions are very frequent and always happen when: 1) Logging the directory for the first time in the current transaction, as none of the items exist in the log tree yet; 2) Logging new dir index keys, because the offset for new dir index keys comes from a monotonically increasing counter. This means if we keep adding dentries to a directory, through creation of new files and sub-directories or by adding new links or renaming from some other directory into the one we are logging, all the new dir index keys have a new offset that is greater than the offset of any previously logged index keys, so we can insert them in batches into the log tree. For dir item keys, since their offset depends on the result of an hash function against the dentry's name, unless the directory is being logged for the first time in the current transaction, the chances being able to insert the items in the log using batches is pretty much random and not predictable, as it depends on the names of the dentries, but still happens often enough. So change directory logging to keep track of consecutive directory items that don't exist yet in the log and batch insert them. This patch is part of a patchset comprised of the following 5 patches: btrfs: remove root argument from btrfs_log_inode() and its callees btrfs: remove redundant log root assignment from log_dir_items() btrfs: factor out the copying loop of dir items from log_dir_items() btrfs: insert items in batches when logging a directory when possible btrfs: keep track of the last logged keys when logging a directory This is patch 4/5. The change log of the last patch (5/5) has performance results. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-16 13:32:13 +03:00
if (batch_size > 0) {
int ret;
ret = flush_dir_items_batch(trans, log, src, dst_path,
batch_start, batch_size);
if (ret < 0)
return ret;
}
return last_found ? 1 : 0;
}
/*
* log all the items included in the current transaction for a given
* directory. This also creates the range items in the log tree required
* to replay anything deleted before the fsync
*/
static noinline int log_dir_items(struct btrfs_trans_handle *trans,
struct btrfs_inode *inode,
struct btrfs_path *path,
btrfs: only copy dir index keys when logging a directory Currently, when logging a directory, we copy both dir items and dir index items from the fs/subvolume tree to the log tree. Both items have exactly the same data (same struct btrfs_dir_item), the difference lies in the key values, where a dir index key contains the index number of a directory entry while the dir item key does not, as it's used for doing fast lookups of an entry by name, while the former is used for sorting entries when listing a directory. We can exploit that and log only the dir index items, since they contain all the information needed to correctly add, replace and delete directory entries when replaying a log tree. Logging only the dir index items is also backward and forward compatible: an unpatched kernel (without this change) can correctly replay a log tree generated by a patched kernel (with this patch), and a patched kernel can correctly replay a log tree generated by an unpatched kernel. The backward compatibility is ensured because: 1) For inserting a new dentry: a dentry is only inserted when we find a new dir index key - we can only insert if we know the dir index offset, which is encoded in the dir index key's offset; 2) For deleting dentries: during log replay, before adding or replacing dentries, we first replay dentry deletions. Whenever we find a dir item key or a dir index key in the subvolume/fs tree that is not logged in a range for which the log tree is authoritative, we do the unlink of the dentry, which removes both the existing dir item key and the dir index key. Therefore logging just dir index keys is enough to ensure dentry deletions are correctly replayed; 3) For dentry replacements: they work when we log only dir index keys and this is mostly due to a combination of 1) and 2). If we replace a dentry with name "foobar" to point from inode A to inode B, then we know the dir index key for the new dentry is different from the old one, as it has an index number (key offset) larger than the old one. This results in replaying a deletion, through replay_dir_deletes(), that causes the old dentry to be removed, both the dir item key and the dir index key, as mentioned at 2). Then when processing the new dir index key, we add the new dentry, adding both a new dir item key and a new index key pointing to inode B, as stated in 1). The forward compatibility, the ability for a patched kernel to replay a log created by an older, unpatched kernel, comes from the changes required for making sure we are able to replay a log that only contains dir index keys - we simply ignore every dir item key we find. So modify directory logging to log only dir index items, and modify the log replay process to ignore dir item keys, from log trees created by an unpatched kernel, and process only with dir index keys. This reduces the amount of logged metadata by about half, and therefore the time spent logging or fsyncing large directories (less CPU time and less IO). The following test script was used to measure this change: #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=1000000 NUM_FILE_DELETES=10000 mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The tests were run on a physical machine, with a non-debug kernel (Debian's default kernel config), for different values of $NUM_NEW_FILES and $NUM_FILE_DELETES, and the results were the following: ** Before patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 8412 ms after adding 1000000 files dir fsync took 500 ms after deleting 10000 files ** After patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 4252 ms after adding 1000000 files (-49.5%) dir fsync took 269 ms after deleting 10000 files (-46.2%) ** Before patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 745 ms after adding 100000 files dir fsync took 59 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 404 ms after adding 100000 files (-45.8%) dir fsync took 31 ms after deleting 1000 files (-47.5%) ** Before patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 67 ms after adding 10000 files dir fsync took 9 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 36 ms after adding 10000 files (-46.3%) dir fsync took 5 ms after deleting 1000 files (-44.4%) ** Before patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 9 ms after adding 1000 files dir fsync took 4 ms after deleting 100 files ** After patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 7 ms after adding 1000 files (-22.2%) dir fsync took 3 ms after deleting 100 files (-25.0%) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-10-25 19:31:53 +03:00
struct btrfs_path *dst_path,
Btrfs: fix metadata inconsistencies after directory fsync We can get into inconsistency between inodes and directory entries after fsyncing a directory. The issue is that while a directory gets the new dentries persisted in the fsync log and replayed at mount time, the link count of the inode that directory entries point to doesn't get updated, staying with an incorrect link count (smaller then the correct value). This later leads to stale file handle errors when accessing (including attempt to delete) some of the links if all the other ones are removed, which also implies impossibility to delete the parent directories, since the dentries can not be removed. Another issue is that (unlike ext3/4, xfs, f2fs, reiserfs, nilfs2), when fsyncing a directory, new files aren't logged (their metadata and dentries) nor any child directories. So this patch fixes this issue too, since it has the same resolution as the incorrect inode link count issue mentioned before. This is very easy to reproduce, and the following excerpt from my test case for xfstests shows how: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create our main test file and directory. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 8K" $SCRATCH_MNT/foo | _filter_xfs_io mkdir $SCRATCH_MNT/mydir # Make sure all metadata and data are durably persisted. sync # Add a hard link to 'foo' inside our test directory and fsync only the # directory. The btrfs fsync implementation had a bug that caused the new # directory entry to be visible after the fsync log replay but, the inode # of our file remained with a link count of 1. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_2 # Add a few more links and new files. # This is just to verify nothing breaks or gives incorrect results after the # fsync log is replayed. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_3 $XFS_IO_PROG -f -c "pwrite -S 0xff 0 64K" $SCRATCH_MNT/hello | _filter_xfs_io ln $SCRATCH_MNT/hello $SCRATCH_MNT/mydir/hello_2 # Add some subdirectories and new files and links to them. This is to verify # that after fsyncing our top level directory 'mydir', all the subdirectories # and their files/links are registered in the fsync log and exist after the # fsync log is replayed. mkdir -p $SCRATCH_MNT/mydir/x/y/z ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/foo_y_link ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/z/foo_z_link touch $SCRATCH_MNT/mydir/x/y/z/qwerty # Now fsync only our top directory. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/mydir # And fsync now our new file named 'hello', just to verify later that it has # the expected content and that the previous fsync on the directory 'mydir' had # no bad influence on this fsync. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/hello # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # Verify the content of our file 'foo' remains the same as before, 8192 bytes, # all with the value 0xaa. echo "File 'foo' content after log replay:" od -t x1 $SCRATCH_MNT/foo # Remove the first name of our inode. Because of the directory fsync bug, the # inode's link count was 1 instead of 5, so removing the 'foo' name ended up # deleting the inode and the other names became stale directory entries (still # visible to applications). Attempting to remove or access the remaining # dentries pointing to that inode resulted in stale file handle errors and # made it impossible to remove the parent directories since it was impossible # for them to become empty. echo "file 'foo' link count after log replay: $(stat -c %h $SCRATCH_MNT/foo)" rm -f $SCRATCH_MNT/foo # Now verify that all files, links and directories created before fsyncing our # directory exist after the fsync log was replayed. [ -f $SCRATCH_MNT/mydir/foo_2 ] || echo "Link mydir/foo_2 is missing" [ -f $SCRATCH_MNT/mydir/foo_3 ] || echo "Link mydir/foo_3 is missing" [ -f $SCRATCH_MNT/hello ] || echo "File hello is missing" [ -f $SCRATCH_MNT/mydir/hello_2 ] || echo "Link mydir/hello_2 is missing" [ -f $SCRATCH_MNT/mydir/x/y/foo_y_link ] || \ echo "Link mydir/x/y/foo_y_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/foo_z_link ] || \ echo "Link mydir/x/y/z/foo_z_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/qwerty ] || \ echo "File mydir/x/y/z/qwerty is missing" # We expect our file here to have a size of 64Kb and all the bytes having the # value 0xff. echo "file 'hello' content after log replay:" od -t x1 $SCRATCH_MNT/hello # Now remove all files/links, under our test directory 'mydir', and verify we # can remove all the directories. rm -f $SCRATCH_MNT/mydir/x/y/z/* rmdir $SCRATCH_MNT/mydir/x/y/z rm -f $SCRATCH_MNT/mydir/x/y/* rmdir $SCRATCH_MNT/mydir/x/y rmdir $SCRATCH_MNT/mydir/x rm -f $SCRATCH_MNT/mydir/* rmdir $SCRATCH_MNT/mydir # An fsck, run by the fstests framework everytime a test finishes, also detected # the inconsistency and printed the following error message: # # root 5 inode 257 errors 2001, no inode item, link count wrong # unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref # unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref status=0 exit The expected golden output for the test is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 5 file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 Which is the output after this patch and when running the test against ext3/4, xfs, f2fs, reiserfs or nilfs2. Without this patch, the test's output is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 1 Link mydir/foo_2 is missing Link mydir/foo_3 is missing Link mydir/x/y/foo_y_link is missing Link mydir/x/y/z/foo_z_link is missing File mydir/x/y/z/qwerty is missing file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y/z': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x': No such file or directory rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_2': Stale file handle rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_3': Stale file handle rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir': Directory not empty Fsck, without this fix, also complains about the wrong link count: root 5 inode 257 errors 2001, no inode item, link count wrong unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref So fix this by logging the inodes that the dentries point to when fsyncing a directory. A test case for xfstests follows. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-03-20 20:19:46 +03:00
struct btrfs_log_ctx *ctx,
u64 min_offset, u64 *last_offset_ret)
{
struct btrfs_key min_key;
struct btrfs_root *root = inode->root;
struct btrfs_root *log = root->log_root;
int err = 0;
int ret;
btrfs: stop copying old dir items when logging a directory When logging a directory, we go over every leaf of the subvolume tree that was changed in the current transaction and copy all its dir index keys to the log tree. That includes copying dir index keys created in past transactions. This is done mostly for simplicity, as after logging the keys we log an item that specifies the start and end ranges of the keys we logged. That item is then used during log replay to figure out which keys need to be deleted - every key in that range that we find in the subvolume tree and is not in the log tree, needs to be deleted. Now that we log only dir index keys, and not dir item keys anymore, when we remove dentries from a directory (due to unlink and rename operations), we can get entire leaves that we changed only for deleting old dir index keys, or that have few dir index keys that are new - this is due to the fact that the offset for new index keys comes from a monotonically increasing counter. We can avoid logging dir index keys from past transactions, and in order to track the deletions, only log range items (BTRFS_DIR_LOG_INDEX_KEY key type) when we find gaps between consecutive index keys. This massively reduces the amount of logged metadata when we have deleted directory entries, even if it's a small percentage of the total number of entries. The reduction comes from both less items that are logged and instead of logging many dir index items (struct btrfs_dir_item), which have a size of 30 bytes plus a file name, we typically log just a few range items (struct btrfs_dir_log_item), which take only 8 bytes each. Even if no entries were deleted from a directory and only new entries were added, we typically still get a reduction on the amount of logged metadata, because it's very likely the first leaf that got the new dir index entries also has several old dir index entries. So change the logging logic to not log dir index keys created in past transactions and log a range item for every gap it finds between each pair of consecutive index keys, to ensure deletions are tracked and replayed on log replay. This patch is part of a patchset comprised of the following patches: 1/4 btrfs: don't log unnecessary boundary keys when logging directory 2/4 btrfs: put initial index value of a directory in a constant 3/4 btrfs: stop copying old dir items when logging a directory 4/4 btrfs: stop trying to log subdirectories created in past transactions The following test was run on a branch without this patchset and on a branch with the first three patches applied: $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=1000000 NUM_FILE_DELETES=10000 MKFS_OPTIONS="-O no-holes -R free-space-tree" MOUNT_OPTIONS="-o ssd" mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync del_inc=$(( $NUM_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The test was run on a non-debug kernel (Debian's default kernel config), and the results were the following for various values of NUM_FILES and NUM_FILE_DELETES: ** before, NUM_FILES = 1 000 000, NUM_FILE_DELETES = 10 000 ** dir fsync took 585 ms after deleting 10000 files ** after, NUM_FILES = 1 000 000, NUM_FILE_DELETES = 10 000 ** dir fsync took 34 ms after deleting 10000 files (-94.2%) ** before, NUM_FILES = 100 000, NUM_FILE_DELETES = 1 000 ** dir fsync took 50 ms after deleting 1000 files ** after, NUM_FILES = 100 000, NUM_FILE_DELETES = 1 000 ** dir fsync took 7 ms after deleting 1000 files (-86.0%) ** before, NUM_FILES = 10 000, NUM_FILE_DELETES = 100 ** dir fsync took 9 ms after deleting 100 files ** after, NUM_FILES = 10 000, NUM_FILE_DELETES = 100 ** dir fsync took 5 ms after deleting 100 files (-44.4%) Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-12-15 15:20:00 +03:00
u64 last_old_dentry_offset = min_offset - 1;
u64 last_offset = (u64)-1;
u64 ino = btrfs_ino(inode);
min_key.objectid = ino;
btrfs: only copy dir index keys when logging a directory Currently, when logging a directory, we copy both dir items and dir index items from the fs/subvolume tree to the log tree. Both items have exactly the same data (same struct btrfs_dir_item), the difference lies in the key values, where a dir index key contains the index number of a directory entry while the dir item key does not, as it's used for doing fast lookups of an entry by name, while the former is used for sorting entries when listing a directory. We can exploit that and log only the dir index items, since they contain all the information needed to correctly add, replace and delete directory entries when replaying a log tree. Logging only the dir index items is also backward and forward compatible: an unpatched kernel (without this change) can correctly replay a log tree generated by a patched kernel (with this patch), and a patched kernel can correctly replay a log tree generated by an unpatched kernel. The backward compatibility is ensured because: 1) For inserting a new dentry: a dentry is only inserted when we find a new dir index key - we can only insert if we know the dir index offset, which is encoded in the dir index key's offset; 2) For deleting dentries: during log replay, before adding or replacing dentries, we first replay dentry deletions. Whenever we find a dir item key or a dir index key in the subvolume/fs tree that is not logged in a range for which the log tree is authoritative, we do the unlink of the dentry, which removes both the existing dir item key and the dir index key. Therefore logging just dir index keys is enough to ensure dentry deletions are correctly replayed; 3) For dentry replacements: they work when we log only dir index keys and this is mostly due to a combination of 1) and 2). If we replace a dentry with name "foobar" to point from inode A to inode B, then we know the dir index key for the new dentry is different from the old one, as it has an index number (key offset) larger than the old one. This results in replaying a deletion, through replay_dir_deletes(), that causes the old dentry to be removed, both the dir item key and the dir index key, as mentioned at 2). Then when processing the new dir index key, we add the new dentry, adding both a new dir item key and a new index key pointing to inode B, as stated in 1). The forward compatibility, the ability for a patched kernel to replay a log created by an older, unpatched kernel, comes from the changes required for making sure we are able to replay a log that only contains dir index keys - we simply ignore every dir item key we find. So modify directory logging to log only dir index items, and modify the log replay process to ignore dir item keys, from log trees created by an unpatched kernel, and process only with dir index keys. This reduces the amount of logged metadata by about half, and therefore the time spent logging or fsyncing large directories (less CPU time and less IO). The following test script was used to measure this change: #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=1000000 NUM_FILE_DELETES=10000 mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The tests were run on a physical machine, with a non-debug kernel (Debian's default kernel config), for different values of $NUM_NEW_FILES and $NUM_FILE_DELETES, and the results were the following: ** Before patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 8412 ms after adding 1000000 files dir fsync took 500 ms after deleting 10000 files ** After patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 4252 ms after adding 1000000 files (-49.5%) dir fsync took 269 ms after deleting 10000 files (-46.2%) ** Before patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 745 ms after adding 100000 files dir fsync took 59 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 404 ms after adding 100000 files (-45.8%) dir fsync took 31 ms after deleting 1000 files (-47.5%) ** Before patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 67 ms after adding 10000 files dir fsync took 9 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 36 ms after adding 10000 files (-46.3%) dir fsync took 5 ms after deleting 1000 files (-44.4%) ** Before patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 9 ms after adding 1000 files dir fsync took 4 ms after deleting 100 files ** After patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 7 ms after adding 1000 files (-22.2%) dir fsync took 3 ms after deleting 100 files (-25.0%) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-10-25 19:31:53 +03:00
min_key.type = BTRFS_DIR_INDEX_KEY;
min_key.offset = min_offset;
ret = btrfs_search_forward(root, &min_key, path, trans->transid);
/*
* we didn't find anything from this transaction, see if there
* is anything at all
*/
btrfs: only copy dir index keys when logging a directory Currently, when logging a directory, we copy both dir items and dir index items from the fs/subvolume tree to the log tree. Both items have exactly the same data (same struct btrfs_dir_item), the difference lies in the key values, where a dir index key contains the index number of a directory entry while the dir item key does not, as it's used for doing fast lookups of an entry by name, while the former is used for sorting entries when listing a directory. We can exploit that and log only the dir index items, since they contain all the information needed to correctly add, replace and delete directory entries when replaying a log tree. Logging only the dir index items is also backward and forward compatible: an unpatched kernel (without this change) can correctly replay a log tree generated by a patched kernel (with this patch), and a patched kernel can correctly replay a log tree generated by an unpatched kernel. The backward compatibility is ensured because: 1) For inserting a new dentry: a dentry is only inserted when we find a new dir index key - we can only insert if we know the dir index offset, which is encoded in the dir index key's offset; 2) For deleting dentries: during log replay, before adding or replacing dentries, we first replay dentry deletions. Whenever we find a dir item key or a dir index key in the subvolume/fs tree that is not logged in a range for which the log tree is authoritative, we do the unlink of the dentry, which removes both the existing dir item key and the dir index key. Therefore logging just dir index keys is enough to ensure dentry deletions are correctly replayed; 3) For dentry replacements: they work when we log only dir index keys and this is mostly due to a combination of 1) and 2). If we replace a dentry with name "foobar" to point from inode A to inode B, then we know the dir index key for the new dentry is different from the old one, as it has an index number (key offset) larger than the old one. This results in replaying a deletion, through replay_dir_deletes(), that causes the old dentry to be removed, both the dir item key and the dir index key, as mentioned at 2). Then when processing the new dir index key, we add the new dentry, adding both a new dir item key and a new index key pointing to inode B, as stated in 1). The forward compatibility, the ability for a patched kernel to replay a log created by an older, unpatched kernel, comes from the changes required for making sure we are able to replay a log that only contains dir index keys - we simply ignore every dir item key we find. So modify directory logging to log only dir index items, and modify the log replay process to ignore dir item keys, from log trees created by an unpatched kernel, and process only with dir index keys. This reduces the amount of logged metadata by about half, and therefore the time spent logging or fsyncing large directories (less CPU time and less IO). The following test script was used to measure this change: #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=1000000 NUM_FILE_DELETES=10000 mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The tests were run on a physical machine, with a non-debug kernel (Debian's default kernel config), for different values of $NUM_NEW_FILES and $NUM_FILE_DELETES, and the results were the following: ** Before patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 8412 ms after adding 1000000 files dir fsync took 500 ms after deleting 10000 files ** After patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 4252 ms after adding 1000000 files (-49.5%) dir fsync took 269 ms after deleting 10000 files (-46.2%) ** Before patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 745 ms after adding 100000 files dir fsync took 59 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 404 ms after adding 100000 files (-45.8%) dir fsync took 31 ms after deleting 1000 files (-47.5%) ** Before patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 67 ms after adding 10000 files dir fsync took 9 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 36 ms after adding 10000 files (-46.3%) dir fsync took 5 ms after deleting 1000 files (-44.4%) ** Before patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 9 ms after adding 1000 files dir fsync took 4 ms after deleting 100 files ** After patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 7 ms after adding 1000 files (-22.2%) dir fsync took 3 ms after deleting 100 files (-25.0%) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-10-25 19:31:53 +03:00
if (ret != 0 || min_key.objectid != ino ||
min_key.type != BTRFS_DIR_INDEX_KEY) {
min_key.objectid = ino;
btrfs: only copy dir index keys when logging a directory Currently, when logging a directory, we copy both dir items and dir index items from the fs/subvolume tree to the log tree. Both items have exactly the same data (same struct btrfs_dir_item), the difference lies in the key values, where a dir index key contains the index number of a directory entry while the dir item key does not, as it's used for doing fast lookups of an entry by name, while the former is used for sorting entries when listing a directory. We can exploit that and log only the dir index items, since they contain all the information needed to correctly add, replace and delete directory entries when replaying a log tree. Logging only the dir index items is also backward and forward compatible: an unpatched kernel (without this change) can correctly replay a log tree generated by a patched kernel (with this patch), and a patched kernel can correctly replay a log tree generated by an unpatched kernel. The backward compatibility is ensured because: 1) For inserting a new dentry: a dentry is only inserted when we find a new dir index key - we can only insert if we know the dir index offset, which is encoded in the dir index key's offset; 2) For deleting dentries: during log replay, before adding or replacing dentries, we first replay dentry deletions. Whenever we find a dir item key or a dir index key in the subvolume/fs tree that is not logged in a range for which the log tree is authoritative, we do the unlink of the dentry, which removes both the existing dir item key and the dir index key. Therefore logging just dir index keys is enough to ensure dentry deletions are correctly replayed; 3) For dentry replacements: they work when we log only dir index keys and this is mostly due to a combination of 1) and 2). If we replace a dentry with name "foobar" to point from inode A to inode B, then we know the dir index key for the new dentry is different from the old one, as it has an index number (key offset) larger than the old one. This results in replaying a deletion, through replay_dir_deletes(), that causes the old dentry to be removed, both the dir item key and the dir index key, as mentioned at 2). Then when processing the new dir index key, we add the new dentry, adding both a new dir item key and a new index key pointing to inode B, as stated in 1). The forward compatibility, the ability for a patched kernel to replay a log created by an older, unpatched kernel, comes from the changes required for making sure we are able to replay a log that only contains dir index keys - we simply ignore every dir item key we find. So modify directory logging to log only dir index items, and modify the log replay process to ignore dir item keys, from log trees created by an unpatched kernel, and process only with dir index keys. This reduces the amount of logged metadata by about half, and therefore the time spent logging or fsyncing large directories (less CPU time and less IO). The following test script was used to measure this change: #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=1000000 NUM_FILE_DELETES=10000 mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The tests were run on a physical machine, with a non-debug kernel (Debian's default kernel config), for different values of $NUM_NEW_FILES and $NUM_FILE_DELETES, and the results were the following: ** Before patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 8412 ms after adding 1000000 files dir fsync took 500 ms after deleting 10000 files ** After patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 4252 ms after adding 1000000 files (-49.5%) dir fsync took 269 ms after deleting 10000 files (-46.2%) ** Before patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 745 ms after adding 100000 files dir fsync took 59 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 404 ms after adding 100000 files (-45.8%) dir fsync took 31 ms after deleting 1000 files (-47.5%) ** Before patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 67 ms after adding 10000 files dir fsync took 9 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 36 ms after adding 10000 files (-46.3%) dir fsync took 5 ms after deleting 1000 files (-44.4%) ** Before patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 9 ms after adding 1000 files dir fsync took 4 ms after deleting 100 files ** After patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 7 ms after adding 1000 files (-22.2%) dir fsync took 3 ms after deleting 100 files (-25.0%) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-10-25 19:31:53 +03:00
min_key.type = BTRFS_DIR_INDEX_KEY;
min_key.offset = (u64)-1;
btrfs_release_path(path);
ret = btrfs_search_slot(NULL, root, &min_key, path, 0, 0);
if (ret < 0) {
btrfs_release_path(path);
return ret;
}
btrfs: only copy dir index keys when logging a directory Currently, when logging a directory, we copy both dir items and dir index items from the fs/subvolume tree to the log tree. Both items have exactly the same data (same struct btrfs_dir_item), the difference lies in the key values, where a dir index key contains the index number of a directory entry while the dir item key does not, as it's used for doing fast lookups of an entry by name, while the former is used for sorting entries when listing a directory. We can exploit that and log only the dir index items, since they contain all the information needed to correctly add, replace and delete directory entries when replaying a log tree. Logging only the dir index items is also backward and forward compatible: an unpatched kernel (without this change) can correctly replay a log tree generated by a patched kernel (with this patch), and a patched kernel can correctly replay a log tree generated by an unpatched kernel. The backward compatibility is ensured because: 1) For inserting a new dentry: a dentry is only inserted when we find a new dir index key - we can only insert if we know the dir index offset, which is encoded in the dir index key's offset; 2) For deleting dentries: during log replay, before adding or replacing dentries, we first replay dentry deletions. Whenever we find a dir item key or a dir index key in the subvolume/fs tree that is not logged in a range for which the log tree is authoritative, we do the unlink of the dentry, which removes both the existing dir item key and the dir index key. Therefore logging just dir index keys is enough to ensure dentry deletions are correctly replayed; 3) For dentry replacements: they work when we log only dir index keys and this is mostly due to a combination of 1) and 2). If we replace a dentry with name "foobar" to point from inode A to inode B, then we know the dir index key for the new dentry is different from the old one, as it has an index number (key offset) larger than the old one. This results in replaying a deletion, through replay_dir_deletes(), that causes the old dentry to be removed, both the dir item key and the dir index key, as mentioned at 2). Then when processing the new dir index key, we add the new dentry, adding both a new dir item key and a new index key pointing to inode B, as stated in 1). The forward compatibility, the ability for a patched kernel to replay a log created by an older, unpatched kernel, comes from the changes required for making sure we are able to replay a log that only contains dir index keys - we simply ignore every dir item key we find. So modify directory logging to log only dir index items, and modify the log replay process to ignore dir item keys, from log trees created by an unpatched kernel, and process only with dir index keys. This reduces the amount of logged metadata by about half, and therefore the time spent logging or fsyncing large directories (less CPU time and less IO). The following test script was used to measure this change: #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=1000000 NUM_FILE_DELETES=10000 mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The tests were run on a physical machine, with a non-debug kernel (Debian's default kernel config), for different values of $NUM_NEW_FILES and $NUM_FILE_DELETES, and the results were the following: ** Before patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 8412 ms after adding 1000000 files dir fsync took 500 ms after deleting 10000 files ** After patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 4252 ms after adding 1000000 files (-49.5%) dir fsync took 269 ms after deleting 10000 files (-46.2%) ** Before patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 745 ms after adding 100000 files dir fsync took 59 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 404 ms after adding 100000 files (-45.8%) dir fsync took 31 ms after deleting 1000 files (-47.5%) ** Before patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 67 ms after adding 10000 files dir fsync took 9 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 36 ms after adding 10000 files (-46.3%) dir fsync took 5 ms after deleting 1000 files (-44.4%) ** Before patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 9 ms after adding 1000 files dir fsync took 4 ms after deleting 100 files ** After patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 7 ms after adding 1000 files (-22.2%) dir fsync took 3 ms after deleting 100 files (-25.0%) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-10-25 19:31:53 +03:00
ret = btrfs_previous_item(root, path, ino, BTRFS_DIR_INDEX_KEY);
/* if ret == 0 there are items for this type,
* create a range to tell us the last key of this type.
* otherwise, there are no items in this directory after
* *min_offset, and we create a range to indicate that.
*/
if (ret == 0) {
struct btrfs_key tmp;
btrfs: stop copying old dir items when logging a directory When logging a directory, we go over every leaf of the subvolume tree that was changed in the current transaction and copy all its dir index keys to the log tree. That includes copying dir index keys created in past transactions. This is done mostly for simplicity, as after logging the keys we log an item that specifies the start and end ranges of the keys we logged. That item is then used during log replay to figure out which keys need to be deleted - every key in that range that we find in the subvolume tree and is not in the log tree, needs to be deleted. Now that we log only dir index keys, and not dir item keys anymore, when we remove dentries from a directory (due to unlink and rename operations), we can get entire leaves that we changed only for deleting old dir index keys, or that have few dir index keys that are new - this is due to the fact that the offset for new index keys comes from a monotonically increasing counter. We can avoid logging dir index keys from past transactions, and in order to track the deletions, only log range items (BTRFS_DIR_LOG_INDEX_KEY key type) when we find gaps between consecutive index keys. This massively reduces the amount of logged metadata when we have deleted directory entries, even if it's a small percentage of the total number of entries. The reduction comes from both less items that are logged and instead of logging many dir index items (struct btrfs_dir_item), which have a size of 30 bytes plus a file name, we typically log just a few range items (struct btrfs_dir_log_item), which take only 8 bytes each. Even if no entries were deleted from a directory and only new entries were added, we typically still get a reduction on the amount of logged metadata, because it's very likely the first leaf that got the new dir index entries also has several old dir index entries. So change the logging logic to not log dir index keys created in past transactions and log a range item for every gap it finds between each pair of consecutive index keys, to ensure deletions are tracked and replayed on log replay. This patch is part of a patchset comprised of the following patches: 1/4 btrfs: don't log unnecessary boundary keys when logging directory 2/4 btrfs: put initial index value of a directory in a constant 3/4 btrfs: stop copying old dir items when logging a directory 4/4 btrfs: stop trying to log subdirectories created in past transactions The following test was run on a branch without this patchset and on a branch with the first three patches applied: $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=1000000 NUM_FILE_DELETES=10000 MKFS_OPTIONS="-O no-holes -R free-space-tree" MOUNT_OPTIONS="-o ssd" mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync del_inc=$(( $NUM_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The test was run on a non-debug kernel (Debian's default kernel config), and the results were the following for various values of NUM_FILES and NUM_FILE_DELETES: ** before, NUM_FILES = 1 000 000, NUM_FILE_DELETES = 10 000 ** dir fsync took 585 ms after deleting 10000 files ** after, NUM_FILES = 1 000 000, NUM_FILE_DELETES = 10 000 ** dir fsync took 34 ms after deleting 10000 files (-94.2%) ** before, NUM_FILES = 100 000, NUM_FILE_DELETES = 1 000 ** dir fsync took 50 ms after deleting 1000 files ** after, NUM_FILES = 100 000, NUM_FILE_DELETES = 1 000 ** dir fsync took 7 ms after deleting 1000 files (-86.0%) ** before, NUM_FILES = 10 000, NUM_FILE_DELETES = 100 ** dir fsync took 9 ms after deleting 100 files ** after, NUM_FILES = 10 000, NUM_FILE_DELETES = 100 ** dir fsync took 5 ms after deleting 100 files (-44.4%) Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-12-15 15:20:00 +03:00
btrfs_item_key_to_cpu(path->nodes[0], &tmp,
path->slots[0]);
btrfs: only copy dir index keys when logging a directory Currently, when logging a directory, we copy both dir items and dir index items from the fs/subvolume tree to the log tree. Both items have exactly the same data (same struct btrfs_dir_item), the difference lies in the key values, where a dir index key contains the index number of a directory entry while the dir item key does not, as it's used for doing fast lookups of an entry by name, while the former is used for sorting entries when listing a directory. We can exploit that and log only the dir index items, since they contain all the information needed to correctly add, replace and delete directory entries when replaying a log tree. Logging only the dir index items is also backward and forward compatible: an unpatched kernel (without this change) can correctly replay a log tree generated by a patched kernel (with this patch), and a patched kernel can correctly replay a log tree generated by an unpatched kernel. The backward compatibility is ensured because: 1) For inserting a new dentry: a dentry is only inserted when we find a new dir index key - we can only insert if we know the dir index offset, which is encoded in the dir index key's offset; 2) For deleting dentries: during log replay, before adding or replacing dentries, we first replay dentry deletions. Whenever we find a dir item key or a dir index key in the subvolume/fs tree that is not logged in a range for which the log tree is authoritative, we do the unlink of the dentry, which removes both the existing dir item key and the dir index key. Therefore logging just dir index keys is enough to ensure dentry deletions are correctly replayed; 3) For dentry replacements: they work when we log only dir index keys and this is mostly due to a combination of 1) and 2). If we replace a dentry with name "foobar" to point from inode A to inode B, then we know the dir index key for the new dentry is different from the old one, as it has an index number (key offset) larger than the old one. This results in replaying a deletion, through replay_dir_deletes(), that causes the old dentry to be removed, both the dir item key and the dir index key, as mentioned at 2). Then when processing the new dir index key, we add the new dentry, adding both a new dir item key and a new index key pointing to inode B, as stated in 1). The forward compatibility, the ability for a patched kernel to replay a log created by an older, unpatched kernel, comes from the changes required for making sure we are able to replay a log that only contains dir index keys - we simply ignore every dir item key we find. So modify directory logging to log only dir index items, and modify the log replay process to ignore dir item keys, from log trees created by an unpatched kernel, and process only with dir index keys. This reduces the amount of logged metadata by about half, and therefore the time spent logging or fsyncing large directories (less CPU time and less IO). The following test script was used to measure this change: #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=1000000 NUM_FILE_DELETES=10000 mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The tests were run on a physical machine, with a non-debug kernel (Debian's default kernel config), for different values of $NUM_NEW_FILES and $NUM_FILE_DELETES, and the results were the following: ** Before patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 8412 ms after adding 1000000 files dir fsync took 500 ms after deleting 10000 files ** After patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 4252 ms after adding 1000000 files (-49.5%) dir fsync took 269 ms after deleting 10000 files (-46.2%) ** Before patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 745 ms after adding 100000 files dir fsync took 59 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 404 ms after adding 100000 files (-45.8%) dir fsync took 31 ms after deleting 1000 files (-47.5%) ** Before patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 67 ms after adding 10000 files dir fsync took 9 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 36 ms after adding 10000 files (-46.3%) dir fsync took 5 ms after deleting 1000 files (-44.4%) ** Before patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 9 ms after adding 1000 files dir fsync took 4 ms after deleting 100 files ** After patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 7 ms after adding 1000 files (-22.2%) dir fsync took 3 ms after deleting 100 files (-25.0%) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-10-25 19:31:53 +03:00
if (tmp.type == BTRFS_DIR_INDEX_KEY)
btrfs: stop copying old dir items when logging a directory When logging a directory, we go over every leaf of the subvolume tree that was changed in the current transaction and copy all its dir index keys to the log tree. That includes copying dir index keys created in past transactions. This is done mostly for simplicity, as after logging the keys we log an item that specifies the start and end ranges of the keys we logged. That item is then used during log replay to figure out which keys need to be deleted - every key in that range that we find in the subvolume tree and is not in the log tree, needs to be deleted. Now that we log only dir index keys, and not dir item keys anymore, when we remove dentries from a directory (due to unlink and rename operations), we can get entire leaves that we changed only for deleting old dir index keys, or that have few dir index keys that are new - this is due to the fact that the offset for new index keys comes from a monotonically increasing counter. We can avoid logging dir index keys from past transactions, and in order to track the deletions, only log range items (BTRFS_DIR_LOG_INDEX_KEY key type) when we find gaps between consecutive index keys. This massively reduces the amount of logged metadata when we have deleted directory entries, even if it's a small percentage of the total number of entries. The reduction comes from both less items that are logged and instead of logging many dir index items (struct btrfs_dir_item), which have a size of 30 bytes plus a file name, we typically log just a few range items (struct btrfs_dir_log_item), which take only 8 bytes each. Even if no entries were deleted from a directory and only new entries were added, we typically still get a reduction on the amount of logged metadata, because it's very likely the first leaf that got the new dir index entries also has several old dir index entries. So change the logging logic to not log dir index keys created in past transactions and log a range item for every gap it finds between each pair of consecutive index keys, to ensure deletions are tracked and replayed on log replay. This patch is part of a patchset comprised of the following patches: 1/4 btrfs: don't log unnecessary boundary keys when logging directory 2/4 btrfs: put initial index value of a directory in a constant 3/4 btrfs: stop copying old dir items when logging a directory 4/4 btrfs: stop trying to log subdirectories created in past transactions The following test was run on a branch without this patchset and on a branch with the first three patches applied: $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=1000000 NUM_FILE_DELETES=10000 MKFS_OPTIONS="-O no-holes -R free-space-tree" MOUNT_OPTIONS="-o ssd" mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync del_inc=$(( $NUM_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The test was run on a non-debug kernel (Debian's default kernel config), and the results were the following for various values of NUM_FILES and NUM_FILE_DELETES: ** before, NUM_FILES = 1 000 000, NUM_FILE_DELETES = 10 000 ** dir fsync took 585 ms after deleting 10000 files ** after, NUM_FILES = 1 000 000, NUM_FILE_DELETES = 10 000 ** dir fsync took 34 ms after deleting 10000 files (-94.2%) ** before, NUM_FILES = 100 000, NUM_FILE_DELETES = 1 000 ** dir fsync took 50 ms after deleting 1000 files ** after, NUM_FILES = 100 000, NUM_FILE_DELETES = 1 000 ** dir fsync took 7 ms after deleting 1000 files (-86.0%) ** before, NUM_FILES = 10 000, NUM_FILE_DELETES = 100 ** dir fsync took 9 ms after deleting 100 files ** after, NUM_FILES = 10 000, NUM_FILE_DELETES = 100 ** dir fsync took 5 ms after deleting 100 files (-44.4%) Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-12-15 15:20:00 +03:00
last_old_dentry_offset = tmp.offset;
}
goto done;
}
/* go backward to find any previous key */
btrfs: only copy dir index keys when logging a directory Currently, when logging a directory, we copy both dir items and dir index items from the fs/subvolume tree to the log tree. Both items have exactly the same data (same struct btrfs_dir_item), the difference lies in the key values, where a dir index key contains the index number of a directory entry while the dir item key does not, as it's used for doing fast lookups of an entry by name, while the former is used for sorting entries when listing a directory. We can exploit that and log only the dir index items, since they contain all the information needed to correctly add, replace and delete directory entries when replaying a log tree. Logging only the dir index items is also backward and forward compatible: an unpatched kernel (without this change) can correctly replay a log tree generated by a patched kernel (with this patch), and a patched kernel can correctly replay a log tree generated by an unpatched kernel. The backward compatibility is ensured because: 1) For inserting a new dentry: a dentry is only inserted when we find a new dir index key - we can only insert if we know the dir index offset, which is encoded in the dir index key's offset; 2) For deleting dentries: during log replay, before adding or replacing dentries, we first replay dentry deletions. Whenever we find a dir item key or a dir index key in the subvolume/fs tree that is not logged in a range for which the log tree is authoritative, we do the unlink of the dentry, which removes both the existing dir item key and the dir index key. Therefore logging just dir index keys is enough to ensure dentry deletions are correctly replayed; 3) For dentry replacements: they work when we log only dir index keys and this is mostly due to a combination of 1) and 2). If we replace a dentry with name "foobar" to point from inode A to inode B, then we know the dir index key for the new dentry is different from the old one, as it has an index number (key offset) larger than the old one. This results in replaying a deletion, through replay_dir_deletes(), that causes the old dentry to be removed, both the dir item key and the dir index key, as mentioned at 2). Then when processing the new dir index key, we add the new dentry, adding both a new dir item key and a new index key pointing to inode B, as stated in 1). The forward compatibility, the ability for a patched kernel to replay a log created by an older, unpatched kernel, comes from the changes required for making sure we are able to replay a log that only contains dir index keys - we simply ignore every dir item key we find. So modify directory logging to log only dir index items, and modify the log replay process to ignore dir item keys, from log trees created by an unpatched kernel, and process only with dir index keys. This reduces the amount of logged metadata by about half, and therefore the time spent logging or fsyncing large directories (less CPU time and less IO). The following test script was used to measure this change: #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=1000000 NUM_FILE_DELETES=10000 mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The tests were run on a physical machine, with a non-debug kernel (Debian's default kernel config), for different values of $NUM_NEW_FILES and $NUM_FILE_DELETES, and the results were the following: ** Before patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 8412 ms after adding 1000000 files dir fsync took 500 ms after deleting 10000 files ** After patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 4252 ms after adding 1000000 files (-49.5%) dir fsync took 269 ms after deleting 10000 files (-46.2%) ** Before patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 745 ms after adding 100000 files dir fsync took 59 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 404 ms after adding 100000 files (-45.8%) dir fsync took 31 ms after deleting 1000 files (-47.5%) ** Before patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 67 ms after adding 10000 files dir fsync took 9 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 36 ms after adding 10000 files (-46.3%) dir fsync took 5 ms after deleting 1000 files (-44.4%) ** Before patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 9 ms after adding 1000 files dir fsync took 4 ms after deleting 100 files ** After patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 7 ms after adding 1000 files (-22.2%) dir fsync took 3 ms after deleting 100 files (-25.0%) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-10-25 19:31:53 +03:00
ret = btrfs_previous_item(root, path, ino, BTRFS_DIR_INDEX_KEY);
if (ret == 0) {
struct btrfs_key tmp;
btrfs: don't log unnecessary boundary keys when logging directory Before we start to log dir index keys from a leaf, we check if there is a previous index key, which normally is at the end of a leaf that was not changed in the current transaction. Then we log that key and set the start of logged range (item of type BTRFS_DIR_LOG_INDEX_KEY) to the offset of that key. This is to ensure that if there were deleted index keys between that key and the first key we are going to log, those deletions are replayed in case we need to replay to the log after a power failure. However we really don't need to log that previous key, we can just set the start of the logged range to that key's offset plus 1. This achieves the same and avoids logging one dir index key. The same logic is performed when we finish logging the index keys of a leaf and we find that the next leaf has index keys and was not changed in the current transaction. We are logging the first key of that next leaf and use its offset as the end of range we log. This is just to ensure that if there were deleted index keys between the last index key we logged and the first key of that next leaf, those index keys are deleted if we end up replaying the log. However that is not necessary, we can avoid logging that first index key of the next leaf and instead set the end of the logged range to match the offset of that index key minus 1. So avoid logging those index keys at the boundaries and adjust the start and end offsets of the logged ranges as described above. This patch is part of a patchset comprised of the following patches: 1/4 btrfs: don't log unnecessary boundary keys when logging directory 2/4 btrfs: put initial index value of a directory in a constant 3/4 btrfs: stop copying old dir items when logging a directory 4/4 btrfs: stop trying to log subdirectories created in past transactions Performance test results are listed in the changelog of patch 3/4. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-12-15 15:19:58 +03:00
btrfs_item_key_to_cpu(path->nodes[0], &tmp, path->slots[0]);
btrfs: don't log unnecessary boundary keys when logging directory Before we start to log dir index keys from a leaf, we check if there is a previous index key, which normally is at the end of a leaf that was not changed in the current transaction. Then we log that key and set the start of logged range (item of type BTRFS_DIR_LOG_INDEX_KEY) to the offset of that key. This is to ensure that if there were deleted index keys between that key and the first key we are going to log, those deletions are replayed in case we need to replay to the log after a power failure. However we really don't need to log that previous key, we can just set the start of the logged range to that key's offset plus 1. This achieves the same and avoids logging one dir index key. The same logic is performed when we finish logging the index keys of a leaf and we find that the next leaf has index keys and was not changed in the current transaction. We are logging the first key of that next leaf and use its offset as the end of range we log. This is just to ensure that if there were deleted index keys between the last index key we logged and the first key of that next leaf, those index keys are deleted if we end up replaying the log. However that is not necessary, we can avoid logging that first index key of the next leaf and instead set the end of the logged range to match the offset of that index key minus 1. So avoid logging those index keys at the boundaries and adjust the start and end offsets of the logged ranges as described above. This patch is part of a patchset comprised of the following patches: 1/4 btrfs: don't log unnecessary boundary keys when logging directory 2/4 btrfs: put initial index value of a directory in a constant 3/4 btrfs: stop copying old dir items when logging a directory 4/4 btrfs: stop trying to log subdirectories created in past transactions Performance test results are listed in the changelog of patch 3/4. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-12-15 15:19:58 +03:00
/*
* The dir index key before the first one we found that needs to
* be logged might be in a previous leaf, and there might be a
* gap between these keys, meaning that we had deletions that
* happened. So the key range item we log (key type
* BTRFS_DIR_LOG_INDEX_KEY) must cover a range that starts at the
* previous key's offset plus 1, so that those deletes are replayed.
*/
if (tmp.type == BTRFS_DIR_INDEX_KEY)
btrfs: stop copying old dir items when logging a directory When logging a directory, we go over every leaf of the subvolume tree that was changed in the current transaction and copy all its dir index keys to the log tree. That includes copying dir index keys created in past transactions. This is done mostly for simplicity, as after logging the keys we log an item that specifies the start and end ranges of the keys we logged. That item is then used during log replay to figure out which keys need to be deleted - every key in that range that we find in the subvolume tree and is not in the log tree, needs to be deleted. Now that we log only dir index keys, and not dir item keys anymore, when we remove dentries from a directory (due to unlink and rename operations), we can get entire leaves that we changed only for deleting old dir index keys, or that have few dir index keys that are new - this is due to the fact that the offset for new index keys comes from a monotonically increasing counter. We can avoid logging dir index keys from past transactions, and in order to track the deletions, only log range items (BTRFS_DIR_LOG_INDEX_KEY key type) when we find gaps between consecutive index keys. This massively reduces the amount of logged metadata when we have deleted directory entries, even if it's a small percentage of the total number of entries. The reduction comes from both less items that are logged and instead of logging many dir index items (struct btrfs_dir_item), which have a size of 30 bytes plus a file name, we typically log just a few range items (struct btrfs_dir_log_item), which take only 8 bytes each. Even if no entries were deleted from a directory and only new entries were added, we typically still get a reduction on the amount of logged metadata, because it's very likely the first leaf that got the new dir index entries also has several old dir index entries. So change the logging logic to not log dir index keys created in past transactions and log a range item for every gap it finds between each pair of consecutive index keys, to ensure deletions are tracked and replayed on log replay. This patch is part of a patchset comprised of the following patches: 1/4 btrfs: don't log unnecessary boundary keys when logging directory 2/4 btrfs: put initial index value of a directory in a constant 3/4 btrfs: stop copying old dir items when logging a directory 4/4 btrfs: stop trying to log subdirectories created in past transactions The following test was run on a branch without this patchset and on a branch with the first three patches applied: $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=1000000 NUM_FILE_DELETES=10000 MKFS_OPTIONS="-O no-holes -R free-space-tree" MOUNT_OPTIONS="-o ssd" mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync del_inc=$(( $NUM_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The test was run on a non-debug kernel (Debian's default kernel config), and the results were the following for various values of NUM_FILES and NUM_FILE_DELETES: ** before, NUM_FILES = 1 000 000, NUM_FILE_DELETES = 10 000 ** dir fsync took 585 ms after deleting 10000 files ** after, NUM_FILES = 1 000 000, NUM_FILE_DELETES = 10 000 ** dir fsync took 34 ms after deleting 10000 files (-94.2%) ** before, NUM_FILES = 100 000, NUM_FILE_DELETES = 1 000 ** dir fsync took 50 ms after deleting 1000 files ** after, NUM_FILES = 100 000, NUM_FILE_DELETES = 1 000 ** dir fsync took 7 ms after deleting 1000 files (-86.0%) ** before, NUM_FILES = 10 000, NUM_FILE_DELETES = 100 ** dir fsync took 9 ms after deleting 100 files ** after, NUM_FILES = 10 000, NUM_FILE_DELETES = 100 ** dir fsync took 5 ms after deleting 100 files (-44.4%) Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-12-15 15:20:00 +03:00
last_old_dentry_offset = tmp.offset;
}
btrfs_release_path(path);
/*
* Find the first key from this transaction again. See the note for
* log_new_dir_dentries, if we're logging a directory recursively we
* won't be holding its i_mutex, which means we can modify the directory
* while we're logging it. If we remove an entry between our first
* search and this search we'll not find the key again and can just
* bail.
*/
btrfs: reschedule if necessary when logging directory items Logging directories with many entries can take a significant amount of time, and in some cases monopolize a cpu/core for a long time if the logging task doesn't happen to block often enough. Johannes and Lu Fengqi reported test case generic/041 triggering a soft lockup when the kernel has CONFIG_SOFTLOCKUP_DETECTOR=y. For this test case we log an inode with 3002 hard links, and because the test removed one hard link before fsyncing the file, the inode logging causes the parent directory do be logged as well, which has 6004 directory items to log (3002 BTRFS_DIR_ITEM_KEY items plus 3002 BTRFS_DIR_INDEX_KEY items), so it can take a significant amount of time and trigger the soft lockup. So just make tree-log.c:log_dir_items() reschedule when necessary, releasing the current search path before doing so and then resume from where it was before the reschedule. The stack trace produced when the soft lockup happens is the following: [10480.277653] watchdog: BUG: soft lockup - CPU#2 stuck for 22s! [xfs_io:28172] [10480.279418] Modules linked in: dm_thin_pool dm_persistent_data (...) [10480.284915] irq event stamp: 29646366 [10480.285987] hardirqs last enabled at (29646365): [<ffffffff85249b66>] __slab_alloc.constprop.0+0x56/0x60 [10480.288482] hardirqs last disabled at (29646366): [<ffffffff8579b00d>] irqentry_enter+0x1d/0x50 [10480.290856] softirqs last enabled at (4612): [<ffffffff85a00323>] __do_softirq+0x323/0x56c [10480.293615] softirqs last disabled at (4483): [<ffffffff85800dbf>] asm_call_on_stack+0xf/0x20 [10480.296428] CPU: 2 PID: 28172 Comm: xfs_io Not tainted 5.9.0-rc4-default+ #1248 [10480.298948] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.12.0-59-gc9ba527-rebuilt.opensuse.org 04/01/2014 [10480.302455] RIP: 0010:__slab_alloc.constprop.0+0x19/0x60 [10480.304151] Code: 86 e8 31 75 21 00 66 66 2e 0f 1f 84 00 00 00 (...) [10480.309558] RSP: 0018:ffffadbe09397a58 EFLAGS: 00000282 [10480.311179] RAX: ffff8a495ab92840 RBX: 0000000000000282 RCX: 0000000000000006 [10480.313242] RDX: 0000000000000000 RSI: 0000000000000000 RDI: ffffffff85249b66 [10480.315260] RBP: ffff8a497d04b740 R08: 0000000000000001 R09: 0000000000000001 [10480.317229] R10: ffff8a497d044800 R11: ffff8a495ab93c40 R12: 0000000000000000 [10480.319169] R13: 0000000000000000 R14: 0000000000000c40 R15: ffffffffc01daf70 [10480.321104] FS: 00007fa1dc5c0e40(0000) GS:ffff8a497da00000(0000) knlGS:0000000000000000 [10480.323559] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [10480.325235] CR2: 00007fa1dc5befb8 CR3: 0000000004f8a006 CR4: 0000000000170ea0 [10480.327259] Call Trace: [10480.328286] ? overwrite_item+0x1f0/0x5a0 [btrfs] [10480.329784] __kmalloc+0x831/0xa20 [10480.331009] ? btrfs_get_32+0xb0/0x1d0 [btrfs] [10480.332464] overwrite_item+0x1f0/0x5a0 [btrfs] [10480.333948] log_dir_items+0x2ee/0x570 [btrfs] [10480.335413] log_directory_changes+0x82/0xd0 [btrfs] [10480.336926] btrfs_log_inode+0xc9b/0xda0 [btrfs] [10480.338374] ? init_once+0x20/0x20 [btrfs] [10480.339711] btrfs_log_inode_parent+0x8d3/0xd10 [btrfs] [10480.341257] ? dget_parent+0x97/0x2e0 [10480.342480] btrfs_log_dentry_safe+0x3a/0x50 [btrfs] [10480.343977] btrfs_sync_file+0x24b/0x5e0 [btrfs] [10480.345381] do_fsync+0x38/0x70 [10480.346483] __x64_sys_fsync+0x10/0x20 [10480.347703] do_syscall_64+0x2d/0x70 [10480.348891] entry_SYSCALL_64_after_hwframe+0x44/0xa9 [10480.350444] RIP: 0033:0x7fa1dc80970b [10480.351642] Code: 0f 05 48 3d 00 f0 ff ff 77 45 c3 0f 1f 40 00 48 (...) [10480.356952] RSP: 002b:00007fffb3d081d0 EFLAGS: 00000293 ORIG_RAX: 000000000000004a [10480.359458] RAX: ffffffffffffffda RBX: 0000562d93d45e40 RCX: 00007fa1dc80970b [10480.361426] RDX: 0000562d93d44ab0 RSI: 0000562d93d45e60 RDI: 0000000000000003 [10480.363367] RBP: 0000000000000001 R08: 0000000000000000 R09: 00007fa1dc7b2a40 [10480.365317] R10: 0000562d93d0e366 R11: 0000000000000293 R12: 0000000000000001 [10480.367299] R13: 0000562d93d45290 R14: 0000562d93d45e40 R15: 0000562d93d45e60 Link: https://lore.kernel.org/linux-btrfs/20180713090216.GC575@fnst.localdomain/ Reported-by: Johannes Thumshirn <johannes.thumshirn@wdc.com> CC: stable@vger.kernel.org # 4.4+ Tested-by: Johannes Thumshirn <johannes.thumshirn@wdc.com> Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-09-14 17:27:50 +03:00
search:
ret = btrfs_search_slot(NULL, root, &min_key, path, 0, 0);
if (ret != 0)
goto done;
/*
* we have a block from this transaction, log every item in it
* from our directory
*/
while (1) {
btrfs: stop copying old dir items when logging a directory When logging a directory, we go over every leaf of the subvolume tree that was changed in the current transaction and copy all its dir index keys to the log tree. That includes copying dir index keys created in past transactions. This is done mostly for simplicity, as after logging the keys we log an item that specifies the start and end ranges of the keys we logged. That item is then used during log replay to figure out which keys need to be deleted - every key in that range that we find in the subvolume tree and is not in the log tree, needs to be deleted. Now that we log only dir index keys, and not dir item keys anymore, when we remove dentries from a directory (due to unlink and rename operations), we can get entire leaves that we changed only for deleting old dir index keys, or that have few dir index keys that are new - this is due to the fact that the offset for new index keys comes from a monotonically increasing counter. We can avoid logging dir index keys from past transactions, and in order to track the deletions, only log range items (BTRFS_DIR_LOG_INDEX_KEY key type) when we find gaps between consecutive index keys. This massively reduces the amount of logged metadata when we have deleted directory entries, even if it's a small percentage of the total number of entries. The reduction comes from both less items that are logged and instead of logging many dir index items (struct btrfs_dir_item), which have a size of 30 bytes plus a file name, we typically log just a few range items (struct btrfs_dir_log_item), which take only 8 bytes each. Even if no entries were deleted from a directory and only new entries were added, we typically still get a reduction on the amount of logged metadata, because it's very likely the first leaf that got the new dir index entries also has several old dir index entries. So change the logging logic to not log dir index keys created in past transactions and log a range item for every gap it finds between each pair of consecutive index keys, to ensure deletions are tracked and replayed on log replay. This patch is part of a patchset comprised of the following patches: 1/4 btrfs: don't log unnecessary boundary keys when logging directory 2/4 btrfs: put initial index value of a directory in a constant 3/4 btrfs: stop copying old dir items when logging a directory 4/4 btrfs: stop trying to log subdirectories created in past transactions The following test was run on a branch without this patchset and on a branch with the first three patches applied: $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=1000000 NUM_FILE_DELETES=10000 MKFS_OPTIONS="-O no-holes -R free-space-tree" MOUNT_OPTIONS="-o ssd" mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync del_inc=$(( $NUM_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The test was run on a non-debug kernel (Debian's default kernel config), and the results were the following for various values of NUM_FILES and NUM_FILE_DELETES: ** before, NUM_FILES = 1 000 000, NUM_FILE_DELETES = 10 000 ** dir fsync took 585 ms after deleting 10000 files ** after, NUM_FILES = 1 000 000, NUM_FILE_DELETES = 10 000 ** dir fsync took 34 ms after deleting 10000 files (-94.2%) ** before, NUM_FILES = 100 000, NUM_FILE_DELETES = 1 000 ** dir fsync took 50 ms after deleting 1000 files ** after, NUM_FILES = 100 000, NUM_FILE_DELETES = 1 000 ** dir fsync took 7 ms after deleting 1000 files (-86.0%) ** before, NUM_FILES = 10 000, NUM_FILE_DELETES = 100 ** dir fsync took 9 ms after deleting 100 files ** after, NUM_FILES = 10 000, NUM_FILE_DELETES = 100 ** dir fsync took 5 ms after deleting 100 files (-44.4%) Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-12-15 15:20:00 +03:00
ret = process_dir_items_leaf(trans, inode, path, dst_path, ctx,
&last_old_dentry_offset);
if (ret != 0) {
if (ret < 0)
err = ret;
goto done;
}
path->slots[0] = btrfs_header_nritems(path->nodes[0]);
/*
* look ahead to the next item and see if it is also
* from this directory and from this transaction
*/
ret = btrfs_next_leaf(root, path);
if (ret) {
if (ret == 1)
last_offset = (u64)-1;
else
err = ret;
goto done;
}
btrfs_item_key_to_cpu(path->nodes[0], &min_key, path->slots[0]);
btrfs: only copy dir index keys when logging a directory Currently, when logging a directory, we copy both dir items and dir index items from the fs/subvolume tree to the log tree. Both items have exactly the same data (same struct btrfs_dir_item), the difference lies in the key values, where a dir index key contains the index number of a directory entry while the dir item key does not, as it's used for doing fast lookups of an entry by name, while the former is used for sorting entries when listing a directory. We can exploit that and log only the dir index items, since they contain all the information needed to correctly add, replace and delete directory entries when replaying a log tree. Logging only the dir index items is also backward and forward compatible: an unpatched kernel (without this change) can correctly replay a log tree generated by a patched kernel (with this patch), and a patched kernel can correctly replay a log tree generated by an unpatched kernel. The backward compatibility is ensured because: 1) For inserting a new dentry: a dentry is only inserted when we find a new dir index key - we can only insert if we know the dir index offset, which is encoded in the dir index key's offset; 2) For deleting dentries: during log replay, before adding or replacing dentries, we first replay dentry deletions. Whenever we find a dir item key or a dir index key in the subvolume/fs tree that is not logged in a range for which the log tree is authoritative, we do the unlink of the dentry, which removes both the existing dir item key and the dir index key. Therefore logging just dir index keys is enough to ensure dentry deletions are correctly replayed; 3) For dentry replacements: they work when we log only dir index keys and this is mostly due to a combination of 1) and 2). If we replace a dentry with name "foobar" to point from inode A to inode B, then we know the dir index key for the new dentry is different from the old one, as it has an index number (key offset) larger than the old one. This results in replaying a deletion, through replay_dir_deletes(), that causes the old dentry to be removed, both the dir item key and the dir index key, as mentioned at 2). Then when processing the new dir index key, we add the new dentry, adding both a new dir item key and a new index key pointing to inode B, as stated in 1). The forward compatibility, the ability for a patched kernel to replay a log created by an older, unpatched kernel, comes from the changes required for making sure we are able to replay a log that only contains dir index keys - we simply ignore every dir item key we find. So modify directory logging to log only dir index items, and modify the log replay process to ignore dir item keys, from log trees created by an unpatched kernel, and process only with dir index keys. This reduces the amount of logged metadata by about half, and therefore the time spent logging or fsyncing large directories (less CPU time and less IO). The following test script was used to measure this change: #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=1000000 NUM_FILE_DELETES=10000 mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The tests were run on a physical machine, with a non-debug kernel (Debian's default kernel config), for different values of $NUM_NEW_FILES and $NUM_FILE_DELETES, and the results were the following: ** Before patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 8412 ms after adding 1000000 files dir fsync took 500 ms after deleting 10000 files ** After patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 4252 ms after adding 1000000 files (-49.5%) dir fsync took 269 ms after deleting 10000 files (-46.2%) ** Before patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 745 ms after adding 100000 files dir fsync took 59 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 404 ms after adding 100000 files (-45.8%) dir fsync took 31 ms after deleting 1000 files (-47.5%) ** Before patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 67 ms after adding 10000 files dir fsync took 9 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 36 ms after adding 10000 files (-46.3%) dir fsync took 5 ms after deleting 1000 files (-44.4%) ** Before patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 9 ms after adding 1000 files dir fsync took 4 ms after deleting 100 files ** After patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 7 ms after adding 1000 files (-22.2%) dir fsync took 3 ms after deleting 100 files (-25.0%) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-10-25 19:31:53 +03:00
if (min_key.objectid != ino || min_key.type != BTRFS_DIR_INDEX_KEY) {
last_offset = (u64)-1;
goto done;
}
if (btrfs_header_generation(path->nodes[0]) != trans->transid) {
btrfs: don't log unnecessary boundary keys when logging directory Before we start to log dir index keys from a leaf, we check if there is a previous index key, which normally is at the end of a leaf that was not changed in the current transaction. Then we log that key and set the start of logged range (item of type BTRFS_DIR_LOG_INDEX_KEY) to the offset of that key. This is to ensure that if there were deleted index keys between that key and the first key we are going to log, those deletions are replayed in case we need to replay to the log after a power failure. However we really don't need to log that previous key, we can just set the start of the logged range to that key's offset plus 1. This achieves the same and avoids logging one dir index key. The same logic is performed when we finish logging the index keys of a leaf and we find that the next leaf has index keys and was not changed in the current transaction. We are logging the first key of that next leaf and use its offset as the end of range we log. This is just to ensure that if there were deleted index keys between the last index key we logged and the first key of that next leaf, those index keys are deleted if we end up replaying the log. However that is not necessary, we can avoid logging that first index key of the next leaf and instead set the end of the logged range to match the offset of that index key minus 1. So avoid logging those index keys at the boundaries and adjust the start and end offsets of the logged ranges as described above. This patch is part of a patchset comprised of the following patches: 1/4 btrfs: don't log unnecessary boundary keys when logging directory 2/4 btrfs: put initial index value of a directory in a constant 3/4 btrfs: stop copying old dir items when logging a directory 4/4 btrfs: stop trying to log subdirectories created in past transactions Performance test results are listed in the changelog of patch 3/4. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-12-15 15:19:58 +03:00
/*
* The next leaf was not changed in the current transaction
* and has at least one dir index key.
* We check for the next key because there might have been
* one or more deletions between the last key we logged and
* that next key. So the key range item we log (key type
* BTRFS_DIR_LOG_INDEX_KEY) must end at the next key's
* offset minus 1, so that those deletes are replayed.
*/
last_offset = min_key.offset - 1;
goto done;
}
if (need_resched()) {
btrfs_release_path(path);
cond_resched();
goto search;
}
}
done:
btrfs_release_path(path);
btrfs_release_path(dst_path);
if (err == 0) {
*last_offset_ret = last_offset;
/*
btrfs: stop copying old dir items when logging a directory When logging a directory, we go over every leaf of the subvolume tree that was changed in the current transaction and copy all its dir index keys to the log tree. That includes copying dir index keys created in past transactions. This is done mostly for simplicity, as after logging the keys we log an item that specifies the start and end ranges of the keys we logged. That item is then used during log replay to figure out which keys need to be deleted - every key in that range that we find in the subvolume tree and is not in the log tree, needs to be deleted. Now that we log only dir index keys, and not dir item keys anymore, when we remove dentries from a directory (due to unlink and rename operations), we can get entire leaves that we changed only for deleting old dir index keys, or that have few dir index keys that are new - this is due to the fact that the offset for new index keys comes from a monotonically increasing counter. We can avoid logging dir index keys from past transactions, and in order to track the deletions, only log range items (BTRFS_DIR_LOG_INDEX_KEY key type) when we find gaps between consecutive index keys. This massively reduces the amount of logged metadata when we have deleted directory entries, even if it's a small percentage of the total number of entries. The reduction comes from both less items that are logged and instead of logging many dir index items (struct btrfs_dir_item), which have a size of 30 bytes plus a file name, we typically log just a few range items (struct btrfs_dir_log_item), which take only 8 bytes each. Even if no entries were deleted from a directory and only new entries were added, we typically still get a reduction on the amount of logged metadata, because it's very likely the first leaf that got the new dir index entries also has several old dir index entries. So change the logging logic to not log dir index keys created in past transactions and log a range item for every gap it finds between each pair of consecutive index keys, to ensure deletions are tracked and replayed on log replay. This patch is part of a patchset comprised of the following patches: 1/4 btrfs: don't log unnecessary boundary keys when logging directory 2/4 btrfs: put initial index value of a directory in a constant 3/4 btrfs: stop copying old dir items when logging a directory 4/4 btrfs: stop trying to log subdirectories created in past transactions The following test was run on a branch without this patchset and on a branch with the first three patches applied: $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=1000000 NUM_FILE_DELETES=10000 MKFS_OPTIONS="-O no-holes -R free-space-tree" MOUNT_OPTIONS="-o ssd" mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync del_inc=$(( $NUM_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The test was run on a non-debug kernel (Debian's default kernel config), and the results were the following for various values of NUM_FILES and NUM_FILE_DELETES: ** before, NUM_FILES = 1 000 000, NUM_FILE_DELETES = 10 000 ** dir fsync took 585 ms after deleting 10000 files ** after, NUM_FILES = 1 000 000, NUM_FILE_DELETES = 10 000 ** dir fsync took 34 ms after deleting 10000 files (-94.2%) ** before, NUM_FILES = 100 000, NUM_FILE_DELETES = 1 000 ** dir fsync took 50 ms after deleting 1000 files ** after, NUM_FILES = 100 000, NUM_FILE_DELETES = 1 000 ** dir fsync took 7 ms after deleting 1000 files (-86.0%) ** before, NUM_FILES = 10 000, NUM_FILE_DELETES = 100 ** dir fsync took 9 ms after deleting 100 files ** after, NUM_FILES = 10 000, NUM_FILE_DELETES = 100 ** dir fsync took 5 ms after deleting 100 files (-44.4%) Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-12-15 15:20:00 +03:00
* In case the leaf was changed in the current transaction but
* all its dir items are from a past transaction, the last item
* in the leaf is a dir item and there's no gap between that last
* dir item and the first one on the next leaf (which did not
* change in the current transaction), then we don't need to log
* a range, last_old_dentry_offset is == to last_offset.
*/
btrfs: stop copying old dir items when logging a directory When logging a directory, we go over every leaf of the subvolume tree that was changed in the current transaction and copy all its dir index keys to the log tree. That includes copying dir index keys created in past transactions. This is done mostly for simplicity, as after logging the keys we log an item that specifies the start and end ranges of the keys we logged. That item is then used during log replay to figure out which keys need to be deleted - every key in that range that we find in the subvolume tree and is not in the log tree, needs to be deleted. Now that we log only dir index keys, and not dir item keys anymore, when we remove dentries from a directory (due to unlink and rename operations), we can get entire leaves that we changed only for deleting old dir index keys, or that have few dir index keys that are new - this is due to the fact that the offset for new index keys comes from a monotonically increasing counter. We can avoid logging dir index keys from past transactions, and in order to track the deletions, only log range items (BTRFS_DIR_LOG_INDEX_KEY key type) when we find gaps between consecutive index keys. This massively reduces the amount of logged metadata when we have deleted directory entries, even if it's a small percentage of the total number of entries. The reduction comes from both less items that are logged and instead of logging many dir index items (struct btrfs_dir_item), which have a size of 30 bytes plus a file name, we typically log just a few range items (struct btrfs_dir_log_item), which take only 8 bytes each. Even if no entries were deleted from a directory and only new entries were added, we typically still get a reduction on the amount of logged metadata, because it's very likely the first leaf that got the new dir index entries also has several old dir index entries. So change the logging logic to not log dir index keys created in past transactions and log a range item for every gap it finds between each pair of consecutive index keys, to ensure deletions are tracked and replayed on log replay. This patch is part of a patchset comprised of the following patches: 1/4 btrfs: don't log unnecessary boundary keys when logging directory 2/4 btrfs: put initial index value of a directory in a constant 3/4 btrfs: stop copying old dir items when logging a directory 4/4 btrfs: stop trying to log subdirectories created in past transactions The following test was run on a branch without this patchset and on a branch with the first three patches applied: $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=1000000 NUM_FILE_DELETES=10000 MKFS_OPTIONS="-O no-holes -R free-space-tree" MOUNT_OPTIONS="-o ssd" mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync del_inc=$(( $NUM_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The test was run on a non-debug kernel (Debian's default kernel config), and the results were the following for various values of NUM_FILES and NUM_FILE_DELETES: ** before, NUM_FILES = 1 000 000, NUM_FILE_DELETES = 10 000 ** dir fsync took 585 ms after deleting 10000 files ** after, NUM_FILES = 1 000 000, NUM_FILE_DELETES = 10 000 ** dir fsync took 34 ms after deleting 10000 files (-94.2%) ** before, NUM_FILES = 100 000, NUM_FILE_DELETES = 1 000 ** dir fsync took 50 ms after deleting 1000 files ** after, NUM_FILES = 100 000, NUM_FILE_DELETES = 1 000 ** dir fsync took 7 ms after deleting 1000 files (-86.0%) ** before, NUM_FILES = 10 000, NUM_FILE_DELETES = 100 ** dir fsync took 9 ms after deleting 100 files ** after, NUM_FILES = 10 000, NUM_FILE_DELETES = 100 ** dir fsync took 5 ms after deleting 100 files (-44.4%) Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-12-15 15:20:00 +03:00
ASSERT(last_old_dentry_offset <= last_offset);
if (last_old_dentry_offset < last_offset) {
ret = insert_dir_log_key(trans, log, path, ino,
last_old_dentry_offset + 1,
last_offset);
if (ret)
err = ret;
}
}
return err;
}
/*
* logging directories is very similar to logging inodes, We find all the items
* from the current transaction and write them to the log.
*
* The recovery code scans the directory in the subvolume, and if it finds a
* key in the range logged that is not present in the log tree, then it means
* that dir entry was unlinked during the transaction.
*
* In order for that scan to work, we must include one key smaller than
* the smallest logged by this transaction and one key larger than the largest
* key logged by this transaction.
*/
static noinline int log_directory_changes(struct btrfs_trans_handle *trans,
struct btrfs_inode *inode,
struct btrfs_path *path,
Btrfs: fix metadata inconsistencies after directory fsync We can get into inconsistency between inodes and directory entries after fsyncing a directory. The issue is that while a directory gets the new dentries persisted in the fsync log and replayed at mount time, the link count of the inode that directory entries point to doesn't get updated, staying with an incorrect link count (smaller then the correct value). This later leads to stale file handle errors when accessing (including attempt to delete) some of the links if all the other ones are removed, which also implies impossibility to delete the parent directories, since the dentries can not be removed. Another issue is that (unlike ext3/4, xfs, f2fs, reiserfs, nilfs2), when fsyncing a directory, new files aren't logged (their metadata and dentries) nor any child directories. So this patch fixes this issue too, since it has the same resolution as the incorrect inode link count issue mentioned before. This is very easy to reproduce, and the following excerpt from my test case for xfstests shows how: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create our main test file and directory. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 8K" $SCRATCH_MNT/foo | _filter_xfs_io mkdir $SCRATCH_MNT/mydir # Make sure all metadata and data are durably persisted. sync # Add a hard link to 'foo' inside our test directory and fsync only the # directory. The btrfs fsync implementation had a bug that caused the new # directory entry to be visible after the fsync log replay but, the inode # of our file remained with a link count of 1. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_2 # Add a few more links and new files. # This is just to verify nothing breaks or gives incorrect results after the # fsync log is replayed. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_3 $XFS_IO_PROG -f -c "pwrite -S 0xff 0 64K" $SCRATCH_MNT/hello | _filter_xfs_io ln $SCRATCH_MNT/hello $SCRATCH_MNT/mydir/hello_2 # Add some subdirectories and new files and links to them. This is to verify # that after fsyncing our top level directory 'mydir', all the subdirectories # and their files/links are registered in the fsync log and exist after the # fsync log is replayed. mkdir -p $SCRATCH_MNT/mydir/x/y/z ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/foo_y_link ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/z/foo_z_link touch $SCRATCH_MNT/mydir/x/y/z/qwerty # Now fsync only our top directory. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/mydir # And fsync now our new file named 'hello', just to verify later that it has # the expected content and that the previous fsync on the directory 'mydir' had # no bad influence on this fsync. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/hello # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # Verify the content of our file 'foo' remains the same as before, 8192 bytes, # all with the value 0xaa. echo "File 'foo' content after log replay:" od -t x1 $SCRATCH_MNT/foo # Remove the first name of our inode. Because of the directory fsync bug, the # inode's link count was 1 instead of 5, so removing the 'foo' name ended up # deleting the inode and the other names became stale directory entries (still # visible to applications). Attempting to remove or access the remaining # dentries pointing to that inode resulted in stale file handle errors and # made it impossible to remove the parent directories since it was impossible # for them to become empty. echo "file 'foo' link count after log replay: $(stat -c %h $SCRATCH_MNT/foo)" rm -f $SCRATCH_MNT/foo # Now verify that all files, links and directories created before fsyncing our # directory exist after the fsync log was replayed. [ -f $SCRATCH_MNT/mydir/foo_2 ] || echo "Link mydir/foo_2 is missing" [ -f $SCRATCH_MNT/mydir/foo_3 ] || echo "Link mydir/foo_3 is missing" [ -f $SCRATCH_MNT/hello ] || echo "File hello is missing" [ -f $SCRATCH_MNT/mydir/hello_2 ] || echo "Link mydir/hello_2 is missing" [ -f $SCRATCH_MNT/mydir/x/y/foo_y_link ] || \ echo "Link mydir/x/y/foo_y_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/foo_z_link ] || \ echo "Link mydir/x/y/z/foo_z_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/qwerty ] || \ echo "File mydir/x/y/z/qwerty is missing" # We expect our file here to have a size of 64Kb and all the bytes having the # value 0xff. echo "file 'hello' content after log replay:" od -t x1 $SCRATCH_MNT/hello # Now remove all files/links, under our test directory 'mydir', and verify we # can remove all the directories. rm -f $SCRATCH_MNT/mydir/x/y/z/* rmdir $SCRATCH_MNT/mydir/x/y/z rm -f $SCRATCH_MNT/mydir/x/y/* rmdir $SCRATCH_MNT/mydir/x/y rmdir $SCRATCH_MNT/mydir/x rm -f $SCRATCH_MNT/mydir/* rmdir $SCRATCH_MNT/mydir # An fsck, run by the fstests framework everytime a test finishes, also detected # the inconsistency and printed the following error message: # # root 5 inode 257 errors 2001, no inode item, link count wrong # unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref # unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref status=0 exit The expected golden output for the test is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 5 file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 Which is the output after this patch and when running the test against ext3/4, xfs, f2fs, reiserfs or nilfs2. Without this patch, the test's output is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 1 Link mydir/foo_2 is missing Link mydir/foo_3 is missing Link mydir/x/y/foo_y_link is missing Link mydir/x/y/z/foo_z_link is missing File mydir/x/y/z/qwerty is missing file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y/z': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x': No such file or directory rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_2': Stale file handle rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_3': Stale file handle rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir': Directory not empty Fsck, without this fix, also complains about the wrong link count: root 5 inode 257 errors 2001, no inode item, link count wrong unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref So fix this by logging the inodes that the dentries point to when fsyncing a directory. A test case for xfstests follows. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-03-20 20:19:46 +03:00
struct btrfs_path *dst_path,
struct btrfs_log_ctx *ctx)
{
u64 min_key;
u64 max_key;
int ret;
btrfs: stop copying old dir items when logging a directory When logging a directory, we go over every leaf of the subvolume tree that was changed in the current transaction and copy all its dir index keys to the log tree. That includes copying dir index keys created in past transactions. This is done mostly for simplicity, as after logging the keys we log an item that specifies the start and end ranges of the keys we logged. That item is then used during log replay to figure out which keys need to be deleted - every key in that range that we find in the subvolume tree and is not in the log tree, needs to be deleted. Now that we log only dir index keys, and not dir item keys anymore, when we remove dentries from a directory (due to unlink and rename operations), we can get entire leaves that we changed only for deleting old dir index keys, or that have few dir index keys that are new - this is due to the fact that the offset for new index keys comes from a monotonically increasing counter. We can avoid logging dir index keys from past transactions, and in order to track the deletions, only log range items (BTRFS_DIR_LOG_INDEX_KEY key type) when we find gaps between consecutive index keys. This massively reduces the amount of logged metadata when we have deleted directory entries, even if it's a small percentage of the total number of entries. The reduction comes from both less items that are logged and instead of logging many dir index items (struct btrfs_dir_item), which have a size of 30 bytes plus a file name, we typically log just a few range items (struct btrfs_dir_log_item), which take only 8 bytes each. Even if no entries were deleted from a directory and only new entries were added, we typically still get a reduction on the amount of logged metadata, because it's very likely the first leaf that got the new dir index entries also has several old dir index entries. So change the logging logic to not log dir index keys created in past transactions and log a range item for every gap it finds between each pair of consecutive index keys, to ensure deletions are tracked and replayed on log replay. This patch is part of a patchset comprised of the following patches: 1/4 btrfs: don't log unnecessary boundary keys when logging directory 2/4 btrfs: put initial index value of a directory in a constant 3/4 btrfs: stop copying old dir items when logging a directory 4/4 btrfs: stop trying to log subdirectories created in past transactions The following test was run on a branch without this patchset and on a branch with the first three patches applied: $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=1000000 NUM_FILE_DELETES=10000 MKFS_OPTIONS="-O no-holes -R free-space-tree" MOUNT_OPTIONS="-o ssd" mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync del_inc=$(( $NUM_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The test was run on a non-debug kernel (Debian's default kernel config), and the results were the following for various values of NUM_FILES and NUM_FILE_DELETES: ** before, NUM_FILES = 1 000 000, NUM_FILE_DELETES = 10 000 ** dir fsync took 585 ms after deleting 10000 files ** after, NUM_FILES = 1 000 000, NUM_FILE_DELETES = 10 000 ** dir fsync took 34 ms after deleting 10000 files (-94.2%) ** before, NUM_FILES = 100 000, NUM_FILE_DELETES = 1 000 ** dir fsync took 50 ms after deleting 1000 files ** after, NUM_FILES = 100 000, NUM_FILE_DELETES = 1 000 ** dir fsync took 7 ms after deleting 1000 files (-86.0%) ** before, NUM_FILES = 10 000, NUM_FILE_DELETES = 100 ** dir fsync took 9 ms after deleting 100 files ** after, NUM_FILES = 10 000, NUM_FILE_DELETES = 100 ** dir fsync took 5 ms after deleting 100 files (-44.4%) Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-12-15 15:20:00 +03:00
min_key = BTRFS_DIR_START_INDEX;
max_key = 0;
btrfs: only copy dir index keys when logging a directory Currently, when logging a directory, we copy both dir items and dir index items from the fs/subvolume tree to the log tree. Both items have exactly the same data (same struct btrfs_dir_item), the difference lies in the key values, where a dir index key contains the index number of a directory entry while the dir item key does not, as it's used for doing fast lookups of an entry by name, while the former is used for sorting entries when listing a directory. We can exploit that and log only the dir index items, since they contain all the information needed to correctly add, replace and delete directory entries when replaying a log tree. Logging only the dir index items is also backward and forward compatible: an unpatched kernel (without this change) can correctly replay a log tree generated by a patched kernel (with this patch), and a patched kernel can correctly replay a log tree generated by an unpatched kernel. The backward compatibility is ensured because: 1) For inserting a new dentry: a dentry is only inserted when we find a new dir index key - we can only insert if we know the dir index offset, which is encoded in the dir index key's offset; 2) For deleting dentries: during log replay, before adding or replacing dentries, we first replay dentry deletions. Whenever we find a dir item key or a dir index key in the subvolume/fs tree that is not logged in a range for which the log tree is authoritative, we do the unlink of the dentry, which removes both the existing dir item key and the dir index key. Therefore logging just dir index keys is enough to ensure dentry deletions are correctly replayed; 3) For dentry replacements: they work when we log only dir index keys and this is mostly due to a combination of 1) and 2). If we replace a dentry with name "foobar" to point from inode A to inode B, then we know the dir index key for the new dentry is different from the old one, as it has an index number (key offset) larger than the old one. This results in replaying a deletion, through replay_dir_deletes(), that causes the old dentry to be removed, both the dir item key and the dir index key, as mentioned at 2). Then when processing the new dir index key, we add the new dentry, adding both a new dir item key and a new index key pointing to inode B, as stated in 1). The forward compatibility, the ability for a patched kernel to replay a log created by an older, unpatched kernel, comes from the changes required for making sure we are able to replay a log that only contains dir index keys - we simply ignore every dir item key we find. So modify directory logging to log only dir index items, and modify the log replay process to ignore dir item keys, from log trees created by an unpatched kernel, and process only with dir index keys. This reduces the amount of logged metadata by about half, and therefore the time spent logging or fsyncing large directories (less CPU time and less IO). The following test script was used to measure this change: #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=1000000 NUM_FILE_DELETES=10000 mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The tests were run on a physical machine, with a non-debug kernel (Debian's default kernel config), for different values of $NUM_NEW_FILES and $NUM_FILE_DELETES, and the results were the following: ** Before patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 8412 ms after adding 1000000 files dir fsync took 500 ms after deleting 10000 files ** After patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 4252 ms after adding 1000000 files (-49.5%) dir fsync took 269 ms after deleting 10000 files (-46.2%) ** Before patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 745 ms after adding 100000 files dir fsync took 59 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 404 ms after adding 100000 files (-45.8%) dir fsync took 31 ms after deleting 1000 files (-47.5%) ** Before patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 67 ms after adding 10000 files dir fsync took 9 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 36 ms after adding 10000 files (-46.3%) dir fsync took 5 ms after deleting 1000 files (-44.4%) ** Before patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 9 ms after adding 1000 files dir fsync took 4 ms after deleting 100 files ** After patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 7 ms after adding 1000 files (-22.2%) dir fsync took 3 ms after deleting 100 files (-25.0%) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-10-25 19:31:53 +03:00
ctx->last_dir_item_offset = inode->last_dir_index_offset;
btrfs: keep track of the last logged keys when logging a directory After the first time we log a directory in the current transaction, for each directory item in a changed leaf of the subvolume tree, we have to check if we previously logged the item, in order to overwrite it in case its data changed or skip it in case its data hasn't changed. Checking if we have logged each item before not only wastes times, but it also adds lock contention on the log tree. So in order to minimize the number of times we do such checks, keep track of the offset of the last key we logged for a directory and, on the next time we log the directory, skip the checks for any new keys that have an offset greater than the offset we have previously saved. This is specially effective for index keys, because the offset for these keys comes from a monotonically increasing counter. This patch is part of a patchset comprised of the following 5 patches: btrfs: remove root argument from btrfs_log_inode() and its callees btrfs: remove redundant log root assignment from log_dir_items() btrfs: factor out the copying loop of dir items from log_dir_items() btrfs: insert items in batches when logging a directory when possible btrfs: keep track of the last logged keys when logging a directory This is patch 5/5. The following test was used on a non-debug kernel to measure the impact it has on a directory fsync: $ cat test-dir-fsync.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=100000 NUM_FILE_DELETES=1000 mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done # fsync the directory, this will log the new dir items and the inodes # they point to, because these are new inodes. start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done # fsync the directory, this will only log dir items, there are no # dentries pointing to new inodes. start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" umount $MNT Test results with NUM_NEW_FILES set to 100 000 and 1 000 000: **** before patchset, 100 000 files, 1000 deletes **** dir fsync took 848 ms after adding 100000 files dir fsync took 175 ms after deleting 1000 files **** after patchset, 100 000 files, 1000 deletes **** dir fsync took 758 ms after adding 100000 files (-11.2%) dir fsync took 63 ms after deleting 1000 files (-94.1%) **** before patchset, 1 000 000 files, 1000 deletes **** dir fsync took 9945 ms after adding 1000000 files dir fsync took 473 ms after deleting 1000 files **** after patchset, 1 000 000 files, 1000 deletes **** dir fsync took 8677 ms after adding 1000000 files (-13.6%) dir fsync took 146 ms after deleting 1000 files (-105.6%) Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-16 13:32:14 +03:00
while (1) {
btrfs: only copy dir index keys when logging a directory Currently, when logging a directory, we copy both dir items and dir index items from the fs/subvolume tree to the log tree. Both items have exactly the same data (same struct btrfs_dir_item), the difference lies in the key values, where a dir index key contains the index number of a directory entry while the dir item key does not, as it's used for doing fast lookups of an entry by name, while the former is used for sorting entries when listing a directory. We can exploit that and log only the dir index items, since they contain all the information needed to correctly add, replace and delete directory entries when replaying a log tree. Logging only the dir index items is also backward and forward compatible: an unpatched kernel (without this change) can correctly replay a log tree generated by a patched kernel (with this patch), and a patched kernel can correctly replay a log tree generated by an unpatched kernel. The backward compatibility is ensured because: 1) For inserting a new dentry: a dentry is only inserted when we find a new dir index key - we can only insert if we know the dir index offset, which is encoded in the dir index key's offset; 2) For deleting dentries: during log replay, before adding or replacing dentries, we first replay dentry deletions. Whenever we find a dir item key or a dir index key in the subvolume/fs tree that is not logged in a range for which the log tree is authoritative, we do the unlink of the dentry, which removes both the existing dir item key and the dir index key. Therefore logging just dir index keys is enough to ensure dentry deletions are correctly replayed; 3) For dentry replacements: they work when we log only dir index keys and this is mostly due to a combination of 1) and 2). If we replace a dentry with name "foobar" to point from inode A to inode B, then we know the dir index key for the new dentry is different from the old one, as it has an index number (key offset) larger than the old one. This results in replaying a deletion, through replay_dir_deletes(), that causes the old dentry to be removed, both the dir item key and the dir index key, as mentioned at 2). Then when processing the new dir index key, we add the new dentry, adding both a new dir item key and a new index key pointing to inode B, as stated in 1). The forward compatibility, the ability for a patched kernel to replay a log created by an older, unpatched kernel, comes from the changes required for making sure we are able to replay a log that only contains dir index keys - we simply ignore every dir item key we find. So modify directory logging to log only dir index items, and modify the log replay process to ignore dir item keys, from log trees created by an unpatched kernel, and process only with dir index keys. This reduces the amount of logged metadata by about half, and therefore the time spent logging or fsyncing large directories (less CPU time and less IO). The following test script was used to measure this change: #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=1000000 NUM_FILE_DELETES=10000 mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The tests were run on a physical machine, with a non-debug kernel (Debian's default kernel config), for different values of $NUM_NEW_FILES and $NUM_FILE_DELETES, and the results were the following: ** Before patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 8412 ms after adding 1000000 files dir fsync took 500 ms after deleting 10000 files ** After patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 4252 ms after adding 1000000 files (-49.5%) dir fsync took 269 ms after deleting 10000 files (-46.2%) ** Before patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 745 ms after adding 100000 files dir fsync took 59 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 404 ms after adding 100000 files (-45.8%) dir fsync took 31 ms after deleting 1000 files (-47.5%) ** Before patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 67 ms after adding 10000 files dir fsync took 9 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 36 ms after adding 10000 files (-46.3%) dir fsync took 5 ms after deleting 1000 files (-44.4%) ** Before patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 9 ms after adding 1000 files dir fsync took 4 ms after deleting 100 files ** After patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 7 ms after adding 1000 files (-22.2%) dir fsync took 3 ms after deleting 100 files (-25.0%) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-10-25 19:31:53 +03:00
ret = log_dir_items(trans, inode, path, dst_path,
ctx, min_key, &max_key);
if (ret)
return ret;
if (max_key == (u64)-1)
break;
min_key = max_key + 1;
}
btrfs: only copy dir index keys when logging a directory Currently, when logging a directory, we copy both dir items and dir index items from the fs/subvolume tree to the log tree. Both items have exactly the same data (same struct btrfs_dir_item), the difference lies in the key values, where a dir index key contains the index number of a directory entry while the dir item key does not, as it's used for doing fast lookups of an entry by name, while the former is used for sorting entries when listing a directory. We can exploit that and log only the dir index items, since they contain all the information needed to correctly add, replace and delete directory entries when replaying a log tree. Logging only the dir index items is also backward and forward compatible: an unpatched kernel (without this change) can correctly replay a log tree generated by a patched kernel (with this patch), and a patched kernel can correctly replay a log tree generated by an unpatched kernel. The backward compatibility is ensured because: 1) For inserting a new dentry: a dentry is only inserted when we find a new dir index key - we can only insert if we know the dir index offset, which is encoded in the dir index key's offset; 2) For deleting dentries: during log replay, before adding or replacing dentries, we first replay dentry deletions. Whenever we find a dir item key or a dir index key in the subvolume/fs tree that is not logged in a range for which the log tree is authoritative, we do the unlink of the dentry, which removes both the existing dir item key and the dir index key. Therefore logging just dir index keys is enough to ensure dentry deletions are correctly replayed; 3) For dentry replacements: they work when we log only dir index keys and this is mostly due to a combination of 1) and 2). If we replace a dentry with name "foobar" to point from inode A to inode B, then we know the dir index key for the new dentry is different from the old one, as it has an index number (key offset) larger than the old one. This results in replaying a deletion, through replay_dir_deletes(), that causes the old dentry to be removed, both the dir item key and the dir index key, as mentioned at 2). Then when processing the new dir index key, we add the new dentry, adding both a new dir item key and a new index key pointing to inode B, as stated in 1). The forward compatibility, the ability for a patched kernel to replay a log created by an older, unpatched kernel, comes from the changes required for making sure we are able to replay a log that only contains dir index keys - we simply ignore every dir item key we find. So modify directory logging to log only dir index items, and modify the log replay process to ignore dir item keys, from log trees created by an unpatched kernel, and process only with dir index keys. This reduces the amount of logged metadata by about half, and therefore the time spent logging or fsyncing large directories (less CPU time and less IO). The following test script was used to measure this change: #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=1000000 NUM_FILE_DELETES=10000 mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The tests were run on a physical machine, with a non-debug kernel (Debian's default kernel config), for different values of $NUM_NEW_FILES and $NUM_FILE_DELETES, and the results were the following: ** Before patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 8412 ms after adding 1000000 files dir fsync took 500 ms after deleting 10000 files ** After patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 4252 ms after adding 1000000 files (-49.5%) dir fsync took 269 ms after deleting 10000 files (-46.2%) ** Before patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 745 ms after adding 100000 files dir fsync took 59 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 404 ms after adding 100000 files (-45.8%) dir fsync took 31 ms after deleting 1000 files (-47.5%) ** Before patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 67 ms after adding 10000 files dir fsync took 9 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 36 ms after adding 10000 files (-46.3%) dir fsync took 5 ms after deleting 1000 files (-44.4%) ** Before patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 9 ms after adding 1000 files dir fsync took 4 ms after deleting 100 files ** After patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 7 ms after adding 1000 files (-22.2%) dir fsync took 3 ms after deleting 100 files (-25.0%) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-10-25 19:31:53 +03:00
inode->last_dir_index_offset = ctx->last_dir_item_offset;
return 0;
}
/*
* a helper function to drop items from the log before we relog an
* inode. max_key_type indicates the highest item type to remove.
* This cannot be run for file data extents because it does not
* free the extents they point to.
*/
btrfs: avoid expensive search when dropping inode items from log Whenever we are logging a directory inode, logging that an inode exists or logging an inode that has changes in its references or xattrs, we attempt to delete items of this inode we may have previously logged (through calls to drop_objectid_items()). That attempt does a btree search for deletion, which is expensive because it always acquires write locks for extent buffers at levels 2, 1 and 0, and it balances any node that is less than half full. Acquiring the write locks can block the task if the extent buffers are already locked or block other tasks attempting to lock them, which is specially bad in case of log trees since they are small due to their short life, with a root node at a level typically not greater than level 2. If we know that we are logging the inode for the first time in the current transaction, we can skip the search. This change does that. This patch is part of a patch set comprised of the following patches: btrfs: check if a log tree exists at inode_logged() btrfs: remove no longer needed checks for NULL log context btrfs: do not log new dentries when logging that a new name exists btrfs: always update the logged transaction when logging new names btrfs: avoid expensive search when dropping inode items from log btrfs: add helper to truncate inode items when logging inode btrfs: avoid expensive search when truncating inode items from the log btrfs: avoid search for logged i_size when logging inode if possible btrfs: avoid attempt to drop extents when logging inode for the first time btrfs: do not commit delayed inode when logging a file in full sync mode This is patch 5/10 and test results are listed in the change log of the last patch in the set. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-08-31 17:30:35 +03:00
static int drop_inode_items(struct btrfs_trans_handle *trans,
struct btrfs_root *log,
struct btrfs_path *path,
btrfs: avoid expensive search when dropping inode items from log Whenever we are logging a directory inode, logging that an inode exists or logging an inode that has changes in its references or xattrs, we attempt to delete items of this inode we may have previously logged (through calls to drop_objectid_items()). That attempt does a btree search for deletion, which is expensive because it always acquires write locks for extent buffers at levels 2, 1 and 0, and it balances any node that is less than half full. Acquiring the write locks can block the task if the extent buffers are already locked or block other tasks attempting to lock them, which is specially bad in case of log trees since they are small due to their short life, with a root node at a level typically not greater than level 2. If we know that we are logging the inode for the first time in the current transaction, we can skip the search. This change does that. This patch is part of a patch set comprised of the following patches: btrfs: check if a log tree exists at inode_logged() btrfs: remove no longer needed checks for NULL log context btrfs: do not log new dentries when logging that a new name exists btrfs: always update the logged transaction when logging new names btrfs: avoid expensive search when dropping inode items from log btrfs: add helper to truncate inode items when logging inode btrfs: avoid expensive search when truncating inode items from the log btrfs: avoid search for logged i_size when logging inode if possible btrfs: avoid attempt to drop extents when logging inode for the first time btrfs: do not commit delayed inode when logging a file in full sync mode This is patch 5/10 and test results are listed in the change log of the last patch in the set. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-08-31 17:30:35 +03:00
struct btrfs_inode *inode,
int max_key_type)
{
int ret;
struct btrfs_key key;
struct btrfs_key found_key;
int start_slot;
btrfs: avoid expensive search when dropping inode items from log Whenever we are logging a directory inode, logging that an inode exists or logging an inode that has changes in its references or xattrs, we attempt to delete items of this inode we may have previously logged (through calls to drop_objectid_items()). That attempt does a btree search for deletion, which is expensive because it always acquires write locks for extent buffers at levels 2, 1 and 0, and it balances any node that is less than half full. Acquiring the write locks can block the task if the extent buffers are already locked or block other tasks attempting to lock them, which is specially bad in case of log trees since they are small due to their short life, with a root node at a level typically not greater than level 2. If we know that we are logging the inode for the first time in the current transaction, we can skip the search. This change does that. This patch is part of a patch set comprised of the following patches: btrfs: check if a log tree exists at inode_logged() btrfs: remove no longer needed checks for NULL log context btrfs: do not log new dentries when logging that a new name exists btrfs: always update the logged transaction when logging new names btrfs: avoid expensive search when dropping inode items from log btrfs: add helper to truncate inode items when logging inode btrfs: avoid expensive search when truncating inode items from the log btrfs: avoid search for logged i_size when logging inode if possible btrfs: avoid attempt to drop extents when logging inode for the first time btrfs: do not commit delayed inode when logging a file in full sync mode This is patch 5/10 and test results are listed in the change log of the last patch in the set. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-08-31 17:30:35 +03:00
key.objectid = btrfs_ino(inode);
key.type = max_key_type;
key.offset = (u64)-1;
while (1) {
ret = btrfs_search_slot(trans, log, &key, path, -1, 1);
BUG_ON(ret == 0); /* Logic error */
if (ret < 0)
break;
if (path->slots[0] == 0)
break;
path->slots[0]--;
btrfs_item_key_to_cpu(path->nodes[0], &found_key,
path->slots[0]);
btrfs: avoid expensive search when dropping inode items from log Whenever we are logging a directory inode, logging that an inode exists or logging an inode that has changes in its references or xattrs, we attempt to delete items of this inode we may have previously logged (through calls to drop_objectid_items()). That attempt does a btree search for deletion, which is expensive because it always acquires write locks for extent buffers at levels 2, 1 and 0, and it balances any node that is less than half full. Acquiring the write locks can block the task if the extent buffers are already locked or block other tasks attempting to lock them, which is specially bad in case of log trees since they are small due to their short life, with a root node at a level typically not greater than level 2. If we know that we are logging the inode for the first time in the current transaction, we can skip the search. This change does that. This patch is part of a patch set comprised of the following patches: btrfs: check if a log tree exists at inode_logged() btrfs: remove no longer needed checks for NULL log context btrfs: do not log new dentries when logging that a new name exists btrfs: always update the logged transaction when logging new names btrfs: avoid expensive search when dropping inode items from log btrfs: add helper to truncate inode items when logging inode btrfs: avoid expensive search when truncating inode items from the log btrfs: avoid search for logged i_size when logging inode if possible btrfs: avoid attempt to drop extents when logging inode for the first time btrfs: do not commit delayed inode when logging a file in full sync mode This is patch 5/10 and test results are listed in the change log of the last patch in the set. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-08-31 17:30:35 +03:00
if (found_key.objectid != key.objectid)
break;
found_key.offset = 0;
found_key.type = 0;
ret = btrfs_bin_search(path->nodes[0], &found_key, &start_slot);
if (ret < 0)
break;
ret = btrfs_del_items(trans, log, path, start_slot,
path->slots[0] - start_slot + 1);
/*
* If start slot isn't 0 then we don't need to re-search, we've
* found the last guy with the objectid in this tree.
*/
if (ret || start_slot != 0)
break;
btrfs_release_path(path);
}
btrfs_release_path(path);
if (ret > 0)
ret = 0;
return ret;
}
static int truncate_inode_items(struct btrfs_trans_handle *trans,
struct btrfs_root *log_root,
struct btrfs_inode *inode,
u64 new_size, u32 min_type)
{
struct btrfs_truncate_control control = {
.new_size = new_size,
.ino = btrfs_ino(inode),
.min_type = min_type,
.skip_ref_updates = true,
};
return btrfs_truncate_inode_items(trans, log_root, &control);
}
static void fill_inode_item(struct btrfs_trans_handle *trans,
struct extent_buffer *leaf,
struct btrfs_inode_item *item,
Btrfs: fix fsync data loss after adding hard link to inode We have a scenario where after the fsync log replay we can lose file data that had been previously fsync'ed if we added an hard link for our inode and after that we sync'ed the fsync log (for example by fsync'ing some other file or directory). This is because when adding an hard link we updated the inode item in the log tree with an i_size value of 0. At that point the new inode item was in memory only and a subsequent fsync log replay would not make us lose the file data. However if after adding the hard link we sync the log tree to disk, by fsync'ing some other file or directory for example, we ended up losing the file data after log replay, because the inode item in the persisted log tree had an an i_size of zero. This is easy to reproduce, and the following excerpt from my test for xfstests shows this: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create one file with data and fsync it. # This made the btrfs fsync log persist the data and the inode metadata with # a correct inode->i_size (4096 bytes). $XFS_IO_PROG -f -c "pwrite -S 0xaa -b 4K 0 4K" -c "fsync" \ $SCRATCH_MNT/foo | _filter_xfs_io # Now add one hard link to our file. This made the btrfs code update the fsync # log, in memory only, with an inode metadata having a size of 0. ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link # Now force persistence of the fsync log to disk, for example, by fsyncing some # other file. touch $SCRATCH_MNT/bar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/bar # Before a power loss or crash, we could read the 4Kb of data from our file as # expected. echo "File content before:" od -t x1 $SCRATCH_MNT/foo # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # After the fsync log replay, because the fsync log had a value of 0 for our # inode's i_size, we couldn't read anymore the 4Kb of data that we previously # wrote and fsync'ed. The size of the file became 0 after the fsync log replay. echo "File content after:" od -t x1 $SCRATCH_MNT/foo Another alternative test, that doesn't need to fsync an inode in the same transaction it was created, is: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create our test file with some data. $XFS_IO_PROG -f -c "pwrite -S 0xaa -b 8K 0 8K" \ $SCRATCH_MNT/foo | _filter_xfs_io # Make sure the file is durably persisted. sync # Append some data to our file, to increase its size. $XFS_IO_PROG -f -c "pwrite -S 0xcc -b 4K 8K 4K" \ $SCRATCH_MNT/foo | _filter_xfs_io # Fsync the file, so from this point on if a crash/power failure happens, our # new data is guaranteed to be there next time the fs is mounted. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foo # Add one hard link to our file. This made btrfs write into the in memory fsync # log a special inode with generation 0 and an i_size of 0 too. Note that this # didn't update the inode in the fsync log on disk. ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link # Now make sure the in memory fsync log is durably persisted. # Creating and fsync'ing another file will do it. touch $SCRATCH_MNT/bar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/bar # As expected, before the crash/power failure, we should be able to read the # 12Kb of file data. echo "File content before:" od -t x1 $SCRATCH_MNT/foo # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # After mounting the fs again, the fsync log was replayed. # The btrfs fsync log replay code didn't update the i_size of the persisted # inode because the inode item in the log had a special generation with a # value of 0 (and it couldn't know the correct i_size, since that inode item # had a 0 i_size too). This made the last 4Kb of file data inaccessible and # effectively lost. echo "File content after:" od -t x1 $SCRATCH_MNT/foo This isn't a new issue/regression. This problem has been around since the log tree code was added in 2008: Btrfs: Add a write ahead tree log to optimize synchronous operations (commit e02119d5a7b4396c5a872582fddc8bd6d305a70a) Test cases for xfstests follow soon. CC: <stable@vger.kernel.org> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-02-13 15:30:56 +03:00
struct inode *inode, int log_inode_only,
u64 logged_isize)
{
struct btrfs_map_token token;
btrfs: add ro compat flags to inodes Currently, inode flags are fully backwards incompatible in btrfs. If we introduce a new inode flag, then tree-checker will detect it and fail. This can even cause us to fail to mount entirely. To make it possible to introduce new flags which can be read-only compatible, like VERITY, we add new ro flags to btrfs without treating them quite so harshly in tree-checker. A read-only file system can survive an unexpected flag, and can be mounted. As for the implementation, it unfortunately gets a little complicated. The on-disk representation of the inode, btrfs_inode_item, has an __le64 for flags but the in-memory representation, btrfs_inode, uses a u32. David Sterba had the nice idea that we could reclaim those wasted 32 bits on disk and use them for the new ro_compat flags. It turns out that the tree-checker code which checks for unknown flags is broken, and ignores the upper 32 bits we are hoping to use. The issue is that the flags use the literal 1 rather than 1ULL, so the flags are signed ints, and one of them is specifically (1 << 31). As a result, the mask which ORs the flags is a negative integer on machines where int is 32 bit twos complement. When tree-checker evaluates the expression: btrfs_inode_flags(leaf, iitem) & ~BTRFS_INODE_FLAG_MASK) The mask is something like 0x80000abc, which gets promoted to u64 with sign extension to 0xffffffff80000abc. Negating that 64 bit mask leaves all the upper bits zeroed, and we can't detect unexpected flags. This suggests that we can't use those bits after all. Luckily, we have good reason to believe that they are zero anyway. Inode flags are metadata, which is always checksummed, so any bit flips that would introduce 1s would cause a checksum failure anyway (excluding the improbable case of the checksum getting corrupted exactly badly). Further, unless the 1 << 31 flag is used, the cast to u64 of the 32 bit inode flag should preserve its value and not add leading zeroes (at least for twos complement). The only place that flag (BTRFS_INODE_ROOT_ITEM_INIT) is used is in a special inode embedded in the root item, and indeed for that inode we see 0xffffffff80000000 as the flags on disk. However, that inode is never seen by tree checker, nor is it used in a context where verity might be meaningful. Theoretically, a future ro flag might cause trouble on that inode, so we should proactively clean up that mess before it does. With the introduction of the new ro flags, keep two separate unsigned masks and check them against the appropriate u32. Since we no longer run afoul of sign extension, this also stops writing out 0xffffffff80000000 in root_item inodes going forward. Signed-off-by: Boris Burkov <boris@bur.io> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-06-30 23:01:48 +03:00
u64 flags;
btrfs_init_map_token(&token, leaf);
if (log_inode_only) {
/* set the generation to zero so the recover code
* can tell the difference between an logging
* just to say 'this inode exists' and a logging
* to say 'update this inode with these values'
*/
btrfs_set_token_inode_generation(&token, item, 0);
btrfs_set_token_inode_size(&token, item, logged_isize);
} else {
btrfs_set_token_inode_generation(&token, item,
BTRFS_I(inode)->generation);
btrfs_set_token_inode_size(&token, item, inode->i_size);
}
btrfs_set_token_inode_uid(&token, item, i_uid_read(inode));
btrfs_set_token_inode_gid(&token, item, i_gid_read(inode));
btrfs_set_token_inode_mode(&token, item, inode->i_mode);
btrfs_set_token_inode_nlink(&token, item, inode->i_nlink);
btrfs_set_token_timespec_sec(&token, &item->atime,
inode->i_atime.tv_sec);
btrfs_set_token_timespec_nsec(&token, &item->atime,
inode->i_atime.tv_nsec);
btrfs_set_token_timespec_sec(&token, &item->mtime,
inode->i_mtime.tv_sec);
btrfs_set_token_timespec_nsec(&token, &item->mtime,
inode->i_mtime.tv_nsec);
btrfs_set_token_timespec_sec(&token, &item->ctime,
inode->i_ctime.tv_sec);
btrfs_set_token_timespec_nsec(&token, &item->ctime,
inode->i_ctime.tv_nsec);
btrfs: stop setting nbytes when filling inode item for logging When we fill an inode item for logging we are setting its nbytes field with the value returned by inode_get_bytes() (a VFS API), however we do not need it because it is not used during log replay. In fact, for fast fsyncs, when we call inode_get_bytes() we may even get an outdated value for nbytes because the nbytes field of the inode is only updated when ordered extents complete, and a fast fsync only waits for writeback to complete, it does not wait for ordered extent completion. So just remove the setup of nbytes and add an explicit comment mentioning why we do not set it. This also avoids adding contention on the inode's i_lock (VFS) with concurrent stat() calls, since that spinlock is used by inode_get_bytes() which is also called by our stat callback (btrfs_getattr()). This patch is part of a patchset comprised of the following patches: btrfs: remove unnecessary directory inode item update when deleting dir entry btrfs: stop setting nbytes when filling inode item for logging btrfs: avoid logging new ancestor inodes when logging new inode btrfs: skip logging directories already logged when logging all parents btrfs: skip logging inodes already logged when logging new entries btrfs: remove unnecessary check_parent_dirs_for_sync() btrfs: make concurrent fsyncs wait less when waiting for a transaction commit Performance results, after applying all patches, are mentioned in the change log of the last patch. Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-01-27 13:34:55 +03:00
/*
* We do not need to set the nbytes field, in fact during a fast fsync
* its value may not even be correct, since a fast fsync does not wait
* for ordered extent completion, which is where we update nbytes, it
* only waits for writeback to complete. During log replay as we find
* file extent items and replay them, we adjust the nbytes field of the
* inode item in subvolume tree as needed (see overwrite_item()).
*/
btrfs_set_token_inode_sequence(&token, item, inode_peek_iversion(inode));
btrfs_set_token_inode_transid(&token, item, trans->transid);
btrfs_set_token_inode_rdev(&token, item, inode->i_rdev);
btrfs: add ro compat flags to inodes Currently, inode flags are fully backwards incompatible in btrfs. If we introduce a new inode flag, then tree-checker will detect it and fail. This can even cause us to fail to mount entirely. To make it possible to introduce new flags which can be read-only compatible, like VERITY, we add new ro flags to btrfs without treating them quite so harshly in tree-checker. A read-only file system can survive an unexpected flag, and can be mounted. As for the implementation, it unfortunately gets a little complicated. The on-disk representation of the inode, btrfs_inode_item, has an __le64 for flags but the in-memory representation, btrfs_inode, uses a u32. David Sterba had the nice idea that we could reclaim those wasted 32 bits on disk and use them for the new ro_compat flags. It turns out that the tree-checker code which checks for unknown flags is broken, and ignores the upper 32 bits we are hoping to use. The issue is that the flags use the literal 1 rather than 1ULL, so the flags are signed ints, and one of them is specifically (1 << 31). As a result, the mask which ORs the flags is a negative integer on machines where int is 32 bit twos complement. When tree-checker evaluates the expression: btrfs_inode_flags(leaf, iitem) & ~BTRFS_INODE_FLAG_MASK) The mask is something like 0x80000abc, which gets promoted to u64 with sign extension to 0xffffffff80000abc. Negating that 64 bit mask leaves all the upper bits zeroed, and we can't detect unexpected flags. This suggests that we can't use those bits after all. Luckily, we have good reason to believe that they are zero anyway. Inode flags are metadata, which is always checksummed, so any bit flips that would introduce 1s would cause a checksum failure anyway (excluding the improbable case of the checksum getting corrupted exactly badly). Further, unless the 1 << 31 flag is used, the cast to u64 of the 32 bit inode flag should preserve its value and not add leading zeroes (at least for twos complement). The only place that flag (BTRFS_INODE_ROOT_ITEM_INIT) is used is in a special inode embedded in the root item, and indeed for that inode we see 0xffffffff80000000 as the flags on disk. However, that inode is never seen by tree checker, nor is it used in a context where verity might be meaningful. Theoretically, a future ro flag might cause trouble on that inode, so we should proactively clean up that mess before it does. With the introduction of the new ro flags, keep two separate unsigned masks and check them against the appropriate u32. Since we no longer run afoul of sign extension, this also stops writing out 0xffffffff80000000 in root_item inodes going forward. Signed-off-by: Boris Burkov <boris@bur.io> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-06-30 23:01:48 +03:00
flags = btrfs_inode_combine_flags(BTRFS_I(inode)->flags,
BTRFS_I(inode)->ro_flags);
btrfs_set_token_inode_flags(&token, item, flags);
btrfs_set_token_inode_block_group(&token, item, 0);
}
static int log_inode_item(struct btrfs_trans_handle *trans,
struct btrfs_root *log, struct btrfs_path *path,
btrfs: avoid unnecessary lock and leaf splits when updating inode in the log During a fast fsync, if we have already fsynced the file before and in the current transaction, we can make the inode item update more efficient and avoid acquiring a write lock on the leaf's parent. To update the inode item we are always using btrfs_insert_empty_item() to get a path pointing to the inode item, which calls btrfs_search_slot() with an "ins_len" argument of 'sizeof(struct btrfs_inode_item) + sizeof(struct btrfs_item)', and that always results in the search taking a write lock on the level 1 node that is the parent of the leaf that contains the inode item. This adds unnecessary lock contention on log trees when we have multiple fsyncs in parallel against inodes in the same subvolume, which has a very significant impact due to the fact that log trees are short lived and their height very rarely goes beyond level 2. Also, by using btrfs_insert_empty_item() when we need to update the inode item, we also end up splitting the leaf of the existing inode item when the leaf has an amount of free space smaller than the size of an inode item. Improve this by using btrfs_seach_slot(), with a 0 "ins_len" argument, when we know the inode item already exists in the log. This avoids these two inefficiencies. The following script, using fio, was used to perform the tests: $ cat fio-test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 MOUNT_OPTIONS="-o ssd" MKFS_OPTIONS="-d single -m single" if [ $# -ne 4 ]; then echo "Use $0 NUM_JOBS FILE_SIZE FSYNC_FREQ BLOCK_SIZE" exit 1 fi NUM_JOBS=$1 FILE_SIZE=$2 FSYNC_FREQ=$3 BLOCK_SIZE=$4 cat <<EOF > /tmp/fio-job.ini [writers] rw=randwrite fsync=$FSYNC_FREQ fallocate=none group_reporting=1 direct=0 bs=$BLOCK_SIZE ioengine=sync size=$FILE_SIZE directory=$MNT numjobs=$NUM_JOBS EOF echo "performance" | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor echo echo "Using config:" echo cat /tmp/fio-job.ini echo echo "mount options: $MOUNT_OPTIONS" echo umount $MNT &> /dev/null mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT fio /tmp/fio-job.ini umount $MNT The tests were done on a physical machine, with 12 cores, 64G of RAM, using a NVMEe device and using a non-debug kernel config (the default one from Debian). The summary line from fio is provided below for each test run. With 8 jobs, file size 256M, fsync frequency of 4 and a block size of 4K: Before: WRITE: bw=28.3MiB/s (29.7MB/s), 28.3MiB/s-28.3MiB/s (29.7MB/s-29.7MB/s), io=2048MiB (2147MB), run=72297-72297msec After: WRITE: bw=28.7MiB/s (30.1MB/s), 28.7MiB/s-28.7MiB/s (30.1MB/s-30.1MB/s), io=2048MiB (2147MB), run=71411-71411msec +1.4% throughput, -1.2% runtime With 16 jobs, file size 256M, fsync frequency of 4 and a block size of 4K: Before: WRITE: bw=40.0MiB/s (42.0MB/s), 40.0MiB/s-40.0MiB/s (42.0MB/s-42.0MB/s), io=4096MiB (4295MB), run=99980-99980msec After: WRITE: bw=40.9MiB/s (42.9MB/s), 40.9MiB/s-40.9MiB/s (42.9MB/s-42.9MB/s), io=4096MiB (4295MB), run=97933-97933msec +2.2% throughput, -2.1% runtime The changes are small but it's possible to be better on faster hardware as in the test machine used disk utilization was pretty much 100% during the whole time the tests were running (observed with 'iostat -xz 1'). The tests also included the previous patch with the subject of: "btrfs: avoid unnecessary log mutex contention when syncing log". So they compared a branch without that patch and without this patch versus a branch with these two patches applied. Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-07-20 18:03:43 +03:00
struct btrfs_inode *inode, bool inode_item_dropped)
{
struct btrfs_inode_item *inode_item;
int ret;
btrfs: avoid unnecessary lock and leaf splits when updating inode in the log During a fast fsync, if we have already fsynced the file before and in the current transaction, we can make the inode item update more efficient and avoid acquiring a write lock on the leaf's parent. To update the inode item we are always using btrfs_insert_empty_item() to get a path pointing to the inode item, which calls btrfs_search_slot() with an "ins_len" argument of 'sizeof(struct btrfs_inode_item) + sizeof(struct btrfs_item)', and that always results in the search taking a write lock on the level 1 node that is the parent of the leaf that contains the inode item. This adds unnecessary lock contention on log trees when we have multiple fsyncs in parallel against inodes in the same subvolume, which has a very significant impact due to the fact that log trees are short lived and their height very rarely goes beyond level 2. Also, by using btrfs_insert_empty_item() when we need to update the inode item, we also end up splitting the leaf of the existing inode item when the leaf has an amount of free space smaller than the size of an inode item. Improve this by using btrfs_seach_slot(), with a 0 "ins_len" argument, when we know the inode item already exists in the log. This avoids these two inefficiencies. The following script, using fio, was used to perform the tests: $ cat fio-test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 MOUNT_OPTIONS="-o ssd" MKFS_OPTIONS="-d single -m single" if [ $# -ne 4 ]; then echo "Use $0 NUM_JOBS FILE_SIZE FSYNC_FREQ BLOCK_SIZE" exit 1 fi NUM_JOBS=$1 FILE_SIZE=$2 FSYNC_FREQ=$3 BLOCK_SIZE=$4 cat <<EOF > /tmp/fio-job.ini [writers] rw=randwrite fsync=$FSYNC_FREQ fallocate=none group_reporting=1 direct=0 bs=$BLOCK_SIZE ioengine=sync size=$FILE_SIZE directory=$MNT numjobs=$NUM_JOBS EOF echo "performance" | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor echo echo "Using config:" echo cat /tmp/fio-job.ini echo echo "mount options: $MOUNT_OPTIONS" echo umount $MNT &> /dev/null mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT fio /tmp/fio-job.ini umount $MNT The tests were done on a physical machine, with 12 cores, 64G of RAM, using a NVMEe device and using a non-debug kernel config (the default one from Debian). The summary line from fio is provided below for each test run. With 8 jobs, file size 256M, fsync frequency of 4 and a block size of 4K: Before: WRITE: bw=28.3MiB/s (29.7MB/s), 28.3MiB/s-28.3MiB/s (29.7MB/s-29.7MB/s), io=2048MiB (2147MB), run=72297-72297msec After: WRITE: bw=28.7MiB/s (30.1MB/s), 28.7MiB/s-28.7MiB/s (30.1MB/s-30.1MB/s), io=2048MiB (2147MB), run=71411-71411msec +1.4% throughput, -1.2% runtime With 16 jobs, file size 256M, fsync frequency of 4 and a block size of 4K: Before: WRITE: bw=40.0MiB/s (42.0MB/s), 40.0MiB/s-40.0MiB/s (42.0MB/s-42.0MB/s), io=4096MiB (4295MB), run=99980-99980msec After: WRITE: bw=40.9MiB/s (42.9MB/s), 40.9MiB/s-40.9MiB/s (42.9MB/s-42.9MB/s), io=4096MiB (4295MB), run=97933-97933msec +2.2% throughput, -2.1% runtime The changes are small but it's possible to be better on faster hardware as in the test machine used disk utilization was pretty much 100% during the whole time the tests were running (observed with 'iostat -xz 1'). The tests also included the previous patch with the subject of: "btrfs: avoid unnecessary log mutex contention when syncing log". So they compared a branch without that patch and without this patch versus a branch with these two patches applied. Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-07-20 18:03:43 +03:00
/*
* If we are doing a fast fsync and the inode was logged before in the
* current transaction, then we know the inode was previously logged and
* it exists in the log tree. For performance reasons, in this case use
* btrfs_search_slot() directly with ins_len set to 0 so that we never
* attempt a write lock on the leaf's parent, which adds unnecessary lock
* contention in case there are concurrent fsyncs for other inodes of the
* same subvolume. Using btrfs_insert_empty_item() when the inode item
* already exists can also result in unnecessarily splitting a leaf.
*/
if (!inode_item_dropped && inode->logged_trans == trans->transid) {
ret = btrfs_search_slot(trans, log, &inode->location, path, 0, 1);
ASSERT(ret <= 0);
if (ret > 0)
ret = -ENOENT;
} else {
/*
* This means it is the first fsync in the current transaction,
* so the inode item is not in the log and we need to insert it.
* We can never get -EEXIST because we are only called for a fast
* fsync and in case an inode eviction happens after the inode was
* logged before in the current transaction, when we load again
* the inode, we set BTRFS_INODE_NEEDS_FULL_SYNC on its runtime
* flags and set ->logged_trans to 0.
*/
ret = btrfs_insert_empty_item(trans, log, path, &inode->location,
sizeof(*inode_item));
ASSERT(ret != -EEXIST);
}
if (ret)
return ret;
inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
struct btrfs_inode_item);
fill_inode_item(trans, path->nodes[0], inode_item, &inode->vfs_inode,
0, 0);
btrfs_release_path(path);
return 0;
}
Btrfs: fix missing data checksums after replaying a log tree When logging a file that has shared extents (reflinked with other files or with itself), we can end up logging multiple checksum items that cover overlapping ranges. This confuses the search for checksums at log replay time causing some checksums to never be added to the fs/subvolume tree. Consider the following example of a file that shares the same extent at offsets 0 and 256Kb: [ bytenr 13893632, offset 64Kb, len 64Kb ] 0 64Kb [ bytenr 13631488, offset 64Kb, len 192Kb ] 64Kb 256Kb [ bytenr 13893632, offset 0, len 256Kb ] 256Kb 512Kb When logging the inode, at tree-log.c:copy_items(), when processing the file extent item at offset 0, we log a checksum item covering the range 13959168 to 14024704, which corresponds to 13893632 + 64Kb and 13893632 + 64Kb + 64Kb, respectively. Later when processing the extent item at offset 256K, we log the checksums for the range from 13893632 to 14155776 (which corresponds to 13893632 + 256Kb). These checksums get merged with the checksum item for the range from 13631488 to 13893632 (13631488 + 256Kb), logged by a previous fsync. So after this we get the two following checksum items in the log tree: (...) item 6 key (EXTENT_CSUM EXTENT_CSUM 13631488) itemoff 3095 itemsize 512 range start 13631488 end 14155776 length 524288 item 7 key (EXTENT_CSUM EXTENT_CSUM 13959168) itemoff 3031 itemsize 64 range start 13959168 end 14024704 length 65536 The first one covers the range from the second one, they overlap. So far this does not cause a problem after replaying the log, because when replaying the file extent item for offset 256K, we copy all the checksums for the extent 13893632 from the log tree to the fs/subvolume tree, since searching for an checksum item for bytenr 13893632 leaves us at the first checksum item, which covers the whole range of the extent. However if we write 64Kb to file offset 256Kb for example, we will not be able to find and copy the checksums for the last 128Kb of the extent at bytenr 13893632, referenced by the file range 384Kb to 512Kb. After writing 64Kb into file offset 256Kb we get the following extent layout for our file: [ bytenr 13893632, offset 64K, len 64Kb ] 0 64Kb [ bytenr 13631488, offset 64Kb, len 192Kb ] 64Kb 256Kb [ bytenr 14155776, offset 0, len 64Kb ] 256Kb 320Kb [ bytenr 13893632, offset 64Kb, len 192Kb ] 320Kb 512Kb After fsync'ing the file, if we have a power failure and then mount the filesystem to replay the log, the following happens: 1) When replaying the file extent item for file offset 320Kb, we lookup for the checksums for the extent range from 13959168 (13893632 + 64Kb) to 14155776 (13893632 + 256Kb), through a call to btrfs_lookup_csums_range(); 2) btrfs_lookup_csums_range() finds the checksum item that starts precisely at offset 13959168 (item 7 in the log tree, shown before); 3) However that checksum item only covers 64Kb of data, and not 192Kb of data; 4) As a result only the checksums for the first 64Kb of data referenced by the file extent item are found and copied to the fs/subvolume tree. The remaining 128Kb of data, file range 384Kb to 512Kb, doesn't get the corresponding data checksums found and copied to the fs/subvolume tree. 5) After replaying the log userspace will not be able to read the file range from 384Kb to 512Kb, because the checksums are missing and resulting in an -EIO error. The following steps reproduce this scenario: $ mkfs.btrfs -f /dev/sdc $ mount /dev/sdc /mnt/sdc $ xfs_io -f -c "pwrite -S 0xa3 0 256K" /mnt/sdc/foobar $ xfs_io -c "fsync" /mnt/sdc/foobar $ xfs_io -c "pwrite -S 0xc7 256K 256K" /mnt/sdc/foobar $ xfs_io -c "reflink /mnt/sdc/foobar 320K 0 64K" /mnt/sdc/foobar $ xfs_io -c "fsync" /mnt/sdc/foobar $ xfs_io -c "pwrite -S 0xe5 256K 64K" /mnt/sdc/foobar $ xfs_io -c "fsync" /mnt/sdc/foobar <power failure> $ mount /dev/sdc /mnt/sdc $ md5sum /mnt/sdc/foobar md5sum: /mnt/sdc/foobar: Input/output error $ dmesg | tail [165305.003464] BTRFS info (device sdc): no csum found for inode 257 start 401408 [165305.004014] BTRFS info (device sdc): no csum found for inode 257 start 405504 [165305.004559] BTRFS info (device sdc): no csum found for inode 257 start 409600 [165305.005101] BTRFS info (device sdc): no csum found for inode 257 start 413696 [165305.005627] BTRFS info (device sdc): no csum found for inode 257 start 417792 [165305.006134] BTRFS info (device sdc): no csum found for inode 257 start 421888 [165305.006625] BTRFS info (device sdc): no csum found for inode 257 start 425984 [165305.007278] BTRFS info (device sdc): no csum found for inode 257 start 430080 [165305.008248] BTRFS warning (device sdc): csum failed root 5 ino 257 off 393216 csum 0x1337385e expected csum 0x00000000 mirror 1 [165305.009550] BTRFS warning (device sdc): csum failed root 5 ino 257 off 393216 csum 0x1337385e expected csum 0x00000000 mirror 1 Fix this simply by deleting first any checksums, from the log tree, for the range of the extent we are logging at copy_items(). This ensures we do not get checksum items in the log tree that have overlapping ranges. This is a long time issue that has been present since we have the clone (and deduplication) ioctl, and can happen both when an extent is shared between different files and within the same file. A test case for fstests follows soon. CC: stable@vger.kernel.org # 4.4+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-12-05 19:58:30 +03:00
static int log_csums(struct btrfs_trans_handle *trans,
btrfs: reduce contention on log trees when logging checksums The possibility of extents being shared (through clone and deduplication operations) requires special care when logging data checksums, to avoid having a log tree with different checksum items that cover ranges which overlap (which resulted in missing checksums after replaying a log tree). Such problems were fixed in the past by the following commits: commit 40e046acbd2f ("Btrfs: fix missing data checksums after replaying a log tree") commit e289f03ea79b ("btrfs: fix corrupt log due to concurrent fsync of inodes with shared extents") Test case generic/588 exercises the scenario solved by the first commit (purely sequential and deterministic) while test case generic/457 often triggered the case fixed by the second commit (not deterministic, requires specific timings under concurrency). The problems were addressed by deleting, from the log tree, any existing checksums before logging the new ones. And also by doing the deletion and logging of the cheksums while locking the checksum range in an extent io tree (root->log_csum_range), to deal with the case where we have concurrent fsyncs against files with shared extents. That however causes more contention on the leaves of a log tree where we store checksums (and all the nodes in the paths leading to them), even when we do not have shared extents, or all the shared extents were created by past transactions. It also adds a bit of contention on the spin lock of the log_csums_range extent io tree of the log root. This change adds a 'last_reflink_trans' field to the inode to keep track of the last transaction where a new extent was shared between inodes (through clone and deduplication operations). It is updated for both the source and destination inodes of reflink operations whenever a new extent (created in the current transaction) becomes shared by the inodes. This field is kept in memory only, not persisted in the inode item, similar to other existing fields (last_unlink_trans, logged_trans). When logging checksums for an extent, if the value of 'last_reflink_trans' is smaller then the current transaction's generation/id, we skip locking the extent range and deletion of checksums from the log tree, since we know we do not have new shared extents. This reduces contention on the log tree's leaves where checksums are stored. The following script, which uses fio, was used to measure the impact of this change: $ cat test-fsync.sh #!/bin/bash DEV=/dev/sdk MNT=/mnt/sdk MOUNT_OPTIONS="-o ssd" MKFS_OPTIONS="-d single -m single" if [ $# -ne 3 ]; then echo "Use $0 NUM_JOBS FILE_SIZE FSYNC_FREQ" exit 1 fi NUM_JOBS=$1 FILE_SIZE=$2 FSYNC_FREQ=$3 cat <<EOF > /tmp/fio-job.ini [writers] rw=write fsync=$FSYNC_FREQ fallocate=none group_reporting=1 direct=0 bs=64k ioengine=sync size=$FILE_SIZE directory=$MNT numjobs=$NUM_JOBS EOF echo "Using config:" echo cat /tmp/fio-job.ini echo mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT fio /tmp/fio-job.ini umount $MNT The tests were performed for different numbers of jobs, file sizes and fsync frequency. A qemu VM using kvm was used, with 8 cores (the host has 12 cores, with cpu governance set to performance mode on all cores), 16GiB of ram (the host has 64GiB) and using a NVMe device directly (without an intermediary filesystem in the host). While running the tests, the host was not used for anything else, to avoid disturbing the tests. The obtained results were the following (the last line of fio's output was pasted). Starting with 16 jobs is where a significant difference is observable in this particular setup and hardware (differences highlighted below). The very small differences for tests with less than 16 jobs are possibly just noise and random. **** 1 job, file size 1G, fsync frequency 1 **** before this change: WRITE: bw=23.8MiB/s (24.9MB/s), 23.8MiB/s-23.8MiB/s (24.9MB/s-24.9MB/s), io=1024MiB (1074MB), run=43075-43075msec after this change: WRITE: bw=24.4MiB/s (25.6MB/s), 24.4MiB/s-24.4MiB/s (25.6MB/s-25.6MB/s), io=1024MiB (1074MB), run=41938-41938msec **** 2 jobs, file size 1G, fsync frequency 1 **** before this change: WRITE: bw=37.7MiB/s (39.5MB/s), 37.7MiB/s-37.7MiB/s (39.5MB/s-39.5MB/s), io=2048MiB (2147MB), run=54351-54351msec after this change: WRITE: bw=37.7MiB/s (39.5MB/s), 37.6MiB/s-37.6MiB/s (39.5MB/s-39.5MB/s), io=2048MiB (2147MB), run=54428-54428msec **** 4 jobs, file size 1G, fsync frequency 1 **** before this change: WRITE: bw=67.5MiB/s (70.8MB/s), 67.5MiB/s-67.5MiB/s (70.8MB/s-70.8MB/s), io=4096MiB (4295MB), run=60669-60669msec after this change: WRITE: bw=68.6MiB/s (71.0MB/s), 68.6MiB/s-68.6MiB/s (71.0MB/s-71.0MB/s), io=4096MiB (4295MB), run=59678-59678msec **** 8 jobs, file size 1G, fsync frequency 1 **** before this change: WRITE: bw=128MiB/s (134MB/s), 128MiB/s-128MiB/s (134MB/s-134MB/s), io=8192MiB (8590MB), run=64048-64048msec after this change: WRITE: bw=129MiB/s (135MB/s), 129MiB/s-129MiB/s (135MB/s-135MB/s), io=8192MiB (8590MB), run=63405-63405msec **** 16 jobs, file size 1G, fsync frequency 1 **** before this change: WRITE: bw=78.5MiB/s (82.3MB/s), 78.5MiB/s-78.5MiB/s (82.3MB/s-82.3MB/s), io=16.0GiB (17.2GB), run=208676-208676msec after this change: WRITE: bw=110MiB/s (115MB/s), 110MiB/s-110MiB/s (115MB/s-115MB/s), io=16.0GiB (17.2GB), run=149295-149295msec (+40.1% throughput, -28.5% runtime) **** 32 jobs, file size 1G, fsync frequency 1 **** before this change: WRITE: bw=58.8MiB/s (61.7MB/s), 58.8MiB/s-58.8MiB/s (61.7MB/s-61.7MB/s), io=32.0GiB (34.4GB), run=557134-557134msec after this change: WRITE: bw=76.1MiB/s (79.8MB/s), 76.1MiB/s-76.1MiB/s (79.8MB/s-79.8MB/s), io=32.0GiB (34.4GB), run=430550-430550msec (+29.4% throughput, -22.7% runtime) **** 64 jobs, file size 512M, fsync frequency 1 **** before this change: WRITE: bw=65.8MiB/s (68.0MB/s), 65.8MiB/s-65.8MiB/s (68.0MB/s-68.0MB/s), io=32.0GiB (34.4GB), run=498055-498055msec after this change: WRITE: bw=85.1MiB/s (89.2MB/s), 85.1MiB/s-85.1MiB/s (89.2MB/s-89.2MB/s), io=32.0GiB (34.4GB), run=385116-385116msec (+29.3% throughput, -22.7% runtime) **** 128 jobs, file size 256M, fsync frequency 1 **** before this change: WRITE: bw=54.7MiB/s (57.3MB/s), 54.7MiB/s-54.7MiB/s (57.3MB/s-57.3MB/s), io=32.0GiB (34.4GB), run=599373-599373msec after this change: WRITE: bw=121MiB/s (126MB/s), 121MiB/s-121MiB/s (126MB/s-126MB/s), io=32.0GiB (34.4GB), run=271907-271907msec (+121.2% throughput, -54.6% runtime) **** 256 jobs, file size 256M, fsync frequency 1 **** before this change: WRITE: bw=69.2MiB/s (72.5MB/s), 69.2MiB/s-69.2MiB/s (72.5MB/s-72.5MB/s), io=64.0GiB (68.7GB), run=947536-947536msec after this change: WRITE: bw=121MiB/s (127MB/s), 121MiB/s-121MiB/s (127MB/s-127MB/s), io=64.0GiB (68.7GB), run=541916-541916msec (+74.9% throughput, -42.8% runtime) **** 512 jobs, file size 128M, fsync frequency 1 **** before this change: WRITE: bw=85.4MiB/s (89.5MB/s), 85.4MiB/s-85.4MiB/s (89.5MB/s-89.5MB/s), io=64.0GiB (68.7GB), run=767734-767734msec after this change: WRITE: bw=141MiB/s (147MB/s), 141MiB/s-141MiB/s (147MB/s-147MB/s), io=64.0GiB (68.7GB), run=466022-466022msec (+65.1% throughput, -39.3% runtime) **** 1024 jobs, file size 128M, fsync frequency 1 **** before this change: WRITE: bw=115MiB/s (120MB/s), 115MiB/s-115MiB/s (120MB/s-120MB/s), io=128GiB (137GB), run=1143775-1143775msec after this change: WRITE: bw=171MiB/s (180MB/s), 171MiB/s-171MiB/s (180MB/s-180MB/s), io=128GiB (137GB), run=764843-764843msec (+48.7% throughput, -33.1% runtime) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-07-15 14:30:43 +03:00
struct btrfs_inode *inode,
Btrfs: fix missing data checksums after replaying a log tree When logging a file that has shared extents (reflinked with other files or with itself), we can end up logging multiple checksum items that cover overlapping ranges. This confuses the search for checksums at log replay time causing some checksums to never be added to the fs/subvolume tree. Consider the following example of a file that shares the same extent at offsets 0 and 256Kb: [ bytenr 13893632, offset 64Kb, len 64Kb ] 0 64Kb [ bytenr 13631488, offset 64Kb, len 192Kb ] 64Kb 256Kb [ bytenr 13893632, offset 0, len 256Kb ] 256Kb 512Kb When logging the inode, at tree-log.c:copy_items(), when processing the file extent item at offset 0, we log a checksum item covering the range 13959168 to 14024704, which corresponds to 13893632 + 64Kb and 13893632 + 64Kb + 64Kb, respectively. Later when processing the extent item at offset 256K, we log the checksums for the range from 13893632 to 14155776 (which corresponds to 13893632 + 256Kb). These checksums get merged with the checksum item for the range from 13631488 to 13893632 (13631488 + 256Kb), logged by a previous fsync. So after this we get the two following checksum items in the log tree: (...) item 6 key (EXTENT_CSUM EXTENT_CSUM 13631488) itemoff 3095 itemsize 512 range start 13631488 end 14155776 length 524288 item 7 key (EXTENT_CSUM EXTENT_CSUM 13959168) itemoff 3031 itemsize 64 range start 13959168 end 14024704 length 65536 The first one covers the range from the second one, they overlap. So far this does not cause a problem after replaying the log, because when replaying the file extent item for offset 256K, we copy all the checksums for the extent 13893632 from the log tree to the fs/subvolume tree, since searching for an checksum item for bytenr 13893632 leaves us at the first checksum item, which covers the whole range of the extent. However if we write 64Kb to file offset 256Kb for example, we will not be able to find and copy the checksums for the last 128Kb of the extent at bytenr 13893632, referenced by the file range 384Kb to 512Kb. After writing 64Kb into file offset 256Kb we get the following extent layout for our file: [ bytenr 13893632, offset 64K, len 64Kb ] 0 64Kb [ bytenr 13631488, offset 64Kb, len 192Kb ] 64Kb 256Kb [ bytenr 14155776, offset 0, len 64Kb ] 256Kb 320Kb [ bytenr 13893632, offset 64Kb, len 192Kb ] 320Kb 512Kb After fsync'ing the file, if we have a power failure and then mount the filesystem to replay the log, the following happens: 1) When replaying the file extent item for file offset 320Kb, we lookup for the checksums for the extent range from 13959168 (13893632 + 64Kb) to 14155776 (13893632 + 256Kb), through a call to btrfs_lookup_csums_range(); 2) btrfs_lookup_csums_range() finds the checksum item that starts precisely at offset 13959168 (item 7 in the log tree, shown before); 3) However that checksum item only covers 64Kb of data, and not 192Kb of data; 4) As a result only the checksums for the first 64Kb of data referenced by the file extent item are found and copied to the fs/subvolume tree. The remaining 128Kb of data, file range 384Kb to 512Kb, doesn't get the corresponding data checksums found and copied to the fs/subvolume tree. 5) After replaying the log userspace will not be able to read the file range from 384Kb to 512Kb, because the checksums are missing and resulting in an -EIO error. The following steps reproduce this scenario: $ mkfs.btrfs -f /dev/sdc $ mount /dev/sdc /mnt/sdc $ xfs_io -f -c "pwrite -S 0xa3 0 256K" /mnt/sdc/foobar $ xfs_io -c "fsync" /mnt/sdc/foobar $ xfs_io -c "pwrite -S 0xc7 256K 256K" /mnt/sdc/foobar $ xfs_io -c "reflink /mnt/sdc/foobar 320K 0 64K" /mnt/sdc/foobar $ xfs_io -c "fsync" /mnt/sdc/foobar $ xfs_io -c "pwrite -S 0xe5 256K 64K" /mnt/sdc/foobar $ xfs_io -c "fsync" /mnt/sdc/foobar <power failure> $ mount /dev/sdc /mnt/sdc $ md5sum /mnt/sdc/foobar md5sum: /mnt/sdc/foobar: Input/output error $ dmesg | tail [165305.003464] BTRFS info (device sdc): no csum found for inode 257 start 401408 [165305.004014] BTRFS info (device sdc): no csum found for inode 257 start 405504 [165305.004559] BTRFS info (device sdc): no csum found for inode 257 start 409600 [165305.005101] BTRFS info (device sdc): no csum found for inode 257 start 413696 [165305.005627] BTRFS info (device sdc): no csum found for inode 257 start 417792 [165305.006134] BTRFS info (device sdc): no csum found for inode 257 start 421888 [165305.006625] BTRFS info (device sdc): no csum found for inode 257 start 425984 [165305.007278] BTRFS info (device sdc): no csum found for inode 257 start 430080 [165305.008248] BTRFS warning (device sdc): csum failed root 5 ino 257 off 393216 csum 0x1337385e expected csum 0x00000000 mirror 1 [165305.009550] BTRFS warning (device sdc): csum failed root 5 ino 257 off 393216 csum 0x1337385e expected csum 0x00000000 mirror 1 Fix this simply by deleting first any checksums, from the log tree, for the range of the extent we are logging at copy_items(). This ensures we do not get checksum items in the log tree that have overlapping ranges. This is a long time issue that has been present since we have the clone (and deduplication) ioctl, and can happen both when an extent is shared between different files and within the same file. A test case for fstests follows soon. CC: stable@vger.kernel.org # 4.4+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-12-05 19:58:30 +03:00
struct btrfs_root *log_root,
struct btrfs_ordered_sum *sums)
{
btrfs: fix corrupt log due to concurrent fsync of inodes with shared extents When we have extents shared amongst different inodes in the same subvolume, if we fsync them in parallel we can end up with checksum items in the log tree that represent ranges which overlap. For example, consider we have inodes A and B, both sharing an extent that covers the logical range from X to X + 64KiB: 1) Task A starts an fsync on inode A; 2) Task B starts an fsync on inode B; 3) Task A calls btrfs_csum_file_blocks(), and the first search in the log tree, through btrfs_lookup_csum(), returns -EFBIG because it finds an existing checksum item that covers the range from X - 64KiB to X; 4) Task A checks that the checksum item has not reached the maximum possible size (MAX_CSUM_ITEMS) and then releases the search path before it does another path search for insertion (through a direct call to btrfs_search_slot()); 5) As soon as task A releases the path and before it does the search for insertion, task B calls btrfs_csum_file_blocks() and gets -EFBIG too, because there is an existing checksum item that has an end offset that matches the start offset (X) of the checksum range we want to log; 6) Task B releases the path; 7) Task A does the path search for insertion (through btrfs_search_slot()) and then verifies that the checksum item that ends at offset X still exists and extends its size to insert the checksums for the range from X to X + 64KiB; 8) Task A releases the path and returns from btrfs_csum_file_blocks(), having inserted the checksums into an existing checksum item that got its size extended. At this point we have one checksum item in the log tree that covers the logical range from X - 64KiB to X + 64KiB; 9) Task B now does a search for insertion using btrfs_search_slot() too, but it finds that the previous checksum item no longer ends at the offset X, it now ends at an of offset X + 64KiB, so it leaves that item untouched. Then it releases the path and calls btrfs_insert_empty_item() that inserts a checksum item with a key offset corresponding to X and a size for inserting a single checksum (4 bytes in case of crc32c). Subsequent iterations end up extending this new checksum item so that it contains the checksums for the range from X to X + 64KiB. So after task B returns from btrfs_csum_file_blocks() we end up with two checksum items in the log tree that have overlapping ranges, one for the range from X - 64KiB to X + 64KiB, and another for the range from X to X + 64KiB. Having checksum items that represent ranges which overlap, regardless of being in the log tree or in the chekcsums tree, can lead to problems where checksums for a file range end up not being found. This type of problem has happened a few times in the past and the following commits fixed them and explain in detail why having checksum items with overlapping ranges is problematic: 27b9a8122ff71a "Btrfs: fix csum tree corruption, duplicate and outdated checksums" b84b8390d6009c "Btrfs: fix file read corruption after extent cloning and fsync" 40e046acbd2f36 "Btrfs: fix missing data checksums after replaying a log tree" Since this specific instance of the problem can only happen when logging inodes, because it is the only case where concurrent attempts to insert checksums for the same range can happen, fix the issue by using an extent io tree as a range lock to serialize checksum insertion during inode logging. This issue could often be reproduced by the test case generic/457 from fstests. When it happens it produces the following trace: BTRFS critical (device dm-0): corrupt leaf: root=18446744073709551610 block=30625792 slot=42, csum end range (15020032) goes beyond the start range (15015936) of the next csum item BTRFS info (device dm-0): leaf 30625792 gen 7 total ptrs 49 free space 2402 owner 18446744073709551610 BTRFS info (device dm-0): refs 1 lock (w:0 r:0 bw:0 br:0 sw:0 sr:0) lock_owner 0 current 15884 item 0 key (18446744073709551606 128 13979648) itemoff 3991 itemsize 4 item 1 key (18446744073709551606 128 13983744) itemoff 3987 itemsize 4 item 2 key (18446744073709551606 128 13987840) itemoff 3983 itemsize 4 item 3 key (18446744073709551606 128 13991936) itemoff 3979 itemsize 4 item 4 key (18446744073709551606 128 13996032) itemoff 3975 itemsize 4 item 5 key (18446744073709551606 128 14000128) itemoff 3971 itemsize 4 (...) BTRFS error (device dm-0): block=30625792 write time tree block corruption detected ------------[ cut here ]------------ WARNING: CPU: 1 PID: 15884 at fs/btrfs/disk-io.c:539 btree_csum_one_bio+0x268/0x2d0 [btrfs] Modules linked in: btrfs dm_thin_pool ... CPU: 1 PID: 15884 Comm: fsx Tainted: G W 5.6.0-rc7-btrfs-next-58 #1 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.12.0-59-gc9ba5276e321-prebuilt.qemu.org 04/01/2014 RIP: 0010:btree_csum_one_bio+0x268/0x2d0 [btrfs] Code: c7 c7 ... RSP: 0018:ffffbb0109e6f8e0 EFLAGS: 00010296 RAX: 0000000000000000 RBX: ffffe1c0847b6080 RCX: 0000000000000000 RDX: 0000000000000000 RSI: ffffffffaa963988 RDI: 0000000000000001 RBP: ffff956a4f4d2000 R08: 0000000000000000 R09: 0000000000000001 R10: 0000000000000526 R11: 0000000000000000 R12: ffff956a5cd28bb0 R13: 0000000000000000 R14: ffff956a649c9388 R15: 000000011ed82000 FS: 00007fb419959e80(0000) GS:ffff956a7aa00000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 0000000000fe6d54 CR3: 0000000138696005 CR4: 00000000003606e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: btree_submit_bio_hook+0x67/0xc0 [btrfs] submit_one_bio+0x31/0x50 [btrfs] btree_write_cache_pages+0x2db/0x4b0 [btrfs] ? __filemap_fdatawrite_range+0xb1/0x110 do_writepages+0x23/0x80 __filemap_fdatawrite_range+0xd2/0x110 btrfs_write_marked_extents+0x15e/0x180 [btrfs] btrfs_sync_log+0x206/0x10a0 [btrfs] ? kmem_cache_free+0x315/0x3b0 ? btrfs_log_inode+0x1e8/0xf90 [btrfs] ? __mutex_unlock_slowpath+0x45/0x2a0 ? lockref_put_or_lock+0x9/0x30 ? dput+0x2d/0x580 ? dput+0xb5/0x580 ? btrfs_sync_file+0x464/0x4d0 [btrfs] btrfs_sync_file+0x464/0x4d0 [btrfs] do_fsync+0x38/0x60 __x64_sys_fsync+0x10/0x20 do_syscall_64+0x5c/0x280 entry_SYSCALL_64_after_hwframe+0x49/0xbe RIP: 0033:0x7fb41953a6d0 Code: 48 3d ... RSP: 002b:00007ffcc86bd218 EFLAGS: 00000246 ORIG_RAX: 000000000000004a RAX: ffffffffffffffda RBX: 000000000000000d RCX: 00007fb41953a6d0 RDX: 0000000000000009 RSI: 0000000000040000 RDI: 0000000000000003 RBP: 0000000000040000 R08: 0000000000000001 R09: 0000000000000009 R10: 0000000000000064 R11: 0000000000000246 R12: 0000556cf4b2c060 R13: 0000000000000100 R14: 0000000000000000 R15: 0000556cf322b420 irq event stamp: 0 hardirqs last enabled at (0): [<0000000000000000>] 0x0 hardirqs last disabled at (0): [<ffffffffa96bdedf>] copy_process+0x74f/0x2020 softirqs last enabled at (0): [<ffffffffa96bdedf>] copy_process+0x74f/0x2020 softirqs last disabled at (0): [<0000000000000000>] 0x0 ---[ end trace d543fc76f5ad7fd8 ]--- In that trace the tree checker detected the overlapping checksum items at the time when we triggered writeback for the log tree when syncing the log. Another trace that can happen is due to BUG_ON() when deleting checksum items while logging an inode: BTRFS critical (device dm-0): slot 81 key (18446744073709551606 128 13635584) new key (18446744073709551606 128 13635584) BTRFS info (device dm-0): leaf 30949376 gen 7 total ptrs 98 free space 8527 owner 18446744073709551610 BTRFS info (device dm-0): refs 4 lock (w:1 r:0 bw:0 br:0 sw:1 sr:0) lock_owner 13473 current 13473 item 0 key (257 1 0) itemoff 16123 itemsize 160 inode generation 7 size 262144 mode 100600 item 1 key (257 12 256) itemoff 16103 itemsize 20 item 2 key (257 108 0) itemoff 16050 itemsize 53 extent data disk bytenr 13631488 nr 4096 extent data offset 0 nr 131072 ram 131072 (...) ------------[ cut here ]------------ kernel BUG at fs/btrfs/ctree.c:3153! invalid opcode: 0000 [#1] PREEMPT SMP DEBUG_PAGEALLOC PTI CPU: 1 PID: 13473 Comm: fsx Not tainted 5.6.0-rc7-btrfs-next-58 #1 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.12.0-59-gc9ba5276e321-prebuilt.qemu.org 04/01/2014 RIP: 0010:btrfs_set_item_key_safe+0x1ea/0x270 [btrfs] Code: 0f b6 ... RSP: 0018:ffff95e3889179d0 EFLAGS: 00010282 RAX: 0000000000000000 RBX: 0000000000000051 RCX: 0000000000000000 RDX: 0000000000000000 RSI: ffffffffb7763988 RDI: 0000000000000001 RBP: fffffffffffffff6 R08: 0000000000000000 R09: 0000000000000001 R10: 00000000000009ef R11: 0000000000000000 R12: ffff8912a8ba5a08 R13: ffff95e388917a06 R14: ffff89138dcf68c8 R15: ffff95e388917ace FS: 00007fe587084e80(0000) GS:ffff8913baa00000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 00007fe587091000 CR3: 0000000126dac005 CR4: 00000000003606e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: btrfs_del_csums+0x2f4/0x540 [btrfs] copy_items+0x4b5/0x560 [btrfs] btrfs_log_inode+0x910/0xf90 [btrfs] btrfs_log_inode_parent+0x2a0/0xe40 [btrfs] ? dget_parent+0x5/0x370 btrfs_log_dentry_safe+0x4a/0x70 [btrfs] btrfs_sync_file+0x42b/0x4d0 [btrfs] __x64_sys_msync+0x199/0x200 do_syscall_64+0x5c/0x280 entry_SYSCALL_64_after_hwframe+0x49/0xbe RIP: 0033:0x7fe586c65760 Code: 00 f7 ... RSP: 002b:00007ffe250f98b8 EFLAGS: 00000246 ORIG_RAX: 000000000000001a RAX: ffffffffffffffda RBX: 00000000000040e1 RCX: 00007fe586c65760 RDX: 0000000000000004 RSI: 0000000000006b51 RDI: 00007fe58708b000 RBP: 0000000000006a70 R08: 0000000000000003 R09: 00007fe58700cb61 R10: 0000000000000100 R11: 0000000000000246 R12: 00000000000000e1 R13: 00007fe58708b000 R14: 0000000000006b51 R15: 0000558de021a420 Modules linked in: dm_log_writes ... ---[ end trace c92a7f447a8515f5 ]--- CC: stable@vger.kernel.org # 4.4+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-05-18 14:14:50 +03:00
const u64 lock_end = sums->bytenr + sums->len - 1;
struct extent_state *cached_state = NULL;
Btrfs: fix missing data checksums after replaying a log tree When logging a file that has shared extents (reflinked with other files or with itself), we can end up logging multiple checksum items that cover overlapping ranges. This confuses the search for checksums at log replay time causing some checksums to never be added to the fs/subvolume tree. Consider the following example of a file that shares the same extent at offsets 0 and 256Kb: [ bytenr 13893632, offset 64Kb, len 64Kb ] 0 64Kb [ bytenr 13631488, offset 64Kb, len 192Kb ] 64Kb 256Kb [ bytenr 13893632, offset 0, len 256Kb ] 256Kb 512Kb When logging the inode, at tree-log.c:copy_items(), when processing the file extent item at offset 0, we log a checksum item covering the range 13959168 to 14024704, which corresponds to 13893632 + 64Kb and 13893632 + 64Kb + 64Kb, respectively. Later when processing the extent item at offset 256K, we log the checksums for the range from 13893632 to 14155776 (which corresponds to 13893632 + 256Kb). These checksums get merged with the checksum item for the range from 13631488 to 13893632 (13631488 + 256Kb), logged by a previous fsync. So after this we get the two following checksum items in the log tree: (...) item 6 key (EXTENT_CSUM EXTENT_CSUM 13631488) itemoff 3095 itemsize 512 range start 13631488 end 14155776 length 524288 item 7 key (EXTENT_CSUM EXTENT_CSUM 13959168) itemoff 3031 itemsize 64 range start 13959168 end 14024704 length 65536 The first one covers the range from the second one, they overlap. So far this does not cause a problem after replaying the log, because when replaying the file extent item for offset 256K, we copy all the checksums for the extent 13893632 from the log tree to the fs/subvolume tree, since searching for an checksum item for bytenr 13893632 leaves us at the first checksum item, which covers the whole range of the extent. However if we write 64Kb to file offset 256Kb for example, we will not be able to find and copy the checksums for the last 128Kb of the extent at bytenr 13893632, referenced by the file range 384Kb to 512Kb. After writing 64Kb into file offset 256Kb we get the following extent layout for our file: [ bytenr 13893632, offset 64K, len 64Kb ] 0 64Kb [ bytenr 13631488, offset 64Kb, len 192Kb ] 64Kb 256Kb [ bytenr 14155776, offset 0, len 64Kb ] 256Kb 320Kb [ bytenr 13893632, offset 64Kb, len 192Kb ] 320Kb 512Kb After fsync'ing the file, if we have a power failure and then mount the filesystem to replay the log, the following happens: 1) When replaying the file extent item for file offset 320Kb, we lookup for the checksums for the extent range from 13959168 (13893632 + 64Kb) to 14155776 (13893632 + 256Kb), through a call to btrfs_lookup_csums_range(); 2) btrfs_lookup_csums_range() finds the checksum item that starts precisely at offset 13959168 (item 7 in the log tree, shown before); 3) However that checksum item only covers 64Kb of data, and not 192Kb of data; 4) As a result only the checksums for the first 64Kb of data referenced by the file extent item are found and copied to the fs/subvolume tree. The remaining 128Kb of data, file range 384Kb to 512Kb, doesn't get the corresponding data checksums found and copied to the fs/subvolume tree. 5) After replaying the log userspace will not be able to read the file range from 384Kb to 512Kb, because the checksums are missing and resulting in an -EIO error. The following steps reproduce this scenario: $ mkfs.btrfs -f /dev/sdc $ mount /dev/sdc /mnt/sdc $ xfs_io -f -c "pwrite -S 0xa3 0 256K" /mnt/sdc/foobar $ xfs_io -c "fsync" /mnt/sdc/foobar $ xfs_io -c "pwrite -S 0xc7 256K 256K" /mnt/sdc/foobar $ xfs_io -c "reflink /mnt/sdc/foobar 320K 0 64K" /mnt/sdc/foobar $ xfs_io -c "fsync" /mnt/sdc/foobar $ xfs_io -c "pwrite -S 0xe5 256K 64K" /mnt/sdc/foobar $ xfs_io -c "fsync" /mnt/sdc/foobar <power failure> $ mount /dev/sdc /mnt/sdc $ md5sum /mnt/sdc/foobar md5sum: /mnt/sdc/foobar: Input/output error $ dmesg | tail [165305.003464] BTRFS info (device sdc): no csum found for inode 257 start 401408 [165305.004014] BTRFS info (device sdc): no csum found for inode 257 start 405504 [165305.004559] BTRFS info (device sdc): no csum found for inode 257 start 409600 [165305.005101] BTRFS info (device sdc): no csum found for inode 257 start 413696 [165305.005627] BTRFS info (device sdc): no csum found for inode 257 start 417792 [165305.006134] BTRFS info (device sdc): no csum found for inode 257 start 421888 [165305.006625] BTRFS info (device sdc): no csum found for inode 257 start 425984 [165305.007278] BTRFS info (device sdc): no csum found for inode 257 start 430080 [165305.008248] BTRFS warning (device sdc): csum failed root 5 ino 257 off 393216 csum 0x1337385e expected csum 0x00000000 mirror 1 [165305.009550] BTRFS warning (device sdc): csum failed root 5 ino 257 off 393216 csum 0x1337385e expected csum 0x00000000 mirror 1 Fix this simply by deleting first any checksums, from the log tree, for the range of the extent we are logging at copy_items(). This ensures we do not get checksum items in the log tree that have overlapping ranges. This is a long time issue that has been present since we have the clone (and deduplication) ioctl, and can happen both when an extent is shared between different files and within the same file. A test case for fstests follows soon. CC: stable@vger.kernel.org # 4.4+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-12-05 19:58:30 +03:00
int ret;
btrfs: reduce contention on log trees when logging checksums The possibility of extents being shared (through clone and deduplication operations) requires special care when logging data checksums, to avoid having a log tree with different checksum items that cover ranges which overlap (which resulted in missing checksums after replaying a log tree). Such problems were fixed in the past by the following commits: commit 40e046acbd2f ("Btrfs: fix missing data checksums after replaying a log tree") commit e289f03ea79b ("btrfs: fix corrupt log due to concurrent fsync of inodes with shared extents") Test case generic/588 exercises the scenario solved by the first commit (purely sequential and deterministic) while test case generic/457 often triggered the case fixed by the second commit (not deterministic, requires specific timings under concurrency). The problems were addressed by deleting, from the log tree, any existing checksums before logging the new ones. And also by doing the deletion and logging of the cheksums while locking the checksum range in an extent io tree (root->log_csum_range), to deal with the case where we have concurrent fsyncs against files with shared extents. That however causes more contention on the leaves of a log tree where we store checksums (and all the nodes in the paths leading to them), even when we do not have shared extents, or all the shared extents were created by past transactions. It also adds a bit of contention on the spin lock of the log_csums_range extent io tree of the log root. This change adds a 'last_reflink_trans' field to the inode to keep track of the last transaction where a new extent was shared between inodes (through clone and deduplication operations). It is updated for both the source and destination inodes of reflink operations whenever a new extent (created in the current transaction) becomes shared by the inodes. This field is kept in memory only, not persisted in the inode item, similar to other existing fields (last_unlink_trans, logged_trans). When logging checksums for an extent, if the value of 'last_reflink_trans' is smaller then the current transaction's generation/id, we skip locking the extent range and deletion of checksums from the log tree, since we know we do not have new shared extents. This reduces contention on the log tree's leaves where checksums are stored. The following script, which uses fio, was used to measure the impact of this change: $ cat test-fsync.sh #!/bin/bash DEV=/dev/sdk MNT=/mnt/sdk MOUNT_OPTIONS="-o ssd" MKFS_OPTIONS="-d single -m single" if [ $# -ne 3 ]; then echo "Use $0 NUM_JOBS FILE_SIZE FSYNC_FREQ" exit 1 fi NUM_JOBS=$1 FILE_SIZE=$2 FSYNC_FREQ=$3 cat <<EOF > /tmp/fio-job.ini [writers] rw=write fsync=$FSYNC_FREQ fallocate=none group_reporting=1 direct=0 bs=64k ioengine=sync size=$FILE_SIZE directory=$MNT numjobs=$NUM_JOBS EOF echo "Using config:" echo cat /tmp/fio-job.ini echo mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT fio /tmp/fio-job.ini umount $MNT The tests were performed for different numbers of jobs, file sizes and fsync frequency. A qemu VM using kvm was used, with 8 cores (the host has 12 cores, with cpu governance set to performance mode on all cores), 16GiB of ram (the host has 64GiB) and using a NVMe device directly (without an intermediary filesystem in the host). While running the tests, the host was not used for anything else, to avoid disturbing the tests. The obtained results were the following (the last line of fio's output was pasted). Starting with 16 jobs is where a significant difference is observable in this particular setup and hardware (differences highlighted below). The very small differences for tests with less than 16 jobs are possibly just noise and random. **** 1 job, file size 1G, fsync frequency 1 **** before this change: WRITE: bw=23.8MiB/s (24.9MB/s), 23.8MiB/s-23.8MiB/s (24.9MB/s-24.9MB/s), io=1024MiB (1074MB), run=43075-43075msec after this change: WRITE: bw=24.4MiB/s (25.6MB/s), 24.4MiB/s-24.4MiB/s (25.6MB/s-25.6MB/s), io=1024MiB (1074MB), run=41938-41938msec **** 2 jobs, file size 1G, fsync frequency 1 **** before this change: WRITE: bw=37.7MiB/s (39.5MB/s), 37.7MiB/s-37.7MiB/s (39.5MB/s-39.5MB/s), io=2048MiB (2147MB), run=54351-54351msec after this change: WRITE: bw=37.7MiB/s (39.5MB/s), 37.6MiB/s-37.6MiB/s (39.5MB/s-39.5MB/s), io=2048MiB (2147MB), run=54428-54428msec **** 4 jobs, file size 1G, fsync frequency 1 **** before this change: WRITE: bw=67.5MiB/s (70.8MB/s), 67.5MiB/s-67.5MiB/s (70.8MB/s-70.8MB/s), io=4096MiB (4295MB), run=60669-60669msec after this change: WRITE: bw=68.6MiB/s (71.0MB/s), 68.6MiB/s-68.6MiB/s (71.0MB/s-71.0MB/s), io=4096MiB (4295MB), run=59678-59678msec **** 8 jobs, file size 1G, fsync frequency 1 **** before this change: WRITE: bw=128MiB/s (134MB/s), 128MiB/s-128MiB/s (134MB/s-134MB/s), io=8192MiB (8590MB), run=64048-64048msec after this change: WRITE: bw=129MiB/s (135MB/s), 129MiB/s-129MiB/s (135MB/s-135MB/s), io=8192MiB (8590MB), run=63405-63405msec **** 16 jobs, file size 1G, fsync frequency 1 **** before this change: WRITE: bw=78.5MiB/s (82.3MB/s), 78.5MiB/s-78.5MiB/s (82.3MB/s-82.3MB/s), io=16.0GiB (17.2GB), run=208676-208676msec after this change: WRITE: bw=110MiB/s (115MB/s), 110MiB/s-110MiB/s (115MB/s-115MB/s), io=16.0GiB (17.2GB), run=149295-149295msec (+40.1% throughput, -28.5% runtime) **** 32 jobs, file size 1G, fsync frequency 1 **** before this change: WRITE: bw=58.8MiB/s (61.7MB/s), 58.8MiB/s-58.8MiB/s (61.7MB/s-61.7MB/s), io=32.0GiB (34.4GB), run=557134-557134msec after this change: WRITE: bw=76.1MiB/s (79.8MB/s), 76.1MiB/s-76.1MiB/s (79.8MB/s-79.8MB/s), io=32.0GiB (34.4GB), run=430550-430550msec (+29.4% throughput, -22.7% runtime) **** 64 jobs, file size 512M, fsync frequency 1 **** before this change: WRITE: bw=65.8MiB/s (68.0MB/s), 65.8MiB/s-65.8MiB/s (68.0MB/s-68.0MB/s), io=32.0GiB (34.4GB), run=498055-498055msec after this change: WRITE: bw=85.1MiB/s (89.2MB/s), 85.1MiB/s-85.1MiB/s (89.2MB/s-89.2MB/s), io=32.0GiB (34.4GB), run=385116-385116msec (+29.3% throughput, -22.7% runtime) **** 128 jobs, file size 256M, fsync frequency 1 **** before this change: WRITE: bw=54.7MiB/s (57.3MB/s), 54.7MiB/s-54.7MiB/s (57.3MB/s-57.3MB/s), io=32.0GiB (34.4GB), run=599373-599373msec after this change: WRITE: bw=121MiB/s (126MB/s), 121MiB/s-121MiB/s (126MB/s-126MB/s), io=32.0GiB (34.4GB), run=271907-271907msec (+121.2% throughput, -54.6% runtime) **** 256 jobs, file size 256M, fsync frequency 1 **** before this change: WRITE: bw=69.2MiB/s (72.5MB/s), 69.2MiB/s-69.2MiB/s (72.5MB/s-72.5MB/s), io=64.0GiB (68.7GB), run=947536-947536msec after this change: WRITE: bw=121MiB/s (127MB/s), 121MiB/s-121MiB/s (127MB/s-127MB/s), io=64.0GiB (68.7GB), run=541916-541916msec (+74.9% throughput, -42.8% runtime) **** 512 jobs, file size 128M, fsync frequency 1 **** before this change: WRITE: bw=85.4MiB/s (89.5MB/s), 85.4MiB/s-85.4MiB/s (89.5MB/s-89.5MB/s), io=64.0GiB (68.7GB), run=767734-767734msec after this change: WRITE: bw=141MiB/s (147MB/s), 141MiB/s-141MiB/s (147MB/s-147MB/s), io=64.0GiB (68.7GB), run=466022-466022msec (+65.1% throughput, -39.3% runtime) **** 1024 jobs, file size 128M, fsync frequency 1 **** before this change: WRITE: bw=115MiB/s (120MB/s), 115MiB/s-115MiB/s (120MB/s-120MB/s), io=128GiB (137GB), run=1143775-1143775msec after this change: WRITE: bw=171MiB/s (180MB/s), 171MiB/s-171MiB/s (180MB/s-180MB/s), io=128GiB (137GB), run=764843-764843msec (+48.7% throughput, -33.1% runtime) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-07-15 14:30:43 +03:00
/*
* If this inode was not used for reflink operations in the current
* transaction with new extents, then do the fast path, no need to
* worry about logging checksum items with overlapping ranges.
*/
if (inode->last_reflink_trans < trans->transid)
return btrfs_csum_file_blocks(trans, log_root, sums);
btrfs: fix corrupt log due to concurrent fsync of inodes with shared extents When we have extents shared amongst different inodes in the same subvolume, if we fsync them in parallel we can end up with checksum items in the log tree that represent ranges which overlap. For example, consider we have inodes A and B, both sharing an extent that covers the logical range from X to X + 64KiB: 1) Task A starts an fsync on inode A; 2) Task B starts an fsync on inode B; 3) Task A calls btrfs_csum_file_blocks(), and the first search in the log tree, through btrfs_lookup_csum(), returns -EFBIG because it finds an existing checksum item that covers the range from X - 64KiB to X; 4) Task A checks that the checksum item has not reached the maximum possible size (MAX_CSUM_ITEMS) and then releases the search path before it does another path search for insertion (through a direct call to btrfs_search_slot()); 5) As soon as task A releases the path and before it does the search for insertion, task B calls btrfs_csum_file_blocks() and gets -EFBIG too, because there is an existing checksum item that has an end offset that matches the start offset (X) of the checksum range we want to log; 6) Task B releases the path; 7) Task A does the path search for insertion (through btrfs_search_slot()) and then verifies that the checksum item that ends at offset X still exists and extends its size to insert the checksums for the range from X to X + 64KiB; 8) Task A releases the path and returns from btrfs_csum_file_blocks(), having inserted the checksums into an existing checksum item that got its size extended. At this point we have one checksum item in the log tree that covers the logical range from X - 64KiB to X + 64KiB; 9) Task B now does a search for insertion using btrfs_search_slot() too, but it finds that the previous checksum item no longer ends at the offset X, it now ends at an of offset X + 64KiB, so it leaves that item untouched. Then it releases the path and calls btrfs_insert_empty_item() that inserts a checksum item with a key offset corresponding to X and a size for inserting a single checksum (4 bytes in case of crc32c). Subsequent iterations end up extending this new checksum item so that it contains the checksums for the range from X to X + 64KiB. So after task B returns from btrfs_csum_file_blocks() we end up with two checksum items in the log tree that have overlapping ranges, one for the range from X - 64KiB to X + 64KiB, and another for the range from X to X + 64KiB. Having checksum items that represent ranges which overlap, regardless of being in the log tree or in the chekcsums tree, can lead to problems where checksums for a file range end up not being found. This type of problem has happened a few times in the past and the following commits fixed them and explain in detail why having checksum items with overlapping ranges is problematic: 27b9a8122ff71a "Btrfs: fix csum tree corruption, duplicate and outdated checksums" b84b8390d6009c "Btrfs: fix file read corruption after extent cloning and fsync" 40e046acbd2f36 "Btrfs: fix missing data checksums after replaying a log tree" Since this specific instance of the problem can only happen when logging inodes, because it is the only case where concurrent attempts to insert checksums for the same range can happen, fix the issue by using an extent io tree as a range lock to serialize checksum insertion during inode logging. This issue could often be reproduced by the test case generic/457 from fstests. When it happens it produces the following trace: BTRFS critical (device dm-0): corrupt leaf: root=18446744073709551610 block=30625792 slot=42, csum end range (15020032) goes beyond the start range (15015936) of the next csum item BTRFS info (device dm-0): leaf 30625792 gen 7 total ptrs 49 free space 2402 owner 18446744073709551610 BTRFS info (device dm-0): refs 1 lock (w:0 r:0 bw:0 br:0 sw:0 sr:0) lock_owner 0 current 15884 item 0 key (18446744073709551606 128 13979648) itemoff 3991 itemsize 4 item 1 key (18446744073709551606 128 13983744) itemoff 3987 itemsize 4 item 2 key (18446744073709551606 128 13987840) itemoff 3983 itemsize 4 item 3 key (18446744073709551606 128 13991936) itemoff 3979 itemsize 4 item 4 key (18446744073709551606 128 13996032) itemoff 3975 itemsize 4 item 5 key (18446744073709551606 128 14000128) itemoff 3971 itemsize 4 (...) BTRFS error (device dm-0): block=30625792 write time tree block corruption detected ------------[ cut here ]------------ WARNING: CPU: 1 PID: 15884 at fs/btrfs/disk-io.c:539 btree_csum_one_bio+0x268/0x2d0 [btrfs] Modules linked in: btrfs dm_thin_pool ... CPU: 1 PID: 15884 Comm: fsx Tainted: G W 5.6.0-rc7-btrfs-next-58 #1 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.12.0-59-gc9ba5276e321-prebuilt.qemu.org 04/01/2014 RIP: 0010:btree_csum_one_bio+0x268/0x2d0 [btrfs] Code: c7 c7 ... RSP: 0018:ffffbb0109e6f8e0 EFLAGS: 00010296 RAX: 0000000000000000 RBX: ffffe1c0847b6080 RCX: 0000000000000000 RDX: 0000000000000000 RSI: ffffffffaa963988 RDI: 0000000000000001 RBP: ffff956a4f4d2000 R08: 0000000000000000 R09: 0000000000000001 R10: 0000000000000526 R11: 0000000000000000 R12: ffff956a5cd28bb0 R13: 0000000000000000 R14: ffff956a649c9388 R15: 000000011ed82000 FS: 00007fb419959e80(0000) GS:ffff956a7aa00000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 0000000000fe6d54 CR3: 0000000138696005 CR4: 00000000003606e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: btree_submit_bio_hook+0x67/0xc0 [btrfs] submit_one_bio+0x31/0x50 [btrfs] btree_write_cache_pages+0x2db/0x4b0 [btrfs] ? __filemap_fdatawrite_range+0xb1/0x110 do_writepages+0x23/0x80 __filemap_fdatawrite_range+0xd2/0x110 btrfs_write_marked_extents+0x15e/0x180 [btrfs] btrfs_sync_log+0x206/0x10a0 [btrfs] ? kmem_cache_free+0x315/0x3b0 ? btrfs_log_inode+0x1e8/0xf90 [btrfs] ? __mutex_unlock_slowpath+0x45/0x2a0 ? lockref_put_or_lock+0x9/0x30 ? dput+0x2d/0x580 ? dput+0xb5/0x580 ? btrfs_sync_file+0x464/0x4d0 [btrfs] btrfs_sync_file+0x464/0x4d0 [btrfs] do_fsync+0x38/0x60 __x64_sys_fsync+0x10/0x20 do_syscall_64+0x5c/0x280 entry_SYSCALL_64_after_hwframe+0x49/0xbe RIP: 0033:0x7fb41953a6d0 Code: 48 3d ... RSP: 002b:00007ffcc86bd218 EFLAGS: 00000246 ORIG_RAX: 000000000000004a RAX: ffffffffffffffda RBX: 000000000000000d RCX: 00007fb41953a6d0 RDX: 0000000000000009 RSI: 0000000000040000 RDI: 0000000000000003 RBP: 0000000000040000 R08: 0000000000000001 R09: 0000000000000009 R10: 0000000000000064 R11: 0000000000000246 R12: 0000556cf4b2c060 R13: 0000000000000100 R14: 0000000000000000 R15: 0000556cf322b420 irq event stamp: 0 hardirqs last enabled at (0): [<0000000000000000>] 0x0 hardirqs last disabled at (0): [<ffffffffa96bdedf>] copy_process+0x74f/0x2020 softirqs last enabled at (0): [<ffffffffa96bdedf>] copy_process+0x74f/0x2020 softirqs last disabled at (0): [<0000000000000000>] 0x0 ---[ end trace d543fc76f5ad7fd8 ]--- In that trace the tree checker detected the overlapping checksum items at the time when we triggered writeback for the log tree when syncing the log. Another trace that can happen is due to BUG_ON() when deleting checksum items while logging an inode: BTRFS critical (device dm-0): slot 81 key (18446744073709551606 128 13635584) new key (18446744073709551606 128 13635584) BTRFS info (device dm-0): leaf 30949376 gen 7 total ptrs 98 free space 8527 owner 18446744073709551610 BTRFS info (device dm-0): refs 4 lock (w:1 r:0 bw:0 br:0 sw:1 sr:0) lock_owner 13473 current 13473 item 0 key (257 1 0) itemoff 16123 itemsize 160 inode generation 7 size 262144 mode 100600 item 1 key (257 12 256) itemoff 16103 itemsize 20 item 2 key (257 108 0) itemoff 16050 itemsize 53 extent data disk bytenr 13631488 nr 4096 extent data offset 0 nr 131072 ram 131072 (...) ------------[ cut here ]------------ kernel BUG at fs/btrfs/ctree.c:3153! invalid opcode: 0000 [#1] PREEMPT SMP DEBUG_PAGEALLOC PTI CPU: 1 PID: 13473 Comm: fsx Not tainted 5.6.0-rc7-btrfs-next-58 #1 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.12.0-59-gc9ba5276e321-prebuilt.qemu.org 04/01/2014 RIP: 0010:btrfs_set_item_key_safe+0x1ea/0x270 [btrfs] Code: 0f b6 ... RSP: 0018:ffff95e3889179d0 EFLAGS: 00010282 RAX: 0000000000000000 RBX: 0000000000000051 RCX: 0000000000000000 RDX: 0000000000000000 RSI: ffffffffb7763988 RDI: 0000000000000001 RBP: fffffffffffffff6 R08: 0000000000000000 R09: 0000000000000001 R10: 00000000000009ef R11: 0000000000000000 R12: ffff8912a8ba5a08 R13: ffff95e388917a06 R14: ffff89138dcf68c8 R15: ffff95e388917ace FS: 00007fe587084e80(0000) GS:ffff8913baa00000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 00007fe587091000 CR3: 0000000126dac005 CR4: 00000000003606e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: btrfs_del_csums+0x2f4/0x540 [btrfs] copy_items+0x4b5/0x560 [btrfs] btrfs_log_inode+0x910/0xf90 [btrfs] btrfs_log_inode_parent+0x2a0/0xe40 [btrfs] ? dget_parent+0x5/0x370 btrfs_log_dentry_safe+0x4a/0x70 [btrfs] btrfs_sync_file+0x42b/0x4d0 [btrfs] __x64_sys_msync+0x199/0x200 do_syscall_64+0x5c/0x280 entry_SYSCALL_64_after_hwframe+0x49/0xbe RIP: 0033:0x7fe586c65760 Code: 00 f7 ... RSP: 002b:00007ffe250f98b8 EFLAGS: 00000246 ORIG_RAX: 000000000000001a RAX: ffffffffffffffda RBX: 00000000000040e1 RCX: 00007fe586c65760 RDX: 0000000000000004 RSI: 0000000000006b51 RDI: 00007fe58708b000 RBP: 0000000000006a70 R08: 0000000000000003 R09: 00007fe58700cb61 R10: 0000000000000100 R11: 0000000000000246 R12: 00000000000000e1 R13: 00007fe58708b000 R14: 0000000000006b51 R15: 0000558de021a420 Modules linked in: dm_log_writes ... ---[ end trace c92a7f447a8515f5 ]--- CC: stable@vger.kernel.org # 4.4+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-05-18 14:14:50 +03:00
/*
* Serialize logging for checksums. This is to avoid racing with the
* same checksum being logged by another task that is logging another
* file which happens to refer to the same extent as well. Such races
* can leave checksum items in the log with overlapping ranges.
*/
ret = lock_extent_bits(&log_root->log_csum_range, sums->bytenr,
lock_end, &cached_state);
if (ret)
return ret;
Btrfs: fix missing data checksums after replaying a log tree When logging a file that has shared extents (reflinked with other files or with itself), we can end up logging multiple checksum items that cover overlapping ranges. This confuses the search for checksums at log replay time causing some checksums to never be added to the fs/subvolume tree. Consider the following example of a file that shares the same extent at offsets 0 and 256Kb: [ bytenr 13893632, offset 64Kb, len 64Kb ] 0 64Kb [ bytenr 13631488, offset 64Kb, len 192Kb ] 64Kb 256Kb [ bytenr 13893632, offset 0, len 256Kb ] 256Kb 512Kb When logging the inode, at tree-log.c:copy_items(), when processing the file extent item at offset 0, we log a checksum item covering the range 13959168 to 14024704, which corresponds to 13893632 + 64Kb and 13893632 + 64Kb + 64Kb, respectively. Later when processing the extent item at offset 256K, we log the checksums for the range from 13893632 to 14155776 (which corresponds to 13893632 + 256Kb). These checksums get merged with the checksum item for the range from 13631488 to 13893632 (13631488 + 256Kb), logged by a previous fsync. So after this we get the two following checksum items in the log tree: (...) item 6 key (EXTENT_CSUM EXTENT_CSUM 13631488) itemoff 3095 itemsize 512 range start 13631488 end 14155776 length 524288 item 7 key (EXTENT_CSUM EXTENT_CSUM 13959168) itemoff 3031 itemsize 64 range start 13959168 end 14024704 length 65536 The first one covers the range from the second one, they overlap. So far this does not cause a problem after replaying the log, because when replaying the file extent item for offset 256K, we copy all the checksums for the extent 13893632 from the log tree to the fs/subvolume tree, since searching for an checksum item for bytenr 13893632 leaves us at the first checksum item, which covers the whole range of the extent. However if we write 64Kb to file offset 256Kb for example, we will not be able to find and copy the checksums for the last 128Kb of the extent at bytenr 13893632, referenced by the file range 384Kb to 512Kb. After writing 64Kb into file offset 256Kb we get the following extent layout for our file: [ bytenr 13893632, offset 64K, len 64Kb ] 0 64Kb [ bytenr 13631488, offset 64Kb, len 192Kb ] 64Kb 256Kb [ bytenr 14155776, offset 0, len 64Kb ] 256Kb 320Kb [ bytenr 13893632, offset 64Kb, len 192Kb ] 320Kb 512Kb After fsync'ing the file, if we have a power failure and then mount the filesystem to replay the log, the following happens: 1) When replaying the file extent item for file offset 320Kb, we lookup for the checksums for the extent range from 13959168 (13893632 + 64Kb) to 14155776 (13893632 + 256Kb), through a call to btrfs_lookup_csums_range(); 2) btrfs_lookup_csums_range() finds the checksum item that starts precisely at offset 13959168 (item 7 in the log tree, shown before); 3) However that checksum item only covers 64Kb of data, and not 192Kb of data; 4) As a result only the checksums for the first 64Kb of data referenced by the file extent item are found and copied to the fs/subvolume tree. The remaining 128Kb of data, file range 384Kb to 512Kb, doesn't get the corresponding data checksums found and copied to the fs/subvolume tree. 5) After replaying the log userspace will not be able to read the file range from 384Kb to 512Kb, because the checksums are missing and resulting in an -EIO error. The following steps reproduce this scenario: $ mkfs.btrfs -f /dev/sdc $ mount /dev/sdc /mnt/sdc $ xfs_io -f -c "pwrite -S 0xa3 0 256K" /mnt/sdc/foobar $ xfs_io -c "fsync" /mnt/sdc/foobar $ xfs_io -c "pwrite -S 0xc7 256K 256K" /mnt/sdc/foobar $ xfs_io -c "reflink /mnt/sdc/foobar 320K 0 64K" /mnt/sdc/foobar $ xfs_io -c "fsync" /mnt/sdc/foobar $ xfs_io -c "pwrite -S 0xe5 256K 64K" /mnt/sdc/foobar $ xfs_io -c "fsync" /mnt/sdc/foobar <power failure> $ mount /dev/sdc /mnt/sdc $ md5sum /mnt/sdc/foobar md5sum: /mnt/sdc/foobar: Input/output error $ dmesg | tail [165305.003464] BTRFS info (device sdc): no csum found for inode 257 start 401408 [165305.004014] BTRFS info (device sdc): no csum found for inode 257 start 405504 [165305.004559] BTRFS info (device sdc): no csum found for inode 257 start 409600 [165305.005101] BTRFS info (device sdc): no csum found for inode 257 start 413696 [165305.005627] BTRFS info (device sdc): no csum found for inode 257 start 417792 [165305.006134] BTRFS info (device sdc): no csum found for inode 257 start 421888 [165305.006625] BTRFS info (device sdc): no csum found for inode 257 start 425984 [165305.007278] BTRFS info (device sdc): no csum found for inode 257 start 430080 [165305.008248] BTRFS warning (device sdc): csum failed root 5 ino 257 off 393216 csum 0x1337385e expected csum 0x00000000 mirror 1 [165305.009550] BTRFS warning (device sdc): csum failed root 5 ino 257 off 393216 csum 0x1337385e expected csum 0x00000000 mirror 1 Fix this simply by deleting first any checksums, from the log tree, for the range of the extent we are logging at copy_items(). This ensures we do not get checksum items in the log tree that have overlapping ranges. This is a long time issue that has been present since we have the clone (and deduplication) ioctl, and can happen both when an extent is shared between different files and within the same file. A test case for fstests follows soon. CC: stable@vger.kernel.org # 4.4+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-12-05 19:58:30 +03:00
/*
* Due to extent cloning, we might have logged a csum item that covers a
* subrange of a cloned extent, and later we can end up logging a csum
* item for a larger subrange of the same extent or the entire range.
* This would leave csum items in the log tree that cover the same range
* and break the searches for checksums in the log tree, resulting in
* some checksums missing in the fs/subvolume tree. So just delete (or
* trim and adjust) any existing csum items in the log for this range.
*/
ret = btrfs_del_csums(trans, log_root, sums->bytenr, sums->len);
btrfs: fix corrupt log due to concurrent fsync of inodes with shared extents When we have extents shared amongst different inodes in the same subvolume, if we fsync them in parallel we can end up with checksum items in the log tree that represent ranges which overlap. For example, consider we have inodes A and B, both sharing an extent that covers the logical range from X to X + 64KiB: 1) Task A starts an fsync on inode A; 2) Task B starts an fsync on inode B; 3) Task A calls btrfs_csum_file_blocks(), and the first search in the log tree, through btrfs_lookup_csum(), returns -EFBIG because it finds an existing checksum item that covers the range from X - 64KiB to X; 4) Task A checks that the checksum item has not reached the maximum possible size (MAX_CSUM_ITEMS) and then releases the search path before it does another path search for insertion (through a direct call to btrfs_search_slot()); 5) As soon as task A releases the path and before it does the search for insertion, task B calls btrfs_csum_file_blocks() and gets -EFBIG too, because there is an existing checksum item that has an end offset that matches the start offset (X) of the checksum range we want to log; 6) Task B releases the path; 7) Task A does the path search for insertion (through btrfs_search_slot()) and then verifies that the checksum item that ends at offset X still exists and extends its size to insert the checksums for the range from X to X + 64KiB; 8) Task A releases the path and returns from btrfs_csum_file_blocks(), having inserted the checksums into an existing checksum item that got its size extended. At this point we have one checksum item in the log tree that covers the logical range from X - 64KiB to X + 64KiB; 9) Task B now does a search for insertion using btrfs_search_slot() too, but it finds that the previous checksum item no longer ends at the offset X, it now ends at an of offset X + 64KiB, so it leaves that item untouched. Then it releases the path and calls btrfs_insert_empty_item() that inserts a checksum item with a key offset corresponding to X and a size for inserting a single checksum (4 bytes in case of crc32c). Subsequent iterations end up extending this new checksum item so that it contains the checksums for the range from X to X + 64KiB. So after task B returns from btrfs_csum_file_blocks() we end up with two checksum items in the log tree that have overlapping ranges, one for the range from X - 64KiB to X + 64KiB, and another for the range from X to X + 64KiB. Having checksum items that represent ranges which overlap, regardless of being in the log tree or in the chekcsums tree, can lead to problems where checksums for a file range end up not being found. This type of problem has happened a few times in the past and the following commits fixed them and explain in detail why having checksum items with overlapping ranges is problematic: 27b9a8122ff71a "Btrfs: fix csum tree corruption, duplicate and outdated checksums" b84b8390d6009c "Btrfs: fix file read corruption after extent cloning and fsync" 40e046acbd2f36 "Btrfs: fix missing data checksums after replaying a log tree" Since this specific instance of the problem can only happen when logging inodes, because it is the only case where concurrent attempts to insert checksums for the same range can happen, fix the issue by using an extent io tree as a range lock to serialize checksum insertion during inode logging. This issue could often be reproduced by the test case generic/457 from fstests. When it happens it produces the following trace: BTRFS critical (device dm-0): corrupt leaf: root=18446744073709551610 block=30625792 slot=42, csum end range (15020032) goes beyond the start range (15015936) of the next csum item BTRFS info (device dm-0): leaf 30625792 gen 7 total ptrs 49 free space 2402 owner 18446744073709551610 BTRFS info (device dm-0): refs 1 lock (w:0 r:0 bw:0 br:0 sw:0 sr:0) lock_owner 0 current 15884 item 0 key (18446744073709551606 128 13979648) itemoff 3991 itemsize 4 item 1 key (18446744073709551606 128 13983744) itemoff 3987 itemsize 4 item 2 key (18446744073709551606 128 13987840) itemoff 3983 itemsize 4 item 3 key (18446744073709551606 128 13991936) itemoff 3979 itemsize 4 item 4 key (18446744073709551606 128 13996032) itemoff 3975 itemsize 4 item 5 key (18446744073709551606 128 14000128) itemoff 3971 itemsize 4 (...) BTRFS error (device dm-0): block=30625792 write time tree block corruption detected ------------[ cut here ]------------ WARNING: CPU: 1 PID: 15884 at fs/btrfs/disk-io.c:539 btree_csum_one_bio+0x268/0x2d0 [btrfs] Modules linked in: btrfs dm_thin_pool ... CPU: 1 PID: 15884 Comm: fsx Tainted: G W 5.6.0-rc7-btrfs-next-58 #1 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.12.0-59-gc9ba5276e321-prebuilt.qemu.org 04/01/2014 RIP: 0010:btree_csum_one_bio+0x268/0x2d0 [btrfs] Code: c7 c7 ... RSP: 0018:ffffbb0109e6f8e0 EFLAGS: 00010296 RAX: 0000000000000000 RBX: ffffe1c0847b6080 RCX: 0000000000000000 RDX: 0000000000000000 RSI: ffffffffaa963988 RDI: 0000000000000001 RBP: ffff956a4f4d2000 R08: 0000000000000000 R09: 0000000000000001 R10: 0000000000000526 R11: 0000000000000000 R12: ffff956a5cd28bb0 R13: 0000000000000000 R14: ffff956a649c9388 R15: 000000011ed82000 FS: 00007fb419959e80(0000) GS:ffff956a7aa00000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 0000000000fe6d54 CR3: 0000000138696005 CR4: 00000000003606e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: btree_submit_bio_hook+0x67/0xc0 [btrfs] submit_one_bio+0x31/0x50 [btrfs] btree_write_cache_pages+0x2db/0x4b0 [btrfs] ? __filemap_fdatawrite_range+0xb1/0x110 do_writepages+0x23/0x80 __filemap_fdatawrite_range+0xd2/0x110 btrfs_write_marked_extents+0x15e/0x180 [btrfs] btrfs_sync_log+0x206/0x10a0 [btrfs] ? kmem_cache_free+0x315/0x3b0 ? btrfs_log_inode+0x1e8/0xf90 [btrfs] ? __mutex_unlock_slowpath+0x45/0x2a0 ? lockref_put_or_lock+0x9/0x30 ? dput+0x2d/0x580 ? dput+0xb5/0x580 ? btrfs_sync_file+0x464/0x4d0 [btrfs] btrfs_sync_file+0x464/0x4d0 [btrfs] do_fsync+0x38/0x60 __x64_sys_fsync+0x10/0x20 do_syscall_64+0x5c/0x280 entry_SYSCALL_64_after_hwframe+0x49/0xbe RIP: 0033:0x7fb41953a6d0 Code: 48 3d ... RSP: 002b:00007ffcc86bd218 EFLAGS: 00000246 ORIG_RAX: 000000000000004a RAX: ffffffffffffffda RBX: 000000000000000d RCX: 00007fb41953a6d0 RDX: 0000000000000009 RSI: 0000000000040000 RDI: 0000000000000003 RBP: 0000000000040000 R08: 0000000000000001 R09: 0000000000000009 R10: 0000000000000064 R11: 0000000000000246 R12: 0000556cf4b2c060 R13: 0000000000000100 R14: 0000000000000000 R15: 0000556cf322b420 irq event stamp: 0 hardirqs last enabled at (0): [<0000000000000000>] 0x0 hardirqs last disabled at (0): [<ffffffffa96bdedf>] copy_process+0x74f/0x2020 softirqs last enabled at (0): [<ffffffffa96bdedf>] copy_process+0x74f/0x2020 softirqs last disabled at (0): [<0000000000000000>] 0x0 ---[ end trace d543fc76f5ad7fd8 ]--- In that trace the tree checker detected the overlapping checksum items at the time when we triggered writeback for the log tree when syncing the log. Another trace that can happen is due to BUG_ON() when deleting checksum items while logging an inode: BTRFS critical (device dm-0): slot 81 key (18446744073709551606 128 13635584) new key (18446744073709551606 128 13635584) BTRFS info (device dm-0): leaf 30949376 gen 7 total ptrs 98 free space 8527 owner 18446744073709551610 BTRFS info (device dm-0): refs 4 lock (w:1 r:0 bw:0 br:0 sw:1 sr:0) lock_owner 13473 current 13473 item 0 key (257 1 0) itemoff 16123 itemsize 160 inode generation 7 size 262144 mode 100600 item 1 key (257 12 256) itemoff 16103 itemsize 20 item 2 key (257 108 0) itemoff 16050 itemsize 53 extent data disk bytenr 13631488 nr 4096 extent data offset 0 nr 131072 ram 131072 (...) ------------[ cut here ]------------ kernel BUG at fs/btrfs/ctree.c:3153! invalid opcode: 0000 [#1] PREEMPT SMP DEBUG_PAGEALLOC PTI CPU: 1 PID: 13473 Comm: fsx Not tainted 5.6.0-rc7-btrfs-next-58 #1 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.12.0-59-gc9ba5276e321-prebuilt.qemu.org 04/01/2014 RIP: 0010:btrfs_set_item_key_safe+0x1ea/0x270 [btrfs] Code: 0f b6 ... RSP: 0018:ffff95e3889179d0 EFLAGS: 00010282 RAX: 0000000000000000 RBX: 0000000000000051 RCX: 0000000000000000 RDX: 0000000000000000 RSI: ffffffffb7763988 RDI: 0000000000000001 RBP: fffffffffffffff6 R08: 0000000000000000 R09: 0000000000000001 R10: 00000000000009ef R11: 0000000000000000 R12: ffff8912a8ba5a08 R13: ffff95e388917a06 R14: ffff89138dcf68c8 R15: ffff95e388917ace FS: 00007fe587084e80(0000) GS:ffff8913baa00000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 00007fe587091000 CR3: 0000000126dac005 CR4: 00000000003606e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: btrfs_del_csums+0x2f4/0x540 [btrfs] copy_items+0x4b5/0x560 [btrfs] btrfs_log_inode+0x910/0xf90 [btrfs] btrfs_log_inode_parent+0x2a0/0xe40 [btrfs] ? dget_parent+0x5/0x370 btrfs_log_dentry_safe+0x4a/0x70 [btrfs] btrfs_sync_file+0x42b/0x4d0 [btrfs] __x64_sys_msync+0x199/0x200 do_syscall_64+0x5c/0x280 entry_SYSCALL_64_after_hwframe+0x49/0xbe RIP: 0033:0x7fe586c65760 Code: 00 f7 ... RSP: 002b:00007ffe250f98b8 EFLAGS: 00000246 ORIG_RAX: 000000000000001a RAX: ffffffffffffffda RBX: 00000000000040e1 RCX: 00007fe586c65760 RDX: 0000000000000004 RSI: 0000000000006b51 RDI: 00007fe58708b000 RBP: 0000000000006a70 R08: 0000000000000003 R09: 00007fe58700cb61 R10: 0000000000000100 R11: 0000000000000246 R12: 00000000000000e1 R13: 00007fe58708b000 R14: 0000000000006b51 R15: 0000558de021a420 Modules linked in: dm_log_writes ... ---[ end trace c92a7f447a8515f5 ]--- CC: stable@vger.kernel.org # 4.4+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-05-18 14:14:50 +03:00
if (!ret)
ret = btrfs_csum_file_blocks(trans, log_root, sums);
Btrfs: fix missing data checksums after replaying a log tree When logging a file that has shared extents (reflinked with other files or with itself), we can end up logging multiple checksum items that cover overlapping ranges. This confuses the search for checksums at log replay time causing some checksums to never be added to the fs/subvolume tree. Consider the following example of a file that shares the same extent at offsets 0 and 256Kb: [ bytenr 13893632, offset 64Kb, len 64Kb ] 0 64Kb [ bytenr 13631488, offset 64Kb, len 192Kb ] 64Kb 256Kb [ bytenr 13893632, offset 0, len 256Kb ] 256Kb 512Kb When logging the inode, at tree-log.c:copy_items(), when processing the file extent item at offset 0, we log a checksum item covering the range 13959168 to 14024704, which corresponds to 13893632 + 64Kb and 13893632 + 64Kb + 64Kb, respectively. Later when processing the extent item at offset 256K, we log the checksums for the range from 13893632 to 14155776 (which corresponds to 13893632 + 256Kb). These checksums get merged with the checksum item for the range from 13631488 to 13893632 (13631488 + 256Kb), logged by a previous fsync. So after this we get the two following checksum items in the log tree: (...) item 6 key (EXTENT_CSUM EXTENT_CSUM 13631488) itemoff 3095 itemsize 512 range start 13631488 end 14155776 length 524288 item 7 key (EXTENT_CSUM EXTENT_CSUM 13959168) itemoff 3031 itemsize 64 range start 13959168 end 14024704 length 65536 The first one covers the range from the second one, they overlap. So far this does not cause a problem after replaying the log, because when replaying the file extent item for offset 256K, we copy all the checksums for the extent 13893632 from the log tree to the fs/subvolume tree, since searching for an checksum item for bytenr 13893632 leaves us at the first checksum item, which covers the whole range of the extent. However if we write 64Kb to file offset 256Kb for example, we will not be able to find and copy the checksums for the last 128Kb of the extent at bytenr 13893632, referenced by the file range 384Kb to 512Kb. After writing 64Kb into file offset 256Kb we get the following extent layout for our file: [ bytenr 13893632, offset 64K, len 64Kb ] 0 64Kb [ bytenr 13631488, offset 64Kb, len 192Kb ] 64Kb 256Kb [ bytenr 14155776, offset 0, len 64Kb ] 256Kb 320Kb [ bytenr 13893632, offset 64Kb, len 192Kb ] 320Kb 512Kb After fsync'ing the file, if we have a power failure and then mount the filesystem to replay the log, the following happens: 1) When replaying the file extent item for file offset 320Kb, we lookup for the checksums for the extent range from 13959168 (13893632 + 64Kb) to 14155776 (13893632 + 256Kb), through a call to btrfs_lookup_csums_range(); 2) btrfs_lookup_csums_range() finds the checksum item that starts precisely at offset 13959168 (item 7 in the log tree, shown before); 3) However that checksum item only covers 64Kb of data, and not 192Kb of data; 4) As a result only the checksums for the first 64Kb of data referenced by the file extent item are found and copied to the fs/subvolume tree. The remaining 128Kb of data, file range 384Kb to 512Kb, doesn't get the corresponding data checksums found and copied to the fs/subvolume tree. 5) After replaying the log userspace will not be able to read the file range from 384Kb to 512Kb, because the checksums are missing and resulting in an -EIO error. The following steps reproduce this scenario: $ mkfs.btrfs -f /dev/sdc $ mount /dev/sdc /mnt/sdc $ xfs_io -f -c "pwrite -S 0xa3 0 256K" /mnt/sdc/foobar $ xfs_io -c "fsync" /mnt/sdc/foobar $ xfs_io -c "pwrite -S 0xc7 256K 256K" /mnt/sdc/foobar $ xfs_io -c "reflink /mnt/sdc/foobar 320K 0 64K" /mnt/sdc/foobar $ xfs_io -c "fsync" /mnt/sdc/foobar $ xfs_io -c "pwrite -S 0xe5 256K 64K" /mnt/sdc/foobar $ xfs_io -c "fsync" /mnt/sdc/foobar <power failure> $ mount /dev/sdc /mnt/sdc $ md5sum /mnt/sdc/foobar md5sum: /mnt/sdc/foobar: Input/output error $ dmesg | tail [165305.003464] BTRFS info (device sdc): no csum found for inode 257 start 401408 [165305.004014] BTRFS info (device sdc): no csum found for inode 257 start 405504 [165305.004559] BTRFS info (device sdc): no csum found for inode 257 start 409600 [165305.005101] BTRFS info (device sdc): no csum found for inode 257 start 413696 [165305.005627] BTRFS info (device sdc): no csum found for inode 257 start 417792 [165305.006134] BTRFS info (device sdc): no csum found for inode 257 start 421888 [165305.006625] BTRFS info (device sdc): no csum found for inode 257 start 425984 [165305.007278] BTRFS info (device sdc): no csum found for inode 257 start 430080 [165305.008248] BTRFS warning (device sdc): csum failed root 5 ino 257 off 393216 csum 0x1337385e expected csum 0x00000000 mirror 1 [165305.009550] BTRFS warning (device sdc): csum failed root 5 ino 257 off 393216 csum 0x1337385e expected csum 0x00000000 mirror 1 Fix this simply by deleting first any checksums, from the log tree, for the range of the extent we are logging at copy_items(). This ensures we do not get checksum items in the log tree that have overlapping ranges. This is a long time issue that has been present since we have the clone (and deduplication) ioctl, and can happen both when an extent is shared between different files and within the same file. A test case for fstests follows soon. CC: stable@vger.kernel.org # 4.4+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-12-05 19:58:30 +03:00
btrfs: fix corrupt log due to concurrent fsync of inodes with shared extents When we have extents shared amongst different inodes in the same subvolume, if we fsync them in parallel we can end up with checksum items in the log tree that represent ranges which overlap. For example, consider we have inodes A and B, both sharing an extent that covers the logical range from X to X + 64KiB: 1) Task A starts an fsync on inode A; 2) Task B starts an fsync on inode B; 3) Task A calls btrfs_csum_file_blocks(), and the first search in the log tree, through btrfs_lookup_csum(), returns -EFBIG because it finds an existing checksum item that covers the range from X - 64KiB to X; 4) Task A checks that the checksum item has not reached the maximum possible size (MAX_CSUM_ITEMS) and then releases the search path before it does another path search for insertion (through a direct call to btrfs_search_slot()); 5) As soon as task A releases the path and before it does the search for insertion, task B calls btrfs_csum_file_blocks() and gets -EFBIG too, because there is an existing checksum item that has an end offset that matches the start offset (X) of the checksum range we want to log; 6) Task B releases the path; 7) Task A does the path search for insertion (through btrfs_search_slot()) and then verifies that the checksum item that ends at offset X still exists and extends its size to insert the checksums for the range from X to X + 64KiB; 8) Task A releases the path and returns from btrfs_csum_file_blocks(), having inserted the checksums into an existing checksum item that got its size extended. At this point we have one checksum item in the log tree that covers the logical range from X - 64KiB to X + 64KiB; 9) Task B now does a search for insertion using btrfs_search_slot() too, but it finds that the previous checksum item no longer ends at the offset X, it now ends at an of offset X + 64KiB, so it leaves that item untouched. Then it releases the path and calls btrfs_insert_empty_item() that inserts a checksum item with a key offset corresponding to X and a size for inserting a single checksum (4 bytes in case of crc32c). Subsequent iterations end up extending this new checksum item so that it contains the checksums for the range from X to X + 64KiB. So after task B returns from btrfs_csum_file_blocks() we end up with two checksum items in the log tree that have overlapping ranges, one for the range from X - 64KiB to X + 64KiB, and another for the range from X to X + 64KiB. Having checksum items that represent ranges which overlap, regardless of being in the log tree or in the chekcsums tree, can lead to problems where checksums for a file range end up not being found. This type of problem has happened a few times in the past and the following commits fixed them and explain in detail why having checksum items with overlapping ranges is problematic: 27b9a8122ff71a "Btrfs: fix csum tree corruption, duplicate and outdated checksums" b84b8390d6009c "Btrfs: fix file read corruption after extent cloning and fsync" 40e046acbd2f36 "Btrfs: fix missing data checksums after replaying a log tree" Since this specific instance of the problem can only happen when logging inodes, because it is the only case where concurrent attempts to insert checksums for the same range can happen, fix the issue by using an extent io tree as a range lock to serialize checksum insertion during inode logging. This issue could often be reproduced by the test case generic/457 from fstests. When it happens it produces the following trace: BTRFS critical (device dm-0): corrupt leaf: root=18446744073709551610 block=30625792 slot=42, csum end range (15020032) goes beyond the start range (15015936) of the next csum item BTRFS info (device dm-0): leaf 30625792 gen 7 total ptrs 49 free space 2402 owner 18446744073709551610 BTRFS info (device dm-0): refs 1 lock (w:0 r:0 bw:0 br:0 sw:0 sr:0) lock_owner 0 current 15884 item 0 key (18446744073709551606 128 13979648) itemoff 3991 itemsize 4 item 1 key (18446744073709551606 128 13983744) itemoff 3987 itemsize 4 item 2 key (18446744073709551606 128 13987840) itemoff 3983 itemsize 4 item 3 key (18446744073709551606 128 13991936) itemoff 3979 itemsize 4 item 4 key (18446744073709551606 128 13996032) itemoff 3975 itemsize 4 item 5 key (18446744073709551606 128 14000128) itemoff 3971 itemsize 4 (...) BTRFS error (device dm-0): block=30625792 write time tree block corruption detected ------------[ cut here ]------------ WARNING: CPU: 1 PID: 15884 at fs/btrfs/disk-io.c:539 btree_csum_one_bio+0x268/0x2d0 [btrfs] Modules linked in: btrfs dm_thin_pool ... CPU: 1 PID: 15884 Comm: fsx Tainted: G W 5.6.0-rc7-btrfs-next-58 #1 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.12.0-59-gc9ba5276e321-prebuilt.qemu.org 04/01/2014 RIP: 0010:btree_csum_one_bio+0x268/0x2d0 [btrfs] Code: c7 c7 ... RSP: 0018:ffffbb0109e6f8e0 EFLAGS: 00010296 RAX: 0000000000000000 RBX: ffffe1c0847b6080 RCX: 0000000000000000 RDX: 0000000000000000 RSI: ffffffffaa963988 RDI: 0000000000000001 RBP: ffff956a4f4d2000 R08: 0000000000000000 R09: 0000000000000001 R10: 0000000000000526 R11: 0000000000000000 R12: ffff956a5cd28bb0 R13: 0000000000000000 R14: ffff956a649c9388 R15: 000000011ed82000 FS: 00007fb419959e80(0000) GS:ffff956a7aa00000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 0000000000fe6d54 CR3: 0000000138696005 CR4: 00000000003606e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: btree_submit_bio_hook+0x67/0xc0 [btrfs] submit_one_bio+0x31/0x50 [btrfs] btree_write_cache_pages+0x2db/0x4b0 [btrfs] ? __filemap_fdatawrite_range+0xb1/0x110 do_writepages+0x23/0x80 __filemap_fdatawrite_range+0xd2/0x110 btrfs_write_marked_extents+0x15e/0x180 [btrfs] btrfs_sync_log+0x206/0x10a0 [btrfs] ? kmem_cache_free+0x315/0x3b0 ? btrfs_log_inode+0x1e8/0xf90 [btrfs] ? __mutex_unlock_slowpath+0x45/0x2a0 ? lockref_put_or_lock+0x9/0x30 ? dput+0x2d/0x580 ? dput+0xb5/0x580 ? btrfs_sync_file+0x464/0x4d0 [btrfs] btrfs_sync_file+0x464/0x4d0 [btrfs] do_fsync+0x38/0x60 __x64_sys_fsync+0x10/0x20 do_syscall_64+0x5c/0x280 entry_SYSCALL_64_after_hwframe+0x49/0xbe RIP: 0033:0x7fb41953a6d0 Code: 48 3d ... RSP: 002b:00007ffcc86bd218 EFLAGS: 00000246 ORIG_RAX: 000000000000004a RAX: ffffffffffffffda RBX: 000000000000000d RCX: 00007fb41953a6d0 RDX: 0000000000000009 RSI: 0000000000040000 RDI: 0000000000000003 RBP: 0000000000040000 R08: 0000000000000001 R09: 0000000000000009 R10: 0000000000000064 R11: 0000000000000246 R12: 0000556cf4b2c060 R13: 0000000000000100 R14: 0000000000000000 R15: 0000556cf322b420 irq event stamp: 0 hardirqs last enabled at (0): [<0000000000000000>] 0x0 hardirqs last disabled at (0): [<ffffffffa96bdedf>] copy_process+0x74f/0x2020 softirqs last enabled at (0): [<ffffffffa96bdedf>] copy_process+0x74f/0x2020 softirqs last disabled at (0): [<0000000000000000>] 0x0 ---[ end trace d543fc76f5ad7fd8 ]--- In that trace the tree checker detected the overlapping checksum items at the time when we triggered writeback for the log tree when syncing the log. Another trace that can happen is due to BUG_ON() when deleting checksum items while logging an inode: BTRFS critical (device dm-0): slot 81 key (18446744073709551606 128 13635584) new key (18446744073709551606 128 13635584) BTRFS info (device dm-0): leaf 30949376 gen 7 total ptrs 98 free space 8527 owner 18446744073709551610 BTRFS info (device dm-0): refs 4 lock (w:1 r:0 bw:0 br:0 sw:1 sr:0) lock_owner 13473 current 13473 item 0 key (257 1 0) itemoff 16123 itemsize 160 inode generation 7 size 262144 mode 100600 item 1 key (257 12 256) itemoff 16103 itemsize 20 item 2 key (257 108 0) itemoff 16050 itemsize 53 extent data disk bytenr 13631488 nr 4096 extent data offset 0 nr 131072 ram 131072 (...) ------------[ cut here ]------------ kernel BUG at fs/btrfs/ctree.c:3153! invalid opcode: 0000 [#1] PREEMPT SMP DEBUG_PAGEALLOC PTI CPU: 1 PID: 13473 Comm: fsx Not tainted 5.6.0-rc7-btrfs-next-58 #1 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.12.0-59-gc9ba5276e321-prebuilt.qemu.org 04/01/2014 RIP: 0010:btrfs_set_item_key_safe+0x1ea/0x270 [btrfs] Code: 0f b6 ... RSP: 0018:ffff95e3889179d0 EFLAGS: 00010282 RAX: 0000000000000000 RBX: 0000000000000051 RCX: 0000000000000000 RDX: 0000000000000000 RSI: ffffffffb7763988 RDI: 0000000000000001 RBP: fffffffffffffff6 R08: 0000000000000000 R09: 0000000000000001 R10: 00000000000009ef R11: 0000000000000000 R12: ffff8912a8ba5a08 R13: ffff95e388917a06 R14: ffff89138dcf68c8 R15: ffff95e388917ace FS: 00007fe587084e80(0000) GS:ffff8913baa00000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 00007fe587091000 CR3: 0000000126dac005 CR4: 00000000003606e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: btrfs_del_csums+0x2f4/0x540 [btrfs] copy_items+0x4b5/0x560 [btrfs] btrfs_log_inode+0x910/0xf90 [btrfs] btrfs_log_inode_parent+0x2a0/0xe40 [btrfs] ? dget_parent+0x5/0x370 btrfs_log_dentry_safe+0x4a/0x70 [btrfs] btrfs_sync_file+0x42b/0x4d0 [btrfs] __x64_sys_msync+0x199/0x200 do_syscall_64+0x5c/0x280 entry_SYSCALL_64_after_hwframe+0x49/0xbe RIP: 0033:0x7fe586c65760 Code: 00 f7 ... RSP: 002b:00007ffe250f98b8 EFLAGS: 00000246 ORIG_RAX: 000000000000001a RAX: ffffffffffffffda RBX: 00000000000040e1 RCX: 00007fe586c65760 RDX: 0000000000000004 RSI: 0000000000006b51 RDI: 00007fe58708b000 RBP: 0000000000006a70 R08: 0000000000000003 R09: 00007fe58700cb61 R10: 0000000000000100 R11: 0000000000000246 R12: 00000000000000e1 R13: 00007fe58708b000 R14: 0000000000006b51 R15: 0000558de021a420 Modules linked in: dm_log_writes ... ---[ end trace c92a7f447a8515f5 ]--- CC: stable@vger.kernel.org # 4.4+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-05-18 14:14:50 +03:00
unlock_extent_cached(&log_root->log_csum_range, sums->bytenr, lock_end,
&cached_state);
return ret;
Btrfs: fix missing data checksums after replaying a log tree When logging a file that has shared extents (reflinked with other files or with itself), we can end up logging multiple checksum items that cover overlapping ranges. This confuses the search for checksums at log replay time causing some checksums to never be added to the fs/subvolume tree. Consider the following example of a file that shares the same extent at offsets 0 and 256Kb: [ bytenr 13893632, offset 64Kb, len 64Kb ] 0 64Kb [ bytenr 13631488, offset 64Kb, len 192Kb ] 64Kb 256Kb [ bytenr 13893632, offset 0, len 256Kb ] 256Kb 512Kb When logging the inode, at tree-log.c:copy_items(), when processing the file extent item at offset 0, we log a checksum item covering the range 13959168 to 14024704, which corresponds to 13893632 + 64Kb and 13893632 + 64Kb + 64Kb, respectively. Later when processing the extent item at offset 256K, we log the checksums for the range from 13893632 to 14155776 (which corresponds to 13893632 + 256Kb). These checksums get merged with the checksum item for the range from 13631488 to 13893632 (13631488 + 256Kb), logged by a previous fsync. So after this we get the two following checksum items in the log tree: (...) item 6 key (EXTENT_CSUM EXTENT_CSUM 13631488) itemoff 3095 itemsize 512 range start 13631488 end 14155776 length 524288 item 7 key (EXTENT_CSUM EXTENT_CSUM 13959168) itemoff 3031 itemsize 64 range start 13959168 end 14024704 length 65536 The first one covers the range from the second one, they overlap. So far this does not cause a problem after replaying the log, because when replaying the file extent item for offset 256K, we copy all the checksums for the extent 13893632 from the log tree to the fs/subvolume tree, since searching for an checksum item for bytenr 13893632 leaves us at the first checksum item, which covers the whole range of the extent. However if we write 64Kb to file offset 256Kb for example, we will not be able to find and copy the checksums for the last 128Kb of the extent at bytenr 13893632, referenced by the file range 384Kb to 512Kb. After writing 64Kb into file offset 256Kb we get the following extent layout for our file: [ bytenr 13893632, offset 64K, len 64Kb ] 0 64Kb [ bytenr 13631488, offset 64Kb, len 192Kb ] 64Kb 256Kb [ bytenr 14155776, offset 0, len 64Kb ] 256Kb 320Kb [ bytenr 13893632, offset 64Kb, len 192Kb ] 320Kb 512Kb After fsync'ing the file, if we have a power failure and then mount the filesystem to replay the log, the following happens: 1) When replaying the file extent item for file offset 320Kb, we lookup for the checksums for the extent range from 13959168 (13893632 + 64Kb) to 14155776 (13893632 + 256Kb), through a call to btrfs_lookup_csums_range(); 2) btrfs_lookup_csums_range() finds the checksum item that starts precisely at offset 13959168 (item 7 in the log tree, shown before); 3) However that checksum item only covers 64Kb of data, and not 192Kb of data; 4) As a result only the checksums for the first 64Kb of data referenced by the file extent item are found and copied to the fs/subvolume tree. The remaining 128Kb of data, file range 384Kb to 512Kb, doesn't get the corresponding data checksums found and copied to the fs/subvolume tree. 5) After replaying the log userspace will not be able to read the file range from 384Kb to 512Kb, because the checksums are missing and resulting in an -EIO error. The following steps reproduce this scenario: $ mkfs.btrfs -f /dev/sdc $ mount /dev/sdc /mnt/sdc $ xfs_io -f -c "pwrite -S 0xa3 0 256K" /mnt/sdc/foobar $ xfs_io -c "fsync" /mnt/sdc/foobar $ xfs_io -c "pwrite -S 0xc7 256K 256K" /mnt/sdc/foobar $ xfs_io -c "reflink /mnt/sdc/foobar 320K 0 64K" /mnt/sdc/foobar $ xfs_io -c "fsync" /mnt/sdc/foobar $ xfs_io -c "pwrite -S 0xe5 256K 64K" /mnt/sdc/foobar $ xfs_io -c "fsync" /mnt/sdc/foobar <power failure> $ mount /dev/sdc /mnt/sdc $ md5sum /mnt/sdc/foobar md5sum: /mnt/sdc/foobar: Input/output error $ dmesg | tail [165305.003464] BTRFS info (device sdc): no csum found for inode 257 start 401408 [165305.004014] BTRFS info (device sdc): no csum found for inode 257 start 405504 [165305.004559] BTRFS info (device sdc): no csum found for inode 257 start 409600 [165305.005101] BTRFS info (device sdc): no csum found for inode 257 start 413696 [165305.005627] BTRFS info (device sdc): no csum found for inode 257 start 417792 [165305.006134] BTRFS info (device sdc): no csum found for inode 257 start 421888 [165305.006625] BTRFS info (device sdc): no csum found for inode 257 start 425984 [165305.007278] BTRFS info (device sdc): no csum found for inode 257 start 430080 [165305.008248] BTRFS warning (device sdc): csum failed root 5 ino 257 off 393216 csum 0x1337385e expected csum 0x00000000 mirror 1 [165305.009550] BTRFS warning (device sdc): csum failed root 5 ino 257 off 393216 csum 0x1337385e expected csum 0x00000000 mirror 1 Fix this simply by deleting first any checksums, from the log tree, for the range of the extent we are logging at copy_items(). This ensures we do not get checksum items in the log tree that have overlapping ranges. This is a long time issue that has been present since we have the clone (and deduplication) ioctl, and can happen both when an extent is shared between different files and within the same file. A test case for fstests follows soon. CC: stable@vger.kernel.org # 4.4+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-12-05 19:58:30 +03:00
}
static noinline int copy_items(struct btrfs_trans_handle *trans,
struct btrfs_inode *inode,
struct btrfs_path *dst_path,
Btrfs: fix missing hole after hole punching and fsync when using NO_HOLES When using the NO_HOLES feature, if we punch a hole into a file and then fsync it, there are cases where a subsequent fsync will miss the fact that a hole was punched, resulting in the holes not existing after replaying the log tree. Essentially these cases all imply that, tree-log.c:copy_items(), is not invoked for the leafs that delimit holes, because nothing changed those leafs in the current transaction. And it's precisely copy_items() where we currenly detect and log holes, which works as long as the holes are between file extent items in the input leaf or between the beginning of input leaf and the previous leaf or between the last item in the leaf and the next leaf. First example where we miss a hole: *) The extent items of the inode span multiple leafs; *) The punched hole covers a range that affects only the extent items of the first leaf; *) The fsync operation is done in full mode (BTRFS_INODE_NEEDS_FULL_SYNC is set in the inode's runtime flags). That results in the hole not existing after replaying the log tree. For example, if the fs/subvolume tree has the following layout for a particular inode: Leaf N, generation 10: [ ... INODE_ITEM INODE_REF EXTENT_ITEM (0 64K) EXTENT_ITEM (64K 128K) ] Leaf N + 1, generation 10: [ EXTENT_ITEM (128K 64K) ... ] If at transaction 11 we punch a hole coverting the range [0, 128K[, we end up dropping the two extent items from leaf N, but we don't touch the other leaf, so we end up in the following state: Leaf N, generation 11: [ ... INODE_ITEM INODE_REF ] Leaf N + 1, generation 10: [ EXTENT_ITEM (128K 64K) ... ] A full fsync after punching the hole will only process leaf N because it was modified in the current transaction, but not leaf N + 1, since it was not modified in the current transaction (generation 10 and not 11). As a result the fsync will not log any holes, because it didn't process any leaf with extent items. Second example where we will miss a hole: *) An inode as its items spanning 5 (or more) leafs; *) A hole is punched and it covers only the extents items of the 3rd leaf. This resulsts in deleting the entire leaf and not touching any of the other leafs. So the only leaf that is modified in the current transaction, when punching the hole, is the first leaf, which contains the inode item. During the full fsync, the only leaf that is passed to copy_items() is that first leaf, and that's not enough for the hole detection code in copy_items() to determine there's a hole between the last file extent item in the 2nd leaf and the first file extent item in the 3rd leaf (which was the 4th leaf before punching the hole). Fix this by scanning all leafs and punch holes as necessary when doing a full fsync (less common than a non-full fsync) when the NO_HOLES feature is enabled. The lack of explicit file extent items to mark holes makes it necessary to scan existing extents to determine if holes exist. A test case for fstests follows soon. Fixes: 16e7549f045d33 ("Btrfs: incompatible format change to remove hole extents") CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-11-19 15:07:33 +03:00
struct btrfs_path *src_path,
Btrfs: fix fsync data loss after adding hard link to inode We have a scenario where after the fsync log replay we can lose file data that had been previously fsync'ed if we added an hard link for our inode and after that we sync'ed the fsync log (for example by fsync'ing some other file or directory). This is because when adding an hard link we updated the inode item in the log tree with an i_size value of 0. At that point the new inode item was in memory only and a subsequent fsync log replay would not make us lose the file data. However if after adding the hard link we sync the log tree to disk, by fsync'ing some other file or directory for example, we ended up losing the file data after log replay, because the inode item in the persisted log tree had an an i_size of zero. This is easy to reproduce, and the following excerpt from my test for xfstests shows this: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create one file with data and fsync it. # This made the btrfs fsync log persist the data and the inode metadata with # a correct inode->i_size (4096 bytes). $XFS_IO_PROG -f -c "pwrite -S 0xaa -b 4K 0 4K" -c "fsync" \ $SCRATCH_MNT/foo | _filter_xfs_io # Now add one hard link to our file. This made the btrfs code update the fsync # log, in memory only, with an inode metadata having a size of 0. ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link # Now force persistence of the fsync log to disk, for example, by fsyncing some # other file. touch $SCRATCH_MNT/bar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/bar # Before a power loss or crash, we could read the 4Kb of data from our file as # expected. echo "File content before:" od -t x1 $SCRATCH_MNT/foo # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # After the fsync log replay, because the fsync log had a value of 0 for our # inode's i_size, we couldn't read anymore the 4Kb of data that we previously # wrote and fsync'ed. The size of the file became 0 after the fsync log replay. echo "File content after:" od -t x1 $SCRATCH_MNT/foo Another alternative test, that doesn't need to fsync an inode in the same transaction it was created, is: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create our test file with some data. $XFS_IO_PROG -f -c "pwrite -S 0xaa -b 8K 0 8K" \ $SCRATCH_MNT/foo | _filter_xfs_io # Make sure the file is durably persisted. sync # Append some data to our file, to increase its size. $XFS_IO_PROG -f -c "pwrite -S 0xcc -b 4K 8K 4K" \ $SCRATCH_MNT/foo | _filter_xfs_io # Fsync the file, so from this point on if a crash/power failure happens, our # new data is guaranteed to be there next time the fs is mounted. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foo # Add one hard link to our file. This made btrfs write into the in memory fsync # log a special inode with generation 0 and an i_size of 0 too. Note that this # didn't update the inode in the fsync log on disk. ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link # Now make sure the in memory fsync log is durably persisted. # Creating and fsync'ing another file will do it. touch $SCRATCH_MNT/bar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/bar # As expected, before the crash/power failure, we should be able to read the # 12Kb of file data. echo "File content before:" od -t x1 $SCRATCH_MNT/foo # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # After mounting the fs again, the fsync log was replayed. # The btrfs fsync log replay code didn't update the i_size of the persisted # inode because the inode item in the log had a special generation with a # value of 0 (and it couldn't know the correct i_size, since that inode item # had a 0 i_size too). This made the last 4Kb of file data inaccessible and # effectively lost. echo "File content after:" od -t x1 $SCRATCH_MNT/foo This isn't a new issue/regression. This problem has been around since the log tree code was added in 2008: Btrfs: Add a write ahead tree log to optimize synchronous operations (commit e02119d5a7b4396c5a872582fddc8bd6d305a70a) Test cases for xfstests follow soon. CC: <stable@vger.kernel.org> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-02-13 15:30:56 +03:00
int start_slot, int nr, int inode_only,
u64 logged_isize)
{
struct btrfs_fs_info *fs_info = trans->fs_info;
unsigned long src_offset;
unsigned long dst_offset;
struct btrfs_root *log = inode->root->log_root;
struct btrfs_file_extent_item *extent;
struct btrfs_inode_item *inode_item;
struct extent_buffer *src = src_path->nodes[0];
int ret;
struct btrfs_key *ins_keys;
u32 *ins_sizes;
btrfs: loop only once over data sizes array when inserting an item batch When inserting a batch of items into a btree, we end up looping over the data sizes array 3 times: 1) Once in the caller of btrfs_insert_empty_items(), when it populates the array with the data sizes for each item; 2) Once at btrfs_insert_empty_items() to sum the elements of the data sizes array and compute the total data size; 3) And then once again at setup_items_for_insert(), where we do exactly the same as what we do at btrfs_insert_empty_items(), to compute the total data size. That is not bad for small arrays, but when the arrays have hundreds of elements, the time spent on looping is not negligible. For example when doing batch inserts of delayed items for dir index items or when logging a directory, it's common to have 200 to 260 dir index items in a single batch when using a leaf size of 16K and using file names between 8 and 12 characters. For a 64K leaf size, multiply that by 4. Taking into account that during directory logging or when flushing delayed dir index items we can have many of those large batches, the time spent on the looping adds up quickly. It's also more important to avoid it at setup_items_for_insert(), since we are holding a write lock on a leaf and, in some cases, on upper nodes of the btree, which causes us to block other tasks that want to access the leaf and nodes for longer than necessary. So change the code so that setup_items_for_insert() and btrfs_insert_empty_items() no longer compute the total data size, and instead rely on the caller to supply it. This makes us loop over the array only once, where we can both populate the data size array and compute the total data size, taking advantage of spatial and temporal locality. To make this more manageable, use a structure to contain all the relevant details for a batch of items (keys array, data sizes array, total data size, number of items), and use it as an argument for btrfs_insert_empty_items() and setup_items_for_insert(). This patch is part of a small patchset that is comprised of the following patches: btrfs: loop only once over data sizes array when inserting an item batch btrfs: unexport setup_items_for_insert() btrfs: use single bulk copy operations when logging directories This is patch 1/3 and performance results, and the specific tests, are included in the changelog of patch 3/3. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-24 14:28:13 +03:00
struct btrfs_item_batch batch;
char *ins_data;
int i;
Btrfs: move data checksumming into a dedicated tree Btrfs stores checksums for each data block. Until now, they have been stored in the subvolume trees, indexed by the inode that is referencing the data block. This means that when we read the inode, we've probably read in at least some checksums as well. But, this has a few problems: * The checksums are indexed by logical offset in the file. When compression is on, this means we have to do the expensive checksumming on the uncompressed data. It would be faster if we could checksum the compressed data instead. * If we implement encryption, we'll be checksumming the plain text and storing that on disk. This is significantly less secure. * For either compression or encryption, we have to get the plain text back before we can verify the checksum as correct. This makes the raid layer balancing and extent moving much more expensive. * It makes the front end caching code more complex, as we have touch the subvolume and inodes as we cache extents. * There is potentitally one copy of the checksum in each subvolume referencing an extent. The solution used here is to store the extent checksums in a dedicated tree. This allows us to index the checksums by phyiscal extent start and length. It means: * The checksum is against the data stored on disk, after any compression or encryption is done. * The checksum is stored in a central location, and can be verified without following back references, or reading inodes. This makes compression significantly faster by reducing the amount of data that needs to be checksummed. It will also allow much faster raid management code in general. The checksums are indexed by a key with a fixed objectid (a magic value in ctree.h) and offset set to the starting byte of the extent. This allows us to copy the checksum items into the fsync log tree directly (or any other tree), without having to invent a second format for them. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-12-09 00:58:54 +03:00
struct list_head ordered_sums;
int skip_csum = inode->flags & BTRFS_INODE_NODATASUM;
Btrfs: move data checksumming into a dedicated tree Btrfs stores checksums for each data block. Until now, they have been stored in the subvolume trees, indexed by the inode that is referencing the data block. This means that when we read the inode, we've probably read in at least some checksums as well. But, this has a few problems: * The checksums are indexed by logical offset in the file. When compression is on, this means we have to do the expensive checksumming on the uncompressed data. It would be faster if we could checksum the compressed data instead. * If we implement encryption, we'll be checksumming the plain text and storing that on disk. This is significantly less secure. * For either compression or encryption, we have to get the plain text back before we can verify the checksum as correct. This makes the raid layer balancing and extent moving much more expensive. * It makes the front end caching code more complex, as we have touch the subvolume and inodes as we cache extents. * There is potentitally one copy of the checksum in each subvolume referencing an extent. The solution used here is to store the extent checksums in a dedicated tree. This allows us to index the checksums by phyiscal extent start and length. It means: * The checksum is against the data stored on disk, after any compression or encryption is done. * The checksum is stored in a central location, and can be verified without following back references, or reading inodes. This makes compression significantly faster by reducing the amount of data that needs to be checksummed. It will also allow much faster raid management code in general. The checksums are indexed by a key with a fixed objectid (a magic value in ctree.h) and offset set to the starting byte of the extent. This allows us to copy the checksum items into the fsync log tree directly (or any other tree), without having to invent a second format for them. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-12-09 00:58:54 +03:00
INIT_LIST_HEAD(&ordered_sums);
ins_data = kmalloc(nr * sizeof(struct btrfs_key) +
nr * sizeof(u32), GFP_NOFS);
if (!ins_data)
return -ENOMEM;
ins_sizes = (u32 *)ins_data;
ins_keys = (struct btrfs_key *)(ins_data + nr * sizeof(u32));
btrfs: loop only once over data sizes array when inserting an item batch When inserting a batch of items into a btree, we end up looping over the data sizes array 3 times: 1) Once in the caller of btrfs_insert_empty_items(), when it populates the array with the data sizes for each item; 2) Once at btrfs_insert_empty_items() to sum the elements of the data sizes array and compute the total data size; 3) And then once again at setup_items_for_insert(), where we do exactly the same as what we do at btrfs_insert_empty_items(), to compute the total data size. That is not bad for small arrays, but when the arrays have hundreds of elements, the time spent on looping is not negligible. For example when doing batch inserts of delayed items for dir index items or when logging a directory, it's common to have 200 to 260 dir index items in a single batch when using a leaf size of 16K and using file names between 8 and 12 characters. For a 64K leaf size, multiply that by 4. Taking into account that during directory logging or when flushing delayed dir index items we can have many of those large batches, the time spent on the looping adds up quickly. It's also more important to avoid it at setup_items_for_insert(), since we are holding a write lock on a leaf and, in some cases, on upper nodes of the btree, which causes us to block other tasks that want to access the leaf and nodes for longer than necessary. So change the code so that setup_items_for_insert() and btrfs_insert_empty_items() no longer compute the total data size, and instead rely on the caller to supply it. This makes us loop over the array only once, where we can both populate the data size array and compute the total data size, taking advantage of spatial and temporal locality. To make this more manageable, use a structure to contain all the relevant details for a batch of items (keys array, data sizes array, total data size, number of items), and use it as an argument for btrfs_insert_empty_items() and setup_items_for_insert(). This patch is part of a small patchset that is comprised of the following patches: btrfs: loop only once over data sizes array when inserting an item batch btrfs: unexport setup_items_for_insert() btrfs: use single bulk copy operations when logging directories This is patch 1/3 and performance results, and the specific tests, are included in the changelog of patch 3/3. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-24 14:28:13 +03:00
batch.keys = ins_keys;
batch.data_sizes = ins_sizes;
batch.total_data_size = 0;
batch.nr = nr;
for (i = 0; i < nr; i++) {
ins_sizes[i] = btrfs_item_size(src, i + start_slot);
btrfs: loop only once over data sizes array when inserting an item batch When inserting a batch of items into a btree, we end up looping over the data sizes array 3 times: 1) Once in the caller of btrfs_insert_empty_items(), when it populates the array with the data sizes for each item; 2) Once at btrfs_insert_empty_items() to sum the elements of the data sizes array and compute the total data size; 3) And then once again at setup_items_for_insert(), where we do exactly the same as what we do at btrfs_insert_empty_items(), to compute the total data size. That is not bad for small arrays, but when the arrays have hundreds of elements, the time spent on looping is not negligible. For example when doing batch inserts of delayed items for dir index items or when logging a directory, it's common to have 200 to 260 dir index items in a single batch when using a leaf size of 16K and using file names between 8 and 12 characters. For a 64K leaf size, multiply that by 4. Taking into account that during directory logging or when flushing delayed dir index items we can have many of those large batches, the time spent on the looping adds up quickly. It's also more important to avoid it at setup_items_for_insert(), since we are holding a write lock on a leaf and, in some cases, on upper nodes of the btree, which causes us to block other tasks that want to access the leaf and nodes for longer than necessary. So change the code so that setup_items_for_insert() and btrfs_insert_empty_items() no longer compute the total data size, and instead rely on the caller to supply it. This makes us loop over the array only once, where we can both populate the data size array and compute the total data size, taking advantage of spatial and temporal locality. To make this more manageable, use a structure to contain all the relevant details for a batch of items (keys array, data sizes array, total data size, number of items), and use it as an argument for btrfs_insert_empty_items() and setup_items_for_insert(). This patch is part of a small patchset that is comprised of the following patches: btrfs: loop only once over data sizes array when inserting an item batch btrfs: unexport setup_items_for_insert() btrfs: use single bulk copy operations when logging directories This is patch 1/3 and performance results, and the specific tests, are included in the changelog of patch 3/3. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-24 14:28:13 +03:00
batch.total_data_size += ins_sizes[i];
btrfs_item_key_to_cpu(src, ins_keys + i, i + start_slot);
}
btrfs: loop only once over data sizes array when inserting an item batch When inserting a batch of items into a btree, we end up looping over the data sizes array 3 times: 1) Once in the caller of btrfs_insert_empty_items(), when it populates the array with the data sizes for each item; 2) Once at btrfs_insert_empty_items() to sum the elements of the data sizes array and compute the total data size; 3) And then once again at setup_items_for_insert(), where we do exactly the same as what we do at btrfs_insert_empty_items(), to compute the total data size. That is not bad for small arrays, but when the arrays have hundreds of elements, the time spent on looping is not negligible. For example when doing batch inserts of delayed items for dir index items or when logging a directory, it's common to have 200 to 260 dir index items in a single batch when using a leaf size of 16K and using file names between 8 and 12 characters. For a 64K leaf size, multiply that by 4. Taking into account that during directory logging or when flushing delayed dir index items we can have many of those large batches, the time spent on the looping adds up quickly. It's also more important to avoid it at setup_items_for_insert(), since we are holding a write lock on a leaf and, in some cases, on upper nodes of the btree, which causes us to block other tasks that want to access the leaf and nodes for longer than necessary. So change the code so that setup_items_for_insert() and btrfs_insert_empty_items() no longer compute the total data size, and instead rely on the caller to supply it. This makes us loop over the array only once, where we can both populate the data size array and compute the total data size, taking advantage of spatial and temporal locality. To make this more manageable, use a structure to contain all the relevant details for a batch of items (keys array, data sizes array, total data size, number of items), and use it as an argument for btrfs_insert_empty_items() and setup_items_for_insert(). This patch is part of a small patchset that is comprised of the following patches: btrfs: loop only once over data sizes array when inserting an item batch btrfs: unexport setup_items_for_insert() btrfs: use single bulk copy operations when logging directories This is patch 1/3 and performance results, and the specific tests, are included in the changelog of patch 3/3. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-09-24 14:28:13 +03:00
ret = btrfs_insert_empty_items(trans, log, dst_path, &batch);
if (ret) {
kfree(ins_data);
return ret;
}
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 18:45:14 +04:00
for (i = 0; i < nr; i++, dst_path->slots[0]++) {
dst_offset = btrfs_item_ptr_offset(dst_path->nodes[0],
dst_path->slots[0]);
src_offset = btrfs_item_ptr_offset(src, start_slot + i);
if (ins_keys[i].type == BTRFS_INODE_ITEM_KEY) {
inode_item = btrfs_item_ptr(dst_path->nodes[0],
dst_path->slots[0],
struct btrfs_inode_item);
fill_inode_item(trans, dst_path->nodes[0], inode_item,
&inode->vfs_inode,
inode_only == LOG_INODE_EXISTS,
Btrfs: fix fsync data loss after adding hard link to inode We have a scenario where after the fsync log replay we can lose file data that had been previously fsync'ed if we added an hard link for our inode and after that we sync'ed the fsync log (for example by fsync'ing some other file or directory). This is because when adding an hard link we updated the inode item in the log tree with an i_size value of 0. At that point the new inode item was in memory only and a subsequent fsync log replay would not make us lose the file data. However if after adding the hard link we sync the log tree to disk, by fsync'ing some other file or directory for example, we ended up losing the file data after log replay, because the inode item in the persisted log tree had an an i_size of zero. This is easy to reproduce, and the following excerpt from my test for xfstests shows this: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create one file with data and fsync it. # This made the btrfs fsync log persist the data and the inode metadata with # a correct inode->i_size (4096 bytes). $XFS_IO_PROG -f -c "pwrite -S 0xaa -b 4K 0 4K" -c "fsync" \ $SCRATCH_MNT/foo | _filter_xfs_io # Now add one hard link to our file. This made the btrfs code update the fsync # log, in memory only, with an inode metadata having a size of 0. ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link # Now force persistence of the fsync log to disk, for example, by fsyncing some # other file. touch $SCRATCH_MNT/bar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/bar # Before a power loss or crash, we could read the 4Kb of data from our file as # expected. echo "File content before:" od -t x1 $SCRATCH_MNT/foo # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # After the fsync log replay, because the fsync log had a value of 0 for our # inode's i_size, we couldn't read anymore the 4Kb of data that we previously # wrote and fsync'ed. The size of the file became 0 after the fsync log replay. echo "File content after:" od -t x1 $SCRATCH_MNT/foo Another alternative test, that doesn't need to fsync an inode in the same transaction it was created, is: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create our test file with some data. $XFS_IO_PROG -f -c "pwrite -S 0xaa -b 8K 0 8K" \ $SCRATCH_MNT/foo | _filter_xfs_io # Make sure the file is durably persisted. sync # Append some data to our file, to increase its size. $XFS_IO_PROG -f -c "pwrite -S 0xcc -b 4K 8K 4K" \ $SCRATCH_MNT/foo | _filter_xfs_io # Fsync the file, so from this point on if a crash/power failure happens, our # new data is guaranteed to be there next time the fs is mounted. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foo # Add one hard link to our file. This made btrfs write into the in memory fsync # log a special inode with generation 0 and an i_size of 0 too. Note that this # didn't update the inode in the fsync log on disk. ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link # Now make sure the in memory fsync log is durably persisted. # Creating and fsync'ing another file will do it. touch $SCRATCH_MNT/bar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/bar # As expected, before the crash/power failure, we should be able to read the # 12Kb of file data. echo "File content before:" od -t x1 $SCRATCH_MNT/foo # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # After mounting the fs again, the fsync log was replayed. # The btrfs fsync log replay code didn't update the i_size of the persisted # inode because the inode item in the log had a special generation with a # value of 0 (and it couldn't know the correct i_size, since that inode item # had a 0 i_size too). This made the last 4Kb of file data inaccessible and # effectively lost. echo "File content after:" od -t x1 $SCRATCH_MNT/foo This isn't a new issue/regression. This problem has been around since the log tree code was added in 2008: Btrfs: Add a write ahead tree log to optimize synchronous operations (commit e02119d5a7b4396c5a872582fddc8bd6d305a70a) Test cases for xfstests follow soon. CC: <stable@vger.kernel.org> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-02-13 15:30:56 +03:00
logged_isize);
} else {
copy_extent_buffer(dst_path->nodes[0], src, dst_offset,
src_offset, ins_sizes[i]);
}
/* take a reference on file data extents so that truncates
* or deletes of this inode don't have to relog the inode
* again
*/
if (ins_keys[i].type == BTRFS_EXTENT_DATA_KEY &&
!skip_csum) {
int found_type;
extent = btrfs_item_ptr(src, start_slot + i,
struct btrfs_file_extent_item);
Btrfs: do not flush csum items of unchanged file data during treelog The current code relogs the entire inode every time during fsync log, and it is much better suited to small files rather than large ones. During my performance test, the fsync performace of large files sucks, and we can ascribe this to the tremendous amount of csum infos of the large ones, cause we have to flush all of these csum infos into log trees even when there are only _one_ change in the whole file data. Apparently, to optimize fsync, we need to create a filter to skip the unnecessary csum ones, that is, the corresponding file data remains unchanged before this fsync. Here I have some test results to show, I use sysbench to do "random write + fsync". === sysbench --test=fileio --num-threads=1 --file-num=2 --file-block-size=4K --file-total-size=8G --file-test-mode=rndwr --file-io-mode=sync --file-extra-flags= [prepare, run] === Sysbench args: - Number of threads: 1 - Extra file open flags: 0 - 2 files, 4Gb each - Block size 4Kb - Number of random requests for random IO: 10000 - Read/Write ratio for combined random IO test: 1.50 - Periodic FSYNC enabled, calling fsync() each 100 requests. - Calling fsync() at the end of test, Enabled. - Using synchronous I/O mode - Doing random write test Sysbench results: === Operations performed: 0 Read, 10000 Write, 200 Other = 10200 Total Read 0b Written 39.062Mb Total transferred 39.062Mb === a) without patch: (*SPEED* : 451.01Kb/sec) 112.75 Requests/sec executed b) with patch: (*SPEED* : 4.7533Mb/sec) 1216.84 Requests/sec executed PS: I've made a _sub transid_ stuff patch, but it does not perform as effectively as this patch, and I'm wanderring where the problem is and trying to improve it more. Signed-off-by: Liu Bo <liubo2009@cn.fujitsu.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2011-05-06 06:36:09 +04:00
if (btrfs_file_extent_generation(src, extent) < trans->transid)
continue;
found_type = btrfs_file_extent_type(src, extent);
if (found_type == BTRFS_FILE_EXTENT_REG) {
struct btrfs_root *csum_root;
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 18:45:14 +04:00
u64 ds, dl, cs, cl;
ds = btrfs_file_extent_disk_bytenr(src,
extent);
/* ds == 0 is a hole */
if (ds == 0)
continue;
dl = btrfs_file_extent_disk_num_bytes(src,
extent);
cs = btrfs_file_extent_offset(src, extent);
cl = btrfs_file_extent_num_bytes(src,
extent);
if (btrfs_file_extent_compression(src,
extent)) {
cs = 0;
cl = dl;
}
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 18:45:14 +04:00
csum_root = btrfs_csum_root(fs_info, ds);
ret = btrfs_lookup_csums_range(csum_root,
Btrfs: Mixed back reference (FORWARD ROLLING FORMAT CHANGE) This commit introduces a new kind of back reference for btrfs metadata. Once a filesystem has been mounted with this commit, IT WILL NO LONGER BE MOUNTABLE BY OLDER KERNELS. When a tree block in subvolume tree is cow'd, the reference counts of all extents it points to are increased by one. At transaction commit time, the old root of the subvolume is recorded in a "dead root" data structure, and the btree it points to is later walked, dropping reference counts and freeing any blocks where the reference count goes to 0. The increments done during cow and decrements done after commit cancel out, and the walk is a very expensive way to go about freeing the blocks that are no longer referenced by the new btree root. This commit reduces the transaction overhead by avoiding the need for dead root records. When a non-shared tree block is cow'd, we free the old block at once, and the new block inherits old block's references. When a tree block with reference count > 1 is cow'd, we increase the reference counts of all extents the new block points to by one, and decrease the old block's reference count by one. This dead tree avoidance code removes the need to modify the reference counts of lower level extents when a non-shared tree block is cow'd. But we still need to update back ref for all pointers in the block. This is because the location of the block is recorded in the back ref item. We can solve this by introducing a new type of back ref. The new back ref provides information about pointer's key, level and in which tree the pointer lives. This information allow us to find the pointer by searching the tree. The shortcoming of the new back ref is that it only works for pointers in tree blocks referenced by their owner trees. This is mostly a problem for snapshots, where resolving one of these fuzzy back references would be O(number_of_snapshots) and quite slow. The solution used here is to use the fuzzy back references in the common case where a given tree block is only referenced by one root, and use the full back references when multiple roots have a reference on a given block. This commit adds per subvolume red-black tree to keep trace of cached inodes. The red-black tree helps the balancing code to find cached inodes whose inode numbers within a given range. This commit improves the balancing code by introducing several data structures to keep the state of balancing. The most important one is the back ref cache. It caches how the upper level tree blocks are referenced. This greatly reduce the overhead of checking back ref. The improved balancing code scales significantly better with a large number of snapshots. This is a very large commit and was written in a number of pieces. But, they depend heavily on the disk format change and were squashed together to make sure git bisect didn't end up in a bad state wrt space balancing or the format change. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-06-10 18:45:14 +04:00
ds + cs, ds + cs + cl - 1,
&ordered_sums, 0);
if (ret)
break;
}
}
}
btrfs_mark_buffer_dirty(dst_path->nodes[0]);
btrfs_release_path(dst_path);
kfree(ins_data);
Btrfs: move data checksumming into a dedicated tree Btrfs stores checksums for each data block. Until now, they have been stored in the subvolume trees, indexed by the inode that is referencing the data block. This means that when we read the inode, we've probably read in at least some checksums as well. But, this has a few problems: * The checksums are indexed by logical offset in the file. When compression is on, this means we have to do the expensive checksumming on the uncompressed data. It would be faster if we could checksum the compressed data instead. * If we implement encryption, we'll be checksumming the plain text and storing that on disk. This is significantly less secure. * For either compression or encryption, we have to get the plain text back before we can verify the checksum as correct. This makes the raid layer balancing and extent moving much more expensive. * It makes the front end caching code more complex, as we have touch the subvolume and inodes as we cache extents. * There is potentitally one copy of the checksum in each subvolume referencing an extent. The solution used here is to store the extent checksums in a dedicated tree. This allows us to index the checksums by phyiscal extent start and length. It means: * The checksum is against the data stored on disk, after any compression or encryption is done. * The checksum is stored in a central location, and can be verified without following back references, or reading inodes. This makes compression significantly faster by reducing the amount of data that needs to be checksummed. It will also allow much faster raid management code in general. The checksums are indexed by a key with a fixed objectid (a magic value in ctree.h) and offset set to the starting byte of the extent. This allows us to copy the checksum items into the fsync log tree directly (or any other tree), without having to invent a second format for them. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-12-09 00:58:54 +03:00
/*
* we have to do this after the loop above to avoid changing the
* log tree while trying to change the log tree.
*/
while (!list_empty(&ordered_sums)) {
Btrfs: move data checksumming into a dedicated tree Btrfs stores checksums for each data block. Until now, they have been stored in the subvolume trees, indexed by the inode that is referencing the data block. This means that when we read the inode, we've probably read in at least some checksums as well. But, this has a few problems: * The checksums are indexed by logical offset in the file. When compression is on, this means we have to do the expensive checksumming on the uncompressed data. It would be faster if we could checksum the compressed data instead. * If we implement encryption, we'll be checksumming the plain text and storing that on disk. This is significantly less secure. * For either compression or encryption, we have to get the plain text back before we can verify the checksum as correct. This makes the raid layer balancing and extent moving much more expensive. * It makes the front end caching code more complex, as we have touch the subvolume and inodes as we cache extents. * There is potentitally one copy of the checksum in each subvolume referencing an extent. The solution used here is to store the extent checksums in a dedicated tree. This allows us to index the checksums by phyiscal extent start and length. It means: * The checksum is against the data stored on disk, after any compression or encryption is done. * The checksum is stored in a central location, and can be verified without following back references, or reading inodes. This makes compression significantly faster by reducing the amount of data that needs to be checksummed. It will also allow much faster raid management code in general. The checksums are indexed by a key with a fixed objectid (a magic value in ctree.h) and offset set to the starting byte of the extent. This allows us to copy the checksum items into the fsync log tree directly (or any other tree), without having to invent a second format for them. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-12-09 00:58:54 +03:00
struct btrfs_ordered_sum *sums = list_entry(ordered_sums.next,
struct btrfs_ordered_sum,
list);
if (!ret)
btrfs: reduce contention on log trees when logging checksums The possibility of extents being shared (through clone and deduplication operations) requires special care when logging data checksums, to avoid having a log tree with different checksum items that cover ranges which overlap (which resulted in missing checksums after replaying a log tree). Such problems were fixed in the past by the following commits: commit 40e046acbd2f ("Btrfs: fix missing data checksums after replaying a log tree") commit e289f03ea79b ("btrfs: fix corrupt log due to concurrent fsync of inodes with shared extents") Test case generic/588 exercises the scenario solved by the first commit (purely sequential and deterministic) while test case generic/457 often triggered the case fixed by the second commit (not deterministic, requires specific timings under concurrency). The problems were addressed by deleting, from the log tree, any existing checksums before logging the new ones. And also by doing the deletion and logging of the cheksums while locking the checksum range in an extent io tree (root->log_csum_range), to deal with the case where we have concurrent fsyncs against files with shared extents. That however causes more contention on the leaves of a log tree where we store checksums (and all the nodes in the paths leading to them), even when we do not have shared extents, or all the shared extents were created by past transactions. It also adds a bit of contention on the spin lock of the log_csums_range extent io tree of the log root. This change adds a 'last_reflink_trans' field to the inode to keep track of the last transaction where a new extent was shared between inodes (through clone and deduplication operations). It is updated for both the source and destination inodes of reflink operations whenever a new extent (created in the current transaction) becomes shared by the inodes. This field is kept in memory only, not persisted in the inode item, similar to other existing fields (last_unlink_trans, logged_trans). When logging checksums for an extent, if the value of 'last_reflink_trans' is smaller then the current transaction's generation/id, we skip locking the extent range and deletion of checksums from the log tree, since we know we do not have new shared extents. This reduces contention on the log tree's leaves where checksums are stored. The following script, which uses fio, was used to measure the impact of this change: $ cat test-fsync.sh #!/bin/bash DEV=/dev/sdk MNT=/mnt/sdk MOUNT_OPTIONS="-o ssd" MKFS_OPTIONS="-d single -m single" if [ $# -ne 3 ]; then echo "Use $0 NUM_JOBS FILE_SIZE FSYNC_FREQ" exit 1 fi NUM_JOBS=$1 FILE_SIZE=$2 FSYNC_FREQ=$3 cat <<EOF > /tmp/fio-job.ini [writers] rw=write fsync=$FSYNC_FREQ fallocate=none group_reporting=1 direct=0 bs=64k ioengine=sync size=$FILE_SIZE directory=$MNT numjobs=$NUM_JOBS EOF echo "Using config:" echo cat /tmp/fio-job.ini echo mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT fio /tmp/fio-job.ini umount $MNT The tests were performed for different numbers of jobs, file sizes and fsync frequency. A qemu VM using kvm was used, with 8 cores (the host has 12 cores, with cpu governance set to performance mode on all cores), 16GiB of ram (the host has 64GiB) and using a NVMe device directly (without an intermediary filesystem in the host). While running the tests, the host was not used for anything else, to avoid disturbing the tests. The obtained results were the following (the last line of fio's output was pasted). Starting with 16 jobs is where a significant difference is observable in this particular setup and hardware (differences highlighted below). The very small differences for tests with less than 16 jobs are possibly just noise and random. **** 1 job, file size 1G, fsync frequency 1 **** before this change: WRITE: bw=23.8MiB/s (24.9MB/s), 23.8MiB/s-23.8MiB/s (24.9MB/s-24.9MB/s), io=1024MiB (1074MB), run=43075-43075msec after this change: WRITE: bw=24.4MiB/s (25.6MB/s), 24.4MiB/s-24.4MiB/s (25.6MB/s-25.6MB/s), io=1024MiB (1074MB), run=41938-41938msec **** 2 jobs, file size 1G, fsync frequency 1 **** before this change: WRITE: bw=37.7MiB/s (39.5MB/s), 37.7MiB/s-37.7MiB/s (39.5MB/s-39.5MB/s), io=2048MiB (2147MB), run=54351-54351msec after this change: WRITE: bw=37.7MiB/s (39.5MB/s), 37.6MiB/s-37.6MiB/s (39.5MB/s-39.5MB/s), io=2048MiB (2147MB), run=54428-54428msec **** 4 jobs, file size 1G, fsync frequency 1 **** before this change: WRITE: bw=67.5MiB/s (70.8MB/s), 67.5MiB/s-67.5MiB/s (70.8MB/s-70.8MB/s), io=4096MiB (4295MB), run=60669-60669msec after this change: WRITE: bw=68.6MiB/s (71.0MB/s), 68.6MiB/s-68.6MiB/s (71.0MB/s-71.0MB/s), io=4096MiB (4295MB), run=59678-59678msec **** 8 jobs, file size 1G, fsync frequency 1 **** before this change: WRITE: bw=128MiB/s (134MB/s), 128MiB/s-128MiB/s (134MB/s-134MB/s), io=8192MiB (8590MB), run=64048-64048msec after this change: WRITE: bw=129MiB/s (135MB/s), 129MiB/s-129MiB/s (135MB/s-135MB/s), io=8192MiB (8590MB), run=63405-63405msec **** 16 jobs, file size 1G, fsync frequency 1 **** before this change: WRITE: bw=78.5MiB/s (82.3MB/s), 78.5MiB/s-78.5MiB/s (82.3MB/s-82.3MB/s), io=16.0GiB (17.2GB), run=208676-208676msec after this change: WRITE: bw=110MiB/s (115MB/s), 110MiB/s-110MiB/s (115MB/s-115MB/s), io=16.0GiB (17.2GB), run=149295-149295msec (+40.1% throughput, -28.5% runtime) **** 32 jobs, file size 1G, fsync frequency 1 **** before this change: WRITE: bw=58.8MiB/s (61.7MB/s), 58.8MiB/s-58.8MiB/s (61.7MB/s-61.7MB/s), io=32.0GiB (34.4GB), run=557134-557134msec after this change: WRITE: bw=76.1MiB/s (79.8MB/s), 76.1MiB/s-76.1MiB/s (79.8MB/s-79.8MB/s), io=32.0GiB (34.4GB), run=430550-430550msec (+29.4% throughput, -22.7% runtime) **** 64 jobs, file size 512M, fsync frequency 1 **** before this change: WRITE: bw=65.8MiB/s (68.0MB/s), 65.8MiB/s-65.8MiB/s (68.0MB/s-68.0MB/s), io=32.0GiB (34.4GB), run=498055-498055msec after this change: WRITE: bw=85.1MiB/s (89.2MB/s), 85.1MiB/s-85.1MiB/s (89.2MB/s-89.2MB/s), io=32.0GiB (34.4GB), run=385116-385116msec (+29.3% throughput, -22.7% runtime) **** 128 jobs, file size 256M, fsync frequency 1 **** before this change: WRITE: bw=54.7MiB/s (57.3MB/s), 54.7MiB/s-54.7MiB/s (57.3MB/s-57.3MB/s), io=32.0GiB (34.4GB), run=599373-599373msec after this change: WRITE: bw=121MiB/s (126MB/s), 121MiB/s-121MiB/s (126MB/s-126MB/s), io=32.0GiB (34.4GB), run=271907-271907msec (+121.2% throughput, -54.6% runtime) **** 256 jobs, file size 256M, fsync frequency 1 **** before this change: WRITE: bw=69.2MiB/s (72.5MB/s), 69.2MiB/s-69.2MiB/s (72.5MB/s-72.5MB/s), io=64.0GiB (68.7GB), run=947536-947536msec after this change: WRITE: bw=121MiB/s (127MB/s), 121MiB/s-121MiB/s (127MB/s-127MB/s), io=64.0GiB (68.7GB), run=541916-541916msec (+74.9% throughput, -42.8% runtime) **** 512 jobs, file size 128M, fsync frequency 1 **** before this change: WRITE: bw=85.4MiB/s (89.5MB/s), 85.4MiB/s-85.4MiB/s (89.5MB/s-89.5MB/s), io=64.0GiB (68.7GB), run=767734-767734msec after this change: WRITE: bw=141MiB/s (147MB/s), 141MiB/s-141MiB/s (147MB/s-147MB/s), io=64.0GiB (68.7GB), run=466022-466022msec (+65.1% throughput, -39.3% runtime) **** 1024 jobs, file size 128M, fsync frequency 1 **** before this change: WRITE: bw=115MiB/s (120MB/s), 115MiB/s-115MiB/s (120MB/s-120MB/s), io=128GiB (137GB), run=1143775-1143775msec after this change: WRITE: bw=171MiB/s (180MB/s), 171MiB/s-171MiB/s (180MB/s-180MB/s), io=128GiB (137GB), run=764843-764843msec (+48.7% throughput, -33.1% runtime) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-07-15 14:30:43 +03:00
ret = log_csums(trans, inode, log, sums);
Btrfs: move data checksumming into a dedicated tree Btrfs stores checksums for each data block. Until now, they have been stored in the subvolume trees, indexed by the inode that is referencing the data block. This means that when we read the inode, we've probably read in at least some checksums as well. But, this has a few problems: * The checksums are indexed by logical offset in the file. When compression is on, this means we have to do the expensive checksumming on the uncompressed data. It would be faster if we could checksum the compressed data instead. * If we implement encryption, we'll be checksumming the plain text and storing that on disk. This is significantly less secure. * For either compression or encryption, we have to get the plain text back before we can verify the checksum as correct. This makes the raid layer balancing and extent moving much more expensive. * It makes the front end caching code more complex, as we have touch the subvolume and inodes as we cache extents. * There is potentitally one copy of the checksum in each subvolume referencing an extent. The solution used here is to store the extent checksums in a dedicated tree. This allows us to index the checksums by phyiscal extent start and length. It means: * The checksum is against the data stored on disk, after any compression or encryption is done. * The checksum is stored in a central location, and can be verified without following back references, or reading inodes. This makes compression significantly faster by reducing the amount of data that needs to be checksummed. It will also allow much faster raid management code in general. The checksums are indexed by a key with a fixed objectid (a magic value in ctree.h) and offset set to the starting byte of the extent. This allows us to copy the checksum items into the fsync log tree directly (or any other tree), without having to invent a second format for them. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2008-12-09 00:58:54 +03:00
list_del(&sums->list);
kfree(sums);
}
return ret;
}
static int extent_cmp(void *priv, const struct list_head *a,
const struct list_head *b)
Btrfs: turbo charge fsync At least for the vm workload. Currently on fsync we will 1) Truncate all items in the log tree for the given inode if they exist and 2) Copy all items for a given inode into the log The problem with this is that for things like VMs you can have lots of extents from the fragmented writing behavior, and worst yet you may have only modified a few extents, not the entire thing. This patch fixes this problem by tracking which transid modified our extent, and then when we do the tree logging we find all of the extents we've modified in our current transaction, sort them and commit them. We also only truncate up to the xattrs of the inode and copy that stuff in normally, and then just drop any extents in the range we have that exist in the log already. Here are some numbers of a 50 meg fio job that does random writes and fsync()s after every write Original Patched SATA drive 82KB/s 140KB/s Fusion drive 431KB/s 2532KB/s So around 2-6 times faster depending on your hardware. There are a few corner cases, for example if you truncate at all we have to do it the old way since there is no way to be sure what is in the log is ok. This probably could be done smarter, but if you write-fsync-truncate-write-fsync you deserve what you get. All this work is in RAM of course so if your inode gets evicted from cache and you read it in and fsync it we'll do it the slow way if we are still in the same transaction that we last modified the inode in. The biggest cool part of this is that it requires no changes to the recovery code, so if you fsync with this patch and crash and load an old kernel, it will run the recovery and be a-ok. I have tested this pretty thoroughly with an fsync tester and everything comes back fine, as well as xfstests. Thanks, Signed-off-by: Josef Bacik <jbacik@fusionio.com>
2012-08-17 21:14:17 +04:00
{
const struct extent_map *em1, *em2;
Btrfs: turbo charge fsync At least for the vm workload. Currently on fsync we will 1) Truncate all items in the log tree for the given inode if they exist and 2) Copy all items for a given inode into the log The problem with this is that for things like VMs you can have lots of extents from the fragmented writing behavior, and worst yet you may have only modified a few extents, not the entire thing. This patch fixes this problem by tracking which transid modified our extent, and then when we do the tree logging we find all of the extents we've modified in our current transaction, sort them and commit them. We also only truncate up to the xattrs of the inode and copy that stuff in normally, and then just drop any extents in the range we have that exist in the log already. Here are some numbers of a 50 meg fio job that does random writes and fsync()s after every write Original Patched SATA drive 82KB/s 140KB/s Fusion drive 431KB/s 2532KB/s So around 2-6 times faster depending on your hardware. There are a few corner cases, for example if you truncate at all we have to do it the old way since there is no way to be sure what is in the log is ok. This probably could be done smarter, but if you write-fsync-truncate-write-fsync you deserve what you get. All this work is in RAM of course so if your inode gets evicted from cache and you read it in and fsync it we'll do it the slow way if we are still in the same transaction that we last modified the inode in. The biggest cool part of this is that it requires no changes to the recovery code, so if you fsync with this patch and crash and load an old kernel, it will run the recovery and be a-ok. I have tested this pretty thoroughly with an fsync tester and everything comes back fine, as well as xfstests. Thanks, Signed-off-by: Josef Bacik <jbacik@fusionio.com>
2012-08-17 21:14:17 +04:00
em1 = list_entry(a, struct extent_map, list);
em2 = list_entry(b, struct extent_map, list);
if (em1->start < em2->start)
return -1;
else if (em1->start > em2->start)
return 1;
return 0;
}
static int log_extent_csums(struct btrfs_trans_handle *trans,
struct btrfs_inode *inode,
struct btrfs_root *log_root,
btrfs: make fast fsyncs wait only for writeback Currently regardless of a full or a fast fsync we always wait for ordered extents to complete, and then start logging the inode after that. However for fast fsyncs we can just wait for the writeback to complete, we don't need to wait for the ordered extents to complete since we use the list of modified extents maps to figure out which extents we must log and we can get their checksums directly from the ordered extents that are still in flight, otherwise look them up from the checksums tree. Until commit b5e6c3e170b770 ("btrfs: always wait on ordered extents at fsync time"), for fast fsyncs, we used to start logging without even waiting for the writeback to complete first, we would wait for it to complete after logging, while holding a transaction open, which lead to performance issues when using cgroups and probably for other cases too, as wait for IO while holding a transaction handle should be avoided as much as possible. After that, for fast fsyncs, we started to wait for ordered extents to complete before starting to log, which adds some latency to fsyncs and we even got at least one report about a performance drop which bisected to that particular change: https://lore.kernel.org/linux-btrfs/20181109215148.GF23260@techsingularity.net/ This change makes fast fsyncs only wait for writeback to finish before starting to log the inode, instead of waiting for both the writeback to finish and for the ordered extents to complete. This brings back part of the logic we had that extracts checksums from in flight ordered extents, which are not yet in the checksums tree, and making sure transaction commits wait for the completion of ordered extents previously logged (by far most of the time they have already completed by the time a transaction commit starts, resulting in no wait at all), to avoid any data loss if an ordered extent completes after the transaction used to log an inode is committed, followed by a power failure. When there are no other tasks accessing the checksums and the subvolume btrees, the ordered extent completion is pretty fast, typically taking 100 to 200 microseconds only in my observations. However when there are other tasks accessing these btrees, ordered extent completion can take a lot more time due to lock contention on nodes and leaves of these btrees. I've seen cases over 2 milliseconds, which starts to be significant. In particular when we do have concurrent fsyncs against different files there is a lot of contention on the checksums btree, since we have many tasks writing the checksums into the btree and other tasks that already started the logging phase are doing lookups for checksums in the btree. This change also turns all ranged fsyncs into full ranged fsyncs, which is something we already did when not using the NO_HOLES features or when doing a full fsync. This is to guarantee we never miss checksums due to writeback having been triggered only for a part of an extent, and we end up logging the full extent but only checksums for the written range, which results in missing checksums after log replay. Allowing ranged fsyncs to operate again only in the original range, when using the NO_HOLES feature and doing a fast fsync is doable but requires some non trivial changes to the writeback path, which can always be worked on later if needed, but I don't think they are a very common use case. Several tests were performed using fio for different numbers of concurrent jobs, each writing and fsyncing its own file, for both sequential and random file writes. The tests were run on bare metal, no virtualization, on a box with 12 cores (Intel i7-8700), 64Gb of RAM and a NVMe device, with a kernel configuration that is the default of typical distributions (debian in this case), without debug options enabled (kasan, kmemleak, slub debug, debug of page allocations, lock debugging, etc). The following script that calls fio was used: $ cat test-fsync.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/btrfs MOUNT_OPTIONS="-o ssd -o space_cache=v2" MKFS_OPTIONS="-d single -m single" if [ $# -ne 5 ]; then echo "Use $0 NUM_JOBS FILE_SIZE FSYNC_FREQ BLOCK_SIZE [write|randwrite]" exit 1 fi NUM_JOBS=$1 FILE_SIZE=$2 FSYNC_FREQ=$3 BLOCK_SIZE=$4 WRITE_MODE=$5 if [ "$WRITE_MODE" != "write" ] && [ "$WRITE_MODE" != "randwrite" ]; then echo "Invalid WRITE_MODE, must be 'write' or 'randwrite'" exit 1 fi cat <<EOF > /tmp/fio-job.ini [writers] rw=$WRITE_MODE fsync=$FSYNC_FREQ fallocate=none group_reporting=1 direct=0 bs=$BLOCK_SIZE ioengine=sync size=$FILE_SIZE directory=$MNT numjobs=$NUM_JOBS EOF echo "performance" | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor echo echo "Using config:" echo cat /tmp/fio-job.ini echo umount $MNT &> /dev/null mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT fio /tmp/fio-job.ini umount $MNT The results were the following: ************************* *** sequential writes *** ************************* ==== 1 job, 8GiB file, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=36.6MiB/s (38.4MB/s), 36.6MiB/s-36.6MiB/s (38.4MB/s-38.4MB/s), io=8192MiB (8590MB), run=223689-223689msec After patch: WRITE: bw=40.2MiB/s (42.1MB/s), 40.2MiB/s-40.2MiB/s (42.1MB/s-42.1MB/s), io=8192MiB (8590MB), run=203980-203980msec (+9.8%, -8.8% runtime) ==== 2 jobs, 4GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=35.8MiB/s (37.5MB/s), 35.8MiB/s-35.8MiB/s (37.5MB/s-37.5MB/s), io=8192MiB (8590MB), run=228950-228950msec After patch: WRITE: bw=43.5MiB/s (45.6MB/s), 43.5MiB/s-43.5MiB/s (45.6MB/s-45.6MB/s), io=8192MiB (8590MB), run=188272-188272msec (+21.5% throughput, -17.8% runtime) ==== 4 jobs, 2GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=50.1MiB/s (52.6MB/s), 50.1MiB/s-50.1MiB/s (52.6MB/s-52.6MB/s), io=8192MiB (8590MB), run=163446-163446msec After patch: WRITE: bw=64.5MiB/s (67.6MB/s), 64.5MiB/s-64.5MiB/s (67.6MB/s-67.6MB/s), io=8192MiB (8590MB), run=126987-126987msec (+28.7% throughput, -22.3% runtime) ==== 8 jobs, 1GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=64.0MiB/s (68.1MB/s), 64.0MiB/s-64.0MiB/s (68.1MB/s-68.1MB/s), io=8192MiB (8590MB), run=126075-126075msec After patch: WRITE: bw=86.8MiB/s (91.0MB/s), 86.8MiB/s-86.8MiB/s (91.0MB/s-91.0MB/s), io=8192MiB (8590MB), run=94358-94358msec (+35.6% throughput, -25.2% runtime) ==== 16 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=79.8MiB/s (83.6MB/s), 79.8MiB/s-79.8MiB/s (83.6MB/s-83.6MB/s), io=8192MiB (8590MB), run=102694-102694msec After patch: WRITE: bw=107MiB/s (112MB/s), 107MiB/s-107MiB/s (112MB/s-112MB/s), io=8192MiB (8590MB), run=76446-76446msec (+34.1% throughput, -25.6% runtime) ==== 32 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=93.2MiB/s (97.7MB/s), 93.2MiB/s-93.2MiB/s (97.7MB/s-97.7MB/s), io=16.0GiB (17.2GB), run=175836-175836msec After patch: WRITE: bw=111MiB/s (117MB/s), 111MiB/s-111MiB/s (117MB/s-117MB/s), io=16.0GiB (17.2GB), run=147001-147001msec (+19.1% throughput, -16.4% runtime) ==== 64 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=108MiB/s (114MB/s), 108MiB/s-108MiB/s (114MB/s-114MB/s), io=32.0GiB (34.4GB), run=302656-302656msec After patch: WRITE: bw=133MiB/s (140MB/s), 133MiB/s-133MiB/s (140MB/s-140MB/s), io=32.0GiB (34.4GB), run=246003-246003msec (+23.1% throughput, -18.7% runtime) ************************ *** random writes *** ************************ ==== 1 job, 8GiB file, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=11.5MiB/s (12.0MB/s), 11.5MiB/s-11.5MiB/s (12.0MB/s-12.0MB/s), io=8192MiB (8590MB), run=714281-714281msec After patch: WRITE: bw=11.6MiB/s (12.2MB/s), 11.6MiB/s-11.6MiB/s (12.2MB/s-12.2MB/s), io=8192MiB (8590MB), run=705959-705959msec (+0.9% throughput, -1.7% runtime) ==== 2 jobs, 4GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=12.8MiB/s (13.5MB/s), 12.8MiB/s-12.8MiB/s (13.5MB/s-13.5MB/s), io=8192MiB (8590MB), run=638101-638101msec After patch: WRITE: bw=13.1MiB/s (13.7MB/s), 13.1MiB/s-13.1MiB/s (13.7MB/s-13.7MB/s), io=8192MiB (8590MB), run=625374-625374msec (+2.3% throughput, -2.0% runtime) ==== 4 jobs, 2GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=15.4MiB/s (16.2MB/s), 15.4MiB/s-15.4MiB/s (16.2MB/s-16.2MB/s), io=8192MiB (8590MB), run=531146-531146msec After patch: WRITE: bw=17.8MiB/s (18.7MB/s), 17.8MiB/s-17.8MiB/s (18.7MB/s-18.7MB/s), io=8192MiB (8590MB), run=460431-460431msec (+15.6% throughput, -13.3% runtime) ==== 8 jobs, 1GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=19.9MiB/s (20.8MB/s), 19.9MiB/s-19.9MiB/s (20.8MB/s-20.8MB/s), io=8192MiB (8590MB), run=412664-412664msec After patch: WRITE: bw=22.2MiB/s (23.3MB/s), 22.2MiB/s-22.2MiB/s (23.3MB/s-23.3MB/s), io=8192MiB (8590MB), run=368589-368589msec (+11.6% throughput, -10.7% runtime) ==== 16 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=29.3MiB/s (30.7MB/s), 29.3MiB/s-29.3MiB/s (30.7MB/s-30.7MB/s), io=8192MiB (8590MB), run=279924-279924msec After patch: WRITE: bw=30.4MiB/s (31.9MB/s), 30.4MiB/s-30.4MiB/s (31.9MB/s-31.9MB/s), io=8192MiB (8590MB), run=269258-269258msec (+3.8% throughput, -3.8% runtime) ==== 32 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=36.9MiB/s (38.7MB/s), 36.9MiB/s-36.9MiB/s (38.7MB/s-38.7MB/s), io=16.0GiB (17.2GB), run=443581-443581msec After patch: WRITE: bw=41.6MiB/s (43.6MB/s), 41.6MiB/s-41.6MiB/s (43.6MB/s-43.6MB/s), io=16.0GiB (17.2GB), run=394114-394114msec (+12.7% throughput, -11.2% runtime) ==== 64 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=45.9MiB/s (48.1MB/s), 45.9MiB/s-45.9MiB/s (48.1MB/s-48.1MB/s), io=32.0GiB (34.4GB), run=714614-714614msec After patch: WRITE: bw=48.8MiB/s (51.1MB/s), 48.8MiB/s-48.8MiB/s (51.1MB/s-51.1MB/s), io=32.0GiB (34.4GB), run=672087-672087msec (+6.3% throughput, -6.0% runtime) Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-08-11 14:43:58 +03:00
const struct extent_map *em,
struct btrfs_log_ctx *ctx)
Btrfs: turbo charge fsync At least for the vm workload. Currently on fsync we will 1) Truncate all items in the log tree for the given inode if they exist and 2) Copy all items for a given inode into the log The problem with this is that for things like VMs you can have lots of extents from the fragmented writing behavior, and worst yet you may have only modified a few extents, not the entire thing. This patch fixes this problem by tracking which transid modified our extent, and then when we do the tree logging we find all of the extents we've modified in our current transaction, sort them and commit them. We also only truncate up to the xattrs of the inode and copy that stuff in normally, and then just drop any extents in the range we have that exist in the log already. Here are some numbers of a 50 meg fio job that does random writes and fsync()s after every write Original Patched SATA drive 82KB/s 140KB/s Fusion drive 431KB/s 2532KB/s So around 2-6 times faster depending on your hardware. There are a few corner cases, for example if you truncate at all we have to do it the old way since there is no way to be sure what is in the log is ok. This probably could be done smarter, but if you write-fsync-truncate-write-fsync you deserve what you get. All this work is in RAM of course so if your inode gets evicted from cache and you read it in and fsync it we'll do it the slow way if we are still in the same transaction that we last modified the inode in. The biggest cool part of this is that it requires no changes to the recovery code, so if you fsync with this patch and crash and load an old kernel, it will run the recovery and be a-ok. I have tested this pretty thoroughly with an fsync tester and everything comes back fine, as well as xfstests. Thanks, Signed-off-by: Josef Bacik <jbacik@fusionio.com>
2012-08-17 21:14:17 +04:00
{
btrfs: make fast fsyncs wait only for writeback Currently regardless of a full or a fast fsync we always wait for ordered extents to complete, and then start logging the inode after that. However for fast fsyncs we can just wait for the writeback to complete, we don't need to wait for the ordered extents to complete since we use the list of modified extents maps to figure out which extents we must log and we can get their checksums directly from the ordered extents that are still in flight, otherwise look them up from the checksums tree. Until commit b5e6c3e170b770 ("btrfs: always wait on ordered extents at fsync time"), for fast fsyncs, we used to start logging without even waiting for the writeback to complete first, we would wait for it to complete after logging, while holding a transaction open, which lead to performance issues when using cgroups and probably for other cases too, as wait for IO while holding a transaction handle should be avoided as much as possible. After that, for fast fsyncs, we started to wait for ordered extents to complete before starting to log, which adds some latency to fsyncs and we even got at least one report about a performance drop which bisected to that particular change: https://lore.kernel.org/linux-btrfs/20181109215148.GF23260@techsingularity.net/ This change makes fast fsyncs only wait for writeback to finish before starting to log the inode, instead of waiting for both the writeback to finish and for the ordered extents to complete. This brings back part of the logic we had that extracts checksums from in flight ordered extents, which are not yet in the checksums tree, and making sure transaction commits wait for the completion of ordered extents previously logged (by far most of the time they have already completed by the time a transaction commit starts, resulting in no wait at all), to avoid any data loss if an ordered extent completes after the transaction used to log an inode is committed, followed by a power failure. When there are no other tasks accessing the checksums and the subvolume btrees, the ordered extent completion is pretty fast, typically taking 100 to 200 microseconds only in my observations. However when there are other tasks accessing these btrees, ordered extent completion can take a lot more time due to lock contention on nodes and leaves of these btrees. I've seen cases over 2 milliseconds, which starts to be significant. In particular when we do have concurrent fsyncs against different files there is a lot of contention on the checksums btree, since we have many tasks writing the checksums into the btree and other tasks that already started the logging phase are doing lookups for checksums in the btree. This change also turns all ranged fsyncs into full ranged fsyncs, which is something we already did when not using the NO_HOLES features or when doing a full fsync. This is to guarantee we never miss checksums due to writeback having been triggered only for a part of an extent, and we end up logging the full extent but only checksums for the written range, which results in missing checksums after log replay. Allowing ranged fsyncs to operate again only in the original range, when using the NO_HOLES feature and doing a fast fsync is doable but requires some non trivial changes to the writeback path, which can always be worked on later if needed, but I don't think they are a very common use case. Several tests were performed using fio for different numbers of concurrent jobs, each writing and fsyncing its own file, for both sequential and random file writes. The tests were run on bare metal, no virtualization, on a box with 12 cores (Intel i7-8700), 64Gb of RAM and a NVMe device, with a kernel configuration that is the default of typical distributions (debian in this case), without debug options enabled (kasan, kmemleak, slub debug, debug of page allocations, lock debugging, etc). The following script that calls fio was used: $ cat test-fsync.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/btrfs MOUNT_OPTIONS="-o ssd -o space_cache=v2" MKFS_OPTIONS="-d single -m single" if [ $# -ne 5 ]; then echo "Use $0 NUM_JOBS FILE_SIZE FSYNC_FREQ BLOCK_SIZE [write|randwrite]" exit 1 fi NUM_JOBS=$1 FILE_SIZE=$2 FSYNC_FREQ=$3 BLOCK_SIZE=$4 WRITE_MODE=$5 if [ "$WRITE_MODE" != "write" ] && [ "$WRITE_MODE" != "randwrite" ]; then echo "Invalid WRITE_MODE, must be 'write' or 'randwrite'" exit 1 fi cat <<EOF > /tmp/fio-job.ini [writers] rw=$WRITE_MODE fsync=$FSYNC_FREQ fallocate=none group_reporting=1 direct=0 bs=$BLOCK_SIZE ioengine=sync size=$FILE_SIZE directory=$MNT numjobs=$NUM_JOBS EOF echo "performance" | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor echo echo "Using config:" echo cat /tmp/fio-job.ini echo umount $MNT &> /dev/null mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT fio /tmp/fio-job.ini umount $MNT The results were the following: ************************* *** sequential writes *** ************************* ==== 1 job, 8GiB file, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=36.6MiB/s (38.4MB/s), 36.6MiB/s-36.6MiB/s (38.4MB/s-38.4MB/s), io=8192MiB (8590MB), run=223689-223689msec After patch: WRITE: bw=40.2MiB/s (42.1MB/s), 40.2MiB/s-40.2MiB/s (42.1MB/s-42.1MB/s), io=8192MiB (8590MB), run=203980-203980msec (+9.8%, -8.8% runtime) ==== 2 jobs, 4GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=35.8MiB/s (37.5MB/s), 35.8MiB/s-35.8MiB/s (37.5MB/s-37.5MB/s), io=8192MiB (8590MB), run=228950-228950msec After patch: WRITE: bw=43.5MiB/s (45.6MB/s), 43.5MiB/s-43.5MiB/s (45.6MB/s-45.6MB/s), io=8192MiB (8590MB), run=188272-188272msec (+21.5% throughput, -17.8% runtime) ==== 4 jobs, 2GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=50.1MiB/s (52.6MB/s), 50.1MiB/s-50.1MiB/s (52.6MB/s-52.6MB/s), io=8192MiB (8590MB), run=163446-163446msec After patch: WRITE: bw=64.5MiB/s (67.6MB/s), 64.5MiB/s-64.5MiB/s (67.6MB/s-67.6MB/s), io=8192MiB (8590MB), run=126987-126987msec (+28.7% throughput, -22.3% runtime) ==== 8 jobs, 1GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=64.0MiB/s (68.1MB/s), 64.0MiB/s-64.0MiB/s (68.1MB/s-68.1MB/s), io=8192MiB (8590MB), run=126075-126075msec After patch: WRITE: bw=86.8MiB/s (91.0MB/s), 86.8MiB/s-86.8MiB/s (91.0MB/s-91.0MB/s), io=8192MiB (8590MB), run=94358-94358msec (+35.6% throughput, -25.2% runtime) ==== 16 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=79.8MiB/s (83.6MB/s), 79.8MiB/s-79.8MiB/s (83.6MB/s-83.6MB/s), io=8192MiB (8590MB), run=102694-102694msec After patch: WRITE: bw=107MiB/s (112MB/s), 107MiB/s-107MiB/s (112MB/s-112MB/s), io=8192MiB (8590MB), run=76446-76446msec (+34.1% throughput, -25.6% runtime) ==== 32 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=93.2MiB/s (97.7MB/s), 93.2MiB/s-93.2MiB/s (97.7MB/s-97.7MB/s), io=16.0GiB (17.2GB), run=175836-175836msec After patch: WRITE: bw=111MiB/s (117MB/s), 111MiB/s-111MiB/s (117MB/s-117MB/s), io=16.0GiB (17.2GB), run=147001-147001msec (+19.1% throughput, -16.4% runtime) ==== 64 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=108MiB/s (114MB/s), 108MiB/s-108MiB/s (114MB/s-114MB/s), io=32.0GiB (34.4GB), run=302656-302656msec After patch: WRITE: bw=133MiB/s (140MB/s), 133MiB/s-133MiB/s (140MB/s-140MB/s), io=32.0GiB (34.4GB), run=246003-246003msec (+23.1% throughput, -18.7% runtime) ************************ *** random writes *** ************************ ==== 1 job, 8GiB file, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=11.5MiB/s (12.0MB/s), 11.5MiB/s-11.5MiB/s (12.0MB/s-12.0MB/s), io=8192MiB (8590MB), run=714281-714281msec After patch: WRITE: bw=11.6MiB/s (12.2MB/s), 11.6MiB/s-11.6MiB/s (12.2MB/s-12.2MB/s), io=8192MiB (8590MB), run=705959-705959msec (+0.9% throughput, -1.7% runtime) ==== 2 jobs, 4GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=12.8MiB/s (13.5MB/s), 12.8MiB/s-12.8MiB/s (13.5MB/s-13.5MB/s), io=8192MiB (8590MB), run=638101-638101msec After patch: WRITE: bw=13.1MiB/s (13.7MB/s), 13.1MiB/s-13.1MiB/s (13.7MB/s-13.7MB/s), io=8192MiB (8590MB), run=625374-625374msec (+2.3% throughput, -2.0% runtime) ==== 4 jobs, 2GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=15.4MiB/s (16.2MB/s), 15.4MiB/s-15.4MiB/s (16.2MB/s-16.2MB/s), io=8192MiB (8590MB), run=531146-531146msec After patch: WRITE: bw=17.8MiB/s (18.7MB/s), 17.8MiB/s-17.8MiB/s (18.7MB/s-18.7MB/s), io=8192MiB (8590MB), run=460431-460431msec (+15.6% throughput, -13.3% runtime) ==== 8 jobs, 1GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=19.9MiB/s (20.8MB/s), 19.9MiB/s-19.9MiB/s (20.8MB/s-20.8MB/s), io=8192MiB (8590MB), run=412664-412664msec After patch: WRITE: bw=22.2MiB/s (23.3MB/s), 22.2MiB/s-22.2MiB/s (23.3MB/s-23.3MB/s), io=8192MiB (8590MB), run=368589-368589msec (+11.6% throughput, -10.7% runtime) ==== 16 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=29.3MiB/s (30.7MB/s), 29.3MiB/s-29.3MiB/s (30.7MB/s-30.7MB/s), io=8192MiB (8590MB), run=279924-279924msec After patch: WRITE: bw=30.4MiB/s (31.9MB/s), 30.4MiB/s-30.4MiB/s (31.9MB/s-31.9MB/s), io=8192MiB (8590MB), run=269258-269258msec (+3.8% throughput, -3.8% runtime) ==== 32 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=36.9MiB/s (38.7MB/s), 36.9MiB/s-36.9MiB/s (38.7MB/s-38.7MB/s), io=16.0GiB (17.2GB), run=443581-443581msec After patch: WRITE: bw=41.6MiB/s (43.6MB/s), 41.6MiB/s-41.6MiB/s (43.6MB/s-43.6MB/s), io=16.0GiB (17.2GB), run=394114-394114msec (+12.7% throughput, -11.2% runtime) ==== 64 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=45.9MiB/s (48.1MB/s), 45.9MiB/s-45.9MiB/s (48.1MB/s-48.1MB/s), io=32.0GiB (34.4GB), run=714614-714614msec After patch: WRITE: bw=48.8MiB/s (51.1MB/s), 48.8MiB/s-48.8MiB/s (51.1MB/s-51.1MB/s), io=32.0GiB (34.4GB), run=672087-672087msec (+6.3% throughput, -6.0% runtime) Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-08-11 14:43:58 +03:00
struct btrfs_ordered_extent *ordered;
struct btrfs_root *csum_root;
u64 csum_offset;
u64 csum_len;
btrfs: make fast fsyncs wait only for writeback Currently regardless of a full or a fast fsync we always wait for ordered extents to complete, and then start logging the inode after that. However for fast fsyncs we can just wait for the writeback to complete, we don't need to wait for the ordered extents to complete since we use the list of modified extents maps to figure out which extents we must log and we can get their checksums directly from the ordered extents that are still in flight, otherwise look them up from the checksums tree. Until commit b5e6c3e170b770 ("btrfs: always wait on ordered extents at fsync time"), for fast fsyncs, we used to start logging without even waiting for the writeback to complete first, we would wait for it to complete after logging, while holding a transaction open, which lead to performance issues when using cgroups and probably for other cases too, as wait for IO while holding a transaction handle should be avoided as much as possible. After that, for fast fsyncs, we started to wait for ordered extents to complete before starting to log, which adds some latency to fsyncs and we even got at least one report about a performance drop which bisected to that particular change: https://lore.kernel.org/linux-btrfs/20181109215148.GF23260@techsingularity.net/ This change makes fast fsyncs only wait for writeback to finish before starting to log the inode, instead of waiting for both the writeback to finish and for the ordered extents to complete. This brings back part of the logic we had that extracts checksums from in flight ordered extents, which are not yet in the checksums tree, and making sure transaction commits wait for the completion of ordered extents previously logged (by far most of the time they have already completed by the time a transaction commit starts, resulting in no wait at all), to avoid any data loss if an ordered extent completes after the transaction used to log an inode is committed, followed by a power failure. When there are no other tasks accessing the checksums and the subvolume btrees, the ordered extent completion is pretty fast, typically taking 100 to 200 microseconds only in my observations. However when there are other tasks accessing these btrees, ordered extent completion can take a lot more time due to lock contention on nodes and leaves of these btrees. I've seen cases over 2 milliseconds, which starts to be significant. In particular when we do have concurrent fsyncs against different files there is a lot of contention on the checksums btree, since we have many tasks writing the checksums into the btree and other tasks that already started the logging phase are doing lookups for checksums in the btree. This change also turns all ranged fsyncs into full ranged fsyncs, which is something we already did when not using the NO_HOLES features or when doing a full fsync. This is to guarantee we never miss checksums due to writeback having been triggered only for a part of an extent, and we end up logging the full extent but only checksums for the written range, which results in missing checksums after log replay. Allowing ranged fsyncs to operate again only in the original range, when using the NO_HOLES feature and doing a fast fsync is doable but requires some non trivial changes to the writeback path, which can always be worked on later if needed, but I don't think they are a very common use case. Several tests were performed using fio for different numbers of concurrent jobs, each writing and fsyncing its own file, for both sequential and random file writes. The tests were run on bare metal, no virtualization, on a box with 12 cores (Intel i7-8700), 64Gb of RAM and a NVMe device, with a kernel configuration that is the default of typical distributions (debian in this case), without debug options enabled (kasan, kmemleak, slub debug, debug of page allocations, lock debugging, etc). The following script that calls fio was used: $ cat test-fsync.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/btrfs MOUNT_OPTIONS="-o ssd -o space_cache=v2" MKFS_OPTIONS="-d single -m single" if [ $# -ne 5 ]; then echo "Use $0 NUM_JOBS FILE_SIZE FSYNC_FREQ BLOCK_SIZE [write|randwrite]" exit 1 fi NUM_JOBS=$1 FILE_SIZE=$2 FSYNC_FREQ=$3 BLOCK_SIZE=$4 WRITE_MODE=$5 if [ "$WRITE_MODE" != "write" ] && [ "$WRITE_MODE" != "randwrite" ]; then echo "Invalid WRITE_MODE, must be 'write' or 'randwrite'" exit 1 fi cat <<EOF > /tmp/fio-job.ini [writers] rw=$WRITE_MODE fsync=$FSYNC_FREQ fallocate=none group_reporting=1 direct=0 bs=$BLOCK_SIZE ioengine=sync size=$FILE_SIZE directory=$MNT numjobs=$NUM_JOBS EOF echo "performance" | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor echo echo "Using config:" echo cat /tmp/fio-job.ini echo umount $MNT &> /dev/null mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT fio /tmp/fio-job.ini umount $MNT The results were the following: ************************* *** sequential writes *** ************************* ==== 1 job, 8GiB file, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=36.6MiB/s (38.4MB/s), 36.6MiB/s-36.6MiB/s (38.4MB/s-38.4MB/s), io=8192MiB (8590MB), run=223689-223689msec After patch: WRITE: bw=40.2MiB/s (42.1MB/s), 40.2MiB/s-40.2MiB/s (42.1MB/s-42.1MB/s), io=8192MiB (8590MB), run=203980-203980msec (+9.8%, -8.8% runtime) ==== 2 jobs, 4GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=35.8MiB/s (37.5MB/s), 35.8MiB/s-35.8MiB/s (37.5MB/s-37.5MB/s), io=8192MiB (8590MB), run=228950-228950msec After patch: WRITE: bw=43.5MiB/s (45.6MB/s), 43.5MiB/s-43.5MiB/s (45.6MB/s-45.6MB/s), io=8192MiB (8590MB), run=188272-188272msec (+21.5% throughput, -17.8% runtime) ==== 4 jobs, 2GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=50.1MiB/s (52.6MB/s), 50.1MiB/s-50.1MiB/s (52.6MB/s-52.6MB/s), io=8192MiB (8590MB), run=163446-163446msec After patch: WRITE: bw=64.5MiB/s (67.6MB/s), 64.5MiB/s-64.5MiB/s (67.6MB/s-67.6MB/s), io=8192MiB (8590MB), run=126987-126987msec (+28.7% throughput, -22.3% runtime) ==== 8 jobs, 1GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=64.0MiB/s (68.1MB/s), 64.0MiB/s-64.0MiB/s (68.1MB/s-68.1MB/s), io=8192MiB (8590MB), run=126075-126075msec After patch: WRITE: bw=86.8MiB/s (91.0MB/s), 86.8MiB/s-86.8MiB/s (91.0MB/s-91.0MB/s), io=8192MiB (8590MB), run=94358-94358msec (+35.6% throughput, -25.2% runtime) ==== 16 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=79.8MiB/s (83.6MB/s), 79.8MiB/s-79.8MiB/s (83.6MB/s-83.6MB/s), io=8192MiB (8590MB), run=102694-102694msec After patch: WRITE: bw=107MiB/s (112MB/s), 107MiB/s-107MiB/s (112MB/s-112MB/s), io=8192MiB (8590MB), run=76446-76446msec (+34.1% throughput, -25.6% runtime) ==== 32 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=93.2MiB/s (97.7MB/s), 93.2MiB/s-93.2MiB/s (97.7MB/s-97.7MB/s), io=16.0GiB (17.2GB), run=175836-175836msec After patch: WRITE: bw=111MiB/s (117MB/s), 111MiB/s-111MiB/s (117MB/s-117MB/s), io=16.0GiB (17.2GB), run=147001-147001msec (+19.1% throughput, -16.4% runtime) ==== 64 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=108MiB/s (114MB/s), 108MiB/s-108MiB/s (114MB/s-114MB/s), io=32.0GiB (34.4GB), run=302656-302656msec After patch: WRITE: bw=133MiB/s (140MB/s), 133MiB/s-133MiB/s (140MB/s-140MB/s), io=32.0GiB (34.4GB), run=246003-246003msec (+23.1% throughput, -18.7% runtime) ************************ *** random writes *** ************************ ==== 1 job, 8GiB file, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=11.5MiB/s (12.0MB/s), 11.5MiB/s-11.5MiB/s (12.0MB/s-12.0MB/s), io=8192MiB (8590MB), run=714281-714281msec After patch: WRITE: bw=11.6MiB/s (12.2MB/s), 11.6MiB/s-11.6MiB/s (12.2MB/s-12.2MB/s), io=8192MiB (8590MB), run=705959-705959msec (+0.9% throughput, -1.7% runtime) ==== 2 jobs, 4GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=12.8MiB/s (13.5MB/s), 12.8MiB/s-12.8MiB/s (13.5MB/s-13.5MB/s), io=8192MiB (8590MB), run=638101-638101msec After patch: WRITE: bw=13.1MiB/s (13.7MB/s), 13.1MiB/s-13.1MiB/s (13.7MB/s-13.7MB/s), io=8192MiB (8590MB), run=625374-625374msec (+2.3% throughput, -2.0% runtime) ==== 4 jobs, 2GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=15.4MiB/s (16.2MB/s), 15.4MiB/s-15.4MiB/s (16.2MB/s-16.2MB/s), io=8192MiB (8590MB), run=531146-531146msec After patch: WRITE: bw=17.8MiB/s (18.7MB/s), 17.8MiB/s-17.8MiB/s (18.7MB/s-18.7MB/s), io=8192MiB (8590MB), run=460431-460431msec (+15.6% throughput, -13.3% runtime) ==== 8 jobs, 1GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=19.9MiB/s (20.8MB/s), 19.9MiB/s-19.9MiB/s (20.8MB/s-20.8MB/s), io=8192MiB (8590MB), run=412664-412664msec After patch: WRITE: bw=22.2MiB/s (23.3MB/s), 22.2MiB/s-22.2MiB/s (23.3MB/s-23.3MB/s), io=8192MiB (8590MB), run=368589-368589msec (+11.6% throughput, -10.7% runtime) ==== 16 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=29.3MiB/s (30.7MB/s), 29.3MiB/s-29.3MiB/s (30.7MB/s-30.7MB/s), io=8192MiB (8590MB), run=279924-279924msec After patch: WRITE: bw=30.4MiB/s (31.9MB/s), 30.4MiB/s-30.4MiB/s (31.9MB/s-31.9MB/s), io=8192MiB (8590MB), run=269258-269258msec (+3.8% throughput, -3.8% runtime) ==== 32 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=36.9MiB/s (38.7MB/s), 36.9MiB/s-36.9MiB/s (38.7MB/s-38.7MB/s), io=16.0GiB (17.2GB), run=443581-443581msec After patch: WRITE: bw=41.6MiB/s (43.6MB/s), 41.6MiB/s-41.6MiB/s (43.6MB/s-43.6MB/s), io=16.0GiB (17.2GB), run=394114-394114msec (+12.7% throughput, -11.2% runtime) ==== 64 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=45.9MiB/s (48.1MB/s), 45.9MiB/s-45.9MiB/s (48.1MB/s-48.1MB/s), io=32.0GiB (34.4GB), run=714614-714614msec After patch: WRITE: bw=48.8MiB/s (51.1MB/s), 48.8MiB/s-48.8MiB/s (51.1MB/s-51.1MB/s), io=32.0GiB (34.4GB), run=672087-672087msec (+6.3% throughput, -6.0% runtime) Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-08-11 14:43:58 +03:00
u64 mod_start = em->mod_start;
u64 mod_len = em->mod_len;
Btrfs: fix data corruption after fast fsync and writeback error When we do a fast fsync, we start all ordered operations and then while they're running in parallel we visit the list of modified extent maps and construct their matching file extent items and write them to the log btree. After that, in btrfs_sync_log() we wait for all the ordered operations to finish (via btrfs_wait_logged_extents). The problem with this is that we were completely ignoring errors that can happen in the extent write path, such as -ENOSPC, a temporary -ENOMEM or -EIO errors for example. When such error happens, it means we have parts of the on disk extent that weren't written to, and so we end up logging file extent items that point to these extents that contain garbage/random data - so after a crash/reboot plus log replay, we get our inode's metadata pointing to those extents. This worked in contrast with the full (non-fast) fsync path, where we start all ordered operations, wait for them to finish and then write to the log btree. In this path, after each ordered operation completes we check if it's flagged with an error (BTRFS_ORDERED_IOERR) and return -EIO if so (via btrfs_wait_ordered_range). So if an error happens with any ordered operation, just return a -EIO error to userspace, so that it knows that not all of its previous writes were durably persisted and the application can take proper action (like redo the writes for e.g.) - and definitely not leave any file extent items in the log refer to non fully written extents. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2014-09-05 18:14:39 +04:00
LIST_HEAD(ordered_sums);
int ret = 0;
if (inode->flags & BTRFS_INODE_NODATASUM ||
test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
Btrfs: fix data corruption after fast fsync and writeback error When we do a fast fsync, we start all ordered operations and then while they're running in parallel we visit the list of modified extent maps and construct their matching file extent items and write them to the log btree. After that, in btrfs_sync_log() we wait for all the ordered operations to finish (via btrfs_wait_logged_extents). The problem with this is that we were completely ignoring errors that can happen in the extent write path, such as -ENOSPC, a temporary -ENOMEM or -EIO errors for example. When such error happens, it means we have parts of the on disk extent that weren't written to, and so we end up logging file extent items that point to these extents that contain garbage/random data - so after a crash/reboot plus log replay, we get our inode's metadata pointing to those extents. This worked in contrast with the full (non-fast) fsync path, where we start all ordered operations, wait for them to finish and then write to the log btree. In this path, after each ordered operation completes we check if it's flagged with an error (BTRFS_ORDERED_IOERR) and return -EIO if so (via btrfs_wait_ordered_range). So if an error happens with any ordered operation, just return a -EIO error to userspace, so that it knows that not all of its previous writes were durably persisted and the application can take proper action (like redo the writes for e.g.) - and definitely not leave any file extent items in the log refer to non fully written extents. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2014-09-05 18:14:39 +04:00
em->block_start == EXTENT_MAP_HOLE)
return 0;
Btrfs: turbo charge fsync At least for the vm workload. Currently on fsync we will 1) Truncate all items in the log tree for the given inode if they exist and 2) Copy all items for a given inode into the log The problem with this is that for things like VMs you can have lots of extents from the fragmented writing behavior, and worst yet you may have only modified a few extents, not the entire thing. This patch fixes this problem by tracking which transid modified our extent, and then when we do the tree logging we find all of the extents we've modified in our current transaction, sort them and commit them. We also only truncate up to the xattrs of the inode and copy that stuff in normally, and then just drop any extents in the range we have that exist in the log already. Here are some numbers of a 50 meg fio job that does random writes and fsync()s after every write Original Patched SATA drive 82KB/s 140KB/s Fusion drive 431KB/s 2532KB/s So around 2-6 times faster depending on your hardware. There are a few corner cases, for example if you truncate at all we have to do it the old way since there is no way to be sure what is in the log is ok. This probably could be done smarter, but if you write-fsync-truncate-write-fsync you deserve what you get. All this work is in RAM of course so if your inode gets evicted from cache and you read it in and fsync it we'll do it the slow way if we are still in the same transaction that we last modified the inode in. The biggest cool part of this is that it requires no changes to the recovery code, so if you fsync with this patch and crash and load an old kernel, it will run the recovery and be a-ok. I have tested this pretty thoroughly with an fsync tester and everything comes back fine, as well as xfstests. Thanks, Signed-off-by: Josef Bacik <jbacik@fusionio.com>
2012-08-17 21:14:17 +04:00
btrfs: make fast fsyncs wait only for writeback Currently regardless of a full or a fast fsync we always wait for ordered extents to complete, and then start logging the inode after that. However for fast fsyncs we can just wait for the writeback to complete, we don't need to wait for the ordered extents to complete since we use the list of modified extents maps to figure out which extents we must log and we can get their checksums directly from the ordered extents that are still in flight, otherwise look them up from the checksums tree. Until commit b5e6c3e170b770 ("btrfs: always wait on ordered extents at fsync time"), for fast fsyncs, we used to start logging without even waiting for the writeback to complete first, we would wait for it to complete after logging, while holding a transaction open, which lead to performance issues when using cgroups and probably for other cases too, as wait for IO while holding a transaction handle should be avoided as much as possible. After that, for fast fsyncs, we started to wait for ordered extents to complete before starting to log, which adds some latency to fsyncs and we even got at least one report about a performance drop which bisected to that particular change: https://lore.kernel.org/linux-btrfs/20181109215148.GF23260@techsingularity.net/ This change makes fast fsyncs only wait for writeback to finish before starting to log the inode, instead of waiting for both the writeback to finish and for the ordered extents to complete. This brings back part of the logic we had that extracts checksums from in flight ordered extents, which are not yet in the checksums tree, and making sure transaction commits wait for the completion of ordered extents previously logged (by far most of the time they have already completed by the time a transaction commit starts, resulting in no wait at all), to avoid any data loss if an ordered extent completes after the transaction used to log an inode is committed, followed by a power failure. When there are no other tasks accessing the checksums and the subvolume btrees, the ordered extent completion is pretty fast, typically taking 100 to 200 microseconds only in my observations. However when there are other tasks accessing these btrees, ordered extent completion can take a lot more time due to lock contention on nodes and leaves of these btrees. I've seen cases over 2 milliseconds, which starts to be significant. In particular when we do have concurrent fsyncs against different files there is a lot of contention on the checksums btree, since we have many tasks writing the checksums into the btree and other tasks that already started the logging phase are doing lookups for checksums in the btree. This change also turns all ranged fsyncs into full ranged fsyncs, which is something we already did when not using the NO_HOLES features or when doing a full fsync. This is to guarantee we never miss checksums due to writeback having been triggered only for a part of an extent, and we end up logging the full extent but only checksums for the written range, which results in missing checksums after log replay. Allowing ranged fsyncs to operate again only in the original range, when using the NO_HOLES feature and doing a fast fsync is doable but requires some non trivial changes to the writeback path, which can always be worked on later if needed, but I don't think they are a very common use case. Several tests were performed using fio for different numbers of concurrent jobs, each writing and fsyncing its own file, for both sequential and random file writes. The tests were run on bare metal, no virtualization, on a box with 12 cores (Intel i7-8700), 64Gb of RAM and a NVMe device, with a kernel configuration that is the default of typical distributions (debian in this case), without debug options enabled (kasan, kmemleak, slub debug, debug of page allocations, lock debugging, etc). The following script that calls fio was used: $ cat test-fsync.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/btrfs MOUNT_OPTIONS="-o ssd -o space_cache=v2" MKFS_OPTIONS="-d single -m single" if [ $# -ne 5 ]; then echo "Use $0 NUM_JOBS FILE_SIZE FSYNC_FREQ BLOCK_SIZE [write|randwrite]" exit 1 fi NUM_JOBS=$1 FILE_SIZE=$2 FSYNC_FREQ=$3 BLOCK_SIZE=$4 WRITE_MODE=$5 if [ "$WRITE_MODE" != "write" ] && [ "$WRITE_MODE" != "randwrite" ]; then echo "Invalid WRITE_MODE, must be 'write' or 'randwrite'" exit 1 fi cat <<EOF > /tmp/fio-job.ini [writers] rw=$WRITE_MODE fsync=$FSYNC_FREQ fallocate=none group_reporting=1 direct=0 bs=$BLOCK_SIZE ioengine=sync size=$FILE_SIZE directory=$MNT numjobs=$NUM_JOBS EOF echo "performance" | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor echo echo "Using config:" echo cat /tmp/fio-job.ini echo umount $MNT &> /dev/null mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT fio /tmp/fio-job.ini umount $MNT The results were the following: ************************* *** sequential writes *** ************************* ==== 1 job, 8GiB file, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=36.6MiB/s (38.4MB/s), 36.6MiB/s-36.6MiB/s (38.4MB/s-38.4MB/s), io=8192MiB (8590MB), run=223689-223689msec After patch: WRITE: bw=40.2MiB/s (42.1MB/s), 40.2MiB/s-40.2MiB/s (42.1MB/s-42.1MB/s), io=8192MiB (8590MB), run=203980-203980msec (+9.8%, -8.8% runtime) ==== 2 jobs, 4GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=35.8MiB/s (37.5MB/s), 35.8MiB/s-35.8MiB/s (37.5MB/s-37.5MB/s), io=8192MiB (8590MB), run=228950-228950msec After patch: WRITE: bw=43.5MiB/s (45.6MB/s), 43.5MiB/s-43.5MiB/s (45.6MB/s-45.6MB/s), io=8192MiB (8590MB), run=188272-188272msec (+21.5% throughput, -17.8% runtime) ==== 4 jobs, 2GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=50.1MiB/s (52.6MB/s), 50.1MiB/s-50.1MiB/s (52.6MB/s-52.6MB/s), io=8192MiB (8590MB), run=163446-163446msec After patch: WRITE: bw=64.5MiB/s (67.6MB/s), 64.5MiB/s-64.5MiB/s (67.6MB/s-67.6MB/s), io=8192MiB (8590MB), run=126987-126987msec (+28.7% throughput, -22.3% runtime) ==== 8 jobs, 1GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=64.0MiB/s (68.1MB/s), 64.0MiB/s-64.0MiB/s (68.1MB/s-68.1MB/s), io=8192MiB (8590MB), run=126075-126075msec After patch: WRITE: bw=86.8MiB/s (91.0MB/s), 86.8MiB/s-86.8MiB/s (91.0MB/s-91.0MB/s), io=8192MiB (8590MB), run=94358-94358msec (+35.6% throughput, -25.2% runtime) ==== 16 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=79.8MiB/s (83.6MB/s), 79.8MiB/s-79.8MiB/s (83.6MB/s-83.6MB/s), io=8192MiB (8590MB), run=102694-102694msec After patch: WRITE: bw=107MiB/s (112MB/s), 107MiB/s-107MiB/s (112MB/s-112MB/s), io=8192MiB (8590MB), run=76446-76446msec (+34.1% throughput, -25.6% runtime) ==== 32 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=93.2MiB/s (97.7MB/s), 93.2MiB/s-93.2MiB/s (97.7MB/s-97.7MB/s), io=16.0GiB (17.2GB), run=175836-175836msec After patch: WRITE: bw=111MiB/s (117MB/s), 111MiB/s-111MiB/s (117MB/s-117MB/s), io=16.0GiB (17.2GB), run=147001-147001msec (+19.1% throughput, -16.4% runtime) ==== 64 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=108MiB/s (114MB/s), 108MiB/s-108MiB/s (114MB/s-114MB/s), io=32.0GiB (34.4GB), run=302656-302656msec After patch: WRITE: bw=133MiB/s (140MB/s), 133MiB/s-133MiB/s (140MB/s-140MB/s), io=32.0GiB (34.4GB), run=246003-246003msec (+23.1% throughput, -18.7% runtime) ************************ *** random writes *** ************************ ==== 1 job, 8GiB file, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=11.5MiB/s (12.0MB/s), 11.5MiB/s-11.5MiB/s (12.0MB/s-12.0MB/s), io=8192MiB (8590MB), run=714281-714281msec After patch: WRITE: bw=11.6MiB/s (12.2MB/s), 11.6MiB/s-11.6MiB/s (12.2MB/s-12.2MB/s), io=8192MiB (8590MB), run=705959-705959msec (+0.9% throughput, -1.7% runtime) ==== 2 jobs, 4GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=12.8MiB/s (13.5MB/s), 12.8MiB/s-12.8MiB/s (13.5MB/s-13.5MB/s), io=8192MiB (8590MB), run=638101-638101msec After patch: WRITE: bw=13.1MiB/s (13.7MB/s), 13.1MiB/s-13.1MiB/s (13.7MB/s-13.7MB/s), io=8192MiB (8590MB), run=625374-625374msec (+2.3% throughput, -2.0% runtime) ==== 4 jobs, 2GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=15.4MiB/s (16.2MB/s), 15.4MiB/s-15.4MiB/s (16.2MB/s-16.2MB/s), io=8192MiB (8590MB), run=531146-531146msec After patch: WRITE: bw=17.8MiB/s (18.7MB/s), 17.8MiB/s-17.8MiB/s (18.7MB/s-18.7MB/s), io=8192MiB (8590MB), run=460431-460431msec (+15.6% throughput, -13.3% runtime) ==== 8 jobs, 1GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=19.9MiB/s (20.8MB/s), 19.9MiB/s-19.9MiB/s (20.8MB/s-20.8MB/s), io=8192MiB (8590MB), run=412664-412664msec After patch: WRITE: bw=22.2MiB/s (23.3MB/s), 22.2MiB/s-22.2MiB/s (23.3MB/s-23.3MB/s), io=8192MiB (8590MB), run=368589-368589msec (+11.6% throughput, -10.7% runtime) ==== 16 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=29.3MiB/s (30.7MB/s), 29.3MiB/s-29.3MiB/s (30.7MB/s-30.7MB/s), io=8192MiB (8590MB), run=279924-279924msec After patch: WRITE: bw=30.4MiB/s (31.9MB/s), 30.4MiB/s-30.4MiB/s (31.9MB/s-31.9MB/s), io=8192MiB (8590MB), run=269258-269258msec (+3.8% throughput, -3.8% runtime) ==== 32 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=36.9MiB/s (38.7MB/s), 36.9MiB/s-36.9MiB/s (38.7MB/s-38.7MB/s), io=16.0GiB (17.2GB), run=443581-443581msec After patch: WRITE: bw=41.6MiB/s (43.6MB/s), 41.6MiB/s-41.6MiB/s (43.6MB/s-43.6MB/s), io=16.0GiB (17.2GB), run=394114-394114msec (+12.7% throughput, -11.2% runtime) ==== 64 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=45.9MiB/s (48.1MB/s), 45.9MiB/s-45.9MiB/s (48.1MB/s-48.1MB/s), io=32.0GiB (34.4GB), run=714614-714614msec After patch: WRITE: bw=48.8MiB/s (51.1MB/s), 48.8MiB/s-48.8MiB/s (51.1MB/s-51.1MB/s), io=32.0GiB (34.4GB), run=672087-672087msec (+6.3% throughput, -6.0% runtime) Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-08-11 14:43:58 +03:00
list_for_each_entry(ordered, &ctx->ordered_extents, log_list) {
const u64 ordered_end = ordered->file_offset + ordered->num_bytes;
const u64 mod_end = mod_start + mod_len;
struct btrfs_ordered_sum *sums;
if (mod_len == 0)
break;
if (ordered_end <= mod_start)
continue;
if (mod_end <= ordered->file_offset)
break;
/*
* We are going to copy all the csums on this ordered extent, so
* go ahead and adjust mod_start and mod_len in case this ordered
* extent has already been logged.
*/
if (ordered->file_offset > mod_start) {
if (ordered_end >= mod_end)
mod_len = ordered->file_offset - mod_start;
/*
* If we have this case
*
* |--------- logged extent ---------|
* |----- ordered extent ----|
*
* Just don't mess with mod_start and mod_len, we'll
* just end up logging more csums than we need and it
* will be ok.
*/
} else {
if (ordered_end < mod_end) {
mod_len = mod_end - ordered_end;
mod_start = ordered_end;
} else {
mod_len = 0;
}
}
/*
* To keep us from looping for the above case of an ordered
* extent that falls inside of the logged extent.
*/
if (test_and_set_bit(BTRFS_ORDERED_LOGGED_CSUM, &ordered->flags))
continue;
list_for_each_entry(sums, &ordered->list, list) {
ret = log_csums(trans, inode, log_root, sums);
if (ret)
return ret;
}
}
/* We're done, found all csums in the ordered extents. */
if (mod_len == 0)
return 0;
/* If we're compressed we have to save the entire range of csums. */
if (em->compress_type) {
csum_offset = 0;
Btrfs: fix data corruption after fast fsync and writeback error When we do a fast fsync, we start all ordered operations and then while they're running in parallel we visit the list of modified extent maps and construct their matching file extent items and write them to the log btree. After that, in btrfs_sync_log() we wait for all the ordered operations to finish (via btrfs_wait_logged_extents). The problem with this is that we were completely ignoring errors that can happen in the extent write path, such as -ENOSPC, a temporary -ENOMEM or -EIO errors for example. When such error happens, it means we have parts of the on disk extent that weren't written to, and so we end up logging file extent items that point to these extents that contain garbage/random data - so after a crash/reboot plus log replay, we get our inode's metadata pointing to those extents. This worked in contrast with the full (non-fast) fsync path, where we start all ordered operations, wait for them to finish and then write to the log btree. In this path, after each ordered operation completes we check if it's flagged with an error (BTRFS_ORDERED_IOERR) and return -EIO if so (via btrfs_wait_ordered_range). So if an error happens with any ordered operation, just return a -EIO error to userspace, so that it knows that not all of its previous writes were durably persisted and the application can take proper action (like redo the writes for e.g.) - and definitely not leave any file extent items in the log refer to non fully written extents. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2014-09-05 18:14:39 +04:00
csum_len = max(em->block_len, em->orig_block_len);
} else {
btrfs: make fast fsyncs wait only for writeback Currently regardless of a full or a fast fsync we always wait for ordered extents to complete, and then start logging the inode after that. However for fast fsyncs we can just wait for the writeback to complete, we don't need to wait for the ordered extents to complete since we use the list of modified extents maps to figure out which extents we must log and we can get their checksums directly from the ordered extents that are still in flight, otherwise look them up from the checksums tree. Until commit b5e6c3e170b770 ("btrfs: always wait on ordered extents at fsync time"), for fast fsyncs, we used to start logging without even waiting for the writeback to complete first, we would wait for it to complete after logging, while holding a transaction open, which lead to performance issues when using cgroups and probably for other cases too, as wait for IO while holding a transaction handle should be avoided as much as possible. After that, for fast fsyncs, we started to wait for ordered extents to complete before starting to log, which adds some latency to fsyncs and we even got at least one report about a performance drop which bisected to that particular change: https://lore.kernel.org/linux-btrfs/20181109215148.GF23260@techsingularity.net/ This change makes fast fsyncs only wait for writeback to finish before starting to log the inode, instead of waiting for both the writeback to finish and for the ordered extents to complete. This brings back part of the logic we had that extracts checksums from in flight ordered extents, which are not yet in the checksums tree, and making sure transaction commits wait for the completion of ordered extents previously logged (by far most of the time they have already completed by the time a transaction commit starts, resulting in no wait at all), to avoid any data loss if an ordered extent completes after the transaction used to log an inode is committed, followed by a power failure. When there are no other tasks accessing the checksums and the subvolume btrees, the ordered extent completion is pretty fast, typically taking 100 to 200 microseconds only in my observations. However when there are other tasks accessing these btrees, ordered extent completion can take a lot more time due to lock contention on nodes and leaves of these btrees. I've seen cases over 2 milliseconds, which starts to be significant. In particular when we do have concurrent fsyncs against different files there is a lot of contention on the checksums btree, since we have many tasks writing the checksums into the btree and other tasks that already started the logging phase are doing lookups for checksums in the btree. This change also turns all ranged fsyncs into full ranged fsyncs, which is something we already did when not using the NO_HOLES features or when doing a full fsync. This is to guarantee we never miss checksums due to writeback having been triggered only for a part of an extent, and we end up logging the full extent but only checksums for the written range, which results in missing checksums after log replay. Allowing ranged fsyncs to operate again only in the original range, when using the NO_HOLES feature and doing a fast fsync is doable but requires some non trivial changes to the writeback path, which can always be worked on later if needed, but I don't think they are a very common use case. Several tests were performed using fio for different numbers of concurrent jobs, each writing and fsyncing its own file, for both sequential and random file writes. The tests were run on bare metal, no virtualization, on a box with 12 cores (Intel i7-8700), 64Gb of RAM and a NVMe device, with a kernel configuration that is the default of typical distributions (debian in this case), without debug options enabled (kasan, kmemleak, slub debug, debug of page allocations, lock debugging, etc). The following script that calls fio was used: $ cat test-fsync.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/btrfs MOUNT_OPTIONS="-o ssd -o space_cache=v2" MKFS_OPTIONS="-d single -m single" if [ $# -ne 5 ]; then echo "Use $0 NUM_JOBS FILE_SIZE FSYNC_FREQ BLOCK_SIZE [write|randwrite]" exit 1 fi NUM_JOBS=$1 FILE_SIZE=$2 FSYNC_FREQ=$3 BLOCK_SIZE=$4 WRITE_MODE=$5 if [ "$WRITE_MODE" != "write" ] && [ "$WRITE_MODE" != "randwrite" ]; then echo "Invalid WRITE_MODE, must be 'write' or 'randwrite'" exit 1 fi cat <<EOF > /tmp/fio-job.ini [writers] rw=$WRITE_MODE fsync=$FSYNC_FREQ fallocate=none group_reporting=1 direct=0 bs=$BLOCK_SIZE ioengine=sync size=$FILE_SIZE directory=$MNT numjobs=$NUM_JOBS EOF echo "performance" | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor echo echo "Using config:" echo cat /tmp/fio-job.ini echo umount $MNT &> /dev/null mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT fio /tmp/fio-job.ini umount $MNT The results were the following: ************************* *** sequential writes *** ************************* ==== 1 job, 8GiB file, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=36.6MiB/s (38.4MB/s), 36.6MiB/s-36.6MiB/s (38.4MB/s-38.4MB/s), io=8192MiB (8590MB), run=223689-223689msec After patch: WRITE: bw=40.2MiB/s (42.1MB/s), 40.2MiB/s-40.2MiB/s (42.1MB/s-42.1MB/s), io=8192MiB (8590MB), run=203980-203980msec (+9.8%, -8.8% runtime) ==== 2 jobs, 4GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=35.8MiB/s (37.5MB/s), 35.8MiB/s-35.8MiB/s (37.5MB/s-37.5MB/s), io=8192MiB (8590MB), run=228950-228950msec After patch: WRITE: bw=43.5MiB/s (45.6MB/s), 43.5MiB/s-43.5MiB/s (45.6MB/s-45.6MB/s), io=8192MiB (8590MB), run=188272-188272msec (+21.5% throughput, -17.8% runtime) ==== 4 jobs, 2GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=50.1MiB/s (52.6MB/s), 50.1MiB/s-50.1MiB/s (52.6MB/s-52.6MB/s), io=8192MiB (8590MB), run=163446-163446msec After patch: WRITE: bw=64.5MiB/s (67.6MB/s), 64.5MiB/s-64.5MiB/s (67.6MB/s-67.6MB/s), io=8192MiB (8590MB), run=126987-126987msec (+28.7% throughput, -22.3% runtime) ==== 8 jobs, 1GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=64.0MiB/s (68.1MB/s), 64.0MiB/s-64.0MiB/s (68.1MB/s-68.1MB/s), io=8192MiB (8590MB), run=126075-126075msec After patch: WRITE: bw=86.8MiB/s (91.0MB/s), 86.8MiB/s-86.8MiB/s (91.0MB/s-91.0MB/s), io=8192MiB (8590MB), run=94358-94358msec (+35.6% throughput, -25.2% runtime) ==== 16 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=79.8MiB/s (83.6MB/s), 79.8MiB/s-79.8MiB/s (83.6MB/s-83.6MB/s), io=8192MiB (8590MB), run=102694-102694msec After patch: WRITE: bw=107MiB/s (112MB/s), 107MiB/s-107MiB/s (112MB/s-112MB/s), io=8192MiB (8590MB), run=76446-76446msec (+34.1% throughput, -25.6% runtime) ==== 32 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=93.2MiB/s (97.7MB/s), 93.2MiB/s-93.2MiB/s (97.7MB/s-97.7MB/s), io=16.0GiB (17.2GB), run=175836-175836msec After patch: WRITE: bw=111MiB/s (117MB/s), 111MiB/s-111MiB/s (117MB/s-117MB/s), io=16.0GiB (17.2GB), run=147001-147001msec (+19.1% throughput, -16.4% runtime) ==== 64 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=108MiB/s (114MB/s), 108MiB/s-108MiB/s (114MB/s-114MB/s), io=32.0GiB (34.4GB), run=302656-302656msec After patch: WRITE: bw=133MiB/s (140MB/s), 133MiB/s-133MiB/s (140MB/s-140MB/s), io=32.0GiB (34.4GB), run=246003-246003msec (+23.1% throughput, -18.7% runtime) ************************ *** random writes *** ************************ ==== 1 job, 8GiB file, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=11.5MiB/s (12.0MB/s), 11.5MiB/s-11.5MiB/s (12.0MB/s-12.0MB/s), io=8192MiB (8590MB), run=714281-714281msec After patch: WRITE: bw=11.6MiB/s (12.2MB/s), 11.6MiB/s-11.6MiB/s (12.2MB/s-12.2MB/s), io=8192MiB (8590MB), run=705959-705959msec (+0.9% throughput, -1.7% runtime) ==== 2 jobs, 4GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=12.8MiB/s (13.5MB/s), 12.8MiB/s-12.8MiB/s (13.5MB/s-13.5MB/s), io=8192MiB (8590MB), run=638101-638101msec After patch: WRITE: bw=13.1MiB/s (13.7MB/s), 13.1MiB/s-13.1MiB/s (13.7MB/s-13.7MB/s), io=8192MiB (8590MB), run=625374-625374msec (+2.3% throughput, -2.0% runtime) ==== 4 jobs, 2GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=15.4MiB/s (16.2MB/s), 15.4MiB/s-15.4MiB/s (16.2MB/s-16.2MB/s), io=8192MiB (8590MB), run=531146-531146msec After patch: WRITE: bw=17.8MiB/s (18.7MB/s), 17.8MiB/s-17.8MiB/s (18.7MB/s-18.7MB/s), io=8192MiB (8590MB), run=460431-460431msec (+15.6% throughput, -13.3% runtime) ==== 8 jobs, 1GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=19.9MiB/s (20.8MB/s), 19.9MiB/s-19.9MiB/s (20.8MB/s-20.8MB/s), io=8192MiB (8590MB), run=412664-412664msec After patch: WRITE: bw=22.2MiB/s (23.3MB/s), 22.2MiB/s-22.2MiB/s (23.3MB/s-23.3MB/s), io=8192MiB (8590MB), run=368589-368589msec (+11.6% throughput, -10.7% runtime) ==== 16 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=29.3MiB/s (30.7MB/s), 29.3MiB/s-29.3MiB/s (30.7MB/s-30.7MB/s), io=8192MiB (8590MB), run=279924-279924msec After patch: WRITE: bw=30.4MiB/s (31.9MB/s), 30.4MiB/s-30.4MiB/s (31.9MB/s-31.9MB/s), io=8192MiB (8590MB), run=269258-269258msec (+3.8% throughput, -3.8% runtime) ==== 32 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=36.9MiB/s (38.7MB/s), 36.9MiB/s-36.9MiB/s (38.7MB/s-38.7MB/s), io=16.0GiB (17.2GB), run=443581-443581msec After patch: WRITE: bw=41.6MiB/s (43.6MB/s), 41.6MiB/s-41.6MiB/s (43.6MB/s-43.6MB/s), io=16.0GiB (17.2GB), run=394114-394114msec (+12.7% throughput, -11.2% runtime) ==== 64 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=45.9MiB/s (48.1MB/s), 45.9MiB/s-45.9MiB/s (48.1MB/s-48.1MB/s), io=32.0GiB (34.4GB), run=714614-714614msec After patch: WRITE: bw=48.8MiB/s (51.1MB/s), 48.8MiB/s-48.8MiB/s (51.1MB/s-51.1MB/s), io=32.0GiB (34.4GB), run=672087-672087msec (+6.3% throughput, -6.0% runtime) Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-08-11 14:43:58 +03:00
csum_offset = mod_start - em->start;
csum_len = mod_len;
}
/* block start is already adjusted for the file extent offset. */
csum_root = btrfs_csum_root(trans->fs_info, em->block_start);
ret = btrfs_lookup_csums_range(csum_root,
em->block_start + csum_offset,
em->block_start + csum_offset +
csum_len - 1, &ordered_sums, 0);
if (ret)
return ret;
Btrfs: turbo charge fsync At least for the vm workload. Currently on fsync we will 1) Truncate all items in the log tree for the given inode if they exist and 2) Copy all items for a given inode into the log The problem with this is that for things like VMs you can have lots of extents from the fragmented writing behavior, and worst yet you may have only modified a few extents, not the entire thing. This patch fixes this problem by tracking which transid modified our extent, and then when we do the tree logging we find all of the extents we've modified in our current transaction, sort them and commit them. We also only truncate up to the xattrs of the inode and copy that stuff in normally, and then just drop any extents in the range we have that exist in the log already. Here are some numbers of a 50 meg fio job that does random writes and fsync()s after every write Original Patched SATA drive 82KB/s 140KB/s Fusion drive 431KB/s 2532KB/s So around 2-6 times faster depending on your hardware. There are a few corner cases, for example if you truncate at all we have to do it the old way since there is no way to be sure what is in the log is ok. This probably could be done smarter, but if you write-fsync-truncate-write-fsync you deserve what you get. All this work is in RAM of course so if your inode gets evicted from cache and you read it in and fsync it we'll do it the slow way if we are still in the same transaction that we last modified the inode in. The biggest cool part of this is that it requires no changes to the recovery code, so if you fsync with this patch and crash and load an old kernel, it will run the recovery and be a-ok. I have tested this pretty thoroughly with an fsync tester and everything comes back fine, as well as xfstests. Thanks, Signed-off-by: Josef Bacik <jbacik@fusionio.com>
2012-08-17 21:14:17 +04:00
while (!list_empty(&ordered_sums)) {
struct btrfs_ordered_sum *sums = list_entry(ordered_sums.next,
struct btrfs_ordered_sum,
list);
if (!ret)
btrfs: reduce contention on log trees when logging checksums The possibility of extents being shared (through clone and deduplication operations) requires special care when logging data checksums, to avoid having a log tree with different checksum items that cover ranges which overlap (which resulted in missing checksums after replaying a log tree). Such problems were fixed in the past by the following commits: commit 40e046acbd2f ("Btrfs: fix missing data checksums after replaying a log tree") commit e289f03ea79b ("btrfs: fix corrupt log due to concurrent fsync of inodes with shared extents") Test case generic/588 exercises the scenario solved by the first commit (purely sequential and deterministic) while test case generic/457 often triggered the case fixed by the second commit (not deterministic, requires specific timings under concurrency). The problems were addressed by deleting, from the log tree, any existing checksums before logging the new ones. And also by doing the deletion and logging of the cheksums while locking the checksum range in an extent io tree (root->log_csum_range), to deal with the case where we have concurrent fsyncs against files with shared extents. That however causes more contention on the leaves of a log tree where we store checksums (and all the nodes in the paths leading to them), even when we do not have shared extents, or all the shared extents were created by past transactions. It also adds a bit of contention on the spin lock of the log_csums_range extent io tree of the log root. This change adds a 'last_reflink_trans' field to the inode to keep track of the last transaction where a new extent was shared between inodes (through clone and deduplication operations). It is updated for both the source and destination inodes of reflink operations whenever a new extent (created in the current transaction) becomes shared by the inodes. This field is kept in memory only, not persisted in the inode item, similar to other existing fields (last_unlink_trans, logged_trans). When logging checksums for an extent, if the value of 'last_reflink_trans' is smaller then the current transaction's generation/id, we skip locking the extent range and deletion of checksums from the log tree, since we know we do not have new shared extents. This reduces contention on the log tree's leaves where checksums are stored. The following script, which uses fio, was used to measure the impact of this change: $ cat test-fsync.sh #!/bin/bash DEV=/dev/sdk MNT=/mnt/sdk MOUNT_OPTIONS="-o ssd" MKFS_OPTIONS="-d single -m single" if [ $# -ne 3 ]; then echo "Use $0 NUM_JOBS FILE_SIZE FSYNC_FREQ" exit 1 fi NUM_JOBS=$1 FILE_SIZE=$2 FSYNC_FREQ=$3 cat <<EOF > /tmp/fio-job.ini [writers] rw=write fsync=$FSYNC_FREQ fallocate=none group_reporting=1 direct=0 bs=64k ioengine=sync size=$FILE_SIZE directory=$MNT numjobs=$NUM_JOBS EOF echo "Using config:" echo cat /tmp/fio-job.ini echo mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT fio /tmp/fio-job.ini umount $MNT The tests were performed for different numbers of jobs, file sizes and fsync frequency. A qemu VM using kvm was used, with 8 cores (the host has 12 cores, with cpu governance set to performance mode on all cores), 16GiB of ram (the host has 64GiB) and using a NVMe device directly (without an intermediary filesystem in the host). While running the tests, the host was not used for anything else, to avoid disturbing the tests. The obtained results were the following (the last line of fio's output was pasted). Starting with 16 jobs is where a significant difference is observable in this particular setup and hardware (differences highlighted below). The very small differences for tests with less than 16 jobs are possibly just noise and random. **** 1 job, file size 1G, fsync frequency 1 **** before this change: WRITE: bw=23.8MiB/s (24.9MB/s), 23.8MiB/s-23.8MiB/s (24.9MB/s-24.9MB/s), io=1024MiB (1074MB), run=43075-43075msec after this change: WRITE: bw=24.4MiB/s (25.6MB/s), 24.4MiB/s-24.4MiB/s (25.6MB/s-25.6MB/s), io=1024MiB (1074MB), run=41938-41938msec **** 2 jobs, file size 1G, fsync frequency 1 **** before this change: WRITE: bw=37.7MiB/s (39.5MB/s), 37.7MiB/s-37.7MiB/s (39.5MB/s-39.5MB/s), io=2048MiB (2147MB), run=54351-54351msec after this change: WRITE: bw=37.7MiB/s (39.5MB/s), 37.6MiB/s-37.6MiB/s (39.5MB/s-39.5MB/s), io=2048MiB (2147MB), run=54428-54428msec **** 4 jobs, file size 1G, fsync frequency 1 **** before this change: WRITE: bw=67.5MiB/s (70.8MB/s), 67.5MiB/s-67.5MiB/s (70.8MB/s-70.8MB/s), io=4096MiB (4295MB), run=60669-60669msec after this change: WRITE: bw=68.6MiB/s (71.0MB/s), 68.6MiB/s-68.6MiB/s (71.0MB/s-71.0MB/s), io=4096MiB (4295MB), run=59678-59678msec **** 8 jobs, file size 1G, fsync frequency 1 **** before this change: WRITE: bw=128MiB/s (134MB/s), 128MiB/s-128MiB/s (134MB/s-134MB/s), io=8192MiB (8590MB), run=64048-64048msec after this change: WRITE: bw=129MiB/s (135MB/s), 129MiB/s-129MiB/s (135MB/s-135MB/s), io=8192MiB (8590MB), run=63405-63405msec **** 16 jobs, file size 1G, fsync frequency 1 **** before this change: WRITE: bw=78.5MiB/s (82.3MB/s), 78.5MiB/s-78.5MiB/s (82.3MB/s-82.3MB/s), io=16.0GiB (17.2GB), run=208676-208676msec after this change: WRITE: bw=110MiB/s (115MB/s), 110MiB/s-110MiB/s (115MB/s-115MB/s), io=16.0GiB (17.2GB), run=149295-149295msec (+40.1% throughput, -28.5% runtime) **** 32 jobs, file size 1G, fsync frequency 1 **** before this change: WRITE: bw=58.8MiB/s (61.7MB/s), 58.8MiB/s-58.8MiB/s (61.7MB/s-61.7MB/s), io=32.0GiB (34.4GB), run=557134-557134msec after this change: WRITE: bw=76.1MiB/s (79.8MB/s), 76.1MiB/s-76.1MiB/s (79.8MB/s-79.8MB/s), io=32.0GiB (34.4GB), run=430550-430550msec (+29.4% throughput, -22.7% runtime) **** 64 jobs, file size 512M, fsync frequency 1 **** before this change: WRITE: bw=65.8MiB/s (68.0MB/s), 65.8MiB/s-65.8MiB/s (68.0MB/s-68.0MB/s), io=32.0GiB (34.4GB), run=498055-498055msec after this change: WRITE: bw=85.1MiB/s (89.2MB/s), 85.1MiB/s-85.1MiB/s (89.2MB/s-89.2MB/s), io=32.0GiB (34.4GB), run=385116-385116msec (+29.3% throughput, -22.7% runtime) **** 128 jobs, file size 256M, fsync frequency 1 **** before this change: WRITE: bw=54.7MiB/s (57.3MB/s), 54.7MiB/s-54.7MiB/s (57.3MB/s-57.3MB/s), io=32.0GiB (34.4GB), run=599373-599373msec after this change: WRITE: bw=121MiB/s (126MB/s), 121MiB/s-121MiB/s (126MB/s-126MB/s), io=32.0GiB (34.4GB), run=271907-271907msec (+121.2% throughput, -54.6% runtime) **** 256 jobs, file size 256M, fsync frequency 1 **** before this change: WRITE: bw=69.2MiB/s (72.5MB/s), 69.2MiB/s-69.2MiB/s (72.5MB/s-72.5MB/s), io=64.0GiB (68.7GB), run=947536-947536msec after this change: WRITE: bw=121MiB/s (127MB/s), 121MiB/s-121MiB/s (127MB/s-127MB/s), io=64.0GiB (68.7GB), run=541916-541916msec (+74.9% throughput, -42.8% runtime) **** 512 jobs, file size 128M, fsync frequency 1 **** before this change: WRITE: bw=85.4MiB/s (89.5MB/s), 85.4MiB/s-85.4MiB/s (89.5MB/s-89.5MB/s), io=64.0GiB (68.7GB), run=767734-767734msec after this change: WRITE: bw=141MiB/s (147MB/s), 141MiB/s-141MiB/s (147MB/s-147MB/s), io=64.0GiB (68.7GB), run=466022-466022msec (+65.1% throughput, -39.3% runtime) **** 1024 jobs, file size 128M, fsync frequency 1 **** before this change: WRITE: bw=115MiB/s (120MB/s), 115MiB/s-115MiB/s (120MB/s-120MB/s), io=128GiB (137GB), run=1143775-1143775msec after this change: WRITE: bw=171MiB/s (180MB/s), 171MiB/s-171MiB/s (180MB/s-180MB/s), io=128GiB (137GB), run=764843-764843msec (+48.7% throughput, -33.1% runtime) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-07-15 14:30:43 +03:00
ret = log_csums(trans, inode, log_root, sums);
list_del(&sums->list);
kfree(sums);
Btrfs: turbo charge fsync At least for the vm workload. Currently on fsync we will 1) Truncate all items in the log tree for the given inode if they exist and 2) Copy all items for a given inode into the log The problem with this is that for things like VMs you can have lots of extents from the fragmented writing behavior, and worst yet you may have only modified a few extents, not the entire thing. This patch fixes this problem by tracking which transid modified our extent, and then when we do the tree logging we find all of the extents we've modified in our current transaction, sort them and commit them. We also only truncate up to the xattrs of the inode and copy that stuff in normally, and then just drop any extents in the range we have that exist in the log already. Here are some numbers of a 50 meg fio job that does random writes and fsync()s after every write Original Patched SATA drive 82KB/s 140KB/s Fusion drive 431KB/s 2532KB/s So around 2-6 times faster depending on your hardware. There are a few corner cases, for example if you truncate at all we have to do it the old way since there is no way to be sure what is in the log is ok. This probably could be done smarter, but if you write-fsync-truncate-write-fsync you deserve what you get. All this work is in RAM of course so if your inode gets evicted from cache and you read it in and fsync it we'll do it the slow way if we are still in the same transaction that we last modified the inode in. The biggest cool part of this is that it requires no changes to the recovery code, so if you fsync with this patch and crash and load an old kernel, it will run the recovery and be a-ok. I have tested this pretty thoroughly with an fsync tester and everything comes back fine, as well as xfstests. Thanks, Signed-off-by: Josef Bacik <jbacik@fusionio.com>
2012-08-17 21:14:17 +04:00
}
return ret;
Btrfs: turbo charge fsync At least for the vm workload. Currently on fsync we will 1) Truncate all items in the log tree for the given inode if they exist and 2) Copy all items for a given inode into the log The problem with this is that for things like VMs you can have lots of extents from the fragmented writing behavior, and worst yet you may have only modified a few extents, not the entire thing. This patch fixes this problem by tracking which transid modified our extent, and then when we do the tree logging we find all of the extents we've modified in our current transaction, sort them and commit them. We also only truncate up to the xattrs of the inode and copy that stuff in normally, and then just drop any extents in the range we have that exist in the log already. Here are some numbers of a 50 meg fio job that does random writes and fsync()s after every write Original Patched SATA drive 82KB/s 140KB/s Fusion drive 431KB/s 2532KB/s So around 2-6 times faster depending on your hardware. There are a few corner cases, for example if you truncate at all we have to do it the old way since there is no way to be sure what is in the log is ok. This probably could be done smarter, but if you write-fsync-truncate-write-fsync you deserve what you get. All this work is in RAM of course so if your inode gets evicted from cache and you read it in and fsync it we'll do it the slow way if we are still in the same transaction that we last modified the inode in. The biggest cool part of this is that it requires no changes to the recovery code, so if you fsync with this patch and crash and load an old kernel, it will run the recovery and be a-ok. I have tested this pretty thoroughly with an fsync tester and everything comes back fine, as well as xfstests. Thanks, Signed-off-by: Josef Bacik <jbacik@fusionio.com>
2012-08-17 21:14:17 +04:00
}
Btrfs: fix data corruption after fast fsync and writeback error When we do a fast fsync, we start all ordered operations and then while they're running in parallel we visit the list of modified extent maps and construct their matching file extent items and write them to the log btree. After that, in btrfs_sync_log() we wait for all the ordered operations to finish (via btrfs_wait_logged_extents). The problem with this is that we were completely ignoring errors that can happen in the extent write path, such as -ENOSPC, a temporary -ENOMEM or -EIO errors for example. When such error happens, it means we have parts of the on disk extent that weren't written to, and so we end up logging file extent items that point to these extents that contain garbage/random data - so after a crash/reboot plus log replay, we get our inode's metadata pointing to those extents. This worked in contrast with the full (non-fast) fsync path, where we start all ordered operations, wait for them to finish and then write to the log btree. In this path, after each ordered operation completes we check if it's flagged with an error (BTRFS_ORDERED_IOERR) and return -EIO if so (via btrfs_wait_ordered_range). So if an error happens with any ordered operation, just return a -EIO error to userspace, so that it knows that not all of its previous writes were durably persisted and the application can take proper action (like redo the writes for e.g.) - and definitely not leave any file extent items in the log refer to non fully written extents. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2014-09-05 18:14:39 +04:00
static int log_one_extent(struct btrfs_trans_handle *trans,
struct btrfs_inode *inode,
Btrfs: fix data corruption after fast fsync and writeback error When we do a fast fsync, we start all ordered operations and then while they're running in parallel we visit the list of modified extent maps and construct their matching file extent items and write them to the log btree. After that, in btrfs_sync_log() we wait for all the ordered operations to finish (via btrfs_wait_logged_extents). The problem with this is that we were completely ignoring errors that can happen in the extent write path, such as -ENOSPC, a temporary -ENOMEM or -EIO errors for example. When such error happens, it means we have parts of the on disk extent that weren't written to, and so we end up logging file extent items that point to these extents that contain garbage/random data - so after a crash/reboot plus log replay, we get our inode's metadata pointing to those extents. This worked in contrast with the full (non-fast) fsync path, where we start all ordered operations, wait for them to finish and then write to the log btree. In this path, after each ordered operation completes we check if it's flagged with an error (BTRFS_ORDERED_IOERR) and return -EIO if so (via btrfs_wait_ordered_range). So if an error happens with any ordered operation, just return a -EIO error to userspace, so that it knows that not all of its previous writes were durably persisted and the application can take proper action (like redo the writes for e.g.) - and definitely not leave any file extent items in the log refer to non fully written extents. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2014-09-05 18:14:39 +04:00
const struct extent_map *em,
struct btrfs_path *path,
struct btrfs_log_ctx *ctx)
{
struct btrfs_drop_extents_args drop_args = { 0 };
struct btrfs_root *log = inode->root->log_root;
Btrfs: fix data corruption after fast fsync and writeback error When we do a fast fsync, we start all ordered operations and then while they're running in parallel we visit the list of modified extent maps and construct their matching file extent items and write them to the log btree. After that, in btrfs_sync_log() we wait for all the ordered operations to finish (via btrfs_wait_logged_extents). The problem with this is that we were completely ignoring errors that can happen in the extent write path, such as -ENOSPC, a temporary -ENOMEM or -EIO errors for example. When such error happens, it means we have parts of the on disk extent that weren't written to, and so we end up logging file extent items that point to these extents that contain garbage/random data - so after a crash/reboot plus log replay, we get our inode's metadata pointing to those extents. This worked in contrast with the full (non-fast) fsync path, where we start all ordered operations, wait for them to finish and then write to the log btree. In this path, after each ordered operation completes we check if it's flagged with an error (BTRFS_ORDERED_IOERR) and return -EIO if so (via btrfs_wait_ordered_range). So if an error happens with any ordered operation, just return a -EIO error to userspace, so that it knows that not all of its previous writes were durably persisted and the application can take proper action (like redo the writes for e.g.) - and definitely not leave any file extent items in the log refer to non fully written extents. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2014-09-05 18:14:39 +04:00
struct btrfs_file_extent_item *fi;
struct extent_buffer *leaf;
struct btrfs_map_token token;
struct btrfs_key key;
u64 extent_offset = em->start - em->orig_start;
u64 block_len;
int ret;
btrfs: make fast fsyncs wait only for writeback Currently regardless of a full or a fast fsync we always wait for ordered extents to complete, and then start logging the inode after that. However for fast fsyncs we can just wait for the writeback to complete, we don't need to wait for the ordered extents to complete since we use the list of modified extents maps to figure out which extents we must log and we can get their checksums directly from the ordered extents that are still in flight, otherwise look them up from the checksums tree. Until commit b5e6c3e170b770 ("btrfs: always wait on ordered extents at fsync time"), for fast fsyncs, we used to start logging without even waiting for the writeback to complete first, we would wait for it to complete after logging, while holding a transaction open, which lead to performance issues when using cgroups and probably for other cases too, as wait for IO while holding a transaction handle should be avoided as much as possible. After that, for fast fsyncs, we started to wait for ordered extents to complete before starting to log, which adds some latency to fsyncs and we even got at least one report about a performance drop which bisected to that particular change: https://lore.kernel.org/linux-btrfs/20181109215148.GF23260@techsingularity.net/ This change makes fast fsyncs only wait for writeback to finish before starting to log the inode, instead of waiting for both the writeback to finish and for the ordered extents to complete. This brings back part of the logic we had that extracts checksums from in flight ordered extents, which are not yet in the checksums tree, and making sure transaction commits wait for the completion of ordered extents previously logged (by far most of the time they have already completed by the time a transaction commit starts, resulting in no wait at all), to avoid any data loss if an ordered extent completes after the transaction used to log an inode is committed, followed by a power failure. When there are no other tasks accessing the checksums and the subvolume btrees, the ordered extent completion is pretty fast, typically taking 100 to 200 microseconds only in my observations. However when there are other tasks accessing these btrees, ordered extent completion can take a lot more time due to lock contention on nodes and leaves of these btrees. I've seen cases over 2 milliseconds, which starts to be significant. In particular when we do have concurrent fsyncs against different files there is a lot of contention on the checksums btree, since we have many tasks writing the checksums into the btree and other tasks that already started the logging phase are doing lookups for checksums in the btree. This change also turns all ranged fsyncs into full ranged fsyncs, which is something we already did when not using the NO_HOLES features or when doing a full fsync. This is to guarantee we never miss checksums due to writeback having been triggered only for a part of an extent, and we end up logging the full extent but only checksums for the written range, which results in missing checksums after log replay. Allowing ranged fsyncs to operate again only in the original range, when using the NO_HOLES feature and doing a fast fsync is doable but requires some non trivial changes to the writeback path, which can always be worked on later if needed, but I don't think they are a very common use case. Several tests were performed using fio for different numbers of concurrent jobs, each writing and fsyncing its own file, for both sequential and random file writes. The tests were run on bare metal, no virtualization, on a box with 12 cores (Intel i7-8700), 64Gb of RAM and a NVMe device, with a kernel configuration that is the default of typical distributions (debian in this case), without debug options enabled (kasan, kmemleak, slub debug, debug of page allocations, lock debugging, etc). The following script that calls fio was used: $ cat test-fsync.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/btrfs MOUNT_OPTIONS="-o ssd -o space_cache=v2" MKFS_OPTIONS="-d single -m single" if [ $# -ne 5 ]; then echo "Use $0 NUM_JOBS FILE_SIZE FSYNC_FREQ BLOCK_SIZE [write|randwrite]" exit 1 fi NUM_JOBS=$1 FILE_SIZE=$2 FSYNC_FREQ=$3 BLOCK_SIZE=$4 WRITE_MODE=$5 if [ "$WRITE_MODE" != "write" ] && [ "$WRITE_MODE" != "randwrite" ]; then echo "Invalid WRITE_MODE, must be 'write' or 'randwrite'" exit 1 fi cat <<EOF > /tmp/fio-job.ini [writers] rw=$WRITE_MODE fsync=$FSYNC_FREQ fallocate=none group_reporting=1 direct=0 bs=$BLOCK_SIZE ioengine=sync size=$FILE_SIZE directory=$MNT numjobs=$NUM_JOBS EOF echo "performance" | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor echo echo "Using config:" echo cat /tmp/fio-job.ini echo umount $MNT &> /dev/null mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT fio /tmp/fio-job.ini umount $MNT The results were the following: ************************* *** sequential writes *** ************************* ==== 1 job, 8GiB file, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=36.6MiB/s (38.4MB/s), 36.6MiB/s-36.6MiB/s (38.4MB/s-38.4MB/s), io=8192MiB (8590MB), run=223689-223689msec After patch: WRITE: bw=40.2MiB/s (42.1MB/s), 40.2MiB/s-40.2MiB/s (42.1MB/s-42.1MB/s), io=8192MiB (8590MB), run=203980-203980msec (+9.8%, -8.8% runtime) ==== 2 jobs, 4GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=35.8MiB/s (37.5MB/s), 35.8MiB/s-35.8MiB/s (37.5MB/s-37.5MB/s), io=8192MiB (8590MB), run=228950-228950msec After patch: WRITE: bw=43.5MiB/s (45.6MB/s), 43.5MiB/s-43.5MiB/s (45.6MB/s-45.6MB/s), io=8192MiB (8590MB), run=188272-188272msec (+21.5% throughput, -17.8% runtime) ==== 4 jobs, 2GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=50.1MiB/s (52.6MB/s), 50.1MiB/s-50.1MiB/s (52.6MB/s-52.6MB/s), io=8192MiB (8590MB), run=163446-163446msec After patch: WRITE: bw=64.5MiB/s (67.6MB/s), 64.5MiB/s-64.5MiB/s (67.6MB/s-67.6MB/s), io=8192MiB (8590MB), run=126987-126987msec (+28.7% throughput, -22.3% runtime) ==== 8 jobs, 1GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=64.0MiB/s (68.1MB/s), 64.0MiB/s-64.0MiB/s (68.1MB/s-68.1MB/s), io=8192MiB (8590MB), run=126075-126075msec After patch: WRITE: bw=86.8MiB/s (91.0MB/s), 86.8MiB/s-86.8MiB/s (91.0MB/s-91.0MB/s), io=8192MiB (8590MB), run=94358-94358msec (+35.6% throughput, -25.2% runtime) ==== 16 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=79.8MiB/s (83.6MB/s), 79.8MiB/s-79.8MiB/s (83.6MB/s-83.6MB/s), io=8192MiB (8590MB), run=102694-102694msec After patch: WRITE: bw=107MiB/s (112MB/s), 107MiB/s-107MiB/s (112MB/s-112MB/s), io=8192MiB (8590MB), run=76446-76446msec (+34.1% throughput, -25.6% runtime) ==== 32 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=93.2MiB/s (97.7MB/s), 93.2MiB/s-93.2MiB/s (97.7MB/s-97.7MB/s), io=16.0GiB (17.2GB), run=175836-175836msec After patch: WRITE: bw=111MiB/s (117MB/s), 111MiB/s-111MiB/s (117MB/s-117MB/s), io=16.0GiB (17.2GB), run=147001-147001msec (+19.1% throughput, -16.4% runtime) ==== 64 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=108MiB/s (114MB/s), 108MiB/s-108MiB/s (114MB/s-114MB/s), io=32.0GiB (34.4GB), run=302656-302656msec After patch: WRITE: bw=133MiB/s (140MB/s), 133MiB/s-133MiB/s (140MB/s-140MB/s), io=32.0GiB (34.4GB), run=246003-246003msec (+23.1% throughput, -18.7% runtime) ************************ *** random writes *** ************************ ==== 1 job, 8GiB file, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=11.5MiB/s (12.0MB/s), 11.5MiB/s-11.5MiB/s (12.0MB/s-12.0MB/s), io=8192MiB (8590MB), run=714281-714281msec After patch: WRITE: bw=11.6MiB/s (12.2MB/s), 11.6MiB/s-11.6MiB/s (12.2MB/s-12.2MB/s), io=8192MiB (8590MB), run=705959-705959msec (+0.9% throughput, -1.7% runtime) ==== 2 jobs, 4GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=12.8MiB/s (13.5MB/s), 12.8MiB/s-12.8MiB/s (13.5MB/s-13.5MB/s), io=8192MiB (8590MB), run=638101-638101msec After patch: WRITE: bw=13.1MiB/s (13.7MB/s), 13.1MiB/s-13.1MiB/s (13.7MB/s-13.7MB/s), io=8192MiB (8590MB), run=625374-625374msec (+2.3% throughput, -2.0% runtime) ==== 4 jobs, 2GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=15.4MiB/s (16.2MB/s), 15.4MiB/s-15.4MiB/s (16.2MB/s-16.2MB/s), io=8192MiB (8590MB), run=531146-531146msec After patch: WRITE: bw=17.8MiB/s (18.7MB/s), 17.8MiB/s-17.8MiB/s (18.7MB/s-18.7MB/s), io=8192MiB (8590MB), run=460431-460431msec (+15.6% throughput, -13.3% runtime) ==== 8 jobs, 1GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=19.9MiB/s (20.8MB/s), 19.9MiB/s-19.9MiB/s (20.8MB/s-20.8MB/s), io=8192MiB (8590MB), run=412664-412664msec After patch: WRITE: bw=22.2MiB/s (23.3MB/s), 22.2MiB/s-22.2MiB/s (23.3MB/s-23.3MB/s), io=8192MiB (8590MB), run=368589-368589msec (+11.6% throughput, -10.7% runtime) ==== 16 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=29.3MiB/s (30.7MB/s), 29.3MiB/s-29.3MiB/s (30.7MB/s-30.7MB/s), io=8192MiB (8590MB), run=279924-279924msec After patch: WRITE: bw=30.4MiB/s (31.9MB/s), 30.4MiB/s-30.4MiB/s (31.9MB/s-31.9MB/s), io=8192MiB (8590MB), run=269258-269258msec (+3.8% throughput, -3.8% runtime) ==== 32 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=36.9MiB/s (38.7MB/s), 36.9MiB/s-36.9MiB/s (38.7MB/s-38.7MB/s), io=16.0GiB (17.2GB), run=443581-443581msec After patch: WRITE: bw=41.6MiB/s (43.6MB/s), 41.6MiB/s-41.6MiB/s (43.6MB/s-43.6MB/s), io=16.0GiB (17.2GB), run=394114-394114msec (+12.7% throughput, -11.2% runtime) ==== 64 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=45.9MiB/s (48.1MB/s), 45.9MiB/s-45.9MiB/s (48.1MB/s-48.1MB/s), io=32.0GiB (34.4GB), run=714614-714614msec After patch: WRITE: bw=48.8MiB/s (51.1MB/s), 48.8MiB/s-48.8MiB/s (51.1MB/s-51.1MB/s), io=32.0GiB (34.4GB), run=672087-672087msec (+6.3% throughput, -6.0% runtime) Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-08-11 14:43:58 +03:00
ret = log_extent_csums(trans, inode, log, em, ctx);
Btrfs: fix data corruption after fast fsync and writeback error When we do a fast fsync, we start all ordered operations and then while they're running in parallel we visit the list of modified extent maps and construct their matching file extent items and write them to the log btree. After that, in btrfs_sync_log() we wait for all the ordered operations to finish (via btrfs_wait_logged_extents). The problem with this is that we were completely ignoring errors that can happen in the extent write path, such as -ENOSPC, a temporary -ENOMEM or -EIO errors for example. When such error happens, it means we have parts of the on disk extent that weren't written to, and so we end up logging file extent items that point to these extents that contain garbage/random data - so after a crash/reboot plus log replay, we get our inode's metadata pointing to those extents. This worked in contrast with the full (non-fast) fsync path, where we start all ordered operations, wait for them to finish and then write to the log btree. In this path, after each ordered operation completes we check if it's flagged with an error (BTRFS_ORDERED_IOERR) and return -EIO if so (via btrfs_wait_ordered_range). So if an error happens with any ordered operation, just return a -EIO error to userspace, so that it knows that not all of its previous writes were durably persisted and the application can take proper action (like redo the writes for e.g.) - and definitely not leave any file extent items in the log refer to non fully written extents. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2014-09-05 18:14:39 +04:00
if (ret)
return ret;
btrfs: avoid attempt to drop extents when logging inode for the first time When logging an extent, in the fast fsync path, we always attempt do drop or trim any existing extents with a range that match or overlap the range of the extent we are about to log. We do that through a call to btrfs_drop_extents(). However this is not needed when we are logging the inode for the first time in the current transaction, since we have no inode items of the inode in the log tree. Calling btrfs_drop_extents() does a deletion search on the log tree, which is expensive when we have concurrent tasks accessing the log tree because a deletion search always acquires a write lock on the extent buffers at levels 2, 1 and 0, adding significant lock contention, specially taking into account the height of a log tree rarely (if ever) goes beyond 2 or 3, due to its short life. So skip the call to btrfs_drop_extents() when the inode was not previously logged in the current transaction. This patch is part of a patch set comprised of the following patches: btrfs: check if a log tree exists at inode_logged() btrfs: remove no longer needed checks for NULL log context btrfs: do not log new dentries when logging that a new name exists btrfs: always update the logged transaction when logging new names btrfs: avoid expensive search when dropping inode items from log btrfs: add helper to truncate inode items when logging inode btrfs: avoid expensive search when truncating inode items from the log btrfs: avoid search for logged i_size when logging inode if possible btrfs: avoid attempt to drop extents when logging inode for the first time btrfs: do not commit delayed inode when logging a file in full sync mode This is patch 9/10 and test results are listed in the change log of the last patch in the set. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-08-31 17:30:39 +03:00
/*
* If this is the first time we are logging the inode in the current
* transaction, we can avoid btrfs_drop_extents(), which is expensive
* because it does a deletion search, which always acquires write locks
* for extent buffers at levels 2, 1 and 0. This not only wastes time
* but also adds significant contention in a log tree, since log trees
* are small, with a root at level 2 or 3 at most, due to their short
* life span.
*/
btrfs: avoid inode logging during rename and link when possible During a rename or link operation, we need to determine if an inode was previously logged or not, and if it was, do some update to the logged inode. We used to rely exclusively on the logged_trans field of struct btrfs_inode to determine that, but that was not reliable because the value of that field is not persisted in the inode item, so it's lost when an inode is evicted and loaded back again. That led to several issues in the past, such as not persisting deletions (such as the case fixed by commit 803f0f64d17769 ("Btrfs: fix fsync not persisting dentry deletions due to inode evictions")), or resulting in losing a file after an inode eviction followed by a rename (commit ecc64fab7d49c6 ("btrfs: fix lost inode on log replay after mix of fsync, rename and inode eviction")), besides other issues. So the inode_logged() helper was introduced and used to determine if an inode was possibly logged before in the current transaction, with the caveat that it could return false positives, in the sense that even if an inode was not logged before in the current transaction, it could still return true, but never to return false in case the inode was logged. >From a functional point of view that is fine, but from a performance perspective it can introduce significant latencies to rename and link operations, as they will end up doing inode logging even when it is not necessary. Recently on a 5.15 kernel, an openSUSE Tumbleweed user reported package installations and upgrades, with the zypper tool, were often taking a long time to complete. With strace it could be observed that zypper was spending about 99% of its time on rename operations, and then with further analysis we checked that directory logging was happening too frequently. Taking into account that installation/upgrade of some of the packages needed a few thousand file renames, the slowdown was very noticeable for the user. The issue was caused indirectly due to an excessive number of inode evictions on a 5.15 kernel, about 100x more compared to a 5.13, 5.14 or a 5.16-rc8 kernel. While triggering the inode evictions if something outside btrfs' control, btrfs could still behave better by eliminating the false positives from the inode_logged() helper. So change inode_logged() to actually eliminate such false positives caused by inode eviction and when an inode was never logged since the filesystem was mounted, as both cases relate to when the logged_trans field of struct btrfs_inode has a value of zero. When it can not determine if the inode was logged based only on the logged_trans value, lookup for the existence of the inode item in the log tree - if it's there then we known the inode was logged, if it's not there then it can not have been logged in the current transaction. Once we determine if the inode was logged, update the logged_trans value to avoid future calls to have to search in the log tree again. Alternatively, we could start storing logged_trans in the on disk inode item structure (struct btrfs_inode_item) in the unused space it still has, but that would be a bit odd because: 1) We only care about logged_trans since the filesystem was mounted, we don't care about its value from a previous mount. Having it persisted in the inode item structure would not make the best use of the precious unused space; 2) In order to get logged_trans persisted before inode eviction, we would have to update the delayed inode when we finish logging the inode and update its logged_trans in struct btrfs_inode, which makes it a bit cumbersome since we need to check if the delayed inode exists, if not create it and populate it and deal with any errors (-ENOMEM mostly). This change is part of a patchset comprised of the following patches: 1/5 btrfs: add helper to delete a dir entry from a log tree 2/5 btrfs: pass the dentry to btrfs_log_new_name() instead of the inode 3/5 btrfs: avoid logging all directory changes during renames 4/5 btrfs: stop doing unnecessary log updates during a rename 5/5 btrfs: avoid inode logging during rename and link when possible The following test script mimics part of what the zypper tool does during package installations/upgrades. It does not triggers inode evictions, but it's similar because it triggers false positives from the inode_logged() helper, because the inodes have a logged_trans of 0, there's a log tree due to a fsync of an unrelated file and the directory inode has its last_trans field set to the current transaction: $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=10000 mkfs.btrfs -f $DEV mount $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync # Now do some change to an unrelated file and fsync it. # This is just to create a log tree to make sure that inode_logged() # does not return false when called against "testdir". xfs_io -f -c "pwrite 0 4K" -c "fsync" $MNT/foo # Do some change to testdir. This is to make sure inode_logged() # will return true when called against "testdir", because its # logged_trans is 0, it was changed in the current transaction # and there's a log tree. echo -n > $MNT/testdir/file_$((NUM_FILES + 1)) echo "Renaming $NUM_FILES files..." start=$(date +%s%N) for ((i = 1; i <= $NUM_FILES; i++)); do mv $MNT/testdir/file_$i $MNT/testdir/file_$i-RPMDELETE done end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "Renames took $dur milliseconds" umount $MNT Testing this change on a box using a non-debug kernel (Debian's default kernel config) gave the following results: NUM_FILES=10000, before patchset: 27837 ms NUM_FILES=10000, after patches 1/5 to 4/5 applied: 9236 ms (-66.8%) NUM_FILES=10000, after whole patchset applied: 8902 ms (-68.0%) NUM_FILES=5000, before patchset: 9127 ms NUM_FILES=5000, after patches 1/5 to 4/5 applied: 4640 ms (-49.2%) NUM_FILES=5000, after whole patchset applied: 4441 ms (-51.3%) NUM_FILES=2000, before patchset: 2528 ms NUM_FILES=2000, after patches 1/5 to 4/5 applied: 1983 ms (-21.6%) NUM_FILES=2000, after whole patchset applied: 1747 ms (-30.9%) NUM_FILES=1000, before patchset: 1085 ms NUM_FILES=1000, after patches 1/5 to 4/5 applied: 893 ms (-17.7%) NUM_FILES=1000, after whole patchset applied: 867 ms (-20.1%) Running dbench on the same physical machine with the following script: $ cat run-dbench.sh #!/bin/bash NUM_JOBS=$(nproc --all) DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 MOUNT_OPTIONS="-o ssd" MKFS_OPTIONS="-O no-holes -R free-space-tree" echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -D $MNT -t 120 $NUM_JOBS umount $MNT Before patchset: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 3761352 0.032 143.843 Close 2762770 0.002 2.273 Rename 159304 0.291 67.037 Unlink 759784 0.207 143.998 Deltree 72 4.028 15.977 Mkdir 36 0.003 0.006 Qpathinfo 3409780 0.013 9.678 Qfileinfo 596772 0.001 0.878 Qfsinfo 625189 0.003 1.245 Sfileinfo 306443 0.006 1.840 Find 1318106 0.063 19.798 WriteX 1871137 0.021 8.532 ReadX 5897325 0.003 3.567 LockX 12252 0.003 0.258 UnlockX 12252 0.002 0.100 Flush 263666 3.327 155.632 Throughput 980.047 MB/sec 12 clients 12 procs max_latency=155.636 ms After whole patchset applied: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 4195584 0.033 107.742 Close 3081932 0.002 1.935 Rename 177641 0.218 14.905 Unlink 847333 0.166 107.822 Deltree 118 5.315 15.247 Mkdir 59 0.004 0.048 Qpathinfo 3802612 0.014 10.302 Qfileinfo 666748 0.001 1.034 Qfsinfo 697329 0.003 0.944 Sfileinfo 341712 0.006 2.099 Find 1470365 0.065 9.359 WriteX 2093921 0.021 8.087 ReadX 6576234 0.003 3.407 LockX 13660 0.003 0.308 UnlockX 13660 0.002 0.114 Flush 294090 2.906 115.539 Throughput 1093.11 MB/sec 12 clients 12 procs max_latency=115.544 ms +11.5% throughput -25.8% max latency rename max latency -77.8% Link: https://bugzilla.opensuse.org/show_bug.cgi?id=1193549 Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-01-20 14:00:10 +03:00
if (ctx->logged_before) {
btrfs: avoid attempt to drop extents when logging inode for the first time When logging an extent, in the fast fsync path, we always attempt do drop or trim any existing extents with a range that match or overlap the range of the extent we are about to log. We do that through a call to btrfs_drop_extents(). However this is not needed when we are logging the inode for the first time in the current transaction, since we have no inode items of the inode in the log tree. Calling btrfs_drop_extents() does a deletion search on the log tree, which is expensive when we have concurrent tasks accessing the log tree because a deletion search always acquires a write lock on the extent buffers at levels 2, 1 and 0, adding significant lock contention, specially taking into account the height of a log tree rarely (if ever) goes beyond 2 or 3, due to its short life. So skip the call to btrfs_drop_extents() when the inode was not previously logged in the current transaction. This patch is part of a patch set comprised of the following patches: btrfs: check if a log tree exists at inode_logged() btrfs: remove no longer needed checks for NULL log context btrfs: do not log new dentries when logging that a new name exists btrfs: always update the logged transaction when logging new names btrfs: avoid expensive search when dropping inode items from log btrfs: add helper to truncate inode items when logging inode btrfs: avoid expensive search when truncating inode items from the log btrfs: avoid search for logged i_size when logging inode if possible btrfs: avoid attempt to drop extents when logging inode for the first time btrfs: do not commit delayed inode when logging a file in full sync mode This is patch 9/10 and test results are listed in the change log of the last patch in the set. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-08-31 17:30:39 +03:00
drop_args.path = path;
drop_args.start = em->start;
drop_args.end = em->start + em->len;
drop_args.replace_extent = true;
drop_args.extent_item_size = sizeof(*fi);
ret = btrfs_drop_extents(trans, log, inode, &drop_args);
if (ret)
return ret;
}
Btrfs: fix data corruption after fast fsync and writeback error When we do a fast fsync, we start all ordered operations and then while they're running in parallel we visit the list of modified extent maps and construct their matching file extent items and write them to the log btree. After that, in btrfs_sync_log() we wait for all the ordered operations to finish (via btrfs_wait_logged_extents). The problem with this is that we were completely ignoring errors that can happen in the extent write path, such as -ENOSPC, a temporary -ENOMEM or -EIO errors for example. When such error happens, it means we have parts of the on disk extent that weren't written to, and so we end up logging file extent items that point to these extents that contain garbage/random data - so after a crash/reboot plus log replay, we get our inode's metadata pointing to those extents. This worked in contrast with the full (non-fast) fsync path, where we start all ordered operations, wait for them to finish and then write to the log btree. In this path, after each ordered operation completes we check if it's flagged with an error (BTRFS_ORDERED_IOERR) and return -EIO if so (via btrfs_wait_ordered_range). So if an error happens with any ordered operation, just return a -EIO error to userspace, so that it knows that not all of its previous writes were durably persisted and the application can take proper action (like redo the writes for e.g.) - and definitely not leave any file extent items in the log refer to non fully written extents. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2014-09-05 18:14:39 +04:00
if (!drop_args.extent_inserted) {
key.objectid = btrfs_ino(inode);
Btrfs: fix data corruption after fast fsync and writeback error When we do a fast fsync, we start all ordered operations and then while they're running in parallel we visit the list of modified extent maps and construct their matching file extent items and write them to the log btree. After that, in btrfs_sync_log() we wait for all the ordered operations to finish (via btrfs_wait_logged_extents). The problem with this is that we were completely ignoring errors that can happen in the extent write path, such as -ENOSPC, a temporary -ENOMEM or -EIO errors for example. When such error happens, it means we have parts of the on disk extent that weren't written to, and so we end up logging file extent items that point to these extents that contain garbage/random data - so after a crash/reboot plus log replay, we get our inode's metadata pointing to those extents. This worked in contrast with the full (non-fast) fsync path, where we start all ordered operations, wait for them to finish and then write to the log btree. In this path, after each ordered operation completes we check if it's flagged with an error (BTRFS_ORDERED_IOERR) and return -EIO if so (via btrfs_wait_ordered_range). So if an error happens with any ordered operation, just return a -EIO error to userspace, so that it knows that not all of its previous writes were durably persisted and the application can take proper action (like redo the writes for e.g.) - and definitely not leave any file extent items in the log refer to non fully written extents. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2014-09-05 18:14:39 +04:00
key.type = BTRFS_EXTENT_DATA_KEY;
key.offset = em->start;
ret = btrfs_insert_empty_item(trans, log, path, &key,
sizeof(*fi));
if (ret)
return ret;
}
leaf = path->nodes[0];
btrfs_init_map_token(&token, leaf);
Btrfs: fix data corruption after fast fsync and writeback error When we do a fast fsync, we start all ordered operations and then while they're running in parallel we visit the list of modified extent maps and construct their matching file extent items and write them to the log btree. After that, in btrfs_sync_log() we wait for all the ordered operations to finish (via btrfs_wait_logged_extents). The problem with this is that we were completely ignoring errors that can happen in the extent write path, such as -ENOSPC, a temporary -ENOMEM or -EIO errors for example. When such error happens, it means we have parts of the on disk extent that weren't written to, and so we end up logging file extent items that point to these extents that contain garbage/random data - so after a crash/reboot plus log replay, we get our inode's metadata pointing to those extents. This worked in contrast with the full (non-fast) fsync path, where we start all ordered operations, wait for them to finish and then write to the log btree. In this path, after each ordered operation completes we check if it's flagged with an error (BTRFS_ORDERED_IOERR) and return -EIO if so (via btrfs_wait_ordered_range). So if an error happens with any ordered operation, just return a -EIO error to userspace, so that it knows that not all of its previous writes were durably persisted and the application can take proper action (like redo the writes for e.g.) - and definitely not leave any file extent items in the log refer to non fully written extents. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2014-09-05 18:14:39 +04:00
fi = btrfs_item_ptr(leaf, path->slots[0],
struct btrfs_file_extent_item);
btrfs_set_token_file_extent_generation(&token, fi, trans->transid);
Btrfs: fix data corruption after fast fsync and writeback error When we do a fast fsync, we start all ordered operations and then while they're running in parallel we visit the list of modified extent maps and construct their matching file extent items and write them to the log btree. After that, in btrfs_sync_log() we wait for all the ordered operations to finish (via btrfs_wait_logged_extents). The problem with this is that we were completely ignoring errors that can happen in the extent write path, such as -ENOSPC, a temporary -ENOMEM or -EIO errors for example. When such error happens, it means we have parts of the on disk extent that weren't written to, and so we end up logging file extent items that point to these extents that contain garbage/random data - so after a crash/reboot plus log replay, we get our inode's metadata pointing to those extents. This worked in contrast with the full (non-fast) fsync path, where we start all ordered operations, wait for them to finish and then write to the log btree. In this path, after each ordered operation completes we check if it's flagged with an error (BTRFS_ORDERED_IOERR) and return -EIO if so (via btrfs_wait_ordered_range). So if an error happens with any ordered operation, just return a -EIO error to userspace, so that it knows that not all of its previous writes were durably persisted and the application can take proper action (like redo the writes for e.g.) - and definitely not leave any file extent items in the log refer to non fully written extents. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2014-09-05 18:14:39 +04:00
if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
btrfs_set_token_file_extent_type(&token, fi,
BTRFS_FILE_EXTENT_PREALLOC);
Btrfs: fix data corruption after fast fsync and writeback error When we do a fast fsync, we start all ordered operations and then while they're running in parallel we visit the list of modified extent maps and construct their matching file extent items and write them to the log btree. After that, in btrfs_sync_log() we wait for all the ordered operations to finish (via btrfs_wait_logged_extents). The problem with this is that we were completely ignoring errors that can happen in the extent write path, such as -ENOSPC, a temporary -ENOMEM or -EIO errors for example. When such error happens, it means we have parts of the on disk extent that weren't written to, and so we end up logging file extent items that point to these extents that contain garbage/random data - so after a crash/reboot plus log replay, we get our inode's metadata pointing to those extents. This worked in contrast with the full (non-fast) fsync path, where we start all ordered operations, wait for them to finish and then write to the log btree. In this path, after each ordered operation completes we check if it's flagged with an error (BTRFS_ORDERED_IOERR) and return -EIO if so (via btrfs_wait_ordered_range). So if an error happens with any ordered operation, just return a -EIO error to userspace, so that it knows that not all of its previous writes were durably persisted and the application can take proper action (like redo the writes for e.g.) - and definitely not leave any file extent items in the log refer to non fully written extents. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2014-09-05 18:14:39 +04:00
else
btrfs_set_token_file_extent_type(&token, fi,
BTRFS_FILE_EXTENT_REG);
Btrfs: fix data corruption after fast fsync and writeback error When we do a fast fsync, we start all ordered operations and then while they're running in parallel we visit the list of modified extent maps and construct their matching file extent items and write them to the log btree. After that, in btrfs_sync_log() we wait for all the ordered operations to finish (via btrfs_wait_logged_extents). The problem with this is that we were completely ignoring errors that can happen in the extent write path, such as -ENOSPC, a temporary -ENOMEM or -EIO errors for example. When such error happens, it means we have parts of the on disk extent that weren't written to, and so we end up logging file extent items that point to these extents that contain garbage/random data - so after a crash/reboot plus log replay, we get our inode's metadata pointing to those extents. This worked in contrast with the full (non-fast) fsync path, where we start all ordered operations, wait for them to finish and then write to the log btree. In this path, after each ordered operation completes we check if it's flagged with an error (BTRFS_ORDERED_IOERR) and return -EIO if so (via btrfs_wait_ordered_range). So if an error happens with any ordered operation, just return a -EIO error to userspace, so that it knows that not all of its previous writes were durably persisted and the application can take proper action (like redo the writes for e.g.) - and definitely not leave any file extent items in the log refer to non fully written extents. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2014-09-05 18:14:39 +04:00
block_len = max(em->block_len, em->orig_block_len);
if (em->compress_type != BTRFS_COMPRESS_NONE) {
btrfs_set_token_file_extent_disk_bytenr(&token, fi,
em->block_start);
btrfs_set_token_file_extent_disk_num_bytes(&token, fi, block_len);
Btrfs: fix data corruption after fast fsync and writeback error When we do a fast fsync, we start all ordered operations and then while they're running in parallel we visit the list of modified extent maps and construct their matching file extent items and write them to the log btree. After that, in btrfs_sync_log() we wait for all the ordered operations to finish (via btrfs_wait_logged_extents). The problem with this is that we were completely ignoring errors that can happen in the extent write path, such as -ENOSPC, a temporary -ENOMEM or -EIO errors for example. When such error happens, it means we have parts of the on disk extent that weren't written to, and so we end up logging file extent items that point to these extents that contain garbage/random data - so after a crash/reboot plus log replay, we get our inode's metadata pointing to those extents. This worked in contrast with the full (non-fast) fsync path, where we start all ordered operations, wait for them to finish and then write to the log btree. In this path, after each ordered operation completes we check if it's flagged with an error (BTRFS_ORDERED_IOERR) and return -EIO if so (via btrfs_wait_ordered_range). So if an error happens with any ordered operation, just return a -EIO error to userspace, so that it knows that not all of its previous writes were durably persisted and the application can take proper action (like redo the writes for e.g.) - and definitely not leave any file extent items in the log refer to non fully written extents. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2014-09-05 18:14:39 +04:00
} else if (em->block_start < EXTENT_MAP_LAST_BYTE) {
btrfs_set_token_file_extent_disk_bytenr(&token, fi,
Btrfs: fix data corruption after fast fsync and writeback error When we do a fast fsync, we start all ordered operations and then while they're running in parallel we visit the list of modified extent maps and construct their matching file extent items and write them to the log btree. After that, in btrfs_sync_log() we wait for all the ordered operations to finish (via btrfs_wait_logged_extents). The problem with this is that we were completely ignoring errors that can happen in the extent write path, such as -ENOSPC, a temporary -ENOMEM or -EIO errors for example. When such error happens, it means we have parts of the on disk extent that weren't written to, and so we end up logging file extent items that point to these extents that contain garbage/random data - so after a crash/reboot plus log replay, we get our inode's metadata pointing to those extents. This worked in contrast with the full (non-fast) fsync path, where we start all ordered operations, wait for them to finish and then write to the log btree. In this path, after each ordered operation completes we check if it's flagged with an error (BTRFS_ORDERED_IOERR) and return -EIO if so (via btrfs_wait_ordered_range). So if an error happens with any ordered operation, just return a -EIO error to userspace, so that it knows that not all of its previous writes were durably persisted and the application can take proper action (like redo the writes for e.g.) - and definitely not leave any file extent items in the log refer to non fully written extents. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2014-09-05 18:14:39 +04:00
em->block_start -
extent_offset);
btrfs_set_token_file_extent_disk_num_bytes(&token, fi, block_len);
Btrfs: fix data corruption after fast fsync and writeback error When we do a fast fsync, we start all ordered operations and then while they're running in parallel we visit the list of modified extent maps and construct their matching file extent items and write them to the log btree. After that, in btrfs_sync_log() we wait for all the ordered operations to finish (via btrfs_wait_logged_extents). The problem with this is that we were completely ignoring errors that can happen in the extent write path, such as -ENOSPC, a temporary -ENOMEM or -EIO errors for example. When such error happens, it means we have parts of the on disk extent that weren't written to, and so we end up logging file extent items that point to these extents that contain garbage/random data - so after a crash/reboot plus log replay, we get our inode's metadata pointing to those extents. This worked in contrast with the full (non-fast) fsync path, where we start all ordered operations, wait for them to finish and then write to the log btree. In this path, after each ordered operation completes we check if it's flagged with an error (BTRFS_ORDERED_IOERR) and return -EIO if so (via btrfs_wait_ordered_range). So if an error happens with any ordered operation, just return a -EIO error to userspace, so that it knows that not all of its previous writes were durably persisted and the application can take proper action (like redo the writes for e.g.) - and definitely not leave any file extent items in the log refer to non fully written extents. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2014-09-05 18:14:39 +04:00
} else {
btrfs_set_token_file_extent_disk_bytenr(&token, fi, 0);
btrfs_set_token_file_extent_disk_num_bytes(&token, fi, 0);
Btrfs: fix data corruption after fast fsync and writeback error When we do a fast fsync, we start all ordered operations and then while they're running in parallel we visit the list of modified extent maps and construct their matching file extent items and write them to the log btree. After that, in btrfs_sync_log() we wait for all the ordered operations to finish (via btrfs_wait_logged_extents). The problem with this is that we were completely ignoring errors that can happen in the extent write path, such as -ENOSPC, a temporary -ENOMEM or -EIO errors for example. When such error happens, it means we have parts of the on disk extent that weren't written to, and so we end up logging file extent items that point to these extents that contain garbage/random data - so after a crash/reboot plus log replay, we get our inode's metadata pointing to those extents. This worked in contrast with the full (non-fast) fsync path, where we start all ordered operations, wait for them to finish and then write to the log btree. In this path, after each ordered operation completes we check if it's flagged with an error (BTRFS_ORDERED_IOERR) and return -EIO if so (via btrfs_wait_ordered_range). So if an error happens with any ordered operation, just return a -EIO error to userspace, so that it knows that not all of its previous writes were durably persisted and the application can take proper action (like redo the writes for e.g.) - and definitely not leave any file extent items in the log refer to non fully written extents. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2014-09-05 18:14:39 +04:00
}
btrfs_set_token_file_extent_offset(&token, fi, extent_offset);
btrfs_set_token_file_extent_num_bytes(&token, fi, em->len);
btrfs_set_token_file_extent_ram_bytes(&token, fi, em->ram_bytes);
btrfs_set_token_file_extent_compression(&token, fi, em->compress_type);
btrfs_set_token_file_extent_encryption(&token, fi, 0);
btrfs_set_token_file_extent_other_encoding(&token, fi, 0);
Btrfs: fix data corruption after fast fsync and writeback error When we do a fast fsync, we start all ordered operations and then while they're running in parallel we visit the list of modified extent maps and construct their matching file extent items and write them to the log btree. After that, in btrfs_sync_log() we wait for all the ordered operations to finish (via btrfs_wait_logged_extents). The problem with this is that we were completely ignoring errors that can happen in the extent write path, such as -ENOSPC, a temporary -ENOMEM or -EIO errors for example. When such error happens, it means we have parts of the on disk extent that weren't written to, and so we end up logging file extent items that point to these extents that contain garbage/random data - so after a crash/reboot plus log replay, we get our inode's metadata pointing to those extents. This worked in contrast with the full (non-fast) fsync path, where we start all ordered operations, wait for them to finish and then write to the log btree. In this path, after each ordered operation completes we check if it's flagged with an error (BTRFS_ORDERED_IOERR) and return -EIO if so (via btrfs_wait_ordered_range). So if an error happens with any ordered operation, just return a -EIO error to userspace, so that it knows that not all of its previous writes were durably persisted and the application can take proper action (like redo the writes for e.g.) - and definitely not leave any file extent items in the log refer to non fully written extents. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2014-09-05 18:14:39 +04:00
btrfs_mark_buffer_dirty(leaf);
btrfs_release_path(path);
return ret;
}
Btrfs: fix duplicate extents after fsync of file with prealloc extents In commit 471d557afed1 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay"), on fsync, we started to always log all prealloc extents beyond an inode's i_size in order to avoid losing them after a power failure. However under some cases this can lead to the log replay code to create duplicate extent items, with different lengths, in the extent tree. That happens because, as of that commit, we can now log extent items based on extent maps that are not on the "modified" list of extent maps of the inode's extent map tree. Logging extent items based on extent maps is used during the fast fsync path to save time and for this to work reliably it requires that the extent maps are not merged with other adjacent extent maps - having the extent maps in the list of modified extents gives such guarantee. Consider the following example, captured during a long run of fsstress, which illustrates this problem. We have inode 271, in the filesystem tree (root 5), for which all of the following operations and discussion apply to. A buffered write starts at offset 312391 with a length of 933471 bytes (end offset at 1245862). At this point we have, for this inode, the following extent maps with the their field values: em A, start 0, orig_start 0, len 40960, block_start 18446744073709551613, block_len 0, orig_block_len 0 em B, start 40960, orig_start 40960, len 376832, block_start 1106399232, block_len 376832, orig_block_len 376832 em C, start 417792, orig_start 417792, len 782336, block_start 18446744073709551613, block_len 0, orig_block_len 0 em D, start 1200128, orig_start 1200128, len 835584, block_start 1106776064, block_len 835584, orig_block_len 835584 em E, start 2035712, orig_start 2035712, len 245760, block_start 1107611648, block_len 245760, orig_block_len 245760 Extent map A corresponds to a hole and extent maps D and E correspond to preallocated extents. Extent map D ends where extent map E begins (1106776064 + 835584 = 1107611648), but these extent maps were not merged because they are in the inode's list of modified extent maps. An fsync against this inode is made, which triggers the fast path (BTRFS_INODE_NEEDS_FULL_SYNC is not set). This fsync triggers writeback of the data previously written using buffered IO, and when the respective ordered extent finishes, btrfs_drop_extents() is called against the (aligned) range 311296..1249279. This causes a split of extent map D at btrfs_drop_extent_cache(), replacing extent map D with a new extent map D', also added to the list of modified extents, with the following values: em D', start 1249280, orig_start of 1200128, block_start 1106825216 (= 1106776064 + 1249280 - 1200128), orig_block_len 835584, block_len 786432 (835584 - (1249280 - 1200128)) Then, during the fast fsync, btrfs_log_changed_extents() is called and extent maps D' and E are removed from the list of modified extents. The flag EXTENT_FLAG_LOGGING is also set on them. After the extents are logged clear_em_logging() is called on each of them, and that makes extent map E to be merged with extent map D' (try_merge_map()), resulting in D' being deleted and E adjusted to: em E, start 1249280, orig_start 1200128, len 1032192, block_start 1106825216, block_len 1032192, orig_block_len 245760 A direct IO write at offset 1847296 and length of 360448 bytes (end offset at 2207744) starts, and at that moment the following extent maps exist for our inode: em A, start 0, orig_start 0, len 40960, block_start 18446744073709551613, block_len 0, orig_block_len 0 em B, start 40960, orig_start 40960, len 270336, block_start 1106399232, block_len 270336, orig_block_len 376832 em C, start 311296, orig_start 311296, len 937984, block_start 1112842240, block_len 937984, orig_block_len 937984 em E (prealloc), start 1249280, orig_start 1200128, len 1032192, block_start 1106825216, block_len 1032192, orig_block_len 245760 The dio write results in drop_extent_cache() being called twice. The first time for a range that starts at offset 1847296 and ends at offset 2035711 (length of 188416), which results in a double split of extent map E, replacing it with two new extent maps: em F, start 1249280, orig_start 1200128, block_start 1106825216, block_len 598016, orig_block_len 598016 em G, start 2035712, orig_start 1200128, block_start 1107611648, block_len 245760, orig_block_len 1032192 It also creates a new extent map that represents a part of the requested IO (through create_io_em()): em H, start 1847296, len 188416, block_start 1107423232, block_len 188416 The second call to drop_extent_cache() has a range with a start offset of 2035712 and end offset of 2207743 (length of 172032). This leads to replacing extent map G with a new extent map I with the following values: em I, start 2207744, orig_start 1200128, block_start 1107783680, block_len 73728, orig_block_len 1032192 It also creates a new extent map that represents the second part of the requested IO (through create_io_em()): em J, start 2035712, len 172032, block_start 1107611648, block_len 172032 The dio write set the inode's i_size to 2207744 bytes. After the dio write the inode has the following extent maps: em A, start 0, orig_start 0, len 40960, block_start 18446744073709551613, block_len 0, orig_block_len 0 em B, start 40960, orig_start 40960, len 270336, block_start 1106399232, block_len 270336, orig_block_len 376832 em C, start 311296, orig_start 311296, len 937984, block_start 1112842240, block_len 937984, orig_block_len 937984 em F, start 1249280, orig_start 1200128, len 598016, block_start 1106825216, block_len 598016, orig_block_len 598016 em H, start 1847296, orig_start 1200128, len 188416, block_start 1107423232, block_len 188416, orig_block_len 835584 em J, start 2035712, orig_start 2035712, len 172032, block_start 1107611648, block_len 172032, orig_block_len 245760 em I, start 2207744, orig_start 1200128, len 73728, block_start 1107783680, block_len 73728, orig_block_len 1032192 Now do some change to the file, like adding a xattr for example and then fsync it again. This triggers a fast fsync path, and as of commit 471d557afed1 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay"), we use the extent map I to log a file extent item because it's a prealloc extent and it starts at an offset matching the inode's i_size. However when we log it, we create a file extent item with a value for the disk byte location that is wrong, as can be seen from the following output of "btrfs inspect-internal dump-tree": item 1 key (271 EXTENT_DATA 2207744) itemoff 3782 itemsize 53 generation 22 type 2 (prealloc) prealloc data disk byte 1106776064 nr 1032192 prealloc data offset 1007616 nr 73728 Here the disk byte value corresponds to calculation based on some fields from the extent map I: 1106776064 = block_start (1107783680) - 1007616 (extent_offset) extent_offset = 2207744 (start) - 1200128 (orig_start) = 1007616 The disk byte value of 1106776064 clashes with disk byte values of the file extent items at offsets 1249280 and 1847296 in the fs tree: item 6 key (271 EXTENT_DATA 1249280) itemoff 3568 itemsize 53 generation 20 type 2 (prealloc) prealloc data disk byte 1106776064 nr 835584 prealloc data offset 49152 nr 598016 item 7 key (271 EXTENT_DATA 1847296) itemoff 3515 itemsize 53 generation 20 type 1 (regular) extent data disk byte 1106776064 nr 835584 extent data offset 647168 nr 188416 ram 835584 extent compression 0 (none) item 8 key (271 EXTENT_DATA 2035712) itemoff 3462 itemsize 53 generation 20 type 1 (regular) extent data disk byte 1107611648 nr 245760 extent data offset 0 nr 172032 ram 245760 extent compression 0 (none) item 9 key (271 EXTENT_DATA 2207744) itemoff 3409 itemsize 53 generation 20 type 2 (prealloc) prealloc data disk byte 1107611648 nr 245760 prealloc data offset 172032 nr 73728 Instead of the disk byte value of 1106776064, the value of 1107611648 should have been logged. Also the data offset value should have been 172032 and not 1007616. After a log replay we end up getting two extent items in the extent tree with different lengths, one of 835584, which is correct and existed before the log replay, and another one of 1032192 which is wrong and is based on the logged file extent item: item 12 key (1106776064 EXTENT_ITEM 835584) itemoff 3406 itemsize 53 refs 2 gen 15 flags DATA extent data backref root 5 objectid 271 offset 1200128 count 2 item 13 key (1106776064 EXTENT_ITEM 1032192) itemoff 3353 itemsize 53 refs 1 gen 22 flags DATA extent data backref root 5 objectid 271 offset 1200128 count 1 Obviously this leads to many problems and a filesystem check reports many errors: (...) checking extents Extent back ref already exists for 1106776064 parent 0 root 5 owner 271 offset 1200128 num_refs 1 extent item 1106776064 has multiple extent items ref mismatch on [1106776064 835584] extent item 2, found 3 Incorrect local backref count on 1106776064 root 5 owner 271 offset 1200128 found 2 wanted 1 back 0x55b1d0ad7680 Backref 1106776064 root 5 owner 271 offset 1200128 num_refs 0 not found in extent tree Incorrect local backref count on 1106776064 root 5 owner 271 offset 1200128 found 1 wanted 0 back 0x55b1d0ad4e70 Backref bytes do not match extent backref, bytenr=1106776064, ref bytes=835584, backref bytes=1032192 backpointer mismatch on [1106776064 835584] checking free space cache block group 1103101952 has wrong amount of free space failed to load free space cache for block group 1103101952 checking fs roots (...) So fix this by logging the prealloc extents beyond the inode's i_size based on searches in the subvolume tree instead of the extent maps. Fixes: 471d557afed1 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay") CC: stable@vger.kernel.org # 4.14+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2018-05-09 18:01:46 +03:00
/*
* Log all prealloc extents beyond the inode's i_size to make sure we do not
btrfs: fix lost prealloc extents beyond eof after full fsync When doing a full fsync, if we have prealloc extents beyond (or at) eof, and the leaves that contain them were not modified in the current transaction, we end up not logging them. This results in losing those extents when we replay the log after a power failure, since the inode is truncated to the current value of the logged i_size. Just like for the fast fsync path, we need to always log all prealloc extents starting at or beyond i_size. The fast fsync case was fixed in commit 471d557afed155 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay") but it missed the full fsync path. The problem exists since the very early days, when the log tree was added by commit e02119d5a7b439 ("Btrfs: Add a write ahead tree log to optimize synchronous operations"). Example reproducer: $ mkfs.btrfs -f /dev/sdc $ mount /dev/sdc /mnt # Create our test file with many file extent items, so that they span # several leaves of metadata, even if the node/page size is 64K. Use # direct IO and not fsync/O_SYNC because it's both faster and it avoids # clearing the full sync flag from the inode - we want the fsync below # to trigger the slow full sync code path. $ xfs_io -f -d -c "pwrite -b 4K 0 16M" /mnt/foo # Now add two preallocated extents to our file without extending the # file's size. One right at i_size, and another further beyond, leaving # a gap between the two prealloc extents. $ xfs_io -c "falloc -k 16M 1M" /mnt/foo $ xfs_io -c "falloc -k 20M 1M" /mnt/foo # Make sure everything is durably persisted and the transaction is # committed. This makes all created extents to have a generation lower # than the generation of the transaction used by the next write and # fsync. sync # Now overwrite only the first extent, which will result in modifying # only the first leaf of metadata for our inode. Then fsync it. This # fsync will use the slow code path (inode full sync bit is set) because # it's the first fsync since the inode was created/loaded. $ xfs_io -c "pwrite 0 4K" -c "fsync" /mnt/foo # Extent list before power failure. $ xfs_io -c "fiemap -v" /mnt/foo /mnt/foo: EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS 0: [0..7]: 2178048..2178055 8 0x0 1: [8..16383]: 26632..43007 16376 0x0 2: [16384..32767]: 2156544..2172927 16384 0x0 3: [32768..34815]: 2172928..2174975 2048 0x800 4: [34816..40959]: hole 6144 5: [40960..43007]: 2174976..2177023 2048 0x801 <power fail> # Mount fs again, trigger log replay. $ mount /dev/sdc /mnt # Extent list after power failure and log replay. $ xfs_io -c "fiemap -v" /mnt/foo /mnt/foo: EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS 0: [0..7]: 2178048..2178055 8 0x0 1: [8..16383]: 26632..43007 16376 0x0 2: [16384..32767]: 2156544..2172927 16384 0x1 # The prealloc extents at file offsets 16M and 20M are missing. So fix this by calling btrfs_log_prealloc_extents() when we are doing a full fsync, so that we always log all prealloc extents beyond eof. A test case for fstests will follow soon. CC: stable@vger.kernel.org # 4.19+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-02-17 15:12:02 +03:00
* lose them after doing a full/fast fsync and replaying the log. We scan the
Btrfs: fix duplicate extents after fsync of file with prealloc extents In commit 471d557afed1 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay"), on fsync, we started to always log all prealloc extents beyond an inode's i_size in order to avoid losing them after a power failure. However under some cases this can lead to the log replay code to create duplicate extent items, with different lengths, in the extent tree. That happens because, as of that commit, we can now log extent items based on extent maps that are not on the "modified" list of extent maps of the inode's extent map tree. Logging extent items based on extent maps is used during the fast fsync path to save time and for this to work reliably it requires that the extent maps are not merged with other adjacent extent maps - having the extent maps in the list of modified extents gives such guarantee. Consider the following example, captured during a long run of fsstress, which illustrates this problem. We have inode 271, in the filesystem tree (root 5), for which all of the following operations and discussion apply to. A buffered write starts at offset 312391 with a length of 933471 bytes (end offset at 1245862). At this point we have, for this inode, the following extent maps with the their field values: em A, start 0, orig_start 0, len 40960, block_start 18446744073709551613, block_len 0, orig_block_len 0 em B, start 40960, orig_start 40960, len 376832, block_start 1106399232, block_len 376832, orig_block_len 376832 em C, start 417792, orig_start 417792, len 782336, block_start 18446744073709551613, block_len 0, orig_block_len 0 em D, start 1200128, orig_start 1200128, len 835584, block_start 1106776064, block_len 835584, orig_block_len 835584 em E, start 2035712, orig_start 2035712, len 245760, block_start 1107611648, block_len 245760, orig_block_len 245760 Extent map A corresponds to a hole and extent maps D and E correspond to preallocated extents. Extent map D ends where extent map E begins (1106776064 + 835584 = 1107611648), but these extent maps were not merged because they are in the inode's list of modified extent maps. An fsync against this inode is made, which triggers the fast path (BTRFS_INODE_NEEDS_FULL_SYNC is not set). This fsync triggers writeback of the data previously written using buffered IO, and when the respective ordered extent finishes, btrfs_drop_extents() is called against the (aligned) range 311296..1249279. This causes a split of extent map D at btrfs_drop_extent_cache(), replacing extent map D with a new extent map D', also added to the list of modified extents, with the following values: em D', start 1249280, orig_start of 1200128, block_start 1106825216 (= 1106776064 + 1249280 - 1200128), orig_block_len 835584, block_len 786432 (835584 - (1249280 - 1200128)) Then, during the fast fsync, btrfs_log_changed_extents() is called and extent maps D' and E are removed from the list of modified extents. The flag EXTENT_FLAG_LOGGING is also set on them. After the extents are logged clear_em_logging() is called on each of them, and that makes extent map E to be merged with extent map D' (try_merge_map()), resulting in D' being deleted and E adjusted to: em E, start 1249280, orig_start 1200128, len 1032192, block_start 1106825216, block_len 1032192, orig_block_len 245760 A direct IO write at offset 1847296 and length of 360448 bytes (end offset at 2207744) starts, and at that moment the following extent maps exist for our inode: em A, start 0, orig_start 0, len 40960, block_start 18446744073709551613, block_len 0, orig_block_len 0 em B, start 40960, orig_start 40960, len 270336, block_start 1106399232, block_len 270336, orig_block_len 376832 em C, start 311296, orig_start 311296, len 937984, block_start 1112842240, block_len 937984, orig_block_len 937984 em E (prealloc), start 1249280, orig_start 1200128, len 1032192, block_start 1106825216, block_len 1032192, orig_block_len 245760 The dio write results in drop_extent_cache() being called twice. The first time for a range that starts at offset 1847296 and ends at offset 2035711 (length of 188416), which results in a double split of extent map E, replacing it with two new extent maps: em F, start 1249280, orig_start 1200128, block_start 1106825216, block_len 598016, orig_block_len 598016 em G, start 2035712, orig_start 1200128, block_start 1107611648, block_len 245760, orig_block_len 1032192 It also creates a new extent map that represents a part of the requested IO (through create_io_em()): em H, start 1847296, len 188416, block_start 1107423232, block_len 188416 The second call to drop_extent_cache() has a range with a start offset of 2035712 and end offset of 2207743 (length of 172032). This leads to replacing extent map G with a new extent map I with the following values: em I, start 2207744, orig_start 1200128, block_start 1107783680, block_len 73728, orig_block_len 1032192 It also creates a new extent map that represents the second part of the requested IO (through create_io_em()): em J, start 2035712, len 172032, block_start 1107611648, block_len 172032 The dio write set the inode's i_size to 2207744 bytes. After the dio write the inode has the following extent maps: em A, start 0, orig_start 0, len 40960, block_start 18446744073709551613, block_len 0, orig_block_len 0 em B, start 40960, orig_start 40960, len 270336, block_start 1106399232, block_len 270336, orig_block_len 376832 em C, start 311296, orig_start 311296, len 937984, block_start 1112842240, block_len 937984, orig_block_len 937984 em F, start 1249280, orig_start 1200128, len 598016, block_start 1106825216, block_len 598016, orig_block_len 598016 em H, start 1847296, orig_start 1200128, len 188416, block_start 1107423232, block_len 188416, orig_block_len 835584 em J, start 2035712, orig_start 2035712, len 172032, block_start 1107611648, block_len 172032, orig_block_len 245760 em I, start 2207744, orig_start 1200128, len 73728, block_start 1107783680, block_len 73728, orig_block_len 1032192 Now do some change to the file, like adding a xattr for example and then fsync it again. This triggers a fast fsync path, and as of commit 471d557afed1 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay"), we use the extent map I to log a file extent item because it's a prealloc extent and it starts at an offset matching the inode's i_size. However when we log it, we create a file extent item with a value for the disk byte location that is wrong, as can be seen from the following output of "btrfs inspect-internal dump-tree": item 1 key (271 EXTENT_DATA 2207744) itemoff 3782 itemsize 53 generation 22 type 2 (prealloc) prealloc data disk byte 1106776064 nr 1032192 prealloc data offset 1007616 nr 73728 Here the disk byte value corresponds to calculation based on some fields from the extent map I: 1106776064 = block_start (1107783680) - 1007616 (extent_offset) extent_offset = 2207744 (start) - 1200128 (orig_start) = 1007616 The disk byte value of 1106776064 clashes with disk byte values of the file extent items at offsets 1249280 and 1847296 in the fs tree: item 6 key (271 EXTENT_DATA 1249280) itemoff 3568 itemsize 53 generation 20 type 2 (prealloc) prealloc data disk byte 1106776064 nr 835584 prealloc data offset 49152 nr 598016 item 7 key (271 EXTENT_DATA 1847296) itemoff 3515 itemsize 53 generation 20 type 1 (regular) extent data disk byte 1106776064 nr 835584 extent data offset 647168 nr 188416 ram 835584 extent compression 0 (none) item 8 key (271 EXTENT_DATA 2035712) itemoff 3462 itemsize 53 generation 20 type 1 (regular) extent data disk byte 1107611648 nr 245760 extent data offset 0 nr 172032 ram 245760 extent compression 0 (none) item 9 key (271 EXTENT_DATA 2207744) itemoff 3409 itemsize 53 generation 20 type 2 (prealloc) prealloc data disk byte 1107611648 nr 245760 prealloc data offset 172032 nr 73728 Instead of the disk byte value of 1106776064, the value of 1107611648 should have been logged. Also the data offset value should have been 172032 and not 1007616. After a log replay we end up getting two extent items in the extent tree with different lengths, one of 835584, which is correct and existed before the log replay, and another one of 1032192 which is wrong and is based on the logged file extent item: item 12 key (1106776064 EXTENT_ITEM 835584) itemoff 3406 itemsize 53 refs 2 gen 15 flags DATA extent data backref root 5 objectid 271 offset 1200128 count 2 item 13 key (1106776064 EXTENT_ITEM 1032192) itemoff 3353 itemsize 53 refs 1 gen 22 flags DATA extent data backref root 5 objectid 271 offset 1200128 count 1 Obviously this leads to many problems and a filesystem check reports many errors: (...) checking extents Extent back ref already exists for 1106776064 parent 0 root 5 owner 271 offset 1200128 num_refs 1 extent item 1106776064 has multiple extent items ref mismatch on [1106776064 835584] extent item 2, found 3 Incorrect local backref count on 1106776064 root 5 owner 271 offset 1200128 found 2 wanted 1 back 0x55b1d0ad7680 Backref 1106776064 root 5 owner 271 offset 1200128 num_refs 0 not found in extent tree Incorrect local backref count on 1106776064 root 5 owner 271 offset 1200128 found 1 wanted 0 back 0x55b1d0ad4e70 Backref bytes do not match extent backref, bytenr=1106776064, ref bytes=835584, backref bytes=1032192 backpointer mismatch on [1106776064 835584] checking free space cache block group 1103101952 has wrong amount of free space failed to load free space cache for block group 1103101952 checking fs roots (...) So fix this by logging the prealloc extents beyond the inode's i_size based on searches in the subvolume tree instead of the extent maps. Fixes: 471d557afed1 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay") CC: stable@vger.kernel.org # 4.14+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2018-05-09 18:01:46 +03:00
* subvolume's root instead of iterating the inode's extent map tree because
* otherwise we can log incorrect extent items based on extent map conversion.
* That can happen due to the fact that extent maps are merged when they
* are not in the extent map tree's list of modified extents.
*/
static int btrfs_log_prealloc_extents(struct btrfs_trans_handle *trans,
struct btrfs_inode *inode,
struct btrfs_path *path)
{
struct btrfs_root *root = inode->root;
struct btrfs_key key;
const u64 i_size = i_size_read(&inode->vfs_inode);
const u64 ino = btrfs_ino(inode);
struct btrfs_path *dst_path = NULL;
Btrfs: fix missing hole after hole punching and fsync when using NO_HOLES When using the NO_HOLES feature, if we punch a hole into a file and then fsync it, there are cases where a subsequent fsync will miss the fact that a hole was punched, resulting in the holes not existing after replaying the log tree. Essentially these cases all imply that, tree-log.c:copy_items(), is not invoked for the leafs that delimit holes, because nothing changed those leafs in the current transaction. And it's precisely copy_items() where we currenly detect and log holes, which works as long as the holes are between file extent items in the input leaf or between the beginning of input leaf and the previous leaf or between the last item in the leaf and the next leaf. First example where we miss a hole: *) The extent items of the inode span multiple leafs; *) The punched hole covers a range that affects only the extent items of the first leaf; *) The fsync operation is done in full mode (BTRFS_INODE_NEEDS_FULL_SYNC is set in the inode's runtime flags). That results in the hole not existing after replaying the log tree. For example, if the fs/subvolume tree has the following layout for a particular inode: Leaf N, generation 10: [ ... INODE_ITEM INODE_REF EXTENT_ITEM (0 64K) EXTENT_ITEM (64K 128K) ] Leaf N + 1, generation 10: [ EXTENT_ITEM (128K 64K) ... ] If at transaction 11 we punch a hole coverting the range [0, 128K[, we end up dropping the two extent items from leaf N, but we don't touch the other leaf, so we end up in the following state: Leaf N, generation 11: [ ... INODE_ITEM INODE_REF ] Leaf N + 1, generation 10: [ EXTENT_ITEM (128K 64K) ... ] A full fsync after punching the hole will only process leaf N because it was modified in the current transaction, but not leaf N + 1, since it was not modified in the current transaction (generation 10 and not 11). As a result the fsync will not log any holes, because it didn't process any leaf with extent items. Second example where we will miss a hole: *) An inode as its items spanning 5 (or more) leafs; *) A hole is punched and it covers only the extents items of the 3rd leaf. This resulsts in deleting the entire leaf and not touching any of the other leafs. So the only leaf that is modified in the current transaction, when punching the hole, is the first leaf, which contains the inode item. During the full fsync, the only leaf that is passed to copy_items() is that first leaf, and that's not enough for the hole detection code in copy_items() to determine there's a hole between the last file extent item in the 2nd leaf and the first file extent item in the 3rd leaf (which was the 4th leaf before punching the hole). Fix this by scanning all leafs and punch holes as necessary when doing a full fsync (less common than a non-full fsync) when the NO_HOLES feature is enabled. The lack of explicit file extent items to mark holes makes it necessary to scan existing extents to determine if holes exist. A test case for fstests follows soon. Fixes: 16e7549f045d33 ("Btrfs: incompatible format change to remove hole extents") CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-11-19 15:07:33 +03:00
bool dropped_extents = false;
btrfs: fix partial loss of prealloc extent past i_size after fsync When we have an inode with a prealloc extent that starts at an offset lower than the i_size and there is another prealloc extent that starts at an offset beyond i_size, we can end up losing part of the first prealloc extent (the part that starts at i_size) and have an implicit hole if we fsync the file and then have a power failure. Consider the following example with comments explaining how and why it happens. $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt # Create our test file with 2 consecutive prealloc extents, each with a # size of 128Kb, and covering the range from 0 to 256Kb, with a file # size of 0. $ xfs_io -f -c "falloc -k 0 128K" /mnt/foo $ xfs_io -c "falloc -k 128K 128K" /mnt/foo # Fsync the file to record both extents in the log tree. $ xfs_io -c "fsync" /mnt/foo # Now do a redudant extent allocation for the range from 0 to 64Kb. # This will merely increase the file size from 0 to 64Kb. Instead we # could also do a truncate to set the file size to 64Kb. $ xfs_io -c "falloc 0 64K" /mnt/foo # Fsync the file, so we update the inode item in the log tree with the # new file size (64Kb). This also ends up setting the number of bytes # for the first prealloc extent to 64Kb. This is done by the truncation # at btrfs_log_prealloc_extents(). # This means that if a power failure happens after this, a write into # the file range 64Kb to 128Kb will not use the prealloc extent and # will result in allocation of a new extent. $ xfs_io -c "fsync" /mnt/foo # Now set the file size to 256K with a truncate and then fsync the file. # Since no changes happened to the extents, the fsync only updates the # i_size in the inode item at the log tree. This results in an implicit # hole for the file range from 64Kb to 128Kb, something which fsck will # complain when not using the NO_HOLES feature if we replay the log # after a power failure. $ xfs_io -c "truncate 256K" -c "fsync" /mnt/foo So instead of always truncating the log to the inode's current i_size at btrfs_log_prealloc_extents(), check first if there's a prealloc extent that starts at an offset lower than the i_size and with a length that crosses the i_size - if there is one, just make sure we truncate to a size that corresponds to the end offset of that prealloc extent, so that we don't lose the part of that extent that starts at i_size if a power failure happens. A test case for fstests follows soon. Fixes: 31d11b83b96f ("Btrfs: fix duplicate extents after fsync of file with prealloc extents") CC: stable@vger.kernel.org # 4.14+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-04-23 18:30:53 +03:00
u64 truncate_offset = i_size;
struct extent_buffer *leaf;
int slot;
Btrfs: fix duplicate extents after fsync of file with prealloc extents In commit 471d557afed1 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay"), on fsync, we started to always log all prealloc extents beyond an inode's i_size in order to avoid losing them after a power failure. However under some cases this can lead to the log replay code to create duplicate extent items, with different lengths, in the extent tree. That happens because, as of that commit, we can now log extent items based on extent maps that are not on the "modified" list of extent maps of the inode's extent map tree. Logging extent items based on extent maps is used during the fast fsync path to save time and for this to work reliably it requires that the extent maps are not merged with other adjacent extent maps - having the extent maps in the list of modified extents gives such guarantee. Consider the following example, captured during a long run of fsstress, which illustrates this problem. We have inode 271, in the filesystem tree (root 5), for which all of the following operations and discussion apply to. A buffered write starts at offset 312391 with a length of 933471 bytes (end offset at 1245862). At this point we have, for this inode, the following extent maps with the their field values: em A, start 0, orig_start 0, len 40960, block_start 18446744073709551613, block_len 0, orig_block_len 0 em B, start 40960, orig_start 40960, len 376832, block_start 1106399232, block_len 376832, orig_block_len 376832 em C, start 417792, orig_start 417792, len 782336, block_start 18446744073709551613, block_len 0, orig_block_len 0 em D, start 1200128, orig_start 1200128, len 835584, block_start 1106776064, block_len 835584, orig_block_len 835584 em E, start 2035712, orig_start 2035712, len 245760, block_start 1107611648, block_len 245760, orig_block_len 245760 Extent map A corresponds to a hole and extent maps D and E correspond to preallocated extents. Extent map D ends where extent map E begins (1106776064 + 835584 = 1107611648), but these extent maps were not merged because they are in the inode's list of modified extent maps. An fsync against this inode is made, which triggers the fast path (BTRFS_INODE_NEEDS_FULL_SYNC is not set). This fsync triggers writeback of the data previously written using buffered IO, and when the respective ordered extent finishes, btrfs_drop_extents() is called against the (aligned) range 311296..1249279. This causes a split of extent map D at btrfs_drop_extent_cache(), replacing extent map D with a new extent map D', also added to the list of modified extents, with the following values: em D', start 1249280, orig_start of 1200128, block_start 1106825216 (= 1106776064 + 1249280 - 1200128), orig_block_len 835584, block_len 786432 (835584 - (1249280 - 1200128)) Then, during the fast fsync, btrfs_log_changed_extents() is called and extent maps D' and E are removed from the list of modified extents. The flag EXTENT_FLAG_LOGGING is also set on them. After the extents are logged clear_em_logging() is called on each of them, and that makes extent map E to be merged with extent map D' (try_merge_map()), resulting in D' being deleted and E adjusted to: em E, start 1249280, orig_start 1200128, len 1032192, block_start 1106825216, block_len 1032192, orig_block_len 245760 A direct IO write at offset 1847296 and length of 360448 bytes (end offset at 2207744) starts, and at that moment the following extent maps exist for our inode: em A, start 0, orig_start 0, len 40960, block_start 18446744073709551613, block_len 0, orig_block_len 0 em B, start 40960, orig_start 40960, len 270336, block_start 1106399232, block_len 270336, orig_block_len 376832 em C, start 311296, orig_start 311296, len 937984, block_start 1112842240, block_len 937984, orig_block_len 937984 em E (prealloc), start 1249280, orig_start 1200128, len 1032192, block_start 1106825216, block_len 1032192, orig_block_len 245760 The dio write results in drop_extent_cache() being called twice. The first time for a range that starts at offset 1847296 and ends at offset 2035711 (length of 188416), which results in a double split of extent map E, replacing it with two new extent maps: em F, start 1249280, orig_start 1200128, block_start 1106825216, block_len 598016, orig_block_len 598016 em G, start 2035712, orig_start 1200128, block_start 1107611648, block_len 245760, orig_block_len 1032192 It also creates a new extent map that represents a part of the requested IO (through create_io_em()): em H, start 1847296, len 188416, block_start 1107423232, block_len 188416 The second call to drop_extent_cache() has a range with a start offset of 2035712 and end offset of 2207743 (length of 172032). This leads to replacing extent map G with a new extent map I with the following values: em I, start 2207744, orig_start 1200128, block_start 1107783680, block_len 73728, orig_block_len 1032192 It also creates a new extent map that represents the second part of the requested IO (through create_io_em()): em J, start 2035712, len 172032, block_start 1107611648, block_len 172032 The dio write set the inode's i_size to 2207744 bytes. After the dio write the inode has the following extent maps: em A, start 0, orig_start 0, len 40960, block_start 18446744073709551613, block_len 0, orig_block_len 0 em B, start 40960, orig_start 40960, len 270336, block_start 1106399232, block_len 270336, orig_block_len 376832 em C, start 311296, orig_start 311296, len 937984, block_start 1112842240, block_len 937984, orig_block_len 937984 em F, start 1249280, orig_start 1200128, len 598016, block_start 1106825216, block_len 598016, orig_block_len 598016 em H, start 1847296, orig_start 1200128, len 188416, block_start 1107423232, block_len 188416, orig_block_len 835584 em J, start 2035712, orig_start 2035712, len 172032, block_start 1107611648, block_len 172032, orig_block_len 245760 em I, start 2207744, orig_start 1200128, len 73728, block_start 1107783680, block_len 73728, orig_block_len 1032192 Now do some change to the file, like adding a xattr for example and then fsync it again. This triggers a fast fsync path, and as of commit 471d557afed1 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay"), we use the extent map I to log a file extent item because it's a prealloc extent and it starts at an offset matching the inode's i_size. However when we log it, we create a file extent item with a value for the disk byte location that is wrong, as can be seen from the following output of "btrfs inspect-internal dump-tree": item 1 key (271 EXTENT_DATA 2207744) itemoff 3782 itemsize 53 generation 22 type 2 (prealloc) prealloc data disk byte 1106776064 nr 1032192 prealloc data offset 1007616 nr 73728 Here the disk byte value corresponds to calculation based on some fields from the extent map I: 1106776064 = block_start (1107783680) - 1007616 (extent_offset) extent_offset = 2207744 (start) - 1200128 (orig_start) = 1007616 The disk byte value of 1106776064 clashes with disk byte values of the file extent items at offsets 1249280 and 1847296 in the fs tree: item 6 key (271 EXTENT_DATA 1249280) itemoff 3568 itemsize 53 generation 20 type 2 (prealloc) prealloc data disk byte 1106776064 nr 835584 prealloc data offset 49152 nr 598016 item 7 key (271 EXTENT_DATA 1847296) itemoff 3515 itemsize 53 generation 20 type 1 (regular) extent data disk byte 1106776064 nr 835584 extent data offset 647168 nr 188416 ram 835584 extent compression 0 (none) item 8 key (271 EXTENT_DATA 2035712) itemoff 3462 itemsize 53 generation 20 type 1 (regular) extent data disk byte 1107611648 nr 245760 extent data offset 0 nr 172032 ram 245760 extent compression 0 (none) item 9 key (271 EXTENT_DATA 2207744) itemoff 3409 itemsize 53 generation 20 type 2 (prealloc) prealloc data disk byte 1107611648 nr 245760 prealloc data offset 172032 nr 73728 Instead of the disk byte value of 1106776064, the value of 1107611648 should have been logged. Also the data offset value should have been 172032 and not 1007616. After a log replay we end up getting two extent items in the extent tree with different lengths, one of 835584, which is correct and existed before the log replay, and another one of 1032192 which is wrong and is based on the logged file extent item: item 12 key (1106776064 EXTENT_ITEM 835584) itemoff 3406 itemsize 53 refs 2 gen 15 flags DATA extent data backref root 5 objectid 271 offset 1200128 count 2 item 13 key (1106776064 EXTENT_ITEM 1032192) itemoff 3353 itemsize 53 refs 1 gen 22 flags DATA extent data backref root 5 objectid 271 offset 1200128 count 1 Obviously this leads to many problems and a filesystem check reports many errors: (...) checking extents Extent back ref already exists for 1106776064 parent 0 root 5 owner 271 offset 1200128 num_refs 1 extent item 1106776064 has multiple extent items ref mismatch on [1106776064 835584] extent item 2, found 3 Incorrect local backref count on 1106776064 root 5 owner 271 offset 1200128 found 2 wanted 1 back 0x55b1d0ad7680 Backref 1106776064 root 5 owner 271 offset 1200128 num_refs 0 not found in extent tree Incorrect local backref count on 1106776064 root 5 owner 271 offset 1200128 found 1 wanted 0 back 0x55b1d0ad4e70 Backref bytes do not match extent backref, bytenr=1106776064, ref bytes=835584, backref bytes=1032192 backpointer mismatch on [1106776064 835584] checking free space cache block group 1103101952 has wrong amount of free space failed to load free space cache for block group 1103101952 checking fs roots (...) So fix this by logging the prealloc extents beyond the inode's i_size based on searches in the subvolume tree instead of the extent maps. Fixes: 471d557afed1 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay") CC: stable@vger.kernel.org # 4.14+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2018-05-09 18:01:46 +03:00
int ins_nr = 0;
int start_slot;
int ret;
if (!(inode->flags & BTRFS_INODE_PREALLOC))
return 0;
key.objectid = ino;
key.type = BTRFS_EXTENT_DATA_KEY;
key.offset = i_size;
ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
if (ret < 0)
goto out;
btrfs: fix partial loss of prealloc extent past i_size after fsync When we have an inode with a prealloc extent that starts at an offset lower than the i_size and there is another prealloc extent that starts at an offset beyond i_size, we can end up losing part of the first prealloc extent (the part that starts at i_size) and have an implicit hole if we fsync the file and then have a power failure. Consider the following example with comments explaining how and why it happens. $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt # Create our test file with 2 consecutive prealloc extents, each with a # size of 128Kb, and covering the range from 0 to 256Kb, with a file # size of 0. $ xfs_io -f -c "falloc -k 0 128K" /mnt/foo $ xfs_io -c "falloc -k 128K 128K" /mnt/foo # Fsync the file to record both extents in the log tree. $ xfs_io -c "fsync" /mnt/foo # Now do a redudant extent allocation for the range from 0 to 64Kb. # This will merely increase the file size from 0 to 64Kb. Instead we # could also do a truncate to set the file size to 64Kb. $ xfs_io -c "falloc 0 64K" /mnt/foo # Fsync the file, so we update the inode item in the log tree with the # new file size (64Kb). This also ends up setting the number of bytes # for the first prealloc extent to 64Kb. This is done by the truncation # at btrfs_log_prealloc_extents(). # This means that if a power failure happens after this, a write into # the file range 64Kb to 128Kb will not use the prealloc extent and # will result in allocation of a new extent. $ xfs_io -c "fsync" /mnt/foo # Now set the file size to 256K with a truncate and then fsync the file. # Since no changes happened to the extents, the fsync only updates the # i_size in the inode item at the log tree. This results in an implicit # hole for the file range from 64Kb to 128Kb, something which fsck will # complain when not using the NO_HOLES feature if we replay the log # after a power failure. $ xfs_io -c "truncate 256K" -c "fsync" /mnt/foo So instead of always truncating the log to the inode's current i_size at btrfs_log_prealloc_extents(), check first if there's a prealloc extent that starts at an offset lower than the i_size and with a length that crosses the i_size - if there is one, just make sure we truncate to a size that corresponds to the end offset of that prealloc extent, so that we don't lose the part of that extent that starts at i_size if a power failure happens. A test case for fstests follows soon. Fixes: 31d11b83b96f ("Btrfs: fix duplicate extents after fsync of file with prealloc extents") CC: stable@vger.kernel.org # 4.14+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-04-23 18:30:53 +03:00
/*
* We must check if there is a prealloc extent that starts before the
* i_size and crosses the i_size boundary. This is to ensure later we
* truncate down to the end of that extent and not to the i_size, as
* otherwise we end up losing part of the prealloc extent after a log
* replay and with an implicit hole if there is another prealloc extent
* that starts at an offset beyond i_size.
*/
ret = btrfs_previous_item(root, path, ino, BTRFS_EXTENT_DATA_KEY);
if (ret < 0)
goto out;
if (ret == 0) {
struct btrfs_file_extent_item *ei;
leaf = path->nodes[0];
slot = path->slots[0];
ei = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
if (btrfs_file_extent_type(leaf, ei) ==
BTRFS_FILE_EXTENT_PREALLOC) {
u64 extent_end;
btrfs_item_key_to_cpu(leaf, &key, slot);
extent_end = key.offset +
btrfs_file_extent_num_bytes(leaf, ei);
if (extent_end > i_size)
truncate_offset = extent_end;
}
} else {
ret = 0;
}
Btrfs: fix duplicate extents after fsync of file with prealloc extents In commit 471d557afed1 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay"), on fsync, we started to always log all prealloc extents beyond an inode's i_size in order to avoid losing them after a power failure. However under some cases this can lead to the log replay code to create duplicate extent items, with different lengths, in the extent tree. That happens because, as of that commit, we can now log extent items based on extent maps that are not on the "modified" list of extent maps of the inode's extent map tree. Logging extent items based on extent maps is used during the fast fsync path to save time and for this to work reliably it requires that the extent maps are not merged with other adjacent extent maps - having the extent maps in the list of modified extents gives such guarantee. Consider the following example, captured during a long run of fsstress, which illustrates this problem. We have inode 271, in the filesystem tree (root 5), for which all of the following operations and discussion apply to. A buffered write starts at offset 312391 with a length of 933471 bytes (end offset at 1245862). At this point we have, for this inode, the following extent maps with the their field values: em A, start 0, orig_start 0, len 40960, block_start 18446744073709551613, block_len 0, orig_block_len 0 em B, start 40960, orig_start 40960, len 376832, block_start 1106399232, block_len 376832, orig_block_len 376832 em C, start 417792, orig_start 417792, len 782336, block_start 18446744073709551613, block_len 0, orig_block_len 0 em D, start 1200128, orig_start 1200128, len 835584, block_start 1106776064, block_len 835584, orig_block_len 835584 em E, start 2035712, orig_start 2035712, len 245760, block_start 1107611648, block_len 245760, orig_block_len 245760 Extent map A corresponds to a hole and extent maps D and E correspond to preallocated extents. Extent map D ends where extent map E begins (1106776064 + 835584 = 1107611648), but these extent maps were not merged because they are in the inode's list of modified extent maps. An fsync against this inode is made, which triggers the fast path (BTRFS_INODE_NEEDS_FULL_SYNC is not set). This fsync triggers writeback of the data previously written using buffered IO, and when the respective ordered extent finishes, btrfs_drop_extents() is called against the (aligned) range 311296..1249279. This causes a split of extent map D at btrfs_drop_extent_cache(), replacing extent map D with a new extent map D', also added to the list of modified extents, with the following values: em D', start 1249280, orig_start of 1200128, block_start 1106825216 (= 1106776064 + 1249280 - 1200128), orig_block_len 835584, block_len 786432 (835584 - (1249280 - 1200128)) Then, during the fast fsync, btrfs_log_changed_extents() is called and extent maps D' and E are removed from the list of modified extents. The flag EXTENT_FLAG_LOGGING is also set on them. After the extents are logged clear_em_logging() is called on each of them, and that makes extent map E to be merged with extent map D' (try_merge_map()), resulting in D' being deleted and E adjusted to: em E, start 1249280, orig_start 1200128, len 1032192, block_start 1106825216, block_len 1032192, orig_block_len 245760 A direct IO write at offset 1847296 and length of 360448 bytes (end offset at 2207744) starts, and at that moment the following extent maps exist for our inode: em A, start 0, orig_start 0, len 40960, block_start 18446744073709551613, block_len 0, orig_block_len 0 em B, start 40960, orig_start 40960, len 270336, block_start 1106399232, block_len 270336, orig_block_len 376832 em C, start 311296, orig_start 311296, len 937984, block_start 1112842240, block_len 937984, orig_block_len 937984 em E (prealloc), start 1249280, orig_start 1200128, len 1032192, block_start 1106825216, block_len 1032192, orig_block_len 245760 The dio write results in drop_extent_cache() being called twice. The first time for a range that starts at offset 1847296 and ends at offset 2035711 (length of 188416), which results in a double split of extent map E, replacing it with two new extent maps: em F, start 1249280, orig_start 1200128, block_start 1106825216, block_len 598016, orig_block_len 598016 em G, start 2035712, orig_start 1200128, block_start 1107611648, block_len 245760, orig_block_len 1032192 It also creates a new extent map that represents a part of the requested IO (through create_io_em()): em H, start 1847296, len 188416, block_start 1107423232, block_len 188416 The second call to drop_extent_cache() has a range with a start offset of 2035712 and end offset of 2207743 (length of 172032). This leads to replacing extent map G with a new extent map I with the following values: em I, start 2207744, orig_start 1200128, block_start 1107783680, block_len 73728, orig_block_len 1032192 It also creates a new extent map that represents the second part of the requested IO (through create_io_em()): em J, start 2035712, len 172032, block_start 1107611648, block_len 172032 The dio write set the inode's i_size to 2207744 bytes. After the dio write the inode has the following extent maps: em A, start 0, orig_start 0, len 40960, block_start 18446744073709551613, block_len 0, orig_block_len 0 em B, start 40960, orig_start 40960, len 270336, block_start 1106399232, block_len 270336, orig_block_len 376832 em C, start 311296, orig_start 311296, len 937984, block_start 1112842240, block_len 937984, orig_block_len 937984 em F, start 1249280, orig_start 1200128, len 598016, block_start 1106825216, block_len 598016, orig_block_len 598016 em H, start 1847296, orig_start 1200128, len 188416, block_start 1107423232, block_len 188416, orig_block_len 835584 em J, start 2035712, orig_start 2035712, len 172032, block_start 1107611648, block_len 172032, orig_block_len 245760 em I, start 2207744, orig_start 1200128, len 73728, block_start 1107783680, block_len 73728, orig_block_len 1032192 Now do some change to the file, like adding a xattr for example and then fsync it again. This triggers a fast fsync path, and as of commit 471d557afed1 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay"), we use the extent map I to log a file extent item because it's a prealloc extent and it starts at an offset matching the inode's i_size. However when we log it, we create a file extent item with a value for the disk byte location that is wrong, as can be seen from the following output of "btrfs inspect-internal dump-tree": item 1 key (271 EXTENT_DATA 2207744) itemoff 3782 itemsize 53 generation 22 type 2 (prealloc) prealloc data disk byte 1106776064 nr 1032192 prealloc data offset 1007616 nr 73728 Here the disk byte value corresponds to calculation based on some fields from the extent map I: 1106776064 = block_start (1107783680) - 1007616 (extent_offset) extent_offset = 2207744 (start) - 1200128 (orig_start) = 1007616 The disk byte value of 1106776064 clashes with disk byte values of the file extent items at offsets 1249280 and 1847296 in the fs tree: item 6 key (271 EXTENT_DATA 1249280) itemoff 3568 itemsize 53 generation 20 type 2 (prealloc) prealloc data disk byte 1106776064 nr 835584 prealloc data offset 49152 nr 598016 item 7 key (271 EXTENT_DATA 1847296) itemoff 3515 itemsize 53 generation 20 type 1 (regular) extent data disk byte 1106776064 nr 835584 extent data offset 647168 nr 188416 ram 835584 extent compression 0 (none) item 8 key (271 EXTENT_DATA 2035712) itemoff 3462 itemsize 53 generation 20 type 1 (regular) extent data disk byte 1107611648 nr 245760 extent data offset 0 nr 172032 ram 245760 extent compression 0 (none) item 9 key (271 EXTENT_DATA 2207744) itemoff 3409 itemsize 53 generation 20 type 2 (prealloc) prealloc data disk byte 1107611648 nr 245760 prealloc data offset 172032 nr 73728 Instead of the disk byte value of 1106776064, the value of 1107611648 should have been logged. Also the data offset value should have been 172032 and not 1007616. After a log replay we end up getting two extent items in the extent tree with different lengths, one of 835584, which is correct and existed before the log replay, and another one of 1032192 which is wrong and is based on the logged file extent item: item 12 key (1106776064 EXTENT_ITEM 835584) itemoff 3406 itemsize 53 refs 2 gen 15 flags DATA extent data backref root 5 objectid 271 offset 1200128 count 2 item 13 key (1106776064 EXTENT_ITEM 1032192) itemoff 3353 itemsize 53 refs 1 gen 22 flags DATA extent data backref root 5 objectid 271 offset 1200128 count 1 Obviously this leads to many problems and a filesystem check reports many errors: (...) checking extents Extent back ref already exists for 1106776064 parent 0 root 5 owner 271 offset 1200128 num_refs 1 extent item 1106776064 has multiple extent items ref mismatch on [1106776064 835584] extent item 2, found 3 Incorrect local backref count on 1106776064 root 5 owner 271 offset 1200128 found 2 wanted 1 back 0x55b1d0ad7680 Backref 1106776064 root 5 owner 271 offset 1200128 num_refs 0 not found in extent tree Incorrect local backref count on 1106776064 root 5 owner 271 offset 1200128 found 1 wanted 0 back 0x55b1d0ad4e70 Backref bytes do not match extent backref, bytenr=1106776064, ref bytes=835584, backref bytes=1032192 backpointer mismatch on [1106776064 835584] checking free space cache block group 1103101952 has wrong amount of free space failed to load free space cache for block group 1103101952 checking fs roots (...) So fix this by logging the prealloc extents beyond the inode's i_size based on searches in the subvolume tree instead of the extent maps. Fixes: 471d557afed1 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay") CC: stable@vger.kernel.org # 4.14+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2018-05-09 18:01:46 +03:00
while (true) {
btrfs: fix partial loss of prealloc extent past i_size after fsync When we have an inode with a prealloc extent that starts at an offset lower than the i_size and there is another prealloc extent that starts at an offset beyond i_size, we can end up losing part of the first prealloc extent (the part that starts at i_size) and have an implicit hole if we fsync the file and then have a power failure. Consider the following example with comments explaining how and why it happens. $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt # Create our test file with 2 consecutive prealloc extents, each with a # size of 128Kb, and covering the range from 0 to 256Kb, with a file # size of 0. $ xfs_io -f -c "falloc -k 0 128K" /mnt/foo $ xfs_io -c "falloc -k 128K 128K" /mnt/foo # Fsync the file to record both extents in the log tree. $ xfs_io -c "fsync" /mnt/foo # Now do a redudant extent allocation for the range from 0 to 64Kb. # This will merely increase the file size from 0 to 64Kb. Instead we # could also do a truncate to set the file size to 64Kb. $ xfs_io -c "falloc 0 64K" /mnt/foo # Fsync the file, so we update the inode item in the log tree with the # new file size (64Kb). This also ends up setting the number of bytes # for the first prealloc extent to 64Kb. This is done by the truncation # at btrfs_log_prealloc_extents(). # This means that if a power failure happens after this, a write into # the file range 64Kb to 128Kb will not use the prealloc extent and # will result in allocation of a new extent. $ xfs_io -c "fsync" /mnt/foo # Now set the file size to 256K with a truncate and then fsync the file. # Since no changes happened to the extents, the fsync only updates the # i_size in the inode item at the log tree. This results in an implicit # hole for the file range from 64Kb to 128Kb, something which fsck will # complain when not using the NO_HOLES feature if we replay the log # after a power failure. $ xfs_io -c "truncate 256K" -c "fsync" /mnt/foo So instead of always truncating the log to the inode's current i_size at btrfs_log_prealloc_extents(), check first if there's a prealloc extent that starts at an offset lower than the i_size and with a length that crosses the i_size - if there is one, just make sure we truncate to a size that corresponds to the end offset of that prealloc extent, so that we don't lose the part of that extent that starts at i_size if a power failure happens. A test case for fstests follows soon. Fixes: 31d11b83b96f ("Btrfs: fix duplicate extents after fsync of file with prealloc extents") CC: stable@vger.kernel.org # 4.14+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-04-23 18:30:53 +03:00
leaf = path->nodes[0];
slot = path->slots[0];
Btrfs: fix duplicate extents after fsync of file with prealloc extents In commit 471d557afed1 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay"), on fsync, we started to always log all prealloc extents beyond an inode's i_size in order to avoid losing them after a power failure. However under some cases this can lead to the log replay code to create duplicate extent items, with different lengths, in the extent tree. That happens because, as of that commit, we can now log extent items based on extent maps that are not on the "modified" list of extent maps of the inode's extent map tree. Logging extent items based on extent maps is used during the fast fsync path to save time and for this to work reliably it requires that the extent maps are not merged with other adjacent extent maps - having the extent maps in the list of modified extents gives such guarantee. Consider the following example, captured during a long run of fsstress, which illustrates this problem. We have inode 271, in the filesystem tree (root 5), for which all of the following operations and discussion apply to. A buffered write starts at offset 312391 with a length of 933471 bytes (end offset at 1245862). At this point we have, for this inode, the following extent maps with the their field values: em A, start 0, orig_start 0, len 40960, block_start 18446744073709551613, block_len 0, orig_block_len 0 em B, start 40960, orig_start 40960, len 376832, block_start 1106399232, block_len 376832, orig_block_len 376832 em C, start 417792, orig_start 417792, len 782336, block_start 18446744073709551613, block_len 0, orig_block_len 0 em D, start 1200128, orig_start 1200128, len 835584, block_start 1106776064, block_len 835584, orig_block_len 835584 em E, start 2035712, orig_start 2035712, len 245760, block_start 1107611648, block_len 245760, orig_block_len 245760 Extent map A corresponds to a hole and extent maps D and E correspond to preallocated extents. Extent map D ends where extent map E begins (1106776064 + 835584 = 1107611648), but these extent maps were not merged because they are in the inode's list of modified extent maps. An fsync against this inode is made, which triggers the fast path (BTRFS_INODE_NEEDS_FULL_SYNC is not set). This fsync triggers writeback of the data previously written using buffered IO, and when the respective ordered extent finishes, btrfs_drop_extents() is called against the (aligned) range 311296..1249279. This causes a split of extent map D at btrfs_drop_extent_cache(), replacing extent map D with a new extent map D', also added to the list of modified extents, with the following values: em D', start 1249280, orig_start of 1200128, block_start 1106825216 (= 1106776064 + 1249280 - 1200128), orig_block_len 835584, block_len 786432 (835584 - (1249280 - 1200128)) Then, during the fast fsync, btrfs_log_changed_extents() is called and extent maps D' and E are removed from the list of modified extents. The flag EXTENT_FLAG_LOGGING is also set on them. After the extents are logged clear_em_logging() is called on each of them, and that makes extent map E to be merged with extent map D' (try_merge_map()), resulting in D' being deleted and E adjusted to: em E, start 1249280, orig_start 1200128, len 1032192, block_start 1106825216, block_len 1032192, orig_block_len 245760 A direct IO write at offset 1847296 and length of 360448 bytes (end offset at 2207744) starts, and at that moment the following extent maps exist for our inode: em A, start 0, orig_start 0, len 40960, block_start 18446744073709551613, block_len 0, orig_block_len 0 em B, start 40960, orig_start 40960, len 270336, block_start 1106399232, block_len 270336, orig_block_len 376832 em C, start 311296, orig_start 311296, len 937984, block_start 1112842240, block_len 937984, orig_block_len 937984 em E (prealloc), start 1249280, orig_start 1200128, len 1032192, block_start 1106825216, block_len 1032192, orig_block_len 245760 The dio write results in drop_extent_cache() being called twice. The first time for a range that starts at offset 1847296 and ends at offset 2035711 (length of 188416), which results in a double split of extent map E, replacing it with two new extent maps: em F, start 1249280, orig_start 1200128, block_start 1106825216, block_len 598016, orig_block_len 598016 em G, start 2035712, orig_start 1200128, block_start 1107611648, block_len 245760, orig_block_len 1032192 It also creates a new extent map that represents a part of the requested IO (through create_io_em()): em H, start 1847296, len 188416, block_start 1107423232, block_len 188416 The second call to drop_extent_cache() has a range with a start offset of 2035712 and end offset of 2207743 (length of 172032). This leads to replacing extent map G with a new extent map I with the following values: em I, start 2207744, orig_start 1200128, block_start 1107783680, block_len 73728, orig_block_len 1032192 It also creates a new extent map that represents the second part of the requested IO (through create_io_em()): em J, start 2035712, len 172032, block_start 1107611648, block_len 172032 The dio write set the inode's i_size to 2207744 bytes. After the dio write the inode has the following extent maps: em A, start 0, orig_start 0, len 40960, block_start 18446744073709551613, block_len 0, orig_block_len 0 em B, start 40960, orig_start 40960, len 270336, block_start 1106399232, block_len 270336, orig_block_len 376832 em C, start 311296, orig_start 311296, len 937984, block_start 1112842240, block_len 937984, orig_block_len 937984 em F, start 1249280, orig_start 1200128, len 598016, block_start 1106825216, block_len 598016, orig_block_len 598016 em H, start 1847296, orig_start 1200128, len 188416, block_start 1107423232, block_len 188416, orig_block_len 835584 em J, start 2035712, orig_start 2035712, len 172032, block_start 1107611648, block_len 172032, orig_block_len 245760 em I, start 2207744, orig_start 1200128, len 73728, block_start 1107783680, block_len 73728, orig_block_len 1032192 Now do some change to the file, like adding a xattr for example and then fsync it again. This triggers a fast fsync path, and as of commit 471d557afed1 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay"), we use the extent map I to log a file extent item because it's a prealloc extent and it starts at an offset matching the inode's i_size. However when we log it, we create a file extent item with a value for the disk byte location that is wrong, as can be seen from the following output of "btrfs inspect-internal dump-tree": item 1 key (271 EXTENT_DATA 2207744) itemoff 3782 itemsize 53 generation 22 type 2 (prealloc) prealloc data disk byte 1106776064 nr 1032192 prealloc data offset 1007616 nr 73728 Here the disk byte value corresponds to calculation based on some fields from the extent map I: 1106776064 = block_start (1107783680) - 1007616 (extent_offset) extent_offset = 2207744 (start) - 1200128 (orig_start) = 1007616 The disk byte value of 1106776064 clashes with disk byte values of the file extent items at offsets 1249280 and 1847296 in the fs tree: item 6 key (271 EXTENT_DATA 1249280) itemoff 3568 itemsize 53 generation 20 type 2 (prealloc) prealloc data disk byte 1106776064 nr 835584 prealloc data offset 49152 nr 598016 item 7 key (271 EXTENT_DATA 1847296) itemoff 3515 itemsize 53 generation 20 type 1 (regular) extent data disk byte 1106776064 nr 835584 extent data offset 647168 nr 188416 ram 835584 extent compression 0 (none) item 8 key (271 EXTENT_DATA 2035712) itemoff 3462 itemsize 53 generation 20 type 1 (regular) extent data disk byte 1107611648 nr 245760 extent data offset 0 nr 172032 ram 245760 extent compression 0 (none) item 9 key (271 EXTENT_DATA 2207744) itemoff 3409 itemsize 53 generation 20 type 2 (prealloc) prealloc data disk byte 1107611648 nr 245760 prealloc data offset 172032 nr 73728 Instead of the disk byte value of 1106776064, the value of 1107611648 should have been logged. Also the data offset value should have been 172032 and not 1007616. After a log replay we end up getting two extent items in the extent tree with different lengths, one of 835584, which is correct and existed before the log replay, and another one of 1032192 which is wrong and is based on the logged file extent item: item 12 key (1106776064 EXTENT_ITEM 835584) itemoff 3406 itemsize 53 refs 2 gen 15 flags DATA extent data backref root 5 objectid 271 offset 1200128 count 2 item 13 key (1106776064 EXTENT_ITEM 1032192) itemoff 3353 itemsize 53 refs 1 gen 22 flags DATA extent data backref root 5 objectid 271 offset 1200128 count 1 Obviously this leads to many problems and a filesystem check reports many errors: (...) checking extents Extent back ref already exists for 1106776064 parent 0 root 5 owner 271 offset 1200128 num_refs 1 extent item 1106776064 has multiple extent items ref mismatch on [1106776064 835584] extent item 2, found 3 Incorrect local backref count on 1106776064 root 5 owner 271 offset 1200128 found 2 wanted 1 back 0x55b1d0ad7680 Backref 1106776064 root 5 owner 271 offset 1200128 num_refs 0 not found in extent tree Incorrect local backref count on 1106776064 root 5 owner 271 offset 1200128 found 1 wanted 0 back 0x55b1d0ad4e70 Backref bytes do not match extent backref, bytenr=1106776064, ref bytes=835584, backref bytes=1032192 backpointer mismatch on [1106776064 835584] checking free space cache block group 1103101952 has wrong amount of free space failed to load free space cache for block group 1103101952 checking fs roots (...) So fix this by logging the prealloc extents beyond the inode's i_size based on searches in the subvolume tree instead of the extent maps. Fixes: 471d557afed1 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay") CC: stable@vger.kernel.org # 4.14+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2018-05-09 18:01:46 +03:00
if (slot >= btrfs_header_nritems(leaf)) {
if (ins_nr > 0) {
ret = copy_items(trans, inode, dst_path, path,
Btrfs: fix missing hole after hole punching and fsync when using NO_HOLES When using the NO_HOLES feature, if we punch a hole into a file and then fsync it, there are cases where a subsequent fsync will miss the fact that a hole was punched, resulting in the holes not existing after replaying the log tree. Essentially these cases all imply that, tree-log.c:copy_items(), is not invoked for the leafs that delimit holes, because nothing changed those leafs in the current transaction. And it's precisely copy_items() where we currenly detect and log holes, which works as long as the holes are between file extent items in the input leaf or between the beginning of input leaf and the previous leaf or between the last item in the leaf and the next leaf. First example where we miss a hole: *) The extent items of the inode span multiple leafs; *) The punched hole covers a range that affects only the extent items of the first leaf; *) The fsync operation is done in full mode (BTRFS_INODE_NEEDS_FULL_SYNC is set in the inode's runtime flags). That results in the hole not existing after replaying the log tree. For example, if the fs/subvolume tree has the following layout for a particular inode: Leaf N, generation 10: [ ... INODE_ITEM INODE_REF EXTENT_ITEM (0 64K) EXTENT_ITEM (64K 128K) ] Leaf N + 1, generation 10: [ EXTENT_ITEM (128K 64K) ... ] If at transaction 11 we punch a hole coverting the range [0, 128K[, we end up dropping the two extent items from leaf N, but we don't touch the other leaf, so we end up in the following state: Leaf N, generation 11: [ ... INODE_ITEM INODE_REF ] Leaf N + 1, generation 10: [ EXTENT_ITEM (128K 64K) ... ] A full fsync after punching the hole will only process leaf N because it was modified in the current transaction, but not leaf N + 1, since it was not modified in the current transaction (generation 10 and not 11). As a result the fsync will not log any holes, because it didn't process any leaf with extent items. Second example where we will miss a hole: *) An inode as its items spanning 5 (or more) leafs; *) A hole is punched and it covers only the extents items of the 3rd leaf. This resulsts in deleting the entire leaf and not touching any of the other leafs. So the only leaf that is modified in the current transaction, when punching the hole, is the first leaf, which contains the inode item. During the full fsync, the only leaf that is passed to copy_items() is that first leaf, and that's not enough for the hole detection code in copy_items() to determine there's a hole between the last file extent item in the 2nd leaf and the first file extent item in the 3rd leaf (which was the 4th leaf before punching the hole). Fix this by scanning all leafs and punch holes as necessary when doing a full fsync (less common than a non-full fsync) when the NO_HOLES feature is enabled. The lack of explicit file extent items to mark holes makes it necessary to scan existing extents to determine if holes exist. A test case for fstests follows soon. Fixes: 16e7549f045d33 ("Btrfs: incompatible format change to remove hole extents") CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-11-19 15:07:33 +03:00
start_slot, ins_nr, 1, 0);
Btrfs: fix duplicate extents after fsync of file with prealloc extents In commit 471d557afed1 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay"), on fsync, we started to always log all prealloc extents beyond an inode's i_size in order to avoid losing them after a power failure. However under some cases this can lead to the log replay code to create duplicate extent items, with different lengths, in the extent tree. That happens because, as of that commit, we can now log extent items based on extent maps that are not on the "modified" list of extent maps of the inode's extent map tree. Logging extent items based on extent maps is used during the fast fsync path to save time and for this to work reliably it requires that the extent maps are not merged with other adjacent extent maps - having the extent maps in the list of modified extents gives such guarantee. Consider the following example, captured during a long run of fsstress, which illustrates this problem. We have inode 271, in the filesystem tree (root 5), for which all of the following operations and discussion apply to. A buffered write starts at offset 312391 with a length of 933471 bytes (end offset at 1245862). At this point we have, for this inode, the following extent maps with the their field values: em A, start 0, orig_start 0, len 40960, block_start 18446744073709551613, block_len 0, orig_block_len 0 em B, start 40960, orig_start 40960, len 376832, block_start 1106399232, block_len 376832, orig_block_len 376832 em C, start 417792, orig_start 417792, len 782336, block_start 18446744073709551613, block_len 0, orig_block_len 0 em D, start 1200128, orig_start 1200128, len 835584, block_start 1106776064, block_len 835584, orig_block_len 835584 em E, start 2035712, orig_start 2035712, len 245760, block_start 1107611648, block_len 245760, orig_block_len 245760 Extent map A corresponds to a hole and extent maps D and E correspond to preallocated extents. Extent map D ends where extent map E begins (1106776064 + 835584 = 1107611648), but these extent maps were not merged because they are in the inode's list of modified extent maps. An fsync against this inode is made, which triggers the fast path (BTRFS_INODE_NEEDS_FULL_SYNC is not set). This fsync triggers writeback of the data previously written using buffered IO, and when the respective ordered extent finishes, btrfs_drop_extents() is called against the (aligned) range 311296..1249279. This causes a split of extent map D at btrfs_drop_extent_cache(), replacing extent map D with a new extent map D', also added to the list of modified extents, with the following values: em D', start 1249280, orig_start of 1200128, block_start 1106825216 (= 1106776064 + 1249280 - 1200128), orig_block_len 835584, block_len 786432 (835584 - (1249280 - 1200128)) Then, during the fast fsync, btrfs_log_changed_extents() is called and extent maps D' and E are removed from the list of modified extents. The flag EXTENT_FLAG_LOGGING is also set on them. After the extents are logged clear_em_logging() is called on each of them, and that makes extent map E to be merged with extent map D' (try_merge_map()), resulting in D' being deleted and E adjusted to: em E, start 1249280, orig_start 1200128, len 1032192, block_start 1106825216, block_len 1032192, orig_block_len 245760 A direct IO write at offset 1847296 and length of 360448 bytes (end offset at 2207744) starts, and at that moment the following extent maps exist for our inode: em A, start 0, orig_start 0, len 40960, block_start 18446744073709551613, block_len 0, orig_block_len 0 em B, start 40960, orig_start 40960, len 270336, block_start 1106399232, block_len 270336, orig_block_len 376832 em C, start 311296, orig_start 311296, len 937984, block_start 1112842240, block_len 937984, orig_block_len 937984 em E (prealloc), start 1249280, orig_start 1200128, len 1032192, block_start 1106825216, block_len 1032192, orig_block_len 245760 The dio write results in drop_extent_cache() being called twice. The first time for a range that starts at offset 1847296 and ends at offset 2035711 (length of 188416), which results in a double split of extent map E, replacing it with two new extent maps: em F, start 1249280, orig_start 1200128, block_start 1106825216, block_len 598016, orig_block_len 598016 em G, start 2035712, orig_start 1200128, block_start 1107611648, block_len 245760, orig_block_len 1032192 It also creates a new extent map that represents a part of the requested IO (through create_io_em()): em H, start 1847296, len 188416, block_start 1107423232, block_len 188416 The second call to drop_extent_cache() has a range with a start offset of 2035712 and end offset of 2207743 (length of 172032). This leads to replacing extent map G with a new extent map I with the following values: em I, start 2207744, orig_start 1200128, block_start 1107783680, block_len 73728, orig_block_len 1032192 It also creates a new extent map that represents the second part of the requested IO (through create_io_em()): em J, start 2035712, len 172032, block_start 1107611648, block_len 172032 The dio write set the inode's i_size to 2207744 bytes. After the dio write the inode has the following extent maps: em A, start 0, orig_start 0, len 40960, block_start 18446744073709551613, block_len 0, orig_block_len 0 em B, start 40960, orig_start 40960, len 270336, block_start 1106399232, block_len 270336, orig_block_len 376832 em C, start 311296, orig_start 311296, len 937984, block_start 1112842240, block_len 937984, orig_block_len 937984 em F, start 1249280, orig_start 1200128, len 598016, block_start 1106825216, block_len 598016, orig_block_len 598016 em H, start 1847296, orig_start 1200128, len 188416, block_start 1107423232, block_len 188416, orig_block_len 835584 em J, start 2035712, orig_start 2035712, len 172032, block_start 1107611648, block_len 172032, orig_block_len 245760 em I, start 2207744, orig_start 1200128, len 73728, block_start 1107783680, block_len 73728, orig_block_len 1032192 Now do some change to the file, like adding a xattr for example and then fsync it again. This triggers a fast fsync path, and as of commit 471d557afed1 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay"), we use the extent map I to log a file extent item because it's a prealloc extent and it starts at an offset matching the inode's i_size. However when we log it, we create a file extent item with a value for the disk byte location that is wrong, as can be seen from the following output of "btrfs inspect-internal dump-tree": item 1 key (271 EXTENT_DATA 2207744) itemoff 3782 itemsize 53 generation 22 type 2 (prealloc) prealloc data disk byte 1106776064 nr 1032192 prealloc data offset 1007616 nr 73728 Here the disk byte value corresponds to calculation based on some fields from the extent map I: 1106776064 = block_start (1107783680) - 1007616 (extent_offset) extent_offset = 2207744 (start) - 1200128 (orig_start) = 1007616 The disk byte value of 1106776064 clashes with disk byte values of the file extent items at offsets 1249280 and 1847296 in the fs tree: item 6 key (271 EXTENT_DATA 1249280) itemoff 3568 itemsize 53 generation 20 type 2 (prealloc) prealloc data disk byte 1106776064 nr 835584 prealloc data offset 49152 nr 598016 item 7 key (271 EXTENT_DATA 1847296) itemoff 3515 itemsize 53 generation 20 type 1 (regular) extent data disk byte 1106776064 nr 835584 extent data offset 647168 nr 188416 ram 835584 extent compression 0 (none) item 8 key (271 EXTENT_DATA 2035712) itemoff 3462 itemsize 53 generation 20 type 1 (regular) extent data disk byte 1107611648 nr 245760 extent data offset 0 nr 172032 ram 245760 extent compression 0 (none) item 9 key (271 EXTENT_DATA 2207744) itemoff 3409 itemsize 53 generation 20 type 2 (prealloc) prealloc data disk byte 1107611648 nr 245760 prealloc data offset 172032 nr 73728 Instead of the disk byte value of 1106776064, the value of 1107611648 should have been logged. Also the data offset value should have been 172032 and not 1007616. After a log replay we end up getting two extent items in the extent tree with different lengths, one of 835584, which is correct and existed before the log replay, and another one of 1032192 which is wrong and is based on the logged file extent item: item 12 key (1106776064 EXTENT_ITEM 835584) itemoff 3406 itemsize 53 refs 2 gen 15 flags DATA extent data backref root 5 objectid 271 offset 1200128 count 2 item 13 key (1106776064 EXTENT_ITEM 1032192) itemoff 3353 itemsize 53 refs 1 gen 22 flags DATA extent data backref root 5 objectid 271 offset 1200128 count 1 Obviously this leads to many problems and a filesystem check reports many errors: (...) checking extents Extent back ref already exists for 1106776064 parent 0 root 5 owner 271 offset 1200128 num_refs 1 extent item 1106776064 has multiple extent items ref mismatch on [1106776064 835584] extent item 2, found 3 Incorrect local backref count on 1106776064 root 5 owner 271 offset 1200128 found 2 wanted 1 back 0x55b1d0ad7680 Backref 1106776064 root 5 owner 271 offset 1200128 num_refs 0 not found in extent tree Incorrect local backref count on 1106776064 root 5 owner 271 offset 1200128 found 1 wanted 0 back 0x55b1d0ad4e70 Backref bytes do not match extent backref, bytenr=1106776064, ref bytes=835584, backref bytes=1032192 backpointer mismatch on [1106776064 835584] checking free space cache block group 1103101952 has wrong amount of free space failed to load free space cache for block group 1103101952 checking fs roots (...) So fix this by logging the prealloc extents beyond the inode's i_size based on searches in the subvolume tree instead of the extent maps. Fixes: 471d557afed1 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay") CC: stable@vger.kernel.org # 4.14+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2018-05-09 18:01:46 +03:00
if (ret < 0)
goto out;
ins_nr = 0;
}
ret = btrfs_next_leaf(root, path);
if (ret < 0)
goto out;
if (ret > 0) {
ret = 0;
break;
}
continue;
}
btrfs_item_key_to_cpu(leaf, &key, slot);
if (key.objectid > ino)
break;
if (WARN_ON_ONCE(key.objectid < ino) ||
key.type < BTRFS_EXTENT_DATA_KEY ||
key.offset < i_size) {
path->slots[0]++;
continue;
}
Btrfs: fix missing hole after hole punching and fsync when using NO_HOLES When using the NO_HOLES feature, if we punch a hole into a file and then fsync it, there are cases where a subsequent fsync will miss the fact that a hole was punched, resulting in the holes not existing after replaying the log tree. Essentially these cases all imply that, tree-log.c:copy_items(), is not invoked for the leafs that delimit holes, because nothing changed those leafs in the current transaction. And it's precisely copy_items() where we currenly detect and log holes, which works as long as the holes are between file extent items in the input leaf or between the beginning of input leaf and the previous leaf or between the last item in the leaf and the next leaf. First example where we miss a hole: *) The extent items of the inode span multiple leafs; *) The punched hole covers a range that affects only the extent items of the first leaf; *) The fsync operation is done in full mode (BTRFS_INODE_NEEDS_FULL_SYNC is set in the inode's runtime flags). That results in the hole not existing after replaying the log tree. For example, if the fs/subvolume tree has the following layout for a particular inode: Leaf N, generation 10: [ ... INODE_ITEM INODE_REF EXTENT_ITEM (0 64K) EXTENT_ITEM (64K 128K) ] Leaf N + 1, generation 10: [ EXTENT_ITEM (128K 64K) ... ] If at transaction 11 we punch a hole coverting the range [0, 128K[, we end up dropping the two extent items from leaf N, but we don't touch the other leaf, so we end up in the following state: Leaf N, generation 11: [ ... INODE_ITEM INODE_REF ] Leaf N + 1, generation 10: [ EXTENT_ITEM (128K 64K) ... ] A full fsync after punching the hole will only process leaf N because it was modified in the current transaction, but not leaf N + 1, since it was not modified in the current transaction (generation 10 and not 11). As a result the fsync will not log any holes, because it didn't process any leaf with extent items. Second example where we will miss a hole: *) An inode as its items spanning 5 (or more) leafs; *) A hole is punched and it covers only the extents items of the 3rd leaf. This resulsts in deleting the entire leaf and not touching any of the other leafs. So the only leaf that is modified in the current transaction, when punching the hole, is the first leaf, which contains the inode item. During the full fsync, the only leaf that is passed to copy_items() is that first leaf, and that's not enough for the hole detection code in copy_items() to determine there's a hole between the last file extent item in the 2nd leaf and the first file extent item in the 3rd leaf (which was the 4th leaf before punching the hole). Fix this by scanning all leafs and punch holes as necessary when doing a full fsync (less common than a non-full fsync) when the NO_HOLES feature is enabled. The lack of explicit file extent items to mark holes makes it necessary to scan existing extents to determine if holes exist. A test case for fstests follows soon. Fixes: 16e7549f045d33 ("Btrfs: incompatible format change to remove hole extents") CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-11-19 15:07:33 +03:00
if (!dropped_extents) {
Btrfs: fix duplicate extents after fsync of file with prealloc extents In commit 471d557afed1 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay"), on fsync, we started to always log all prealloc extents beyond an inode's i_size in order to avoid losing them after a power failure. However under some cases this can lead to the log replay code to create duplicate extent items, with different lengths, in the extent tree. That happens because, as of that commit, we can now log extent items based on extent maps that are not on the "modified" list of extent maps of the inode's extent map tree. Logging extent items based on extent maps is used during the fast fsync path to save time and for this to work reliably it requires that the extent maps are not merged with other adjacent extent maps - having the extent maps in the list of modified extents gives such guarantee. Consider the following example, captured during a long run of fsstress, which illustrates this problem. We have inode 271, in the filesystem tree (root 5), for which all of the following operations and discussion apply to. A buffered write starts at offset 312391 with a length of 933471 bytes (end offset at 1245862). At this point we have, for this inode, the following extent maps with the their field values: em A, start 0, orig_start 0, len 40960, block_start 18446744073709551613, block_len 0, orig_block_len 0 em B, start 40960, orig_start 40960, len 376832, block_start 1106399232, block_len 376832, orig_block_len 376832 em C, start 417792, orig_start 417792, len 782336, block_start 18446744073709551613, block_len 0, orig_block_len 0 em D, start 1200128, orig_start 1200128, len 835584, block_start 1106776064, block_len 835584, orig_block_len 835584 em E, start 2035712, orig_start 2035712, len 245760, block_start 1107611648, block_len 245760, orig_block_len 245760 Extent map A corresponds to a hole and extent maps D and E correspond to preallocated extents. Extent map D ends where extent map E begins (1106776064 + 835584 = 1107611648), but these extent maps were not merged because they are in the inode's list of modified extent maps. An fsync against this inode is made, which triggers the fast path (BTRFS_INODE_NEEDS_FULL_SYNC is not set). This fsync triggers writeback of the data previously written using buffered IO, and when the respective ordered extent finishes, btrfs_drop_extents() is called against the (aligned) range 311296..1249279. This causes a split of extent map D at btrfs_drop_extent_cache(), replacing extent map D with a new extent map D', also added to the list of modified extents, with the following values: em D', start 1249280, orig_start of 1200128, block_start 1106825216 (= 1106776064 + 1249280 - 1200128), orig_block_len 835584, block_len 786432 (835584 - (1249280 - 1200128)) Then, during the fast fsync, btrfs_log_changed_extents() is called and extent maps D' and E are removed from the list of modified extents. The flag EXTENT_FLAG_LOGGING is also set on them. After the extents are logged clear_em_logging() is called on each of them, and that makes extent map E to be merged with extent map D' (try_merge_map()), resulting in D' being deleted and E adjusted to: em E, start 1249280, orig_start 1200128, len 1032192, block_start 1106825216, block_len 1032192, orig_block_len 245760 A direct IO write at offset 1847296 and length of 360448 bytes (end offset at 2207744) starts, and at that moment the following extent maps exist for our inode: em A, start 0, orig_start 0, len 40960, block_start 18446744073709551613, block_len 0, orig_block_len 0 em B, start 40960, orig_start 40960, len 270336, block_start 1106399232, block_len 270336, orig_block_len 376832 em C, start 311296, orig_start 311296, len 937984, block_start 1112842240, block_len 937984, orig_block_len 937984 em E (prealloc), start 1249280, orig_start 1200128, len 1032192, block_start 1106825216, block_len 1032192, orig_block_len 245760 The dio write results in drop_extent_cache() being called twice. The first time for a range that starts at offset 1847296 and ends at offset 2035711 (length of 188416), which results in a double split of extent map E, replacing it with two new extent maps: em F, start 1249280, orig_start 1200128, block_start 1106825216, block_len 598016, orig_block_len 598016 em G, start 2035712, orig_start 1200128, block_start 1107611648, block_len 245760, orig_block_len 1032192 It also creates a new extent map that represents a part of the requested IO (through create_io_em()): em H, start 1847296, len 188416, block_start 1107423232, block_len 188416 The second call to drop_extent_cache() has a range with a start offset of 2035712 and end offset of 2207743 (length of 172032). This leads to replacing extent map G with a new extent map I with the following values: em I, start 2207744, orig_start 1200128, block_start 1107783680, block_len 73728, orig_block_len 1032192 It also creates a new extent map that represents the second part of the requested IO (through create_io_em()): em J, start 2035712, len 172032, block_start 1107611648, block_len 172032 The dio write set the inode's i_size to 2207744 bytes. After the dio write the inode has the following extent maps: em A, start 0, orig_start 0, len 40960, block_start 18446744073709551613, block_len 0, orig_block_len 0 em B, start 40960, orig_start 40960, len 270336, block_start 1106399232, block_len 270336, orig_block_len 376832 em C, start 311296, orig_start 311296, len 937984, block_start 1112842240, block_len 937984, orig_block_len 937984 em F, start 1249280, orig_start 1200128, len 598016, block_start 1106825216, block_len 598016, orig_block_len 598016 em H, start 1847296, orig_start 1200128, len 188416, block_start 1107423232, block_len 188416, orig_block_len 835584 em J, start 2035712, orig_start 2035712, len 172032, block_start 1107611648, block_len 172032, orig_block_len 245760 em I, start 2207744, orig_start 1200128, len 73728, block_start 1107783680, block_len 73728, orig_block_len 1032192 Now do some change to the file, like adding a xattr for example and then fsync it again. This triggers a fast fsync path, and as of commit 471d557afed1 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay"), we use the extent map I to log a file extent item because it's a prealloc extent and it starts at an offset matching the inode's i_size. However when we log it, we create a file extent item with a value for the disk byte location that is wrong, as can be seen from the following output of "btrfs inspect-internal dump-tree": item 1 key (271 EXTENT_DATA 2207744) itemoff 3782 itemsize 53 generation 22 type 2 (prealloc) prealloc data disk byte 1106776064 nr 1032192 prealloc data offset 1007616 nr 73728 Here the disk byte value corresponds to calculation based on some fields from the extent map I: 1106776064 = block_start (1107783680) - 1007616 (extent_offset) extent_offset = 2207744 (start) - 1200128 (orig_start) = 1007616 The disk byte value of 1106776064 clashes with disk byte values of the file extent items at offsets 1249280 and 1847296 in the fs tree: item 6 key (271 EXTENT_DATA 1249280) itemoff 3568 itemsize 53 generation 20 type 2 (prealloc) prealloc data disk byte 1106776064 nr 835584 prealloc data offset 49152 nr 598016 item 7 key (271 EXTENT_DATA 1847296) itemoff 3515 itemsize 53 generation 20 type 1 (regular) extent data disk byte 1106776064 nr 835584 extent data offset 647168 nr 188416 ram 835584 extent compression 0 (none) item 8 key (271 EXTENT_DATA 2035712) itemoff 3462 itemsize 53 generation 20 type 1 (regular) extent data disk byte 1107611648 nr 245760 extent data offset 0 nr 172032 ram 245760 extent compression 0 (none) item 9 key (271 EXTENT_DATA 2207744) itemoff 3409 itemsize 53 generation 20 type 2 (prealloc) prealloc data disk byte 1107611648 nr 245760 prealloc data offset 172032 nr 73728 Instead of the disk byte value of 1106776064, the value of 1107611648 should have been logged. Also the data offset value should have been 172032 and not 1007616. After a log replay we end up getting two extent items in the extent tree with different lengths, one of 835584, which is correct and existed before the log replay, and another one of 1032192 which is wrong and is based on the logged file extent item: item 12 key (1106776064 EXTENT_ITEM 835584) itemoff 3406 itemsize 53 refs 2 gen 15 flags DATA extent data backref root 5 objectid 271 offset 1200128 count 2 item 13 key (1106776064 EXTENT_ITEM 1032192) itemoff 3353 itemsize 53 refs 1 gen 22 flags DATA extent data backref root 5 objectid 271 offset 1200128 count 1 Obviously this leads to many problems and a filesystem check reports many errors: (...) checking extents Extent back ref already exists for 1106776064 parent 0 root 5 owner 271 offset 1200128 num_refs 1 extent item 1106776064 has multiple extent items ref mismatch on [1106776064 835584] extent item 2, found 3 Incorrect local backref count on 1106776064 root 5 owner 271 offset 1200128 found 2 wanted 1 back 0x55b1d0ad7680 Backref 1106776064 root 5 owner 271 offset 1200128 num_refs 0 not found in extent tree Incorrect local backref count on 1106776064 root 5 owner 271 offset 1200128 found 1 wanted 0 back 0x55b1d0ad4e70 Backref bytes do not match extent backref, bytenr=1106776064, ref bytes=835584, backref bytes=1032192 backpointer mismatch on [1106776064 835584] checking free space cache block group 1103101952 has wrong amount of free space failed to load free space cache for block group 1103101952 checking fs roots (...) So fix this by logging the prealloc extents beyond the inode's i_size based on searches in the subvolume tree instead of the extent maps. Fixes: 471d557afed1 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay") CC: stable@vger.kernel.org # 4.14+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2018-05-09 18:01:46 +03:00
/*
* Avoid logging extent items logged in past fsync calls
* and leading to duplicate keys in the log tree.
*/
ret = truncate_inode_items(trans, root->log_root, inode,
truncate_offset,
BTRFS_EXTENT_DATA_KEY);
Btrfs: fix duplicate extents after fsync of file with prealloc extents In commit 471d557afed1 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay"), on fsync, we started to always log all prealloc extents beyond an inode's i_size in order to avoid losing them after a power failure. However under some cases this can lead to the log replay code to create duplicate extent items, with different lengths, in the extent tree. That happens because, as of that commit, we can now log extent items based on extent maps that are not on the "modified" list of extent maps of the inode's extent map tree. Logging extent items based on extent maps is used during the fast fsync path to save time and for this to work reliably it requires that the extent maps are not merged with other adjacent extent maps - having the extent maps in the list of modified extents gives such guarantee. Consider the following example, captured during a long run of fsstress, which illustrates this problem. We have inode 271, in the filesystem tree (root 5), for which all of the following operations and discussion apply to. A buffered write starts at offset 312391 with a length of 933471 bytes (end offset at 1245862). At this point we have, for this inode, the following extent maps with the their field values: em A, start 0, orig_start 0, len 40960, block_start 18446744073709551613, block_len 0, orig_block_len 0 em B, start 40960, orig_start 40960, len 376832, block_start 1106399232, block_len 376832, orig_block_len 376832 em C, start 417792, orig_start 417792, len 782336, block_start 18446744073709551613, block_len 0, orig_block_len 0 em D, start 1200128, orig_start 1200128, len 835584, block_start 1106776064, block_len 835584, orig_block_len 835584 em E, start 2035712, orig_start 2035712, len 245760, block_start 1107611648, block_len 245760, orig_block_len 245760 Extent map A corresponds to a hole and extent maps D and E correspond to preallocated extents. Extent map D ends where extent map E begins (1106776064 + 835584 = 1107611648), but these extent maps were not merged because they are in the inode's list of modified extent maps. An fsync against this inode is made, which triggers the fast path (BTRFS_INODE_NEEDS_FULL_SYNC is not set). This fsync triggers writeback of the data previously written using buffered IO, and when the respective ordered extent finishes, btrfs_drop_extents() is called against the (aligned) range 311296..1249279. This causes a split of extent map D at btrfs_drop_extent_cache(), replacing extent map D with a new extent map D', also added to the list of modified extents, with the following values: em D', start 1249280, orig_start of 1200128, block_start 1106825216 (= 1106776064 + 1249280 - 1200128), orig_block_len 835584, block_len 786432 (835584 - (1249280 - 1200128)) Then, during the fast fsync, btrfs_log_changed_extents() is called and extent maps D' and E are removed from the list of modified extents. The flag EXTENT_FLAG_LOGGING is also set on them. After the extents are logged clear_em_logging() is called on each of them, and that makes extent map E to be merged with extent map D' (try_merge_map()), resulting in D' being deleted and E adjusted to: em E, start 1249280, orig_start 1200128, len 1032192, block_start 1106825216, block_len 1032192, orig_block_len 245760 A direct IO write at offset 1847296 and length of 360448 bytes (end offset at 2207744) starts, and at that moment the following extent maps exist for our inode: em A, start 0, orig_start 0, len 40960, block_start 18446744073709551613, block_len 0, orig_block_len 0 em B, start 40960, orig_start 40960, len 270336, block_start 1106399232, block_len 270336, orig_block_len 376832 em C, start 311296, orig_start 311296, len 937984, block_start 1112842240, block_len 937984, orig_block_len 937984 em E (prealloc), start 1249280, orig_start 1200128, len 1032192, block_start 1106825216, block_len 1032192, orig_block_len 245760 The dio write results in drop_extent_cache() being called twice. The first time for a range that starts at offset 1847296 and ends at offset 2035711 (length of 188416), which results in a double split of extent map E, replacing it with two new extent maps: em F, start 1249280, orig_start 1200128, block_start 1106825216, block_len 598016, orig_block_len 598016 em G, start 2035712, orig_start 1200128, block_start 1107611648, block_len 245760, orig_block_len 1032192 It also creates a new extent map that represents a part of the requested IO (through create_io_em()): em H, start 1847296, len 188416, block_start 1107423232, block_len 188416 The second call to drop_extent_cache() has a range with a start offset of 2035712 and end offset of 2207743 (length of 172032). This leads to replacing extent map G with a new extent map I with the following values: em I, start 2207744, orig_start 1200128, block_start 1107783680, block_len 73728, orig_block_len 1032192 It also creates a new extent map that represents the second part of the requested IO (through create_io_em()): em J, start 2035712, len 172032, block_start 1107611648, block_len 172032 The dio write set the inode's i_size to 2207744 bytes. After the dio write the inode has the following extent maps: em A, start 0, orig_start 0, len 40960, block_start 18446744073709551613, block_len 0, orig_block_len 0 em B, start 40960, orig_start 40960, len 270336, block_start 1106399232, block_len 270336, orig_block_len 376832 em C, start 311296, orig_start 311296, len 937984, block_start 1112842240, block_len 937984, orig_block_len 937984 em F, start 1249280, orig_start 1200128, len 598016, block_start 1106825216, block_len 598016, orig_block_len 598016 em H, start 1847296, orig_start 1200128, len 188416, block_start 1107423232, block_len 188416, orig_block_len 835584 em J, start 2035712, orig_start 2035712, len 172032, block_start 1107611648, block_len 172032, orig_block_len 245760 em I, start 2207744, orig_start 1200128, len 73728, block_start 1107783680, block_len 73728, orig_block_len 1032192 Now do some change to the file, like adding a xattr for example and then fsync it again. This triggers a fast fsync path, and as of commit 471d557afed1 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay"), we use the extent map I to log a file extent item because it's a prealloc extent and it starts at an offset matching the inode's i_size. However when we log it, we create a file extent item with a value for the disk byte location that is wrong, as can be seen from the following output of "btrfs inspect-internal dump-tree": item 1 key (271 EXTENT_DATA 2207744) itemoff 3782 itemsize 53 generation 22 type 2 (prealloc) prealloc data disk byte 1106776064 nr 1032192 prealloc data offset 1007616 nr 73728 Here the disk byte value corresponds to calculation based on some fields from the extent map I: 1106776064 = block_start (1107783680) - 1007616 (extent_offset) extent_offset = 2207744 (start) - 1200128 (orig_start) = 1007616 The disk byte value of 1106776064 clashes with disk byte values of the file extent items at offsets 1249280 and 1847296 in the fs tree: item 6 key (271 EXTENT_DATA 1249280) itemoff 3568 itemsize 53 generation 20 type 2 (prealloc) prealloc data disk byte 1106776064 nr 835584 prealloc data offset 49152 nr 598016 item 7 key (271 EXTENT_DATA 1847296) itemoff 3515 itemsize 53 generation 20 type 1 (regular) extent data disk byte 1106776064 nr 835584 extent data offset 647168 nr 188416 ram 835584 extent compression 0 (none) item 8 key (271 EXTENT_DATA 2035712) itemoff 3462 itemsize 53 generation 20 type 1 (regular) extent data disk byte 1107611648 nr 245760 extent data offset 0 nr 172032 ram 245760 extent compression 0 (none) item 9 key (271 EXTENT_DATA 2207744) itemoff 3409 itemsize 53 generation 20 type 2 (prealloc) prealloc data disk byte 1107611648 nr 245760 prealloc data offset 172032 nr 73728 Instead of the disk byte value of 1106776064, the value of 1107611648 should have been logged. Also the data offset value should have been 172032 and not 1007616. After a log replay we end up getting two extent items in the extent tree with different lengths, one of 835584, which is correct and existed before the log replay, and another one of 1032192 which is wrong and is based on the logged file extent item: item 12 key (1106776064 EXTENT_ITEM 835584) itemoff 3406 itemsize 53 refs 2 gen 15 flags DATA extent data backref root 5 objectid 271 offset 1200128 count 2 item 13 key (1106776064 EXTENT_ITEM 1032192) itemoff 3353 itemsize 53 refs 1 gen 22 flags DATA extent data backref root 5 objectid 271 offset 1200128 count 1 Obviously this leads to many problems and a filesystem check reports many errors: (...) checking extents Extent back ref already exists for 1106776064 parent 0 root 5 owner 271 offset 1200128 num_refs 1 extent item 1106776064 has multiple extent items ref mismatch on [1106776064 835584] extent item 2, found 3 Incorrect local backref count on 1106776064 root 5 owner 271 offset 1200128 found 2 wanted 1 back 0x55b1d0ad7680 Backref 1106776064 root 5 owner 271 offset 1200128 num_refs 0 not found in extent tree Incorrect local backref count on 1106776064 root 5 owner 271 offset 1200128 found 1 wanted 0 back 0x55b1d0ad4e70 Backref bytes do not match extent backref, bytenr=1106776064, ref bytes=835584, backref bytes=1032192 backpointer mismatch on [1106776064 835584] checking free space cache block group 1103101952 has wrong amount of free space failed to load free space cache for block group 1103101952 checking fs roots (...) So fix this by logging the prealloc extents beyond the inode's i_size based on searches in the subvolume tree instead of the extent maps. Fixes: 471d557afed1 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay") CC: stable@vger.kernel.org # 4.14+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2018-05-09 18:01:46 +03:00
if (ret)
goto out;
Btrfs: fix missing hole after hole punching and fsync when using NO_HOLES When using the NO_HOLES feature, if we punch a hole into a file and then fsync it, there are cases where a subsequent fsync will miss the fact that a hole was punched, resulting in the holes not existing after replaying the log tree. Essentially these cases all imply that, tree-log.c:copy_items(), is not invoked for the leafs that delimit holes, because nothing changed those leafs in the current transaction. And it's precisely copy_items() where we currenly detect and log holes, which works as long as the holes are between file extent items in the input leaf or between the beginning of input leaf and the previous leaf or between the last item in the leaf and the next leaf. First example where we miss a hole: *) The extent items of the inode span multiple leafs; *) The punched hole covers a range that affects only the extent items of the first leaf; *) The fsync operation is done in full mode (BTRFS_INODE_NEEDS_FULL_SYNC is set in the inode's runtime flags). That results in the hole not existing after replaying the log tree. For example, if the fs/subvolume tree has the following layout for a particular inode: Leaf N, generation 10: [ ... INODE_ITEM INODE_REF EXTENT_ITEM (0 64K) EXTENT_ITEM (64K 128K) ] Leaf N + 1, generation 10: [ EXTENT_ITEM (128K 64K) ... ] If at transaction 11 we punch a hole coverting the range [0, 128K[, we end up dropping the two extent items from leaf N, but we don't touch the other leaf, so we end up in the following state: Leaf N, generation 11: [ ... INODE_ITEM INODE_REF ] Leaf N + 1, generation 10: [ EXTENT_ITEM (128K 64K) ... ] A full fsync after punching the hole will only process leaf N because it was modified in the current transaction, but not leaf N + 1, since it was not modified in the current transaction (generation 10 and not 11). As a result the fsync will not log any holes, because it didn't process any leaf with extent items. Second example where we will miss a hole: *) An inode as its items spanning 5 (or more) leafs; *) A hole is punched and it covers only the extents items of the 3rd leaf. This resulsts in deleting the entire leaf and not touching any of the other leafs. So the only leaf that is modified in the current transaction, when punching the hole, is the first leaf, which contains the inode item. During the full fsync, the only leaf that is passed to copy_items() is that first leaf, and that's not enough for the hole detection code in copy_items() to determine there's a hole between the last file extent item in the 2nd leaf and the first file extent item in the 3rd leaf (which was the 4th leaf before punching the hole). Fix this by scanning all leafs and punch holes as necessary when doing a full fsync (less common than a non-full fsync) when the NO_HOLES feature is enabled. The lack of explicit file extent items to mark holes makes it necessary to scan existing extents to determine if holes exist. A test case for fstests follows soon. Fixes: 16e7549f045d33 ("Btrfs: incompatible format change to remove hole extents") CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-11-19 15:07:33 +03:00
dropped_extents = true;
Btrfs: fix duplicate extents after fsync of file with prealloc extents In commit 471d557afed1 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay"), on fsync, we started to always log all prealloc extents beyond an inode's i_size in order to avoid losing them after a power failure. However under some cases this can lead to the log replay code to create duplicate extent items, with different lengths, in the extent tree. That happens because, as of that commit, we can now log extent items based on extent maps that are not on the "modified" list of extent maps of the inode's extent map tree. Logging extent items based on extent maps is used during the fast fsync path to save time and for this to work reliably it requires that the extent maps are not merged with other adjacent extent maps - having the extent maps in the list of modified extents gives such guarantee. Consider the following example, captured during a long run of fsstress, which illustrates this problem. We have inode 271, in the filesystem tree (root 5), for which all of the following operations and discussion apply to. A buffered write starts at offset 312391 with a length of 933471 bytes (end offset at 1245862). At this point we have, for this inode, the following extent maps with the their field values: em A, start 0, orig_start 0, len 40960, block_start 18446744073709551613, block_len 0, orig_block_len 0 em B, start 40960, orig_start 40960, len 376832, block_start 1106399232, block_len 376832, orig_block_len 376832 em C, start 417792, orig_start 417792, len 782336, block_start 18446744073709551613, block_len 0, orig_block_len 0 em D, start 1200128, orig_start 1200128, len 835584, block_start 1106776064, block_len 835584, orig_block_len 835584 em E, start 2035712, orig_start 2035712, len 245760, block_start 1107611648, block_len 245760, orig_block_len 245760 Extent map A corresponds to a hole and extent maps D and E correspond to preallocated extents. Extent map D ends where extent map E begins (1106776064 + 835584 = 1107611648), but these extent maps were not merged because they are in the inode's list of modified extent maps. An fsync against this inode is made, which triggers the fast path (BTRFS_INODE_NEEDS_FULL_SYNC is not set). This fsync triggers writeback of the data previously written using buffered IO, and when the respective ordered extent finishes, btrfs_drop_extents() is called against the (aligned) range 311296..1249279. This causes a split of extent map D at btrfs_drop_extent_cache(), replacing extent map D with a new extent map D', also added to the list of modified extents, with the following values: em D', start 1249280, orig_start of 1200128, block_start 1106825216 (= 1106776064 + 1249280 - 1200128), orig_block_len 835584, block_len 786432 (835584 - (1249280 - 1200128)) Then, during the fast fsync, btrfs_log_changed_extents() is called and extent maps D' and E are removed from the list of modified extents. The flag EXTENT_FLAG_LOGGING is also set on them. After the extents are logged clear_em_logging() is called on each of them, and that makes extent map E to be merged with extent map D' (try_merge_map()), resulting in D' being deleted and E adjusted to: em E, start 1249280, orig_start 1200128, len 1032192, block_start 1106825216, block_len 1032192, orig_block_len 245760 A direct IO write at offset 1847296 and length of 360448 bytes (end offset at 2207744) starts, and at that moment the following extent maps exist for our inode: em A, start 0, orig_start 0, len 40960, block_start 18446744073709551613, block_len 0, orig_block_len 0 em B, start 40960, orig_start 40960, len 270336, block_start 1106399232, block_len 270336, orig_block_len 376832 em C, start 311296, orig_start 311296, len 937984, block_start 1112842240, block_len 937984, orig_block_len 937984 em E (prealloc), start 1249280, orig_start 1200128, len 1032192, block_start 1106825216, block_len 1032192, orig_block_len 245760 The dio write results in drop_extent_cache() being called twice. The first time for a range that starts at offset 1847296 and ends at offset 2035711 (length of 188416), which results in a double split of extent map E, replacing it with two new extent maps: em F, start 1249280, orig_start 1200128, block_start 1106825216, block_len 598016, orig_block_len 598016 em G, start 2035712, orig_start 1200128, block_start 1107611648, block_len 245760, orig_block_len 1032192 It also creates a new extent map that represents a part of the requested IO (through create_io_em()): em H, start 1847296, len 188416, block_start 1107423232, block_len 188416 The second call to drop_extent_cache() has a range with a start offset of 2035712 and end offset of 2207743 (length of 172032). This leads to replacing extent map G with a new extent map I with the following values: em I, start 2207744, orig_start 1200128, block_start 1107783680, block_len 73728, orig_block_len 1032192 It also creates a new extent map that represents the second part of the requested IO (through create_io_em()): em J, start 2035712, len 172032, block_start 1107611648, block_len 172032 The dio write set the inode's i_size to 2207744 bytes. After the dio write the inode has the following extent maps: em A, start 0, orig_start 0, len 40960, block_start 18446744073709551613, block_len 0, orig_block_len 0 em B, start 40960, orig_start 40960, len 270336, block_start 1106399232, block_len 270336, orig_block_len 376832 em C, start 311296, orig_start 311296, len 937984, block_start 1112842240, block_len 937984, orig_block_len 937984 em F, start 1249280, orig_start 1200128, len 598016, block_start 1106825216, block_len 598016, orig_block_len 598016 em H, start 1847296, orig_start 1200128, len 188416, block_start 1107423232, block_len 188416, orig_block_len 835584 em J, start 2035712, orig_start 2035712, len 172032, block_start 1107611648, block_len 172032, orig_block_len 245760 em I, start 2207744, orig_start 1200128, len 73728, block_start 1107783680, block_len 73728, orig_block_len 1032192 Now do some change to the file, like adding a xattr for example and then fsync it again. This triggers a fast fsync path, and as of commit 471d557afed1 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay"), we use the extent map I to log a file extent item because it's a prealloc extent and it starts at an offset matching the inode's i_size. However when we log it, we create a file extent item with a value for the disk byte location that is wrong, as can be seen from the following output of "btrfs inspect-internal dump-tree": item 1 key (271 EXTENT_DATA 2207744) itemoff 3782 itemsize 53 generation 22 type 2 (prealloc) prealloc data disk byte 1106776064 nr 1032192 prealloc data offset 1007616 nr 73728 Here the disk byte value corresponds to calculation based on some fields from the extent map I: 1106776064 = block_start (1107783680) - 1007616 (extent_offset) extent_offset = 2207744 (start) - 1200128 (orig_start) = 1007616 The disk byte value of 1106776064 clashes with disk byte values of the file extent items at offsets 1249280 and 1847296 in the fs tree: item 6 key (271 EXTENT_DATA 1249280) itemoff 3568 itemsize 53 generation 20 type 2 (prealloc) prealloc data disk byte 1106776064 nr 835584 prealloc data offset 49152 nr 598016 item 7 key (271 EXTENT_DATA 1847296) itemoff 3515 itemsize 53 generation 20 type 1 (regular) extent data disk byte 1106776064 nr 835584 extent data offset 647168 nr 188416 ram 835584 extent compression 0 (none) item 8 key (271 EXTENT_DATA 2035712) itemoff 3462 itemsize 53 generation 20 type 1 (regular) extent data disk byte 1107611648 nr 245760 extent data offset 0 nr 172032 ram 245760 extent compression 0 (none) item 9 key (271 EXTENT_DATA 2207744) itemoff 3409 itemsize 53 generation 20 type 2 (prealloc) prealloc data disk byte 1107611648 nr 245760 prealloc data offset 172032 nr 73728 Instead of the disk byte value of 1106776064, the value of 1107611648 should have been logged. Also the data offset value should have been 172032 and not 1007616. After a log replay we end up getting two extent items in the extent tree with different lengths, one of 835584, which is correct and existed before the log replay, and another one of 1032192 which is wrong and is based on the logged file extent item: item 12 key (1106776064 EXTENT_ITEM 835584) itemoff 3406 itemsize 53 refs 2 gen 15 flags DATA extent data backref root 5 objectid 271 offset 1200128 count 2 item 13 key (1106776064 EXTENT_ITEM 1032192) itemoff 3353 itemsize 53 refs 1 gen 22 flags DATA extent data backref root 5 objectid 271 offset 1200128 count 1 Obviously this leads to many problems and a filesystem check reports many errors: (...) checking extents Extent back ref already exists for 1106776064 parent 0 root 5 owner 271 offset 1200128 num_refs 1 extent item 1106776064 has multiple extent items ref mismatch on [1106776064 835584] extent item 2, found 3 Incorrect local backref count on 1106776064 root 5 owner 271 offset 1200128 found 2 wanted 1 back 0x55b1d0ad7680 Backref 1106776064 root 5 owner 271 offset 1200128 num_refs 0 not found in extent tree Incorrect local backref count on 1106776064 root 5 owner 271 offset 1200128 found 1 wanted 0 back 0x55b1d0ad4e70 Backref bytes do not match extent backref, bytenr=1106776064, ref bytes=835584, backref bytes=1032192 backpointer mismatch on [1106776064 835584] checking free space cache block group 1103101952 has wrong amount of free space failed to load free space cache for block group 1103101952 checking fs roots (...) So fix this by logging the prealloc extents beyond the inode's i_size based on searches in the subvolume tree instead of the extent maps. Fixes: 471d557afed1 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay") CC: stable@vger.kernel.org # 4.14+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2018-05-09 18:01:46 +03:00
}
if (ins_nr == 0)
start_slot = slot;
ins_nr++;
path->slots[0]++;
if (!dst_path) {
dst_path = btrfs_alloc_path();
if (!dst_path) {
ret = -ENOMEM;
goto out;
}
}
}
if (ins_nr > 0)
Btrfs: fix missing hole after hole punching and fsync when using NO_HOLES When using the NO_HOLES feature, if we punch a hole into a file and then fsync it, there are cases where a subsequent fsync will miss the fact that a hole was punched, resulting in the holes not existing after replaying the log tree. Essentially these cases all imply that, tree-log.c:copy_items(), is not invoked for the leafs that delimit holes, because nothing changed those leafs in the current transaction. And it's precisely copy_items() where we currenly detect and log holes, which works as long as the holes are between file extent items in the input leaf or between the beginning of input leaf and the previous leaf or between the last item in the leaf and the next leaf. First example where we miss a hole: *) The extent items of the inode span multiple leafs; *) The punched hole covers a range that affects only the extent items of the first leaf; *) The fsync operation is done in full mode (BTRFS_INODE_NEEDS_FULL_SYNC is set in the inode's runtime flags). That results in the hole not existing after replaying the log tree. For example, if the fs/subvolume tree has the following layout for a particular inode: Leaf N, generation 10: [ ... INODE_ITEM INODE_REF EXTENT_ITEM (0 64K) EXTENT_ITEM (64K 128K) ] Leaf N + 1, generation 10: [ EXTENT_ITEM (128K 64K) ... ] If at transaction 11 we punch a hole coverting the range [0, 128K[, we end up dropping the two extent items from leaf N, but we don't touch the other leaf, so we end up in the following state: Leaf N, generation 11: [ ... INODE_ITEM INODE_REF ] Leaf N + 1, generation 10: [ EXTENT_ITEM (128K 64K) ... ] A full fsync after punching the hole will only process leaf N because it was modified in the current transaction, but not leaf N + 1, since it was not modified in the current transaction (generation 10 and not 11). As a result the fsync will not log any holes, because it didn't process any leaf with extent items. Second example where we will miss a hole: *) An inode as its items spanning 5 (or more) leafs; *) A hole is punched and it covers only the extents items of the 3rd leaf. This resulsts in deleting the entire leaf and not touching any of the other leafs. So the only leaf that is modified in the current transaction, when punching the hole, is the first leaf, which contains the inode item. During the full fsync, the only leaf that is passed to copy_items() is that first leaf, and that's not enough for the hole detection code in copy_items() to determine there's a hole between the last file extent item in the 2nd leaf and the first file extent item in the 3rd leaf (which was the 4th leaf before punching the hole). Fix this by scanning all leafs and punch holes as necessary when doing a full fsync (less common than a non-full fsync) when the NO_HOLES feature is enabled. The lack of explicit file extent items to mark holes makes it necessary to scan existing extents to determine if holes exist. A test case for fstests follows soon. Fixes: 16e7549f045d33 ("Btrfs: incompatible format change to remove hole extents") CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-11-19 15:07:33 +03:00
ret = copy_items(trans, inode, dst_path, path,
Btrfs: fix duplicate extents after fsync of file with prealloc extents In commit 471d557afed1 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay"), on fsync, we started to always log all prealloc extents beyond an inode's i_size in order to avoid losing them after a power failure. However under some cases this can lead to the log replay code to create duplicate extent items, with different lengths, in the extent tree. That happens because, as of that commit, we can now log extent items based on extent maps that are not on the "modified" list of extent maps of the inode's extent map tree. Logging extent items based on extent maps is used during the fast fsync path to save time and for this to work reliably it requires that the extent maps are not merged with other adjacent extent maps - having the extent maps in the list of modified extents gives such guarantee. Consider the following example, captured during a long run of fsstress, which illustrates this problem. We have inode 271, in the filesystem tree (root 5), for which all of the following operations and discussion apply to. A buffered write starts at offset 312391 with a length of 933471 bytes (end offset at 1245862). At this point we have, for this inode, the following extent maps with the their field values: em A, start 0, orig_start 0, len 40960, block_start 18446744073709551613, block_len 0, orig_block_len 0 em B, start 40960, orig_start 40960, len 376832, block_start 1106399232, block_len 376832, orig_block_len 376832 em C, start 417792, orig_start 417792, len 782336, block_start 18446744073709551613, block_len 0, orig_block_len 0 em D, start 1200128, orig_start 1200128, len 835584, block_start 1106776064, block_len 835584, orig_block_len 835584 em E, start 2035712, orig_start 2035712, len 245760, block_start 1107611648, block_len 245760, orig_block_len 245760 Extent map A corresponds to a hole and extent maps D and E correspond to preallocated extents. Extent map D ends where extent map E begins (1106776064 + 835584 = 1107611648), but these extent maps were not merged because they are in the inode's list of modified extent maps. An fsync against this inode is made, which triggers the fast path (BTRFS_INODE_NEEDS_FULL_SYNC is not set). This fsync triggers writeback of the data previously written using buffered IO, and when the respective ordered extent finishes, btrfs_drop_extents() is called against the (aligned) range 311296..1249279. This causes a split of extent map D at btrfs_drop_extent_cache(), replacing extent map D with a new extent map D', also added to the list of modified extents, with the following values: em D', start 1249280, orig_start of 1200128, block_start 1106825216 (= 1106776064 + 1249280 - 1200128), orig_block_len 835584, block_len 786432 (835584 - (1249280 - 1200128)) Then, during the fast fsync, btrfs_log_changed_extents() is called and extent maps D' and E are removed from the list of modified extents. The flag EXTENT_FLAG_LOGGING is also set on them. After the extents are logged clear_em_logging() is called on each of them, and that makes extent map E to be merged with extent map D' (try_merge_map()), resulting in D' being deleted and E adjusted to: em E, start 1249280, orig_start 1200128, len 1032192, block_start 1106825216, block_len 1032192, orig_block_len 245760 A direct IO write at offset 1847296 and length of 360448 bytes (end offset at 2207744) starts, and at that moment the following extent maps exist for our inode: em A, start 0, orig_start 0, len 40960, block_start 18446744073709551613, block_len 0, orig_block_len 0 em B, start 40960, orig_start 40960, len 270336, block_start 1106399232, block_len 270336, orig_block_len 376832 em C, start 311296, orig_start 311296, len 937984, block_start 1112842240, block_len 937984, orig_block_len 937984 em E (prealloc), start 1249280, orig_start 1200128, len 1032192, block_start 1106825216, block_len 1032192, orig_block_len 245760 The dio write results in drop_extent_cache() being called twice. The first time for a range that starts at offset 1847296 and ends at offset 2035711 (length of 188416), which results in a double split of extent map E, replacing it with two new extent maps: em F, start 1249280, orig_start 1200128, block_start 1106825216, block_len 598016, orig_block_len 598016 em G, start 2035712, orig_start 1200128, block_start 1107611648, block_len 245760, orig_block_len 1032192 It also creates a new extent map that represents a part of the requested IO (through create_io_em()): em H, start 1847296, len 188416, block_start 1107423232, block_len 188416 The second call to drop_extent_cache() has a range with a start offset of 2035712 and end offset of 2207743 (length of 172032). This leads to replacing extent map G with a new extent map I with the following values: em I, start 2207744, orig_start 1200128, block_start 1107783680, block_len 73728, orig_block_len 1032192 It also creates a new extent map that represents the second part of the requested IO (through create_io_em()): em J, start 2035712, len 172032, block_start 1107611648, block_len 172032 The dio write set the inode's i_size to 2207744 bytes. After the dio write the inode has the following extent maps: em A, start 0, orig_start 0, len 40960, block_start 18446744073709551613, block_len 0, orig_block_len 0 em B, start 40960, orig_start 40960, len 270336, block_start 1106399232, block_len 270336, orig_block_len 376832 em C, start 311296, orig_start 311296, len 937984, block_start 1112842240, block_len 937984, orig_block_len 937984 em F, start 1249280, orig_start 1200128, len 598016, block_start 1106825216, block_len 598016, orig_block_len 598016 em H, start 1847296, orig_start 1200128, len 188416, block_start 1107423232, block_len 188416, orig_block_len 835584 em J, start 2035712, orig_start 2035712, len 172032, block_start 1107611648, block_len 172032, orig_block_len 245760 em I, start 2207744, orig_start 1200128, len 73728, block_start 1107783680, block_len 73728, orig_block_len 1032192 Now do some change to the file, like adding a xattr for example and then fsync it again. This triggers a fast fsync path, and as of commit 471d557afed1 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay"), we use the extent map I to log a file extent item because it's a prealloc extent and it starts at an offset matching the inode's i_size. However when we log it, we create a file extent item with a value for the disk byte location that is wrong, as can be seen from the following output of "btrfs inspect-internal dump-tree": item 1 key (271 EXTENT_DATA 2207744) itemoff 3782 itemsize 53 generation 22 type 2 (prealloc) prealloc data disk byte 1106776064 nr 1032192 prealloc data offset 1007616 nr 73728 Here the disk byte value corresponds to calculation based on some fields from the extent map I: 1106776064 = block_start (1107783680) - 1007616 (extent_offset) extent_offset = 2207744 (start) - 1200128 (orig_start) = 1007616 The disk byte value of 1106776064 clashes with disk byte values of the file extent items at offsets 1249280 and 1847296 in the fs tree: item 6 key (271 EXTENT_DATA 1249280) itemoff 3568 itemsize 53 generation 20 type 2 (prealloc) prealloc data disk byte 1106776064 nr 835584 prealloc data offset 49152 nr 598016 item 7 key (271 EXTENT_DATA 1847296) itemoff 3515 itemsize 53 generation 20 type 1 (regular) extent data disk byte 1106776064 nr 835584 extent data offset 647168 nr 188416 ram 835584 extent compression 0 (none) item 8 key (271 EXTENT_DATA 2035712) itemoff 3462 itemsize 53 generation 20 type 1 (regular) extent data disk byte 1107611648 nr 245760 extent data offset 0 nr 172032 ram 245760 extent compression 0 (none) item 9 key (271 EXTENT_DATA 2207744) itemoff 3409 itemsize 53 generation 20 type 2 (prealloc) prealloc data disk byte 1107611648 nr 245760 prealloc data offset 172032 nr 73728 Instead of the disk byte value of 1106776064, the value of 1107611648 should have been logged. Also the data offset value should have been 172032 and not 1007616. After a log replay we end up getting two extent items in the extent tree with different lengths, one of 835584, which is correct and existed before the log replay, and another one of 1032192 which is wrong and is based on the logged file extent item: item 12 key (1106776064 EXTENT_ITEM 835584) itemoff 3406 itemsize 53 refs 2 gen 15 flags DATA extent data backref root 5 objectid 271 offset 1200128 count 2 item 13 key (1106776064 EXTENT_ITEM 1032192) itemoff 3353 itemsize 53 refs 1 gen 22 flags DATA extent data backref root 5 objectid 271 offset 1200128 count 1 Obviously this leads to many problems and a filesystem check reports many errors: (...) checking extents Extent back ref already exists for 1106776064 parent 0 root 5 owner 271 offset 1200128 num_refs 1 extent item 1106776064 has multiple extent items ref mismatch on [1106776064 835584] extent item 2, found 3 Incorrect local backref count on 1106776064 root 5 owner 271 offset 1200128 found 2 wanted 1 back 0x55b1d0ad7680 Backref 1106776064 root 5 owner 271 offset 1200128 num_refs 0 not found in extent tree Incorrect local backref count on 1106776064 root 5 owner 271 offset 1200128 found 1 wanted 0 back 0x55b1d0ad4e70 Backref bytes do not match extent backref, bytenr=1106776064, ref bytes=835584, backref bytes=1032192 backpointer mismatch on [1106776064 835584] checking free space cache block group 1103101952 has wrong amount of free space failed to load free space cache for block group 1103101952 checking fs roots (...) So fix this by logging the prealloc extents beyond the inode's i_size based on searches in the subvolume tree instead of the extent maps. Fixes: 471d557afed1 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay") CC: stable@vger.kernel.org # 4.14+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2018-05-09 18:01:46 +03:00
start_slot, ins_nr, 1, 0);
out:
btrfs_release_path(path);
btrfs_free_path(dst_path);
return ret;
}
Btrfs: turbo charge fsync At least for the vm workload. Currently on fsync we will 1) Truncate all items in the log tree for the given inode if they exist and 2) Copy all items for a given inode into the log The problem with this is that for things like VMs you can have lots of extents from the fragmented writing behavior, and worst yet you may have only modified a few extents, not the entire thing. This patch fixes this problem by tracking which transid modified our extent, and then when we do the tree logging we find all of the extents we've modified in our current transaction, sort them and commit them. We also only truncate up to the xattrs of the inode and copy that stuff in normally, and then just drop any extents in the range we have that exist in the log already. Here are some numbers of a 50 meg fio job that does random writes and fsync()s after every write Original Patched SATA drive 82KB/s 140KB/s Fusion drive 431KB/s 2532KB/s So around 2-6 times faster depending on your hardware. There are a few corner cases, for example if you truncate at all we have to do it the old way since there is no way to be sure what is in the log is ok. This probably could be done smarter, but if you write-fsync-truncate-write-fsync you deserve what you get. All this work is in RAM of course so if your inode gets evicted from cache and you read it in and fsync it we'll do it the slow way if we are still in the same transaction that we last modified the inode in. The biggest cool part of this is that it requires no changes to the recovery code, so if you fsync with this patch and crash and load an old kernel, it will run the recovery and be a-ok. I have tested this pretty thoroughly with an fsync tester and everything comes back fine, as well as xfstests. Thanks, Signed-off-by: Josef Bacik <jbacik@fusionio.com>
2012-08-17 21:14:17 +04:00
static int btrfs_log_changed_extents(struct btrfs_trans_handle *trans,
struct btrfs_inode *inode,
struct btrfs_path *path,
btrfs: make fast fsyncs wait only for writeback Currently regardless of a full or a fast fsync we always wait for ordered extents to complete, and then start logging the inode after that. However for fast fsyncs we can just wait for the writeback to complete, we don't need to wait for the ordered extents to complete since we use the list of modified extents maps to figure out which extents we must log and we can get their checksums directly from the ordered extents that are still in flight, otherwise look them up from the checksums tree. Until commit b5e6c3e170b770 ("btrfs: always wait on ordered extents at fsync time"), for fast fsyncs, we used to start logging without even waiting for the writeback to complete first, we would wait for it to complete after logging, while holding a transaction open, which lead to performance issues when using cgroups and probably for other cases too, as wait for IO while holding a transaction handle should be avoided as much as possible. After that, for fast fsyncs, we started to wait for ordered extents to complete before starting to log, which adds some latency to fsyncs and we even got at least one report about a performance drop which bisected to that particular change: https://lore.kernel.org/linux-btrfs/20181109215148.GF23260@techsingularity.net/ This change makes fast fsyncs only wait for writeback to finish before starting to log the inode, instead of waiting for both the writeback to finish and for the ordered extents to complete. This brings back part of the logic we had that extracts checksums from in flight ordered extents, which are not yet in the checksums tree, and making sure transaction commits wait for the completion of ordered extents previously logged (by far most of the time they have already completed by the time a transaction commit starts, resulting in no wait at all), to avoid any data loss if an ordered extent completes after the transaction used to log an inode is committed, followed by a power failure. When there are no other tasks accessing the checksums and the subvolume btrees, the ordered extent completion is pretty fast, typically taking 100 to 200 microseconds only in my observations. However when there are other tasks accessing these btrees, ordered extent completion can take a lot more time due to lock contention on nodes and leaves of these btrees. I've seen cases over 2 milliseconds, which starts to be significant. In particular when we do have concurrent fsyncs against different files there is a lot of contention on the checksums btree, since we have many tasks writing the checksums into the btree and other tasks that already started the logging phase are doing lookups for checksums in the btree. This change also turns all ranged fsyncs into full ranged fsyncs, which is something we already did when not using the NO_HOLES features or when doing a full fsync. This is to guarantee we never miss checksums due to writeback having been triggered only for a part of an extent, and we end up logging the full extent but only checksums for the written range, which results in missing checksums after log replay. Allowing ranged fsyncs to operate again only in the original range, when using the NO_HOLES feature and doing a fast fsync is doable but requires some non trivial changes to the writeback path, which can always be worked on later if needed, but I don't think they are a very common use case. Several tests were performed using fio for different numbers of concurrent jobs, each writing and fsyncing its own file, for both sequential and random file writes. The tests were run on bare metal, no virtualization, on a box with 12 cores (Intel i7-8700), 64Gb of RAM and a NVMe device, with a kernel configuration that is the default of typical distributions (debian in this case), without debug options enabled (kasan, kmemleak, slub debug, debug of page allocations, lock debugging, etc). The following script that calls fio was used: $ cat test-fsync.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/btrfs MOUNT_OPTIONS="-o ssd -o space_cache=v2" MKFS_OPTIONS="-d single -m single" if [ $# -ne 5 ]; then echo "Use $0 NUM_JOBS FILE_SIZE FSYNC_FREQ BLOCK_SIZE [write|randwrite]" exit 1 fi NUM_JOBS=$1 FILE_SIZE=$2 FSYNC_FREQ=$3 BLOCK_SIZE=$4 WRITE_MODE=$5 if [ "$WRITE_MODE" != "write" ] && [ "$WRITE_MODE" != "randwrite" ]; then echo "Invalid WRITE_MODE, must be 'write' or 'randwrite'" exit 1 fi cat <<EOF > /tmp/fio-job.ini [writers] rw=$WRITE_MODE fsync=$FSYNC_FREQ fallocate=none group_reporting=1 direct=0 bs=$BLOCK_SIZE ioengine=sync size=$FILE_SIZE directory=$MNT numjobs=$NUM_JOBS EOF echo "performance" | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor echo echo "Using config:" echo cat /tmp/fio-job.ini echo umount $MNT &> /dev/null mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT fio /tmp/fio-job.ini umount $MNT The results were the following: ************************* *** sequential writes *** ************************* ==== 1 job, 8GiB file, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=36.6MiB/s (38.4MB/s), 36.6MiB/s-36.6MiB/s (38.4MB/s-38.4MB/s), io=8192MiB (8590MB), run=223689-223689msec After patch: WRITE: bw=40.2MiB/s (42.1MB/s), 40.2MiB/s-40.2MiB/s (42.1MB/s-42.1MB/s), io=8192MiB (8590MB), run=203980-203980msec (+9.8%, -8.8% runtime) ==== 2 jobs, 4GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=35.8MiB/s (37.5MB/s), 35.8MiB/s-35.8MiB/s (37.5MB/s-37.5MB/s), io=8192MiB (8590MB), run=228950-228950msec After patch: WRITE: bw=43.5MiB/s (45.6MB/s), 43.5MiB/s-43.5MiB/s (45.6MB/s-45.6MB/s), io=8192MiB (8590MB), run=188272-188272msec (+21.5% throughput, -17.8% runtime) ==== 4 jobs, 2GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=50.1MiB/s (52.6MB/s), 50.1MiB/s-50.1MiB/s (52.6MB/s-52.6MB/s), io=8192MiB (8590MB), run=163446-163446msec After patch: WRITE: bw=64.5MiB/s (67.6MB/s), 64.5MiB/s-64.5MiB/s (67.6MB/s-67.6MB/s), io=8192MiB (8590MB), run=126987-126987msec (+28.7% throughput, -22.3% runtime) ==== 8 jobs, 1GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=64.0MiB/s (68.1MB/s), 64.0MiB/s-64.0MiB/s (68.1MB/s-68.1MB/s), io=8192MiB (8590MB), run=126075-126075msec After patch: WRITE: bw=86.8MiB/s (91.0MB/s), 86.8MiB/s-86.8MiB/s (91.0MB/s-91.0MB/s), io=8192MiB (8590MB), run=94358-94358msec (+35.6% throughput, -25.2% runtime) ==== 16 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=79.8MiB/s (83.6MB/s), 79.8MiB/s-79.8MiB/s (83.6MB/s-83.6MB/s), io=8192MiB (8590MB), run=102694-102694msec After patch: WRITE: bw=107MiB/s (112MB/s), 107MiB/s-107MiB/s (112MB/s-112MB/s), io=8192MiB (8590MB), run=76446-76446msec (+34.1% throughput, -25.6% runtime) ==== 32 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=93.2MiB/s (97.7MB/s), 93.2MiB/s-93.2MiB/s (97.7MB/s-97.7MB/s), io=16.0GiB (17.2GB), run=175836-175836msec After patch: WRITE: bw=111MiB/s (117MB/s), 111MiB/s-111MiB/s (117MB/s-117MB/s), io=16.0GiB (17.2GB), run=147001-147001msec (+19.1% throughput, -16.4% runtime) ==== 64 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=108MiB/s (114MB/s), 108MiB/s-108MiB/s (114MB/s-114MB/s), io=32.0GiB (34.4GB), run=302656-302656msec After patch: WRITE: bw=133MiB/s (140MB/s), 133MiB/s-133MiB/s (140MB/s-140MB/s), io=32.0GiB (34.4GB), run=246003-246003msec (+23.1% throughput, -18.7% runtime) ************************ *** random writes *** ************************ ==== 1 job, 8GiB file, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=11.5MiB/s (12.0MB/s), 11.5MiB/s-11.5MiB/s (12.0MB/s-12.0MB/s), io=8192MiB (8590MB), run=714281-714281msec After patch: WRITE: bw=11.6MiB/s (12.2MB/s), 11.6MiB/s-11.6MiB/s (12.2MB/s-12.2MB/s), io=8192MiB (8590MB), run=705959-705959msec (+0.9% throughput, -1.7% runtime) ==== 2 jobs, 4GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=12.8MiB/s (13.5MB/s), 12.8MiB/s-12.8MiB/s (13.5MB/s-13.5MB/s), io=8192MiB (8590MB), run=638101-638101msec After patch: WRITE: bw=13.1MiB/s (13.7MB/s), 13.1MiB/s-13.1MiB/s (13.7MB/s-13.7MB/s), io=8192MiB (8590MB), run=625374-625374msec (+2.3% throughput, -2.0% runtime) ==== 4 jobs, 2GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=15.4MiB/s (16.2MB/s), 15.4MiB/s-15.4MiB/s (16.2MB/s-16.2MB/s), io=8192MiB (8590MB), run=531146-531146msec After patch: WRITE: bw=17.8MiB/s (18.7MB/s), 17.8MiB/s-17.8MiB/s (18.7MB/s-18.7MB/s), io=8192MiB (8590MB), run=460431-460431msec (+15.6% throughput, -13.3% runtime) ==== 8 jobs, 1GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=19.9MiB/s (20.8MB/s), 19.9MiB/s-19.9MiB/s (20.8MB/s-20.8MB/s), io=8192MiB (8590MB), run=412664-412664msec After patch: WRITE: bw=22.2MiB/s (23.3MB/s), 22.2MiB/s-22.2MiB/s (23.3MB/s-23.3MB/s), io=8192MiB (8590MB), run=368589-368589msec (+11.6% throughput, -10.7% runtime) ==== 16 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=29.3MiB/s (30.7MB/s), 29.3MiB/s-29.3MiB/s (30.7MB/s-30.7MB/s), io=8192MiB (8590MB), run=279924-279924msec After patch: WRITE: bw=30.4MiB/s (31.9MB/s), 30.4MiB/s-30.4MiB/s (31.9MB/s-31.9MB/s), io=8192MiB (8590MB), run=269258-269258msec (+3.8% throughput, -3.8% runtime) ==== 32 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=36.9MiB/s (38.7MB/s), 36.9MiB/s-36.9MiB/s (38.7MB/s-38.7MB/s), io=16.0GiB (17.2GB), run=443581-443581msec After patch: WRITE: bw=41.6MiB/s (43.6MB/s), 41.6MiB/s-41.6MiB/s (43.6MB/s-43.6MB/s), io=16.0GiB (17.2GB), run=394114-394114msec (+12.7% throughput, -11.2% runtime) ==== 64 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=45.9MiB/s (48.1MB/s), 45.9MiB/s-45.9MiB/s (48.1MB/s-48.1MB/s), io=32.0GiB (34.4GB), run=714614-714614msec After patch: WRITE: bw=48.8MiB/s (51.1MB/s), 48.8MiB/s-48.8MiB/s (51.1MB/s-51.1MB/s), io=32.0GiB (34.4GB), run=672087-672087msec (+6.3% throughput, -6.0% runtime) Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-08-11 14:43:58 +03:00
struct btrfs_log_ctx *ctx)
Btrfs: turbo charge fsync At least for the vm workload. Currently on fsync we will 1) Truncate all items in the log tree for the given inode if they exist and 2) Copy all items for a given inode into the log The problem with this is that for things like VMs you can have lots of extents from the fragmented writing behavior, and worst yet you may have only modified a few extents, not the entire thing. This patch fixes this problem by tracking which transid modified our extent, and then when we do the tree logging we find all of the extents we've modified in our current transaction, sort them and commit them. We also only truncate up to the xattrs of the inode and copy that stuff in normally, and then just drop any extents in the range we have that exist in the log already. Here are some numbers of a 50 meg fio job that does random writes and fsync()s after every write Original Patched SATA drive 82KB/s 140KB/s Fusion drive 431KB/s 2532KB/s So around 2-6 times faster depending on your hardware. There are a few corner cases, for example if you truncate at all we have to do it the old way since there is no way to be sure what is in the log is ok. This probably could be done smarter, but if you write-fsync-truncate-write-fsync you deserve what you get. All this work is in RAM of course so if your inode gets evicted from cache and you read it in and fsync it we'll do it the slow way if we are still in the same transaction that we last modified the inode in. The biggest cool part of this is that it requires no changes to the recovery code, so if you fsync with this patch and crash and load an old kernel, it will run the recovery and be a-ok. I have tested this pretty thoroughly with an fsync tester and everything comes back fine, as well as xfstests. Thanks, Signed-off-by: Josef Bacik <jbacik@fusionio.com>
2012-08-17 21:14:17 +04:00
{
btrfs: make fast fsyncs wait only for writeback Currently regardless of a full or a fast fsync we always wait for ordered extents to complete, and then start logging the inode after that. However for fast fsyncs we can just wait for the writeback to complete, we don't need to wait for the ordered extents to complete since we use the list of modified extents maps to figure out which extents we must log and we can get their checksums directly from the ordered extents that are still in flight, otherwise look them up from the checksums tree. Until commit b5e6c3e170b770 ("btrfs: always wait on ordered extents at fsync time"), for fast fsyncs, we used to start logging without even waiting for the writeback to complete first, we would wait for it to complete after logging, while holding a transaction open, which lead to performance issues when using cgroups and probably for other cases too, as wait for IO while holding a transaction handle should be avoided as much as possible. After that, for fast fsyncs, we started to wait for ordered extents to complete before starting to log, which adds some latency to fsyncs and we even got at least one report about a performance drop which bisected to that particular change: https://lore.kernel.org/linux-btrfs/20181109215148.GF23260@techsingularity.net/ This change makes fast fsyncs only wait for writeback to finish before starting to log the inode, instead of waiting for both the writeback to finish and for the ordered extents to complete. This brings back part of the logic we had that extracts checksums from in flight ordered extents, which are not yet in the checksums tree, and making sure transaction commits wait for the completion of ordered extents previously logged (by far most of the time they have already completed by the time a transaction commit starts, resulting in no wait at all), to avoid any data loss if an ordered extent completes after the transaction used to log an inode is committed, followed by a power failure. When there are no other tasks accessing the checksums and the subvolume btrees, the ordered extent completion is pretty fast, typically taking 100 to 200 microseconds only in my observations. However when there are other tasks accessing these btrees, ordered extent completion can take a lot more time due to lock contention on nodes and leaves of these btrees. I've seen cases over 2 milliseconds, which starts to be significant. In particular when we do have concurrent fsyncs against different files there is a lot of contention on the checksums btree, since we have many tasks writing the checksums into the btree and other tasks that already started the logging phase are doing lookups for checksums in the btree. This change also turns all ranged fsyncs into full ranged fsyncs, which is something we already did when not using the NO_HOLES features or when doing a full fsync. This is to guarantee we never miss checksums due to writeback having been triggered only for a part of an extent, and we end up logging the full extent but only checksums for the written range, which results in missing checksums after log replay. Allowing ranged fsyncs to operate again only in the original range, when using the NO_HOLES feature and doing a fast fsync is doable but requires some non trivial changes to the writeback path, which can always be worked on later if needed, but I don't think they are a very common use case. Several tests were performed using fio for different numbers of concurrent jobs, each writing and fsyncing its own file, for both sequential and random file writes. The tests were run on bare metal, no virtualization, on a box with 12 cores (Intel i7-8700), 64Gb of RAM and a NVMe device, with a kernel configuration that is the default of typical distributions (debian in this case), without debug options enabled (kasan, kmemleak, slub debug, debug of page allocations, lock debugging, etc). The following script that calls fio was used: $ cat test-fsync.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/btrfs MOUNT_OPTIONS="-o ssd -o space_cache=v2" MKFS_OPTIONS="-d single -m single" if [ $# -ne 5 ]; then echo "Use $0 NUM_JOBS FILE_SIZE FSYNC_FREQ BLOCK_SIZE [write|randwrite]" exit 1 fi NUM_JOBS=$1 FILE_SIZE=$2 FSYNC_FREQ=$3 BLOCK_SIZE=$4 WRITE_MODE=$5 if [ "$WRITE_MODE" != "write" ] && [ "$WRITE_MODE" != "randwrite" ]; then echo "Invalid WRITE_MODE, must be 'write' or 'randwrite'" exit 1 fi cat <<EOF > /tmp/fio-job.ini [writers] rw=$WRITE_MODE fsync=$FSYNC_FREQ fallocate=none group_reporting=1 direct=0 bs=$BLOCK_SIZE ioengine=sync size=$FILE_SIZE directory=$MNT numjobs=$NUM_JOBS EOF echo "performance" | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor echo echo "Using config:" echo cat /tmp/fio-job.ini echo umount $MNT &> /dev/null mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT fio /tmp/fio-job.ini umount $MNT The results were the following: ************************* *** sequential writes *** ************************* ==== 1 job, 8GiB file, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=36.6MiB/s (38.4MB/s), 36.6MiB/s-36.6MiB/s (38.4MB/s-38.4MB/s), io=8192MiB (8590MB), run=223689-223689msec After patch: WRITE: bw=40.2MiB/s (42.1MB/s), 40.2MiB/s-40.2MiB/s (42.1MB/s-42.1MB/s), io=8192MiB (8590MB), run=203980-203980msec (+9.8%, -8.8% runtime) ==== 2 jobs, 4GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=35.8MiB/s (37.5MB/s), 35.8MiB/s-35.8MiB/s (37.5MB/s-37.5MB/s), io=8192MiB (8590MB), run=228950-228950msec After patch: WRITE: bw=43.5MiB/s (45.6MB/s), 43.5MiB/s-43.5MiB/s (45.6MB/s-45.6MB/s), io=8192MiB (8590MB), run=188272-188272msec (+21.5% throughput, -17.8% runtime) ==== 4 jobs, 2GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=50.1MiB/s (52.6MB/s), 50.1MiB/s-50.1MiB/s (52.6MB/s-52.6MB/s), io=8192MiB (8590MB), run=163446-163446msec After patch: WRITE: bw=64.5MiB/s (67.6MB/s), 64.5MiB/s-64.5MiB/s (67.6MB/s-67.6MB/s), io=8192MiB (8590MB), run=126987-126987msec (+28.7% throughput, -22.3% runtime) ==== 8 jobs, 1GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=64.0MiB/s (68.1MB/s), 64.0MiB/s-64.0MiB/s (68.1MB/s-68.1MB/s), io=8192MiB (8590MB), run=126075-126075msec After patch: WRITE: bw=86.8MiB/s (91.0MB/s), 86.8MiB/s-86.8MiB/s (91.0MB/s-91.0MB/s), io=8192MiB (8590MB), run=94358-94358msec (+35.6% throughput, -25.2% runtime) ==== 16 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=79.8MiB/s (83.6MB/s), 79.8MiB/s-79.8MiB/s (83.6MB/s-83.6MB/s), io=8192MiB (8590MB), run=102694-102694msec After patch: WRITE: bw=107MiB/s (112MB/s), 107MiB/s-107MiB/s (112MB/s-112MB/s), io=8192MiB (8590MB), run=76446-76446msec (+34.1% throughput, -25.6% runtime) ==== 32 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=93.2MiB/s (97.7MB/s), 93.2MiB/s-93.2MiB/s (97.7MB/s-97.7MB/s), io=16.0GiB (17.2GB), run=175836-175836msec After patch: WRITE: bw=111MiB/s (117MB/s), 111MiB/s-111MiB/s (117MB/s-117MB/s), io=16.0GiB (17.2GB), run=147001-147001msec (+19.1% throughput, -16.4% runtime) ==== 64 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=108MiB/s (114MB/s), 108MiB/s-108MiB/s (114MB/s-114MB/s), io=32.0GiB (34.4GB), run=302656-302656msec After patch: WRITE: bw=133MiB/s (140MB/s), 133MiB/s-133MiB/s (140MB/s-140MB/s), io=32.0GiB (34.4GB), run=246003-246003msec (+23.1% throughput, -18.7% runtime) ************************ *** random writes *** ************************ ==== 1 job, 8GiB file, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=11.5MiB/s (12.0MB/s), 11.5MiB/s-11.5MiB/s (12.0MB/s-12.0MB/s), io=8192MiB (8590MB), run=714281-714281msec After patch: WRITE: bw=11.6MiB/s (12.2MB/s), 11.6MiB/s-11.6MiB/s (12.2MB/s-12.2MB/s), io=8192MiB (8590MB), run=705959-705959msec (+0.9% throughput, -1.7% runtime) ==== 2 jobs, 4GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=12.8MiB/s (13.5MB/s), 12.8MiB/s-12.8MiB/s (13.5MB/s-13.5MB/s), io=8192MiB (8590MB), run=638101-638101msec After patch: WRITE: bw=13.1MiB/s (13.7MB/s), 13.1MiB/s-13.1MiB/s (13.7MB/s-13.7MB/s), io=8192MiB (8590MB), run=625374-625374msec (+2.3% throughput, -2.0% runtime) ==== 4 jobs, 2GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=15.4MiB/s (16.2MB/s), 15.4MiB/s-15.4MiB/s (16.2MB/s-16.2MB/s), io=8192MiB (8590MB), run=531146-531146msec After patch: WRITE: bw=17.8MiB/s (18.7MB/s), 17.8MiB/s-17.8MiB/s (18.7MB/s-18.7MB/s), io=8192MiB (8590MB), run=460431-460431msec (+15.6% throughput, -13.3% runtime) ==== 8 jobs, 1GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=19.9MiB/s (20.8MB/s), 19.9MiB/s-19.9MiB/s (20.8MB/s-20.8MB/s), io=8192MiB (8590MB), run=412664-412664msec After patch: WRITE: bw=22.2MiB/s (23.3MB/s), 22.2MiB/s-22.2MiB/s (23.3MB/s-23.3MB/s), io=8192MiB (8590MB), run=368589-368589msec (+11.6% throughput, -10.7% runtime) ==== 16 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=29.3MiB/s (30.7MB/s), 29.3MiB/s-29.3MiB/s (30.7MB/s-30.7MB/s), io=8192MiB (8590MB), run=279924-279924msec After patch: WRITE: bw=30.4MiB/s (31.9MB/s), 30.4MiB/s-30.4MiB/s (31.9MB/s-31.9MB/s), io=8192MiB (8590MB), run=269258-269258msec (+3.8% throughput, -3.8% runtime) ==== 32 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=36.9MiB/s (38.7MB/s), 36.9MiB/s-36.9MiB/s (38.7MB/s-38.7MB/s), io=16.0GiB (17.2GB), run=443581-443581msec After patch: WRITE: bw=41.6MiB/s (43.6MB/s), 41.6MiB/s-41.6MiB/s (43.6MB/s-43.6MB/s), io=16.0GiB (17.2GB), run=394114-394114msec (+12.7% throughput, -11.2% runtime) ==== 64 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=45.9MiB/s (48.1MB/s), 45.9MiB/s-45.9MiB/s (48.1MB/s-48.1MB/s), io=32.0GiB (34.4GB), run=714614-714614msec After patch: WRITE: bw=48.8MiB/s (51.1MB/s), 48.8MiB/s-48.8MiB/s (51.1MB/s-51.1MB/s), io=32.0GiB (34.4GB), run=672087-672087msec (+6.3% throughput, -6.0% runtime) Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-08-11 14:43:58 +03:00
struct btrfs_ordered_extent *ordered;
struct btrfs_ordered_extent *tmp;
Btrfs: turbo charge fsync At least for the vm workload. Currently on fsync we will 1) Truncate all items in the log tree for the given inode if they exist and 2) Copy all items for a given inode into the log The problem with this is that for things like VMs you can have lots of extents from the fragmented writing behavior, and worst yet you may have only modified a few extents, not the entire thing. This patch fixes this problem by tracking which transid modified our extent, and then when we do the tree logging we find all of the extents we've modified in our current transaction, sort them and commit them. We also only truncate up to the xattrs of the inode and copy that stuff in normally, and then just drop any extents in the range we have that exist in the log already. Here are some numbers of a 50 meg fio job that does random writes and fsync()s after every write Original Patched SATA drive 82KB/s 140KB/s Fusion drive 431KB/s 2532KB/s So around 2-6 times faster depending on your hardware. There are a few corner cases, for example if you truncate at all we have to do it the old way since there is no way to be sure what is in the log is ok. This probably could be done smarter, but if you write-fsync-truncate-write-fsync you deserve what you get. All this work is in RAM of course so if your inode gets evicted from cache and you read it in and fsync it we'll do it the slow way if we are still in the same transaction that we last modified the inode in. The biggest cool part of this is that it requires no changes to the recovery code, so if you fsync with this patch and crash and load an old kernel, it will run the recovery and be a-ok. I have tested this pretty thoroughly with an fsync tester and everything comes back fine, as well as xfstests. Thanks, Signed-off-by: Josef Bacik <jbacik@fusionio.com>
2012-08-17 21:14:17 +04:00
struct extent_map *em, *n;
struct list_head extents;
struct extent_map_tree *tree = &inode->extent_tree;
Btrfs: turbo charge fsync At least for the vm workload. Currently on fsync we will 1) Truncate all items in the log tree for the given inode if they exist and 2) Copy all items for a given inode into the log The problem with this is that for things like VMs you can have lots of extents from the fragmented writing behavior, and worst yet you may have only modified a few extents, not the entire thing. This patch fixes this problem by tracking which transid modified our extent, and then when we do the tree logging we find all of the extents we've modified in our current transaction, sort them and commit them. We also only truncate up to the xattrs of the inode and copy that stuff in normally, and then just drop any extents in the range we have that exist in the log already. Here are some numbers of a 50 meg fio job that does random writes and fsync()s after every write Original Patched SATA drive 82KB/s 140KB/s Fusion drive 431KB/s 2532KB/s So around 2-6 times faster depending on your hardware. There are a few corner cases, for example if you truncate at all we have to do it the old way since there is no way to be sure what is in the log is ok. This probably could be done smarter, but if you write-fsync-truncate-write-fsync you deserve what you get. All this work is in RAM of course so if your inode gets evicted from cache and you read it in and fsync it we'll do it the slow way if we are still in the same transaction that we last modified the inode in. The biggest cool part of this is that it requires no changes to the recovery code, so if you fsync with this patch and crash and load an old kernel, it will run the recovery and be a-ok. I have tested this pretty thoroughly with an fsync tester and everything comes back fine, as well as xfstests. Thanks, Signed-off-by: Josef Bacik <jbacik@fusionio.com>
2012-08-17 21:14:17 +04:00
int ret = 0;
int num = 0;
Btrfs: turbo charge fsync At least for the vm workload. Currently on fsync we will 1) Truncate all items in the log tree for the given inode if they exist and 2) Copy all items for a given inode into the log The problem with this is that for things like VMs you can have lots of extents from the fragmented writing behavior, and worst yet you may have only modified a few extents, not the entire thing. This patch fixes this problem by tracking which transid modified our extent, and then when we do the tree logging we find all of the extents we've modified in our current transaction, sort them and commit them. We also only truncate up to the xattrs of the inode and copy that stuff in normally, and then just drop any extents in the range we have that exist in the log already. Here are some numbers of a 50 meg fio job that does random writes and fsync()s after every write Original Patched SATA drive 82KB/s 140KB/s Fusion drive 431KB/s 2532KB/s So around 2-6 times faster depending on your hardware. There are a few corner cases, for example if you truncate at all we have to do it the old way since there is no way to be sure what is in the log is ok. This probably could be done smarter, but if you write-fsync-truncate-write-fsync you deserve what you get. All this work is in RAM of course so if your inode gets evicted from cache and you read it in and fsync it we'll do it the slow way if we are still in the same transaction that we last modified the inode in. The biggest cool part of this is that it requires no changes to the recovery code, so if you fsync with this patch and crash and load an old kernel, it will run the recovery and be a-ok. I have tested this pretty thoroughly with an fsync tester and everything comes back fine, as well as xfstests. Thanks, Signed-off-by: Josef Bacik <jbacik@fusionio.com>
2012-08-17 21:14:17 +04:00
INIT_LIST_HEAD(&extents);
write_lock(&tree->lock);
list_for_each_entry_safe(em, n, &tree->modified_extents, list) {
list_del_init(&em->list);
/*
* Just an arbitrary number, this can be really CPU intensive
* once we start getting a lot of extents, and really once we
* have a bunch of extents we just want to commit since it will
* be faster.
*/
if (++num > 32768) {
list_del_init(&tree->modified_extents);
ret = -EFBIG;
goto process;
}
btrfs: fix race that results in logging old extents during a fast fsync When logging the extents of an inode during a fast fsync, we have a time window where we can log extents that are from the previous transaction and already persisted. This only makes us waste time unnecessarily. The following sequence of steps shows how this can happen: 1) We are at transaction 1000; 2) An ordered extent E from inode I completes, that is it has gone through btrfs_finish_ordered_io(), and it set the extent maps' generation to 1000 when we unpin the extent, which is the generation of the current transaction; 3) The commit for transaction 1000 starts by task A; 4) The task committing transaction 1000 sets the transaction state to unblocked, writes the dirty extent buffers and the super blocks, then unlocks tree_log_mutex; 5) Some change is made to inode I, resulting in creation of a new transaction with a generation of 1001; 6) The transaction 1000 commit starts unpinning extents. At this point fs_info->last_trans_committed still has a value of 999; 7) Task B starts an fsync on inode I, and when it gets to btrfs_log_changed_extents() sees the extent map for extent E in the list of modified extents. It sees the extent map has a generation of 1000 and fs_info->last_trans_committed has a value of 999, so it proceeds to logging the respective file extent item and all the checksums covering its range. So we end up wasting time since the extent was already persisted and is reachable through the trees pointed to by the super block committed by transaction 1000. So just fix this by comparing the extent maps generation against the generation of the transaction handle - if it is smaller then the id in the handle, we know the extent was already persisted and we do not need to log it. This patch belongs to a patch set that is comprised of the following patches: btrfs: fix race causing unnecessary inode logging during link and rename btrfs: fix race that results in logging old extents during a fast fsync btrfs: fix race that causes unnecessary logging of ancestor inodes btrfs: fix race that makes inode logging fallback to transaction commit btrfs: fix race leading to unnecessary transaction commit when logging inode btrfs: do not block inode logging for so long during transaction commit Performance results are mentioned in the change log of the last patch. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-11-25 15:19:24 +03:00
if (em->generation < trans->transid)
Btrfs: turbo charge fsync At least for the vm workload. Currently on fsync we will 1) Truncate all items in the log tree for the given inode if they exist and 2) Copy all items for a given inode into the log The problem with this is that for things like VMs you can have lots of extents from the fragmented writing behavior, and worst yet you may have only modified a few extents, not the entire thing. This patch fixes this problem by tracking which transid modified our extent, and then when we do the tree logging we find all of the extents we've modified in our current transaction, sort them and commit them. We also only truncate up to the xattrs of the inode and copy that stuff in normally, and then just drop any extents in the range we have that exist in the log already. Here are some numbers of a 50 meg fio job that does random writes and fsync()s after every write Original Patched SATA drive 82KB/s 140KB/s Fusion drive 431KB/s 2532KB/s So around 2-6 times faster depending on your hardware. There are a few corner cases, for example if you truncate at all we have to do it the old way since there is no way to be sure what is in the log is ok. This probably could be done smarter, but if you write-fsync-truncate-write-fsync you deserve what you get. All this work is in RAM of course so if your inode gets evicted from cache and you read it in and fsync it we'll do it the slow way if we are still in the same transaction that we last modified the inode in. The biggest cool part of this is that it requires no changes to the recovery code, so if you fsync with this patch and crash and load an old kernel, it will run the recovery and be a-ok. I have tested this pretty thoroughly with an fsync tester and everything comes back fine, as well as xfstests. Thanks, Signed-off-by: Josef Bacik <jbacik@fusionio.com>
2012-08-17 21:14:17 +04:00
continue;
Btrfs: fix duplicate extents after fsync of file with prealloc extents In commit 471d557afed1 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay"), on fsync, we started to always log all prealloc extents beyond an inode's i_size in order to avoid losing them after a power failure. However under some cases this can lead to the log replay code to create duplicate extent items, with different lengths, in the extent tree. That happens because, as of that commit, we can now log extent items based on extent maps that are not on the "modified" list of extent maps of the inode's extent map tree. Logging extent items based on extent maps is used during the fast fsync path to save time and for this to work reliably it requires that the extent maps are not merged with other adjacent extent maps - having the extent maps in the list of modified extents gives such guarantee. Consider the following example, captured during a long run of fsstress, which illustrates this problem. We have inode 271, in the filesystem tree (root 5), for which all of the following operations and discussion apply to. A buffered write starts at offset 312391 with a length of 933471 bytes (end offset at 1245862). At this point we have, for this inode, the following extent maps with the their field values: em A, start 0, orig_start 0, len 40960, block_start 18446744073709551613, block_len 0, orig_block_len 0 em B, start 40960, orig_start 40960, len 376832, block_start 1106399232, block_len 376832, orig_block_len 376832 em C, start 417792, orig_start 417792, len 782336, block_start 18446744073709551613, block_len 0, orig_block_len 0 em D, start 1200128, orig_start 1200128, len 835584, block_start 1106776064, block_len 835584, orig_block_len 835584 em E, start 2035712, orig_start 2035712, len 245760, block_start 1107611648, block_len 245760, orig_block_len 245760 Extent map A corresponds to a hole and extent maps D and E correspond to preallocated extents. Extent map D ends where extent map E begins (1106776064 + 835584 = 1107611648), but these extent maps were not merged because they are in the inode's list of modified extent maps. An fsync against this inode is made, which triggers the fast path (BTRFS_INODE_NEEDS_FULL_SYNC is not set). This fsync triggers writeback of the data previously written using buffered IO, and when the respective ordered extent finishes, btrfs_drop_extents() is called against the (aligned) range 311296..1249279. This causes a split of extent map D at btrfs_drop_extent_cache(), replacing extent map D with a new extent map D', also added to the list of modified extents, with the following values: em D', start 1249280, orig_start of 1200128, block_start 1106825216 (= 1106776064 + 1249280 - 1200128), orig_block_len 835584, block_len 786432 (835584 - (1249280 - 1200128)) Then, during the fast fsync, btrfs_log_changed_extents() is called and extent maps D' and E are removed from the list of modified extents. The flag EXTENT_FLAG_LOGGING is also set on them. After the extents are logged clear_em_logging() is called on each of them, and that makes extent map E to be merged with extent map D' (try_merge_map()), resulting in D' being deleted and E adjusted to: em E, start 1249280, orig_start 1200128, len 1032192, block_start 1106825216, block_len 1032192, orig_block_len 245760 A direct IO write at offset 1847296 and length of 360448 bytes (end offset at 2207744) starts, and at that moment the following extent maps exist for our inode: em A, start 0, orig_start 0, len 40960, block_start 18446744073709551613, block_len 0, orig_block_len 0 em B, start 40960, orig_start 40960, len 270336, block_start 1106399232, block_len 270336, orig_block_len 376832 em C, start 311296, orig_start 311296, len 937984, block_start 1112842240, block_len 937984, orig_block_len 937984 em E (prealloc), start 1249280, orig_start 1200128, len 1032192, block_start 1106825216, block_len 1032192, orig_block_len 245760 The dio write results in drop_extent_cache() being called twice. The first time for a range that starts at offset 1847296 and ends at offset 2035711 (length of 188416), which results in a double split of extent map E, replacing it with two new extent maps: em F, start 1249280, orig_start 1200128, block_start 1106825216, block_len 598016, orig_block_len 598016 em G, start 2035712, orig_start 1200128, block_start 1107611648, block_len 245760, orig_block_len 1032192 It also creates a new extent map that represents a part of the requested IO (through create_io_em()): em H, start 1847296, len 188416, block_start 1107423232, block_len 188416 The second call to drop_extent_cache() has a range with a start offset of 2035712 and end offset of 2207743 (length of 172032). This leads to replacing extent map G with a new extent map I with the following values: em I, start 2207744, orig_start 1200128, block_start 1107783680, block_len 73728, orig_block_len 1032192 It also creates a new extent map that represents the second part of the requested IO (through create_io_em()): em J, start 2035712, len 172032, block_start 1107611648, block_len 172032 The dio write set the inode's i_size to 2207744 bytes. After the dio write the inode has the following extent maps: em A, start 0, orig_start 0, len 40960, block_start 18446744073709551613, block_len 0, orig_block_len 0 em B, start 40960, orig_start 40960, len 270336, block_start 1106399232, block_len 270336, orig_block_len 376832 em C, start 311296, orig_start 311296, len 937984, block_start 1112842240, block_len 937984, orig_block_len 937984 em F, start 1249280, orig_start 1200128, len 598016, block_start 1106825216, block_len 598016, orig_block_len 598016 em H, start 1847296, orig_start 1200128, len 188416, block_start 1107423232, block_len 188416, orig_block_len 835584 em J, start 2035712, orig_start 2035712, len 172032, block_start 1107611648, block_len 172032, orig_block_len 245760 em I, start 2207744, orig_start 1200128, len 73728, block_start 1107783680, block_len 73728, orig_block_len 1032192 Now do some change to the file, like adding a xattr for example and then fsync it again. This triggers a fast fsync path, and as of commit 471d557afed1 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay"), we use the extent map I to log a file extent item because it's a prealloc extent and it starts at an offset matching the inode's i_size. However when we log it, we create a file extent item with a value for the disk byte location that is wrong, as can be seen from the following output of "btrfs inspect-internal dump-tree": item 1 key (271 EXTENT_DATA 2207744) itemoff 3782 itemsize 53 generation 22 type 2 (prealloc) prealloc data disk byte 1106776064 nr 1032192 prealloc data offset 1007616 nr 73728 Here the disk byte value corresponds to calculation based on some fields from the extent map I: 1106776064 = block_start (1107783680) - 1007616 (extent_offset) extent_offset = 2207744 (start) - 1200128 (orig_start) = 1007616 The disk byte value of 1106776064 clashes with disk byte values of the file extent items at offsets 1249280 and 1847296 in the fs tree: item 6 key (271 EXTENT_DATA 1249280) itemoff 3568 itemsize 53 generation 20 type 2 (prealloc) prealloc data disk byte 1106776064 nr 835584 prealloc data offset 49152 nr 598016 item 7 key (271 EXTENT_DATA 1847296) itemoff 3515 itemsize 53 generation 20 type 1 (regular) extent data disk byte 1106776064 nr 835584 extent data offset 647168 nr 188416 ram 835584 extent compression 0 (none) item 8 key (271 EXTENT_DATA 2035712) itemoff 3462 itemsize 53 generation 20 type 1 (regular) extent data disk byte 1107611648 nr 245760 extent data offset 0 nr 172032 ram 245760 extent compression 0 (none) item 9 key (271 EXTENT_DATA 2207744) itemoff 3409 itemsize 53 generation 20 type 2 (prealloc) prealloc data disk byte 1107611648 nr 245760 prealloc data offset 172032 nr 73728 Instead of the disk byte value of 1106776064, the value of 1107611648 should have been logged. Also the data offset value should have been 172032 and not 1007616. After a log replay we end up getting two extent items in the extent tree with different lengths, one of 835584, which is correct and existed before the log replay, and another one of 1032192 which is wrong and is based on the logged file extent item: item 12 key (1106776064 EXTENT_ITEM 835584) itemoff 3406 itemsize 53 refs 2 gen 15 flags DATA extent data backref root 5 objectid 271 offset 1200128 count 2 item 13 key (1106776064 EXTENT_ITEM 1032192) itemoff 3353 itemsize 53 refs 1 gen 22 flags DATA extent data backref root 5 objectid 271 offset 1200128 count 1 Obviously this leads to many problems and a filesystem check reports many errors: (...) checking extents Extent back ref already exists for 1106776064 parent 0 root 5 owner 271 offset 1200128 num_refs 1 extent item 1106776064 has multiple extent items ref mismatch on [1106776064 835584] extent item 2, found 3 Incorrect local backref count on 1106776064 root 5 owner 271 offset 1200128 found 2 wanted 1 back 0x55b1d0ad7680 Backref 1106776064 root 5 owner 271 offset 1200128 num_refs 0 not found in extent tree Incorrect local backref count on 1106776064 root 5 owner 271 offset 1200128 found 1 wanted 0 back 0x55b1d0ad4e70 Backref bytes do not match extent backref, bytenr=1106776064, ref bytes=835584, backref bytes=1032192 backpointer mismatch on [1106776064 835584] checking free space cache block group 1103101952 has wrong amount of free space failed to load free space cache for block group 1103101952 checking fs roots (...) So fix this by logging the prealloc extents beyond the inode's i_size based on searches in the subvolume tree instead of the extent maps. Fixes: 471d557afed1 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay") CC: stable@vger.kernel.org # 4.14+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2018-05-09 18:01:46 +03:00
/* We log prealloc extents beyond eof later. */
if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) &&
em->start >= i_size_read(&inode->vfs_inode))
continue;
/* Need a ref to keep it from getting evicted from cache */
refcount_inc(&em->refs);
set_bit(EXTENT_FLAG_LOGGING, &em->flags);
Btrfs: turbo charge fsync At least for the vm workload. Currently on fsync we will 1) Truncate all items in the log tree for the given inode if they exist and 2) Copy all items for a given inode into the log The problem with this is that for things like VMs you can have lots of extents from the fragmented writing behavior, and worst yet you may have only modified a few extents, not the entire thing. This patch fixes this problem by tracking which transid modified our extent, and then when we do the tree logging we find all of the extents we've modified in our current transaction, sort them and commit them. We also only truncate up to the xattrs of the inode and copy that stuff in normally, and then just drop any extents in the range we have that exist in the log already. Here are some numbers of a 50 meg fio job that does random writes and fsync()s after every write Original Patched SATA drive 82KB/s 140KB/s Fusion drive 431KB/s 2532KB/s So around 2-6 times faster depending on your hardware. There are a few corner cases, for example if you truncate at all we have to do it the old way since there is no way to be sure what is in the log is ok. This probably could be done smarter, but if you write-fsync-truncate-write-fsync you deserve what you get. All this work is in RAM of course so if your inode gets evicted from cache and you read it in and fsync it we'll do it the slow way if we are still in the same transaction that we last modified the inode in. The biggest cool part of this is that it requires no changes to the recovery code, so if you fsync with this patch and crash and load an old kernel, it will run the recovery and be a-ok. I have tested this pretty thoroughly with an fsync tester and everything comes back fine, as well as xfstests. Thanks, Signed-off-by: Josef Bacik <jbacik@fusionio.com>
2012-08-17 21:14:17 +04:00
list_add_tail(&em->list, &extents);
num++;
Btrfs: turbo charge fsync At least for the vm workload. Currently on fsync we will 1) Truncate all items in the log tree for the given inode if they exist and 2) Copy all items for a given inode into the log The problem with this is that for things like VMs you can have lots of extents from the fragmented writing behavior, and worst yet you may have only modified a few extents, not the entire thing. This patch fixes this problem by tracking which transid modified our extent, and then when we do the tree logging we find all of the extents we've modified in our current transaction, sort them and commit them. We also only truncate up to the xattrs of the inode and copy that stuff in normally, and then just drop any extents in the range we have that exist in the log already. Here are some numbers of a 50 meg fio job that does random writes and fsync()s after every write Original Patched SATA drive 82KB/s 140KB/s Fusion drive 431KB/s 2532KB/s So around 2-6 times faster depending on your hardware. There are a few corner cases, for example if you truncate at all we have to do it the old way since there is no way to be sure what is in the log is ok. This probably could be done smarter, but if you write-fsync-truncate-write-fsync you deserve what you get. All this work is in RAM of course so if your inode gets evicted from cache and you read it in and fsync it we'll do it the slow way if we are still in the same transaction that we last modified the inode in. The biggest cool part of this is that it requires no changes to the recovery code, so if you fsync with this patch and crash and load an old kernel, it will run the recovery and be a-ok. I have tested this pretty thoroughly with an fsync tester and everything comes back fine, as well as xfstests. Thanks, Signed-off-by: Josef Bacik <jbacik@fusionio.com>
2012-08-17 21:14:17 +04:00
}
list_sort(NULL, &extents, extent_cmp);
process:
Btrfs: turbo charge fsync At least for the vm workload. Currently on fsync we will 1) Truncate all items in the log tree for the given inode if they exist and 2) Copy all items for a given inode into the log The problem with this is that for things like VMs you can have lots of extents from the fragmented writing behavior, and worst yet you may have only modified a few extents, not the entire thing. This patch fixes this problem by tracking which transid modified our extent, and then when we do the tree logging we find all of the extents we've modified in our current transaction, sort them and commit them. We also only truncate up to the xattrs of the inode and copy that stuff in normally, and then just drop any extents in the range we have that exist in the log already. Here are some numbers of a 50 meg fio job that does random writes and fsync()s after every write Original Patched SATA drive 82KB/s 140KB/s Fusion drive 431KB/s 2532KB/s So around 2-6 times faster depending on your hardware. There are a few corner cases, for example if you truncate at all we have to do it the old way since there is no way to be sure what is in the log is ok. This probably could be done smarter, but if you write-fsync-truncate-write-fsync you deserve what you get. All this work is in RAM of course so if your inode gets evicted from cache and you read it in and fsync it we'll do it the slow way if we are still in the same transaction that we last modified the inode in. The biggest cool part of this is that it requires no changes to the recovery code, so if you fsync with this patch and crash and load an old kernel, it will run the recovery and be a-ok. I have tested this pretty thoroughly with an fsync tester and everything comes back fine, as well as xfstests. Thanks, Signed-off-by: Josef Bacik <jbacik@fusionio.com>
2012-08-17 21:14:17 +04:00
while (!list_empty(&extents)) {
em = list_entry(extents.next, struct extent_map, list);
list_del_init(&em->list);
/*
* If we had an error we just need to delete everybody from our
* private list.
*/
if (ret) {
clear_em_logging(tree, em);
free_extent_map(em);
Btrfs: turbo charge fsync At least for the vm workload. Currently on fsync we will 1) Truncate all items in the log tree for the given inode if they exist and 2) Copy all items for a given inode into the log The problem with this is that for things like VMs you can have lots of extents from the fragmented writing behavior, and worst yet you may have only modified a few extents, not the entire thing. This patch fixes this problem by tracking which transid modified our extent, and then when we do the tree logging we find all of the extents we've modified in our current transaction, sort them and commit them. We also only truncate up to the xattrs of the inode and copy that stuff in normally, and then just drop any extents in the range we have that exist in the log already. Here are some numbers of a 50 meg fio job that does random writes and fsync()s after every write Original Patched SATA drive 82KB/s 140KB/s Fusion drive 431KB/s 2532KB/s So around 2-6 times faster depending on your hardware. There are a few corner cases, for example if you truncate at all we have to do it the old way since there is no way to be sure what is in the log is ok. This probably could be done smarter, but if you write-fsync-truncate-write-fsync you deserve what you get. All this work is in RAM of course so if your inode gets evicted from cache and you read it in and fsync it we'll do it the slow way if we are still in the same transaction that we last modified the inode in. The biggest cool part of this is that it requires no changes to the recovery code, so if you fsync with this patch and crash and load an old kernel, it will run the recovery and be a-ok. I have tested this pretty thoroughly with an fsync tester and everything comes back fine, as well as xfstests. Thanks, Signed-off-by: Josef Bacik <jbacik@fusionio.com>
2012-08-17 21:14:17 +04:00
continue;
}
write_unlock(&tree->lock);
Btrfs: turbo charge fsync At least for the vm workload. Currently on fsync we will 1) Truncate all items in the log tree for the given inode if they exist and 2) Copy all items for a given inode into the log The problem with this is that for things like VMs you can have lots of extents from the fragmented writing behavior, and worst yet you may have only modified a few extents, not the entire thing. This patch fixes this problem by tracking which transid modified our extent, and then when we do the tree logging we find all of the extents we've modified in our current transaction, sort them and commit them. We also only truncate up to the xattrs of the inode and copy that stuff in normally, and then just drop any extents in the range we have that exist in the log already. Here are some numbers of a 50 meg fio job that does random writes and fsync()s after every write Original Patched SATA drive 82KB/s 140KB/s Fusion drive 431KB/s 2532KB/s So around 2-6 times faster depending on your hardware. There are a few corner cases, for example if you truncate at all we have to do it the old way since there is no way to be sure what is in the log is ok. This probably could be done smarter, but if you write-fsync-truncate-write-fsync you deserve what you get. All this work is in RAM of course so if your inode gets evicted from cache and you read it in and fsync it we'll do it the slow way if we are still in the same transaction that we last modified the inode in. The biggest cool part of this is that it requires no changes to the recovery code, so if you fsync with this patch and crash and load an old kernel, it will run the recovery and be a-ok. I have tested this pretty thoroughly with an fsync tester and everything comes back fine, as well as xfstests. Thanks, Signed-off-by: Josef Bacik <jbacik@fusionio.com>
2012-08-17 21:14:17 +04:00
ret = log_one_extent(trans, inode, em, path, ctx);
write_lock(&tree->lock);
clear_em_logging(tree, em);
free_extent_map(em);
Btrfs: turbo charge fsync At least for the vm workload. Currently on fsync we will 1) Truncate all items in the log tree for the given inode if they exist and 2) Copy all items for a given inode into the log The problem with this is that for things like VMs you can have lots of extents from the fragmented writing behavior, and worst yet you may have only modified a few extents, not the entire thing. This patch fixes this problem by tracking which transid modified our extent, and then when we do the tree logging we find all of the extents we've modified in our current transaction, sort them and commit them. We also only truncate up to the xattrs of the inode and copy that stuff in normally, and then just drop any extents in the range we have that exist in the log already. Here are some numbers of a 50 meg fio job that does random writes and fsync()s after every write Original Patched SATA drive 82KB/s 140KB/s Fusion drive 431KB/s 2532KB/s So around 2-6 times faster depending on your hardware. There are a few corner cases, for example if you truncate at all we have to do it the old way since there is no way to be sure what is in the log is ok. This probably could be done smarter, but if you write-fsync-truncate-write-fsync you deserve what you get. All this work is in RAM of course so if your inode gets evicted from cache and you read it in and fsync it we'll do it the slow way if we are still in the same transaction that we last modified the inode in. The biggest cool part of this is that it requires no changes to the recovery code, so if you fsync with this patch and crash and load an old kernel, it will run the recovery and be a-ok. I have tested this pretty thoroughly with an fsync tester and everything comes back fine, as well as xfstests. Thanks, Signed-off-by: Josef Bacik <jbacik@fusionio.com>
2012-08-17 21:14:17 +04:00
}
WARN_ON(!list_empty(&extents));
write_unlock(&tree->lock);
Btrfs: turbo charge fsync At least for the vm workload. Currently on fsync we will 1) Truncate all items in the log tree for the given inode if they exist and 2) Copy all items for a given inode into the log The problem with this is that for things like VMs you can have lots of extents from the fragmented writing behavior, and worst yet you may have only modified a few extents, not the entire thing. This patch fixes this problem by tracking which transid modified our extent, and then when we do the tree logging we find all of the extents we've modified in our current transaction, sort them and commit them. We also only truncate up to the xattrs of the inode and copy that stuff in normally, and then just drop any extents in the range we have that exist in the log already. Here are some numbers of a 50 meg fio job that does random writes and fsync()s after every write Original Patched SATA drive 82KB/s 140KB/s Fusion drive 431KB/s 2532KB/s So around 2-6 times faster depending on your hardware. There are a few corner cases, for example if you truncate at all we have to do it the old way since there is no way to be sure what is in the log is ok. This probably could be done smarter, but if you write-fsync-truncate-write-fsync you deserve what you get. All this work is in RAM of course so if your inode gets evicted from cache and you read it in and fsync it we'll do it the slow way if we are still in the same transaction that we last modified the inode in. The biggest cool part of this is that it requires no changes to the recovery code, so if you fsync with this patch and crash and load an old kernel, it will run the recovery and be a-ok. I have tested this pretty thoroughly with an fsync tester and everything comes back fine, as well as xfstests. Thanks, Signed-off-by: Josef Bacik <jbacik@fusionio.com>
2012-08-17 21:14:17 +04:00
Btrfs: fix duplicate extents after fsync of file with prealloc extents In commit 471d557afed1 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay"), on fsync, we started to always log all prealloc extents beyond an inode's i_size in order to avoid losing them after a power failure. However under some cases this can lead to the log replay code to create duplicate extent items, with different lengths, in the extent tree. That happens because, as of that commit, we can now log extent items based on extent maps that are not on the "modified" list of extent maps of the inode's extent map tree. Logging extent items based on extent maps is used during the fast fsync path to save time and for this to work reliably it requires that the extent maps are not merged with other adjacent extent maps - having the extent maps in the list of modified extents gives such guarantee. Consider the following example, captured during a long run of fsstress, which illustrates this problem. We have inode 271, in the filesystem tree (root 5), for which all of the following operations and discussion apply to. A buffered write starts at offset 312391 with a length of 933471 bytes (end offset at 1245862). At this point we have, for this inode, the following extent maps with the their field values: em A, start 0, orig_start 0, len 40960, block_start 18446744073709551613, block_len 0, orig_block_len 0 em B, start 40960, orig_start 40960, len 376832, block_start 1106399232, block_len 376832, orig_block_len 376832 em C, start 417792, orig_start 417792, len 782336, block_start 18446744073709551613, block_len 0, orig_block_len 0 em D, start 1200128, orig_start 1200128, len 835584, block_start 1106776064, block_len 835584, orig_block_len 835584 em E, start 2035712, orig_start 2035712, len 245760, block_start 1107611648, block_len 245760, orig_block_len 245760 Extent map A corresponds to a hole and extent maps D and E correspond to preallocated extents. Extent map D ends where extent map E begins (1106776064 + 835584 = 1107611648), but these extent maps were not merged because they are in the inode's list of modified extent maps. An fsync against this inode is made, which triggers the fast path (BTRFS_INODE_NEEDS_FULL_SYNC is not set). This fsync triggers writeback of the data previously written using buffered IO, and when the respective ordered extent finishes, btrfs_drop_extents() is called against the (aligned) range 311296..1249279. This causes a split of extent map D at btrfs_drop_extent_cache(), replacing extent map D with a new extent map D', also added to the list of modified extents, with the following values: em D', start 1249280, orig_start of 1200128, block_start 1106825216 (= 1106776064 + 1249280 - 1200128), orig_block_len 835584, block_len 786432 (835584 - (1249280 - 1200128)) Then, during the fast fsync, btrfs_log_changed_extents() is called and extent maps D' and E are removed from the list of modified extents. The flag EXTENT_FLAG_LOGGING is also set on them. After the extents are logged clear_em_logging() is called on each of them, and that makes extent map E to be merged with extent map D' (try_merge_map()), resulting in D' being deleted and E adjusted to: em E, start 1249280, orig_start 1200128, len 1032192, block_start 1106825216, block_len 1032192, orig_block_len 245760 A direct IO write at offset 1847296 and length of 360448 bytes (end offset at 2207744) starts, and at that moment the following extent maps exist for our inode: em A, start 0, orig_start 0, len 40960, block_start 18446744073709551613, block_len 0, orig_block_len 0 em B, start 40960, orig_start 40960, len 270336, block_start 1106399232, block_len 270336, orig_block_len 376832 em C, start 311296, orig_start 311296, len 937984, block_start 1112842240, block_len 937984, orig_block_len 937984 em E (prealloc), start 1249280, orig_start 1200128, len 1032192, block_start 1106825216, block_len 1032192, orig_block_len 245760 The dio write results in drop_extent_cache() being called twice. The first time for a range that starts at offset 1847296 and ends at offset 2035711 (length of 188416), which results in a double split of extent map E, replacing it with two new extent maps: em F, start 1249280, orig_start 1200128, block_start 1106825216, block_len 598016, orig_block_len 598016 em G, start 2035712, orig_start 1200128, block_start 1107611648, block_len 245760, orig_block_len 1032192 It also creates a new extent map that represents a part of the requested IO (through create_io_em()): em H, start 1847296, len 188416, block_start 1107423232, block_len 188416 The second call to drop_extent_cache() has a range with a start offset of 2035712 and end offset of 2207743 (length of 172032). This leads to replacing extent map G with a new extent map I with the following values: em I, start 2207744, orig_start 1200128, block_start 1107783680, block_len 73728, orig_block_len 1032192 It also creates a new extent map that represents the second part of the requested IO (through create_io_em()): em J, start 2035712, len 172032, block_start 1107611648, block_len 172032 The dio write set the inode's i_size to 2207744 bytes. After the dio write the inode has the following extent maps: em A, start 0, orig_start 0, len 40960, block_start 18446744073709551613, block_len 0, orig_block_len 0 em B, start 40960, orig_start 40960, len 270336, block_start 1106399232, block_len 270336, orig_block_len 376832 em C, start 311296, orig_start 311296, len 937984, block_start 1112842240, block_len 937984, orig_block_len 937984 em F, start 1249280, orig_start 1200128, len 598016, block_start 1106825216, block_len 598016, orig_block_len 598016 em H, start 1847296, orig_start 1200128, len 188416, block_start 1107423232, block_len 188416, orig_block_len 835584 em J, start 2035712, orig_start 2035712, len 172032, block_start 1107611648, block_len 172032, orig_block_len 245760 em I, start 2207744, orig_start 1200128, len 73728, block_start 1107783680, block_len 73728, orig_block_len 1032192 Now do some change to the file, like adding a xattr for example and then fsync it again. This triggers a fast fsync path, and as of commit 471d557afed1 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay"), we use the extent map I to log a file extent item because it's a prealloc extent and it starts at an offset matching the inode's i_size. However when we log it, we create a file extent item with a value for the disk byte location that is wrong, as can be seen from the following output of "btrfs inspect-internal dump-tree": item 1 key (271 EXTENT_DATA 2207744) itemoff 3782 itemsize 53 generation 22 type 2 (prealloc) prealloc data disk byte 1106776064 nr 1032192 prealloc data offset 1007616 nr 73728 Here the disk byte value corresponds to calculation based on some fields from the extent map I: 1106776064 = block_start (1107783680) - 1007616 (extent_offset) extent_offset = 2207744 (start) - 1200128 (orig_start) = 1007616 The disk byte value of 1106776064 clashes with disk byte values of the file extent items at offsets 1249280 and 1847296 in the fs tree: item 6 key (271 EXTENT_DATA 1249280) itemoff 3568 itemsize 53 generation 20 type 2 (prealloc) prealloc data disk byte 1106776064 nr 835584 prealloc data offset 49152 nr 598016 item 7 key (271 EXTENT_DATA 1847296) itemoff 3515 itemsize 53 generation 20 type 1 (regular) extent data disk byte 1106776064 nr 835584 extent data offset 647168 nr 188416 ram 835584 extent compression 0 (none) item 8 key (271 EXTENT_DATA 2035712) itemoff 3462 itemsize 53 generation 20 type 1 (regular) extent data disk byte 1107611648 nr 245760 extent data offset 0 nr 172032 ram 245760 extent compression 0 (none) item 9 key (271 EXTENT_DATA 2207744) itemoff 3409 itemsize 53 generation 20 type 2 (prealloc) prealloc data disk byte 1107611648 nr 245760 prealloc data offset 172032 nr 73728 Instead of the disk byte value of 1106776064, the value of 1107611648 should have been logged. Also the data offset value should have been 172032 and not 1007616. After a log replay we end up getting two extent items in the extent tree with different lengths, one of 835584, which is correct and existed before the log replay, and another one of 1032192 which is wrong and is based on the logged file extent item: item 12 key (1106776064 EXTENT_ITEM 835584) itemoff 3406 itemsize 53 refs 2 gen 15 flags DATA extent data backref root 5 objectid 271 offset 1200128 count 2 item 13 key (1106776064 EXTENT_ITEM 1032192) itemoff 3353 itemsize 53 refs 1 gen 22 flags DATA extent data backref root 5 objectid 271 offset 1200128 count 1 Obviously this leads to many problems and a filesystem check reports many errors: (...) checking extents Extent back ref already exists for 1106776064 parent 0 root 5 owner 271 offset 1200128 num_refs 1 extent item 1106776064 has multiple extent items ref mismatch on [1106776064 835584] extent item 2, found 3 Incorrect local backref count on 1106776064 root 5 owner 271 offset 1200128 found 2 wanted 1 back 0x55b1d0ad7680 Backref 1106776064 root 5 owner 271 offset 1200128 num_refs 0 not found in extent tree Incorrect local backref count on 1106776064 root 5 owner 271 offset 1200128 found 1 wanted 0 back 0x55b1d0ad4e70 Backref bytes do not match extent backref, bytenr=1106776064, ref bytes=835584, backref bytes=1032192 backpointer mismatch on [1106776064 835584] checking free space cache block group 1103101952 has wrong amount of free space failed to load free space cache for block group 1103101952 checking fs roots (...) So fix this by logging the prealloc extents beyond the inode's i_size based on searches in the subvolume tree instead of the extent maps. Fixes: 471d557afed1 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay") CC: stable@vger.kernel.org # 4.14+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2018-05-09 18:01:46 +03:00
if (!ret)
ret = btrfs_log_prealloc_extents(trans, inode, path);
btrfs: make fast fsyncs wait only for writeback Currently regardless of a full or a fast fsync we always wait for ordered extents to complete, and then start logging the inode after that. However for fast fsyncs we can just wait for the writeback to complete, we don't need to wait for the ordered extents to complete since we use the list of modified extents maps to figure out which extents we must log and we can get their checksums directly from the ordered extents that are still in flight, otherwise look them up from the checksums tree. Until commit b5e6c3e170b770 ("btrfs: always wait on ordered extents at fsync time"), for fast fsyncs, we used to start logging without even waiting for the writeback to complete first, we would wait for it to complete after logging, while holding a transaction open, which lead to performance issues when using cgroups and probably for other cases too, as wait for IO while holding a transaction handle should be avoided as much as possible. After that, for fast fsyncs, we started to wait for ordered extents to complete before starting to log, which adds some latency to fsyncs and we even got at least one report about a performance drop which bisected to that particular change: https://lore.kernel.org/linux-btrfs/20181109215148.GF23260@techsingularity.net/ This change makes fast fsyncs only wait for writeback to finish before starting to log the inode, instead of waiting for both the writeback to finish and for the ordered extents to complete. This brings back part of the logic we had that extracts checksums from in flight ordered extents, which are not yet in the checksums tree, and making sure transaction commits wait for the completion of ordered extents previously logged (by far most of the time they have already completed by the time a transaction commit starts, resulting in no wait at all), to avoid any data loss if an ordered extent completes after the transaction used to log an inode is committed, followed by a power failure. When there are no other tasks accessing the checksums and the subvolume btrees, the ordered extent completion is pretty fast, typically taking 100 to 200 microseconds only in my observations. However when there are other tasks accessing these btrees, ordered extent completion can take a lot more time due to lock contention on nodes and leaves of these btrees. I've seen cases over 2 milliseconds, which starts to be significant. In particular when we do have concurrent fsyncs against different files there is a lot of contention on the checksums btree, since we have many tasks writing the checksums into the btree and other tasks that already started the logging phase are doing lookups for checksums in the btree. This change also turns all ranged fsyncs into full ranged fsyncs, which is something we already did when not using the NO_HOLES features or when doing a full fsync. This is to guarantee we never miss checksums due to writeback having been triggered only for a part of an extent, and we end up logging the full extent but only checksums for the written range, which results in missing checksums after log replay. Allowing ranged fsyncs to operate again only in the original range, when using the NO_HOLES feature and doing a fast fsync is doable but requires some non trivial changes to the writeback path, which can always be worked on later if needed, but I don't think they are a very common use case. Several tests were performed using fio for different numbers of concurrent jobs, each writing and fsyncing its own file, for both sequential and random file writes. The tests were run on bare metal, no virtualization, on a box with 12 cores (Intel i7-8700), 64Gb of RAM and a NVMe device, with a kernel configuration that is the default of typical distributions (debian in this case), without debug options enabled (kasan, kmemleak, slub debug, debug of page allocations, lock debugging, etc). The following script that calls fio was used: $ cat test-fsync.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/btrfs MOUNT_OPTIONS="-o ssd -o space_cache=v2" MKFS_OPTIONS="-d single -m single" if [ $# -ne 5 ]; then echo "Use $0 NUM_JOBS FILE_SIZE FSYNC_FREQ BLOCK_SIZE [write|randwrite]" exit 1 fi NUM_JOBS=$1 FILE_SIZE=$2 FSYNC_FREQ=$3 BLOCK_SIZE=$4 WRITE_MODE=$5 if [ "$WRITE_MODE" != "write" ] && [ "$WRITE_MODE" != "randwrite" ]; then echo "Invalid WRITE_MODE, must be 'write' or 'randwrite'" exit 1 fi cat <<EOF > /tmp/fio-job.ini [writers] rw=$WRITE_MODE fsync=$FSYNC_FREQ fallocate=none group_reporting=1 direct=0 bs=$BLOCK_SIZE ioengine=sync size=$FILE_SIZE directory=$MNT numjobs=$NUM_JOBS EOF echo "performance" | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor echo echo "Using config:" echo cat /tmp/fio-job.ini echo umount $MNT &> /dev/null mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT fio /tmp/fio-job.ini umount $MNT The results were the following: ************************* *** sequential writes *** ************************* ==== 1 job, 8GiB file, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=36.6MiB/s (38.4MB/s), 36.6MiB/s-36.6MiB/s (38.4MB/s-38.4MB/s), io=8192MiB (8590MB), run=223689-223689msec After patch: WRITE: bw=40.2MiB/s (42.1MB/s), 40.2MiB/s-40.2MiB/s (42.1MB/s-42.1MB/s), io=8192MiB (8590MB), run=203980-203980msec (+9.8%, -8.8% runtime) ==== 2 jobs, 4GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=35.8MiB/s (37.5MB/s), 35.8MiB/s-35.8MiB/s (37.5MB/s-37.5MB/s), io=8192MiB (8590MB), run=228950-228950msec After patch: WRITE: bw=43.5MiB/s (45.6MB/s), 43.5MiB/s-43.5MiB/s (45.6MB/s-45.6MB/s), io=8192MiB (8590MB), run=188272-188272msec (+21.5% throughput, -17.8% runtime) ==== 4 jobs, 2GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=50.1MiB/s (52.6MB/s), 50.1MiB/s-50.1MiB/s (52.6MB/s-52.6MB/s), io=8192MiB (8590MB), run=163446-163446msec After patch: WRITE: bw=64.5MiB/s (67.6MB/s), 64.5MiB/s-64.5MiB/s (67.6MB/s-67.6MB/s), io=8192MiB (8590MB), run=126987-126987msec (+28.7% throughput, -22.3% runtime) ==== 8 jobs, 1GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=64.0MiB/s (68.1MB/s), 64.0MiB/s-64.0MiB/s (68.1MB/s-68.1MB/s), io=8192MiB (8590MB), run=126075-126075msec After patch: WRITE: bw=86.8MiB/s (91.0MB/s), 86.8MiB/s-86.8MiB/s (91.0MB/s-91.0MB/s), io=8192MiB (8590MB), run=94358-94358msec (+35.6% throughput, -25.2% runtime) ==== 16 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=79.8MiB/s (83.6MB/s), 79.8MiB/s-79.8MiB/s (83.6MB/s-83.6MB/s), io=8192MiB (8590MB), run=102694-102694msec After patch: WRITE: bw=107MiB/s (112MB/s), 107MiB/s-107MiB/s (112MB/s-112MB/s), io=8192MiB (8590MB), run=76446-76446msec (+34.1% throughput, -25.6% runtime) ==== 32 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=93.2MiB/s (97.7MB/s), 93.2MiB/s-93.2MiB/s (97.7MB/s-97.7MB/s), io=16.0GiB (17.2GB), run=175836-175836msec After patch: WRITE: bw=111MiB/s (117MB/s), 111MiB/s-111MiB/s (117MB/s-117MB/s), io=16.0GiB (17.2GB), run=147001-147001msec (+19.1% throughput, -16.4% runtime) ==== 64 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=108MiB/s (114MB/s), 108MiB/s-108MiB/s (114MB/s-114MB/s), io=32.0GiB (34.4GB), run=302656-302656msec After patch: WRITE: bw=133MiB/s (140MB/s), 133MiB/s-133MiB/s (140MB/s-140MB/s), io=32.0GiB (34.4GB), run=246003-246003msec (+23.1% throughput, -18.7% runtime) ************************ *** random writes *** ************************ ==== 1 job, 8GiB file, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=11.5MiB/s (12.0MB/s), 11.5MiB/s-11.5MiB/s (12.0MB/s-12.0MB/s), io=8192MiB (8590MB), run=714281-714281msec After patch: WRITE: bw=11.6MiB/s (12.2MB/s), 11.6MiB/s-11.6MiB/s (12.2MB/s-12.2MB/s), io=8192MiB (8590MB), run=705959-705959msec (+0.9% throughput, -1.7% runtime) ==== 2 jobs, 4GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=12.8MiB/s (13.5MB/s), 12.8MiB/s-12.8MiB/s (13.5MB/s-13.5MB/s), io=8192MiB (8590MB), run=638101-638101msec After patch: WRITE: bw=13.1MiB/s (13.7MB/s), 13.1MiB/s-13.1MiB/s (13.7MB/s-13.7MB/s), io=8192MiB (8590MB), run=625374-625374msec (+2.3% throughput, -2.0% runtime) ==== 4 jobs, 2GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=15.4MiB/s (16.2MB/s), 15.4MiB/s-15.4MiB/s (16.2MB/s-16.2MB/s), io=8192MiB (8590MB), run=531146-531146msec After patch: WRITE: bw=17.8MiB/s (18.7MB/s), 17.8MiB/s-17.8MiB/s (18.7MB/s-18.7MB/s), io=8192MiB (8590MB), run=460431-460431msec (+15.6% throughput, -13.3% runtime) ==== 8 jobs, 1GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=19.9MiB/s (20.8MB/s), 19.9MiB/s-19.9MiB/s (20.8MB/s-20.8MB/s), io=8192MiB (8590MB), run=412664-412664msec After patch: WRITE: bw=22.2MiB/s (23.3MB/s), 22.2MiB/s-22.2MiB/s (23.3MB/s-23.3MB/s), io=8192MiB (8590MB), run=368589-368589msec (+11.6% throughput, -10.7% runtime) ==== 16 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=29.3MiB/s (30.7MB/s), 29.3MiB/s-29.3MiB/s (30.7MB/s-30.7MB/s), io=8192MiB (8590MB), run=279924-279924msec After patch: WRITE: bw=30.4MiB/s (31.9MB/s), 30.4MiB/s-30.4MiB/s (31.9MB/s-31.9MB/s), io=8192MiB (8590MB), run=269258-269258msec (+3.8% throughput, -3.8% runtime) ==== 32 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=36.9MiB/s (38.7MB/s), 36.9MiB/s-36.9MiB/s (38.7MB/s-38.7MB/s), io=16.0GiB (17.2GB), run=443581-443581msec After patch: WRITE: bw=41.6MiB/s (43.6MB/s), 41.6MiB/s-41.6MiB/s (43.6MB/s-43.6MB/s), io=16.0GiB (17.2GB), run=394114-394114msec (+12.7% throughput, -11.2% runtime) ==== 64 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=45.9MiB/s (48.1MB/s), 45.9MiB/s-45.9MiB/s (48.1MB/s-48.1MB/s), io=32.0GiB (34.4GB), run=714614-714614msec After patch: WRITE: bw=48.8MiB/s (51.1MB/s), 48.8MiB/s-48.8MiB/s (51.1MB/s-51.1MB/s), io=32.0GiB (34.4GB), run=672087-672087msec (+6.3% throughput, -6.0% runtime) Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-08-11 14:43:58 +03:00
if (ret)
return ret;
Btrfs: fix duplicate extents after fsync of file with prealloc extents In commit 471d557afed1 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay"), on fsync, we started to always log all prealloc extents beyond an inode's i_size in order to avoid losing them after a power failure. However under some cases this can lead to the log replay code to create duplicate extent items, with different lengths, in the extent tree. That happens because, as of that commit, we can now log extent items based on extent maps that are not on the "modified" list of extent maps of the inode's extent map tree. Logging extent items based on extent maps is used during the fast fsync path to save time and for this to work reliably it requires that the extent maps are not merged with other adjacent extent maps - having the extent maps in the list of modified extents gives such guarantee. Consider the following example, captured during a long run of fsstress, which illustrates this problem. We have inode 271, in the filesystem tree (root 5), for which all of the following operations and discussion apply to. A buffered write starts at offset 312391 with a length of 933471 bytes (end offset at 1245862). At this point we have, for this inode, the following extent maps with the their field values: em A, start 0, orig_start 0, len 40960, block_start 18446744073709551613, block_len 0, orig_block_len 0 em B, start 40960, orig_start 40960, len 376832, block_start 1106399232, block_len 376832, orig_block_len 376832 em C, start 417792, orig_start 417792, len 782336, block_start 18446744073709551613, block_len 0, orig_block_len 0 em D, start 1200128, orig_start 1200128, len 835584, block_start 1106776064, block_len 835584, orig_block_len 835584 em E, start 2035712, orig_start 2035712, len 245760, block_start 1107611648, block_len 245760, orig_block_len 245760 Extent map A corresponds to a hole and extent maps D and E correspond to preallocated extents. Extent map D ends where extent map E begins (1106776064 + 835584 = 1107611648), but these extent maps were not merged because they are in the inode's list of modified extent maps. An fsync against this inode is made, which triggers the fast path (BTRFS_INODE_NEEDS_FULL_SYNC is not set). This fsync triggers writeback of the data previously written using buffered IO, and when the respective ordered extent finishes, btrfs_drop_extents() is called against the (aligned) range 311296..1249279. This causes a split of extent map D at btrfs_drop_extent_cache(), replacing extent map D with a new extent map D', also added to the list of modified extents, with the following values: em D', start 1249280, orig_start of 1200128, block_start 1106825216 (= 1106776064 + 1249280 - 1200128), orig_block_len 835584, block_len 786432 (835584 - (1249280 - 1200128)) Then, during the fast fsync, btrfs_log_changed_extents() is called and extent maps D' and E are removed from the list of modified extents. The flag EXTENT_FLAG_LOGGING is also set on them. After the extents are logged clear_em_logging() is called on each of them, and that makes extent map E to be merged with extent map D' (try_merge_map()), resulting in D' being deleted and E adjusted to: em E, start 1249280, orig_start 1200128, len 1032192, block_start 1106825216, block_len 1032192, orig_block_len 245760 A direct IO write at offset 1847296 and length of 360448 bytes (end offset at 2207744) starts, and at that moment the following extent maps exist for our inode: em A, start 0, orig_start 0, len 40960, block_start 18446744073709551613, block_len 0, orig_block_len 0 em B, start 40960, orig_start 40960, len 270336, block_start 1106399232, block_len 270336, orig_block_len 376832 em C, start 311296, orig_start 311296, len 937984, block_start 1112842240, block_len 937984, orig_block_len 937984 em E (prealloc), start 1249280, orig_start 1200128, len 1032192, block_start 1106825216, block_len 1032192, orig_block_len 245760 The dio write results in drop_extent_cache() being called twice. The first time for a range that starts at offset 1847296 and ends at offset 2035711 (length of 188416), which results in a double split of extent map E, replacing it with two new extent maps: em F, start 1249280, orig_start 1200128, block_start 1106825216, block_len 598016, orig_block_len 598016 em G, start 2035712, orig_start 1200128, block_start 1107611648, block_len 245760, orig_block_len 1032192 It also creates a new extent map that represents a part of the requested IO (through create_io_em()): em H, start 1847296, len 188416, block_start 1107423232, block_len 188416 The second call to drop_extent_cache() has a range with a start offset of 2035712 and end offset of 2207743 (length of 172032). This leads to replacing extent map G with a new extent map I with the following values: em I, start 2207744, orig_start 1200128, block_start 1107783680, block_len 73728, orig_block_len 1032192 It also creates a new extent map that represents the second part of the requested IO (through create_io_em()): em J, start 2035712, len 172032, block_start 1107611648, block_len 172032 The dio write set the inode's i_size to 2207744 bytes. After the dio write the inode has the following extent maps: em A, start 0, orig_start 0, len 40960, block_start 18446744073709551613, block_len 0, orig_block_len 0 em B, start 40960, orig_start 40960, len 270336, block_start 1106399232, block_len 270336, orig_block_len 376832 em C, start 311296, orig_start 311296, len 937984, block_start 1112842240, block_len 937984, orig_block_len 937984 em F, start 1249280, orig_start 1200128, len 598016, block_start 1106825216, block_len 598016, orig_block_len 598016 em H, start 1847296, orig_start 1200128, len 188416, block_start 1107423232, block_len 188416, orig_block_len 835584 em J, start 2035712, orig_start 2035712, len 172032, block_start 1107611648, block_len 172032, orig_block_len 245760 em I, start 2207744, orig_start 1200128, len 73728, block_start 1107783680, block_len 73728, orig_block_len 1032192 Now do some change to the file, like adding a xattr for example and then fsync it again. This triggers a fast fsync path, and as of commit 471d557afed1 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay"), we use the extent map I to log a file extent item because it's a prealloc extent and it starts at an offset matching the inode's i_size. However when we log it, we create a file extent item with a value for the disk byte location that is wrong, as can be seen from the following output of "btrfs inspect-internal dump-tree": item 1 key (271 EXTENT_DATA 2207744) itemoff 3782 itemsize 53 generation 22 type 2 (prealloc) prealloc data disk byte 1106776064 nr 1032192 prealloc data offset 1007616 nr 73728 Here the disk byte value corresponds to calculation based on some fields from the extent map I: 1106776064 = block_start (1107783680) - 1007616 (extent_offset) extent_offset = 2207744 (start) - 1200128 (orig_start) = 1007616 The disk byte value of 1106776064 clashes with disk byte values of the file extent items at offsets 1249280 and 1847296 in the fs tree: item 6 key (271 EXTENT_DATA 1249280) itemoff 3568 itemsize 53 generation 20 type 2 (prealloc) prealloc data disk byte 1106776064 nr 835584 prealloc data offset 49152 nr 598016 item 7 key (271 EXTENT_DATA 1847296) itemoff 3515 itemsize 53 generation 20 type 1 (regular) extent data disk byte 1106776064 nr 835584 extent data offset 647168 nr 188416 ram 835584 extent compression 0 (none) item 8 key (271 EXTENT_DATA 2035712) itemoff 3462 itemsize 53 generation 20 type 1 (regular) extent data disk byte 1107611648 nr 245760 extent data offset 0 nr 172032 ram 245760 extent compression 0 (none) item 9 key (271 EXTENT_DATA 2207744) itemoff 3409 itemsize 53 generation 20 type 2 (prealloc) prealloc data disk byte 1107611648 nr 245760 prealloc data offset 172032 nr 73728 Instead of the disk byte value of 1106776064, the value of 1107611648 should have been logged. Also the data offset value should have been 172032 and not 1007616. After a log replay we end up getting two extent items in the extent tree with different lengths, one of 835584, which is correct and existed before the log replay, and another one of 1032192 which is wrong and is based on the logged file extent item: item 12 key (1106776064 EXTENT_ITEM 835584) itemoff 3406 itemsize 53 refs 2 gen 15 flags DATA extent data backref root 5 objectid 271 offset 1200128 count 2 item 13 key (1106776064 EXTENT_ITEM 1032192) itemoff 3353 itemsize 53 refs 1 gen 22 flags DATA extent data backref root 5 objectid 271 offset 1200128 count 1 Obviously this leads to many problems and a filesystem check reports many errors: (...) checking extents Extent back ref already exists for 1106776064 parent 0 root 5 owner 271 offset 1200128 num_refs 1 extent item 1106776064 has multiple extent items ref mismatch on [1106776064 835584] extent item 2, found 3 Incorrect local backref count on 1106776064 root 5 owner 271 offset 1200128 found 2 wanted 1 back 0x55b1d0ad7680 Backref 1106776064 root 5 owner 271 offset 1200128 num_refs 0 not found in extent tree Incorrect local backref count on 1106776064 root 5 owner 271 offset 1200128 found 1 wanted 0 back 0x55b1d0ad4e70 Backref bytes do not match extent backref, bytenr=1106776064, ref bytes=835584, backref bytes=1032192 backpointer mismatch on [1106776064 835584] checking free space cache block group 1103101952 has wrong amount of free space failed to load free space cache for block group 1103101952 checking fs roots (...) So fix this by logging the prealloc extents beyond the inode's i_size based on searches in the subvolume tree instead of the extent maps. Fixes: 471d557afed1 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay") CC: stable@vger.kernel.org # 4.14+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2018-05-09 18:01:46 +03:00
btrfs: make fast fsyncs wait only for writeback Currently regardless of a full or a fast fsync we always wait for ordered extents to complete, and then start logging the inode after that. However for fast fsyncs we can just wait for the writeback to complete, we don't need to wait for the ordered extents to complete since we use the list of modified extents maps to figure out which extents we must log and we can get their checksums directly from the ordered extents that are still in flight, otherwise look them up from the checksums tree. Until commit b5e6c3e170b770 ("btrfs: always wait on ordered extents at fsync time"), for fast fsyncs, we used to start logging without even waiting for the writeback to complete first, we would wait for it to complete after logging, while holding a transaction open, which lead to performance issues when using cgroups and probably for other cases too, as wait for IO while holding a transaction handle should be avoided as much as possible. After that, for fast fsyncs, we started to wait for ordered extents to complete before starting to log, which adds some latency to fsyncs and we even got at least one report about a performance drop which bisected to that particular change: https://lore.kernel.org/linux-btrfs/20181109215148.GF23260@techsingularity.net/ This change makes fast fsyncs only wait for writeback to finish before starting to log the inode, instead of waiting for both the writeback to finish and for the ordered extents to complete. This brings back part of the logic we had that extracts checksums from in flight ordered extents, which are not yet in the checksums tree, and making sure transaction commits wait for the completion of ordered extents previously logged (by far most of the time they have already completed by the time a transaction commit starts, resulting in no wait at all), to avoid any data loss if an ordered extent completes after the transaction used to log an inode is committed, followed by a power failure. When there are no other tasks accessing the checksums and the subvolume btrees, the ordered extent completion is pretty fast, typically taking 100 to 200 microseconds only in my observations. However when there are other tasks accessing these btrees, ordered extent completion can take a lot more time due to lock contention on nodes and leaves of these btrees. I've seen cases over 2 milliseconds, which starts to be significant. In particular when we do have concurrent fsyncs against different files there is a lot of contention on the checksums btree, since we have many tasks writing the checksums into the btree and other tasks that already started the logging phase are doing lookups for checksums in the btree. This change also turns all ranged fsyncs into full ranged fsyncs, which is something we already did when not using the NO_HOLES features or when doing a full fsync. This is to guarantee we never miss checksums due to writeback having been triggered only for a part of an extent, and we end up logging the full extent but only checksums for the written range, which results in missing checksums after log replay. Allowing ranged fsyncs to operate again only in the original range, when using the NO_HOLES feature and doing a fast fsync is doable but requires some non trivial changes to the writeback path, which can always be worked on later if needed, but I don't think they are a very common use case. Several tests were performed using fio for different numbers of concurrent jobs, each writing and fsyncing its own file, for both sequential and random file writes. The tests were run on bare metal, no virtualization, on a box with 12 cores (Intel i7-8700), 64Gb of RAM and a NVMe device, with a kernel configuration that is the default of typical distributions (debian in this case), without debug options enabled (kasan, kmemleak, slub debug, debug of page allocations, lock debugging, etc). The following script that calls fio was used: $ cat test-fsync.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/btrfs MOUNT_OPTIONS="-o ssd -o space_cache=v2" MKFS_OPTIONS="-d single -m single" if [ $# -ne 5 ]; then echo "Use $0 NUM_JOBS FILE_SIZE FSYNC_FREQ BLOCK_SIZE [write|randwrite]" exit 1 fi NUM_JOBS=$1 FILE_SIZE=$2 FSYNC_FREQ=$3 BLOCK_SIZE=$4 WRITE_MODE=$5 if [ "$WRITE_MODE" != "write" ] && [ "$WRITE_MODE" != "randwrite" ]; then echo "Invalid WRITE_MODE, must be 'write' or 'randwrite'" exit 1 fi cat <<EOF > /tmp/fio-job.ini [writers] rw=$WRITE_MODE fsync=$FSYNC_FREQ fallocate=none group_reporting=1 direct=0 bs=$BLOCK_SIZE ioengine=sync size=$FILE_SIZE directory=$MNT numjobs=$NUM_JOBS EOF echo "performance" | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor echo echo "Using config:" echo cat /tmp/fio-job.ini echo umount $MNT &> /dev/null mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT fio /tmp/fio-job.ini umount $MNT The results were the following: ************************* *** sequential writes *** ************************* ==== 1 job, 8GiB file, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=36.6MiB/s (38.4MB/s), 36.6MiB/s-36.6MiB/s (38.4MB/s-38.4MB/s), io=8192MiB (8590MB), run=223689-223689msec After patch: WRITE: bw=40.2MiB/s (42.1MB/s), 40.2MiB/s-40.2MiB/s (42.1MB/s-42.1MB/s), io=8192MiB (8590MB), run=203980-203980msec (+9.8%, -8.8% runtime) ==== 2 jobs, 4GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=35.8MiB/s (37.5MB/s), 35.8MiB/s-35.8MiB/s (37.5MB/s-37.5MB/s), io=8192MiB (8590MB), run=228950-228950msec After patch: WRITE: bw=43.5MiB/s (45.6MB/s), 43.5MiB/s-43.5MiB/s (45.6MB/s-45.6MB/s), io=8192MiB (8590MB), run=188272-188272msec (+21.5% throughput, -17.8% runtime) ==== 4 jobs, 2GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=50.1MiB/s (52.6MB/s), 50.1MiB/s-50.1MiB/s (52.6MB/s-52.6MB/s), io=8192MiB (8590MB), run=163446-163446msec After patch: WRITE: bw=64.5MiB/s (67.6MB/s), 64.5MiB/s-64.5MiB/s (67.6MB/s-67.6MB/s), io=8192MiB (8590MB), run=126987-126987msec (+28.7% throughput, -22.3% runtime) ==== 8 jobs, 1GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=64.0MiB/s (68.1MB/s), 64.0MiB/s-64.0MiB/s (68.1MB/s-68.1MB/s), io=8192MiB (8590MB), run=126075-126075msec After patch: WRITE: bw=86.8MiB/s (91.0MB/s), 86.8MiB/s-86.8MiB/s (91.0MB/s-91.0MB/s), io=8192MiB (8590MB), run=94358-94358msec (+35.6% throughput, -25.2% runtime) ==== 16 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=79.8MiB/s (83.6MB/s), 79.8MiB/s-79.8MiB/s (83.6MB/s-83.6MB/s), io=8192MiB (8590MB), run=102694-102694msec After patch: WRITE: bw=107MiB/s (112MB/s), 107MiB/s-107MiB/s (112MB/s-112MB/s), io=8192MiB (8590MB), run=76446-76446msec (+34.1% throughput, -25.6% runtime) ==== 32 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=93.2MiB/s (97.7MB/s), 93.2MiB/s-93.2MiB/s (97.7MB/s-97.7MB/s), io=16.0GiB (17.2GB), run=175836-175836msec After patch: WRITE: bw=111MiB/s (117MB/s), 111MiB/s-111MiB/s (117MB/s-117MB/s), io=16.0GiB (17.2GB), run=147001-147001msec (+19.1% throughput, -16.4% runtime) ==== 64 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=108MiB/s (114MB/s), 108MiB/s-108MiB/s (114MB/s-114MB/s), io=32.0GiB (34.4GB), run=302656-302656msec After patch: WRITE: bw=133MiB/s (140MB/s), 133MiB/s-133MiB/s (140MB/s-140MB/s), io=32.0GiB (34.4GB), run=246003-246003msec (+23.1% throughput, -18.7% runtime) ************************ *** random writes *** ************************ ==== 1 job, 8GiB file, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=11.5MiB/s (12.0MB/s), 11.5MiB/s-11.5MiB/s (12.0MB/s-12.0MB/s), io=8192MiB (8590MB), run=714281-714281msec After patch: WRITE: bw=11.6MiB/s (12.2MB/s), 11.6MiB/s-11.6MiB/s (12.2MB/s-12.2MB/s), io=8192MiB (8590MB), run=705959-705959msec (+0.9% throughput, -1.7% runtime) ==== 2 jobs, 4GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=12.8MiB/s (13.5MB/s), 12.8MiB/s-12.8MiB/s (13.5MB/s-13.5MB/s), io=8192MiB (8590MB), run=638101-638101msec After patch: WRITE: bw=13.1MiB/s (13.7MB/s), 13.1MiB/s-13.1MiB/s (13.7MB/s-13.7MB/s), io=8192MiB (8590MB), run=625374-625374msec (+2.3% throughput, -2.0% runtime) ==== 4 jobs, 2GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=15.4MiB/s (16.2MB/s), 15.4MiB/s-15.4MiB/s (16.2MB/s-16.2MB/s), io=8192MiB (8590MB), run=531146-531146msec After patch: WRITE: bw=17.8MiB/s (18.7MB/s), 17.8MiB/s-17.8MiB/s (18.7MB/s-18.7MB/s), io=8192MiB (8590MB), run=460431-460431msec (+15.6% throughput, -13.3% runtime) ==== 8 jobs, 1GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=19.9MiB/s (20.8MB/s), 19.9MiB/s-19.9MiB/s (20.8MB/s-20.8MB/s), io=8192MiB (8590MB), run=412664-412664msec After patch: WRITE: bw=22.2MiB/s (23.3MB/s), 22.2MiB/s-22.2MiB/s (23.3MB/s-23.3MB/s), io=8192MiB (8590MB), run=368589-368589msec (+11.6% throughput, -10.7% runtime) ==== 16 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=29.3MiB/s (30.7MB/s), 29.3MiB/s-29.3MiB/s (30.7MB/s-30.7MB/s), io=8192MiB (8590MB), run=279924-279924msec After patch: WRITE: bw=30.4MiB/s (31.9MB/s), 30.4MiB/s-30.4MiB/s (31.9MB/s-31.9MB/s), io=8192MiB (8590MB), run=269258-269258msec (+3.8% throughput, -3.8% runtime) ==== 32 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=36.9MiB/s (38.7MB/s), 36.9MiB/s-36.9MiB/s (38.7MB/s-38.7MB/s), io=16.0GiB (17.2GB), run=443581-443581msec After patch: WRITE: bw=41.6MiB/s (43.6MB/s), 41.6MiB/s-41.6MiB/s (43.6MB/s-43.6MB/s), io=16.0GiB (17.2GB), run=394114-394114msec (+12.7% throughput, -11.2% runtime) ==== 64 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=45.9MiB/s (48.1MB/s), 45.9MiB/s-45.9MiB/s (48.1MB/s-48.1MB/s), io=32.0GiB (34.4GB), run=714614-714614msec After patch: WRITE: bw=48.8MiB/s (51.1MB/s), 48.8MiB/s-48.8MiB/s (51.1MB/s-51.1MB/s), io=32.0GiB (34.4GB), run=672087-672087msec (+6.3% throughput, -6.0% runtime) Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-08-11 14:43:58 +03:00
/*
* We have logged all extents successfully, now make sure the commit of
* the current transaction waits for the ordered extents to complete
* before it commits and wipes out the log trees, otherwise we would
* lose data if an ordered extents completes after the transaction
* commits and a power failure happens after the transaction commit.
*/
list_for_each_entry_safe(ordered, tmp, &ctx->ordered_extents, log_list) {
list_del_init(&ordered->log_list);
set_bit(BTRFS_ORDERED_LOGGED, &ordered->flags);
if (!test_bit(BTRFS_ORDERED_COMPLETE, &ordered->flags)) {
spin_lock_irq(&inode->ordered_tree.lock);
if (!test_bit(BTRFS_ORDERED_COMPLETE, &ordered->flags)) {
set_bit(BTRFS_ORDERED_PENDING, &ordered->flags);
atomic_inc(&trans->transaction->pending_ordered);
}
spin_unlock_irq(&inode->ordered_tree.lock);
}
btrfs_put_ordered_extent(ordered);
}
return 0;
Btrfs: turbo charge fsync At least for the vm workload. Currently on fsync we will 1) Truncate all items in the log tree for the given inode if they exist and 2) Copy all items for a given inode into the log The problem with this is that for things like VMs you can have lots of extents from the fragmented writing behavior, and worst yet you may have only modified a few extents, not the entire thing. This patch fixes this problem by tracking which transid modified our extent, and then when we do the tree logging we find all of the extents we've modified in our current transaction, sort them and commit them. We also only truncate up to the xattrs of the inode and copy that stuff in normally, and then just drop any extents in the range we have that exist in the log already. Here are some numbers of a 50 meg fio job that does random writes and fsync()s after every write Original Patched SATA drive 82KB/s 140KB/s Fusion drive 431KB/s 2532KB/s So around 2-6 times faster depending on your hardware. There are a few corner cases, for example if you truncate at all we have to do it the old way since there is no way to be sure what is in the log is ok. This probably could be done smarter, but if you write-fsync-truncate-write-fsync you deserve what you get. All this work is in RAM of course so if your inode gets evicted from cache and you read it in and fsync it we'll do it the slow way if we are still in the same transaction that we last modified the inode in. The biggest cool part of this is that it requires no changes to the recovery code, so if you fsync with this patch and crash and load an old kernel, it will run the recovery and be a-ok. I have tested this pretty thoroughly with an fsync tester and everything comes back fine, as well as xfstests. Thanks, Signed-off-by: Josef Bacik <jbacik@fusionio.com>
2012-08-17 21:14:17 +04:00
}
static int logged_inode_size(struct btrfs_root *log, struct btrfs_inode *inode,
Btrfs: fix fsync data loss after adding hard link to inode We have a scenario where after the fsync log replay we can lose file data that had been previously fsync'ed if we added an hard link for our inode and after that we sync'ed the fsync log (for example by fsync'ing some other file or directory). This is because when adding an hard link we updated the inode item in the log tree with an i_size value of 0. At that point the new inode item was in memory only and a subsequent fsync log replay would not make us lose the file data. However if after adding the hard link we sync the log tree to disk, by fsync'ing some other file or directory for example, we ended up losing the file data after log replay, because the inode item in the persisted log tree had an an i_size of zero. This is easy to reproduce, and the following excerpt from my test for xfstests shows this: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create one file with data and fsync it. # This made the btrfs fsync log persist the data and the inode metadata with # a correct inode->i_size (4096 bytes). $XFS_IO_PROG -f -c "pwrite -S 0xaa -b 4K 0 4K" -c "fsync" \ $SCRATCH_MNT/foo | _filter_xfs_io # Now add one hard link to our file. This made the btrfs code update the fsync # log, in memory only, with an inode metadata having a size of 0. ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link # Now force persistence of the fsync log to disk, for example, by fsyncing some # other file. touch $SCRATCH_MNT/bar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/bar # Before a power loss or crash, we could read the 4Kb of data from our file as # expected. echo "File content before:" od -t x1 $SCRATCH_MNT/foo # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # After the fsync log replay, because the fsync log had a value of 0 for our # inode's i_size, we couldn't read anymore the 4Kb of data that we previously # wrote and fsync'ed. The size of the file became 0 after the fsync log replay. echo "File content after:" od -t x1 $SCRATCH_MNT/foo Another alternative test, that doesn't need to fsync an inode in the same transaction it was created, is: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create our test file with some data. $XFS_IO_PROG -f -c "pwrite -S 0xaa -b 8K 0 8K" \ $SCRATCH_MNT/foo | _filter_xfs_io # Make sure the file is durably persisted. sync # Append some data to our file, to increase its size. $XFS_IO_PROG -f -c "pwrite -S 0xcc -b 4K 8K 4K" \ $SCRATCH_MNT/foo | _filter_xfs_io # Fsync the file, so from this point on if a crash/power failure happens, our # new data is guaranteed to be there next time the fs is mounted. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foo # Add one hard link to our file. This made btrfs write into the in memory fsync # log a special inode with generation 0 and an i_size of 0 too. Note that this # didn't update the inode in the fsync log on disk. ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link # Now make sure the in memory fsync log is durably persisted. # Creating and fsync'ing another file will do it. touch $SCRATCH_MNT/bar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/bar # As expected, before the crash/power failure, we should be able to read the # 12Kb of file data. echo "File content before:" od -t x1 $SCRATCH_MNT/foo # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # After mounting the fs again, the fsync log was replayed. # The btrfs fsync log replay code didn't update the i_size of the persisted # inode because the inode item in the log had a special generation with a # value of 0 (and it couldn't know the correct i_size, since that inode item # had a 0 i_size too). This made the last 4Kb of file data inaccessible and # effectively lost. echo "File content after:" od -t x1 $SCRATCH_MNT/foo This isn't a new issue/regression. This problem has been around since the log tree code was added in 2008: Btrfs: Add a write ahead tree log to optimize synchronous operations (commit e02119d5a7b4396c5a872582fddc8bd6d305a70a) Test cases for xfstests follow soon. CC: <stable@vger.kernel.org> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-02-13 15:30:56 +03:00
struct btrfs_path *path, u64 *size_ret)
{
struct btrfs_key key;
int ret;
key.objectid = btrfs_ino(inode);
Btrfs: fix fsync data loss after adding hard link to inode We have a scenario where after the fsync log replay we can lose file data that had been previously fsync'ed if we added an hard link for our inode and after that we sync'ed the fsync log (for example by fsync'ing some other file or directory). This is because when adding an hard link we updated the inode item in the log tree with an i_size value of 0. At that point the new inode item was in memory only and a subsequent fsync log replay would not make us lose the file data. However if after adding the hard link we sync the log tree to disk, by fsync'ing some other file or directory for example, we ended up losing the file data after log replay, because the inode item in the persisted log tree had an an i_size of zero. This is easy to reproduce, and the following excerpt from my test for xfstests shows this: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create one file with data and fsync it. # This made the btrfs fsync log persist the data and the inode metadata with # a correct inode->i_size (4096 bytes). $XFS_IO_PROG -f -c "pwrite -S 0xaa -b 4K 0 4K" -c "fsync" \ $SCRATCH_MNT/foo | _filter_xfs_io # Now add one hard link to our file. This made the btrfs code update the fsync # log, in memory only, with an inode metadata having a size of 0. ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link # Now force persistence of the fsync log to disk, for example, by fsyncing some # other file. touch $SCRATCH_MNT/bar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/bar # Before a power loss or crash, we could read the 4Kb of data from our file as # expected. echo "File content before:" od -t x1 $SCRATCH_MNT/foo # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # After the fsync log replay, because the fsync log had a value of 0 for our # inode's i_size, we couldn't read anymore the 4Kb of data that we previously # wrote and fsync'ed. The size of the file became 0 after the fsync log replay. echo "File content after:" od -t x1 $SCRATCH_MNT/foo Another alternative test, that doesn't need to fsync an inode in the same transaction it was created, is: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create our test file with some data. $XFS_IO_PROG -f -c "pwrite -S 0xaa -b 8K 0 8K" \ $SCRATCH_MNT/foo | _filter_xfs_io # Make sure the file is durably persisted. sync # Append some data to our file, to increase its size. $XFS_IO_PROG -f -c "pwrite -S 0xcc -b 4K 8K 4K" \ $SCRATCH_MNT/foo | _filter_xfs_io # Fsync the file, so from this point on if a crash/power failure happens, our # new data is guaranteed to be there next time the fs is mounted. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foo # Add one hard link to our file. This made btrfs write into the in memory fsync # log a special inode with generation 0 and an i_size of 0 too. Note that this # didn't update the inode in the fsync log on disk. ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link # Now make sure the in memory fsync log is durably persisted. # Creating and fsync'ing another file will do it. touch $SCRATCH_MNT/bar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/bar # As expected, before the crash/power failure, we should be able to read the # 12Kb of file data. echo "File content before:" od -t x1 $SCRATCH_MNT/foo # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # After mounting the fs again, the fsync log was replayed. # The btrfs fsync log replay code didn't update the i_size of the persisted # inode because the inode item in the log had a special generation with a # value of 0 (and it couldn't know the correct i_size, since that inode item # had a 0 i_size too). This made the last 4Kb of file data inaccessible and # effectively lost. echo "File content after:" od -t x1 $SCRATCH_MNT/foo This isn't a new issue/regression. This problem has been around since the log tree code was added in 2008: Btrfs: Add a write ahead tree log to optimize synchronous operations (commit e02119d5a7b4396c5a872582fddc8bd6d305a70a) Test cases for xfstests follow soon. CC: <stable@vger.kernel.org> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-02-13 15:30:56 +03:00
key.type = BTRFS_INODE_ITEM_KEY;
key.offset = 0;
ret = btrfs_search_slot(NULL, log, &key, path, 0, 0);
if (ret < 0) {
return ret;
} else if (ret > 0) {
Btrfs: fix metadata inconsistencies after directory fsync We can get into inconsistency between inodes and directory entries after fsyncing a directory. The issue is that while a directory gets the new dentries persisted in the fsync log and replayed at mount time, the link count of the inode that directory entries point to doesn't get updated, staying with an incorrect link count (smaller then the correct value). This later leads to stale file handle errors when accessing (including attempt to delete) some of the links if all the other ones are removed, which also implies impossibility to delete the parent directories, since the dentries can not be removed. Another issue is that (unlike ext3/4, xfs, f2fs, reiserfs, nilfs2), when fsyncing a directory, new files aren't logged (their metadata and dentries) nor any child directories. So this patch fixes this issue too, since it has the same resolution as the incorrect inode link count issue mentioned before. This is very easy to reproduce, and the following excerpt from my test case for xfstests shows how: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create our main test file and directory. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 8K" $SCRATCH_MNT/foo | _filter_xfs_io mkdir $SCRATCH_MNT/mydir # Make sure all metadata and data are durably persisted. sync # Add a hard link to 'foo' inside our test directory and fsync only the # directory. The btrfs fsync implementation had a bug that caused the new # directory entry to be visible after the fsync log replay but, the inode # of our file remained with a link count of 1. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_2 # Add a few more links and new files. # This is just to verify nothing breaks or gives incorrect results after the # fsync log is replayed. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_3 $XFS_IO_PROG -f -c "pwrite -S 0xff 0 64K" $SCRATCH_MNT/hello | _filter_xfs_io ln $SCRATCH_MNT/hello $SCRATCH_MNT/mydir/hello_2 # Add some subdirectories and new files and links to them. This is to verify # that after fsyncing our top level directory 'mydir', all the subdirectories # and their files/links are registered in the fsync log and exist after the # fsync log is replayed. mkdir -p $SCRATCH_MNT/mydir/x/y/z ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/foo_y_link ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/z/foo_z_link touch $SCRATCH_MNT/mydir/x/y/z/qwerty # Now fsync only our top directory. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/mydir # And fsync now our new file named 'hello', just to verify later that it has # the expected content and that the previous fsync on the directory 'mydir' had # no bad influence on this fsync. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/hello # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # Verify the content of our file 'foo' remains the same as before, 8192 bytes, # all with the value 0xaa. echo "File 'foo' content after log replay:" od -t x1 $SCRATCH_MNT/foo # Remove the first name of our inode. Because of the directory fsync bug, the # inode's link count was 1 instead of 5, so removing the 'foo' name ended up # deleting the inode and the other names became stale directory entries (still # visible to applications). Attempting to remove or access the remaining # dentries pointing to that inode resulted in stale file handle errors and # made it impossible to remove the parent directories since it was impossible # for them to become empty. echo "file 'foo' link count after log replay: $(stat -c %h $SCRATCH_MNT/foo)" rm -f $SCRATCH_MNT/foo # Now verify that all files, links and directories created before fsyncing our # directory exist after the fsync log was replayed. [ -f $SCRATCH_MNT/mydir/foo_2 ] || echo "Link mydir/foo_2 is missing" [ -f $SCRATCH_MNT/mydir/foo_3 ] || echo "Link mydir/foo_3 is missing" [ -f $SCRATCH_MNT/hello ] || echo "File hello is missing" [ -f $SCRATCH_MNT/mydir/hello_2 ] || echo "Link mydir/hello_2 is missing" [ -f $SCRATCH_MNT/mydir/x/y/foo_y_link ] || \ echo "Link mydir/x/y/foo_y_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/foo_z_link ] || \ echo "Link mydir/x/y/z/foo_z_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/qwerty ] || \ echo "File mydir/x/y/z/qwerty is missing" # We expect our file here to have a size of 64Kb and all the bytes having the # value 0xff. echo "file 'hello' content after log replay:" od -t x1 $SCRATCH_MNT/hello # Now remove all files/links, under our test directory 'mydir', and verify we # can remove all the directories. rm -f $SCRATCH_MNT/mydir/x/y/z/* rmdir $SCRATCH_MNT/mydir/x/y/z rm -f $SCRATCH_MNT/mydir/x/y/* rmdir $SCRATCH_MNT/mydir/x/y rmdir $SCRATCH_MNT/mydir/x rm -f $SCRATCH_MNT/mydir/* rmdir $SCRATCH_MNT/mydir # An fsck, run by the fstests framework everytime a test finishes, also detected # the inconsistency and printed the following error message: # # root 5 inode 257 errors 2001, no inode item, link count wrong # unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref # unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref status=0 exit The expected golden output for the test is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 5 file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 Which is the output after this patch and when running the test against ext3/4, xfs, f2fs, reiserfs or nilfs2. Without this patch, the test's output is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 1 Link mydir/foo_2 is missing Link mydir/foo_3 is missing Link mydir/x/y/foo_y_link is missing Link mydir/x/y/z/foo_z_link is missing File mydir/x/y/z/qwerty is missing file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y/z': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x': No such file or directory rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_2': Stale file handle rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_3': Stale file handle rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir': Directory not empty Fsck, without this fix, also complains about the wrong link count: root 5 inode 257 errors 2001, no inode item, link count wrong unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref So fix this by logging the inodes that the dentries point to when fsyncing a directory. A test case for xfstests follows. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-03-20 20:19:46 +03:00
*size_ret = 0;
Btrfs: fix fsync data loss after adding hard link to inode We have a scenario where after the fsync log replay we can lose file data that had been previously fsync'ed if we added an hard link for our inode and after that we sync'ed the fsync log (for example by fsync'ing some other file or directory). This is because when adding an hard link we updated the inode item in the log tree with an i_size value of 0. At that point the new inode item was in memory only and a subsequent fsync log replay would not make us lose the file data. However if after adding the hard link we sync the log tree to disk, by fsync'ing some other file or directory for example, we ended up losing the file data after log replay, because the inode item in the persisted log tree had an an i_size of zero. This is easy to reproduce, and the following excerpt from my test for xfstests shows this: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create one file with data and fsync it. # This made the btrfs fsync log persist the data and the inode metadata with # a correct inode->i_size (4096 bytes). $XFS_IO_PROG -f -c "pwrite -S 0xaa -b 4K 0 4K" -c "fsync" \ $SCRATCH_MNT/foo | _filter_xfs_io # Now add one hard link to our file. This made the btrfs code update the fsync # log, in memory only, with an inode metadata having a size of 0. ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link # Now force persistence of the fsync log to disk, for example, by fsyncing some # other file. touch $SCRATCH_MNT/bar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/bar # Before a power loss or crash, we could read the 4Kb of data from our file as # expected. echo "File content before:" od -t x1 $SCRATCH_MNT/foo # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # After the fsync log replay, because the fsync log had a value of 0 for our # inode's i_size, we couldn't read anymore the 4Kb of data that we previously # wrote and fsync'ed. The size of the file became 0 after the fsync log replay. echo "File content after:" od -t x1 $SCRATCH_MNT/foo Another alternative test, that doesn't need to fsync an inode in the same transaction it was created, is: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create our test file with some data. $XFS_IO_PROG -f -c "pwrite -S 0xaa -b 8K 0 8K" \ $SCRATCH_MNT/foo | _filter_xfs_io # Make sure the file is durably persisted. sync # Append some data to our file, to increase its size. $XFS_IO_PROG -f -c "pwrite -S 0xcc -b 4K 8K 4K" \ $SCRATCH_MNT/foo | _filter_xfs_io # Fsync the file, so from this point on if a crash/power failure happens, our # new data is guaranteed to be there next time the fs is mounted. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foo # Add one hard link to our file. This made btrfs write into the in memory fsync # log a special inode with generation 0 and an i_size of 0 too. Note that this # didn't update the inode in the fsync log on disk. ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link # Now make sure the in memory fsync log is durably persisted. # Creating and fsync'ing another file will do it. touch $SCRATCH_MNT/bar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/bar # As expected, before the crash/power failure, we should be able to read the # 12Kb of file data. echo "File content before:" od -t x1 $SCRATCH_MNT/foo # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # After mounting the fs again, the fsync log was replayed. # The btrfs fsync log replay code didn't update the i_size of the persisted # inode because the inode item in the log had a special generation with a # value of 0 (and it couldn't know the correct i_size, since that inode item # had a 0 i_size too). This made the last 4Kb of file data inaccessible and # effectively lost. echo "File content after:" od -t x1 $SCRATCH_MNT/foo This isn't a new issue/regression. This problem has been around since the log tree code was added in 2008: Btrfs: Add a write ahead tree log to optimize synchronous operations (commit e02119d5a7b4396c5a872582fddc8bd6d305a70a) Test cases for xfstests follow soon. CC: <stable@vger.kernel.org> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-02-13 15:30:56 +03:00
} else {
struct btrfs_inode_item *item;
item = btrfs_item_ptr(path->nodes[0], path->slots[0],
struct btrfs_inode_item);
*size_ret = btrfs_inode_size(path->nodes[0], item);
Btrfs: fix incorrect file size after shrinking truncate and fsync If we do a shrinking truncate against an inode which is already present in the respective log tree and then rename it, as part of logging the new name we end up logging an inode item that reflects the old size of the file (the one which we previously logged) and not the new smaller size. The decision to preserve the size previously logged was added by commit 1a4bcf470c886b ("Btrfs: fix fsync data loss after adding hard link to inode") in order to avoid data loss after replaying the log. However that decision is only needed for the case the logged inode size is smaller then the current size of the inode, as explained in that commit's change log. If the current size of the inode is smaller then the previously logged size, we know a shrinking truncate happened and therefore need to use that smaller size. Example to trigger the problem: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ xfs_io -f -c "pwrite -S 0xab 0 8000" /mnt/foo $ xfs_io -c "fsync" /mnt/foo $ xfs_io -c "truncate 3000" /mnt/foo $ mv /mnt/foo /mnt/bar $ xfs_io -c "fsync" /mnt/bar <power failure> $ mount /dev/sdb /mnt $ od -t x1 -A d /mnt/bar 0000000 ab ab ab ab ab ab ab ab ab ab ab ab ab ab ab ab * 0008000 Once we rename the file, we log its name (and inode item), and because the inode was already logged before in the current transaction, we log it with a size of 8000 bytes because that is the size we previously logged (with the first fsync). As part of the rename, besides logging the inode, we do also sync the log, which is done since commit d4682ba03ef618 ("Btrfs: sync log after logging new name"), so the next fsync against our inode is effectively a no-op, since no new changes happened since the rename operation. Even if did not sync the log during the rename operation, the same problem (fize size of 8000 bytes instead of 3000 bytes) would be visible after replaying the log if the log ended up getting synced to disk through some other means, such as for example by fsyncing some other modified file. In the example above the fsync after the rename operation is there just because not every filesystem may guarantee logging/journalling the inode (and syncing the log/journal) during the rename operation, for example it is needed for f2fs, but not for ext4 and xfs. Fix this scenario by, when logging a new name (which is triggered by rename and link operations), using the current size of the inode instead of the previously logged inode size. A test case for fstests follows soon. Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=202695 CC: stable@vger.kernel.org # 4.4+ Reported-by: Seulbae Kim <seulbae@gatech.edu> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-03-04 17:06:12 +03:00
/*
* If the in-memory inode's i_size is smaller then the inode
* size stored in the btree, return the inode's i_size, so
* that we get a correct inode size after replaying the log
* when before a power failure we had a shrinking truncate
* followed by addition of a new name (rename / new hard link).
* Otherwise return the inode size from the btree, to avoid
* data loss when replaying a log due to previously doing a
* write that expands the inode's size and logging a new name
* immediately after.
*/
if (*size_ret > inode->vfs_inode.i_size)
*size_ret = inode->vfs_inode.i_size;
Btrfs: fix fsync data loss after adding hard link to inode We have a scenario where after the fsync log replay we can lose file data that had been previously fsync'ed if we added an hard link for our inode and after that we sync'ed the fsync log (for example by fsync'ing some other file or directory). This is because when adding an hard link we updated the inode item in the log tree with an i_size value of 0. At that point the new inode item was in memory only and a subsequent fsync log replay would not make us lose the file data. However if after adding the hard link we sync the log tree to disk, by fsync'ing some other file or directory for example, we ended up losing the file data after log replay, because the inode item in the persisted log tree had an an i_size of zero. This is easy to reproduce, and the following excerpt from my test for xfstests shows this: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create one file with data and fsync it. # This made the btrfs fsync log persist the data and the inode metadata with # a correct inode->i_size (4096 bytes). $XFS_IO_PROG -f -c "pwrite -S 0xaa -b 4K 0 4K" -c "fsync" \ $SCRATCH_MNT/foo | _filter_xfs_io # Now add one hard link to our file. This made the btrfs code update the fsync # log, in memory only, with an inode metadata having a size of 0. ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link # Now force persistence of the fsync log to disk, for example, by fsyncing some # other file. touch $SCRATCH_MNT/bar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/bar # Before a power loss or crash, we could read the 4Kb of data from our file as # expected. echo "File content before:" od -t x1 $SCRATCH_MNT/foo # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # After the fsync log replay, because the fsync log had a value of 0 for our # inode's i_size, we couldn't read anymore the 4Kb of data that we previously # wrote and fsync'ed. The size of the file became 0 after the fsync log replay. echo "File content after:" od -t x1 $SCRATCH_MNT/foo Another alternative test, that doesn't need to fsync an inode in the same transaction it was created, is: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create our test file with some data. $XFS_IO_PROG -f -c "pwrite -S 0xaa -b 8K 0 8K" \ $SCRATCH_MNT/foo | _filter_xfs_io # Make sure the file is durably persisted. sync # Append some data to our file, to increase its size. $XFS_IO_PROG -f -c "pwrite -S 0xcc -b 4K 8K 4K" \ $SCRATCH_MNT/foo | _filter_xfs_io # Fsync the file, so from this point on if a crash/power failure happens, our # new data is guaranteed to be there next time the fs is mounted. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foo # Add one hard link to our file. This made btrfs write into the in memory fsync # log a special inode with generation 0 and an i_size of 0 too. Note that this # didn't update the inode in the fsync log on disk. ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link # Now make sure the in memory fsync log is durably persisted. # Creating and fsync'ing another file will do it. touch $SCRATCH_MNT/bar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/bar # As expected, before the crash/power failure, we should be able to read the # 12Kb of file data. echo "File content before:" od -t x1 $SCRATCH_MNT/foo # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # After mounting the fs again, the fsync log was replayed. # The btrfs fsync log replay code didn't update the i_size of the persisted # inode because the inode item in the log had a special generation with a # value of 0 (and it couldn't know the correct i_size, since that inode item # had a 0 i_size too). This made the last 4Kb of file data inaccessible and # effectively lost. echo "File content after:" od -t x1 $SCRATCH_MNT/foo This isn't a new issue/regression. This problem has been around since the log tree code was added in 2008: Btrfs: Add a write ahead tree log to optimize synchronous operations (commit e02119d5a7b4396c5a872582fddc8bd6d305a70a) Test cases for xfstests follow soon. CC: <stable@vger.kernel.org> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-02-13 15:30:56 +03:00
}
btrfs_release_path(path);
return 0;
}
Btrfs: fix fsync xattr loss in the fast fsync path After commit 4f764e515361 ("Btrfs: remove deleted xattrs on fsync log replay"), we can end up in a situation where during log replay we end up deleting xattrs that were never deleted when their file was last fsynced. This happens in the fast fsync path (flag BTRFS_INODE_NEEDS_FULL_SYNC is not set in the inode) if the inode has the flag BTRFS_INODE_COPY_EVERYTHING set, the xattr was added in a past transaction and the leaf where the xattr is located was not updated (COWed or created) in the current transaction. In this scenario the xattr item never ends up in the log tree and therefore at log replay time, which makes the replay code delete the xattr from the fs/subvol tree as it thinks that xattr was deleted prior to the last fsync. Fix this by always logging all xattrs, which is the simplest and most reliable way to detect deleted xattrs and replay the deletes at log replay time. This issue is reproducible with the following test case for fstests: seq=`basename $0` seqres=$RESULT_DIR/$seq echo "QA output created by $seq" here=`pwd` tmp=/tmp/$$ status=1 # failure is the default! _cleanup() { _cleanup_flakey rm -f $tmp.* } trap "_cleanup; exit \$status" 0 1 2 3 15 # get standard environment, filters and checks . ./common/rc . ./common/filter . ./common/dmflakey . ./common/attr # real QA test starts here # We create a lot of xattrs for a single file. Only btrfs and xfs are currently # able to store such a large mount of xattrs per file, other filesystems such # as ext3/4 and f2fs for example, fail with ENOSPC even if we attempt to add # less than 1000 xattrs with very small values. _supported_fs btrfs xfs _supported_os Linux _need_to_be_root _require_scratch _require_dm_flakey _require_attrs _require_metadata_journaling $SCRATCH_DEV rm -f $seqres.full _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create the test file with some initial data and make sure everything is # durably persisted. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 32k" $SCRATCH_MNT/foo | _filter_xfs_io sync # Add many small xattrs to our file. # We create such a large amount because it's needed to trigger the issue found # in btrfs - we need to have an amount that causes the fs to have at least 3 # btree leafs with xattrs stored in them, and it must work on any leaf size # (maximum leaf/node size is 64Kb). num_xattrs=2000 for ((i = 1; i <= $num_xattrs; i++)); do name="user.attr_$(printf "%04d" $i)" $SETFATTR_PROG -n $name -v "val_$(printf "%04d" $i)" $SCRATCH_MNT/foo done # Sync the filesystem to force a commit of the current btrfs transaction, this # is a necessary condition to trigger the bug on btrfs. sync # Now update our file's data and fsync the file. # After a successful fsync, if the fsync log/journal is replayed we expect to # see all the xattrs we added before with the same values (and the updated file # data of course). Btrfs used to delete some of these xattrs when it replayed # its fsync log/journal. $XFS_IO_PROG -c "pwrite -S 0xbb 8K 16K" \ -c "fsync" \ $SCRATCH_MNT/foo | _filter_xfs_io # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again and mount. This makes the fs replay its fsync log. _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey echo "File content after crash and log replay:" od -t x1 $SCRATCH_MNT/foo echo "File xattrs after crash and log replay:" for ((i = 1; i <= $num_xattrs; i++)); do name="user.attr_$(printf "%04d" $i)" echo -n "$name=" $GETFATTR_PROG --absolute-names -n $name --only-values $SCRATCH_MNT/foo echo done status=0 exit The golden output expects all xattrs to be available, and with the correct values, after the fsync log is replayed. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-06-20 02:44:51 +03:00
/*
* At the moment we always log all xattrs. This is to figure out at log replay
* time which xattrs must have their deletion replayed. If a xattr is missing
* in the log tree and exists in the fs/subvol tree, we delete it. This is
* because if a xattr is deleted, the inode is fsynced and a power failure
* happens, causing the log to be replayed the next time the fs is mounted,
* we want the xattr to not exist anymore (same behaviour as other filesystems
* with a journal, ext3/4, xfs, f2fs, etc).
*/
static int btrfs_log_all_xattrs(struct btrfs_trans_handle *trans,
struct btrfs_inode *inode,
Btrfs: fix fsync xattr loss in the fast fsync path After commit 4f764e515361 ("Btrfs: remove deleted xattrs on fsync log replay"), we can end up in a situation where during log replay we end up deleting xattrs that were never deleted when their file was last fsynced. This happens in the fast fsync path (flag BTRFS_INODE_NEEDS_FULL_SYNC is not set in the inode) if the inode has the flag BTRFS_INODE_COPY_EVERYTHING set, the xattr was added in a past transaction and the leaf where the xattr is located was not updated (COWed or created) in the current transaction. In this scenario the xattr item never ends up in the log tree and therefore at log replay time, which makes the replay code delete the xattr from the fs/subvol tree as it thinks that xattr was deleted prior to the last fsync. Fix this by always logging all xattrs, which is the simplest and most reliable way to detect deleted xattrs and replay the deletes at log replay time. This issue is reproducible with the following test case for fstests: seq=`basename $0` seqres=$RESULT_DIR/$seq echo "QA output created by $seq" here=`pwd` tmp=/tmp/$$ status=1 # failure is the default! _cleanup() { _cleanup_flakey rm -f $tmp.* } trap "_cleanup; exit \$status" 0 1 2 3 15 # get standard environment, filters and checks . ./common/rc . ./common/filter . ./common/dmflakey . ./common/attr # real QA test starts here # We create a lot of xattrs for a single file. Only btrfs and xfs are currently # able to store such a large mount of xattrs per file, other filesystems such # as ext3/4 and f2fs for example, fail with ENOSPC even if we attempt to add # less than 1000 xattrs with very small values. _supported_fs btrfs xfs _supported_os Linux _need_to_be_root _require_scratch _require_dm_flakey _require_attrs _require_metadata_journaling $SCRATCH_DEV rm -f $seqres.full _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create the test file with some initial data and make sure everything is # durably persisted. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 32k" $SCRATCH_MNT/foo | _filter_xfs_io sync # Add many small xattrs to our file. # We create such a large amount because it's needed to trigger the issue found # in btrfs - we need to have an amount that causes the fs to have at least 3 # btree leafs with xattrs stored in them, and it must work on any leaf size # (maximum leaf/node size is 64Kb). num_xattrs=2000 for ((i = 1; i <= $num_xattrs; i++)); do name="user.attr_$(printf "%04d" $i)" $SETFATTR_PROG -n $name -v "val_$(printf "%04d" $i)" $SCRATCH_MNT/foo done # Sync the filesystem to force a commit of the current btrfs transaction, this # is a necessary condition to trigger the bug on btrfs. sync # Now update our file's data and fsync the file. # After a successful fsync, if the fsync log/journal is replayed we expect to # see all the xattrs we added before with the same values (and the updated file # data of course). Btrfs used to delete some of these xattrs when it replayed # its fsync log/journal. $XFS_IO_PROG -c "pwrite -S 0xbb 8K 16K" \ -c "fsync" \ $SCRATCH_MNT/foo | _filter_xfs_io # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again and mount. This makes the fs replay its fsync log. _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey echo "File content after crash and log replay:" od -t x1 $SCRATCH_MNT/foo echo "File xattrs after crash and log replay:" for ((i = 1; i <= $num_xattrs; i++)); do name="user.attr_$(printf "%04d" $i)" echo -n "$name=" $GETFATTR_PROG --absolute-names -n $name --only-values $SCRATCH_MNT/foo echo done status=0 exit The golden output expects all xattrs to be available, and with the correct values, after the fsync log is replayed. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-06-20 02:44:51 +03:00
struct btrfs_path *path,
struct btrfs_path *dst_path)
{
struct btrfs_root *root = inode->root;
Btrfs: fix fsync xattr loss in the fast fsync path After commit 4f764e515361 ("Btrfs: remove deleted xattrs on fsync log replay"), we can end up in a situation where during log replay we end up deleting xattrs that were never deleted when their file was last fsynced. This happens in the fast fsync path (flag BTRFS_INODE_NEEDS_FULL_SYNC is not set in the inode) if the inode has the flag BTRFS_INODE_COPY_EVERYTHING set, the xattr was added in a past transaction and the leaf where the xattr is located was not updated (COWed or created) in the current transaction. In this scenario the xattr item never ends up in the log tree and therefore at log replay time, which makes the replay code delete the xattr from the fs/subvol tree as it thinks that xattr was deleted prior to the last fsync. Fix this by always logging all xattrs, which is the simplest and most reliable way to detect deleted xattrs and replay the deletes at log replay time. This issue is reproducible with the following test case for fstests: seq=`basename $0` seqres=$RESULT_DIR/$seq echo "QA output created by $seq" here=`pwd` tmp=/tmp/$$ status=1 # failure is the default! _cleanup() { _cleanup_flakey rm -f $tmp.* } trap "_cleanup; exit \$status" 0 1 2 3 15 # get standard environment, filters and checks . ./common/rc . ./common/filter . ./common/dmflakey . ./common/attr # real QA test starts here # We create a lot of xattrs for a single file. Only btrfs and xfs are currently # able to store such a large mount of xattrs per file, other filesystems such # as ext3/4 and f2fs for example, fail with ENOSPC even if we attempt to add # less than 1000 xattrs with very small values. _supported_fs btrfs xfs _supported_os Linux _need_to_be_root _require_scratch _require_dm_flakey _require_attrs _require_metadata_journaling $SCRATCH_DEV rm -f $seqres.full _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create the test file with some initial data and make sure everything is # durably persisted. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 32k" $SCRATCH_MNT/foo | _filter_xfs_io sync # Add many small xattrs to our file. # We create such a large amount because it's needed to trigger the issue found # in btrfs - we need to have an amount that causes the fs to have at least 3 # btree leafs with xattrs stored in them, and it must work on any leaf size # (maximum leaf/node size is 64Kb). num_xattrs=2000 for ((i = 1; i <= $num_xattrs; i++)); do name="user.attr_$(printf "%04d" $i)" $SETFATTR_PROG -n $name -v "val_$(printf "%04d" $i)" $SCRATCH_MNT/foo done # Sync the filesystem to force a commit of the current btrfs transaction, this # is a necessary condition to trigger the bug on btrfs. sync # Now update our file's data and fsync the file. # After a successful fsync, if the fsync log/journal is replayed we expect to # see all the xattrs we added before with the same values (and the updated file # data of course). Btrfs used to delete some of these xattrs when it replayed # its fsync log/journal. $XFS_IO_PROG -c "pwrite -S 0xbb 8K 16K" \ -c "fsync" \ $SCRATCH_MNT/foo | _filter_xfs_io # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again and mount. This makes the fs replay its fsync log. _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey echo "File content after crash and log replay:" od -t x1 $SCRATCH_MNT/foo echo "File xattrs after crash and log replay:" for ((i = 1; i <= $num_xattrs; i++)); do name="user.attr_$(printf "%04d" $i)" echo -n "$name=" $GETFATTR_PROG --absolute-names -n $name --only-values $SCRATCH_MNT/foo echo done status=0 exit The golden output expects all xattrs to be available, and with the correct values, after the fsync log is replayed. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-06-20 02:44:51 +03:00
int ret;
struct btrfs_key key;
const u64 ino = btrfs_ino(inode);
Btrfs: fix fsync xattr loss in the fast fsync path After commit 4f764e515361 ("Btrfs: remove deleted xattrs on fsync log replay"), we can end up in a situation where during log replay we end up deleting xattrs that were never deleted when their file was last fsynced. This happens in the fast fsync path (flag BTRFS_INODE_NEEDS_FULL_SYNC is not set in the inode) if the inode has the flag BTRFS_INODE_COPY_EVERYTHING set, the xattr was added in a past transaction and the leaf where the xattr is located was not updated (COWed or created) in the current transaction. In this scenario the xattr item never ends up in the log tree and therefore at log replay time, which makes the replay code delete the xattr from the fs/subvol tree as it thinks that xattr was deleted prior to the last fsync. Fix this by always logging all xattrs, which is the simplest and most reliable way to detect deleted xattrs and replay the deletes at log replay time. This issue is reproducible with the following test case for fstests: seq=`basename $0` seqres=$RESULT_DIR/$seq echo "QA output created by $seq" here=`pwd` tmp=/tmp/$$ status=1 # failure is the default! _cleanup() { _cleanup_flakey rm -f $tmp.* } trap "_cleanup; exit \$status" 0 1 2 3 15 # get standard environment, filters and checks . ./common/rc . ./common/filter . ./common/dmflakey . ./common/attr # real QA test starts here # We create a lot of xattrs for a single file. Only btrfs and xfs are currently # able to store such a large mount of xattrs per file, other filesystems such # as ext3/4 and f2fs for example, fail with ENOSPC even if we attempt to add # less than 1000 xattrs with very small values. _supported_fs btrfs xfs _supported_os Linux _need_to_be_root _require_scratch _require_dm_flakey _require_attrs _require_metadata_journaling $SCRATCH_DEV rm -f $seqres.full _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create the test file with some initial data and make sure everything is # durably persisted. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 32k" $SCRATCH_MNT/foo | _filter_xfs_io sync # Add many small xattrs to our file. # We create such a large amount because it's needed to trigger the issue found # in btrfs - we need to have an amount that causes the fs to have at least 3 # btree leafs with xattrs stored in them, and it must work on any leaf size # (maximum leaf/node size is 64Kb). num_xattrs=2000 for ((i = 1; i <= $num_xattrs; i++)); do name="user.attr_$(printf "%04d" $i)" $SETFATTR_PROG -n $name -v "val_$(printf "%04d" $i)" $SCRATCH_MNT/foo done # Sync the filesystem to force a commit of the current btrfs transaction, this # is a necessary condition to trigger the bug on btrfs. sync # Now update our file's data and fsync the file. # After a successful fsync, if the fsync log/journal is replayed we expect to # see all the xattrs we added before with the same values (and the updated file # data of course). Btrfs used to delete some of these xattrs when it replayed # its fsync log/journal. $XFS_IO_PROG -c "pwrite -S 0xbb 8K 16K" \ -c "fsync" \ $SCRATCH_MNT/foo | _filter_xfs_io # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again and mount. This makes the fs replay its fsync log. _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey echo "File content after crash and log replay:" od -t x1 $SCRATCH_MNT/foo echo "File xattrs after crash and log replay:" for ((i = 1; i <= $num_xattrs; i++)); do name="user.attr_$(printf "%04d" $i)" echo -n "$name=" $GETFATTR_PROG --absolute-names -n $name --only-values $SCRATCH_MNT/foo echo done status=0 exit The golden output expects all xattrs to be available, and with the correct values, after the fsync log is replayed. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-06-20 02:44:51 +03:00
int ins_nr = 0;
int start_slot = 0;
btrfs: skip unnecessary searches for xattrs when logging an inode Every time we log an inode we lookup in the fs/subvol tree for xattrs and if we have any, log them into the log tree. However it is very common to have inodes without any xattrs, so doing the search wastes times, but more importantly it adds contention on the fs/subvol tree locks, either making the logging code block and wait for tree locks or making the logging code making other concurrent operations block and wait. The most typical use cases where xattrs are used are when capabilities or ACLs are defined for an inode, or when SELinux is enabled. This change makes the logging code detect when an inode does not have xattrs and skip the xattrs search the next time the inode is logged, unless the inode is evicted and loaded again or a xattr is added to the inode. Therefore skipping the search for xattrs on inodes that don't ever have xattrs and are fsynced with some frequency. The following script that calls dbench was used to measure the impact of this change on a VM with 8 CPUs, 16Gb of ram, using a raw NVMe device directly (no intermediary filesystem on the host) and using a non-debug kernel (default configuration on Debian distributions): $ cat test.sh #!/bin/bash DEV=/dev/sdk MNT=/mnt/sdk MOUNT_OPTIONS="-o ssd" mkfs.btrfs -f -m single -d single $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -D $MNT -t 200 40 umount $MNT The results before this change: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 5761605 0.172 312.057 Close 4232452 0.002 10.927 Rename 243937 1.406 277.344 Unlink 1163456 0.631 298.402 Deltree 160 11.581 221.107 Mkdir 80 0.003 0.005 Qpathinfo 5221410 0.065 122.309 Qfileinfo 915432 0.001 3.333 Qfsinfo 957555 0.003 3.992 Sfileinfo 469244 0.023 20.494 Find 2018865 0.448 123.659 WriteX 2874851 0.049 118.529 ReadX 9030579 0.004 21.654 LockX 18754 0.003 4.423 UnlockX 18754 0.002 0.331 Flush 403792 10.944 359.494 Throughput 908.444 MB/sec 40 clients 40 procs max_latency=359.500 ms The results after this change: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 6442521 0.159 230.693 Close 4732357 0.002 10.972 Rename 272809 1.293 227.398 Unlink 1301059 0.563 218.500 Deltree 160 7.796 54.887 Mkdir 80 0.008 0.478 Qpathinfo 5839452 0.047 124.330 Qfileinfo 1023199 0.001 4.996 Qfsinfo 1070760 0.003 5.709 Sfileinfo 524790 0.033 21.765 Find 2257658 0.314 125.611 WriteX 3211520 0.040 232.135 ReadX 10098969 0.004 25.340 LockX 20974 0.003 1.569 UnlockX 20974 0.002 3.475 Flush 451553 10.287 331.037 Throughput 1011.77 MB/sec 40 clients 40 procs max_latency=331.045 ms +10.8% throughput, -8.2% max latency Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-11-13 14:21:49 +03:00
bool found_xattrs = false;
if (test_bit(BTRFS_INODE_NO_XATTRS, &inode->runtime_flags))
return 0;
Btrfs: fix fsync xattr loss in the fast fsync path After commit 4f764e515361 ("Btrfs: remove deleted xattrs on fsync log replay"), we can end up in a situation where during log replay we end up deleting xattrs that were never deleted when their file was last fsynced. This happens in the fast fsync path (flag BTRFS_INODE_NEEDS_FULL_SYNC is not set in the inode) if the inode has the flag BTRFS_INODE_COPY_EVERYTHING set, the xattr was added in a past transaction and the leaf where the xattr is located was not updated (COWed or created) in the current transaction. In this scenario the xattr item never ends up in the log tree and therefore at log replay time, which makes the replay code delete the xattr from the fs/subvol tree as it thinks that xattr was deleted prior to the last fsync. Fix this by always logging all xattrs, which is the simplest and most reliable way to detect deleted xattrs and replay the deletes at log replay time. This issue is reproducible with the following test case for fstests: seq=`basename $0` seqres=$RESULT_DIR/$seq echo "QA output created by $seq" here=`pwd` tmp=/tmp/$$ status=1 # failure is the default! _cleanup() { _cleanup_flakey rm -f $tmp.* } trap "_cleanup; exit \$status" 0 1 2 3 15 # get standard environment, filters and checks . ./common/rc . ./common/filter . ./common/dmflakey . ./common/attr # real QA test starts here # We create a lot of xattrs for a single file. Only btrfs and xfs are currently # able to store such a large mount of xattrs per file, other filesystems such # as ext3/4 and f2fs for example, fail with ENOSPC even if we attempt to add # less than 1000 xattrs with very small values. _supported_fs btrfs xfs _supported_os Linux _need_to_be_root _require_scratch _require_dm_flakey _require_attrs _require_metadata_journaling $SCRATCH_DEV rm -f $seqres.full _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create the test file with some initial data and make sure everything is # durably persisted. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 32k" $SCRATCH_MNT/foo | _filter_xfs_io sync # Add many small xattrs to our file. # We create such a large amount because it's needed to trigger the issue found # in btrfs - we need to have an amount that causes the fs to have at least 3 # btree leafs with xattrs stored in them, and it must work on any leaf size # (maximum leaf/node size is 64Kb). num_xattrs=2000 for ((i = 1; i <= $num_xattrs; i++)); do name="user.attr_$(printf "%04d" $i)" $SETFATTR_PROG -n $name -v "val_$(printf "%04d" $i)" $SCRATCH_MNT/foo done # Sync the filesystem to force a commit of the current btrfs transaction, this # is a necessary condition to trigger the bug on btrfs. sync # Now update our file's data and fsync the file. # After a successful fsync, if the fsync log/journal is replayed we expect to # see all the xattrs we added before with the same values (and the updated file # data of course). Btrfs used to delete some of these xattrs when it replayed # its fsync log/journal. $XFS_IO_PROG -c "pwrite -S 0xbb 8K 16K" \ -c "fsync" \ $SCRATCH_MNT/foo | _filter_xfs_io # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again and mount. This makes the fs replay its fsync log. _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey echo "File content after crash and log replay:" od -t x1 $SCRATCH_MNT/foo echo "File xattrs after crash and log replay:" for ((i = 1; i <= $num_xattrs; i++)); do name="user.attr_$(printf "%04d" $i)" echo -n "$name=" $GETFATTR_PROG --absolute-names -n $name --only-values $SCRATCH_MNT/foo echo done status=0 exit The golden output expects all xattrs to be available, and with the correct values, after the fsync log is replayed. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-06-20 02:44:51 +03:00
key.objectid = ino;
key.type = BTRFS_XATTR_ITEM_KEY;
key.offset = 0;
ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
if (ret < 0)
return ret;
while (true) {
int slot = path->slots[0];
struct extent_buffer *leaf = path->nodes[0];
int nritems = btrfs_header_nritems(leaf);
if (slot >= nritems) {
if (ins_nr > 0) {
ret = copy_items(trans, inode, dst_path, path,
Btrfs: fix missing hole after hole punching and fsync when using NO_HOLES When using the NO_HOLES feature, if we punch a hole into a file and then fsync it, there are cases where a subsequent fsync will miss the fact that a hole was punched, resulting in the holes not existing after replaying the log tree. Essentially these cases all imply that, tree-log.c:copy_items(), is not invoked for the leafs that delimit holes, because nothing changed those leafs in the current transaction. And it's precisely copy_items() where we currenly detect and log holes, which works as long as the holes are between file extent items in the input leaf or between the beginning of input leaf and the previous leaf or between the last item in the leaf and the next leaf. First example where we miss a hole: *) The extent items of the inode span multiple leafs; *) The punched hole covers a range that affects only the extent items of the first leaf; *) The fsync operation is done in full mode (BTRFS_INODE_NEEDS_FULL_SYNC is set in the inode's runtime flags). That results in the hole not existing after replaying the log tree. For example, if the fs/subvolume tree has the following layout for a particular inode: Leaf N, generation 10: [ ... INODE_ITEM INODE_REF EXTENT_ITEM (0 64K) EXTENT_ITEM (64K 128K) ] Leaf N + 1, generation 10: [ EXTENT_ITEM (128K 64K) ... ] If at transaction 11 we punch a hole coverting the range [0, 128K[, we end up dropping the two extent items from leaf N, but we don't touch the other leaf, so we end up in the following state: Leaf N, generation 11: [ ... INODE_ITEM INODE_REF ] Leaf N + 1, generation 10: [ EXTENT_ITEM (128K 64K) ... ] A full fsync after punching the hole will only process leaf N because it was modified in the current transaction, but not leaf N + 1, since it was not modified in the current transaction (generation 10 and not 11). As a result the fsync will not log any holes, because it didn't process any leaf with extent items. Second example where we will miss a hole: *) An inode as its items spanning 5 (or more) leafs; *) A hole is punched and it covers only the extents items of the 3rd leaf. This resulsts in deleting the entire leaf and not touching any of the other leafs. So the only leaf that is modified in the current transaction, when punching the hole, is the first leaf, which contains the inode item. During the full fsync, the only leaf that is passed to copy_items() is that first leaf, and that's not enough for the hole detection code in copy_items() to determine there's a hole between the last file extent item in the 2nd leaf and the first file extent item in the 3rd leaf (which was the 4th leaf before punching the hole). Fix this by scanning all leafs and punch holes as necessary when doing a full fsync (less common than a non-full fsync) when the NO_HOLES feature is enabled. The lack of explicit file extent items to mark holes makes it necessary to scan existing extents to determine if holes exist. A test case for fstests follows soon. Fixes: 16e7549f045d33 ("Btrfs: incompatible format change to remove hole extents") CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-11-19 15:07:33 +03:00
start_slot, ins_nr, 1, 0);
Btrfs: fix fsync xattr loss in the fast fsync path After commit 4f764e515361 ("Btrfs: remove deleted xattrs on fsync log replay"), we can end up in a situation where during log replay we end up deleting xattrs that were never deleted when their file was last fsynced. This happens in the fast fsync path (flag BTRFS_INODE_NEEDS_FULL_SYNC is not set in the inode) if the inode has the flag BTRFS_INODE_COPY_EVERYTHING set, the xattr was added in a past transaction and the leaf where the xattr is located was not updated (COWed or created) in the current transaction. In this scenario the xattr item never ends up in the log tree and therefore at log replay time, which makes the replay code delete the xattr from the fs/subvol tree as it thinks that xattr was deleted prior to the last fsync. Fix this by always logging all xattrs, which is the simplest and most reliable way to detect deleted xattrs and replay the deletes at log replay time. This issue is reproducible with the following test case for fstests: seq=`basename $0` seqres=$RESULT_DIR/$seq echo "QA output created by $seq" here=`pwd` tmp=/tmp/$$ status=1 # failure is the default! _cleanup() { _cleanup_flakey rm -f $tmp.* } trap "_cleanup; exit \$status" 0 1 2 3 15 # get standard environment, filters and checks . ./common/rc . ./common/filter . ./common/dmflakey . ./common/attr # real QA test starts here # We create a lot of xattrs for a single file. Only btrfs and xfs are currently # able to store such a large mount of xattrs per file, other filesystems such # as ext3/4 and f2fs for example, fail with ENOSPC even if we attempt to add # less than 1000 xattrs with very small values. _supported_fs btrfs xfs _supported_os Linux _need_to_be_root _require_scratch _require_dm_flakey _require_attrs _require_metadata_journaling $SCRATCH_DEV rm -f $seqres.full _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create the test file with some initial data and make sure everything is # durably persisted. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 32k" $SCRATCH_MNT/foo | _filter_xfs_io sync # Add many small xattrs to our file. # We create such a large amount because it's needed to trigger the issue found # in btrfs - we need to have an amount that causes the fs to have at least 3 # btree leafs with xattrs stored in them, and it must work on any leaf size # (maximum leaf/node size is 64Kb). num_xattrs=2000 for ((i = 1; i <= $num_xattrs; i++)); do name="user.attr_$(printf "%04d" $i)" $SETFATTR_PROG -n $name -v "val_$(printf "%04d" $i)" $SCRATCH_MNT/foo done # Sync the filesystem to force a commit of the current btrfs transaction, this # is a necessary condition to trigger the bug on btrfs. sync # Now update our file's data and fsync the file. # After a successful fsync, if the fsync log/journal is replayed we expect to # see all the xattrs we added before with the same values (and the updated file # data of course). Btrfs used to delete some of these xattrs when it replayed # its fsync log/journal. $XFS_IO_PROG -c "pwrite -S 0xbb 8K 16K" \ -c "fsync" \ $SCRATCH_MNT/foo | _filter_xfs_io # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again and mount. This makes the fs replay its fsync log. _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey echo "File content after crash and log replay:" od -t x1 $SCRATCH_MNT/foo echo "File xattrs after crash and log replay:" for ((i = 1; i <= $num_xattrs; i++)); do name="user.attr_$(printf "%04d" $i)" echo -n "$name=" $GETFATTR_PROG --absolute-names -n $name --only-values $SCRATCH_MNT/foo echo done status=0 exit The golden output expects all xattrs to be available, and with the correct values, after the fsync log is replayed. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-06-20 02:44:51 +03:00
if (ret < 0)
return ret;
ins_nr = 0;
}
ret = btrfs_next_leaf(root, path);
if (ret < 0)
return ret;
else if (ret > 0)
break;
continue;
}
btrfs_item_key_to_cpu(leaf, &key, slot);
if (key.objectid != ino || key.type != BTRFS_XATTR_ITEM_KEY)
break;
if (ins_nr == 0)
start_slot = slot;
ins_nr++;
path->slots[0]++;
btrfs: skip unnecessary searches for xattrs when logging an inode Every time we log an inode we lookup in the fs/subvol tree for xattrs and if we have any, log them into the log tree. However it is very common to have inodes without any xattrs, so doing the search wastes times, but more importantly it adds contention on the fs/subvol tree locks, either making the logging code block and wait for tree locks or making the logging code making other concurrent operations block and wait. The most typical use cases where xattrs are used are when capabilities or ACLs are defined for an inode, or when SELinux is enabled. This change makes the logging code detect when an inode does not have xattrs and skip the xattrs search the next time the inode is logged, unless the inode is evicted and loaded again or a xattr is added to the inode. Therefore skipping the search for xattrs on inodes that don't ever have xattrs and are fsynced with some frequency. The following script that calls dbench was used to measure the impact of this change on a VM with 8 CPUs, 16Gb of ram, using a raw NVMe device directly (no intermediary filesystem on the host) and using a non-debug kernel (default configuration on Debian distributions): $ cat test.sh #!/bin/bash DEV=/dev/sdk MNT=/mnt/sdk MOUNT_OPTIONS="-o ssd" mkfs.btrfs -f -m single -d single $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -D $MNT -t 200 40 umount $MNT The results before this change: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 5761605 0.172 312.057 Close 4232452 0.002 10.927 Rename 243937 1.406 277.344 Unlink 1163456 0.631 298.402 Deltree 160 11.581 221.107 Mkdir 80 0.003 0.005 Qpathinfo 5221410 0.065 122.309 Qfileinfo 915432 0.001 3.333 Qfsinfo 957555 0.003 3.992 Sfileinfo 469244 0.023 20.494 Find 2018865 0.448 123.659 WriteX 2874851 0.049 118.529 ReadX 9030579 0.004 21.654 LockX 18754 0.003 4.423 UnlockX 18754 0.002 0.331 Flush 403792 10.944 359.494 Throughput 908.444 MB/sec 40 clients 40 procs max_latency=359.500 ms The results after this change: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 6442521 0.159 230.693 Close 4732357 0.002 10.972 Rename 272809 1.293 227.398 Unlink 1301059 0.563 218.500 Deltree 160 7.796 54.887 Mkdir 80 0.008 0.478 Qpathinfo 5839452 0.047 124.330 Qfileinfo 1023199 0.001 4.996 Qfsinfo 1070760 0.003 5.709 Sfileinfo 524790 0.033 21.765 Find 2257658 0.314 125.611 WriteX 3211520 0.040 232.135 ReadX 10098969 0.004 25.340 LockX 20974 0.003 1.569 UnlockX 20974 0.002 3.475 Flush 451553 10.287 331.037 Throughput 1011.77 MB/sec 40 clients 40 procs max_latency=331.045 ms +10.8% throughput, -8.2% max latency Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-11-13 14:21:49 +03:00
found_xattrs = true;
Btrfs: fix fsync xattr loss in the fast fsync path After commit 4f764e515361 ("Btrfs: remove deleted xattrs on fsync log replay"), we can end up in a situation where during log replay we end up deleting xattrs that were never deleted when their file was last fsynced. This happens in the fast fsync path (flag BTRFS_INODE_NEEDS_FULL_SYNC is not set in the inode) if the inode has the flag BTRFS_INODE_COPY_EVERYTHING set, the xattr was added in a past transaction and the leaf where the xattr is located was not updated (COWed or created) in the current transaction. In this scenario the xattr item never ends up in the log tree and therefore at log replay time, which makes the replay code delete the xattr from the fs/subvol tree as it thinks that xattr was deleted prior to the last fsync. Fix this by always logging all xattrs, which is the simplest and most reliable way to detect deleted xattrs and replay the deletes at log replay time. This issue is reproducible with the following test case for fstests: seq=`basename $0` seqres=$RESULT_DIR/$seq echo "QA output created by $seq" here=`pwd` tmp=/tmp/$$ status=1 # failure is the default! _cleanup() { _cleanup_flakey rm -f $tmp.* } trap "_cleanup; exit \$status" 0 1 2 3 15 # get standard environment, filters and checks . ./common/rc . ./common/filter . ./common/dmflakey . ./common/attr # real QA test starts here # We create a lot of xattrs for a single file. Only btrfs and xfs are currently # able to store such a large mount of xattrs per file, other filesystems such # as ext3/4 and f2fs for example, fail with ENOSPC even if we attempt to add # less than 1000 xattrs with very small values. _supported_fs btrfs xfs _supported_os Linux _need_to_be_root _require_scratch _require_dm_flakey _require_attrs _require_metadata_journaling $SCRATCH_DEV rm -f $seqres.full _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create the test file with some initial data and make sure everything is # durably persisted. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 32k" $SCRATCH_MNT/foo | _filter_xfs_io sync # Add many small xattrs to our file. # We create such a large amount because it's needed to trigger the issue found # in btrfs - we need to have an amount that causes the fs to have at least 3 # btree leafs with xattrs stored in them, and it must work on any leaf size # (maximum leaf/node size is 64Kb). num_xattrs=2000 for ((i = 1; i <= $num_xattrs; i++)); do name="user.attr_$(printf "%04d" $i)" $SETFATTR_PROG -n $name -v "val_$(printf "%04d" $i)" $SCRATCH_MNT/foo done # Sync the filesystem to force a commit of the current btrfs transaction, this # is a necessary condition to trigger the bug on btrfs. sync # Now update our file's data and fsync the file. # After a successful fsync, if the fsync log/journal is replayed we expect to # see all the xattrs we added before with the same values (and the updated file # data of course). Btrfs used to delete some of these xattrs when it replayed # its fsync log/journal. $XFS_IO_PROG -c "pwrite -S 0xbb 8K 16K" \ -c "fsync" \ $SCRATCH_MNT/foo | _filter_xfs_io # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again and mount. This makes the fs replay its fsync log. _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey echo "File content after crash and log replay:" od -t x1 $SCRATCH_MNT/foo echo "File xattrs after crash and log replay:" for ((i = 1; i <= $num_xattrs; i++)); do name="user.attr_$(printf "%04d" $i)" echo -n "$name=" $GETFATTR_PROG --absolute-names -n $name --only-values $SCRATCH_MNT/foo echo done status=0 exit The golden output expects all xattrs to be available, and with the correct values, after the fsync log is replayed. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-06-20 02:44:51 +03:00
cond_resched();
}
if (ins_nr > 0) {
ret = copy_items(trans, inode, dst_path, path,
Btrfs: fix missing hole after hole punching and fsync when using NO_HOLES When using the NO_HOLES feature, if we punch a hole into a file and then fsync it, there are cases where a subsequent fsync will miss the fact that a hole was punched, resulting in the holes not existing after replaying the log tree. Essentially these cases all imply that, tree-log.c:copy_items(), is not invoked for the leafs that delimit holes, because nothing changed those leafs in the current transaction. And it's precisely copy_items() where we currenly detect and log holes, which works as long as the holes are between file extent items in the input leaf or between the beginning of input leaf and the previous leaf or between the last item in the leaf and the next leaf. First example where we miss a hole: *) The extent items of the inode span multiple leafs; *) The punched hole covers a range that affects only the extent items of the first leaf; *) The fsync operation is done in full mode (BTRFS_INODE_NEEDS_FULL_SYNC is set in the inode's runtime flags). That results in the hole not existing after replaying the log tree. For example, if the fs/subvolume tree has the following layout for a particular inode: Leaf N, generation 10: [ ... INODE_ITEM INODE_REF EXTENT_ITEM (0 64K) EXTENT_ITEM (64K 128K) ] Leaf N + 1, generation 10: [ EXTENT_ITEM (128K 64K) ... ] If at transaction 11 we punch a hole coverting the range [0, 128K[, we end up dropping the two extent items from leaf N, but we don't touch the other leaf, so we end up in the following state: Leaf N, generation 11: [ ... INODE_ITEM INODE_REF ] Leaf N + 1, generation 10: [ EXTENT_ITEM (128K 64K) ... ] A full fsync after punching the hole will only process leaf N because it was modified in the current transaction, but not leaf N + 1, since it was not modified in the current transaction (generation 10 and not 11). As a result the fsync will not log any holes, because it didn't process any leaf with extent items. Second example where we will miss a hole: *) An inode as its items spanning 5 (or more) leafs; *) A hole is punched and it covers only the extents items of the 3rd leaf. This resulsts in deleting the entire leaf and not touching any of the other leafs. So the only leaf that is modified in the current transaction, when punching the hole, is the first leaf, which contains the inode item. During the full fsync, the only leaf that is passed to copy_items() is that first leaf, and that's not enough for the hole detection code in copy_items() to determine there's a hole between the last file extent item in the 2nd leaf and the first file extent item in the 3rd leaf (which was the 4th leaf before punching the hole). Fix this by scanning all leafs and punch holes as necessary when doing a full fsync (less common than a non-full fsync) when the NO_HOLES feature is enabled. The lack of explicit file extent items to mark holes makes it necessary to scan existing extents to determine if holes exist. A test case for fstests follows soon. Fixes: 16e7549f045d33 ("Btrfs: incompatible format change to remove hole extents") CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-11-19 15:07:33 +03:00
start_slot, ins_nr, 1, 0);
Btrfs: fix fsync xattr loss in the fast fsync path After commit 4f764e515361 ("Btrfs: remove deleted xattrs on fsync log replay"), we can end up in a situation where during log replay we end up deleting xattrs that were never deleted when their file was last fsynced. This happens in the fast fsync path (flag BTRFS_INODE_NEEDS_FULL_SYNC is not set in the inode) if the inode has the flag BTRFS_INODE_COPY_EVERYTHING set, the xattr was added in a past transaction and the leaf where the xattr is located was not updated (COWed or created) in the current transaction. In this scenario the xattr item never ends up in the log tree and therefore at log replay time, which makes the replay code delete the xattr from the fs/subvol tree as it thinks that xattr was deleted prior to the last fsync. Fix this by always logging all xattrs, which is the simplest and most reliable way to detect deleted xattrs and replay the deletes at log replay time. This issue is reproducible with the following test case for fstests: seq=`basename $0` seqres=$RESULT_DIR/$seq echo "QA output created by $seq" here=`pwd` tmp=/tmp/$$ status=1 # failure is the default! _cleanup() { _cleanup_flakey rm -f $tmp.* } trap "_cleanup; exit \$status" 0 1 2 3 15 # get standard environment, filters and checks . ./common/rc . ./common/filter . ./common/dmflakey . ./common/attr # real QA test starts here # We create a lot of xattrs for a single file. Only btrfs and xfs are currently # able to store such a large mount of xattrs per file, other filesystems such # as ext3/4 and f2fs for example, fail with ENOSPC even if we attempt to add # less than 1000 xattrs with very small values. _supported_fs btrfs xfs _supported_os Linux _need_to_be_root _require_scratch _require_dm_flakey _require_attrs _require_metadata_journaling $SCRATCH_DEV rm -f $seqres.full _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create the test file with some initial data and make sure everything is # durably persisted. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 32k" $SCRATCH_MNT/foo | _filter_xfs_io sync # Add many small xattrs to our file. # We create such a large amount because it's needed to trigger the issue found # in btrfs - we need to have an amount that causes the fs to have at least 3 # btree leafs with xattrs stored in them, and it must work on any leaf size # (maximum leaf/node size is 64Kb). num_xattrs=2000 for ((i = 1; i <= $num_xattrs; i++)); do name="user.attr_$(printf "%04d" $i)" $SETFATTR_PROG -n $name -v "val_$(printf "%04d" $i)" $SCRATCH_MNT/foo done # Sync the filesystem to force a commit of the current btrfs transaction, this # is a necessary condition to trigger the bug on btrfs. sync # Now update our file's data and fsync the file. # After a successful fsync, if the fsync log/journal is replayed we expect to # see all the xattrs we added before with the same values (and the updated file # data of course). Btrfs used to delete some of these xattrs when it replayed # its fsync log/journal. $XFS_IO_PROG -c "pwrite -S 0xbb 8K 16K" \ -c "fsync" \ $SCRATCH_MNT/foo | _filter_xfs_io # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again and mount. This makes the fs replay its fsync log. _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey echo "File content after crash and log replay:" od -t x1 $SCRATCH_MNT/foo echo "File xattrs after crash and log replay:" for ((i = 1; i <= $num_xattrs; i++)); do name="user.attr_$(printf "%04d" $i)" echo -n "$name=" $GETFATTR_PROG --absolute-names -n $name --only-values $SCRATCH_MNT/foo echo done status=0 exit The golden output expects all xattrs to be available, and with the correct values, after the fsync log is replayed. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-06-20 02:44:51 +03:00
if (ret < 0)
return ret;
}
btrfs: skip unnecessary searches for xattrs when logging an inode Every time we log an inode we lookup in the fs/subvol tree for xattrs and if we have any, log them into the log tree. However it is very common to have inodes without any xattrs, so doing the search wastes times, but more importantly it adds contention on the fs/subvol tree locks, either making the logging code block and wait for tree locks or making the logging code making other concurrent operations block and wait. The most typical use cases where xattrs are used are when capabilities or ACLs are defined for an inode, or when SELinux is enabled. This change makes the logging code detect when an inode does not have xattrs and skip the xattrs search the next time the inode is logged, unless the inode is evicted and loaded again or a xattr is added to the inode. Therefore skipping the search for xattrs on inodes that don't ever have xattrs and are fsynced with some frequency. The following script that calls dbench was used to measure the impact of this change on a VM with 8 CPUs, 16Gb of ram, using a raw NVMe device directly (no intermediary filesystem on the host) and using a non-debug kernel (default configuration on Debian distributions): $ cat test.sh #!/bin/bash DEV=/dev/sdk MNT=/mnt/sdk MOUNT_OPTIONS="-o ssd" mkfs.btrfs -f -m single -d single $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -D $MNT -t 200 40 umount $MNT The results before this change: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 5761605 0.172 312.057 Close 4232452 0.002 10.927 Rename 243937 1.406 277.344 Unlink 1163456 0.631 298.402 Deltree 160 11.581 221.107 Mkdir 80 0.003 0.005 Qpathinfo 5221410 0.065 122.309 Qfileinfo 915432 0.001 3.333 Qfsinfo 957555 0.003 3.992 Sfileinfo 469244 0.023 20.494 Find 2018865 0.448 123.659 WriteX 2874851 0.049 118.529 ReadX 9030579 0.004 21.654 LockX 18754 0.003 4.423 UnlockX 18754 0.002 0.331 Flush 403792 10.944 359.494 Throughput 908.444 MB/sec 40 clients 40 procs max_latency=359.500 ms The results after this change: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 6442521 0.159 230.693 Close 4732357 0.002 10.972 Rename 272809 1.293 227.398 Unlink 1301059 0.563 218.500 Deltree 160 7.796 54.887 Mkdir 80 0.008 0.478 Qpathinfo 5839452 0.047 124.330 Qfileinfo 1023199 0.001 4.996 Qfsinfo 1070760 0.003 5.709 Sfileinfo 524790 0.033 21.765 Find 2257658 0.314 125.611 WriteX 3211520 0.040 232.135 ReadX 10098969 0.004 25.340 LockX 20974 0.003 1.569 UnlockX 20974 0.002 3.475 Flush 451553 10.287 331.037 Throughput 1011.77 MB/sec 40 clients 40 procs max_latency=331.045 ms +10.8% throughput, -8.2% max latency Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-11-13 14:21:49 +03:00
if (!found_xattrs)
set_bit(BTRFS_INODE_NO_XATTRS, &inode->runtime_flags);
Btrfs: fix fsync xattr loss in the fast fsync path After commit 4f764e515361 ("Btrfs: remove deleted xattrs on fsync log replay"), we can end up in a situation where during log replay we end up deleting xattrs that were never deleted when their file was last fsynced. This happens in the fast fsync path (flag BTRFS_INODE_NEEDS_FULL_SYNC is not set in the inode) if the inode has the flag BTRFS_INODE_COPY_EVERYTHING set, the xattr was added in a past transaction and the leaf where the xattr is located was not updated (COWed or created) in the current transaction. In this scenario the xattr item never ends up in the log tree and therefore at log replay time, which makes the replay code delete the xattr from the fs/subvol tree as it thinks that xattr was deleted prior to the last fsync. Fix this by always logging all xattrs, which is the simplest and most reliable way to detect deleted xattrs and replay the deletes at log replay time. This issue is reproducible with the following test case for fstests: seq=`basename $0` seqres=$RESULT_DIR/$seq echo "QA output created by $seq" here=`pwd` tmp=/tmp/$$ status=1 # failure is the default! _cleanup() { _cleanup_flakey rm -f $tmp.* } trap "_cleanup; exit \$status" 0 1 2 3 15 # get standard environment, filters and checks . ./common/rc . ./common/filter . ./common/dmflakey . ./common/attr # real QA test starts here # We create a lot of xattrs for a single file. Only btrfs and xfs are currently # able to store such a large mount of xattrs per file, other filesystems such # as ext3/4 and f2fs for example, fail with ENOSPC even if we attempt to add # less than 1000 xattrs with very small values. _supported_fs btrfs xfs _supported_os Linux _need_to_be_root _require_scratch _require_dm_flakey _require_attrs _require_metadata_journaling $SCRATCH_DEV rm -f $seqres.full _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create the test file with some initial data and make sure everything is # durably persisted. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 32k" $SCRATCH_MNT/foo | _filter_xfs_io sync # Add many small xattrs to our file. # We create such a large amount because it's needed to trigger the issue found # in btrfs - we need to have an amount that causes the fs to have at least 3 # btree leafs with xattrs stored in them, and it must work on any leaf size # (maximum leaf/node size is 64Kb). num_xattrs=2000 for ((i = 1; i <= $num_xattrs; i++)); do name="user.attr_$(printf "%04d" $i)" $SETFATTR_PROG -n $name -v "val_$(printf "%04d" $i)" $SCRATCH_MNT/foo done # Sync the filesystem to force a commit of the current btrfs transaction, this # is a necessary condition to trigger the bug on btrfs. sync # Now update our file's data and fsync the file. # After a successful fsync, if the fsync log/journal is replayed we expect to # see all the xattrs we added before with the same values (and the updated file # data of course). Btrfs used to delete some of these xattrs when it replayed # its fsync log/journal. $XFS_IO_PROG -c "pwrite -S 0xbb 8K 16K" \ -c "fsync" \ $SCRATCH_MNT/foo | _filter_xfs_io # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again and mount. This makes the fs replay its fsync log. _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey echo "File content after crash and log replay:" od -t x1 $SCRATCH_MNT/foo echo "File xattrs after crash and log replay:" for ((i = 1; i <= $num_xattrs; i++)); do name="user.attr_$(printf "%04d" $i)" echo -n "$name=" $GETFATTR_PROG --absolute-names -n $name --only-values $SCRATCH_MNT/foo echo done status=0 exit The golden output expects all xattrs to be available, and with the correct values, after the fsync log is replayed. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-06-20 02:44:51 +03:00
return 0;
}
Btrfs: fix fsync after truncate when no_holes feature is enabled When we have the no_holes feature enabled, if a we truncate a file to a smaller size, truncate it again but to a size greater than or equals to its original size and fsync it, the log tree will not have any information about the hole covering the range [truncate_1_offset, new_file_size[. Which means if the fsync log is replayed, the file will remain with the state it had before both truncate operations. Without the no_holes feature this does not happen, since when the inode is logged (full sync flag is set) it will find in the fs/subvol tree a leaf with a generation matching the current transaction id that has an explicit extent item representing the hole. Fix this by adding an explicit extent item representing a hole between the last extent and the inode's i_size if we are doing a full sync. The issue is easy to reproduce with the following test case for fstests: . ./common/rc . ./common/filter . ./common/dmflakey _need_to_be_root _supported_fs generic _supported_os Linux _require_scratch _require_dm_flakey # This test was motivated by an issue found in btrfs when the btrfs # no-holes feature is enabled (introduced in kernel 3.14). So enable # the feature if the fs being tested is btrfs. if [ $FSTYP == "btrfs" ]; then _require_btrfs_fs_feature "no_holes" _require_btrfs_mkfs_feature "no-holes" MKFS_OPTIONS="$MKFS_OPTIONS -O no-holes" fi rm -f $seqres.full _scratch_mkfs >>$seqres.full 2>&1 _init_flakey _mount_flakey # Create our test files and make sure everything is durably persisted. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 64K" \ -c "pwrite -S 0xbb 64K 61K" \ $SCRATCH_MNT/foo | _filter_xfs_io $XFS_IO_PROG -f -c "pwrite -S 0xee 0 64K" \ -c "pwrite -S 0xff 64K 61K" \ $SCRATCH_MNT/bar | _filter_xfs_io sync # Now truncate our file foo to a smaller size (64Kb) and then truncate # it to the size it had before the shrinking truncate (125Kb). Then # fsync our file. If a power failure happens after the fsync, we expect # our file to have a size of 125Kb, with the first 64Kb of data having # the value 0xaa and the second 61Kb of data having the value 0x00. $XFS_IO_PROG -c "truncate 64K" \ -c "truncate 125K" \ -c "fsync" \ $SCRATCH_MNT/foo # Do something similar to our file bar, but the first truncation sets # the file size to 0 and the second truncation expands the size to the # double of what it was initially. $XFS_IO_PROG -c "truncate 0" \ -c "truncate 253K" \ -c "fsync" \ $SCRATCH_MNT/bar _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again, mount to trigger log replay and validate file # contents. _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # We expect foo to have a size of 125Kb, the first 64Kb of data all # having the value 0xaa and the remaining 61Kb to be a hole (all bytes # with value 0x00). echo "File foo content after log replay:" od -t x1 $SCRATCH_MNT/foo # We expect bar to have a size of 253Kb and no extents (any byte read # from bar has the value 0x00). echo "File bar content after log replay:" od -t x1 $SCRATCH_MNT/bar status=0 exit The expected file contents in the golden output are: File foo content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0200000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 * 0372000 File bar content after log replay: 0000000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 * 0772000 Without this fix, their contents are: File foo content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0200000 bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb * 0372000 File bar content after log replay: 0000000 ee ee ee ee ee ee ee ee ee ee ee ee ee ee ee ee * 0200000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0372000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 * 0772000 A test case submission for fstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Reviewed-by: Liu Bo <bo.li.liu@oracle.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-06-25 06:17:46 +03:00
/*
Btrfs: fix missing hole after hole punching and fsync when using NO_HOLES When using the NO_HOLES feature, if we punch a hole into a file and then fsync it, there are cases where a subsequent fsync will miss the fact that a hole was punched, resulting in the holes not existing after replaying the log tree. Essentially these cases all imply that, tree-log.c:copy_items(), is not invoked for the leafs that delimit holes, because nothing changed those leafs in the current transaction. And it's precisely copy_items() where we currenly detect and log holes, which works as long as the holes are between file extent items in the input leaf or between the beginning of input leaf and the previous leaf or between the last item in the leaf and the next leaf. First example where we miss a hole: *) The extent items of the inode span multiple leafs; *) The punched hole covers a range that affects only the extent items of the first leaf; *) The fsync operation is done in full mode (BTRFS_INODE_NEEDS_FULL_SYNC is set in the inode's runtime flags). That results in the hole not existing after replaying the log tree. For example, if the fs/subvolume tree has the following layout for a particular inode: Leaf N, generation 10: [ ... INODE_ITEM INODE_REF EXTENT_ITEM (0 64K) EXTENT_ITEM (64K 128K) ] Leaf N + 1, generation 10: [ EXTENT_ITEM (128K 64K) ... ] If at transaction 11 we punch a hole coverting the range [0, 128K[, we end up dropping the two extent items from leaf N, but we don't touch the other leaf, so we end up in the following state: Leaf N, generation 11: [ ... INODE_ITEM INODE_REF ] Leaf N + 1, generation 10: [ EXTENT_ITEM (128K 64K) ... ] A full fsync after punching the hole will only process leaf N because it was modified in the current transaction, but not leaf N + 1, since it was not modified in the current transaction (generation 10 and not 11). As a result the fsync will not log any holes, because it didn't process any leaf with extent items. Second example where we will miss a hole: *) An inode as its items spanning 5 (or more) leafs; *) A hole is punched and it covers only the extents items of the 3rd leaf. This resulsts in deleting the entire leaf and not touching any of the other leafs. So the only leaf that is modified in the current transaction, when punching the hole, is the first leaf, which contains the inode item. During the full fsync, the only leaf that is passed to copy_items() is that first leaf, and that's not enough for the hole detection code in copy_items() to determine there's a hole between the last file extent item in the 2nd leaf and the first file extent item in the 3rd leaf (which was the 4th leaf before punching the hole). Fix this by scanning all leafs and punch holes as necessary when doing a full fsync (less common than a non-full fsync) when the NO_HOLES feature is enabled. The lack of explicit file extent items to mark holes makes it necessary to scan existing extents to determine if holes exist. A test case for fstests follows soon. Fixes: 16e7549f045d33 ("Btrfs: incompatible format change to remove hole extents") CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-11-19 15:07:33 +03:00
* When using the NO_HOLES feature if we punched a hole that causes the
* deletion of entire leafs or all the extent items of the first leaf (the one
* that contains the inode item and references) we may end up not processing
* any extents, because there are no leafs with a generation matching the
* current transaction that have extent items for our inode. So we need to find
* if any holes exist and then log them. We also need to log holes after any
* truncate operation that changes the inode's size.
Btrfs: fix fsync after truncate when no_holes feature is enabled When we have the no_holes feature enabled, if a we truncate a file to a smaller size, truncate it again but to a size greater than or equals to its original size and fsync it, the log tree will not have any information about the hole covering the range [truncate_1_offset, new_file_size[. Which means if the fsync log is replayed, the file will remain with the state it had before both truncate operations. Without the no_holes feature this does not happen, since when the inode is logged (full sync flag is set) it will find in the fs/subvol tree a leaf with a generation matching the current transaction id that has an explicit extent item representing the hole. Fix this by adding an explicit extent item representing a hole between the last extent and the inode's i_size if we are doing a full sync. The issue is easy to reproduce with the following test case for fstests: . ./common/rc . ./common/filter . ./common/dmflakey _need_to_be_root _supported_fs generic _supported_os Linux _require_scratch _require_dm_flakey # This test was motivated by an issue found in btrfs when the btrfs # no-holes feature is enabled (introduced in kernel 3.14). So enable # the feature if the fs being tested is btrfs. if [ $FSTYP == "btrfs" ]; then _require_btrfs_fs_feature "no_holes" _require_btrfs_mkfs_feature "no-holes" MKFS_OPTIONS="$MKFS_OPTIONS -O no-holes" fi rm -f $seqres.full _scratch_mkfs >>$seqres.full 2>&1 _init_flakey _mount_flakey # Create our test files and make sure everything is durably persisted. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 64K" \ -c "pwrite -S 0xbb 64K 61K" \ $SCRATCH_MNT/foo | _filter_xfs_io $XFS_IO_PROG -f -c "pwrite -S 0xee 0 64K" \ -c "pwrite -S 0xff 64K 61K" \ $SCRATCH_MNT/bar | _filter_xfs_io sync # Now truncate our file foo to a smaller size (64Kb) and then truncate # it to the size it had before the shrinking truncate (125Kb). Then # fsync our file. If a power failure happens after the fsync, we expect # our file to have a size of 125Kb, with the first 64Kb of data having # the value 0xaa and the second 61Kb of data having the value 0x00. $XFS_IO_PROG -c "truncate 64K" \ -c "truncate 125K" \ -c "fsync" \ $SCRATCH_MNT/foo # Do something similar to our file bar, but the first truncation sets # the file size to 0 and the second truncation expands the size to the # double of what it was initially. $XFS_IO_PROG -c "truncate 0" \ -c "truncate 253K" \ -c "fsync" \ $SCRATCH_MNT/bar _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again, mount to trigger log replay and validate file # contents. _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # We expect foo to have a size of 125Kb, the first 64Kb of data all # having the value 0xaa and the remaining 61Kb to be a hole (all bytes # with value 0x00). echo "File foo content after log replay:" od -t x1 $SCRATCH_MNT/foo # We expect bar to have a size of 253Kb and no extents (any byte read # from bar has the value 0x00). echo "File bar content after log replay:" od -t x1 $SCRATCH_MNT/bar status=0 exit The expected file contents in the golden output are: File foo content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0200000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 * 0372000 File bar content after log replay: 0000000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 * 0772000 Without this fix, their contents are: File foo content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0200000 bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb * 0372000 File bar content after log replay: 0000000 ee ee ee ee ee ee ee ee ee ee ee ee ee ee ee ee * 0200000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0372000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 * 0772000 A test case submission for fstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Reviewed-by: Liu Bo <bo.li.liu@oracle.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-06-25 06:17:46 +03:00
*/
Btrfs: fix missing hole after hole punching and fsync when using NO_HOLES When using the NO_HOLES feature, if we punch a hole into a file and then fsync it, there are cases where a subsequent fsync will miss the fact that a hole was punched, resulting in the holes not existing after replaying the log tree. Essentially these cases all imply that, tree-log.c:copy_items(), is not invoked for the leafs that delimit holes, because nothing changed those leafs in the current transaction. And it's precisely copy_items() where we currenly detect and log holes, which works as long as the holes are between file extent items in the input leaf or between the beginning of input leaf and the previous leaf or between the last item in the leaf and the next leaf. First example where we miss a hole: *) The extent items of the inode span multiple leafs; *) The punched hole covers a range that affects only the extent items of the first leaf; *) The fsync operation is done in full mode (BTRFS_INODE_NEEDS_FULL_SYNC is set in the inode's runtime flags). That results in the hole not existing after replaying the log tree. For example, if the fs/subvolume tree has the following layout for a particular inode: Leaf N, generation 10: [ ... INODE_ITEM INODE_REF EXTENT_ITEM (0 64K) EXTENT_ITEM (64K 128K) ] Leaf N + 1, generation 10: [ EXTENT_ITEM (128K 64K) ... ] If at transaction 11 we punch a hole coverting the range [0, 128K[, we end up dropping the two extent items from leaf N, but we don't touch the other leaf, so we end up in the following state: Leaf N, generation 11: [ ... INODE_ITEM INODE_REF ] Leaf N + 1, generation 10: [ EXTENT_ITEM (128K 64K) ... ] A full fsync after punching the hole will only process leaf N because it was modified in the current transaction, but not leaf N + 1, since it was not modified in the current transaction (generation 10 and not 11). As a result the fsync will not log any holes, because it didn't process any leaf with extent items. Second example where we will miss a hole: *) An inode as its items spanning 5 (or more) leafs; *) A hole is punched and it covers only the extents items of the 3rd leaf. This resulsts in deleting the entire leaf and not touching any of the other leafs. So the only leaf that is modified in the current transaction, when punching the hole, is the first leaf, which contains the inode item. During the full fsync, the only leaf that is passed to copy_items() is that first leaf, and that's not enough for the hole detection code in copy_items() to determine there's a hole between the last file extent item in the 2nd leaf and the first file extent item in the 3rd leaf (which was the 4th leaf before punching the hole). Fix this by scanning all leafs and punch holes as necessary when doing a full fsync (less common than a non-full fsync) when the NO_HOLES feature is enabled. The lack of explicit file extent items to mark holes makes it necessary to scan existing extents to determine if holes exist. A test case for fstests follows soon. Fixes: 16e7549f045d33 ("Btrfs: incompatible format change to remove hole extents") CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-11-19 15:07:33 +03:00
static int btrfs_log_holes(struct btrfs_trans_handle *trans,
struct btrfs_inode *inode,
btrfs: make full fsyncs always operate on the entire file again This is a revert of commit 0a8068a3dd4294 ("btrfs: make ranged full fsyncs more efficient"), with updated comment in btrfs_sync_file. Commit 0a8068a3dd4294 ("btrfs: make ranged full fsyncs more efficient") made full fsyncs operate on the given range only as it assumed it was safe when using the NO_HOLES feature, since the hole detection was simplified some time ago and no longer was a source for races with ordered extent completion of adjacent file ranges. However it's still not safe to have a full fsync only operate on the given range, because extent maps for new extents might not be present in memory due to inode eviction or extent cloning. Consider the following example: 1) We are currently at transaction N; 2) We write to the file range [0, 1MiB); 3) Writeback finishes for the whole range and ordered extents complete, while we are still at transaction N; 4) The inode is evicted; 5) We open the file for writing, causing the inode to be loaded to memory again, which sets the 'full sync' bit on its flags. At this point the inode's list of modified extent maps is empty (figuring out which extents were created in the current transaction and were not yet logged by an fsync is expensive, that's why we set the 'full sync' bit when loading an inode); 6) We write to the file range [512KiB, 768KiB); 7) We do a ranged fsync (such as msync()) for file range [512KiB, 768KiB). This correctly flushes this range and logs its extent into the log tree. When the writeback started an extent map for range [512KiB, 768KiB) was added to the inode's list of modified extents, and when the fsync() finishes logging it removes that extent map from the list of modified extent maps. This fsync also clears the 'full sync' bit; 8) We do a regular fsync() (full ranged). This fsync() ends up doing nothing because the inode's list of modified extents is empty and no other changes happened since the previous ranged fsync(), so it just returns success (0) and we end up never logging extents for the file ranges [0, 512KiB) and [768KiB, 1MiB). Another scenario where this can happen is if we replace steps 2 to 4 with cloning from another file into our test file, as that sets the 'full sync' bit in our inode's flags and does not populate its list of modified extent maps. This was causing test case generic/457 to fail sporadically when using the NO_HOLES feature, as it exercised this later case where the inode has the 'full sync' bit set and has no extent maps in memory to represent the new extents due to extent cloning. Fix this by reverting commit 0a8068a3dd4294 ("btrfs: make ranged full fsyncs more efficient") since there is no easy way to work around it. Fixes: 0a8068a3dd4294 ("btrfs: make ranged full fsyncs more efficient") Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-04-07 13:37:44 +03:00
struct btrfs_path *path)
Btrfs: fix fsync after truncate when no_holes feature is enabled When we have the no_holes feature enabled, if a we truncate a file to a smaller size, truncate it again but to a size greater than or equals to its original size and fsync it, the log tree will not have any information about the hole covering the range [truncate_1_offset, new_file_size[. Which means if the fsync log is replayed, the file will remain with the state it had before both truncate operations. Without the no_holes feature this does not happen, since when the inode is logged (full sync flag is set) it will find in the fs/subvol tree a leaf with a generation matching the current transaction id that has an explicit extent item representing the hole. Fix this by adding an explicit extent item representing a hole between the last extent and the inode's i_size if we are doing a full sync. The issue is easy to reproduce with the following test case for fstests: . ./common/rc . ./common/filter . ./common/dmflakey _need_to_be_root _supported_fs generic _supported_os Linux _require_scratch _require_dm_flakey # This test was motivated by an issue found in btrfs when the btrfs # no-holes feature is enabled (introduced in kernel 3.14). So enable # the feature if the fs being tested is btrfs. if [ $FSTYP == "btrfs" ]; then _require_btrfs_fs_feature "no_holes" _require_btrfs_mkfs_feature "no-holes" MKFS_OPTIONS="$MKFS_OPTIONS -O no-holes" fi rm -f $seqres.full _scratch_mkfs >>$seqres.full 2>&1 _init_flakey _mount_flakey # Create our test files and make sure everything is durably persisted. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 64K" \ -c "pwrite -S 0xbb 64K 61K" \ $SCRATCH_MNT/foo | _filter_xfs_io $XFS_IO_PROG -f -c "pwrite -S 0xee 0 64K" \ -c "pwrite -S 0xff 64K 61K" \ $SCRATCH_MNT/bar | _filter_xfs_io sync # Now truncate our file foo to a smaller size (64Kb) and then truncate # it to the size it had before the shrinking truncate (125Kb). Then # fsync our file. If a power failure happens after the fsync, we expect # our file to have a size of 125Kb, with the first 64Kb of data having # the value 0xaa and the second 61Kb of data having the value 0x00. $XFS_IO_PROG -c "truncate 64K" \ -c "truncate 125K" \ -c "fsync" \ $SCRATCH_MNT/foo # Do something similar to our file bar, but the first truncation sets # the file size to 0 and the second truncation expands the size to the # double of what it was initially. $XFS_IO_PROG -c "truncate 0" \ -c "truncate 253K" \ -c "fsync" \ $SCRATCH_MNT/bar _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again, mount to trigger log replay and validate file # contents. _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # We expect foo to have a size of 125Kb, the first 64Kb of data all # having the value 0xaa and the remaining 61Kb to be a hole (all bytes # with value 0x00). echo "File foo content after log replay:" od -t x1 $SCRATCH_MNT/foo # We expect bar to have a size of 253Kb and no extents (any byte read # from bar has the value 0x00). echo "File bar content after log replay:" od -t x1 $SCRATCH_MNT/bar status=0 exit The expected file contents in the golden output are: File foo content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0200000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 * 0372000 File bar content after log replay: 0000000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 * 0772000 Without this fix, their contents are: File foo content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0200000 bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb * 0372000 File bar content after log replay: 0000000 ee ee ee ee ee ee ee ee ee ee ee ee ee ee ee ee * 0200000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0372000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 * 0772000 A test case submission for fstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Reviewed-by: Liu Bo <bo.li.liu@oracle.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-06-25 06:17:46 +03:00
{
struct btrfs_root *root = inode->root;
struct btrfs_fs_info *fs_info = root->fs_info;
Btrfs: fix fsync after truncate when no_holes feature is enabled When we have the no_holes feature enabled, if a we truncate a file to a smaller size, truncate it again but to a size greater than or equals to its original size and fsync it, the log tree will not have any information about the hole covering the range [truncate_1_offset, new_file_size[. Which means if the fsync log is replayed, the file will remain with the state it had before both truncate operations. Without the no_holes feature this does not happen, since when the inode is logged (full sync flag is set) it will find in the fs/subvol tree a leaf with a generation matching the current transaction id that has an explicit extent item representing the hole. Fix this by adding an explicit extent item representing a hole between the last extent and the inode's i_size if we are doing a full sync. The issue is easy to reproduce with the following test case for fstests: . ./common/rc . ./common/filter . ./common/dmflakey _need_to_be_root _supported_fs generic _supported_os Linux _require_scratch _require_dm_flakey # This test was motivated by an issue found in btrfs when the btrfs # no-holes feature is enabled (introduced in kernel 3.14). So enable # the feature if the fs being tested is btrfs. if [ $FSTYP == "btrfs" ]; then _require_btrfs_fs_feature "no_holes" _require_btrfs_mkfs_feature "no-holes" MKFS_OPTIONS="$MKFS_OPTIONS -O no-holes" fi rm -f $seqres.full _scratch_mkfs >>$seqres.full 2>&1 _init_flakey _mount_flakey # Create our test files and make sure everything is durably persisted. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 64K" \ -c "pwrite -S 0xbb 64K 61K" \ $SCRATCH_MNT/foo | _filter_xfs_io $XFS_IO_PROG -f -c "pwrite -S 0xee 0 64K" \ -c "pwrite -S 0xff 64K 61K" \ $SCRATCH_MNT/bar | _filter_xfs_io sync # Now truncate our file foo to a smaller size (64Kb) and then truncate # it to the size it had before the shrinking truncate (125Kb). Then # fsync our file. If a power failure happens after the fsync, we expect # our file to have a size of 125Kb, with the first 64Kb of data having # the value 0xaa and the second 61Kb of data having the value 0x00. $XFS_IO_PROG -c "truncate 64K" \ -c "truncate 125K" \ -c "fsync" \ $SCRATCH_MNT/foo # Do something similar to our file bar, but the first truncation sets # the file size to 0 and the second truncation expands the size to the # double of what it was initially. $XFS_IO_PROG -c "truncate 0" \ -c "truncate 253K" \ -c "fsync" \ $SCRATCH_MNT/bar _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again, mount to trigger log replay and validate file # contents. _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # We expect foo to have a size of 125Kb, the first 64Kb of data all # having the value 0xaa and the remaining 61Kb to be a hole (all bytes # with value 0x00). echo "File foo content after log replay:" od -t x1 $SCRATCH_MNT/foo # We expect bar to have a size of 253Kb and no extents (any byte read # from bar has the value 0x00). echo "File bar content after log replay:" od -t x1 $SCRATCH_MNT/bar status=0 exit The expected file contents in the golden output are: File foo content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0200000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 * 0372000 File bar content after log replay: 0000000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 * 0772000 Without this fix, their contents are: File foo content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0200000 bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb * 0372000 File bar content after log replay: 0000000 ee ee ee ee ee ee ee ee ee ee ee ee ee ee ee ee * 0200000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0372000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 * 0772000 A test case submission for fstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Reviewed-by: Liu Bo <bo.li.liu@oracle.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-06-25 06:17:46 +03:00
struct btrfs_key key;
const u64 ino = btrfs_ino(inode);
const u64 i_size = i_size_read(&inode->vfs_inode);
btrfs: make full fsyncs always operate on the entire file again This is a revert of commit 0a8068a3dd4294 ("btrfs: make ranged full fsyncs more efficient"), with updated comment in btrfs_sync_file. Commit 0a8068a3dd4294 ("btrfs: make ranged full fsyncs more efficient") made full fsyncs operate on the given range only as it assumed it was safe when using the NO_HOLES feature, since the hole detection was simplified some time ago and no longer was a source for races with ordered extent completion of adjacent file ranges. However it's still not safe to have a full fsync only operate on the given range, because extent maps for new extents might not be present in memory due to inode eviction or extent cloning. Consider the following example: 1) We are currently at transaction N; 2) We write to the file range [0, 1MiB); 3) Writeback finishes for the whole range and ordered extents complete, while we are still at transaction N; 4) The inode is evicted; 5) We open the file for writing, causing the inode to be loaded to memory again, which sets the 'full sync' bit on its flags. At this point the inode's list of modified extent maps is empty (figuring out which extents were created in the current transaction and were not yet logged by an fsync is expensive, that's why we set the 'full sync' bit when loading an inode); 6) We write to the file range [512KiB, 768KiB); 7) We do a ranged fsync (such as msync()) for file range [512KiB, 768KiB). This correctly flushes this range and logs its extent into the log tree. When the writeback started an extent map for range [512KiB, 768KiB) was added to the inode's list of modified extents, and when the fsync() finishes logging it removes that extent map from the list of modified extent maps. This fsync also clears the 'full sync' bit; 8) We do a regular fsync() (full ranged). This fsync() ends up doing nothing because the inode's list of modified extents is empty and no other changes happened since the previous ranged fsync(), so it just returns success (0) and we end up never logging extents for the file ranges [0, 512KiB) and [768KiB, 1MiB). Another scenario where this can happen is if we replace steps 2 to 4 with cloning from another file into our test file, as that sets the 'full sync' bit in our inode's flags and does not populate its list of modified extent maps. This was causing test case generic/457 to fail sporadically when using the NO_HOLES feature, as it exercised this later case where the inode has the 'full sync' bit set and has no extent maps in memory to represent the new extents due to extent cloning. Fix this by reverting commit 0a8068a3dd4294 ("btrfs: make ranged full fsyncs more efficient") since there is no easy way to work around it. Fixes: 0a8068a3dd4294 ("btrfs: make ranged full fsyncs more efficient") Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-04-07 13:37:44 +03:00
u64 prev_extent_end = 0;
Btrfs: fix missing hole after hole punching and fsync when using NO_HOLES When using the NO_HOLES feature, if we punch a hole into a file and then fsync it, there are cases where a subsequent fsync will miss the fact that a hole was punched, resulting in the holes not existing after replaying the log tree. Essentially these cases all imply that, tree-log.c:copy_items(), is not invoked for the leafs that delimit holes, because nothing changed those leafs in the current transaction. And it's precisely copy_items() where we currenly detect and log holes, which works as long as the holes are between file extent items in the input leaf or between the beginning of input leaf and the previous leaf or between the last item in the leaf and the next leaf. First example where we miss a hole: *) The extent items of the inode span multiple leafs; *) The punched hole covers a range that affects only the extent items of the first leaf; *) The fsync operation is done in full mode (BTRFS_INODE_NEEDS_FULL_SYNC is set in the inode's runtime flags). That results in the hole not existing after replaying the log tree. For example, if the fs/subvolume tree has the following layout for a particular inode: Leaf N, generation 10: [ ... INODE_ITEM INODE_REF EXTENT_ITEM (0 64K) EXTENT_ITEM (64K 128K) ] Leaf N + 1, generation 10: [ EXTENT_ITEM (128K 64K) ... ] If at transaction 11 we punch a hole coverting the range [0, 128K[, we end up dropping the two extent items from leaf N, but we don't touch the other leaf, so we end up in the following state: Leaf N, generation 11: [ ... INODE_ITEM INODE_REF ] Leaf N + 1, generation 10: [ EXTENT_ITEM (128K 64K) ... ] A full fsync after punching the hole will only process leaf N because it was modified in the current transaction, but not leaf N + 1, since it was not modified in the current transaction (generation 10 and not 11). As a result the fsync will not log any holes, because it didn't process any leaf with extent items. Second example where we will miss a hole: *) An inode as its items spanning 5 (or more) leafs; *) A hole is punched and it covers only the extents items of the 3rd leaf. This resulsts in deleting the entire leaf and not touching any of the other leafs. So the only leaf that is modified in the current transaction, when punching the hole, is the first leaf, which contains the inode item. During the full fsync, the only leaf that is passed to copy_items() is that first leaf, and that's not enough for the hole detection code in copy_items() to determine there's a hole between the last file extent item in the 2nd leaf and the first file extent item in the 3rd leaf (which was the 4th leaf before punching the hole). Fix this by scanning all leafs and punch holes as necessary when doing a full fsync (less common than a non-full fsync) when the NO_HOLES feature is enabled. The lack of explicit file extent items to mark holes makes it necessary to scan existing extents to determine if holes exist. A test case for fstests follows soon. Fixes: 16e7549f045d33 ("Btrfs: incompatible format change to remove hole extents") CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-11-19 15:07:33 +03:00
int ret;
Btrfs: fix fsync after truncate when no_holes feature is enabled When we have the no_holes feature enabled, if a we truncate a file to a smaller size, truncate it again but to a size greater than or equals to its original size and fsync it, the log tree will not have any information about the hole covering the range [truncate_1_offset, new_file_size[. Which means if the fsync log is replayed, the file will remain with the state it had before both truncate operations. Without the no_holes feature this does not happen, since when the inode is logged (full sync flag is set) it will find in the fs/subvol tree a leaf with a generation matching the current transaction id that has an explicit extent item representing the hole. Fix this by adding an explicit extent item representing a hole between the last extent and the inode's i_size if we are doing a full sync. The issue is easy to reproduce with the following test case for fstests: . ./common/rc . ./common/filter . ./common/dmflakey _need_to_be_root _supported_fs generic _supported_os Linux _require_scratch _require_dm_flakey # This test was motivated by an issue found in btrfs when the btrfs # no-holes feature is enabled (introduced in kernel 3.14). So enable # the feature if the fs being tested is btrfs. if [ $FSTYP == "btrfs" ]; then _require_btrfs_fs_feature "no_holes" _require_btrfs_mkfs_feature "no-holes" MKFS_OPTIONS="$MKFS_OPTIONS -O no-holes" fi rm -f $seqres.full _scratch_mkfs >>$seqres.full 2>&1 _init_flakey _mount_flakey # Create our test files and make sure everything is durably persisted. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 64K" \ -c "pwrite -S 0xbb 64K 61K" \ $SCRATCH_MNT/foo | _filter_xfs_io $XFS_IO_PROG -f -c "pwrite -S 0xee 0 64K" \ -c "pwrite -S 0xff 64K 61K" \ $SCRATCH_MNT/bar | _filter_xfs_io sync # Now truncate our file foo to a smaller size (64Kb) and then truncate # it to the size it had before the shrinking truncate (125Kb). Then # fsync our file. If a power failure happens after the fsync, we expect # our file to have a size of 125Kb, with the first 64Kb of data having # the value 0xaa and the second 61Kb of data having the value 0x00. $XFS_IO_PROG -c "truncate 64K" \ -c "truncate 125K" \ -c "fsync" \ $SCRATCH_MNT/foo # Do something similar to our file bar, but the first truncation sets # the file size to 0 and the second truncation expands the size to the # double of what it was initially. $XFS_IO_PROG -c "truncate 0" \ -c "truncate 253K" \ -c "fsync" \ $SCRATCH_MNT/bar _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again, mount to trigger log replay and validate file # contents. _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # We expect foo to have a size of 125Kb, the first 64Kb of data all # having the value 0xaa and the remaining 61Kb to be a hole (all bytes # with value 0x00). echo "File foo content after log replay:" od -t x1 $SCRATCH_MNT/foo # We expect bar to have a size of 253Kb and no extents (any byte read # from bar has the value 0x00). echo "File bar content after log replay:" od -t x1 $SCRATCH_MNT/bar status=0 exit The expected file contents in the golden output are: File foo content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0200000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 * 0372000 File bar content after log replay: 0000000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 * 0772000 Without this fix, their contents are: File foo content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0200000 bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb * 0372000 File bar content after log replay: 0000000 ee ee ee ee ee ee ee ee ee ee ee ee ee ee ee ee * 0200000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0372000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 * 0772000 A test case submission for fstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Reviewed-by: Liu Bo <bo.li.liu@oracle.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-06-25 06:17:46 +03:00
Btrfs: fix missing hole after hole punching and fsync when using NO_HOLES When using the NO_HOLES feature, if we punch a hole into a file and then fsync it, there are cases where a subsequent fsync will miss the fact that a hole was punched, resulting in the holes not existing after replaying the log tree. Essentially these cases all imply that, tree-log.c:copy_items(), is not invoked for the leafs that delimit holes, because nothing changed those leafs in the current transaction. And it's precisely copy_items() where we currenly detect and log holes, which works as long as the holes are between file extent items in the input leaf or between the beginning of input leaf and the previous leaf or between the last item in the leaf and the next leaf. First example where we miss a hole: *) The extent items of the inode span multiple leafs; *) The punched hole covers a range that affects only the extent items of the first leaf; *) The fsync operation is done in full mode (BTRFS_INODE_NEEDS_FULL_SYNC is set in the inode's runtime flags). That results in the hole not existing after replaying the log tree. For example, if the fs/subvolume tree has the following layout for a particular inode: Leaf N, generation 10: [ ... INODE_ITEM INODE_REF EXTENT_ITEM (0 64K) EXTENT_ITEM (64K 128K) ] Leaf N + 1, generation 10: [ EXTENT_ITEM (128K 64K) ... ] If at transaction 11 we punch a hole coverting the range [0, 128K[, we end up dropping the two extent items from leaf N, but we don't touch the other leaf, so we end up in the following state: Leaf N, generation 11: [ ... INODE_ITEM INODE_REF ] Leaf N + 1, generation 10: [ EXTENT_ITEM (128K 64K) ... ] A full fsync after punching the hole will only process leaf N because it was modified in the current transaction, but not leaf N + 1, since it was not modified in the current transaction (generation 10 and not 11). As a result the fsync will not log any holes, because it didn't process any leaf with extent items. Second example where we will miss a hole: *) An inode as its items spanning 5 (or more) leafs; *) A hole is punched and it covers only the extents items of the 3rd leaf. This resulsts in deleting the entire leaf and not touching any of the other leafs. So the only leaf that is modified in the current transaction, when punching the hole, is the first leaf, which contains the inode item. During the full fsync, the only leaf that is passed to copy_items() is that first leaf, and that's not enough for the hole detection code in copy_items() to determine there's a hole between the last file extent item in the 2nd leaf and the first file extent item in the 3rd leaf (which was the 4th leaf before punching the hole). Fix this by scanning all leafs and punch holes as necessary when doing a full fsync (less common than a non-full fsync) when the NO_HOLES feature is enabled. The lack of explicit file extent items to mark holes makes it necessary to scan existing extents to determine if holes exist. A test case for fstests follows soon. Fixes: 16e7549f045d33 ("Btrfs: incompatible format change to remove hole extents") CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-11-19 15:07:33 +03:00
if (!btrfs_fs_incompat(fs_info, NO_HOLES) || i_size == 0)
Btrfs: fix fsync after truncate when no_holes feature is enabled When we have the no_holes feature enabled, if a we truncate a file to a smaller size, truncate it again but to a size greater than or equals to its original size and fsync it, the log tree will not have any information about the hole covering the range [truncate_1_offset, new_file_size[. Which means if the fsync log is replayed, the file will remain with the state it had before both truncate operations. Without the no_holes feature this does not happen, since when the inode is logged (full sync flag is set) it will find in the fs/subvol tree a leaf with a generation matching the current transaction id that has an explicit extent item representing the hole. Fix this by adding an explicit extent item representing a hole between the last extent and the inode's i_size if we are doing a full sync. The issue is easy to reproduce with the following test case for fstests: . ./common/rc . ./common/filter . ./common/dmflakey _need_to_be_root _supported_fs generic _supported_os Linux _require_scratch _require_dm_flakey # This test was motivated by an issue found in btrfs when the btrfs # no-holes feature is enabled (introduced in kernel 3.14). So enable # the feature if the fs being tested is btrfs. if [ $FSTYP == "btrfs" ]; then _require_btrfs_fs_feature "no_holes" _require_btrfs_mkfs_feature "no-holes" MKFS_OPTIONS="$MKFS_OPTIONS -O no-holes" fi rm -f $seqres.full _scratch_mkfs >>$seqres.full 2>&1 _init_flakey _mount_flakey # Create our test files and make sure everything is durably persisted. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 64K" \ -c "pwrite -S 0xbb 64K 61K" \ $SCRATCH_MNT/foo | _filter_xfs_io $XFS_IO_PROG -f -c "pwrite -S 0xee 0 64K" \ -c "pwrite -S 0xff 64K 61K" \ $SCRATCH_MNT/bar | _filter_xfs_io sync # Now truncate our file foo to a smaller size (64Kb) and then truncate # it to the size it had before the shrinking truncate (125Kb). Then # fsync our file. If a power failure happens after the fsync, we expect # our file to have a size of 125Kb, with the first 64Kb of data having # the value 0xaa and the second 61Kb of data having the value 0x00. $XFS_IO_PROG -c "truncate 64K" \ -c "truncate 125K" \ -c "fsync" \ $SCRATCH_MNT/foo # Do something similar to our file bar, but the first truncation sets # the file size to 0 and the second truncation expands the size to the # double of what it was initially. $XFS_IO_PROG -c "truncate 0" \ -c "truncate 253K" \ -c "fsync" \ $SCRATCH_MNT/bar _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again, mount to trigger log replay and validate file # contents. _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # We expect foo to have a size of 125Kb, the first 64Kb of data all # having the value 0xaa and the remaining 61Kb to be a hole (all bytes # with value 0x00). echo "File foo content after log replay:" od -t x1 $SCRATCH_MNT/foo # We expect bar to have a size of 253Kb and no extents (any byte read # from bar has the value 0x00). echo "File bar content after log replay:" od -t x1 $SCRATCH_MNT/bar status=0 exit The expected file contents in the golden output are: File foo content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0200000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 * 0372000 File bar content after log replay: 0000000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 * 0772000 Without this fix, their contents are: File foo content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0200000 bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb * 0372000 File bar content after log replay: 0000000 ee ee ee ee ee ee ee ee ee ee ee ee ee ee ee ee * 0200000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0372000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 * 0772000 A test case submission for fstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Reviewed-by: Liu Bo <bo.li.liu@oracle.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-06-25 06:17:46 +03:00
return 0;
key.objectid = ino;
key.type = BTRFS_EXTENT_DATA_KEY;
btrfs: make full fsyncs always operate on the entire file again This is a revert of commit 0a8068a3dd4294 ("btrfs: make ranged full fsyncs more efficient"), with updated comment in btrfs_sync_file. Commit 0a8068a3dd4294 ("btrfs: make ranged full fsyncs more efficient") made full fsyncs operate on the given range only as it assumed it was safe when using the NO_HOLES feature, since the hole detection was simplified some time ago and no longer was a source for races with ordered extent completion of adjacent file ranges. However it's still not safe to have a full fsync only operate on the given range, because extent maps for new extents might not be present in memory due to inode eviction or extent cloning. Consider the following example: 1) We are currently at transaction N; 2) We write to the file range [0, 1MiB); 3) Writeback finishes for the whole range and ordered extents complete, while we are still at transaction N; 4) The inode is evicted; 5) We open the file for writing, causing the inode to be loaded to memory again, which sets the 'full sync' bit on its flags. At this point the inode's list of modified extent maps is empty (figuring out which extents were created in the current transaction and were not yet logged by an fsync is expensive, that's why we set the 'full sync' bit when loading an inode); 6) We write to the file range [512KiB, 768KiB); 7) We do a ranged fsync (such as msync()) for file range [512KiB, 768KiB). This correctly flushes this range and logs its extent into the log tree. When the writeback started an extent map for range [512KiB, 768KiB) was added to the inode's list of modified extents, and when the fsync() finishes logging it removes that extent map from the list of modified extent maps. This fsync also clears the 'full sync' bit; 8) We do a regular fsync() (full ranged). This fsync() ends up doing nothing because the inode's list of modified extents is empty and no other changes happened since the previous ranged fsync(), so it just returns success (0) and we end up never logging extents for the file ranges [0, 512KiB) and [768KiB, 1MiB). Another scenario where this can happen is if we replace steps 2 to 4 with cloning from another file into our test file, as that sets the 'full sync' bit in our inode's flags and does not populate its list of modified extent maps. This was causing test case generic/457 to fail sporadically when using the NO_HOLES feature, as it exercised this later case where the inode has the 'full sync' bit set and has no extent maps in memory to represent the new extents due to extent cloning. Fix this by reverting commit 0a8068a3dd4294 ("btrfs: make ranged full fsyncs more efficient") since there is no easy way to work around it. Fixes: 0a8068a3dd4294 ("btrfs: make ranged full fsyncs more efficient") Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-04-07 13:37:44 +03:00
key.offset = 0;
Btrfs: fix fsync after truncate when no_holes feature is enabled When we have the no_holes feature enabled, if a we truncate a file to a smaller size, truncate it again but to a size greater than or equals to its original size and fsync it, the log tree will not have any information about the hole covering the range [truncate_1_offset, new_file_size[. Which means if the fsync log is replayed, the file will remain with the state it had before both truncate operations. Without the no_holes feature this does not happen, since when the inode is logged (full sync flag is set) it will find in the fs/subvol tree a leaf with a generation matching the current transaction id that has an explicit extent item representing the hole. Fix this by adding an explicit extent item representing a hole between the last extent and the inode's i_size if we are doing a full sync. The issue is easy to reproduce with the following test case for fstests: . ./common/rc . ./common/filter . ./common/dmflakey _need_to_be_root _supported_fs generic _supported_os Linux _require_scratch _require_dm_flakey # This test was motivated by an issue found in btrfs when the btrfs # no-holes feature is enabled (introduced in kernel 3.14). So enable # the feature if the fs being tested is btrfs. if [ $FSTYP == "btrfs" ]; then _require_btrfs_fs_feature "no_holes" _require_btrfs_mkfs_feature "no-holes" MKFS_OPTIONS="$MKFS_OPTIONS -O no-holes" fi rm -f $seqres.full _scratch_mkfs >>$seqres.full 2>&1 _init_flakey _mount_flakey # Create our test files and make sure everything is durably persisted. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 64K" \ -c "pwrite -S 0xbb 64K 61K" \ $SCRATCH_MNT/foo | _filter_xfs_io $XFS_IO_PROG -f -c "pwrite -S 0xee 0 64K" \ -c "pwrite -S 0xff 64K 61K" \ $SCRATCH_MNT/bar | _filter_xfs_io sync # Now truncate our file foo to a smaller size (64Kb) and then truncate # it to the size it had before the shrinking truncate (125Kb). Then # fsync our file. If a power failure happens after the fsync, we expect # our file to have a size of 125Kb, with the first 64Kb of data having # the value 0xaa and the second 61Kb of data having the value 0x00. $XFS_IO_PROG -c "truncate 64K" \ -c "truncate 125K" \ -c "fsync" \ $SCRATCH_MNT/foo # Do something similar to our file bar, but the first truncation sets # the file size to 0 and the second truncation expands the size to the # double of what it was initially. $XFS_IO_PROG -c "truncate 0" \ -c "truncate 253K" \ -c "fsync" \ $SCRATCH_MNT/bar _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again, mount to trigger log replay and validate file # contents. _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # We expect foo to have a size of 125Kb, the first 64Kb of data all # having the value 0xaa and the remaining 61Kb to be a hole (all bytes # with value 0x00). echo "File foo content after log replay:" od -t x1 $SCRATCH_MNT/foo # We expect bar to have a size of 253Kb and no extents (any byte read # from bar has the value 0x00). echo "File bar content after log replay:" od -t x1 $SCRATCH_MNT/bar status=0 exit The expected file contents in the golden output are: File foo content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0200000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 * 0372000 File bar content after log replay: 0000000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 * 0772000 Without this fix, their contents are: File foo content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0200000 bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb * 0372000 File bar content after log replay: 0000000 ee ee ee ee ee ee ee ee ee ee ee ee ee ee ee ee * 0200000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0372000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 * 0772000 A test case submission for fstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Reviewed-by: Liu Bo <bo.li.liu@oracle.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-06-25 06:17:46 +03:00
ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
if (ret < 0)
return ret;
Btrfs: fix missing hole after hole punching and fsync when using NO_HOLES When using the NO_HOLES feature, if we punch a hole into a file and then fsync it, there are cases where a subsequent fsync will miss the fact that a hole was punched, resulting in the holes not existing after replaying the log tree. Essentially these cases all imply that, tree-log.c:copy_items(), is not invoked for the leafs that delimit holes, because nothing changed those leafs in the current transaction. And it's precisely copy_items() where we currenly detect and log holes, which works as long as the holes are between file extent items in the input leaf or between the beginning of input leaf and the previous leaf or between the last item in the leaf and the next leaf. First example where we miss a hole: *) The extent items of the inode span multiple leafs; *) The punched hole covers a range that affects only the extent items of the first leaf; *) The fsync operation is done in full mode (BTRFS_INODE_NEEDS_FULL_SYNC is set in the inode's runtime flags). That results in the hole not existing after replaying the log tree. For example, if the fs/subvolume tree has the following layout for a particular inode: Leaf N, generation 10: [ ... INODE_ITEM INODE_REF EXTENT_ITEM (0 64K) EXTENT_ITEM (64K 128K) ] Leaf N + 1, generation 10: [ EXTENT_ITEM (128K 64K) ... ] If at transaction 11 we punch a hole coverting the range [0, 128K[, we end up dropping the two extent items from leaf N, but we don't touch the other leaf, so we end up in the following state: Leaf N, generation 11: [ ... INODE_ITEM INODE_REF ] Leaf N + 1, generation 10: [ EXTENT_ITEM (128K 64K) ... ] A full fsync after punching the hole will only process leaf N because it was modified in the current transaction, but not leaf N + 1, since it was not modified in the current transaction (generation 10 and not 11). As a result the fsync will not log any holes, because it didn't process any leaf with extent items. Second example where we will miss a hole: *) An inode as its items spanning 5 (or more) leafs; *) A hole is punched and it covers only the extents items of the 3rd leaf. This resulsts in deleting the entire leaf and not touching any of the other leafs. So the only leaf that is modified in the current transaction, when punching the hole, is the first leaf, which contains the inode item. During the full fsync, the only leaf that is passed to copy_items() is that first leaf, and that's not enough for the hole detection code in copy_items() to determine there's a hole between the last file extent item in the 2nd leaf and the first file extent item in the 3rd leaf (which was the 4th leaf before punching the hole). Fix this by scanning all leafs and punch holes as necessary when doing a full fsync (less common than a non-full fsync) when the NO_HOLES feature is enabled. The lack of explicit file extent items to mark holes makes it necessary to scan existing extents to determine if holes exist. A test case for fstests follows soon. Fixes: 16e7549f045d33 ("Btrfs: incompatible format change to remove hole extents") CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-11-19 15:07:33 +03:00
while (true) {
struct extent_buffer *leaf = path->nodes[0];
Btrfs: fix fsync after truncate when no_holes feature is enabled When we have the no_holes feature enabled, if a we truncate a file to a smaller size, truncate it again but to a size greater than or equals to its original size and fsync it, the log tree will not have any information about the hole covering the range [truncate_1_offset, new_file_size[. Which means if the fsync log is replayed, the file will remain with the state it had before both truncate operations. Without the no_holes feature this does not happen, since when the inode is logged (full sync flag is set) it will find in the fs/subvol tree a leaf with a generation matching the current transaction id that has an explicit extent item representing the hole. Fix this by adding an explicit extent item representing a hole between the last extent and the inode's i_size if we are doing a full sync. The issue is easy to reproduce with the following test case for fstests: . ./common/rc . ./common/filter . ./common/dmflakey _need_to_be_root _supported_fs generic _supported_os Linux _require_scratch _require_dm_flakey # This test was motivated by an issue found in btrfs when the btrfs # no-holes feature is enabled (introduced in kernel 3.14). So enable # the feature if the fs being tested is btrfs. if [ $FSTYP == "btrfs" ]; then _require_btrfs_fs_feature "no_holes" _require_btrfs_mkfs_feature "no-holes" MKFS_OPTIONS="$MKFS_OPTIONS -O no-holes" fi rm -f $seqres.full _scratch_mkfs >>$seqres.full 2>&1 _init_flakey _mount_flakey # Create our test files and make sure everything is durably persisted. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 64K" \ -c "pwrite -S 0xbb 64K 61K" \ $SCRATCH_MNT/foo | _filter_xfs_io $XFS_IO_PROG -f -c "pwrite -S 0xee 0 64K" \ -c "pwrite -S 0xff 64K 61K" \ $SCRATCH_MNT/bar | _filter_xfs_io sync # Now truncate our file foo to a smaller size (64Kb) and then truncate # it to the size it had before the shrinking truncate (125Kb). Then # fsync our file. If a power failure happens after the fsync, we expect # our file to have a size of 125Kb, with the first 64Kb of data having # the value 0xaa and the second 61Kb of data having the value 0x00. $XFS_IO_PROG -c "truncate 64K" \ -c "truncate 125K" \ -c "fsync" \ $SCRATCH_MNT/foo # Do something similar to our file bar, but the first truncation sets # the file size to 0 and the second truncation expands the size to the # double of what it was initially. $XFS_IO_PROG -c "truncate 0" \ -c "truncate 253K" \ -c "fsync" \ $SCRATCH_MNT/bar _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again, mount to trigger log replay and validate file # contents. _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # We expect foo to have a size of 125Kb, the first 64Kb of data all # having the value 0xaa and the remaining 61Kb to be a hole (all bytes # with value 0x00). echo "File foo content after log replay:" od -t x1 $SCRATCH_MNT/foo # We expect bar to have a size of 253Kb and no extents (any byte read # from bar has the value 0x00). echo "File bar content after log replay:" od -t x1 $SCRATCH_MNT/bar status=0 exit The expected file contents in the golden output are: File foo content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0200000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 * 0372000 File bar content after log replay: 0000000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 * 0772000 Without this fix, their contents are: File foo content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0200000 bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb * 0372000 File bar content after log replay: 0000000 ee ee ee ee ee ee ee ee ee ee ee ee ee ee ee ee * 0200000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0372000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 * 0772000 A test case submission for fstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Reviewed-by: Liu Bo <bo.li.liu@oracle.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-06-25 06:17:46 +03:00
Btrfs: fix missing hole after hole punching and fsync when using NO_HOLES When using the NO_HOLES feature, if we punch a hole into a file and then fsync it, there are cases where a subsequent fsync will miss the fact that a hole was punched, resulting in the holes not existing after replaying the log tree. Essentially these cases all imply that, tree-log.c:copy_items(), is not invoked for the leafs that delimit holes, because nothing changed those leafs in the current transaction. And it's precisely copy_items() where we currenly detect and log holes, which works as long as the holes are between file extent items in the input leaf or between the beginning of input leaf and the previous leaf or between the last item in the leaf and the next leaf. First example where we miss a hole: *) The extent items of the inode span multiple leafs; *) The punched hole covers a range that affects only the extent items of the first leaf; *) The fsync operation is done in full mode (BTRFS_INODE_NEEDS_FULL_SYNC is set in the inode's runtime flags). That results in the hole not existing after replaying the log tree. For example, if the fs/subvolume tree has the following layout for a particular inode: Leaf N, generation 10: [ ... INODE_ITEM INODE_REF EXTENT_ITEM (0 64K) EXTENT_ITEM (64K 128K) ] Leaf N + 1, generation 10: [ EXTENT_ITEM (128K 64K) ... ] If at transaction 11 we punch a hole coverting the range [0, 128K[, we end up dropping the two extent items from leaf N, but we don't touch the other leaf, so we end up in the following state: Leaf N, generation 11: [ ... INODE_ITEM INODE_REF ] Leaf N + 1, generation 10: [ EXTENT_ITEM (128K 64K) ... ] A full fsync after punching the hole will only process leaf N because it was modified in the current transaction, but not leaf N + 1, since it was not modified in the current transaction (generation 10 and not 11). As a result the fsync will not log any holes, because it didn't process any leaf with extent items. Second example where we will miss a hole: *) An inode as its items spanning 5 (or more) leafs; *) A hole is punched and it covers only the extents items of the 3rd leaf. This resulsts in deleting the entire leaf and not touching any of the other leafs. So the only leaf that is modified in the current transaction, when punching the hole, is the first leaf, which contains the inode item. During the full fsync, the only leaf that is passed to copy_items() is that first leaf, and that's not enough for the hole detection code in copy_items() to determine there's a hole between the last file extent item in the 2nd leaf and the first file extent item in the 3rd leaf (which was the 4th leaf before punching the hole). Fix this by scanning all leafs and punch holes as necessary when doing a full fsync (less common than a non-full fsync) when the NO_HOLES feature is enabled. The lack of explicit file extent items to mark holes makes it necessary to scan existing extents to determine if holes exist. A test case for fstests follows soon. Fixes: 16e7549f045d33 ("Btrfs: incompatible format change to remove hole extents") CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-11-19 15:07:33 +03:00
if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) {
ret = btrfs_next_leaf(root, path);
if (ret < 0)
return ret;
if (ret > 0) {
ret = 0;
break;
}
leaf = path->nodes[0];
}
btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
if (key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY)
break;
/* We have a hole, log it. */
if (prev_extent_end < key.offset) {
btrfs: make full fsyncs always operate on the entire file again This is a revert of commit 0a8068a3dd4294 ("btrfs: make ranged full fsyncs more efficient"), with updated comment in btrfs_sync_file. Commit 0a8068a3dd4294 ("btrfs: make ranged full fsyncs more efficient") made full fsyncs operate on the given range only as it assumed it was safe when using the NO_HOLES feature, since the hole detection was simplified some time ago and no longer was a source for races with ordered extent completion of adjacent file ranges. However it's still not safe to have a full fsync only operate on the given range, because extent maps for new extents might not be present in memory due to inode eviction or extent cloning. Consider the following example: 1) We are currently at transaction N; 2) We write to the file range [0, 1MiB); 3) Writeback finishes for the whole range and ordered extents complete, while we are still at transaction N; 4) The inode is evicted; 5) We open the file for writing, causing the inode to be loaded to memory again, which sets the 'full sync' bit on its flags. At this point the inode's list of modified extent maps is empty (figuring out which extents were created in the current transaction and were not yet logged by an fsync is expensive, that's why we set the 'full sync' bit when loading an inode); 6) We write to the file range [512KiB, 768KiB); 7) We do a ranged fsync (such as msync()) for file range [512KiB, 768KiB). This correctly flushes this range and logs its extent into the log tree. When the writeback started an extent map for range [512KiB, 768KiB) was added to the inode's list of modified extents, and when the fsync() finishes logging it removes that extent map from the list of modified extent maps. This fsync also clears the 'full sync' bit; 8) We do a regular fsync() (full ranged). This fsync() ends up doing nothing because the inode's list of modified extents is empty and no other changes happened since the previous ranged fsync(), so it just returns success (0) and we end up never logging extents for the file ranges [0, 512KiB) and [768KiB, 1MiB). Another scenario where this can happen is if we replace steps 2 to 4 with cloning from another file into our test file, as that sets the 'full sync' bit in our inode's flags and does not populate its list of modified extent maps. This was causing test case generic/457 to fail sporadically when using the NO_HOLES feature, as it exercised this later case where the inode has the 'full sync' bit set and has no extent maps in memory to represent the new extents due to extent cloning. Fix this by reverting commit 0a8068a3dd4294 ("btrfs: make ranged full fsyncs more efficient") since there is no easy way to work around it. Fixes: 0a8068a3dd4294 ("btrfs: make ranged full fsyncs more efficient") Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-04-07 13:37:44 +03:00
const u64 hole_len = key.offset - prev_extent_end;
Btrfs: fix missing hole after hole punching and fsync when using NO_HOLES When using the NO_HOLES feature, if we punch a hole into a file and then fsync it, there are cases where a subsequent fsync will miss the fact that a hole was punched, resulting in the holes not existing after replaying the log tree. Essentially these cases all imply that, tree-log.c:copy_items(), is not invoked for the leafs that delimit holes, because nothing changed those leafs in the current transaction. And it's precisely copy_items() where we currenly detect and log holes, which works as long as the holes are between file extent items in the input leaf or between the beginning of input leaf and the previous leaf or between the last item in the leaf and the next leaf. First example where we miss a hole: *) The extent items of the inode span multiple leafs; *) The punched hole covers a range that affects only the extent items of the first leaf; *) The fsync operation is done in full mode (BTRFS_INODE_NEEDS_FULL_SYNC is set in the inode's runtime flags). That results in the hole not existing after replaying the log tree. For example, if the fs/subvolume tree has the following layout for a particular inode: Leaf N, generation 10: [ ... INODE_ITEM INODE_REF EXTENT_ITEM (0 64K) EXTENT_ITEM (64K 128K) ] Leaf N + 1, generation 10: [ EXTENT_ITEM (128K 64K) ... ] If at transaction 11 we punch a hole coverting the range [0, 128K[, we end up dropping the two extent items from leaf N, but we don't touch the other leaf, so we end up in the following state: Leaf N, generation 11: [ ... INODE_ITEM INODE_REF ] Leaf N + 1, generation 10: [ EXTENT_ITEM (128K 64K) ... ] A full fsync after punching the hole will only process leaf N because it was modified in the current transaction, but not leaf N + 1, since it was not modified in the current transaction (generation 10 and not 11). As a result the fsync will not log any holes, because it didn't process any leaf with extent items. Second example where we will miss a hole: *) An inode as its items spanning 5 (or more) leafs; *) A hole is punched and it covers only the extents items of the 3rd leaf. This resulsts in deleting the entire leaf and not touching any of the other leafs. So the only leaf that is modified in the current transaction, when punching the hole, is the first leaf, which contains the inode item. During the full fsync, the only leaf that is passed to copy_items() is that first leaf, and that's not enough for the hole detection code in copy_items() to determine there's a hole between the last file extent item in the 2nd leaf and the first file extent item in the 3rd leaf (which was the 4th leaf before punching the hole). Fix this by scanning all leafs and punch holes as necessary when doing a full fsync (less common than a non-full fsync) when the NO_HOLES feature is enabled. The lack of explicit file extent items to mark holes makes it necessary to scan existing extents to determine if holes exist. A test case for fstests follows soon. Fixes: 16e7549f045d33 ("Btrfs: incompatible format change to remove hole extents") CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-11-19 15:07:33 +03:00
/*
* Release the path to avoid deadlocks with other code
* paths that search the root while holding locks on
* leafs from the log root.
*/
btrfs_release_path(path);
ret = btrfs_insert_file_extent(trans, root->log_root,
ino, prev_extent_end, 0,
0, hole_len, 0, hole_len,
0, 0, 0);
if (ret < 0)
return ret;
/*
* Search for the same key again in the root. Since it's
* an extent item and we are holding the inode lock, the
* key must still exist. If it doesn't just emit warning
* and return an error to fall back to a transaction
* commit.
*/
ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
if (ret < 0)
return ret;
if (WARN_ON(ret > 0))
return -ENOENT;
leaf = path->nodes[0];
}
Btrfs: fix fsync after truncate when no_holes feature is enabled When we have the no_holes feature enabled, if a we truncate a file to a smaller size, truncate it again but to a size greater than or equals to its original size and fsync it, the log tree will not have any information about the hole covering the range [truncate_1_offset, new_file_size[. Which means if the fsync log is replayed, the file will remain with the state it had before both truncate operations. Without the no_holes feature this does not happen, since when the inode is logged (full sync flag is set) it will find in the fs/subvol tree a leaf with a generation matching the current transaction id that has an explicit extent item representing the hole. Fix this by adding an explicit extent item representing a hole between the last extent and the inode's i_size if we are doing a full sync. The issue is easy to reproduce with the following test case for fstests: . ./common/rc . ./common/filter . ./common/dmflakey _need_to_be_root _supported_fs generic _supported_os Linux _require_scratch _require_dm_flakey # This test was motivated by an issue found in btrfs when the btrfs # no-holes feature is enabled (introduced in kernel 3.14). So enable # the feature if the fs being tested is btrfs. if [ $FSTYP == "btrfs" ]; then _require_btrfs_fs_feature "no_holes" _require_btrfs_mkfs_feature "no-holes" MKFS_OPTIONS="$MKFS_OPTIONS -O no-holes" fi rm -f $seqres.full _scratch_mkfs >>$seqres.full 2>&1 _init_flakey _mount_flakey # Create our test files and make sure everything is durably persisted. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 64K" \ -c "pwrite -S 0xbb 64K 61K" \ $SCRATCH_MNT/foo | _filter_xfs_io $XFS_IO_PROG -f -c "pwrite -S 0xee 0 64K" \ -c "pwrite -S 0xff 64K 61K" \ $SCRATCH_MNT/bar | _filter_xfs_io sync # Now truncate our file foo to a smaller size (64Kb) and then truncate # it to the size it had before the shrinking truncate (125Kb). Then # fsync our file. If a power failure happens after the fsync, we expect # our file to have a size of 125Kb, with the first 64Kb of data having # the value 0xaa and the second 61Kb of data having the value 0x00. $XFS_IO_PROG -c "truncate 64K" \ -c "truncate 125K" \ -c "fsync" \ $SCRATCH_MNT/foo # Do something similar to our file bar, but the first truncation sets # the file size to 0 and the second truncation expands the size to the # double of what it was initially. $XFS_IO_PROG -c "truncate 0" \ -c "truncate 253K" \ -c "fsync" \ $SCRATCH_MNT/bar _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again, mount to trigger log replay and validate file # contents. _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # We expect foo to have a size of 125Kb, the first 64Kb of data all # having the value 0xaa and the remaining 61Kb to be a hole (all bytes # with value 0x00). echo "File foo content after log replay:" od -t x1 $SCRATCH_MNT/foo # We expect bar to have a size of 253Kb and no extents (any byte read # from bar has the value 0x00). echo "File bar content after log replay:" od -t x1 $SCRATCH_MNT/bar status=0 exit The expected file contents in the golden output are: File foo content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0200000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 * 0372000 File bar content after log replay: 0000000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 * 0772000 Without this fix, their contents are: File foo content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0200000 bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb * 0372000 File bar content after log replay: 0000000 ee ee ee ee ee ee ee ee ee ee ee ee ee ee ee ee * 0200000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0372000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 * 0772000 A test case submission for fstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Reviewed-by: Liu Bo <bo.li.liu@oracle.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-06-25 06:17:46 +03:00
btrfs: make full fsyncs always operate on the entire file again This is a revert of commit 0a8068a3dd4294 ("btrfs: make ranged full fsyncs more efficient"), with updated comment in btrfs_sync_file. Commit 0a8068a3dd4294 ("btrfs: make ranged full fsyncs more efficient") made full fsyncs operate on the given range only as it assumed it was safe when using the NO_HOLES feature, since the hole detection was simplified some time ago and no longer was a source for races with ordered extent completion of adjacent file ranges. However it's still not safe to have a full fsync only operate on the given range, because extent maps for new extents might not be present in memory due to inode eviction or extent cloning. Consider the following example: 1) We are currently at transaction N; 2) We write to the file range [0, 1MiB); 3) Writeback finishes for the whole range and ordered extents complete, while we are still at transaction N; 4) The inode is evicted; 5) We open the file for writing, causing the inode to be loaded to memory again, which sets the 'full sync' bit on its flags. At this point the inode's list of modified extent maps is empty (figuring out which extents were created in the current transaction and were not yet logged by an fsync is expensive, that's why we set the 'full sync' bit when loading an inode); 6) We write to the file range [512KiB, 768KiB); 7) We do a ranged fsync (such as msync()) for file range [512KiB, 768KiB). This correctly flushes this range and logs its extent into the log tree. When the writeback started an extent map for range [512KiB, 768KiB) was added to the inode's list of modified extents, and when the fsync() finishes logging it removes that extent map from the list of modified extent maps. This fsync also clears the 'full sync' bit; 8) We do a regular fsync() (full ranged). This fsync() ends up doing nothing because the inode's list of modified extents is empty and no other changes happened since the previous ranged fsync(), so it just returns success (0) and we end up never logging extents for the file ranges [0, 512KiB) and [768KiB, 1MiB). Another scenario where this can happen is if we replace steps 2 to 4 with cloning from another file into our test file, as that sets the 'full sync' bit in our inode's flags and does not populate its list of modified extent maps. This was causing test case generic/457 to fail sporadically when using the NO_HOLES feature, as it exercised this later case where the inode has the 'full sync' bit set and has no extent maps in memory to represent the new extents due to extent cloning. Fix this by reverting commit 0a8068a3dd4294 ("btrfs: make ranged full fsyncs more efficient") since there is no easy way to work around it. Fixes: 0a8068a3dd4294 ("btrfs: make ranged full fsyncs more efficient") Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-04-07 13:37:44 +03:00
prev_extent_end = btrfs_file_extent_end(path);
Btrfs: fix missing hole after hole punching and fsync when using NO_HOLES When using the NO_HOLES feature, if we punch a hole into a file and then fsync it, there are cases where a subsequent fsync will miss the fact that a hole was punched, resulting in the holes not existing after replaying the log tree. Essentially these cases all imply that, tree-log.c:copy_items(), is not invoked for the leafs that delimit holes, because nothing changed those leafs in the current transaction. And it's precisely copy_items() where we currenly detect and log holes, which works as long as the holes are between file extent items in the input leaf or between the beginning of input leaf and the previous leaf or between the last item in the leaf and the next leaf. First example where we miss a hole: *) The extent items of the inode span multiple leafs; *) The punched hole covers a range that affects only the extent items of the first leaf; *) The fsync operation is done in full mode (BTRFS_INODE_NEEDS_FULL_SYNC is set in the inode's runtime flags). That results in the hole not existing after replaying the log tree. For example, if the fs/subvolume tree has the following layout for a particular inode: Leaf N, generation 10: [ ... INODE_ITEM INODE_REF EXTENT_ITEM (0 64K) EXTENT_ITEM (64K 128K) ] Leaf N + 1, generation 10: [ EXTENT_ITEM (128K 64K) ... ] If at transaction 11 we punch a hole coverting the range [0, 128K[, we end up dropping the two extent items from leaf N, but we don't touch the other leaf, so we end up in the following state: Leaf N, generation 11: [ ... INODE_ITEM INODE_REF ] Leaf N + 1, generation 10: [ EXTENT_ITEM (128K 64K) ... ] A full fsync after punching the hole will only process leaf N because it was modified in the current transaction, but not leaf N + 1, since it was not modified in the current transaction (generation 10 and not 11). As a result the fsync will not log any holes, because it didn't process any leaf with extent items. Second example where we will miss a hole: *) An inode as its items spanning 5 (or more) leafs; *) A hole is punched and it covers only the extents items of the 3rd leaf. This resulsts in deleting the entire leaf and not touching any of the other leafs. So the only leaf that is modified in the current transaction, when punching the hole, is the first leaf, which contains the inode item. During the full fsync, the only leaf that is passed to copy_items() is that first leaf, and that's not enough for the hole detection code in copy_items() to determine there's a hole between the last file extent item in the 2nd leaf and the first file extent item in the 3rd leaf (which was the 4th leaf before punching the hole). Fix this by scanning all leafs and punch holes as necessary when doing a full fsync (less common than a non-full fsync) when the NO_HOLES feature is enabled. The lack of explicit file extent items to mark holes makes it necessary to scan existing extents to determine if holes exist. A test case for fstests follows soon. Fixes: 16e7549f045d33 ("Btrfs: incompatible format change to remove hole extents") CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-11-19 15:07:33 +03:00
path->slots[0]++;
cond_resched();
Btrfs: fix fsync after truncate when no_holes feature is enabled When we have the no_holes feature enabled, if a we truncate a file to a smaller size, truncate it again but to a size greater than or equals to its original size and fsync it, the log tree will not have any information about the hole covering the range [truncate_1_offset, new_file_size[. Which means if the fsync log is replayed, the file will remain with the state it had before both truncate operations. Without the no_holes feature this does not happen, since when the inode is logged (full sync flag is set) it will find in the fs/subvol tree a leaf with a generation matching the current transaction id that has an explicit extent item representing the hole. Fix this by adding an explicit extent item representing a hole between the last extent and the inode's i_size if we are doing a full sync. The issue is easy to reproduce with the following test case for fstests: . ./common/rc . ./common/filter . ./common/dmflakey _need_to_be_root _supported_fs generic _supported_os Linux _require_scratch _require_dm_flakey # This test was motivated by an issue found in btrfs when the btrfs # no-holes feature is enabled (introduced in kernel 3.14). So enable # the feature if the fs being tested is btrfs. if [ $FSTYP == "btrfs" ]; then _require_btrfs_fs_feature "no_holes" _require_btrfs_mkfs_feature "no-holes" MKFS_OPTIONS="$MKFS_OPTIONS -O no-holes" fi rm -f $seqres.full _scratch_mkfs >>$seqres.full 2>&1 _init_flakey _mount_flakey # Create our test files and make sure everything is durably persisted. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 64K" \ -c "pwrite -S 0xbb 64K 61K" \ $SCRATCH_MNT/foo | _filter_xfs_io $XFS_IO_PROG -f -c "pwrite -S 0xee 0 64K" \ -c "pwrite -S 0xff 64K 61K" \ $SCRATCH_MNT/bar | _filter_xfs_io sync # Now truncate our file foo to a smaller size (64Kb) and then truncate # it to the size it had before the shrinking truncate (125Kb). Then # fsync our file. If a power failure happens after the fsync, we expect # our file to have a size of 125Kb, with the first 64Kb of data having # the value 0xaa and the second 61Kb of data having the value 0x00. $XFS_IO_PROG -c "truncate 64K" \ -c "truncate 125K" \ -c "fsync" \ $SCRATCH_MNT/foo # Do something similar to our file bar, but the first truncation sets # the file size to 0 and the second truncation expands the size to the # double of what it was initially. $XFS_IO_PROG -c "truncate 0" \ -c "truncate 253K" \ -c "fsync" \ $SCRATCH_MNT/bar _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again, mount to trigger log replay and validate file # contents. _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # We expect foo to have a size of 125Kb, the first 64Kb of data all # having the value 0xaa and the remaining 61Kb to be a hole (all bytes # with value 0x00). echo "File foo content after log replay:" od -t x1 $SCRATCH_MNT/foo # We expect bar to have a size of 253Kb and no extents (any byte read # from bar has the value 0x00). echo "File bar content after log replay:" od -t x1 $SCRATCH_MNT/bar status=0 exit The expected file contents in the golden output are: File foo content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0200000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 * 0372000 File bar content after log replay: 0000000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 * 0772000 Without this fix, their contents are: File foo content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0200000 bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb * 0372000 File bar content after log replay: 0000000 ee ee ee ee ee ee ee ee ee ee ee ee ee ee ee ee * 0200000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0372000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 * 0772000 A test case submission for fstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Reviewed-by: Liu Bo <bo.li.liu@oracle.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-06-25 06:17:46 +03:00
}
btrfs: make full fsyncs always operate on the entire file again This is a revert of commit 0a8068a3dd4294 ("btrfs: make ranged full fsyncs more efficient"), with updated comment in btrfs_sync_file. Commit 0a8068a3dd4294 ("btrfs: make ranged full fsyncs more efficient") made full fsyncs operate on the given range only as it assumed it was safe when using the NO_HOLES feature, since the hole detection was simplified some time ago and no longer was a source for races with ordered extent completion of adjacent file ranges. However it's still not safe to have a full fsync only operate on the given range, because extent maps for new extents might not be present in memory due to inode eviction or extent cloning. Consider the following example: 1) We are currently at transaction N; 2) We write to the file range [0, 1MiB); 3) Writeback finishes for the whole range and ordered extents complete, while we are still at transaction N; 4) The inode is evicted; 5) We open the file for writing, causing the inode to be loaded to memory again, which sets the 'full sync' bit on its flags. At this point the inode's list of modified extent maps is empty (figuring out which extents were created in the current transaction and were not yet logged by an fsync is expensive, that's why we set the 'full sync' bit when loading an inode); 6) We write to the file range [512KiB, 768KiB); 7) We do a ranged fsync (such as msync()) for file range [512KiB, 768KiB). This correctly flushes this range and logs its extent into the log tree. When the writeback started an extent map for range [512KiB, 768KiB) was added to the inode's list of modified extents, and when the fsync() finishes logging it removes that extent map from the list of modified extent maps. This fsync also clears the 'full sync' bit; 8) We do a regular fsync() (full ranged). This fsync() ends up doing nothing because the inode's list of modified extents is empty and no other changes happened since the previous ranged fsync(), so it just returns success (0) and we end up never logging extents for the file ranges [0, 512KiB) and [768KiB, 1MiB). Another scenario where this can happen is if we replace steps 2 to 4 with cloning from another file into our test file, as that sets the 'full sync' bit in our inode's flags and does not populate its list of modified extent maps. This was causing test case generic/457 to fail sporadically when using the NO_HOLES feature, as it exercised this later case where the inode has the 'full sync' bit set and has no extent maps in memory to represent the new extents due to extent cloning. Fix this by reverting commit 0a8068a3dd4294 ("btrfs: make ranged full fsyncs more efficient") since there is no easy way to work around it. Fixes: 0a8068a3dd4294 ("btrfs: make ranged full fsyncs more efficient") Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-04-07 13:37:44 +03:00
if (prev_extent_end < i_size) {
Btrfs: fix missing hole after hole punching and fsync when using NO_HOLES When using the NO_HOLES feature, if we punch a hole into a file and then fsync it, there are cases where a subsequent fsync will miss the fact that a hole was punched, resulting in the holes not existing after replaying the log tree. Essentially these cases all imply that, tree-log.c:copy_items(), is not invoked for the leafs that delimit holes, because nothing changed those leafs in the current transaction. And it's precisely copy_items() where we currenly detect and log holes, which works as long as the holes are between file extent items in the input leaf or between the beginning of input leaf and the previous leaf or between the last item in the leaf and the next leaf. First example where we miss a hole: *) The extent items of the inode span multiple leafs; *) The punched hole covers a range that affects only the extent items of the first leaf; *) The fsync operation is done in full mode (BTRFS_INODE_NEEDS_FULL_SYNC is set in the inode's runtime flags). That results in the hole not existing after replaying the log tree. For example, if the fs/subvolume tree has the following layout for a particular inode: Leaf N, generation 10: [ ... INODE_ITEM INODE_REF EXTENT_ITEM (0 64K) EXTENT_ITEM (64K 128K) ] Leaf N + 1, generation 10: [ EXTENT_ITEM (128K 64K) ... ] If at transaction 11 we punch a hole coverting the range [0, 128K[, we end up dropping the two extent items from leaf N, but we don't touch the other leaf, so we end up in the following state: Leaf N, generation 11: [ ... INODE_ITEM INODE_REF ] Leaf N + 1, generation 10: [ EXTENT_ITEM (128K 64K) ... ] A full fsync after punching the hole will only process leaf N because it was modified in the current transaction, but not leaf N + 1, since it was not modified in the current transaction (generation 10 and not 11). As a result the fsync will not log any holes, because it didn't process any leaf with extent items. Second example where we will miss a hole: *) An inode as its items spanning 5 (or more) leafs; *) A hole is punched and it covers only the extents items of the 3rd leaf. This resulsts in deleting the entire leaf and not touching any of the other leafs. So the only leaf that is modified in the current transaction, when punching the hole, is the first leaf, which contains the inode item. During the full fsync, the only leaf that is passed to copy_items() is that first leaf, and that's not enough for the hole detection code in copy_items() to determine there's a hole between the last file extent item in the 2nd leaf and the first file extent item in the 3rd leaf (which was the 4th leaf before punching the hole). Fix this by scanning all leafs and punch holes as necessary when doing a full fsync (less common than a non-full fsync) when the NO_HOLES feature is enabled. The lack of explicit file extent items to mark holes makes it necessary to scan existing extents to determine if holes exist. A test case for fstests follows soon. Fixes: 16e7549f045d33 ("Btrfs: incompatible format change to remove hole extents") CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-11-19 15:07:33 +03:00
u64 hole_len;
Btrfs: fix fsync after truncate when no_holes feature is enabled When we have the no_holes feature enabled, if a we truncate a file to a smaller size, truncate it again but to a size greater than or equals to its original size and fsync it, the log tree will not have any information about the hole covering the range [truncate_1_offset, new_file_size[. Which means if the fsync log is replayed, the file will remain with the state it had before both truncate operations. Without the no_holes feature this does not happen, since when the inode is logged (full sync flag is set) it will find in the fs/subvol tree a leaf with a generation matching the current transaction id that has an explicit extent item representing the hole. Fix this by adding an explicit extent item representing a hole between the last extent and the inode's i_size if we are doing a full sync. The issue is easy to reproduce with the following test case for fstests: . ./common/rc . ./common/filter . ./common/dmflakey _need_to_be_root _supported_fs generic _supported_os Linux _require_scratch _require_dm_flakey # This test was motivated by an issue found in btrfs when the btrfs # no-holes feature is enabled (introduced in kernel 3.14). So enable # the feature if the fs being tested is btrfs. if [ $FSTYP == "btrfs" ]; then _require_btrfs_fs_feature "no_holes" _require_btrfs_mkfs_feature "no-holes" MKFS_OPTIONS="$MKFS_OPTIONS -O no-holes" fi rm -f $seqres.full _scratch_mkfs >>$seqres.full 2>&1 _init_flakey _mount_flakey # Create our test files and make sure everything is durably persisted. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 64K" \ -c "pwrite -S 0xbb 64K 61K" \ $SCRATCH_MNT/foo | _filter_xfs_io $XFS_IO_PROG -f -c "pwrite -S 0xee 0 64K" \ -c "pwrite -S 0xff 64K 61K" \ $SCRATCH_MNT/bar | _filter_xfs_io sync # Now truncate our file foo to a smaller size (64Kb) and then truncate # it to the size it had before the shrinking truncate (125Kb). Then # fsync our file. If a power failure happens after the fsync, we expect # our file to have a size of 125Kb, with the first 64Kb of data having # the value 0xaa and the second 61Kb of data having the value 0x00. $XFS_IO_PROG -c "truncate 64K" \ -c "truncate 125K" \ -c "fsync" \ $SCRATCH_MNT/foo # Do something similar to our file bar, but the first truncation sets # the file size to 0 and the second truncation expands the size to the # double of what it was initially. $XFS_IO_PROG -c "truncate 0" \ -c "truncate 253K" \ -c "fsync" \ $SCRATCH_MNT/bar _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again, mount to trigger log replay and validate file # contents. _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # We expect foo to have a size of 125Kb, the first 64Kb of data all # having the value 0xaa and the remaining 61Kb to be a hole (all bytes # with value 0x00). echo "File foo content after log replay:" od -t x1 $SCRATCH_MNT/foo # We expect bar to have a size of 253Kb and no extents (any byte read # from bar has the value 0x00). echo "File bar content after log replay:" od -t x1 $SCRATCH_MNT/bar status=0 exit The expected file contents in the golden output are: File foo content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0200000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 * 0372000 File bar content after log replay: 0000000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 * 0772000 Without this fix, their contents are: File foo content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0200000 bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb * 0372000 File bar content after log replay: 0000000 ee ee ee ee ee ee ee ee ee ee ee ee ee ee ee ee * 0200000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0372000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 * 0772000 A test case submission for fstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Reviewed-by: Liu Bo <bo.li.liu@oracle.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-06-25 06:17:46 +03:00
Btrfs: fix missing hole after hole punching and fsync when using NO_HOLES When using the NO_HOLES feature, if we punch a hole into a file and then fsync it, there are cases where a subsequent fsync will miss the fact that a hole was punched, resulting in the holes not existing after replaying the log tree. Essentially these cases all imply that, tree-log.c:copy_items(), is not invoked for the leafs that delimit holes, because nothing changed those leafs in the current transaction. And it's precisely copy_items() where we currenly detect and log holes, which works as long as the holes are between file extent items in the input leaf or between the beginning of input leaf and the previous leaf or between the last item in the leaf and the next leaf. First example where we miss a hole: *) The extent items of the inode span multiple leafs; *) The punched hole covers a range that affects only the extent items of the first leaf; *) The fsync operation is done in full mode (BTRFS_INODE_NEEDS_FULL_SYNC is set in the inode's runtime flags). That results in the hole not existing after replaying the log tree. For example, if the fs/subvolume tree has the following layout for a particular inode: Leaf N, generation 10: [ ... INODE_ITEM INODE_REF EXTENT_ITEM (0 64K) EXTENT_ITEM (64K 128K) ] Leaf N + 1, generation 10: [ EXTENT_ITEM (128K 64K) ... ] If at transaction 11 we punch a hole coverting the range [0, 128K[, we end up dropping the two extent items from leaf N, but we don't touch the other leaf, so we end up in the following state: Leaf N, generation 11: [ ... INODE_ITEM INODE_REF ] Leaf N + 1, generation 10: [ EXTENT_ITEM (128K 64K) ... ] A full fsync after punching the hole will only process leaf N because it was modified in the current transaction, but not leaf N + 1, since it was not modified in the current transaction (generation 10 and not 11). As a result the fsync will not log any holes, because it didn't process any leaf with extent items. Second example where we will miss a hole: *) An inode as its items spanning 5 (or more) leafs; *) A hole is punched and it covers only the extents items of the 3rd leaf. This resulsts in deleting the entire leaf and not touching any of the other leafs. So the only leaf that is modified in the current transaction, when punching the hole, is the first leaf, which contains the inode item. During the full fsync, the only leaf that is passed to copy_items() is that first leaf, and that's not enough for the hole detection code in copy_items() to determine there's a hole between the last file extent item in the 2nd leaf and the first file extent item in the 3rd leaf (which was the 4th leaf before punching the hole). Fix this by scanning all leafs and punch holes as necessary when doing a full fsync (less common than a non-full fsync) when the NO_HOLES feature is enabled. The lack of explicit file extent items to mark holes makes it necessary to scan existing extents to determine if holes exist. A test case for fstests follows soon. Fixes: 16e7549f045d33 ("Btrfs: incompatible format change to remove hole extents") CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-11-19 15:07:33 +03:00
btrfs_release_path(path);
btrfs: make full fsyncs always operate on the entire file again This is a revert of commit 0a8068a3dd4294 ("btrfs: make ranged full fsyncs more efficient"), with updated comment in btrfs_sync_file. Commit 0a8068a3dd4294 ("btrfs: make ranged full fsyncs more efficient") made full fsyncs operate on the given range only as it assumed it was safe when using the NO_HOLES feature, since the hole detection was simplified some time ago and no longer was a source for races with ordered extent completion of adjacent file ranges. However it's still not safe to have a full fsync only operate on the given range, because extent maps for new extents might not be present in memory due to inode eviction or extent cloning. Consider the following example: 1) We are currently at transaction N; 2) We write to the file range [0, 1MiB); 3) Writeback finishes for the whole range and ordered extents complete, while we are still at transaction N; 4) The inode is evicted; 5) We open the file for writing, causing the inode to be loaded to memory again, which sets the 'full sync' bit on its flags. At this point the inode's list of modified extent maps is empty (figuring out which extents were created in the current transaction and were not yet logged by an fsync is expensive, that's why we set the 'full sync' bit when loading an inode); 6) We write to the file range [512KiB, 768KiB); 7) We do a ranged fsync (such as msync()) for file range [512KiB, 768KiB). This correctly flushes this range and logs its extent into the log tree. When the writeback started an extent map for range [512KiB, 768KiB) was added to the inode's list of modified extents, and when the fsync() finishes logging it removes that extent map from the list of modified extent maps. This fsync also clears the 'full sync' bit; 8) We do a regular fsync() (full ranged). This fsync() ends up doing nothing because the inode's list of modified extents is empty and no other changes happened since the previous ranged fsync(), so it just returns success (0) and we end up never logging extents for the file ranges [0, 512KiB) and [768KiB, 1MiB). Another scenario where this can happen is if we replace steps 2 to 4 with cloning from another file into our test file, as that sets the 'full sync' bit in our inode's flags and does not populate its list of modified extent maps. This was causing test case generic/457 to fail sporadically when using the NO_HOLES feature, as it exercised this later case where the inode has the 'full sync' bit set and has no extent maps in memory to represent the new extents due to extent cloning. Fix this by reverting commit 0a8068a3dd4294 ("btrfs: make ranged full fsyncs more efficient") since there is no easy way to work around it. Fixes: 0a8068a3dd4294 ("btrfs: make ranged full fsyncs more efficient") Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-04-07 13:37:44 +03:00
hole_len = ALIGN(i_size - prev_extent_end, fs_info->sectorsize);
Btrfs: fix missing hole after hole punching and fsync when using NO_HOLES When using the NO_HOLES feature, if we punch a hole into a file and then fsync it, there are cases where a subsequent fsync will miss the fact that a hole was punched, resulting in the holes not existing after replaying the log tree. Essentially these cases all imply that, tree-log.c:copy_items(), is not invoked for the leafs that delimit holes, because nothing changed those leafs in the current transaction. And it's precisely copy_items() where we currenly detect and log holes, which works as long as the holes are between file extent items in the input leaf or between the beginning of input leaf and the previous leaf or between the last item in the leaf and the next leaf. First example where we miss a hole: *) The extent items of the inode span multiple leafs; *) The punched hole covers a range that affects only the extent items of the first leaf; *) The fsync operation is done in full mode (BTRFS_INODE_NEEDS_FULL_SYNC is set in the inode's runtime flags). That results in the hole not existing after replaying the log tree. For example, if the fs/subvolume tree has the following layout for a particular inode: Leaf N, generation 10: [ ... INODE_ITEM INODE_REF EXTENT_ITEM (0 64K) EXTENT_ITEM (64K 128K) ] Leaf N + 1, generation 10: [ EXTENT_ITEM (128K 64K) ... ] If at transaction 11 we punch a hole coverting the range [0, 128K[, we end up dropping the two extent items from leaf N, but we don't touch the other leaf, so we end up in the following state: Leaf N, generation 11: [ ... INODE_ITEM INODE_REF ] Leaf N + 1, generation 10: [ EXTENT_ITEM (128K 64K) ... ] A full fsync after punching the hole will only process leaf N because it was modified in the current transaction, but not leaf N + 1, since it was not modified in the current transaction (generation 10 and not 11). As a result the fsync will not log any holes, because it didn't process any leaf with extent items. Second example where we will miss a hole: *) An inode as its items spanning 5 (or more) leafs; *) A hole is punched and it covers only the extents items of the 3rd leaf. This resulsts in deleting the entire leaf and not touching any of the other leafs. So the only leaf that is modified in the current transaction, when punching the hole, is the first leaf, which contains the inode item. During the full fsync, the only leaf that is passed to copy_items() is that first leaf, and that's not enough for the hole detection code in copy_items() to determine there's a hole between the last file extent item in the 2nd leaf and the first file extent item in the 3rd leaf (which was the 4th leaf before punching the hole). Fix this by scanning all leafs and punch holes as necessary when doing a full fsync (less common than a non-full fsync) when the NO_HOLES feature is enabled. The lack of explicit file extent items to mark holes makes it necessary to scan existing extents to determine if holes exist. A test case for fstests follows soon. Fixes: 16e7549f045d33 ("Btrfs: incompatible format change to remove hole extents") CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-11-19 15:07:33 +03:00
ret = btrfs_insert_file_extent(trans, root->log_root,
ino, prev_extent_end, 0, 0,
hole_len, 0, hole_len,
0, 0, 0);
if (ret < 0)
return ret;
}
return 0;
Btrfs: fix fsync after truncate when no_holes feature is enabled When we have the no_holes feature enabled, if a we truncate a file to a smaller size, truncate it again but to a size greater than or equals to its original size and fsync it, the log tree will not have any information about the hole covering the range [truncate_1_offset, new_file_size[. Which means if the fsync log is replayed, the file will remain with the state it had before both truncate operations. Without the no_holes feature this does not happen, since when the inode is logged (full sync flag is set) it will find in the fs/subvol tree a leaf with a generation matching the current transaction id that has an explicit extent item representing the hole. Fix this by adding an explicit extent item representing a hole between the last extent and the inode's i_size if we are doing a full sync. The issue is easy to reproduce with the following test case for fstests: . ./common/rc . ./common/filter . ./common/dmflakey _need_to_be_root _supported_fs generic _supported_os Linux _require_scratch _require_dm_flakey # This test was motivated by an issue found in btrfs when the btrfs # no-holes feature is enabled (introduced in kernel 3.14). So enable # the feature if the fs being tested is btrfs. if [ $FSTYP == "btrfs" ]; then _require_btrfs_fs_feature "no_holes" _require_btrfs_mkfs_feature "no-holes" MKFS_OPTIONS="$MKFS_OPTIONS -O no-holes" fi rm -f $seqres.full _scratch_mkfs >>$seqres.full 2>&1 _init_flakey _mount_flakey # Create our test files and make sure everything is durably persisted. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 64K" \ -c "pwrite -S 0xbb 64K 61K" \ $SCRATCH_MNT/foo | _filter_xfs_io $XFS_IO_PROG -f -c "pwrite -S 0xee 0 64K" \ -c "pwrite -S 0xff 64K 61K" \ $SCRATCH_MNT/bar | _filter_xfs_io sync # Now truncate our file foo to a smaller size (64Kb) and then truncate # it to the size it had before the shrinking truncate (125Kb). Then # fsync our file. If a power failure happens after the fsync, we expect # our file to have a size of 125Kb, with the first 64Kb of data having # the value 0xaa and the second 61Kb of data having the value 0x00. $XFS_IO_PROG -c "truncate 64K" \ -c "truncate 125K" \ -c "fsync" \ $SCRATCH_MNT/foo # Do something similar to our file bar, but the first truncation sets # the file size to 0 and the second truncation expands the size to the # double of what it was initially. $XFS_IO_PROG -c "truncate 0" \ -c "truncate 253K" \ -c "fsync" \ $SCRATCH_MNT/bar _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again, mount to trigger log replay and validate file # contents. _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # We expect foo to have a size of 125Kb, the first 64Kb of data all # having the value 0xaa and the remaining 61Kb to be a hole (all bytes # with value 0x00). echo "File foo content after log replay:" od -t x1 $SCRATCH_MNT/foo # We expect bar to have a size of 253Kb and no extents (any byte read # from bar has the value 0x00). echo "File bar content after log replay:" od -t x1 $SCRATCH_MNT/bar status=0 exit The expected file contents in the golden output are: File foo content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0200000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 * 0372000 File bar content after log replay: 0000000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 * 0772000 Without this fix, their contents are: File foo content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0200000 bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb * 0372000 File bar content after log replay: 0000000 ee ee ee ee ee ee ee ee ee ee ee ee ee ee ee ee * 0200000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0372000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 * 0772000 A test case submission for fstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Reviewed-by: Liu Bo <bo.li.liu@oracle.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-06-25 06:17:46 +03:00
}
Btrfs: fix file/data loss caused by fsync after rename and new inode If we rename an inode A (be it a file or a directory), create a new inode B with the old name of inode A and under the same parent directory, fsync inode B and then power fail, at log tree replay time we end up removing inode A completely. If inode A is a directory then all its files are gone too. Example scenarios where this happens: This is reproducible with the following steps, taken from a couple of test cases written for fstests which are going to be submitted upstream soon: # Scenario 1 mkfs.btrfs -f /dev/sdc mount /dev/sdc /mnt mkdir -p /mnt/a/x echo "hello" > /mnt/a/x/foo echo "world" > /mnt/a/x/bar sync mv /mnt/a/x /mnt/a/y mkdir /mnt/a/x xfs_io -c fsync /mnt/a/x <power failure happens> The next time the fs is mounted, log tree replay happens and the directory "y" does not exist nor do the files "foo" and "bar" exist anywhere (neither in "y" nor in "x", nor the root nor anywhere). # Scenario 2 mkfs.btrfs -f /dev/sdc mount /dev/sdc /mnt mkdir /mnt/a echo "hello" > /mnt/a/foo sync mv /mnt/a/foo /mnt/a/bar echo "world" > /mnt/a/foo xfs_io -c fsync /mnt/a/foo <power failure happens> The next time the fs is mounted, log tree replay happens and the file "bar" does not exists anymore. A file with the name "foo" exists and it matches the second file we created. Another related problem that does not involve file/data loss is when a new inode is created with the name of a deleted snapshot and we fsync it: mkfs.btrfs -f /dev/sdc mount /dev/sdc /mnt mkdir /mnt/testdir btrfs subvolume snapshot /mnt /mnt/testdir/snap btrfs subvolume delete /mnt/testdir/snap rmdir /mnt/testdir mkdir /mnt/testdir xfs_io -c fsync /mnt/testdir # or fsync some file inside /mnt/testdir <power failure> The next time the fs is mounted the log replay procedure fails because it attempts to delete the snapshot entry (which has dir item key type of BTRFS_ROOT_ITEM_KEY) as if it were a regular (non-root) entry, resulting in the following error that causes mount to fail: [52174.510532] BTRFS info (device dm-0): failed to delete reference to snap, inode 257 parent 257 [52174.512570] ------------[ cut here ]------------ [52174.513278] WARNING: CPU: 12 PID: 28024 at fs/btrfs/inode.c:3986 __btrfs_unlink_inode+0x178/0x351 [btrfs]() [52174.514681] BTRFS: Transaction aborted (error -2) [52174.515630] Modules linked in: btrfs dm_flakey dm_mod overlay crc32c_generic ppdev xor raid6_pq acpi_cpufreq parport_pc tpm_tis sg parport tpm evdev i2c_piix4 proc [52174.521568] CPU: 12 PID: 28024 Comm: mount Tainted: G W 4.5.0-rc6-btrfs-next-27+ #1 [52174.522805] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS by qemu-project.org 04/01/2014 [52174.524053] 0000000000000000 ffff8801df2a7710 ffffffff81264e93 ffff8801df2a7758 [52174.524053] 0000000000000009 ffff8801df2a7748 ffffffff81051618 ffffffffa03591cd [52174.524053] 00000000fffffffe ffff88015e6e5000 ffff88016dbc3c88 ffff88016dbc3c88 [52174.524053] Call Trace: [52174.524053] [<ffffffff81264e93>] dump_stack+0x67/0x90 [52174.524053] [<ffffffff81051618>] warn_slowpath_common+0x99/0xb2 [52174.524053] [<ffffffffa03591cd>] ? __btrfs_unlink_inode+0x178/0x351 [btrfs] [52174.524053] [<ffffffff81051679>] warn_slowpath_fmt+0x48/0x50 [52174.524053] [<ffffffffa03591cd>] __btrfs_unlink_inode+0x178/0x351 [btrfs] [52174.524053] [<ffffffff8118f5e9>] ? iput+0xb0/0x284 [52174.524053] [<ffffffffa0359fe8>] btrfs_unlink_inode+0x1c/0x3d [btrfs] [52174.524053] [<ffffffffa038631e>] check_item_in_log+0x1fe/0x29b [btrfs] [52174.524053] [<ffffffffa0386522>] replay_dir_deletes+0x167/0x1cf [btrfs] [52174.524053] [<ffffffffa038739e>] fixup_inode_link_count+0x289/0x2aa [btrfs] [52174.524053] [<ffffffffa038748a>] fixup_inode_link_counts+0xcb/0x105 [btrfs] [52174.524053] [<ffffffffa038a5ec>] btrfs_recover_log_trees+0x258/0x32c [btrfs] [52174.524053] [<ffffffffa03885b2>] ? replay_one_extent+0x511/0x511 [btrfs] [52174.524053] [<ffffffffa034f288>] open_ctree+0x1dd4/0x21b9 [btrfs] [52174.524053] [<ffffffffa032b753>] btrfs_mount+0x97e/0xaed [btrfs] [52174.524053] [<ffffffff8108e1b7>] ? trace_hardirqs_on+0xd/0xf [52174.524053] [<ffffffff8117bafa>] mount_fs+0x67/0x131 [52174.524053] [<ffffffff81193003>] vfs_kern_mount+0x6c/0xde [52174.524053] [<ffffffffa032af81>] btrfs_mount+0x1ac/0xaed [btrfs] [52174.524053] [<ffffffff8108e1b7>] ? trace_hardirqs_on+0xd/0xf [52174.524053] [<ffffffff8108c262>] ? lockdep_init_map+0xb9/0x1b3 [52174.524053] [<ffffffff8117bafa>] mount_fs+0x67/0x131 [52174.524053] [<ffffffff81193003>] vfs_kern_mount+0x6c/0xde [52174.524053] [<ffffffff8119590f>] do_mount+0x8a6/0x9e8 [52174.524053] [<ffffffff811358dd>] ? strndup_user+0x3f/0x59 [52174.524053] [<ffffffff81195c65>] SyS_mount+0x77/0x9f [52174.524053] [<ffffffff814935d7>] entry_SYSCALL_64_fastpath+0x12/0x6b [52174.561288] ---[ end trace 6b53049efb1a3ea6 ]--- Fix this by forcing a transaction commit when such cases happen. This means we check in the commit root of the subvolume tree if there was any other inode with the same reference when the inode we are fsync'ing is a new inode (created in the current transaction). Test cases for fstests, covering all the scenarios given above, were submitted upstream for fstests: * fstests: generic test for fsync after renaming directory https://patchwork.kernel.org/patch/8694281/ * fstests: generic test for fsync after renaming file https://patchwork.kernel.org/patch/8694301/ * fstests: add btrfs test for fsync after snapshot deletion https://patchwork.kernel.org/patch/8670671/ Cc: stable@vger.kernel.org Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2016-03-31 01:37:21 +03:00
/*
* When we are logging a new inode X, check if it doesn't have a reference that
* matches the reference from some other inode Y created in a past transaction
* and that was renamed in the current transaction. If we don't do this, then at
* log replay time we can lose inode Y (and all its files if it's a directory):
*
* mkdir /mnt/x
* echo "hello world" > /mnt/x/foobar
* sync
* mv /mnt/x /mnt/y
* mkdir /mnt/x # or touch /mnt/x
* xfs_io -c fsync /mnt/x
* <power fail>
* mount fs, trigger log replay
*
* After the log replay procedure, we would lose the first directory and all its
* files (file foobar).
* For the case where inode Y is not a directory we simply end up losing it:
*
* echo "123" > /mnt/foo
* sync
* mv /mnt/foo /mnt/bar
* echo "abc" > /mnt/foo
* xfs_io -c fsync /mnt/foo
* <power fail>
*
* We also need this for cases where a snapshot entry is replaced by some other
* entry (file or directory) otherwise we end up with an unreplayable log due to
* attempts to delete the snapshot entry (entry of type BTRFS_ROOT_ITEM_KEY) as
* if it were a regular entry:
*
* mkdir /mnt/x
* btrfs subvolume snapshot /mnt /mnt/x/snap
* btrfs subvolume delete /mnt/x/snap
* rmdir /mnt/x
* mkdir /mnt/x
* fsync /mnt/x or fsync some new file inside it
* <power fail>
*
* The snapshot delete, rmdir of x, mkdir of a new x and the fsync all happen in
* the same transaction.
*/
static int btrfs_check_ref_name_override(struct extent_buffer *eb,
const int slot,
const struct btrfs_key *key,
struct btrfs_inode *inode,
Btrfs: fix fsync after succession of renames and unlink/rmdir After a succession of renames operations of different files and unlinking one of them, if we fsync one of the renamed files we can end up with a log that will either fail to replay at mount time or result in a filesystem that is in an inconsistent state. One example scenario: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/fname1 $ touch /mnt/testdir/fname2 $ sync $ mv /mnt/testdir/fname1 /mnt/testdir/fname3 $ rm -f /mnt/testdir/fname2 $ ln /mnt/testdir/fname3 /mnt/testdir/fname2 $ touch /mnt/testdir/fname1 $ xfs_io -c "fsync" /mnt/testdir/fname1 <power failure> $ mount /dev/sdb /mnt $ umount /mnt $ btrfs check /dev/sdb [1/7] checking root items [2/7] checking extents [3/7] checking free space cache [4/7] checking fs roots root 5 inode 259 errors 2, no orphan item ERROR: errors found in fs roots Opening filesystem to check... Checking filesystem on /dev/sdc UUID: 20e4abb8-5a19-4492-8bb4-6084125c2d0d found 393216 bytes used, error(s) found total csum bytes: 0 total tree bytes: 131072 total fs tree bytes: 32768 total extent tree bytes: 16384 btree space waste bytes: 122986 file data blocks allocated: 262144 referenced 262144 On a kernel without the first patch in this series, titled "[PATCH] Btrfs: fix fsync after succession of renames of different files", we get instead an error when mounting the filesystem due to failure of replaying the log: $ mount /dev/sdb /mnt mount: mount /dev/sdb on /mnt failed: File exists Fix this by logging the parent directory of an inode whenever we find an inode that no longer exists (was unlinked in the current transaction), during the procedure which finds inodes that have old names that collide with new names of other inodes. A test case for fstests follows soon. Reviewed-by: Nikolay Borisov <nborisov@suse.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-02-13 15:14:09 +03:00
u64 *other_ino, u64 *other_parent)
Btrfs: fix file/data loss caused by fsync after rename and new inode If we rename an inode A (be it a file or a directory), create a new inode B with the old name of inode A and under the same parent directory, fsync inode B and then power fail, at log tree replay time we end up removing inode A completely. If inode A is a directory then all its files are gone too. Example scenarios where this happens: This is reproducible with the following steps, taken from a couple of test cases written for fstests which are going to be submitted upstream soon: # Scenario 1 mkfs.btrfs -f /dev/sdc mount /dev/sdc /mnt mkdir -p /mnt/a/x echo "hello" > /mnt/a/x/foo echo "world" > /mnt/a/x/bar sync mv /mnt/a/x /mnt/a/y mkdir /mnt/a/x xfs_io -c fsync /mnt/a/x <power failure happens> The next time the fs is mounted, log tree replay happens and the directory "y" does not exist nor do the files "foo" and "bar" exist anywhere (neither in "y" nor in "x", nor the root nor anywhere). # Scenario 2 mkfs.btrfs -f /dev/sdc mount /dev/sdc /mnt mkdir /mnt/a echo "hello" > /mnt/a/foo sync mv /mnt/a/foo /mnt/a/bar echo "world" > /mnt/a/foo xfs_io -c fsync /mnt/a/foo <power failure happens> The next time the fs is mounted, log tree replay happens and the file "bar" does not exists anymore. A file with the name "foo" exists and it matches the second file we created. Another related problem that does not involve file/data loss is when a new inode is created with the name of a deleted snapshot and we fsync it: mkfs.btrfs -f /dev/sdc mount /dev/sdc /mnt mkdir /mnt/testdir btrfs subvolume snapshot /mnt /mnt/testdir/snap btrfs subvolume delete /mnt/testdir/snap rmdir /mnt/testdir mkdir /mnt/testdir xfs_io -c fsync /mnt/testdir # or fsync some file inside /mnt/testdir <power failure> The next time the fs is mounted the log replay procedure fails because it attempts to delete the snapshot entry (which has dir item key type of BTRFS_ROOT_ITEM_KEY) as if it were a regular (non-root) entry, resulting in the following error that causes mount to fail: [52174.510532] BTRFS info (device dm-0): failed to delete reference to snap, inode 257 parent 257 [52174.512570] ------------[ cut here ]------------ [52174.513278] WARNING: CPU: 12 PID: 28024 at fs/btrfs/inode.c:3986 __btrfs_unlink_inode+0x178/0x351 [btrfs]() [52174.514681] BTRFS: Transaction aborted (error -2) [52174.515630] Modules linked in: btrfs dm_flakey dm_mod overlay crc32c_generic ppdev xor raid6_pq acpi_cpufreq parport_pc tpm_tis sg parport tpm evdev i2c_piix4 proc [52174.521568] CPU: 12 PID: 28024 Comm: mount Tainted: G W 4.5.0-rc6-btrfs-next-27+ #1 [52174.522805] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS by qemu-project.org 04/01/2014 [52174.524053] 0000000000000000 ffff8801df2a7710 ffffffff81264e93 ffff8801df2a7758 [52174.524053] 0000000000000009 ffff8801df2a7748 ffffffff81051618 ffffffffa03591cd [52174.524053] 00000000fffffffe ffff88015e6e5000 ffff88016dbc3c88 ffff88016dbc3c88 [52174.524053] Call Trace: [52174.524053] [<ffffffff81264e93>] dump_stack+0x67/0x90 [52174.524053] [<ffffffff81051618>] warn_slowpath_common+0x99/0xb2 [52174.524053] [<ffffffffa03591cd>] ? __btrfs_unlink_inode+0x178/0x351 [btrfs] [52174.524053] [<ffffffff81051679>] warn_slowpath_fmt+0x48/0x50 [52174.524053] [<ffffffffa03591cd>] __btrfs_unlink_inode+0x178/0x351 [btrfs] [52174.524053] [<ffffffff8118f5e9>] ? iput+0xb0/0x284 [52174.524053] [<ffffffffa0359fe8>] btrfs_unlink_inode+0x1c/0x3d [btrfs] [52174.524053] [<ffffffffa038631e>] check_item_in_log+0x1fe/0x29b [btrfs] [52174.524053] [<ffffffffa0386522>] replay_dir_deletes+0x167/0x1cf [btrfs] [52174.524053] [<ffffffffa038739e>] fixup_inode_link_count+0x289/0x2aa [btrfs] [52174.524053] [<ffffffffa038748a>] fixup_inode_link_counts+0xcb/0x105 [btrfs] [52174.524053] [<ffffffffa038a5ec>] btrfs_recover_log_trees+0x258/0x32c [btrfs] [52174.524053] [<ffffffffa03885b2>] ? replay_one_extent+0x511/0x511 [btrfs] [52174.524053] [<ffffffffa034f288>] open_ctree+0x1dd4/0x21b9 [btrfs] [52174.524053] [<ffffffffa032b753>] btrfs_mount+0x97e/0xaed [btrfs] [52174.524053] [<ffffffff8108e1b7>] ? trace_hardirqs_on+0xd/0xf [52174.524053] [<ffffffff8117bafa>] mount_fs+0x67/0x131 [52174.524053] [<ffffffff81193003>] vfs_kern_mount+0x6c/0xde [52174.524053] [<ffffffffa032af81>] btrfs_mount+0x1ac/0xaed [btrfs] [52174.524053] [<ffffffff8108e1b7>] ? trace_hardirqs_on+0xd/0xf [52174.524053] [<ffffffff8108c262>] ? lockdep_init_map+0xb9/0x1b3 [52174.524053] [<ffffffff8117bafa>] mount_fs+0x67/0x131 [52174.524053] [<ffffffff81193003>] vfs_kern_mount+0x6c/0xde [52174.524053] [<ffffffff8119590f>] do_mount+0x8a6/0x9e8 [52174.524053] [<ffffffff811358dd>] ? strndup_user+0x3f/0x59 [52174.524053] [<ffffffff81195c65>] SyS_mount+0x77/0x9f [52174.524053] [<ffffffff814935d7>] entry_SYSCALL_64_fastpath+0x12/0x6b [52174.561288] ---[ end trace 6b53049efb1a3ea6 ]--- Fix this by forcing a transaction commit when such cases happen. This means we check in the commit root of the subvolume tree if there was any other inode with the same reference when the inode we are fsync'ing is a new inode (created in the current transaction). Test cases for fstests, covering all the scenarios given above, were submitted upstream for fstests: * fstests: generic test for fsync after renaming directory https://patchwork.kernel.org/patch/8694281/ * fstests: generic test for fsync after renaming file https://patchwork.kernel.org/patch/8694301/ * fstests: add btrfs test for fsync after snapshot deletion https://patchwork.kernel.org/patch/8670671/ Cc: stable@vger.kernel.org Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2016-03-31 01:37:21 +03:00
{
int ret;
struct btrfs_path *search_path;
char *name = NULL;
u32 name_len = 0;
u32 item_size = btrfs_item_size(eb, slot);
Btrfs: fix file/data loss caused by fsync after rename and new inode If we rename an inode A (be it a file or a directory), create a new inode B with the old name of inode A and under the same parent directory, fsync inode B and then power fail, at log tree replay time we end up removing inode A completely. If inode A is a directory then all its files are gone too. Example scenarios where this happens: This is reproducible with the following steps, taken from a couple of test cases written for fstests which are going to be submitted upstream soon: # Scenario 1 mkfs.btrfs -f /dev/sdc mount /dev/sdc /mnt mkdir -p /mnt/a/x echo "hello" > /mnt/a/x/foo echo "world" > /mnt/a/x/bar sync mv /mnt/a/x /mnt/a/y mkdir /mnt/a/x xfs_io -c fsync /mnt/a/x <power failure happens> The next time the fs is mounted, log tree replay happens and the directory "y" does not exist nor do the files "foo" and "bar" exist anywhere (neither in "y" nor in "x", nor the root nor anywhere). # Scenario 2 mkfs.btrfs -f /dev/sdc mount /dev/sdc /mnt mkdir /mnt/a echo "hello" > /mnt/a/foo sync mv /mnt/a/foo /mnt/a/bar echo "world" > /mnt/a/foo xfs_io -c fsync /mnt/a/foo <power failure happens> The next time the fs is mounted, log tree replay happens and the file "bar" does not exists anymore. A file with the name "foo" exists and it matches the second file we created. Another related problem that does not involve file/data loss is when a new inode is created with the name of a deleted snapshot and we fsync it: mkfs.btrfs -f /dev/sdc mount /dev/sdc /mnt mkdir /mnt/testdir btrfs subvolume snapshot /mnt /mnt/testdir/snap btrfs subvolume delete /mnt/testdir/snap rmdir /mnt/testdir mkdir /mnt/testdir xfs_io -c fsync /mnt/testdir # or fsync some file inside /mnt/testdir <power failure> The next time the fs is mounted the log replay procedure fails because it attempts to delete the snapshot entry (which has dir item key type of BTRFS_ROOT_ITEM_KEY) as if it were a regular (non-root) entry, resulting in the following error that causes mount to fail: [52174.510532] BTRFS info (device dm-0): failed to delete reference to snap, inode 257 parent 257 [52174.512570] ------------[ cut here ]------------ [52174.513278] WARNING: CPU: 12 PID: 28024 at fs/btrfs/inode.c:3986 __btrfs_unlink_inode+0x178/0x351 [btrfs]() [52174.514681] BTRFS: Transaction aborted (error -2) [52174.515630] Modules linked in: btrfs dm_flakey dm_mod overlay crc32c_generic ppdev xor raid6_pq acpi_cpufreq parport_pc tpm_tis sg parport tpm evdev i2c_piix4 proc [52174.521568] CPU: 12 PID: 28024 Comm: mount Tainted: G W 4.5.0-rc6-btrfs-next-27+ #1 [52174.522805] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS by qemu-project.org 04/01/2014 [52174.524053] 0000000000000000 ffff8801df2a7710 ffffffff81264e93 ffff8801df2a7758 [52174.524053] 0000000000000009 ffff8801df2a7748 ffffffff81051618 ffffffffa03591cd [52174.524053] 00000000fffffffe ffff88015e6e5000 ffff88016dbc3c88 ffff88016dbc3c88 [52174.524053] Call Trace: [52174.524053] [<ffffffff81264e93>] dump_stack+0x67/0x90 [52174.524053] [<ffffffff81051618>] warn_slowpath_common+0x99/0xb2 [52174.524053] [<ffffffffa03591cd>] ? __btrfs_unlink_inode+0x178/0x351 [btrfs] [52174.524053] [<ffffffff81051679>] warn_slowpath_fmt+0x48/0x50 [52174.524053] [<ffffffffa03591cd>] __btrfs_unlink_inode+0x178/0x351 [btrfs] [52174.524053] [<ffffffff8118f5e9>] ? iput+0xb0/0x284 [52174.524053] [<ffffffffa0359fe8>] btrfs_unlink_inode+0x1c/0x3d [btrfs] [52174.524053] [<ffffffffa038631e>] check_item_in_log+0x1fe/0x29b [btrfs] [52174.524053] [<ffffffffa0386522>] replay_dir_deletes+0x167/0x1cf [btrfs] [52174.524053] [<ffffffffa038739e>] fixup_inode_link_count+0x289/0x2aa [btrfs] [52174.524053] [<ffffffffa038748a>] fixup_inode_link_counts+0xcb/0x105 [btrfs] [52174.524053] [<ffffffffa038a5ec>] btrfs_recover_log_trees+0x258/0x32c [btrfs] [52174.524053] [<ffffffffa03885b2>] ? replay_one_extent+0x511/0x511 [btrfs] [52174.524053] [<ffffffffa034f288>] open_ctree+0x1dd4/0x21b9 [btrfs] [52174.524053] [<ffffffffa032b753>] btrfs_mount+0x97e/0xaed [btrfs] [52174.524053] [<ffffffff8108e1b7>] ? trace_hardirqs_on+0xd/0xf [52174.524053] [<ffffffff8117bafa>] mount_fs+0x67/0x131 [52174.524053] [<ffffffff81193003>] vfs_kern_mount+0x6c/0xde [52174.524053] [<ffffffffa032af81>] btrfs_mount+0x1ac/0xaed [btrfs] [52174.524053] [<ffffffff8108e1b7>] ? trace_hardirqs_on+0xd/0xf [52174.524053] [<ffffffff8108c262>] ? lockdep_init_map+0xb9/0x1b3 [52174.524053] [<ffffffff8117bafa>] mount_fs+0x67/0x131 [52174.524053] [<ffffffff81193003>] vfs_kern_mount+0x6c/0xde [52174.524053] [<ffffffff8119590f>] do_mount+0x8a6/0x9e8 [52174.524053] [<ffffffff811358dd>] ? strndup_user+0x3f/0x59 [52174.524053] [<ffffffff81195c65>] SyS_mount+0x77/0x9f [52174.524053] [<ffffffff814935d7>] entry_SYSCALL_64_fastpath+0x12/0x6b [52174.561288] ---[ end trace 6b53049efb1a3ea6 ]--- Fix this by forcing a transaction commit when such cases happen. This means we check in the commit root of the subvolume tree if there was any other inode with the same reference when the inode we are fsync'ing is a new inode (created in the current transaction). Test cases for fstests, covering all the scenarios given above, were submitted upstream for fstests: * fstests: generic test for fsync after renaming directory https://patchwork.kernel.org/patch/8694281/ * fstests: generic test for fsync after renaming file https://patchwork.kernel.org/patch/8694301/ * fstests: add btrfs test for fsync after snapshot deletion https://patchwork.kernel.org/patch/8670671/ Cc: stable@vger.kernel.org Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2016-03-31 01:37:21 +03:00
u32 cur_offset = 0;
unsigned long ptr = btrfs_item_ptr_offset(eb, slot);
search_path = btrfs_alloc_path();
if (!search_path)
return -ENOMEM;
search_path->search_commit_root = 1;
search_path->skip_locking = 1;
while (cur_offset < item_size) {
u64 parent;
u32 this_name_len;
u32 this_len;
unsigned long name_ptr;
struct btrfs_dir_item *di;
if (key->type == BTRFS_INODE_REF_KEY) {
struct btrfs_inode_ref *iref;
iref = (struct btrfs_inode_ref *)(ptr + cur_offset);
parent = key->offset;
this_name_len = btrfs_inode_ref_name_len(eb, iref);
name_ptr = (unsigned long)(iref + 1);
this_len = sizeof(*iref) + this_name_len;
} else {
struct btrfs_inode_extref *extref;
extref = (struct btrfs_inode_extref *)(ptr +
cur_offset);
parent = btrfs_inode_extref_parent(eb, extref);
this_name_len = btrfs_inode_extref_name_len(eb, extref);
name_ptr = (unsigned long)&extref->name;
this_len = sizeof(*extref) + this_name_len;
}
if (this_name_len > name_len) {
char *new_name;
new_name = krealloc(name, this_name_len, GFP_NOFS);
if (!new_name) {
ret = -ENOMEM;
goto out;
}
name_len = this_name_len;
name = new_name;
}
read_extent_buffer(eb, name, name_ptr, this_name_len);
di = btrfs_lookup_dir_item(NULL, inode->root, search_path,
parent, name, this_name_len, 0);
Btrfs: fix file/data loss caused by fsync after rename and new inode If we rename an inode A (be it a file or a directory), create a new inode B with the old name of inode A and under the same parent directory, fsync inode B and then power fail, at log tree replay time we end up removing inode A completely. If inode A is a directory then all its files are gone too. Example scenarios where this happens: This is reproducible with the following steps, taken from a couple of test cases written for fstests which are going to be submitted upstream soon: # Scenario 1 mkfs.btrfs -f /dev/sdc mount /dev/sdc /mnt mkdir -p /mnt/a/x echo "hello" > /mnt/a/x/foo echo "world" > /mnt/a/x/bar sync mv /mnt/a/x /mnt/a/y mkdir /mnt/a/x xfs_io -c fsync /mnt/a/x <power failure happens> The next time the fs is mounted, log tree replay happens and the directory "y" does not exist nor do the files "foo" and "bar" exist anywhere (neither in "y" nor in "x", nor the root nor anywhere). # Scenario 2 mkfs.btrfs -f /dev/sdc mount /dev/sdc /mnt mkdir /mnt/a echo "hello" > /mnt/a/foo sync mv /mnt/a/foo /mnt/a/bar echo "world" > /mnt/a/foo xfs_io -c fsync /mnt/a/foo <power failure happens> The next time the fs is mounted, log tree replay happens and the file "bar" does not exists anymore. A file with the name "foo" exists and it matches the second file we created. Another related problem that does not involve file/data loss is when a new inode is created with the name of a deleted snapshot and we fsync it: mkfs.btrfs -f /dev/sdc mount /dev/sdc /mnt mkdir /mnt/testdir btrfs subvolume snapshot /mnt /mnt/testdir/snap btrfs subvolume delete /mnt/testdir/snap rmdir /mnt/testdir mkdir /mnt/testdir xfs_io -c fsync /mnt/testdir # or fsync some file inside /mnt/testdir <power failure> The next time the fs is mounted the log replay procedure fails because it attempts to delete the snapshot entry (which has dir item key type of BTRFS_ROOT_ITEM_KEY) as if it were a regular (non-root) entry, resulting in the following error that causes mount to fail: [52174.510532] BTRFS info (device dm-0): failed to delete reference to snap, inode 257 parent 257 [52174.512570] ------------[ cut here ]------------ [52174.513278] WARNING: CPU: 12 PID: 28024 at fs/btrfs/inode.c:3986 __btrfs_unlink_inode+0x178/0x351 [btrfs]() [52174.514681] BTRFS: Transaction aborted (error -2) [52174.515630] Modules linked in: btrfs dm_flakey dm_mod overlay crc32c_generic ppdev xor raid6_pq acpi_cpufreq parport_pc tpm_tis sg parport tpm evdev i2c_piix4 proc [52174.521568] CPU: 12 PID: 28024 Comm: mount Tainted: G W 4.5.0-rc6-btrfs-next-27+ #1 [52174.522805] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS by qemu-project.org 04/01/2014 [52174.524053] 0000000000000000 ffff8801df2a7710 ffffffff81264e93 ffff8801df2a7758 [52174.524053] 0000000000000009 ffff8801df2a7748 ffffffff81051618 ffffffffa03591cd [52174.524053] 00000000fffffffe ffff88015e6e5000 ffff88016dbc3c88 ffff88016dbc3c88 [52174.524053] Call Trace: [52174.524053] [<ffffffff81264e93>] dump_stack+0x67/0x90 [52174.524053] [<ffffffff81051618>] warn_slowpath_common+0x99/0xb2 [52174.524053] [<ffffffffa03591cd>] ? __btrfs_unlink_inode+0x178/0x351 [btrfs] [52174.524053] [<ffffffff81051679>] warn_slowpath_fmt+0x48/0x50 [52174.524053] [<ffffffffa03591cd>] __btrfs_unlink_inode+0x178/0x351 [btrfs] [52174.524053] [<ffffffff8118f5e9>] ? iput+0xb0/0x284 [52174.524053] [<ffffffffa0359fe8>] btrfs_unlink_inode+0x1c/0x3d [btrfs] [52174.524053] [<ffffffffa038631e>] check_item_in_log+0x1fe/0x29b [btrfs] [52174.524053] [<ffffffffa0386522>] replay_dir_deletes+0x167/0x1cf [btrfs] [52174.524053] [<ffffffffa038739e>] fixup_inode_link_count+0x289/0x2aa [btrfs] [52174.524053] [<ffffffffa038748a>] fixup_inode_link_counts+0xcb/0x105 [btrfs] [52174.524053] [<ffffffffa038a5ec>] btrfs_recover_log_trees+0x258/0x32c [btrfs] [52174.524053] [<ffffffffa03885b2>] ? replay_one_extent+0x511/0x511 [btrfs] [52174.524053] [<ffffffffa034f288>] open_ctree+0x1dd4/0x21b9 [btrfs] [52174.524053] [<ffffffffa032b753>] btrfs_mount+0x97e/0xaed [btrfs] [52174.524053] [<ffffffff8108e1b7>] ? trace_hardirqs_on+0xd/0xf [52174.524053] [<ffffffff8117bafa>] mount_fs+0x67/0x131 [52174.524053] [<ffffffff81193003>] vfs_kern_mount+0x6c/0xde [52174.524053] [<ffffffffa032af81>] btrfs_mount+0x1ac/0xaed [btrfs] [52174.524053] [<ffffffff8108e1b7>] ? trace_hardirqs_on+0xd/0xf [52174.524053] [<ffffffff8108c262>] ? lockdep_init_map+0xb9/0x1b3 [52174.524053] [<ffffffff8117bafa>] mount_fs+0x67/0x131 [52174.524053] [<ffffffff81193003>] vfs_kern_mount+0x6c/0xde [52174.524053] [<ffffffff8119590f>] do_mount+0x8a6/0x9e8 [52174.524053] [<ffffffff811358dd>] ? strndup_user+0x3f/0x59 [52174.524053] [<ffffffff81195c65>] SyS_mount+0x77/0x9f [52174.524053] [<ffffffff814935d7>] entry_SYSCALL_64_fastpath+0x12/0x6b [52174.561288] ---[ end trace 6b53049efb1a3ea6 ]--- Fix this by forcing a transaction commit when such cases happen. This means we check in the commit root of the subvolume tree if there was any other inode with the same reference when the inode we are fsync'ing is a new inode (created in the current transaction). Test cases for fstests, covering all the scenarios given above, were submitted upstream for fstests: * fstests: generic test for fsync after renaming directory https://patchwork.kernel.org/patch/8694281/ * fstests: generic test for fsync after renaming file https://patchwork.kernel.org/patch/8694301/ * fstests: add btrfs test for fsync after snapshot deletion https://patchwork.kernel.org/patch/8670671/ Cc: stable@vger.kernel.org Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2016-03-31 01:37:21 +03:00
if (di && !IS_ERR(di)) {
Btrfs: improve performance on fsync against new inode after rename/unlink With commit 56f23fdbb600 ("Btrfs: fix file/data loss caused by fsync after rename and new inode") we got simple fix for a functional issue when the following sequence of actions is done: at transaction N create file A at directory D at transaction N + M (where M >= 1) move/rename existing file A from directory D to directory E create a new file named A at directory D fsync the new file power fail The solution was to simply detect such scenario and fallback to a full transaction commit when we detect it. However this turned out to had a significant impact on throughput (and a bit on latency too) for benchmarks using the dbench tool, which simulates real workloads from smbd (Samba) servers. For example on a test vm (with a debug kernel): Unpatched: Throughput 19.1572 MB/sec 32 clients 32 procs max_latency=1005.229 ms Patched: Throughput 23.7015 MB/sec 32 clients 32 procs max_latency=809.206 ms The patched results (this patch is applied) are similar to the results of a kernel with the commit 56f23fdbb600 ("Btrfs: fix file/data loss caused by fsync after rename and new inode") reverted. This change avoids the fallback to a transaction commit and instead makes sure all the names of the conflicting inode (the one that had a name in a past transaction that matches the name of the new file in the same parent directory) are logged so that at log replay time we don't lose neither the new file nor the old file, and the old file gets the name it was renamed to. This also ends up avoiding a full transaction commit for a similar case that involves an unlink instead of a rename of the old file: at transaction N create file A at directory D at transaction N + M (where M >= 1) remove file A create a new file named A at directory D fsync the new file power fail Signed-off-by: Filipe Manana <fdmanana@suse.com>
2016-06-06 18:11:13 +03:00
struct btrfs_key di_key;
btrfs_dir_item_key_to_cpu(search_path->nodes[0],
di, &di_key);
if (di_key.type == BTRFS_INODE_ITEM_KEY) {
Btrfs: fix fsync after succession of renames of different files After a succession of rename operations of different files and fsyncing one of them, such that each file gets a new name that corresponds to an old name of another file, we can end up with a log that will cause a failure when attempted to replay at mount time (an EEXIST error). We currently have correct behaviour when such succession of renames involves only two files, but if there are more files involved, we end up not logging all the inodes that are needed, therefore resulting in a failure when attempting to replay the log. Example: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/fname1 $ touch /mnt/testdir/fname2 $ sync $ mv /mnt/testdir/fname1 /mnt/testdir/fname3 $ mv /mnt/testdir/fname2 /mnt/testdir/fname4 $ ln /mnt/testdir/fname3 /mnt/testdir/fname2 $ touch /mnt/testdir/fname1 $ xfs_io -c "fsync" /mnt/testdir/fname1 <power failure> $ mount /dev/sdb /mnt mount: mount /dev/sdb on /mnt failed: File exists So fix this by checking all inode dependencies when logging an inode. That is, if one logged inode A has a new name that matches the old name of some other inode B, check if inode B has a new name that matches the old name of some other inode C, and so on. This fix is implemented not by doing any recursive function calls but by using an iterative method using a linked list that is used in a first-in-first-out fashion. A test case for fstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-02-13 15:14:03 +03:00
if (di_key.objectid != key->objectid) {
ret = 1;
*other_ino = di_key.objectid;
Btrfs: fix fsync after succession of renames and unlink/rmdir After a succession of renames operations of different files and unlinking one of them, if we fsync one of the renamed files we can end up with a log that will either fail to replay at mount time or result in a filesystem that is in an inconsistent state. One example scenario: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/fname1 $ touch /mnt/testdir/fname2 $ sync $ mv /mnt/testdir/fname1 /mnt/testdir/fname3 $ rm -f /mnt/testdir/fname2 $ ln /mnt/testdir/fname3 /mnt/testdir/fname2 $ touch /mnt/testdir/fname1 $ xfs_io -c "fsync" /mnt/testdir/fname1 <power failure> $ mount /dev/sdb /mnt $ umount /mnt $ btrfs check /dev/sdb [1/7] checking root items [2/7] checking extents [3/7] checking free space cache [4/7] checking fs roots root 5 inode 259 errors 2, no orphan item ERROR: errors found in fs roots Opening filesystem to check... Checking filesystem on /dev/sdc UUID: 20e4abb8-5a19-4492-8bb4-6084125c2d0d found 393216 bytes used, error(s) found total csum bytes: 0 total tree bytes: 131072 total fs tree bytes: 32768 total extent tree bytes: 16384 btree space waste bytes: 122986 file data blocks allocated: 262144 referenced 262144 On a kernel without the first patch in this series, titled "[PATCH] Btrfs: fix fsync after succession of renames of different files", we get instead an error when mounting the filesystem due to failure of replaying the log: $ mount /dev/sdb /mnt mount: mount /dev/sdb on /mnt failed: File exists Fix this by logging the parent directory of an inode whenever we find an inode that no longer exists (was unlinked in the current transaction), during the procedure which finds inodes that have old names that collide with new names of other inodes. A test case for fstests follows soon. Reviewed-by: Nikolay Borisov <nborisov@suse.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-02-13 15:14:09 +03:00
*other_parent = parent;
Btrfs: fix fsync after succession of renames of different files After a succession of rename operations of different files and fsyncing one of them, such that each file gets a new name that corresponds to an old name of another file, we can end up with a log that will cause a failure when attempted to replay at mount time (an EEXIST error). We currently have correct behaviour when such succession of renames involves only two files, but if there are more files involved, we end up not logging all the inodes that are needed, therefore resulting in a failure when attempting to replay the log. Example: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/fname1 $ touch /mnt/testdir/fname2 $ sync $ mv /mnt/testdir/fname1 /mnt/testdir/fname3 $ mv /mnt/testdir/fname2 /mnt/testdir/fname4 $ ln /mnt/testdir/fname3 /mnt/testdir/fname2 $ touch /mnt/testdir/fname1 $ xfs_io -c "fsync" /mnt/testdir/fname1 <power failure> $ mount /dev/sdb /mnt mount: mount /dev/sdb on /mnt failed: File exists So fix this by checking all inode dependencies when logging an inode. That is, if one logged inode A has a new name that matches the old name of some other inode B, check if inode B has a new name that matches the old name of some other inode C, and so on. This fix is implemented not by doing any recursive function calls but by using an iterative method using a linked list that is used in a first-in-first-out fashion. A test case for fstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-02-13 15:14:03 +03:00
} else {
ret = 0;
}
Btrfs: improve performance on fsync against new inode after rename/unlink With commit 56f23fdbb600 ("Btrfs: fix file/data loss caused by fsync after rename and new inode") we got simple fix for a functional issue when the following sequence of actions is done: at transaction N create file A at directory D at transaction N + M (where M >= 1) move/rename existing file A from directory D to directory E create a new file named A at directory D fsync the new file power fail The solution was to simply detect such scenario and fallback to a full transaction commit when we detect it. However this turned out to had a significant impact on throughput (and a bit on latency too) for benchmarks using the dbench tool, which simulates real workloads from smbd (Samba) servers. For example on a test vm (with a debug kernel): Unpatched: Throughput 19.1572 MB/sec 32 clients 32 procs max_latency=1005.229 ms Patched: Throughput 23.7015 MB/sec 32 clients 32 procs max_latency=809.206 ms The patched results (this patch is applied) are similar to the results of a kernel with the commit 56f23fdbb600 ("Btrfs: fix file/data loss caused by fsync after rename and new inode") reverted. This change avoids the fallback to a transaction commit and instead makes sure all the names of the conflicting inode (the one that had a name in a past transaction that matches the name of the new file in the same parent directory) are logged so that at log replay time we don't lose neither the new file nor the old file, and the old file gets the name it was renamed to. This also ends up avoiding a full transaction commit for a similar case that involves an unlink instead of a rename of the old file: at transaction N create file A at directory D at transaction N + M (where M >= 1) remove file A create a new file named A at directory D fsync the new file power fail Signed-off-by: Filipe Manana <fdmanana@suse.com>
2016-06-06 18:11:13 +03:00
} else {
ret = -EAGAIN;
}
Btrfs: fix file/data loss caused by fsync after rename and new inode If we rename an inode A (be it a file or a directory), create a new inode B with the old name of inode A and under the same parent directory, fsync inode B and then power fail, at log tree replay time we end up removing inode A completely. If inode A is a directory then all its files are gone too. Example scenarios where this happens: This is reproducible with the following steps, taken from a couple of test cases written for fstests which are going to be submitted upstream soon: # Scenario 1 mkfs.btrfs -f /dev/sdc mount /dev/sdc /mnt mkdir -p /mnt/a/x echo "hello" > /mnt/a/x/foo echo "world" > /mnt/a/x/bar sync mv /mnt/a/x /mnt/a/y mkdir /mnt/a/x xfs_io -c fsync /mnt/a/x <power failure happens> The next time the fs is mounted, log tree replay happens and the directory "y" does not exist nor do the files "foo" and "bar" exist anywhere (neither in "y" nor in "x", nor the root nor anywhere). # Scenario 2 mkfs.btrfs -f /dev/sdc mount /dev/sdc /mnt mkdir /mnt/a echo "hello" > /mnt/a/foo sync mv /mnt/a/foo /mnt/a/bar echo "world" > /mnt/a/foo xfs_io -c fsync /mnt/a/foo <power failure happens> The next time the fs is mounted, log tree replay happens and the file "bar" does not exists anymore. A file with the name "foo" exists and it matches the second file we created. Another related problem that does not involve file/data loss is when a new inode is created with the name of a deleted snapshot and we fsync it: mkfs.btrfs -f /dev/sdc mount /dev/sdc /mnt mkdir /mnt/testdir btrfs subvolume snapshot /mnt /mnt/testdir/snap btrfs subvolume delete /mnt/testdir/snap rmdir /mnt/testdir mkdir /mnt/testdir xfs_io -c fsync /mnt/testdir # or fsync some file inside /mnt/testdir <power failure> The next time the fs is mounted the log replay procedure fails because it attempts to delete the snapshot entry (which has dir item key type of BTRFS_ROOT_ITEM_KEY) as if it were a regular (non-root) entry, resulting in the following error that causes mount to fail: [52174.510532] BTRFS info (device dm-0): failed to delete reference to snap, inode 257 parent 257 [52174.512570] ------------[ cut here ]------------ [52174.513278] WARNING: CPU: 12 PID: 28024 at fs/btrfs/inode.c:3986 __btrfs_unlink_inode+0x178/0x351 [btrfs]() [52174.514681] BTRFS: Transaction aborted (error -2) [52174.515630] Modules linked in: btrfs dm_flakey dm_mod overlay crc32c_generic ppdev xor raid6_pq acpi_cpufreq parport_pc tpm_tis sg parport tpm evdev i2c_piix4 proc [52174.521568] CPU: 12 PID: 28024 Comm: mount Tainted: G W 4.5.0-rc6-btrfs-next-27+ #1 [52174.522805] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS by qemu-project.org 04/01/2014 [52174.524053] 0000000000000000 ffff8801df2a7710 ffffffff81264e93 ffff8801df2a7758 [52174.524053] 0000000000000009 ffff8801df2a7748 ffffffff81051618 ffffffffa03591cd [52174.524053] 00000000fffffffe ffff88015e6e5000 ffff88016dbc3c88 ffff88016dbc3c88 [52174.524053] Call Trace: [52174.524053] [<ffffffff81264e93>] dump_stack+0x67/0x90 [52174.524053] [<ffffffff81051618>] warn_slowpath_common+0x99/0xb2 [52174.524053] [<ffffffffa03591cd>] ? __btrfs_unlink_inode+0x178/0x351 [btrfs] [52174.524053] [<ffffffff81051679>] warn_slowpath_fmt+0x48/0x50 [52174.524053] [<ffffffffa03591cd>] __btrfs_unlink_inode+0x178/0x351 [btrfs] [52174.524053] [<ffffffff8118f5e9>] ? iput+0xb0/0x284 [52174.524053] [<ffffffffa0359fe8>] btrfs_unlink_inode+0x1c/0x3d [btrfs] [52174.524053] [<ffffffffa038631e>] check_item_in_log+0x1fe/0x29b [btrfs] [52174.524053] [<ffffffffa0386522>] replay_dir_deletes+0x167/0x1cf [btrfs] [52174.524053] [<ffffffffa038739e>] fixup_inode_link_count+0x289/0x2aa [btrfs] [52174.524053] [<ffffffffa038748a>] fixup_inode_link_counts+0xcb/0x105 [btrfs] [52174.524053] [<ffffffffa038a5ec>] btrfs_recover_log_trees+0x258/0x32c [btrfs] [52174.524053] [<ffffffffa03885b2>] ? replay_one_extent+0x511/0x511 [btrfs] [52174.524053] [<ffffffffa034f288>] open_ctree+0x1dd4/0x21b9 [btrfs] [52174.524053] [<ffffffffa032b753>] btrfs_mount+0x97e/0xaed [btrfs] [52174.524053] [<ffffffff8108e1b7>] ? trace_hardirqs_on+0xd/0xf [52174.524053] [<ffffffff8117bafa>] mount_fs+0x67/0x131 [52174.524053] [<ffffffff81193003>] vfs_kern_mount+0x6c/0xde [52174.524053] [<ffffffffa032af81>] btrfs_mount+0x1ac/0xaed [btrfs] [52174.524053] [<ffffffff8108e1b7>] ? trace_hardirqs_on+0xd/0xf [52174.524053] [<ffffffff8108c262>] ? lockdep_init_map+0xb9/0x1b3 [52174.524053] [<ffffffff8117bafa>] mount_fs+0x67/0x131 [52174.524053] [<ffffffff81193003>] vfs_kern_mount+0x6c/0xde [52174.524053] [<ffffffff8119590f>] do_mount+0x8a6/0x9e8 [52174.524053] [<ffffffff811358dd>] ? strndup_user+0x3f/0x59 [52174.524053] [<ffffffff81195c65>] SyS_mount+0x77/0x9f [52174.524053] [<ffffffff814935d7>] entry_SYSCALL_64_fastpath+0x12/0x6b [52174.561288] ---[ end trace 6b53049efb1a3ea6 ]--- Fix this by forcing a transaction commit when such cases happen. This means we check in the commit root of the subvolume tree if there was any other inode with the same reference when the inode we are fsync'ing is a new inode (created in the current transaction). Test cases for fstests, covering all the scenarios given above, were submitted upstream for fstests: * fstests: generic test for fsync after renaming directory https://patchwork.kernel.org/patch/8694281/ * fstests: generic test for fsync after renaming file https://patchwork.kernel.org/patch/8694301/ * fstests: add btrfs test for fsync after snapshot deletion https://patchwork.kernel.org/patch/8670671/ Cc: stable@vger.kernel.org Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2016-03-31 01:37:21 +03:00
goto out;
} else if (IS_ERR(di)) {
ret = PTR_ERR(di);
goto out;
}
btrfs_release_path(search_path);
cur_offset += this_len;
}
ret = 0;
out:
btrfs_free_path(search_path);
kfree(name);
return ret;
}
Btrfs: fix fsync after succession of renames of different files After a succession of rename operations of different files and fsyncing one of them, such that each file gets a new name that corresponds to an old name of another file, we can end up with a log that will cause a failure when attempted to replay at mount time (an EEXIST error). We currently have correct behaviour when such succession of renames involves only two files, but if there are more files involved, we end up not logging all the inodes that are needed, therefore resulting in a failure when attempting to replay the log. Example: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/fname1 $ touch /mnt/testdir/fname2 $ sync $ mv /mnt/testdir/fname1 /mnt/testdir/fname3 $ mv /mnt/testdir/fname2 /mnt/testdir/fname4 $ ln /mnt/testdir/fname3 /mnt/testdir/fname2 $ touch /mnt/testdir/fname1 $ xfs_io -c "fsync" /mnt/testdir/fname1 <power failure> $ mount /dev/sdb /mnt mount: mount /dev/sdb on /mnt failed: File exists So fix this by checking all inode dependencies when logging an inode. That is, if one logged inode A has a new name that matches the old name of some other inode B, check if inode B has a new name that matches the old name of some other inode C, and so on. This fix is implemented not by doing any recursive function calls but by using an iterative method using a linked list that is used in a first-in-first-out fashion. A test case for fstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-02-13 15:14:03 +03:00
struct btrfs_ino_list {
u64 ino;
Btrfs: fix fsync after succession of renames and unlink/rmdir After a succession of renames operations of different files and unlinking one of them, if we fsync one of the renamed files we can end up with a log that will either fail to replay at mount time or result in a filesystem that is in an inconsistent state. One example scenario: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/fname1 $ touch /mnt/testdir/fname2 $ sync $ mv /mnt/testdir/fname1 /mnt/testdir/fname3 $ rm -f /mnt/testdir/fname2 $ ln /mnt/testdir/fname3 /mnt/testdir/fname2 $ touch /mnt/testdir/fname1 $ xfs_io -c "fsync" /mnt/testdir/fname1 <power failure> $ mount /dev/sdb /mnt $ umount /mnt $ btrfs check /dev/sdb [1/7] checking root items [2/7] checking extents [3/7] checking free space cache [4/7] checking fs roots root 5 inode 259 errors 2, no orphan item ERROR: errors found in fs roots Opening filesystem to check... Checking filesystem on /dev/sdc UUID: 20e4abb8-5a19-4492-8bb4-6084125c2d0d found 393216 bytes used, error(s) found total csum bytes: 0 total tree bytes: 131072 total fs tree bytes: 32768 total extent tree bytes: 16384 btree space waste bytes: 122986 file data blocks allocated: 262144 referenced 262144 On a kernel without the first patch in this series, titled "[PATCH] Btrfs: fix fsync after succession of renames of different files", we get instead an error when mounting the filesystem due to failure of replaying the log: $ mount /dev/sdb /mnt mount: mount /dev/sdb on /mnt failed: File exists Fix this by logging the parent directory of an inode whenever we find an inode that no longer exists (was unlinked in the current transaction), during the procedure which finds inodes that have old names that collide with new names of other inodes. A test case for fstests follows soon. Reviewed-by: Nikolay Borisov <nborisov@suse.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-02-13 15:14:09 +03:00
u64 parent;
Btrfs: fix fsync after succession of renames of different files After a succession of rename operations of different files and fsyncing one of them, such that each file gets a new name that corresponds to an old name of another file, we can end up with a log that will cause a failure when attempted to replay at mount time (an EEXIST error). We currently have correct behaviour when such succession of renames involves only two files, but if there are more files involved, we end up not logging all the inodes that are needed, therefore resulting in a failure when attempting to replay the log. Example: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/fname1 $ touch /mnt/testdir/fname2 $ sync $ mv /mnt/testdir/fname1 /mnt/testdir/fname3 $ mv /mnt/testdir/fname2 /mnt/testdir/fname4 $ ln /mnt/testdir/fname3 /mnt/testdir/fname2 $ touch /mnt/testdir/fname1 $ xfs_io -c "fsync" /mnt/testdir/fname1 <power failure> $ mount /dev/sdb /mnt mount: mount /dev/sdb on /mnt failed: File exists So fix this by checking all inode dependencies when logging an inode. That is, if one logged inode A has a new name that matches the old name of some other inode B, check if inode B has a new name that matches the old name of some other inode C, and so on. This fix is implemented not by doing any recursive function calls but by using an iterative method using a linked list that is used in a first-in-first-out fashion. A test case for fstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-02-13 15:14:03 +03:00
struct list_head list;
};
static int log_conflicting_inodes(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_path *path,
struct btrfs_log_ctx *ctx,
Btrfs: fix fsync after succession of renames and unlink/rmdir After a succession of renames operations of different files and unlinking one of them, if we fsync one of the renamed files we can end up with a log that will either fail to replay at mount time or result in a filesystem that is in an inconsistent state. One example scenario: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/fname1 $ touch /mnt/testdir/fname2 $ sync $ mv /mnt/testdir/fname1 /mnt/testdir/fname3 $ rm -f /mnt/testdir/fname2 $ ln /mnt/testdir/fname3 /mnt/testdir/fname2 $ touch /mnt/testdir/fname1 $ xfs_io -c "fsync" /mnt/testdir/fname1 <power failure> $ mount /dev/sdb /mnt $ umount /mnt $ btrfs check /dev/sdb [1/7] checking root items [2/7] checking extents [3/7] checking free space cache [4/7] checking fs roots root 5 inode 259 errors 2, no orphan item ERROR: errors found in fs roots Opening filesystem to check... Checking filesystem on /dev/sdc UUID: 20e4abb8-5a19-4492-8bb4-6084125c2d0d found 393216 bytes used, error(s) found total csum bytes: 0 total tree bytes: 131072 total fs tree bytes: 32768 total extent tree bytes: 16384 btree space waste bytes: 122986 file data blocks allocated: 262144 referenced 262144 On a kernel without the first patch in this series, titled "[PATCH] Btrfs: fix fsync after succession of renames of different files", we get instead an error when mounting the filesystem due to failure of replaying the log: $ mount /dev/sdb /mnt mount: mount /dev/sdb on /mnt failed: File exists Fix this by logging the parent directory of an inode whenever we find an inode that no longer exists (was unlinked in the current transaction), during the procedure which finds inodes that have old names that collide with new names of other inodes. A test case for fstests follows soon. Reviewed-by: Nikolay Borisov <nborisov@suse.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-02-13 15:14:09 +03:00
u64 ino, u64 parent)
Btrfs: fix fsync after succession of renames of different files After a succession of rename operations of different files and fsyncing one of them, such that each file gets a new name that corresponds to an old name of another file, we can end up with a log that will cause a failure when attempted to replay at mount time (an EEXIST error). We currently have correct behaviour when such succession of renames involves only two files, but if there are more files involved, we end up not logging all the inodes that are needed, therefore resulting in a failure when attempting to replay the log. Example: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/fname1 $ touch /mnt/testdir/fname2 $ sync $ mv /mnt/testdir/fname1 /mnt/testdir/fname3 $ mv /mnt/testdir/fname2 /mnt/testdir/fname4 $ ln /mnt/testdir/fname3 /mnt/testdir/fname2 $ touch /mnt/testdir/fname1 $ xfs_io -c "fsync" /mnt/testdir/fname1 <power failure> $ mount /dev/sdb /mnt mount: mount /dev/sdb on /mnt failed: File exists So fix this by checking all inode dependencies when logging an inode. That is, if one logged inode A has a new name that matches the old name of some other inode B, check if inode B has a new name that matches the old name of some other inode C, and so on. This fix is implemented not by doing any recursive function calls but by using an iterative method using a linked list that is used in a first-in-first-out fashion. A test case for fstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-02-13 15:14:03 +03:00
{
struct btrfs_ino_list *ino_elem;
LIST_HEAD(inode_list);
int ret = 0;
ino_elem = kmalloc(sizeof(*ino_elem), GFP_NOFS);
if (!ino_elem)
return -ENOMEM;
ino_elem->ino = ino;
Btrfs: fix fsync after succession of renames and unlink/rmdir After a succession of renames operations of different files and unlinking one of them, if we fsync one of the renamed files we can end up with a log that will either fail to replay at mount time or result in a filesystem that is in an inconsistent state. One example scenario: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/fname1 $ touch /mnt/testdir/fname2 $ sync $ mv /mnt/testdir/fname1 /mnt/testdir/fname3 $ rm -f /mnt/testdir/fname2 $ ln /mnt/testdir/fname3 /mnt/testdir/fname2 $ touch /mnt/testdir/fname1 $ xfs_io -c "fsync" /mnt/testdir/fname1 <power failure> $ mount /dev/sdb /mnt $ umount /mnt $ btrfs check /dev/sdb [1/7] checking root items [2/7] checking extents [3/7] checking free space cache [4/7] checking fs roots root 5 inode 259 errors 2, no orphan item ERROR: errors found in fs roots Opening filesystem to check... Checking filesystem on /dev/sdc UUID: 20e4abb8-5a19-4492-8bb4-6084125c2d0d found 393216 bytes used, error(s) found total csum bytes: 0 total tree bytes: 131072 total fs tree bytes: 32768 total extent tree bytes: 16384 btree space waste bytes: 122986 file data blocks allocated: 262144 referenced 262144 On a kernel without the first patch in this series, titled "[PATCH] Btrfs: fix fsync after succession of renames of different files", we get instead an error when mounting the filesystem due to failure of replaying the log: $ mount /dev/sdb /mnt mount: mount /dev/sdb on /mnt failed: File exists Fix this by logging the parent directory of an inode whenever we find an inode that no longer exists (was unlinked in the current transaction), during the procedure which finds inodes that have old names that collide with new names of other inodes. A test case for fstests follows soon. Reviewed-by: Nikolay Borisov <nborisov@suse.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-02-13 15:14:09 +03:00
ino_elem->parent = parent;
Btrfs: fix fsync after succession of renames of different files After a succession of rename operations of different files and fsyncing one of them, such that each file gets a new name that corresponds to an old name of another file, we can end up with a log that will cause a failure when attempted to replay at mount time (an EEXIST error). We currently have correct behaviour when such succession of renames involves only two files, but if there are more files involved, we end up not logging all the inodes that are needed, therefore resulting in a failure when attempting to replay the log. Example: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/fname1 $ touch /mnt/testdir/fname2 $ sync $ mv /mnt/testdir/fname1 /mnt/testdir/fname3 $ mv /mnt/testdir/fname2 /mnt/testdir/fname4 $ ln /mnt/testdir/fname3 /mnt/testdir/fname2 $ touch /mnt/testdir/fname1 $ xfs_io -c "fsync" /mnt/testdir/fname1 <power failure> $ mount /dev/sdb /mnt mount: mount /dev/sdb on /mnt failed: File exists So fix this by checking all inode dependencies when logging an inode. That is, if one logged inode A has a new name that matches the old name of some other inode B, check if inode B has a new name that matches the old name of some other inode C, and so on. This fix is implemented not by doing any recursive function calls but by using an iterative method using a linked list that is used in a first-in-first-out fashion. A test case for fstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-02-13 15:14:03 +03:00
list_add_tail(&ino_elem->list, &inode_list);
while (!list_empty(&inode_list)) {
struct btrfs_fs_info *fs_info = root->fs_info;
struct btrfs_key key;
struct inode *inode;
ino_elem = list_first_entry(&inode_list, struct btrfs_ino_list,
list);
ino = ino_elem->ino;
Btrfs: fix fsync after succession of renames and unlink/rmdir After a succession of renames operations of different files and unlinking one of them, if we fsync one of the renamed files we can end up with a log that will either fail to replay at mount time or result in a filesystem that is in an inconsistent state. One example scenario: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/fname1 $ touch /mnt/testdir/fname2 $ sync $ mv /mnt/testdir/fname1 /mnt/testdir/fname3 $ rm -f /mnt/testdir/fname2 $ ln /mnt/testdir/fname3 /mnt/testdir/fname2 $ touch /mnt/testdir/fname1 $ xfs_io -c "fsync" /mnt/testdir/fname1 <power failure> $ mount /dev/sdb /mnt $ umount /mnt $ btrfs check /dev/sdb [1/7] checking root items [2/7] checking extents [3/7] checking free space cache [4/7] checking fs roots root 5 inode 259 errors 2, no orphan item ERROR: errors found in fs roots Opening filesystem to check... Checking filesystem on /dev/sdc UUID: 20e4abb8-5a19-4492-8bb4-6084125c2d0d found 393216 bytes used, error(s) found total csum bytes: 0 total tree bytes: 131072 total fs tree bytes: 32768 total extent tree bytes: 16384 btree space waste bytes: 122986 file data blocks allocated: 262144 referenced 262144 On a kernel without the first patch in this series, titled "[PATCH] Btrfs: fix fsync after succession of renames of different files", we get instead an error when mounting the filesystem due to failure of replaying the log: $ mount /dev/sdb /mnt mount: mount /dev/sdb on /mnt failed: File exists Fix this by logging the parent directory of an inode whenever we find an inode that no longer exists (was unlinked in the current transaction), during the procedure which finds inodes that have old names that collide with new names of other inodes. A test case for fstests follows soon. Reviewed-by: Nikolay Borisov <nborisov@suse.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-02-13 15:14:09 +03:00
parent = ino_elem->parent;
Btrfs: fix fsync after succession of renames of different files After a succession of rename operations of different files and fsyncing one of them, such that each file gets a new name that corresponds to an old name of another file, we can end up with a log that will cause a failure when attempted to replay at mount time (an EEXIST error). We currently have correct behaviour when such succession of renames involves only two files, but if there are more files involved, we end up not logging all the inodes that are needed, therefore resulting in a failure when attempting to replay the log. Example: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/fname1 $ touch /mnt/testdir/fname2 $ sync $ mv /mnt/testdir/fname1 /mnt/testdir/fname3 $ mv /mnt/testdir/fname2 /mnt/testdir/fname4 $ ln /mnt/testdir/fname3 /mnt/testdir/fname2 $ touch /mnt/testdir/fname1 $ xfs_io -c "fsync" /mnt/testdir/fname1 <power failure> $ mount /dev/sdb /mnt mount: mount /dev/sdb on /mnt failed: File exists So fix this by checking all inode dependencies when logging an inode. That is, if one logged inode A has a new name that matches the old name of some other inode B, check if inode B has a new name that matches the old name of some other inode C, and so on. This fix is implemented not by doing any recursive function calls but by using an iterative method using a linked list that is used in a first-in-first-out fashion. A test case for fstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-02-13 15:14:03 +03:00
list_del(&ino_elem->list);
kfree(ino_elem);
if (ret)
continue;
btrfs_release_path(path);
inode = btrfs_iget(fs_info->sb, ino, root);
Btrfs: fix fsync after succession of renames of different files After a succession of rename operations of different files and fsyncing one of them, such that each file gets a new name that corresponds to an old name of another file, we can end up with a log that will cause a failure when attempted to replay at mount time (an EEXIST error). We currently have correct behaviour when such succession of renames involves only two files, but if there are more files involved, we end up not logging all the inodes that are needed, therefore resulting in a failure when attempting to replay the log. Example: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/fname1 $ touch /mnt/testdir/fname2 $ sync $ mv /mnt/testdir/fname1 /mnt/testdir/fname3 $ mv /mnt/testdir/fname2 /mnt/testdir/fname4 $ ln /mnt/testdir/fname3 /mnt/testdir/fname2 $ touch /mnt/testdir/fname1 $ xfs_io -c "fsync" /mnt/testdir/fname1 <power failure> $ mount /dev/sdb /mnt mount: mount /dev/sdb on /mnt failed: File exists So fix this by checking all inode dependencies when logging an inode. That is, if one logged inode A has a new name that matches the old name of some other inode B, check if inode B has a new name that matches the old name of some other inode C, and so on. This fix is implemented not by doing any recursive function calls but by using an iterative method using a linked list that is used in a first-in-first-out fashion. A test case for fstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-02-13 15:14:03 +03:00
/*
* If the other inode that had a conflicting dir entry was
Btrfs: fix fsync after succession of renames and unlink/rmdir After a succession of renames operations of different files and unlinking one of them, if we fsync one of the renamed files we can end up with a log that will either fail to replay at mount time or result in a filesystem that is in an inconsistent state. One example scenario: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/fname1 $ touch /mnt/testdir/fname2 $ sync $ mv /mnt/testdir/fname1 /mnt/testdir/fname3 $ rm -f /mnt/testdir/fname2 $ ln /mnt/testdir/fname3 /mnt/testdir/fname2 $ touch /mnt/testdir/fname1 $ xfs_io -c "fsync" /mnt/testdir/fname1 <power failure> $ mount /dev/sdb /mnt $ umount /mnt $ btrfs check /dev/sdb [1/7] checking root items [2/7] checking extents [3/7] checking free space cache [4/7] checking fs roots root 5 inode 259 errors 2, no orphan item ERROR: errors found in fs roots Opening filesystem to check... Checking filesystem on /dev/sdc UUID: 20e4abb8-5a19-4492-8bb4-6084125c2d0d found 393216 bytes used, error(s) found total csum bytes: 0 total tree bytes: 131072 total fs tree bytes: 32768 total extent tree bytes: 16384 btree space waste bytes: 122986 file data blocks allocated: 262144 referenced 262144 On a kernel without the first patch in this series, titled "[PATCH] Btrfs: fix fsync after succession of renames of different files", we get instead an error when mounting the filesystem due to failure of replaying the log: $ mount /dev/sdb /mnt mount: mount /dev/sdb on /mnt failed: File exists Fix this by logging the parent directory of an inode whenever we find an inode that no longer exists (was unlinked in the current transaction), during the procedure which finds inodes that have old names that collide with new names of other inodes. A test case for fstests follows soon. Reviewed-by: Nikolay Borisov <nborisov@suse.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-02-13 15:14:09 +03:00
* deleted in the current transaction, we need to log its parent
* directory.
Btrfs: fix fsync after succession of renames of different files After a succession of rename operations of different files and fsyncing one of them, such that each file gets a new name that corresponds to an old name of another file, we can end up with a log that will cause a failure when attempted to replay at mount time (an EEXIST error). We currently have correct behaviour when such succession of renames involves only two files, but if there are more files involved, we end up not logging all the inodes that are needed, therefore resulting in a failure when attempting to replay the log. Example: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/fname1 $ touch /mnt/testdir/fname2 $ sync $ mv /mnt/testdir/fname1 /mnt/testdir/fname3 $ mv /mnt/testdir/fname2 /mnt/testdir/fname4 $ ln /mnt/testdir/fname3 /mnt/testdir/fname2 $ touch /mnt/testdir/fname1 $ xfs_io -c "fsync" /mnt/testdir/fname1 <power failure> $ mount /dev/sdb /mnt mount: mount /dev/sdb on /mnt failed: File exists So fix this by checking all inode dependencies when logging an inode. That is, if one logged inode A has a new name that matches the old name of some other inode B, check if inode B has a new name that matches the old name of some other inode C, and so on. This fix is implemented not by doing any recursive function calls but by using an iterative method using a linked list that is used in a first-in-first-out fashion. A test case for fstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-02-13 15:14:03 +03:00
*/
if (IS_ERR(inode)) {
ret = PTR_ERR(inode);
Btrfs: fix fsync after succession of renames and unlink/rmdir After a succession of renames operations of different files and unlinking one of them, if we fsync one of the renamed files we can end up with a log that will either fail to replay at mount time or result in a filesystem that is in an inconsistent state. One example scenario: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/fname1 $ touch /mnt/testdir/fname2 $ sync $ mv /mnt/testdir/fname1 /mnt/testdir/fname3 $ rm -f /mnt/testdir/fname2 $ ln /mnt/testdir/fname3 /mnt/testdir/fname2 $ touch /mnt/testdir/fname1 $ xfs_io -c "fsync" /mnt/testdir/fname1 <power failure> $ mount /dev/sdb /mnt $ umount /mnt $ btrfs check /dev/sdb [1/7] checking root items [2/7] checking extents [3/7] checking free space cache [4/7] checking fs roots root 5 inode 259 errors 2, no orphan item ERROR: errors found in fs roots Opening filesystem to check... Checking filesystem on /dev/sdc UUID: 20e4abb8-5a19-4492-8bb4-6084125c2d0d found 393216 bytes used, error(s) found total csum bytes: 0 total tree bytes: 131072 total fs tree bytes: 32768 total extent tree bytes: 16384 btree space waste bytes: 122986 file data blocks allocated: 262144 referenced 262144 On a kernel without the first patch in this series, titled "[PATCH] Btrfs: fix fsync after succession of renames of different files", we get instead an error when mounting the filesystem due to failure of replaying the log: $ mount /dev/sdb /mnt mount: mount /dev/sdb on /mnt failed: File exists Fix this by logging the parent directory of an inode whenever we find an inode that no longer exists (was unlinked in the current transaction), during the procedure which finds inodes that have old names that collide with new names of other inodes. A test case for fstests follows soon. Reviewed-by: Nikolay Borisov <nborisov@suse.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-02-13 15:14:09 +03:00
if (ret == -ENOENT) {
inode = btrfs_iget(fs_info->sb, parent, root);
Btrfs: fix fsync after succession of renames and unlink/rmdir After a succession of renames operations of different files and unlinking one of them, if we fsync one of the renamed files we can end up with a log that will either fail to replay at mount time or result in a filesystem that is in an inconsistent state. One example scenario: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/fname1 $ touch /mnt/testdir/fname2 $ sync $ mv /mnt/testdir/fname1 /mnt/testdir/fname3 $ rm -f /mnt/testdir/fname2 $ ln /mnt/testdir/fname3 /mnt/testdir/fname2 $ touch /mnt/testdir/fname1 $ xfs_io -c "fsync" /mnt/testdir/fname1 <power failure> $ mount /dev/sdb /mnt $ umount /mnt $ btrfs check /dev/sdb [1/7] checking root items [2/7] checking extents [3/7] checking free space cache [4/7] checking fs roots root 5 inode 259 errors 2, no orphan item ERROR: errors found in fs roots Opening filesystem to check... Checking filesystem on /dev/sdc UUID: 20e4abb8-5a19-4492-8bb4-6084125c2d0d found 393216 bytes used, error(s) found total csum bytes: 0 total tree bytes: 131072 total fs tree bytes: 32768 total extent tree bytes: 16384 btree space waste bytes: 122986 file data blocks allocated: 262144 referenced 262144 On a kernel without the first patch in this series, titled "[PATCH] Btrfs: fix fsync after succession of renames of different files", we get instead an error when mounting the filesystem due to failure of replaying the log: $ mount /dev/sdb /mnt mount: mount /dev/sdb on /mnt failed: File exists Fix this by logging the parent directory of an inode whenever we find an inode that no longer exists (was unlinked in the current transaction), during the procedure which finds inodes that have old names that collide with new names of other inodes. A test case for fstests follows soon. Reviewed-by: Nikolay Borisov <nborisov@suse.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-02-13 15:14:09 +03:00
if (IS_ERR(inode)) {
ret = PTR_ERR(inode);
} else {
ret = btrfs_log_inode(trans,
Btrfs: fix fsync after succession of renames and unlink/rmdir After a succession of renames operations of different files and unlinking one of them, if we fsync one of the renamed files we can end up with a log that will either fail to replay at mount time or result in a filesystem that is in an inconsistent state. One example scenario: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/fname1 $ touch /mnt/testdir/fname2 $ sync $ mv /mnt/testdir/fname1 /mnt/testdir/fname3 $ rm -f /mnt/testdir/fname2 $ ln /mnt/testdir/fname3 /mnt/testdir/fname2 $ touch /mnt/testdir/fname1 $ xfs_io -c "fsync" /mnt/testdir/fname1 <power failure> $ mount /dev/sdb /mnt $ umount /mnt $ btrfs check /dev/sdb [1/7] checking root items [2/7] checking extents [3/7] checking free space cache [4/7] checking fs roots root 5 inode 259 errors 2, no orphan item ERROR: errors found in fs roots Opening filesystem to check... Checking filesystem on /dev/sdc UUID: 20e4abb8-5a19-4492-8bb4-6084125c2d0d found 393216 bytes used, error(s) found total csum bytes: 0 total tree bytes: 131072 total fs tree bytes: 32768 total extent tree bytes: 16384 btree space waste bytes: 122986 file data blocks allocated: 262144 referenced 262144 On a kernel without the first patch in this series, titled "[PATCH] Btrfs: fix fsync after succession of renames of different files", we get instead an error when mounting the filesystem due to failure of replaying the log: $ mount /dev/sdb /mnt mount: mount /dev/sdb on /mnt failed: File exists Fix this by logging the parent directory of an inode whenever we find an inode that no longer exists (was unlinked in the current transaction), during the procedure which finds inodes that have old names that collide with new names of other inodes. A test case for fstests follows soon. Reviewed-by: Nikolay Borisov <nborisov@suse.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-02-13 15:14:09 +03:00
BTRFS_I(inode),
LOG_OTHER_INODE_ALL,
btrfs: make fast fsyncs wait only for writeback Currently regardless of a full or a fast fsync we always wait for ordered extents to complete, and then start logging the inode after that. However for fast fsyncs we can just wait for the writeback to complete, we don't need to wait for the ordered extents to complete since we use the list of modified extents maps to figure out which extents we must log and we can get their checksums directly from the ordered extents that are still in flight, otherwise look them up from the checksums tree. Until commit b5e6c3e170b770 ("btrfs: always wait on ordered extents at fsync time"), for fast fsyncs, we used to start logging without even waiting for the writeback to complete first, we would wait for it to complete after logging, while holding a transaction open, which lead to performance issues when using cgroups and probably for other cases too, as wait for IO while holding a transaction handle should be avoided as much as possible. After that, for fast fsyncs, we started to wait for ordered extents to complete before starting to log, which adds some latency to fsyncs and we even got at least one report about a performance drop which bisected to that particular change: https://lore.kernel.org/linux-btrfs/20181109215148.GF23260@techsingularity.net/ This change makes fast fsyncs only wait for writeback to finish before starting to log the inode, instead of waiting for both the writeback to finish and for the ordered extents to complete. This brings back part of the logic we had that extracts checksums from in flight ordered extents, which are not yet in the checksums tree, and making sure transaction commits wait for the completion of ordered extents previously logged (by far most of the time they have already completed by the time a transaction commit starts, resulting in no wait at all), to avoid any data loss if an ordered extent completes after the transaction used to log an inode is committed, followed by a power failure. When there are no other tasks accessing the checksums and the subvolume btrees, the ordered extent completion is pretty fast, typically taking 100 to 200 microseconds only in my observations. However when there are other tasks accessing these btrees, ordered extent completion can take a lot more time due to lock contention on nodes and leaves of these btrees. I've seen cases over 2 milliseconds, which starts to be significant. In particular when we do have concurrent fsyncs against different files there is a lot of contention on the checksums btree, since we have many tasks writing the checksums into the btree and other tasks that already started the logging phase are doing lookups for checksums in the btree. This change also turns all ranged fsyncs into full ranged fsyncs, which is something we already did when not using the NO_HOLES features or when doing a full fsync. This is to guarantee we never miss checksums due to writeback having been triggered only for a part of an extent, and we end up logging the full extent but only checksums for the written range, which results in missing checksums after log replay. Allowing ranged fsyncs to operate again only in the original range, when using the NO_HOLES feature and doing a fast fsync is doable but requires some non trivial changes to the writeback path, which can always be worked on later if needed, but I don't think they are a very common use case. Several tests were performed using fio for different numbers of concurrent jobs, each writing and fsyncing its own file, for both sequential and random file writes. The tests were run on bare metal, no virtualization, on a box with 12 cores (Intel i7-8700), 64Gb of RAM and a NVMe device, with a kernel configuration that is the default of typical distributions (debian in this case), without debug options enabled (kasan, kmemleak, slub debug, debug of page allocations, lock debugging, etc). The following script that calls fio was used: $ cat test-fsync.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/btrfs MOUNT_OPTIONS="-o ssd -o space_cache=v2" MKFS_OPTIONS="-d single -m single" if [ $# -ne 5 ]; then echo "Use $0 NUM_JOBS FILE_SIZE FSYNC_FREQ BLOCK_SIZE [write|randwrite]" exit 1 fi NUM_JOBS=$1 FILE_SIZE=$2 FSYNC_FREQ=$3 BLOCK_SIZE=$4 WRITE_MODE=$5 if [ "$WRITE_MODE" != "write" ] && [ "$WRITE_MODE" != "randwrite" ]; then echo "Invalid WRITE_MODE, must be 'write' or 'randwrite'" exit 1 fi cat <<EOF > /tmp/fio-job.ini [writers] rw=$WRITE_MODE fsync=$FSYNC_FREQ fallocate=none group_reporting=1 direct=0 bs=$BLOCK_SIZE ioengine=sync size=$FILE_SIZE directory=$MNT numjobs=$NUM_JOBS EOF echo "performance" | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor echo echo "Using config:" echo cat /tmp/fio-job.ini echo umount $MNT &> /dev/null mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT fio /tmp/fio-job.ini umount $MNT The results were the following: ************************* *** sequential writes *** ************************* ==== 1 job, 8GiB file, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=36.6MiB/s (38.4MB/s), 36.6MiB/s-36.6MiB/s (38.4MB/s-38.4MB/s), io=8192MiB (8590MB), run=223689-223689msec After patch: WRITE: bw=40.2MiB/s (42.1MB/s), 40.2MiB/s-40.2MiB/s (42.1MB/s-42.1MB/s), io=8192MiB (8590MB), run=203980-203980msec (+9.8%, -8.8% runtime) ==== 2 jobs, 4GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=35.8MiB/s (37.5MB/s), 35.8MiB/s-35.8MiB/s (37.5MB/s-37.5MB/s), io=8192MiB (8590MB), run=228950-228950msec After patch: WRITE: bw=43.5MiB/s (45.6MB/s), 43.5MiB/s-43.5MiB/s (45.6MB/s-45.6MB/s), io=8192MiB (8590MB), run=188272-188272msec (+21.5% throughput, -17.8% runtime) ==== 4 jobs, 2GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=50.1MiB/s (52.6MB/s), 50.1MiB/s-50.1MiB/s (52.6MB/s-52.6MB/s), io=8192MiB (8590MB), run=163446-163446msec After patch: WRITE: bw=64.5MiB/s (67.6MB/s), 64.5MiB/s-64.5MiB/s (67.6MB/s-67.6MB/s), io=8192MiB (8590MB), run=126987-126987msec (+28.7% throughput, -22.3% runtime) ==== 8 jobs, 1GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=64.0MiB/s (68.1MB/s), 64.0MiB/s-64.0MiB/s (68.1MB/s-68.1MB/s), io=8192MiB (8590MB), run=126075-126075msec After patch: WRITE: bw=86.8MiB/s (91.0MB/s), 86.8MiB/s-86.8MiB/s (91.0MB/s-91.0MB/s), io=8192MiB (8590MB), run=94358-94358msec (+35.6% throughput, -25.2% runtime) ==== 16 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=79.8MiB/s (83.6MB/s), 79.8MiB/s-79.8MiB/s (83.6MB/s-83.6MB/s), io=8192MiB (8590MB), run=102694-102694msec After patch: WRITE: bw=107MiB/s (112MB/s), 107MiB/s-107MiB/s (112MB/s-112MB/s), io=8192MiB (8590MB), run=76446-76446msec (+34.1% throughput, -25.6% runtime) ==== 32 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=93.2MiB/s (97.7MB/s), 93.2MiB/s-93.2MiB/s (97.7MB/s-97.7MB/s), io=16.0GiB (17.2GB), run=175836-175836msec After patch: WRITE: bw=111MiB/s (117MB/s), 111MiB/s-111MiB/s (117MB/s-117MB/s), io=16.0GiB (17.2GB), run=147001-147001msec (+19.1% throughput, -16.4% runtime) ==== 64 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=108MiB/s (114MB/s), 108MiB/s-108MiB/s (114MB/s-114MB/s), io=32.0GiB (34.4GB), run=302656-302656msec After patch: WRITE: bw=133MiB/s (140MB/s), 133MiB/s-133MiB/s (140MB/s-140MB/s), io=32.0GiB (34.4GB), run=246003-246003msec (+23.1% throughput, -18.7% runtime) ************************ *** random writes *** ************************ ==== 1 job, 8GiB file, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=11.5MiB/s (12.0MB/s), 11.5MiB/s-11.5MiB/s (12.0MB/s-12.0MB/s), io=8192MiB (8590MB), run=714281-714281msec After patch: WRITE: bw=11.6MiB/s (12.2MB/s), 11.6MiB/s-11.6MiB/s (12.2MB/s-12.2MB/s), io=8192MiB (8590MB), run=705959-705959msec (+0.9% throughput, -1.7% runtime) ==== 2 jobs, 4GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=12.8MiB/s (13.5MB/s), 12.8MiB/s-12.8MiB/s (13.5MB/s-13.5MB/s), io=8192MiB (8590MB), run=638101-638101msec After patch: WRITE: bw=13.1MiB/s (13.7MB/s), 13.1MiB/s-13.1MiB/s (13.7MB/s-13.7MB/s), io=8192MiB (8590MB), run=625374-625374msec (+2.3% throughput, -2.0% runtime) ==== 4 jobs, 2GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=15.4MiB/s (16.2MB/s), 15.4MiB/s-15.4MiB/s (16.2MB/s-16.2MB/s), io=8192MiB (8590MB), run=531146-531146msec After patch: WRITE: bw=17.8MiB/s (18.7MB/s), 17.8MiB/s-17.8MiB/s (18.7MB/s-18.7MB/s), io=8192MiB (8590MB), run=460431-460431msec (+15.6% throughput, -13.3% runtime) ==== 8 jobs, 1GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=19.9MiB/s (20.8MB/s), 19.9MiB/s-19.9MiB/s (20.8MB/s-20.8MB/s), io=8192MiB (8590MB), run=412664-412664msec After patch: WRITE: bw=22.2MiB/s (23.3MB/s), 22.2MiB/s-22.2MiB/s (23.3MB/s-23.3MB/s), io=8192MiB (8590MB), run=368589-368589msec (+11.6% throughput, -10.7% runtime) ==== 16 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=29.3MiB/s (30.7MB/s), 29.3MiB/s-29.3MiB/s (30.7MB/s-30.7MB/s), io=8192MiB (8590MB), run=279924-279924msec After patch: WRITE: bw=30.4MiB/s (31.9MB/s), 30.4MiB/s-30.4MiB/s (31.9MB/s-31.9MB/s), io=8192MiB (8590MB), run=269258-269258msec (+3.8% throughput, -3.8% runtime) ==== 32 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=36.9MiB/s (38.7MB/s), 36.9MiB/s-36.9MiB/s (38.7MB/s-38.7MB/s), io=16.0GiB (17.2GB), run=443581-443581msec After patch: WRITE: bw=41.6MiB/s (43.6MB/s), 41.6MiB/s-41.6MiB/s (43.6MB/s-43.6MB/s), io=16.0GiB (17.2GB), run=394114-394114msec (+12.7% throughput, -11.2% runtime) ==== 64 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=45.9MiB/s (48.1MB/s), 45.9MiB/s-45.9MiB/s (48.1MB/s-48.1MB/s), io=32.0GiB (34.4GB), run=714614-714614msec After patch: WRITE: bw=48.8MiB/s (51.1MB/s), 48.8MiB/s-48.8MiB/s (51.1MB/s-51.1MB/s), io=32.0GiB (34.4GB), run=672087-672087msec (+6.3% throughput, -6.0% runtime) Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-08-11 14:43:58 +03:00
ctx);
Btrfs: fix assertion failure during fsync and use of stale transaction Sometimes when fsync'ing a file we need to log that other inodes exist and when we need to do that we acquire a reference on the inodes and then drop that reference using iput() after logging them. That generally is not a problem except if we end up doing the final iput() (dropping the last reference) on the inode and that inode has a link count of 0, which can happen in a very short time window if the logging path gets a reference on the inode while it's being unlinked. In that case we end up getting the eviction callback, btrfs_evict_inode(), invoked through the iput() call chain which needs to drop all of the inode's items from its subvolume btree, and in order to do that, it needs to join a transaction at the helper function evict_refill_and_join(). However because the task previously started a transaction at the fsync handler, btrfs_sync_file(), it has current->journal_info already pointing to a transaction handle and therefore evict_refill_and_join() will get that transaction handle from btrfs_join_transaction(). From this point on, two different problems can happen: 1) evict_refill_and_join() will often change the transaction handle's block reserve (->block_rsv) and set its ->bytes_reserved field to a value greater than 0. If evict_refill_and_join() never commits the transaction, the eviction handler ends up decreasing the reference count (->use_count) of the transaction handle through the call to btrfs_end_transaction(), and after that point we have a transaction handle with a NULL ->block_rsv (which is the value prior to the transaction join from evict_refill_and_join()) and a ->bytes_reserved value greater than 0. If after the eviction/iput completes the inode logging path hits an error or it decides that it must fallback to a transaction commit, the btrfs fsync handle, btrfs_sync_file(), gets a non-zero value from btrfs_log_dentry_safe(), and because of that non-zero value it tries to commit the transaction using a handle with a NULL ->block_rsv and a non-zero ->bytes_reserved value. This makes the transaction commit hit an assertion failure at btrfs_trans_release_metadata() because ->bytes_reserved is not zero but the ->block_rsv is NULL. The produced stack trace for that is like the following: [192922.917158] assertion failed: !trans->bytes_reserved, file: fs/btrfs/transaction.c, line: 816 [192922.917553] ------------[ cut here ]------------ [192922.917922] kernel BUG at fs/btrfs/ctree.h:3532! [192922.918310] invalid opcode: 0000 [#1] SMP DEBUG_PAGEALLOC PTI [192922.918666] CPU: 2 PID: 883 Comm: fsstress Tainted: G W 5.1.4-btrfs-next-47 #1 [192922.919035] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.11.2-0-gf9626ccb91-prebuilt.qemu-project.org 04/01/2014 [192922.919801] RIP: 0010:assfail.constprop.25+0x18/0x1a [btrfs] (...) [192922.920925] RSP: 0018:ffffaebdc8a27da8 EFLAGS: 00010286 [192922.921315] RAX: 0000000000000051 RBX: ffff95c9c16a41c0 RCX: 0000000000000000 [192922.921692] RDX: 0000000000000000 RSI: ffff95cab6b16838 RDI: ffff95cab6b16838 [192922.922066] RBP: ffff95c9c16a41c0 R08: 0000000000000000 R09: 0000000000000000 [192922.922442] R10: ffffaebdc8a27e70 R11: 0000000000000000 R12: ffff95ca731a0980 [192922.922820] R13: 0000000000000000 R14: ffff95ca84c73338 R15: ffff95ca731a0ea8 [192922.923200] FS: 00007f337eda4e80(0000) GS:ffff95cab6b00000(0000) knlGS:0000000000000000 [192922.923579] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [192922.923948] CR2: 00007f337edad000 CR3: 00000001e00f6002 CR4: 00000000003606e0 [192922.924329] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [192922.924711] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [192922.925105] Call Trace: [192922.925505] btrfs_trans_release_metadata+0x10c/0x170 [btrfs] [192922.925911] btrfs_commit_transaction+0x3e/0xaf0 [btrfs] [192922.926324] btrfs_sync_file+0x44c/0x490 [btrfs] [192922.926731] do_fsync+0x38/0x60 [192922.927138] __x64_sys_fdatasync+0x13/0x20 [192922.927543] do_syscall_64+0x60/0x1c0 [192922.927939] entry_SYSCALL_64_after_hwframe+0x49/0xbe (...) [192922.934077] ---[ end trace f00808b12068168f ]--- 2) If evict_refill_and_join() decides to commit the transaction, it will be able to do it, since the nested transaction join only increments the transaction handle's ->use_count reference counter and it does not prevent the transaction from getting committed. This means that after eviction completes, the fsync logging path will be using a transaction handle that refers to an already committed transaction. What happens when using such a stale transaction can be unpredictable, we are at least having a use-after-free on the transaction handle itself, since the transaction commit will call kmem_cache_free() against the handle regardless of its ->use_count value, or we can end up silently losing all the updates to the log tree after that iput() in the logging path, or using a transaction handle that in the meanwhile was allocated to another task for a new transaction, etc, pretty much unpredictable what can happen. In order to fix both of them, instead of using iput() during logging, use btrfs_add_delayed_iput(), so that the logging path of fsync never drops the last reference on an inode, that step is offloaded to a safe context (usually the cleaner kthread). The assertion failure issue was sporadically triggered by the test case generic/475 from fstests, which loads the dm error target while fsstress is running, which lead to fsync failing while logging inodes with -EIO errors and then trying later to commit the transaction, triggering the assertion failure. CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-09-10 17:26:49 +03:00
btrfs_add_delayed_iput(inode);
Btrfs: fix fsync after succession of renames and unlink/rmdir After a succession of renames operations of different files and unlinking one of them, if we fsync one of the renamed files we can end up with a log that will either fail to replay at mount time or result in a filesystem that is in an inconsistent state. One example scenario: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/fname1 $ touch /mnt/testdir/fname2 $ sync $ mv /mnt/testdir/fname1 /mnt/testdir/fname3 $ rm -f /mnt/testdir/fname2 $ ln /mnt/testdir/fname3 /mnt/testdir/fname2 $ touch /mnt/testdir/fname1 $ xfs_io -c "fsync" /mnt/testdir/fname1 <power failure> $ mount /dev/sdb /mnt $ umount /mnt $ btrfs check /dev/sdb [1/7] checking root items [2/7] checking extents [3/7] checking free space cache [4/7] checking fs roots root 5 inode 259 errors 2, no orphan item ERROR: errors found in fs roots Opening filesystem to check... Checking filesystem on /dev/sdc UUID: 20e4abb8-5a19-4492-8bb4-6084125c2d0d found 393216 bytes used, error(s) found total csum bytes: 0 total tree bytes: 131072 total fs tree bytes: 32768 total extent tree bytes: 16384 btree space waste bytes: 122986 file data blocks allocated: 262144 referenced 262144 On a kernel without the first patch in this series, titled "[PATCH] Btrfs: fix fsync after succession of renames of different files", we get instead an error when mounting the filesystem due to failure of replaying the log: $ mount /dev/sdb /mnt mount: mount /dev/sdb on /mnt failed: File exists Fix this by logging the parent directory of an inode whenever we find an inode that no longer exists (was unlinked in the current transaction), during the procedure which finds inodes that have old names that collide with new names of other inodes. A test case for fstests follows soon. Reviewed-by: Nikolay Borisov <nborisov@suse.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-02-13 15:14:09 +03:00
}
}
Btrfs: fix fsync after succession of renames of different files After a succession of rename operations of different files and fsyncing one of them, such that each file gets a new name that corresponds to an old name of another file, we can end up with a log that will cause a failure when attempted to replay at mount time (an EEXIST error). We currently have correct behaviour when such succession of renames involves only two files, but if there are more files involved, we end up not logging all the inodes that are needed, therefore resulting in a failure when attempting to replay the log. Example: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/fname1 $ touch /mnt/testdir/fname2 $ sync $ mv /mnt/testdir/fname1 /mnt/testdir/fname3 $ mv /mnt/testdir/fname2 /mnt/testdir/fname4 $ ln /mnt/testdir/fname3 /mnt/testdir/fname2 $ touch /mnt/testdir/fname1 $ xfs_io -c "fsync" /mnt/testdir/fname1 <power failure> $ mount /dev/sdb /mnt mount: mount /dev/sdb on /mnt failed: File exists So fix this by checking all inode dependencies when logging an inode. That is, if one logged inode A has a new name that matches the old name of some other inode B, check if inode B has a new name that matches the old name of some other inode C, and so on. This fix is implemented not by doing any recursive function calls but by using an iterative method using a linked list that is used in a first-in-first-out fashion. A test case for fstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-02-13 15:14:03 +03:00
continue;
}
Btrfs: fix infinite loop during fsync after rename operations Recently fsstress (from fstests) sporadically started to trigger an infinite loop during fsync operations. This turned out to be because support for the rename exchange and whiteout operations was added to fsstress in fstests. These operations, unlike any others in fsstress, cause file names to be reused, whence triggering this issue. However it's not necessary to use rename exchange and rename whiteout operations trigger this issue, simple rename operations and file creations are enough to trigger the issue. The issue boils down to when we are logging inodes that conflict (that had the name of any inode we need to log during the fsync operation), we keep logging them even if they were already logged before, and after that we check if there's any other inode that conflicts with them and then add it again to the list of inodes to log. Skipping already logged inodes fixes the issue. Consider the following example: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir # inode 257 $ touch /mnt/testdir/zz # inode 258 $ ln /mnt/testdir/zz /mnt/testdir/zz_link $ touch /mnt/testdir/a # inode 259 $ sync # The following 3 renames achieve the same result as a rename exchange # operation (<rename_exchange> /mnt/testdir/zz_link to /mnt/testdir/a). $ mv /mnt/testdir/a /mnt/testdir/a/tmp $ mv /mnt/testdir/zz_link /mnt/testdir/a $ mv /mnt/testdir/a/tmp /mnt/testdir/zz_link # The following rename and file creation give the same result as a # rename whiteout operation (<rename_whiteout> zz to a2). $ mv /mnt/testdir/zz /mnt/testdir/a2 $ touch /mnt/testdir/zz # inode 260 $ xfs_io -c fsync /mnt/testdir/zz --> results in the infinite loop The following steps happen: 1) When logging inode 260, we find that its reference named "zz" was used by inode 258 in the previous transaction (through the commit root), so inode 258 is added to the list of conflicting indoes that need to be logged; 2) After logging inode 258, we find that its reference named "a" was used by inode 259 in the previous transaction, and therefore we add inode 259 to the list of conflicting inodes to be logged; 3) After logging inode 259, we find that its reference named "zz_link" was used by inode 258 in the previous transaction - we add inode 258 to the list of conflicting inodes to log, again - we had already logged it before at step 3. After logging it again, we find again that inode 259 conflicts with him, and we add again 259 to the list, etc - we end up repeating all the previous steps. So fix this by skipping logging of conflicting inodes that were already logged. Fixes: 6b5fc433a7ad67 ("Btrfs: fix fsync after succession of renames of different files") CC: stable@vger.kernel.org # 5.1+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-01-15 16:21:35 +03:00
/*
* If the inode was already logged skip it - otherwise we can
* hit an infinite loop. Example:
*
* From the commit root (previous transaction) we have the
* following inodes:
*
* inode 257 a directory
* inode 258 with references "zz" and "zz_link" on inode 257
* inode 259 with reference "a" on inode 257
*
* And in the current (uncommitted) transaction we have:
*
* inode 257 a directory, unchanged
* inode 258 with references "a" and "a2" on inode 257
* inode 259 with reference "zz_link" on inode 257
* inode 261 with reference "zz" on inode 257
*
* When logging inode 261 the following infinite loop could
* happen if we don't skip already logged inodes:
*
* - we detect inode 258 as a conflicting inode, with inode 261
* on reference "zz", and log it;
*
* - we detect inode 259 as a conflicting inode, with inode 258
* on reference "a", and log it;
*
* - we detect inode 258 as a conflicting inode, with inode 259
* on reference "zz_link", and log it - again! After this we
* repeat the above steps forever.
*/
spin_lock(&BTRFS_I(inode)->lock);
/*
* Check the inode's logged_trans only instead of
* btrfs_inode_in_log(). This is because the last_log_commit of
* the inode is not updated when we only log that it exists (see
* btrfs_log_inode()).
Btrfs: fix infinite loop during fsync after rename operations Recently fsstress (from fstests) sporadically started to trigger an infinite loop during fsync operations. This turned out to be because support for the rename exchange and whiteout operations was added to fsstress in fstests. These operations, unlike any others in fsstress, cause file names to be reused, whence triggering this issue. However it's not necessary to use rename exchange and rename whiteout operations trigger this issue, simple rename operations and file creations are enough to trigger the issue. The issue boils down to when we are logging inodes that conflict (that had the name of any inode we need to log during the fsync operation), we keep logging them even if they were already logged before, and after that we check if there's any other inode that conflicts with them and then add it again to the list of inodes to log. Skipping already logged inodes fixes the issue. Consider the following example: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir # inode 257 $ touch /mnt/testdir/zz # inode 258 $ ln /mnt/testdir/zz /mnt/testdir/zz_link $ touch /mnt/testdir/a # inode 259 $ sync # The following 3 renames achieve the same result as a rename exchange # operation (<rename_exchange> /mnt/testdir/zz_link to /mnt/testdir/a). $ mv /mnt/testdir/a /mnt/testdir/a/tmp $ mv /mnt/testdir/zz_link /mnt/testdir/a $ mv /mnt/testdir/a/tmp /mnt/testdir/zz_link # The following rename and file creation give the same result as a # rename whiteout operation (<rename_whiteout> zz to a2). $ mv /mnt/testdir/zz /mnt/testdir/a2 $ touch /mnt/testdir/zz # inode 260 $ xfs_io -c fsync /mnt/testdir/zz --> results in the infinite loop The following steps happen: 1) When logging inode 260, we find that its reference named "zz" was used by inode 258 in the previous transaction (through the commit root), so inode 258 is added to the list of conflicting indoes that need to be logged; 2) After logging inode 258, we find that its reference named "a" was used by inode 259 in the previous transaction, and therefore we add inode 259 to the list of conflicting inodes to be logged; 3) After logging inode 259, we find that its reference named "zz_link" was used by inode 258 in the previous transaction - we add inode 258 to the list of conflicting inodes to log, again - we had already logged it before at step 3. After logging it again, we find again that inode 259 conflicts with him, and we add again 259 to the list, etc - we end up repeating all the previous steps. So fix this by skipping logging of conflicting inodes that were already logged. Fixes: 6b5fc433a7ad67 ("Btrfs: fix fsync after succession of renames of different files") CC: stable@vger.kernel.org # 5.1+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-01-15 16:21:35 +03:00
*/
if (BTRFS_I(inode)->logged_trans == trans->transid) {
spin_unlock(&BTRFS_I(inode)->lock);
btrfs_add_delayed_iput(inode);
continue;
}
spin_unlock(&BTRFS_I(inode)->lock);
Btrfs: fix fsync after succession of renames of different files After a succession of rename operations of different files and fsyncing one of them, such that each file gets a new name that corresponds to an old name of another file, we can end up with a log that will cause a failure when attempted to replay at mount time (an EEXIST error). We currently have correct behaviour when such succession of renames involves only two files, but if there are more files involved, we end up not logging all the inodes that are needed, therefore resulting in a failure when attempting to replay the log. Example: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/fname1 $ touch /mnt/testdir/fname2 $ sync $ mv /mnt/testdir/fname1 /mnt/testdir/fname3 $ mv /mnt/testdir/fname2 /mnt/testdir/fname4 $ ln /mnt/testdir/fname3 /mnt/testdir/fname2 $ touch /mnt/testdir/fname1 $ xfs_io -c "fsync" /mnt/testdir/fname1 <power failure> $ mount /dev/sdb /mnt mount: mount /dev/sdb on /mnt failed: File exists So fix this by checking all inode dependencies when logging an inode. That is, if one logged inode A has a new name that matches the old name of some other inode B, check if inode B has a new name that matches the old name of some other inode C, and so on. This fix is implemented not by doing any recursive function calls but by using an iterative method using a linked list that is used in a first-in-first-out fashion. A test case for fstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-02-13 15:14:03 +03:00
/*
* We are safe logging the other inode without acquiring its
* lock as long as we log with the LOG_INODE_EXISTS mode. We
* are safe against concurrent renames of the other inode as
* well because during a rename we pin the log and update the
* log with the new name before we unpin it.
*/
ret = btrfs_log_inode(trans, BTRFS_I(inode), LOG_OTHER_INODE, ctx);
Btrfs: fix fsync after succession of renames of different files After a succession of rename operations of different files and fsyncing one of them, such that each file gets a new name that corresponds to an old name of another file, we can end up with a log that will cause a failure when attempted to replay at mount time (an EEXIST error). We currently have correct behaviour when such succession of renames involves only two files, but if there are more files involved, we end up not logging all the inodes that are needed, therefore resulting in a failure when attempting to replay the log. Example: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/fname1 $ touch /mnt/testdir/fname2 $ sync $ mv /mnt/testdir/fname1 /mnt/testdir/fname3 $ mv /mnt/testdir/fname2 /mnt/testdir/fname4 $ ln /mnt/testdir/fname3 /mnt/testdir/fname2 $ touch /mnt/testdir/fname1 $ xfs_io -c "fsync" /mnt/testdir/fname1 <power failure> $ mount /dev/sdb /mnt mount: mount /dev/sdb on /mnt failed: File exists So fix this by checking all inode dependencies when logging an inode. That is, if one logged inode A has a new name that matches the old name of some other inode B, check if inode B has a new name that matches the old name of some other inode C, and so on. This fix is implemented not by doing any recursive function calls but by using an iterative method using a linked list that is used in a first-in-first-out fashion. A test case for fstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-02-13 15:14:03 +03:00
if (ret) {
Btrfs: fix assertion failure during fsync and use of stale transaction Sometimes when fsync'ing a file we need to log that other inodes exist and when we need to do that we acquire a reference on the inodes and then drop that reference using iput() after logging them. That generally is not a problem except if we end up doing the final iput() (dropping the last reference) on the inode and that inode has a link count of 0, which can happen in a very short time window if the logging path gets a reference on the inode while it's being unlinked. In that case we end up getting the eviction callback, btrfs_evict_inode(), invoked through the iput() call chain which needs to drop all of the inode's items from its subvolume btree, and in order to do that, it needs to join a transaction at the helper function evict_refill_and_join(). However because the task previously started a transaction at the fsync handler, btrfs_sync_file(), it has current->journal_info already pointing to a transaction handle and therefore evict_refill_and_join() will get that transaction handle from btrfs_join_transaction(). From this point on, two different problems can happen: 1) evict_refill_and_join() will often change the transaction handle's block reserve (->block_rsv) and set its ->bytes_reserved field to a value greater than 0. If evict_refill_and_join() never commits the transaction, the eviction handler ends up decreasing the reference count (->use_count) of the transaction handle through the call to btrfs_end_transaction(), and after that point we have a transaction handle with a NULL ->block_rsv (which is the value prior to the transaction join from evict_refill_and_join()) and a ->bytes_reserved value greater than 0. If after the eviction/iput completes the inode logging path hits an error or it decides that it must fallback to a transaction commit, the btrfs fsync handle, btrfs_sync_file(), gets a non-zero value from btrfs_log_dentry_safe(), and because of that non-zero value it tries to commit the transaction using a handle with a NULL ->block_rsv and a non-zero ->bytes_reserved value. This makes the transaction commit hit an assertion failure at btrfs_trans_release_metadata() because ->bytes_reserved is not zero but the ->block_rsv is NULL. The produced stack trace for that is like the following: [192922.917158] assertion failed: !trans->bytes_reserved, file: fs/btrfs/transaction.c, line: 816 [192922.917553] ------------[ cut here ]------------ [192922.917922] kernel BUG at fs/btrfs/ctree.h:3532! [192922.918310] invalid opcode: 0000 [#1] SMP DEBUG_PAGEALLOC PTI [192922.918666] CPU: 2 PID: 883 Comm: fsstress Tainted: G W 5.1.4-btrfs-next-47 #1 [192922.919035] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.11.2-0-gf9626ccb91-prebuilt.qemu-project.org 04/01/2014 [192922.919801] RIP: 0010:assfail.constprop.25+0x18/0x1a [btrfs] (...) [192922.920925] RSP: 0018:ffffaebdc8a27da8 EFLAGS: 00010286 [192922.921315] RAX: 0000000000000051 RBX: ffff95c9c16a41c0 RCX: 0000000000000000 [192922.921692] RDX: 0000000000000000 RSI: ffff95cab6b16838 RDI: ffff95cab6b16838 [192922.922066] RBP: ffff95c9c16a41c0 R08: 0000000000000000 R09: 0000000000000000 [192922.922442] R10: ffffaebdc8a27e70 R11: 0000000000000000 R12: ffff95ca731a0980 [192922.922820] R13: 0000000000000000 R14: ffff95ca84c73338 R15: ffff95ca731a0ea8 [192922.923200] FS: 00007f337eda4e80(0000) GS:ffff95cab6b00000(0000) knlGS:0000000000000000 [192922.923579] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [192922.923948] CR2: 00007f337edad000 CR3: 00000001e00f6002 CR4: 00000000003606e0 [192922.924329] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [192922.924711] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [192922.925105] Call Trace: [192922.925505] btrfs_trans_release_metadata+0x10c/0x170 [btrfs] [192922.925911] btrfs_commit_transaction+0x3e/0xaf0 [btrfs] [192922.926324] btrfs_sync_file+0x44c/0x490 [btrfs] [192922.926731] do_fsync+0x38/0x60 [192922.927138] __x64_sys_fdatasync+0x13/0x20 [192922.927543] do_syscall_64+0x60/0x1c0 [192922.927939] entry_SYSCALL_64_after_hwframe+0x49/0xbe (...) [192922.934077] ---[ end trace f00808b12068168f ]--- 2) If evict_refill_and_join() decides to commit the transaction, it will be able to do it, since the nested transaction join only increments the transaction handle's ->use_count reference counter and it does not prevent the transaction from getting committed. This means that after eviction completes, the fsync logging path will be using a transaction handle that refers to an already committed transaction. What happens when using such a stale transaction can be unpredictable, we are at least having a use-after-free on the transaction handle itself, since the transaction commit will call kmem_cache_free() against the handle regardless of its ->use_count value, or we can end up silently losing all the updates to the log tree after that iput() in the logging path, or using a transaction handle that in the meanwhile was allocated to another task for a new transaction, etc, pretty much unpredictable what can happen. In order to fix both of them, instead of using iput() during logging, use btrfs_add_delayed_iput(), so that the logging path of fsync never drops the last reference on an inode, that step is offloaded to a safe context (usually the cleaner kthread). The assertion failure issue was sporadically triggered by the test case generic/475 from fstests, which loads the dm error target while fsstress is running, which lead to fsync failing while logging inodes with -EIO errors and then trying later to commit the transaction, triggering the assertion failure. CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-09-10 17:26:49 +03:00
btrfs_add_delayed_iput(inode);
Btrfs: fix fsync after succession of renames of different files After a succession of rename operations of different files and fsyncing one of them, such that each file gets a new name that corresponds to an old name of another file, we can end up with a log that will cause a failure when attempted to replay at mount time (an EEXIST error). We currently have correct behaviour when such succession of renames involves only two files, but if there are more files involved, we end up not logging all the inodes that are needed, therefore resulting in a failure when attempting to replay the log. Example: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/fname1 $ touch /mnt/testdir/fname2 $ sync $ mv /mnt/testdir/fname1 /mnt/testdir/fname3 $ mv /mnt/testdir/fname2 /mnt/testdir/fname4 $ ln /mnt/testdir/fname3 /mnt/testdir/fname2 $ touch /mnt/testdir/fname1 $ xfs_io -c "fsync" /mnt/testdir/fname1 <power failure> $ mount /dev/sdb /mnt mount: mount /dev/sdb on /mnt failed: File exists So fix this by checking all inode dependencies when logging an inode. That is, if one logged inode A has a new name that matches the old name of some other inode B, check if inode B has a new name that matches the old name of some other inode C, and so on. This fix is implemented not by doing any recursive function calls but by using an iterative method using a linked list that is used in a first-in-first-out fashion. A test case for fstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-02-13 15:14:03 +03:00
continue;
}
key.objectid = ino;
key.type = BTRFS_INODE_REF_KEY;
key.offset = 0;
ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
if (ret < 0) {
Btrfs: fix assertion failure during fsync and use of stale transaction Sometimes when fsync'ing a file we need to log that other inodes exist and when we need to do that we acquire a reference on the inodes and then drop that reference using iput() after logging them. That generally is not a problem except if we end up doing the final iput() (dropping the last reference) on the inode and that inode has a link count of 0, which can happen in a very short time window if the logging path gets a reference on the inode while it's being unlinked. In that case we end up getting the eviction callback, btrfs_evict_inode(), invoked through the iput() call chain which needs to drop all of the inode's items from its subvolume btree, and in order to do that, it needs to join a transaction at the helper function evict_refill_and_join(). However because the task previously started a transaction at the fsync handler, btrfs_sync_file(), it has current->journal_info already pointing to a transaction handle and therefore evict_refill_and_join() will get that transaction handle from btrfs_join_transaction(). From this point on, two different problems can happen: 1) evict_refill_and_join() will often change the transaction handle's block reserve (->block_rsv) and set its ->bytes_reserved field to a value greater than 0. If evict_refill_and_join() never commits the transaction, the eviction handler ends up decreasing the reference count (->use_count) of the transaction handle through the call to btrfs_end_transaction(), and after that point we have a transaction handle with a NULL ->block_rsv (which is the value prior to the transaction join from evict_refill_and_join()) and a ->bytes_reserved value greater than 0. If after the eviction/iput completes the inode logging path hits an error or it decides that it must fallback to a transaction commit, the btrfs fsync handle, btrfs_sync_file(), gets a non-zero value from btrfs_log_dentry_safe(), and because of that non-zero value it tries to commit the transaction using a handle with a NULL ->block_rsv and a non-zero ->bytes_reserved value. This makes the transaction commit hit an assertion failure at btrfs_trans_release_metadata() because ->bytes_reserved is not zero but the ->block_rsv is NULL. The produced stack trace for that is like the following: [192922.917158] assertion failed: !trans->bytes_reserved, file: fs/btrfs/transaction.c, line: 816 [192922.917553] ------------[ cut here ]------------ [192922.917922] kernel BUG at fs/btrfs/ctree.h:3532! [192922.918310] invalid opcode: 0000 [#1] SMP DEBUG_PAGEALLOC PTI [192922.918666] CPU: 2 PID: 883 Comm: fsstress Tainted: G W 5.1.4-btrfs-next-47 #1 [192922.919035] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.11.2-0-gf9626ccb91-prebuilt.qemu-project.org 04/01/2014 [192922.919801] RIP: 0010:assfail.constprop.25+0x18/0x1a [btrfs] (...) [192922.920925] RSP: 0018:ffffaebdc8a27da8 EFLAGS: 00010286 [192922.921315] RAX: 0000000000000051 RBX: ffff95c9c16a41c0 RCX: 0000000000000000 [192922.921692] RDX: 0000000000000000 RSI: ffff95cab6b16838 RDI: ffff95cab6b16838 [192922.922066] RBP: ffff95c9c16a41c0 R08: 0000000000000000 R09: 0000000000000000 [192922.922442] R10: ffffaebdc8a27e70 R11: 0000000000000000 R12: ffff95ca731a0980 [192922.922820] R13: 0000000000000000 R14: ffff95ca84c73338 R15: ffff95ca731a0ea8 [192922.923200] FS: 00007f337eda4e80(0000) GS:ffff95cab6b00000(0000) knlGS:0000000000000000 [192922.923579] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [192922.923948] CR2: 00007f337edad000 CR3: 00000001e00f6002 CR4: 00000000003606e0 [192922.924329] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [192922.924711] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [192922.925105] Call Trace: [192922.925505] btrfs_trans_release_metadata+0x10c/0x170 [btrfs] [192922.925911] btrfs_commit_transaction+0x3e/0xaf0 [btrfs] [192922.926324] btrfs_sync_file+0x44c/0x490 [btrfs] [192922.926731] do_fsync+0x38/0x60 [192922.927138] __x64_sys_fdatasync+0x13/0x20 [192922.927543] do_syscall_64+0x60/0x1c0 [192922.927939] entry_SYSCALL_64_after_hwframe+0x49/0xbe (...) [192922.934077] ---[ end trace f00808b12068168f ]--- 2) If evict_refill_and_join() decides to commit the transaction, it will be able to do it, since the nested transaction join only increments the transaction handle's ->use_count reference counter and it does not prevent the transaction from getting committed. This means that after eviction completes, the fsync logging path will be using a transaction handle that refers to an already committed transaction. What happens when using such a stale transaction can be unpredictable, we are at least having a use-after-free on the transaction handle itself, since the transaction commit will call kmem_cache_free() against the handle regardless of its ->use_count value, or we can end up silently losing all the updates to the log tree after that iput() in the logging path, or using a transaction handle that in the meanwhile was allocated to another task for a new transaction, etc, pretty much unpredictable what can happen. In order to fix both of them, instead of using iput() during logging, use btrfs_add_delayed_iput(), so that the logging path of fsync never drops the last reference on an inode, that step is offloaded to a safe context (usually the cleaner kthread). The assertion failure issue was sporadically triggered by the test case generic/475 from fstests, which loads the dm error target while fsstress is running, which lead to fsync failing while logging inodes with -EIO errors and then trying later to commit the transaction, triggering the assertion failure. CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-09-10 17:26:49 +03:00
btrfs_add_delayed_iput(inode);
Btrfs: fix fsync after succession of renames of different files After a succession of rename operations of different files and fsyncing one of them, such that each file gets a new name that corresponds to an old name of another file, we can end up with a log that will cause a failure when attempted to replay at mount time (an EEXIST error). We currently have correct behaviour when such succession of renames involves only two files, but if there are more files involved, we end up not logging all the inodes that are needed, therefore resulting in a failure when attempting to replay the log. Example: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/fname1 $ touch /mnt/testdir/fname2 $ sync $ mv /mnt/testdir/fname1 /mnt/testdir/fname3 $ mv /mnt/testdir/fname2 /mnt/testdir/fname4 $ ln /mnt/testdir/fname3 /mnt/testdir/fname2 $ touch /mnt/testdir/fname1 $ xfs_io -c "fsync" /mnt/testdir/fname1 <power failure> $ mount /dev/sdb /mnt mount: mount /dev/sdb on /mnt failed: File exists So fix this by checking all inode dependencies when logging an inode. That is, if one logged inode A has a new name that matches the old name of some other inode B, check if inode B has a new name that matches the old name of some other inode C, and so on. This fix is implemented not by doing any recursive function calls but by using an iterative method using a linked list that is used in a first-in-first-out fashion. A test case for fstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-02-13 15:14:03 +03:00
continue;
}
while (true) {
struct extent_buffer *leaf = path->nodes[0];
int slot = path->slots[0];
u64 other_ino = 0;
Btrfs: fix fsync after succession of renames and unlink/rmdir After a succession of renames operations of different files and unlinking one of them, if we fsync one of the renamed files we can end up with a log that will either fail to replay at mount time or result in a filesystem that is in an inconsistent state. One example scenario: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/fname1 $ touch /mnt/testdir/fname2 $ sync $ mv /mnt/testdir/fname1 /mnt/testdir/fname3 $ rm -f /mnt/testdir/fname2 $ ln /mnt/testdir/fname3 /mnt/testdir/fname2 $ touch /mnt/testdir/fname1 $ xfs_io -c "fsync" /mnt/testdir/fname1 <power failure> $ mount /dev/sdb /mnt $ umount /mnt $ btrfs check /dev/sdb [1/7] checking root items [2/7] checking extents [3/7] checking free space cache [4/7] checking fs roots root 5 inode 259 errors 2, no orphan item ERROR: errors found in fs roots Opening filesystem to check... Checking filesystem on /dev/sdc UUID: 20e4abb8-5a19-4492-8bb4-6084125c2d0d found 393216 bytes used, error(s) found total csum bytes: 0 total tree bytes: 131072 total fs tree bytes: 32768 total extent tree bytes: 16384 btree space waste bytes: 122986 file data blocks allocated: 262144 referenced 262144 On a kernel without the first patch in this series, titled "[PATCH] Btrfs: fix fsync after succession of renames of different files", we get instead an error when mounting the filesystem due to failure of replaying the log: $ mount /dev/sdb /mnt mount: mount /dev/sdb on /mnt failed: File exists Fix this by logging the parent directory of an inode whenever we find an inode that no longer exists (was unlinked in the current transaction), during the procedure which finds inodes that have old names that collide with new names of other inodes. A test case for fstests follows soon. Reviewed-by: Nikolay Borisov <nborisov@suse.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-02-13 15:14:09 +03:00
u64 other_parent = 0;
Btrfs: fix fsync after succession of renames of different files After a succession of rename operations of different files and fsyncing one of them, such that each file gets a new name that corresponds to an old name of another file, we can end up with a log that will cause a failure when attempted to replay at mount time (an EEXIST error). We currently have correct behaviour when such succession of renames involves only two files, but if there are more files involved, we end up not logging all the inodes that are needed, therefore resulting in a failure when attempting to replay the log. Example: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/fname1 $ touch /mnt/testdir/fname2 $ sync $ mv /mnt/testdir/fname1 /mnt/testdir/fname3 $ mv /mnt/testdir/fname2 /mnt/testdir/fname4 $ ln /mnt/testdir/fname3 /mnt/testdir/fname2 $ touch /mnt/testdir/fname1 $ xfs_io -c "fsync" /mnt/testdir/fname1 <power failure> $ mount /dev/sdb /mnt mount: mount /dev/sdb on /mnt failed: File exists So fix this by checking all inode dependencies when logging an inode. That is, if one logged inode A has a new name that matches the old name of some other inode B, check if inode B has a new name that matches the old name of some other inode C, and so on. This fix is implemented not by doing any recursive function calls but by using an iterative method using a linked list that is used in a first-in-first-out fashion. A test case for fstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-02-13 15:14:03 +03:00
if (slot >= btrfs_header_nritems(leaf)) {
ret = btrfs_next_leaf(root, path);
if (ret < 0) {
break;
} else if (ret > 0) {
ret = 0;
break;
}
continue;
}
btrfs_item_key_to_cpu(leaf, &key, slot);
if (key.objectid != ino ||
(key.type != BTRFS_INODE_REF_KEY &&
key.type != BTRFS_INODE_EXTREF_KEY)) {
ret = 0;
break;
}
ret = btrfs_check_ref_name_override(leaf, slot, &key,
Btrfs: fix fsync after succession of renames and unlink/rmdir After a succession of renames operations of different files and unlinking one of them, if we fsync one of the renamed files we can end up with a log that will either fail to replay at mount time or result in a filesystem that is in an inconsistent state. One example scenario: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/fname1 $ touch /mnt/testdir/fname2 $ sync $ mv /mnt/testdir/fname1 /mnt/testdir/fname3 $ rm -f /mnt/testdir/fname2 $ ln /mnt/testdir/fname3 /mnt/testdir/fname2 $ touch /mnt/testdir/fname1 $ xfs_io -c "fsync" /mnt/testdir/fname1 <power failure> $ mount /dev/sdb /mnt $ umount /mnt $ btrfs check /dev/sdb [1/7] checking root items [2/7] checking extents [3/7] checking free space cache [4/7] checking fs roots root 5 inode 259 errors 2, no orphan item ERROR: errors found in fs roots Opening filesystem to check... Checking filesystem on /dev/sdc UUID: 20e4abb8-5a19-4492-8bb4-6084125c2d0d found 393216 bytes used, error(s) found total csum bytes: 0 total tree bytes: 131072 total fs tree bytes: 32768 total extent tree bytes: 16384 btree space waste bytes: 122986 file data blocks allocated: 262144 referenced 262144 On a kernel without the first patch in this series, titled "[PATCH] Btrfs: fix fsync after succession of renames of different files", we get instead an error when mounting the filesystem due to failure of replaying the log: $ mount /dev/sdb /mnt mount: mount /dev/sdb on /mnt failed: File exists Fix this by logging the parent directory of an inode whenever we find an inode that no longer exists (was unlinked in the current transaction), during the procedure which finds inodes that have old names that collide with new names of other inodes. A test case for fstests follows soon. Reviewed-by: Nikolay Borisov <nborisov@suse.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-02-13 15:14:09 +03:00
BTRFS_I(inode), &other_ino,
&other_parent);
Btrfs: fix fsync after succession of renames of different files After a succession of rename operations of different files and fsyncing one of them, such that each file gets a new name that corresponds to an old name of another file, we can end up with a log that will cause a failure when attempted to replay at mount time (an EEXIST error). We currently have correct behaviour when such succession of renames involves only two files, but if there are more files involved, we end up not logging all the inodes that are needed, therefore resulting in a failure when attempting to replay the log. Example: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/fname1 $ touch /mnt/testdir/fname2 $ sync $ mv /mnt/testdir/fname1 /mnt/testdir/fname3 $ mv /mnt/testdir/fname2 /mnt/testdir/fname4 $ ln /mnt/testdir/fname3 /mnt/testdir/fname2 $ touch /mnt/testdir/fname1 $ xfs_io -c "fsync" /mnt/testdir/fname1 <power failure> $ mount /dev/sdb /mnt mount: mount /dev/sdb on /mnt failed: File exists So fix this by checking all inode dependencies when logging an inode. That is, if one logged inode A has a new name that matches the old name of some other inode B, check if inode B has a new name that matches the old name of some other inode C, and so on. This fix is implemented not by doing any recursive function calls but by using an iterative method using a linked list that is used in a first-in-first-out fashion. A test case for fstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-02-13 15:14:03 +03:00
if (ret < 0)
break;
if (ret > 0) {
ino_elem = kmalloc(sizeof(*ino_elem), GFP_NOFS);
if (!ino_elem) {
ret = -ENOMEM;
break;
}
ino_elem->ino = other_ino;
Btrfs: fix fsync after succession of renames and unlink/rmdir After a succession of renames operations of different files and unlinking one of them, if we fsync one of the renamed files we can end up with a log that will either fail to replay at mount time or result in a filesystem that is in an inconsistent state. One example scenario: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/fname1 $ touch /mnt/testdir/fname2 $ sync $ mv /mnt/testdir/fname1 /mnt/testdir/fname3 $ rm -f /mnt/testdir/fname2 $ ln /mnt/testdir/fname3 /mnt/testdir/fname2 $ touch /mnt/testdir/fname1 $ xfs_io -c "fsync" /mnt/testdir/fname1 <power failure> $ mount /dev/sdb /mnt $ umount /mnt $ btrfs check /dev/sdb [1/7] checking root items [2/7] checking extents [3/7] checking free space cache [4/7] checking fs roots root 5 inode 259 errors 2, no orphan item ERROR: errors found in fs roots Opening filesystem to check... Checking filesystem on /dev/sdc UUID: 20e4abb8-5a19-4492-8bb4-6084125c2d0d found 393216 bytes used, error(s) found total csum bytes: 0 total tree bytes: 131072 total fs tree bytes: 32768 total extent tree bytes: 16384 btree space waste bytes: 122986 file data blocks allocated: 262144 referenced 262144 On a kernel without the first patch in this series, titled "[PATCH] Btrfs: fix fsync after succession of renames of different files", we get instead an error when mounting the filesystem due to failure of replaying the log: $ mount /dev/sdb /mnt mount: mount /dev/sdb on /mnt failed: File exists Fix this by logging the parent directory of an inode whenever we find an inode that no longer exists (was unlinked in the current transaction), during the procedure which finds inodes that have old names that collide with new names of other inodes. A test case for fstests follows soon. Reviewed-by: Nikolay Borisov <nborisov@suse.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-02-13 15:14:09 +03:00
ino_elem->parent = other_parent;
Btrfs: fix fsync after succession of renames of different files After a succession of rename operations of different files and fsyncing one of them, such that each file gets a new name that corresponds to an old name of another file, we can end up with a log that will cause a failure when attempted to replay at mount time (an EEXIST error). We currently have correct behaviour when such succession of renames involves only two files, but if there are more files involved, we end up not logging all the inodes that are needed, therefore resulting in a failure when attempting to replay the log. Example: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/fname1 $ touch /mnt/testdir/fname2 $ sync $ mv /mnt/testdir/fname1 /mnt/testdir/fname3 $ mv /mnt/testdir/fname2 /mnt/testdir/fname4 $ ln /mnt/testdir/fname3 /mnt/testdir/fname2 $ touch /mnt/testdir/fname1 $ xfs_io -c "fsync" /mnt/testdir/fname1 <power failure> $ mount /dev/sdb /mnt mount: mount /dev/sdb on /mnt failed: File exists So fix this by checking all inode dependencies when logging an inode. That is, if one logged inode A has a new name that matches the old name of some other inode B, check if inode B has a new name that matches the old name of some other inode C, and so on. This fix is implemented not by doing any recursive function calls but by using an iterative method using a linked list that is used in a first-in-first-out fashion. A test case for fstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-02-13 15:14:03 +03:00
list_add_tail(&ino_elem->list, &inode_list);
ret = 0;
}
path->slots[0]++;
}
Btrfs: fix assertion failure during fsync and use of stale transaction Sometimes when fsync'ing a file we need to log that other inodes exist and when we need to do that we acquire a reference on the inodes and then drop that reference using iput() after logging them. That generally is not a problem except if we end up doing the final iput() (dropping the last reference) on the inode and that inode has a link count of 0, which can happen in a very short time window if the logging path gets a reference on the inode while it's being unlinked. In that case we end up getting the eviction callback, btrfs_evict_inode(), invoked through the iput() call chain which needs to drop all of the inode's items from its subvolume btree, and in order to do that, it needs to join a transaction at the helper function evict_refill_and_join(). However because the task previously started a transaction at the fsync handler, btrfs_sync_file(), it has current->journal_info already pointing to a transaction handle and therefore evict_refill_and_join() will get that transaction handle from btrfs_join_transaction(). From this point on, two different problems can happen: 1) evict_refill_and_join() will often change the transaction handle's block reserve (->block_rsv) and set its ->bytes_reserved field to a value greater than 0. If evict_refill_and_join() never commits the transaction, the eviction handler ends up decreasing the reference count (->use_count) of the transaction handle through the call to btrfs_end_transaction(), and after that point we have a transaction handle with a NULL ->block_rsv (which is the value prior to the transaction join from evict_refill_and_join()) and a ->bytes_reserved value greater than 0. If after the eviction/iput completes the inode logging path hits an error or it decides that it must fallback to a transaction commit, the btrfs fsync handle, btrfs_sync_file(), gets a non-zero value from btrfs_log_dentry_safe(), and because of that non-zero value it tries to commit the transaction using a handle with a NULL ->block_rsv and a non-zero ->bytes_reserved value. This makes the transaction commit hit an assertion failure at btrfs_trans_release_metadata() because ->bytes_reserved is not zero but the ->block_rsv is NULL. The produced stack trace for that is like the following: [192922.917158] assertion failed: !trans->bytes_reserved, file: fs/btrfs/transaction.c, line: 816 [192922.917553] ------------[ cut here ]------------ [192922.917922] kernel BUG at fs/btrfs/ctree.h:3532! [192922.918310] invalid opcode: 0000 [#1] SMP DEBUG_PAGEALLOC PTI [192922.918666] CPU: 2 PID: 883 Comm: fsstress Tainted: G W 5.1.4-btrfs-next-47 #1 [192922.919035] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.11.2-0-gf9626ccb91-prebuilt.qemu-project.org 04/01/2014 [192922.919801] RIP: 0010:assfail.constprop.25+0x18/0x1a [btrfs] (...) [192922.920925] RSP: 0018:ffffaebdc8a27da8 EFLAGS: 00010286 [192922.921315] RAX: 0000000000000051 RBX: ffff95c9c16a41c0 RCX: 0000000000000000 [192922.921692] RDX: 0000000000000000 RSI: ffff95cab6b16838 RDI: ffff95cab6b16838 [192922.922066] RBP: ffff95c9c16a41c0 R08: 0000000000000000 R09: 0000000000000000 [192922.922442] R10: ffffaebdc8a27e70 R11: 0000000000000000 R12: ffff95ca731a0980 [192922.922820] R13: 0000000000000000 R14: ffff95ca84c73338 R15: ffff95ca731a0ea8 [192922.923200] FS: 00007f337eda4e80(0000) GS:ffff95cab6b00000(0000) knlGS:0000000000000000 [192922.923579] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [192922.923948] CR2: 00007f337edad000 CR3: 00000001e00f6002 CR4: 00000000003606e0 [192922.924329] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [192922.924711] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [192922.925105] Call Trace: [192922.925505] btrfs_trans_release_metadata+0x10c/0x170 [btrfs] [192922.925911] btrfs_commit_transaction+0x3e/0xaf0 [btrfs] [192922.926324] btrfs_sync_file+0x44c/0x490 [btrfs] [192922.926731] do_fsync+0x38/0x60 [192922.927138] __x64_sys_fdatasync+0x13/0x20 [192922.927543] do_syscall_64+0x60/0x1c0 [192922.927939] entry_SYSCALL_64_after_hwframe+0x49/0xbe (...) [192922.934077] ---[ end trace f00808b12068168f ]--- 2) If evict_refill_and_join() decides to commit the transaction, it will be able to do it, since the nested transaction join only increments the transaction handle's ->use_count reference counter and it does not prevent the transaction from getting committed. This means that after eviction completes, the fsync logging path will be using a transaction handle that refers to an already committed transaction. What happens when using such a stale transaction can be unpredictable, we are at least having a use-after-free on the transaction handle itself, since the transaction commit will call kmem_cache_free() against the handle regardless of its ->use_count value, or we can end up silently losing all the updates to the log tree after that iput() in the logging path, or using a transaction handle that in the meanwhile was allocated to another task for a new transaction, etc, pretty much unpredictable what can happen. In order to fix both of them, instead of using iput() during logging, use btrfs_add_delayed_iput(), so that the logging path of fsync never drops the last reference on an inode, that step is offloaded to a safe context (usually the cleaner kthread). The assertion failure issue was sporadically triggered by the test case generic/475 from fstests, which loads the dm error target while fsstress is running, which lead to fsync failing while logging inodes with -EIO errors and then trying later to commit the transaction, triggering the assertion failure. CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-09-10 17:26:49 +03:00
btrfs_add_delayed_iput(inode);
Btrfs: fix fsync after succession of renames of different files After a succession of rename operations of different files and fsyncing one of them, such that each file gets a new name that corresponds to an old name of another file, we can end up with a log that will cause a failure when attempted to replay at mount time (an EEXIST error). We currently have correct behaviour when such succession of renames involves only two files, but if there are more files involved, we end up not logging all the inodes that are needed, therefore resulting in a failure when attempting to replay the log. Example: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/fname1 $ touch /mnt/testdir/fname2 $ sync $ mv /mnt/testdir/fname1 /mnt/testdir/fname3 $ mv /mnt/testdir/fname2 /mnt/testdir/fname4 $ ln /mnt/testdir/fname3 /mnt/testdir/fname2 $ touch /mnt/testdir/fname1 $ xfs_io -c "fsync" /mnt/testdir/fname1 <power failure> $ mount /dev/sdb /mnt mount: mount /dev/sdb on /mnt failed: File exists So fix this by checking all inode dependencies when logging an inode. That is, if one logged inode A has a new name that matches the old name of some other inode B, check if inode B has a new name that matches the old name of some other inode C, and so on. This fix is implemented not by doing any recursive function calls but by using an iterative method using a linked list that is used in a first-in-first-out fashion. A test case for fstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-02-13 15:14:03 +03:00
}
return ret;
}
static int copy_inode_items_to_log(struct btrfs_trans_handle *trans,
struct btrfs_inode *inode,
struct btrfs_key *min_key,
const struct btrfs_key *max_key,
struct btrfs_path *path,
struct btrfs_path *dst_path,
const u64 logged_isize,
const bool recursive_logging,
const int inode_only,
struct btrfs_log_ctx *ctx,
bool *need_log_inode_item)
{
btrfs: fix lost prealloc extents beyond eof after full fsync When doing a full fsync, if we have prealloc extents beyond (or at) eof, and the leaves that contain them were not modified in the current transaction, we end up not logging them. This results in losing those extents when we replay the log after a power failure, since the inode is truncated to the current value of the logged i_size. Just like for the fast fsync path, we need to always log all prealloc extents starting at or beyond i_size. The fast fsync case was fixed in commit 471d557afed155 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay") but it missed the full fsync path. The problem exists since the very early days, when the log tree was added by commit e02119d5a7b439 ("Btrfs: Add a write ahead tree log to optimize synchronous operations"). Example reproducer: $ mkfs.btrfs -f /dev/sdc $ mount /dev/sdc /mnt # Create our test file with many file extent items, so that they span # several leaves of metadata, even if the node/page size is 64K. Use # direct IO and not fsync/O_SYNC because it's both faster and it avoids # clearing the full sync flag from the inode - we want the fsync below # to trigger the slow full sync code path. $ xfs_io -f -d -c "pwrite -b 4K 0 16M" /mnt/foo # Now add two preallocated extents to our file without extending the # file's size. One right at i_size, and another further beyond, leaving # a gap between the two prealloc extents. $ xfs_io -c "falloc -k 16M 1M" /mnt/foo $ xfs_io -c "falloc -k 20M 1M" /mnt/foo # Make sure everything is durably persisted and the transaction is # committed. This makes all created extents to have a generation lower # than the generation of the transaction used by the next write and # fsync. sync # Now overwrite only the first extent, which will result in modifying # only the first leaf of metadata for our inode. Then fsync it. This # fsync will use the slow code path (inode full sync bit is set) because # it's the first fsync since the inode was created/loaded. $ xfs_io -c "pwrite 0 4K" -c "fsync" /mnt/foo # Extent list before power failure. $ xfs_io -c "fiemap -v" /mnt/foo /mnt/foo: EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS 0: [0..7]: 2178048..2178055 8 0x0 1: [8..16383]: 26632..43007 16376 0x0 2: [16384..32767]: 2156544..2172927 16384 0x0 3: [32768..34815]: 2172928..2174975 2048 0x800 4: [34816..40959]: hole 6144 5: [40960..43007]: 2174976..2177023 2048 0x801 <power fail> # Mount fs again, trigger log replay. $ mount /dev/sdc /mnt # Extent list after power failure and log replay. $ xfs_io -c "fiemap -v" /mnt/foo /mnt/foo: EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS 0: [0..7]: 2178048..2178055 8 0x0 1: [8..16383]: 26632..43007 16376 0x0 2: [16384..32767]: 2156544..2172927 16384 0x1 # The prealloc extents at file offsets 16M and 20M are missing. So fix this by calling btrfs_log_prealloc_extents() when we are doing a full fsync, so that we always log all prealloc extents beyond eof. A test case for fstests will follow soon. CC: stable@vger.kernel.org # 4.19+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-02-17 15:12:02 +03:00
const u64 i_size = i_size_read(&inode->vfs_inode);
struct btrfs_root *root = inode->root;
int ins_start_slot = 0;
int ins_nr = 0;
int ret;
while (1) {
ret = btrfs_search_forward(root, min_key, path, trans->transid);
if (ret < 0)
return ret;
if (ret > 0) {
ret = 0;
break;
}
again:
/* Note, ins_nr might be > 0 here, cleanup outside the loop */
if (min_key->objectid != max_key->objectid)
break;
if (min_key->type > max_key->type)
break;
btrfs: fix lost prealloc extents beyond eof after full fsync When doing a full fsync, if we have prealloc extents beyond (or at) eof, and the leaves that contain them were not modified in the current transaction, we end up not logging them. This results in losing those extents when we replay the log after a power failure, since the inode is truncated to the current value of the logged i_size. Just like for the fast fsync path, we need to always log all prealloc extents starting at or beyond i_size. The fast fsync case was fixed in commit 471d557afed155 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay") but it missed the full fsync path. The problem exists since the very early days, when the log tree was added by commit e02119d5a7b439 ("Btrfs: Add a write ahead tree log to optimize synchronous operations"). Example reproducer: $ mkfs.btrfs -f /dev/sdc $ mount /dev/sdc /mnt # Create our test file with many file extent items, so that they span # several leaves of metadata, even if the node/page size is 64K. Use # direct IO and not fsync/O_SYNC because it's both faster and it avoids # clearing the full sync flag from the inode - we want the fsync below # to trigger the slow full sync code path. $ xfs_io -f -d -c "pwrite -b 4K 0 16M" /mnt/foo # Now add two preallocated extents to our file without extending the # file's size. One right at i_size, and another further beyond, leaving # a gap between the two prealloc extents. $ xfs_io -c "falloc -k 16M 1M" /mnt/foo $ xfs_io -c "falloc -k 20M 1M" /mnt/foo # Make sure everything is durably persisted and the transaction is # committed. This makes all created extents to have a generation lower # than the generation of the transaction used by the next write and # fsync. sync # Now overwrite only the first extent, which will result in modifying # only the first leaf of metadata for our inode. Then fsync it. This # fsync will use the slow code path (inode full sync bit is set) because # it's the first fsync since the inode was created/loaded. $ xfs_io -c "pwrite 0 4K" -c "fsync" /mnt/foo # Extent list before power failure. $ xfs_io -c "fiemap -v" /mnt/foo /mnt/foo: EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS 0: [0..7]: 2178048..2178055 8 0x0 1: [8..16383]: 26632..43007 16376 0x0 2: [16384..32767]: 2156544..2172927 16384 0x0 3: [32768..34815]: 2172928..2174975 2048 0x800 4: [34816..40959]: hole 6144 5: [40960..43007]: 2174976..2177023 2048 0x801 <power fail> # Mount fs again, trigger log replay. $ mount /dev/sdc /mnt # Extent list after power failure and log replay. $ xfs_io -c "fiemap -v" /mnt/foo /mnt/foo: EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS 0: [0..7]: 2178048..2178055 8 0x0 1: [8..16383]: 26632..43007 16376 0x0 2: [16384..32767]: 2156544..2172927 16384 0x1 # The prealloc extents at file offsets 16M and 20M are missing. So fix this by calling btrfs_log_prealloc_extents() when we are doing a full fsync, so that we always log all prealloc extents beyond eof. A test case for fstests will follow soon. CC: stable@vger.kernel.org # 4.19+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-02-17 15:12:02 +03:00
if (min_key->type == BTRFS_INODE_ITEM_KEY) {
*need_log_inode_item = false;
btrfs: fix lost prealloc extents beyond eof after full fsync When doing a full fsync, if we have prealloc extents beyond (or at) eof, and the leaves that contain them were not modified in the current transaction, we end up not logging them. This results in losing those extents when we replay the log after a power failure, since the inode is truncated to the current value of the logged i_size. Just like for the fast fsync path, we need to always log all prealloc extents starting at or beyond i_size. The fast fsync case was fixed in commit 471d557afed155 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay") but it missed the full fsync path. The problem exists since the very early days, when the log tree was added by commit e02119d5a7b439 ("Btrfs: Add a write ahead tree log to optimize synchronous operations"). Example reproducer: $ mkfs.btrfs -f /dev/sdc $ mount /dev/sdc /mnt # Create our test file with many file extent items, so that they span # several leaves of metadata, even if the node/page size is 64K. Use # direct IO and not fsync/O_SYNC because it's both faster and it avoids # clearing the full sync flag from the inode - we want the fsync below # to trigger the slow full sync code path. $ xfs_io -f -d -c "pwrite -b 4K 0 16M" /mnt/foo # Now add two preallocated extents to our file without extending the # file's size. One right at i_size, and another further beyond, leaving # a gap between the two prealloc extents. $ xfs_io -c "falloc -k 16M 1M" /mnt/foo $ xfs_io -c "falloc -k 20M 1M" /mnt/foo # Make sure everything is durably persisted and the transaction is # committed. This makes all created extents to have a generation lower # than the generation of the transaction used by the next write and # fsync. sync # Now overwrite only the first extent, which will result in modifying # only the first leaf of metadata for our inode. Then fsync it. This # fsync will use the slow code path (inode full sync bit is set) because # it's the first fsync since the inode was created/loaded. $ xfs_io -c "pwrite 0 4K" -c "fsync" /mnt/foo # Extent list before power failure. $ xfs_io -c "fiemap -v" /mnt/foo /mnt/foo: EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS 0: [0..7]: 2178048..2178055 8 0x0 1: [8..16383]: 26632..43007 16376 0x0 2: [16384..32767]: 2156544..2172927 16384 0x0 3: [32768..34815]: 2172928..2174975 2048 0x800 4: [34816..40959]: hole 6144 5: [40960..43007]: 2174976..2177023 2048 0x801 <power fail> # Mount fs again, trigger log replay. $ mount /dev/sdc /mnt # Extent list after power failure and log replay. $ xfs_io -c "fiemap -v" /mnt/foo /mnt/foo: EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS 0: [0..7]: 2178048..2178055 8 0x0 1: [8..16383]: 26632..43007 16376 0x0 2: [16384..32767]: 2156544..2172927 16384 0x1 # The prealloc extents at file offsets 16M and 20M are missing. So fix this by calling btrfs_log_prealloc_extents() when we are doing a full fsync, so that we always log all prealloc extents beyond eof. A test case for fstests will follow soon. CC: stable@vger.kernel.org # 4.19+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-02-17 15:12:02 +03:00
} else if (min_key->type == BTRFS_EXTENT_DATA_KEY &&
min_key->offset >= i_size) {
/*
* Extents at and beyond eof are logged with
* btrfs_log_prealloc_extents().
* Only regular files have BTRFS_EXTENT_DATA_KEY keys,
* and no keys greater than that, so bail out.
*/
break;
} else if ((min_key->type == BTRFS_INODE_REF_KEY ||
min_key->type == BTRFS_INODE_EXTREF_KEY) &&
inode->generation == trans->transid &&
!recursive_logging) {
u64 other_ino = 0;
u64 other_parent = 0;
ret = btrfs_check_ref_name_override(path->nodes[0],
path->slots[0], min_key, inode,
&other_ino, &other_parent);
if (ret < 0) {
return ret;
} else if (ret > 0 &&
other_ino != btrfs_ino(BTRFS_I(ctx->inode))) {
if (ins_nr > 0) {
ins_nr++;
} else {
ins_nr = 1;
ins_start_slot = path->slots[0];
}
ret = copy_items(trans, inode, dst_path, path,
ins_start_slot, ins_nr,
inode_only, logged_isize);
if (ret < 0)
return ret;
ins_nr = 0;
ret = log_conflicting_inodes(trans, root, path,
ctx, other_ino, other_parent);
if (ret)
return ret;
btrfs_release_path(path);
goto next_key;
}
btrfs: fix lost prealloc extents beyond eof after full fsync When doing a full fsync, if we have prealloc extents beyond (or at) eof, and the leaves that contain them were not modified in the current transaction, we end up not logging them. This results in losing those extents when we replay the log after a power failure, since the inode is truncated to the current value of the logged i_size. Just like for the fast fsync path, we need to always log all prealloc extents starting at or beyond i_size. The fast fsync case was fixed in commit 471d557afed155 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay") but it missed the full fsync path. The problem exists since the very early days, when the log tree was added by commit e02119d5a7b439 ("Btrfs: Add a write ahead tree log to optimize synchronous operations"). Example reproducer: $ mkfs.btrfs -f /dev/sdc $ mount /dev/sdc /mnt # Create our test file with many file extent items, so that they span # several leaves of metadata, even if the node/page size is 64K. Use # direct IO and not fsync/O_SYNC because it's both faster and it avoids # clearing the full sync flag from the inode - we want the fsync below # to trigger the slow full sync code path. $ xfs_io -f -d -c "pwrite -b 4K 0 16M" /mnt/foo # Now add two preallocated extents to our file without extending the # file's size. One right at i_size, and another further beyond, leaving # a gap between the two prealloc extents. $ xfs_io -c "falloc -k 16M 1M" /mnt/foo $ xfs_io -c "falloc -k 20M 1M" /mnt/foo # Make sure everything is durably persisted and the transaction is # committed. This makes all created extents to have a generation lower # than the generation of the transaction used by the next write and # fsync. sync # Now overwrite only the first extent, which will result in modifying # only the first leaf of metadata for our inode. Then fsync it. This # fsync will use the slow code path (inode full sync bit is set) because # it's the first fsync since the inode was created/loaded. $ xfs_io -c "pwrite 0 4K" -c "fsync" /mnt/foo # Extent list before power failure. $ xfs_io -c "fiemap -v" /mnt/foo /mnt/foo: EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS 0: [0..7]: 2178048..2178055 8 0x0 1: [8..16383]: 26632..43007 16376 0x0 2: [16384..32767]: 2156544..2172927 16384 0x0 3: [32768..34815]: 2172928..2174975 2048 0x800 4: [34816..40959]: hole 6144 5: [40960..43007]: 2174976..2177023 2048 0x801 <power fail> # Mount fs again, trigger log replay. $ mount /dev/sdc /mnt # Extent list after power failure and log replay. $ xfs_io -c "fiemap -v" /mnt/foo /mnt/foo: EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS 0: [0..7]: 2178048..2178055 8 0x0 1: [8..16383]: 26632..43007 16376 0x0 2: [16384..32767]: 2156544..2172927 16384 0x1 # The prealloc extents at file offsets 16M and 20M are missing. So fix this by calling btrfs_log_prealloc_extents() when we are doing a full fsync, so that we always log all prealloc extents beyond eof. A test case for fstests will follow soon. CC: stable@vger.kernel.org # 4.19+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-02-17 15:12:02 +03:00
} else if (min_key->type == BTRFS_XATTR_ITEM_KEY) {
/* Skip xattrs, logged later with btrfs_log_all_xattrs() */
if (ins_nr == 0)
goto next_slot;
ret = copy_items(trans, inode, dst_path, path,
ins_start_slot,
ins_nr, inode_only, logged_isize);
if (ret < 0)
return ret;
ins_nr = 0;
goto next_slot;
}
if (ins_nr && ins_start_slot + ins_nr == path->slots[0]) {
ins_nr++;
goto next_slot;
} else if (!ins_nr) {
ins_start_slot = path->slots[0];
ins_nr = 1;
goto next_slot;
}
ret = copy_items(trans, inode, dst_path, path, ins_start_slot,
ins_nr, inode_only, logged_isize);
if (ret < 0)
return ret;
ins_nr = 1;
ins_start_slot = path->slots[0];
next_slot:
path->slots[0]++;
if (path->slots[0] < btrfs_header_nritems(path->nodes[0])) {
btrfs_item_key_to_cpu(path->nodes[0], min_key,
path->slots[0]);
goto again;
}
if (ins_nr) {
ret = copy_items(trans, inode, dst_path, path,
ins_start_slot, ins_nr, inode_only,
logged_isize);
if (ret < 0)
return ret;
ins_nr = 0;
}
btrfs_release_path(path);
next_key:
if (min_key->offset < (u64)-1) {
min_key->offset++;
} else if (min_key->type < max_key->type) {
min_key->type++;
min_key->offset = 0;
} else {
break;
}
}
btrfs: fix lost prealloc extents beyond eof after full fsync When doing a full fsync, if we have prealloc extents beyond (or at) eof, and the leaves that contain them were not modified in the current transaction, we end up not logging them. This results in losing those extents when we replay the log after a power failure, since the inode is truncated to the current value of the logged i_size. Just like for the fast fsync path, we need to always log all prealloc extents starting at or beyond i_size. The fast fsync case was fixed in commit 471d557afed155 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay") but it missed the full fsync path. The problem exists since the very early days, when the log tree was added by commit e02119d5a7b439 ("Btrfs: Add a write ahead tree log to optimize synchronous operations"). Example reproducer: $ mkfs.btrfs -f /dev/sdc $ mount /dev/sdc /mnt # Create our test file with many file extent items, so that they span # several leaves of metadata, even if the node/page size is 64K. Use # direct IO and not fsync/O_SYNC because it's both faster and it avoids # clearing the full sync flag from the inode - we want the fsync below # to trigger the slow full sync code path. $ xfs_io -f -d -c "pwrite -b 4K 0 16M" /mnt/foo # Now add two preallocated extents to our file without extending the # file's size. One right at i_size, and another further beyond, leaving # a gap between the two prealloc extents. $ xfs_io -c "falloc -k 16M 1M" /mnt/foo $ xfs_io -c "falloc -k 20M 1M" /mnt/foo # Make sure everything is durably persisted and the transaction is # committed. This makes all created extents to have a generation lower # than the generation of the transaction used by the next write and # fsync. sync # Now overwrite only the first extent, which will result in modifying # only the first leaf of metadata for our inode. Then fsync it. This # fsync will use the slow code path (inode full sync bit is set) because # it's the first fsync since the inode was created/loaded. $ xfs_io -c "pwrite 0 4K" -c "fsync" /mnt/foo # Extent list before power failure. $ xfs_io -c "fiemap -v" /mnt/foo /mnt/foo: EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS 0: [0..7]: 2178048..2178055 8 0x0 1: [8..16383]: 26632..43007 16376 0x0 2: [16384..32767]: 2156544..2172927 16384 0x0 3: [32768..34815]: 2172928..2174975 2048 0x800 4: [34816..40959]: hole 6144 5: [40960..43007]: 2174976..2177023 2048 0x801 <power fail> # Mount fs again, trigger log replay. $ mount /dev/sdc /mnt # Extent list after power failure and log replay. $ xfs_io -c "fiemap -v" /mnt/foo /mnt/foo: EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS 0: [0..7]: 2178048..2178055 8 0x0 1: [8..16383]: 26632..43007 16376 0x0 2: [16384..32767]: 2156544..2172927 16384 0x1 # The prealloc extents at file offsets 16M and 20M are missing. So fix this by calling btrfs_log_prealloc_extents() when we are doing a full fsync, so that we always log all prealloc extents beyond eof. A test case for fstests will follow soon. CC: stable@vger.kernel.org # 4.19+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-02-17 15:12:02 +03:00
if (ins_nr) {
ret = copy_items(trans, inode, dst_path, path, ins_start_slot,
ins_nr, inode_only, logged_isize);
btrfs: fix lost prealloc extents beyond eof after full fsync When doing a full fsync, if we have prealloc extents beyond (or at) eof, and the leaves that contain them were not modified in the current transaction, we end up not logging them. This results in losing those extents when we replay the log after a power failure, since the inode is truncated to the current value of the logged i_size. Just like for the fast fsync path, we need to always log all prealloc extents starting at or beyond i_size. The fast fsync case was fixed in commit 471d557afed155 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay") but it missed the full fsync path. The problem exists since the very early days, when the log tree was added by commit e02119d5a7b439 ("Btrfs: Add a write ahead tree log to optimize synchronous operations"). Example reproducer: $ mkfs.btrfs -f /dev/sdc $ mount /dev/sdc /mnt # Create our test file with many file extent items, so that they span # several leaves of metadata, even if the node/page size is 64K. Use # direct IO and not fsync/O_SYNC because it's both faster and it avoids # clearing the full sync flag from the inode - we want the fsync below # to trigger the slow full sync code path. $ xfs_io -f -d -c "pwrite -b 4K 0 16M" /mnt/foo # Now add two preallocated extents to our file without extending the # file's size. One right at i_size, and another further beyond, leaving # a gap between the two prealloc extents. $ xfs_io -c "falloc -k 16M 1M" /mnt/foo $ xfs_io -c "falloc -k 20M 1M" /mnt/foo # Make sure everything is durably persisted and the transaction is # committed. This makes all created extents to have a generation lower # than the generation of the transaction used by the next write and # fsync. sync # Now overwrite only the first extent, which will result in modifying # only the first leaf of metadata for our inode. Then fsync it. This # fsync will use the slow code path (inode full sync bit is set) because # it's the first fsync since the inode was created/loaded. $ xfs_io -c "pwrite 0 4K" -c "fsync" /mnt/foo # Extent list before power failure. $ xfs_io -c "fiemap -v" /mnt/foo /mnt/foo: EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS 0: [0..7]: 2178048..2178055 8 0x0 1: [8..16383]: 26632..43007 16376 0x0 2: [16384..32767]: 2156544..2172927 16384 0x0 3: [32768..34815]: 2172928..2174975 2048 0x800 4: [34816..40959]: hole 6144 5: [40960..43007]: 2174976..2177023 2048 0x801 <power fail> # Mount fs again, trigger log replay. $ mount /dev/sdc /mnt # Extent list after power failure and log replay. $ xfs_io -c "fiemap -v" /mnt/foo /mnt/foo: EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS 0: [0..7]: 2178048..2178055 8 0x0 1: [8..16383]: 26632..43007 16376 0x0 2: [16384..32767]: 2156544..2172927 16384 0x1 # The prealloc extents at file offsets 16M and 20M are missing. So fix this by calling btrfs_log_prealloc_extents() when we are doing a full fsync, so that we always log all prealloc extents beyond eof. A test case for fstests will follow soon. CC: stable@vger.kernel.org # 4.19+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-02-17 15:12:02 +03:00
if (ret)
return ret;
}
if (inode_only == LOG_INODE_ALL && S_ISREG(inode->vfs_inode.i_mode)) {
/*
* Release the path because otherwise we might attempt to double
* lock the same leaf with btrfs_log_prealloc_extents() below.
*/
btrfs_release_path(path);
ret = btrfs_log_prealloc_extents(trans, inode, dst_path);
}
return ret;
}
/* log a single inode in the tree log.
* At least one parent directory for this inode must exist in the tree
* or be logged already.
*
* Any items from this inode changed by the current transaction are copied
* to the log tree. An extra reference is taken on any extents in this
* file, allowing us to avoid a whole pile of corner cases around logging
* blocks that have been removed from the tree.
*
* See LOG_INODE_ALL and related defines for a description of what inode_only
* does.
*
* This handles both files and directories.
*/
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 17:24:20 +03:00
static int btrfs_log_inode(struct btrfs_trans_handle *trans,
struct btrfs_inode *inode,
int inode_only,
Btrfs: fix data corruption after fast fsync and writeback error When we do a fast fsync, we start all ordered operations and then while they're running in parallel we visit the list of modified extent maps and construct their matching file extent items and write them to the log btree. After that, in btrfs_sync_log() we wait for all the ordered operations to finish (via btrfs_wait_logged_extents). The problem with this is that we were completely ignoring errors that can happen in the extent write path, such as -ENOSPC, a temporary -ENOMEM or -EIO errors for example. When such error happens, it means we have parts of the on disk extent that weren't written to, and so we end up logging file extent items that point to these extents that contain garbage/random data - so after a crash/reboot plus log replay, we get our inode's metadata pointing to those extents. This worked in contrast with the full (non-fast) fsync path, where we start all ordered operations, wait for them to finish and then write to the log btree. In this path, after each ordered operation completes we check if it's flagged with an error (BTRFS_ORDERED_IOERR) and return -EIO if so (via btrfs_wait_ordered_range). So if an error happens with any ordered operation, just return a -EIO error to userspace, so that it knows that not all of its previous writes were durably persisted and the application can take proper action (like redo the writes for e.g.) - and definitely not leave any file extent items in the log refer to non fully written extents. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2014-09-05 18:14:39 +04:00
struct btrfs_log_ctx *ctx)
{
struct btrfs_path *path;
struct btrfs_path *dst_path;
struct btrfs_key min_key;
struct btrfs_key max_key;
struct btrfs_root *log = inode->root->log_root;
int ret;
Btrfs: turbo charge fsync At least for the vm workload. Currently on fsync we will 1) Truncate all items in the log tree for the given inode if they exist and 2) Copy all items for a given inode into the log The problem with this is that for things like VMs you can have lots of extents from the fragmented writing behavior, and worst yet you may have only modified a few extents, not the entire thing. This patch fixes this problem by tracking which transid modified our extent, and then when we do the tree logging we find all of the extents we've modified in our current transaction, sort them and commit them. We also only truncate up to the xattrs of the inode and copy that stuff in normally, and then just drop any extents in the range we have that exist in the log already. Here are some numbers of a 50 meg fio job that does random writes and fsync()s after every write Original Patched SATA drive 82KB/s 140KB/s Fusion drive 431KB/s 2532KB/s So around 2-6 times faster depending on your hardware. There are a few corner cases, for example if you truncate at all we have to do it the old way since there is no way to be sure what is in the log is ok. This probably could be done smarter, but if you write-fsync-truncate-write-fsync you deserve what you get. All this work is in RAM of course so if your inode gets evicted from cache and you read it in and fsync it we'll do it the slow way if we are still in the same transaction that we last modified the inode in. The biggest cool part of this is that it requires no changes to the recovery code, so if you fsync with this patch and crash and load an old kernel, it will run the recovery and be a-ok. I have tested this pretty thoroughly with an fsync tester and everything comes back fine, as well as xfstests. Thanks, Signed-off-by: Josef Bacik <jbacik@fusionio.com>
2012-08-17 21:14:17 +04:00
bool fast_search = false;
u64 ino = btrfs_ino(inode);
struct extent_map_tree *em_tree = &inode->extent_tree;
Btrfs: fix fsync data loss after adding hard link to inode We have a scenario where after the fsync log replay we can lose file data that had been previously fsync'ed if we added an hard link for our inode and after that we sync'ed the fsync log (for example by fsync'ing some other file or directory). This is because when adding an hard link we updated the inode item in the log tree with an i_size value of 0. At that point the new inode item was in memory only and a subsequent fsync log replay would not make us lose the file data. However if after adding the hard link we sync the log tree to disk, by fsync'ing some other file or directory for example, we ended up losing the file data after log replay, because the inode item in the persisted log tree had an an i_size of zero. This is easy to reproduce, and the following excerpt from my test for xfstests shows this: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create one file with data and fsync it. # This made the btrfs fsync log persist the data and the inode metadata with # a correct inode->i_size (4096 bytes). $XFS_IO_PROG -f -c "pwrite -S 0xaa -b 4K 0 4K" -c "fsync" \ $SCRATCH_MNT/foo | _filter_xfs_io # Now add one hard link to our file. This made the btrfs code update the fsync # log, in memory only, with an inode metadata having a size of 0. ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link # Now force persistence of the fsync log to disk, for example, by fsyncing some # other file. touch $SCRATCH_MNT/bar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/bar # Before a power loss or crash, we could read the 4Kb of data from our file as # expected. echo "File content before:" od -t x1 $SCRATCH_MNT/foo # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # After the fsync log replay, because the fsync log had a value of 0 for our # inode's i_size, we couldn't read anymore the 4Kb of data that we previously # wrote and fsync'ed. The size of the file became 0 after the fsync log replay. echo "File content after:" od -t x1 $SCRATCH_MNT/foo Another alternative test, that doesn't need to fsync an inode in the same transaction it was created, is: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create our test file with some data. $XFS_IO_PROG -f -c "pwrite -S 0xaa -b 8K 0 8K" \ $SCRATCH_MNT/foo | _filter_xfs_io # Make sure the file is durably persisted. sync # Append some data to our file, to increase its size. $XFS_IO_PROG -f -c "pwrite -S 0xcc -b 4K 8K 4K" \ $SCRATCH_MNT/foo | _filter_xfs_io # Fsync the file, so from this point on if a crash/power failure happens, our # new data is guaranteed to be there next time the fs is mounted. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foo # Add one hard link to our file. This made btrfs write into the in memory fsync # log a special inode with generation 0 and an i_size of 0 too. Note that this # didn't update the inode in the fsync log on disk. ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link # Now make sure the in memory fsync log is durably persisted. # Creating and fsync'ing another file will do it. touch $SCRATCH_MNT/bar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/bar # As expected, before the crash/power failure, we should be able to read the # 12Kb of file data. echo "File content before:" od -t x1 $SCRATCH_MNT/foo # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # After mounting the fs again, the fsync log was replayed. # The btrfs fsync log replay code didn't update the i_size of the persisted # inode because the inode item in the log had a special generation with a # value of 0 (and it couldn't know the correct i_size, since that inode item # had a 0 i_size too). This made the last 4Kb of file data inaccessible and # effectively lost. echo "File content after:" od -t x1 $SCRATCH_MNT/foo This isn't a new issue/regression. This problem has been around since the log tree code was added in 2008: Btrfs: Add a write ahead tree log to optimize synchronous operations (commit e02119d5a7b4396c5a872582fddc8bd6d305a70a) Test cases for xfstests follow soon. CC: <stable@vger.kernel.org> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-02-13 15:30:56 +03:00
u64 logged_isize = 0;
Btrfs: fix fsync data loss after append write If we do an append write to a file (which increases its inode's i_size) that does not have the flag BTRFS_INODE_NEEDS_FULL_SYNC set in its inode, and the previous transaction added a new hard link to the file, which sets the flag BTRFS_INODE_COPY_EVERYTHING in the file's inode, and then fsync the file, the inode's new i_size isn't logged. This has the consequence that after the fsync log is replayed, the file size remains what it was before the append write operation, which means users/applications will not be able to read the data that was successsfully fsync'ed before. This happens because neither the inode item nor the delayed inode get their i_size updated when the append write is made - doing so would require starting a transaction in the buffered write path, something that we do not do intentionally for performance reasons. Fix this by making sure that when the flag BTRFS_INODE_COPY_EVERYTHING is set the inode is logged with its current i_size (log the in-memory inode into the log tree). This issue is not a recent regression and is easy to reproduce with the following test case for fstests: seq=`basename $0` seqres=$RESULT_DIR/$seq echo "QA output created by $seq" here=`pwd` tmp=/tmp/$$ status=1 # failure is the default! _cleanup() { _cleanup_flakey rm -f $tmp.* } trap "_cleanup; exit \$status" 0 1 2 3 15 # get standard environment, filters and checks . ./common/rc . ./common/filter . ./common/dmflakey # real QA test starts here _supported_fs generic _supported_os Linux _need_to_be_root _require_scratch _require_dm_flakey _require_metadata_journaling $SCRATCH_DEV _crash_and_mount() { # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again and mount. This makes the fs replay its fsync log. _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey } rm -f $seqres.full _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create the test file with some initial data and then fsync it. # The fsync here is only needed to trigger the issue in btrfs, as it causes the # the flag BTRFS_INODE_NEEDS_FULL_SYNC to be removed from the btrfs inode. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 32k" \ -c "fsync" \ $SCRATCH_MNT/foo | _filter_xfs_io sync # Add a hard link to our file. # On btrfs this sets the flag BTRFS_INODE_COPY_EVERYTHING on the btrfs inode, # which is a necessary condition to trigger the issue. ln $SCRATCH_MNT/foo $SCRATCH_MNT/bar # Sync the filesystem to force a commit of the current btrfs transaction, this # is a necessary condition to trigger the bug on btrfs. sync # Now append more data to our file, increasing its size, and fsync the file. # In btrfs because the inode flag BTRFS_INODE_COPY_EVERYTHING was set and the # write path did not update the inode item in the btree nor the delayed inode # item (in memory struture) in the current transaction (created by the fsync # handler), the fsync did not record the inode's new i_size in the fsync # log/journal. This made the data unavailable after the fsync log/journal is # replayed. $XFS_IO_PROG -c "pwrite -S 0xbb 32K 32K" \ -c "fsync" \ $SCRATCH_MNT/foo | _filter_xfs_io echo "File content after fsync and before crash:" od -t x1 $SCRATCH_MNT/foo _crash_and_mount echo "File content after crash and log replay:" od -t x1 $SCRATCH_MNT/foo status=0 exit The expected file output before and after the crash/power failure expects the appended data to be available, which is: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0100000 bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb * 0200000 Cc: stable@vger.kernel.org Signed-off-by: Filipe Manana <fdmanana@suse.com> Reviewed-by: Liu Bo <bo.li.liu@oracle.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-06-17 14:49:23 +03:00
bool need_log_inode_item = true;
bool xattrs_logged = false;
Btrfs: fix fsync after succession of renames and unlink/rmdir After a succession of renames operations of different files and unlinking one of them, if we fsync one of the renamed files we can end up with a log that will either fail to replay at mount time or result in a filesystem that is in an inconsistent state. One example scenario: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/fname1 $ touch /mnt/testdir/fname2 $ sync $ mv /mnt/testdir/fname1 /mnt/testdir/fname3 $ rm -f /mnt/testdir/fname2 $ ln /mnt/testdir/fname3 /mnt/testdir/fname2 $ touch /mnt/testdir/fname1 $ xfs_io -c "fsync" /mnt/testdir/fname1 <power failure> $ mount /dev/sdb /mnt $ umount /mnt $ btrfs check /dev/sdb [1/7] checking root items [2/7] checking extents [3/7] checking free space cache [4/7] checking fs roots root 5 inode 259 errors 2, no orphan item ERROR: errors found in fs roots Opening filesystem to check... Checking filesystem on /dev/sdc UUID: 20e4abb8-5a19-4492-8bb4-6084125c2d0d found 393216 bytes used, error(s) found total csum bytes: 0 total tree bytes: 131072 total fs tree bytes: 32768 total extent tree bytes: 16384 btree space waste bytes: 122986 file data blocks allocated: 262144 referenced 262144 On a kernel without the first patch in this series, titled "[PATCH] Btrfs: fix fsync after succession of renames of different files", we get instead an error when mounting the filesystem due to failure of replaying the log: $ mount /dev/sdb /mnt mount: mount /dev/sdb on /mnt failed: File exists Fix this by logging the parent directory of an inode whenever we find an inode that no longer exists (was unlinked in the current transaction), during the procedure which finds inodes that have old names that collide with new names of other inodes. A test case for fstests follows soon. Reviewed-by: Nikolay Borisov <nborisov@suse.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-02-13 15:14:09 +03:00
bool recursive_logging = false;
btrfs: avoid unnecessary lock and leaf splits when updating inode in the log During a fast fsync, if we have already fsynced the file before and in the current transaction, we can make the inode item update more efficient and avoid acquiring a write lock on the leaf's parent. To update the inode item we are always using btrfs_insert_empty_item() to get a path pointing to the inode item, which calls btrfs_search_slot() with an "ins_len" argument of 'sizeof(struct btrfs_inode_item) + sizeof(struct btrfs_item)', and that always results in the search taking a write lock on the level 1 node that is the parent of the leaf that contains the inode item. This adds unnecessary lock contention on log trees when we have multiple fsyncs in parallel against inodes in the same subvolume, which has a very significant impact due to the fact that log trees are short lived and their height very rarely goes beyond level 2. Also, by using btrfs_insert_empty_item() when we need to update the inode item, we also end up splitting the leaf of the existing inode item when the leaf has an amount of free space smaller than the size of an inode item. Improve this by using btrfs_seach_slot(), with a 0 "ins_len" argument, when we know the inode item already exists in the log. This avoids these two inefficiencies. The following script, using fio, was used to perform the tests: $ cat fio-test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 MOUNT_OPTIONS="-o ssd" MKFS_OPTIONS="-d single -m single" if [ $# -ne 4 ]; then echo "Use $0 NUM_JOBS FILE_SIZE FSYNC_FREQ BLOCK_SIZE" exit 1 fi NUM_JOBS=$1 FILE_SIZE=$2 FSYNC_FREQ=$3 BLOCK_SIZE=$4 cat <<EOF > /tmp/fio-job.ini [writers] rw=randwrite fsync=$FSYNC_FREQ fallocate=none group_reporting=1 direct=0 bs=$BLOCK_SIZE ioengine=sync size=$FILE_SIZE directory=$MNT numjobs=$NUM_JOBS EOF echo "performance" | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor echo echo "Using config:" echo cat /tmp/fio-job.ini echo echo "mount options: $MOUNT_OPTIONS" echo umount $MNT &> /dev/null mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT fio /tmp/fio-job.ini umount $MNT The tests were done on a physical machine, with 12 cores, 64G of RAM, using a NVMEe device and using a non-debug kernel config (the default one from Debian). The summary line from fio is provided below for each test run. With 8 jobs, file size 256M, fsync frequency of 4 and a block size of 4K: Before: WRITE: bw=28.3MiB/s (29.7MB/s), 28.3MiB/s-28.3MiB/s (29.7MB/s-29.7MB/s), io=2048MiB (2147MB), run=72297-72297msec After: WRITE: bw=28.7MiB/s (30.1MB/s), 28.7MiB/s-28.7MiB/s (30.1MB/s-30.1MB/s), io=2048MiB (2147MB), run=71411-71411msec +1.4% throughput, -1.2% runtime With 16 jobs, file size 256M, fsync frequency of 4 and a block size of 4K: Before: WRITE: bw=40.0MiB/s (42.0MB/s), 40.0MiB/s-40.0MiB/s (42.0MB/s-42.0MB/s), io=4096MiB (4295MB), run=99980-99980msec After: WRITE: bw=40.9MiB/s (42.9MB/s), 40.9MiB/s-40.9MiB/s (42.9MB/s-42.9MB/s), io=4096MiB (4295MB), run=97933-97933msec +2.2% throughput, -2.1% runtime The changes are small but it's possible to be better on faster hardware as in the test machine used disk utilization was pretty much 100% during the whole time the tests were running (observed with 'iostat -xz 1'). The tests also included the previous patch with the subject of: "btrfs: avoid unnecessary log mutex contention when syncing log". So they compared a branch without that patch and without this patch versus a branch with these two patches applied. Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-07-20 18:03:43 +03:00
bool inode_item_dropped = true;
btrfs: avoid inode logging during rename and link when possible During a rename or link operation, we need to determine if an inode was previously logged or not, and if it was, do some update to the logged inode. We used to rely exclusively on the logged_trans field of struct btrfs_inode to determine that, but that was not reliable because the value of that field is not persisted in the inode item, so it's lost when an inode is evicted and loaded back again. That led to several issues in the past, such as not persisting deletions (such as the case fixed by commit 803f0f64d17769 ("Btrfs: fix fsync not persisting dentry deletions due to inode evictions")), or resulting in losing a file after an inode eviction followed by a rename (commit ecc64fab7d49c6 ("btrfs: fix lost inode on log replay after mix of fsync, rename and inode eviction")), besides other issues. So the inode_logged() helper was introduced and used to determine if an inode was possibly logged before in the current transaction, with the caveat that it could return false positives, in the sense that even if an inode was not logged before in the current transaction, it could still return true, but never to return false in case the inode was logged. >From a functional point of view that is fine, but from a performance perspective it can introduce significant latencies to rename and link operations, as they will end up doing inode logging even when it is not necessary. Recently on a 5.15 kernel, an openSUSE Tumbleweed user reported package installations and upgrades, with the zypper tool, were often taking a long time to complete. With strace it could be observed that zypper was spending about 99% of its time on rename operations, and then with further analysis we checked that directory logging was happening too frequently. Taking into account that installation/upgrade of some of the packages needed a few thousand file renames, the slowdown was very noticeable for the user. The issue was caused indirectly due to an excessive number of inode evictions on a 5.15 kernel, about 100x more compared to a 5.13, 5.14 or a 5.16-rc8 kernel. While triggering the inode evictions if something outside btrfs' control, btrfs could still behave better by eliminating the false positives from the inode_logged() helper. So change inode_logged() to actually eliminate such false positives caused by inode eviction and when an inode was never logged since the filesystem was mounted, as both cases relate to when the logged_trans field of struct btrfs_inode has a value of zero. When it can not determine if the inode was logged based only on the logged_trans value, lookup for the existence of the inode item in the log tree - if it's there then we known the inode was logged, if it's not there then it can not have been logged in the current transaction. Once we determine if the inode was logged, update the logged_trans value to avoid future calls to have to search in the log tree again. Alternatively, we could start storing logged_trans in the on disk inode item structure (struct btrfs_inode_item) in the unused space it still has, but that would be a bit odd because: 1) We only care about logged_trans since the filesystem was mounted, we don't care about its value from a previous mount. Having it persisted in the inode item structure would not make the best use of the precious unused space; 2) In order to get logged_trans persisted before inode eviction, we would have to update the delayed inode when we finish logging the inode and update its logged_trans in struct btrfs_inode, which makes it a bit cumbersome since we need to check if the delayed inode exists, if not create it and populate it and deal with any errors (-ENOMEM mostly). This change is part of a patchset comprised of the following patches: 1/5 btrfs: add helper to delete a dir entry from a log tree 2/5 btrfs: pass the dentry to btrfs_log_new_name() instead of the inode 3/5 btrfs: avoid logging all directory changes during renames 4/5 btrfs: stop doing unnecessary log updates during a rename 5/5 btrfs: avoid inode logging during rename and link when possible The following test script mimics part of what the zypper tool does during package installations/upgrades. It does not triggers inode evictions, but it's similar because it triggers false positives from the inode_logged() helper, because the inodes have a logged_trans of 0, there's a log tree due to a fsync of an unrelated file and the directory inode has its last_trans field set to the current transaction: $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=10000 mkfs.btrfs -f $DEV mount $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync # Now do some change to an unrelated file and fsync it. # This is just to create a log tree to make sure that inode_logged() # does not return false when called against "testdir". xfs_io -f -c "pwrite 0 4K" -c "fsync" $MNT/foo # Do some change to testdir. This is to make sure inode_logged() # will return true when called against "testdir", because its # logged_trans is 0, it was changed in the current transaction # and there's a log tree. echo -n > $MNT/testdir/file_$((NUM_FILES + 1)) echo "Renaming $NUM_FILES files..." start=$(date +%s%N) for ((i = 1; i <= $NUM_FILES; i++)); do mv $MNT/testdir/file_$i $MNT/testdir/file_$i-RPMDELETE done end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "Renames took $dur milliseconds" umount $MNT Testing this change on a box using a non-debug kernel (Debian's default kernel config) gave the following results: NUM_FILES=10000, before patchset: 27837 ms NUM_FILES=10000, after patches 1/5 to 4/5 applied: 9236 ms (-66.8%) NUM_FILES=10000, after whole patchset applied: 8902 ms (-68.0%) NUM_FILES=5000, before patchset: 9127 ms NUM_FILES=5000, after patches 1/5 to 4/5 applied: 4640 ms (-49.2%) NUM_FILES=5000, after whole patchset applied: 4441 ms (-51.3%) NUM_FILES=2000, before patchset: 2528 ms NUM_FILES=2000, after patches 1/5 to 4/5 applied: 1983 ms (-21.6%) NUM_FILES=2000, after whole patchset applied: 1747 ms (-30.9%) NUM_FILES=1000, before patchset: 1085 ms NUM_FILES=1000, after patches 1/5 to 4/5 applied: 893 ms (-17.7%) NUM_FILES=1000, after whole patchset applied: 867 ms (-20.1%) Running dbench on the same physical machine with the following script: $ cat run-dbench.sh #!/bin/bash NUM_JOBS=$(nproc --all) DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 MOUNT_OPTIONS="-o ssd" MKFS_OPTIONS="-O no-holes -R free-space-tree" echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -D $MNT -t 120 $NUM_JOBS umount $MNT Before patchset: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 3761352 0.032 143.843 Close 2762770 0.002 2.273 Rename 159304 0.291 67.037 Unlink 759784 0.207 143.998 Deltree 72 4.028 15.977 Mkdir 36 0.003 0.006 Qpathinfo 3409780 0.013 9.678 Qfileinfo 596772 0.001 0.878 Qfsinfo 625189 0.003 1.245 Sfileinfo 306443 0.006 1.840 Find 1318106 0.063 19.798 WriteX 1871137 0.021 8.532 ReadX 5897325 0.003 3.567 LockX 12252 0.003 0.258 UnlockX 12252 0.002 0.100 Flush 263666 3.327 155.632 Throughput 980.047 MB/sec 12 clients 12 procs max_latency=155.636 ms After whole patchset applied: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 4195584 0.033 107.742 Close 3081932 0.002 1.935 Rename 177641 0.218 14.905 Unlink 847333 0.166 107.822 Deltree 118 5.315 15.247 Mkdir 59 0.004 0.048 Qpathinfo 3802612 0.014 10.302 Qfileinfo 666748 0.001 1.034 Qfsinfo 697329 0.003 0.944 Sfileinfo 341712 0.006 2.099 Find 1470365 0.065 9.359 WriteX 2093921 0.021 8.087 ReadX 6576234 0.003 3.407 LockX 13660 0.003 0.308 UnlockX 13660 0.002 0.114 Flush 294090 2.906 115.539 Throughput 1093.11 MB/sec 12 clients 12 procs max_latency=115.544 ms +11.5% throughput -25.8% max latency rename max latency -77.8% Link: https://bugzilla.opensuse.org/show_bug.cgi?id=1193549 Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-01-20 14:00:10 +03:00
const bool orig_logged_before = ctx->logged_before;
path = btrfs_alloc_path();
if (!path)
return -ENOMEM;
dst_path = btrfs_alloc_path();
if (!dst_path) {
btrfs_free_path(path);
return -ENOMEM;
}
min_key.objectid = ino;
min_key.type = BTRFS_INODE_ITEM_KEY;
min_key.offset = 0;
max_key.objectid = ino;
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 17:24:20 +03:00
Btrfs: turbo charge fsync At least for the vm workload. Currently on fsync we will 1) Truncate all items in the log tree for the given inode if they exist and 2) Copy all items for a given inode into the log The problem with this is that for things like VMs you can have lots of extents from the fragmented writing behavior, and worst yet you may have only modified a few extents, not the entire thing. This patch fixes this problem by tracking which transid modified our extent, and then when we do the tree logging we find all of the extents we've modified in our current transaction, sort them and commit them. We also only truncate up to the xattrs of the inode and copy that stuff in normally, and then just drop any extents in the range we have that exist in the log already. Here are some numbers of a 50 meg fio job that does random writes and fsync()s after every write Original Patched SATA drive 82KB/s 140KB/s Fusion drive 431KB/s 2532KB/s So around 2-6 times faster depending on your hardware. There are a few corner cases, for example if you truncate at all we have to do it the old way since there is no way to be sure what is in the log is ok. This probably could be done smarter, but if you write-fsync-truncate-write-fsync you deserve what you get. All this work is in RAM of course so if your inode gets evicted from cache and you read it in and fsync it we'll do it the slow way if we are still in the same transaction that we last modified the inode in. The biggest cool part of this is that it requires no changes to the recovery code, so if you fsync with this patch and crash and load an old kernel, it will run the recovery and be a-ok. I have tested this pretty thoroughly with an fsync tester and everything comes back fine, as well as xfstests. Thanks, Signed-off-by: Josef Bacik <jbacik@fusionio.com>
2012-08-17 21:14:17 +04:00
/* today the code can only do partial logging of directories */
if (S_ISDIR(inode->vfs_inode.i_mode) ||
(!test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
&inode->runtime_flags) &&
inode_only >= LOG_INODE_EXISTS))
max_key.type = BTRFS_XATTR_ITEM_KEY;
else
max_key.type = (u8)-1;
max_key.offset = (u64)-1;
Btrfs: fix fsync when extend references are added to an inode If we added an extended reference to an inode and fsync'ed it, the log replay code would make our inode have an incorrect link count, which was lower then the expected/correct count. This resulted in stale directory index entries after deleting some of the hard links, and any access to the dangling directory entries resulted in -ESTALE errors because the entries pointed to inode items that don't exist anymore. This is easy to reproduce with the test case I made for xfstests, and the bulk of that test is: _scratch_mkfs "-O extref" >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create a test file with 3001 hard links. This number is large enough to # make btrfs start using extrefs at some point even if the fs has the maximum # possible leaf/node size (64Kb). echo "hello world" > $SCRATCH_MNT/foo for i in `seq 1 3000`; do ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link_`printf "%04d" $i` done # Make sure all metadata and data are durably persisted. sync # Add one more link to the inode that ends up being a btrfs extref and fsync # the inode. ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link_3001 $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foo # Simulate a crash/power loss. This makes sure the next mount # will see an fsync log and will replay that log. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # Now after the fsync log replay btrfs left our inode with a wrong link count N, # which was smaller than the correct link count M (N < M). # So after removing N hard links, the remaining M - N directory entries were # still visible to user space but it was impossible to do anything with them # because they pointed to an inode that didn't exist anymore. This resulted in # stale file handle errors (-ESTALE) when accessing those dentries for example. # # So remove all hard links except the first one and then attempt to read the # file, to verify we don't get an -ESTALE error when accessing the inodel # # The btrfs fsck tool also detected the incorrect inode link count and it # reported an error message like the following: # # root 5 inode 257 errors 2001, no inode item, link count wrong # unresolved ref dir 256 index 2978 namelen 13 name foo_link_2976 filetype 1 errors 4, no inode ref # # The fstests framework automatically calls fsck after a test is run, so we # don't need to call fsck explicitly here. rm -f $SCRATCH_MNT/foo_link_* cat $SCRATCH_MNT/foo status=0 exit So make sure an fsync always flushes the delayed inode item, so that the fsync log contains it (needed in order to trigger the link count fixup code) and fix the extref counting function, which always return -ENOENT to its caller (and made it assume there were always 0 extrefs). This issue has been present since the introduction of the extrefs feature (2012). A test case for xfstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-01-13 19:40:04 +03:00
/*
btrfs: only commit delayed items at fsync if we are logging a directory When logging an inode we are committing its delayed items if either the inode is a directory or if it is a new inode, created in the current transaction. We need to do it for directories, since new directory indexes are stored as delayed items of the inode and when logging a directory we need to be able to access all indexes from the fs/subvolume tree in order to figure out which index ranges need to be logged. However for new inodes that are not directories, we do not need to do it because the only type of delayed item they can have is the inode item, and we are guaranteed to always log an up to date version of the inode item: *) for a full fsync we do it by committing the delayed inode and then copying the item from the fs/subvolume tree with copy_inode_items_to_log(); *) for a fast fsync we always log the inode item based on the contents of the in-memory struct btrfs_inode. We guarantee this is always done since commit e4545de5b035c7 ("Btrfs: fix fsync data loss after append write"). So stop running delayed items for a new inodes that are not directories, since that forces committing the delayed inode into the fs/subvolume tree, wasting time and adding contention to the tree when a full fsync is not required. We will only do it in case a fast fsync is needed. This patch is part of a series that has the following patches: 1/4 btrfs: only commit the delayed inode when doing a full fsync 2/4 btrfs: only commit delayed items at fsync if we are logging a directory 3/4 btrfs: stop incremening log_batch for the log root tree when syncing log 4/4 btrfs: remove no longer needed use of log_writers for the log root tree After the entire patchset applied I saw about 12% decrease on max latency reported by dbench. The test was done on a qemu vm, with 8 cores, 16Gb of ram, using kvm and using a raw NVMe device directly (no intermediary fs on the host). The test was invoked like the following: mkfs.btrfs -f /dev/sdk mount -o ssd -o nospace_cache /dev/sdk /mnt/sdk dbench -D /mnt/sdk -t 300 8 umount /mnt/dsk CC: stable@vger.kernel.org # 5.4+ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-07-02 14:32:20 +03:00
* Only run delayed items if we are a directory. We want to make sure
* all directory indexes hit the fs/subvolume tree so we can find them
* and figure out which index ranges have to be logged.
Btrfs: fix fsync when extend references are added to an inode If we added an extended reference to an inode and fsync'ed it, the log replay code would make our inode have an incorrect link count, which was lower then the expected/correct count. This resulted in stale directory index entries after deleting some of the hard links, and any access to the dangling directory entries resulted in -ESTALE errors because the entries pointed to inode items that don't exist anymore. This is easy to reproduce with the test case I made for xfstests, and the bulk of that test is: _scratch_mkfs "-O extref" >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create a test file with 3001 hard links. This number is large enough to # make btrfs start using extrefs at some point even if the fs has the maximum # possible leaf/node size (64Kb). echo "hello world" > $SCRATCH_MNT/foo for i in `seq 1 3000`; do ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link_`printf "%04d" $i` done # Make sure all metadata and data are durably persisted. sync # Add one more link to the inode that ends up being a btrfs extref and fsync # the inode. ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link_3001 $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foo # Simulate a crash/power loss. This makes sure the next mount # will see an fsync log and will replay that log. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # Now after the fsync log replay btrfs left our inode with a wrong link count N, # which was smaller than the correct link count M (N < M). # So after removing N hard links, the remaining M - N directory entries were # still visible to user space but it was impossible to do anything with them # because they pointed to an inode that didn't exist anymore. This resulted in # stale file handle errors (-ESTALE) when accessing those dentries for example. # # So remove all hard links except the first one and then attempt to read the # file, to verify we don't get an -ESTALE error when accessing the inodel # # The btrfs fsck tool also detected the incorrect inode link count and it # reported an error message like the following: # # root 5 inode 257 errors 2001, no inode item, link count wrong # unresolved ref dir 256 index 2978 namelen 13 name foo_link_2976 filetype 1 errors 4, no inode ref # # The fstests framework automatically calls fsck after a test is run, so we # don't need to call fsck explicitly here. rm -f $SCRATCH_MNT/foo_link_* cat $SCRATCH_MNT/foo status=0 exit So make sure an fsync always flushes the delayed inode item, so that the fsync log contains it (needed in order to trigger the link count fixup code) and fix the extref counting function, which always return -ENOENT to its caller (and made it assume there were always 0 extrefs). This issue has been present since the introduction of the extrefs feature (2012). A test case for xfstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-01-13 19:40:04 +03:00
*/
btrfs: do not commit delayed inode when logging a file in full sync mode When logging a regular file in full sync mode, we are currently committing its delayed inode item. This is to ensure that we never miss copying the inode item, with its most up to date data, into the log tree. However that is not necessary since commit e4545de5b035 ("Btrfs: fix fsync data loss after append write"), because even if we don't find the leaf with the inode item when looking for leaves that changed in the current transaction, we end up logging the inode item later using the in-memory content. In case we find the leaf containing the inode item, we already end up using the in-memory inode for filling the inode item in the log tree, and not the inode item that is in the fs/subvolume tree, as it might be not up to date (copy_items() -> fill_inode_item()). So don't commit the delayed inode item, which brings a couple of benefits: 1) Avoid writing the inode item to the fs/subvolume btree, saving time and reducing lock contention on the btree; 2) In case no other item for the inode was changed, added or deleted in the same leaf where the inode item is located, we ended up copying all the items in that leaf to the log tree - it's harmless from a functional point of view, but it wastes time and log tree space. This patch is part of a patch set comprised of the following patches: btrfs: check if a log tree exists at inode_logged() btrfs: remove no longer needed checks for NULL log context btrfs: do not log new dentries when logging that a new name exists btrfs: always update the logged transaction when logging new names btrfs: avoid expensive search when dropping inode items from log btrfs: add helper to truncate inode items when logging inode btrfs: avoid expensive search when truncating inode items from the log btrfs: avoid search for logged i_size when logging inode if possible btrfs: avoid attempt to drop extents when logging inode for the first time btrfs: do not commit delayed inode when logging a file in full sync mode This is patch 10/10 and the following test results compare a branch with the whole patch set applied versus a branch without any of the patches applied. The following script was used to test dbench with 8 and 16 jobs on a machine with 12 cores, 64G of RAM, a NVME device and using a non-debug kernel config (Debian's default): $ cat test.sh #!/bin/bash if [ $# -ne 1 ]; then echo "Use $0 NUM_JOBS" exit 1 fi NUM_JOBS=$1 DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 MOUNT_OPTIONS="-o ssd" MKFS_OPTIONS="-m single -d single" echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -D $MNT -t 120 $NUM_JOBS umount $MNT The results were the following: 8 jobs, before patchset: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 4113896 0.009 238.665 Close 3021699 0.001 0.590 Rename 174215 0.082 238.733 Unlink 830977 0.049 238.642 Deltree 96 2.232 8.022 Mkdir 48 0.003 0.005 Qpathinfo 3729013 0.005 2.672 Qfileinfo 653206 0.001 0.152 Qfsinfo 683866 0.002 0.526 Sfileinfo 335055 0.004 1.571 Find 1441800 0.016 4.288 WriteX 2049644 0.010 3.982 ReadX 6449786 0.003 0.969 LockX 13400 0.002 0.043 UnlockX 13400 0.001 0.075 Flush 288349 2.521 245.516 Throughput 1075.73 MB/sec 8 clients 8 procs max_latency=245.520 ms 8 jobs, after patchset: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 4154282 0.009 156.675 Close 3051450 0.001 0.843 Rename 175912 0.072 4.444 Unlink 839067 0.048 66.050 Deltree 96 2.131 5.979 Mkdir 48 0.002 0.004 Qpathinfo 3765575 0.005 3.079 Qfileinfo 659582 0.001 0.099 Qfsinfo 690474 0.002 0.155 Sfileinfo 338366 0.004 1.419 Find 1455816 0.016 3.423 WriteX 2069538 0.010 4.328 ReadX 6512429 0.003 0.840 LockX 13530 0.002 0.078 UnlockX 13530 0.001 0.051 Flush 291158 2.500 163.468 Throughput 1105.45 MB/sec 8 clients 8 procs max_latency=163.474 ms +2.7% throughput, -40.1% max latency 16 jobs, before patchset: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 5457602 0.033 337.098 Close 4008979 0.002 2.018 Rename 231051 0.323 254.054 Unlink 1102209 0.202 337.243 Deltree 160 6.521 31.720 Mkdir 80 0.003 0.007 Qpathinfo 4946147 0.014 6.988 Qfileinfo 867440 0.001 1.642 Qfsinfo 907081 0.003 1.821 Sfileinfo 444433 0.005 2.053 Find 1912506 0.067 7.854 WriteX 2724852 0.018 7.428 ReadX 8553883 0.003 2.059 LockX 17770 0.003 0.350 UnlockX 17770 0.002 0.627 Flush 382533 2.810 353.691 Throughput 1413.09 MB/sec 16 clients 16 procs max_latency=353.696 ms 16 jobs, after patchset: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 5393156 0.034 303.181 Close 3961986 0.002 1.502 Rename 228359 0.320 253.379 Unlink 1088920 0.206 303.409 Deltree 160 6.419 30.088 Mkdir 80 0.003 0.004 Qpathinfo 4887967 0.015 7.722 Qfileinfo 857408 0.001 1.651 Qfsinfo 896343 0.002 2.147 Sfileinfo 439317 0.005 4.298 Find 1890018 0.073 8.347 WriteX 2693356 0.018 6.373 ReadX 8453485 0.003 3.836 LockX 17562 0.003 0.486 UnlockX 17562 0.002 0.635 Flush 378023 2.802 315.904 Throughput 1454.46 MB/sec 16 clients 16 procs max_latency=315.910 ms +2.9% throughput, -11.3% max latency Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-08-31 17:30:40 +03:00
if (S_ISDIR(inode->vfs_inode.i_mode)) {
ret = btrfs_commit_inode_delayed_items(trans, inode);
if (ret)
btrfs: do not commit delayed inode when logging a file in full sync mode When logging a regular file in full sync mode, we are currently committing its delayed inode item. This is to ensure that we never miss copying the inode item, with its most up to date data, into the log tree. However that is not necessary since commit e4545de5b035 ("Btrfs: fix fsync data loss after append write"), because even if we don't find the leaf with the inode item when looking for leaves that changed in the current transaction, we end up logging the inode item later using the in-memory content. In case we find the leaf containing the inode item, we already end up using the in-memory inode for filling the inode item in the log tree, and not the inode item that is in the fs/subvolume tree, as it might be not up to date (copy_items() -> fill_inode_item()). So don't commit the delayed inode item, which brings a couple of benefits: 1) Avoid writing the inode item to the fs/subvolume btree, saving time and reducing lock contention on the btree; 2) In case no other item for the inode was changed, added or deleted in the same leaf where the inode item is located, we ended up copying all the items in that leaf to the log tree - it's harmless from a functional point of view, but it wastes time and log tree space. This patch is part of a patch set comprised of the following patches: btrfs: check if a log tree exists at inode_logged() btrfs: remove no longer needed checks for NULL log context btrfs: do not log new dentries when logging that a new name exists btrfs: always update the logged transaction when logging new names btrfs: avoid expensive search when dropping inode items from log btrfs: add helper to truncate inode items when logging inode btrfs: avoid expensive search when truncating inode items from the log btrfs: avoid search for logged i_size when logging inode if possible btrfs: avoid attempt to drop extents when logging inode for the first time btrfs: do not commit delayed inode when logging a file in full sync mode This is patch 10/10 and the following test results compare a branch with the whole patch set applied versus a branch without any of the patches applied. The following script was used to test dbench with 8 and 16 jobs on a machine with 12 cores, 64G of RAM, a NVME device and using a non-debug kernel config (Debian's default): $ cat test.sh #!/bin/bash if [ $# -ne 1 ]; then echo "Use $0 NUM_JOBS" exit 1 fi NUM_JOBS=$1 DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 MOUNT_OPTIONS="-o ssd" MKFS_OPTIONS="-m single -d single" echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -D $MNT -t 120 $NUM_JOBS umount $MNT The results were the following: 8 jobs, before patchset: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 4113896 0.009 238.665 Close 3021699 0.001 0.590 Rename 174215 0.082 238.733 Unlink 830977 0.049 238.642 Deltree 96 2.232 8.022 Mkdir 48 0.003 0.005 Qpathinfo 3729013 0.005 2.672 Qfileinfo 653206 0.001 0.152 Qfsinfo 683866 0.002 0.526 Sfileinfo 335055 0.004 1.571 Find 1441800 0.016 4.288 WriteX 2049644 0.010 3.982 ReadX 6449786 0.003 0.969 LockX 13400 0.002 0.043 UnlockX 13400 0.001 0.075 Flush 288349 2.521 245.516 Throughput 1075.73 MB/sec 8 clients 8 procs max_latency=245.520 ms 8 jobs, after patchset: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 4154282 0.009 156.675 Close 3051450 0.001 0.843 Rename 175912 0.072 4.444 Unlink 839067 0.048 66.050 Deltree 96 2.131 5.979 Mkdir 48 0.002 0.004 Qpathinfo 3765575 0.005 3.079 Qfileinfo 659582 0.001 0.099 Qfsinfo 690474 0.002 0.155 Sfileinfo 338366 0.004 1.419 Find 1455816 0.016 3.423 WriteX 2069538 0.010 4.328 ReadX 6512429 0.003 0.840 LockX 13530 0.002 0.078 UnlockX 13530 0.001 0.051 Flush 291158 2.500 163.468 Throughput 1105.45 MB/sec 8 clients 8 procs max_latency=163.474 ms +2.7% throughput, -40.1% max latency 16 jobs, before patchset: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 5457602 0.033 337.098 Close 4008979 0.002 2.018 Rename 231051 0.323 254.054 Unlink 1102209 0.202 337.243 Deltree 160 6.521 31.720 Mkdir 80 0.003 0.007 Qpathinfo 4946147 0.014 6.988 Qfileinfo 867440 0.001 1.642 Qfsinfo 907081 0.003 1.821 Sfileinfo 444433 0.005 2.053 Find 1912506 0.067 7.854 WriteX 2724852 0.018 7.428 ReadX 8553883 0.003 2.059 LockX 17770 0.003 0.350 UnlockX 17770 0.002 0.627 Flush 382533 2.810 353.691 Throughput 1413.09 MB/sec 16 clients 16 procs max_latency=353.696 ms 16 jobs, after patchset: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 5393156 0.034 303.181 Close 3961986 0.002 1.502 Rename 228359 0.320 253.379 Unlink 1088920 0.206 303.409 Deltree 160 6.419 30.088 Mkdir 80 0.003 0.004 Qpathinfo 4887967 0.015 7.722 Qfileinfo 857408 0.001 1.651 Qfsinfo 896343 0.002 2.147 Sfileinfo 439317 0.005 4.298 Find 1890018 0.073 8.347 WriteX 2693356 0.018 6.373 ReadX 8453485 0.003 3.836 LockX 17562 0.003 0.486 UnlockX 17562 0.002 0.635 Flush 378023 2.802 315.904 Throughput 1454.46 MB/sec 16 clients 16 procs max_latency=315.910 ms +2.9% throughput, -11.3% max latency Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-08-31 17:30:40 +03:00
goto out;
btrfs: implement delayed inode items operation Changelog V5 -> V6: - Fix oom when the memory load is high, by storing the delayed nodes into the root's radix tree, and letting btrfs inodes go. Changelog V4 -> V5: - Fix the race on adding the delayed node to the inode, which is spotted by Chris Mason. - Merge Chris Mason's incremental patch into this patch. - Fix deadlock between readdir() and memory fault, which is reported by Itaru Kitayama. Changelog V3 -> V4: - Fix nested lock, which is reported by Itaru Kitayama, by updating space cache inode in time. Changelog V2 -> V3: - Fix the race between the delayed worker and the task which does delayed items balance, which is reported by Tsutomu Itoh. - Modify the patch address David Sterba's comment. - Fix the bug of the cpu recursion spinlock, reported by Chris Mason Changelog V1 -> V2: - break up the global rb-tree, use a list to manage the delayed nodes, which is created for every directory and file, and used to manage the delayed directory name index items and the delayed inode item. - introduce a worker to deal with the delayed nodes. Compare with Ext3/4, the performance of file creation and deletion on btrfs is very poor. the reason is that btrfs must do a lot of b+ tree insertions, such as inode item, directory name item, directory name index and so on. If we can do some delayed b+ tree insertion or deletion, we can improve the performance, so we made this patch which implemented delayed directory name index insertion/deletion and delayed inode update. Implementation: - introduce a delayed root object into the filesystem, that use two lists to manage the delayed nodes which are created for every file/directory. One is used to manage all the delayed nodes that have delayed items. And the other is used to manage the delayed nodes which is waiting to be dealt with by the work thread. - Every delayed node has two rb-tree, one is used to manage the directory name index which is going to be inserted into b+ tree, and the other is used to manage the directory name index which is going to be deleted from b+ tree. - introduce a worker to deal with the delayed operation. This worker is used to deal with the works of the delayed directory name index items insertion and deletion and the delayed inode update. When the delayed items is beyond the lower limit, we create works for some delayed nodes and insert them into the work queue of the worker, and then go back. When the delayed items is beyond the upper bound, we create works for all the delayed nodes that haven't been dealt with, and insert them into the work queue of the worker, and then wait for that the untreated items is below some threshold value. - When we want to insert a directory name index into b+ tree, we just add the information into the delayed inserting rb-tree. And then we check the number of the delayed items and do delayed items balance. (The balance policy is above.) - When we want to delete a directory name index from the b+ tree, we search it in the inserting rb-tree at first. If we look it up, just drop it. If not, add the key of it into the delayed deleting rb-tree. Similar to the delayed inserting rb-tree, we also check the number of the delayed items and do delayed items balance. (The same to inserting manipulation) - When we want to update the metadata of some inode, we cached the data of the inode into the delayed node. the worker will flush it into the b+ tree after dealing with the delayed insertion and deletion. - We will move the delayed node to the tail of the list after we access the delayed node, By this way, we can cache more delayed items and merge more inode updates. - If we want to commit transaction, we will deal with all the delayed node. - the delayed node will be freed when we free the btrfs inode. - Before we log the inode items, we commit all the directory name index items and the delayed inode update. I did a quick test by the benchmark tool[1] and found we can improve the performance of file creation by ~15%, and file deletion by ~20%. Before applying this patch: Create files: Total files: 50000 Total time: 1.096108 Average time: 0.000022 Delete files: Total files: 50000 Total time: 1.510403 Average time: 0.000030 After applying this patch: Create files: Total files: 50000 Total time: 0.932899 Average time: 0.000019 Delete files: Total files: 50000 Total time: 1.215732 Average time: 0.000024 [1] http://marc.info/?l=linux-btrfs&m=128212635122920&q=p3 Many thanks for Kitayama-san's help! Signed-off-by: Miao Xie <miaox@cn.fujitsu.com> Reviewed-by: David Sterba <dave@jikos.cz> Tested-by: Tsutomu Itoh <t-itoh@jp.fujitsu.com> Tested-by: Itaru Kitayama <kitayama@cl.bb4u.ne.jp> Signed-off-by: Chris Mason <chris.mason@oracle.com>
2011-04-22 14:12:22 +04:00
}
Btrfs: fix fsync after succession of renames and unlink/rmdir After a succession of renames operations of different files and unlinking one of them, if we fsync one of the renamed files we can end up with a log that will either fail to replay at mount time or result in a filesystem that is in an inconsistent state. One example scenario: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/testdir $ touch /mnt/testdir/fname1 $ touch /mnt/testdir/fname2 $ sync $ mv /mnt/testdir/fname1 /mnt/testdir/fname3 $ rm -f /mnt/testdir/fname2 $ ln /mnt/testdir/fname3 /mnt/testdir/fname2 $ touch /mnt/testdir/fname1 $ xfs_io -c "fsync" /mnt/testdir/fname1 <power failure> $ mount /dev/sdb /mnt $ umount /mnt $ btrfs check /dev/sdb [1/7] checking root items [2/7] checking extents [3/7] checking free space cache [4/7] checking fs roots root 5 inode 259 errors 2, no orphan item ERROR: errors found in fs roots Opening filesystem to check... Checking filesystem on /dev/sdc UUID: 20e4abb8-5a19-4492-8bb4-6084125c2d0d found 393216 bytes used, error(s) found total csum bytes: 0 total tree bytes: 131072 total fs tree bytes: 32768 total extent tree bytes: 16384 btree space waste bytes: 122986 file data blocks allocated: 262144 referenced 262144 On a kernel without the first patch in this series, titled "[PATCH] Btrfs: fix fsync after succession of renames of different files", we get instead an error when mounting the filesystem due to failure of replaying the log: $ mount /dev/sdb /mnt mount: mount /dev/sdb on /mnt failed: File exists Fix this by logging the parent directory of an inode whenever we find an inode that no longer exists (was unlinked in the current transaction), during the procedure which finds inodes that have old names that collide with new names of other inodes. A test case for fstests follows soon. Reviewed-by: Nikolay Borisov <nborisov@suse.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-02-13 15:14:09 +03:00
if (inode_only == LOG_OTHER_INODE || inode_only == LOG_OTHER_INODE_ALL) {
recursive_logging = true;
if (inode_only == LOG_OTHER_INODE)
inode_only = LOG_INODE_EXISTS;
else
inode_only = LOG_INODE_ALL;
mutex_lock_nested(&inode->log_mutex, SINGLE_DEPTH_NESTING);
} else {
mutex_lock(&inode->log_mutex);
}
btrfs: avoid inode logging during rename and link when possible During a rename or link operation, we need to determine if an inode was previously logged or not, and if it was, do some update to the logged inode. We used to rely exclusively on the logged_trans field of struct btrfs_inode to determine that, but that was not reliable because the value of that field is not persisted in the inode item, so it's lost when an inode is evicted and loaded back again. That led to several issues in the past, such as not persisting deletions (such as the case fixed by commit 803f0f64d17769 ("Btrfs: fix fsync not persisting dentry deletions due to inode evictions")), or resulting in losing a file after an inode eviction followed by a rename (commit ecc64fab7d49c6 ("btrfs: fix lost inode on log replay after mix of fsync, rename and inode eviction")), besides other issues. So the inode_logged() helper was introduced and used to determine if an inode was possibly logged before in the current transaction, with the caveat that it could return false positives, in the sense that even if an inode was not logged before in the current transaction, it could still return true, but never to return false in case the inode was logged. >From a functional point of view that is fine, but from a performance perspective it can introduce significant latencies to rename and link operations, as they will end up doing inode logging even when it is not necessary. Recently on a 5.15 kernel, an openSUSE Tumbleweed user reported package installations and upgrades, with the zypper tool, were often taking a long time to complete. With strace it could be observed that zypper was spending about 99% of its time on rename operations, and then with further analysis we checked that directory logging was happening too frequently. Taking into account that installation/upgrade of some of the packages needed a few thousand file renames, the slowdown was very noticeable for the user. The issue was caused indirectly due to an excessive number of inode evictions on a 5.15 kernel, about 100x more compared to a 5.13, 5.14 or a 5.16-rc8 kernel. While triggering the inode evictions if something outside btrfs' control, btrfs could still behave better by eliminating the false positives from the inode_logged() helper. So change inode_logged() to actually eliminate such false positives caused by inode eviction and when an inode was never logged since the filesystem was mounted, as both cases relate to when the logged_trans field of struct btrfs_inode has a value of zero. When it can not determine if the inode was logged based only on the logged_trans value, lookup for the existence of the inode item in the log tree - if it's there then we known the inode was logged, if it's not there then it can not have been logged in the current transaction. Once we determine if the inode was logged, update the logged_trans value to avoid future calls to have to search in the log tree again. Alternatively, we could start storing logged_trans in the on disk inode item structure (struct btrfs_inode_item) in the unused space it still has, but that would be a bit odd because: 1) We only care about logged_trans since the filesystem was mounted, we don't care about its value from a previous mount. Having it persisted in the inode item structure would not make the best use of the precious unused space; 2) In order to get logged_trans persisted before inode eviction, we would have to update the delayed inode when we finish logging the inode and update its logged_trans in struct btrfs_inode, which makes it a bit cumbersome since we need to check if the delayed inode exists, if not create it and populate it and deal with any errors (-ENOMEM mostly). This change is part of a patchset comprised of the following patches: 1/5 btrfs: add helper to delete a dir entry from a log tree 2/5 btrfs: pass the dentry to btrfs_log_new_name() instead of the inode 3/5 btrfs: avoid logging all directory changes during renames 4/5 btrfs: stop doing unnecessary log updates during a rename 5/5 btrfs: avoid inode logging during rename and link when possible The following test script mimics part of what the zypper tool does during package installations/upgrades. It does not triggers inode evictions, but it's similar because it triggers false positives from the inode_logged() helper, because the inodes have a logged_trans of 0, there's a log tree due to a fsync of an unrelated file and the directory inode has its last_trans field set to the current transaction: $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=10000 mkfs.btrfs -f $DEV mount $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync # Now do some change to an unrelated file and fsync it. # This is just to create a log tree to make sure that inode_logged() # does not return false when called against "testdir". xfs_io -f -c "pwrite 0 4K" -c "fsync" $MNT/foo # Do some change to testdir. This is to make sure inode_logged() # will return true when called against "testdir", because its # logged_trans is 0, it was changed in the current transaction # and there's a log tree. echo -n > $MNT/testdir/file_$((NUM_FILES + 1)) echo "Renaming $NUM_FILES files..." start=$(date +%s%N) for ((i = 1; i <= $NUM_FILES; i++)); do mv $MNT/testdir/file_$i $MNT/testdir/file_$i-RPMDELETE done end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "Renames took $dur milliseconds" umount $MNT Testing this change on a box using a non-debug kernel (Debian's default kernel config) gave the following results: NUM_FILES=10000, before patchset: 27837 ms NUM_FILES=10000, after patches 1/5 to 4/5 applied: 9236 ms (-66.8%) NUM_FILES=10000, after whole patchset applied: 8902 ms (-68.0%) NUM_FILES=5000, before patchset: 9127 ms NUM_FILES=5000, after patches 1/5 to 4/5 applied: 4640 ms (-49.2%) NUM_FILES=5000, after whole patchset applied: 4441 ms (-51.3%) NUM_FILES=2000, before patchset: 2528 ms NUM_FILES=2000, after patches 1/5 to 4/5 applied: 1983 ms (-21.6%) NUM_FILES=2000, after whole patchset applied: 1747 ms (-30.9%) NUM_FILES=1000, before patchset: 1085 ms NUM_FILES=1000, after patches 1/5 to 4/5 applied: 893 ms (-17.7%) NUM_FILES=1000, after whole patchset applied: 867 ms (-20.1%) Running dbench on the same physical machine with the following script: $ cat run-dbench.sh #!/bin/bash NUM_JOBS=$(nproc --all) DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 MOUNT_OPTIONS="-o ssd" MKFS_OPTIONS="-O no-holes -R free-space-tree" echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -D $MNT -t 120 $NUM_JOBS umount $MNT Before patchset: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 3761352 0.032 143.843 Close 2762770 0.002 2.273 Rename 159304 0.291 67.037 Unlink 759784 0.207 143.998 Deltree 72 4.028 15.977 Mkdir 36 0.003 0.006 Qpathinfo 3409780 0.013 9.678 Qfileinfo 596772 0.001 0.878 Qfsinfo 625189 0.003 1.245 Sfileinfo 306443 0.006 1.840 Find 1318106 0.063 19.798 WriteX 1871137 0.021 8.532 ReadX 5897325 0.003 3.567 LockX 12252 0.003 0.258 UnlockX 12252 0.002 0.100 Flush 263666 3.327 155.632 Throughput 980.047 MB/sec 12 clients 12 procs max_latency=155.636 ms After whole patchset applied: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 4195584 0.033 107.742 Close 3081932 0.002 1.935 Rename 177641 0.218 14.905 Unlink 847333 0.166 107.822 Deltree 118 5.315 15.247 Mkdir 59 0.004 0.048 Qpathinfo 3802612 0.014 10.302 Qfileinfo 666748 0.001 1.034 Qfsinfo 697329 0.003 0.944 Sfileinfo 341712 0.006 2.099 Find 1470365 0.065 9.359 WriteX 2093921 0.021 8.087 ReadX 6576234 0.003 3.407 LockX 13660 0.003 0.308 UnlockX 13660 0.002 0.114 Flush 294090 2.906 115.539 Throughput 1093.11 MB/sec 12 clients 12 procs max_latency=115.544 ms +11.5% throughput -25.8% max latency rename max latency -77.8% Link: https://bugzilla.opensuse.org/show_bug.cgi?id=1193549 Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-01-20 14:00:10 +03:00
/*
* Before logging the inode item, cache the value returned by
* inode_logged(), because after that we have the need to figure out if
* the inode was previously logged in this transaction.
*/
ret = inode_logged(trans, inode, path);
if (ret < 0)
btrfs: avoid inode logging during rename and link when possible During a rename or link operation, we need to determine if an inode was previously logged or not, and if it was, do some update to the logged inode. We used to rely exclusively on the logged_trans field of struct btrfs_inode to determine that, but that was not reliable because the value of that field is not persisted in the inode item, so it's lost when an inode is evicted and loaded back again. That led to several issues in the past, such as not persisting deletions (such as the case fixed by commit 803f0f64d17769 ("Btrfs: fix fsync not persisting dentry deletions due to inode evictions")), or resulting in losing a file after an inode eviction followed by a rename (commit ecc64fab7d49c6 ("btrfs: fix lost inode on log replay after mix of fsync, rename and inode eviction")), besides other issues. So the inode_logged() helper was introduced and used to determine if an inode was possibly logged before in the current transaction, with the caveat that it could return false positives, in the sense that even if an inode was not logged before in the current transaction, it could still return true, but never to return false in case the inode was logged. >From a functional point of view that is fine, but from a performance perspective it can introduce significant latencies to rename and link operations, as they will end up doing inode logging even when it is not necessary. Recently on a 5.15 kernel, an openSUSE Tumbleweed user reported package installations and upgrades, with the zypper tool, were often taking a long time to complete. With strace it could be observed that zypper was spending about 99% of its time on rename operations, and then with further analysis we checked that directory logging was happening too frequently. Taking into account that installation/upgrade of some of the packages needed a few thousand file renames, the slowdown was very noticeable for the user. The issue was caused indirectly due to an excessive number of inode evictions on a 5.15 kernel, about 100x more compared to a 5.13, 5.14 or a 5.16-rc8 kernel. While triggering the inode evictions if something outside btrfs' control, btrfs could still behave better by eliminating the false positives from the inode_logged() helper. So change inode_logged() to actually eliminate such false positives caused by inode eviction and when an inode was never logged since the filesystem was mounted, as both cases relate to when the logged_trans field of struct btrfs_inode has a value of zero. When it can not determine if the inode was logged based only on the logged_trans value, lookup for the existence of the inode item in the log tree - if it's there then we known the inode was logged, if it's not there then it can not have been logged in the current transaction. Once we determine if the inode was logged, update the logged_trans value to avoid future calls to have to search in the log tree again. Alternatively, we could start storing logged_trans in the on disk inode item structure (struct btrfs_inode_item) in the unused space it still has, but that would be a bit odd because: 1) We only care about logged_trans since the filesystem was mounted, we don't care about its value from a previous mount. Having it persisted in the inode item structure would not make the best use of the precious unused space; 2) In order to get logged_trans persisted before inode eviction, we would have to update the delayed inode when we finish logging the inode and update its logged_trans in struct btrfs_inode, which makes it a bit cumbersome since we need to check if the delayed inode exists, if not create it and populate it and deal with any errors (-ENOMEM mostly). This change is part of a patchset comprised of the following patches: 1/5 btrfs: add helper to delete a dir entry from a log tree 2/5 btrfs: pass the dentry to btrfs_log_new_name() instead of the inode 3/5 btrfs: avoid logging all directory changes during renames 4/5 btrfs: stop doing unnecessary log updates during a rename 5/5 btrfs: avoid inode logging during rename and link when possible The following test script mimics part of what the zypper tool does during package installations/upgrades. It does not triggers inode evictions, but it's similar because it triggers false positives from the inode_logged() helper, because the inodes have a logged_trans of 0, there's a log tree due to a fsync of an unrelated file and the directory inode has its last_trans field set to the current transaction: $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=10000 mkfs.btrfs -f $DEV mount $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync # Now do some change to an unrelated file and fsync it. # This is just to create a log tree to make sure that inode_logged() # does not return false when called against "testdir". xfs_io -f -c "pwrite 0 4K" -c "fsync" $MNT/foo # Do some change to testdir. This is to make sure inode_logged() # will return true when called against "testdir", because its # logged_trans is 0, it was changed in the current transaction # and there's a log tree. echo -n > $MNT/testdir/file_$((NUM_FILES + 1)) echo "Renaming $NUM_FILES files..." start=$(date +%s%N) for ((i = 1; i <= $NUM_FILES; i++)); do mv $MNT/testdir/file_$i $MNT/testdir/file_$i-RPMDELETE done end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "Renames took $dur milliseconds" umount $MNT Testing this change on a box using a non-debug kernel (Debian's default kernel config) gave the following results: NUM_FILES=10000, before patchset: 27837 ms NUM_FILES=10000, after patches 1/5 to 4/5 applied: 9236 ms (-66.8%) NUM_FILES=10000, after whole patchset applied: 8902 ms (-68.0%) NUM_FILES=5000, before patchset: 9127 ms NUM_FILES=5000, after patches 1/5 to 4/5 applied: 4640 ms (-49.2%) NUM_FILES=5000, after whole patchset applied: 4441 ms (-51.3%) NUM_FILES=2000, before patchset: 2528 ms NUM_FILES=2000, after patches 1/5 to 4/5 applied: 1983 ms (-21.6%) NUM_FILES=2000, after whole patchset applied: 1747 ms (-30.9%) NUM_FILES=1000, before patchset: 1085 ms NUM_FILES=1000, after patches 1/5 to 4/5 applied: 893 ms (-17.7%) NUM_FILES=1000, after whole patchset applied: 867 ms (-20.1%) Running dbench on the same physical machine with the following script: $ cat run-dbench.sh #!/bin/bash NUM_JOBS=$(nproc --all) DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 MOUNT_OPTIONS="-o ssd" MKFS_OPTIONS="-O no-holes -R free-space-tree" echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -D $MNT -t 120 $NUM_JOBS umount $MNT Before patchset: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 3761352 0.032 143.843 Close 2762770 0.002 2.273 Rename 159304 0.291 67.037 Unlink 759784 0.207 143.998 Deltree 72 4.028 15.977 Mkdir 36 0.003 0.006 Qpathinfo 3409780 0.013 9.678 Qfileinfo 596772 0.001 0.878 Qfsinfo 625189 0.003 1.245 Sfileinfo 306443 0.006 1.840 Find 1318106 0.063 19.798 WriteX 1871137 0.021 8.532 ReadX 5897325 0.003 3.567 LockX 12252 0.003 0.258 UnlockX 12252 0.002 0.100 Flush 263666 3.327 155.632 Throughput 980.047 MB/sec 12 clients 12 procs max_latency=155.636 ms After whole patchset applied: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 4195584 0.033 107.742 Close 3081932 0.002 1.935 Rename 177641 0.218 14.905 Unlink 847333 0.166 107.822 Deltree 118 5.315 15.247 Mkdir 59 0.004 0.048 Qpathinfo 3802612 0.014 10.302 Qfileinfo 666748 0.001 1.034 Qfsinfo 697329 0.003 0.944 Sfileinfo 341712 0.006 2.099 Find 1470365 0.065 9.359 WriteX 2093921 0.021 8.087 ReadX 6576234 0.003 3.407 LockX 13660 0.003 0.308 UnlockX 13660 0.002 0.114 Flush 294090 2.906 115.539 Throughput 1093.11 MB/sec 12 clients 12 procs max_latency=115.544 ms +11.5% throughput -25.8% max latency rename max latency -77.8% Link: https://bugzilla.opensuse.org/show_bug.cgi?id=1193549 Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-01-20 14:00:10 +03:00
goto out_unlock;
ctx->logged_before = (ret == 1);
ret = 0;
btrfs: avoid inode logging during rename and link when possible During a rename or link operation, we need to determine if an inode was previously logged or not, and if it was, do some update to the logged inode. We used to rely exclusively on the logged_trans field of struct btrfs_inode to determine that, but that was not reliable because the value of that field is not persisted in the inode item, so it's lost when an inode is evicted and loaded back again. That led to several issues in the past, such as not persisting deletions (such as the case fixed by commit 803f0f64d17769 ("Btrfs: fix fsync not persisting dentry deletions due to inode evictions")), or resulting in losing a file after an inode eviction followed by a rename (commit ecc64fab7d49c6 ("btrfs: fix lost inode on log replay after mix of fsync, rename and inode eviction")), besides other issues. So the inode_logged() helper was introduced and used to determine if an inode was possibly logged before in the current transaction, with the caveat that it could return false positives, in the sense that even if an inode was not logged before in the current transaction, it could still return true, but never to return false in case the inode was logged. >From a functional point of view that is fine, but from a performance perspective it can introduce significant latencies to rename and link operations, as they will end up doing inode logging even when it is not necessary. Recently on a 5.15 kernel, an openSUSE Tumbleweed user reported package installations and upgrades, with the zypper tool, were often taking a long time to complete. With strace it could be observed that zypper was spending about 99% of its time on rename operations, and then with further analysis we checked that directory logging was happening too frequently. Taking into account that installation/upgrade of some of the packages needed a few thousand file renames, the slowdown was very noticeable for the user. The issue was caused indirectly due to an excessive number of inode evictions on a 5.15 kernel, about 100x more compared to a 5.13, 5.14 or a 5.16-rc8 kernel. While triggering the inode evictions if something outside btrfs' control, btrfs could still behave better by eliminating the false positives from the inode_logged() helper. So change inode_logged() to actually eliminate such false positives caused by inode eviction and when an inode was never logged since the filesystem was mounted, as both cases relate to when the logged_trans field of struct btrfs_inode has a value of zero. When it can not determine if the inode was logged based only on the logged_trans value, lookup for the existence of the inode item in the log tree - if it's there then we known the inode was logged, if it's not there then it can not have been logged in the current transaction. Once we determine if the inode was logged, update the logged_trans value to avoid future calls to have to search in the log tree again. Alternatively, we could start storing logged_trans in the on disk inode item structure (struct btrfs_inode_item) in the unused space it still has, but that would be a bit odd because: 1) We only care about logged_trans since the filesystem was mounted, we don't care about its value from a previous mount. Having it persisted in the inode item structure would not make the best use of the precious unused space; 2) In order to get logged_trans persisted before inode eviction, we would have to update the delayed inode when we finish logging the inode and update its logged_trans in struct btrfs_inode, which makes it a bit cumbersome since we need to check if the delayed inode exists, if not create it and populate it and deal with any errors (-ENOMEM mostly). This change is part of a patchset comprised of the following patches: 1/5 btrfs: add helper to delete a dir entry from a log tree 2/5 btrfs: pass the dentry to btrfs_log_new_name() instead of the inode 3/5 btrfs: avoid logging all directory changes during renames 4/5 btrfs: stop doing unnecessary log updates during a rename 5/5 btrfs: avoid inode logging during rename and link when possible The following test script mimics part of what the zypper tool does during package installations/upgrades. It does not triggers inode evictions, but it's similar because it triggers false positives from the inode_logged() helper, because the inodes have a logged_trans of 0, there's a log tree due to a fsync of an unrelated file and the directory inode has its last_trans field set to the current transaction: $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=10000 mkfs.btrfs -f $DEV mount $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync # Now do some change to an unrelated file and fsync it. # This is just to create a log tree to make sure that inode_logged() # does not return false when called against "testdir". xfs_io -f -c "pwrite 0 4K" -c "fsync" $MNT/foo # Do some change to testdir. This is to make sure inode_logged() # will return true when called against "testdir", because its # logged_trans is 0, it was changed in the current transaction # and there's a log tree. echo -n > $MNT/testdir/file_$((NUM_FILES + 1)) echo "Renaming $NUM_FILES files..." start=$(date +%s%N) for ((i = 1; i <= $NUM_FILES; i++)); do mv $MNT/testdir/file_$i $MNT/testdir/file_$i-RPMDELETE done end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "Renames took $dur milliseconds" umount $MNT Testing this change on a box using a non-debug kernel (Debian's default kernel config) gave the following results: NUM_FILES=10000, before patchset: 27837 ms NUM_FILES=10000, after patches 1/5 to 4/5 applied: 9236 ms (-66.8%) NUM_FILES=10000, after whole patchset applied: 8902 ms (-68.0%) NUM_FILES=5000, before patchset: 9127 ms NUM_FILES=5000, after patches 1/5 to 4/5 applied: 4640 ms (-49.2%) NUM_FILES=5000, after whole patchset applied: 4441 ms (-51.3%) NUM_FILES=2000, before patchset: 2528 ms NUM_FILES=2000, after patches 1/5 to 4/5 applied: 1983 ms (-21.6%) NUM_FILES=2000, after whole patchset applied: 1747 ms (-30.9%) NUM_FILES=1000, before patchset: 1085 ms NUM_FILES=1000, after patches 1/5 to 4/5 applied: 893 ms (-17.7%) NUM_FILES=1000, after whole patchset applied: 867 ms (-20.1%) Running dbench on the same physical machine with the following script: $ cat run-dbench.sh #!/bin/bash NUM_JOBS=$(nproc --all) DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 MOUNT_OPTIONS="-o ssd" MKFS_OPTIONS="-O no-holes -R free-space-tree" echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -D $MNT -t 120 $NUM_JOBS umount $MNT Before patchset: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 3761352 0.032 143.843 Close 2762770 0.002 2.273 Rename 159304 0.291 67.037 Unlink 759784 0.207 143.998 Deltree 72 4.028 15.977 Mkdir 36 0.003 0.006 Qpathinfo 3409780 0.013 9.678 Qfileinfo 596772 0.001 0.878 Qfsinfo 625189 0.003 1.245 Sfileinfo 306443 0.006 1.840 Find 1318106 0.063 19.798 WriteX 1871137 0.021 8.532 ReadX 5897325 0.003 3.567 LockX 12252 0.003 0.258 UnlockX 12252 0.002 0.100 Flush 263666 3.327 155.632 Throughput 980.047 MB/sec 12 clients 12 procs max_latency=155.636 ms After whole patchset applied: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 4195584 0.033 107.742 Close 3081932 0.002 1.935 Rename 177641 0.218 14.905 Unlink 847333 0.166 107.822 Deltree 118 5.315 15.247 Mkdir 59 0.004 0.048 Qpathinfo 3802612 0.014 10.302 Qfileinfo 666748 0.001 1.034 Qfsinfo 697329 0.003 0.944 Sfileinfo 341712 0.006 2.099 Find 1470365 0.065 9.359 WriteX 2093921 0.021 8.087 ReadX 6576234 0.003 3.407 LockX 13660 0.003 0.308 UnlockX 13660 0.002 0.114 Flush 294090 2.906 115.539 Throughput 1093.11 MB/sec 12 clients 12 procs max_latency=115.544 ms +11.5% throughput -25.8% max latency rename max latency -77.8% Link: https://bugzilla.opensuse.org/show_bug.cgi?id=1193549 Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-01-20 14:00:10 +03:00
btrfs: remove unnecessary check_parent_dirs_for_sync() Whenever we fsync an inode, if it is a directory, a regular file that was created in the current transaction or has last_unlink_trans set to the generation of the current transaction, we check if any of its ancestor inodes (and the inode itself if it is a directory) can not be logged and need a fallback to a full transaction commit - if so, we return with a value of 1 in order to fallback to a transaction commit. However we often do not need to fallback to a transaction commit because: 1) The ancestor inode is not an immediate parent, and therefore there is not an explicit request to log it and it is not needed neither to guarantee the consistency of the inode originally asked to be logged (fsynced) nor its immediate parent; 2) The ancestor inode was already logged before, in which case any link, unlink or rename operation updates the log as needed. So for these two cases we can avoid an unnecessary transaction commit. Therefore remove check_parent_dirs_for_sync() and add a check at the top of btrfs_log_inode() to make us fallback immediately to a transaction commit when we are logging a directory inode that can not be logged and needs a full transaction commit. All we need to protect is the case where after renaming a file someone fsyncs only the old directory, which would result is losing the renamed file after a log replay. This patch is part of a patchset comprised of the following patches: btrfs: remove unnecessary directory inode item update when deleting dir entry btrfs: stop setting nbytes when filling inode item for logging btrfs: avoid logging new ancestor inodes when logging new inode btrfs: skip logging directories already logged when logging all parents btrfs: skip logging inodes already logged when logging new entries btrfs: remove unnecessary check_parent_dirs_for_sync() btrfs: make concurrent fsyncs wait less when waiting for a transaction commit Performance results, after applying all patches, are mentioned in the change log of the last patch. Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-01-27 13:34:59 +03:00
/*
* This is for cases where logging a directory could result in losing a
* a file after replaying the log. For example, if we move a file from a
* directory A to a directory B, then fsync directory A, we have no way
* to known the file was moved from A to B, so logging just A would
* result in losing the file after a log replay.
*/
if (S_ISDIR(inode->vfs_inode.i_mode) &&
inode_only == LOG_INODE_ALL &&
inode->last_unlink_trans >= trans->transid) {
btrfs_set_log_full_commit(trans);
ret = 1;
btrfs: remove unnecessary check_parent_dirs_for_sync() Whenever we fsync an inode, if it is a directory, a regular file that was created in the current transaction or has last_unlink_trans set to the generation of the current transaction, we check if any of its ancestor inodes (and the inode itself if it is a directory) can not be logged and need a fallback to a full transaction commit - if so, we return with a value of 1 in order to fallback to a transaction commit. However we often do not need to fallback to a transaction commit because: 1) The ancestor inode is not an immediate parent, and therefore there is not an explicit request to log it and it is not needed neither to guarantee the consistency of the inode originally asked to be logged (fsynced) nor its immediate parent; 2) The ancestor inode was already logged before, in which case any link, unlink or rename operation updates the log as needed. So for these two cases we can avoid an unnecessary transaction commit. Therefore remove check_parent_dirs_for_sync() and add a check at the top of btrfs_log_inode() to make us fallback immediately to a transaction commit when we are logging a directory inode that can not be logged and needs a full transaction commit. All we need to protect is the case where after renaming a file someone fsyncs only the old directory, which would result is losing the renamed file after a log replay. This patch is part of a patchset comprised of the following patches: btrfs: remove unnecessary directory inode item update when deleting dir entry btrfs: stop setting nbytes when filling inode item for logging btrfs: avoid logging new ancestor inodes when logging new inode btrfs: skip logging directories already logged when logging all parents btrfs: skip logging inodes already logged when logging new entries btrfs: remove unnecessary check_parent_dirs_for_sync() btrfs: make concurrent fsyncs wait less when waiting for a transaction commit Performance results, after applying all patches, are mentioned in the change log of the last patch. Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-01-27 13:34:59 +03:00
goto out_unlock;
}
/*
* a brute force approach to making sure we get the most uptodate
* copies of everything.
*/
if (S_ISDIR(inode->vfs_inode.i_mode)) {
int max_key_type = BTRFS_DIR_LOG_INDEX_KEY;
clear_bit(BTRFS_INODE_COPY_EVERYTHING, &inode->runtime_flags);
Btrfs: remove deleted xattrs on fsync log replay If we deleted xattrs from a file and fsynced the file, after a log replay the xattrs would remain associated to the file. This was an unexpected behaviour and differs from what other filesystems do, such as for example xfs and ext3/4. Fix this by, on fsync log replay, check if every xattr in the fs/subvol tree (that belongs to a logged inode) has a matching xattr in the log, and if it does not, delete it from the fs/subvol tree. This is a similar approach to what we do for dentries when we replay a directory from the fsync log. This issue is trivial to reproduce, and the following excerpt from my test for xfstests triggers the issue: _crash_and_mount() { # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey } rm -f $seqres.full _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create out test file and add 3 xattrs to it. touch $SCRATCH_MNT/foobar $SETFATTR_PROG -n user.attr1 -v val1 $SCRATCH_MNT/foobar $SETFATTR_PROG -n user.attr2 -v val2 $SCRATCH_MNT/foobar $SETFATTR_PROG -n user.attr3 -v val3 $SCRATCH_MNT/foobar # Make sure everything is durably persisted. sync # Now delete the second xattr and fsync the inode. $SETFATTR_PROG -x user.attr2 $SCRATCH_MNT/foobar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foobar _crash_and_mount # After the fsync log is replayed, the file should have only 2 xattrs, the ones # named user.attr1 and user.attr3. The btrfs fsync log replay bug left the file # with the 3 xattrs that we had before deleting the second one and fsyncing the # file. echo "xattr names and values after first fsync log replay:" $GETFATTR_PROG --absolute-names --dump $SCRATCH_MNT/foobar | _filter_scratch # Now write some data to our file, fsync it, remove the first xattr, add a new # hard link to our file and commit the fsync log by fsyncing some other new # file. This is to verify that after log replay our first xattr does not exist # anymore. echo "hello world!" >> $SCRATCH_MNT/foobar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foobar $SETFATTR_PROG -x user.attr1 $SCRATCH_MNT/foobar ln $SCRATCH_MNT/foobar $SCRATCH_MNT/foobar_link touch $SCRATCH_MNT/qwerty $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/qwerty _crash_and_mount # Now only the xattr with name user.attr3 should be set in our file. echo "xattr names and values after second fsync log replay:" $GETFATTR_PROG --absolute-names --dump $SCRATCH_MNT/foobar | _filter_scratch status=0 exit The expected golden output, which is produced with this patch applied or when testing against xfs or ext3/4, is: xattr names and values after first fsync log replay: # file: SCRATCH_MNT/foobar user.attr1="val1" user.attr3="val3" xattr names and values after second fsync log replay: # file: SCRATCH_MNT/foobar user.attr3="val3" Without this patch applied, the output is: xattr names and values after first fsync log replay: # file: SCRATCH_MNT/foobar user.attr1="val1" user.attr2="val2" user.attr3="val3" xattr names and values after second fsync log replay: # file: SCRATCH_MNT/foobar user.attr1="val1" user.attr2="val2" user.attr3="val3" A patch with a test case for xfstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-02-23 22:53:35 +03:00
if (inode_only == LOG_INODE_EXISTS)
max_key_type = BTRFS_XATTR_ITEM_KEY;
btrfs: avoid inode logging during rename and link when possible During a rename or link operation, we need to determine if an inode was previously logged or not, and if it was, do some update to the logged inode. We used to rely exclusively on the logged_trans field of struct btrfs_inode to determine that, but that was not reliable because the value of that field is not persisted in the inode item, so it's lost when an inode is evicted and loaded back again. That led to several issues in the past, such as not persisting deletions (such as the case fixed by commit 803f0f64d17769 ("Btrfs: fix fsync not persisting dentry deletions due to inode evictions")), or resulting in losing a file after an inode eviction followed by a rename (commit ecc64fab7d49c6 ("btrfs: fix lost inode on log replay after mix of fsync, rename and inode eviction")), besides other issues. So the inode_logged() helper was introduced and used to determine if an inode was possibly logged before in the current transaction, with the caveat that it could return false positives, in the sense that even if an inode was not logged before in the current transaction, it could still return true, but never to return false in case the inode was logged. >From a functional point of view that is fine, but from a performance perspective it can introduce significant latencies to rename and link operations, as they will end up doing inode logging even when it is not necessary. Recently on a 5.15 kernel, an openSUSE Tumbleweed user reported package installations and upgrades, with the zypper tool, were often taking a long time to complete. With strace it could be observed that zypper was spending about 99% of its time on rename operations, and then with further analysis we checked that directory logging was happening too frequently. Taking into account that installation/upgrade of some of the packages needed a few thousand file renames, the slowdown was very noticeable for the user. The issue was caused indirectly due to an excessive number of inode evictions on a 5.15 kernel, about 100x more compared to a 5.13, 5.14 or a 5.16-rc8 kernel. While triggering the inode evictions if something outside btrfs' control, btrfs could still behave better by eliminating the false positives from the inode_logged() helper. So change inode_logged() to actually eliminate such false positives caused by inode eviction and when an inode was never logged since the filesystem was mounted, as both cases relate to when the logged_trans field of struct btrfs_inode has a value of zero. When it can not determine if the inode was logged based only on the logged_trans value, lookup for the existence of the inode item in the log tree - if it's there then we known the inode was logged, if it's not there then it can not have been logged in the current transaction. Once we determine if the inode was logged, update the logged_trans value to avoid future calls to have to search in the log tree again. Alternatively, we could start storing logged_trans in the on disk inode item structure (struct btrfs_inode_item) in the unused space it still has, but that would be a bit odd because: 1) We only care about logged_trans since the filesystem was mounted, we don't care about its value from a previous mount. Having it persisted in the inode item structure would not make the best use of the precious unused space; 2) In order to get logged_trans persisted before inode eviction, we would have to update the delayed inode when we finish logging the inode and update its logged_trans in struct btrfs_inode, which makes it a bit cumbersome since we need to check if the delayed inode exists, if not create it and populate it and deal with any errors (-ENOMEM mostly). This change is part of a patchset comprised of the following patches: 1/5 btrfs: add helper to delete a dir entry from a log tree 2/5 btrfs: pass the dentry to btrfs_log_new_name() instead of the inode 3/5 btrfs: avoid logging all directory changes during renames 4/5 btrfs: stop doing unnecessary log updates during a rename 5/5 btrfs: avoid inode logging during rename and link when possible The following test script mimics part of what the zypper tool does during package installations/upgrades. It does not triggers inode evictions, but it's similar because it triggers false positives from the inode_logged() helper, because the inodes have a logged_trans of 0, there's a log tree due to a fsync of an unrelated file and the directory inode has its last_trans field set to the current transaction: $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=10000 mkfs.btrfs -f $DEV mount $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync # Now do some change to an unrelated file and fsync it. # This is just to create a log tree to make sure that inode_logged() # does not return false when called against "testdir". xfs_io -f -c "pwrite 0 4K" -c "fsync" $MNT/foo # Do some change to testdir. This is to make sure inode_logged() # will return true when called against "testdir", because its # logged_trans is 0, it was changed in the current transaction # and there's a log tree. echo -n > $MNT/testdir/file_$((NUM_FILES + 1)) echo "Renaming $NUM_FILES files..." start=$(date +%s%N) for ((i = 1; i <= $NUM_FILES; i++)); do mv $MNT/testdir/file_$i $MNT/testdir/file_$i-RPMDELETE done end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "Renames took $dur milliseconds" umount $MNT Testing this change on a box using a non-debug kernel (Debian's default kernel config) gave the following results: NUM_FILES=10000, before patchset: 27837 ms NUM_FILES=10000, after patches 1/5 to 4/5 applied: 9236 ms (-66.8%) NUM_FILES=10000, after whole patchset applied: 8902 ms (-68.0%) NUM_FILES=5000, before patchset: 9127 ms NUM_FILES=5000, after patches 1/5 to 4/5 applied: 4640 ms (-49.2%) NUM_FILES=5000, after whole patchset applied: 4441 ms (-51.3%) NUM_FILES=2000, before patchset: 2528 ms NUM_FILES=2000, after patches 1/5 to 4/5 applied: 1983 ms (-21.6%) NUM_FILES=2000, after whole patchset applied: 1747 ms (-30.9%) NUM_FILES=1000, before patchset: 1085 ms NUM_FILES=1000, after patches 1/5 to 4/5 applied: 893 ms (-17.7%) NUM_FILES=1000, after whole patchset applied: 867 ms (-20.1%) Running dbench on the same physical machine with the following script: $ cat run-dbench.sh #!/bin/bash NUM_JOBS=$(nproc --all) DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 MOUNT_OPTIONS="-o ssd" MKFS_OPTIONS="-O no-holes -R free-space-tree" echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -D $MNT -t 120 $NUM_JOBS umount $MNT Before patchset: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 3761352 0.032 143.843 Close 2762770 0.002 2.273 Rename 159304 0.291 67.037 Unlink 759784 0.207 143.998 Deltree 72 4.028 15.977 Mkdir 36 0.003 0.006 Qpathinfo 3409780 0.013 9.678 Qfileinfo 596772 0.001 0.878 Qfsinfo 625189 0.003 1.245 Sfileinfo 306443 0.006 1.840 Find 1318106 0.063 19.798 WriteX 1871137 0.021 8.532 ReadX 5897325 0.003 3.567 LockX 12252 0.003 0.258 UnlockX 12252 0.002 0.100 Flush 263666 3.327 155.632 Throughput 980.047 MB/sec 12 clients 12 procs max_latency=155.636 ms After whole patchset applied: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 4195584 0.033 107.742 Close 3081932 0.002 1.935 Rename 177641 0.218 14.905 Unlink 847333 0.166 107.822 Deltree 118 5.315 15.247 Mkdir 59 0.004 0.048 Qpathinfo 3802612 0.014 10.302 Qfileinfo 666748 0.001 1.034 Qfsinfo 697329 0.003 0.944 Sfileinfo 341712 0.006 2.099 Find 1470365 0.065 9.359 WriteX 2093921 0.021 8.087 ReadX 6576234 0.003 3.407 LockX 13660 0.003 0.308 UnlockX 13660 0.002 0.114 Flush 294090 2.906 115.539 Throughput 1093.11 MB/sec 12 clients 12 procs max_latency=115.544 ms +11.5% throughput -25.8% max latency rename max latency -77.8% Link: https://bugzilla.opensuse.org/show_bug.cgi?id=1193549 Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-01-20 14:00:10 +03:00
if (ctx->logged_before)
ret = drop_inode_items(trans, log, path, inode,
max_key_type);
} else {
btrfs: avoid inode logging during rename and link when possible During a rename or link operation, we need to determine if an inode was previously logged or not, and if it was, do some update to the logged inode. We used to rely exclusively on the logged_trans field of struct btrfs_inode to determine that, but that was not reliable because the value of that field is not persisted in the inode item, so it's lost when an inode is evicted and loaded back again. That led to several issues in the past, such as not persisting deletions (such as the case fixed by commit 803f0f64d17769 ("Btrfs: fix fsync not persisting dentry deletions due to inode evictions")), or resulting in losing a file after an inode eviction followed by a rename (commit ecc64fab7d49c6 ("btrfs: fix lost inode on log replay after mix of fsync, rename and inode eviction")), besides other issues. So the inode_logged() helper was introduced and used to determine if an inode was possibly logged before in the current transaction, with the caveat that it could return false positives, in the sense that even if an inode was not logged before in the current transaction, it could still return true, but never to return false in case the inode was logged. >From a functional point of view that is fine, but from a performance perspective it can introduce significant latencies to rename and link operations, as they will end up doing inode logging even when it is not necessary. Recently on a 5.15 kernel, an openSUSE Tumbleweed user reported package installations and upgrades, with the zypper tool, were often taking a long time to complete. With strace it could be observed that zypper was spending about 99% of its time on rename operations, and then with further analysis we checked that directory logging was happening too frequently. Taking into account that installation/upgrade of some of the packages needed a few thousand file renames, the slowdown was very noticeable for the user. The issue was caused indirectly due to an excessive number of inode evictions on a 5.15 kernel, about 100x more compared to a 5.13, 5.14 or a 5.16-rc8 kernel. While triggering the inode evictions if something outside btrfs' control, btrfs could still behave better by eliminating the false positives from the inode_logged() helper. So change inode_logged() to actually eliminate such false positives caused by inode eviction and when an inode was never logged since the filesystem was mounted, as both cases relate to when the logged_trans field of struct btrfs_inode has a value of zero. When it can not determine if the inode was logged based only on the logged_trans value, lookup for the existence of the inode item in the log tree - if it's there then we known the inode was logged, if it's not there then it can not have been logged in the current transaction. Once we determine if the inode was logged, update the logged_trans value to avoid future calls to have to search in the log tree again. Alternatively, we could start storing logged_trans in the on disk inode item structure (struct btrfs_inode_item) in the unused space it still has, but that would be a bit odd because: 1) We only care about logged_trans since the filesystem was mounted, we don't care about its value from a previous mount. Having it persisted in the inode item structure would not make the best use of the precious unused space; 2) In order to get logged_trans persisted before inode eviction, we would have to update the delayed inode when we finish logging the inode and update its logged_trans in struct btrfs_inode, which makes it a bit cumbersome since we need to check if the delayed inode exists, if not create it and populate it and deal with any errors (-ENOMEM mostly). This change is part of a patchset comprised of the following patches: 1/5 btrfs: add helper to delete a dir entry from a log tree 2/5 btrfs: pass the dentry to btrfs_log_new_name() instead of the inode 3/5 btrfs: avoid logging all directory changes during renames 4/5 btrfs: stop doing unnecessary log updates during a rename 5/5 btrfs: avoid inode logging during rename and link when possible The following test script mimics part of what the zypper tool does during package installations/upgrades. It does not triggers inode evictions, but it's similar because it triggers false positives from the inode_logged() helper, because the inodes have a logged_trans of 0, there's a log tree due to a fsync of an unrelated file and the directory inode has its last_trans field set to the current transaction: $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=10000 mkfs.btrfs -f $DEV mount $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync # Now do some change to an unrelated file and fsync it. # This is just to create a log tree to make sure that inode_logged() # does not return false when called against "testdir". xfs_io -f -c "pwrite 0 4K" -c "fsync" $MNT/foo # Do some change to testdir. This is to make sure inode_logged() # will return true when called against "testdir", because its # logged_trans is 0, it was changed in the current transaction # and there's a log tree. echo -n > $MNT/testdir/file_$((NUM_FILES + 1)) echo "Renaming $NUM_FILES files..." start=$(date +%s%N) for ((i = 1; i <= $NUM_FILES; i++)); do mv $MNT/testdir/file_$i $MNT/testdir/file_$i-RPMDELETE done end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "Renames took $dur milliseconds" umount $MNT Testing this change on a box using a non-debug kernel (Debian's default kernel config) gave the following results: NUM_FILES=10000, before patchset: 27837 ms NUM_FILES=10000, after patches 1/5 to 4/5 applied: 9236 ms (-66.8%) NUM_FILES=10000, after whole patchset applied: 8902 ms (-68.0%) NUM_FILES=5000, before patchset: 9127 ms NUM_FILES=5000, after patches 1/5 to 4/5 applied: 4640 ms (-49.2%) NUM_FILES=5000, after whole patchset applied: 4441 ms (-51.3%) NUM_FILES=2000, before patchset: 2528 ms NUM_FILES=2000, after patches 1/5 to 4/5 applied: 1983 ms (-21.6%) NUM_FILES=2000, after whole patchset applied: 1747 ms (-30.9%) NUM_FILES=1000, before patchset: 1085 ms NUM_FILES=1000, after patches 1/5 to 4/5 applied: 893 ms (-17.7%) NUM_FILES=1000, after whole patchset applied: 867 ms (-20.1%) Running dbench on the same physical machine with the following script: $ cat run-dbench.sh #!/bin/bash NUM_JOBS=$(nproc --all) DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 MOUNT_OPTIONS="-o ssd" MKFS_OPTIONS="-O no-holes -R free-space-tree" echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -D $MNT -t 120 $NUM_JOBS umount $MNT Before patchset: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 3761352 0.032 143.843 Close 2762770 0.002 2.273 Rename 159304 0.291 67.037 Unlink 759784 0.207 143.998 Deltree 72 4.028 15.977 Mkdir 36 0.003 0.006 Qpathinfo 3409780 0.013 9.678 Qfileinfo 596772 0.001 0.878 Qfsinfo 625189 0.003 1.245 Sfileinfo 306443 0.006 1.840 Find 1318106 0.063 19.798 WriteX 1871137 0.021 8.532 ReadX 5897325 0.003 3.567 LockX 12252 0.003 0.258 UnlockX 12252 0.002 0.100 Flush 263666 3.327 155.632 Throughput 980.047 MB/sec 12 clients 12 procs max_latency=155.636 ms After whole patchset applied: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 4195584 0.033 107.742 Close 3081932 0.002 1.935 Rename 177641 0.218 14.905 Unlink 847333 0.166 107.822 Deltree 118 5.315 15.247 Mkdir 59 0.004 0.048 Qpathinfo 3802612 0.014 10.302 Qfileinfo 666748 0.001 1.034 Qfsinfo 697329 0.003 0.944 Sfileinfo 341712 0.006 2.099 Find 1470365 0.065 9.359 WriteX 2093921 0.021 8.087 ReadX 6576234 0.003 3.407 LockX 13660 0.003 0.308 UnlockX 13660 0.002 0.114 Flush 294090 2.906 115.539 Throughput 1093.11 MB/sec 12 clients 12 procs max_latency=115.544 ms +11.5% throughput -25.8% max latency rename max latency -77.8% Link: https://bugzilla.opensuse.org/show_bug.cgi?id=1193549 Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-01-20 14:00:10 +03:00
if (inode_only == LOG_INODE_EXISTS && ctx->logged_before) {
Btrfs: fix fsync data loss after adding hard link to inode We have a scenario where after the fsync log replay we can lose file data that had been previously fsync'ed if we added an hard link for our inode and after that we sync'ed the fsync log (for example by fsync'ing some other file or directory). This is because when adding an hard link we updated the inode item in the log tree with an i_size value of 0. At that point the new inode item was in memory only and a subsequent fsync log replay would not make us lose the file data. However if after adding the hard link we sync the log tree to disk, by fsync'ing some other file or directory for example, we ended up losing the file data after log replay, because the inode item in the persisted log tree had an an i_size of zero. This is easy to reproduce, and the following excerpt from my test for xfstests shows this: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create one file with data and fsync it. # This made the btrfs fsync log persist the data and the inode metadata with # a correct inode->i_size (4096 bytes). $XFS_IO_PROG -f -c "pwrite -S 0xaa -b 4K 0 4K" -c "fsync" \ $SCRATCH_MNT/foo | _filter_xfs_io # Now add one hard link to our file. This made the btrfs code update the fsync # log, in memory only, with an inode metadata having a size of 0. ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link # Now force persistence of the fsync log to disk, for example, by fsyncing some # other file. touch $SCRATCH_MNT/bar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/bar # Before a power loss or crash, we could read the 4Kb of data from our file as # expected. echo "File content before:" od -t x1 $SCRATCH_MNT/foo # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # After the fsync log replay, because the fsync log had a value of 0 for our # inode's i_size, we couldn't read anymore the 4Kb of data that we previously # wrote and fsync'ed. The size of the file became 0 after the fsync log replay. echo "File content after:" od -t x1 $SCRATCH_MNT/foo Another alternative test, that doesn't need to fsync an inode in the same transaction it was created, is: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create our test file with some data. $XFS_IO_PROG -f -c "pwrite -S 0xaa -b 8K 0 8K" \ $SCRATCH_MNT/foo | _filter_xfs_io # Make sure the file is durably persisted. sync # Append some data to our file, to increase its size. $XFS_IO_PROG -f -c "pwrite -S 0xcc -b 4K 8K 4K" \ $SCRATCH_MNT/foo | _filter_xfs_io # Fsync the file, so from this point on if a crash/power failure happens, our # new data is guaranteed to be there next time the fs is mounted. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foo # Add one hard link to our file. This made btrfs write into the in memory fsync # log a special inode with generation 0 and an i_size of 0 too. Note that this # didn't update the inode in the fsync log on disk. ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link # Now make sure the in memory fsync log is durably persisted. # Creating and fsync'ing another file will do it. touch $SCRATCH_MNT/bar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/bar # As expected, before the crash/power failure, we should be able to read the # 12Kb of file data. echo "File content before:" od -t x1 $SCRATCH_MNT/foo # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # After mounting the fs again, the fsync log was replayed. # The btrfs fsync log replay code didn't update the i_size of the persisted # inode because the inode item in the log had a special generation with a # value of 0 (and it couldn't know the correct i_size, since that inode item # had a 0 i_size too). This made the last 4Kb of file data inaccessible and # effectively lost. echo "File content after:" od -t x1 $SCRATCH_MNT/foo This isn't a new issue/regression. This problem has been around since the log tree code was added in 2008: Btrfs: Add a write ahead tree log to optimize synchronous operations (commit e02119d5a7b4396c5a872582fddc8bd6d305a70a) Test cases for xfstests follow soon. CC: <stable@vger.kernel.org> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-02-13 15:30:56 +03:00
/*
* Make sure the new inode item we write to the log has
* the same isize as the current one (if it exists).
* This is necessary to prevent data loss after log
* replay, and also to prevent doing a wrong expanding
* truncate - for e.g. create file, write 4K into offset
* 0, fsync, write 4K into offset 4096, add hard link,
* fsync some other file (to sync log), power fail - if
* we use the inode's current i_size, after log replay
* we get a 8Kb file, with the last 4Kb extent as a hole
* (zeroes), as if an expanding truncate happened,
* instead of getting a file of 4Kb only.
*/
ret = logged_inode_size(log, inode, path, &logged_isize);
if (ret)
Btrfs: fix fsync data loss after adding hard link to inode We have a scenario where after the fsync log replay we can lose file data that had been previously fsync'ed if we added an hard link for our inode and after that we sync'ed the fsync log (for example by fsync'ing some other file or directory). This is because when adding an hard link we updated the inode item in the log tree with an i_size value of 0. At that point the new inode item was in memory only and a subsequent fsync log replay would not make us lose the file data. However if after adding the hard link we sync the log tree to disk, by fsync'ing some other file or directory for example, we ended up losing the file data after log replay, because the inode item in the persisted log tree had an an i_size of zero. This is easy to reproduce, and the following excerpt from my test for xfstests shows this: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create one file with data and fsync it. # This made the btrfs fsync log persist the data and the inode metadata with # a correct inode->i_size (4096 bytes). $XFS_IO_PROG -f -c "pwrite -S 0xaa -b 4K 0 4K" -c "fsync" \ $SCRATCH_MNT/foo | _filter_xfs_io # Now add one hard link to our file. This made the btrfs code update the fsync # log, in memory only, with an inode metadata having a size of 0. ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link # Now force persistence of the fsync log to disk, for example, by fsyncing some # other file. touch $SCRATCH_MNT/bar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/bar # Before a power loss or crash, we could read the 4Kb of data from our file as # expected. echo "File content before:" od -t x1 $SCRATCH_MNT/foo # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # After the fsync log replay, because the fsync log had a value of 0 for our # inode's i_size, we couldn't read anymore the 4Kb of data that we previously # wrote and fsync'ed. The size of the file became 0 after the fsync log replay. echo "File content after:" od -t x1 $SCRATCH_MNT/foo Another alternative test, that doesn't need to fsync an inode in the same transaction it was created, is: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create our test file with some data. $XFS_IO_PROG -f -c "pwrite -S 0xaa -b 8K 0 8K" \ $SCRATCH_MNT/foo | _filter_xfs_io # Make sure the file is durably persisted. sync # Append some data to our file, to increase its size. $XFS_IO_PROG -f -c "pwrite -S 0xcc -b 4K 8K 4K" \ $SCRATCH_MNT/foo | _filter_xfs_io # Fsync the file, so from this point on if a crash/power failure happens, our # new data is guaranteed to be there next time the fs is mounted. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foo # Add one hard link to our file. This made btrfs write into the in memory fsync # log a special inode with generation 0 and an i_size of 0 too. Note that this # didn't update the inode in the fsync log on disk. ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link # Now make sure the in memory fsync log is durably persisted. # Creating and fsync'ing another file will do it. touch $SCRATCH_MNT/bar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/bar # As expected, before the crash/power failure, we should be able to read the # 12Kb of file data. echo "File content before:" od -t x1 $SCRATCH_MNT/foo # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # After mounting the fs again, the fsync log was replayed. # The btrfs fsync log replay code didn't update the i_size of the persisted # inode because the inode item in the log had a special generation with a # value of 0 (and it couldn't know the correct i_size, since that inode item # had a 0 i_size too). This made the last 4Kb of file data inaccessible and # effectively lost. echo "File content after:" od -t x1 $SCRATCH_MNT/foo This isn't a new issue/regression. This problem has been around since the log tree code was added in 2008: Btrfs: Add a write ahead tree log to optimize synchronous operations (commit e02119d5a7b4396c5a872582fddc8bd6d305a70a) Test cases for xfstests follow soon. CC: <stable@vger.kernel.org> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-02-13 15:30:56 +03:00
goto out_unlock;
}
Btrfs: don't remove extents and xattrs when logging new names If we are recording in the tree log that an inode has new names (new hard links were added), we would drop items, belonging to the inode, that we shouldn't: 1) When the flag BTRFS_INODE_COPY_EVERYTHING is set in the inode's runtime flags, we ended up dropping all the extent and xattr items that were previously logged. This was done only in memory, since logging a new name doesn't imply syncing the log; 2) When the flag BTRFS_INODE_COPY_EVERYTHING is set in the inode's runtime flags, we ended up dropping all the xattr items that were previously logged. Like the case before, this was done only in memory because logging a new name doesn't imply syncing the log. This led to some surprises in scenarios such as the following: 1) write some extents to an inode; 2) fsync the inode; 3) truncate the inode or delete/modify some of its xattrs 4) add a new hard link for that inode 5) fsync some other file, to force the log tree to be durably persisted 6) power failure happens The next time the fs is mounted, the fsync log replay code is executed, and the resulting file doesn't have the content it had when the last fsync against it was performed, instead if has a content matching what it had when the last transaction commit happened. So change the behaviour such that when a new name is logged, only the inode item and reference items are processed. This is easy to reproduce with the test I just made for xfstests, whose main body is: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create our test file with some data. $XFS_IO_PROG -f -c "pwrite -S 0xaa -b 8K 0 8K" \ $SCRATCH_MNT/foo | _filter_xfs_io # Make sure the file is durably persisted. sync # Append some data to our file, to increase its size. $XFS_IO_PROG -f -c "pwrite -S 0xcc -b 4K 8K 4K" \ $SCRATCH_MNT/foo | _filter_xfs_io # Fsync the file, so from this point on if a crash/power failure happens, our # new data is guaranteed to be there next time the fs is mounted. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foo # Now shrink our file to 5000 bytes. $XFS_IO_PROG -c "truncate 5000" $SCRATCH_MNT/foo # Now do an expanding truncate to a size larger than what we had when we last # fsync'ed our file. This is just to verify that after power failure and # replaying the fsync log, our file matches what it was when we last fsync'ed # it - 12Kb size, first 8Kb of data had a value of 0xaa and the last 4Kb of # data had a value of 0xcc. $XFS_IO_PROG -c "truncate 32K" $SCRATCH_MNT/foo # Add one hard link to our file. This made btrfs drop all of our file's # metadata from the fsync log, including the metadata relative to the # extent we just wrote and fsync'ed. This change was made only to the fsync # log in memory, so adding the hard link alone doesn't change the persisted # fsync log. This happened because the previous truncates set the runtime # flag BTRFS_INODE_NEEDS_FULL_SYNC in the btrfs inode structure. ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link # Now make sure the in memory fsync log is durably persisted. # Creating and fsync'ing another file will do it. # After this our persisted fsync log will no longer have metadata for our file # foo that points to the extent we wrote and fsync'ed before. touch $SCRATCH_MNT/bar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/bar # As expected, before the crash/power failure, we should be able to see a file # with a size of 32Kb, with its first 5000 bytes having the value 0xaa and all # the remaining bytes with value 0x00. echo "File content before:" od -t x1 $SCRATCH_MNT/foo # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # After mounting the fs again, the fsync log was replayed. # The expected result is to see a file with a size of 12Kb, with its first 8Kb # of data having the value 0xaa and its last 4Kb of data having a value of 0xcc. # The btrfs bug used to leave the file as it used te be as of the last # transaction commit - that is, with a size of 8Kb with all bytes having a # value of 0xaa. echo "File content after:" od -t x1 $SCRATCH_MNT/foo The test case for xfstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-02-13 19:56:14 +03:00
if (test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
&inode->runtime_flags)) {
Btrfs: don't remove extents and xattrs when logging new names If we are recording in the tree log that an inode has new names (new hard links were added), we would drop items, belonging to the inode, that we shouldn't: 1) When the flag BTRFS_INODE_COPY_EVERYTHING is set in the inode's runtime flags, we ended up dropping all the extent and xattr items that were previously logged. This was done only in memory, since logging a new name doesn't imply syncing the log; 2) When the flag BTRFS_INODE_COPY_EVERYTHING is set in the inode's runtime flags, we ended up dropping all the xattr items that were previously logged. Like the case before, this was done only in memory because logging a new name doesn't imply syncing the log. This led to some surprises in scenarios such as the following: 1) write some extents to an inode; 2) fsync the inode; 3) truncate the inode or delete/modify some of its xattrs 4) add a new hard link for that inode 5) fsync some other file, to force the log tree to be durably persisted 6) power failure happens The next time the fs is mounted, the fsync log replay code is executed, and the resulting file doesn't have the content it had when the last fsync against it was performed, instead if has a content matching what it had when the last transaction commit happened. So change the behaviour such that when a new name is logged, only the inode item and reference items are processed. This is easy to reproduce with the test I just made for xfstests, whose main body is: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create our test file with some data. $XFS_IO_PROG -f -c "pwrite -S 0xaa -b 8K 0 8K" \ $SCRATCH_MNT/foo | _filter_xfs_io # Make sure the file is durably persisted. sync # Append some data to our file, to increase its size. $XFS_IO_PROG -f -c "pwrite -S 0xcc -b 4K 8K 4K" \ $SCRATCH_MNT/foo | _filter_xfs_io # Fsync the file, so from this point on if a crash/power failure happens, our # new data is guaranteed to be there next time the fs is mounted. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foo # Now shrink our file to 5000 bytes. $XFS_IO_PROG -c "truncate 5000" $SCRATCH_MNT/foo # Now do an expanding truncate to a size larger than what we had when we last # fsync'ed our file. This is just to verify that after power failure and # replaying the fsync log, our file matches what it was when we last fsync'ed # it - 12Kb size, first 8Kb of data had a value of 0xaa and the last 4Kb of # data had a value of 0xcc. $XFS_IO_PROG -c "truncate 32K" $SCRATCH_MNT/foo # Add one hard link to our file. This made btrfs drop all of our file's # metadata from the fsync log, including the metadata relative to the # extent we just wrote and fsync'ed. This change was made only to the fsync # log in memory, so adding the hard link alone doesn't change the persisted # fsync log. This happened because the previous truncates set the runtime # flag BTRFS_INODE_NEEDS_FULL_SYNC in the btrfs inode structure. ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link # Now make sure the in memory fsync log is durably persisted. # Creating and fsync'ing another file will do it. # After this our persisted fsync log will no longer have metadata for our file # foo that points to the extent we wrote and fsync'ed before. touch $SCRATCH_MNT/bar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/bar # As expected, before the crash/power failure, we should be able to see a file # with a size of 32Kb, with its first 5000 bytes having the value 0xaa and all # the remaining bytes with value 0x00. echo "File content before:" od -t x1 $SCRATCH_MNT/foo # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # After mounting the fs again, the fsync log was replayed. # The expected result is to see a file with a size of 12Kb, with its first 8Kb # of data having the value 0xaa and its last 4Kb of data having a value of 0xcc. # The btrfs bug used to leave the file as it used te be as of the last # transaction commit - that is, with a size of 8Kb with all bytes having a # value of 0xaa. echo "File content after:" od -t x1 $SCRATCH_MNT/foo The test case for xfstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-02-13 19:56:14 +03:00
if (inode_only == LOG_INODE_EXISTS) {
Btrfs: remove deleted xattrs on fsync log replay If we deleted xattrs from a file and fsynced the file, after a log replay the xattrs would remain associated to the file. This was an unexpected behaviour and differs from what other filesystems do, such as for example xfs and ext3/4. Fix this by, on fsync log replay, check if every xattr in the fs/subvol tree (that belongs to a logged inode) has a matching xattr in the log, and if it does not, delete it from the fs/subvol tree. This is a similar approach to what we do for dentries when we replay a directory from the fsync log. This issue is trivial to reproduce, and the following excerpt from my test for xfstests triggers the issue: _crash_and_mount() { # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey } rm -f $seqres.full _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create out test file and add 3 xattrs to it. touch $SCRATCH_MNT/foobar $SETFATTR_PROG -n user.attr1 -v val1 $SCRATCH_MNT/foobar $SETFATTR_PROG -n user.attr2 -v val2 $SCRATCH_MNT/foobar $SETFATTR_PROG -n user.attr3 -v val3 $SCRATCH_MNT/foobar # Make sure everything is durably persisted. sync # Now delete the second xattr and fsync the inode. $SETFATTR_PROG -x user.attr2 $SCRATCH_MNT/foobar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foobar _crash_and_mount # After the fsync log is replayed, the file should have only 2 xattrs, the ones # named user.attr1 and user.attr3. The btrfs fsync log replay bug left the file # with the 3 xattrs that we had before deleting the second one and fsyncing the # file. echo "xattr names and values after first fsync log replay:" $GETFATTR_PROG --absolute-names --dump $SCRATCH_MNT/foobar | _filter_scratch # Now write some data to our file, fsync it, remove the first xattr, add a new # hard link to our file and commit the fsync log by fsyncing some other new # file. This is to verify that after log replay our first xattr does not exist # anymore. echo "hello world!" >> $SCRATCH_MNT/foobar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foobar $SETFATTR_PROG -x user.attr1 $SCRATCH_MNT/foobar ln $SCRATCH_MNT/foobar $SCRATCH_MNT/foobar_link touch $SCRATCH_MNT/qwerty $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/qwerty _crash_and_mount # Now only the xattr with name user.attr3 should be set in our file. echo "xattr names and values after second fsync log replay:" $GETFATTR_PROG --absolute-names --dump $SCRATCH_MNT/foobar | _filter_scratch status=0 exit The expected golden output, which is produced with this patch applied or when testing against xfs or ext3/4, is: xattr names and values after first fsync log replay: # file: SCRATCH_MNT/foobar user.attr1="val1" user.attr3="val3" xattr names and values after second fsync log replay: # file: SCRATCH_MNT/foobar user.attr3="val3" Without this patch applied, the output is: xattr names and values after first fsync log replay: # file: SCRATCH_MNT/foobar user.attr1="val1" user.attr2="val2" user.attr3="val3" xattr names and values after second fsync log replay: # file: SCRATCH_MNT/foobar user.attr1="val1" user.attr2="val2" user.attr3="val3" A patch with a test case for xfstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-02-23 22:53:35 +03:00
max_key.type = BTRFS_XATTR_ITEM_KEY;
btrfs: avoid inode logging during rename and link when possible During a rename or link operation, we need to determine if an inode was previously logged or not, and if it was, do some update to the logged inode. We used to rely exclusively on the logged_trans field of struct btrfs_inode to determine that, but that was not reliable because the value of that field is not persisted in the inode item, so it's lost when an inode is evicted and loaded back again. That led to several issues in the past, such as not persisting deletions (such as the case fixed by commit 803f0f64d17769 ("Btrfs: fix fsync not persisting dentry deletions due to inode evictions")), or resulting in losing a file after an inode eviction followed by a rename (commit ecc64fab7d49c6 ("btrfs: fix lost inode on log replay after mix of fsync, rename and inode eviction")), besides other issues. So the inode_logged() helper was introduced and used to determine if an inode was possibly logged before in the current transaction, with the caveat that it could return false positives, in the sense that even if an inode was not logged before in the current transaction, it could still return true, but never to return false in case the inode was logged. >From a functional point of view that is fine, but from a performance perspective it can introduce significant latencies to rename and link operations, as they will end up doing inode logging even when it is not necessary. Recently on a 5.15 kernel, an openSUSE Tumbleweed user reported package installations and upgrades, with the zypper tool, were often taking a long time to complete. With strace it could be observed that zypper was spending about 99% of its time on rename operations, and then with further analysis we checked that directory logging was happening too frequently. Taking into account that installation/upgrade of some of the packages needed a few thousand file renames, the slowdown was very noticeable for the user. The issue was caused indirectly due to an excessive number of inode evictions on a 5.15 kernel, about 100x more compared to a 5.13, 5.14 or a 5.16-rc8 kernel. While triggering the inode evictions if something outside btrfs' control, btrfs could still behave better by eliminating the false positives from the inode_logged() helper. So change inode_logged() to actually eliminate such false positives caused by inode eviction and when an inode was never logged since the filesystem was mounted, as both cases relate to when the logged_trans field of struct btrfs_inode has a value of zero. When it can not determine if the inode was logged based only on the logged_trans value, lookup for the existence of the inode item in the log tree - if it's there then we known the inode was logged, if it's not there then it can not have been logged in the current transaction. Once we determine if the inode was logged, update the logged_trans value to avoid future calls to have to search in the log tree again. Alternatively, we could start storing logged_trans in the on disk inode item structure (struct btrfs_inode_item) in the unused space it still has, but that would be a bit odd because: 1) We only care about logged_trans since the filesystem was mounted, we don't care about its value from a previous mount. Having it persisted in the inode item structure would not make the best use of the precious unused space; 2) In order to get logged_trans persisted before inode eviction, we would have to update the delayed inode when we finish logging the inode and update its logged_trans in struct btrfs_inode, which makes it a bit cumbersome since we need to check if the delayed inode exists, if not create it and populate it and deal with any errors (-ENOMEM mostly). This change is part of a patchset comprised of the following patches: 1/5 btrfs: add helper to delete a dir entry from a log tree 2/5 btrfs: pass the dentry to btrfs_log_new_name() instead of the inode 3/5 btrfs: avoid logging all directory changes during renames 4/5 btrfs: stop doing unnecessary log updates during a rename 5/5 btrfs: avoid inode logging during rename and link when possible The following test script mimics part of what the zypper tool does during package installations/upgrades. It does not triggers inode evictions, but it's similar because it triggers false positives from the inode_logged() helper, because the inodes have a logged_trans of 0, there's a log tree due to a fsync of an unrelated file and the directory inode has its last_trans field set to the current transaction: $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=10000 mkfs.btrfs -f $DEV mount $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync # Now do some change to an unrelated file and fsync it. # This is just to create a log tree to make sure that inode_logged() # does not return false when called against "testdir". xfs_io -f -c "pwrite 0 4K" -c "fsync" $MNT/foo # Do some change to testdir. This is to make sure inode_logged() # will return true when called against "testdir", because its # logged_trans is 0, it was changed in the current transaction # and there's a log tree. echo -n > $MNT/testdir/file_$((NUM_FILES + 1)) echo "Renaming $NUM_FILES files..." start=$(date +%s%N) for ((i = 1; i <= $NUM_FILES; i++)); do mv $MNT/testdir/file_$i $MNT/testdir/file_$i-RPMDELETE done end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "Renames took $dur milliseconds" umount $MNT Testing this change on a box using a non-debug kernel (Debian's default kernel config) gave the following results: NUM_FILES=10000, before patchset: 27837 ms NUM_FILES=10000, after patches 1/5 to 4/5 applied: 9236 ms (-66.8%) NUM_FILES=10000, after whole patchset applied: 8902 ms (-68.0%) NUM_FILES=5000, before patchset: 9127 ms NUM_FILES=5000, after patches 1/5 to 4/5 applied: 4640 ms (-49.2%) NUM_FILES=5000, after whole patchset applied: 4441 ms (-51.3%) NUM_FILES=2000, before patchset: 2528 ms NUM_FILES=2000, after patches 1/5 to 4/5 applied: 1983 ms (-21.6%) NUM_FILES=2000, after whole patchset applied: 1747 ms (-30.9%) NUM_FILES=1000, before patchset: 1085 ms NUM_FILES=1000, after patches 1/5 to 4/5 applied: 893 ms (-17.7%) NUM_FILES=1000, after whole patchset applied: 867 ms (-20.1%) Running dbench on the same physical machine with the following script: $ cat run-dbench.sh #!/bin/bash NUM_JOBS=$(nproc --all) DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 MOUNT_OPTIONS="-o ssd" MKFS_OPTIONS="-O no-holes -R free-space-tree" echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -D $MNT -t 120 $NUM_JOBS umount $MNT Before patchset: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 3761352 0.032 143.843 Close 2762770 0.002 2.273 Rename 159304 0.291 67.037 Unlink 759784 0.207 143.998 Deltree 72 4.028 15.977 Mkdir 36 0.003 0.006 Qpathinfo 3409780 0.013 9.678 Qfileinfo 596772 0.001 0.878 Qfsinfo 625189 0.003 1.245 Sfileinfo 306443 0.006 1.840 Find 1318106 0.063 19.798 WriteX 1871137 0.021 8.532 ReadX 5897325 0.003 3.567 LockX 12252 0.003 0.258 UnlockX 12252 0.002 0.100 Flush 263666 3.327 155.632 Throughput 980.047 MB/sec 12 clients 12 procs max_latency=155.636 ms After whole patchset applied: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 4195584 0.033 107.742 Close 3081932 0.002 1.935 Rename 177641 0.218 14.905 Unlink 847333 0.166 107.822 Deltree 118 5.315 15.247 Mkdir 59 0.004 0.048 Qpathinfo 3802612 0.014 10.302 Qfileinfo 666748 0.001 1.034 Qfsinfo 697329 0.003 0.944 Sfileinfo 341712 0.006 2.099 Find 1470365 0.065 9.359 WriteX 2093921 0.021 8.087 ReadX 6576234 0.003 3.407 LockX 13660 0.003 0.308 UnlockX 13660 0.002 0.114 Flush 294090 2.906 115.539 Throughput 1093.11 MB/sec 12 clients 12 procs max_latency=115.544 ms +11.5% throughput -25.8% max latency rename max latency -77.8% Link: https://bugzilla.opensuse.org/show_bug.cgi?id=1193549 Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-01-20 14:00:10 +03:00
if (ctx->logged_before)
ret = drop_inode_items(trans, log, path,
inode, max_key.type);
Btrfs: don't remove extents and xattrs when logging new names If we are recording in the tree log that an inode has new names (new hard links were added), we would drop items, belonging to the inode, that we shouldn't: 1) When the flag BTRFS_INODE_COPY_EVERYTHING is set in the inode's runtime flags, we ended up dropping all the extent and xattr items that were previously logged. This was done only in memory, since logging a new name doesn't imply syncing the log; 2) When the flag BTRFS_INODE_COPY_EVERYTHING is set in the inode's runtime flags, we ended up dropping all the xattr items that were previously logged. Like the case before, this was done only in memory because logging a new name doesn't imply syncing the log. This led to some surprises in scenarios such as the following: 1) write some extents to an inode; 2) fsync the inode; 3) truncate the inode or delete/modify some of its xattrs 4) add a new hard link for that inode 5) fsync some other file, to force the log tree to be durably persisted 6) power failure happens The next time the fs is mounted, the fsync log replay code is executed, and the resulting file doesn't have the content it had when the last fsync against it was performed, instead if has a content matching what it had when the last transaction commit happened. So change the behaviour such that when a new name is logged, only the inode item and reference items are processed. This is easy to reproduce with the test I just made for xfstests, whose main body is: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create our test file with some data. $XFS_IO_PROG -f -c "pwrite -S 0xaa -b 8K 0 8K" \ $SCRATCH_MNT/foo | _filter_xfs_io # Make sure the file is durably persisted. sync # Append some data to our file, to increase its size. $XFS_IO_PROG -f -c "pwrite -S 0xcc -b 4K 8K 4K" \ $SCRATCH_MNT/foo | _filter_xfs_io # Fsync the file, so from this point on if a crash/power failure happens, our # new data is guaranteed to be there next time the fs is mounted. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foo # Now shrink our file to 5000 bytes. $XFS_IO_PROG -c "truncate 5000" $SCRATCH_MNT/foo # Now do an expanding truncate to a size larger than what we had when we last # fsync'ed our file. This is just to verify that after power failure and # replaying the fsync log, our file matches what it was when we last fsync'ed # it - 12Kb size, first 8Kb of data had a value of 0xaa and the last 4Kb of # data had a value of 0xcc. $XFS_IO_PROG -c "truncate 32K" $SCRATCH_MNT/foo # Add one hard link to our file. This made btrfs drop all of our file's # metadata from the fsync log, including the metadata relative to the # extent we just wrote and fsync'ed. This change was made only to the fsync # log in memory, so adding the hard link alone doesn't change the persisted # fsync log. This happened because the previous truncates set the runtime # flag BTRFS_INODE_NEEDS_FULL_SYNC in the btrfs inode structure. ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link # Now make sure the in memory fsync log is durably persisted. # Creating and fsync'ing another file will do it. # After this our persisted fsync log will no longer have metadata for our file # foo that points to the extent we wrote and fsync'ed before. touch $SCRATCH_MNT/bar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/bar # As expected, before the crash/power failure, we should be able to see a file # with a size of 32Kb, with its first 5000 bytes having the value 0xaa and all # the remaining bytes with value 0x00. echo "File content before:" od -t x1 $SCRATCH_MNT/foo # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # After mounting the fs again, the fsync log was replayed. # The expected result is to see a file with a size of 12Kb, with its first 8Kb # of data having the value 0xaa and its last 4Kb of data having a value of 0xcc. # The btrfs bug used to leave the file as it used te be as of the last # transaction commit - that is, with a size of 8Kb with all bytes having a # value of 0xaa. echo "File content after:" od -t x1 $SCRATCH_MNT/foo The test case for xfstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-02-13 19:56:14 +03:00
} else {
clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
&inode->runtime_flags);
Btrfs: don't remove extents and xattrs when logging new names If we are recording in the tree log that an inode has new names (new hard links were added), we would drop items, belonging to the inode, that we shouldn't: 1) When the flag BTRFS_INODE_COPY_EVERYTHING is set in the inode's runtime flags, we ended up dropping all the extent and xattr items that were previously logged. This was done only in memory, since logging a new name doesn't imply syncing the log; 2) When the flag BTRFS_INODE_COPY_EVERYTHING is set in the inode's runtime flags, we ended up dropping all the xattr items that were previously logged. Like the case before, this was done only in memory because logging a new name doesn't imply syncing the log. This led to some surprises in scenarios such as the following: 1) write some extents to an inode; 2) fsync the inode; 3) truncate the inode or delete/modify some of its xattrs 4) add a new hard link for that inode 5) fsync some other file, to force the log tree to be durably persisted 6) power failure happens The next time the fs is mounted, the fsync log replay code is executed, and the resulting file doesn't have the content it had when the last fsync against it was performed, instead if has a content matching what it had when the last transaction commit happened. So change the behaviour such that when a new name is logged, only the inode item and reference items are processed. This is easy to reproduce with the test I just made for xfstests, whose main body is: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create our test file with some data. $XFS_IO_PROG -f -c "pwrite -S 0xaa -b 8K 0 8K" \ $SCRATCH_MNT/foo | _filter_xfs_io # Make sure the file is durably persisted. sync # Append some data to our file, to increase its size. $XFS_IO_PROG -f -c "pwrite -S 0xcc -b 4K 8K 4K" \ $SCRATCH_MNT/foo | _filter_xfs_io # Fsync the file, so from this point on if a crash/power failure happens, our # new data is guaranteed to be there next time the fs is mounted. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foo # Now shrink our file to 5000 bytes. $XFS_IO_PROG -c "truncate 5000" $SCRATCH_MNT/foo # Now do an expanding truncate to a size larger than what we had when we last # fsync'ed our file. This is just to verify that after power failure and # replaying the fsync log, our file matches what it was when we last fsync'ed # it - 12Kb size, first 8Kb of data had a value of 0xaa and the last 4Kb of # data had a value of 0xcc. $XFS_IO_PROG -c "truncate 32K" $SCRATCH_MNT/foo # Add one hard link to our file. This made btrfs drop all of our file's # metadata from the fsync log, including the metadata relative to the # extent we just wrote and fsync'ed. This change was made only to the fsync # log in memory, so adding the hard link alone doesn't change the persisted # fsync log. This happened because the previous truncates set the runtime # flag BTRFS_INODE_NEEDS_FULL_SYNC in the btrfs inode structure. ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link # Now make sure the in memory fsync log is durably persisted. # Creating and fsync'ing another file will do it. # After this our persisted fsync log will no longer have metadata for our file # foo that points to the extent we wrote and fsync'ed before. touch $SCRATCH_MNT/bar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/bar # As expected, before the crash/power failure, we should be able to see a file # with a size of 32Kb, with its first 5000 bytes having the value 0xaa and all # the remaining bytes with value 0x00. echo "File content before:" od -t x1 $SCRATCH_MNT/foo # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # After mounting the fs again, the fsync log was replayed. # The expected result is to see a file with a size of 12Kb, with its first 8Kb # of data having the value 0xaa and its last 4Kb of data having a value of 0xcc. # The btrfs bug used to leave the file as it used te be as of the last # transaction commit - that is, with a size of 8Kb with all bytes having a # value of 0xaa. echo "File content after:" od -t x1 $SCRATCH_MNT/foo The test case for xfstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-02-13 19:56:14 +03:00
clear_bit(BTRFS_INODE_COPY_EVERYTHING,
&inode->runtime_flags);
btrfs: avoid inode logging during rename and link when possible During a rename or link operation, we need to determine if an inode was previously logged or not, and if it was, do some update to the logged inode. We used to rely exclusively on the logged_trans field of struct btrfs_inode to determine that, but that was not reliable because the value of that field is not persisted in the inode item, so it's lost when an inode is evicted and loaded back again. That led to several issues in the past, such as not persisting deletions (such as the case fixed by commit 803f0f64d17769 ("Btrfs: fix fsync not persisting dentry deletions due to inode evictions")), or resulting in losing a file after an inode eviction followed by a rename (commit ecc64fab7d49c6 ("btrfs: fix lost inode on log replay after mix of fsync, rename and inode eviction")), besides other issues. So the inode_logged() helper was introduced and used to determine if an inode was possibly logged before in the current transaction, with the caveat that it could return false positives, in the sense that even if an inode was not logged before in the current transaction, it could still return true, but never to return false in case the inode was logged. >From a functional point of view that is fine, but from a performance perspective it can introduce significant latencies to rename and link operations, as they will end up doing inode logging even when it is not necessary. Recently on a 5.15 kernel, an openSUSE Tumbleweed user reported package installations and upgrades, with the zypper tool, were often taking a long time to complete. With strace it could be observed that zypper was spending about 99% of its time on rename operations, and then with further analysis we checked that directory logging was happening too frequently. Taking into account that installation/upgrade of some of the packages needed a few thousand file renames, the slowdown was very noticeable for the user. The issue was caused indirectly due to an excessive number of inode evictions on a 5.15 kernel, about 100x more compared to a 5.13, 5.14 or a 5.16-rc8 kernel. While triggering the inode evictions if something outside btrfs' control, btrfs could still behave better by eliminating the false positives from the inode_logged() helper. So change inode_logged() to actually eliminate such false positives caused by inode eviction and when an inode was never logged since the filesystem was mounted, as both cases relate to when the logged_trans field of struct btrfs_inode has a value of zero. When it can not determine if the inode was logged based only on the logged_trans value, lookup for the existence of the inode item in the log tree - if it's there then we known the inode was logged, if it's not there then it can not have been logged in the current transaction. Once we determine if the inode was logged, update the logged_trans value to avoid future calls to have to search in the log tree again. Alternatively, we could start storing logged_trans in the on disk inode item structure (struct btrfs_inode_item) in the unused space it still has, but that would be a bit odd because: 1) We only care about logged_trans since the filesystem was mounted, we don't care about its value from a previous mount. Having it persisted in the inode item structure would not make the best use of the precious unused space; 2) In order to get logged_trans persisted before inode eviction, we would have to update the delayed inode when we finish logging the inode and update its logged_trans in struct btrfs_inode, which makes it a bit cumbersome since we need to check if the delayed inode exists, if not create it and populate it and deal with any errors (-ENOMEM mostly). This change is part of a patchset comprised of the following patches: 1/5 btrfs: add helper to delete a dir entry from a log tree 2/5 btrfs: pass the dentry to btrfs_log_new_name() instead of the inode 3/5 btrfs: avoid logging all directory changes during renames 4/5 btrfs: stop doing unnecessary log updates during a rename 5/5 btrfs: avoid inode logging during rename and link when possible The following test script mimics part of what the zypper tool does during package installations/upgrades. It does not triggers inode evictions, but it's similar because it triggers false positives from the inode_logged() helper, because the inodes have a logged_trans of 0, there's a log tree due to a fsync of an unrelated file and the directory inode has its last_trans field set to the current transaction: $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=10000 mkfs.btrfs -f $DEV mount $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync # Now do some change to an unrelated file and fsync it. # This is just to create a log tree to make sure that inode_logged() # does not return false when called against "testdir". xfs_io -f -c "pwrite 0 4K" -c "fsync" $MNT/foo # Do some change to testdir. This is to make sure inode_logged() # will return true when called against "testdir", because its # logged_trans is 0, it was changed in the current transaction # and there's a log tree. echo -n > $MNT/testdir/file_$((NUM_FILES + 1)) echo "Renaming $NUM_FILES files..." start=$(date +%s%N) for ((i = 1; i <= $NUM_FILES; i++)); do mv $MNT/testdir/file_$i $MNT/testdir/file_$i-RPMDELETE done end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "Renames took $dur milliseconds" umount $MNT Testing this change on a box using a non-debug kernel (Debian's default kernel config) gave the following results: NUM_FILES=10000, before patchset: 27837 ms NUM_FILES=10000, after patches 1/5 to 4/5 applied: 9236 ms (-66.8%) NUM_FILES=10000, after whole patchset applied: 8902 ms (-68.0%) NUM_FILES=5000, before patchset: 9127 ms NUM_FILES=5000, after patches 1/5 to 4/5 applied: 4640 ms (-49.2%) NUM_FILES=5000, after whole patchset applied: 4441 ms (-51.3%) NUM_FILES=2000, before patchset: 2528 ms NUM_FILES=2000, after patches 1/5 to 4/5 applied: 1983 ms (-21.6%) NUM_FILES=2000, after whole patchset applied: 1747 ms (-30.9%) NUM_FILES=1000, before patchset: 1085 ms NUM_FILES=1000, after patches 1/5 to 4/5 applied: 893 ms (-17.7%) NUM_FILES=1000, after whole patchset applied: 867 ms (-20.1%) Running dbench on the same physical machine with the following script: $ cat run-dbench.sh #!/bin/bash NUM_JOBS=$(nproc --all) DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 MOUNT_OPTIONS="-o ssd" MKFS_OPTIONS="-O no-holes -R free-space-tree" echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -D $MNT -t 120 $NUM_JOBS umount $MNT Before patchset: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 3761352 0.032 143.843 Close 2762770 0.002 2.273 Rename 159304 0.291 67.037 Unlink 759784 0.207 143.998 Deltree 72 4.028 15.977 Mkdir 36 0.003 0.006 Qpathinfo 3409780 0.013 9.678 Qfileinfo 596772 0.001 0.878 Qfsinfo 625189 0.003 1.245 Sfileinfo 306443 0.006 1.840 Find 1318106 0.063 19.798 WriteX 1871137 0.021 8.532 ReadX 5897325 0.003 3.567 LockX 12252 0.003 0.258 UnlockX 12252 0.002 0.100 Flush 263666 3.327 155.632 Throughput 980.047 MB/sec 12 clients 12 procs max_latency=155.636 ms After whole patchset applied: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 4195584 0.033 107.742 Close 3081932 0.002 1.935 Rename 177641 0.218 14.905 Unlink 847333 0.166 107.822 Deltree 118 5.315 15.247 Mkdir 59 0.004 0.048 Qpathinfo 3802612 0.014 10.302 Qfileinfo 666748 0.001 1.034 Qfsinfo 697329 0.003 0.944 Sfileinfo 341712 0.006 2.099 Find 1470365 0.065 9.359 WriteX 2093921 0.021 8.087 ReadX 6576234 0.003 3.407 LockX 13660 0.003 0.308 UnlockX 13660 0.002 0.114 Flush 294090 2.906 115.539 Throughput 1093.11 MB/sec 12 clients 12 procs max_latency=115.544 ms +11.5% throughput -25.8% max latency rename max latency -77.8% Link: https://bugzilla.opensuse.org/show_bug.cgi?id=1193549 Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-01-20 14:00:10 +03:00
if (ctx->logged_before)
btrfs: avoid expensive search when truncating inode items from the log Whenever we are logging a file inode in full sync mode we call btrfs_truncate_inode_items() to delete items of the inode we may have previously logged. That results in doing a btree search for deletion, which is expensive because it always acquires write locks for extent buffers at levels 2, 1 and 0, and it balances any node that is less than half full. Acquiring the write locks can block the task if the extent buffers are already locked by another task or block other tasks attempting to lock them, which is specially bad in case of log trees since they are small due to their short life, with a root node at a level typically not greater than level 2. If we know that we are logging the inode for the first time in the current transaction, we can skip the call to btrfs_truncate_inode_items(), avoiding the deletion search. This change does that. This patch is part of a patch set comprised of the following patches: btrfs: check if a log tree exists at inode_logged() btrfs: remove no longer needed checks for NULL log context btrfs: do not log new dentries when logging that a new name exists btrfs: always update the logged transaction when logging new names btrfs: avoid expensive search when dropping inode items from log btrfs: add helper to truncate inode items when logging inode btrfs: avoid expensive search when truncating inode items from the log btrfs: avoid search for logged i_size when logging inode if possible btrfs: avoid attempt to drop extents when logging inode for the first time btrfs: do not commit delayed inode when logging a file in full sync mode This is patch 7/10 and test results are listed in the change log of the last patch in the set. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-08-31 17:30:37 +03:00
ret = truncate_inode_items(trans, log,
inode, 0, 0);
Btrfs: don't remove extents and xattrs when logging new names If we are recording in the tree log that an inode has new names (new hard links were added), we would drop items, belonging to the inode, that we shouldn't: 1) When the flag BTRFS_INODE_COPY_EVERYTHING is set in the inode's runtime flags, we ended up dropping all the extent and xattr items that were previously logged. This was done only in memory, since logging a new name doesn't imply syncing the log; 2) When the flag BTRFS_INODE_COPY_EVERYTHING is set in the inode's runtime flags, we ended up dropping all the xattr items that were previously logged. Like the case before, this was done only in memory because logging a new name doesn't imply syncing the log. This led to some surprises in scenarios such as the following: 1) write some extents to an inode; 2) fsync the inode; 3) truncate the inode or delete/modify some of its xattrs 4) add a new hard link for that inode 5) fsync some other file, to force the log tree to be durably persisted 6) power failure happens The next time the fs is mounted, the fsync log replay code is executed, and the resulting file doesn't have the content it had when the last fsync against it was performed, instead if has a content matching what it had when the last transaction commit happened. So change the behaviour such that when a new name is logged, only the inode item and reference items are processed. This is easy to reproduce with the test I just made for xfstests, whose main body is: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create our test file with some data. $XFS_IO_PROG -f -c "pwrite -S 0xaa -b 8K 0 8K" \ $SCRATCH_MNT/foo | _filter_xfs_io # Make sure the file is durably persisted. sync # Append some data to our file, to increase its size. $XFS_IO_PROG -f -c "pwrite -S 0xcc -b 4K 8K 4K" \ $SCRATCH_MNT/foo | _filter_xfs_io # Fsync the file, so from this point on if a crash/power failure happens, our # new data is guaranteed to be there next time the fs is mounted. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foo # Now shrink our file to 5000 bytes. $XFS_IO_PROG -c "truncate 5000" $SCRATCH_MNT/foo # Now do an expanding truncate to a size larger than what we had when we last # fsync'ed our file. This is just to verify that after power failure and # replaying the fsync log, our file matches what it was when we last fsync'ed # it - 12Kb size, first 8Kb of data had a value of 0xaa and the last 4Kb of # data had a value of 0xcc. $XFS_IO_PROG -c "truncate 32K" $SCRATCH_MNT/foo # Add one hard link to our file. This made btrfs drop all of our file's # metadata from the fsync log, including the metadata relative to the # extent we just wrote and fsync'ed. This change was made only to the fsync # log in memory, so adding the hard link alone doesn't change the persisted # fsync log. This happened because the previous truncates set the runtime # flag BTRFS_INODE_NEEDS_FULL_SYNC in the btrfs inode structure. ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link # Now make sure the in memory fsync log is durably persisted. # Creating and fsync'ing another file will do it. # After this our persisted fsync log will no longer have metadata for our file # foo that points to the extent we wrote and fsync'ed before. touch $SCRATCH_MNT/bar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/bar # As expected, before the crash/power failure, we should be able to see a file # with a size of 32Kb, with its first 5000 bytes having the value 0xaa and all # the remaining bytes with value 0x00. echo "File content before:" od -t x1 $SCRATCH_MNT/foo # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # After mounting the fs again, the fsync log was replayed. # The expected result is to see a file with a size of 12Kb, with its first 8Kb # of data having the value 0xaa and its last 4Kb of data having a value of 0xcc. # The btrfs bug used to leave the file as it used te be as of the last # transaction commit - that is, with a size of 8Kb with all bytes having a # value of 0xaa. echo "File content after:" od -t x1 $SCRATCH_MNT/foo The test case for xfstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-02-13 19:56:14 +03:00
}
Btrfs: remove deleted xattrs on fsync log replay If we deleted xattrs from a file and fsynced the file, after a log replay the xattrs would remain associated to the file. This was an unexpected behaviour and differs from what other filesystems do, such as for example xfs and ext3/4. Fix this by, on fsync log replay, check if every xattr in the fs/subvol tree (that belongs to a logged inode) has a matching xattr in the log, and if it does not, delete it from the fs/subvol tree. This is a similar approach to what we do for dentries when we replay a directory from the fsync log. This issue is trivial to reproduce, and the following excerpt from my test for xfstests triggers the issue: _crash_and_mount() { # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey } rm -f $seqres.full _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create out test file and add 3 xattrs to it. touch $SCRATCH_MNT/foobar $SETFATTR_PROG -n user.attr1 -v val1 $SCRATCH_MNT/foobar $SETFATTR_PROG -n user.attr2 -v val2 $SCRATCH_MNT/foobar $SETFATTR_PROG -n user.attr3 -v val3 $SCRATCH_MNT/foobar # Make sure everything is durably persisted. sync # Now delete the second xattr and fsync the inode. $SETFATTR_PROG -x user.attr2 $SCRATCH_MNT/foobar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foobar _crash_and_mount # After the fsync log is replayed, the file should have only 2 xattrs, the ones # named user.attr1 and user.attr3. The btrfs fsync log replay bug left the file # with the 3 xattrs that we had before deleting the second one and fsyncing the # file. echo "xattr names and values after first fsync log replay:" $GETFATTR_PROG --absolute-names --dump $SCRATCH_MNT/foobar | _filter_scratch # Now write some data to our file, fsync it, remove the first xattr, add a new # hard link to our file and commit the fsync log by fsyncing some other new # file. This is to verify that after log replay our first xattr does not exist # anymore. echo "hello world!" >> $SCRATCH_MNT/foobar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foobar $SETFATTR_PROG -x user.attr1 $SCRATCH_MNT/foobar ln $SCRATCH_MNT/foobar $SCRATCH_MNT/foobar_link touch $SCRATCH_MNT/qwerty $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/qwerty _crash_and_mount # Now only the xattr with name user.attr3 should be set in our file. echo "xattr names and values after second fsync log replay:" $GETFATTR_PROG --absolute-names --dump $SCRATCH_MNT/foobar | _filter_scratch status=0 exit The expected golden output, which is produced with this patch applied or when testing against xfs or ext3/4, is: xattr names and values after first fsync log replay: # file: SCRATCH_MNT/foobar user.attr1="val1" user.attr3="val3" xattr names and values after second fsync log replay: # file: SCRATCH_MNT/foobar user.attr3="val3" Without this patch applied, the output is: xattr names and values after first fsync log replay: # file: SCRATCH_MNT/foobar user.attr1="val1" user.attr2="val2" user.attr3="val3" xattr names and values after second fsync log replay: # file: SCRATCH_MNT/foobar user.attr1="val1" user.attr2="val2" user.attr3="val3" A patch with a test case for xfstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-02-23 22:53:35 +03:00
} else if (test_and_clear_bit(BTRFS_INODE_COPY_EVERYTHING,
&inode->runtime_flags) ||
inode_only == LOG_INODE_EXISTS) {
Btrfs: remove deleted xattrs on fsync log replay If we deleted xattrs from a file and fsynced the file, after a log replay the xattrs would remain associated to the file. This was an unexpected behaviour and differs from what other filesystems do, such as for example xfs and ext3/4. Fix this by, on fsync log replay, check if every xattr in the fs/subvol tree (that belongs to a logged inode) has a matching xattr in the log, and if it does not, delete it from the fs/subvol tree. This is a similar approach to what we do for dentries when we replay a directory from the fsync log. This issue is trivial to reproduce, and the following excerpt from my test for xfstests triggers the issue: _crash_and_mount() { # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey } rm -f $seqres.full _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create out test file and add 3 xattrs to it. touch $SCRATCH_MNT/foobar $SETFATTR_PROG -n user.attr1 -v val1 $SCRATCH_MNT/foobar $SETFATTR_PROG -n user.attr2 -v val2 $SCRATCH_MNT/foobar $SETFATTR_PROG -n user.attr3 -v val3 $SCRATCH_MNT/foobar # Make sure everything is durably persisted. sync # Now delete the second xattr and fsync the inode. $SETFATTR_PROG -x user.attr2 $SCRATCH_MNT/foobar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foobar _crash_and_mount # After the fsync log is replayed, the file should have only 2 xattrs, the ones # named user.attr1 and user.attr3. The btrfs fsync log replay bug left the file # with the 3 xattrs that we had before deleting the second one and fsyncing the # file. echo "xattr names and values after first fsync log replay:" $GETFATTR_PROG --absolute-names --dump $SCRATCH_MNT/foobar | _filter_scratch # Now write some data to our file, fsync it, remove the first xattr, add a new # hard link to our file and commit the fsync log by fsyncing some other new # file. This is to verify that after log replay our first xattr does not exist # anymore. echo "hello world!" >> $SCRATCH_MNT/foobar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foobar $SETFATTR_PROG -x user.attr1 $SCRATCH_MNT/foobar ln $SCRATCH_MNT/foobar $SCRATCH_MNT/foobar_link touch $SCRATCH_MNT/qwerty $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/qwerty _crash_and_mount # Now only the xattr with name user.attr3 should be set in our file. echo "xattr names and values after second fsync log replay:" $GETFATTR_PROG --absolute-names --dump $SCRATCH_MNT/foobar | _filter_scratch status=0 exit The expected golden output, which is produced with this patch applied or when testing against xfs or ext3/4, is: xattr names and values after first fsync log replay: # file: SCRATCH_MNT/foobar user.attr1="val1" user.attr3="val3" xattr names and values after second fsync log replay: # file: SCRATCH_MNT/foobar user.attr3="val3" Without this patch applied, the output is: xattr names and values after first fsync log replay: # file: SCRATCH_MNT/foobar user.attr1="val1" user.attr2="val2" user.attr3="val3" xattr names and values after second fsync log replay: # file: SCRATCH_MNT/foobar user.attr1="val1" user.attr2="val2" user.attr3="val3" A patch with a test case for xfstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-02-23 22:53:35 +03:00
if (inode_only == LOG_INODE_ALL)
fast_search = true;
Btrfs: remove deleted xattrs on fsync log replay If we deleted xattrs from a file and fsynced the file, after a log replay the xattrs would remain associated to the file. This was an unexpected behaviour and differs from what other filesystems do, such as for example xfs and ext3/4. Fix this by, on fsync log replay, check if every xattr in the fs/subvol tree (that belongs to a logged inode) has a matching xattr in the log, and if it does not, delete it from the fs/subvol tree. This is a similar approach to what we do for dentries when we replay a directory from the fsync log. This issue is trivial to reproduce, and the following excerpt from my test for xfstests triggers the issue: _crash_and_mount() { # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey } rm -f $seqres.full _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create out test file and add 3 xattrs to it. touch $SCRATCH_MNT/foobar $SETFATTR_PROG -n user.attr1 -v val1 $SCRATCH_MNT/foobar $SETFATTR_PROG -n user.attr2 -v val2 $SCRATCH_MNT/foobar $SETFATTR_PROG -n user.attr3 -v val3 $SCRATCH_MNT/foobar # Make sure everything is durably persisted. sync # Now delete the second xattr and fsync the inode. $SETFATTR_PROG -x user.attr2 $SCRATCH_MNT/foobar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foobar _crash_and_mount # After the fsync log is replayed, the file should have only 2 xattrs, the ones # named user.attr1 and user.attr3. The btrfs fsync log replay bug left the file # with the 3 xattrs that we had before deleting the second one and fsyncing the # file. echo "xattr names and values after first fsync log replay:" $GETFATTR_PROG --absolute-names --dump $SCRATCH_MNT/foobar | _filter_scratch # Now write some data to our file, fsync it, remove the first xattr, add a new # hard link to our file and commit the fsync log by fsyncing some other new # file. This is to verify that after log replay our first xattr does not exist # anymore. echo "hello world!" >> $SCRATCH_MNT/foobar $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foobar $SETFATTR_PROG -x user.attr1 $SCRATCH_MNT/foobar ln $SCRATCH_MNT/foobar $SCRATCH_MNT/foobar_link touch $SCRATCH_MNT/qwerty $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/qwerty _crash_and_mount # Now only the xattr with name user.attr3 should be set in our file. echo "xattr names and values after second fsync log replay:" $GETFATTR_PROG --absolute-names --dump $SCRATCH_MNT/foobar | _filter_scratch status=0 exit The expected golden output, which is produced with this patch applied or when testing against xfs or ext3/4, is: xattr names and values after first fsync log replay: # file: SCRATCH_MNT/foobar user.attr1="val1" user.attr3="val3" xattr names and values after second fsync log replay: # file: SCRATCH_MNT/foobar user.attr3="val3" Without this patch applied, the output is: xattr names and values after first fsync log replay: # file: SCRATCH_MNT/foobar user.attr1="val1" user.attr2="val2" user.attr3="val3" xattr names and values after second fsync log replay: # file: SCRATCH_MNT/foobar user.attr1="val1" user.attr2="val2" user.attr3="val3" A patch with a test case for xfstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-02-23 22:53:35 +03:00
max_key.type = BTRFS_XATTR_ITEM_KEY;
btrfs: avoid inode logging during rename and link when possible During a rename or link operation, we need to determine if an inode was previously logged or not, and if it was, do some update to the logged inode. We used to rely exclusively on the logged_trans field of struct btrfs_inode to determine that, but that was not reliable because the value of that field is not persisted in the inode item, so it's lost when an inode is evicted and loaded back again. That led to several issues in the past, such as not persisting deletions (such as the case fixed by commit 803f0f64d17769 ("Btrfs: fix fsync not persisting dentry deletions due to inode evictions")), or resulting in losing a file after an inode eviction followed by a rename (commit ecc64fab7d49c6 ("btrfs: fix lost inode on log replay after mix of fsync, rename and inode eviction")), besides other issues. So the inode_logged() helper was introduced and used to determine if an inode was possibly logged before in the current transaction, with the caveat that it could return false positives, in the sense that even if an inode was not logged before in the current transaction, it could still return true, but never to return false in case the inode was logged. >From a functional point of view that is fine, but from a performance perspective it can introduce significant latencies to rename and link operations, as they will end up doing inode logging even when it is not necessary. Recently on a 5.15 kernel, an openSUSE Tumbleweed user reported package installations and upgrades, with the zypper tool, were often taking a long time to complete. With strace it could be observed that zypper was spending about 99% of its time on rename operations, and then with further analysis we checked that directory logging was happening too frequently. Taking into account that installation/upgrade of some of the packages needed a few thousand file renames, the slowdown was very noticeable for the user. The issue was caused indirectly due to an excessive number of inode evictions on a 5.15 kernel, about 100x more compared to a 5.13, 5.14 or a 5.16-rc8 kernel. While triggering the inode evictions if something outside btrfs' control, btrfs could still behave better by eliminating the false positives from the inode_logged() helper. So change inode_logged() to actually eliminate such false positives caused by inode eviction and when an inode was never logged since the filesystem was mounted, as both cases relate to when the logged_trans field of struct btrfs_inode has a value of zero. When it can not determine if the inode was logged based only on the logged_trans value, lookup for the existence of the inode item in the log tree - if it's there then we known the inode was logged, if it's not there then it can not have been logged in the current transaction. Once we determine if the inode was logged, update the logged_trans value to avoid future calls to have to search in the log tree again. Alternatively, we could start storing logged_trans in the on disk inode item structure (struct btrfs_inode_item) in the unused space it still has, but that would be a bit odd because: 1) We only care about logged_trans since the filesystem was mounted, we don't care about its value from a previous mount. Having it persisted in the inode item structure would not make the best use of the precious unused space; 2) In order to get logged_trans persisted before inode eviction, we would have to update the delayed inode when we finish logging the inode and update its logged_trans in struct btrfs_inode, which makes it a bit cumbersome since we need to check if the delayed inode exists, if not create it and populate it and deal with any errors (-ENOMEM mostly). This change is part of a patchset comprised of the following patches: 1/5 btrfs: add helper to delete a dir entry from a log tree 2/5 btrfs: pass the dentry to btrfs_log_new_name() instead of the inode 3/5 btrfs: avoid logging all directory changes during renames 4/5 btrfs: stop doing unnecessary log updates during a rename 5/5 btrfs: avoid inode logging during rename and link when possible The following test script mimics part of what the zypper tool does during package installations/upgrades. It does not triggers inode evictions, but it's similar because it triggers false positives from the inode_logged() helper, because the inodes have a logged_trans of 0, there's a log tree due to a fsync of an unrelated file and the directory inode has its last_trans field set to the current transaction: $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=10000 mkfs.btrfs -f $DEV mount $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync # Now do some change to an unrelated file and fsync it. # This is just to create a log tree to make sure that inode_logged() # does not return false when called against "testdir". xfs_io -f -c "pwrite 0 4K" -c "fsync" $MNT/foo # Do some change to testdir. This is to make sure inode_logged() # will return true when called against "testdir", because its # logged_trans is 0, it was changed in the current transaction # and there's a log tree. echo -n > $MNT/testdir/file_$((NUM_FILES + 1)) echo "Renaming $NUM_FILES files..." start=$(date +%s%N) for ((i = 1; i <= $NUM_FILES; i++)); do mv $MNT/testdir/file_$i $MNT/testdir/file_$i-RPMDELETE done end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "Renames took $dur milliseconds" umount $MNT Testing this change on a box using a non-debug kernel (Debian's default kernel config) gave the following results: NUM_FILES=10000, before patchset: 27837 ms NUM_FILES=10000, after patches 1/5 to 4/5 applied: 9236 ms (-66.8%) NUM_FILES=10000, after whole patchset applied: 8902 ms (-68.0%) NUM_FILES=5000, before patchset: 9127 ms NUM_FILES=5000, after patches 1/5 to 4/5 applied: 4640 ms (-49.2%) NUM_FILES=5000, after whole patchset applied: 4441 ms (-51.3%) NUM_FILES=2000, before patchset: 2528 ms NUM_FILES=2000, after patches 1/5 to 4/5 applied: 1983 ms (-21.6%) NUM_FILES=2000, after whole patchset applied: 1747 ms (-30.9%) NUM_FILES=1000, before patchset: 1085 ms NUM_FILES=1000, after patches 1/5 to 4/5 applied: 893 ms (-17.7%) NUM_FILES=1000, after whole patchset applied: 867 ms (-20.1%) Running dbench on the same physical machine with the following script: $ cat run-dbench.sh #!/bin/bash NUM_JOBS=$(nproc --all) DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 MOUNT_OPTIONS="-o ssd" MKFS_OPTIONS="-O no-holes -R free-space-tree" echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -D $MNT -t 120 $NUM_JOBS umount $MNT Before patchset: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 3761352 0.032 143.843 Close 2762770 0.002 2.273 Rename 159304 0.291 67.037 Unlink 759784 0.207 143.998 Deltree 72 4.028 15.977 Mkdir 36 0.003 0.006 Qpathinfo 3409780 0.013 9.678 Qfileinfo 596772 0.001 0.878 Qfsinfo 625189 0.003 1.245 Sfileinfo 306443 0.006 1.840 Find 1318106 0.063 19.798 WriteX 1871137 0.021 8.532 ReadX 5897325 0.003 3.567 LockX 12252 0.003 0.258 UnlockX 12252 0.002 0.100 Flush 263666 3.327 155.632 Throughput 980.047 MB/sec 12 clients 12 procs max_latency=155.636 ms After whole patchset applied: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 4195584 0.033 107.742 Close 3081932 0.002 1.935 Rename 177641 0.218 14.905 Unlink 847333 0.166 107.822 Deltree 118 5.315 15.247 Mkdir 59 0.004 0.048 Qpathinfo 3802612 0.014 10.302 Qfileinfo 666748 0.001 1.034 Qfsinfo 697329 0.003 0.944 Sfileinfo 341712 0.006 2.099 Find 1470365 0.065 9.359 WriteX 2093921 0.021 8.087 ReadX 6576234 0.003 3.407 LockX 13660 0.003 0.308 UnlockX 13660 0.002 0.114 Flush 294090 2.906 115.539 Throughput 1093.11 MB/sec 12 clients 12 procs max_latency=115.544 ms +11.5% throughput -25.8% max latency rename max latency -77.8% Link: https://bugzilla.opensuse.org/show_bug.cgi?id=1193549 Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-01-20 14:00:10 +03:00
if (ctx->logged_before)
ret = drop_inode_items(trans, log, path, inode,
max_key.type);
} else {
if (inode_only == LOG_INODE_ALL)
fast_search = true;
btrfs: avoid unnecessary lock and leaf splits when updating inode in the log During a fast fsync, if we have already fsynced the file before and in the current transaction, we can make the inode item update more efficient and avoid acquiring a write lock on the leaf's parent. To update the inode item we are always using btrfs_insert_empty_item() to get a path pointing to the inode item, which calls btrfs_search_slot() with an "ins_len" argument of 'sizeof(struct btrfs_inode_item) + sizeof(struct btrfs_item)', and that always results in the search taking a write lock on the level 1 node that is the parent of the leaf that contains the inode item. This adds unnecessary lock contention on log trees when we have multiple fsyncs in parallel against inodes in the same subvolume, which has a very significant impact due to the fact that log trees are short lived and their height very rarely goes beyond level 2. Also, by using btrfs_insert_empty_item() when we need to update the inode item, we also end up splitting the leaf of the existing inode item when the leaf has an amount of free space smaller than the size of an inode item. Improve this by using btrfs_seach_slot(), with a 0 "ins_len" argument, when we know the inode item already exists in the log. This avoids these two inefficiencies. The following script, using fio, was used to perform the tests: $ cat fio-test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 MOUNT_OPTIONS="-o ssd" MKFS_OPTIONS="-d single -m single" if [ $# -ne 4 ]; then echo "Use $0 NUM_JOBS FILE_SIZE FSYNC_FREQ BLOCK_SIZE" exit 1 fi NUM_JOBS=$1 FILE_SIZE=$2 FSYNC_FREQ=$3 BLOCK_SIZE=$4 cat <<EOF > /tmp/fio-job.ini [writers] rw=randwrite fsync=$FSYNC_FREQ fallocate=none group_reporting=1 direct=0 bs=$BLOCK_SIZE ioengine=sync size=$FILE_SIZE directory=$MNT numjobs=$NUM_JOBS EOF echo "performance" | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor echo echo "Using config:" echo cat /tmp/fio-job.ini echo echo "mount options: $MOUNT_OPTIONS" echo umount $MNT &> /dev/null mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT fio /tmp/fio-job.ini umount $MNT The tests were done on a physical machine, with 12 cores, 64G of RAM, using a NVMEe device and using a non-debug kernel config (the default one from Debian). The summary line from fio is provided below for each test run. With 8 jobs, file size 256M, fsync frequency of 4 and a block size of 4K: Before: WRITE: bw=28.3MiB/s (29.7MB/s), 28.3MiB/s-28.3MiB/s (29.7MB/s-29.7MB/s), io=2048MiB (2147MB), run=72297-72297msec After: WRITE: bw=28.7MiB/s (30.1MB/s), 28.7MiB/s-28.7MiB/s (30.1MB/s-30.1MB/s), io=2048MiB (2147MB), run=71411-71411msec +1.4% throughput, -1.2% runtime With 16 jobs, file size 256M, fsync frequency of 4 and a block size of 4K: Before: WRITE: bw=40.0MiB/s (42.0MB/s), 40.0MiB/s-40.0MiB/s (42.0MB/s-42.0MB/s), io=4096MiB (4295MB), run=99980-99980msec After: WRITE: bw=40.9MiB/s (42.9MB/s), 40.9MiB/s-40.9MiB/s (42.9MB/s-42.9MB/s), io=4096MiB (4295MB), run=97933-97933msec +2.2% throughput, -2.1% runtime The changes are small but it's possible to be better on faster hardware as in the test machine used disk utilization was pretty much 100% during the whole time the tests were running (observed with 'iostat -xz 1'). The tests also included the previous patch with the subject of: "btrfs: avoid unnecessary log mutex contention when syncing log". So they compared a branch without that patch and without this patch versus a branch with these two patches applied. Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-07-20 18:03:43 +03:00
inode_item_dropped = false;
goto log_extents;
Btrfs: turbo charge fsync At least for the vm workload. Currently on fsync we will 1) Truncate all items in the log tree for the given inode if they exist and 2) Copy all items for a given inode into the log The problem with this is that for things like VMs you can have lots of extents from the fragmented writing behavior, and worst yet you may have only modified a few extents, not the entire thing. This patch fixes this problem by tracking which transid modified our extent, and then when we do the tree logging we find all of the extents we've modified in our current transaction, sort them and commit them. We also only truncate up to the xattrs of the inode and copy that stuff in normally, and then just drop any extents in the range we have that exist in the log already. Here are some numbers of a 50 meg fio job that does random writes and fsync()s after every write Original Patched SATA drive 82KB/s 140KB/s Fusion drive 431KB/s 2532KB/s So around 2-6 times faster depending on your hardware. There are a few corner cases, for example if you truncate at all we have to do it the old way since there is no way to be sure what is in the log is ok. This probably could be done smarter, but if you write-fsync-truncate-write-fsync you deserve what you get. All this work is in RAM of course so if your inode gets evicted from cache and you read it in and fsync it we'll do it the slow way if we are still in the same transaction that we last modified the inode in. The biggest cool part of this is that it requires no changes to the recovery code, so if you fsync with this patch and crash and load an old kernel, it will run the recovery and be a-ok. I have tested this pretty thoroughly with an fsync tester and everything comes back fine, as well as xfstests. Thanks, Signed-off-by: Josef Bacik <jbacik@fusionio.com>
2012-08-17 21:14:17 +04:00
}
}
if (ret)
goto out_unlock;
ret = copy_inode_items_to_log(trans, inode, &min_key, &max_key,
path, dst_path, logged_isize,
btrfs: make full fsyncs always operate on the entire file again This is a revert of commit 0a8068a3dd4294 ("btrfs: make ranged full fsyncs more efficient"), with updated comment in btrfs_sync_file. Commit 0a8068a3dd4294 ("btrfs: make ranged full fsyncs more efficient") made full fsyncs operate on the given range only as it assumed it was safe when using the NO_HOLES feature, since the hole detection was simplified some time ago and no longer was a source for races with ordered extent completion of adjacent file ranges. However it's still not safe to have a full fsync only operate on the given range, because extent maps for new extents might not be present in memory due to inode eviction or extent cloning. Consider the following example: 1) We are currently at transaction N; 2) We write to the file range [0, 1MiB); 3) Writeback finishes for the whole range and ordered extents complete, while we are still at transaction N; 4) The inode is evicted; 5) We open the file for writing, causing the inode to be loaded to memory again, which sets the 'full sync' bit on its flags. At this point the inode's list of modified extent maps is empty (figuring out which extents were created in the current transaction and were not yet logged by an fsync is expensive, that's why we set the 'full sync' bit when loading an inode); 6) We write to the file range [512KiB, 768KiB); 7) We do a ranged fsync (such as msync()) for file range [512KiB, 768KiB). This correctly flushes this range and logs its extent into the log tree. When the writeback started an extent map for range [512KiB, 768KiB) was added to the inode's list of modified extents, and when the fsync() finishes logging it removes that extent map from the list of modified extent maps. This fsync also clears the 'full sync' bit; 8) We do a regular fsync() (full ranged). This fsync() ends up doing nothing because the inode's list of modified extents is empty and no other changes happened since the previous ranged fsync(), so it just returns success (0) and we end up never logging extents for the file ranges [0, 512KiB) and [768KiB, 1MiB). Another scenario where this can happen is if we replace steps 2 to 4 with cloning from another file into our test file, as that sets the 'full sync' bit in our inode's flags and does not populate its list of modified extent maps. This was causing test case generic/457 to fail sporadically when using the NO_HOLES feature, as it exercised this later case where the inode has the 'full sync' bit set and has no extent maps in memory to represent the new extents due to extent cloning. Fix this by reverting commit 0a8068a3dd4294 ("btrfs: make ranged full fsyncs more efficient") since there is no easy way to work around it. Fixes: 0a8068a3dd4294 ("btrfs: make ranged full fsyncs more efficient") Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-04-07 13:37:44 +03:00
recursive_logging, inode_only, ctx,
&need_log_inode_item);
if (ret)
goto out_unlock;
Btrfs: turbo charge fsync At least for the vm workload. Currently on fsync we will 1) Truncate all items in the log tree for the given inode if they exist and 2) Copy all items for a given inode into the log The problem with this is that for things like VMs you can have lots of extents from the fragmented writing behavior, and worst yet you may have only modified a few extents, not the entire thing. This patch fixes this problem by tracking which transid modified our extent, and then when we do the tree logging we find all of the extents we've modified in our current transaction, sort them and commit them. We also only truncate up to the xattrs of the inode and copy that stuff in normally, and then just drop any extents in the range we have that exist in the log already. Here are some numbers of a 50 meg fio job that does random writes and fsync()s after every write Original Patched SATA drive 82KB/s 140KB/s Fusion drive 431KB/s 2532KB/s So around 2-6 times faster depending on your hardware. There are a few corner cases, for example if you truncate at all we have to do it the old way since there is no way to be sure what is in the log is ok. This probably could be done smarter, but if you write-fsync-truncate-write-fsync you deserve what you get. All this work is in RAM of course so if your inode gets evicted from cache and you read it in and fsync it we'll do it the slow way if we are still in the same transaction that we last modified the inode in. The biggest cool part of this is that it requires no changes to the recovery code, so if you fsync with this patch and crash and load an old kernel, it will run the recovery and be a-ok. I have tested this pretty thoroughly with an fsync tester and everything comes back fine, as well as xfstests. Thanks, Signed-off-by: Josef Bacik <jbacik@fusionio.com>
2012-08-17 21:14:17 +04:00
Btrfs: fix fsync xattr loss in the fast fsync path After commit 4f764e515361 ("Btrfs: remove deleted xattrs on fsync log replay"), we can end up in a situation where during log replay we end up deleting xattrs that were never deleted when their file was last fsynced. This happens in the fast fsync path (flag BTRFS_INODE_NEEDS_FULL_SYNC is not set in the inode) if the inode has the flag BTRFS_INODE_COPY_EVERYTHING set, the xattr was added in a past transaction and the leaf where the xattr is located was not updated (COWed or created) in the current transaction. In this scenario the xattr item never ends up in the log tree and therefore at log replay time, which makes the replay code delete the xattr from the fs/subvol tree as it thinks that xattr was deleted prior to the last fsync. Fix this by always logging all xattrs, which is the simplest and most reliable way to detect deleted xattrs and replay the deletes at log replay time. This issue is reproducible with the following test case for fstests: seq=`basename $0` seqres=$RESULT_DIR/$seq echo "QA output created by $seq" here=`pwd` tmp=/tmp/$$ status=1 # failure is the default! _cleanup() { _cleanup_flakey rm -f $tmp.* } trap "_cleanup; exit \$status" 0 1 2 3 15 # get standard environment, filters and checks . ./common/rc . ./common/filter . ./common/dmflakey . ./common/attr # real QA test starts here # We create a lot of xattrs for a single file. Only btrfs and xfs are currently # able to store such a large mount of xattrs per file, other filesystems such # as ext3/4 and f2fs for example, fail with ENOSPC even if we attempt to add # less than 1000 xattrs with very small values. _supported_fs btrfs xfs _supported_os Linux _need_to_be_root _require_scratch _require_dm_flakey _require_attrs _require_metadata_journaling $SCRATCH_DEV rm -f $seqres.full _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create the test file with some initial data and make sure everything is # durably persisted. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 32k" $SCRATCH_MNT/foo | _filter_xfs_io sync # Add many small xattrs to our file. # We create such a large amount because it's needed to trigger the issue found # in btrfs - we need to have an amount that causes the fs to have at least 3 # btree leafs with xattrs stored in them, and it must work on any leaf size # (maximum leaf/node size is 64Kb). num_xattrs=2000 for ((i = 1; i <= $num_xattrs; i++)); do name="user.attr_$(printf "%04d" $i)" $SETFATTR_PROG -n $name -v "val_$(printf "%04d" $i)" $SCRATCH_MNT/foo done # Sync the filesystem to force a commit of the current btrfs transaction, this # is a necessary condition to trigger the bug on btrfs. sync # Now update our file's data and fsync the file. # After a successful fsync, if the fsync log/journal is replayed we expect to # see all the xattrs we added before with the same values (and the updated file # data of course). Btrfs used to delete some of these xattrs when it replayed # its fsync log/journal. $XFS_IO_PROG -c "pwrite -S 0xbb 8K 16K" \ -c "fsync" \ $SCRATCH_MNT/foo | _filter_xfs_io # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again and mount. This makes the fs replay its fsync log. _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey echo "File content after crash and log replay:" od -t x1 $SCRATCH_MNT/foo echo "File xattrs after crash and log replay:" for ((i = 1; i <= $num_xattrs; i++)); do name="user.attr_$(printf "%04d" $i)" echo -n "$name=" $GETFATTR_PROG --absolute-names -n $name --only-values $SCRATCH_MNT/foo echo done status=0 exit The golden output expects all xattrs to be available, and with the correct values, after the fsync log is replayed. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-06-20 02:44:51 +03:00
btrfs_release_path(path);
btrfs_release_path(dst_path);
ret = btrfs_log_all_xattrs(trans, inode, path, dst_path);
if (ret)
Btrfs: fix fsync xattr loss in the fast fsync path After commit 4f764e515361 ("Btrfs: remove deleted xattrs on fsync log replay"), we can end up in a situation where during log replay we end up deleting xattrs that were never deleted when their file was last fsynced. This happens in the fast fsync path (flag BTRFS_INODE_NEEDS_FULL_SYNC is not set in the inode) if the inode has the flag BTRFS_INODE_COPY_EVERYTHING set, the xattr was added in a past transaction and the leaf where the xattr is located was not updated (COWed or created) in the current transaction. In this scenario the xattr item never ends up in the log tree and therefore at log replay time, which makes the replay code delete the xattr from the fs/subvol tree as it thinks that xattr was deleted prior to the last fsync. Fix this by always logging all xattrs, which is the simplest and most reliable way to detect deleted xattrs and replay the deletes at log replay time. This issue is reproducible with the following test case for fstests: seq=`basename $0` seqres=$RESULT_DIR/$seq echo "QA output created by $seq" here=`pwd` tmp=/tmp/$$ status=1 # failure is the default! _cleanup() { _cleanup_flakey rm -f $tmp.* } trap "_cleanup; exit \$status" 0 1 2 3 15 # get standard environment, filters and checks . ./common/rc . ./common/filter . ./common/dmflakey . ./common/attr # real QA test starts here # We create a lot of xattrs for a single file. Only btrfs and xfs are currently # able to store such a large mount of xattrs per file, other filesystems such # as ext3/4 and f2fs for example, fail with ENOSPC even if we attempt to add # less than 1000 xattrs with very small values. _supported_fs btrfs xfs _supported_os Linux _need_to_be_root _require_scratch _require_dm_flakey _require_attrs _require_metadata_journaling $SCRATCH_DEV rm -f $seqres.full _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create the test file with some initial data and make sure everything is # durably persisted. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 32k" $SCRATCH_MNT/foo | _filter_xfs_io sync # Add many small xattrs to our file. # We create such a large amount because it's needed to trigger the issue found # in btrfs - we need to have an amount that causes the fs to have at least 3 # btree leafs with xattrs stored in them, and it must work on any leaf size # (maximum leaf/node size is 64Kb). num_xattrs=2000 for ((i = 1; i <= $num_xattrs; i++)); do name="user.attr_$(printf "%04d" $i)" $SETFATTR_PROG -n $name -v "val_$(printf "%04d" $i)" $SCRATCH_MNT/foo done # Sync the filesystem to force a commit of the current btrfs transaction, this # is a necessary condition to trigger the bug on btrfs. sync # Now update our file's data and fsync the file. # After a successful fsync, if the fsync log/journal is replayed we expect to # see all the xattrs we added before with the same values (and the updated file # data of course). Btrfs used to delete some of these xattrs when it replayed # its fsync log/journal. $XFS_IO_PROG -c "pwrite -S 0xbb 8K 16K" \ -c "fsync" \ $SCRATCH_MNT/foo | _filter_xfs_io # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again and mount. This makes the fs replay its fsync log. _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey echo "File content after crash and log replay:" od -t x1 $SCRATCH_MNT/foo echo "File xattrs after crash and log replay:" for ((i = 1; i <= $num_xattrs; i++)); do name="user.attr_$(printf "%04d" $i)" echo -n "$name=" $GETFATTR_PROG --absolute-names -n $name --only-values $SCRATCH_MNT/foo echo done status=0 exit The golden output expects all xattrs to be available, and with the correct values, after the fsync log is replayed. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-06-20 02:44:51 +03:00
goto out_unlock;
xattrs_logged = true;
Btrfs: fix fsync after truncate when no_holes feature is enabled When we have the no_holes feature enabled, if a we truncate a file to a smaller size, truncate it again but to a size greater than or equals to its original size and fsync it, the log tree will not have any information about the hole covering the range [truncate_1_offset, new_file_size[. Which means if the fsync log is replayed, the file will remain with the state it had before both truncate operations. Without the no_holes feature this does not happen, since when the inode is logged (full sync flag is set) it will find in the fs/subvol tree a leaf with a generation matching the current transaction id that has an explicit extent item representing the hole. Fix this by adding an explicit extent item representing a hole between the last extent and the inode's i_size if we are doing a full sync. The issue is easy to reproduce with the following test case for fstests: . ./common/rc . ./common/filter . ./common/dmflakey _need_to_be_root _supported_fs generic _supported_os Linux _require_scratch _require_dm_flakey # This test was motivated by an issue found in btrfs when the btrfs # no-holes feature is enabled (introduced in kernel 3.14). So enable # the feature if the fs being tested is btrfs. if [ $FSTYP == "btrfs" ]; then _require_btrfs_fs_feature "no_holes" _require_btrfs_mkfs_feature "no-holes" MKFS_OPTIONS="$MKFS_OPTIONS -O no-holes" fi rm -f $seqres.full _scratch_mkfs >>$seqres.full 2>&1 _init_flakey _mount_flakey # Create our test files and make sure everything is durably persisted. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 64K" \ -c "pwrite -S 0xbb 64K 61K" \ $SCRATCH_MNT/foo | _filter_xfs_io $XFS_IO_PROG -f -c "pwrite -S 0xee 0 64K" \ -c "pwrite -S 0xff 64K 61K" \ $SCRATCH_MNT/bar | _filter_xfs_io sync # Now truncate our file foo to a smaller size (64Kb) and then truncate # it to the size it had before the shrinking truncate (125Kb). Then # fsync our file. If a power failure happens after the fsync, we expect # our file to have a size of 125Kb, with the first 64Kb of data having # the value 0xaa and the second 61Kb of data having the value 0x00. $XFS_IO_PROG -c "truncate 64K" \ -c "truncate 125K" \ -c "fsync" \ $SCRATCH_MNT/foo # Do something similar to our file bar, but the first truncation sets # the file size to 0 and the second truncation expands the size to the # double of what it was initially. $XFS_IO_PROG -c "truncate 0" \ -c "truncate 253K" \ -c "fsync" \ $SCRATCH_MNT/bar _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again, mount to trigger log replay and validate file # contents. _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # We expect foo to have a size of 125Kb, the first 64Kb of data all # having the value 0xaa and the remaining 61Kb to be a hole (all bytes # with value 0x00). echo "File foo content after log replay:" od -t x1 $SCRATCH_MNT/foo # We expect bar to have a size of 253Kb and no extents (any byte read # from bar has the value 0x00). echo "File bar content after log replay:" od -t x1 $SCRATCH_MNT/bar status=0 exit The expected file contents in the golden output are: File foo content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0200000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 * 0372000 File bar content after log replay: 0000000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 * 0772000 Without this fix, their contents are: File foo content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0200000 bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb * 0372000 File bar content after log replay: 0000000 ee ee ee ee ee ee ee ee ee ee ee ee ee ee ee ee * 0200000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0372000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 * 0772000 A test case submission for fstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Reviewed-by: Liu Bo <bo.li.liu@oracle.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-06-25 06:17:46 +03:00
if (max_key.type >= BTRFS_EXTENT_DATA_KEY && !fast_search) {
btrfs_release_path(path);
btrfs_release_path(dst_path);
ret = btrfs_log_holes(trans, inode, path);
if (ret)
Btrfs: fix fsync after truncate when no_holes feature is enabled When we have the no_holes feature enabled, if a we truncate a file to a smaller size, truncate it again but to a size greater than or equals to its original size and fsync it, the log tree will not have any information about the hole covering the range [truncate_1_offset, new_file_size[. Which means if the fsync log is replayed, the file will remain with the state it had before both truncate operations. Without the no_holes feature this does not happen, since when the inode is logged (full sync flag is set) it will find in the fs/subvol tree a leaf with a generation matching the current transaction id that has an explicit extent item representing the hole. Fix this by adding an explicit extent item representing a hole between the last extent and the inode's i_size if we are doing a full sync. The issue is easy to reproduce with the following test case for fstests: . ./common/rc . ./common/filter . ./common/dmflakey _need_to_be_root _supported_fs generic _supported_os Linux _require_scratch _require_dm_flakey # This test was motivated by an issue found in btrfs when the btrfs # no-holes feature is enabled (introduced in kernel 3.14). So enable # the feature if the fs being tested is btrfs. if [ $FSTYP == "btrfs" ]; then _require_btrfs_fs_feature "no_holes" _require_btrfs_mkfs_feature "no-holes" MKFS_OPTIONS="$MKFS_OPTIONS -O no-holes" fi rm -f $seqres.full _scratch_mkfs >>$seqres.full 2>&1 _init_flakey _mount_flakey # Create our test files and make sure everything is durably persisted. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 64K" \ -c "pwrite -S 0xbb 64K 61K" \ $SCRATCH_MNT/foo | _filter_xfs_io $XFS_IO_PROG -f -c "pwrite -S 0xee 0 64K" \ -c "pwrite -S 0xff 64K 61K" \ $SCRATCH_MNT/bar | _filter_xfs_io sync # Now truncate our file foo to a smaller size (64Kb) and then truncate # it to the size it had before the shrinking truncate (125Kb). Then # fsync our file. If a power failure happens after the fsync, we expect # our file to have a size of 125Kb, with the first 64Kb of data having # the value 0xaa and the second 61Kb of data having the value 0x00. $XFS_IO_PROG -c "truncate 64K" \ -c "truncate 125K" \ -c "fsync" \ $SCRATCH_MNT/foo # Do something similar to our file bar, but the first truncation sets # the file size to 0 and the second truncation expands the size to the # double of what it was initially. $XFS_IO_PROG -c "truncate 0" \ -c "truncate 253K" \ -c "fsync" \ $SCRATCH_MNT/bar _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again, mount to trigger log replay and validate file # contents. _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # We expect foo to have a size of 125Kb, the first 64Kb of data all # having the value 0xaa and the remaining 61Kb to be a hole (all bytes # with value 0x00). echo "File foo content after log replay:" od -t x1 $SCRATCH_MNT/foo # We expect bar to have a size of 253Kb and no extents (any byte read # from bar has the value 0x00). echo "File bar content after log replay:" od -t x1 $SCRATCH_MNT/bar status=0 exit The expected file contents in the golden output are: File foo content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0200000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 * 0372000 File bar content after log replay: 0000000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 * 0772000 Without this fix, their contents are: File foo content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0200000 bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb * 0372000 File bar content after log replay: 0000000 ee ee ee ee ee ee ee ee ee ee ee ee ee ee ee ee * 0200000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0372000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 * 0772000 A test case submission for fstests follows soon. Signed-off-by: Filipe Manana <fdmanana@suse.com> Reviewed-by: Liu Bo <bo.li.liu@oracle.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-06-25 06:17:46 +03:00
goto out_unlock;
}
log_extents:
btrfs_release_path(path);
btrfs_release_path(dst_path);
Btrfs: fix fsync data loss after append write If we do an append write to a file (which increases its inode's i_size) that does not have the flag BTRFS_INODE_NEEDS_FULL_SYNC set in its inode, and the previous transaction added a new hard link to the file, which sets the flag BTRFS_INODE_COPY_EVERYTHING in the file's inode, and then fsync the file, the inode's new i_size isn't logged. This has the consequence that after the fsync log is replayed, the file size remains what it was before the append write operation, which means users/applications will not be able to read the data that was successsfully fsync'ed before. This happens because neither the inode item nor the delayed inode get their i_size updated when the append write is made - doing so would require starting a transaction in the buffered write path, something that we do not do intentionally for performance reasons. Fix this by making sure that when the flag BTRFS_INODE_COPY_EVERYTHING is set the inode is logged with its current i_size (log the in-memory inode into the log tree). This issue is not a recent regression and is easy to reproduce with the following test case for fstests: seq=`basename $0` seqres=$RESULT_DIR/$seq echo "QA output created by $seq" here=`pwd` tmp=/tmp/$$ status=1 # failure is the default! _cleanup() { _cleanup_flakey rm -f $tmp.* } trap "_cleanup; exit \$status" 0 1 2 3 15 # get standard environment, filters and checks . ./common/rc . ./common/filter . ./common/dmflakey # real QA test starts here _supported_fs generic _supported_os Linux _need_to_be_root _require_scratch _require_dm_flakey _require_metadata_journaling $SCRATCH_DEV _crash_and_mount() { # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again and mount. This makes the fs replay its fsync log. _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey } rm -f $seqres.full _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create the test file with some initial data and then fsync it. # The fsync here is only needed to trigger the issue in btrfs, as it causes the # the flag BTRFS_INODE_NEEDS_FULL_SYNC to be removed from the btrfs inode. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 32k" \ -c "fsync" \ $SCRATCH_MNT/foo | _filter_xfs_io sync # Add a hard link to our file. # On btrfs this sets the flag BTRFS_INODE_COPY_EVERYTHING on the btrfs inode, # which is a necessary condition to trigger the issue. ln $SCRATCH_MNT/foo $SCRATCH_MNT/bar # Sync the filesystem to force a commit of the current btrfs transaction, this # is a necessary condition to trigger the bug on btrfs. sync # Now append more data to our file, increasing its size, and fsync the file. # In btrfs because the inode flag BTRFS_INODE_COPY_EVERYTHING was set and the # write path did not update the inode item in the btree nor the delayed inode # item (in memory struture) in the current transaction (created by the fsync # handler), the fsync did not record the inode's new i_size in the fsync # log/journal. This made the data unavailable after the fsync log/journal is # replayed. $XFS_IO_PROG -c "pwrite -S 0xbb 32K 32K" \ -c "fsync" \ $SCRATCH_MNT/foo | _filter_xfs_io echo "File content after fsync and before crash:" od -t x1 $SCRATCH_MNT/foo _crash_and_mount echo "File content after crash and log replay:" od -t x1 $SCRATCH_MNT/foo status=0 exit The expected file output before and after the crash/power failure expects the appended data to be available, which is: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0100000 bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb * 0200000 Cc: stable@vger.kernel.org Signed-off-by: Filipe Manana <fdmanana@suse.com> Reviewed-by: Liu Bo <bo.li.liu@oracle.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-06-17 14:49:23 +03:00
if (need_log_inode_item) {
ret = log_inode_item(trans, log, dst_path, inode, inode_item_dropped);
if (ret)
btrfs: avoid unnecessary logging of xattrs during fast fsyncs When logging an inode we always log all its xattrs, so that we are able to figure out which ones should be deleted during log replay. However this is unnecessary when we are doing a fast fsync and no xattrs were added, changed or deleted since the last time we logged the inode in the current transaction. So skip the logging of xattrs when the inode was previously logged in the current transaction and no xattrs were added, changed or deleted. If any changes to xattrs happened, than the inode has BTRFS_INODE_COPY_EVERYTHING set in its runtime flags and the xattrs get logged. This saves time on scanning for xattrs, allocating memory, COWing log tree extent buffers and adding more lock contention on the extent buffers when there are multiple tasks logging in parallel. The use of xattrs is common when using ACLs, some applications, or when using security modules like SELinux where every inode gets a security xattr added to it. The following test script, using fio, was used on a box with 12 cores, 64G of RAM, a NVMe device and the default non-debug kernel config from Debian. It uses 8 concurrent jobs each writing in blocks of 64K to its own 4G file, each file with a single xattr of 50 bytes (about the same size for an ACL or SELinux xattr), doing random buffered writes with an fsync after each write. $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/test MOUNT_OPTIONS="-o ssd" MKFS_OPTIONS="-d single -m single" NUM_JOBS=8 FILE_SIZE=4G cat <<EOF > /tmp/fio-job.ini [writers] rw=randwrite fsync=1 fallocate=none group_reporting=1 direct=0 bs=64K ioengine=sync size=$FILE_SIZE directory=$MNT numjobs=$NUM_JOBS EOF echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f $MKFS_OPTIONS $DEV > /dev/null mount $MOUNT_OPTIONS $DEV $MNT echo "Creating files before fio runs, each with 1 xattr of 50 bytes" for ((i = 0; i < $NUM_JOBS; i++)); do path="$MNT/writers.$i.0" truncate -s $FILE_SIZE $path setfattr -n user.xa1 -v $(printf '%0.sX' $(seq 50)) $path done fio /tmp/fio-job.ini umount $MNT fio output before this change: WRITE: bw=120MiB/s (126MB/s), 120MiB/s-120MiB/s (126MB/s-126MB/s), io=32.0GiB (34.4GB), run=272145-272145msec fio output after this change: WRITE: bw=142MiB/s (149MB/s), 142MiB/s-142MiB/s (149MB/s-149MB/s), io=32.0GiB (34.4GB), run=230408-230408msec +16.8% throughput, -16.6% runtime Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-05-28 13:37:32 +03:00
goto out_unlock;
/*
* If we are doing a fast fsync and the inode was logged before
* in this transaction, we don't need to log the xattrs because
* they were logged before. If xattrs were added, changed or
* deleted since the last time we logged the inode, then we have
* already logged them because the inode had the runtime flag
* BTRFS_INODE_COPY_EVERYTHING set.
*/
if (!xattrs_logged && inode->logged_trans < trans->transid) {
ret = btrfs_log_all_xattrs(trans, inode, path, dst_path);
if (ret)
btrfs: avoid unnecessary logging of xattrs during fast fsyncs When logging an inode we always log all its xattrs, so that we are able to figure out which ones should be deleted during log replay. However this is unnecessary when we are doing a fast fsync and no xattrs were added, changed or deleted since the last time we logged the inode in the current transaction. So skip the logging of xattrs when the inode was previously logged in the current transaction and no xattrs were added, changed or deleted. If any changes to xattrs happened, than the inode has BTRFS_INODE_COPY_EVERYTHING set in its runtime flags and the xattrs get logged. This saves time on scanning for xattrs, allocating memory, COWing log tree extent buffers and adding more lock contention on the extent buffers when there are multiple tasks logging in parallel. The use of xattrs is common when using ACLs, some applications, or when using security modules like SELinux where every inode gets a security xattr added to it. The following test script, using fio, was used on a box with 12 cores, 64G of RAM, a NVMe device and the default non-debug kernel config from Debian. It uses 8 concurrent jobs each writing in blocks of 64K to its own 4G file, each file with a single xattr of 50 bytes (about the same size for an ACL or SELinux xattr), doing random buffered writes with an fsync after each write. $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/test MOUNT_OPTIONS="-o ssd" MKFS_OPTIONS="-d single -m single" NUM_JOBS=8 FILE_SIZE=4G cat <<EOF > /tmp/fio-job.ini [writers] rw=randwrite fsync=1 fallocate=none group_reporting=1 direct=0 bs=64K ioengine=sync size=$FILE_SIZE directory=$MNT numjobs=$NUM_JOBS EOF echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f $MKFS_OPTIONS $DEV > /dev/null mount $MOUNT_OPTIONS $DEV $MNT echo "Creating files before fio runs, each with 1 xattr of 50 bytes" for ((i = 0; i < $NUM_JOBS; i++)); do path="$MNT/writers.$i.0" truncate -s $FILE_SIZE $path setfattr -n user.xa1 -v $(printf '%0.sX' $(seq 50)) $path done fio /tmp/fio-job.ini umount $MNT fio output before this change: WRITE: bw=120MiB/s (126MB/s), 120MiB/s-120MiB/s (126MB/s-126MB/s), io=32.0GiB (34.4GB), run=272145-272145msec fio output after this change: WRITE: bw=142MiB/s (149MB/s), 142MiB/s-142MiB/s (149MB/s-149MB/s), io=32.0GiB (34.4GB), run=230408-230408msec +16.8% throughput, -16.6% runtime Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-05-28 13:37:32 +03:00
goto out_unlock;
btrfs_release_path(path);
}
Btrfs: fix fsync data loss after append write If we do an append write to a file (which increases its inode's i_size) that does not have the flag BTRFS_INODE_NEEDS_FULL_SYNC set in its inode, and the previous transaction added a new hard link to the file, which sets the flag BTRFS_INODE_COPY_EVERYTHING in the file's inode, and then fsync the file, the inode's new i_size isn't logged. This has the consequence that after the fsync log is replayed, the file size remains what it was before the append write operation, which means users/applications will not be able to read the data that was successsfully fsync'ed before. This happens because neither the inode item nor the delayed inode get their i_size updated when the append write is made - doing so would require starting a transaction in the buffered write path, something that we do not do intentionally for performance reasons. Fix this by making sure that when the flag BTRFS_INODE_COPY_EVERYTHING is set the inode is logged with its current i_size (log the in-memory inode into the log tree). This issue is not a recent regression and is easy to reproduce with the following test case for fstests: seq=`basename $0` seqres=$RESULT_DIR/$seq echo "QA output created by $seq" here=`pwd` tmp=/tmp/$$ status=1 # failure is the default! _cleanup() { _cleanup_flakey rm -f $tmp.* } trap "_cleanup; exit \$status" 0 1 2 3 15 # get standard environment, filters and checks . ./common/rc . ./common/filter . ./common/dmflakey # real QA test starts here _supported_fs generic _supported_os Linux _need_to_be_root _require_scratch _require_dm_flakey _require_metadata_journaling $SCRATCH_DEV _crash_and_mount() { # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again and mount. This makes the fs replay its fsync log. _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey } rm -f $seqres.full _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create the test file with some initial data and then fsync it. # The fsync here is only needed to trigger the issue in btrfs, as it causes the # the flag BTRFS_INODE_NEEDS_FULL_SYNC to be removed from the btrfs inode. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 32k" \ -c "fsync" \ $SCRATCH_MNT/foo | _filter_xfs_io sync # Add a hard link to our file. # On btrfs this sets the flag BTRFS_INODE_COPY_EVERYTHING on the btrfs inode, # which is a necessary condition to trigger the issue. ln $SCRATCH_MNT/foo $SCRATCH_MNT/bar # Sync the filesystem to force a commit of the current btrfs transaction, this # is a necessary condition to trigger the bug on btrfs. sync # Now append more data to our file, increasing its size, and fsync the file. # In btrfs because the inode flag BTRFS_INODE_COPY_EVERYTHING was set and the # write path did not update the inode item in the btree nor the delayed inode # item (in memory struture) in the current transaction (created by the fsync # handler), the fsync did not record the inode's new i_size in the fsync # log/journal. This made the data unavailable after the fsync log/journal is # replayed. $XFS_IO_PROG -c "pwrite -S 0xbb 32K 32K" \ -c "fsync" \ $SCRATCH_MNT/foo | _filter_xfs_io echo "File content after fsync and before crash:" od -t x1 $SCRATCH_MNT/foo _crash_and_mount echo "File content after crash and log replay:" od -t x1 $SCRATCH_MNT/foo status=0 exit The expected file output before and after the crash/power failure expects the appended data to be available, which is: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0100000 bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb * 0200000 Cc: stable@vger.kernel.org Signed-off-by: Filipe Manana <fdmanana@suse.com> Reviewed-by: Liu Bo <bo.li.liu@oracle.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-06-17 14:49:23 +03:00
}
Btrfs: turbo charge fsync At least for the vm workload. Currently on fsync we will 1) Truncate all items in the log tree for the given inode if they exist and 2) Copy all items for a given inode into the log The problem with this is that for things like VMs you can have lots of extents from the fragmented writing behavior, and worst yet you may have only modified a few extents, not the entire thing. This patch fixes this problem by tracking which transid modified our extent, and then when we do the tree logging we find all of the extents we've modified in our current transaction, sort them and commit them. We also only truncate up to the xattrs of the inode and copy that stuff in normally, and then just drop any extents in the range we have that exist in the log already. Here are some numbers of a 50 meg fio job that does random writes and fsync()s after every write Original Patched SATA drive 82KB/s 140KB/s Fusion drive 431KB/s 2532KB/s So around 2-6 times faster depending on your hardware. There are a few corner cases, for example if you truncate at all we have to do it the old way since there is no way to be sure what is in the log is ok. This probably could be done smarter, but if you write-fsync-truncate-write-fsync you deserve what you get. All this work is in RAM of course so if your inode gets evicted from cache and you read it in and fsync it we'll do it the slow way if we are still in the same transaction that we last modified the inode in. The biggest cool part of this is that it requires no changes to the recovery code, so if you fsync with this patch and crash and load an old kernel, it will run the recovery and be a-ok. I have tested this pretty thoroughly with an fsync tester and everything comes back fine, as well as xfstests. Thanks, Signed-off-by: Josef Bacik <jbacik@fusionio.com>
2012-08-17 21:14:17 +04:00
if (fast_search) {
ret = btrfs_log_changed_extents(trans, inode, dst_path, ctx);
if (ret)
Btrfs: turbo charge fsync At least for the vm workload. Currently on fsync we will 1) Truncate all items in the log tree for the given inode if they exist and 2) Copy all items for a given inode into the log The problem with this is that for things like VMs you can have lots of extents from the fragmented writing behavior, and worst yet you may have only modified a few extents, not the entire thing. This patch fixes this problem by tracking which transid modified our extent, and then when we do the tree logging we find all of the extents we've modified in our current transaction, sort them and commit them. We also only truncate up to the xattrs of the inode and copy that stuff in normally, and then just drop any extents in the range we have that exist in the log already. Here are some numbers of a 50 meg fio job that does random writes and fsync()s after every write Original Patched SATA drive 82KB/s 140KB/s Fusion drive 431KB/s 2532KB/s So around 2-6 times faster depending on your hardware. There are a few corner cases, for example if you truncate at all we have to do it the old way since there is no way to be sure what is in the log is ok. This probably could be done smarter, but if you write-fsync-truncate-write-fsync you deserve what you get. All this work is in RAM of course so if your inode gets evicted from cache and you read it in and fsync it we'll do it the slow way if we are still in the same transaction that we last modified the inode in. The biggest cool part of this is that it requires no changes to the recovery code, so if you fsync with this patch and crash and load an old kernel, it will run the recovery and be a-ok. I have tested this pretty thoroughly with an fsync tester and everything comes back fine, as well as xfstests. Thanks, Signed-off-by: Josef Bacik <jbacik@fusionio.com>
2012-08-17 21:14:17 +04:00
goto out_unlock;
} else if (inode_only == LOG_INODE_ALL) {
struct extent_map *em, *n;
write_lock(&em_tree->lock);
btrfs: make fast fsyncs wait only for writeback Currently regardless of a full or a fast fsync we always wait for ordered extents to complete, and then start logging the inode after that. However for fast fsyncs we can just wait for the writeback to complete, we don't need to wait for the ordered extents to complete since we use the list of modified extents maps to figure out which extents we must log and we can get their checksums directly from the ordered extents that are still in flight, otherwise look them up from the checksums tree. Until commit b5e6c3e170b770 ("btrfs: always wait on ordered extents at fsync time"), for fast fsyncs, we used to start logging without even waiting for the writeback to complete first, we would wait for it to complete after logging, while holding a transaction open, which lead to performance issues when using cgroups and probably for other cases too, as wait for IO while holding a transaction handle should be avoided as much as possible. After that, for fast fsyncs, we started to wait for ordered extents to complete before starting to log, which adds some latency to fsyncs and we even got at least one report about a performance drop which bisected to that particular change: https://lore.kernel.org/linux-btrfs/20181109215148.GF23260@techsingularity.net/ This change makes fast fsyncs only wait for writeback to finish before starting to log the inode, instead of waiting for both the writeback to finish and for the ordered extents to complete. This brings back part of the logic we had that extracts checksums from in flight ordered extents, which are not yet in the checksums tree, and making sure transaction commits wait for the completion of ordered extents previously logged (by far most of the time they have already completed by the time a transaction commit starts, resulting in no wait at all), to avoid any data loss if an ordered extent completes after the transaction used to log an inode is committed, followed by a power failure. When there are no other tasks accessing the checksums and the subvolume btrees, the ordered extent completion is pretty fast, typically taking 100 to 200 microseconds only in my observations. However when there are other tasks accessing these btrees, ordered extent completion can take a lot more time due to lock contention on nodes and leaves of these btrees. I've seen cases over 2 milliseconds, which starts to be significant. In particular when we do have concurrent fsyncs against different files there is a lot of contention on the checksums btree, since we have many tasks writing the checksums into the btree and other tasks that already started the logging phase are doing lookups for checksums in the btree. This change also turns all ranged fsyncs into full ranged fsyncs, which is something we already did when not using the NO_HOLES features or when doing a full fsync. This is to guarantee we never miss checksums due to writeback having been triggered only for a part of an extent, and we end up logging the full extent but only checksums for the written range, which results in missing checksums after log replay. Allowing ranged fsyncs to operate again only in the original range, when using the NO_HOLES feature and doing a fast fsync is doable but requires some non trivial changes to the writeback path, which can always be worked on later if needed, but I don't think they are a very common use case. Several tests were performed using fio for different numbers of concurrent jobs, each writing and fsyncing its own file, for both sequential and random file writes. The tests were run on bare metal, no virtualization, on a box with 12 cores (Intel i7-8700), 64Gb of RAM and a NVMe device, with a kernel configuration that is the default of typical distributions (debian in this case), without debug options enabled (kasan, kmemleak, slub debug, debug of page allocations, lock debugging, etc). The following script that calls fio was used: $ cat test-fsync.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/btrfs MOUNT_OPTIONS="-o ssd -o space_cache=v2" MKFS_OPTIONS="-d single -m single" if [ $# -ne 5 ]; then echo "Use $0 NUM_JOBS FILE_SIZE FSYNC_FREQ BLOCK_SIZE [write|randwrite]" exit 1 fi NUM_JOBS=$1 FILE_SIZE=$2 FSYNC_FREQ=$3 BLOCK_SIZE=$4 WRITE_MODE=$5 if [ "$WRITE_MODE" != "write" ] && [ "$WRITE_MODE" != "randwrite" ]; then echo "Invalid WRITE_MODE, must be 'write' or 'randwrite'" exit 1 fi cat <<EOF > /tmp/fio-job.ini [writers] rw=$WRITE_MODE fsync=$FSYNC_FREQ fallocate=none group_reporting=1 direct=0 bs=$BLOCK_SIZE ioengine=sync size=$FILE_SIZE directory=$MNT numjobs=$NUM_JOBS EOF echo "performance" | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor echo echo "Using config:" echo cat /tmp/fio-job.ini echo umount $MNT &> /dev/null mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT fio /tmp/fio-job.ini umount $MNT The results were the following: ************************* *** sequential writes *** ************************* ==== 1 job, 8GiB file, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=36.6MiB/s (38.4MB/s), 36.6MiB/s-36.6MiB/s (38.4MB/s-38.4MB/s), io=8192MiB (8590MB), run=223689-223689msec After patch: WRITE: bw=40.2MiB/s (42.1MB/s), 40.2MiB/s-40.2MiB/s (42.1MB/s-42.1MB/s), io=8192MiB (8590MB), run=203980-203980msec (+9.8%, -8.8% runtime) ==== 2 jobs, 4GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=35.8MiB/s (37.5MB/s), 35.8MiB/s-35.8MiB/s (37.5MB/s-37.5MB/s), io=8192MiB (8590MB), run=228950-228950msec After patch: WRITE: bw=43.5MiB/s (45.6MB/s), 43.5MiB/s-43.5MiB/s (45.6MB/s-45.6MB/s), io=8192MiB (8590MB), run=188272-188272msec (+21.5% throughput, -17.8% runtime) ==== 4 jobs, 2GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=50.1MiB/s (52.6MB/s), 50.1MiB/s-50.1MiB/s (52.6MB/s-52.6MB/s), io=8192MiB (8590MB), run=163446-163446msec After patch: WRITE: bw=64.5MiB/s (67.6MB/s), 64.5MiB/s-64.5MiB/s (67.6MB/s-67.6MB/s), io=8192MiB (8590MB), run=126987-126987msec (+28.7% throughput, -22.3% runtime) ==== 8 jobs, 1GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=64.0MiB/s (68.1MB/s), 64.0MiB/s-64.0MiB/s (68.1MB/s-68.1MB/s), io=8192MiB (8590MB), run=126075-126075msec After patch: WRITE: bw=86.8MiB/s (91.0MB/s), 86.8MiB/s-86.8MiB/s (91.0MB/s-91.0MB/s), io=8192MiB (8590MB), run=94358-94358msec (+35.6% throughput, -25.2% runtime) ==== 16 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=79.8MiB/s (83.6MB/s), 79.8MiB/s-79.8MiB/s (83.6MB/s-83.6MB/s), io=8192MiB (8590MB), run=102694-102694msec After patch: WRITE: bw=107MiB/s (112MB/s), 107MiB/s-107MiB/s (112MB/s-112MB/s), io=8192MiB (8590MB), run=76446-76446msec (+34.1% throughput, -25.6% runtime) ==== 32 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=93.2MiB/s (97.7MB/s), 93.2MiB/s-93.2MiB/s (97.7MB/s-97.7MB/s), io=16.0GiB (17.2GB), run=175836-175836msec After patch: WRITE: bw=111MiB/s (117MB/s), 111MiB/s-111MiB/s (117MB/s-117MB/s), io=16.0GiB (17.2GB), run=147001-147001msec (+19.1% throughput, -16.4% runtime) ==== 64 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=108MiB/s (114MB/s), 108MiB/s-108MiB/s (114MB/s-114MB/s), io=32.0GiB (34.4GB), run=302656-302656msec After patch: WRITE: bw=133MiB/s (140MB/s), 133MiB/s-133MiB/s (140MB/s-140MB/s), io=32.0GiB (34.4GB), run=246003-246003msec (+23.1% throughput, -18.7% runtime) ************************ *** random writes *** ************************ ==== 1 job, 8GiB file, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=11.5MiB/s (12.0MB/s), 11.5MiB/s-11.5MiB/s (12.0MB/s-12.0MB/s), io=8192MiB (8590MB), run=714281-714281msec After patch: WRITE: bw=11.6MiB/s (12.2MB/s), 11.6MiB/s-11.6MiB/s (12.2MB/s-12.2MB/s), io=8192MiB (8590MB), run=705959-705959msec (+0.9% throughput, -1.7% runtime) ==== 2 jobs, 4GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=12.8MiB/s (13.5MB/s), 12.8MiB/s-12.8MiB/s (13.5MB/s-13.5MB/s), io=8192MiB (8590MB), run=638101-638101msec After patch: WRITE: bw=13.1MiB/s (13.7MB/s), 13.1MiB/s-13.1MiB/s (13.7MB/s-13.7MB/s), io=8192MiB (8590MB), run=625374-625374msec (+2.3% throughput, -2.0% runtime) ==== 4 jobs, 2GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=15.4MiB/s (16.2MB/s), 15.4MiB/s-15.4MiB/s (16.2MB/s-16.2MB/s), io=8192MiB (8590MB), run=531146-531146msec After patch: WRITE: bw=17.8MiB/s (18.7MB/s), 17.8MiB/s-17.8MiB/s (18.7MB/s-18.7MB/s), io=8192MiB (8590MB), run=460431-460431msec (+15.6% throughput, -13.3% runtime) ==== 8 jobs, 1GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=19.9MiB/s (20.8MB/s), 19.9MiB/s-19.9MiB/s (20.8MB/s-20.8MB/s), io=8192MiB (8590MB), run=412664-412664msec After patch: WRITE: bw=22.2MiB/s (23.3MB/s), 22.2MiB/s-22.2MiB/s (23.3MB/s-23.3MB/s), io=8192MiB (8590MB), run=368589-368589msec (+11.6% throughput, -10.7% runtime) ==== 16 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=29.3MiB/s (30.7MB/s), 29.3MiB/s-29.3MiB/s (30.7MB/s-30.7MB/s), io=8192MiB (8590MB), run=279924-279924msec After patch: WRITE: bw=30.4MiB/s (31.9MB/s), 30.4MiB/s-30.4MiB/s (31.9MB/s-31.9MB/s), io=8192MiB (8590MB), run=269258-269258msec (+3.8% throughput, -3.8% runtime) ==== 32 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=36.9MiB/s (38.7MB/s), 36.9MiB/s-36.9MiB/s (38.7MB/s-38.7MB/s), io=16.0GiB (17.2GB), run=443581-443581msec After patch: WRITE: bw=41.6MiB/s (43.6MB/s), 41.6MiB/s-41.6MiB/s (43.6MB/s-43.6MB/s), io=16.0GiB (17.2GB), run=394114-394114msec (+12.7% throughput, -11.2% runtime) ==== 64 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=45.9MiB/s (48.1MB/s), 45.9MiB/s-45.9MiB/s (48.1MB/s-48.1MB/s), io=32.0GiB (34.4GB), run=714614-714614msec After patch: WRITE: bw=48.8MiB/s (51.1MB/s), 48.8MiB/s-48.8MiB/s (51.1MB/s-51.1MB/s), io=32.0GiB (34.4GB), run=672087-672087msec (+6.3% throughput, -6.0% runtime) Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-08-11 14:43:58 +03:00
list_for_each_entry_safe(em, n, &em_tree->modified_extents, list)
list_del_init(&em->list);
write_unlock(&em_tree->lock);
Btrfs: turbo charge fsync At least for the vm workload. Currently on fsync we will 1) Truncate all items in the log tree for the given inode if they exist and 2) Copy all items for a given inode into the log The problem with this is that for things like VMs you can have lots of extents from the fragmented writing behavior, and worst yet you may have only modified a few extents, not the entire thing. This patch fixes this problem by tracking which transid modified our extent, and then when we do the tree logging we find all of the extents we've modified in our current transaction, sort them and commit them. We also only truncate up to the xattrs of the inode and copy that stuff in normally, and then just drop any extents in the range we have that exist in the log already. Here are some numbers of a 50 meg fio job that does random writes and fsync()s after every write Original Patched SATA drive 82KB/s 140KB/s Fusion drive 431KB/s 2532KB/s So around 2-6 times faster depending on your hardware. There are a few corner cases, for example if you truncate at all we have to do it the old way since there is no way to be sure what is in the log is ok. This probably could be done smarter, but if you write-fsync-truncate-write-fsync you deserve what you get. All this work is in RAM of course so if your inode gets evicted from cache and you read it in and fsync it we'll do it the slow way if we are still in the same transaction that we last modified the inode in. The biggest cool part of this is that it requires no changes to the recovery code, so if you fsync with this patch and crash and load an old kernel, it will run the recovery and be a-ok. I have tested this pretty thoroughly with an fsync tester and everything comes back fine, as well as xfstests. Thanks, Signed-off-by: Josef Bacik <jbacik@fusionio.com>
2012-08-17 21:14:17 +04:00
}
if (inode_only == LOG_INODE_ALL && S_ISDIR(inode->vfs_inode.i_mode)) {
ret = log_directory_changes(trans, inode, path, dst_path, ctx);
if (ret)
goto out_unlock;
}
btrfs: always update the logged transaction when logging new names When we are logging a new name for an inode, due to a link or rename operation, if the inode has ancestor inodes that are new, created in the current transaction, we need to log that these inodes exist. To ensure that a subsequent explicit fsync on one of these ancestor inodes does sync the log, we don't set the logged_trans field of these inodes. This was done in commit 75b463d2b47aef ("btrfs: do not commit logs and transactions during link and rename operations"), to avoid syncing a log after a rename or link operation. In order to allow for future changes to do some optimizations, change this behaviour to always update the logged_trans of any logged inode and don't update the last_log_commit of the inode if we are logging that it exists. This accomplishes that same objective with simpler logic, allowing for some optimizations in the next patches. So just do that simplification. This patch is part of a patch set comprised of the following patches: btrfs: check if a log tree exists at inode_logged() btrfs: remove no longer needed checks for NULL log context btrfs: do not log new dentries when logging that a new name exists btrfs: always update the logged transaction when logging new names btrfs: avoid expensive search when dropping inode items from log btrfs: add helper to truncate inode items when logging inode btrfs: avoid expensive search when truncating inode items from the log btrfs: avoid search for logged i_size when logging inode if possible btrfs: avoid attempt to drop extents when logging inode for the first time btrfs: do not commit delayed inode when logging a file in full sync mode This is patch 4/10 and test results are listed in the change log of the last patch in the set. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-08-31 17:30:34 +03:00
spin_lock(&inode->lock);
inode->logged_trans = trans->transid;
Btrfs: fix data loss after inode eviction, renaming it, and fsync it When we log an inode, regardless of logging it completely or only that it exists, we always update it as logged (logged_trans and last_log_commit fields of the inode are updated). This is generally fine and avoids future attempts to log it from having to do repeated work that brings no value. However, if we write data to a file, then evict its inode after all the dealloc was flushed (and ordered extents completed), rename the file and fsync it, we end up not logging the new extents, since the rename may result in logging that the inode exists in case the parent directory was logged before. The following reproducer shows and explains how this can happen: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/dir $ touch /mnt/dir/foo $ touch /mnt/dir/bar # Do a direct IO write instead of a buffered write because with a # buffered write we would need to make sure dealloc gets flushed and # complete before we do the inode eviction later, and we can not do that # from user space with call to things such as sync(2) since that results # in a transaction commit as well. $ xfs_io -d -c "pwrite -S 0xd3 0 4K" /mnt/dir/bar # Keep the directory dir in use while we evict inodes. We want our file # bar's inode to be evicted but we don't want our directory's inode to # be evicted (if it were evicted too, we would not be able to reproduce # the issue since the first fsync below, of file foo, would result in a # transaction commit. $ ( cd /mnt/dir; while true; do :; done ) & $ pid=$! # Wait a bit to give time for the background process to chdir. $ sleep 0.1 # Evict all inodes, except the inode for the directory dir because it is # currently in use by our background process. $ echo 2 > /proc/sys/vm/drop_caches # fsync file foo, which ends up persisting information about the parent # directory because it is a new inode. $ xfs_io -c fsync /mnt/dir/foo # Rename bar, this results in logging that this inode exists (inode item, # names, xattrs) because the parent directory is in the log. $ mv /mnt/dir/bar /mnt/dir/baz # Now fsync baz, which ends up doing absolutely nothing because of the # rename operation which logged that the inode exists only. $ xfs_io -c fsync /mnt/dir/baz <power failure> $ mount /dev/sdb /mnt $ od -t x1 -A d /mnt/dir/baz 0000000 --> Empty file, data we wrote is missing. Fix this by not updating last_sub_trans of an inode when we are logging only that it exists and the inode was not yet logged since it was loaded from disk (full_sync bit set), this is enough to make btrfs_inode_in_log() return false for this scenario and make us log the inode. The logged_trans of the inode is still always setsince that alone is used to track if names need to be deleted as part of unlink operations. Fixes: 257c62e1bce03e ("Btrfs: avoid tree log commit when there are no changes") CC: stable@vger.kernel.org # 4.4+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-06-07 13:25:24 +03:00
/*
btrfs: always update the logged transaction when logging new names When we are logging a new name for an inode, due to a link or rename operation, if the inode has ancestor inodes that are new, created in the current transaction, we need to log that these inodes exist. To ensure that a subsequent explicit fsync on one of these ancestor inodes does sync the log, we don't set the logged_trans field of these inodes. This was done in commit 75b463d2b47aef ("btrfs: do not commit logs and transactions during link and rename operations"), to avoid syncing a log after a rename or link operation. In order to allow for future changes to do some optimizations, change this behaviour to always update the logged_trans of any logged inode and don't update the last_log_commit of the inode if we are logging that it exists. This accomplishes that same objective with simpler logic, allowing for some optimizations in the next patches. So just do that simplification. This patch is part of a patch set comprised of the following patches: btrfs: check if a log tree exists at inode_logged() btrfs: remove no longer needed checks for NULL log context btrfs: do not log new dentries when logging that a new name exists btrfs: always update the logged transaction when logging new names btrfs: avoid expensive search when dropping inode items from log btrfs: add helper to truncate inode items when logging inode btrfs: avoid expensive search when truncating inode items from the log btrfs: avoid search for logged i_size when logging inode if possible btrfs: avoid attempt to drop extents when logging inode for the first time btrfs: do not commit delayed inode when logging a file in full sync mode This is patch 4/10 and test results are listed in the change log of the last patch in the set. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-08-31 17:30:34 +03:00
* Don't update last_log_commit if we logged that an inode exists.
* We do this for three reasons:
*
* 1) We might have had buffered writes to this inode that were
* flushed and had their ordered extents completed in this
* transaction, but we did not previously log the inode with
* LOG_INODE_ALL. Later the inode was evicted and after that
* it was loaded again and this LOG_INODE_EXISTS log operation
* happened. We must make sure that if an explicit fsync against
* the inode is performed later, it logs the new extents, an
* updated inode item, etc, and syncs the log. The same logic
* applies to direct IO writes instead of buffered writes.
*
* 2) When we log the inode with LOG_INODE_EXISTS, its inode item
* is logged with an i_size of 0 or whatever value was logged
* before. If later the i_size of the inode is increased by a
* truncate operation, the log is synced through an fsync of
* some other inode and then finally an explicit fsync against
* this inode is made, we must make sure this fsync logs the
* inode with the new i_size, the hole between old i_size and
* the new i_size, and syncs the log.
*
* 3) If we are logging that an ancestor inode exists as part of
* logging a new name from a link or rename operation, don't update
* its last_log_commit - otherwise if an explicit fsync is made
* against an ancestor, the fsync considers the inode in the log
* and doesn't sync the log, resulting in the ancestor missing after
* a power failure unless the log was synced as part of an fsync
* against any other unrelated inode.
Btrfs: fix data loss after inode eviction, renaming it, and fsync it When we log an inode, regardless of logging it completely or only that it exists, we always update it as logged (logged_trans and last_log_commit fields of the inode are updated). This is generally fine and avoids future attempts to log it from having to do repeated work that brings no value. However, if we write data to a file, then evict its inode after all the dealloc was flushed (and ordered extents completed), rename the file and fsync it, we end up not logging the new extents, since the rename may result in logging that the inode exists in case the parent directory was logged before. The following reproducer shows and explains how this can happen: $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/dir $ touch /mnt/dir/foo $ touch /mnt/dir/bar # Do a direct IO write instead of a buffered write because with a # buffered write we would need to make sure dealloc gets flushed and # complete before we do the inode eviction later, and we can not do that # from user space with call to things such as sync(2) since that results # in a transaction commit as well. $ xfs_io -d -c "pwrite -S 0xd3 0 4K" /mnt/dir/bar # Keep the directory dir in use while we evict inodes. We want our file # bar's inode to be evicted but we don't want our directory's inode to # be evicted (if it were evicted too, we would not be able to reproduce # the issue since the first fsync below, of file foo, would result in a # transaction commit. $ ( cd /mnt/dir; while true; do :; done ) & $ pid=$! # Wait a bit to give time for the background process to chdir. $ sleep 0.1 # Evict all inodes, except the inode for the directory dir because it is # currently in use by our background process. $ echo 2 > /proc/sys/vm/drop_caches # fsync file foo, which ends up persisting information about the parent # directory because it is a new inode. $ xfs_io -c fsync /mnt/dir/foo # Rename bar, this results in logging that this inode exists (inode item, # names, xattrs) because the parent directory is in the log. $ mv /mnt/dir/bar /mnt/dir/baz # Now fsync baz, which ends up doing absolutely nothing because of the # rename operation which logged that the inode exists only. $ xfs_io -c fsync /mnt/dir/baz <power failure> $ mount /dev/sdb /mnt $ od -t x1 -A d /mnt/dir/baz 0000000 --> Empty file, data we wrote is missing. Fix this by not updating last_sub_trans of an inode when we are logging only that it exists and the inode was not yet logged since it was loaded from disk (full_sync bit set), this is enough to make btrfs_inode_in_log() return false for this scenario and make us log the inode. The logged_trans of the inode is still always setsince that alone is used to track if names need to be deleted as part of unlink operations. Fixes: 257c62e1bce03e ("Btrfs: avoid tree log commit when there are no changes") CC: stable@vger.kernel.org # 4.4+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-06-07 13:25:24 +03:00
*/
btrfs: always update the logged transaction when logging new names When we are logging a new name for an inode, due to a link or rename operation, if the inode has ancestor inodes that are new, created in the current transaction, we need to log that these inodes exist. To ensure that a subsequent explicit fsync on one of these ancestor inodes does sync the log, we don't set the logged_trans field of these inodes. This was done in commit 75b463d2b47aef ("btrfs: do not commit logs and transactions during link and rename operations"), to avoid syncing a log after a rename or link operation. In order to allow for future changes to do some optimizations, change this behaviour to always update the logged_trans of any logged inode and don't update the last_log_commit of the inode if we are logging that it exists. This accomplishes that same objective with simpler logic, allowing for some optimizations in the next patches. So just do that simplification. This patch is part of a patch set comprised of the following patches: btrfs: check if a log tree exists at inode_logged() btrfs: remove no longer needed checks for NULL log context btrfs: do not log new dentries when logging that a new name exists btrfs: always update the logged transaction when logging new names btrfs: avoid expensive search when dropping inode items from log btrfs: add helper to truncate inode items when logging inode btrfs: avoid expensive search when truncating inode items from the log btrfs: avoid search for logged i_size when logging inode if possible btrfs: avoid attempt to drop extents when logging inode for the first time btrfs: do not commit delayed inode when logging a file in full sync mode This is patch 4/10 and test results are listed in the change log of the last patch in the set. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-08-31 17:30:34 +03:00
if (inode_only != LOG_INODE_EXISTS)
inode->last_log_commit = inode->last_sub_trans;
spin_unlock(&inode->lock);
out_unlock:
mutex_unlock(&inode->log_mutex);
btrfs: do not commit delayed inode when logging a file in full sync mode When logging a regular file in full sync mode, we are currently committing its delayed inode item. This is to ensure that we never miss copying the inode item, with its most up to date data, into the log tree. However that is not necessary since commit e4545de5b035 ("Btrfs: fix fsync data loss after append write"), because even if we don't find the leaf with the inode item when looking for leaves that changed in the current transaction, we end up logging the inode item later using the in-memory content. In case we find the leaf containing the inode item, we already end up using the in-memory inode for filling the inode item in the log tree, and not the inode item that is in the fs/subvolume tree, as it might be not up to date (copy_items() -> fill_inode_item()). So don't commit the delayed inode item, which brings a couple of benefits: 1) Avoid writing the inode item to the fs/subvolume btree, saving time and reducing lock contention on the btree; 2) In case no other item for the inode was changed, added or deleted in the same leaf where the inode item is located, we ended up copying all the items in that leaf to the log tree - it's harmless from a functional point of view, but it wastes time and log tree space. This patch is part of a patch set comprised of the following patches: btrfs: check if a log tree exists at inode_logged() btrfs: remove no longer needed checks for NULL log context btrfs: do not log new dentries when logging that a new name exists btrfs: always update the logged transaction when logging new names btrfs: avoid expensive search when dropping inode items from log btrfs: add helper to truncate inode items when logging inode btrfs: avoid expensive search when truncating inode items from the log btrfs: avoid search for logged i_size when logging inode if possible btrfs: avoid attempt to drop extents when logging inode for the first time btrfs: do not commit delayed inode when logging a file in full sync mode This is patch 10/10 and the following test results compare a branch with the whole patch set applied versus a branch without any of the patches applied. The following script was used to test dbench with 8 and 16 jobs on a machine with 12 cores, 64G of RAM, a NVME device and using a non-debug kernel config (Debian's default): $ cat test.sh #!/bin/bash if [ $# -ne 1 ]; then echo "Use $0 NUM_JOBS" exit 1 fi NUM_JOBS=$1 DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 MOUNT_OPTIONS="-o ssd" MKFS_OPTIONS="-m single -d single" echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -D $MNT -t 120 $NUM_JOBS umount $MNT The results were the following: 8 jobs, before patchset: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 4113896 0.009 238.665 Close 3021699 0.001 0.590 Rename 174215 0.082 238.733 Unlink 830977 0.049 238.642 Deltree 96 2.232 8.022 Mkdir 48 0.003 0.005 Qpathinfo 3729013 0.005 2.672 Qfileinfo 653206 0.001 0.152 Qfsinfo 683866 0.002 0.526 Sfileinfo 335055 0.004 1.571 Find 1441800 0.016 4.288 WriteX 2049644 0.010 3.982 ReadX 6449786 0.003 0.969 LockX 13400 0.002 0.043 UnlockX 13400 0.001 0.075 Flush 288349 2.521 245.516 Throughput 1075.73 MB/sec 8 clients 8 procs max_latency=245.520 ms 8 jobs, after patchset: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 4154282 0.009 156.675 Close 3051450 0.001 0.843 Rename 175912 0.072 4.444 Unlink 839067 0.048 66.050 Deltree 96 2.131 5.979 Mkdir 48 0.002 0.004 Qpathinfo 3765575 0.005 3.079 Qfileinfo 659582 0.001 0.099 Qfsinfo 690474 0.002 0.155 Sfileinfo 338366 0.004 1.419 Find 1455816 0.016 3.423 WriteX 2069538 0.010 4.328 ReadX 6512429 0.003 0.840 LockX 13530 0.002 0.078 UnlockX 13530 0.001 0.051 Flush 291158 2.500 163.468 Throughput 1105.45 MB/sec 8 clients 8 procs max_latency=163.474 ms +2.7% throughput, -40.1% max latency 16 jobs, before patchset: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 5457602 0.033 337.098 Close 4008979 0.002 2.018 Rename 231051 0.323 254.054 Unlink 1102209 0.202 337.243 Deltree 160 6.521 31.720 Mkdir 80 0.003 0.007 Qpathinfo 4946147 0.014 6.988 Qfileinfo 867440 0.001 1.642 Qfsinfo 907081 0.003 1.821 Sfileinfo 444433 0.005 2.053 Find 1912506 0.067 7.854 WriteX 2724852 0.018 7.428 ReadX 8553883 0.003 2.059 LockX 17770 0.003 0.350 UnlockX 17770 0.002 0.627 Flush 382533 2.810 353.691 Throughput 1413.09 MB/sec 16 clients 16 procs max_latency=353.696 ms 16 jobs, after patchset: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 5393156 0.034 303.181 Close 3961986 0.002 1.502 Rename 228359 0.320 253.379 Unlink 1088920 0.206 303.409 Deltree 160 6.419 30.088 Mkdir 80 0.003 0.004 Qpathinfo 4887967 0.015 7.722 Qfileinfo 857408 0.001 1.651 Qfsinfo 896343 0.002 2.147 Sfileinfo 439317 0.005 4.298 Find 1890018 0.073 8.347 WriteX 2693356 0.018 6.373 ReadX 8453485 0.003 3.836 LockX 17562 0.003 0.486 UnlockX 17562 0.002 0.635 Flush 378023 2.802 315.904 Throughput 1454.46 MB/sec 16 clients 16 procs max_latency=315.910 ms +2.9% throughput, -11.3% max latency Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-08-31 17:30:40 +03:00
out:
btrfs_free_path(path);
btrfs_free_path(dst_path);
btrfs: avoid inode logging during rename and link when possible During a rename or link operation, we need to determine if an inode was previously logged or not, and if it was, do some update to the logged inode. We used to rely exclusively on the logged_trans field of struct btrfs_inode to determine that, but that was not reliable because the value of that field is not persisted in the inode item, so it's lost when an inode is evicted and loaded back again. That led to several issues in the past, such as not persisting deletions (such as the case fixed by commit 803f0f64d17769 ("Btrfs: fix fsync not persisting dentry deletions due to inode evictions")), or resulting in losing a file after an inode eviction followed by a rename (commit ecc64fab7d49c6 ("btrfs: fix lost inode on log replay after mix of fsync, rename and inode eviction")), besides other issues. So the inode_logged() helper was introduced and used to determine if an inode was possibly logged before in the current transaction, with the caveat that it could return false positives, in the sense that even if an inode was not logged before in the current transaction, it could still return true, but never to return false in case the inode was logged. >From a functional point of view that is fine, but from a performance perspective it can introduce significant latencies to rename and link operations, as they will end up doing inode logging even when it is not necessary. Recently on a 5.15 kernel, an openSUSE Tumbleweed user reported package installations and upgrades, with the zypper tool, were often taking a long time to complete. With strace it could be observed that zypper was spending about 99% of its time on rename operations, and then with further analysis we checked that directory logging was happening too frequently. Taking into account that installation/upgrade of some of the packages needed a few thousand file renames, the slowdown was very noticeable for the user. The issue was caused indirectly due to an excessive number of inode evictions on a 5.15 kernel, about 100x more compared to a 5.13, 5.14 or a 5.16-rc8 kernel. While triggering the inode evictions if something outside btrfs' control, btrfs could still behave better by eliminating the false positives from the inode_logged() helper. So change inode_logged() to actually eliminate such false positives caused by inode eviction and when an inode was never logged since the filesystem was mounted, as both cases relate to when the logged_trans field of struct btrfs_inode has a value of zero. When it can not determine if the inode was logged based only on the logged_trans value, lookup for the existence of the inode item in the log tree - if it's there then we known the inode was logged, if it's not there then it can not have been logged in the current transaction. Once we determine if the inode was logged, update the logged_trans value to avoid future calls to have to search in the log tree again. Alternatively, we could start storing logged_trans in the on disk inode item structure (struct btrfs_inode_item) in the unused space it still has, but that would be a bit odd because: 1) We only care about logged_trans since the filesystem was mounted, we don't care about its value from a previous mount. Having it persisted in the inode item structure would not make the best use of the precious unused space; 2) In order to get logged_trans persisted before inode eviction, we would have to update the delayed inode when we finish logging the inode and update its logged_trans in struct btrfs_inode, which makes it a bit cumbersome since we need to check if the delayed inode exists, if not create it and populate it and deal with any errors (-ENOMEM mostly). This change is part of a patchset comprised of the following patches: 1/5 btrfs: add helper to delete a dir entry from a log tree 2/5 btrfs: pass the dentry to btrfs_log_new_name() instead of the inode 3/5 btrfs: avoid logging all directory changes during renames 4/5 btrfs: stop doing unnecessary log updates during a rename 5/5 btrfs: avoid inode logging during rename and link when possible The following test script mimics part of what the zypper tool does during package installations/upgrades. It does not triggers inode evictions, but it's similar because it triggers false positives from the inode_logged() helper, because the inodes have a logged_trans of 0, there's a log tree due to a fsync of an unrelated file and the directory inode has its last_trans field set to the current transaction: $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=10000 mkfs.btrfs -f $DEV mount $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync # Now do some change to an unrelated file and fsync it. # This is just to create a log tree to make sure that inode_logged() # does not return false when called against "testdir". xfs_io -f -c "pwrite 0 4K" -c "fsync" $MNT/foo # Do some change to testdir. This is to make sure inode_logged() # will return true when called against "testdir", because its # logged_trans is 0, it was changed in the current transaction # and there's a log tree. echo -n > $MNT/testdir/file_$((NUM_FILES + 1)) echo "Renaming $NUM_FILES files..." start=$(date +%s%N) for ((i = 1; i <= $NUM_FILES; i++)); do mv $MNT/testdir/file_$i $MNT/testdir/file_$i-RPMDELETE done end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "Renames took $dur milliseconds" umount $MNT Testing this change on a box using a non-debug kernel (Debian's default kernel config) gave the following results: NUM_FILES=10000, before patchset: 27837 ms NUM_FILES=10000, after patches 1/5 to 4/5 applied: 9236 ms (-66.8%) NUM_FILES=10000, after whole patchset applied: 8902 ms (-68.0%) NUM_FILES=5000, before patchset: 9127 ms NUM_FILES=5000, after patches 1/5 to 4/5 applied: 4640 ms (-49.2%) NUM_FILES=5000, after whole patchset applied: 4441 ms (-51.3%) NUM_FILES=2000, before patchset: 2528 ms NUM_FILES=2000, after patches 1/5 to 4/5 applied: 1983 ms (-21.6%) NUM_FILES=2000, after whole patchset applied: 1747 ms (-30.9%) NUM_FILES=1000, before patchset: 1085 ms NUM_FILES=1000, after patches 1/5 to 4/5 applied: 893 ms (-17.7%) NUM_FILES=1000, after whole patchset applied: 867 ms (-20.1%) Running dbench on the same physical machine with the following script: $ cat run-dbench.sh #!/bin/bash NUM_JOBS=$(nproc --all) DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 MOUNT_OPTIONS="-o ssd" MKFS_OPTIONS="-O no-holes -R free-space-tree" echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -D $MNT -t 120 $NUM_JOBS umount $MNT Before patchset: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 3761352 0.032 143.843 Close 2762770 0.002 2.273 Rename 159304 0.291 67.037 Unlink 759784 0.207 143.998 Deltree 72 4.028 15.977 Mkdir 36 0.003 0.006 Qpathinfo 3409780 0.013 9.678 Qfileinfo 596772 0.001 0.878 Qfsinfo 625189 0.003 1.245 Sfileinfo 306443 0.006 1.840 Find 1318106 0.063 19.798 WriteX 1871137 0.021 8.532 ReadX 5897325 0.003 3.567 LockX 12252 0.003 0.258 UnlockX 12252 0.002 0.100 Flush 263666 3.327 155.632 Throughput 980.047 MB/sec 12 clients 12 procs max_latency=155.636 ms After whole patchset applied: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 4195584 0.033 107.742 Close 3081932 0.002 1.935 Rename 177641 0.218 14.905 Unlink 847333 0.166 107.822 Deltree 118 5.315 15.247 Mkdir 59 0.004 0.048 Qpathinfo 3802612 0.014 10.302 Qfileinfo 666748 0.001 1.034 Qfsinfo 697329 0.003 0.944 Sfileinfo 341712 0.006 2.099 Find 1470365 0.065 9.359 WriteX 2093921 0.021 8.087 ReadX 6576234 0.003 3.407 LockX 13660 0.003 0.308 UnlockX 13660 0.002 0.114 Flush 294090 2.906 115.539 Throughput 1093.11 MB/sec 12 clients 12 procs max_latency=115.544 ms +11.5% throughput -25.8% max latency rename max latency -77.8% Link: https://bugzilla.opensuse.org/show_bug.cgi?id=1193549 Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-01-20 14:00:10 +03:00
if (recursive_logging)
ctx->logged_before = orig_logged_before;
return ret;
}
/*
* Check if we need to log an inode. This is used in contexts where while
* logging an inode we need to log another inode (either that it exists or in
* full mode). This is used instead of btrfs_inode_in_log() because the later
* requires the inode to be in the log and have the log transaction committed,
* while here we do not care if the log transaction was already committed - our
* caller will commit the log later - and we want to avoid logging an inode
* multiple times when multiple tasks have joined the same log transaction.
*/
static bool need_log_inode(struct btrfs_trans_handle *trans,
struct btrfs_inode *inode)
{
/*
* If a directory was not modified, no dentries added or removed, we can
* and should avoid logging it.
*/
if (S_ISDIR(inode->vfs_inode.i_mode) && inode->last_trans < trans->transid)
return false;
/*
* If this inode does not have new/updated/deleted xattrs since the last
* time it was logged and is flagged as logged in the current transaction,
* we can skip logging it. As for new/deleted names, those are updated in
* the log by link/unlink/rename operations.
* In case the inode was logged and then evicted and reloaded, its
* logged_trans will be 0, in which case we have to fully log it since
* logged_trans is a transient field, not persisted.
*/
if (inode->logged_trans == trans->transid &&
!test_bit(BTRFS_INODE_COPY_EVERYTHING, &inode->runtime_flags))
return false;
return true;
}
Btrfs: fix metadata inconsistencies after directory fsync We can get into inconsistency between inodes and directory entries after fsyncing a directory. The issue is that while a directory gets the new dentries persisted in the fsync log and replayed at mount time, the link count of the inode that directory entries point to doesn't get updated, staying with an incorrect link count (smaller then the correct value). This later leads to stale file handle errors when accessing (including attempt to delete) some of the links if all the other ones are removed, which also implies impossibility to delete the parent directories, since the dentries can not be removed. Another issue is that (unlike ext3/4, xfs, f2fs, reiserfs, nilfs2), when fsyncing a directory, new files aren't logged (their metadata and dentries) nor any child directories. So this patch fixes this issue too, since it has the same resolution as the incorrect inode link count issue mentioned before. This is very easy to reproduce, and the following excerpt from my test case for xfstests shows how: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create our main test file and directory. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 8K" $SCRATCH_MNT/foo | _filter_xfs_io mkdir $SCRATCH_MNT/mydir # Make sure all metadata and data are durably persisted. sync # Add a hard link to 'foo' inside our test directory and fsync only the # directory. The btrfs fsync implementation had a bug that caused the new # directory entry to be visible after the fsync log replay but, the inode # of our file remained with a link count of 1. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_2 # Add a few more links and new files. # This is just to verify nothing breaks or gives incorrect results after the # fsync log is replayed. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_3 $XFS_IO_PROG -f -c "pwrite -S 0xff 0 64K" $SCRATCH_MNT/hello | _filter_xfs_io ln $SCRATCH_MNT/hello $SCRATCH_MNT/mydir/hello_2 # Add some subdirectories and new files and links to them. This is to verify # that after fsyncing our top level directory 'mydir', all the subdirectories # and their files/links are registered in the fsync log and exist after the # fsync log is replayed. mkdir -p $SCRATCH_MNT/mydir/x/y/z ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/foo_y_link ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/z/foo_z_link touch $SCRATCH_MNT/mydir/x/y/z/qwerty # Now fsync only our top directory. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/mydir # And fsync now our new file named 'hello', just to verify later that it has # the expected content and that the previous fsync on the directory 'mydir' had # no bad influence on this fsync. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/hello # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # Verify the content of our file 'foo' remains the same as before, 8192 bytes, # all with the value 0xaa. echo "File 'foo' content after log replay:" od -t x1 $SCRATCH_MNT/foo # Remove the first name of our inode. Because of the directory fsync bug, the # inode's link count was 1 instead of 5, so removing the 'foo' name ended up # deleting the inode and the other names became stale directory entries (still # visible to applications). Attempting to remove or access the remaining # dentries pointing to that inode resulted in stale file handle errors and # made it impossible to remove the parent directories since it was impossible # for them to become empty. echo "file 'foo' link count after log replay: $(stat -c %h $SCRATCH_MNT/foo)" rm -f $SCRATCH_MNT/foo # Now verify that all files, links and directories created before fsyncing our # directory exist after the fsync log was replayed. [ -f $SCRATCH_MNT/mydir/foo_2 ] || echo "Link mydir/foo_2 is missing" [ -f $SCRATCH_MNT/mydir/foo_3 ] || echo "Link mydir/foo_3 is missing" [ -f $SCRATCH_MNT/hello ] || echo "File hello is missing" [ -f $SCRATCH_MNT/mydir/hello_2 ] || echo "Link mydir/hello_2 is missing" [ -f $SCRATCH_MNT/mydir/x/y/foo_y_link ] || \ echo "Link mydir/x/y/foo_y_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/foo_z_link ] || \ echo "Link mydir/x/y/z/foo_z_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/qwerty ] || \ echo "File mydir/x/y/z/qwerty is missing" # We expect our file here to have a size of 64Kb and all the bytes having the # value 0xff. echo "file 'hello' content after log replay:" od -t x1 $SCRATCH_MNT/hello # Now remove all files/links, under our test directory 'mydir', and verify we # can remove all the directories. rm -f $SCRATCH_MNT/mydir/x/y/z/* rmdir $SCRATCH_MNT/mydir/x/y/z rm -f $SCRATCH_MNT/mydir/x/y/* rmdir $SCRATCH_MNT/mydir/x/y rmdir $SCRATCH_MNT/mydir/x rm -f $SCRATCH_MNT/mydir/* rmdir $SCRATCH_MNT/mydir # An fsck, run by the fstests framework everytime a test finishes, also detected # the inconsistency and printed the following error message: # # root 5 inode 257 errors 2001, no inode item, link count wrong # unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref # unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref status=0 exit The expected golden output for the test is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 5 file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 Which is the output after this patch and when running the test against ext3/4, xfs, f2fs, reiserfs or nilfs2. Without this patch, the test's output is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 1 Link mydir/foo_2 is missing Link mydir/foo_3 is missing Link mydir/x/y/foo_y_link is missing Link mydir/x/y/z/foo_z_link is missing File mydir/x/y/z/qwerty is missing file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y/z': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x': No such file or directory rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_2': Stale file handle rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_3': Stale file handle rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir': Directory not empty Fsck, without this fix, also complains about the wrong link count: root 5 inode 257 errors 2001, no inode item, link count wrong unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref So fix this by logging the inodes that the dentries point to when fsyncing a directory. A test case for xfstests follows. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-03-20 20:19:46 +03:00
struct btrfs_dir_list {
u64 ino;
struct list_head list;
};
/*
* Log the inodes of the new dentries of a directory. See log_dir_items() for
* details about the why it is needed.
* This is a recursive operation - if an existing dentry corresponds to a
* directory, that directory's new entries are logged too (same behaviour as
* ext3/4, xfs, f2fs, reiserfs, nilfs2). Note that when logging the inodes
* the dentries point to we do not lock their i_mutex, otherwise lockdep
* complains about the following circular lock dependency / possible deadlock:
*
* CPU0 CPU1
* ---- ----
* lock(&type->i_mutex_dir_key#3/2);
* lock(sb_internal#2);
* lock(&type->i_mutex_dir_key#3/2);
* lock(&sb->s_type->i_mutex_key#14);
*
* Where sb_internal is the lock (a counter that works as a lock) acquired by
* sb_start_intwrite() in btrfs_start_transaction().
* Not locking i_mutex of the inodes is still safe because:
*
* 1) For regular files we log with a mode of LOG_INODE_EXISTS. It's possible
* that while logging the inode new references (names) are added or removed
* from the inode, leaving the logged inode item with a link count that does
* not match the number of logged inode reference items. This is fine because
* at log replay time we compute the real number of links and correct the
* link count in the inode item (see replay_one_buffer() and
* link_to_fixup_dir());
*
* 2) For directories we log with a mode of LOG_INODE_ALL. It's possible that
btrfs: only copy dir index keys when logging a directory Currently, when logging a directory, we copy both dir items and dir index items from the fs/subvolume tree to the log tree. Both items have exactly the same data (same struct btrfs_dir_item), the difference lies in the key values, where a dir index key contains the index number of a directory entry while the dir item key does not, as it's used for doing fast lookups of an entry by name, while the former is used for sorting entries when listing a directory. We can exploit that and log only the dir index items, since they contain all the information needed to correctly add, replace and delete directory entries when replaying a log tree. Logging only the dir index items is also backward and forward compatible: an unpatched kernel (without this change) can correctly replay a log tree generated by a patched kernel (with this patch), and a patched kernel can correctly replay a log tree generated by an unpatched kernel. The backward compatibility is ensured because: 1) For inserting a new dentry: a dentry is only inserted when we find a new dir index key - we can only insert if we know the dir index offset, which is encoded in the dir index key's offset; 2) For deleting dentries: during log replay, before adding or replacing dentries, we first replay dentry deletions. Whenever we find a dir item key or a dir index key in the subvolume/fs tree that is not logged in a range for which the log tree is authoritative, we do the unlink of the dentry, which removes both the existing dir item key and the dir index key. Therefore logging just dir index keys is enough to ensure dentry deletions are correctly replayed; 3) For dentry replacements: they work when we log only dir index keys and this is mostly due to a combination of 1) and 2). If we replace a dentry with name "foobar" to point from inode A to inode B, then we know the dir index key for the new dentry is different from the old one, as it has an index number (key offset) larger than the old one. This results in replaying a deletion, through replay_dir_deletes(), that causes the old dentry to be removed, both the dir item key and the dir index key, as mentioned at 2). Then when processing the new dir index key, we add the new dentry, adding both a new dir item key and a new index key pointing to inode B, as stated in 1). The forward compatibility, the ability for a patched kernel to replay a log created by an older, unpatched kernel, comes from the changes required for making sure we are able to replay a log that only contains dir index keys - we simply ignore every dir item key we find. So modify directory logging to log only dir index items, and modify the log replay process to ignore dir item keys, from log trees created by an unpatched kernel, and process only with dir index keys. This reduces the amount of logged metadata by about half, and therefore the time spent logging or fsyncing large directories (less CPU time and less IO). The following test script was used to measure this change: #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=1000000 NUM_FILE_DELETES=10000 mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The tests were run on a physical machine, with a non-debug kernel (Debian's default kernel config), for different values of $NUM_NEW_FILES and $NUM_FILE_DELETES, and the results were the following: ** Before patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 8412 ms after adding 1000000 files dir fsync took 500 ms after deleting 10000 files ** After patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 4252 ms after adding 1000000 files (-49.5%) dir fsync took 269 ms after deleting 10000 files (-46.2%) ** Before patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 745 ms after adding 100000 files dir fsync took 59 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 404 ms after adding 100000 files (-45.8%) dir fsync took 31 ms after deleting 1000 files (-47.5%) ** Before patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 67 ms after adding 10000 files dir fsync took 9 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 36 ms after adding 10000 files (-46.3%) dir fsync took 5 ms after deleting 1000 files (-44.4%) ** Before patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 9 ms after adding 1000 files dir fsync took 4 ms after deleting 100 files ** After patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 7 ms after adding 1000 files (-22.2%) dir fsync took 3 ms after deleting 100 files (-25.0%) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-10-25 19:31:53 +03:00
* while logging the inode's items new index items (key type
* BTRFS_DIR_INDEX_KEY) are added to fs/subvol tree and the logged inode item
Btrfs: fix metadata inconsistencies after directory fsync We can get into inconsistency between inodes and directory entries after fsyncing a directory. The issue is that while a directory gets the new dentries persisted in the fsync log and replayed at mount time, the link count of the inode that directory entries point to doesn't get updated, staying with an incorrect link count (smaller then the correct value). This later leads to stale file handle errors when accessing (including attempt to delete) some of the links if all the other ones are removed, which also implies impossibility to delete the parent directories, since the dentries can not be removed. Another issue is that (unlike ext3/4, xfs, f2fs, reiserfs, nilfs2), when fsyncing a directory, new files aren't logged (their metadata and dentries) nor any child directories. So this patch fixes this issue too, since it has the same resolution as the incorrect inode link count issue mentioned before. This is very easy to reproduce, and the following excerpt from my test case for xfstests shows how: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create our main test file and directory. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 8K" $SCRATCH_MNT/foo | _filter_xfs_io mkdir $SCRATCH_MNT/mydir # Make sure all metadata and data are durably persisted. sync # Add a hard link to 'foo' inside our test directory and fsync only the # directory. The btrfs fsync implementation had a bug that caused the new # directory entry to be visible after the fsync log replay but, the inode # of our file remained with a link count of 1. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_2 # Add a few more links and new files. # This is just to verify nothing breaks or gives incorrect results after the # fsync log is replayed. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_3 $XFS_IO_PROG -f -c "pwrite -S 0xff 0 64K" $SCRATCH_MNT/hello | _filter_xfs_io ln $SCRATCH_MNT/hello $SCRATCH_MNT/mydir/hello_2 # Add some subdirectories and new files and links to them. This is to verify # that after fsyncing our top level directory 'mydir', all the subdirectories # and their files/links are registered in the fsync log and exist after the # fsync log is replayed. mkdir -p $SCRATCH_MNT/mydir/x/y/z ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/foo_y_link ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/z/foo_z_link touch $SCRATCH_MNT/mydir/x/y/z/qwerty # Now fsync only our top directory. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/mydir # And fsync now our new file named 'hello', just to verify later that it has # the expected content and that the previous fsync on the directory 'mydir' had # no bad influence on this fsync. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/hello # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # Verify the content of our file 'foo' remains the same as before, 8192 bytes, # all with the value 0xaa. echo "File 'foo' content after log replay:" od -t x1 $SCRATCH_MNT/foo # Remove the first name of our inode. Because of the directory fsync bug, the # inode's link count was 1 instead of 5, so removing the 'foo' name ended up # deleting the inode and the other names became stale directory entries (still # visible to applications). Attempting to remove or access the remaining # dentries pointing to that inode resulted in stale file handle errors and # made it impossible to remove the parent directories since it was impossible # for them to become empty. echo "file 'foo' link count after log replay: $(stat -c %h $SCRATCH_MNT/foo)" rm -f $SCRATCH_MNT/foo # Now verify that all files, links and directories created before fsyncing our # directory exist after the fsync log was replayed. [ -f $SCRATCH_MNT/mydir/foo_2 ] || echo "Link mydir/foo_2 is missing" [ -f $SCRATCH_MNT/mydir/foo_3 ] || echo "Link mydir/foo_3 is missing" [ -f $SCRATCH_MNT/hello ] || echo "File hello is missing" [ -f $SCRATCH_MNT/mydir/hello_2 ] || echo "Link mydir/hello_2 is missing" [ -f $SCRATCH_MNT/mydir/x/y/foo_y_link ] || \ echo "Link mydir/x/y/foo_y_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/foo_z_link ] || \ echo "Link mydir/x/y/z/foo_z_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/qwerty ] || \ echo "File mydir/x/y/z/qwerty is missing" # We expect our file here to have a size of 64Kb and all the bytes having the # value 0xff. echo "file 'hello' content after log replay:" od -t x1 $SCRATCH_MNT/hello # Now remove all files/links, under our test directory 'mydir', and verify we # can remove all the directories. rm -f $SCRATCH_MNT/mydir/x/y/z/* rmdir $SCRATCH_MNT/mydir/x/y/z rm -f $SCRATCH_MNT/mydir/x/y/* rmdir $SCRATCH_MNT/mydir/x/y rmdir $SCRATCH_MNT/mydir/x rm -f $SCRATCH_MNT/mydir/* rmdir $SCRATCH_MNT/mydir # An fsck, run by the fstests framework everytime a test finishes, also detected # the inconsistency and printed the following error message: # # root 5 inode 257 errors 2001, no inode item, link count wrong # unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref # unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref status=0 exit The expected golden output for the test is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 5 file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 Which is the output after this patch and when running the test against ext3/4, xfs, f2fs, reiserfs or nilfs2. Without this patch, the test's output is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 1 Link mydir/foo_2 is missing Link mydir/foo_3 is missing Link mydir/x/y/foo_y_link is missing Link mydir/x/y/z/foo_z_link is missing File mydir/x/y/z/qwerty is missing file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y/z': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x': No such file or directory rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_2': Stale file handle rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_3': Stale file handle rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir': Directory not empty Fsck, without this fix, also complains about the wrong link count: root 5 inode 257 errors 2001, no inode item, link count wrong unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref So fix this by logging the inodes that the dentries point to when fsyncing a directory. A test case for xfstests follows. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-03-20 20:19:46 +03:00
* has a size that doesn't match the sum of the lengths of all the logged
btrfs: only copy dir index keys when logging a directory Currently, when logging a directory, we copy both dir items and dir index items from the fs/subvolume tree to the log tree. Both items have exactly the same data (same struct btrfs_dir_item), the difference lies in the key values, where a dir index key contains the index number of a directory entry while the dir item key does not, as it's used for doing fast lookups of an entry by name, while the former is used for sorting entries when listing a directory. We can exploit that and log only the dir index items, since they contain all the information needed to correctly add, replace and delete directory entries when replaying a log tree. Logging only the dir index items is also backward and forward compatible: an unpatched kernel (without this change) can correctly replay a log tree generated by a patched kernel (with this patch), and a patched kernel can correctly replay a log tree generated by an unpatched kernel. The backward compatibility is ensured because: 1) For inserting a new dentry: a dentry is only inserted when we find a new dir index key - we can only insert if we know the dir index offset, which is encoded in the dir index key's offset; 2) For deleting dentries: during log replay, before adding or replacing dentries, we first replay dentry deletions. Whenever we find a dir item key or a dir index key in the subvolume/fs tree that is not logged in a range for which the log tree is authoritative, we do the unlink of the dentry, which removes both the existing dir item key and the dir index key. Therefore logging just dir index keys is enough to ensure dentry deletions are correctly replayed; 3) For dentry replacements: they work when we log only dir index keys and this is mostly due to a combination of 1) and 2). If we replace a dentry with name "foobar" to point from inode A to inode B, then we know the dir index key for the new dentry is different from the old one, as it has an index number (key offset) larger than the old one. This results in replaying a deletion, through replay_dir_deletes(), that causes the old dentry to be removed, both the dir item key and the dir index key, as mentioned at 2). Then when processing the new dir index key, we add the new dentry, adding both a new dir item key and a new index key pointing to inode B, as stated in 1). The forward compatibility, the ability for a patched kernel to replay a log created by an older, unpatched kernel, comes from the changes required for making sure we are able to replay a log that only contains dir index keys - we simply ignore every dir item key we find. So modify directory logging to log only dir index items, and modify the log replay process to ignore dir item keys, from log trees created by an unpatched kernel, and process only with dir index keys. This reduces the amount of logged metadata by about half, and therefore the time spent logging or fsyncing large directories (less CPU time and less IO). The following test script was used to measure this change: #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=1000000 NUM_FILE_DELETES=10000 mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The tests were run on a physical machine, with a non-debug kernel (Debian's default kernel config), for different values of $NUM_NEW_FILES and $NUM_FILE_DELETES, and the results were the following: ** Before patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 8412 ms after adding 1000000 files dir fsync took 500 ms after deleting 10000 files ** After patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 4252 ms after adding 1000000 files (-49.5%) dir fsync took 269 ms after deleting 10000 files (-46.2%) ** Before patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 745 ms after adding 100000 files dir fsync took 59 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 404 ms after adding 100000 files (-45.8%) dir fsync took 31 ms after deleting 1000 files (-47.5%) ** Before patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 67 ms after adding 10000 files dir fsync took 9 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 36 ms after adding 10000 files (-46.3%) dir fsync took 5 ms after deleting 1000 files (-44.4%) ** Before patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 9 ms after adding 1000 files dir fsync took 4 ms after deleting 100 files ** After patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 7 ms after adding 1000 files (-22.2%) dir fsync took 3 ms after deleting 100 files (-25.0%) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-10-25 19:31:53 +03:00
* names - this is ok, not a problem, because at log replay time we set the
* directory's i_size to the correct value (see replay_one_name() and
* do_overwrite_item()).
Btrfs: fix metadata inconsistencies after directory fsync We can get into inconsistency between inodes and directory entries after fsyncing a directory. The issue is that while a directory gets the new dentries persisted in the fsync log and replayed at mount time, the link count of the inode that directory entries point to doesn't get updated, staying with an incorrect link count (smaller then the correct value). This later leads to stale file handle errors when accessing (including attempt to delete) some of the links if all the other ones are removed, which also implies impossibility to delete the parent directories, since the dentries can not be removed. Another issue is that (unlike ext3/4, xfs, f2fs, reiserfs, nilfs2), when fsyncing a directory, new files aren't logged (their metadata and dentries) nor any child directories. So this patch fixes this issue too, since it has the same resolution as the incorrect inode link count issue mentioned before. This is very easy to reproduce, and the following excerpt from my test case for xfstests shows how: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create our main test file and directory. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 8K" $SCRATCH_MNT/foo | _filter_xfs_io mkdir $SCRATCH_MNT/mydir # Make sure all metadata and data are durably persisted. sync # Add a hard link to 'foo' inside our test directory and fsync only the # directory. The btrfs fsync implementation had a bug that caused the new # directory entry to be visible after the fsync log replay but, the inode # of our file remained with a link count of 1. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_2 # Add a few more links and new files. # This is just to verify nothing breaks or gives incorrect results after the # fsync log is replayed. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_3 $XFS_IO_PROG -f -c "pwrite -S 0xff 0 64K" $SCRATCH_MNT/hello | _filter_xfs_io ln $SCRATCH_MNT/hello $SCRATCH_MNT/mydir/hello_2 # Add some subdirectories and new files and links to them. This is to verify # that after fsyncing our top level directory 'mydir', all the subdirectories # and their files/links are registered in the fsync log and exist after the # fsync log is replayed. mkdir -p $SCRATCH_MNT/mydir/x/y/z ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/foo_y_link ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/z/foo_z_link touch $SCRATCH_MNT/mydir/x/y/z/qwerty # Now fsync only our top directory. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/mydir # And fsync now our new file named 'hello', just to verify later that it has # the expected content and that the previous fsync on the directory 'mydir' had # no bad influence on this fsync. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/hello # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # Verify the content of our file 'foo' remains the same as before, 8192 bytes, # all with the value 0xaa. echo "File 'foo' content after log replay:" od -t x1 $SCRATCH_MNT/foo # Remove the first name of our inode. Because of the directory fsync bug, the # inode's link count was 1 instead of 5, so removing the 'foo' name ended up # deleting the inode and the other names became stale directory entries (still # visible to applications). Attempting to remove or access the remaining # dentries pointing to that inode resulted in stale file handle errors and # made it impossible to remove the parent directories since it was impossible # for them to become empty. echo "file 'foo' link count after log replay: $(stat -c %h $SCRATCH_MNT/foo)" rm -f $SCRATCH_MNT/foo # Now verify that all files, links and directories created before fsyncing our # directory exist after the fsync log was replayed. [ -f $SCRATCH_MNT/mydir/foo_2 ] || echo "Link mydir/foo_2 is missing" [ -f $SCRATCH_MNT/mydir/foo_3 ] || echo "Link mydir/foo_3 is missing" [ -f $SCRATCH_MNT/hello ] || echo "File hello is missing" [ -f $SCRATCH_MNT/mydir/hello_2 ] || echo "Link mydir/hello_2 is missing" [ -f $SCRATCH_MNT/mydir/x/y/foo_y_link ] || \ echo "Link mydir/x/y/foo_y_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/foo_z_link ] || \ echo "Link mydir/x/y/z/foo_z_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/qwerty ] || \ echo "File mydir/x/y/z/qwerty is missing" # We expect our file here to have a size of 64Kb and all the bytes having the # value 0xff. echo "file 'hello' content after log replay:" od -t x1 $SCRATCH_MNT/hello # Now remove all files/links, under our test directory 'mydir', and verify we # can remove all the directories. rm -f $SCRATCH_MNT/mydir/x/y/z/* rmdir $SCRATCH_MNT/mydir/x/y/z rm -f $SCRATCH_MNT/mydir/x/y/* rmdir $SCRATCH_MNT/mydir/x/y rmdir $SCRATCH_MNT/mydir/x rm -f $SCRATCH_MNT/mydir/* rmdir $SCRATCH_MNT/mydir # An fsck, run by the fstests framework everytime a test finishes, also detected # the inconsistency and printed the following error message: # # root 5 inode 257 errors 2001, no inode item, link count wrong # unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref # unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref status=0 exit The expected golden output for the test is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 5 file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 Which is the output after this patch and when running the test against ext3/4, xfs, f2fs, reiserfs or nilfs2. Without this patch, the test's output is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 1 Link mydir/foo_2 is missing Link mydir/foo_3 is missing Link mydir/x/y/foo_y_link is missing Link mydir/x/y/z/foo_z_link is missing File mydir/x/y/z/qwerty is missing file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y/z': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x': No such file or directory rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_2': Stale file handle rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_3': Stale file handle rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir': Directory not empty Fsck, without this fix, also complains about the wrong link count: root 5 inode 257 errors 2001, no inode item, link count wrong unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref So fix this by logging the inodes that the dentries point to when fsyncing a directory. A test case for xfstests follows. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-03-20 20:19:46 +03:00
*/
static int log_new_dir_dentries(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_inode *start_inode,
Btrfs: fix metadata inconsistencies after directory fsync We can get into inconsistency between inodes and directory entries after fsyncing a directory. The issue is that while a directory gets the new dentries persisted in the fsync log and replayed at mount time, the link count of the inode that directory entries point to doesn't get updated, staying with an incorrect link count (smaller then the correct value). This later leads to stale file handle errors when accessing (including attempt to delete) some of the links if all the other ones are removed, which also implies impossibility to delete the parent directories, since the dentries can not be removed. Another issue is that (unlike ext3/4, xfs, f2fs, reiserfs, nilfs2), when fsyncing a directory, new files aren't logged (their metadata and dentries) nor any child directories. So this patch fixes this issue too, since it has the same resolution as the incorrect inode link count issue mentioned before. This is very easy to reproduce, and the following excerpt from my test case for xfstests shows how: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create our main test file and directory. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 8K" $SCRATCH_MNT/foo | _filter_xfs_io mkdir $SCRATCH_MNT/mydir # Make sure all metadata and data are durably persisted. sync # Add a hard link to 'foo' inside our test directory and fsync only the # directory. The btrfs fsync implementation had a bug that caused the new # directory entry to be visible after the fsync log replay but, the inode # of our file remained with a link count of 1. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_2 # Add a few more links and new files. # This is just to verify nothing breaks or gives incorrect results after the # fsync log is replayed. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_3 $XFS_IO_PROG -f -c "pwrite -S 0xff 0 64K" $SCRATCH_MNT/hello | _filter_xfs_io ln $SCRATCH_MNT/hello $SCRATCH_MNT/mydir/hello_2 # Add some subdirectories and new files and links to them. This is to verify # that after fsyncing our top level directory 'mydir', all the subdirectories # and their files/links are registered in the fsync log and exist after the # fsync log is replayed. mkdir -p $SCRATCH_MNT/mydir/x/y/z ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/foo_y_link ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/z/foo_z_link touch $SCRATCH_MNT/mydir/x/y/z/qwerty # Now fsync only our top directory. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/mydir # And fsync now our new file named 'hello', just to verify later that it has # the expected content and that the previous fsync on the directory 'mydir' had # no bad influence on this fsync. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/hello # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # Verify the content of our file 'foo' remains the same as before, 8192 bytes, # all with the value 0xaa. echo "File 'foo' content after log replay:" od -t x1 $SCRATCH_MNT/foo # Remove the first name of our inode. Because of the directory fsync bug, the # inode's link count was 1 instead of 5, so removing the 'foo' name ended up # deleting the inode and the other names became stale directory entries (still # visible to applications). Attempting to remove or access the remaining # dentries pointing to that inode resulted in stale file handle errors and # made it impossible to remove the parent directories since it was impossible # for them to become empty. echo "file 'foo' link count after log replay: $(stat -c %h $SCRATCH_MNT/foo)" rm -f $SCRATCH_MNT/foo # Now verify that all files, links and directories created before fsyncing our # directory exist after the fsync log was replayed. [ -f $SCRATCH_MNT/mydir/foo_2 ] || echo "Link mydir/foo_2 is missing" [ -f $SCRATCH_MNT/mydir/foo_3 ] || echo "Link mydir/foo_3 is missing" [ -f $SCRATCH_MNT/hello ] || echo "File hello is missing" [ -f $SCRATCH_MNT/mydir/hello_2 ] || echo "Link mydir/hello_2 is missing" [ -f $SCRATCH_MNT/mydir/x/y/foo_y_link ] || \ echo "Link mydir/x/y/foo_y_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/foo_z_link ] || \ echo "Link mydir/x/y/z/foo_z_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/qwerty ] || \ echo "File mydir/x/y/z/qwerty is missing" # We expect our file here to have a size of 64Kb and all the bytes having the # value 0xff. echo "file 'hello' content after log replay:" od -t x1 $SCRATCH_MNT/hello # Now remove all files/links, under our test directory 'mydir', and verify we # can remove all the directories. rm -f $SCRATCH_MNT/mydir/x/y/z/* rmdir $SCRATCH_MNT/mydir/x/y/z rm -f $SCRATCH_MNT/mydir/x/y/* rmdir $SCRATCH_MNT/mydir/x/y rmdir $SCRATCH_MNT/mydir/x rm -f $SCRATCH_MNT/mydir/* rmdir $SCRATCH_MNT/mydir # An fsck, run by the fstests framework everytime a test finishes, also detected # the inconsistency and printed the following error message: # # root 5 inode 257 errors 2001, no inode item, link count wrong # unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref # unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref status=0 exit The expected golden output for the test is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 5 file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 Which is the output after this patch and when running the test against ext3/4, xfs, f2fs, reiserfs or nilfs2. Without this patch, the test's output is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 1 Link mydir/foo_2 is missing Link mydir/foo_3 is missing Link mydir/x/y/foo_y_link is missing Link mydir/x/y/z/foo_z_link is missing File mydir/x/y/z/qwerty is missing file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y/z': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x': No such file or directory rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_2': Stale file handle rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_3': Stale file handle rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir': Directory not empty Fsck, without this fix, also complains about the wrong link count: root 5 inode 257 errors 2001, no inode item, link count wrong unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref So fix this by logging the inodes that the dentries point to when fsyncing a directory. A test case for xfstests follows. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-03-20 20:19:46 +03:00
struct btrfs_log_ctx *ctx)
{
struct btrfs_fs_info *fs_info = root->fs_info;
Btrfs: fix metadata inconsistencies after directory fsync We can get into inconsistency between inodes and directory entries after fsyncing a directory. The issue is that while a directory gets the new dentries persisted in the fsync log and replayed at mount time, the link count of the inode that directory entries point to doesn't get updated, staying with an incorrect link count (smaller then the correct value). This later leads to stale file handle errors when accessing (including attempt to delete) some of the links if all the other ones are removed, which also implies impossibility to delete the parent directories, since the dentries can not be removed. Another issue is that (unlike ext3/4, xfs, f2fs, reiserfs, nilfs2), when fsyncing a directory, new files aren't logged (their metadata and dentries) nor any child directories. So this patch fixes this issue too, since it has the same resolution as the incorrect inode link count issue mentioned before. This is very easy to reproduce, and the following excerpt from my test case for xfstests shows how: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create our main test file and directory. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 8K" $SCRATCH_MNT/foo | _filter_xfs_io mkdir $SCRATCH_MNT/mydir # Make sure all metadata and data are durably persisted. sync # Add a hard link to 'foo' inside our test directory and fsync only the # directory. The btrfs fsync implementation had a bug that caused the new # directory entry to be visible after the fsync log replay but, the inode # of our file remained with a link count of 1. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_2 # Add a few more links and new files. # This is just to verify nothing breaks or gives incorrect results after the # fsync log is replayed. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_3 $XFS_IO_PROG -f -c "pwrite -S 0xff 0 64K" $SCRATCH_MNT/hello | _filter_xfs_io ln $SCRATCH_MNT/hello $SCRATCH_MNT/mydir/hello_2 # Add some subdirectories and new files and links to them. This is to verify # that after fsyncing our top level directory 'mydir', all the subdirectories # and their files/links are registered in the fsync log and exist after the # fsync log is replayed. mkdir -p $SCRATCH_MNT/mydir/x/y/z ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/foo_y_link ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/z/foo_z_link touch $SCRATCH_MNT/mydir/x/y/z/qwerty # Now fsync only our top directory. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/mydir # And fsync now our new file named 'hello', just to verify later that it has # the expected content and that the previous fsync on the directory 'mydir' had # no bad influence on this fsync. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/hello # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # Verify the content of our file 'foo' remains the same as before, 8192 bytes, # all with the value 0xaa. echo "File 'foo' content after log replay:" od -t x1 $SCRATCH_MNT/foo # Remove the first name of our inode. Because of the directory fsync bug, the # inode's link count was 1 instead of 5, so removing the 'foo' name ended up # deleting the inode and the other names became stale directory entries (still # visible to applications). Attempting to remove or access the remaining # dentries pointing to that inode resulted in stale file handle errors and # made it impossible to remove the parent directories since it was impossible # for them to become empty. echo "file 'foo' link count after log replay: $(stat -c %h $SCRATCH_MNT/foo)" rm -f $SCRATCH_MNT/foo # Now verify that all files, links and directories created before fsyncing our # directory exist after the fsync log was replayed. [ -f $SCRATCH_MNT/mydir/foo_2 ] || echo "Link mydir/foo_2 is missing" [ -f $SCRATCH_MNT/mydir/foo_3 ] || echo "Link mydir/foo_3 is missing" [ -f $SCRATCH_MNT/hello ] || echo "File hello is missing" [ -f $SCRATCH_MNT/mydir/hello_2 ] || echo "Link mydir/hello_2 is missing" [ -f $SCRATCH_MNT/mydir/x/y/foo_y_link ] || \ echo "Link mydir/x/y/foo_y_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/foo_z_link ] || \ echo "Link mydir/x/y/z/foo_z_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/qwerty ] || \ echo "File mydir/x/y/z/qwerty is missing" # We expect our file here to have a size of 64Kb and all the bytes having the # value 0xff. echo "file 'hello' content after log replay:" od -t x1 $SCRATCH_MNT/hello # Now remove all files/links, under our test directory 'mydir', and verify we # can remove all the directories. rm -f $SCRATCH_MNT/mydir/x/y/z/* rmdir $SCRATCH_MNT/mydir/x/y/z rm -f $SCRATCH_MNT/mydir/x/y/* rmdir $SCRATCH_MNT/mydir/x/y rmdir $SCRATCH_MNT/mydir/x rm -f $SCRATCH_MNT/mydir/* rmdir $SCRATCH_MNT/mydir # An fsck, run by the fstests framework everytime a test finishes, also detected # the inconsistency and printed the following error message: # # root 5 inode 257 errors 2001, no inode item, link count wrong # unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref # unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref status=0 exit The expected golden output for the test is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 5 file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 Which is the output after this patch and when running the test against ext3/4, xfs, f2fs, reiserfs or nilfs2. Without this patch, the test's output is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 1 Link mydir/foo_2 is missing Link mydir/foo_3 is missing Link mydir/x/y/foo_y_link is missing Link mydir/x/y/z/foo_z_link is missing File mydir/x/y/z/qwerty is missing file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y/z': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x': No such file or directory rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_2': Stale file handle rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_3': Stale file handle rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir': Directory not empty Fsck, without this fix, also complains about the wrong link count: root 5 inode 257 errors 2001, no inode item, link count wrong unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref So fix this by logging the inodes that the dentries point to when fsyncing a directory. A test case for xfstests follows. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-03-20 20:19:46 +03:00
struct btrfs_path *path;
LIST_HEAD(dir_list);
struct btrfs_dir_list *dir_elem;
int ret = 0;
btrfs: do not log new dentries when logging that a new name exists When logging a new name for an inode, due to a link or rename operation, we don't need to log all new dentries of the parent directories and their subdirectories. We only want to log the names of the inode and that any new parent directories exist. So in this case don't trigger logging of the new dentries, that is only need when doing an explicit fsync on a directory or on a file which requires logging its parent directories. This avoids unnecessary work and reduces contention on the extent buffers of a log tree. This patch is part of a patch set comprised of the following patches: btrfs: check if a log tree exists at inode_logged() btrfs: remove no longer needed checks for NULL log context btrfs: do not log new dentries when logging that a new name exists btrfs: always update the logged transaction when logging new names btrfs: avoid expensive search when dropping inode items from log btrfs: add helper to truncate inode items when logging inode btrfs: avoid expensive search when truncating inode items from the log btrfs: avoid search for logged i_size when logging inode if possible btrfs: avoid attempt to drop extents when logging inode for the first time btrfs: do not commit delayed inode when logging a file in full sync mode This is patch 3/10 and test results are listed in the change log of the last patch in the set. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-08-31 17:30:33 +03:00
/*
* If we are logging a new name, as part of a link or rename operation,
* don't bother logging new dentries, as we just want to log the names
* of an inode and that any new parents exist.
*/
if (ctx->logging_new_name)
return 0;
Btrfs: fix metadata inconsistencies after directory fsync We can get into inconsistency between inodes and directory entries after fsyncing a directory. The issue is that while a directory gets the new dentries persisted in the fsync log and replayed at mount time, the link count of the inode that directory entries point to doesn't get updated, staying with an incorrect link count (smaller then the correct value). This later leads to stale file handle errors when accessing (including attempt to delete) some of the links if all the other ones are removed, which also implies impossibility to delete the parent directories, since the dentries can not be removed. Another issue is that (unlike ext3/4, xfs, f2fs, reiserfs, nilfs2), when fsyncing a directory, new files aren't logged (their metadata and dentries) nor any child directories. So this patch fixes this issue too, since it has the same resolution as the incorrect inode link count issue mentioned before. This is very easy to reproduce, and the following excerpt from my test case for xfstests shows how: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create our main test file and directory. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 8K" $SCRATCH_MNT/foo | _filter_xfs_io mkdir $SCRATCH_MNT/mydir # Make sure all metadata and data are durably persisted. sync # Add a hard link to 'foo' inside our test directory and fsync only the # directory. The btrfs fsync implementation had a bug that caused the new # directory entry to be visible after the fsync log replay but, the inode # of our file remained with a link count of 1. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_2 # Add a few more links and new files. # This is just to verify nothing breaks or gives incorrect results after the # fsync log is replayed. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_3 $XFS_IO_PROG -f -c "pwrite -S 0xff 0 64K" $SCRATCH_MNT/hello | _filter_xfs_io ln $SCRATCH_MNT/hello $SCRATCH_MNT/mydir/hello_2 # Add some subdirectories and new files and links to them. This is to verify # that after fsyncing our top level directory 'mydir', all the subdirectories # and their files/links are registered in the fsync log and exist after the # fsync log is replayed. mkdir -p $SCRATCH_MNT/mydir/x/y/z ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/foo_y_link ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/z/foo_z_link touch $SCRATCH_MNT/mydir/x/y/z/qwerty # Now fsync only our top directory. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/mydir # And fsync now our new file named 'hello', just to verify later that it has # the expected content and that the previous fsync on the directory 'mydir' had # no bad influence on this fsync. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/hello # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # Verify the content of our file 'foo' remains the same as before, 8192 bytes, # all with the value 0xaa. echo "File 'foo' content after log replay:" od -t x1 $SCRATCH_MNT/foo # Remove the first name of our inode. Because of the directory fsync bug, the # inode's link count was 1 instead of 5, so removing the 'foo' name ended up # deleting the inode and the other names became stale directory entries (still # visible to applications). Attempting to remove or access the remaining # dentries pointing to that inode resulted in stale file handle errors and # made it impossible to remove the parent directories since it was impossible # for them to become empty. echo "file 'foo' link count after log replay: $(stat -c %h $SCRATCH_MNT/foo)" rm -f $SCRATCH_MNT/foo # Now verify that all files, links and directories created before fsyncing our # directory exist after the fsync log was replayed. [ -f $SCRATCH_MNT/mydir/foo_2 ] || echo "Link mydir/foo_2 is missing" [ -f $SCRATCH_MNT/mydir/foo_3 ] || echo "Link mydir/foo_3 is missing" [ -f $SCRATCH_MNT/hello ] || echo "File hello is missing" [ -f $SCRATCH_MNT/mydir/hello_2 ] || echo "Link mydir/hello_2 is missing" [ -f $SCRATCH_MNT/mydir/x/y/foo_y_link ] || \ echo "Link mydir/x/y/foo_y_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/foo_z_link ] || \ echo "Link mydir/x/y/z/foo_z_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/qwerty ] || \ echo "File mydir/x/y/z/qwerty is missing" # We expect our file here to have a size of 64Kb and all the bytes having the # value 0xff. echo "file 'hello' content after log replay:" od -t x1 $SCRATCH_MNT/hello # Now remove all files/links, under our test directory 'mydir', and verify we # can remove all the directories. rm -f $SCRATCH_MNT/mydir/x/y/z/* rmdir $SCRATCH_MNT/mydir/x/y/z rm -f $SCRATCH_MNT/mydir/x/y/* rmdir $SCRATCH_MNT/mydir/x/y rmdir $SCRATCH_MNT/mydir/x rm -f $SCRATCH_MNT/mydir/* rmdir $SCRATCH_MNT/mydir # An fsck, run by the fstests framework everytime a test finishes, also detected # the inconsistency and printed the following error message: # # root 5 inode 257 errors 2001, no inode item, link count wrong # unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref # unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref status=0 exit The expected golden output for the test is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 5 file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 Which is the output after this patch and when running the test against ext3/4, xfs, f2fs, reiserfs or nilfs2. Without this patch, the test's output is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 1 Link mydir/foo_2 is missing Link mydir/foo_3 is missing Link mydir/x/y/foo_y_link is missing Link mydir/x/y/z/foo_z_link is missing File mydir/x/y/z/qwerty is missing file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y/z': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x': No such file or directory rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_2': Stale file handle rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_3': Stale file handle rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir': Directory not empty Fsck, without this fix, also complains about the wrong link count: root 5 inode 257 errors 2001, no inode item, link count wrong unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref So fix this by logging the inodes that the dentries point to when fsyncing a directory. A test case for xfstests follows. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-03-20 20:19:46 +03:00
path = btrfs_alloc_path();
if (!path)
return -ENOMEM;
dir_elem = kmalloc(sizeof(*dir_elem), GFP_NOFS);
if (!dir_elem) {
btrfs_free_path(path);
return -ENOMEM;
}
dir_elem->ino = btrfs_ino(start_inode);
Btrfs: fix metadata inconsistencies after directory fsync We can get into inconsistency between inodes and directory entries after fsyncing a directory. The issue is that while a directory gets the new dentries persisted in the fsync log and replayed at mount time, the link count of the inode that directory entries point to doesn't get updated, staying with an incorrect link count (smaller then the correct value). This later leads to stale file handle errors when accessing (including attempt to delete) some of the links if all the other ones are removed, which also implies impossibility to delete the parent directories, since the dentries can not be removed. Another issue is that (unlike ext3/4, xfs, f2fs, reiserfs, nilfs2), when fsyncing a directory, new files aren't logged (their metadata and dentries) nor any child directories. So this patch fixes this issue too, since it has the same resolution as the incorrect inode link count issue mentioned before. This is very easy to reproduce, and the following excerpt from my test case for xfstests shows how: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create our main test file and directory. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 8K" $SCRATCH_MNT/foo | _filter_xfs_io mkdir $SCRATCH_MNT/mydir # Make sure all metadata and data are durably persisted. sync # Add a hard link to 'foo' inside our test directory and fsync only the # directory. The btrfs fsync implementation had a bug that caused the new # directory entry to be visible after the fsync log replay but, the inode # of our file remained with a link count of 1. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_2 # Add a few more links and new files. # This is just to verify nothing breaks or gives incorrect results after the # fsync log is replayed. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_3 $XFS_IO_PROG -f -c "pwrite -S 0xff 0 64K" $SCRATCH_MNT/hello | _filter_xfs_io ln $SCRATCH_MNT/hello $SCRATCH_MNT/mydir/hello_2 # Add some subdirectories and new files and links to them. This is to verify # that after fsyncing our top level directory 'mydir', all the subdirectories # and their files/links are registered in the fsync log and exist after the # fsync log is replayed. mkdir -p $SCRATCH_MNT/mydir/x/y/z ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/foo_y_link ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/z/foo_z_link touch $SCRATCH_MNT/mydir/x/y/z/qwerty # Now fsync only our top directory. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/mydir # And fsync now our new file named 'hello', just to verify later that it has # the expected content and that the previous fsync on the directory 'mydir' had # no bad influence on this fsync. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/hello # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # Verify the content of our file 'foo' remains the same as before, 8192 bytes, # all with the value 0xaa. echo "File 'foo' content after log replay:" od -t x1 $SCRATCH_MNT/foo # Remove the first name of our inode. Because of the directory fsync bug, the # inode's link count was 1 instead of 5, so removing the 'foo' name ended up # deleting the inode and the other names became stale directory entries (still # visible to applications). Attempting to remove or access the remaining # dentries pointing to that inode resulted in stale file handle errors and # made it impossible to remove the parent directories since it was impossible # for them to become empty. echo "file 'foo' link count after log replay: $(stat -c %h $SCRATCH_MNT/foo)" rm -f $SCRATCH_MNT/foo # Now verify that all files, links and directories created before fsyncing our # directory exist after the fsync log was replayed. [ -f $SCRATCH_MNT/mydir/foo_2 ] || echo "Link mydir/foo_2 is missing" [ -f $SCRATCH_MNT/mydir/foo_3 ] || echo "Link mydir/foo_3 is missing" [ -f $SCRATCH_MNT/hello ] || echo "File hello is missing" [ -f $SCRATCH_MNT/mydir/hello_2 ] || echo "Link mydir/hello_2 is missing" [ -f $SCRATCH_MNT/mydir/x/y/foo_y_link ] || \ echo "Link mydir/x/y/foo_y_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/foo_z_link ] || \ echo "Link mydir/x/y/z/foo_z_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/qwerty ] || \ echo "File mydir/x/y/z/qwerty is missing" # We expect our file here to have a size of 64Kb and all the bytes having the # value 0xff. echo "file 'hello' content after log replay:" od -t x1 $SCRATCH_MNT/hello # Now remove all files/links, under our test directory 'mydir', and verify we # can remove all the directories. rm -f $SCRATCH_MNT/mydir/x/y/z/* rmdir $SCRATCH_MNT/mydir/x/y/z rm -f $SCRATCH_MNT/mydir/x/y/* rmdir $SCRATCH_MNT/mydir/x/y rmdir $SCRATCH_MNT/mydir/x rm -f $SCRATCH_MNT/mydir/* rmdir $SCRATCH_MNT/mydir # An fsck, run by the fstests framework everytime a test finishes, also detected # the inconsistency and printed the following error message: # # root 5 inode 257 errors 2001, no inode item, link count wrong # unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref # unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref status=0 exit The expected golden output for the test is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 5 file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 Which is the output after this patch and when running the test against ext3/4, xfs, f2fs, reiserfs or nilfs2. Without this patch, the test's output is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 1 Link mydir/foo_2 is missing Link mydir/foo_3 is missing Link mydir/x/y/foo_y_link is missing Link mydir/x/y/z/foo_z_link is missing File mydir/x/y/z/qwerty is missing file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y/z': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x': No such file or directory rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_2': Stale file handle rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_3': Stale file handle rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir': Directory not empty Fsck, without this fix, also complains about the wrong link count: root 5 inode 257 errors 2001, no inode item, link count wrong unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref So fix this by logging the inodes that the dentries point to when fsyncing a directory. A test case for xfstests follows. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-03-20 20:19:46 +03:00
list_add_tail(&dir_elem->list, &dir_list);
while (!list_empty(&dir_list)) {
struct extent_buffer *leaf;
struct btrfs_key min_key;
int nritems;
int i;
dir_elem = list_first_entry(&dir_list, struct btrfs_dir_list,
list);
if (ret)
goto next_dir_inode;
min_key.objectid = dir_elem->ino;
btrfs: only copy dir index keys when logging a directory Currently, when logging a directory, we copy both dir items and dir index items from the fs/subvolume tree to the log tree. Both items have exactly the same data (same struct btrfs_dir_item), the difference lies in the key values, where a dir index key contains the index number of a directory entry while the dir item key does not, as it's used for doing fast lookups of an entry by name, while the former is used for sorting entries when listing a directory. We can exploit that and log only the dir index items, since they contain all the information needed to correctly add, replace and delete directory entries when replaying a log tree. Logging only the dir index items is also backward and forward compatible: an unpatched kernel (without this change) can correctly replay a log tree generated by a patched kernel (with this patch), and a patched kernel can correctly replay a log tree generated by an unpatched kernel. The backward compatibility is ensured because: 1) For inserting a new dentry: a dentry is only inserted when we find a new dir index key - we can only insert if we know the dir index offset, which is encoded in the dir index key's offset; 2) For deleting dentries: during log replay, before adding or replacing dentries, we first replay dentry deletions. Whenever we find a dir item key or a dir index key in the subvolume/fs tree that is not logged in a range for which the log tree is authoritative, we do the unlink of the dentry, which removes both the existing dir item key and the dir index key. Therefore logging just dir index keys is enough to ensure dentry deletions are correctly replayed; 3) For dentry replacements: they work when we log only dir index keys and this is mostly due to a combination of 1) and 2). If we replace a dentry with name "foobar" to point from inode A to inode B, then we know the dir index key for the new dentry is different from the old one, as it has an index number (key offset) larger than the old one. This results in replaying a deletion, through replay_dir_deletes(), that causes the old dentry to be removed, both the dir item key and the dir index key, as mentioned at 2). Then when processing the new dir index key, we add the new dentry, adding both a new dir item key and a new index key pointing to inode B, as stated in 1). The forward compatibility, the ability for a patched kernel to replay a log created by an older, unpatched kernel, comes from the changes required for making sure we are able to replay a log that only contains dir index keys - we simply ignore every dir item key we find. So modify directory logging to log only dir index items, and modify the log replay process to ignore dir item keys, from log trees created by an unpatched kernel, and process only with dir index keys. This reduces the amount of logged metadata by about half, and therefore the time spent logging or fsyncing large directories (less CPU time and less IO). The following test script was used to measure this change: #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=1000000 NUM_FILE_DELETES=10000 mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The tests were run on a physical machine, with a non-debug kernel (Debian's default kernel config), for different values of $NUM_NEW_FILES and $NUM_FILE_DELETES, and the results were the following: ** Before patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 8412 ms after adding 1000000 files dir fsync took 500 ms after deleting 10000 files ** After patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 4252 ms after adding 1000000 files (-49.5%) dir fsync took 269 ms after deleting 10000 files (-46.2%) ** Before patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 745 ms after adding 100000 files dir fsync took 59 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 404 ms after adding 100000 files (-45.8%) dir fsync took 31 ms after deleting 1000 files (-47.5%) ** Before patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 67 ms after adding 10000 files dir fsync took 9 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 36 ms after adding 10000 files (-46.3%) dir fsync took 5 ms after deleting 1000 files (-44.4%) ** Before patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 9 ms after adding 1000 files dir fsync took 4 ms after deleting 100 files ** After patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 7 ms after adding 1000 files (-22.2%) dir fsync took 3 ms after deleting 100 files (-25.0%) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-10-25 19:31:53 +03:00
min_key.type = BTRFS_DIR_INDEX_KEY;
Btrfs: fix metadata inconsistencies after directory fsync We can get into inconsistency between inodes and directory entries after fsyncing a directory. The issue is that while a directory gets the new dentries persisted in the fsync log and replayed at mount time, the link count of the inode that directory entries point to doesn't get updated, staying with an incorrect link count (smaller then the correct value). This later leads to stale file handle errors when accessing (including attempt to delete) some of the links if all the other ones are removed, which also implies impossibility to delete the parent directories, since the dentries can not be removed. Another issue is that (unlike ext3/4, xfs, f2fs, reiserfs, nilfs2), when fsyncing a directory, new files aren't logged (their metadata and dentries) nor any child directories. So this patch fixes this issue too, since it has the same resolution as the incorrect inode link count issue mentioned before. This is very easy to reproduce, and the following excerpt from my test case for xfstests shows how: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create our main test file and directory. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 8K" $SCRATCH_MNT/foo | _filter_xfs_io mkdir $SCRATCH_MNT/mydir # Make sure all metadata and data are durably persisted. sync # Add a hard link to 'foo' inside our test directory and fsync only the # directory. The btrfs fsync implementation had a bug that caused the new # directory entry to be visible after the fsync log replay but, the inode # of our file remained with a link count of 1. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_2 # Add a few more links and new files. # This is just to verify nothing breaks or gives incorrect results after the # fsync log is replayed. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_3 $XFS_IO_PROG -f -c "pwrite -S 0xff 0 64K" $SCRATCH_MNT/hello | _filter_xfs_io ln $SCRATCH_MNT/hello $SCRATCH_MNT/mydir/hello_2 # Add some subdirectories and new files and links to them. This is to verify # that after fsyncing our top level directory 'mydir', all the subdirectories # and their files/links are registered in the fsync log and exist after the # fsync log is replayed. mkdir -p $SCRATCH_MNT/mydir/x/y/z ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/foo_y_link ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/z/foo_z_link touch $SCRATCH_MNT/mydir/x/y/z/qwerty # Now fsync only our top directory. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/mydir # And fsync now our new file named 'hello', just to verify later that it has # the expected content and that the previous fsync on the directory 'mydir' had # no bad influence on this fsync. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/hello # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # Verify the content of our file 'foo' remains the same as before, 8192 bytes, # all with the value 0xaa. echo "File 'foo' content after log replay:" od -t x1 $SCRATCH_MNT/foo # Remove the first name of our inode. Because of the directory fsync bug, the # inode's link count was 1 instead of 5, so removing the 'foo' name ended up # deleting the inode and the other names became stale directory entries (still # visible to applications). Attempting to remove or access the remaining # dentries pointing to that inode resulted in stale file handle errors and # made it impossible to remove the parent directories since it was impossible # for them to become empty. echo "file 'foo' link count after log replay: $(stat -c %h $SCRATCH_MNT/foo)" rm -f $SCRATCH_MNT/foo # Now verify that all files, links and directories created before fsyncing our # directory exist after the fsync log was replayed. [ -f $SCRATCH_MNT/mydir/foo_2 ] || echo "Link mydir/foo_2 is missing" [ -f $SCRATCH_MNT/mydir/foo_3 ] || echo "Link mydir/foo_3 is missing" [ -f $SCRATCH_MNT/hello ] || echo "File hello is missing" [ -f $SCRATCH_MNT/mydir/hello_2 ] || echo "Link mydir/hello_2 is missing" [ -f $SCRATCH_MNT/mydir/x/y/foo_y_link ] || \ echo "Link mydir/x/y/foo_y_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/foo_z_link ] || \ echo "Link mydir/x/y/z/foo_z_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/qwerty ] || \ echo "File mydir/x/y/z/qwerty is missing" # We expect our file here to have a size of 64Kb and all the bytes having the # value 0xff. echo "file 'hello' content after log replay:" od -t x1 $SCRATCH_MNT/hello # Now remove all files/links, under our test directory 'mydir', and verify we # can remove all the directories. rm -f $SCRATCH_MNT/mydir/x/y/z/* rmdir $SCRATCH_MNT/mydir/x/y/z rm -f $SCRATCH_MNT/mydir/x/y/* rmdir $SCRATCH_MNT/mydir/x/y rmdir $SCRATCH_MNT/mydir/x rm -f $SCRATCH_MNT/mydir/* rmdir $SCRATCH_MNT/mydir # An fsck, run by the fstests framework everytime a test finishes, also detected # the inconsistency and printed the following error message: # # root 5 inode 257 errors 2001, no inode item, link count wrong # unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref # unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref status=0 exit The expected golden output for the test is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 5 file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 Which is the output after this patch and when running the test against ext3/4, xfs, f2fs, reiserfs or nilfs2. Without this patch, the test's output is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 1 Link mydir/foo_2 is missing Link mydir/foo_3 is missing Link mydir/x/y/foo_y_link is missing Link mydir/x/y/z/foo_z_link is missing File mydir/x/y/z/qwerty is missing file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y/z': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x': No such file or directory rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_2': Stale file handle rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_3': Stale file handle rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir': Directory not empty Fsck, without this fix, also complains about the wrong link count: root 5 inode 257 errors 2001, no inode item, link count wrong unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref So fix this by logging the inodes that the dentries point to when fsyncing a directory. A test case for xfstests follows. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-03-20 20:19:46 +03:00
min_key.offset = 0;
again:
btrfs_release_path(path);
btrfs: stop copying old dir items when logging a directory When logging a directory, we go over every leaf of the subvolume tree that was changed in the current transaction and copy all its dir index keys to the log tree. That includes copying dir index keys created in past transactions. This is done mostly for simplicity, as after logging the keys we log an item that specifies the start and end ranges of the keys we logged. That item is then used during log replay to figure out which keys need to be deleted - every key in that range that we find in the subvolume tree and is not in the log tree, needs to be deleted. Now that we log only dir index keys, and not dir item keys anymore, when we remove dentries from a directory (due to unlink and rename operations), we can get entire leaves that we changed only for deleting old dir index keys, or that have few dir index keys that are new - this is due to the fact that the offset for new index keys comes from a monotonically increasing counter. We can avoid logging dir index keys from past transactions, and in order to track the deletions, only log range items (BTRFS_DIR_LOG_INDEX_KEY key type) when we find gaps between consecutive index keys. This massively reduces the amount of logged metadata when we have deleted directory entries, even if it's a small percentage of the total number of entries. The reduction comes from both less items that are logged and instead of logging many dir index items (struct btrfs_dir_item), which have a size of 30 bytes plus a file name, we typically log just a few range items (struct btrfs_dir_log_item), which take only 8 bytes each. Even if no entries were deleted from a directory and only new entries were added, we typically still get a reduction on the amount of logged metadata, because it's very likely the first leaf that got the new dir index entries also has several old dir index entries. So change the logging logic to not log dir index keys created in past transactions and log a range item for every gap it finds between each pair of consecutive index keys, to ensure deletions are tracked and replayed on log replay. This patch is part of a patchset comprised of the following patches: 1/4 btrfs: don't log unnecessary boundary keys when logging directory 2/4 btrfs: put initial index value of a directory in a constant 3/4 btrfs: stop copying old dir items when logging a directory 4/4 btrfs: stop trying to log subdirectories created in past transactions The following test was run on a branch without this patchset and on a branch with the first three patches applied: $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=1000000 NUM_FILE_DELETES=10000 MKFS_OPTIONS="-O no-holes -R free-space-tree" MOUNT_OPTIONS="-o ssd" mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync del_inc=$(( $NUM_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The test was run on a non-debug kernel (Debian's default kernel config), and the results were the following for various values of NUM_FILES and NUM_FILE_DELETES: ** before, NUM_FILES = 1 000 000, NUM_FILE_DELETES = 10 000 ** dir fsync took 585 ms after deleting 10000 files ** after, NUM_FILES = 1 000 000, NUM_FILE_DELETES = 10 000 ** dir fsync took 34 ms after deleting 10000 files (-94.2%) ** before, NUM_FILES = 100 000, NUM_FILE_DELETES = 1 000 ** dir fsync took 50 ms after deleting 1000 files ** after, NUM_FILES = 100 000, NUM_FILE_DELETES = 1 000 ** dir fsync took 7 ms after deleting 1000 files (-86.0%) ** before, NUM_FILES = 10 000, NUM_FILE_DELETES = 100 ** dir fsync took 9 ms after deleting 100 files ** after, NUM_FILES = 10 000, NUM_FILE_DELETES = 100 ** dir fsync took 5 ms after deleting 100 files (-44.4%) Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-12-15 15:20:00 +03:00
ret = btrfs_search_forward(root, &min_key, path, trans->transid);
Btrfs: fix metadata inconsistencies after directory fsync We can get into inconsistency between inodes and directory entries after fsyncing a directory. The issue is that while a directory gets the new dentries persisted in the fsync log and replayed at mount time, the link count of the inode that directory entries point to doesn't get updated, staying with an incorrect link count (smaller then the correct value). This later leads to stale file handle errors when accessing (including attempt to delete) some of the links if all the other ones are removed, which also implies impossibility to delete the parent directories, since the dentries can not be removed. Another issue is that (unlike ext3/4, xfs, f2fs, reiserfs, nilfs2), when fsyncing a directory, new files aren't logged (their metadata and dentries) nor any child directories. So this patch fixes this issue too, since it has the same resolution as the incorrect inode link count issue mentioned before. This is very easy to reproduce, and the following excerpt from my test case for xfstests shows how: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create our main test file and directory. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 8K" $SCRATCH_MNT/foo | _filter_xfs_io mkdir $SCRATCH_MNT/mydir # Make sure all metadata and data are durably persisted. sync # Add a hard link to 'foo' inside our test directory and fsync only the # directory. The btrfs fsync implementation had a bug that caused the new # directory entry to be visible after the fsync log replay but, the inode # of our file remained with a link count of 1. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_2 # Add a few more links and new files. # This is just to verify nothing breaks or gives incorrect results after the # fsync log is replayed. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_3 $XFS_IO_PROG -f -c "pwrite -S 0xff 0 64K" $SCRATCH_MNT/hello | _filter_xfs_io ln $SCRATCH_MNT/hello $SCRATCH_MNT/mydir/hello_2 # Add some subdirectories and new files and links to them. This is to verify # that after fsyncing our top level directory 'mydir', all the subdirectories # and their files/links are registered in the fsync log and exist after the # fsync log is replayed. mkdir -p $SCRATCH_MNT/mydir/x/y/z ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/foo_y_link ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/z/foo_z_link touch $SCRATCH_MNT/mydir/x/y/z/qwerty # Now fsync only our top directory. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/mydir # And fsync now our new file named 'hello', just to verify later that it has # the expected content and that the previous fsync on the directory 'mydir' had # no bad influence on this fsync. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/hello # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # Verify the content of our file 'foo' remains the same as before, 8192 bytes, # all with the value 0xaa. echo "File 'foo' content after log replay:" od -t x1 $SCRATCH_MNT/foo # Remove the first name of our inode. Because of the directory fsync bug, the # inode's link count was 1 instead of 5, so removing the 'foo' name ended up # deleting the inode and the other names became stale directory entries (still # visible to applications). Attempting to remove or access the remaining # dentries pointing to that inode resulted in stale file handle errors and # made it impossible to remove the parent directories since it was impossible # for them to become empty. echo "file 'foo' link count after log replay: $(stat -c %h $SCRATCH_MNT/foo)" rm -f $SCRATCH_MNT/foo # Now verify that all files, links and directories created before fsyncing our # directory exist after the fsync log was replayed. [ -f $SCRATCH_MNT/mydir/foo_2 ] || echo "Link mydir/foo_2 is missing" [ -f $SCRATCH_MNT/mydir/foo_3 ] || echo "Link mydir/foo_3 is missing" [ -f $SCRATCH_MNT/hello ] || echo "File hello is missing" [ -f $SCRATCH_MNT/mydir/hello_2 ] || echo "Link mydir/hello_2 is missing" [ -f $SCRATCH_MNT/mydir/x/y/foo_y_link ] || \ echo "Link mydir/x/y/foo_y_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/foo_z_link ] || \ echo "Link mydir/x/y/z/foo_z_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/qwerty ] || \ echo "File mydir/x/y/z/qwerty is missing" # We expect our file here to have a size of 64Kb and all the bytes having the # value 0xff. echo "file 'hello' content after log replay:" od -t x1 $SCRATCH_MNT/hello # Now remove all files/links, under our test directory 'mydir', and verify we # can remove all the directories. rm -f $SCRATCH_MNT/mydir/x/y/z/* rmdir $SCRATCH_MNT/mydir/x/y/z rm -f $SCRATCH_MNT/mydir/x/y/* rmdir $SCRATCH_MNT/mydir/x/y rmdir $SCRATCH_MNT/mydir/x rm -f $SCRATCH_MNT/mydir/* rmdir $SCRATCH_MNT/mydir # An fsck, run by the fstests framework everytime a test finishes, also detected # the inconsistency and printed the following error message: # # root 5 inode 257 errors 2001, no inode item, link count wrong # unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref # unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref status=0 exit The expected golden output for the test is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 5 file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 Which is the output after this patch and when running the test against ext3/4, xfs, f2fs, reiserfs or nilfs2. Without this patch, the test's output is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 1 Link mydir/foo_2 is missing Link mydir/foo_3 is missing Link mydir/x/y/foo_y_link is missing Link mydir/x/y/z/foo_z_link is missing File mydir/x/y/z/qwerty is missing file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y/z': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x': No such file or directory rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_2': Stale file handle rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_3': Stale file handle rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir': Directory not empty Fsck, without this fix, also complains about the wrong link count: root 5 inode 257 errors 2001, no inode item, link count wrong unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref So fix this by logging the inodes that the dentries point to when fsyncing a directory. A test case for xfstests follows. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-03-20 20:19:46 +03:00
if (ret < 0) {
goto next_dir_inode;
} else if (ret > 0) {
ret = 0;
goto next_dir_inode;
}
leaf = path->nodes[0];
nritems = btrfs_header_nritems(leaf);
for (i = path->slots[0]; i < nritems; i++) {
struct btrfs_dir_item *di;
struct btrfs_key di_key;
struct inode *di_inode;
struct btrfs_dir_list *new_dir_elem;
int log_mode = LOG_INODE_EXISTS;
int type;
btrfs_item_key_to_cpu(leaf, &min_key, i);
if (min_key.objectid != dir_elem->ino ||
btrfs: only copy dir index keys when logging a directory Currently, when logging a directory, we copy both dir items and dir index items from the fs/subvolume tree to the log tree. Both items have exactly the same data (same struct btrfs_dir_item), the difference lies in the key values, where a dir index key contains the index number of a directory entry while the dir item key does not, as it's used for doing fast lookups of an entry by name, while the former is used for sorting entries when listing a directory. We can exploit that and log only the dir index items, since they contain all the information needed to correctly add, replace and delete directory entries when replaying a log tree. Logging only the dir index items is also backward and forward compatible: an unpatched kernel (without this change) can correctly replay a log tree generated by a patched kernel (with this patch), and a patched kernel can correctly replay a log tree generated by an unpatched kernel. The backward compatibility is ensured because: 1) For inserting a new dentry: a dentry is only inserted when we find a new dir index key - we can only insert if we know the dir index offset, which is encoded in the dir index key's offset; 2) For deleting dentries: during log replay, before adding or replacing dentries, we first replay dentry deletions. Whenever we find a dir item key or a dir index key in the subvolume/fs tree that is not logged in a range for which the log tree is authoritative, we do the unlink of the dentry, which removes both the existing dir item key and the dir index key. Therefore logging just dir index keys is enough to ensure dentry deletions are correctly replayed; 3) For dentry replacements: they work when we log only dir index keys and this is mostly due to a combination of 1) and 2). If we replace a dentry with name "foobar" to point from inode A to inode B, then we know the dir index key for the new dentry is different from the old one, as it has an index number (key offset) larger than the old one. This results in replaying a deletion, through replay_dir_deletes(), that causes the old dentry to be removed, both the dir item key and the dir index key, as mentioned at 2). Then when processing the new dir index key, we add the new dentry, adding both a new dir item key and a new index key pointing to inode B, as stated in 1). The forward compatibility, the ability for a patched kernel to replay a log created by an older, unpatched kernel, comes from the changes required for making sure we are able to replay a log that only contains dir index keys - we simply ignore every dir item key we find. So modify directory logging to log only dir index items, and modify the log replay process to ignore dir item keys, from log trees created by an unpatched kernel, and process only with dir index keys. This reduces the amount of logged metadata by about half, and therefore the time spent logging or fsyncing large directories (less CPU time and less IO). The following test script was used to measure this change: #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_NEW_FILES=1000000 NUM_FILE_DELETES=10000 mkfs.btrfs -f $DEV mount -o ssd $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_NEW_FILES; i++)); do echo -n > $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after adding $NUM_NEW_FILES files" # sync to force transaction commit and wipeout the log. sync del_inc=$(( $NUM_NEW_FILES / $NUM_FILE_DELETES )) for ((i = 1; i <= $NUM_NEW_FILES; i += $del_inc)); do rm -f $MNT/testdir/file_$i done start=$(date +%s%N) xfs_io -c "fsync" $MNT/testdir end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "dir fsync took $dur ms after deleting $NUM_FILE_DELETES files" echo umount $MNT The tests were run on a physical machine, with a non-debug kernel (Debian's default kernel config), for different values of $NUM_NEW_FILES and $NUM_FILE_DELETES, and the results were the following: ** Before patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 8412 ms after adding 1000000 files dir fsync took 500 ms after deleting 10000 files ** After patch, NUM_NEW_FILES = 1 000 000, NUM_DELETE_FILES = 10 000 ** dir fsync took 4252 ms after adding 1000000 files (-49.5%) dir fsync took 269 ms after deleting 10000 files (-46.2%) ** Before patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 745 ms after adding 100000 files dir fsync took 59 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 100 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 404 ms after adding 100000 files (-45.8%) dir fsync took 31 ms after deleting 1000 files (-47.5%) ** Before patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 67 ms after adding 10000 files dir fsync took 9 ms after deleting 1000 files ** After patch, NUM_NEW_FILES = 10 000, NUM_DELETE_FILES = 1 000 ** dir fsync took 36 ms after adding 10000 files (-46.3%) dir fsync took 5 ms after deleting 1000 files (-44.4%) ** Before patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 9 ms after adding 1000 files dir fsync took 4 ms after deleting 100 files ** After patch, NUM_NEW_FILES = 1 000, NUM_DELETE_FILES = 100 ** dir fsync took 7 ms after adding 1000 files (-22.2%) dir fsync took 3 ms after deleting 100 files (-25.0%) Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-10-25 19:31:53 +03:00
min_key.type != BTRFS_DIR_INDEX_KEY)
Btrfs: fix metadata inconsistencies after directory fsync We can get into inconsistency between inodes and directory entries after fsyncing a directory. The issue is that while a directory gets the new dentries persisted in the fsync log and replayed at mount time, the link count of the inode that directory entries point to doesn't get updated, staying with an incorrect link count (smaller then the correct value). This later leads to stale file handle errors when accessing (including attempt to delete) some of the links if all the other ones are removed, which also implies impossibility to delete the parent directories, since the dentries can not be removed. Another issue is that (unlike ext3/4, xfs, f2fs, reiserfs, nilfs2), when fsyncing a directory, new files aren't logged (their metadata and dentries) nor any child directories. So this patch fixes this issue too, since it has the same resolution as the incorrect inode link count issue mentioned before. This is very easy to reproduce, and the following excerpt from my test case for xfstests shows how: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create our main test file and directory. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 8K" $SCRATCH_MNT/foo | _filter_xfs_io mkdir $SCRATCH_MNT/mydir # Make sure all metadata and data are durably persisted. sync # Add a hard link to 'foo' inside our test directory and fsync only the # directory. The btrfs fsync implementation had a bug that caused the new # directory entry to be visible after the fsync log replay but, the inode # of our file remained with a link count of 1. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_2 # Add a few more links and new files. # This is just to verify nothing breaks or gives incorrect results after the # fsync log is replayed. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_3 $XFS_IO_PROG -f -c "pwrite -S 0xff 0 64K" $SCRATCH_MNT/hello | _filter_xfs_io ln $SCRATCH_MNT/hello $SCRATCH_MNT/mydir/hello_2 # Add some subdirectories and new files and links to them. This is to verify # that after fsyncing our top level directory 'mydir', all the subdirectories # and their files/links are registered in the fsync log and exist after the # fsync log is replayed. mkdir -p $SCRATCH_MNT/mydir/x/y/z ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/foo_y_link ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/z/foo_z_link touch $SCRATCH_MNT/mydir/x/y/z/qwerty # Now fsync only our top directory. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/mydir # And fsync now our new file named 'hello', just to verify later that it has # the expected content and that the previous fsync on the directory 'mydir' had # no bad influence on this fsync. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/hello # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # Verify the content of our file 'foo' remains the same as before, 8192 bytes, # all with the value 0xaa. echo "File 'foo' content after log replay:" od -t x1 $SCRATCH_MNT/foo # Remove the first name of our inode. Because of the directory fsync bug, the # inode's link count was 1 instead of 5, so removing the 'foo' name ended up # deleting the inode and the other names became stale directory entries (still # visible to applications). Attempting to remove or access the remaining # dentries pointing to that inode resulted in stale file handle errors and # made it impossible to remove the parent directories since it was impossible # for them to become empty. echo "file 'foo' link count after log replay: $(stat -c %h $SCRATCH_MNT/foo)" rm -f $SCRATCH_MNT/foo # Now verify that all files, links and directories created before fsyncing our # directory exist after the fsync log was replayed. [ -f $SCRATCH_MNT/mydir/foo_2 ] || echo "Link mydir/foo_2 is missing" [ -f $SCRATCH_MNT/mydir/foo_3 ] || echo "Link mydir/foo_3 is missing" [ -f $SCRATCH_MNT/hello ] || echo "File hello is missing" [ -f $SCRATCH_MNT/mydir/hello_2 ] || echo "Link mydir/hello_2 is missing" [ -f $SCRATCH_MNT/mydir/x/y/foo_y_link ] || \ echo "Link mydir/x/y/foo_y_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/foo_z_link ] || \ echo "Link mydir/x/y/z/foo_z_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/qwerty ] || \ echo "File mydir/x/y/z/qwerty is missing" # We expect our file here to have a size of 64Kb and all the bytes having the # value 0xff. echo "file 'hello' content after log replay:" od -t x1 $SCRATCH_MNT/hello # Now remove all files/links, under our test directory 'mydir', and verify we # can remove all the directories. rm -f $SCRATCH_MNT/mydir/x/y/z/* rmdir $SCRATCH_MNT/mydir/x/y/z rm -f $SCRATCH_MNT/mydir/x/y/* rmdir $SCRATCH_MNT/mydir/x/y rmdir $SCRATCH_MNT/mydir/x rm -f $SCRATCH_MNT/mydir/* rmdir $SCRATCH_MNT/mydir # An fsck, run by the fstests framework everytime a test finishes, also detected # the inconsistency and printed the following error message: # # root 5 inode 257 errors 2001, no inode item, link count wrong # unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref # unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref status=0 exit The expected golden output for the test is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 5 file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 Which is the output after this patch and when running the test against ext3/4, xfs, f2fs, reiserfs or nilfs2. Without this patch, the test's output is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 1 Link mydir/foo_2 is missing Link mydir/foo_3 is missing Link mydir/x/y/foo_y_link is missing Link mydir/x/y/z/foo_z_link is missing File mydir/x/y/z/qwerty is missing file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y/z': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x': No such file or directory rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_2': Stale file handle rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_3': Stale file handle rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir': Directory not empty Fsck, without this fix, also complains about the wrong link count: root 5 inode 257 errors 2001, no inode item, link count wrong unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref So fix this by logging the inodes that the dentries point to when fsyncing a directory. A test case for xfstests follows. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-03-20 20:19:46 +03:00
goto next_dir_inode;
di = btrfs_item_ptr(leaf, i, struct btrfs_dir_item);
type = btrfs_dir_type(leaf, di);
btrfs: stop trying to log subdirectories created in past transactions When logging a directory we are trying to log subdirectories that were changed in the current transaction and created in a past transaction. This type of behaviour was introduced by commit 2f2ff0ee5e4303 ("Btrfs: fix metadata inconsistencies after directory fsync"), to fix some metadata inconsistencies that in the meanwhile no longer need this behaviour due to numerous other changes that happened throughout the years. This behaviour, besides not needed anymore, it's also undesirable because: 1) It's not reliable because it's only triggered for the directories of dentries (dir items) that happen to be present on a leaf that was changed in the current transaction. If a dentry that points to a directory resides on a leaf that was not changed in the current transaction, then it's not logged, as at log_dir_items() and log_new_dir_dentries() we use btrfs_search_forward(); 2) It's not required by posix or any standard, it's undefined territory. The only way to guarantee a subdirectory is logged, it to explicitly fsync it; Making the behaviour guaranteed would require scanning all directory items, check which point to a directory, and then fsync each subdirectory which was modified in the current transaction. This could be very expensive for large directories with many subdirectories and/or large subdirectories. So remove that obsolete logic. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-12-15 15:20:01 +03:00
if (btrfs_dir_transid(leaf, di) < trans->transid)
Btrfs: fix metadata inconsistencies after directory fsync We can get into inconsistency between inodes and directory entries after fsyncing a directory. The issue is that while a directory gets the new dentries persisted in the fsync log and replayed at mount time, the link count of the inode that directory entries point to doesn't get updated, staying with an incorrect link count (smaller then the correct value). This later leads to stale file handle errors when accessing (including attempt to delete) some of the links if all the other ones are removed, which also implies impossibility to delete the parent directories, since the dentries can not be removed. Another issue is that (unlike ext3/4, xfs, f2fs, reiserfs, nilfs2), when fsyncing a directory, new files aren't logged (their metadata and dentries) nor any child directories. So this patch fixes this issue too, since it has the same resolution as the incorrect inode link count issue mentioned before. This is very easy to reproduce, and the following excerpt from my test case for xfstests shows how: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create our main test file and directory. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 8K" $SCRATCH_MNT/foo | _filter_xfs_io mkdir $SCRATCH_MNT/mydir # Make sure all metadata and data are durably persisted. sync # Add a hard link to 'foo' inside our test directory and fsync only the # directory. The btrfs fsync implementation had a bug that caused the new # directory entry to be visible after the fsync log replay but, the inode # of our file remained with a link count of 1. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_2 # Add a few more links and new files. # This is just to verify nothing breaks or gives incorrect results after the # fsync log is replayed. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_3 $XFS_IO_PROG -f -c "pwrite -S 0xff 0 64K" $SCRATCH_MNT/hello | _filter_xfs_io ln $SCRATCH_MNT/hello $SCRATCH_MNT/mydir/hello_2 # Add some subdirectories and new files and links to them. This is to verify # that after fsyncing our top level directory 'mydir', all the subdirectories # and their files/links are registered in the fsync log and exist after the # fsync log is replayed. mkdir -p $SCRATCH_MNT/mydir/x/y/z ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/foo_y_link ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/z/foo_z_link touch $SCRATCH_MNT/mydir/x/y/z/qwerty # Now fsync only our top directory. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/mydir # And fsync now our new file named 'hello', just to verify later that it has # the expected content and that the previous fsync on the directory 'mydir' had # no bad influence on this fsync. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/hello # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # Verify the content of our file 'foo' remains the same as before, 8192 bytes, # all with the value 0xaa. echo "File 'foo' content after log replay:" od -t x1 $SCRATCH_MNT/foo # Remove the first name of our inode. Because of the directory fsync bug, the # inode's link count was 1 instead of 5, so removing the 'foo' name ended up # deleting the inode and the other names became stale directory entries (still # visible to applications). Attempting to remove or access the remaining # dentries pointing to that inode resulted in stale file handle errors and # made it impossible to remove the parent directories since it was impossible # for them to become empty. echo "file 'foo' link count after log replay: $(stat -c %h $SCRATCH_MNT/foo)" rm -f $SCRATCH_MNT/foo # Now verify that all files, links and directories created before fsyncing our # directory exist after the fsync log was replayed. [ -f $SCRATCH_MNT/mydir/foo_2 ] || echo "Link mydir/foo_2 is missing" [ -f $SCRATCH_MNT/mydir/foo_3 ] || echo "Link mydir/foo_3 is missing" [ -f $SCRATCH_MNT/hello ] || echo "File hello is missing" [ -f $SCRATCH_MNT/mydir/hello_2 ] || echo "Link mydir/hello_2 is missing" [ -f $SCRATCH_MNT/mydir/x/y/foo_y_link ] || \ echo "Link mydir/x/y/foo_y_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/foo_z_link ] || \ echo "Link mydir/x/y/z/foo_z_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/qwerty ] || \ echo "File mydir/x/y/z/qwerty is missing" # We expect our file here to have a size of 64Kb and all the bytes having the # value 0xff. echo "file 'hello' content after log replay:" od -t x1 $SCRATCH_MNT/hello # Now remove all files/links, under our test directory 'mydir', and verify we # can remove all the directories. rm -f $SCRATCH_MNT/mydir/x/y/z/* rmdir $SCRATCH_MNT/mydir/x/y/z rm -f $SCRATCH_MNT/mydir/x/y/* rmdir $SCRATCH_MNT/mydir/x/y rmdir $SCRATCH_MNT/mydir/x rm -f $SCRATCH_MNT/mydir/* rmdir $SCRATCH_MNT/mydir # An fsck, run by the fstests framework everytime a test finishes, also detected # the inconsistency and printed the following error message: # # root 5 inode 257 errors 2001, no inode item, link count wrong # unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref # unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref status=0 exit The expected golden output for the test is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 5 file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 Which is the output after this patch and when running the test against ext3/4, xfs, f2fs, reiserfs or nilfs2. Without this patch, the test's output is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 1 Link mydir/foo_2 is missing Link mydir/foo_3 is missing Link mydir/x/y/foo_y_link is missing Link mydir/x/y/z/foo_z_link is missing File mydir/x/y/z/qwerty is missing file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y/z': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x': No such file or directory rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_2': Stale file handle rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_3': Stale file handle rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir': Directory not empty Fsck, without this fix, also complains about the wrong link count: root 5 inode 257 errors 2001, no inode item, link count wrong unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref So fix this by logging the inodes that the dentries point to when fsyncing a directory. A test case for xfstests follows. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-03-20 20:19:46 +03:00
continue;
btrfs_dir_item_key_to_cpu(leaf, di, &di_key);
if (di_key.type == BTRFS_ROOT_ITEM_KEY)
continue;
Btrfs: fix deadlock caused by fsync when logging directory entries While logging new directory entries, at tree-log.c:log_new_dir_dentries(), after we call btrfs_search_forward() we get a leaf with a read lock on it, and without unlocking that leaf we can end up calling btrfs_iget() to get an inode pointer. The later (btrfs_iget()) can end up doing a read-only search on the same tree again, if the inode is not in memory already, which ends up causing a deadlock if some other task in the meanwhile started a write search on the tree and is attempting to write lock the same leaf that btrfs_search_forward() locked while holding write locks on upper levels of the tree blocking the read search from btrfs_iget(). In this scenario we get a deadlock. So fix this by releasing the search path before calling btrfs_iget() at tree-log.c:log_new_dir_dentries(). Example trace of such deadlock: [ 4077.478852] kworker/u24:10 D ffff88107fc90640 0 14431 2 0x00000000 [ 4077.486752] Workqueue: btrfs-endio-write btrfs_endio_write_helper [btrfs] [ 4077.494346] ffff880ffa56bad0 0000000000000046 0000000000009000 ffff880ffa56bfd8 [ 4077.502629] ffff880ffa56bfd8 ffff881016ce21c0 ffffffffa06ecb26 ffff88101a5d6138 [ 4077.510915] ffff880ebb5173b0 ffff880ffa56baf8 ffff880ebb517410 ffff881016ce21c0 [ 4077.519202] Call Trace: [ 4077.528752] [<ffffffffa06ed5ed>] ? btrfs_tree_lock+0xdd/0x2f0 [btrfs] [ 4077.536049] [<ffffffff81053680>] ? wake_up_atomic_t+0x30/0x30 [ 4077.542574] [<ffffffffa068cc1f>] ? btrfs_search_slot+0x79f/0xb10 [btrfs] [ 4077.550171] [<ffffffffa06a5073>] ? btrfs_lookup_file_extent+0x33/0x40 [btrfs] [ 4077.558252] [<ffffffffa06c600b>] ? __btrfs_drop_extents+0x13b/0xdf0 [btrfs] [ 4077.566140] [<ffffffffa06fc9e2>] ? add_delayed_data_ref+0xe2/0x150 [btrfs] [ 4077.573928] [<ffffffffa06fd629>] ? btrfs_add_delayed_data_ref+0x149/0x1d0 [btrfs] [ 4077.582399] [<ffffffffa06cf3c0>] ? __set_extent_bit+0x4c0/0x5c0 [btrfs] [ 4077.589896] [<ffffffffa06b4a64>] ? insert_reserved_file_extent.constprop.75+0xa4/0x320 [btrfs] [ 4077.599632] [<ffffffffa06b206d>] ? start_transaction+0x8d/0x470 [btrfs] [ 4077.607134] [<ffffffffa06bab57>] ? btrfs_finish_ordered_io+0x2e7/0x600 [btrfs] [ 4077.615329] [<ffffffff8104cbc2>] ? process_one_work+0x142/0x3d0 [ 4077.622043] [<ffffffff8104d729>] ? worker_thread+0x109/0x3b0 [ 4077.628459] [<ffffffff8104d620>] ? manage_workers.isra.26+0x270/0x270 [ 4077.635759] [<ffffffff81052b0f>] ? kthread+0xaf/0xc0 [ 4077.641404] [<ffffffff81052a60>] ? kthread_create_on_node+0x110/0x110 [ 4077.648696] [<ffffffff814a9ac8>] ? ret_from_fork+0x58/0x90 [ 4077.654926] [<ffffffff81052a60>] ? kthread_create_on_node+0x110/0x110 [ 4078.358087] kworker/u24:15 D ffff88107fcd0640 0 14436 2 0x00000000 [ 4078.365981] Workqueue: btrfs-endio-write btrfs_endio_write_helper [btrfs] [ 4078.373574] ffff880ffa57fad0 0000000000000046 0000000000009000 ffff880ffa57ffd8 [ 4078.381864] ffff880ffa57ffd8 ffff88103004d0a0 ffffffffa06ecb26 ffff88101a5d6138 [ 4078.390163] ffff880fbeffc298 ffff880ffa57faf8 ffff880fbeffc2f8 ffff88103004d0a0 [ 4078.398466] Call Trace: [ 4078.408019] [<ffffffffa06ed5ed>] ? btrfs_tree_lock+0xdd/0x2f0 [btrfs] [ 4078.415322] [<ffffffff81053680>] ? wake_up_atomic_t+0x30/0x30 [ 4078.421844] [<ffffffffa068cc1f>] ? btrfs_search_slot+0x79f/0xb10 [btrfs] [ 4078.429438] [<ffffffffa06a5073>] ? btrfs_lookup_file_extent+0x33/0x40 [btrfs] [ 4078.437518] [<ffffffffa06c600b>] ? __btrfs_drop_extents+0x13b/0xdf0 [btrfs] [ 4078.445404] [<ffffffffa06fc9e2>] ? add_delayed_data_ref+0xe2/0x150 [btrfs] [ 4078.453194] [<ffffffffa06fd629>] ? btrfs_add_delayed_data_ref+0x149/0x1d0 [btrfs] [ 4078.461663] [<ffffffffa06cf3c0>] ? __set_extent_bit+0x4c0/0x5c0 [btrfs] [ 4078.469161] [<ffffffffa06b4a64>] ? insert_reserved_file_extent.constprop.75+0xa4/0x320 [btrfs] [ 4078.478893] [<ffffffffa06b206d>] ? start_transaction+0x8d/0x470 [btrfs] [ 4078.486388] [<ffffffffa06bab57>] ? btrfs_finish_ordered_io+0x2e7/0x600 [btrfs] [ 4078.494561] [<ffffffff8104cbc2>] ? process_one_work+0x142/0x3d0 [ 4078.501278] [<ffffffff8104a507>] ? pwq_activate_delayed_work+0x27/0x40 [ 4078.508673] [<ffffffff8104d729>] ? worker_thread+0x109/0x3b0 [ 4078.515098] [<ffffffff8104d620>] ? manage_workers.isra.26+0x270/0x270 [ 4078.522396] [<ffffffff81052b0f>] ? kthread+0xaf/0xc0 [ 4078.528032] [<ffffffff81052a60>] ? kthread_create_on_node+0x110/0x110 [ 4078.535325] [<ffffffff814a9ac8>] ? ret_from_fork+0x58/0x90 [ 4078.541552] [<ffffffff81052a60>] ? kthread_create_on_node+0x110/0x110 [ 4079.355824] user-space-program D ffff88107fd30640 0 32020 1 0x00000000 [ 4079.363716] ffff880eae8eba10 0000000000000086 0000000000009000 ffff880eae8ebfd8 [ 4079.372003] ffff880eae8ebfd8 ffff881016c162c0 ffffffffa06ecb26 ffff88101a5d6138 [ 4079.380294] ffff880fbed4b4c8 ffff880eae8eba38 ffff880fbed4b528 ffff881016c162c0 [ 4079.388586] Call Trace: [ 4079.398134] [<ffffffffa06ed595>] ? btrfs_tree_lock+0x85/0x2f0 [btrfs] [ 4079.405431] [<ffffffff81053680>] ? wake_up_atomic_t+0x30/0x30 [ 4079.411955] [<ffffffffa06876fb>] ? btrfs_lock_root_node+0x2b/0x40 [btrfs] [ 4079.419644] [<ffffffffa068ce83>] ? btrfs_search_slot+0xa03/0xb10 [btrfs] [ 4079.427237] [<ffffffffa06aba52>] ? btrfs_buffer_uptodate+0x52/0x70 [btrfs] [ 4079.435041] [<ffffffffa0689b60>] ? generic_bin_search.constprop.38+0x80/0x190 [btrfs] [ 4079.443897] [<ffffffffa068ea44>] ? btrfs_insert_empty_items+0x74/0xd0 [btrfs] [ 4079.451975] [<ffffffffa072c443>] ? copy_items+0x128/0x850 [btrfs] [ 4079.458890] [<ffffffffa072da10>] ? btrfs_log_inode+0x629/0xbf3 [btrfs] [ 4079.466292] [<ffffffffa06f34a1>] ? btrfs_log_inode_parent+0xc61/0xf30 [btrfs] [ 4079.474373] [<ffffffffa06f45a9>] ? btrfs_log_dentry_safe+0x59/0x80 [btrfs] [ 4079.482161] [<ffffffffa06c298d>] ? btrfs_sync_file+0x20d/0x330 [btrfs] [ 4079.489558] [<ffffffff8112777c>] ? do_fsync+0x4c/0x80 [ 4079.495300] [<ffffffff81127a0a>] ? SyS_fdatasync+0xa/0x10 [ 4079.501422] [<ffffffff814a9b72>] ? system_call_fastpath+0x16/0x1b [ 4079.508334] user-space-program D ffff88107fc30640 0 32021 1 0x00000004 [ 4079.516226] ffff880eae8efbf8 0000000000000086 0000000000009000 ffff880eae8effd8 [ 4079.524513] ffff880eae8effd8 ffff881030279610 ffffffffa06ecb26 ffff88101a5d6138 [ 4079.532802] ffff880ebb671d88 ffff880eae8efc20 ffff880ebb671de8 ffff881030279610 [ 4079.541092] Call Trace: [ 4079.550642] [<ffffffffa06ed595>] ? btrfs_tree_lock+0x85/0x2f0 [btrfs] [ 4079.557941] [<ffffffff81053680>] ? wake_up_atomic_t+0x30/0x30 [ 4079.564463] [<ffffffffa068cc1f>] ? btrfs_search_slot+0x79f/0xb10 [btrfs] [ 4079.572058] [<ffffffffa06bb7d8>] ? btrfs_truncate_inode_items+0x168/0xb90 [btrfs] [ 4079.580526] [<ffffffffa06b04be>] ? join_transaction.isra.15+0x1e/0x3a0 [btrfs] [ 4079.588701] [<ffffffffa06b206d>] ? start_transaction+0x8d/0x470 [btrfs] [ 4079.596196] [<ffffffffa0690ac6>] ? block_rsv_add_bytes+0x16/0x50 [btrfs] [ 4079.603789] [<ffffffffa06bc2e9>] ? btrfs_truncate+0xe9/0x2e0 [btrfs] [ 4079.610994] [<ffffffffa06bd00b>] ? btrfs_setattr+0x30b/0x410 [btrfs] [ 4079.618197] [<ffffffff81117c1c>] ? notify_change+0x1dc/0x680 [ 4079.624625] [<ffffffff8123c8a4>] ? aa_path_perm+0xd4/0x160 [ 4079.630854] [<ffffffff810f4fcb>] ? do_truncate+0x5b/0x90 [ 4079.636889] [<ffffffff810f59fa>] ? do_sys_ftruncate.constprop.15+0x10a/0x160 [ 4079.644869] [<ffffffff8110d87b>] ? SyS_fcntl+0x5b/0x570 [ 4079.650805] [<ffffffff814a9b72>] ? system_call_fastpath+0x16/0x1b [ 4080.410607] user-space-program D ffff88107fc70640 0 32028 12639 0x00000004 [ 4080.418489] ffff880eaeccbbe0 0000000000000086 0000000000009000 ffff880eaeccbfd8 [ 4080.426778] ffff880eaeccbfd8 ffff880f317ef1e0 ffffffffa06ecb26 ffff88101a5d6138 [ 4080.435067] ffff880ef7e93928 ffff880f317ef1e0 ffff880eaeccbc08 ffff880f317ef1e0 [ 4080.443353] Call Trace: [ 4080.452920] [<ffffffffa06ed15d>] ? btrfs_tree_read_lock+0xdd/0x190 [btrfs] [ 4080.460703] [<ffffffff81053680>] ? wake_up_atomic_t+0x30/0x30 [ 4080.467225] [<ffffffffa06876bb>] ? btrfs_read_lock_root_node+0x2b/0x40 [btrfs] [ 4080.475400] [<ffffffffa068cc81>] ? btrfs_search_slot+0x801/0xb10 [btrfs] [ 4080.482994] [<ffffffffa06b2df0>] ? btrfs_clean_one_deleted_snapshot+0xe0/0xe0 [btrfs] [ 4080.491857] [<ffffffffa06a70a6>] ? btrfs_lookup_inode+0x26/0x90 [btrfs] [ 4080.499353] [<ffffffff810ec42f>] ? kmem_cache_alloc+0xaf/0xc0 [ 4080.505879] [<ffffffffa06bd905>] ? btrfs_iget+0xd5/0x5d0 [btrfs] [ 4080.512696] [<ffffffffa06caf04>] ? btrfs_get_token_64+0x104/0x120 [btrfs] [ 4080.520387] [<ffffffffa06f341f>] ? btrfs_log_inode_parent+0xbdf/0xf30 [btrfs] [ 4080.528469] [<ffffffffa06f45a9>] ? btrfs_log_dentry_safe+0x59/0x80 [btrfs] [ 4080.536258] [<ffffffffa06c298d>] ? btrfs_sync_file+0x20d/0x330 [btrfs] [ 4080.543657] [<ffffffff8112777c>] ? do_fsync+0x4c/0x80 [ 4080.549399] [<ffffffff81127a0a>] ? SyS_fdatasync+0xa/0x10 [ 4080.555534] [<ffffffff814a9b72>] ? system_call_fastpath+0x16/0x1b Signed-off-by: Robbie Ko <robbieko@synology.com> Reviewed-by: Filipe Manana <fdmanana@suse.com> Fixes: 2f2ff0ee5e43 (Btrfs: fix metadata inconsistencies after directory fsync) Cc: stable@vger.kernel.org # 4.1+ Signed-off-by: Filipe Manana <fdmanana@suse.com> [Modified changelog for clarity and correctness]
2016-10-28 05:48:26 +03:00
btrfs_release_path(path);
di_inode = btrfs_iget(fs_info->sb, di_key.objectid, root);
Btrfs: fix metadata inconsistencies after directory fsync We can get into inconsistency between inodes and directory entries after fsyncing a directory. The issue is that while a directory gets the new dentries persisted in the fsync log and replayed at mount time, the link count of the inode that directory entries point to doesn't get updated, staying with an incorrect link count (smaller then the correct value). This later leads to stale file handle errors when accessing (including attempt to delete) some of the links if all the other ones are removed, which also implies impossibility to delete the parent directories, since the dentries can not be removed. Another issue is that (unlike ext3/4, xfs, f2fs, reiserfs, nilfs2), when fsyncing a directory, new files aren't logged (their metadata and dentries) nor any child directories. So this patch fixes this issue too, since it has the same resolution as the incorrect inode link count issue mentioned before. This is very easy to reproduce, and the following excerpt from my test case for xfstests shows how: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create our main test file and directory. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 8K" $SCRATCH_MNT/foo | _filter_xfs_io mkdir $SCRATCH_MNT/mydir # Make sure all metadata and data are durably persisted. sync # Add a hard link to 'foo' inside our test directory and fsync only the # directory. The btrfs fsync implementation had a bug that caused the new # directory entry to be visible after the fsync log replay but, the inode # of our file remained with a link count of 1. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_2 # Add a few more links and new files. # This is just to verify nothing breaks or gives incorrect results after the # fsync log is replayed. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_3 $XFS_IO_PROG -f -c "pwrite -S 0xff 0 64K" $SCRATCH_MNT/hello | _filter_xfs_io ln $SCRATCH_MNT/hello $SCRATCH_MNT/mydir/hello_2 # Add some subdirectories and new files and links to them. This is to verify # that after fsyncing our top level directory 'mydir', all the subdirectories # and their files/links are registered in the fsync log and exist after the # fsync log is replayed. mkdir -p $SCRATCH_MNT/mydir/x/y/z ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/foo_y_link ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/z/foo_z_link touch $SCRATCH_MNT/mydir/x/y/z/qwerty # Now fsync only our top directory. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/mydir # And fsync now our new file named 'hello', just to verify later that it has # the expected content and that the previous fsync on the directory 'mydir' had # no bad influence on this fsync. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/hello # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # Verify the content of our file 'foo' remains the same as before, 8192 bytes, # all with the value 0xaa. echo "File 'foo' content after log replay:" od -t x1 $SCRATCH_MNT/foo # Remove the first name of our inode. Because of the directory fsync bug, the # inode's link count was 1 instead of 5, so removing the 'foo' name ended up # deleting the inode and the other names became stale directory entries (still # visible to applications). Attempting to remove or access the remaining # dentries pointing to that inode resulted in stale file handle errors and # made it impossible to remove the parent directories since it was impossible # for them to become empty. echo "file 'foo' link count after log replay: $(stat -c %h $SCRATCH_MNT/foo)" rm -f $SCRATCH_MNT/foo # Now verify that all files, links and directories created before fsyncing our # directory exist after the fsync log was replayed. [ -f $SCRATCH_MNT/mydir/foo_2 ] || echo "Link mydir/foo_2 is missing" [ -f $SCRATCH_MNT/mydir/foo_3 ] || echo "Link mydir/foo_3 is missing" [ -f $SCRATCH_MNT/hello ] || echo "File hello is missing" [ -f $SCRATCH_MNT/mydir/hello_2 ] || echo "Link mydir/hello_2 is missing" [ -f $SCRATCH_MNT/mydir/x/y/foo_y_link ] || \ echo "Link mydir/x/y/foo_y_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/foo_z_link ] || \ echo "Link mydir/x/y/z/foo_z_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/qwerty ] || \ echo "File mydir/x/y/z/qwerty is missing" # We expect our file here to have a size of 64Kb and all the bytes having the # value 0xff. echo "file 'hello' content after log replay:" od -t x1 $SCRATCH_MNT/hello # Now remove all files/links, under our test directory 'mydir', and verify we # can remove all the directories. rm -f $SCRATCH_MNT/mydir/x/y/z/* rmdir $SCRATCH_MNT/mydir/x/y/z rm -f $SCRATCH_MNT/mydir/x/y/* rmdir $SCRATCH_MNT/mydir/x/y rmdir $SCRATCH_MNT/mydir/x rm -f $SCRATCH_MNT/mydir/* rmdir $SCRATCH_MNT/mydir # An fsck, run by the fstests framework everytime a test finishes, also detected # the inconsistency and printed the following error message: # # root 5 inode 257 errors 2001, no inode item, link count wrong # unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref # unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref status=0 exit The expected golden output for the test is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 5 file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 Which is the output after this patch and when running the test against ext3/4, xfs, f2fs, reiserfs or nilfs2. Without this patch, the test's output is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 1 Link mydir/foo_2 is missing Link mydir/foo_3 is missing Link mydir/x/y/foo_y_link is missing Link mydir/x/y/z/foo_z_link is missing File mydir/x/y/z/qwerty is missing file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y/z': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x': No such file or directory rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_2': Stale file handle rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_3': Stale file handle rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir': Directory not empty Fsck, without this fix, also complains about the wrong link count: root 5 inode 257 errors 2001, no inode item, link count wrong unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref So fix this by logging the inodes that the dentries point to when fsyncing a directory. A test case for xfstests follows. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-03-20 20:19:46 +03:00
if (IS_ERR(di_inode)) {
ret = PTR_ERR(di_inode);
goto next_dir_inode;
}
btrfs: skip logging inodes already logged when logging new entries When logging new directory entries of a directory, we log the inodes of new dentries and the inodes of dentries pointing to directories that may have been created in past transactions. For the case of directories we log in full mode, which can be particularly expensive for large directories. We do use btrfs_inode_in_log() to skip already logged inodes, however for that helper to return true, it requires that the log transaction used to log the inode to be already committed. This means that when we have more than one task using the same log transaction we can end up logging an inode multiple times, which is a waste of time and not necessary since the log will be committed by one of the tasks and the others will wait for the log transaction to be committed before returning to user space. So simply replace the use of btrfs_inode_in_log() with the new helper function need_log_inode(), introduced in a previous commit. This patch is part of a patchset comprised of the following patches: btrfs: remove unnecessary directory inode item update when deleting dir entry btrfs: stop setting nbytes when filling inode item for logging btrfs: avoid logging new ancestor inodes when logging new inode btrfs: skip logging directories already logged when logging all parents btrfs: skip logging inodes already logged when logging new entries btrfs: remove unnecessary check_parent_dirs_for_sync() btrfs: make concurrent fsyncs wait less when waiting for a transaction commit Performance results, after applying all patches, are mentioned in the change log of the last patch. Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-01-27 13:34:58 +03:00
if (!need_log_inode(trans, BTRFS_I(di_inode))) {
Btrfs: fix assertion failure during fsync and use of stale transaction Sometimes when fsync'ing a file we need to log that other inodes exist and when we need to do that we acquire a reference on the inodes and then drop that reference using iput() after logging them. That generally is not a problem except if we end up doing the final iput() (dropping the last reference) on the inode and that inode has a link count of 0, which can happen in a very short time window if the logging path gets a reference on the inode while it's being unlinked. In that case we end up getting the eviction callback, btrfs_evict_inode(), invoked through the iput() call chain which needs to drop all of the inode's items from its subvolume btree, and in order to do that, it needs to join a transaction at the helper function evict_refill_and_join(). However because the task previously started a transaction at the fsync handler, btrfs_sync_file(), it has current->journal_info already pointing to a transaction handle and therefore evict_refill_and_join() will get that transaction handle from btrfs_join_transaction(). From this point on, two different problems can happen: 1) evict_refill_and_join() will often change the transaction handle's block reserve (->block_rsv) and set its ->bytes_reserved field to a value greater than 0. If evict_refill_and_join() never commits the transaction, the eviction handler ends up decreasing the reference count (->use_count) of the transaction handle through the call to btrfs_end_transaction(), and after that point we have a transaction handle with a NULL ->block_rsv (which is the value prior to the transaction join from evict_refill_and_join()) and a ->bytes_reserved value greater than 0. If after the eviction/iput completes the inode logging path hits an error or it decides that it must fallback to a transaction commit, the btrfs fsync handle, btrfs_sync_file(), gets a non-zero value from btrfs_log_dentry_safe(), and because of that non-zero value it tries to commit the transaction using a handle with a NULL ->block_rsv and a non-zero ->bytes_reserved value. This makes the transaction commit hit an assertion failure at btrfs_trans_release_metadata() because ->bytes_reserved is not zero but the ->block_rsv is NULL. The produced stack trace for that is like the following: [192922.917158] assertion failed: !trans->bytes_reserved, file: fs/btrfs/transaction.c, line: 816 [192922.917553] ------------[ cut here ]------------ [192922.917922] kernel BUG at fs/btrfs/ctree.h:3532! [192922.918310] invalid opcode: 0000 [#1] SMP DEBUG_PAGEALLOC PTI [192922.918666] CPU: 2 PID: 883 Comm: fsstress Tainted: G W 5.1.4-btrfs-next-47 #1 [192922.919035] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.11.2-0-gf9626ccb91-prebuilt.qemu-project.org 04/01/2014 [192922.919801] RIP: 0010:assfail.constprop.25+0x18/0x1a [btrfs] (...) [192922.920925] RSP: 0018:ffffaebdc8a27da8 EFLAGS: 00010286 [192922.921315] RAX: 0000000000000051 RBX: ffff95c9c16a41c0 RCX: 0000000000000000 [192922.921692] RDX: 0000000000000000 RSI: ffff95cab6b16838 RDI: ffff95cab6b16838 [192922.922066] RBP: ffff95c9c16a41c0 R08: 0000000000000000 R09: 0000000000000000 [192922.922442] R10: ffffaebdc8a27e70 R11: 0000000000000000 R12: ffff95ca731a0980 [192922.922820] R13: 0000000000000000 R14: ffff95ca84c73338 R15: ffff95ca731a0ea8 [192922.923200] FS: 00007f337eda4e80(0000) GS:ffff95cab6b00000(0000) knlGS:0000000000000000 [192922.923579] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [192922.923948] CR2: 00007f337edad000 CR3: 00000001e00f6002 CR4: 00000000003606e0 [192922.924329] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [192922.924711] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [192922.925105] Call Trace: [192922.925505] btrfs_trans_release_metadata+0x10c/0x170 [btrfs] [192922.925911] btrfs_commit_transaction+0x3e/0xaf0 [btrfs] [192922.926324] btrfs_sync_file+0x44c/0x490 [btrfs] [192922.926731] do_fsync+0x38/0x60 [192922.927138] __x64_sys_fdatasync+0x13/0x20 [192922.927543] do_syscall_64+0x60/0x1c0 [192922.927939] entry_SYSCALL_64_after_hwframe+0x49/0xbe (...) [192922.934077] ---[ end trace f00808b12068168f ]--- 2) If evict_refill_and_join() decides to commit the transaction, it will be able to do it, since the nested transaction join only increments the transaction handle's ->use_count reference counter and it does not prevent the transaction from getting committed. This means that after eviction completes, the fsync logging path will be using a transaction handle that refers to an already committed transaction. What happens when using such a stale transaction can be unpredictable, we are at least having a use-after-free on the transaction handle itself, since the transaction commit will call kmem_cache_free() against the handle regardless of its ->use_count value, or we can end up silently losing all the updates to the log tree after that iput() in the logging path, or using a transaction handle that in the meanwhile was allocated to another task for a new transaction, etc, pretty much unpredictable what can happen. In order to fix both of them, instead of using iput() during logging, use btrfs_add_delayed_iput(), so that the logging path of fsync never drops the last reference on an inode, that step is offloaded to a safe context (usually the cleaner kthread). The assertion failure issue was sporadically triggered by the test case generic/475 from fstests, which loads the dm error target while fsstress is running, which lead to fsync failing while logging inodes with -EIO errors and then trying later to commit the transaction, triggering the assertion failure. CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-09-10 17:26:49 +03:00
btrfs_add_delayed_iput(di_inode);
Btrfs: fix deadlock caused by fsync when logging directory entries While logging new directory entries, at tree-log.c:log_new_dir_dentries(), after we call btrfs_search_forward() we get a leaf with a read lock on it, and without unlocking that leaf we can end up calling btrfs_iget() to get an inode pointer. The later (btrfs_iget()) can end up doing a read-only search on the same tree again, if the inode is not in memory already, which ends up causing a deadlock if some other task in the meanwhile started a write search on the tree and is attempting to write lock the same leaf that btrfs_search_forward() locked while holding write locks on upper levels of the tree blocking the read search from btrfs_iget(). In this scenario we get a deadlock. So fix this by releasing the search path before calling btrfs_iget() at tree-log.c:log_new_dir_dentries(). Example trace of such deadlock: [ 4077.478852] kworker/u24:10 D ffff88107fc90640 0 14431 2 0x00000000 [ 4077.486752] Workqueue: btrfs-endio-write btrfs_endio_write_helper [btrfs] [ 4077.494346] ffff880ffa56bad0 0000000000000046 0000000000009000 ffff880ffa56bfd8 [ 4077.502629] ffff880ffa56bfd8 ffff881016ce21c0 ffffffffa06ecb26 ffff88101a5d6138 [ 4077.510915] ffff880ebb5173b0 ffff880ffa56baf8 ffff880ebb517410 ffff881016ce21c0 [ 4077.519202] Call Trace: [ 4077.528752] [<ffffffffa06ed5ed>] ? btrfs_tree_lock+0xdd/0x2f0 [btrfs] [ 4077.536049] [<ffffffff81053680>] ? wake_up_atomic_t+0x30/0x30 [ 4077.542574] [<ffffffffa068cc1f>] ? btrfs_search_slot+0x79f/0xb10 [btrfs] [ 4077.550171] [<ffffffffa06a5073>] ? btrfs_lookup_file_extent+0x33/0x40 [btrfs] [ 4077.558252] [<ffffffffa06c600b>] ? __btrfs_drop_extents+0x13b/0xdf0 [btrfs] [ 4077.566140] [<ffffffffa06fc9e2>] ? add_delayed_data_ref+0xe2/0x150 [btrfs] [ 4077.573928] [<ffffffffa06fd629>] ? btrfs_add_delayed_data_ref+0x149/0x1d0 [btrfs] [ 4077.582399] [<ffffffffa06cf3c0>] ? __set_extent_bit+0x4c0/0x5c0 [btrfs] [ 4077.589896] [<ffffffffa06b4a64>] ? insert_reserved_file_extent.constprop.75+0xa4/0x320 [btrfs] [ 4077.599632] [<ffffffffa06b206d>] ? start_transaction+0x8d/0x470 [btrfs] [ 4077.607134] [<ffffffffa06bab57>] ? btrfs_finish_ordered_io+0x2e7/0x600 [btrfs] [ 4077.615329] [<ffffffff8104cbc2>] ? process_one_work+0x142/0x3d0 [ 4077.622043] [<ffffffff8104d729>] ? worker_thread+0x109/0x3b0 [ 4077.628459] [<ffffffff8104d620>] ? manage_workers.isra.26+0x270/0x270 [ 4077.635759] [<ffffffff81052b0f>] ? kthread+0xaf/0xc0 [ 4077.641404] [<ffffffff81052a60>] ? kthread_create_on_node+0x110/0x110 [ 4077.648696] [<ffffffff814a9ac8>] ? ret_from_fork+0x58/0x90 [ 4077.654926] [<ffffffff81052a60>] ? kthread_create_on_node+0x110/0x110 [ 4078.358087] kworker/u24:15 D ffff88107fcd0640 0 14436 2 0x00000000 [ 4078.365981] Workqueue: btrfs-endio-write btrfs_endio_write_helper [btrfs] [ 4078.373574] ffff880ffa57fad0 0000000000000046 0000000000009000 ffff880ffa57ffd8 [ 4078.381864] ffff880ffa57ffd8 ffff88103004d0a0 ffffffffa06ecb26 ffff88101a5d6138 [ 4078.390163] ffff880fbeffc298 ffff880ffa57faf8 ffff880fbeffc2f8 ffff88103004d0a0 [ 4078.398466] Call Trace: [ 4078.408019] [<ffffffffa06ed5ed>] ? btrfs_tree_lock+0xdd/0x2f0 [btrfs] [ 4078.415322] [<ffffffff81053680>] ? wake_up_atomic_t+0x30/0x30 [ 4078.421844] [<ffffffffa068cc1f>] ? btrfs_search_slot+0x79f/0xb10 [btrfs] [ 4078.429438] [<ffffffffa06a5073>] ? btrfs_lookup_file_extent+0x33/0x40 [btrfs] [ 4078.437518] [<ffffffffa06c600b>] ? __btrfs_drop_extents+0x13b/0xdf0 [btrfs] [ 4078.445404] [<ffffffffa06fc9e2>] ? add_delayed_data_ref+0xe2/0x150 [btrfs] [ 4078.453194] [<ffffffffa06fd629>] ? btrfs_add_delayed_data_ref+0x149/0x1d0 [btrfs] [ 4078.461663] [<ffffffffa06cf3c0>] ? __set_extent_bit+0x4c0/0x5c0 [btrfs] [ 4078.469161] [<ffffffffa06b4a64>] ? insert_reserved_file_extent.constprop.75+0xa4/0x320 [btrfs] [ 4078.478893] [<ffffffffa06b206d>] ? start_transaction+0x8d/0x470 [btrfs] [ 4078.486388] [<ffffffffa06bab57>] ? btrfs_finish_ordered_io+0x2e7/0x600 [btrfs] [ 4078.494561] [<ffffffff8104cbc2>] ? process_one_work+0x142/0x3d0 [ 4078.501278] [<ffffffff8104a507>] ? pwq_activate_delayed_work+0x27/0x40 [ 4078.508673] [<ffffffff8104d729>] ? worker_thread+0x109/0x3b0 [ 4078.515098] [<ffffffff8104d620>] ? manage_workers.isra.26+0x270/0x270 [ 4078.522396] [<ffffffff81052b0f>] ? kthread+0xaf/0xc0 [ 4078.528032] [<ffffffff81052a60>] ? kthread_create_on_node+0x110/0x110 [ 4078.535325] [<ffffffff814a9ac8>] ? ret_from_fork+0x58/0x90 [ 4078.541552] [<ffffffff81052a60>] ? kthread_create_on_node+0x110/0x110 [ 4079.355824] user-space-program D ffff88107fd30640 0 32020 1 0x00000000 [ 4079.363716] ffff880eae8eba10 0000000000000086 0000000000009000 ffff880eae8ebfd8 [ 4079.372003] ffff880eae8ebfd8 ffff881016c162c0 ffffffffa06ecb26 ffff88101a5d6138 [ 4079.380294] ffff880fbed4b4c8 ffff880eae8eba38 ffff880fbed4b528 ffff881016c162c0 [ 4079.388586] Call Trace: [ 4079.398134] [<ffffffffa06ed595>] ? btrfs_tree_lock+0x85/0x2f0 [btrfs] [ 4079.405431] [<ffffffff81053680>] ? wake_up_atomic_t+0x30/0x30 [ 4079.411955] [<ffffffffa06876fb>] ? btrfs_lock_root_node+0x2b/0x40 [btrfs] [ 4079.419644] [<ffffffffa068ce83>] ? btrfs_search_slot+0xa03/0xb10 [btrfs] [ 4079.427237] [<ffffffffa06aba52>] ? btrfs_buffer_uptodate+0x52/0x70 [btrfs] [ 4079.435041] [<ffffffffa0689b60>] ? generic_bin_search.constprop.38+0x80/0x190 [btrfs] [ 4079.443897] [<ffffffffa068ea44>] ? btrfs_insert_empty_items+0x74/0xd0 [btrfs] [ 4079.451975] [<ffffffffa072c443>] ? copy_items+0x128/0x850 [btrfs] [ 4079.458890] [<ffffffffa072da10>] ? btrfs_log_inode+0x629/0xbf3 [btrfs] [ 4079.466292] [<ffffffffa06f34a1>] ? btrfs_log_inode_parent+0xc61/0xf30 [btrfs] [ 4079.474373] [<ffffffffa06f45a9>] ? btrfs_log_dentry_safe+0x59/0x80 [btrfs] [ 4079.482161] [<ffffffffa06c298d>] ? btrfs_sync_file+0x20d/0x330 [btrfs] [ 4079.489558] [<ffffffff8112777c>] ? do_fsync+0x4c/0x80 [ 4079.495300] [<ffffffff81127a0a>] ? SyS_fdatasync+0xa/0x10 [ 4079.501422] [<ffffffff814a9b72>] ? system_call_fastpath+0x16/0x1b [ 4079.508334] user-space-program D ffff88107fc30640 0 32021 1 0x00000004 [ 4079.516226] ffff880eae8efbf8 0000000000000086 0000000000009000 ffff880eae8effd8 [ 4079.524513] ffff880eae8effd8 ffff881030279610 ffffffffa06ecb26 ffff88101a5d6138 [ 4079.532802] ffff880ebb671d88 ffff880eae8efc20 ffff880ebb671de8 ffff881030279610 [ 4079.541092] Call Trace: [ 4079.550642] [<ffffffffa06ed595>] ? btrfs_tree_lock+0x85/0x2f0 [btrfs] [ 4079.557941] [<ffffffff81053680>] ? wake_up_atomic_t+0x30/0x30 [ 4079.564463] [<ffffffffa068cc1f>] ? btrfs_search_slot+0x79f/0xb10 [btrfs] [ 4079.572058] [<ffffffffa06bb7d8>] ? btrfs_truncate_inode_items+0x168/0xb90 [btrfs] [ 4079.580526] [<ffffffffa06b04be>] ? join_transaction.isra.15+0x1e/0x3a0 [btrfs] [ 4079.588701] [<ffffffffa06b206d>] ? start_transaction+0x8d/0x470 [btrfs] [ 4079.596196] [<ffffffffa0690ac6>] ? block_rsv_add_bytes+0x16/0x50 [btrfs] [ 4079.603789] [<ffffffffa06bc2e9>] ? btrfs_truncate+0xe9/0x2e0 [btrfs] [ 4079.610994] [<ffffffffa06bd00b>] ? btrfs_setattr+0x30b/0x410 [btrfs] [ 4079.618197] [<ffffffff81117c1c>] ? notify_change+0x1dc/0x680 [ 4079.624625] [<ffffffff8123c8a4>] ? aa_path_perm+0xd4/0x160 [ 4079.630854] [<ffffffff810f4fcb>] ? do_truncate+0x5b/0x90 [ 4079.636889] [<ffffffff810f59fa>] ? do_sys_ftruncate.constprop.15+0x10a/0x160 [ 4079.644869] [<ffffffff8110d87b>] ? SyS_fcntl+0x5b/0x570 [ 4079.650805] [<ffffffff814a9b72>] ? system_call_fastpath+0x16/0x1b [ 4080.410607] user-space-program D ffff88107fc70640 0 32028 12639 0x00000004 [ 4080.418489] ffff880eaeccbbe0 0000000000000086 0000000000009000 ffff880eaeccbfd8 [ 4080.426778] ffff880eaeccbfd8 ffff880f317ef1e0 ffffffffa06ecb26 ffff88101a5d6138 [ 4080.435067] ffff880ef7e93928 ffff880f317ef1e0 ffff880eaeccbc08 ffff880f317ef1e0 [ 4080.443353] Call Trace: [ 4080.452920] [<ffffffffa06ed15d>] ? btrfs_tree_read_lock+0xdd/0x190 [btrfs] [ 4080.460703] [<ffffffff81053680>] ? wake_up_atomic_t+0x30/0x30 [ 4080.467225] [<ffffffffa06876bb>] ? btrfs_read_lock_root_node+0x2b/0x40 [btrfs] [ 4080.475400] [<ffffffffa068cc81>] ? btrfs_search_slot+0x801/0xb10 [btrfs] [ 4080.482994] [<ffffffffa06b2df0>] ? btrfs_clean_one_deleted_snapshot+0xe0/0xe0 [btrfs] [ 4080.491857] [<ffffffffa06a70a6>] ? btrfs_lookup_inode+0x26/0x90 [btrfs] [ 4080.499353] [<ffffffff810ec42f>] ? kmem_cache_alloc+0xaf/0xc0 [ 4080.505879] [<ffffffffa06bd905>] ? btrfs_iget+0xd5/0x5d0 [btrfs] [ 4080.512696] [<ffffffffa06caf04>] ? btrfs_get_token_64+0x104/0x120 [btrfs] [ 4080.520387] [<ffffffffa06f341f>] ? btrfs_log_inode_parent+0xbdf/0xf30 [btrfs] [ 4080.528469] [<ffffffffa06f45a9>] ? btrfs_log_dentry_safe+0x59/0x80 [btrfs] [ 4080.536258] [<ffffffffa06c298d>] ? btrfs_sync_file+0x20d/0x330 [btrfs] [ 4080.543657] [<ffffffff8112777c>] ? do_fsync+0x4c/0x80 [ 4080.549399] [<ffffffff81127a0a>] ? SyS_fdatasync+0xa/0x10 [ 4080.555534] [<ffffffff814a9b72>] ? system_call_fastpath+0x16/0x1b Signed-off-by: Robbie Ko <robbieko@synology.com> Reviewed-by: Filipe Manana <fdmanana@suse.com> Fixes: 2f2ff0ee5e43 (Btrfs: fix metadata inconsistencies after directory fsync) Cc: stable@vger.kernel.org # 4.1+ Signed-off-by: Filipe Manana <fdmanana@suse.com> [Modified changelog for clarity and correctness]
2016-10-28 05:48:26 +03:00
break;
Btrfs: fix metadata inconsistencies after directory fsync We can get into inconsistency between inodes and directory entries after fsyncing a directory. The issue is that while a directory gets the new dentries persisted in the fsync log and replayed at mount time, the link count of the inode that directory entries point to doesn't get updated, staying with an incorrect link count (smaller then the correct value). This later leads to stale file handle errors when accessing (including attempt to delete) some of the links if all the other ones are removed, which also implies impossibility to delete the parent directories, since the dentries can not be removed. Another issue is that (unlike ext3/4, xfs, f2fs, reiserfs, nilfs2), when fsyncing a directory, new files aren't logged (their metadata and dentries) nor any child directories. So this patch fixes this issue too, since it has the same resolution as the incorrect inode link count issue mentioned before. This is very easy to reproduce, and the following excerpt from my test case for xfstests shows how: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create our main test file and directory. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 8K" $SCRATCH_MNT/foo | _filter_xfs_io mkdir $SCRATCH_MNT/mydir # Make sure all metadata and data are durably persisted. sync # Add a hard link to 'foo' inside our test directory and fsync only the # directory. The btrfs fsync implementation had a bug that caused the new # directory entry to be visible after the fsync log replay but, the inode # of our file remained with a link count of 1. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_2 # Add a few more links and new files. # This is just to verify nothing breaks or gives incorrect results after the # fsync log is replayed. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_3 $XFS_IO_PROG -f -c "pwrite -S 0xff 0 64K" $SCRATCH_MNT/hello | _filter_xfs_io ln $SCRATCH_MNT/hello $SCRATCH_MNT/mydir/hello_2 # Add some subdirectories and new files and links to them. This is to verify # that after fsyncing our top level directory 'mydir', all the subdirectories # and their files/links are registered in the fsync log and exist after the # fsync log is replayed. mkdir -p $SCRATCH_MNT/mydir/x/y/z ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/foo_y_link ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/z/foo_z_link touch $SCRATCH_MNT/mydir/x/y/z/qwerty # Now fsync only our top directory. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/mydir # And fsync now our new file named 'hello', just to verify later that it has # the expected content and that the previous fsync on the directory 'mydir' had # no bad influence on this fsync. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/hello # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # Verify the content of our file 'foo' remains the same as before, 8192 bytes, # all with the value 0xaa. echo "File 'foo' content after log replay:" od -t x1 $SCRATCH_MNT/foo # Remove the first name of our inode. Because of the directory fsync bug, the # inode's link count was 1 instead of 5, so removing the 'foo' name ended up # deleting the inode and the other names became stale directory entries (still # visible to applications). Attempting to remove or access the remaining # dentries pointing to that inode resulted in stale file handle errors and # made it impossible to remove the parent directories since it was impossible # for them to become empty. echo "file 'foo' link count after log replay: $(stat -c %h $SCRATCH_MNT/foo)" rm -f $SCRATCH_MNT/foo # Now verify that all files, links and directories created before fsyncing our # directory exist after the fsync log was replayed. [ -f $SCRATCH_MNT/mydir/foo_2 ] || echo "Link mydir/foo_2 is missing" [ -f $SCRATCH_MNT/mydir/foo_3 ] || echo "Link mydir/foo_3 is missing" [ -f $SCRATCH_MNT/hello ] || echo "File hello is missing" [ -f $SCRATCH_MNT/mydir/hello_2 ] || echo "Link mydir/hello_2 is missing" [ -f $SCRATCH_MNT/mydir/x/y/foo_y_link ] || \ echo "Link mydir/x/y/foo_y_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/foo_z_link ] || \ echo "Link mydir/x/y/z/foo_z_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/qwerty ] || \ echo "File mydir/x/y/z/qwerty is missing" # We expect our file here to have a size of 64Kb and all the bytes having the # value 0xff. echo "file 'hello' content after log replay:" od -t x1 $SCRATCH_MNT/hello # Now remove all files/links, under our test directory 'mydir', and verify we # can remove all the directories. rm -f $SCRATCH_MNT/mydir/x/y/z/* rmdir $SCRATCH_MNT/mydir/x/y/z rm -f $SCRATCH_MNT/mydir/x/y/* rmdir $SCRATCH_MNT/mydir/x/y rmdir $SCRATCH_MNT/mydir/x rm -f $SCRATCH_MNT/mydir/* rmdir $SCRATCH_MNT/mydir # An fsck, run by the fstests framework everytime a test finishes, also detected # the inconsistency and printed the following error message: # # root 5 inode 257 errors 2001, no inode item, link count wrong # unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref # unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref status=0 exit The expected golden output for the test is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 5 file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 Which is the output after this patch and when running the test against ext3/4, xfs, f2fs, reiserfs or nilfs2. Without this patch, the test's output is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 1 Link mydir/foo_2 is missing Link mydir/foo_3 is missing Link mydir/x/y/foo_y_link is missing Link mydir/x/y/z/foo_z_link is missing File mydir/x/y/z/qwerty is missing file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y/z': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x': No such file or directory rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_2': Stale file handle rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_3': Stale file handle rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir': Directory not empty Fsck, without this fix, also complains about the wrong link count: root 5 inode 257 errors 2001, no inode item, link count wrong unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref So fix this by logging the inodes that the dentries point to when fsyncing a directory. A test case for xfstests follows. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-03-20 20:19:46 +03:00
}
ctx->log_new_dentries = false;
if (type == BTRFS_FT_DIR || type == BTRFS_FT_SYMLINK)
Btrfs: fix metadata inconsistencies after directory fsync We can get into inconsistency between inodes and directory entries after fsyncing a directory. The issue is that while a directory gets the new dentries persisted in the fsync log and replayed at mount time, the link count of the inode that directory entries point to doesn't get updated, staying with an incorrect link count (smaller then the correct value). This later leads to stale file handle errors when accessing (including attempt to delete) some of the links if all the other ones are removed, which also implies impossibility to delete the parent directories, since the dentries can not be removed. Another issue is that (unlike ext3/4, xfs, f2fs, reiserfs, nilfs2), when fsyncing a directory, new files aren't logged (their metadata and dentries) nor any child directories. So this patch fixes this issue too, since it has the same resolution as the incorrect inode link count issue mentioned before. This is very easy to reproduce, and the following excerpt from my test case for xfstests shows how: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create our main test file and directory. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 8K" $SCRATCH_MNT/foo | _filter_xfs_io mkdir $SCRATCH_MNT/mydir # Make sure all metadata and data are durably persisted. sync # Add a hard link to 'foo' inside our test directory and fsync only the # directory. The btrfs fsync implementation had a bug that caused the new # directory entry to be visible after the fsync log replay but, the inode # of our file remained with a link count of 1. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_2 # Add a few more links and new files. # This is just to verify nothing breaks or gives incorrect results after the # fsync log is replayed. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_3 $XFS_IO_PROG -f -c "pwrite -S 0xff 0 64K" $SCRATCH_MNT/hello | _filter_xfs_io ln $SCRATCH_MNT/hello $SCRATCH_MNT/mydir/hello_2 # Add some subdirectories and new files and links to them. This is to verify # that after fsyncing our top level directory 'mydir', all the subdirectories # and their files/links are registered in the fsync log and exist after the # fsync log is replayed. mkdir -p $SCRATCH_MNT/mydir/x/y/z ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/foo_y_link ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/z/foo_z_link touch $SCRATCH_MNT/mydir/x/y/z/qwerty # Now fsync only our top directory. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/mydir # And fsync now our new file named 'hello', just to verify later that it has # the expected content and that the previous fsync on the directory 'mydir' had # no bad influence on this fsync. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/hello # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # Verify the content of our file 'foo' remains the same as before, 8192 bytes, # all with the value 0xaa. echo "File 'foo' content after log replay:" od -t x1 $SCRATCH_MNT/foo # Remove the first name of our inode. Because of the directory fsync bug, the # inode's link count was 1 instead of 5, so removing the 'foo' name ended up # deleting the inode and the other names became stale directory entries (still # visible to applications). Attempting to remove or access the remaining # dentries pointing to that inode resulted in stale file handle errors and # made it impossible to remove the parent directories since it was impossible # for them to become empty. echo "file 'foo' link count after log replay: $(stat -c %h $SCRATCH_MNT/foo)" rm -f $SCRATCH_MNT/foo # Now verify that all files, links and directories created before fsyncing our # directory exist after the fsync log was replayed. [ -f $SCRATCH_MNT/mydir/foo_2 ] || echo "Link mydir/foo_2 is missing" [ -f $SCRATCH_MNT/mydir/foo_3 ] || echo "Link mydir/foo_3 is missing" [ -f $SCRATCH_MNT/hello ] || echo "File hello is missing" [ -f $SCRATCH_MNT/mydir/hello_2 ] || echo "Link mydir/hello_2 is missing" [ -f $SCRATCH_MNT/mydir/x/y/foo_y_link ] || \ echo "Link mydir/x/y/foo_y_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/foo_z_link ] || \ echo "Link mydir/x/y/z/foo_z_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/qwerty ] || \ echo "File mydir/x/y/z/qwerty is missing" # We expect our file here to have a size of 64Kb and all the bytes having the # value 0xff. echo "file 'hello' content after log replay:" od -t x1 $SCRATCH_MNT/hello # Now remove all files/links, under our test directory 'mydir', and verify we # can remove all the directories. rm -f $SCRATCH_MNT/mydir/x/y/z/* rmdir $SCRATCH_MNT/mydir/x/y/z rm -f $SCRATCH_MNT/mydir/x/y/* rmdir $SCRATCH_MNT/mydir/x/y rmdir $SCRATCH_MNT/mydir/x rm -f $SCRATCH_MNT/mydir/* rmdir $SCRATCH_MNT/mydir # An fsck, run by the fstests framework everytime a test finishes, also detected # the inconsistency and printed the following error message: # # root 5 inode 257 errors 2001, no inode item, link count wrong # unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref # unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref status=0 exit The expected golden output for the test is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 5 file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 Which is the output after this patch and when running the test against ext3/4, xfs, f2fs, reiserfs or nilfs2. Without this patch, the test's output is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 1 Link mydir/foo_2 is missing Link mydir/foo_3 is missing Link mydir/x/y/foo_y_link is missing Link mydir/x/y/z/foo_z_link is missing File mydir/x/y/z/qwerty is missing file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y/z': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x': No such file or directory rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_2': Stale file handle rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_3': Stale file handle rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir': Directory not empty Fsck, without this fix, also complains about the wrong link count: root 5 inode 257 errors 2001, no inode item, link count wrong unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref So fix this by logging the inodes that the dentries point to when fsyncing a directory. A test case for xfstests follows. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-03-20 20:19:46 +03:00
log_mode = LOG_INODE_ALL;
ret = btrfs_log_inode(trans, BTRFS_I(di_inode),
btrfs: make fast fsyncs wait only for writeback Currently regardless of a full or a fast fsync we always wait for ordered extents to complete, and then start logging the inode after that. However for fast fsyncs we can just wait for the writeback to complete, we don't need to wait for the ordered extents to complete since we use the list of modified extents maps to figure out which extents we must log and we can get their checksums directly from the ordered extents that are still in flight, otherwise look them up from the checksums tree. Until commit b5e6c3e170b770 ("btrfs: always wait on ordered extents at fsync time"), for fast fsyncs, we used to start logging without even waiting for the writeback to complete first, we would wait for it to complete after logging, while holding a transaction open, which lead to performance issues when using cgroups and probably for other cases too, as wait for IO while holding a transaction handle should be avoided as much as possible. After that, for fast fsyncs, we started to wait for ordered extents to complete before starting to log, which adds some latency to fsyncs and we even got at least one report about a performance drop which bisected to that particular change: https://lore.kernel.org/linux-btrfs/20181109215148.GF23260@techsingularity.net/ This change makes fast fsyncs only wait for writeback to finish before starting to log the inode, instead of waiting for both the writeback to finish and for the ordered extents to complete. This brings back part of the logic we had that extracts checksums from in flight ordered extents, which are not yet in the checksums tree, and making sure transaction commits wait for the completion of ordered extents previously logged (by far most of the time they have already completed by the time a transaction commit starts, resulting in no wait at all), to avoid any data loss if an ordered extent completes after the transaction used to log an inode is committed, followed by a power failure. When there are no other tasks accessing the checksums and the subvolume btrees, the ordered extent completion is pretty fast, typically taking 100 to 200 microseconds only in my observations. However when there are other tasks accessing these btrees, ordered extent completion can take a lot more time due to lock contention on nodes and leaves of these btrees. I've seen cases over 2 milliseconds, which starts to be significant. In particular when we do have concurrent fsyncs against different files there is a lot of contention on the checksums btree, since we have many tasks writing the checksums into the btree and other tasks that already started the logging phase are doing lookups for checksums in the btree. This change also turns all ranged fsyncs into full ranged fsyncs, which is something we already did when not using the NO_HOLES features or when doing a full fsync. This is to guarantee we never miss checksums due to writeback having been triggered only for a part of an extent, and we end up logging the full extent but only checksums for the written range, which results in missing checksums after log replay. Allowing ranged fsyncs to operate again only in the original range, when using the NO_HOLES feature and doing a fast fsync is doable but requires some non trivial changes to the writeback path, which can always be worked on later if needed, but I don't think they are a very common use case. Several tests were performed using fio for different numbers of concurrent jobs, each writing and fsyncing its own file, for both sequential and random file writes. The tests were run on bare metal, no virtualization, on a box with 12 cores (Intel i7-8700), 64Gb of RAM and a NVMe device, with a kernel configuration that is the default of typical distributions (debian in this case), without debug options enabled (kasan, kmemleak, slub debug, debug of page allocations, lock debugging, etc). The following script that calls fio was used: $ cat test-fsync.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/btrfs MOUNT_OPTIONS="-o ssd -o space_cache=v2" MKFS_OPTIONS="-d single -m single" if [ $# -ne 5 ]; then echo "Use $0 NUM_JOBS FILE_SIZE FSYNC_FREQ BLOCK_SIZE [write|randwrite]" exit 1 fi NUM_JOBS=$1 FILE_SIZE=$2 FSYNC_FREQ=$3 BLOCK_SIZE=$4 WRITE_MODE=$5 if [ "$WRITE_MODE" != "write" ] && [ "$WRITE_MODE" != "randwrite" ]; then echo "Invalid WRITE_MODE, must be 'write' or 'randwrite'" exit 1 fi cat <<EOF > /tmp/fio-job.ini [writers] rw=$WRITE_MODE fsync=$FSYNC_FREQ fallocate=none group_reporting=1 direct=0 bs=$BLOCK_SIZE ioengine=sync size=$FILE_SIZE directory=$MNT numjobs=$NUM_JOBS EOF echo "performance" | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor echo echo "Using config:" echo cat /tmp/fio-job.ini echo umount $MNT &> /dev/null mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT fio /tmp/fio-job.ini umount $MNT The results were the following: ************************* *** sequential writes *** ************************* ==== 1 job, 8GiB file, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=36.6MiB/s (38.4MB/s), 36.6MiB/s-36.6MiB/s (38.4MB/s-38.4MB/s), io=8192MiB (8590MB), run=223689-223689msec After patch: WRITE: bw=40.2MiB/s (42.1MB/s), 40.2MiB/s-40.2MiB/s (42.1MB/s-42.1MB/s), io=8192MiB (8590MB), run=203980-203980msec (+9.8%, -8.8% runtime) ==== 2 jobs, 4GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=35.8MiB/s (37.5MB/s), 35.8MiB/s-35.8MiB/s (37.5MB/s-37.5MB/s), io=8192MiB (8590MB), run=228950-228950msec After patch: WRITE: bw=43.5MiB/s (45.6MB/s), 43.5MiB/s-43.5MiB/s (45.6MB/s-45.6MB/s), io=8192MiB (8590MB), run=188272-188272msec (+21.5% throughput, -17.8% runtime) ==== 4 jobs, 2GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=50.1MiB/s (52.6MB/s), 50.1MiB/s-50.1MiB/s (52.6MB/s-52.6MB/s), io=8192MiB (8590MB), run=163446-163446msec After patch: WRITE: bw=64.5MiB/s (67.6MB/s), 64.5MiB/s-64.5MiB/s (67.6MB/s-67.6MB/s), io=8192MiB (8590MB), run=126987-126987msec (+28.7% throughput, -22.3% runtime) ==== 8 jobs, 1GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=64.0MiB/s (68.1MB/s), 64.0MiB/s-64.0MiB/s (68.1MB/s-68.1MB/s), io=8192MiB (8590MB), run=126075-126075msec After patch: WRITE: bw=86.8MiB/s (91.0MB/s), 86.8MiB/s-86.8MiB/s (91.0MB/s-91.0MB/s), io=8192MiB (8590MB), run=94358-94358msec (+35.6% throughput, -25.2% runtime) ==== 16 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=79.8MiB/s (83.6MB/s), 79.8MiB/s-79.8MiB/s (83.6MB/s-83.6MB/s), io=8192MiB (8590MB), run=102694-102694msec After patch: WRITE: bw=107MiB/s (112MB/s), 107MiB/s-107MiB/s (112MB/s-112MB/s), io=8192MiB (8590MB), run=76446-76446msec (+34.1% throughput, -25.6% runtime) ==== 32 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=93.2MiB/s (97.7MB/s), 93.2MiB/s-93.2MiB/s (97.7MB/s-97.7MB/s), io=16.0GiB (17.2GB), run=175836-175836msec After patch: WRITE: bw=111MiB/s (117MB/s), 111MiB/s-111MiB/s (117MB/s-117MB/s), io=16.0GiB (17.2GB), run=147001-147001msec (+19.1% throughput, -16.4% runtime) ==== 64 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=108MiB/s (114MB/s), 108MiB/s-108MiB/s (114MB/s-114MB/s), io=32.0GiB (34.4GB), run=302656-302656msec After patch: WRITE: bw=133MiB/s (140MB/s), 133MiB/s-133MiB/s (140MB/s-140MB/s), io=32.0GiB (34.4GB), run=246003-246003msec (+23.1% throughput, -18.7% runtime) ************************ *** random writes *** ************************ ==== 1 job, 8GiB file, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=11.5MiB/s (12.0MB/s), 11.5MiB/s-11.5MiB/s (12.0MB/s-12.0MB/s), io=8192MiB (8590MB), run=714281-714281msec After patch: WRITE: bw=11.6MiB/s (12.2MB/s), 11.6MiB/s-11.6MiB/s (12.2MB/s-12.2MB/s), io=8192MiB (8590MB), run=705959-705959msec (+0.9% throughput, -1.7% runtime) ==== 2 jobs, 4GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=12.8MiB/s (13.5MB/s), 12.8MiB/s-12.8MiB/s (13.5MB/s-13.5MB/s), io=8192MiB (8590MB), run=638101-638101msec After patch: WRITE: bw=13.1MiB/s (13.7MB/s), 13.1MiB/s-13.1MiB/s (13.7MB/s-13.7MB/s), io=8192MiB (8590MB), run=625374-625374msec (+2.3% throughput, -2.0% runtime) ==== 4 jobs, 2GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=15.4MiB/s (16.2MB/s), 15.4MiB/s-15.4MiB/s (16.2MB/s-16.2MB/s), io=8192MiB (8590MB), run=531146-531146msec After patch: WRITE: bw=17.8MiB/s (18.7MB/s), 17.8MiB/s-17.8MiB/s (18.7MB/s-18.7MB/s), io=8192MiB (8590MB), run=460431-460431msec (+15.6% throughput, -13.3% runtime) ==== 8 jobs, 1GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=19.9MiB/s (20.8MB/s), 19.9MiB/s-19.9MiB/s (20.8MB/s-20.8MB/s), io=8192MiB (8590MB), run=412664-412664msec After patch: WRITE: bw=22.2MiB/s (23.3MB/s), 22.2MiB/s-22.2MiB/s (23.3MB/s-23.3MB/s), io=8192MiB (8590MB), run=368589-368589msec (+11.6% throughput, -10.7% runtime) ==== 16 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=29.3MiB/s (30.7MB/s), 29.3MiB/s-29.3MiB/s (30.7MB/s-30.7MB/s), io=8192MiB (8590MB), run=279924-279924msec After patch: WRITE: bw=30.4MiB/s (31.9MB/s), 30.4MiB/s-30.4MiB/s (31.9MB/s-31.9MB/s), io=8192MiB (8590MB), run=269258-269258msec (+3.8% throughput, -3.8% runtime) ==== 32 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=36.9MiB/s (38.7MB/s), 36.9MiB/s-36.9MiB/s (38.7MB/s-38.7MB/s), io=16.0GiB (17.2GB), run=443581-443581msec After patch: WRITE: bw=41.6MiB/s (43.6MB/s), 41.6MiB/s-41.6MiB/s (43.6MB/s-43.6MB/s), io=16.0GiB (17.2GB), run=394114-394114msec (+12.7% throughput, -11.2% runtime) ==== 64 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=45.9MiB/s (48.1MB/s), 45.9MiB/s-45.9MiB/s (48.1MB/s-48.1MB/s), io=32.0GiB (34.4GB), run=714614-714614msec After patch: WRITE: bw=48.8MiB/s (51.1MB/s), 48.8MiB/s-48.8MiB/s (51.1MB/s-51.1MB/s), io=32.0GiB (34.4GB), run=672087-672087msec (+6.3% throughput, -6.0% runtime) Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-08-11 14:43:58 +03:00
log_mode, ctx);
Btrfs: fix assertion failure during fsync and use of stale transaction Sometimes when fsync'ing a file we need to log that other inodes exist and when we need to do that we acquire a reference on the inodes and then drop that reference using iput() after logging them. That generally is not a problem except if we end up doing the final iput() (dropping the last reference) on the inode and that inode has a link count of 0, which can happen in a very short time window if the logging path gets a reference on the inode while it's being unlinked. In that case we end up getting the eviction callback, btrfs_evict_inode(), invoked through the iput() call chain which needs to drop all of the inode's items from its subvolume btree, and in order to do that, it needs to join a transaction at the helper function evict_refill_and_join(). However because the task previously started a transaction at the fsync handler, btrfs_sync_file(), it has current->journal_info already pointing to a transaction handle and therefore evict_refill_and_join() will get that transaction handle from btrfs_join_transaction(). From this point on, two different problems can happen: 1) evict_refill_and_join() will often change the transaction handle's block reserve (->block_rsv) and set its ->bytes_reserved field to a value greater than 0. If evict_refill_and_join() never commits the transaction, the eviction handler ends up decreasing the reference count (->use_count) of the transaction handle through the call to btrfs_end_transaction(), and after that point we have a transaction handle with a NULL ->block_rsv (which is the value prior to the transaction join from evict_refill_and_join()) and a ->bytes_reserved value greater than 0. If after the eviction/iput completes the inode logging path hits an error or it decides that it must fallback to a transaction commit, the btrfs fsync handle, btrfs_sync_file(), gets a non-zero value from btrfs_log_dentry_safe(), and because of that non-zero value it tries to commit the transaction using a handle with a NULL ->block_rsv and a non-zero ->bytes_reserved value. This makes the transaction commit hit an assertion failure at btrfs_trans_release_metadata() because ->bytes_reserved is not zero but the ->block_rsv is NULL. The produced stack trace for that is like the following: [192922.917158] assertion failed: !trans->bytes_reserved, file: fs/btrfs/transaction.c, line: 816 [192922.917553] ------------[ cut here ]------------ [192922.917922] kernel BUG at fs/btrfs/ctree.h:3532! [192922.918310] invalid opcode: 0000 [#1] SMP DEBUG_PAGEALLOC PTI [192922.918666] CPU: 2 PID: 883 Comm: fsstress Tainted: G W 5.1.4-btrfs-next-47 #1 [192922.919035] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.11.2-0-gf9626ccb91-prebuilt.qemu-project.org 04/01/2014 [192922.919801] RIP: 0010:assfail.constprop.25+0x18/0x1a [btrfs] (...) [192922.920925] RSP: 0018:ffffaebdc8a27da8 EFLAGS: 00010286 [192922.921315] RAX: 0000000000000051 RBX: ffff95c9c16a41c0 RCX: 0000000000000000 [192922.921692] RDX: 0000000000000000 RSI: ffff95cab6b16838 RDI: ffff95cab6b16838 [192922.922066] RBP: ffff95c9c16a41c0 R08: 0000000000000000 R09: 0000000000000000 [192922.922442] R10: ffffaebdc8a27e70 R11: 0000000000000000 R12: ffff95ca731a0980 [192922.922820] R13: 0000000000000000 R14: ffff95ca84c73338 R15: ffff95ca731a0ea8 [192922.923200] FS: 00007f337eda4e80(0000) GS:ffff95cab6b00000(0000) knlGS:0000000000000000 [192922.923579] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [192922.923948] CR2: 00007f337edad000 CR3: 00000001e00f6002 CR4: 00000000003606e0 [192922.924329] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [192922.924711] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [192922.925105] Call Trace: [192922.925505] btrfs_trans_release_metadata+0x10c/0x170 [btrfs] [192922.925911] btrfs_commit_transaction+0x3e/0xaf0 [btrfs] [192922.926324] btrfs_sync_file+0x44c/0x490 [btrfs] [192922.926731] do_fsync+0x38/0x60 [192922.927138] __x64_sys_fdatasync+0x13/0x20 [192922.927543] do_syscall_64+0x60/0x1c0 [192922.927939] entry_SYSCALL_64_after_hwframe+0x49/0xbe (...) [192922.934077] ---[ end trace f00808b12068168f ]--- 2) If evict_refill_and_join() decides to commit the transaction, it will be able to do it, since the nested transaction join only increments the transaction handle's ->use_count reference counter and it does not prevent the transaction from getting committed. This means that after eviction completes, the fsync logging path will be using a transaction handle that refers to an already committed transaction. What happens when using such a stale transaction can be unpredictable, we are at least having a use-after-free on the transaction handle itself, since the transaction commit will call kmem_cache_free() against the handle regardless of its ->use_count value, or we can end up silently losing all the updates to the log tree after that iput() in the logging path, or using a transaction handle that in the meanwhile was allocated to another task for a new transaction, etc, pretty much unpredictable what can happen. In order to fix both of them, instead of using iput() during logging, use btrfs_add_delayed_iput(), so that the logging path of fsync never drops the last reference on an inode, that step is offloaded to a safe context (usually the cleaner kthread). The assertion failure issue was sporadically triggered by the test case generic/475 from fstests, which loads the dm error target while fsstress is running, which lead to fsync failing while logging inodes with -EIO errors and then trying later to commit the transaction, triggering the assertion failure. CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-09-10 17:26:49 +03:00
btrfs_add_delayed_iput(di_inode);
Btrfs: fix metadata inconsistencies after directory fsync We can get into inconsistency between inodes and directory entries after fsyncing a directory. The issue is that while a directory gets the new dentries persisted in the fsync log and replayed at mount time, the link count of the inode that directory entries point to doesn't get updated, staying with an incorrect link count (smaller then the correct value). This later leads to stale file handle errors when accessing (including attempt to delete) some of the links if all the other ones are removed, which also implies impossibility to delete the parent directories, since the dentries can not be removed. Another issue is that (unlike ext3/4, xfs, f2fs, reiserfs, nilfs2), when fsyncing a directory, new files aren't logged (their metadata and dentries) nor any child directories. So this patch fixes this issue too, since it has the same resolution as the incorrect inode link count issue mentioned before. This is very easy to reproduce, and the following excerpt from my test case for xfstests shows how: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create our main test file and directory. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 8K" $SCRATCH_MNT/foo | _filter_xfs_io mkdir $SCRATCH_MNT/mydir # Make sure all metadata and data are durably persisted. sync # Add a hard link to 'foo' inside our test directory and fsync only the # directory. The btrfs fsync implementation had a bug that caused the new # directory entry to be visible after the fsync log replay but, the inode # of our file remained with a link count of 1. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_2 # Add a few more links and new files. # This is just to verify nothing breaks or gives incorrect results after the # fsync log is replayed. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_3 $XFS_IO_PROG -f -c "pwrite -S 0xff 0 64K" $SCRATCH_MNT/hello | _filter_xfs_io ln $SCRATCH_MNT/hello $SCRATCH_MNT/mydir/hello_2 # Add some subdirectories and new files and links to them. This is to verify # that after fsyncing our top level directory 'mydir', all the subdirectories # and their files/links are registered in the fsync log and exist after the # fsync log is replayed. mkdir -p $SCRATCH_MNT/mydir/x/y/z ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/foo_y_link ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/z/foo_z_link touch $SCRATCH_MNT/mydir/x/y/z/qwerty # Now fsync only our top directory. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/mydir # And fsync now our new file named 'hello', just to verify later that it has # the expected content and that the previous fsync on the directory 'mydir' had # no bad influence on this fsync. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/hello # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # Verify the content of our file 'foo' remains the same as before, 8192 bytes, # all with the value 0xaa. echo "File 'foo' content after log replay:" od -t x1 $SCRATCH_MNT/foo # Remove the first name of our inode. Because of the directory fsync bug, the # inode's link count was 1 instead of 5, so removing the 'foo' name ended up # deleting the inode and the other names became stale directory entries (still # visible to applications). Attempting to remove or access the remaining # dentries pointing to that inode resulted in stale file handle errors and # made it impossible to remove the parent directories since it was impossible # for them to become empty. echo "file 'foo' link count after log replay: $(stat -c %h $SCRATCH_MNT/foo)" rm -f $SCRATCH_MNT/foo # Now verify that all files, links and directories created before fsyncing our # directory exist after the fsync log was replayed. [ -f $SCRATCH_MNT/mydir/foo_2 ] || echo "Link mydir/foo_2 is missing" [ -f $SCRATCH_MNT/mydir/foo_3 ] || echo "Link mydir/foo_3 is missing" [ -f $SCRATCH_MNT/hello ] || echo "File hello is missing" [ -f $SCRATCH_MNT/mydir/hello_2 ] || echo "Link mydir/hello_2 is missing" [ -f $SCRATCH_MNT/mydir/x/y/foo_y_link ] || \ echo "Link mydir/x/y/foo_y_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/foo_z_link ] || \ echo "Link mydir/x/y/z/foo_z_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/qwerty ] || \ echo "File mydir/x/y/z/qwerty is missing" # We expect our file here to have a size of 64Kb and all the bytes having the # value 0xff. echo "file 'hello' content after log replay:" od -t x1 $SCRATCH_MNT/hello # Now remove all files/links, under our test directory 'mydir', and verify we # can remove all the directories. rm -f $SCRATCH_MNT/mydir/x/y/z/* rmdir $SCRATCH_MNT/mydir/x/y/z rm -f $SCRATCH_MNT/mydir/x/y/* rmdir $SCRATCH_MNT/mydir/x/y rmdir $SCRATCH_MNT/mydir/x rm -f $SCRATCH_MNT/mydir/* rmdir $SCRATCH_MNT/mydir # An fsck, run by the fstests framework everytime a test finishes, also detected # the inconsistency and printed the following error message: # # root 5 inode 257 errors 2001, no inode item, link count wrong # unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref # unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref status=0 exit The expected golden output for the test is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 5 file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 Which is the output after this patch and when running the test against ext3/4, xfs, f2fs, reiserfs or nilfs2. Without this patch, the test's output is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 1 Link mydir/foo_2 is missing Link mydir/foo_3 is missing Link mydir/x/y/foo_y_link is missing Link mydir/x/y/z/foo_z_link is missing File mydir/x/y/z/qwerty is missing file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y/z': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x': No such file or directory rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_2': Stale file handle rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_3': Stale file handle rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir': Directory not empty Fsck, without this fix, also complains about the wrong link count: root 5 inode 257 errors 2001, no inode item, link count wrong unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref So fix this by logging the inodes that the dentries point to when fsyncing a directory. A test case for xfstests follows. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-03-20 20:19:46 +03:00
if (ret)
goto next_dir_inode;
if (ctx->log_new_dentries) {
new_dir_elem = kmalloc(sizeof(*new_dir_elem),
GFP_NOFS);
if (!new_dir_elem) {
ret = -ENOMEM;
goto next_dir_inode;
}
new_dir_elem->ino = di_key.objectid;
list_add_tail(&new_dir_elem->list, &dir_list);
}
break;
}
if (min_key.offset < (u64)-1) {
min_key.offset++;
goto again;
}
next_dir_inode:
list_del(&dir_elem->list);
kfree(dir_elem);
}
btrfs_free_path(path);
return ret;
}
Btrfs: fix stale dir entries after removing a link and fsync We have one more case where after a log tree is replayed we get inconsistent metadata leading to stale directory entries, due to some directories having entries pointing to some inode while the inode does not have a matching BTRFS_INODE_[REF|EXTREF]_KEY item. To trigger the problem we need to have a file with multiple hard links belonging to different parent directories. Then if one of those hard links is removed and we fsync the file using one of its other links that belongs to a different parent directory, we end up not logging the fact that the removed hard link doesn't exists anymore in the parent directory. Simple reproducer: seq=`basename $0` seqres=$RESULT_DIR/$seq echo "QA output created by $seq" tmp=/tmp/$$ status=1 # failure is the default! trap "_cleanup; exit \$status" 0 1 2 3 15 _cleanup() { _cleanup_flakey rm -f $tmp.* } # get standard environment, filters and checks . ./common/rc . ./common/filter . ./common/dmflakey # real QA test starts here _need_to_be_root _supported_fs generic _supported_os Linux _require_scratch _require_dm_flakey _require_metadata_journaling $SCRATCH_DEV rm -f $seqres.full _scratch_mkfs >>$seqres.full 2>&1 _init_flakey _mount_flakey # Create our test directory and file. mkdir $SCRATCH_MNT/testdir touch $SCRATCH_MNT/foo ln $SCRATCH_MNT/foo $SCRATCH_MNT/testdir/foo2 ln $SCRATCH_MNT/foo $SCRATCH_MNT/testdir/foo3 # Make sure everything done so far is durably persisted. sync # Now we remove one of our file's hardlinks in the directory testdir. unlink $SCRATCH_MNT/testdir/foo3 # We now fsync our file using the "foo" link, which has a parent that # is not the directory "testdir". $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foo # Silently drop all writes and unmount to simulate a crash/power # failure. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again, mount to trigger journal/log replay. _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # After the journal/log is replayed we expect to not see the "foo3" # link anymore and we should be able to remove all names in the # directory "testdir" and then remove it (no stale directory entries # left after the journal/log replay). echo "Entries in testdir:" ls -1 $SCRATCH_MNT/testdir rm -f $SCRATCH_MNT/testdir/* rmdir $SCRATCH_MNT/testdir _unmount_flakey status=0 exit The test fails with: $ ./check generic/107 FSTYP -- btrfs PLATFORM -- Linux/x86_64 debian3 4.1.0-rc6-btrfs-next-11+ MKFS_OPTIONS -- /dev/sdc MOUNT_OPTIONS -- /dev/sdc /home/fdmanana/btrfs-tests/scratch_1 generic/107 3s ... - output mismatch (see .../results/generic/107.out.bad) --- tests/generic/107.out 2015-08-01 01:39:45.807462161 +0100 +++ /home/fdmanana/git/hub/xfstests/results//generic/107.out.bad @@ -1,3 +1,5 @@ QA output created by 107 Entries in testdir: foo2 +foo3 +rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/testdir': Directory not empty ... _check_btrfs_filesystem: filesystem on /dev/sdc is inconsistent \ (see /home/fdmanana/git/hub/xfstests/results//generic/107.full) _check_dmesg: something found in dmesg (see .../results/generic/107.dmesg) Ran: generic/107 Failures: generic/107 Failed 1 of 1 tests $ cat /home/fdmanana/git/hub/xfstests/results//generic/107.full (...) checking fs roots root 5 inode 257 errors 200, dir isize wrong unresolved ref dir 257 index 3 namelen 4 name foo3 filetype 1 errors 5, no dir item, no inode ref (...) And produces the following warning in dmesg: [127298.759064] BTRFS info (device dm-0): failed to delete reference to foo3, inode 258 parent 257 [127298.762081] ------------[ cut here ]------------ [127298.763311] WARNING: CPU: 10 PID: 7891 at fs/btrfs/inode.c:3956 __btrfs_unlink_inode+0x182/0x35a [btrfs]() [127298.767327] BTRFS: Transaction aborted (error -2) (...) [127298.788611] Call Trace: [127298.789137] [<ffffffff8145f077>] dump_stack+0x4f/0x7b [127298.790090] [<ffffffff81095de5>] ? console_unlock+0x356/0x3a2 [127298.791157] [<ffffffff8104b3b0>] warn_slowpath_common+0xa1/0xbb [127298.792323] [<ffffffffa065ad09>] ? __btrfs_unlink_inode+0x182/0x35a [btrfs] [127298.793633] [<ffffffff8104b410>] warn_slowpath_fmt+0x46/0x48 [127298.794699] [<ffffffffa065ad09>] __btrfs_unlink_inode+0x182/0x35a [btrfs] [127298.797640] [<ffffffffa065be8f>] btrfs_unlink_inode+0x1e/0x40 [btrfs] [127298.798876] [<ffffffffa065bf11>] btrfs_unlink+0x60/0x9b [btrfs] [127298.800154] [<ffffffff8116fb48>] vfs_unlink+0x9c/0xed [127298.801303] [<ffffffff81173481>] do_unlinkat+0x12b/0x1fb [127298.802450] [<ffffffff81253855>] ? lockdep_sys_exit_thunk+0x12/0x14 [127298.803797] [<ffffffff81174056>] SyS_unlinkat+0x29/0x2b [127298.805017] [<ffffffff81465197>] system_call_fastpath+0x12/0x6f [127298.806310] ---[ end trace bbfddacb7aaada7b ]--- [127298.807325] BTRFS warning (device dm-0): __btrfs_unlink_inode:3956: Aborting unused transaction(No such entry). So fix this by logging all parent inodes, current and old ones, to make sure we do not get stale entries after log replay. This is not a simple solution such as triggering a full transaction commit because it would imply full transaction commit when an inode is fsynced in the same transaction that modified it and reloaded it after eviction (because its last_unlink_trans is set to the same value as its last_trans as of the commit with the title "Btrfs: fix stale dir entries after unlink, inode eviction and fsync"), and it would also make fstest generic/066 fail since one of the fsyncs triggers a full commit and the next fsync will not find the inode in the log anymore (therefore not removing the xattr). Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-08-05 18:49:08 +03:00
static int btrfs_log_all_parents(struct btrfs_trans_handle *trans,
struct btrfs_inode *inode,
Btrfs: fix stale dir entries after removing a link and fsync We have one more case where after a log tree is replayed we get inconsistent metadata leading to stale directory entries, due to some directories having entries pointing to some inode while the inode does not have a matching BTRFS_INODE_[REF|EXTREF]_KEY item. To trigger the problem we need to have a file with multiple hard links belonging to different parent directories. Then if one of those hard links is removed and we fsync the file using one of its other links that belongs to a different parent directory, we end up not logging the fact that the removed hard link doesn't exists anymore in the parent directory. Simple reproducer: seq=`basename $0` seqres=$RESULT_DIR/$seq echo "QA output created by $seq" tmp=/tmp/$$ status=1 # failure is the default! trap "_cleanup; exit \$status" 0 1 2 3 15 _cleanup() { _cleanup_flakey rm -f $tmp.* } # get standard environment, filters and checks . ./common/rc . ./common/filter . ./common/dmflakey # real QA test starts here _need_to_be_root _supported_fs generic _supported_os Linux _require_scratch _require_dm_flakey _require_metadata_journaling $SCRATCH_DEV rm -f $seqres.full _scratch_mkfs >>$seqres.full 2>&1 _init_flakey _mount_flakey # Create our test directory and file. mkdir $SCRATCH_MNT/testdir touch $SCRATCH_MNT/foo ln $SCRATCH_MNT/foo $SCRATCH_MNT/testdir/foo2 ln $SCRATCH_MNT/foo $SCRATCH_MNT/testdir/foo3 # Make sure everything done so far is durably persisted. sync # Now we remove one of our file's hardlinks in the directory testdir. unlink $SCRATCH_MNT/testdir/foo3 # We now fsync our file using the "foo" link, which has a parent that # is not the directory "testdir". $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foo # Silently drop all writes and unmount to simulate a crash/power # failure. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again, mount to trigger journal/log replay. _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # After the journal/log is replayed we expect to not see the "foo3" # link anymore and we should be able to remove all names in the # directory "testdir" and then remove it (no stale directory entries # left after the journal/log replay). echo "Entries in testdir:" ls -1 $SCRATCH_MNT/testdir rm -f $SCRATCH_MNT/testdir/* rmdir $SCRATCH_MNT/testdir _unmount_flakey status=0 exit The test fails with: $ ./check generic/107 FSTYP -- btrfs PLATFORM -- Linux/x86_64 debian3 4.1.0-rc6-btrfs-next-11+ MKFS_OPTIONS -- /dev/sdc MOUNT_OPTIONS -- /dev/sdc /home/fdmanana/btrfs-tests/scratch_1 generic/107 3s ... - output mismatch (see .../results/generic/107.out.bad) --- tests/generic/107.out 2015-08-01 01:39:45.807462161 +0100 +++ /home/fdmanana/git/hub/xfstests/results//generic/107.out.bad @@ -1,3 +1,5 @@ QA output created by 107 Entries in testdir: foo2 +foo3 +rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/testdir': Directory not empty ... _check_btrfs_filesystem: filesystem on /dev/sdc is inconsistent \ (see /home/fdmanana/git/hub/xfstests/results//generic/107.full) _check_dmesg: something found in dmesg (see .../results/generic/107.dmesg) Ran: generic/107 Failures: generic/107 Failed 1 of 1 tests $ cat /home/fdmanana/git/hub/xfstests/results//generic/107.full (...) checking fs roots root 5 inode 257 errors 200, dir isize wrong unresolved ref dir 257 index 3 namelen 4 name foo3 filetype 1 errors 5, no dir item, no inode ref (...) And produces the following warning in dmesg: [127298.759064] BTRFS info (device dm-0): failed to delete reference to foo3, inode 258 parent 257 [127298.762081] ------------[ cut here ]------------ [127298.763311] WARNING: CPU: 10 PID: 7891 at fs/btrfs/inode.c:3956 __btrfs_unlink_inode+0x182/0x35a [btrfs]() [127298.767327] BTRFS: Transaction aborted (error -2) (...) [127298.788611] Call Trace: [127298.789137] [<ffffffff8145f077>] dump_stack+0x4f/0x7b [127298.790090] [<ffffffff81095de5>] ? console_unlock+0x356/0x3a2 [127298.791157] [<ffffffff8104b3b0>] warn_slowpath_common+0xa1/0xbb [127298.792323] [<ffffffffa065ad09>] ? __btrfs_unlink_inode+0x182/0x35a [btrfs] [127298.793633] [<ffffffff8104b410>] warn_slowpath_fmt+0x46/0x48 [127298.794699] [<ffffffffa065ad09>] __btrfs_unlink_inode+0x182/0x35a [btrfs] [127298.797640] [<ffffffffa065be8f>] btrfs_unlink_inode+0x1e/0x40 [btrfs] [127298.798876] [<ffffffffa065bf11>] btrfs_unlink+0x60/0x9b [btrfs] [127298.800154] [<ffffffff8116fb48>] vfs_unlink+0x9c/0xed [127298.801303] [<ffffffff81173481>] do_unlinkat+0x12b/0x1fb [127298.802450] [<ffffffff81253855>] ? lockdep_sys_exit_thunk+0x12/0x14 [127298.803797] [<ffffffff81174056>] SyS_unlinkat+0x29/0x2b [127298.805017] [<ffffffff81465197>] system_call_fastpath+0x12/0x6f [127298.806310] ---[ end trace bbfddacb7aaada7b ]--- [127298.807325] BTRFS warning (device dm-0): __btrfs_unlink_inode:3956: Aborting unused transaction(No such entry). So fix this by logging all parent inodes, current and old ones, to make sure we do not get stale entries after log replay. This is not a simple solution such as triggering a full transaction commit because it would imply full transaction commit when an inode is fsynced in the same transaction that modified it and reloaded it after eviction (because its last_unlink_trans is set to the same value as its last_trans as of the commit with the title "Btrfs: fix stale dir entries after unlink, inode eviction and fsync"), and it would also make fstest generic/066 fail since one of the fsyncs triggers a full commit and the next fsync will not find the inode in the log anymore (therefore not removing the xattr). Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-08-05 18:49:08 +03:00
struct btrfs_log_ctx *ctx)
{
struct btrfs_fs_info *fs_info = trans->fs_info;
Btrfs: fix stale dir entries after removing a link and fsync We have one more case where after a log tree is replayed we get inconsistent metadata leading to stale directory entries, due to some directories having entries pointing to some inode while the inode does not have a matching BTRFS_INODE_[REF|EXTREF]_KEY item. To trigger the problem we need to have a file with multiple hard links belonging to different parent directories. Then if one of those hard links is removed and we fsync the file using one of its other links that belongs to a different parent directory, we end up not logging the fact that the removed hard link doesn't exists anymore in the parent directory. Simple reproducer: seq=`basename $0` seqres=$RESULT_DIR/$seq echo "QA output created by $seq" tmp=/tmp/$$ status=1 # failure is the default! trap "_cleanup; exit \$status" 0 1 2 3 15 _cleanup() { _cleanup_flakey rm -f $tmp.* } # get standard environment, filters and checks . ./common/rc . ./common/filter . ./common/dmflakey # real QA test starts here _need_to_be_root _supported_fs generic _supported_os Linux _require_scratch _require_dm_flakey _require_metadata_journaling $SCRATCH_DEV rm -f $seqres.full _scratch_mkfs >>$seqres.full 2>&1 _init_flakey _mount_flakey # Create our test directory and file. mkdir $SCRATCH_MNT/testdir touch $SCRATCH_MNT/foo ln $SCRATCH_MNT/foo $SCRATCH_MNT/testdir/foo2 ln $SCRATCH_MNT/foo $SCRATCH_MNT/testdir/foo3 # Make sure everything done so far is durably persisted. sync # Now we remove one of our file's hardlinks in the directory testdir. unlink $SCRATCH_MNT/testdir/foo3 # We now fsync our file using the "foo" link, which has a parent that # is not the directory "testdir". $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foo # Silently drop all writes and unmount to simulate a crash/power # failure. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again, mount to trigger journal/log replay. _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # After the journal/log is replayed we expect to not see the "foo3" # link anymore and we should be able to remove all names in the # directory "testdir" and then remove it (no stale directory entries # left after the journal/log replay). echo "Entries in testdir:" ls -1 $SCRATCH_MNT/testdir rm -f $SCRATCH_MNT/testdir/* rmdir $SCRATCH_MNT/testdir _unmount_flakey status=0 exit The test fails with: $ ./check generic/107 FSTYP -- btrfs PLATFORM -- Linux/x86_64 debian3 4.1.0-rc6-btrfs-next-11+ MKFS_OPTIONS -- /dev/sdc MOUNT_OPTIONS -- /dev/sdc /home/fdmanana/btrfs-tests/scratch_1 generic/107 3s ... - output mismatch (see .../results/generic/107.out.bad) --- tests/generic/107.out 2015-08-01 01:39:45.807462161 +0100 +++ /home/fdmanana/git/hub/xfstests/results//generic/107.out.bad @@ -1,3 +1,5 @@ QA output created by 107 Entries in testdir: foo2 +foo3 +rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/testdir': Directory not empty ... _check_btrfs_filesystem: filesystem on /dev/sdc is inconsistent \ (see /home/fdmanana/git/hub/xfstests/results//generic/107.full) _check_dmesg: something found in dmesg (see .../results/generic/107.dmesg) Ran: generic/107 Failures: generic/107 Failed 1 of 1 tests $ cat /home/fdmanana/git/hub/xfstests/results//generic/107.full (...) checking fs roots root 5 inode 257 errors 200, dir isize wrong unresolved ref dir 257 index 3 namelen 4 name foo3 filetype 1 errors 5, no dir item, no inode ref (...) And produces the following warning in dmesg: [127298.759064] BTRFS info (device dm-0): failed to delete reference to foo3, inode 258 parent 257 [127298.762081] ------------[ cut here ]------------ [127298.763311] WARNING: CPU: 10 PID: 7891 at fs/btrfs/inode.c:3956 __btrfs_unlink_inode+0x182/0x35a [btrfs]() [127298.767327] BTRFS: Transaction aborted (error -2) (...) [127298.788611] Call Trace: [127298.789137] [<ffffffff8145f077>] dump_stack+0x4f/0x7b [127298.790090] [<ffffffff81095de5>] ? console_unlock+0x356/0x3a2 [127298.791157] [<ffffffff8104b3b0>] warn_slowpath_common+0xa1/0xbb [127298.792323] [<ffffffffa065ad09>] ? __btrfs_unlink_inode+0x182/0x35a [btrfs] [127298.793633] [<ffffffff8104b410>] warn_slowpath_fmt+0x46/0x48 [127298.794699] [<ffffffffa065ad09>] __btrfs_unlink_inode+0x182/0x35a [btrfs] [127298.797640] [<ffffffffa065be8f>] btrfs_unlink_inode+0x1e/0x40 [btrfs] [127298.798876] [<ffffffffa065bf11>] btrfs_unlink+0x60/0x9b [btrfs] [127298.800154] [<ffffffff8116fb48>] vfs_unlink+0x9c/0xed [127298.801303] [<ffffffff81173481>] do_unlinkat+0x12b/0x1fb [127298.802450] [<ffffffff81253855>] ? lockdep_sys_exit_thunk+0x12/0x14 [127298.803797] [<ffffffff81174056>] SyS_unlinkat+0x29/0x2b [127298.805017] [<ffffffff81465197>] system_call_fastpath+0x12/0x6f [127298.806310] ---[ end trace bbfddacb7aaada7b ]--- [127298.807325] BTRFS warning (device dm-0): __btrfs_unlink_inode:3956: Aborting unused transaction(No such entry). So fix this by logging all parent inodes, current and old ones, to make sure we do not get stale entries after log replay. This is not a simple solution such as triggering a full transaction commit because it would imply full transaction commit when an inode is fsynced in the same transaction that modified it and reloaded it after eviction (because its last_unlink_trans is set to the same value as its last_trans as of the commit with the title "Btrfs: fix stale dir entries after unlink, inode eviction and fsync"), and it would also make fstest generic/066 fail since one of the fsyncs triggers a full commit and the next fsync will not find the inode in the log anymore (therefore not removing the xattr). Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-08-05 18:49:08 +03:00
int ret;
struct btrfs_path *path;
struct btrfs_key key;
struct btrfs_root *root = inode->root;
const u64 ino = btrfs_ino(inode);
Btrfs: fix stale dir entries after removing a link and fsync We have one more case where after a log tree is replayed we get inconsistent metadata leading to stale directory entries, due to some directories having entries pointing to some inode while the inode does not have a matching BTRFS_INODE_[REF|EXTREF]_KEY item. To trigger the problem we need to have a file with multiple hard links belonging to different parent directories. Then if one of those hard links is removed and we fsync the file using one of its other links that belongs to a different parent directory, we end up not logging the fact that the removed hard link doesn't exists anymore in the parent directory. Simple reproducer: seq=`basename $0` seqres=$RESULT_DIR/$seq echo "QA output created by $seq" tmp=/tmp/$$ status=1 # failure is the default! trap "_cleanup; exit \$status" 0 1 2 3 15 _cleanup() { _cleanup_flakey rm -f $tmp.* } # get standard environment, filters and checks . ./common/rc . ./common/filter . ./common/dmflakey # real QA test starts here _need_to_be_root _supported_fs generic _supported_os Linux _require_scratch _require_dm_flakey _require_metadata_journaling $SCRATCH_DEV rm -f $seqres.full _scratch_mkfs >>$seqres.full 2>&1 _init_flakey _mount_flakey # Create our test directory and file. mkdir $SCRATCH_MNT/testdir touch $SCRATCH_MNT/foo ln $SCRATCH_MNT/foo $SCRATCH_MNT/testdir/foo2 ln $SCRATCH_MNT/foo $SCRATCH_MNT/testdir/foo3 # Make sure everything done so far is durably persisted. sync # Now we remove one of our file's hardlinks in the directory testdir. unlink $SCRATCH_MNT/testdir/foo3 # We now fsync our file using the "foo" link, which has a parent that # is not the directory "testdir". $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foo # Silently drop all writes and unmount to simulate a crash/power # failure. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again, mount to trigger journal/log replay. _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # After the journal/log is replayed we expect to not see the "foo3" # link anymore and we should be able to remove all names in the # directory "testdir" and then remove it (no stale directory entries # left after the journal/log replay). echo "Entries in testdir:" ls -1 $SCRATCH_MNT/testdir rm -f $SCRATCH_MNT/testdir/* rmdir $SCRATCH_MNT/testdir _unmount_flakey status=0 exit The test fails with: $ ./check generic/107 FSTYP -- btrfs PLATFORM -- Linux/x86_64 debian3 4.1.0-rc6-btrfs-next-11+ MKFS_OPTIONS -- /dev/sdc MOUNT_OPTIONS -- /dev/sdc /home/fdmanana/btrfs-tests/scratch_1 generic/107 3s ... - output mismatch (see .../results/generic/107.out.bad) --- tests/generic/107.out 2015-08-01 01:39:45.807462161 +0100 +++ /home/fdmanana/git/hub/xfstests/results//generic/107.out.bad @@ -1,3 +1,5 @@ QA output created by 107 Entries in testdir: foo2 +foo3 +rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/testdir': Directory not empty ... _check_btrfs_filesystem: filesystem on /dev/sdc is inconsistent \ (see /home/fdmanana/git/hub/xfstests/results//generic/107.full) _check_dmesg: something found in dmesg (see .../results/generic/107.dmesg) Ran: generic/107 Failures: generic/107 Failed 1 of 1 tests $ cat /home/fdmanana/git/hub/xfstests/results//generic/107.full (...) checking fs roots root 5 inode 257 errors 200, dir isize wrong unresolved ref dir 257 index 3 namelen 4 name foo3 filetype 1 errors 5, no dir item, no inode ref (...) And produces the following warning in dmesg: [127298.759064] BTRFS info (device dm-0): failed to delete reference to foo3, inode 258 parent 257 [127298.762081] ------------[ cut here ]------------ [127298.763311] WARNING: CPU: 10 PID: 7891 at fs/btrfs/inode.c:3956 __btrfs_unlink_inode+0x182/0x35a [btrfs]() [127298.767327] BTRFS: Transaction aborted (error -2) (...) [127298.788611] Call Trace: [127298.789137] [<ffffffff8145f077>] dump_stack+0x4f/0x7b [127298.790090] [<ffffffff81095de5>] ? console_unlock+0x356/0x3a2 [127298.791157] [<ffffffff8104b3b0>] warn_slowpath_common+0xa1/0xbb [127298.792323] [<ffffffffa065ad09>] ? __btrfs_unlink_inode+0x182/0x35a [btrfs] [127298.793633] [<ffffffff8104b410>] warn_slowpath_fmt+0x46/0x48 [127298.794699] [<ffffffffa065ad09>] __btrfs_unlink_inode+0x182/0x35a [btrfs] [127298.797640] [<ffffffffa065be8f>] btrfs_unlink_inode+0x1e/0x40 [btrfs] [127298.798876] [<ffffffffa065bf11>] btrfs_unlink+0x60/0x9b [btrfs] [127298.800154] [<ffffffff8116fb48>] vfs_unlink+0x9c/0xed [127298.801303] [<ffffffff81173481>] do_unlinkat+0x12b/0x1fb [127298.802450] [<ffffffff81253855>] ? lockdep_sys_exit_thunk+0x12/0x14 [127298.803797] [<ffffffff81174056>] SyS_unlinkat+0x29/0x2b [127298.805017] [<ffffffff81465197>] system_call_fastpath+0x12/0x6f [127298.806310] ---[ end trace bbfddacb7aaada7b ]--- [127298.807325] BTRFS warning (device dm-0): __btrfs_unlink_inode:3956: Aborting unused transaction(No such entry). So fix this by logging all parent inodes, current and old ones, to make sure we do not get stale entries after log replay. This is not a simple solution such as triggering a full transaction commit because it would imply full transaction commit when an inode is fsynced in the same transaction that modified it and reloaded it after eviction (because its last_unlink_trans is set to the same value as its last_trans as of the commit with the title "Btrfs: fix stale dir entries after unlink, inode eviction and fsync"), and it would also make fstest generic/066 fail since one of the fsyncs triggers a full commit and the next fsync will not find the inode in the log anymore (therefore not removing the xattr). Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-08-05 18:49:08 +03:00
path = btrfs_alloc_path();
if (!path)
return -ENOMEM;
path->skip_locking = 1;
path->search_commit_root = 1;
key.objectid = ino;
key.type = BTRFS_INODE_REF_KEY;
key.offset = 0;
ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
if (ret < 0)
goto out;
while (true) {
struct extent_buffer *leaf = path->nodes[0];
int slot = path->slots[0];
u32 cur_offset = 0;
u32 item_size;
unsigned long ptr;
if (slot >= btrfs_header_nritems(leaf)) {
ret = btrfs_next_leaf(root, path);
if (ret < 0)
goto out;
else if (ret > 0)
break;
continue;
}
btrfs_item_key_to_cpu(leaf, &key, slot);
/* BTRFS_INODE_EXTREF_KEY is BTRFS_INODE_REF_KEY + 1 */
if (key.objectid != ino || key.type > BTRFS_INODE_EXTREF_KEY)
break;
item_size = btrfs_item_size(leaf, slot);
Btrfs: fix stale dir entries after removing a link and fsync We have one more case where after a log tree is replayed we get inconsistent metadata leading to stale directory entries, due to some directories having entries pointing to some inode while the inode does not have a matching BTRFS_INODE_[REF|EXTREF]_KEY item. To trigger the problem we need to have a file with multiple hard links belonging to different parent directories. Then if one of those hard links is removed and we fsync the file using one of its other links that belongs to a different parent directory, we end up not logging the fact that the removed hard link doesn't exists anymore in the parent directory. Simple reproducer: seq=`basename $0` seqres=$RESULT_DIR/$seq echo "QA output created by $seq" tmp=/tmp/$$ status=1 # failure is the default! trap "_cleanup; exit \$status" 0 1 2 3 15 _cleanup() { _cleanup_flakey rm -f $tmp.* } # get standard environment, filters and checks . ./common/rc . ./common/filter . ./common/dmflakey # real QA test starts here _need_to_be_root _supported_fs generic _supported_os Linux _require_scratch _require_dm_flakey _require_metadata_journaling $SCRATCH_DEV rm -f $seqres.full _scratch_mkfs >>$seqres.full 2>&1 _init_flakey _mount_flakey # Create our test directory and file. mkdir $SCRATCH_MNT/testdir touch $SCRATCH_MNT/foo ln $SCRATCH_MNT/foo $SCRATCH_MNT/testdir/foo2 ln $SCRATCH_MNT/foo $SCRATCH_MNT/testdir/foo3 # Make sure everything done so far is durably persisted. sync # Now we remove one of our file's hardlinks in the directory testdir. unlink $SCRATCH_MNT/testdir/foo3 # We now fsync our file using the "foo" link, which has a parent that # is not the directory "testdir". $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foo # Silently drop all writes and unmount to simulate a crash/power # failure. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again, mount to trigger journal/log replay. _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # After the journal/log is replayed we expect to not see the "foo3" # link anymore and we should be able to remove all names in the # directory "testdir" and then remove it (no stale directory entries # left after the journal/log replay). echo "Entries in testdir:" ls -1 $SCRATCH_MNT/testdir rm -f $SCRATCH_MNT/testdir/* rmdir $SCRATCH_MNT/testdir _unmount_flakey status=0 exit The test fails with: $ ./check generic/107 FSTYP -- btrfs PLATFORM -- Linux/x86_64 debian3 4.1.0-rc6-btrfs-next-11+ MKFS_OPTIONS -- /dev/sdc MOUNT_OPTIONS -- /dev/sdc /home/fdmanana/btrfs-tests/scratch_1 generic/107 3s ... - output mismatch (see .../results/generic/107.out.bad) --- tests/generic/107.out 2015-08-01 01:39:45.807462161 +0100 +++ /home/fdmanana/git/hub/xfstests/results//generic/107.out.bad @@ -1,3 +1,5 @@ QA output created by 107 Entries in testdir: foo2 +foo3 +rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/testdir': Directory not empty ... _check_btrfs_filesystem: filesystem on /dev/sdc is inconsistent \ (see /home/fdmanana/git/hub/xfstests/results//generic/107.full) _check_dmesg: something found in dmesg (see .../results/generic/107.dmesg) Ran: generic/107 Failures: generic/107 Failed 1 of 1 tests $ cat /home/fdmanana/git/hub/xfstests/results//generic/107.full (...) checking fs roots root 5 inode 257 errors 200, dir isize wrong unresolved ref dir 257 index 3 namelen 4 name foo3 filetype 1 errors 5, no dir item, no inode ref (...) And produces the following warning in dmesg: [127298.759064] BTRFS info (device dm-0): failed to delete reference to foo3, inode 258 parent 257 [127298.762081] ------------[ cut here ]------------ [127298.763311] WARNING: CPU: 10 PID: 7891 at fs/btrfs/inode.c:3956 __btrfs_unlink_inode+0x182/0x35a [btrfs]() [127298.767327] BTRFS: Transaction aborted (error -2) (...) [127298.788611] Call Trace: [127298.789137] [<ffffffff8145f077>] dump_stack+0x4f/0x7b [127298.790090] [<ffffffff81095de5>] ? console_unlock+0x356/0x3a2 [127298.791157] [<ffffffff8104b3b0>] warn_slowpath_common+0xa1/0xbb [127298.792323] [<ffffffffa065ad09>] ? __btrfs_unlink_inode+0x182/0x35a [btrfs] [127298.793633] [<ffffffff8104b410>] warn_slowpath_fmt+0x46/0x48 [127298.794699] [<ffffffffa065ad09>] __btrfs_unlink_inode+0x182/0x35a [btrfs] [127298.797640] [<ffffffffa065be8f>] btrfs_unlink_inode+0x1e/0x40 [btrfs] [127298.798876] [<ffffffffa065bf11>] btrfs_unlink+0x60/0x9b [btrfs] [127298.800154] [<ffffffff8116fb48>] vfs_unlink+0x9c/0xed [127298.801303] [<ffffffff81173481>] do_unlinkat+0x12b/0x1fb [127298.802450] [<ffffffff81253855>] ? lockdep_sys_exit_thunk+0x12/0x14 [127298.803797] [<ffffffff81174056>] SyS_unlinkat+0x29/0x2b [127298.805017] [<ffffffff81465197>] system_call_fastpath+0x12/0x6f [127298.806310] ---[ end trace bbfddacb7aaada7b ]--- [127298.807325] BTRFS warning (device dm-0): __btrfs_unlink_inode:3956: Aborting unused transaction(No such entry). So fix this by logging all parent inodes, current and old ones, to make sure we do not get stale entries after log replay. This is not a simple solution such as triggering a full transaction commit because it would imply full transaction commit when an inode is fsynced in the same transaction that modified it and reloaded it after eviction (because its last_unlink_trans is set to the same value as its last_trans as of the commit with the title "Btrfs: fix stale dir entries after unlink, inode eviction and fsync"), and it would also make fstest generic/066 fail since one of the fsyncs triggers a full commit and the next fsync will not find the inode in the log anymore (therefore not removing the xattr). Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-08-05 18:49:08 +03:00
ptr = btrfs_item_ptr_offset(leaf, slot);
while (cur_offset < item_size) {
struct btrfs_key inode_key;
struct inode *dir_inode;
inode_key.type = BTRFS_INODE_ITEM_KEY;
inode_key.offset = 0;
if (key.type == BTRFS_INODE_EXTREF_KEY) {
struct btrfs_inode_extref *extref;
extref = (struct btrfs_inode_extref *)
(ptr + cur_offset);
inode_key.objectid = btrfs_inode_extref_parent(
leaf, extref);
cur_offset += sizeof(*extref);
cur_offset += btrfs_inode_extref_name_len(leaf,
extref);
} else {
inode_key.objectid = key.offset;
cur_offset = item_size;
}
dir_inode = btrfs_iget(fs_info->sb, inode_key.objectid,
root);
/*
* If the parent inode was deleted, return an error to
* fallback to a transaction commit. This is to prevent
* getting an inode that was moved from one parent A to
* a parent B, got its former parent A deleted and then
* it got fsync'ed, from existing at both parents after
* a log replay (and the old parent still existing).
* Example:
*
* mkdir /mnt/A
* mkdir /mnt/B
* touch /mnt/B/bar
* sync
* mv /mnt/B/bar /mnt/A/bar
* mv -T /mnt/A /mnt/B
* fsync /mnt/B/bar
* <power fail>
*
* If we ignore the old parent B which got deleted,
* after a log replay we would have file bar linked
* at both parents and the old parent B would still
* exist.
*/
if (IS_ERR(dir_inode)) {
ret = PTR_ERR(dir_inode);
goto out;
}
Btrfs: fix stale dir entries after removing a link and fsync We have one more case where after a log tree is replayed we get inconsistent metadata leading to stale directory entries, due to some directories having entries pointing to some inode while the inode does not have a matching BTRFS_INODE_[REF|EXTREF]_KEY item. To trigger the problem we need to have a file with multiple hard links belonging to different parent directories. Then if one of those hard links is removed and we fsync the file using one of its other links that belongs to a different parent directory, we end up not logging the fact that the removed hard link doesn't exists anymore in the parent directory. Simple reproducer: seq=`basename $0` seqres=$RESULT_DIR/$seq echo "QA output created by $seq" tmp=/tmp/$$ status=1 # failure is the default! trap "_cleanup; exit \$status" 0 1 2 3 15 _cleanup() { _cleanup_flakey rm -f $tmp.* } # get standard environment, filters and checks . ./common/rc . ./common/filter . ./common/dmflakey # real QA test starts here _need_to_be_root _supported_fs generic _supported_os Linux _require_scratch _require_dm_flakey _require_metadata_journaling $SCRATCH_DEV rm -f $seqres.full _scratch_mkfs >>$seqres.full 2>&1 _init_flakey _mount_flakey # Create our test directory and file. mkdir $SCRATCH_MNT/testdir touch $SCRATCH_MNT/foo ln $SCRATCH_MNT/foo $SCRATCH_MNT/testdir/foo2 ln $SCRATCH_MNT/foo $SCRATCH_MNT/testdir/foo3 # Make sure everything done so far is durably persisted. sync # Now we remove one of our file's hardlinks in the directory testdir. unlink $SCRATCH_MNT/testdir/foo3 # We now fsync our file using the "foo" link, which has a parent that # is not the directory "testdir". $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foo # Silently drop all writes and unmount to simulate a crash/power # failure. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again, mount to trigger journal/log replay. _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # After the journal/log is replayed we expect to not see the "foo3" # link anymore and we should be able to remove all names in the # directory "testdir" and then remove it (no stale directory entries # left after the journal/log replay). echo "Entries in testdir:" ls -1 $SCRATCH_MNT/testdir rm -f $SCRATCH_MNT/testdir/* rmdir $SCRATCH_MNT/testdir _unmount_flakey status=0 exit The test fails with: $ ./check generic/107 FSTYP -- btrfs PLATFORM -- Linux/x86_64 debian3 4.1.0-rc6-btrfs-next-11+ MKFS_OPTIONS -- /dev/sdc MOUNT_OPTIONS -- /dev/sdc /home/fdmanana/btrfs-tests/scratch_1 generic/107 3s ... - output mismatch (see .../results/generic/107.out.bad) --- tests/generic/107.out 2015-08-01 01:39:45.807462161 +0100 +++ /home/fdmanana/git/hub/xfstests/results//generic/107.out.bad @@ -1,3 +1,5 @@ QA output created by 107 Entries in testdir: foo2 +foo3 +rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/testdir': Directory not empty ... _check_btrfs_filesystem: filesystem on /dev/sdc is inconsistent \ (see /home/fdmanana/git/hub/xfstests/results//generic/107.full) _check_dmesg: something found in dmesg (see .../results/generic/107.dmesg) Ran: generic/107 Failures: generic/107 Failed 1 of 1 tests $ cat /home/fdmanana/git/hub/xfstests/results//generic/107.full (...) checking fs roots root 5 inode 257 errors 200, dir isize wrong unresolved ref dir 257 index 3 namelen 4 name foo3 filetype 1 errors 5, no dir item, no inode ref (...) And produces the following warning in dmesg: [127298.759064] BTRFS info (device dm-0): failed to delete reference to foo3, inode 258 parent 257 [127298.762081] ------------[ cut here ]------------ [127298.763311] WARNING: CPU: 10 PID: 7891 at fs/btrfs/inode.c:3956 __btrfs_unlink_inode+0x182/0x35a [btrfs]() [127298.767327] BTRFS: Transaction aborted (error -2) (...) [127298.788611] Call Trace: [127298.789137] [<ffffffff8145f077>] dump_stack+0x4f/0x7b [127298.790090] [<ffffffff81095de5>] ? console_unlock+0x356/0x3a2 [127298.791157] [<ffffffff8104b3b0>] warn_slowpath_common+0xa1/0xbb [127298.792323] [<ffffffffa065ad09>] ? __btrfs_unlink_inode+0x182/0x35a [btrfs] [127298.793633] [<ffffffff8104b410>] warn_slowpath_fmt+0x46/0x48 [127298.794699] [<ffffffffa065ad09>] __btrfs_unlink_inode+0x182/0x35a [btrfs] [127298.797640] [<ffffffffa065be8f>] btrfs_unlink_inode+0x1e/0x40 [btrfs] [127298.798876] [<ffffffffa065bf11>] btrfs_unlink+0x60/0x9b [btrfs] [127298.800154] [<ffffffff8116fb48>] vfs_unlink+0x9c/0xed [127298.801303] [<ffffffff81173481>] do_unlinkat+0x12b/0x1fb [127298.802450] [<ffffffff81253855>] ? lockdep_sys_exit_thunk+0x12/0x14 [127298.803797] [<ffffffff81174056>] SyS_unlinkat+0x29/0x2b [127298.805017] [<ffffffff81465197>] system_call_fastpath+0x12/0x6f [127298.806310] ---[ end trace bbfddacb7aaada7b ]--- [127298.807325] BTRFS warning (device dm-0): __btrfs_unlink_inode:3956: Aborting unused transaction(No such entry). So fix this by logging all parent inodes, current and old ones, to make sure we do not get stale entries after log replay. This is not a simple solution such as triggering a full transaction commit because it would imply full transaction commit when an inode is fsynced in the same transaction that modified it and reloaded it after eviction (because its last_unlink_trans is set to the same value as its last_trans as of the commit with the title "Btrfs: fix stale dir entries after unlink, inode eviction and fsync"), and it would also make fstest generic/066 fail since one of the fsyncs triggers a full commit and the next fsync will not find the inode in the log anymore (therefore not removing the xattr). Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-08-05 18:49:08 +03:00
btrfs: skip logging directories already logged when logging all parents Some times when we fsync an inode we need to do a full log of all its ancestors (due to unlink, link or rename operations), which can be an expensive operation, specially if the directories are large. However if we find an ancestor directory inode that is already logged in the current transaction, and has no inserted/updated/deleted xattrs since it was last logged, we can skip logging the directory again. We are safe to skip that since we know that for logged directories, any link, unlink or rename operations that implicate the directory will update the log as necessary. So use the helper need_log_dir(), introduced in a previous commit, to detect already logged directories that can be skipped. This patch is part of a patchset comprised of the following patches: btrfs: remove unnecessary directory inode item update when deleting dir entry btrfs: stop setting nbytes when filling inode item for logging btrfs: avoid logging new ancestor inodes when logging new inode btrfs: skip logging directories already logged when logging all parents btrfs: skip logging inodes already logged when logging new entries btrfs: remove unnecessary check_parent_dirs_for_sync() btrfs: make concurrent fsyncs wait less when waiting for a transaction commit Performance results, after applying all patches, are mentioned in the change log of the last patch. Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-01-27 13:34:57 +03:00
if (!need_log_inode(trans, BTRFS_I(dir_inode))) {
btrfs_add_delayed_iput(dir_inode);
continue;
}
ctx->log_new_dentries = false;
ret = btrfs_log_inode(trans, BTRFS_I(dir_inode),
btrfs: make fast fsyncs wait only for writeback Currently regardless of a full or a fast fsync we always wait for ordered extents to complete, and then start logging the inode after that. However for fast fsyncs we can just wait for the writeback to complete, we don't need to wait for the ordered extents to complete since we use the list of modified extents maps to figure out which extents we must log and we can get their checksums directly from the ordered extents that are still in flight, otherwise look them up from the checksums tree. Until commit b5e6c3e170b770 ("btrfs: always wait on ordered extents at fsync time"), for fast fsyncs, we used to start logging without even waiting for the writeback to complete first, we would wait for it to complete after logging, while holding a transaction open, which lead to performance issues when using cgroups and probably for other cases too, as wait for IO while holding a transaction handle should be avoided as much as possible. After that, for fast fsyncs, we started to wait for ordered extents to complete before starting to log, which adds some latency to fsyncs and we even got at least one report about a performance drop which bisected to that particular change: https://lore.kernel.org/linux-btrfs/20181109215148.GF23260@techsingularity.net/ This change makes fast fsyncs only wait for writeback to finish before starting to log the inode, instead of waiting for both the writeback to finish and for the ordered extents to complete. This brings back part of the logic we had that extracts checksums from in flight ordered extents, which are not yet in the checksums tree, and making sure transaction commits wait for the completion of ordered extents previously logged (by far most of the time they have already completed by the time a transaction commit starts, resulting in no wait at all), to avoid any data loss if an ordered extent completes after the transaction used to log an inode is committed, followed by a power failure. When there are no other tasks accessing the checksums and the subvolume btrees, the ordered extent completion is pretty fast, typically taking 100 to 200 microseconds only in my observations. However when there are other tasks accessing these btrees, ordered extent completion can take a lot more time due to lock contention on nodes and leaves of these btrees. I've seen cases over 2 milliseconds, which starts to be significant. In particular when we do have concurrent fsyncs against different files there is a lot of contention on the checksums btree, since we have many tasks writing the checksums into the btree and other tasks that already started the logging phase are doing lookups for checksums in the btree. This change also turns all ranged fsyncs into full ranged fsyncs, which is something we already did when not using the NO_HOLES features or when doing a full fsync. This is to guarantee we never miss checksums due to writeback having been triggered only for a part of an extent, and we end up logging the full extent but only checksums for the written range, which results in missing checksums after log replay. Allowing ranged fsyncs to operate again only in the original range, when using the NO_HOLES feature and doing a fast fsync is doable but requires some non trivial changes to the writeback path, which can always be worked on later if needed, but I don't think they are a very common use case. Several tests were performed using fio for different numbers of concurrent jobs, each writing and fsyncing its own file, for both sequential and random file writes. The tests were run on bare metal, no virtualization, on a box with 12 cores (Intel i7-8700), 64Gb of RAM and a NVMe device, with a kernel configuration that is the default of typical distributions (debian in this case), without debug options enabled (kasan, kmemleak, slub debug, debug of page allocations, lock debugging, etc). The following script that calls fio was used: $ cat test-fsync.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/btrfs MOUNT_OPTIONS="-o ssd -o space_cache=v2" MKFS_OPTIONS="-d single -m single" if [ $# -ne 5 ]; then echo "Use $0 NUM_JOBS FILE_SIZE FSYNC_FREQ BLOCK_SIZE [write|randwrite]" exit 1 fi NUM_JOBS=$1 FILE_SIZE=$2 FSYNC_FREQ=$3 BLOCK_SIZE=$4 WRITE_MODE=$5 if [ "$WRITE_MODE" != "write" ] && [ "$WRITE_MODE" != "randwrite" ]; then echo "Invalid WRITE_MODE, must be 'write' or 'randwrite'" exit 1 fi cat <<EOF > /tmp/fio-job.ini [writers] rw=$WRITE_MODE fsync=$FSYNC_FREQ fallocate=none group_reporting=1 direct=0 bs=$BLOCK_SIZE ioengine=sync size=$FILE_SIZE directory=$MNT numjobs=$NUM_JOBS EOF echo "performance" | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor echo echo "Using config:" echo cat /tmp/fio-job.ini echo umount $MNT &> /dev/null mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT fio /tmp/fio-job.ini umount $MNT The results were the following: ************************* *** sequential writes *** ************************* ==== 1 job, 8GiB file, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=36.6MiB/s (38.4MB/s), 36.6MiB/s-36.6MiB/s (38.4MB/s-38.4MB/s), io=8192MiB (8590MB), run=223689-223689msec After patch: WRITE: bw=40.2MiB/s (42.1MB/s), 40.2MiB/s-40.2MiB/s (42.1MB/s-42.1MB/s), io=8192MiB (8590MB), run=203980-203980msec (+9.8%, -8.8% runtime) ==== 2 jobs, 4GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=35.8MiB/s (37.5MB/s), 35.8MiB/s-35.8MiB/s (37.5MB/s-37.5MB/s), io=8192MiB (8590MB), run=228950-228950msec After patch: WRITE: bw=43.5MiB/s (45.6MB/s), 43.5MiB/s-43.5MiB/s (45.6MB/s-45.6MB/s), io=8192MiB (8590MB), run=188272-188272msec (+21.5% throughput, -17.8% runtime) ==== 4 jobs, 2GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=50.1MiB/s (52.6MB/s), 50.1MiB/s-50.1MiB/s (52.6MB/s-52.6MB/s), io=8192MiB (8590MB), run=163446-163446msec After patch: WRITE: bw=64.5MiB/s (67.6MB/s), 64.5MiB/s-64.5MiB/s (67.6MB/s-67.6MB/s), io=8192MiB (8590MB), run=126987-126987msec (+28.7% throughput, -22.3% runtime) ==== 8 jobs, 1GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=64.0MiB/s (68.1MB/s), 64.0MiB/s-64.0MiB/s (68.1MB/s-68.1MB/s), io=8192MiB (8590MB), run=126075-126075msec After patch: WRITE: bw=86.8MiB/s (91.0MB/s), 86.8MiB/s-86.8MiB/s (91.0MB/s-91.0MB/s), io=8192MiB (8590MB), run=94358-94358msec (+35.6% throughput, -25.2% runtime) ==== 16 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=79.8MiB/s (83.6MB/s), 79.8MiB/s-79.8MiB/s (83.6MB/s-83.6MB/s), io=8192MiB (8590MB), run=102694-102694msec After patch: WRITE: bw=107MiB/s (112MB/s), 107MiB/s-107MiB/s (112MB/s-112MB/s), io=8192MiB (8590MB), run=76446-76446msec (+34.1% throughput, -25.6% runtime) ==== 32 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=93.2MiB/s (97.7MB/s), 93.2MiB/s-93.2MiB/s (97.7MB/s-97.7MB/s), io=16.0GiB (17.2GB), run=175836-175836msec After patch: WRITE: bw=111MiB/s (117MB/s), 111MiB/s-111MiB/s (117MB/s-117MB/s), io=16.0GiB (17.2GB), run=147001-147001msec (+19.1% throughput, -16.4% runtime) ==== 64 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=108MiB/s (114MB/s), 108MiB/s-108MiB/s (114MB/s-114MB/s), io=32.0GiB (34.4GB), run=302656-302656msec After patch: WRITE: bw=133MiB/s (140MB/s), 133MiB/s-133MiB/s (140MB/s-140MB/s), io=32.0GiB (34.4GB), run=246003-246003msec (+23.1% throughput, -18.7% runtime) ************************ *** random writes *** ************************ ==== 1 job, 8GiB file, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=11.5MiB/s (12.0MB/s), 11.5MiB/s-11.5MiB/s (12.0MB/s-12.0MB/s), io=8192MiB (8590MB), run=714281-714281msec After patch: WRITE: bw=11.6MiB/s (12.2MB/s), 11.6MiB/s-11.6MiB/s (12.2MB/s-12.2MB/s), io=8192MiB (8590MB), run=705959-705959msec (+0.9% throughput, -1.7% runtime) ==== 2 jobs, 4GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=12.8MiB/s (13.5MB/s), 12.8MiB/s-12.8MiB/s (13.5MB/s-13.5MB/s), io=8192MiB (8590MB), run=638101-638101msec After patch: WRITE: bw=13.1MiB/s (13.7MB/s), 13.1MiB/s-13.1MiB/s (13.7MB/s-13.7MB/s), io=8192MiB (8590MB), run=625374-625374msec (+2.3% throughput, -2.0% runtime) ==== 4 jobs, 2GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=15.4MiB/s (16.2MB/s), 15.4MiB/s-15.4MiB/s (16.2MB/s-16.2MB/s), io=8192MiB (8590MB), run=531146-531146msec After patch: WRITE: bw=17.8MiB/s (18.7MB/s), 17.8MiB/s-17.8MiB/s (18.7MB/s-18.7MB/s), io=8192MiB (8590MB), run=460431-460431msec (+15.6% throughput, -13.3% runtime) ==== 8 jobs, 1GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=19.9MiB/s (20.8MB/s), 19.9MiB/s-19.9MiB/s (20.8MB/s-20.8MB/s), io=8192MiB (8590MB), run=412664-412664msec After patch: WRITE: bw=22.2MiB/s (23.3MB/s), 22.2MiB/s-22.2MiB/s (23.3MB/s-23.3MB/s), io=8192MiB (8590MB), run=368589-368589msec (+11.6% throughput, -10.7% runtime) ==== 16 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=29.3MiB/s (30.7MB/s), 29.3MiB/s-29.3MiB/s (30.7MB/s-30.7MB/s), io=8192MiB (8590MB), run=279924-279924msec After patch: WRITE: bw=30.4MiB/s (31.9MB/s), 30.4MiB/s-30.4MiB/s (31.9MB/s-31.9MB/s), io=8192MiB (8590MB), run=269258-269258msec (+3.8% throughput, -3.8% runtime) ==== 32 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=36.9MiB/s (38.7MB/s), 36.9MiB/s-36.9MiB/s (38.7MB/s-38.7MB/s), io=16.0GiB (17.2GB), run=443581-443581msec After patch: WRITE: bw=41.6MiB/s (43.6MB/s), 41.6MiB/s-41.6MiB/s (43.6MB/s-43.6MB/s), io=16.0GiB (17.2GB), run=394114-394114msec (+12.7% throughput, -11.2% runtime) ==== 64 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=45.9MiB/s (48.1MB/s), 45.9MiB/s-45.9MiB/s (48.1MB/s-48.1MB/s), io=32.0GiB (34.4GB), run=714614-714614msec After patch: WRITE: bw=48.8MiB/s (51.1MB/s), 48.8MiB/s-48.8MiB/s (51.1MB/s-51.1MB/s), io=32.0GiB (34.4GB), run=672087-672087msec (+6.3% throughput, -6.0% runtime) Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-08-11 14:43:58 +03:00
LOG_INODE_ALL, ctx);
if (!ret && ctx->log_new_dentries)
Btrfs: fix for incorrect directory entries after fsync log replay If we move a directory to a new parent and later log that parent and don't explicitly log the old parent, when we replay the log we can end up with entries for the moved directory in both the old and new parent directories. Besides being ilegal to have directories with multiple hard links in linux, it also resulted in the leaving the inode item with a link count of 1. A similar issue also happens if we move a regular file - after the log tree is replayed the file has a link in both the old and new parent directories, when it should be only at the new directory. Sample reproducer: $ mkfs.btrfs -f /dev/sdc $ mount /dev/sdc /mnt $ mkdir /mnt/x $ mkdir /mnt/y $ touch /mnt/x/foo $ mkdir /mnt/y/z $ sync $ ln /mnt/x/foo /mnt/x/bar $ mv /mnt/y/z /mnt/x/z < power fail > $ mount /dev/sdc /mnt $ ls -1Ri /mnt /mnt: 257 x 258 y /mnt/x: 259 bar 259 foo 260 z /mnt/x/z: /mnt/y: 260 z /mnt/y/z: $ umount /dev/sdc $ btrfs check /dev/sdc Checking filesystem on /dev/sdc UUID: a67e2c4a-a4b4-4fdc-b015-9d9af1e344be checking extents checking free space cache checking fs roots root 5 inode 260 errors 2000, link count wrong unresolved ref dir 257 index 4 namelen 1 name z filetype 2 errors 0 unresolved ref dir 258 index 2 namelen 1 name z filetype 2 errors 0 (...) Attempting to remove the directory becomes impossible: $ mount /dev/sdc /mnt $ rmdir /mnt/y/z $ ls -lh /mnt/y ls: cannot access /mnt/y/z: No such file or directory total 0 d????????? ? ? ? ? ? z $ rmdir /mnt/x/z rmdir: failed to remove ‘/mnt/x/z’: Stale file handle $ ls -lh /mnt/x ls: cannot access /mnt/x/z: Stale file handle total 0 -rw-r--r-- 2 root root 0 Apr 6 18:06 bar -rw-r--r-- 2 root root 0 Apr 6 18:06 foo d????????? ? ? ? ? ? z So make sure that on rename we set the last_unlink_trans value for our inode, even if it's a directory, to the value of the current transaction's ID and that if the new parent directory is logged that we fallback to a transaction commit. A test case for fstests is being submitted as well. Signed-off-by: Filipe Manana <fdmanana@suse.com>
2016-04-06 19:11:56 +03:00
ret = log_new_dir_dentries(trans, root,
BTRFS_I(dir_inode), ctx);
Btrfs: fix assertion failure during fsync and use of stale transaction Sometimes when fsync'ing a file we need to log that other inodes exist and when we need to do that we acquire a reference on the inodes and then drop that reference using iput() after logging them. That generally is not a problem except if we end up doing the final iput() (dropping the last reference) on the inode and that inode has a link count of 0, which can happen in a very short time window if the logging path gets a reference on the inode while it's being unlinked. In that case we end up getting the eviction callback, btrfs_evict_inode(), invoked through the iput() call chain which needs to drop all of the inode's items from its subvolume btree, and in order to do that, it needs to join a transaction at the helper function evict_refill_and_join(). However because the task previously started a transaction at the fsync handler, btrfs_sync_file(), it has current->journal_info already pointing to a transaction handle and therefore evict_refill_and_join() will get that transaction handle from btrfs_join_transaction(). From this point on, two different problems can happen: 1) evict_refill_and_join() will often change the transaction handle's block reserve (->block_rsv) and set its ->bytes_reserved field to a value greater than 0. If evict_refill_and_join() never commits the transaction, the eviction handler ends up decreasing the reference count (->use_count) of the transaction handle through the call to btrfs_end_transaction(), and after that point we have a transaction handle with a NULL ->block_rsv (which is the value prior to the transaction join from evict_refill_and_join()) and a ->bytes_reserved value greater than 0. If after the eviction/iput completes the inode logging path hits an error or it decides that it must fallback to a transaction commit, the btrfs fsync handle, btrfs_sync_file(), gets a non-zero value from btrfs_log_dentry_safe(), and because of that non-zero value it tries to commit the transaction using a handle with a NULL ->block_rsv and a non-zero ->bytes_reserved value. This makes the transaction commit hit an assertion failure at btrfs_trans_release_metadata() because ->bytes_reserved is not zero but the ->block_rsv is NULL. The produced stack trace for that is like the following: [192922.917158] assertion failed: !trans->bytes_reserved, file: fs/btrfs/transaction.c, line: 816 [192922.917553] ------------[ cut here ]------------ [192922.917922] kernel BUG at fs/btrfs/ctree.h:3532! [192922.918310] invalid opcode: 0000 [#1] SMP DEBUG_PAGEALLOC PTI [192922.918666] CPU: 2 PID: 883 Comm: fsstress Tainted: G W 5.1.4-btrfs-next-47 #1 [192922.919035] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.11.2-0-gf9626ccb91-prebuilt.qemu-project.org 04/01/2014 [192922.919801] RIP: 0010:assfail.constprop.25+0x18/0x1a [btrfs] (...) [192922.920925] RSP: 0018:ffffaebdc8a27da8 EFLAGS: 00010286 [192922.921315] RAX: 0000000000000051 RBX: ffff95c9c16a41c0 RCX: 0000000000000000 [192922.921692] RDX: 0000000000000000 RSI: ffff95cab6b16838 RDI: ffff95cab6b16838 [192922.922066] RBP: ffff95c9c16a41c0 R08: 0000000000000000 R09: 0000000000000000 [192922.922442] R10: ffffaebdc8a27e70 R11: 0000000000000000 R12: ffff95ca731a0980 [192922.922820] R13: 0000000000000000 R14: ffff95ca84c73338 R15: ffff95ca731a0ea8 [192922.923200] FS: 00007f337eda4e80(0000) GS:ffff95cab6b00000(0000) knlGS:0000000000000000 [192922.923579] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [192922.923948] CR2: 00007f337edad000 CR3: 00000001e00f6002 CR4: 00000000003606e0 [192922.924329] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [192922.924711] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [192922.925105] Call Trace: [192922.925505] btrfs_trans_release_metadata+0x10c/0x170 [btrfs] [192922.925911] btrfs_commit_transaction+0x3e/0xaf0 [btrfs] [192922.926324] btrfs_sync_file+0x44c/0x490 [btrfs] [192922.926731] do_fsync+0x38/0x60 [192922.927138] __x64_sys_fdatasync+0x13/0x20 [192922.927543] do_syscall_64+0x60/0x1c0 [192922.927939] entry_SYSCALL_64_after_hwframe+0x49/0xbe (...) [192922.934077] ---[ end trace f00808b12068168f ]--- 2) If evict_refill_and_join() decides to commit the transaction, it will be able to do it, since the nested transaction join only increments the transaction handle's ->use_count reference counter and it does not prevent the transaction from getting committed. This means that after eviction completes, the fsync logging path will be using a transaction handle that refers to an already committed transaction. What happens when using such a stale transaction can be unpredictable, we are at least having a use-after-free on the transaction handle itself, since the transaction commit will call kmem_cache_free() against the handle regardless of its ->use_count value, or we can end up silently losing all the updates to the log tree after that iput() in the logging path, or using a transaction handle that in the meanwhile was allocated to another task for a new transaction, etc, pretty much unpredictable what can happen. In order to fix both of them, instead of using iput() during logging, use btrfs_add_delayed_iput(), so that the logging path of fsync never drops the last reference on an inode, that step is offloaded to a safe context (usually the cleaner kthread). The assertion failure issue was sporadically triggered by the test case generic/475 from fstests, which loads the dm error target while fsstress is running, which lead to fsync failing while logging inodes with -EIO errors and then trying later to commit the transaction, triggering the assertion failure. CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-09-10 17:26:49 +03:00
btrfs_add_delayed_iput(dir_inode);
Btrfs: fix stale dir entries after removing a link and fsync We have one more case where after a log tree is replayed we get inconsistent metadata leading to stale directory entries, due to some directories having entries pointing to some inode while the inode does not have a matching BTRFS_INODE_[REF|EXTREF]_KEY item. To trigger the problem we need to have a file with multiple hard links belonging to different parent directories. Then if one of those hard links is removed and we fsync the file using one of its other links that belongs to a different parent directory, we end up not logging the fact that the removed hard link doesn't exists anymore in the parent directory. Simple reproducer: seq=`basename $0` seqres=$RESULT_DIR/$seq echo "QA output created by $seq" tmp=/tmp/$$ status=1 # failure is the default! trap "_cleanup; exit \$status" 0 1 2 3 15 _cleanup() { _cleanup_flakey rm -f $tmp.* } # get standard environment, filters and checks . ./common/rc . ./common/filter . ./common/dmflakey # real QA test starts here _need_to_be_root _supported_fs generic _supported_os Linux _require_scratch _require_dm_flakey _require_metadata_journaling $SCRATCH_DEV rm -f $seqres.full _scratch_mkfs >>$seqres.full 2>&1 _init_flakey _mount_flakey # Create our test directory and file. mkdir $SCRATCH_MNT/testdir touch $SCRATCH_MNT/foo ln $SCRATCH_MNT/foo $SCRATCH_MNT/testdir/foo2 ln $SCRATCH_MNT/foo $SCRATCH_MNT/testdir/foo3 # Make sure everything done so far is durably persisted. sync # Now we remove one of our file's hardlinks in the directory testdir. unlink $SCRATCH_MNT/testdir/foo3 # We now fsync our file using the "foo" link, which has a parent that # is not the directory "testdir". $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foo # Silently drop all writes and unmount to simulate a crash/power # failure. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again, mount to trigger journal/log replay. _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # After the journal/log is replayed we expect to not see the "foo3" # link anymore and we should be able to remove all names in the # directory "testdir" and then remove it (no stale directory entries # left after the journal/log replay). echo "Entries in testdir:" ls -1 $SCRATCH_MNT/testdir rm -f $SCRATCH_MNT/testdir/* rmdir $SCRATCH_MNT/testdir _unmount_flakey status=0 exit The test fails with: $ ./check generic/107 FSTYP -- btrfs PLATFORM -- Linux/x86_64 debian3 4.1.0-rc6-btrfs-next-11+ MKFS_OPTIONS -- /dev/sdc MOUNT_OPTIONS -- /dev/sdc /home/fdmanana/btrfs-tests/scratch_1 generic/107 3s ... - output mismatch (see .../results/generic/107.out.bad) --- tests/generic/107.out 2015-08-01 01:39:45.807462161 +0100 +++ /home/fdmanana/git/hub/xfstests/results//generic/107.out.bad @@ -1,3 +1,5 @@ QA output created by 107 Entries in testdir: foo2 +foo3 +rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/testdir': Directory not empty ... _check_btrfs_filesystem: filesystem on /dev/sdc is inconsistent \ (see /home/fdmanana/git/hub/xfstests/results//generic/107.full) _check_dmesg: something found in dmesg (see .../results/generic/107.dmesg) Ran: generic/107 Failures: generic/107 Failed 1 of 1 tests $ cat /home/fdmanana/git/hub/xfstests/results//generic/107.full (...) checking fs roots root 5 inode 257 errors 200, dir isize wrong unresolved ref dir 257 index 3 namelen 4 name foo3 filetype 1 errors 5, no dir item, no inode ref (...) And produces the following warning in dmesg: [127298.759064] BTRFS info (device dm-0): failed to delete reference to foo3, inode 258 parent 257 [127298.762081] ------------[ cut here ]------------ [127298.763311] WARNING: CPU: 10 PID: 7891 at fs/btrfs/inode.c:3956 __btrfs_unlink_inode+0x182/0x35a [btrfs]() [127298.767327] BTRFS: Transaction aborted (error -2) (...) [127298.788611] Call Trace: [127298.789137] [<ffffffff8145f077>] dump_stack+0x4f/0x7b [127298.790090] [<ffffffff81095de5>] ? console_unlock+0x356/0x3a2 [127298.791157] [<ffffffff8104b3b0>] warn_slowpath_common+0xa1/0xbb [127298.792323] [<ffffffffa065ad09>] ? __btrfs_unlink_inode+0x182/0x35a [btrfs] [127298.793633] [<ffffffff8104b410>] warn_slowpath_fmt+0x46/0x48 [127298.794699] [<ffffffffa065ad09>] __btrfs_unlink_inode+0x182/0x35a [btrfs] [127298.797640] [<ffffffffa065be8f>] btrfs_unlink_inode+0x1e/0x40 [btrfs] [127298.798876] [<ffffffffa065bf11>] btrfs_unlink+0x60/0x9b [btrfs] [127298.800154] [<ffffffff8116fb48>] vfs_unlink+0x9c/0xed [127298.801303] [<ffffffff81173481>] do_unlinkat+0x12b/0x1fb [127298.802450] [<ffffffff81253855>] ? lockdep_sys_exit_thunk+0x12/0x14 [127298.803797] [<ffffffff81174056>] SyS_unlinkat+0x29/0x2b [127298.805017] [<ffffffff81465197>] system_call_fastpath+0x12/0x6f [127298.806310] ---[ end trace bbfddacb7aaada7b ]--- [127298.807325] BTRFS warning (device dm-0): __btrfs_unlink_inode:3956: Aborting unused transaction(No such entry). So fix this by logging all parent inodes, current and old ones, to make sure we do not get stale entries after log replay. This is not a simple solution such as triggering a full transaction commit because it would imply full transaction commit when an inode is fsynced in the same transaction that modified it and reloaded it after eviction (because its last_unlink_trans is set to the same value as its last_trans as of the commit with the title "Btrfs: fix stale dir entries after unlink, inode eviction and fsync"), and it would also make fstest generic/066 fail since one of the fsyncs triggers a full commit and the next fsync will not find the inode in the log anymore (therefore not removing the xattr). Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-08-05 18:49:08 +03:00
if (ret)
goto out;
}
path->slots[0]++;
}
ret = 0;
out:
btrfs_free_path(path);
return ret;
}
static int log_new_ancestors(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct btrfs_path *path,
struct btrfs_log_ctx *ctx)
{
struct btrfs_key found_key;
btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
while (true) {
struct btrfs_fs_info *fs_info = root->fs_info;
struct extent_buffer *leaf = path->nodes[0];
int slot = path->slots[0];
struct btrfs_key search_key;
struct inode *inode;
u64 ino;
int ret = 0;
btrfs_release_path(path);
ino = found_key.offset;
search_key.objectid = found_key.offset;
search_key.type = BTRFS_INODE_ITEM_KEY;
search_key.offset = 0;
inode = btrfs_iget(fs_info->sb, ino, root);
if (IS_ERR(inode))
return PTR_ERR(inode);
if (BTRFS_I(inode)->generation >= trans->transid &&
need_log_inode(trans, BTRFS_I(inode)))
ret = btrfs_log_inode(trans, BTRFS_I(inode),
btrfs: make fast fsyncs wait only for writeback Currently regardless of a full or a fast fsync we always wait for ordered extents to complete, and then start logging the inode after that. However for fast fsyncs we can just wait for the writeback to complete, we don't need to wait for the ordered extents to complete since we use the list of modified extents maps to figure out which extents we must log and we can get their checksums directly from the ordered extents that are still in flight, otherwise look them up from the checksums tree. Until commit b5e6c3e170b770 ("btrfs: always wait on ordered extents at fsync time"), for fast fsyncs, we used to start logging without even waiting for the writeback to complete first, we would wait for it to complete after logging, while holding a transaction open, which lead to performance issues when using cgroups and probably for other cases too, as wait for IO while holding a transaction handle should be avoided as much as possible. After that, for fast fsyncs, we started to wait for ordered extents to complete before starting to log, which adds some latency to fsyncs and we even got at least one report about a performance drop which bisected to that particular change: https://lore.kernel.org/linux-btrfs/20181109215148.GF23260@techsingularity.net/ This change makes fast fsyncs only wait for writeback to finish before starting to log the inode, instead of waiting for both the writeback to finish and for the ordered extents to complete. This brings back part of the logic we had that extracts checksums from in flight ordered extents, which are not yet in the checksums tree, and making sure transaction commits wait for the completion of ordered extents previously logged (by far most of the time they have already completed by the time a transaction commit starts, resulting in no wait at all), to avoid any data loss if an ordered extent completes after the transaction used to log an inode is committed, followed by a power failure. When there are no other tasks accessing the checksums and the subvolume btrees, the ordered extent completion is pretty fast, typically taking 100 to 200 microseconds only in my observations. However when there are other tasks accessing these btrees, ordered extent completion can take a lot more time due to lock contention on nodes and leaves of these btrees. I've seen cases over 2 milliseconds, which starts to be significant. In particular when we do have concurrent fsyncs against different files there is a lot of contention on the checksums btree, since we have many tasks writing the checksums into the btree and other tasks that already started the logging phase are doing lookups for checksums in the btree. This change also turns all ranged fsyncs into full ranged fsyncs, which is something we already did when not using the NO_HOLES features or when doing a full fsync. This is to guarantee we never miss checksums due to writeback having been triggered only for a part of an extent, and we end up logging the full extent but only checksums for the written range, which results in missing checksums after log replay. Allowing ranged fsyncs to operate again only in the original range, when using the NO_HOLES feature and doing a fast fsync is doable but requires some non trivial changes to the writeback path, which can always be worked on later if needed, but I don't think they are a very common use case. Several tests were performed using fio for different numbers of concurrent jobs, each writing and fsyncing its own file, for both sequential and random file writes. The tests were run on bare metal, no virtualization, on a box with 12 cores (Intel i7-8700), 64Gb of RAM and a NVMe device, with a kernel configuration that is the default of typical distributions (debian in this case), without debug options enabled (kasan, kmemleak, slub debug, debug of page allocations, lock debugging, etc). The following script that calls fio was used: $ cat test-fsync.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/btrfs MOUNT_OPTIONS="-o ssd -o space_cache=v2" MKFS_OPTIONS="-d single -m single" if [ $# -ne 5 ]; then echo "Use $0 NUM_JOBS FILE_SIZE FSYNC_FREQ BLOCK_SIZE [write|randwrite]" exit 1 fi NUM_JOBS=$1 FILE_SIZE=$2 FSYNC_FREQ=$3 BLOCK_SIZE=$4 WRITE_MODE=$5 if [ "$WRITE_MODE" != "write" ] && [ "$WRITE_MODE" != "randwrite" ]; then echo "Invalid WRITE_MODE, must be 'write' or 'randwrite'" exit 1 fi cat <<EOF > /tmp/fio-job.ini [writers] rw=$WRITE_MODE fsync=$FSYNC_FREQ fallocate=none group_reporting=1 direct=0 bs=$BLOCK_SIZE ioengine=sync size=$FILE_SIZE directory=$MNT numjobs=$NUM_JOBS EOF echo "performance" | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor echo echo "Using config:" echo cat /tmp/fio-job.ini echo umount $MNT &> /dev/null mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT fio /tmp/fio-job.ini umount $MNT The results were the following: ************************* *** sequential writes *** ************************* ==== 1 job, 8GiB file, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=36.6MiB/s (38.4MB/s), 36.6MiB/s-36.6MiB/s (38.4MB/s-38.4MB/s), io=8192MiB (8590MB), run=223689-223689msec After patch: WRITE: bw=40.2MiB/s (42.1MB/s), 40.2MiB/s-40.2MiB/s (42.1MB/s-42.1MB/s), io=8192MiB (8590MB), run=203980-203980msec (+9.8%, -8.8% runtime) ==== 2 jobs, 4GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=35.8MiB/s (37.5MB/s), 35.8MiB/s-35.8MiB/s (37.5MB/s-37.5MB/s), io=8192MiB (8590MB), run=228950-228950msec After patch: WRITE: bw=43.5MiB/s (45.6MB/s), 43.5MiB/s-43.5MiB/s (45.6MB/s-45.6MB/s), io=8192MiB (8590MB), run=188272-188272msec (+21.5% throughput, -17.8% runtime) ==== 4 jobs, 2GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=50.1MiB/s (52.6MB/s), 50.1MiB/s-50.1MiB/s (52.6MB/s-52.6MB/s), io=8192MiB (8590MB), run=163446-163446msec After patch: WRITE: bw=64.5MiB/s (67.6MB/s), 64.5MiB/s-64.5MiB/s (67.6MB/s-67.6MB/s), io=8192MiB (8590MB), run=126987-126987msec (+28.7% throughput, -22.3% runtime) ==== 8 jobs, 1GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=64.0MiB/s (68.1MB/s), 64.0MiB/s-64.0MiB/s (68.1MB/s-68.1MB/s), io=8192MiB (8590MB), run=126075-126075msec After patch: WRITE: bw=86.8MiB/s (91.0MB/s), 86.8MiB/s-86.8MiB/s (91.0MB/s-91.0MB/s), io=8192MiB (8590MB), run=94358-94358msec (+35.6% throughput, -25.2% runtime) ==== 16 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=79.8MiB/s (83.6MB/s), 79.8MiB/s-79.8MiB/s (83.6MB/s-83.6MB/s), io=8192MiB (8590MB), run=102694-102694msec After patch: WRITE: bw=107MiB/s (112MB/s), 107MiB/s-107MiB/s (112MB/s-112MB/s), io=8192MiB (8590MB), run=76446-76446msec (+34.1% throughput, -25.6% runtime) ==== 32 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=93.2MiB/s (97.7MB/s), 93.2MiB/s-93.2MiB/s (97.7MB/s-97.7MB/s), io=16.0GiB (17.2GB), run=175836-175836msec After patch: WRITE: bw=111MiB/s (117MB/s), 111MiB/s-111MiB/s (117MB/s-117MB/s), io=16.0GiB (17.2GB), run=147001-147001msec (+19.1% throughput, -16.4% runtime) ==== 64 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=108MiB/s (114MB/s), 108MiB/s-108MiB/s (114MB/s-114MB/s), io=32.0GiB (34.4GB), run=302656-302656msec After patch: WRITE: bw=133MiB/s (140MB/s), 133MiB/s-133MiB/s (140MB/s-140MB/s), io=32.0GiB (34.4GB), run=246003-246003msec (+23.1% throughput, -18.7% runtime) ************************ *** random writes *** ************************ ==== 1 job, 8GiB file, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=11.5MiB/s (12.0MB/s), 11.5MiB/s-11.5MiB/s (12.0MB/s-12.0MB/s), io=8192MiB (8590MB), run=714281-714281msec After patch: WRITE: bw=11.6MiB/s (12.2MB/s), 11.6MiB/s-11.6MiB/s (12.2MB/s-12.2MB/s), io=8192MiB (8590MB), run=705959-705959msec (+0.9% throughput, -1.7% runtime) ==== 2 jobs, 4GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=12.8MiB/s (13.5MB/s), 12.8MiB/s-12.8MiB/s (13.5MB/s-13.5MB/s), io=8192MiB (8590MB), run=638101-638101msec After patch: WRITE: bw=13.1MiB/s (13.7MB/s), 13.1MiB/s-13.1MiB/s (13.7MB/s-13.7MB/s), io=8192MiB (8590MB), run=625374-625374msec (+2.3% throughput, -2.0% runtime) ==== 4 jobs, 2GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=15.4MiB/s (16.2MB/s), 15.4MiB/s-15.4MiB/s (16.2MB/s-16.2MB/s), io=8192MiB (8590MB), run=531146-531146msec After patch: WRITE: bw=17.8MiB/s (18.7MB/s), 17.8MiB/s-17.8MiB/s (18.7MB/s-18.7MB/s), io=8192MiB (8590MB), run=460431-460431msec (+15.6% throughput, -13.3% runtime) ==== 8 jobs, 1GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=19.9MiB/s (20.8MB/s), 19.9MiB/s-19.9MiB/s (20.8MB/s-20.8MB/s), io=8192MiB (8590MB), run=412664-412664msec After patch: WRITE: bw=22.2MiB/s (23.3MB/s), 22.2MiB/s-22.2MiB/s (23.3MB/s-23.3MB/s), io=8192MiB (8590MB), run=368589-368589msec (+11.6% throughput, -10.7% runtime) ==== 16 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=29.3MiB/s (30.7MB/s), 29.3MiB/s-29.3MiB/s (30.7MB/s-30.7MB/s), io=8192MiB (8590MB), run=279924-279924msec After patch: WRITE: bw=30.4MiB/s (31.9MB/s), 30.4MiB/s-30.4MiB/s (31.9MB/s-31.9MB/s), io=8192MiB (8590MB), run=269258-269258msec (+3.8% throughput, -3.8% runtime) ==== 32 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=36.9MiB/s (38.7MB/s), 36.9MiB/s-36.9MiB/s (38.7MB/s-38.7MB/s), io=16.0GiB (17.2GB), run=443581-443581msec After patch: WRITE: bw=41.6MiB/s (43.6MB/s), 41.6MiB/s-41.6MiB/s (43.6MB/s-43.6MB/s), io=16.0GiB (17.2GB), run=394114-394114msec (+12.7% throughput, -11.2% runtime) ==== 64 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=45.9MiB/s (48.1MB/s), 45.9MiB/s-45.9MiB/s (48.1MB/s-48.1MB/s), io=32.0GiB (34.4GB), run=714614-714614msec After patch: WRITE: bw=48.8MiB/s (51.1MB/s), 48.8MiB/s-48.8MiB/s (51.1MB/s-51.1MB/s), io=32.0GiB (34.4GB), run=672087-672087msec (+6.3% throughput, -6.0% runtime) Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-08-11 14:43:58 +03:00
LOG_INODE_EXISTS, ctx);
Btrfs: fix assertion failure during fsync and use of stale transaction Sometimes when fsync'ing a file we need to log that other inodes exist and when we need to do that we acquire a reference on the inodes and then drop that reference using iput() after logging them. That generally is not a problem except if we end up doing the final iput() (dropping the last reference) on the inode and that inode has a link count of 0, which can happen in a very short time window if the logging path gets a reference on the inode while it's being unlinked. In that case we end up getting the eviction callback, btrfs_evict_inode(), invoked through the iput() call chain which needs to drop all of the inode's items from its subvolume btree, and in order to do that, it needs to join a transaction at the helper function evict_refill_and_join(). However because the task previously started a transaction at the fsync handler, btrfs_sync_file(), it has current->journal_info already pointing to a transaction handle and therefore evict_refill_and_join() will get that transaction handle from btrfs_join_transaction(). From this point on, two different problems can happen: 1) evict_refill_and_join() will often change the transaction handle's block reserve (->block_rsv) and set its ->bytes_reserved field to a value greater than 0. If evict_refill_and_join() never commits the transaction, the eviction handler ends up decreasing the reference count (->use_count) of the transaction handle through the call to btrfs_end_transaction(), and after that point we have a transaction handle with a NULL ->block_rsv (which is the value prior to the transaction join from evict_refill_and_join()) and a ->bytes_reserved value greater than 0. If after the eviction/iput completes the inode logging path hits an error or it decides that it must fallback to a transaction commit, the btrfs fsync handle, btrfs_sync_file(), gets a non-zero value from btrfs_log_dentry_safe(), and because of that non-zero value it tries to commit the transaction using a handle with a NULL ->block_rsv and a non-zero ->bytes_reserved value. This makes the transaction commit hit an assertion failure at btrfs_trans_release_metadata() because ->bytes_reserved is not zero but the ->block_rsv is NULL. The produced stack trace for that is like the following: [192922.917158] assertion failed: !trans->bytes_reserved, file: fs/btrfs/transaction.c, line: 816 [192922.917553] ------------[ cut here ]------------ [192922.917922] kernel BUG at fs/btrfs/ctree.h:3532! [192922.918310] invalid opcode: 0000 [#1] SMP DEBUG_PAGEALLOC PTI [192922.918666] CPU: 2 PID: 883 Comm: fsstress Tainted: G W 5.1.4-btrfs-next-47 #1 [192922.919035] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.11.2-0-gf9626ccb91-prebuilt.qemu-project.org 04/01/2014 [192922.919801] RIP: 0010:assfail.constprop.25+0x18/0x1a [btrfs] (...) [192922.920925] RSP: 0018:ffffaebdc8a27da8 EFLAGS: 00010286 [192922.921315] RAX: 0000000000000051 RBX: ffff95c9c16a41c0 RCX: 0000000000000000 [192922.921692] RDX: 0000000000000000 RSI: ffff95cab6b16838 RDI: ffff95cab6b16838 [192922.922066] RBP: ffff95c9c16a41c0 R08: 0000000000000000 R09: 0000000000000000 [192922.922442] R10: ffffaebdc8a27e70 R11: 0000000000000000 R12: ffff95ca731a0980 [192922.922820] R13: 0000000000000000 R14: ffff95ca84c73338 R15: ffff95ca731a0ea8 [192922.923200] FS: 00007f337eda4e80(0000) GS:ffff95cab6b00000(0000) knlGS:0000000000000000 [192922.923579] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [192922.923948] CR2: 00007f337edad000 CR3: 00000001e00f6002 CR4: 00000000003606e0 [192922.924329] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [192922.924711] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [192922.925105] Call Trace: [192922.925505] btrfs_trans_release_metadata+0x10c/0x170 [btrfs] [192922.925911] btrfs_commit_transaction+0x3e/0xaf0 [btrfs] [192922.926324] btrfs_sync_file+0x44c/0x490 [btrfs] [192922.926731] do_fsync+0x38/0x60 [192922.927138] __x64_sys_fdatasync+0x13/0x20 [192922.927543] do_syscall_64+0x60/0x1c0 [192922.927939] entry_SYSCALL_64_after_hwframe+0x49/0xbe (...) [192922.934077] ---[ end trace f00808b12068168f ]--- 2) If evict_refill_and_join() decides to commit the transaction, it will be able to do it, since the nested transaction join only increments the transaction handle's ->use_count reference counter and it does not prevent the transaction from getting committed. This means that after eviction completes, the fsync logging path will be using a transaction handle that refers to an already committed transaction. What happens when using such a stale transaction can be unpredictable, we are at least having a use-after-free on the transaction handle itself, since the transaction commit will call kmem_cache_free() against the handle regardless of its ->use_count value, or we can end up silently losing all the updates to the log tree after that iput() in the logging path, or using a transaction handle that in the meanwhile was allocated to another task for a new transaction, etc, pretty much unpredictable what can happen. In order to fix both of them, instead of using iput() during logging, use btrfs_add_delayed_iput(), so that the logging path of fsync never drops the last reference on an inode, that step is offloaded to a safe context (usually the cleaner kthread). The assertion failure issue was sporadically triggered by the test case generic/475 from fstests, which loads the dm error target while fsstress is running, which lead to fsync failing while logging inodes with -EIO errors and then trying later to commit the transaction, triggering the assertion failure. CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-09-10 17:26:49 +03:00
btrfs_add_delayed_iput(inode);
if (ret)
return ret;
if (search_key.objectid == BTRFS_FIRST_FREE_OBJECTID)
break;
search_key.type = BTRFS_INODE_REF_KEY;
ret = btrfs_search_slot(NULL, root, &search_key, path, 0, 0);
if (ret < 0)
return ret;
leaf = path->nodes[0];
slot = path->slots[0];
if (slot >= btrfs_header_nritems(leaf)) {
ret = btrfs_next_leaf(root, path);
if (ret < 0)
return ret;
else if (ret > 0)
return -ENOENT;
leaf = path->nodes[0];
slot = path->slots[0];
}
btrfs_item_key_to_cpu(leaf, &found_key, slot);
if (found_key.objectid != search_key.objectid ||
found_key.type != BTRFS_INODE_REF_KEY)
return -ENOENT;
}
return 0;
}
static int log_new_ancestors_fast(struct btrfs_trans_handle *trans,
struct btrfs_inode *inode,
struct dentry *parent,
struct btrfs_log_ctx *ctx)
{
struct btrfs_root *root = inode->root;
struct dentry *old_parent = NULL;
struct super_block *sb = inode->vfs_inode.i_sb;
int ret = 0;
while (true) {
if (!parent || d_really_is_negative(parent) ||
sb != parent->d_sb)
break;
inode = BTRFS_I(d_inode(parent));
if (root != inode->root)
break;
if (inode->generation >= trans->transid &&
need_log_inode(trans, inode)) {
ret = btrfs_log_inode(trans, inode,
btrfs: make fast fsyncs wait only for writeback Currently regardless of a full or a fast fsync we always wait for ordered extents to complete, and then start logging the inode after that. However for fast fsyncs we can just wait for the writeback to complete, we don't need to wait for the ordered extents to complete since we use the list of modified extents maps to figure out which extents we must log and we can get their checksums directly from the ordered extents that are still in flight, otherwise look them up from the checksums tree. Until commit b5e6c3e170b770 ("btrfs: always wait on ordered extents at fsync time"), for fast fsyncs, we used to start logging without even waiting for the writeback to complete first, we would wait for it to complete after logging, while holding a transaction open, which lead to performance issues when using cgroups and probably for other cases too, as wait for IO while holding a transaction handle should be avoided as much as possible. After that, for fast fsyncs, we started to wait for ordered extents to complete before starting to log, which adds some latency to fsyncs and we even got at least one report about a performance drop which bisected to that particular change: https://lore.kernel.org/linux-btrfs/20181109215148.GF23260@techsingularity.net/ This change makes fast fsyncs only wait for writeback to finish before starting to log the inode, instead of waiting for both the writeback to finish and for the ordered extents to complete. This brings back part of the logic we had that extracts checksums from in flight ordered extents, which are not yet in the checksums tree, and making sure transaction commits wait for the completion of ordered extents previously logged (by far most of the time they have already completed by the time a transaction commit starts, resulting in no wait at all), to avoid any data loss if an ordered extent completes after the transaction used to log an inode is committed, followed by a power failure. When there are no other tasks accessing the checksums and the subvolume btrees, the ordered extent completion is pretty fast, typically taking 100 to 200 microseconds only in my observations. However when there are other tasks accessing these btrees, ordered extent completion can take a lot more time due to lock contention on nodes and leaves of these btrees. I've seen cases over 2 milliseconds, which starts to be significant. In particular when we do have concurrent fsyncs against different files there is a lot of contention on the checksums btree, since we have many tasks writing the checksums into the btree and other tasks that already started the logging phase are doing lookups for checksums in the btree. This change also turns all ranged fsyncs into full ranged fsyncs, which is something we already did when not using the NO_HOLES features or when doing a full fsync. This is to guarantee we never miss checksums due to writeback having been triggered only for a part of an extent, and we end up logging the full extent but only checksums for the written range, which results in missing checksums after log replay. Allowing ranged fsyncs to operate again only in the original range, when using the NO_HOLES feature and doing a fast fsync is doable but requires some non trivial changes to the writeback path, which can always be worked on later if needed, but I don't think they are a very common use case. Several tests were performed using fio for different numbers of concurrent jobs, each writing and fsyncing its own file, for both sequential and random file writes. The tests were run on bare metal, no virtualization, on a box with 12 cores (Intel i7-8700), 64Gb of RAM and a NVMe device, with a kernel configuration that is the default of typical distributions (debian in this case), without debug options enabled (kasan, kmemleak, slub debug, debug of page allocations, lock debugging, etc). The following script that calls fio was used: $ cat test-fsync.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/btrfs MOUNT_OPTIONS="-o ssd -o space_cache=v2" MKFS_OPTIONS="-d single -m single" if [ $# -ne 5 ]; then echo "Use $0 NUM_JOBS FILE_SIZE FSYNC_FREQ BLOCK_SIZE [write|randwrite]" exit 1 fi NUM_JOBS=$1 FILE_SIZE=$2 FSYNC_FREQ=$3 BLOCK_SIZE=$4 WRITE_MODE=$5 if [ "$WRITE_MODE" != "write" ] && [ "$WRITE_MODE" != "randwrite" ]; then echo "Invalid WRITE_MODE, must be 'write' or 'randwrite'" exit 1 fi cat <<EOF > /tmp/fio-job.ini [writers] rw=$WRITE_MODE fsync=$FSYNC_FREQ fallocate=none group_reporting=1 direct=0 bs=$BLOCK_SIZE ioengine=sync size=$FILE_SIZE directory=$MNT numjobs=$NUM_JOBS EOF echo "performance" | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor echo echo "Using config:" echo cat /tmp/fio-job.ini echo umount $MNT &> /dev/null mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT fio /tmp/fio-job.ini umount $MNT The results were the following: ************************* *** sequential writes *** ************************* ==== 1 job, 8GiB file, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=36.6MiB/s (38.4MB/s), 36.6MiB/s-36.6MiB/s (38.4MB/s-38.4MB/s), io=8192MiB (8590MB), run=223689-223689msec After patch: WRITE: bw=40.2MiB/s (42.1MB/s), 40.2MiB/s-40.2MiB/s (42.1MB/s-42.1MB/s), io=8192MiB (8590MB), run=203980-203980msec (+9.8%, -8.8% runtime) ==== 2 jobs, 4GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=35.8MiB/s (37.5MB/s), 35.8MiB/s-35.8MiB/s (37.5MB/s-37.5MB/s), io=8192MiB (8590MB), run=228950-228950msec After patch: WRITE: bw=43.5MiB/s (45.6MB/s), 43.5MiB/s-43.5MiB/s (45.6MB/s-45.6MB/s), io=8192MiB (8590MB), run=188272-188272msec (+21.5% throughput, -17.8% runtime) ==== 4 jobs, 2GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=50.1MiB/s (52.6MB/s), 50.1MiB/s-50.1MiB/s (52.6MB/s-52.6MB/s), io=8192MiB (8590MB), run=163446-163446msec After patch: WRITE: bw=64.5MiB/s (67.6MB/s), 64.5MiB/s-64.5MiB/s (67.6MB/s-67.6MB/s), io=8192MiB (8590MB), run=126987-126987msec (+28.7% throughput, -22.3% runtime) ==== 8 jobs, 1GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=64.0MiB/s (68.1MB/s), 64.0MiB/s-64.0MiB/s (68.1MB/s-68.1MB/s), io=8192MiB (8590MB), run=126075-126075msec After patch: WRITE: bw=86.8MiB/s (91.0MB/s), 86.8MiB/s-86.8MiB/s (91.0MB/s-91.0MB/s), io=8192MiB (8590MB), run=94358-94358msec (+35.6% throughput, -25.2% runtime) ==== 16 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=79.8MiB/s (83.6MB/s), 79.8MiB/s-79.8MiB/s (83.6MB/s-83.6MB/s), io=8192MiB (8590MB), run=102694-102694msec After patch: WRITE: bw=107MiB/s (112MB/s), 107MiB/s-107MiB/s (112MB/s-112MB/s), io=8192MiB (8590MB), run=76446-76446msec (+34.1% throughput, -25.6% runtime) ==== 32 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=93.2MiB/s (97.7MB/s), 93.2MiB/s-93.2MiB/s (97.7MB/s-97.7MB/s), io=16.0GiB (17.2GB), run=175836-175836msec After patch: WRITE: bw=111MiB/s (117MB/s), 111MiB/s-111MiB/s (117MB/s-117MB/s), io=16.0GiB (17.2GB), run=147001-147001msec (+19.1% throughput, -16.4% runtime) ==== 64 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=108MiB/s (114MB/s), 108MiB/s-108MiB/s (114MB/s-114MB/s), io=32.0GiB (34.4GB), run=302656-302656msec After patch: WRITE: bw=133MiB/s (140MB/s), 133MiB/s-133MiB/s (140MB/s-140MB/s), io=32.0GiB (34.4GB), run=246003-246003msec (+23.1% throughput, -18.7% runtime) ************************ *** random writes *** ************************ ==== 1 job, 8GiB file, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=11.5MiB/s (12.0MB/s), 11.5MiB/s-11.5MiB/s (12.0MB/s-12.0MB/s), io=8192MiB (8590MB), run=714281-714281msec After patch: WRITE: bw=11.6MiB/s (12.2MB/s), 11.6MiB/s-11.6MiB/s (12.2MB/s-12.2MB/s), io=8192MiB (8590MB), run=705959-705959msec (+0.9% throughput, -1.7% runtime) ==== 2 jobs, 4GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=12.8MiB/s (13.5MB/s), 12.8MiB/s-12.8MiB/s (13.5MB/s-13.5MB/s), io=8192MiB (8590MB), run=638101-638101msec After patch: WRITE: bw=13.1MiB/s (13.7MB/s), 13.1MiB/s-13.1MiB/s (13.7MB/s-13.7MB/s), io=8192MiB (8590MB), run=625374-625374msec (+2.3% throughput, -2.0% runtime) ==== 4 jobs, 2GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=15.4MiB/s (16.2MB/s), 15.4MiB/s-15.4MiB/s (16.2MB/s-16.2MB/s), io=8192MiB (8590MB), run=531146-531146msec After patch: WRITE: bw=17.8MiB/s (18.7MB/s), 17.8MiB/s-17.8MiB/s (18.7MB/s-18.7MB/s), io=8192MiB (8590MB), run=460431-460431msec (+15.6% throughput, -13.3% runtime) ==== 8 jobs, 1GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=19.9MiB/s (20.8MB/s), 19.9MiB/s-19.9MiB/s (20.8MB/s-20.8MB/s), io=8192MiB (8590MB), run=412664-412664msec After patch: WRITE: bw=22.2MiB/s (23.3MB/s), 22.2MiB/s-22.2MiB/s (23.3MB/s-23.3MB/s), io=8192MiB (8590MB), run=368589-368589msec (+11.6% throughput, -10.7% runtime) ==== 16 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=29.3MiB/s (30.7MB/s), 29.3MiB/s-29.3MiB/s (30.7MB/s-30.7MB/s), io=8192MiB (8590MB), run=279924-279924msec After patch: WRITE: bw=30.4MiB/s (31.9MB/s), 30.4MiB/s-30.4MiB/s (31.9MB/s-31.9MB/s), io=8192MiB (8590MB), run=269258-269258msec (+3.8% throughput, -3.8% runtime) ==== 32 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=36.9MiB/s (38.7MB/s), 36.9MiB/s-36.9MiB/s (38.7MB/s-38.7MB/s), io=16.0GiB (17.2GB), run=443581-443581msec After patch: WRITE: bw=41.6MiB/s (43.6MB/s), 41.6MiB/s-41.6MiB/s (43.6MB/s-43.6MB/s), io=16.0GiB (17.2GB), run=394114-394114msec (+12.7% throughput, -11.2% runtime) ==== 64 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=45.9MiB/s (48.1MB/s), 45.9MiB/s-45.9MiB/s (48.1MB/s-48.1MB/s), io=32.0GiB (34.4GB), run=714614-714614msec After patch: WRITE: bw=48.8MiB/s (51.1MB/s), 48.8MiB/s-48.8MiB/s (51.1MB/s-51.1MB/s), io=32.0GiB (34.4GB), run=672087-672087msec (+6.3% throughput, -6.0% runtime) Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-08-11 14:43:58 +03:00
LOG_INODE_EXISTS, ctx);
if (ret)
break;
}
if (IS_ROOT(parent))
break;
parent = dget_parent(parent);
dput(old_parent);
old_parent = parent;
}
dput(old_parent);
return ret;
}
static int log_all_new_ancestors(struct btrfs_trans_handle *trans,
struct btrfs_inode *inode,
struct dentry *parent,
struct btrfs_log_ctx *ctx)
{
struct btrfs_root *root = inode->root;
const u64 ino = btrfs_ino(inode);
struct btrfs_path *path;
struct btrfs_key search_key;
int ret;
/*
* For a single hard link case, go through a fast path that does not
* need to iterate the fs/subvolume tree.
*/
if (inode->vfs_inode.i_nlink < 2)
return log_new_ancestors_fast(trans, inode, parent, ctx);
path = btrfs_alloc_path();
if (!path)
return -ENOMEM;
search_key.objectid = ino;
search_key.type = BTRFS_INODE_REF_KEY;
search_key.offset = 0;
again:
ret = btrfs_search_slot(NULL, root, &search_key, path, 0, 0);
if (ret < 0)
goto out;
if (ret == 0)
path->slots[0]++;
while (true) {
struct extent_buffer *leaf = path->nodes[0];
int slot = path->slots[0];
struct btrfs_key found_key;
if (slot >= btrfs_header_nritems(leaf)) {
ret = btrfs_next_leaf(root, path);
if (ret < 0)
goto out;
else if (ret > 0)
break;
continue;
}
btrfs_item_key_to_cpu(leaf, &found_key, slot);
if (found_key.objectid != ino ||
found_key.type > BTRFS_INODE_EXTREF_KEY)
break;
/*
* Don't deal with extended references because they are rare
* cases and too complex to deal with (we would need to keep
* track of which subitem we are processing for each item in
* this loop, etc). So just return some error to fallback to
* a transaction commit.
*/
if (found_key.type == BTRFS_INODE_EXTREF_KEY) {
ret = -EMLINK;
goto out;
}
/*
* Logging ancestors needs to do more searches on the fs/subvol
* tree, so it releases the path as needed to avoid deadlocks.
* Keep track of the last inode ref key and resume from that key
* after logging all new ancestors for the current hard link.
*/
memcpy(&search_key, &found_key, sizeof(search_key));
ret = log_new_ancestors(trans, root, path, ctx);
if (ret)
goto out;
btrfs_release_path(path);
goto again;
}
ret = 0;
out:
btrfs_free_path(path);
return ret;
}
/*
* helper function around btrfs_log_inode to make sure newly created
* parent directories also end up in the log. A minimal inode and backref
* only logging is done of any parent directories that are older than
* the last committed transaction
*/
static int btrfs_log_inode_parent(struct btrfs_trans_handle *trans,
struct btrfs_inode *inode,
struct dentry *parent,
int inode_only,
struct btrfs_log_ctx *ctx)
{
struct btrfs_root *root = inode->root;
struct btrfs_fs_info *fs_info = root->fs_info;
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 17:24:20 +03:00
int ret = 0;
Btrfs: fix metadata inconsistencies after directory fsync We can get into inconsistency between inodes and directory entries after fsyncing a directory. The issue is that while a directory gets the new dentries persisted in the fsync log and replayed at mount time, the link count of the inode that directory entries point to doesn't get updated, staying with an incorrect link count (smaller then the correct value). This later leads to stale file handle errors when accessing (including attempt to delete) some of the links if all the other ones are removed, which also implies impossibility to delete the parent directories, since the dentries can not be removed. Another issue is that (unlike ext3/4, xfs, f2fs, reiserfs, nilfs2), when fsyncing a directory, new files aren't logged (their metadata and dentries) nor any child directories. So this patch fixes this issue too, since it has the same resolution as the incorrect inode link count issue mentioned before. This is very easy to reproduce, and the following excerpt from my test case for xfstests shows how: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create our main test file and directory. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 8K" $SCRATCH_MNT/foo | _filter_xfs_io mkdir $SCRATCH_MNT/mydir # Make sure all metadata and data are durably persisted. sync # Add a hard link to 'foo' inside our test directory and fsync only the # directory. The btrfs fsync implementation had a bug that caused the new # directory entry to be visible after the fsync log replay but, the inode # of our file remained with a link count of 1. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_2 # Add a few more links and new files. # This is just to verify nothing breaks or gives incorrect results after the # fsync log is replayed. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_3 $XFS_IO_PROG -f -c "pwrite -S 0xff 0 64K" $SCRATCH_MNT/hello | _filter_xfs_io ln $SCRATCH_MNT/hello $SCRATCH_MNT/mydir/hello_2 # Add some subdirectories and new files and links to them. This is to verify # that after fsyncing our top level directory 'mydir', all the subdirectories # and their files/links are registered in the fsync log and exist after the # fsync log is replayed. mkdir -p $SCRATCH_MNT/mydir/x/y/z ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/foo_y_link ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/z/foo_z_link touch $SCRATCH_MNT/mydir/x/y/z/qwerty # Now fsync only our top directory. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/mydir # And fsync now our new file named 'hello', just to verify later that it has # the expected content and that the previous fsync on the directory 'mydir' had # no bad influence on this fsync. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/hello # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # Verify the content of our file 'foo' remains the same as before, 8192 bytes, # all with the value 0xaa. echo "File 'foo' content after log replay:" od -t x1 $SCRATCH_MNT/foo # Remove the first name of our inode. Because of the directory fsync bug, the # inode's link count was 1 instead of 5, so removing the 'foo' name ended up # deleting the inode and the other names became stale directory entries (still # visible to applications). Attempting to remove or access the remaining # dentries pointing to that inode resulted in stale file handle errors and # made it impossible to remove the parent directories since it was impossible # for them to become empty. echo "file 'foo' link count after log replay: $(stat -c %h $SCRATCH_MNT/foo)" rm -f $SCRATCH_MNT/foo # Now verify that all files, links and directories created before fsyncing our # directory exist after the fsync log was replayed. [ -f $SCRATCH_MNT/mydir/foo_2 ] || echo "Link mydir/foo_2 is missing" [ -f $SCRATCH_MNT/mydir/foo_3 ] || echo "Link mydir/foo_3 is missing" [ -f $SCRATCH_MNT/hello ] || echo "File hello is missing" [ -f $SCRATCH_MNT/mydir/hello_2 ] || echo "Link mydir/hello_2 is missing" [ -f $SCRATCH_MNT/mydir/x/y/foo_y_link ] || \ echo "Link mydir/x/y/foo_y_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/foo_z_link ] || \ echo "Link mydir/x/y/z/foo_z_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/qwerty ] || \ echo "File mydir/x/y/z/qwerty is missing" # We expect our file here to have a size of 64Kb and all the bytes having the # value 0xff. echo "file 'hello' content after log replay:" od -t x1 $SCRATCH_MNT/hello # Now remove all files/links, under our test directory 'mydir', and verify we # can remove all the directories. rm -f $SCRATCH_MNT/mydir/x/y/z/* rmdir $SCRATCH_MNT/mydir/x/y/z rm -f $SCRATCH_MNT/mydir/x/y/* rmdir $SCRATCH_MNT/mydir/x/y rmdir $SCRATCH_MNT/mydir/x rm -f $SCRATCH_MNT/mydir/* rmdir $SCRATCH_MNT/mydir # An fsck, run by the fstests framework everytime a test finishes, also detected # the inconsistency and printed the following error message: # # root 5 inode 257 errors 2001, no inode item, link count wrong # unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref # unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref status=0 exit The expected golden output for the test is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 5 file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 Which is the output after this patch and when running the test against ext3/4, xfs, f2fs, reiserfs or nilfs2. Without this patch, the test's output is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 1 Link mydir/foo_2 is missing Link mydir/foo_3 is missing Link mydir/x/y/foo_y_link is missing Link mydir/x/y/z/foo_z_link is missing File mydir/x/y/z/qwerty is missing file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y/z': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x': No such file or directory rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_2': Stale file handle rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_3': Stale file handle rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir': Directory not empty Fsck, without this fix, also complains about the wrong link count: root 5 inode 257 errors 2001, no inode item, link count wrong unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref So fix this by logging the inodes that the dentries point to when fsyncing a directory. A test case for xfstests follows. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-03-20 20:19:46 +03:00
bool log_dentries = false;
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 17:24:20 +03:00
if (btrfs_test_opt(fs_info, NOTREELOG)) {
ret = 1;
goto end_no_trans;
}
if (btrfs_root_refs(&root->root_item) == 0) {
ret = 1;
goto end_no_trans;
}
Btrfs: fix warning when replaying log after fsync of a tmpfile When replaying a log which contains a tmpfile (which necessarily has a link count of 0) we end up calling inc_nlink(), at fs/btrfs/tree-log.c:replay_one_buffer(), which produces a warning like the following: [195191.943673] WARNING: CPU: 0 PID: 6924 at fs/inode.c:342 inc_nlink+0x33/0x40 [195191.943723] CPU: 0 PID: 6924 Comm: mount Not tainted 4.19.0-rc6-btrfs-next-38 #1 [195191.943724] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.11.2-0-gf9626ccb91-prebuilt.qemu-project.org 04/01/2014 [195191.943726] RIP: 0010:inc_nlink+0x33/0x40 [195191.943728] RSP: 0018:ffffb96e425e3870 EFLAGS: 00010246 [195191.943730] RAX: 0000000000000000 RBX: ffff8c0d1e6af4f0 RCX: 0000000000000006 [195191.943731] RDX: 0000000000000000 RSI: 0000000000000000 RDI: ffff8c0d1e6af4f0 [195191.943731] RBP: 0000000000000097 R08: 0000000000000001 R09: 0000000000000000 [195191.943732] R10: 0000000000000000 R11: 0000000000000000 R12: ffffb96e425e3a60 [195191.943733] R13: ffff8c0d10cff0c8 R14: ffff8c0d0d515348 R15: ffff8c0d78a1b3f8 [195191.943735] FS: 00007f570ee24480(0000) GS:ffff8c0dfb200000(0000) knlGS:0000000000000000 [195191.943736] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [195191.943737] CR2: 00005593286277c8 CR3: 00000000bb8f2006 CR4: 00000000003606f0 [195191.943739] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [195191.943740] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [195191.943741] Call Trace: [195191.943778] replay_one_buffer+0x797/0x7d0 [btrfs] [195191.943802] walk_up_log_tree+0x1c1/0x250 [btrfs] [195191.943809] ? rcu_read_lock_sched_held+0x3f/0x70 [195191.943825] walk_log_tree+0xae/0x1d0 [btrfs] [195191.943840] btrfs_recover_log_trees+0x1d7/0x4d0 [btrfs] [195191.943856] ? replay_dir_deletes+0x280/0x280 [btrfs] [195191.943870] open_ctree+0x1c3b/0x22a0 [btrfs] [195191.943887] btrfs_mount_root+0x6b4/0x800 [btrfs] [195191.943894] ? rcu_read_lock_sched_held+0x3f/0x70 [195191.943899] ? pcpu_alloc+0x55b/0x7c0 [195191.943906] ? mount_fs+0x3b/0x140 [195191.943908] mount_fs+0x3b/0x140 [195191.943912] ? __init_waitqueue_head+0x36/0x50 [195191.943916] vfs_kern_mount+0x62/0x160 [195191.943927] btrfs_mount+0x134/0x890 [btrfs] [195191.943936] ? rcu_read_lock_sched_held+0x3f/0x70 [195191.943938] ? pcpu_alloc+0x55b/0x7c0 [195191.943943] ? mount_fs+0x3b/0x140 [195191.943952] ? btrfs_remount+0x570/0x570 [btrfs] [195191.943954] mount_fs+0x3b/0x140 [195191.943956] ? __init_waitqueue_head+0x36/0x50 [195191.943960] vfs_kern_mount+0x62/0x160 [195191.943963] do_mount+0x1f9/0xd40 [195191.943967] ? memdup_user+0x4b/0x70 [195191.943971] ksys_mount+0x7e/0xd0 [195191.943974] __x64_sys_mount+0x21/0x30 [195191.943977] do_syscall_64+0x60/0x1b0 [195191.943980] entry_SYSCALL_64_after_hwframe+0x49/0xbe [195191.943983] RIP: 0033:0x7f570e4e524a [195191.943986] RSP: 002b:00007ffd83589478 EFLAGS: 00000206 ORIG_RAX: 00000000000000a5 [195191.943989] RAX: ffffffffffffffda RBX: 0000563f335b2060 RCX: 00007f570e4e524a [195191.943990] RDX: 0000563f335b2240 RSI: 0000563f335b2280 RDI: 0000563f335b2260 [195191.943992] RBP: 0000000000000000 R08: 0000000000000000 R09: 0000000000000020 [195191.943993] R10: 00000000c0ed0000 R11: 0000000000000206 R12: 0000563f335b2260 [195191.943994] R13: 0000563f335b2240 R14: 0000000000000000 R15: 00000000ffffffff [195191.944002] irq event stamp: 8688 [195191.944010] hardirqs last enabled at (8687): [<ffffffff9cb004c3>] console_unlock+0x503/0x640 [195191.944012] hardirqs last disabled at (8688): [<ffffffff9ca037dd>] trace_hardirqs_off_thunk+0x1a/0x1c [195191.944018] softirqs last enabled at (8638): [<ffffffff9cc0a5d1>] __set_page_dirty_nobuffers+0x101/0x150 [195191.944020] softirqs last disabled at (8634): [<ffffffff9cc26bbe>] wb_wakeup_delayed+0x2e/0x60 [195191.944022] ---[ end trace 5d6e873a9a0b811a ]--- This happens because the inode does not have the flag I_LINKABLE set, which is a runtime only flag, not meant to be persisted, set when the inode is created through open(2) if the flag O_EXCL is not passed to it. Except for the warning, there are no other consequences (like corruptions or metadata inconsistencies). Since it's pointless to replay a tmpfile as it would be deleted in a later phase of the log replay procedure (it has a link count of 0), fix this by not logging tmpfiles and if a tmpfile is found in a log (created by a kernel without this change), skip the replay of the inode. A test case for fstests follows soon. Fixes: 471d557afed1 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay") CC: stable@vger.kernel.org # 4.18+ Reported-by: Martin Steigerwald <martin@lichtvoll.de> Link: https://lore.kernel.org/linux-btrfs/3666619.NTnn27ZJZE@merkaba/ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2018-10-08 13:12:55 +03:00
/*
* Skip already logged inodes or inodes corresponding to tmpfiles
* (since logging them is pointless, a link count of 0 means they
* will never be accessible).
*/
btrfs: fix race leading to unpersisted data and metadata on fsync When doing a fast fsync on a file, there is a race which can result in the fsync returning success to user space without logging the inode and without durably persisting new data. The following example shows one possible scenario for this: $ mkfs.btrfs -f /dev/sdc $ mount /dev/sdc /mnt $ touch /mnt/bar $ xfs_io -f -c "pwrite -S 0xab 0 1M" -c "fsync" /mnt/baz # Now we have: # file bar == inode 257 # file baz == inode 258 $ mv /mnt/baz /mnt/foo # Now we have: # file bar == inode 257 # file foo == inode 258 $ xfs_io -c "pwrite -S 0xcd 0 1M" /mnt/foo # fsync bar before foo, it is important to trigger the race. $ xfs_io -c "fsync" /mnt/bar $ xfs_io -c "fsync" /mnt/foo # After this: # inode 257, file bar, is empty # inode 258, file foo, has 1M filled with 0xcd <power failure> # Replay the log: $ mount /dev/sdc /mnt # After this point file foo should have 1M filled with 0xcd and not 0xab The following steps explain how the race happens: 1) Before the first fsync of inode 258, when it has the "baz" name, its ->logged_trans is 0, ->last_sub_trans is 0 and ->last_log_commit is -1. The inode also has the full sync flag set; 2) After the first fsync, we set inode 258 ->logged_trans to 6, which is the generation of the current transaction, and set ->last_log_commit to 0, which is the current value of ->last_sub_trans (done at btrfs_log_inode()). The full sync flag is cleared from the inode during the fsync. The log sub transaction that was committed had an ID of 0 and when we synced the log, at btrfs_sync_log(), we incremented root->log_transid from 0 to 1; 3) During the rename: We update inode 258, through btrfs_update_inode(), and that causes its ->last_sub_trans to be set to 1 (the current log transaction ID), and ->last_log_commit remains with a value of 0. After updating inode 258, because we have previously logged the inode in the previous fsync, we log again the inode through the call to btrfs_log_new_name(). This results in updating the inode's ->last_log_commit from 0 to 1 (the current value of its ->last_sub_trans). The ->last_sub_trans of inode 257 is updated to 1, which is the ID of the next log transaction; 4) Then a buffered write against inode 258 is made. This leaves the value of ->last_sub_trans as 1 (the ID of the current log transaction, stored at root->log_transid); 5) Then an fsync against inode 257 (or any other inode other than 258), happens. This results in committing the log transaction with ID 1, which results in updating root->last_log_commit to 1 and bumping root->log_transid from 1 to 2; 6) Then an fsync against inode 258 starts. We flush delalloc and wait only for writeback to complete, since the full sync flag is not set in the inode's runtime flags - we do not wait for ordered extents to complete. Then, at btrfs_sync_file(), we call btrfs_inode_in_log() before the ordered extent completes. The call returns true: static inline bool btrfs_inode_in_log(...) { bool ret = false; spin_lock(&inode->lock); if (inode->logged_trans == generation && inode->last_sub_trans <= inode->last_log_commit && inode->last_sub_trans <= inode->root->last_log_commit) ret = true; spin_unlock(&inode->lock); return ret; } generation has a value of 6 (fs_info->generation), ->logged_trans also has a value of 6 (set when we logged the inode during the first fsync and when logging it during the rename), ->last_sub_trans has a value of 1, set during the rename (step 3), ->last_log_commit also has a value of 1 (set in step 3) and root->last_log_commit has a value of 1, which was set in step 5 when fsyncing inode 257. As a consequence we don't log the inode, any new extents and do not sync the log, resulting in a data loss if a power failure happens after the fsync and before the current transaction commits. Also, because we do not log the inode, after a power failure the mtime and ctime of the inode do not match those we had before. When the ordered extent completes before we call btrfs_inode_in_log(), then the call returns false and we log the inode and sync the log, since at the end of ordered extent completion we update the inode and set ->last_sub_trans to 2 (the value of root->log_transid) and ->last_log_commit to 1. This problem is found after removing the check for the emptiness of the inode's list of modified extents in the recent commit 209ecbb8585bf6 ("btrfs: remove stale comment and logic from btrfs_inode_in_log()"), added in the 5.13 merge window. However checking the emptiness of the list is not really the way to solve this problem, and was never intended to, because while that solves the problem for COW writes, the problem persists for NOCOW writes because in that case the list is always empty. In the case of NOCOW writes, even though we wait for the writeback to complete before returning from btrfs_sync_file(), we end up not logging the inode, which has a new mtime/ctime, and because we don't sync the log, we never issue disk barriers (send REQ_PREFLUSH to the device) since that only happens when we sync the log (when we write super blocks at btrfs_sync_log()). So effectively, for a NOCOW case, when we return from btrfs_sync_file() to user space, we are not guaranteeing that the data is durably persisted on disk. Also, while the example above uses a rename exchange to show how the problem happens, it is not the only way to trigger it. An alternative could be adding a new hard link to inode 258, since that also results in calling btrfs_log_new_name() and updating the inode in the log. An example reproducer using the addition of a hard link instead of a rename operation: $ mkfs.btrfs -f /dev/sdc $ mount /dev/sdc /mnt $ touch /mnt/bar $ xfs_io -f -c "pwrite -S 0xab 0 1M" -c "fsync" /mnt/foo $ ln /mnt/foo /mnt/foo_link $ xfs_io -c "pwrite -S 0xcd 0 1M" /mnt/foo $ xfs_io -c "fsync" /mnt/bar $ xfs_io -c "fsync" /mnt/foo <power failure> # Replay the log: $ mount /dev/sdc /mnt # After this point file foo often has 1M filled with 0xab and not 0xcd The reasons leading to the final fsync of file foo, inode 258, not persisting the new data are the same as for the previous example with a rename operation. So fix by never skipping logging and log syncing when there are still any ordered extents in flight. To avoid making the conditional if statement that checks if logging an inode is needed harder to read, place all the logic into an helper function with separate if statements to make it more manageable and easier to read. A test case for fstests will follow soon. For NOCOW writes, the problem existed before commit b5e6c3e170b770 ("btrfs: always wait on ordered extents at fsync time"), introduced in kernel 4.19, then it went away with that commit since we started to always wait for ordered extent completion before logging. The problem came back again once the fast fsync path was changed again to avoid waiting for ordered extent completion, in commit 487781796d3022 ("btrfs: make fast fsyncs wait only for writeback"), added in kernel 5.10. However, for COW writes, the race only happens after the recent commit 209ecbb8585bf6 ("btrfs: remove stale comment and logic from btrfs_inode_in_log()"), introduced in the 5.13 merge window. For NOCOW writes, the bug existed before that commit. So tag 5.10+ as the release for stable backports. CC: stable@vger.kernel.org # 5.10+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-04-27 13:27:20 +03:00
if ((btrfs_inode_in_log(inode, trans->transid) &&
list_empty(&ctx->ordered_extents)) ||
Btrfs: fix warning when replaying log after fsync of a tmpfile When replaying a log which contains a tmpfile (which necessarily has a link count of 0) we end up calling inc_nlink(), at fs/btrfs/tree-log.c:replay_one_buffer(), which produces a warning like the following: [195191.943673] WARNING: CPU: 0 PID: 6924 at fs/inode.c:342 inc_nlink+0x33/0x40 [195191.943723] CPU: 0 PID: 6924 Comm: mount Not tainted 4.19.0-rc6-btrfs-next-38 #1 [195191.943724] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.11.2-0-gf9626ccb91-prebuilt.qemu-project.org 04/01/2014 [195191.943726] RIP: 0010:inc_nlink+0x33/0x40 [195191.943728] RSP: 0018:ffffb96e425e3870 EFLAGS: 00010246 [195191.943730] RAX: 0000000000000000 RBX: ffff8c0d1e6af4f0 RCX: 0000000000000006 [195191.943731] RDX: 0000000000000000 RSI: 0000000000000000 RDI: ffff8c0d1e6af4f0 [195191.943731] RBP: 0000000000000097 R08: 0000000000000001 R09: 0000000000000000 [195191.943732] R10: 0000000000000000 R11: 0000000000000000 R12: ffffb96e425e3a60 [195191.943733] R13: ffff8c0d10cff0c8 R14: ffff8c0d0d515348 R15: ffff8c0d78a1b3f8 [195191.943735] FS: 00007f570ee24480(0000) GS:ffff8c0dfb200000(0000) knlGS:0000000000000000 [195191.943736] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [195191.943737] CR2: 00005593286277c8 CR3: 00000000bb8f2006 CR4: 00000000003606f0 [195191.943739] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [195191.943740] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [195191.943741] Call Trace: [195191.943778] replay_one_buffer+0x797/0x7d0 [btrfs] [195191.943802] walk_up_log_tree+0x1c1/0x250 [btrfs] [195191.943809] ? rcu_read_lock_sched_held+0x3f/0x70 [195191.943825] walk_log_tree+0xae/0x1d0 [btrfs] [195191.943840] btrfs_recover_log_trees+0x1d7/0x4d0 [btrfs] [195191.943856] ? replay_dir_deletes+0x280/0x280 [btrfs] [195191.943870] open_ctree+0x1c3b/0x22a0 [btrfs] [195191.943887] btrfs_mount_root+0x6b4/0x800 [btrfs] [195191.943894] ? rcu_read_lock_sched_held+0x3f/0x70 [195191.943899] ? pcpu_alloc+0x55b/0x7c0 [195191.943906] ? mount_fs+0x3b/0x140 [195191.943908] mount_fs+0x3b/0x140 [195191.943912] ? __init_waitqueue_head+0x36/0x50 [195191.943916] vfs_kern_mount+0x62/0x160 [195191.943927] btrfs_mount+0x134/0x890 [btrfs] [195191.943936] ? rcu_read_lock_sched_held+0x3f/0x70 [195191.943938] ? pcpu_alloc+0x55b/0x7c0 [195191.943943] ? mount_fs+0x3b/0x140 [195191.943952] ? btrfs_remount+0x570/0x570 [btrfs] [195191.943954] mount_fs+0x3b/0x140 [195191.943956] ? __init_waitqueue_head+0x36/0x50 [195191.943960] vfs_kern_mount+0x62/0x160 [195191.943963] do_mount+0x1f9/0xd40 [195191.943967] ? memdup_user+0x4b/0x70 [195191.943971] ksys_mount+0x7e/0xd0 [195191.943974] __x64_sys_mount+0x21/0x30 [195191.943977] do_syscall_64+0x60/0x1b0 [195191.943980] entry_SYSCALL_64_after_hwframe+0x49/0xbe [195191.943983] RIP: 0033:0x7f570e4e524a [195191.943986] RSP: 002b:00007ffd83589478 EFLAGS: 00000206 ORIG_RAX: 00000000000000a5 [195191.943989] RAX: ffffffffffffffda RBX: 0000563f335b2060 RCX: 00007f570e4e524a [195191.943990] RDX: 0000563f335b2240 RSI: 0000563f335b2280 RDI: 0000563f335b2260 [195191.943992] RBP: 0000000000000000 R08: 0000000000000000 R09: 0000000000000020 [195191.943993] R10: 00000000c0ed0000 R11: 0000000000000206 R12: 0000563f335b2260 [195191.943994] R13: 0000563f335b2240 R14: 0000000000000000 R15: 00000000ffffffff [195191.944002] irq event stamp: 8688 [195191.944010] hardirqs last enabled at (8687): [<ffffffff9cb004c3>] console_unlock+0x503/0x640 [195191.944012] hardirqs last disabled at (8688): [<ffffffff9ca037dd>] trace_hardirqs_off_thunk+0x1a/0x1c [195191.944018] softirqs last enabled at (8638): [<ffffffff9cc0a5d1>] __set_page_dirty_nobuffers+0x101/0x150 [195191.944020] softirqs last disabled at (8634): [<ffffffff9cc26bbe>] wb_wakeup_delayed+0x2e/0x60 [195191.944022] ---[ end trace 5d6e873a9a0b811a ]--- This happens because the inode does not have the flag I_LINKABLE set, which is a runtime only flag, not meant to be persisted, set when the inode is created through open(2) if the flag O_EXCL is not passed to it. Except for the warning, there are no other consequences (like corruptions or metadata inconsistencies). Since it's pointless to replay a tmpfile as it would be deleted in a later phase of the log replay procedure (it has a link count of 0), fix this by not logging tmpfiles and if a tmpfile is found in a log (created by a kernel without this change), skip the replay of the inode. A test case for fstests follows soon. Fixes: 471d557afed1 ("Btrfs: fix loss of prealloc extents past i_size after fsync log replay") CC: stable@vger.kernel.org # 4.18+ Reported-by: Martin Steigerwald <martin@lichtvoll.de> Link: https://lore.kernel.org/linux-btrfs/3666619.NTnn27ZJZE@merkaba/ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2018-10-08 13:12:55 +03:00
inode->vfs_inode.i_nlink == 0) {
ret = BTRFS_NO_LOG_SYNC;
goto end_no_trans;
}
ret = start_log_trans(trans, root, ctx);
if (ret)
goto end_no_trans;
ret = btrfs_log_inode(trans, inode, inode_only, ctx);
if (ret)
goto end_trans;
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 17:24:20 +03:00
/*
* for regular files, if its inode is already on disk, we don't
* have to worry about the parents at all. This is because
* we can use the last_unlink_trans field to record renames
* and other fun in this file.
*/
if (S_ISREG(inode->vfs_inode.i_mode) &&
btrfs: fix race that makes inode logging fallback to transaction commit When logging an inode and the previous transaction is still committing, we have a time window where we can end up incorrectly think an inode has its last_unlink_trans field with a value greater than the last transaction committed, which results in the logging to fallback to a full transaction commit, which is usually much more expensive than doing a log commit. The race is described by the following steps: 1) We are at transaction 1000; 2) We modify an inode X (a directory) using transaction 1000 and set its last_unlink_trans field to 1000, because for example we removed one of its subdirectories; 3) We create a new inode Y with a dentry in inode X using transaction 1000, so its generation field is set to 1000; 4) The commit for transaction 1000 is started by task A; 5) The task committing transaction 1000 sets the transaction state to unblocked, writes the dirty extent buffers and the super blocks, then unlocks tree_log_mutex; 6) Some task starts a new transaction with a generation of 1001; 7) We do some modification to inode Y (using transaction 1001); 8) The transaction 1000 commit starts unpinning extents. At this point fs_info->last_trans_committed still has a value of 999; 9) Task B starts an fsync on inode Y, and gets a handle for transaction 1001. When it gets to check_parent_dirs_for_sync() it does the checking of the ancestor dentries because the following check does not evaluate to true: if (S_ISREG(inode->vfs_inode.i_mode) && inode->generation <= last_committed && inode->last_unlink_trans <= last_committed) goto out; The generation value for inode Y is 1000 and last_committed, which has the value read from fs_info->last_trans_committed, has a value of 999, so that check evaluates to false and we proceed to check the ancestor inodes. Once we get to the first ancestor, inode X, we call btrfs_must_commit_transaction() on it, which evaluates to true: static bool btrfs_must_commit_transaction(...) { struct btrfs_fs_info *fs_info = inode->root->fs_info; bool ret = false; mutex_lock(&inode->log_mutex); if (inode->last_unlink_trans > fs_info->last_trans_committed) { /* * Make sure any commits to the log are forced to be full * commits. */ btrfs_set_log_full_commit(trans); ret = true; } (...) because inode's X last_unlink_trans has a value of 1000 and fs_info->last_trans_committed still has a value of 999, it returns true to check_parent_dirs_for_sync(), making it return 1 which is propagated up to btrfs_sync_file(), causing it to fallback to a full transaction commit of transaction 1001. We should have not fallen back to commit transaction 1001, since inode X had last_unlink_trans set to 1000 and the super blocks for transaction 1000 were already written. So while not resulting in a functional problem, it leads to a lot more work and higher latencies for a fsync since committing a transaction is usually more expensive than committing a log (if other filesystem changes happened under that transaction). Similar problem happens when logging directories, for the same reason as btrfs_must_commit_transaction() returns true on an inode with its last_unlink_trans having the generation of the previous transaction and that transaction is still committing, unpinning its freed extents. So fix this by comparing last_unlink_trans with the id of the current transaction instead of fs_info->last_trans_committed. This case is often hit when running dbench for a long enough duration, as it does lots of rename and rmdir operations (both update the field last_unlink_trans of an inode) and fsyncs of files and directories. This patch belongs to a patch set that is comprised of the following patches: btrfs: fix race causing unnecessary inode logging during link and rename btrfs: fix race that results in logging old extents during a fast fsync btrfs: fix race that causes unnecessary logging of ancestor inodes btrfs: fix race that makes inode logging fallback to transaction commit btrfs: fix race leading to unnecessary transaction commit when logging inode btrfs: do not block inode logging for so long during transaction commit Performance results are mentioned in the change log of the last patch. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-11-25 15:19:26 +03:00
inode->generation < trans->transid &&
inode->last_unlink_trans < trans->transid) {
ret = 0;
goto end_trans;
}
if (S_ISDIR(inode->vfs_inode.i_mode) && ctx->log_new_dentries)
Btrfs: fix metadata inconsistencies after directory fsync We can get into inconsistency between inodes and directory entries after fsyncing a directory. The issue is that while a directory gets the new dentries persisted in the fsync log and replayed at mount time, the link count of the inode that directory entries point to doesn't get updated, staying with an incorrect link count (smaller then the correct value). This later leads to stale file handle errors when accessing (including attempt to delete) some of the links if all the other ones are removed, which also implies impossibility to delete the parent directories, since the dentries can not be removed. Another issue is that (unlike ext3/4, xfs, f2fs, reiserfs, nilfs2), when fsyncing a directory, new files aren't logged (their metadata and dentries) nor any child directories. So this patch fixes this issue too, since it has the same resolution as the incorrect inode link count issue mentioned before. This is very easy to reproduce, and the following excerpt from my test case for xfstests shows how: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create our main test file and directory. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 8K" $SCRATCH_MNT/foo | _filter_xfs_io mkdir $SCRATCH_MNT/mydir # Make sure all metadata and data are durably persisted. sync # Add a hard link to 'foo' inside our test directory and fsync only the # directory. The btrfs fsync implementation had a bug that caused the new # directory entry to be visible after the fsync log replay but, the inode # of our file remained with a link count of 1. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_2 # Add a few more links and new files. # This is just to verify nothing breaks or gives incorrect results after the # fsync log is replayed. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_3 $XFS_IO_PROG -f -c "pwrite -S 0xff 0 64K" $SCRATCH_MNT/hello | _filter_xfs_io ln $SCRATCH_MNT/hello $SCRATCH_MNT/mydir/hello_2 # Add some subdirectories and new files and links to them. This is to verify # that after fsyncing our top level directory 'mydir', all the subdirectories # and their files/links are registered in the fsync log and exist after the # fsync log is replayed. mkdir -p $SCRATCH_MNT/mydir/x/y/z ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/foo_y_link ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/z/foo_z_link touch $SCRATCH_MNT/mydir/x/y/z/qwerty # Now fsync only our top directory. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/mydir # And fsync now our new file named 'hello', just to verify later that it has # the expected content and that the previous fsync on the directory 'mydir' had # no bad influence on this fsync. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/hello # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # Verify the content of our file 'foo' remains the same as before, 8192 bytes, # all with the value 0xaa. echo "File 'foo' content after log replay:" od -t x1 $SCRATCH_MNT/foo # Remove the first name of our inode. Because of the directory fsync bug, the # inode's link count was 1 instead of 5, so removing the 'foo' name ended up # deleting the inode and the other names became stale directory entries (still # visible to applications). Attempting to remove or access the remaining # dentries pointing to that inode resulted in stale file handle errors and # made it impossible to remove the parent directories since it was impossible # for them to become empty. echo "file 'foo' link count after log replay: $(stat -c %h $SCRATCH_MNT/foo)" rm -f $SCRATCH_MNT/foo # Now verify that all files, links and directories created before fsyncing our # directory exist after the fsync log was replayed. [ -f $SCRATCH_MNT/mydir/foo_2 ] || echo "Link mydir/foo_2 is missing" [ -f $SCRATCH_MNT/mydir/foo_3 ] || echo "Link mydir/foo_3 is missing" [ -f $SCRATCH_MNT/hello ] || echo "File hello is missing" [ -f $SCRATCH_MNT/mydir/hello_2 ] || echo "Link mydir/hello_2 is missing" [ -f $SCRATCH_MNT/mydir/x/y/foo_y_link ] || \ echo "Link mydir/x/y/foo_y_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/foo_z_link ] || \ echo "Link mydir/x/y/z/foo_z_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/qwerty ] || \ echo "File mydir/x/y/z/qwerty is missing" # We expect our file here to have a size of 64Kb and all the bytes having the # value 0xff. echo "file 'hello' content after log replay:" od -t x1 $SCRATCH_MNT/hello # Now remove all files/links, under our test directory 'mydir', and verify we # can remove all the directories. rm -f $SCRATCH_MNT/mydir/x/y/z/* rmdir $SCRATCH_MNT/mydir/x/y/z rm -f $SCRATCH_MNT/mydir/x/y/* rmdir $SCRATCH_MNT/mydir/x/y rmdir $SCRATCH_MNT/mydir/x rm -f $SCRATCH_MNT/mydir/* rmdir $SCRATCH_MNT/mydir # An fsck, run by the fstests framework everytime a test finishes, also detected # the inconsistency and printed the following error message: # # root 5 inode 257 errors 2001, no inode item, link count wrong # unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref # unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref status=0 exit The expected golden output for the test is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 5 file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 Which is the output after this patch and when running the test against ext3/4, xfs, f2fs, reiserfs or nilfs2. Without this patch, the test's output is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 1 Link mydir/foo_2 is missing Link mydir/foo_3 is missing Link mydir/x/y/foo_y_link is missing Link mydir/x/y/z/foo_z_link is missing File mydir/x/y/z/qwerty is missing file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y/z': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x': No such file or directory rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_2': Stale file handle rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_3': Stale file handle rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir': Directory not empty Fsck, without this fix, also complains about the wrong link count: root 5 inode 257 errors 2001, no inode item, link count wrong unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref So fix this by logging the inodes that the dentries point to when fsyncing a directory. A test case for xfstests follows. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-03-20 20:19:46 +03:00
log_dentries = true;
Btrfs: fix stale dir entries after removing a link and fsync We have one more case where after a log tree is replayed we get inconsistent metadata leading to stale directory entries, due to some directories having entries pointing to some inode while the inode does not have a matching BTRFS_INODE_[REF|EXTREF]_KEY item. To trigger the problem we need to have a file with multiple hard links belonging to different parent directories. Then if one of those hard links is removed and we fsync the file using one of its other links that belongs to a different parent directory, we end up not logging the fact that the removed hard link doesn't exists anymore in the parent directory. Simple reproducer: seq=`basename $0` seqres=$RESULT_DIR/$seq echo "QA output created by $seq" tmp=/tmp/$$ status=1 # failure is the default! trap "_cleanup; exit \$status" 0 1 2 3 15 _cleanup() { _cleanup_flakey rm -f $tmp.* } # get standard environment, filters and checks . ./common/rc . ./common/filter . ./common/dmflakey # real QA test starts here _need_to_be_root _supported_fs generic _supported_os Linux _require_scratch _require_dm_flakey _require_metadata_journaling $SCRATCH_DEV rm -f $seqres.full _scratch_mkfs >>$seqres.full 2>&1 _init_flakey _mount_flakey # Create our test directory and file. mkdir $SCRATCH_MNT/testdir touch $SCRATCH_MNT/foo ln $SCRATCH_MNT/foo $SCRATCH_MNT/testdir/foo2 ln $SCRATCH_MNT/foo $SCRATCH_MNT/testdir/foo3 # Make sure everything done so far is durably persisted. sync # Now we remove one of our file's hardlinks in the directory testdir. unlink $SCRATCH_MNT/testdir/foo3 # We now fsync our file using the "foo" link, which has a parent that # is not the directory "testdir". $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foo # Silently drop all writes and unmount to simulate a crash/power # failure. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again, mount to trigger journal/log replay. _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # After the journal/log is replayed we expect to not see the "foo3" # link anymore and we should be able to remove all names in the # directory "testdir" and then remove it (no stale directory entries # left after the journal/log replay). echo "Entries in testdir:" ls -1 $SCRATCH_MNT/testdir rm -f $SCRATCH_MNT/testdir/* rmdir $SCRATCH_MNT/testdir _unmount_flakey status=0 exit The test fails with: $ ./check generic/107 FSTYP -- btrfs PLATFORM -- Linux/x86_64 debian3 4.1.0-rc6-btrfs-next-11+ MKFS_OPTIONS -- /dev/sdc MOUNT_OPTIONS -- /dev/sdc /home/fdmanana/btrfs-tests/scratch_1 generic/107 3s ... - output mismatch (see .../results/generic/107.out.bad) --- tests/generic/107.out 2015-08-01 01:39:45.807462161 +0100 +++ /home/fdmanana/git/hub/xfstests/results//generic/107.out.bad @@ -1,3 +1,5 @@ QA output created by 107 Entries in testdir: foo2 +foo3 +rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/testdir': Directory not empty ... _check_btrfs_filesystem: filesystem on /dev/sdc is inconsistent \ (see /home/fdmanana/git/hub/xfstests/results//generic/107.full) _check_dmesg: something found in dmesg (see .../results/generic/107.dmesg) Ran: generic/107 Failures: generic/107 Failed 1 of 1 tests $ cat /home/fdmanana/git/hub/xfstests/results//generic/107.full (...) checking fs roots root 5 inode 257 errors 200, dir isize wrong unresolved ref dir 257 index 3 namelen 4 name foo3 filetype 1 errors 5, no dir item, no inode ref (...) And produces the following warning in dmesg: [127298.759064] BTRFS info (device dm-0): failed to delete reference to foo3, inode 258 parent 257 [127298.762081] ------------[ cut here ]------------ [127298.763311] WARNING: CPU: 10 PID: 7891 at fs/btrfs/inode.c:3956 __btrfs_unlink_inode+0x182/0x35a [btrfs]() [127298.767327] BTRFS: Transaction aborted (error -2) (...) [127298.788611] Call Trace: [127298.789137] [<ffffffff8145f077>] dump_stack+0x4f/0x7b [127298.790090] [<ffffffff81095de5>] ? console_unlock+0x356/0x3a2 [127298.791157] [<ffffffff8104b3b0>] warn_slowpath_common+0xa1/0xbb [127298.792323] [<ffffffffa065ad09>] ? __btrfs_unlink_inode+0x182/0x35a [btrfs] [127298.793633] [<ffffffff8104b410>] warn_slowpath_fmt+0x46/0x48 [127298.794699] [<ffffffffa065ad09>] __btrfs_unlink_inode+0x182/0x35a [btrfs] [127298.797640] [<ffffffffa065be8f>] btrfs_unlink_inode+0x1e/0x40 [btrfs] [127298.798876] [<ffffffffa065bf11>] btrfs_unlink+0x60/0x9b [btrfs] [127298.800154] [<ffffffff8116fb48>] vfs_unlink+0x9c/0xed [127298.801303] [<ffffffff81173481>] do_unlinkat+0x12b/0x1fb [127298.802450] [<ffffffff81253855>] ? lockdep_sys_exit_thunk+0x12/0x14 [127298.803797] [<ffffffff81174056>] SyS_unlinkat+0x29/0x2b [127298.805017] [<ffffffff81465197>] system_call_fastpath+0x12/0x6f [127298.806310] ---[ end trace bbfddacb7aaada7b ]--- [127298.807325] BTRFS warning (device dm-0): __btrfs_unlink_inode:3956: Aborting unused transaction(No such entry). So fix this by logging all parent inodes, current and old ones, to make sure we do not get stale entries after log replay. This is not a simple solution such as triggering a full transaction commit because it would imply full transaction commit when an inode is fsynced in the same transaction that modified it and reloaded it after eviction (because its last_unlink_trans is set to the same value as its last_trans as of the commit with the title "Btrfs: fix stale dir entries after unlink, inode eviction and fsync"), and it would also make fstest generic/066 fail since one of the fsyncs triggers a full commit and the next fsync will not find the inode in the log anymore (therefore not removing the xattr). Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-08-05 18:49:08 +03:00
/*
* On unlink we must make sure all our current and old parent directory
Btrfs: fix stale dir entries after removing a link and fsync We have one more case where after a log tree is replayed we get inconsistent metadata leading to stale directory entries, due to some directories having entries pointing to some inode while the inode does not have a matching BTRFS_INODE_[REF|EXTREF]_KEY item. To trigger the problem we need to have a file with multiple hard links belonging to different parent directories. Then if one of those hard links is removed and we fsync the file using one of its other links that belongs to a different parent directory, we end up not logging the fact that the removed hard link doesn't exists anymore in the parent directory. Simple reproducer: seq=`basename $0` seqres=$RESULT_DIR/$seq echo "QA output created by $seq" tmp=/tmp/$$ status=1 # failure is the default! trap "_cleanup; exit \$status" 0 1 2 3 15 _cleanup() { _cleanup_flakey rm -f $tmp.* } # get standard environment, filters and checks . ./common/rc . ./common/filter . ./common/dmflakey # real QA test starts here _need_to_be_root _supported_fs generic _supported_os Linux _require_scratch _require_dm_flakey _require_metadata_journaling $SCRATCH_DEV rm -f $seqres.full _scratch_mkfs >>$seqres.full 2>&1 _init_flakey _mount_flakey # Create our test directory and file. mkdir $SCRATCH_MNT/testdir touch $SCRATCH_MNT/foo ln $SCRATCH_MNT/foo $SCRATCH_MNT/testdir/foo2 ln $SCRATCH_MNT/foo $SCRATCH_MNT/testdir/foo3 # Make sure everything done so far is durably persisted. sync # Now we remove one of our file's hardlinks in the directory testdir. unlink $SCRATCH_MNT/testdir/foo3 # We now fsync our file using the "foo" link, which has a parent that # is not the directory "testdir". $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foo # Silently drop all writes and unmount to simulate a crash/power # failure. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again, mount to trigger journal/log replay. _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # After the journal/log is replayed we expect to not see the "foo3" # link anymore and we should be able to remove all names in the # directory "testdir" and then remove it (no stale directory entries # left after the journal/log replay). echo "Entries in testdir:" ls -1 $SCRATCH_MNT/testdir rm -f $SCRATCH_MNT/testdir/* rmdir $SCRATCH_MNT/testdir _unmount_flakey status=0 exit The test fails with: $ ./check generic/107 FSTYP -- btrfs PLATFORM -- Linux/x86_64 debian3 4.1.0-rc6-btrfs-next-11+ MKFS_OPTIONS -- /dev/sdc MOUNT_OPTIONS -- /dev/sdc /home/fdmanana/btrfs-tests/scratch_1 generic/107 3s ... - output mismatch (see .../results/generic/107.out.bad) --- tests/generic/107.out 2015-08-01 01:39:45.807462161 +0100 +++ /home/fdmanana/git/hub/xfstests/results//generic/107.out.bad @@ -1,3 +1,5 @@ QA output created by 107 Entries in testdir: foo2 +foo3 +rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/testdir': Directory not empty ... _check_btrfs_filesystem: filesystem on /dev/sdc is inconsistent \ (see /home/fdmanana/git/hub/xfstests/results//generic/107.full) _check_dmesg: something found in dmesg (see .../results/generic/107.dmesg) Ran: generic/107 Failures: generic/107 Failed 1 of 1 tests $ cat /home/fdmanana/git/hub/xfstests/results//generic/107.full (...) checking fs roots root 5 inode 257 errors 200, dir isize wrong unresolved ref dir 257 index 3 namelen 4 name foo3 filetype 1 errors 5, no dir item, no inode ref (...) And produces the following warning in dmesg: [127298.759064] BTRFS info (device dm-0): failed to delete reference to foo3, inode 258 parent 257 [127298.762081] ------------[ cut here ]------------ [127298.763311] WARNING: CPU: 10 PID: 7891 at fs/btrfs/inode.c:3956 __btrfs_unlink_inode+0x182/0x35a [btrfs]() [127298.767327] BTRFS: Transaction aborted (error -2) (...) [127298.788611] Call Trace: [127298.789137] [<ffffffff8145f077>] dump_stack+0x4f/0x7b [127298.790090] [<ffffffff81095de5>] ? console_unlock+0x356/0x3a2 [127298.791157] [<ffffffff8104b3b0>] warn_slowpath_common+0xa1/0xbb [127298.792323] [<ffffffffa065ad09>] ? __btrfs_unlink_inode+0x182/0x35a [btrfs] [127298.793633] [<ffffffff8104b410>] warn_slowpath_fmt+0x46/0x48 [127298.794699] [<ffffffffa065ad09>] __btrfs_unlink_inode+0x182/0x35a [btrfs] [127298.797640] [<ffffffffa065be8f>] btrfs_unlink_inode+0x1e/0x40 [btrfs] [127298.798876] [<ffffffffa065bf11>] btrfs_unlink+0x60/0x9b [btrfs] [127298.800154] [<ffffffff8116fb48>] vfs_unlink+0x9c/0xed [127298.801303] [<ffffffff81173481>] do_unlinkat+0x12b/0x1fb [127298.802450] [<ffffffff81253855>] ? lockdep_sys_exit_thunk+0x12/0x14 [127298.803797] [<ffffffff81174056>] SyS_unlinkat+0x29/0x2b [127298.805017] [<ffffffff81465197>] system_call_fastpath+0x12/0x6f [127298.806310] ---[ end trace bbfddacb7aaada7b ]--- [127298.807325] BTRFS warning (device dm-0): __btrfs_unlink_inode:3956: Aborting unused transaction(No such entry). So fix this by logging all parent inodes, current and old ones, to make sure we do not get stale entries after log replay. This is not a simple solution such as triggering a full transaction commit because it would imply full transaction commit when an inode is fsynced in the same transaction that modified it and reloaded it after eviction (because its last_unlink_trans is set to the same value as its last_trans as of the commit with the title "Btrfs: fix stale dir entries after unlink, inode eviction and fsync"), and it would also make fstest generic/066 fail since one of the fsyncs triggers a full commit and the next fsync will not find the inode in the log anymore (therefore not removing the xattr). Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-08-05 18:49:08 +03:00
* inodes are fully logged. This is to prevent leaving dangling
* directory index entries in directories that were our parents but are
* not anymore. Not doing this results in old parent directory being
* impossible to delete after log replay (rmdir will always fail with
* error -ENOTEMPTY).
*
* Example 1:
*
* mkdir testdir
* touch testdir/foo
* ln testdir/foo testdir/bar
* sync
* unlink testdir/bar
* xfs_io -c fsync testdir/foo
* <power failure>
* mount fs, triggers log replay
*
* If we don't log the parent directory (testdir), after log replay the
* directory still has an entry pointing to the file inode using the bar
* name, but a matching BTRFS_INODE_[REF|EXTREF]_KEY does not exist and
* the file inode has a link count of 1.
*
* Example 2:
*
* mkdir testdir
* touch foo
* ln foo testdir/foo2
* ln foo testdir/foo3
* sync
* unlink testdir/foo3
* xfs_io -c fsync foo
* <power failure>
* mount fs, triggers log replay
*
* Similar as the first example, after log replay the parent directory
* testdir still has an entry pointing to the inode file with name foo3
* but the file inode does not have a matching BTRFS_INODE_REF_KEY item
* and has a link count of 2.
*/
btrfs: fix race that makes inode logging fallback to transaction commit When logging an inode and the previous transaction is still committing, we have a time window where we can end up incorrectly think an inode has its last_unlink_trans field with a value greater than the last transaction committed, which results in the logging to fallback to a full transaction commit, which is usually much more expensive than doing a log commit. The race is described by the following steps: 1) We are at transaction 1000; 2) We modify an inode X (a directory) using transaction 1000 and set its last_unlink_trans field to 1000, because for example we removed one of its subdirectories; 3) We create a new inode Y with a dentry in inode X using transaction 1000, so its generation field is set to 1000; 4) The commit for transaction 1000 is started by task A; 5) The task committing transaction 1000 sets the transaction state to unblocked, writes the dirty extent buffers and the super blocks, then unlocks tree_log_mutex; 6) Some task starts a new transaction with a generation of 1001; 7) We do some modification to inode Y (using transaction 1001); 8) The transaction 1000 commit starts unpinning extents. At this point fs_info->last_trans_committed still has a value of 999; 9) Task B starts an fsync on inode Y, and gets a handle for transaction 1001. When it gets to check_parent_dirs_for_sync() it does the checking of the ancestor dentries because the following check does not evaluate to true: if (S_ISREG(inode->vfs_inode.i_mode) && inode->generation <= last_committed && inode->last_unlink_trans <= last_committed) goto out; The generation value for inode Y is 1000 and last_committed, which has the value read from fs_info->last_trans_committed, has a value of 999, so that check evaluates to false and we proceed to check the ancestor inodes. Once we get to the first ancestor, inode X, we call btrfs_must_commit_transaction() on it, which evaluates to true: static bool btrfs_must_commit_transaction(...) { struct btrfs_fs_info *fs_info = inode->root->fs_info; bool ret = false; mutex_lock(&inode->log_mutex); if (inode->last_unlink_trans > fs_info->last_trans_committed) { /* * Make sure any commits to the log are forced to be full * commits. */ btrfs_set_log_full_commit(trans); ret = true; } (...) because inode's X last_unlink_trans has a value of 1000 and fs_info->last_trans_committed still has a value of 999, it returns true to check_parent_dirs_for_sync(), making it return 1 which is propagated up to btrfs_sync_file(), causing it to fallback to a full transaction commit of transaction 1001. We should have not fallen back to commit transaction 1001, since inode X had last_unlink_trans set to 1000 and the super blocks for transaction 1000 were already written. So while not resulting in a functional problem, it leads to a lot more work and higher latencies for a fsync since committing a transaction is usually more expensive than committing a log (if other filesystem changes happened under that transaction). Similar problem happens when logging directories, for the same reason as btrfs_must_commit_transaction() returns true on an inode with its last_unlink_trans having the generation of the previous transaction and that transaction is still committing, unpinning its freed extents. So fix this by comparing last_unlink_trans with the id of the current transaction instead of fs_info->last_trans_committed. This case is often hit when running dbench for a long enough duration, as it does lots of rename and rmdir operations (both update the field last_unlink_trans of an inode) and fsyncs of files and directories. This patch belongs to a patch set that is comprised of the following patches: btrfs: fix race causing unnecessary inode logging during link and rename btrfs: fix race that results in logging old extents during a fast fsync btrfs: fix race that causes unnecessary logging of ancestor inodes btrfs: fix race that makes inode logging fallback to transaction commit btrfs: fix race leading to unnecessary transaction commit when logging inode btrfs: do not block inode logging for so long during transaction commit Performance results are mentioned in the change log of the last patch. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-11-25 15:19:26 +03:00
if (inode->last_unlink_trans >= trans->transid) {
ret = btrfs_log_all_parents(trans, inode, ctx);
Btrfs: fix stale dir entries after removing a link and fsync We have one more case where after a log tree is replayed we get inconsistent metadata leading to stale directory entries, due to some directories having entries pointing to some inode while the inode does not have a matching BTRFS_INODE_[REF|EXTREF]_KEY item. To trigger the problem we need to have a file with multiple hard links belonging to different parent directories. Then if one of those hard links is removed and we fsync the file using one of its other links that belongs to a different parent directory, we end up not logging the fact that the removed hard link doesn't exists anymore in the parent directory. Simple reproducer: seq=`basename $0` seqres=$RESULT_DIR/$seq echo "QA output created by $seq" tmp=/tmp/$$ status=1 # failure is the default! trap "_cleanup; exit \$status" 0 1 2 3 15 _cleanup() { _cleanup_flakey rm -f $tmp.* } # get standard environment, filters and checks . ./common/rc . ./common/filter . ./common/dmflakey # real QA test starts here _need_to_be_root _supported_fs generic _supported_os Linux _require_scratch _require_dm_flakey _require_metadata_journaling $SCRATCH_DEV rm -f $seqres.full _scratch_mkfs >>$seqres.full 2>&1 _init_flakey _mount_flakey # Create our test directory and file. mkdir $SCRATCH_MNT/testdir touch $SCRATCH_MNT/foo ln $SCRATCH_MNT/foo $SCRATCH_MNT/testdir/foo2 ln $SCRATCH_MNT/foo $SCRATCH_MNT/testdir/foo3 # Make sure everything done so far is durably persisted. sync # Now we remove one of our file's hardlinks in the directory testdir. unlink $SCRATCH_MNT/testdir/foo3 # We now fsync our file using the "foo" link, which has a parent that # is not the directory "testdir". $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foo # Silently drop all writes and unmount to simulate a crash/power # failure. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey # Allow writes again, mount to trigger journal/log replay. _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # After the journal/log is replayed we expect to not see the "foo3" # link anymore and we should be able to remove all names in the # directory "testdir" and then remove it (no stale directory entries # left after the journal/log replay). echo "Entries in testdir:" ls -1 $SCRATCH_MNT/testdir rm -f $SCRATCH_MNT/testdir/* rmdir $SCRATCH_MNT/testdir _unmount_flakey status=0 exit The test fails with: $ ./check generic/107 FSTYP -- btrfs PLATFORM -- Linux/x86_64 debian3 4.1.0-rc6-btrfs-next-11+ MKFS_OPTIONS -- /dev/sdc MOUNT_OPTIONS -- /dev/sdc /home/fdmanana/btrfs-tests/scratch_1 generic/107 3s ... - output mismatch (see .../results/generic/107.out.bad) --- tests/generic/107.out 2015-08-01 01:39:45.807462161 +0100 +++ /home/fdmanana/git/hub/xfstests/results//generic/107.out.bad @@ -1,3 +1,5 @@ QA output created by 107 Entries in testdir: foo2 +foo3 +rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/testdir': Directory not empty ... _check_btrfs_filesystem: filesystem on /dev/sdc is inconsistent \ (see /home/fdmanana/git/hub/xfstests/results//generic/107.full) _check_dmesg: something found in dmesg (see .../results/generic/107.dmesg) Ran: generic/107 Failures: generic/107 Failed 1 of 1 tests $ cat /home/fdmanana/git/hub/xfstests/results//generic/107.full (...) checking fs roots root 5 inode 257 errors 200, dir isize wrong unresolved ref dir 257 index 3 namelen 4 name foo3 filetype 1 errors 5, no dir item, no inode ref (...) And produces the following warning in dmesg: [127298.759064] BTRFS info (device dm-0): failed to delete reference to foo3, inode 258 parent 257 [127298.762081] ------------[ cut here ]------------ [127298.763311] WARNING: CPU: 10 PID: 7891 at fs/btrfs/inode.c:3956 __btrfs_unlink_inode+0x182/0x35a [btrfs]() [127298.767327] BTRFS: Transaction aborted (error -2) (...) [127298.788611] Call Trace: [127298.789137] [<ffffffff8145f077>] dump_stack+0x4f/0x7b [127298.790090] [<ffffffff81095de5>] ? console_unlock+0x356/0x3a2 [127298.791157] [<ffffffff8104b3b0>] warn_slowpath_common+0xa1/0xbb [127298.792323] [<ffffffffa065ad09>] ? __btrfs_unlink_inode+0x182/0x35a [btrfs] [127298.793633] [<ffffffff8104b410>] warn_slowpath_fmt+0x46/0x48 [127298.794699] [<ffffffffa065ad09>] __btrfs_unlink_inode+0x182/0x35a [btrfs] [127298.797640] [<ffffffffa065be8f>] btrfs_unlink_inode+0x1e/0x40 [btrfs] [127298.798876] [<ffffffffa065bf11>] btrfs_unlink+0x60/0x9b [btrfs] [127298.800154] [<ffffffff8116fb48>] vfs_unlink+0x9c/0xed [127298.801303] [<ffffffff81173481>] do_unlinkat+0x12b/0x1fb [127298.802450] [<ffffffff81253855>] ? lockdep_sys_exit_thunk+0x12/0x14 [127298.803797] [<ffffffff81174056>] SyS_unlinkat+0x29/0x2b [127298.805017] [<ffffffff81465197>] system_call_fastpath+0x12/0x6f [127298.806310] ---[ end trace bbfddacb7aaada7b ]--- [127298.807325] BTRFS warning (device dm-0): __btrfs_unlink_inode:3956: Aborting unused transaction(No such entry). So fix this by logging all parent inodes, current and old ones, to make sure we do not get stale entries after log replay. This is not a simple solution such as triggering a full transaction commit because it would imply full transaction commit when an inode is fsynced in the same transaction that modified it and reloaded it after eviction (because its last_unlink_trans is set to the same value as its last_trans as of the commit with the title "Btrfs: fix stale dir entries after unlink, inode eviction and fsync"), and it would also make fstest generic/066 fail since one of the fsyncs triggers a full commit and the next fsync will not find the inode in the log anymore (therefore not removing the xattr). Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-08-05 18:49:08 +03:00
if (ret)
goto end_trans;
}
ret = log_all_new_ancestors(trans, inode, parent, ctx);
if (ret)
Btrfs: fix fsync of files with multiple hard links in new directories The log tree has a long standing problem that when a file is fsync'ed we only check for new ancestors, created in the current transaction, by following only the hard link for which the fsync was issued. We follow the ancestors using the VFS' dget_parent() API. This means that if we create a new link for a file in a directory that is new (or in an any other new ancestor directory) and then fsync the file using an old hard link, we end up not logging the new ancestor, and on log replay that new hard link and ancestor do not exist. In some cases, involving renames, the file will not exist at all. Example: mkfs.btrfs -f /dev/sdb mount /dev/sdb /mnt mkdir /mnt/A touch /mnt/foo ln /mnt/foo /mnt/A/bar xfs_io -c fsync /mnt/foo <power failure> In this example after log replay only the hard link named 'foo' exists and directory A does not exist, which is unexpected. In other major linux filesystems, such as ext4, xfs and f2fs for example, both hard links exist and so does directory A after mounting again the filesystem. Checking if any new ancestors are new and need to be logged was added in 2009 by commit 12fcfd22fe5b ("Btrfs: tree logging unlink/rename fixes"), however only for the ancestors of the hard link (dentry) for which the fsync was issued, instead of checking for all ancestors for all of the inode's hard links. So fix this by tracking the id of the last transaction where a hard link was created for an inode and then on fsync fallback to a full transaction commit when an inode has more than one hard link and at least one new hard link was created in the current transaction. This is the simplest solution since this is not a common use case (adding frequently hard links for which there's an ancestor created in the current transaction and then fsync the file). In case it ever becomes a common use case, a solution that consists of iterating the fs/subvol btree for each hard link and check if any ancestor is new, could be implemented. This solves many unexpected scenarios reported by Jayashree Mohan and Vijay Chidambaram, and for which there is a new test case for fstests under review. Fixes: 12fcfd22fe5b ("Btrfs: tree logging unlink/rename fixes") CC: stable@vger.kernel.org # 4.4+ Reported-by: Vijay Chidambaram <vvijay03@gmail.com> Reported-by: Jayashree Mohan <jayashree2912@gmail.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2018-11-28 17:54:28 +03:00
goto end_trans;
Btrfs: fix metadata inconsistencies after directory fsync We can get into inconsistency between inodes and directory entries after fsyncing a directory. The issue is that while a directory gets the new dentries persisted in the fsync log and replayed at mount time, the link count of the inode that directory entries point to doesn't get updated, staying with an incorrect link count (smaller then the correct value). This later leads to stale file handle errors when accessing (including attempt to delete) some of the links if all the other ones are removed, which also implies impossibility to delete the parent directories, since the dentries can not be removed. Another issue is that (unlike ext3/4, xfs, f2fs, reiserfs, nilfs2), when fsyncing a directory, new files aren't logged (their metadata and dentries) nor any child directories. So this patch fixes this issue too, since it has the same resolution as the incorrect inode link count issue mentioned before. This is very easy to reproduce, and the following excerpt from my test case for xfstests shows how: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create our main test file and directory. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 8K" $SCRATCH_MNT/foo | _filter_xfs_io mkdir $SCRATCH_MNT/mydir # Make sure all metadata and data are durably persisted. sync # Add a hard link to 'foo' inside our test directory and fsync only the # directory. The btrfs fsync implementation had a bug that caused the new # directory entry to be visible after the fsync log replay but, the inode # of our file remained with a link count of 1. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_2 # Add a few more links and new files. # This is just to verify nothing breaks or gives incorrect results after the # fsync log is replayed. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_3 $XFS_IO_PROG -f -c "pwrite -S 0xff 0 64K" $SCRATCH_MNT/hello | _filter_xfs_io ln $SCRATCH_MNT/hello $SCRATCH_MNT/mydir/hello_2 # Add some subdirectories and new files and links to them. This is to verify # that after fsyncing our top level directory 'mydir', all the subdirectories # and their files/links are registered in the fsync log and exist after the # fsync log is replayed. mkdir -p $SCRATCH_MNT/mydir/x/y/z ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/foo_y_link ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/z/foo_z_link touch $SCRATCH_MNT/mydir/x/y/z/qwerty # Now fsync only our top directory. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/mydir # And fsync now our new file named 'hello', just to verify later that it has # the expected content and that the previous fsync on the directory 'mydir' had # no bad influence on this fsync. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/hello # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # Verify the content of our file 'foo' remains the same as before, 8192 bytes, # all with the value 0xaa. echo "File 'foo' content after log replay:" od -t x1 $SCRATCH_MNT/foo # Remove the first name of our inode. Because of the directory fsync bug, the # inode's link count was 1 instead of 5, so removing the 'foo' name ended up # deleting the inode and the other names became stale directory entries (still # visible to applications). Attempting to remove or access the remaining # dentries pointing to that inode resulted in stale file handle errors and # made it impossible to remove the parent directories since it was impossible # for them to become empty. echo "file 'foo' link count after log replay: $(stat -c %h $SCRATCH_MNT/foo)" rm -f $SCRATCH_MNT/foo # Now verify that all files, links and directories created before fsyncing our # directory exist after the fsync log was replayed. [ -f $SCRATCH_MNT/mydir/foo_2 ] || echo "Link mydir/foo_2 is missing" [ -f $SCRATCH_MNT/mydir/foo_3 ] || echo "Link mydir/foo_3 is missing" [ -f $SCRATCH_MNT/hello ] || echo "File hello is missing" [ -f $SCRATCH_MNT/mydir/hello_2 ] || echo "Link mydir/hello_2 is missing" [ -f $SCRATCH_MNT/mydir/x/y/foo_y_link ] || \ echo "Link mydir/x/y/foo_y_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/foo_z_link ] || \ echo "Link mydir/x/y/z/foo_z_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/qwerty ] || \ echo "File mydir/x/y/z/qwerty is missing" # We expect our file here to have a size of 64Kb and all the bytes having the # value 0xff. echo "file 'hello' content after log replay:" od -t x1 $SCRATCH_MNT/hello # Now remove all files/links, under our test directory 'mydir', and verify we # can remove all the directories. rm -f $SCRATCH_MNT/mydir/x/y/z/* rmdir $SCRATCH_MNT/mydir/x/y/z rm -f $SCRATCH_MNT/mydir/x/y/* rmdir $SCRATCH_MNT/mydir/x/y rmdir $SCRATCH_MNT/mydir/x rm -f $SCRATCH_MNT/mydir/* rmdir $SCRATCH_MNT/mydir # An fsck, run by the fstests framework everytime a test finishes, also detected # the inconsistency and printed the following error message: # # root 5 inode 257 errors 2001, no inode item, link count wrong # unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref # unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref status=0 exit The expected golden output for the test is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 5 file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 Which is the output after this patch and when running the test against ext3/4, xfs, f2fs, reiserfs or nilfs2. Without this patch, the test's output is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 1 Link mydir/foo_2 is missing Link mydir/foo_3 is missing Link mydir/x/y/foo_y_link is missing Link mydir/x/y/z/foo_z_link is missing File mydir/x/y/z/qwerty is missing file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y/z': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x': No such file or directory rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_2': Stale file handle rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_3': Stale file handle rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir': Directory not empty Fsck, without this fix, also complains about the wrong link count: root 5 inode 257 errors 2001, no inode item, link count wrong unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref So fix this by logging the inodes that the dentries point to when fsyncing a directory. A test case for xfstests follows. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-03-20 20:19:46 +03:00
if (log_dentries)
ret = log_new_dir_dentries(trans, root, inode, ctx);
Btrfs: fix metadata inconsistencies after directory fsync We can get into inconsistency between inodes and directory entries after fsyncing a directory. The issue is that while a directory gets the new dentries persisted in the fsync log and replayed at mount time, the link count of the inode that directory entries point to doesn't get updated, staying with an incorrect link count (smaller then the correct value). This later leads to stale file handle errors when accessing (including attempt to delete) some of the links if all the other ones are removed, which also implies impossibility to delete the parent directories, since the dentries can not be removed. Another issue is that (unlike ext3/4, xfs, f2fs, reiserfs, nilfs2), when fsyncing a directory, new files aren't logged (their metadata and dentries) nor any child directories. So this patch fixes this issue too, since it has the same resolution as the incorrect inode link count issue mentioned before. This is very easy to reproduce, and the following excerpt from my test case for xfstests shows how: _scratch_mkfs >> $seqres.full 2>&1 _init_flakey _mount_flakey # Create our main test file and directory. $XFS_IO_PROG -f -c "pwrite -S 0xaa 0 8K" $SCRATCH_MNT/foo | _filter_xfs_io mkdir $SCRATCH_MNT/mydir # Make sure all metadata and data are durably persisted. sync # Add a hard link to 'foo' inside our test directory and fsync only the # directory. The btrfs fsync implementation had a bug that caused the new # directory entry to be visible after the fsync log replay but, the inode # of our file remained with a link count of 1. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_2 # Add a few more links and new files. # This is just to verify nothing breaks or gives incorrect results after the # fsync log is replayed. ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/foo_3 $XFS_IO_PROG -f -c "pwrite -S 0xff 0 64K" $SCRATCH_MNT/hello | _filter_xfs_io ln $SCRATCH_MNT/hello $SCRATCH_MNT/mydir/hello_2 # Add some subdirectories and new files and links to them. This is to verify # that after fsyncing our top level directory 'mydir', all the subdirectories # and their files/links are registered in the fsync log and exist after the # fsync log is replayed. mkdir -p $SCRATCH_MNT/mydir/x/y/z ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/foo_y_link ln $SCRATCH_MNT/foo $SCRATCH_MNT/mydir/x/y/z/foo_z_link touch $SCRATCH_MNT/mydir/x/y/z/qwerty # Now fsync only our top directory. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/mydir # And fsync now our new file named 'hello', just to verify later that it has # the expected content and that the previous fsync on the directory 'mydir' had # no bad influence on this fsync. $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/hello # Simulate a crash/power loss. _load_flakey_table $FLAKEY_DROP_WRITES _unmount_flakey _load_flakey_table $FLAKEY_ALLOW_WRITES _mount_flakey # Verify the content of our file 'foo' remains the same as before, 8192 bytes, # all with the value 0xaa. echo "File 'foo' content after log replay:" od -t x1 $SCRATCH_MNT/foo # Remove the first name of our inode. Because of the directory fsync bug, the # inode's link count was 1 instead of 5, so removing the 'foo' name ended up # deleting the inode and the other names became stale directory entries (still # visible to applications). Attempting to remove or access the remaining # dentries pointing to that inode resulted in stale file handle errors and # made it impossible to remove the parent directories since it was impossible # for them to become empty. echo "file 'foo' link count after log replay: $(stat -c %h $SCRATCH_MNT/foo)" rm -f $SCRATCH_MNT/foo # Now verify that all files, links and directories created before fsyncing our # directory exist after the fsync log was replayed. [ -f $SCRATCH_MNT/mydir/foo_2 ] || echo "Link mydir/foo_2 is missing" [ -f $SCRATCH_MNT/mydir/foo_3 ] || echo "Link mydir/foo_3 is missing" [ -f $SCRATCH_MNT/hello ] || echo "File hello is missing" [ -f $SCRATCH_MNT/mydir/hello_2 ] || echo "Link mydir/hello_2 is missing" [ -f $SCRATCH_MNT/mydir/x/y/foo_y_link ] || \ echo "Link mydir/x/y/foo_y_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/foo_z_link ] || \ echo "Link mydir/x/y/z/foo_z_link is missing" [ -f $SCRATCH_MNT/mydir/x/y/z/qwerty ] || \ echo "File mydir/x/y/z/qwerty is missing" # We expect our file here to have a size of 64Kb and all the bytes having the # value 0xff. echo "file 'hello' content after log replay:" od -t x1 $SCRATCH_MNT/hello # Now remove all files/links, under our test directory 'mydir', and verify we # can remove all the directories. rm -f $SCRATCH_MNT/mydir/x/y/z/* rmdir $SCRATCH_MNT/mydir/x/y/z rm -f $SCRATCH_MNT/mydir/x/y/* rmdir $SCRATCH_MNT/mydir/x/y rmdir $SCRATCH_MNT/mydir/x rm -f $SCRATCH_MNT/mydir/* rmdir $SCRATCH_MNT/mydir # An fsck, run by the fstests framework everytime a test finishes, also detected # the inconsistency and printed the following error message: # # root 5 inode 257 errors 2001, no inode item, link count wrong # unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref # unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref status=0 exit The expected golden output for the test is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 5 file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 Which is the output after this patch and when running the test against ext3/4, xfs, f2fs, reiserfs or nilfs2. Without this patch, the test's output is: wrote 8192/8192 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) wrote 65536/65536 bytes at offset 0 XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec) File 'foo' content after log replay: 0000000 aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa * 0020000 file 'foo' link count after log replay: 1 Link mydir/foo_2 is missing Link mydir/foo_3 is missing Link mydir/x/y/foo_y_link is missing Link mydir/x/y/z/foo_z_link is missing File mydir/x/y/z/qwerty is missing file 'hello' content after log replay: 0000000 ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff * 0200000 rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y/z': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x/y': No such file or directory rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/x': No such file or directory rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_2': Stale file handle rm: cannot remove '/home/fdmanana/btrfs-tests/scratch_1/mydir/foo_3': Stale file handle rmdir: failed to remove '/home/fdmanana/btrfs-tests/scratch_1/mydir': Directory not empty Fsck, without this fix, also complains about the wrong link count: root 5 inode 257 errors 2001, no inode item, link count wrong unresolved ref dir 258 index 2 namelen 5 name foo_2 filetype 1 errors 4, no inode ref unresolved ref dir 258 index 3 namelen 5 name foo_3 filetype 1 errors 4, no inode ref So fix this by logging the inodes that the dentries point to when fsyncing a directory. A test case for xfstests follows. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2015-03-20 20:19:46 +03:00
else
ret = 0;
end_trans:
if (ret < 0) {
btrfs_set_log_full_commit(trans);
ret = 1;
}
if (ret)
btrfs_remove_log_ctx(root, ctx);
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 17:24:20 +03:00
btrfs_end_log_trans(root);
end_no_trans:
return ret;
}
/*
* it is not safe to log dentry if the chunk root has added new
* chunks. This returns 0 if the dentry was logged, and 1 otherwise.
* If this returns 1, you must commit the transaction to safely get your
* data on disk.
*/
int btrfs_log_dentry_safe(struct btrfs_trans_handle *trans,
struct dentry *dentry,
struct btrfs_log_ctx *ctx)
{
struct dentry *parent = dget_parent(dentry);
int ret;
ret = btrfs_log_inode_parent(trans, BTRFS_I(d_inode(dentry)), parent,
btrfs: make fast fsyncs wait only for writeback Currently regardless of a full or a fast fsync we always wait for ordered extents to complete, and then start logging the inode after that. However for fast fsyncs we can just wait for the writeback to complete, we don't need to wait for the ordered extents to complete since we use the list of modified extents maps to figure out which extents we must log and we can get their checksums directly from the ordered extents that are still in flight, otherwise look them up from the checksums tree. Until commit b5e6c3e170b770 ("btrfs: always wait on ordered extents at fsync time"), for fast fsyncs, we used to start logging without even waiting for the writeback to complete first, we would wait for it to complete after logging, while holding a transaction open, which lead to performance issues when using cgroups and probably for other cases too, as wait for IO while holding a transaction handle should be avoided as much as possible. After that, for fast fsyncs, we started to wait for ordered extents to complete before starting to log, which adds some latency to fsyncs and we even got at least one report about a performance drop which bisected to that particular change: https://lore.kernel.org/linux-btrfs/20181109215148.GF23260@techsingularity.net/ This change makes fast fsyncs only wait for writeback to finish before starting to log the inode, instead of waiting for both the writeback to finish and for the ordered extents to complete. This brings back part of the logic we had that extracts checksums from in flight ordered extents, which are not yet in the checksums tree, and making sure transaction commits wait for the completion of ordered extents previously logged (by far most of the time they have already completed by the time a transaction commit starts, resulting in no wait at all), to avoid any data loss if an ordered extent completes after the transaction used to log an inode is committed, followed by a power failure. When there are no other tasks accessing the checksums and the subvolume btrees, the ordered extent completion is pretty fast, typically taking 100 to 200 microseconds only in my observations. However when there are other tasks accessing these btrees, ordered extent completion can take a lot more time due to lock contention on nodes and leaves of these btrees. I've seen cases over 2 milliseconds, which starts to be significant. In particular when we do have concurrent fsyncs against different files there is a lot of contention on the checksums btree, since we have many tasks writing the checksums into the btree and other tasks that already started the logging phase are doing lookups for checksums in the btree. This change also turns all ranged fsyncs into full ranged fsyncs, which is something we already did when not using the NO_HOLES features or when doing a full fsync. This is to guarantee we never miss checksums due to writeback having been triggered only for a part of an extent, and we end up logging the full extent but only checksums for the written range, which results in missing checksums after log replay. Allowing ranged fsyncs to operate again only in the original range, when using the NO_HOLES feature and doing a fast fsync is doable but requires some non trivial changes to the writeback path, which can always be worked on later if needed, but I don't think they are a very common use case. Several tests were performed using fio for different numbers of concurrent jobs, each writing and fsyncing its own file, for both sequential and random file writes. The tests were run on bare metal, no virtualization, on a box with 12 cores (Intel i7-8700), 64Gb of RAM and a NVMe device, with a kernel configuration that is the default of typical distributions (debian in this case), without debug options enabled (kasan, kmemleak, slub debug, debug of page allocations, lock debugging, etc). The following script that calls fio was used: $ cat test-fsync.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/btrfs MOUNT_OPTIONS="-o ssd -o space_cache=v2" MKFS_OPTIONS="-d single -m single" if [ $# -ne 5 ]; then echo "Use $0 NUM_JOBS FILE_SIZE FSYNC_FREQ BLOCK_SIZE [write|randwrite]" exit 1 fi NUM_JOBS=$1 FILE_SIZE=$2 FSYNC_FREQ=$3 BLOCK_SIZE=$4 WRITE_MODE=$5 if [ "$WRITE_MODE" != "write" ] && [ "$WRITE_MODE" != "randwrite" ]; then echo "Invalid WRITE_MODE, must be 'write' or 'randwrite'" exit 1 fi cat <<EOF > /tmp/fio-job.ini [writers] rw=$WRITE_MODE fsync=$FSYNC_FREQ fallocate=none group_reporting=1 direct=0 bs=$BLOCK_SIZE ioengine=sync size=$FILE_SIZE directory=$MNT numjobs=$NUM_JOBS EOF echo "performance" | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor echo echo "Using config:" echo cat /tmp/fio-job.ini echo umount $MNT &> /dev/null mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT fio /tmp/fio-job.ini umount $MNT The results were the following: ************************* *** sequential writes *** ************************* ==== 1 job, 8GiB file, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=36.6MiB/s (38.4MB/s), 36.6MiB/s-36.6MiB/s (38.4MB/s-38.4MB/s), io=8192MiB (8590MB), run=223689-223689msec After patch: WRITE: bw=40.2MiB/s (42.1MB/s), 40.2MiB/s-40.2MiB/s (42.1MB/s-42.1MB/s), io=8192MiB (8590MB), run=203980-203980msec (+9.8%, -8.8% runtime) ==== 2 jobs, 4GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=35.8MiB/s (37.5MB/s), 35.8MiB/s-35.8MiB/s (37.5MB/s-37.5MB/s), io=8192MiB (8590MB), run=228950-228950msec After patch: WRITE: bw=43.5MiB/s (45.6MB/s), 43.5MiB/s-43.5MiB/s (45.6MB/s-45.6MB/s), io=8192MiB (8590MB), run=188272-188272msec (+21.5% throughput, -17.8% runtime) ==== 4 jobs, 2GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=50.1MiB/s (52.6MB/s), 50.1MiB/s-50.1MiB/s (52.6MB/s-52.6MB/s), io=8192MiB (8590MB), run=163446-163446msec After patch: WRITE: bw=64.5MiB/s (67.6MB/s), 64.5MiB/s-64.5MiB/s (67.6MB/s-67.6MB/s), io=8192MiB (8590MB), run=126987-126987msec (+28.7% throughput, -22.3% runtime) ==== 8 jobs, 1GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=64.0MiB/s (68.1MB/s), 64.0MiB/s-64.0MiB/s (68.1MB/s-68.1MB/s), io=8192MiB (8590MB), run=126075-126075msec After patch: WRITE: bw=86.8MiB/s (91.0MB/s), 86.8MiB/s-86.8MiB/s (91.0MB/s-91.0MB/s), io=8192MiB (8590MB), run=94358-94358msec (+35.6% throughput, -25.2% runtime) ==== 16 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=79.8MiB/s (83.6MB/s), 79.8MiB/s-79.8MiB/s (83.6MB/s-83.6MB/s), io=8192MiB (8590MB), run=102694-102694msec After patch: WRITE: bw=107MiB/s (112MB/s), 107MiB/s-107MiB/s (112MB/s-112MB/s), io=8192MiB (8590MB), run=76446-76446msec (+34.1% throughput, -25.6% runtime) ==== 32 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=93.2MiB/s (97.7MB/s), 93.2MiB/s-93.2MiB/s (97.7MB/s-97.7MB/s), io=16.0GiB (17.2GB), run=175836-175836msec After patch: WRITE: bw=111MiB/s (117MB/s), 111MiB/s-111MiB/s (117MB/s-117MB/s), io=16.0GiB (17.2GB), run=147001-147001msec (+19.1% throughput, -16.4% runtime) ==== 64 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=108MiB/s (114MB/s), 108MiB/s-108MiB/s (114MB/s-114MB/s), io=32.0GiB (34.4GB), run=302656-302656msec After patch: WRITE: bw=133MiB/s (140MB/s), 133MiB/s-133MiB/s (140MB/s-140MB/s), io=32.0GiB (34.4GB), run=246003-246003msec (+23.1% throughput, -18.7% runtime) ************************ *** random writes *** ************************ ==== 1 job, 8GiB file, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=11.5MiB/s (12.0MB/s), 11.5MiB/s-11.5MiB/s (12.0MB/s-12.0MB/s), io=8192MiB (8590MB), run=714281-714281msec After patch: WRITE: bw=11.6MiB/s (12.2MB/s), 11.6MiB/s-11.6MiB/s (12.2MB/s-12.2MB/s), io=8192MiB (8590MB), run=705959-705959msec (+0.9% throughput, -1.7% runtime) ==== 2 jobs, 4GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=12.8MiB/s (13.5MB/s), 12.8MiB/s-12.8MiB/s (13.5MB/s-13.5MB/s), io=8192MiB (8590MB), run=638101-638101msec After patch: WRITE: bw=13.1MiB/s (13.7MB/s), 13.1MiB/s-13.1MiB/s (13.7MB/s-13.7MB/s), io=8192MiB (8590MB), run=625374-625374msec (+2.3% throughput, -2.0% runtime) ==== 4 jobs, 2GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=15.4MiB/s (16.2MB/s), 15.4MiB/s-15.4MiB/s (16.2MB/s-16.2MB/s), io=8192MiB (8590MB), run=531146-531146msec After patch: WRITE: bw=17.8MiB/s (18.7MB/s), 17.8MiB/s-17.8MiB/s (18.7MB/s-18.7MB/s), io=8192MiB (8590MB), run=460431-460431msec (+15.6% throughput, -13.3% runtime) ==== 8 jobs, 1GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=19.9MiB/s (20.8MB/s), 19.9MiB/s-19.9MiB/s (20.8MB/s-20.8MB/s), io=8192MiB (8590MB), run=412664-412664msec After patch: WRITE: bw=22.2MiB/s (23.3MB/s), 22.2MiB/s-22.2MiB/s (23.3MB/s-23.3MB/s), io=8192MiB (8590MB), run=368589-368589msec (+11.6% throughput, -10.7% runtime) ==== 16 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=29.3MiB/s (30.7MB/s), 29.3MiB/s-29.3MiB/s (30.7MB/s-30.7MB/s), io=8192MiB (8590MB), run=279924-279924msec After patch: WRITE: bw=30.4MiB/s (31.9MB/s), 30.4MiB/s-30.4MiB/s (31.9MB/s-31.9MB/s), io=8192MiB (8590MB), run=269258-269258msec (+3.8% throughput, -3.8% runtime) ==== 32 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=36.9MiB/s (38.7MB/s), 36.9MiB/s-36.9MiB/s (38.7MB/s-38.7MB/s), io=16.0GiB (17.2GB), run=443581-443581msec After patch: WRITE: bw=41.6MiB/s (43.6MB/s), 41.6MiB/s-41.6MiB/s (43.6MB/s-43.6MB/s), io=16.0GiB (17.2GB), run=394114-394114msec (+12.7% throughput, -11.2% runtime) ==== 64 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=45.9MiB/s (48.1MB/s), 45.9MiB/s-45.9MiB/s (48.1MB/s-48.1MB/s), io=32.0GiB (34.4GB), run=714614-714614msec After patch: WRITE: bw=48.8MiB/s (51.1MB/s), 48.8MiB/s-48.8MiB/s (51.1MB/s-51.1MB/s), io=32.0GiB (34.4GB), run=672087-672087msec (+6.3% throughput, -6.0% runtime) Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-08-11 14:43:58 +03:00
LOG_INODE_ALL, ctx);
dput(parent);
return ret;
}
/*
* should be called during mount to recover any replay any log trees
* from the FS
*/
int btrfs_recover_log_trees(struct btrfs_root *log_root_tree)
{
int ret;
struct btrfs_path *path;
struct btrfs_trans_handle *trans;
struct btrfs_key key;
struct btrfs_key found_key;
struct btrfs_root *log;
struct btrfs_fs_info *fs_info = log_root_tree->fs_info;
struct walk_control wc = {
.process_func = process_one_buffer,
.stage = LOG_WALK_PIN_ONLY,
};
path = btrfs_alloc_path();
if (!path)
return -ENOMEM;
set_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags);
trans = btrfs_start_transaction(fs_info->tree_root, 0);
if (IS_ERR(trans)) {
ret = PTR_ERR(trans);
goto error;
}
wc.trans = trans;
wc.pin = 1;
ret = walk_log_tree(trans, log_root_tree, &wc);
if (ret) {
btrfs_abort_transaction(trans, ret);
goto error;
}
again:
key.objectid = BTRFS_TREE_LOG_OBJECTID;
key.offset = (u64)-1;
key.type = BTRFS_ROOT_ITEM_KEY;
while (1) {
ret = btrfs_search_slot(NULL, log_root_tree, &key, path, 0, 0);
if (ret < 0) {
btrfs_abort_transaction(trans, ret);
goto error;
}
if (ret > 0) {
if (path->slots[0] == 0)
break;
path->slots[0]--;
}
btrfs_item_key_to_cpu(path->nodes[0], &found_key,
path->slots[0]);
btrfs_release_path(path);
if (found_key.objectid != BTRFS_TREE_LOG_OBJECTID)
break;
log = btrfs_read_tree_root(log_root_tree, &found_key);
if (IS_ERR(log)) {
ret = PTR_ERR(log);
btrfs_abort_transaction(trans, ret);
goto error;
}
wc.replay_dest = btrfs_get_fs_root(fs_info, found_key.offset,
true);
if (IS_ERR(wc.replay_dest)) {
ret = PTR_ERR(wc.replay_dest);
/*
* We didn't find the subvol, likely because it was
* deleted. This is ok, simply skip this log and go to
* the next one.
*
* We need to exclude the root because we can't have
* other log replays overwriting this log as we'll read
* it back in a few more times. This will keep our
* block from being modified, and we'll just bail for
* each subsequent pass.
*/
if (ret == -ENOENT)
ret = btrfs_pin_extent_for_log_replay(trans,
log->node->start,
log->node->len);
btrfs_put_root(log);
if (!ret)
goto next;
btrfs_abort_transaction(trans, ret);
goto error;
}
wc.replay_dest->log_root = log;
ret = btrfs_record_root_in_trans(trans, wc.replay_dest);
if (ret)
/* The loop needs to continue due to the root refs */
btrfs_abort_transaction(trans, ret);
else
ret = walk_log_tree(trans, log, &wc);
if (!ret && wc.stage == LOG_WALK_REPLAY_ALL) {
ret = fixup_inode_link_counts(trans, wc.replay_dest,
path);
if (ret)
btrfs_abort_transaction(trans, ret);
}
if (!ret && wc.stage == LOG_WALK_REPLAY_ALL) {
struct btrfs_root *root = wc.replay_dest;
btrfs_release_path(path);
/*
* We have just replayed everything, and the highest
* objectid of fs roots probably has changed in case
* some inode_item's got replayed.
*
* root->objectid_mutex is not acquired as log replay
* could only happen during mount.
*/
ret = btrfs_init_root_free_objectid(root);
if (ret)
btrfs_abort_transaction(trans, ret);
}
wc.replay_dest->log_root = NULL;
btrfs_put_root(wc.replay_dest);
btrfs_put_root(log);
if (ret)
goto error;
next:
if (found_key.offset == 0)
break;
key.offset = found_key.offset - 1;
}
btrfs_release_path(path);
/* step one is to pin it all, step two is to replay just inodes */
if (wc.pin) {
wc.pin = 0;
wc.process_func = replay_one_buffer;
wc.stage = LOG_WALK_REPLAY_INODES;
goto again;
}
/* step three is to replay everything */
if (wc.stage < LOG_WALK_REPLAY_ALL) {
wc.stage++;
goto again;
}
btrfs_free_path(path);
/* step 4: commit the transaction, which also unpins the blocks */
ret = btrfs_commit_transaction(trans);
if (ret)
return ret;
log_root_tree->log_root = NULL;
clear_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags);
btrfs_put_root(log_root_tree);
return 0;
error:
if (wc.trans)
btrfs_end_transaction(wc.trans);
clear_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags);
btrfs_free_path(path);
return ret;
}
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 17:24:20 +03:00
/*
* there are some corner cases where we want to force a full
* commit instead of allowing a directory to be logged.
*
* They revolve around files there were unlinked from the directory, and
* this function updates the parent directory so that a full commit is
* properly done if it is fsync'd later after the unlinks are done.
Btrfs: fix file loss on log replay after renaming a file and fsync We have two cases where we end up deleting a file at log replay time when we should not. For this to happen the file must have been renamed and a directory inode must have been fsynced/logged. Two examples that exercise these two cases are listed below. Case 1) $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir -p /mnt/a/b $ mkdir /mnt/c $ touch /mnt/a/b/foo $ sync $ mv /mnt/a/b/foo /mnt/c/ # Create file bar just to make sure the fsync on directory a/ does # something and it's not a no-op. $ touch /mnt/a/bar $ xfs_io -c "fsync" /mnt/a < power fail / crash > The next time the filesystem is mounted, the log replay procedure deletes file foo. Case 2) $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/a $ mkdir /mnt/b $ mkdir /mnt/c $ touch /mnt/a/foo $ ln /mnt/a/foo /mnt/b/foo_link $ touch /mnt/b/bar $ sync $ unlink /mnt/b/foo_link $ mv /mnt/b/bar /mnt/c/ $ xfs_io -c "fsync" /mnt/a/foo < power fail / crash > The next time the filesystem is mounted, the log replay procedure deletes file bar. The reason why the files are deleted is because when we log inodes other then the fsync target inode, we ignore their last_unlink_trans value and leave the log without enough information to later replay the rename operations. So we need to look at the last_unlink_trans values and fallback to a transaction commit if they are greater than the id of the last committed transaction. So fix this by looking at the last_unlink_trans values and fallback to transaction commits when needed. Also, when logging other inodes (for case 1 we logged descendants of the fsync target inode while for case 2 we logged ascendants) we need to care about concurrent tasks updating the last_unlink_trans of inodes we are logging (which was already an existing problem in check_parent_dirs_for_sync()). Since we can not acquire their inode mutex (vfs' struct inode ->i_mutex), as that causes deadlocks with other concurrent operations that acquire the i_mutex of 2 inodes (other fsyncs or renames for example), we need to serialize on the log_mutex of the inode we are logging. A task setting a new value for an inode's last_unlink_trans must acquire the inode's log_mutex and it must do this update before doing the actual unlink operation (which is already the case except when deleting a snapshot). Conversely the task logging the inode must first log the inode and then check the inode's last_unlink_trans value while holding its log_mutex, as if its value is not greater then the id of the last committed transaction it means it logged a safe state of the inode's items, while if its value is not smaller then the id of the last committed transaction it means the inode state it has logged might not be safe (the concurrent task might have just updated last_unlink_trans but hasn't done yet the unlink operation) and therefore a transaction commit must be done. Test cases for xfstests follow in separate patches. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2016-02-12 14:34:23 +03:00
*
* Must be called before the unlink operations (updates to the subvolume tree,
* inodes, etc) are done.
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 17:24:20 +03:00
*/
void btrfs_record_unlink_dir(struct btrfs_trans_handle *trans,
struct btrfs_inode *dir, struct btrfs_inode *inode,
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 17:24:20 +03:00
int for_rename)
{
/*
* when we're logging a file, if it hasn't been renamed
* or unlinked, and its inode is fully committed on disk,
* we don't have to worry about walking up the directory chain
* to log its parents.
*
* So, we use the last_unlink_trans field to put this transid
* into the file. When the file is logged we check it and
* don't log the parents if the file is fully on disk.
*/
mutex_lock(&inode->log_mutex);
inode->last_unlink_trans = trans->transid;
mutex_unlock(&inode->log_mutex);
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 17:24:20 +03:00
/*
* if this directory was already logged any new
* names for this file/dir will get recorded
*/
if (dir->logged_trans == trans->transid)
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 17:24:20 +03:00
return;
/*
* if the inode we're about to unlink was logged,
* the log will be properly updated for any new names
*/
if (inode->logged_trans == trans->transid)
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 17:24:20 +03:00
return;
/*
* when renaming files across directories, if the directory
* there we're unlinking from gets fsync'd later on, there's
* no way to find the destination directory later and fsync it
* properly. So, we have to be conservative and force commits
* so the new name gets discovered.
*/
if (for_rename)
goto record;
/* we can safely do the unlink without any special recording */
return;
record:
mutex_lock(&dir->log_mutex);
dir->last_unlink_trans = trans->transid;
mutex_unlock(&dir->log_mutex);
Btrfs: fix unreplayable log after snapshot delete + parent dir fsync If we delete a snapshot, fsync its parent directory and crash/power fail before the next transaction commit, on the next mount when we attempt to replay the log tree of the root containing the parent directory we will fail and prevent the filesystem from mounting, which is solvable by wiping out the log trees with the btrfs-zero-log tool but very inconvenient as we will lose any data and metadata fsynced before the parent directory was fsynced. For example: $ mkfs.btrfs -f /dev/sdc $ mount /dev/sdc /mnt $ mkdir /mnt/testdir $ btrfs subvolume snapshot /mnt /mnt/testdir/snap $ btrfs subvolume delete /mnt/testdir/snap $ xfs_io -c "fsync" /mnt/testdir < crash / power failure and reboot > $ mount /dev/sdc /mnt mount: mount(2) failed: No such file or directory And in dmesg/syslog we get the following message and trace: [192066.361162] BTRFS info (device dm-0): failed to delete reference to snap, inode 257 parent 257 [192066.363010] ------------[ cut here ]------------ [192066.365268] WARNING: CPU: 4 PID: 5130 at fs/btrfs/inode.c:3986 __btrfs_unlink_inode+0x17a/0x354 [btrfs]() [192066.367250] BTRFS: Transaction aborted (error -2) [192066.368401] Modules linked in: btrfs dm_flakey dm_mod ppdev sha256_generic xor raid6_pq hmac drbg ansi_cprng aesni_intel acpi_cpufreq tpm_tis aes_x86_64 tpm ablk_helper evdev cryptd sg parport_pc i2c_piix4 psmouse lrw parport i2c_core pcspkr gf128mul processor serio_raw glue_helper button loop autofs4 ext4 crc16 mbcache jbd2 sd_mod sr_mod cdrom ata_generic virtio_scsi ata_piix libata virtio_pci virtio_ring crc32c_intel scsi_mod e1000 virtio floppy [last unloaded: btrfs] [192066.377154] CPU: 4 PID: 5130 Comm: mount Tainted: G W 4.4.0-rc6-btrfs-next-20+ #1 [192066.378875] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS by qemu-project.org 04/01/2014 [192066.380889] 0000000000000000 ffff880143923670 ffffffff81257570 ffff8801439236b8 [192066.382561] ffff8801439236a8 ffffffff8104ec07 ffffffffa039dc2c 00000000fffffffe [192066.384191] ffff8801ed31d000 ffff8801b9fc9c88 ffff8801086875e0 ffff880143923710 [192066.385827] Call Trace: [192066.386373] [<ffffffff81257570>] dump_stack+0x4e/0x79 [192066.387387] [<ffffffff8104ec07>] warn_slowpath_common+0x99/0xb2 [192066.388429] [<ffffffffa039dc2c>] ? __btrfs_unlink_inode+0x17a/0x354 [btrfs] [192066.389236] [<ffffffff8104ec68>] warn_slowpath_fmt+0x48/0x50 [192066.389884] [<ffffffffa039dc2c>] __btrfs_unlink_inode+0x17a/0x354 [btrfs] [192066.390621] [<ffffffff81184b55>] ? iput+0xb0/0x266 [192066.391200] [<ffffffffa039ea25>] btrfs_unlink_inode+0x1c/0x3d [btrfs] [192066.391930] [<ffffffffa03ca623>] check_item_in_log+0x1fe/0x29b [btrfs] [192066.392715] [<ffffffffa03ca827>] replay_dir_deletes+0x167/0x1cf [btrfs] [192066.393510] [<ffffffffa03cccc7>] replay_one_buffer+0x417/0x570 [btrfs] [192066.394241] [<ffffffffa03ca164>] walk_up_log_tree+0x10e/0x1dc [btrfs] [192066.394958] [<ffffffffa03cac72>] walk_log_tree+0xa5/0x190 [btrfs] [192066.395628] [<ffffffffa03ce8b8>] btrfs_recover_log_trees+0x239/0x32c [btrfs] [192066.396790] [<ffffffffa03cc8b0>] ? replay_one_extent+0x50a/0x50a [btrfs] [192066.397891] [<ffffffffa0394041>] open_ctree+0x1d8b/0x2167 [btrfs] [192066.398897] [<ffffffffa03706e1>] btrfs_mount+0x5ef/0x729 [btrfs] [192066.399823] [<ffffffff8108ad98>] ? trace_hardirqs_on+0xd/0xf [192066.400739] [<ffffffff8108959b>] ? lockdep_init_map+0xb9/0x1b3 [192066.401700] [<ffffffff811714b9>] mount_fs+0x67/0x131 [192066.402482] [<ffffffff81188560>] vfs_kern_mount+0x6c/0xde [192066.403930] [<ffffffffa03702bd>] btrfs_mount+0x1cb/0x729 [btrfs] [192066.404831] [<ffffffff8108ad98>] ? trace_hardirqs_on+0xd/0xf [192066.405726] [<ffffffff8108959b>] ? lockdep_init_map+0xb9/0x1b3 [192066.406621] [<ffffffff811714b9>] mount_fs+0x67/0x131 [192066.407401] [<ffffffff81188560>] vfs_kern_mount+0x6c/0xde [192066.408247] [<ffffffff8118ae36>] do_mount+0x893/0x9d2 [192066.409047] [<ffffffff8113009b>] ? strndup_user+0x3f/0x8c [192066.409842] [<ffffffff8118b187>] SyS_mount+0x75/0xa1 [192066.410621] [<ffffffff8147e517>] entry_SYSCALL_64_fastpath+0x12/0x6b [192066.411572] ---[ end trace 2de42126c1e0a0f0 ]--- [192066.412344] BTRFS: error (device dm-0) in __btrfs_unlink_inode:3986: errno=-2 No such entry [192066.413748] BTRFS: error (device dm-0) in btrfs_replay_log:2464: errno=-2 No such entry (Failed to recover log tree) [192066.415458] BTRFS error (device dm-0): cleaner transaction attach returned -30 [192066.444613] BTRFS: open_ctree failed This happens because when we are replaying the log and processing the directory entry pointing to the snapshot in the subvolume tree, we treat its btrfs_dir_item item as having a location with a key type matching BTRFS_INODE_ITEM_KEY, which is wrong because the type matches BTRFS_ROOT_ITEM_KEY and therefore must be processed differently, as the object id refers to a root number and not to an inode in the root containing the parent directory. So fix this by triggering a transaction commit if an fsync against the parent directory is requested after deleting a snapshot. This is the simplest approach for a rare use case. Some alternative that avoids the transaction commit would require more code to explicitly delete the snapshot at log replay time (factoring out common code from ioctl.c: btrfs_ioctl_snap_destroy()), special care at fsync time to remove the log tree of the snapshot's root from the log root of the root of tree roots, amongst other steps. A test case for xfstests that triggers the issue follows. seq=`basename $0` seqres=$RESULT_DIR/$seq echo "QA output created by $seq" tmp=/tmp/$$ status=1 # failure is the default! trap "_cleanup; exit \$status" 0 1 2 3 15 _cleanup() { _cleanup_flakey cd / rm -f $tmp.* } # get standard environment, filters and checks . ./common/rc . ./common/filter . ./common/dmflakey # real QA test starts here _need_to_be_root _supported_fs btrfs _supported_os Linux _require_scratch _require_dm_target flakey _require_metadata_journaling $SCRATCH_DEV rm -f $seqres.full _scratch_mkfs >>$seqres.full 2>&1 _init_flakey _mount_flakey # Create a snapshot at the root of our filesystem (mount point path), delete it, # fsync the mount point path, crash and mount to replay the log. This should # succeed and after the filesystem is mounted the snapshot should not be visible # anymore. _run_btrfs_util_prog subvolume snapshot $SCRATCH_MNT $SCRATCH_MNT/snap1 _run_btrfs_util_prog subvolume delete $SCRATCH_MNT/snap1 $XFS_IO_PROG -c "fsync" $SCRATCH_MNT _flakey_drop_and_remount [ -e $SCRATCH_MNT/snap1 ] && \ echo "Snapshot snap1 still exists after log replay" # Similar scenario as above, but this time the snapshot is created inside a # directory and not directly under the root (mount point path). mkdir $SCRATCH_MNT/testdir _run_btrfs_util_prog subvolume snapshot $SCRATCH_MNT $SCRATCH_MNT/testdir/snap2 _run_btrfs_util_prog subvolume delete $SCRATCH_MNT/testdir/snap2 $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/testdir _flakey_drop_and_remount [ -e $SCRATCH_MNT/testdir/snap2 ] && \ echo "Snapshot snap2 still exists after log replay" _unmount_flakey echo "Silence is golden" status=0 exit Signed-off-by: Filipe Manana <fdmanana@suse.com> Tested-by: Liu Bo <bo.li.liu@oracle.com> Reviewed-by: Liu Bo <bo.li.liu@oracle.com> Signed-off-by: Chris Mason <clm@fb.com>
2016-02-10 13:42:25 +03:00
}
/*
* Make sure that if someone attempts to fsync the parent directory of a deleted
* snapshot, it ends up triggering a transaction commit. This is to guarantee
* that after replaying the log tree of the parent directory's root we will not
* see the snapshot anymore and at log replay time we will not see any log tree
* corresponding to the deleted snapshot's root, which could lead to replaying
* it after replaying the log tree of the parent directory (which would replay
* the snapshot delete operation).
Btrfs: fix file loss on log replay after renaming a file and fsync We have two cases where we end up deleting a file at log replay time when we should not. For this to happen the file must have been renamed and a directory inode must have been fsynced/logged. Two examples that exercise these two cases are listed below. Case 1) $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir -p /mnt/a/b $ mkdir /mnt/c $ touch /mnt/a/b/foo $ sync $ mv /mnt/a/b/foo /mnt/c/ # Create file bar just to make sure the fsync on directory a/ does # something and it's not a no-op. $ touch /mnt/a/bar $ xfs_io -c "fsync" /mnt/a < power fail / crash > The next time the filesystem is mounted, the log replay procedure deletes file foo. Case 2) $ mkfs.btrfs -f /dev/sdb $ mount /dev/sdb /mnt $ mkdir /mnt/a $ mkdir /mnt/b $ mkdir /mnt/c $ touch /mnt/a/foo $ ln /mnt/a/foo /mnt/b/foo_link $ touch /mnt/b/bar $ sync $ unlink /mnt/b/foo_link $ mv /mnt/b/bar /mnt/c/ $ xfs_io -c "fsync" /mnt/a/foo < power fail / crash > The next time the filesystem is mounted, the log replay procedure deletes file bar. The reason why the files are deleted is because when we log inodes other then the fsync target inode, we ignore their last_unlink_trans value and leave the log without enough information to later replay the rename operations. So we need to look at the last_unlink_trans values and fallback to a transaction commit if they are greater than the id of the last committed transaction. So fix this by looking at the last_unlink_trans values and fallback to transaction commits when needed. Also, when logging other inodes (for case 1 we logged descendants of the fsync target inode while for case 2 we logged ascendants) we need to care about concurrent tasks updating the last_unlink_trans of inodes we are logging (which was already an existing problem in check_parent_dirs_for_sync()). Since we can not acquire their inode mutex (vfs' struct inode ->i_mutex), as that causes deadlocks with other concurrent operations that acquire the i_mutex of 2 inodes (other fsyncs or renames for example), we need to serialize on the log_mutex of the inode we are logging. A task setting a new value for an inode's last_unlink_trans must acquire the inode's log_mutex and it must do this update before doing the actual unlink operation (which is already the case except when deleting a snapshot). Conversely the task logging the inode must first log the inode and then check the inode's last_unlink_trans value while holding its log_mutex, as if its value is not greater then the id of the last committed transaction it means it logged a safe state of the inode's items, while if its value is not smaller then the id of the last committed transaction it means the inode state it has logged might not be safe (the concurrent task might have just updated last_unlink_trans but hasn't done yet the unlink operation) and therefore a transaction commit must be done. Test cases for xfstests follow in separate patches. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Chris Mason <clm@fb.com>
2016-02-12 14:34:23 +03:00
*
* Must be called before the actual snapshot destroy operation (updates to the
* parent root and tree of tree roots trees, etc) are done.
Btrfs: fix unreplayable log after snapshot delete + parent dir fsync If we delete a snapshot, fsync its parent directory and crash/power fail before the next transaction commit, on the next mount when we attempt to replay the log tree of the root containing the parent directory we will fail and prevent the filesystem from mounting, which is solvable by wiping out the log trees with the btrfs-zero-log tool but very inconvenient as we will lose any data and metadata fsynced before the parent directory was fsynced. For example: $ mkfs.btrfs -f /dev/sdc $ mount /dev/sdc /mnt $ mkdir /mnt/testdir $ btrfs subvolume snapshot /mnt /mnt/testdir/snap $ btrfs subvolume delete /mnt/testdir/snap $ xfs_io -c "fsync" /mnt/testdir < crash / power failure and reboot > $ mount /dev/sdc /mnt mount: mount(2) failed: No such file or directory And in dmesg/syslog we get the following message and trace: [192066.361162] BTRFS info (device dm-0): failed to delete reference to snap, inode 257 parent 257 [192066.363010] ------------[ cut here ]------------ [192066.365268] WARNING: CPU: 4 PID: 5130 at fs/btrfs/inode.c:3986 __btrfs_unlink_inode+0x17a/0x354 [btrfs]() [192066.367250] BTRFS: Transaction aborted (error -2) [192066.368401] Modules linked in: btrfs dm_flakey dm_mod ppdev sha256_generic xor raid6_pq hmac drbg ansi_cprng aesni_intel acpi_cpufreq tpm_tis aes_x86_64 tpm ablk_helper evdev cryptd sg parport_pc i2c_piix4 psmouse lrw parport i2c_core pcspkr gf128mul processor serio_raw glue_helper button loop autofs4 ext4 crc16 mbcache jbd2 sd_mod sr_mod cdrom ata_generic virtio_scsi ata_piix libata virtio_pci virtio_ring crc32c_intel scsi_mod e1000 virtio floppy [last unloaded: btrfs] [192066.377154] CPU: 4 PID: 5130 Comm: mount Tainted: G W 4.4.0-rc6-btrfs-next-20+ #1 [192066.378875] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS by qemu-project.org 04/01/2014 [192066.380889] 0000000000000000 ffff880143923670 ffffffff81257570 ffff8801439236b8 [192066.382561] ffff8801439236a8 ffffffff8104ec07 ffffffffa039dc2c 00000000fffffffe [192066.384191] ffff8801ed31d000 ffff8801b9fc9c88 ffff8801086875e0 ffff880143923710 [192066.385827] Call Trace: [192066.386373] [<ffffffff81257570>] dump_stack+0x4e/0x79 [192066.387387] [<ffffffff8104ec07>] warn_slowpath_common+0x99/0xb2 [192066.388429] [<ffffffffa039dc2c>] ? __btrfs_unlink_inode+0x17a/0x354 [btrfs] [192066.389236] [<ffffffff8104ec68>] warn_slowpath_fmt+0x48/0x50 [192066.389884] [<ffffffffa039dc2c>] __btrfs_unlink_inode+0x17a/0x354 [btrfs] [192066.390621] [<ffffffff81184b55>] ? iput+0xb0/0x266 [192066.391200] [<ffffffffa039ea25>] btrfs_unlink_inode+0x1c/0x3d [btrfs] [192066.391930] [<ffffffffa03ca623>] check_item_in_log+0x1fe/0x29b [btrfs] [192066.392715] [<ffffffffa03ca827>] replay_dir_deletes+0x167/0x1cf [btrfs] [192066.393510] [<ffffffffa03cccc7>] replay_one_buffer+0x417/0x570 [btrfs] [192066.394241] [<ffffffffa03ca164>] walk_up_log_tree+0x10e/0x1dc [btrfs] [192066.394958] [<ffffffffa03cac72>] walk_log_tree+0xa5/0x190 [btrfs] [192066.395628] [<ffffffffa03ce8b8>] btrfs_recover_log_trees+0x239/0x32c [btrfs] [192066.396790] [<ffffffffa03cc8b0>] ? replay_one_extent+0x50a/0x50a [btrfs] [192066.397891] [<ffffffffa0394041>] open_ctree+0x1d8b/0x2167 [btrfs] [192066.398897] [<ffffffffa03706e1>] btrfs_mount+0x5ef/0x729 [btrfs] [192066.399823] [<ffffffff8108ad98>] ? trace_hardirqs_on+0xd/0xf [192066.400739] [<ffffffff8108959b>] ? lockdep_init_map+0xb9/0x1b3 [192066.401700] [<ffffffff811714b9>] mount_fs+0x67/0x131 [192066.402482] [<ffffffff81188560>] vfs_kern_mount+0x6c/0xde [192066.403930] [<ffffffffa03702bd>] btrfs_mount+0x1cb/0x729 [btrfs] [192066.404831] [<ffffffff8108ad98>] ? trace_hardirqs_on+0xd/0xf [192066.405726] [<ffffffff8108959b>] ? lockdep_init_map+0xb9/0x1b3 [192066.406621] [<ffffffff811714b9>] mount_fs+0x67/0x131 [192066.407401] [<ffffffff81188560>] vfs_kern_mount+0x6c/0xde [192066.408247] [<ffffffff8118ae36>] do_mount+0x893/0x9d2 [192066.409047] [<ffffffff8113009b>] ? strndup_user+0x3f/0x8c [192066.409842] [<ffffffff8118b187>] SyS_mount+0x75/0xa1 [192066.410621] [<ffffffff8147e517>] entry_SYSCALL_64_fastpath+0x12/0x6b [192066.411572] ---[ end trace 2de42126c1e0a0f0 ]--- [192066.412344] BTRFS: error (device dm-0) in __btrfs_unlink_inode:3986: errno=-2 No such entry [192066.413748] BTRFS: error (device dm-0) in btrfs_replay_log:2464: errno=-2 No such entry (Failed to recover log tree) [192066.415458] BTRFS error (device dm-0): cleaner transaction attach returned -30 [192066.444613] BTRFS: open_ctree failed This happens because when we are replaying the log and processing the directory entry pointing to the snapshot in the subvolume tree, we treat its btrfs_dir_item item as having a location with a key type matching BTRFS_INODE_ITEM_KEY, which is wrong because the type matches BTRFS_ROOT_ITEM_KEY and therefore must be processed differently, as the object id refers to a root number and not to an inode in the root containing the parent directory. So fix this by triggering a transaction commit if an fsync against the parent directory is requested after deleting a snapshot. This is the simplest approach for a rare use case. Some alternative that avoids the transaction commit would require more code to explicitly delete the snapshot at log replay time (factoring out common code from ioctl.c: btrfs_ioctl_snap_destroy()), special care at fsync time to remove the log tree of the snapshot's root from the log root of the root of tree roots, amongst other steps. A test case for xfstests that triggers the issue follows. seq=`basename $0` seqres=$RESULT_DIR/$seq echo "QA output created by $seq" tmp=/tmp/$$ status=1 # failure is the default! trap "_cleanup; exit \$status" 0 1 2 3 15 _cleanup() { _cleanup_flakey cd / rm -f $tmp.* } # get standard environment, filters and checks . ./common/rc . ./common/filter . ./common/dmflakey # real QA test starts here _need_to_be_root _supported_fs btrfs _supported_os Linux _require_scratch _require_dm_target flakey _require_metadata_journaling $SCRATCH_DEV rm -f $seqres.full _scratch_mkfs >>$seqres.full 2>&1 _init_flakey _mount_flakey # Create a snapshot at the root of our filesystem (mount point path), delete it, # fsync the mount point path, crash and mount to replay the log. This should # succeed and after the filesystem is mounted the snapshot should not be visible # anymore. _run_btrfs_util_prog subvolume snapshot $SCRATCH_MNT $SCRATCH_MNT/snap1 _run_btrfs_util_prog subvolume delete $SCRATCH_MNT/snap1 $XFS_IO_PROG -c "fsync" $SCRATCH_MNT _flakey_drop_and_remount [ -e $SCRATCH_MNT/snap1 ] && \ echo "Snapshot snap1 still exists after log replay" # Similar scenario as above, but this time the snapshot is created inside a # directory and not directly under the root (mount point path). mkdir $SCRATCH_MNT/testdir _run_btrfs_util_prog subvolume snapshot $SCRATCH_MNT $SCRATCH_MNT/testdir/snap2 _run_btrfs_util_prog subvolume delete $SCRATCH_MNT/testdir/snap2 $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/testdir _flakey_drop_and_remount [ -e $SCRATCH_MNT/testdir/snap2 ] && \ echo "Snapshot snap2 still exists after log replay" _unmount_flakey echo "Silence is golden" status=0 exit Signed-off-by: Filipe Manana <fdmanana@suse.com> Tested-by: Liu Bo <bo.li.liu@oracle.com> Reviewed-by: Liu Bo <bo.li.liu@oracle.com> Signed-off-by: Chris Mason <clm@fb.com>
2016-02-10 13:42:25 +03:00
*/
void btrfs_record_snapshot_destroy(struct btrfs_trans_handle *trans,
struct btrfs_inode *dir)
Btrfs: fix unreplayable log after snapshot delete + parent dir fsync If we delete a snapshot, fsync its parent directory and crash/power fail before the next transaction commit, on the next mount when we attempt to replay the log tree of the root containing the parent directory we will fail and prevent the filesystem from mounting, which is solvable by wiping out the log trees with the btrfs-zero-log tool but very inconvenient as we will lose any data and metadata fsynced before the parent directory was fsynced. For example: $ mkfs.btrfs -f /dev/sdc $ mount /dev/sdc /mnt $ mkdir /mnt/testdir $ btrfs subvolume snapshot /mnt /mnt/testdir/snap $ btrfs subvolume delete /mnt/testdir/snap $ xfs_io -c "fsync" /mnt/testdir < crash / power failure and reboot > $ mount /dev/sdc /mnt mount: mount(2) failed: No such file or directory And in dmesg/syslog we get the following message and trace: [192066.361162] BTRFS info (device dm-0): failed to delete reference to snap, inode 257 parent 257 [192066.363010] ------------[ cut here ]------------ [192066.365268] WARNING: CPU: 4 PID: 5130 at fs/btrfs/inode.c:3986 __btrfs_unlink_inode+0x17a/0x354 [btrfs]() [192066.367250] BTRFS: Transaction aborted (error -2) [192066.368401] Modules linked in: btrfs dm_flakey dm_mod ppdev sha256_generic xor raid6_pq hmac drbg ansi_cprng aesni_intel acpi_cpufreq tpm_tis aes_x86_64 tpm ablk_helper evdev cryptd sg parport_pc i2c_piix4 psmouse lrw parport i2c_core pcspkr gf128mul processor serio_raw glue_helper button loop autofs4 ext4 crc16 mbcache jbd2 sd_mod sr_mod cdrom ata_generic virtio_scsi ata_piix libata virtio_pci virtio_ring crc32c_intel scsi_mod e1000 virtio floppy [last unloaded: btrfs] [192066.377154] CPU: 4 PID: 5130 Comm: mount Tainted: G W 4.4.0-rc6-btrfs-next-20+ #1 [192066.378875] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS by qemu-project.org 04/01/2014 [192066.380889] 0000000000000000 ffff880143923670 ffffffff81257570 ffff8801439236b8 [192066.382561] ffff8801439236a8 ffffffff8104ec07 ffffffffa039dc2c 00000000fffffffe [192066.384191] ffff8801ed31d000 ffff8801b9fc9c88 ffff8801086875e0 ffff880143923710 [192066.385827] Call Trace: [192066.386373] [<ffffffff81257570>] dump_stack+0x4e/0x79 [192066.387387] [<ffffffff8104ec07>] warn_slowpath_common+0x99/0xb2 [192066.388429] [<ffffffffa039dc2c>] ? __btrfs_unlink_inode+0x17a/0x354 [btrfs] [192066.389236] [<ffffffff8104ec68>] warn_slowpath_fmt+0x48/0x50 [192066.389884] [<ffffffffa039dc2c>] __btrfs_unlink_inode+0x17a/0x354 [btrfs] [192066.390621] [<ffffffff81184b55>] ? iput+0xb0/0x266 [192066.391200] [<ffffffffa039ea25>] btrfs_unlink_inode+0x1c/0x3d [btrfs] [192066.391930] [<ffffffffa03ca623>] check_item_in_log+0x1fe/0x29b [btrfs] [192066.392715] [<ffffffffa03ca827>] replay_dir_deletes+0x167/0x1cf [btrfs] [192066.393510] [<ffffffffa03cccc7>] replay_one_buffer+0x417/0x570 [btrfs] [192066.394241] [<ffffffffa03ca164>] walk_up_log_tree+0x10e/0x1dc [btrfs] [192066.394958] [<ffffffffa03cac72>] walk_log_tree+0xa5/0x190 [btrfs] [192066.395628] [<ffffffffa03ce8b8>] btrfs_recover_log_trees+0x239/0x32c [btrfs] [192066.396790] [<ffffffffa03cc8b0>] ? replay_one_extent+0x50a/0x50a [btrfs] [192066.397891] [<ffffffffa0394041>] open_ctree+0x1d8b/0x2167 [btrfs] [192066.398897] [<ffffffffa03706e1>] btrfs_mount+0x5ef/0x729 [btrfs] [192066.399823] [<ffffffff8108ad98>] ? trace_hardirqs_on+0xd/0xf [192066.400739] [<ffffffff8108959b>] ? lockdep_init_map+0xb9/0x1b3 [192066.401700] [<ffffffff811714b9>] mount_fs+0x67/0x131 [192066.402482] [<ffffffff81188560>] vfs_kern_mount+0x6c/0xde [192066.403930] [<ffffffffa03702bd>] btrfs_mount+0x1cb/0x729 [btrfs] [192066.404831] [<ffffffff8108ad98>] ? trace_hardirqs_on+0xd/0xf [192066.405726] [<ffffffff8108959b>] ? lockdep_init_map+0xb9/0x1b3 [192066.406621] [<ffffffff811714b9>] mount_fs+0x67/0x131 [192066.407401] [<ffffffff81188560>] vfs_kern_mount+0x6c/0xde [192066.408247] [<ffffffff8118ae36>] do_mount+0x893/0x9d2 [192066.409047] [<ffffffff8113009b>] ? strndup_user+0x3f/0x8c [192066.409842] [<ffffffff8118b187>] SyS_mount+0x75/0xa1 [192066.410621] [<ffffffff8147e517>] entry_SYSCALL_64_fastpath+0x12/0x6b [192066.411572] ---[ end trace 2de42126c1e0a0f0 ]--- [192066.412344] BTRFS: error (device dm-0) in __btrfs_unlink_inode:3986: errno=-2 No such entry [192066.413748] BTRFS: error (device dm-0) in btrfs_replay_log:2464: errno=-2 No such entry (Failed to recover log tree) [192066.415458] BTRFS error (device dm-0): cleaner transaction attach returned -30 [192066.444613] BTRFS: open_ctree failed This happens because when we are replaying the log and processing the directory entry pointing to the snapshot in the subvolume tree, we treat its btrfs_dir_item item as having a location with a key type matching BTRFS_INODE_ITEM_KEY, which is wrong because the type matches BTRFS_ROOT_ITEM_KEY and therefore must be processed differently, as the object id refers to a root number and not to an inode in the root containing the parent directory. So fix this by triggering a transaction commit if an fsync against the parent directory is requested after deleting a snapshot. This is the simplest approach for a rare use case. Some alternative that avoids the transaction commit would require more code to explicitly delete the snapshot at log replay time (factoring out common code from ioctl.c: btrfs_ioctl_snap_destroy()), special care at fsync time to remove the log tree of the snapshot's root from the log root of the root of tree roots, amongst other steps. A test case for xfstests that triggers the issue follows. seq=`basename $0` seqres=$RESULT_DIR/$seq echo "QA output created by $seq" tmp=/tmp/$$ status=1 # failure is the default! trap "_cleanup; exit \$status" 0 1 2 3 15 _cleanup() { _cleanup_flakey cd / rm -f $tmp.* } # get standard environment, filters and checks . ./common/rc . ./common/filter . ./common/dmflakey # real QA test starts here _need_to_be_root _supported_fs btrfs _supported_os Linux _require_scratch _require_dm_target flakey _require_metadata_journaling $SCRATCH_DEV rm -f $seqres.full _scratch_mkfs >>$seqres.full 2>&1 _init_flakey _mount_flakey # Create a snapshot at the root of our filesystem (mount point path), delete it, # fsync the mount point path, crash and mount to replay the log. This should # succeed and after the filesystem is mounted the snapshot should not be visible # anymore. _run_btrfs_util_prog subvolume snapshot $SCRATCH_MNT $SCRATCH_MNT/snap1 _run_btrfs_util_prog subvolume delete $SCRATCH_MNT/snap1 $XFS_IO_PROG -c "fsync" $SCRATCH_MNT _flakey_drop_and_remount [ -e $SCRATCH_MNT/snap1 ] && \ echo "Snapshot snap1 still exists after log replay" # Similar scenario as above, but this time the snapshot is created inside a # directory and not directly under the root (mount point path). mkdir $SCRATCH_MNT/testdir _run_btrfs_util_prog subvolume snapshot $SCRATCH_MNT $SCRATCH_MNT/testdir/snap2 _run_btrfs_util_prog subvolume delete $SCRATCH_MNT/testdir/snap2 $XFS_IO_PROG -c "fsync" $SCRATCH_MNT/testdir _flakey_drop_and_remount [ -e $SCRATCH_MNT/testdir/snap2 ] && \ echo "Snapshot snap2 still exists after log replay" _unmount_flakey echo "Silence is golden" status=0 exit Signed-off-by: Filipe Manana <fdmanana@suse.com> Tested-by: Liu Bo <bo.li.liu@oracle.com> Reviewed-by: Liu Bo <bo.li.liu@oracle.com> Signed-off-by: Chris Mason <clm@fb.com>
2016-02-10 13:42:25 +03:00
{
mutex_lock(&dir->log_mutex);
dir->last_unlink_trans = trans->transid;
mutex_unlock(&dir->log_mutex);
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 17:24:20 +03:00
}
/**
* Update the log after adding a new name for an inode.
*
* @trans: Transaction handle.
* @old_dentry: The dentry associated with the old name and the old
* parent directory.
* @old_dir: The inode of the previous parent directory for the case
* of a rename. For a link operation, it must be NULL.
btrfs: avoid logging all directory changes during renames When doing a rename of a file, if the file or its old parent directory were logged before, we log the new name of the file and then make sure we log the old parent directory, to ensure that after a log replay the old name of the file is deleted and the new name added. The logging of the old parent directory can take some time, because it will scan all leaves modified in the current transaction, check which directory entries were already logged, copy the ones that were not logged before, etc. In this rename context all we need to do is make sure that the old name of the file is deleted on log replay, so instead of triggering a directory log operation, we can just delete the old directory entry from the log if it's there, or in case it isn't there, just log a range item to signal log replay that the old name must be deleted. So change btrfs_log_new_name() to do that. This scenario is actually not uncommon to trigger, and recently on a 5.15 kernel, an openSUSE Tumbleweed user reported package installations and upgrades, with the zypper tool, were often taking a long time to complete, much more than usual. With strace it could be observed that zypper was spending over 99% of its time on rename operations, and then with further analysis we checked that directory logging was happening too frequently and causing high latencies for the rename operations. Taking into account that installation/upgrade of some of these packages needed about a few thousand file renames, the slowdown was very noticeable for the user. The issue was caused indirectly due to an excessive number of inode evictions on a 5.15 kernel, about 100x more compared to a 5.13, 5.14 or a 5.16-rc8 kernel. After an inode eviction we can't tell for sure, in an efficient way, if an inode was previously logged in the current transaction, so we are pessimistic and assume it was, because in case it was we need to update the logged inode. More details on that in one of the patches in the same series (subject "btrfs: avoid inode logging during rename and link when possible"). Either way, in case the parent directory was logged before, we currently do more work then necessary during a rename, and this change minimizes that amount of work. The following script mimics part of what a package installation/upgrade with zypper does, which is basically renaming a lot of files, in some directory under /usr, to a name with a suffix of "-RPMDELETE": $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=10000 mkfs.btrfs -f $DEV mount $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync # Do some change to testdir and fsync it. echo -n > $MNT/testdir/file_$((NUM_FILES + 1)) xfs_io -c "fsync" $MNT/testdir echo "Renaming $NUM_FILES files..." start=$(date +%s%N) for ((i = 1; i <= $NUM_FILES; i++)); do mv $MNT/testdir/file_$i $MNT/testdir/file_$i-RPMDELETE done end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "Renames took $dur milliseconds" umount $MNT Testing this change on box using a non-debug kernel (Debian's default kernel config) gave the following results: NUM_FILES=10000, before this patch: 27399 ms NUM_FILES=10000, after this patch: 9093 ms (-66.8%) NUM_FILES=5000, before this patch: 9241 ms NUM_FILES=5000, after this patch: 4642 ms (-49.8%) NUM_FILES=2000, before this patch: 2550 ms NUM_FILES=2000, after this patch: 1788 ms (-29.9%) NUM_FILES=1000, before this patch: 1088 ms NUM_FILES=1000, after this patch: 905 ms (-16.9%) Link: https://bugzilla.opensuse.org/show_bug.cgi?id=1193549 Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-01-20 14:00:08 +03:00
* @old_dir_index: The index number associated with the old name, meaningful
* only for rename operations (when @old_dir is not NULL).
* Ignored for link operations.
* @parent: The dentry associated with the directory under which the
* new name is located.
*
* Call this after adding a new name for an inode, as a result of a link or
* rename operation, and it will properly update the log to reflect the new name.
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 17:24:20 +03:00
*/
btrfs: do not commit logs and transactions during link and rename operations Since commit d4682ba03ef618 ("Btrfs: sync log after logging new name") we started to commit logs, and fallback to transaction commits when we failed to log the new names or commit the logs, after link and rename operations when the target inodes (or their parents) were previously logged in the current transaction. This was to avoid losing directories despite an explicit fsync on them when they are ancestors of some inode that got a new named logged, due to a link or rename operation. However that adds the cost of starting IO and waiting for it to complete, which can cause higher latencies for applications. Instead of doing that, just make sure that when we log a new name for an inode we don't mark any of its ancestors as logged, so that if any one does an fsync against any of them, without doing any other change on them, the fsync commits the log. This way we only pay the cost of a log commit (or a transaction commit if something goes wrong or a new block group was created) if the application explicitly asks to fsync any of the parent directories. Using dbench, which mixes several filesystems operations including renames, revealed some significant latency gains. The following script that uses dbench was used to test this: #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/btrfs MOUNT_OPTIONS="-o ssd -o space_cache=v2" MKFS_OPTIONS="-m single -d single" THREADS=16 echo "performance" | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -t 300 -D $MNT $THREADS umount $MNT The test was run on bare metal, no virtualization, on a box with 12 cores (Intel i7-8700), 64Gb of RAM and using a NVMe device, with a kernel configuration that is the default of typical distributions (debian in this case), without debug options enabled (kasan, kmemleak, slub debug, debug of page allocations, lock debugging, etc). Results before this patch: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 10750455 0.011 155.088 Close 7896674 0.001 0.243 Rename 455222 2.158 1101.947 Unlink 2171189 0.067 121.638 Deltree 256 2.425 7.816 Mkdir 128 0.002 0.003 Qpathinfo 9744323 0.006 21.370 Qfileinfo 1707092 0.001 0.146 Qfsinfo 1786756 0.001 11.228 Sfileinfo 875612 0.003 21.263 Find 3767281 0.025 9.617 WriteX 5356924 0.011 211.390 ReadX 16852694 0.003 9.442 LockX 35008 0.002 0.119 UnlockX 35008 0.001 0.138 Flush 753458 4.252 1102.249 Throughput 1128.35 MB/sec 16 clients 16 procs max_latency=1102.255 ms Results after this patch: 16 clients, after Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 11471098 0.012 448.281 Close 8426396 0.001 0.925 Rename 485746 0.123 267.183 Unlink 2316477 0.080 63.433 Deltree 288 2.830 11.144 Mkdir 144 0.003 0.010 Qpathinfo 10397420 0.006 10.288 Qfileinfo 1822039 0.001 0.169 Qfsinfo 1906497 0.002 14.039 Sfileinfo 934433 0.004 2.438 Find 4019879 0.026 10.200 WriteX 5718932 0.011 200.985 ReadX 17981671 0.003 10.036 LockX 37352 0.002 0.076 UnlockX 37352 0.001 0.109 Flush 804018 5.015 778.033 Throughput 1201.98 MB/sec 16 clients 16 procs max_latency=778.036 ms (+6.5% throughput, -29.4% max latency, -75.8% rename latency) Test case generic/498 from fstests tests the scenario that the previously mentioned commit fixed. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-08-11 14:43:48 +03:00
void btrfs_log_new_name(struct btrfs_trans_handle *trans,
struct dentry *old_dentry, struct btrfs_inode *old_dir,
btrfs: avoid logging all directory changes during renames When doing a rename of a file, if the file or its old parent directory were logged before, we log the new name of the file and then make sure we log the old parent directory, to ensure that after a log replay the old name of the file is deleted and the new name added. The logging of the old parent directory can take some time, because it will scan all leaves modified in the current transaction, check which directory entries were already logged, copy the ones that were not logged before, etc. In this rename context all we need to do is make sure that the old name of the file is deleted on log replay, so instead of triggering a directory log operation, we can just delete the old directory entry from the log if it's there, or in case it isn't there, just log a range item to signal log replay that the old name must be deleted. So change btrfs_log_new_name() to do that. This scenario is actually not uncommon to trigger, and recently on a 5.15 kernel, an openSUSE Tumbleweed user reported package installations and upgrades, with the zypper tool, were often taking a long time to complete, much more than usual. With strace it could be observed that zypper was spending over 99% of its time on rename operations, and then with further analysis we checked that directory logging was happening too frequently and causing high latencies for the rename operations. Taking into account that installation/upgrade of some of these packages needed about a few thousand file renames, the slowdown was very noticeable for the user. The issue was caused indirectly due to an excessive number of inode evictions on a 5.15 kernel, about 100x more compared to a 5.13, 5.14 or a 5.16-rc8 kernel. After an inode eviction we can't tell for sure, in an efficient way, if an inode was previously logged in the current transaction, so we are pessimistic and assume it was, because in case it was we need to update the logged inode. More details on that in one of the patches in the same series (subject "btrfs: avoid inode logging during rename and link when possible"). Either way, in case the parent directory was logged before, we currently do more work then necessary during a rename, and this change minimizes that amount of work. The following script mimics part of what a package installation/upgrade with zypper does, which is basically renaming a lot of files, in some directory under /usr, to a name with a suffix of "-RPMDELETE": $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=10000 mkfs.btrfs -f $DEV mount $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync # Do some change to testdir and fsync it. echo -n > $MNT/testdir/file_$((NUM_FILES + 1)) xfs_io -c "fsync" $MNT/testdir echo "Renaming $NUM_FILES files..." start=$(date +%s%N) for ((i = 1; i <= $NUM_FILES; i++)); do mv $MNT/testdir/file_$i $MNT/testdir/file_$i-RPMDELETE done end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "Renames took $dur milliseconds" umount $MNT Testing this change on box using a non-debug kernel (Debian's default kernel config) gave the following results: NUM_FILES=10000, before this patch: 27399 ms NUM_FILES=10000, after this patch: 9093 ms (-66.8%) NUM_FILES=5000, before this patch: 9241 ms NUM_FILES=5000, after this patch: 4642 ms (-49.8%) NUM_FILES=2000, before this patch: 2550 ms NUM_FILES=2000, after this patch: 1788 ms (-29.9%) NUM_FILES=1000, before this patch: 1088 ms NUM_FILES=1000, after this patch: 905 ms (-16.9%) Link: https://bugzilla.opensuse.org/show_bug.cgi?id=1193549 Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-01-20 14:00:08 +03:00
u64 old_dir_index, struct dentry *parent)
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 17:24:20 +03:00
{
struct btrfs_inode *inode = BTRFS_I(d_inode(old_dentry));
btrfs: stop doing unnecessary log updates during a rename During a rename, we call __btrfs_unlink_inode(), which will call btrfs_del_inode_ref_in_log() and btrfs_del_dir_entries_in_log(), in order to remove an inode reference and a directory entry from the log. These are necessary when __btrfs_unlink_inode() is called from the unlink path, but not necessary when it's called from a rename context, because: 1) For the btrfs_del_inode_ref_in_log() call, it's pointless to delete the inode reference related to the old name, because later in the rename path we call btrfs_log_new_name(), which will drop all inode references from the log and copy all inode references from the subvolume tree to the log tree. So we are doing one unnecessary btree operation which adds additional latency and lock contention in case there are other tasks accessing the log tree; 2) For the btrfs_del_dir_entries_in_log() call, we are now doing the equivalent at btrfs_log_new_name() since the previous patch in the series, that has the subject "btrfs: avoid logging all directory changes during renames". In fact, having __btrfs_unlink_inode() call this function not only adds additional latency and lock contention due to the extra btree operation, but also can make btrfs_log_new_name() unnecessarily log a range item to track the deletion of the old name, since it has no way to known that the directory entry related to the old name was previously logged and already deleted by __btrfs_unlink_inode() through its call to btrfs_del_dir_entries_in_log(). So skip those calls at __btrfs_unlink_inode() when we are doing a rename. Skipping them also allows us now to reduce the duration of time we are pinning a log transaction during renames, which is always beneficial as it's not delaying so much other tasks trying to sync the log tree, in particular we end up not holding the log transaction pinned while adding the new name (adding inode ref, directory entry, etc). This change is part of a patchset comprised of the following patches: 1/5 btrfs: add helper to delete a dir entry from a log tree 2/5 btrfs: pass the dentry to btrfs_log_new_name() instead of the inode 3/5 btrfs: avoid logging all directory changes during renames 4/5 btrfs: stop doing unnecessary log updates during a rename 5/5 btrfs: avoid inode logging during rename and link when possible Just like the previous patch in the series, "btrfs: avoid logging all directory changes during renames", the following script mimics part of what a package installation/upgrade with zypper does, which is basically renaming a lot of files, in some directory under /usr, to a name with a suffix of "-RPMDELETE": $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=10000 mkfs.btrfs -f $DEV mount $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync # Do some change to testdir and fsync it. echo -n > $MNT/testdir/file_$((NUM_FILES + 1)) xfs_io -c "fsync" $MNT/testdir echo "Renaming $NUM_FILES files..." start=$(date +%s%N) for ((i = 1; i <= $NUM_FILES; i++)); do mv $MNT/testdir/file_$i $MNT/testdir/file_$i-RPMDELETE done end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "Renames took $dur milliseconds" umount $MNT Testing this change on box a using a non-debug kernel (Debian's default kernel config) gave the following results: NUM_FILES=10000, before patchset: 27399 ms NUM_FILES=10000, after patches 1/5 to 3/5 applied: 9093 ms (-66.8%) NUM_FILES=10000, after patches 1/5 to 4/5 applied: 9016 ms (-67.1%) NUM_FILES=5000, before patchset: 9241 ms NUM_FILES=5000, after patches 1/5 to 3/5 applied: 4642 ms (-49.8%) NUM_FILES=5000, after patches 1/5 to 4/5 applied: 4553 ms (-50.7%) NUM_FILES=2000, before patchset: 2550 ms NUM_FILES=2000, after patches 1/5 to 3/5 applied: 1788 ms (-29.9%) NUM_FILES=2000, after patches 1/5 to 4/5 applied: 1767 ms (-30.7%) NUM_FILES=1000, before patchset: 1088 ms NUM_FILES=1000, after patches 1/5 to 3/5 applied: 905 ms (-16.9%) NUM_FILES=1000, after patches 1/5 to 4/5 applied: 883 ms (-18.8%) The next patch in the series (5/5), also contains dbench results after applying to whole patchset. Link: https://bugzilla.opensuse.org/show_bug.cgi?id=1193549 Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-01-20 14:00:09 +03:00
struct btrfs_root *root = inode->root;
btrfs: do not commit logs and transactions during link and rename operations Since commit d4682ba03ef618 ("Btrfs: sync log after logging new name") we started to commit logs, and fallback to transaction commits when we failed to log the new names or commit the logs, after link and rename operations when the target inodes (or their parents) were previously logged in the current transaction. This was to avoid losing directories despite an explicit fsync on them when they are ancestors of some inode that got a new named logged, due to a link or rename operation. However that adds the cost of starting IO and waiting for it to complete, which can cause higher latencies for applications. Instead of doing that, just make sure that when we log a new name for an inode we don't mark any of its ancestors as logged, so that if any one does an fsync against any of them, without doing any other change on them, the fsync commits the log. This way we only pay the cost of a log commit (or a transaction commit if something goes wrong or a new block group was created) if the application explicitly asks to fsync any of the parent directories. Using dbench, which mixes several filesystems operations including renames, revealed some significant latency gains. The following script that uses dbench was used to test this: #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/btrfs MOUNT_OPTIONS="-o ssd -o space_cache=v2" MKFS_OPTIONS="-m single -d single" THREADS=16 echo "performance" | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -t 300 -D $MNT $THREADS umount $MNT The test was run on bare metal, no virtualization, on a box with 12 cores (Intel i7-8700), 64Gb of RAM and using a NVMe device, with a kernel configuration that is the default of typical distributions (debian in this case), without debug options enabled (kasan, kmemleak, slub debug, debug of page allocations, lock debugging, etc). Results before this patch: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 10750455 0.011 155.088 Close 7896674 0.001 0.243 Rename 455222 2.158 1101.947 Unlink 2171189 0.067 121.638 Deltree 256 2.425 7.816 Mkdir 128 0.002 0.003 Qpathinfo 9744323 0.006 21.370 Qfileinfo 1707092 0.001 0.146 Qfsinfo 1786756 0.001 11.228 Sfileinfo 875612 0.003 21.263 Find 3767281 0.025 9.617 WriteX 5356924 0.011 211.390 ReadX 16852694 0.003 9.442 LockX 35008 0.002 0.119 UnlockX 35008 0.001 0.138 Flush 753458 4.252 1102.249 Throughput 1128.35 MB/sec 16 clients 16 procs max_latency=1102.255 ms Results after this patch: 16 clients, after Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 11471098 0.012 448.281 Close 8426396 0.001 0.925 Rename 485746 0.123 267.183 Unlink 2316477 0.080 63.433 Deltree 288 2.830 11.144 Mkdir 144 0.003 0.010 Qpathinfo 10397420 0.006 10.288 Qfileinfo 1822039 0.001 0.169 Qfsinfo 1906497 0.002 14.039 Sfileinfo 934433 0.004 2.438 Find 4019879 0.026 10.200 WriteX 5718932 0.011 200.985 ReadX 17981671 0.003 10.036 LockX 37352 0.002 0.076 UnlockX 37352 0.001 0.109 Flush 804018 5.015 778.033 Throughput 1201.98 MB/sec 16 clients 16 procs max_latency=778.036 ms (+6.5% throughput, -29.4% max latency, -75.8% rename latency) Test case generic/498 from fstests tests the scenario that the previously mentioned commit fixed. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-08-11 14:43:48 +03:00
struct btrfs_log_ctx ctx;
btrfs: stop doing unnecessary log updates during a rename During a rename, we call __btrfs_unlink_inode(), which will call btrfs_del_inode_ref_in_log() and btrfs_del_dir_entries_in_log(), in order to remove an inode reference and a directory entry from the log. These are necessary when __btrfs_unlink_inode() is called from the unlink path, but not necessary when it's called from a rename context, because: 1) For the btrfs_del_inode_ref_in_log() call, it's pointless to delete the inode reference related to the old name, because later in the rename path we call btrfs_log_new_name(), which will drop all inode references from the log and copy all inode references from the subvolume tree to the log tree. So we are doing one unnecessary btree operation which adds additional latency and lock contention in case there are other tasks accessing the log tree; 2) For the btrfs_del_dir_entries_in_log() call, we are now doing the equivalent at btrfs_log_new_name() since the previous patch in the series, that has the subject "btrfs: avoid logging all directory changes during renames". In fact, having __btrfs_unlink_inode() call this function not only adds additional latency and lock contention due to the extra btree operation, but also can make btrfs_log_new_name() unnecessarily log a range item to track the deletion of the old name, since it has no way to known that the directory entry related to the old name was previously logged and already deleted by __btrfs_unlink_inode() through its call to btrfs_del_dir_entries_in_log(). So skip those calls at __btrfs_unlink_inode() when we are doing a rename. Skipping them also allows us now to reduce the duration of time we are pinning a log transaction during renames, which is always beneficial as it's not delaying so much other tasks trying to sync the log tree, in particular we end up not holding the log transaction pinned while adding the new name (adding inode ref, directory entry, etc). This change is part of a patchset comprised of the following patches: 1/5 btrfs: add helper to delete a dir entry from a log tree 2/5 btrfs: pass the dentry to btrfs_log_new_name() instead of the inode 3/5 btrfs: avoid logging all directory changes during renames 4/5 btrfs: stop doing unnecessary log updates during a rename 5/5 btrfs: avoid inode logging during rename and link when possible Just like the previous patch in the series, "btrfs: avoid logging all directory changes during renames", the following script mimics part of what a package installation/upgrade with zypper does, which is basically renaming a lot of files, in some directory under /usr, to a name with a suffix of "-RPMDELETE": $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=10000 mkfs.btrfs -f $DEV mount $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync # Do some change to testdir and fsync it. echo -n > $MNT/testdir/file_$((NUM_FILES + 1)) xfs_io -c "fsync" $MNT/testdir echo "Renaming $NUM_FILES files..." start=$(date +%s%N) for ((i = 1; i <= $NUM_FILES; i++)); do mv $MNT/testdir/file_$i $MNT/testdir/file_$i-RPMDELETE done end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "Renames took $dur milliseconds" umount $MNT Testing this change on box a using a non-debug kernel (Debian's default kernel config) gave the following results: NUM_FILES=10000, before patchset: 27399 ms NUM_FILES=10000, after patches 1/5 to 3/5 applied: 9093 ms (-66.8%) NUM_FILES=10000, after patches 1/5 to 4/5 applied: 9016 ms (-67.1%) NUM_FILES=5000, before patchset: 9241 ms NUM_FILES=5000, after patches 1/5 to 3/5 applied: 4642 ms (-49.8%) NUM_FILES=5000, after patches 1/5 to 4/5 applied: 4553 ms (-50.7%) NUM_FILES=2000, before patchset: 2550 ms NUM_FILES=2000, after patches 1/5 to 3/5 applied: 1788 ms (-29.9%) NUM_FILES=2000, after patches 1/5 to 4/5 applied: 1767 ms (-30.7%) NUM_FILES=1000, before patchset: 1088 ms NUM_FILES=1000, after patches 1/5 to 3/5 applied: 905 ms (-16.9%) NUM_FILES=1000, after patches 1/5 to 4/5 applied: 883 ms (-18.8%) The next patch in the series (5/5), also contains dbench results after applying to whole patchset. Link: https://bugzilla.opensuse.org/show_bug.cgi?id=1193549 Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-01-20 14:00:09 +03:00
bool log_pinned = false;
btrfs: avoid inode logging during rename and link when possible During a rename or link operation, we need to determine if an inode was previously logged or not, and if it was, do some update to the logged inode. We used to rely exclusively on the logged_trans field of struct btrfs_inode to determine that, but that was not reliable because the value of that field is not persisted in the inode item, so it's lost when an inode is evicted and loaded back again. That led to several issues in the past, such as not persisting deletions (such as the case fixed by commit 803f0f64d17769 ("Btrfs: fix fsync not persisting dentry deletions due to inode evictions")), or resulting in losing a file after an inode eviction followed by a rename (commit ecc64fab7d49c6 ("btrfs: fix lost inode on log replay after mix of fsync, rename and inode eviction")), besides other issues. So the inode_logged() helper was introduced and used to determine if an inode was possibly logged before in the current transaction, with the caveat that it could return false positives, in the sense that even if an inode was not logged before in the current transaction, it could still return true, but never to return false in case the inode was logged. >From a functional point of view that is fine, but from a performance perspective it can introduce significant latencies to rename and link operations, as they will end up doing inode logging even when it is not necessary. Recently on a 5.15 kernel, an openSUSE Tumbleweed user reported package installations and upgrades, with the zypper tool, were often taking a long time to complete. With strace it could be observed that zypper was spending about 99% of its time on rename operations, and then with further analysis we checked that directory logging was happening too frequently. Taking into account that installation/upgrade of some of the packages needed a few thousand file renames, the slowdown was very noticeable for the user. The issue was caused indirectly due to an excessive number of inode evictions on a 5.15 kernel, about 100x more compared to a 5.13, 5.14 or a 5.16-rc8 kernel. While triggering the inode evictions if something outside btrfs' control, btrfs could still behave better by eliminating the false positives from the inode_logged() helper. So change inode_logged() to actually eliminate such false positives caused by inode eviction and when an inode was never logged since the filesystem was mounted, as both cases relate to when the logged_trans field of struct btrfs_inode has a value of zero. When it can not determine if the inode was logged based only on the logged_trans value, lookup for the existence of the inode item in the log tree - if it's there then we known the inode was logged, if it's not there then it can not have been logged in the current transaction. Once we determine if the inode was logged, update the logged_trans value to avoid future calls to have to search in the log tree again. Alternatively, we could start storing logged_trans in the on disk inode item structure (struct btrfs_inode_item) in the unused space it still has, but that would be a bit odd because: 1) We only care about logged_trans since the filesystem was mounted, we don't care about its value from a previous mount. Having it persisted in the inode item structure would not make the best use of the precious unused space; 2) In order to get logged_trans persisted before inode eviction, we would have to update the delayed inode when we finish logging the inode and update its logged_trans in struct btrfs_inode, which makes it a bit cumbersome since we need to check if the delayed inode exists, if not create it and populate it and deal with any errors (-ENOMEM mostly). This change is part of a patchset comprised of the following patches: 1/5 btrfs: add helper to delete a dir entry from a log tree 2/5 btrfs: pass the dentry to btrfs_log_new_name() instead of the inode 3/5 btrfs: avoid logging all directory changes during renames 4/5 btrfs: stop doing unnecessary log updates during a rename 5/5 btrfs: avoid inode logging during rename and link when possible The following test script mimics part of what the zypper tool does during package installations/upgrades. It does not triggers inode evictions, but it's similar because it triggers false positives from the inode_logged() helper, because the inodes have a logged_trans of 0, there's a log tree due to a fsync of an unrelated file and the directory inode has its last_trans field set to the current transaction: $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=10000 mkfs.btrfs -f $DEV mount $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync # Now do some change to an unrelated file and fsync it. # This is just to create a log tree to make sure that inode_logged() # does not return false when called against "testdir". xfs_io -f -c "pwrite 0 4K" -c "fsync" $MNT/foo # Do some change to testdir. This is to make sure inode_logged() # will return true when called against "testdir", because its # logged_trans is 0, it was changed in the current transaction # and there's a log tree. echo -n > $MNT/testdir/file_$((NUM_FILES + 1)) echo "Renaming $NUM_FILES files..." start=$(date +%s%N) for ((i = 1; i <= $NUM_FILES; i++)); do mv $MNT/testdir/file_$i $MNT/testdir/file_$i-RPMDELETE done end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "Renames took $dur milliseconds" umount $MNT Testing this change on a box using a non-debug kernel (Debian's default kernel config) gave the following results: NUM_FILES=10000, before patchset: 27837 ms NUM_FILES=10000, after patches 1/5 to 4/5 applied: 9236 ms (-66.8%) NUM_FILES=10000, after whole patchset applied: 8902 ms (-68.0%) NUM_FILES=5000, before patchset: 9127 ms NUM_FILES=5000, after patches 1/5 to 4/5 applied: 4640 ms (-49.2%) NUM_FILES=5000, after whole patchset applied: 4441 ms (-51.3%) NUM_FILES=2000, before patchset: 2528 ms NUM_FILES=2000, after patches 1/5 to 4/5 applied: 1983 ms (-21.6%) NUM_FILES=2000, after whole patchset applied: 1747 ms (-30.9%) NUM_FILES=1000, before patchset: 1085 ms NUM_FILES=1000, after patches 1/5 to 4/5 applied: 893 ms (-17.7%) NUM_FILES=1000, after whole patchset applied: 867 ms (-20.1%) Running dbench on the same physical machine with the following script: $ cat run-dbench.sh #!/bin/bash NUM_JOBS=$(nproc --all) DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 MOUNT_OPTIONS="-o ssd" MKFS_OPTIONS="-O no-holes -R free-space-tree" echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -D $MNT -t 120 $NUM_JOBS umount $MNT Before patchset: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 3761352 0.032 143.843 Close 2762770 0.002 2.273 Rename 159304 0.291 67.037 Unlink 759784 0.207 143.998 Deltree 72 4.028 15.977 Mkdir 36 0.003 0.006 Qpathinfo 3409780 0.013 9.678 Qfileinfo 596772 0.001 0.878 Qfsinfo 625189 0.003 1.245 Sfileinfo 306443 0.006 1.840 Find 1318106 0.063 19.798 WriteX 1871137 0.021 8.532 ReadX 5897325 0.003 3.567 LockX 12252 0.003 0.258 UnlockX 12252 0.002 0.100 Flush 263666 3.327 155.632 Throughput 980.047 MB/sec 12 clients 12 procs max_latency=155.636 ms After whole patchset applied: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 4195584 0.033 107.742 Close 3081932 0.002 1.935 Rename 177641 0.218 14.905 Unlink 847333 0.166 107.822 Deltree 118 5.315 15.247 Mkdir 59 0.004 0.048 Qpathinfo 3802612 0.014 10.302 Qfileinfo 666748 0.001 1.034 Qfsinfo 697329 0.003 0.944 Sfileinfo 341712 0.006 2.099 Find 1470365 0.065 9.359 WriteX 2093921 0.021 8.087 ReadX 6576234 0.003 3.407 LockX 13660 0.003 0.308 UnlockX 13660 0.002 0.114 Flush 294090 2.906 115.539 Throughput 1093.11 MB/sec 12 clients 12 procs max_latency=115.544 ms +11.5% throughput -25.8% max latency rename max latency -77.8% Link: https://bugzilla.opensuse.org/show_bug.cgi?id=1193549 Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-01-20 14:00:10 +03:00
int ret;
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 17:24:20 +03:00
/*
* this will force the logging code to walk the dentry chain
* up for the file
*/
Btrfs: fix log replay failure after linking special file and fsync If in the same transaction we rename a special file (fifo, character/block device or symbolic link), create a hard link for it having its old name then sync the log, we will end up with a log that can not be replayed and at when attempting to replay it, an EEXIST error is returned and mounting the filesystem fails. Example scenario: $ mkfs.btrfs -f /dev/sdc $ mount /dev/sdc /mnt $ mkdir /mnt/testdir $ mkfifo /mnt/testdir/foo # Make sure everything done so far is durably persisted. $ sync # Create some unrelated file and fsync it, this is just to create a log # tree. The file must be in the same directory as our special file. $ touch /mnt/testdir/f1 $ xfs_io -c "fsync" /mnt/testdir/f1 # Rename our special file and then create a hard link with its old name. $ mv /mnt/testdir/foo /mnt/testdir/bar $ ln /mnt/testdir/bar /mnt/testdir/foo # Create some other unrelated file and fsync it, this is just to persist # the log tree which was modified by the previous rename and link # operations. Alternatively we could have modified file f1 and fsync it. $ touch /mnt/f2 $ xfs_io -c "fsync" /mnt/f2 <power failure> $ mount /dev/sdc /mnt mount: mount /dev/sdc on /mnt failed: File exists This happens because when both the log tree and the subvolume's tree have an entry in the directory "testdir" with the same name, that is, there is one key (258 INODE_REF 257) in the subvolume tree and another one in the log tree (where 258 is the inode number of our special file and 257 is the inode for directory "testdir"). Only the data of those two keys differs, in the subvolume tree the index field for inode reference has a value of 3 while the log tree it has a value of 5. Because the same key exists in both trees, but have different index, the log replay fails with an -EEXIST error when attempting to replay the inode reference from the log tree. Fix this by setting the last_unlink_trans field of the inode (our special file) to the current transaction id when a hard link is created, as this forces logging the parent directory inode, solving the conflict at log replay time. A new generic test case for fstests was also submitted. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2018-02-28 18:55:40 +03:00
if (!S_ISDIR(inode->vfs_inode.i_mode))
inode->last_unlink_trans = trans->transid;
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 17:24:20 +03:00
/*
* if this inode hasn't been logged and directory we're renaming it
* from hasn't been logged, we don't need to log it
*/
btrfs: avoid inode logging during rename and link when possible During a rename or link operation, we need to determine if an inode was previously logged or not, and if it was, do some update to the logged inode. We used to rely exclusively on the logged_trans field of struct btrfs_inode to determine that, but that was not reliable because the value of that field is not persisted in the inode item, so it's lost when an inode is evicted and loaded back again. That led to several issues in the past, such as not persisting deletions (such as the case fixed by commit 803f0f64d17769 ("Btrfs: fix fsync not persisting dentry deletions due to inode evictions")), or resulting in losing a file after an inode eviction followed by a rename (commit ecc64fab7d49c6 ("btrfs: fix lost inode on log replay after mix of fsync, rename and inode eviction")), besides other issues. So the inode_logged() helper was introduced and used to determine if an inode was possibly logged before in the current transaction, with the caveat that it could return false positives, in the sense that even if an inode was not logged before in the current transaction, it could still return true, but never to return false in case the inode was logged. >From a functional point of view that is fine, but from a performance perspective it can introduce significant latencies to rename and link operations, as they will end up doing inode logging even when it is not necessary. Recently on a 5.15 kernel, an openSUSE Tumbleweed user reported package installations and upgrades, with the zypper tool, were often taking a long time to complete. With strace it could be observed that zypper was spending about 99% of its time on rename operations, and then with further analysis we checked that directory logging was happening too frequently. Taking into account that installation/upgrade of some of the packages needed a few thousand file renames, the slowdown was very noticeable for the user. The issue was caused indirectly due to an excessive number of inode evictions on a 5.15 kernel, about 100x more compared to a 5.13, 5.14 or a 5.16-rc8 kernel. While triggering the inode evictions if something outside btrfs' control, btrfs could still behave better by eliminating the false positives from the inode_logged() helper. So change inode_logged() to actually eliminate such false positives caused by inode eviction and when an inode was never logged since the filesystem was mounted, as both cases relate to when the logged_trans field of struct btrfs_inode has a value of zero. When it can not determine if the inode was logged based only on the logged_trans value, lookup for the existence of the inode item in the log tree - if it's there then we known the inode was logged, if it's not there then it can not have been logged in the current transaction. Once we determine if the inode was logged, update the logged_trans value to avoid future calls to have to search in the log tree again. Alternatively, we could start storing logged_trans in the on disk inode item structure (struct btrfs_inode_item) in the unused space it still has, but that would be a bit odd because: 1) We only care about logged_trans since the filesystem was mounted, we don't care about its value from a previous mount. Having it persisted in the inode item structure would not make the best use of the precious unused space; 2) In order to get logged_trans persisted before inode eviction, we would have to update the delayed inode when we finish logging the inode and update its logged_trans in struct btrfs_inode, which makes it a bit cumbersome since we need to check if the delayed inode exists, if not create it and populate it and deal with any errors (-ENOMEM mostly). This change is part of a patchset comprised of the following patches: 1/5 btrfs: add helper to delete a dir entry from a log tree 2/5 btrfs: pass the dentry to btrfs_log_new_name() instead of the inode 3/5 btrfs: avoid logging all directory changes during renames 4/5 btrfs: stop doing unnecessary log updates during a rename 5/5 btrfs: avoid inode logging during rename and link when possible The following test script mimics part of what the zypper tool does during package installations/upgrades. It does not triggers inode evictions, but it's similar because it triggers false positives from the inode_logged() helper, because the inodes have a logged_trans of 0, there's a log tree due to a fsync of an unrelated file and the directory inode has its last_trans field set to the current transaction: $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=10000 mkfs.btrfs -f $DEV mount $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync # Now do some change to an unrelated file and fsync it. # This is just to create a log tree to make sure that inode_logged() # does not return false when called against "testdir". xfs_io -f -c "pwrite 0 4K" -c "fsync" $MNT/foo # Do some change to testdir. This is to make sure inode_logged() # will return true when called against "testdir", because its # logged_trans is 0, it was changed in the current transaction # and there's a log tree. echo -n > $MNT/testdir/file_$((NUM_FILES + 1)) echo "Renaming $NUM_FILES files..." start=$(date +%s%N) for ((i = 1; i <= $NUM_FILES; i++)); do mv $MNT/testdir/file_$i $MNT/testdir/file_$i-RPMDELETE done end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "Renames took $dur milliseconds" umount $MNT Testing this change on a box using a non-debug kernel (Debian's default kernel config) gave the following results: NUM_FILES=10000, before patchset: 27837 ms NUM_FILES=10000, after patches 1/5 to 4/5 applied: 9236 ms (-66.8%) NUM_FILES=10000, after whole patchset applied: 8902 ms (-68.0%) NUM_FILES=5000, before patchset: 9127 ms NUM_FILES=5000, after patches 1/5 to 4/5 applied: 4640 ms (-49.2%) NUM_FILES=5000, after whole patchset applied: 4441 ms (-51.3%) NUM_FILES=2000, before patchset: 2528 ms NUM_FILES=2000, after patches 1/5 to 4/5 applied: 1983 ms (-21.6%) NUM_FILES=2000, after whole patchset applied: 1747 ms (-30.9%) NUM_FILES=1000, before patchset: 1085 ms NUM_FILES=1000, after patches 1/5 to 4/5 applied: 893 ms (-17.7%) NUM_FILES=1000, after whole patchset applied: 867 ms (-20.1%) Running dbench on the same physical machine with the following script: $ cat run-dbench.sh #!/bin/bash NUM_JOBS=$(nproc --all) DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 MOUNT_OPTIONS="-o ssd" MKFS_OPTIONS="-O no-holes -R free-space-tree" echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -D $MNT -t 120 $NUM_JOBS umount $MNT Before patchset: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 3761352 0.032 143.843 Close 2762770 0.002 2.273 Rename 159304 0.291 67.037 Unlink 759784 0.207 143.998 Deltree 72 4.028 15.977 Mkdir 36 0.003 0.006 Qpathinfo 3409780 0.013 9.678 Qfileinfo 596772 0.001 0.878 Qfsinfo 625189 0.003 1.245 Sfileinfo 306443 0.006 1.840 Find 1318106 0.063 19.798 WriteX 1871137 0.021 8.532 ReadX 5897325 0.003 3.567 LockX 12252 0.003 0.258 UnlockX 12252 0.002 0.100 Flush 263666 3.327 155.632 Throughput 980.047 MB/sec 12 clients 12 procs max_latency=155.636 ms After whole patchset applied: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 4195584 0.033 107.742 Close 3081932 0.002 1.935 Rename 177641 0.218 14.905 Unlink 847333 0.166 107.822 Deltree 118 5.315 15.247 Mkdir 59 0.004 0.048 Qpathinfo 3802612 0.014 10.302 Qfileinfo 666748 0.001 1.034 Qfsinfo 697329 0.003 0.944 Sfileinfo 341712 0.006 2.099 Find 1470365 0.065 9.359 WriteX 2093921 0.021 8.087 ReadX 6576234 0.003 3.407 LockX 13660 0.003 0.308 UnlockX 13660 0.002 0.114 Flush 294090 2.906 115.539 Throughput 1093.11 MB/sec 12 clients 12 procs max_latency=115.544 ms +11.5% throughput -25.8% max latency rename max latency -77.8% Link: https://bugzilla.opensuse.org/show_bug.cgi?id=1193549 Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-01-20 14:00:10 +03:00
ret = inode_logged(trans, inode, NULL);
if (ret < 0) {
goto out;
} else if (ret == 0) {
if (!old_dir)
return;
/*
* If the inode was not logged and we are doing a rename (old_dir is not
* NULL), check if old_dir was logged - if it was not we can return and
* do nothing.
*/
ret = inode_logged(trans, old_dir, NULL);
if (ret < 0)
goto out;
else if (ret == 0)
return;
}
ret = 0;
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 17:24:20 +03:00
btrfs: fix removed dentries still existing after log is synced When we move one inode from one directory to another and both the inode and its previous parent directory were logged before, we are not supposed to have the dentry for the old parent if we have a power failure after the log is synced. Only the new dentry is supposed to exist. Generally this works correctly, however there is a scenario where this is not currently working, because the old parent of the file/directory that was moved is not authoritative for a range that includes the dir index and dir item keys of the old dentry. This case is better explained with the following example and reproducer: # The test requires a very specific layout of keys and items in the # fs/subvolume btree to trigger the bug. So we want to make sure that # on whatever platform we are, we have the same leaf/node size. # # Currently in btrfs the node/leaf size can not be smaller than the page # size (but it can be greater than the page size). So use the largest # supported node/leaf size (64K). $ mkfs.btrfs -f -n 65536 /dev/sdc $ mount /dev/sdc /mnt # "testdir" is inode 257. $ mkdir /mnt/testdir $ chmod 755 /mnt/testdir # Create several empty files to have the directory "testdir" with its # items spread over several leaves (7 in this case). $ for ((i = 1; i <= 1200; i++)); do echo -n > /mnt/testdir/file$i done # Create our test directory "dira", inode number 1458, which gets all # its items in leaf 7. # # The BTRFS_DIR_ITEM_KEY item for inode 257 ("testdir") that points to # the entry named "dira" is in leaf 2, while the BTRFS_DIR_INDEX_KEY # item that points to that entry is in leaf 3. # # For this particular filesystem node size (64K), file count and file # names, we endup with the directory entry items from inode 257 in # leaves 2 and 3, as previously mentioned - what matters for triggering # the bug exercised by this test case is that those items are not placed # in leaf 1, they must be placed in a leaf different from the one # containing the inode item for inode 257. # # The corresponding BTRFS_DIR_ITEM_KEY and BTRFS_DIR_INDEX_KEY items for # the parent inode (257) are the following: # # item 460 key (257 DIR_ITEM 3724298081) itemoff 48344 itemsize 34 # location key (1458 INODE_ITEM 0) type DIR # transid 6 data_len 0 name_len 4 # name: dira # # and: # # item 771 key (257 DIR_INDEX 1202) itemoff 36673 itemsize 34 # location key (1458 INODE_ITEM 0) type DIR # transid 6 data_len 0 name_len 4 # name: dira $ mkdir /mnt/testdir/dira # Make sure everything done so far is durably persisted. $ sync # Now do a change to inode 257 ("testdir") that does not result in # COWing leaves 2 and 3 - the leaves that contain the directory items # pointing to inode 1458 (directory "dira"). # # Changing permissions, the owner/group, updating or adding a xattr, # etc, will not change (COW) leaves 2 and 3. So for the sake of # simplicity change the permissions of inode 257, which results in # updating its inode item and therefore change (COW) only leaf 1. $ chmod 700 /mnt/testdir # Now fsync directory inode 257. # # Since only the first leaf was changed/COWed, we log the inode item of # inode 257 and only the dentries found in the first leaf, all have a # key type of BTRFS_DIR_ITEM_KEY, and no keys of type # BTRFS_DIR_INDEX_KEY, because they sort after the former type and none # exist in the first leaf. # # We also log 3 items that represent ranges for dir items and dir # indexes for which the log is authoritative: # # 1) a key of type BTRFS_DIR_LOG_ITEM_KEY, which indicates the log is # authoritative for all BTRFS_DIR_ITEM_KEY keys that have an offset # in the range [0, 2285968570] (the offset here is the crc32c of the # dentry's name). The value 2285968570 corresponds to the offset of # the first key of leaf 2 (which is of type BTRFS_DIR_ITEM_KEY); # # 2) a key of type BTRFS_DIR_LOG_ITEM_KEY, which indicates the log is # authoritative for all BTRFS_DIR_ITEM_KEY keys that have an offset # in the range [4293818216, (u64)-1] (the offset here is the crc32c # of the dentry's name). The value 4293818216 corresponds to the # offset of the highest key of type BTRFS_DIR_ITEM_KEY plus 1 # (4293818215 + 1), which is located in leaf 2; # # 3) a key of type BTRFS_DIR_LOG_INDEX_KEY, with an offset of 1203, # which indicates the log is authoritative for all keys of type # BTRFS_DIR_INDEX_KEY that have an offset in the range # [1203, (u64)-1]. The value 1203 corresponds to the offset of the # last key of type BTRFS_DIR_INDEX_KEY plus 1 (1202 + 1), which is # located in leaf 3; # # Also, because "testdir" is a directory and inode 1458 ("dira") is a # child directory, we log inode 1458 too. $ xfs_io -c "fsync" /mnt/testdir # Now move "dira", inode 1458, to be a child of the root directory # (inode 256). # # Because this inode was previously logged, when "testdir" was fsynced, # the log is updated so that the old inode reference, referring to inode # 257 as the parent, is deleted and the new inode reference, referring # to inode 256 as the parent, is added to the log. $ mv /mnt/testdir/dira /mnt # Now change some file and fsync it. This guarantees the log changes # made by the previous move/rename operation are persisted. We do not # need to do any special modification to the file, just any change to # any file and sync the log. $ xfs_io -c "pwrite -S 0xab 0 64K" -c "fsync" /mnt/testdir/file1 # Simulate a power failure and then mount again the filesystem to # replay the log tree. We want to verify that we are able to mount the # filesystem, meaning log replay was successful, and that directory # inode 1458 ("dira") only has inode 256 (the filesystem's root) as # its parent (and no longer a child of inode 257). # # It used to happen that during log replay we would end up having # inode 1458 (directory "dira") with 2 hard links, being a child of # inode 257 ("testdir") and inode 256 (the filesystem's root). This # resulted in the tree checker detecting the issue and causing the # mount operation to fail (with -EIO). # # This happened because in the log we have the new name/parent for # inode 1458, which results in adding the new dentry with inode 256 # as the parent, but the previous dentry, under inode 257 was never # removed - this is because the ranges for dir items and dir indexes # of inode 257 for which the log is authoritative do not include the # old dir item and dir index for the dentry of inode 257 referring to # inode 1458: # # - for dir items, the log is authoritative for the ranges # [0, 2285968570] and [4293818216, (u64)-1]. The dir item at inode 257 # pointing to inode 1458 has a key of (257 DIR_ITEM 3724298081), as # previously mentioned, so the dir item is not deleted when the log # replay procedure processes the authoritative ranges, as 3724298081 # is outside both ranges; # # - for dir indexes, the log is authoritative for the range # [1203, (u64)-1], and the dir index item of inode 257 pointing to # inode 1458 has a key of (257 DIR_INDEX 1202), as previously # mentioned, so the dir index item is not deleted when the log # replay procedure processes the authoritative range. <power failure> $ mount /dev/sdc /mnt mount: /mnt: can't read superblock on /dev/sdc. $ dmesg (...) [87849.840509] BTRFS info (device sdc): start tree-log replay [87849.875719] BTRFS critical (device sdc): corrupt leaf: root=5 block=30539776 slot=554 ino=1458, invalid nlink: has 2 expect no more than 1 for dir [87849.878084] BTRFS info (device sdc): leaf 30539776 gen 7 total ptrs 557 free space 2092 owner 5 [87849.879516] BTRFS info (device sdc): refs 1 lock_owner 0 current 2099108 [87849.880613] item 0 key (1181 1 0) itemoff 65275 itemsize 160 [87849.881544] inode generation 6 size 0 mode 100644 [87849.882692] item 1 key (1181 12 257) itemoff 65258 itemsize 17 (...) [87850.562549] item 556 key (1458 12 257) itemoff 16017 itemsize 14 [87850.563349] BTRFS error (device dm-0): block=30539776 write time tree block corruption detected [87850.564386] ------------[ cut here ]------------ [87850.564920] WARNING: CPU: 3 PID: 2099108 at fs/btrfs/disk-io.c:465 csum_one_extent_buffer+0xed/0x100 [btrfs] [87850.566129] Modules linked in: btrfs dm_zero dm_snapshot (...) [87850.573789] CPU: 3 PID: 2099108 Comm: mount Not tainted 5.12.0-rc8-btrfs-next-86 #1 (...) [87850.587481] Call Trace: [87850.587768] btree_csum_one_bio+0x244/0x2b0 [btrfs] [87850.588354] ? btrfs_bio_fits_in_stripe+0xd8/0x110 [btrfs] [87850.589003] btrfs_submit_metadata_bio+0xb7/0x100 [btrfs] [87850.589654] submit_one_bio+0x61/0x70 [btrfs] [87850.590248] submit_extent_page+0x91/0x2f0 [btrfs] [87850.590842] write_one_eb+0x175/0x440 [btrfs] [87850.591370] ? find_extent_buffer_nolock+0x1c0/0x1c0 [btrfs] [87850.592036] btree_write_cache_pages+0x1e6/0x610 [btrfs] [87850.592665] ? free_debug_processing+0x1d5/0x240 [87850.593209] do_writepages+0x43/0xf0 [87850.593798] ? __filemap_fdatawrite_range+0xa4/0x100 [87850.594391] __filemap_fdatawrite_range+0xc5/0x100 [87850.595196] btrfs_write_marked_extents+0x68/0x160 [btrfs] [87850.596202] btrfs_write_and_wait_transaction.isra.0+0x4d/0xd0 [btrfs] [87850.597377] btrfs_commit_transaction+0x794/0xca0 [btrfs] [87850.598455] ? _raw_spin_unlock_irqrestore+0x32/0x60 [87850.599305] ? kmem_cache_free+0x15a/0x3d0 [87850.600029] btrfs_recover_log_trees+0x346/0x380 [btrfs] [87850.601021] ? replay_one_extent+0x7d0/0x7d0 [btrfs] [87850.601988] open_ctree+0x13c9/0x1698 [btrfs] [87850.602846] btrfs_mount_root.cold+0x13/0xed [btrfs] [87850.603771] ? kmem_cache_alloc_trace+0x7c9/0x930 [87850.604576] ? vfs_parse_fs_string+0x5d/0xb0 [87850.605293] ? kfree+0x276/0x3f0 [87850.605857] legacy_get_tree+0x30/0x50 [87850.606540] vfs_get_tree+0x28/0xc0 [87850.607163] fc_mount+0xe/0x40 [87850.607695] vfs_kern_mount.part.0+0x71/0x90 [87850.608440] btrfs_mount+0x13b/0x3e0 [btrfs] (...) [87850.629477] ---[ end trace 68802022b99a1ea0 ]--- [87850.630849] BTRFS: error (device sdc) in btrfs_commit_transaction:2381: errno=-5 IO failure (Error while writing out transaction) [87850.632422] BTRFS warning (device sdc): Skipping commit of aborted transaction. [87850.633416] BTRFS: error (device sdc) in cleanup_transaction:1978: errno=-5 IO failure [87850.634553] BTRFS: error (device sdc) in btrfs_replay_log:2431: errno=-5 IO failure (Failed to recover log tree) [87850.637529] BTRFS error (device sdc): open_ctree failed In this example the inode we moved was a directory, so it was easy to detect the problem because directories can only have one hard link and the tree checker immediately detects that. If the moved inode was a file, then the log replay would succeed and we would end up having both the new hard link (/mnt/foo) and the old hard link (/mnt/testdir/foo) present, but only the new one should be present. Fix this by forcing re-logging of the old parent directory when logging the new name during a rename operation. This ensures we end up with a log that is authoritative for a range covering the keys for the old dentry, therefore causing the old dentry do be deleted when replaying the log. A test case for fstests will follow up soon. Fixes: 64d6b281ba4db0 ("btrfs: remove unnecessary check_parent_dirs_for_sync()") CC: stable@vger.kernel.org # 5.12+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-05-12 18:27:16 +03:00
/*
* If we are doing a rename (old_dir is not NULL) from a directory that
btrfs: avoid logging all directory changes during renames When doing a rename of a file, if the file or its old parent directory were logged before, we log the new name of the file and then make sure we log the old parent directory, to ensure that after a log replay the old name of the file is deleted and the new name added. The logging of the old parent directory can take some time, because it will scan all leaves modified in the current transaction, check which directory entries were already logged, copy the ones that were not logged before, etc. In this rename context all we need to do is make sure that the old name of the file is deleted on log replay, so instead of triggering a directory log operation, we can just delete the old directory entry from the log if it's there, or in case it isn't there, just log a range item to signal log replay that the old name must be deleted. So change btrfs_log_new_name() to do that. This scenario is actually not uncommon to trigger, and recently on a 5.15 kernel, an openSUSE Tumbleweed user reported package installations and upgrades, with the zypper tool, were often taking a long time to complete, much more than usual. With strace it could be observed that zypper was spending over 99% of its time on rename operations, and then with further analysis we checked that directory logging was happening too frequently and causing high latencies for the rename operations. Taking into account that installation/upgrade of some of these packages needed about a few thousand file renames, the slowdown was very noticeable for the user. The issue was caused indirectly due to an excessive number of inode evictions on a 5.15 kernel, about 100x more compared to a 5.13, 5.14 or a 5.16-rc8 kernel. After an inode eviction we can't tell for sure, in an efficient way, if an inode was previously logged in the current transaction, so we are pessimistic and assume it was, because in case it was we need to update the logged inode. More details on that in one of the patches in the same series (subject "btrfs: avoid inode logging during rename and link when possible"). Either way, in case the parent directory was logged before, we currently do more work then necessary during a rename, and this change minimizes that amount of work. The following script mimics part of what a package installation/upgrade with zypper does, which is basically renaming a lot of files, in some directory under /usr, to a name with a suffix of "-RPMDELETE": $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=10000 mkfs.btrfs -f $DEV mount $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync # Do some change to testdir and fsync it. echo -n > $MNT/testdir/file_$((NUM_FILES + 1)) xfs_io -c "fsync" $MNT/testdir echo "Renaming $NUM_FILES files..." start=$(date +%s%N) for ((i = 1; i <= $NUM_FILES; i++)); do mv $MNT/testdir/file_$i $MNT/testdir/file_$i-RPMDELETE done end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "Renames took $dur milliseconds" umount $MNT Testing this change on box using a non-debug kernel (Debian's default kernel config) gave the following results: NUM_FILES=10000, before this patch: 27399 ms NUM_FILES=10000, after this patch: 9093 ms (-66.8%) NUM_FILES=5000, before this patch: 9241 ms NUM_FILES=5000, after this patch: 4642 ms (-49.8%) NUM_FILES=2000, before this patch: 2550 ms NUM_FILES=2000, after this patch: 1788 ms (-29.9%) NUM_FILES=1000, before this patch: 1088 ms NUM_FILES=1000, after this patch: 905 ms (-16.9%) Link: https://bugzilla.opensuse.org/show_bug.cgi?id=1193549 Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-01-20 14:00:08 +03:00
* was previously logged, make sure that on log replay we get the old
* dir entry deleted. This is needed because we will also log the new
* name of the renamed inode, so we need to make sure that after log
* replay we don't end up with both the new and old dir entries existing.
btrfs: fix removed dentries still existing after log is synced When we move one inode from one directory to another and both the inode and its previous parent directory were logged before, we are not supposed to have the dentry for the old parent if we have a power failure after the log is synced. Only the new dentry is supposed to exist. Generally this works correctly, however there is a scenario where this is not currently working, because the old parent of the file/directory that was moved is not authoritative for a range that includes the dir index and dir item keys of the old dentry. This case is better explained with the following example and reproducer: # The test requires a very specific layout of keys and items in the # fs/subvolume btree to trigger the bug. So we want to make sure that # on whatever platform we are, we have the same leaf/node size. # # Currently in btrfs the node/leaf size can not be smaller than the page # size (but it can be greater than the page size). So use the largest # supported node/leaf size (64K). $ mkfs.btrfs -f -n 65536 /dev/sdc $ mount /dev/sdc /mnt # "testdir" is inode 257. $ mkdir /mnt/testdir $ chmod 755 /mnt/testdir # Create several empty files to have the directory "testdir" with its # items spread over several leaves (7 in this case). $ for ((i = 1; i <= 1200; i++)); do echo -n > /mnt/testdir/file$i done # Create our test directory "dira", inode number 1458, which gets all # its items in leaf 7. # # The BTRFS_DIR_ITEM_KEY item for inode 257 ("testdir") that points to # the entry named "dira" is in leaf 2, while the BTRFS_DIR_INDEX_KEY # item that points to that entry is in leaf 3. # # For this particular filesystem node size (64K), file count and file # names, we endup with the directory entry items from inode 257 in # leaves 2 and 3, as previously mentioned - what matters for triggering # the bug exercised by this test case is that those items are not placed # in leaf 1, they must be placed in a leaf different from the one # containing the inode item for inode 257. # # The corresponding BTRFS_DIR_ITEM_KEY and BTRFS_DIR_INDEX_KEY items for # the parent inode (257) are the following: # # item 460 key (257 DIR_ITEM 3724298081) itemoff 48344 itemsize 34 # location key (1458 INODE_ITEM 0) type DIR # transid 6 data_len 0 name_len 4 # name: dira # # and: # # item 771 key (257 DIR_INDEX 1202) itemoff 36673 itemsize 34 # location key (1458 INODE_ITEM 0) type DIR # transid 6 data_len 0 name_len 4 # name: dira $ mkdir /mnt/testdir/dira # Make sure everything done so far is durably persisted. $ sync # Now do a change to inode 257 ("testdir") that does not result in # COWing leaves 2 and 3 - the leaves that contain the directory items # pointing to inode 1458 (directory "dira"). # # Changing permissions, the owner/group, updating or adding a xattr, # etc, will not change (COW) leaves 2 and 3. So for the sake of # simplicity change the permissions of inode 257, which results in # updating its inode item and therefore change (COW) only leaf 1. $ chmod 700 /mnt/testdir # Now fsync directory inode 257. # # Since only the first leaf was changed/COWed, we log the inode item of # inode 257 and only the dentries found in the first leaf, all have a # key type of BTRFS_DIR_ITEM_KEY, and no keys of type # BTRFS_DIR_INDEX_KEY, because they sort after the former type and none # exist in the first leaf. # # We also log 3 items that represent ranges for dir items and dir # indexes for which the log is authoritative: # # 1) a key of type BTRFS_DIR_LOG_ITEM_KEY, which indicates the log is # authoritative for all BTRFS_DIR_ITEM_KEY keys that have an offset # in the range [0, 2285968570] (the offset here is the crc32c of the # dentry's name). The value 2285968570 corresponds to the offset of # the first key of leaf 2 (which is of type BTRFS_DIR_ITEM_KEY); # # 2) a key of type BTRFS_DIR_LOG_ITEM_KEY, which indicates the log is # authoritative for all BTRFS_DIR_ITEM_KEY keys that have an offset # in the range [4293818216, (u64)-1] (the offset here is the crc32c # of the dentry's name). The value 4293818216 corresponds to the # offset of the highest key of type BTRFS_DIR_ITEM_KEY plus 1 # (4293818215 + 1), which is located in leaf 2; # # 3) a key of type BTRFS_DIR_LOG_INDEX_KEY, with an offset of 1203, # which indicates the log is authoritative for all keys of type # BTRFS_DIR_INDEX_KEY that have an offset in the range # [1203, (u64)-1]. The value 1203 corresponds to the offset of the # last key of type BTRFS_DIR_INDEX_KEY plus 1 (1202 + 1), which is # located in leaf 3; # # Also, because "testdir" is a directory and inode 1458 ("dira") is a # child directory, we log inode 1458 too. $ xfs_io -c "fsync" /mnt/testdir # Now move "dira", inode 1458, to be a child of the root directory # (inode 256). # # Because this inode was previously logged, when "testdir" was fsynced, # the log is updated so that the old inode reference, referring to inode # 257 as the parent, is deleted and the new inode reference, referring # to inode 256 as the parent, is added to the log. $ mv /mnt/testdir/dira /mnt # Now change some file and fsync it. This guarantees the log changes # made by the previous move/rename operation are persisted. We do not # need to do any special modification to the file, just any change to # any file and sync the log. $ xfs_io -c "pwrite -S 0xab 0 64K" -c "fsync" /mnt/testdir/file1 # Simulate a power failure and then mount again the filesystem to # replay the log tree. We want to verify that we are able to mount the # filesystem, meaning log replay was successful, and that directory # inode 1458 ("dira") only has inode 256 (the filesystem's root) as # its parent (and no longer a child of inode 257). # # It used to happen that during log replay we would end up having # inode 1458 (directory "dira") with 2 hard links, being a child of # inode 257 ("testdir") and inode 256 (the filesystem's root). This # resulted in the tree checker detecting the issue and causing the # mount operation to fail (with -EIO). # # This happened because in the log we have the new name/parent for # inode 1458, which results in adding the new dentry with inode 256 # as the parent, but the previous dentry, under inode 257 was never # removed - this is because the ranges for dir items and dir indexes # of inode 257 for which the log is authoritative do not include the # old dir item and dir index for the dentry of inode 257 referring to # inode 1458: # # - for dir items, the log is authoritative for the ranges # [0, 2285968570] and [4293818216, (u64)-1]. The dir item at inode 257 # pointing to inode 1458 has a key of (257 DIR_ITEM 3724298081), as # previously mentioned, so the dir item is not deleted when the log # replay procedure processes the authoritative ranges, as 3724298081 # is outside both ranges; # # - for dir indexes, the log is authoritative for the range # [1203, (u64)-1], and the dir index item of inode 257 pointing to # inode 1458 has a key of (257 DIR_INDEX 1202), as previously # mentioned, so the dir index item is not deleted when the log # replay procedure processes the authoritative range. <power failure> $ mount /dev/sdc /mnt mount: /mnt: can't read superblock on /dev/sdc. $ dmesg (...) [87849.840509] BTRFS info (device sdc): start tree-log replay [87849.875719] BTRFS critical (device sdc): corrupt leaf: root=5 block=30539776 slot=554 ino=1458, invalid nlink: has 2 expect no more than 1 for dir [87849.878084] BTRFS info (device sdc): leaf 30539776 gen 7 total ptrs 557 free space 2092 owner 5 [87849.879516] BTRFS info (device sdc): refs 1 lock_owner 0 current 2099108 [87849.880613] item 0 key (1181 1 0) itemoff 65275 itemsize 160 [87849.881544] inode generation 6 size 0 mode 100644 [87849.882692] item 1 key (1181 12 257) itemoff 65258 itemsize 17 (...) [87850.562549] item 556 key (1458 12 257) itemoff 16017 itemsize 14 [87850.563349] BTRFS error (device dm-0): block=30539776 write time tree block corruption detected [87850.564386] ------------[ cut here ]------------ [87850.564920] WARNING: CPU: 3 PID: 2099108 at fs/btrfs/disk-io.c:465 csum_one_extent_buffer+0xed/0x100 [btrfs] [87850.566129] Modules linked in: btrfs dm_zero dm_snapshot (...) [87850.573789] CPU: 3 PID: 2099108 Comm: mount Not tainted 5.12.0-rc8-btrfs-next-86 #1 (...) [87850.587481] Call Trace: [87850.587768] btree_csum_one_bio+0x244/0x2b0 [btrfs] [87850.588354] ? btrfs_bio_fits_in_stripe+0xd8/0x110 [btrfs] [87850.589003] btrfs_submit_metadata_bio+0xb7/0x100 [btrfs] [87850.589654] submit_one_bio+0x61/0x70 [btrfs] [87850.590248] submit_extent_page+0x91/0x2f0 [btrfs] [87850.590842] write_one_eb+0x175/0x440 [btrfs] [87850.591370] ? find_extent_buffer_nolock+0x1c0/0x1c0 [btrfs] [87850.592036] btree_write_cache_pages+0x1e6/0x610 [btrfs] [87850.592665] ? free_debug_processing+0x1d5/0x240 [87850.593209] do_writepages+0x43/0xf0 [87850.593798] ? __filemap_fdatawrite_range+0xa4/0x100 [87850.594391] __filemap_fdatawrite_range+0xc5/0x100 [87850.595196] btrfs_write_marked_extents+0x68/0x160 [btrfs] [87850.596202] btrfs_write_and_wait_transaction.isra.0+0x4d/0xd0 [btrfs] [87850.597377] btrfs_commit_transaction+0x794/0xca0 [btrfs] [87850.598455] ? _raw_spin_unlock_irqrestore+0x32/0x60 [87850.599305] ? kmem_cache_free+0x15a/0x3d0 [87850.600029] btrfs_recover_log_trees+0x346/0x380 [btrfs] [87850.601021] ? replay_one_extent+0x7d0/0x7d0 [btrfs] [87850.601988] open_ctree+0x13c9/0x1698 [btrfs] [87850.602846] btrfs_mount_root.cold+0x13/0xed [btrfs] [87850.603771] ? kmem_cache_alloc_trace+0x7c9/0x930 [87850.604576] ? vfs_parse_fs_string+0x5d/0xb0 [87850.605293] ? kfree+0x276/0x3f0 [87850.605857] legacy_get_tree+0x30/0x50 [87850.606540] vfs_get_tree+0x28/0xc0 [87850.607163] fc_mount+0xe/0x40 [87850.607695] vfs_kern_mount.part.0+0x71/0x90 [87850.608440] btrfs_mount+0x13b/0x3e0 [btrfs] (...) [87850.629477] ---[ end trace 68802022b99a1ea0 ]--- [87850.630849] BTRFS: error (device sdc) in btrfs_commit_transaction:2381: errno=-5 IO failure (Error while writing out transaction) [87850.632422] BTRFS warning (device sdc): Skipping commit of aborted transaction. [87850.633416] BTRFS: error (device sdc) in cleanup_transaction:1978: errno=-5 IO failure [87850.634553] BTRFS: error (device sdc) in btrfs_replay_log:2431: errno=-5 IO failure (Failed to recover log tree) [87850.637529] BTRFS error (device sdc): open_ctree failed In this example the inode we moved was a directory, so it was easy to detect the problem because directories can only have one hard link and the tree checker immediately detects that. If the moved inode was a file, then the log replay would succeed and we would end up having both the new hard link (/mnt/foo) and the old hard link (/mnt/testdir/foo) present, but only the new one should be present. Fix this by forcing re-logging of the old parent directory when logging the new name during a rename operation. This ensures we end up with a log that is authoritative for a range covering the keys for the old dentry, therefore causing the old dentry do be deleted when replaying the log. A test case for fstests will follow up soon. Fixes: 64d6b281ba4db0 ("btrfs: remove unnecessary check_parent_dirs_for_sync()") CC: stable@vger.kernel.org # 5.12+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-05-12 18:27:16 +03:00
*/
btrfs: avoid logging all directory changes during renames When doing a rename of a file, if the file or its old parent directory were logged before, we log the new name of the file and then make sure we log the old parent directory, to ensure that after a log replay the old name of the file is deleted and the new name added. The logging of the old parent directory can take some time, because it will scan all leaves modified in the current transaction, check which directory entries were already logged, copy the ones that were not logged before, etc. In this rename context all we need to do is make sure that the old name of the file is deleted on log replay, so instead of triggering a directory log operation, we can just delete the old directory entry from the log if it's there, or in case it isn't there, just log a range item to signal log replay that the old name must be deleted. So change btrfs_log_new_name() to do that. This scenario is actually not uncommon to trigger, and recently on a 5.15 kernel, an openSUSE Tumbleweed user reported package installations and upgrades, with the zypper tool, were often taking a long time to complete, much more than usual. With strace it could be observed that zypper was spending over 99% of its time on rename operations, and then with further analysis we checked that directory logging was happening too frequently and causing high latencies for the rename operations. Taking into account that installation/upgrade of some of these packages needed about a few thousand file renames, the slowdown was very noticeable for the user. The issue was caused indirectly due to an excessive number of inode evictions on a 5.15 kernel, about 100x more compared to a 5.13, 5.14 or a 5.16-rc8 kernel. After an inode eviction we can't tell for sure, in an efficient way, if an inode was previously logged in the current transaction, so we are pessimistic and assume it was, because in case it was we need to update the logged inode. More details on that in one of the patches in the same series (subject "btrfs: avoid inode logging during rename and link when possible"). Either way, in case the parent directory was logged before, we currently do more work then necessary during a rename, and this change minimizes that amount of work. The following script mimics part of what a package installation/upgrade with zypper does, which is basically renaming a lot of files, in some directory under /usr, to a name with a suffix of "-RPMDELETE": $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=10000 mkfs.btrfs -f $DEV mount $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync # Do some change to testdir and fsync it. echo -n > $MNT/testdir/file_$((NUM_FILES + 1)) xfs_io -c "fsync" $MNT/testdir echo "Renaming $NUM_FILES files..." start=$(date +%s%N) for ((i = 1; i <= $NUM_FILES; i++)); do mv $MNT/testdir/file_$i $MNT/testdir/file_$i-RPMDELETE done end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "Renames took $dur milliseconds" umount $MNT Testing this change on box using a non-debug kernel (Debian's default kernel config) gave the following results: NUM_FILES=10000, before this patch: 27399 ms NUM_FILES=10000, after this patch: 9093 ms (-66.8%) NUM_FILES=5000, before this patch: 9241 ms NUM_FILES=5000, after this patch: 4642 ms (-49.8%) NUM_FILES=2000, before this patch: 2550 ms NUM_FILES=2000, after this patch: 1788 ms (-29.9%) NUM_FILES=1000, before this patch: 1088 ms NUM_FILES=1000, after this patch: 905 ms (-16.9%) Link: https://bugzilla.opensuse.org/show_bug.cgi?id=1193549 Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-01-20 14:00:08 +03:00
if (old_dir && old_dir->logged_trans == trans->transid) {
struct btrfs_root *log = old_dir->root->log_root;
struct btrfs_path *path;
ASSERT(old_dir_index >= BTRFS_DIR_START_INDEX);
btrfs: stop doing unnecessary log updates during a rename During a rename, we call __btrfs_unlink_inode(), which will call btrfs_del_inode_ref_in_log() and btrfs_del_dir_entries_in_log(), in order to remove an inode reference and a directory entry from the log. These are necessary when __btrfs_unlink_inode() is called from the unlink path, but not necessary when it's called from a rename context, because: 1) For the btrfs_del_inode_ref_in_log() call, it's pointless to delete the inode reference related to the old name, because later in the rename path we call btrfs_log_new_name(), which will drop all inode references from the log and copy all inode references from the subvolume tree to the log tree. So we are doing one unnecessary btree operation which adds additional latency and lock contention in case there are other tasks accessing the log tree; 2) For the btrfs_del_dir_entries_in_log() call, we are now doing the equivalent at btrfs_log_new_name() since the previous patch in the series, that has the subject "btrfs: avoid logging all directory changes during renames". In fact, having __btrfs_unlink_inode() call this function not only adds additional latency and lock contention due to the extra btree operation, but also can make btrfs_log_new_name() unnecessarily log a range item to track the deletion of the old name, since it has no way to known that the directory entry related to the old name was previously logged and already deleted by __btrfs_unlink_inode() through its call to btrfs_del_dir_entries_in_log(). So skip those calls at __btrfs_unlink_inode() when we are doing a rename. Skipping them also allows us now to reduce the duration of time we are pinning a log transaction during renames, which is always beneficial as it's not delaying so much other tasks trying to sync the log tree, in particular we end up not holding the log transaction pinned while adding the new name (adding inode ref, directory entry, etc). This change is part of a patchset comprised of the following patches: 1/5 btrfs: add helper to delete a dir entry from a log tree 2/5 btrfs: pass the dentry to btrfs_log_new_name() instead of the inode 3/5 btrfs: avoid logging all directory changes during renames 4/5 btrfs: stop doing unnecessary log updates during a rename 5/5 btrfs: avoid inode logging during rename and link when possible Just like the previous patch in the series, "btrfs: avoid logging all directory changes during renames", the following script mimics part of what a package installation/upgrade with zypper does, which is basically renaming a lot of files, in some directory under /usr, to a name with a suffix of "-RPMDELETE": $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=10000 mkfs.btrfs -f $DEV mount $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync # Do some change to testdir and fsync it. echo -n > $MNT/testdir/file_$((NUM_FILES + 1)) xfs_io -c "fsync" $MNT/testdir echo "Renaming $NUM_FILES files..." start=$(date +%s%N) for ((i = 1; i <= $NUM_FILES; i++)); do mv $MNT/testdir/file_$i $MNT/testdir/file_$i-RPMDELETE done end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "Renames took $dur milliseconds" umount $MNT Testing this change on box a using a non-debug kernel (Debian's default kernel config) gave the following results: NUM_FILES=10000, before patchset: 27399 ms NUM_FILES=10000, after patches 1/5 to 3/5 applied: 9093 ms (-66.8%) NUM_FILES=10000, after patches 1/5 to 4/5 applied: 9016 ms (-67.1%) NUM_FILES=5000, before patchset: 9241 ms NUM_FILES=5000, after patches 1/5 to 3/5 applied: 4642 ms (-49.8%) NUM_FILES=5000, after patches 1/5 to 4/5 applied: 4553 ms (-50.7%) NUM_FILES=2000, before patchset: 2550 ms NUM_FILES=2000, after patches 1/5 to 3/5 applied: 1788 ms (-29.9%) NUM_FILES=2000, after patches 1/5 to 4/5 applied: 1767 ms (-30.7%) NUM_FILES=1000, before patchset: 1088 ms NUM_FILES=1000, after patches 1/5 to 3/5 applied: 905 ms (-16.9%) NUM_FILES=1000, after patches 1/5 to 4/5 applied: 883 ms (-18.8%) The next patch in the series (5/5), also contains dbench results after applying to whole patchset. Link: https://bugzilla.opensuse.org/show_bug.cgi?id=1193549 Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-01-20 14:00:09 +03:00
/*
* We have two inodes to update in the log, the old directory and
* the inode that got renamed, so we must pin the log to prevent
* anyone from syncing the log until we have updated both inodes
* in the log.
*/
log_pinned = true;
btrfs_pin_log_trans(root);
btrfs: avoid logging all directory changes during renames When doing a rename of a file, if the file or its old parent directory were logged before, we log the new name of the file and then make sure we log the old parent directory, to ensure that after a log replay the old name of the file is deleted and the new name added. The logging of the old parent directory can take some time, because it will scan all leaves modified in the current transaction, check which directory entries were already logged, copy the ones that were not logged before, etc. In this rename context all we need to do is make sure that the old name of the file is deleted on log replay, so instead of triggering a directory log operation, we can just delete the old directory entry from the log if it's there, or in case it isn't there, just log a range item to signal log replay that the old name must be deleted. So change btrfs_log_new_name() to do that. This scenario is actually not uncommon to trigger, and recently on a 5.15 kernel, an openSUSE Tumbleweed user reported package installations and upgrades, with the zypper tool, were often taking a long time to complete, much more than usual. With strace it could be observed that zypper was spending over 99% of its time on rename operations, and then with further analysis we checked that directory logging was happening too frequently and causing high latencies for the rename operations. Taking into account that installation/upgrade of some of these packages needed about a few thousand file renames, the slowdown was very noticeable for the user. The issue was caused indirectly due to an excessive number of inode evictions on a 5.15 kernel, about 100x more compared to a 5.13, 5.14 or a 5.16-rc8 kernel. After an inode eviction we can't tell for sure, in an efficient way, if an inode was previously logged in the current transaction, so we are pessimistic and assume it was, because in case it was we need to update the logged inode. More details on that in one of the patches in the same series (subject "btrfs: avoid inode logging during rename and link when possible"). Either way, in case the parent directory was logged before, we currently do more work then necessary during a rename, and this change minimizes that amount of work. The following script mimics part of what a package installation/upgrade with zypper does, which is basically renaming a lot of files, in some directory under /usr, to a name with a suffix of "-RPMDELETE": $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=10000 mkfs.btrfs -f $DEV mount $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync # Do some change to testdir and fsync it. echo -n > $MNT/testdir/file_$((NUM_FILES + 1)) xfs_io -c "fsync" $MNT/testdir echo "Renaming $NUM_FILES files..." start=$(date +%s%N) for ((i = 1; i <= $NUM_FILES; i++)); do mv $MNT/testdir/file_$i $MNT/testdir/file_$i-RPMDELETE done end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "Renames took $dur milliseconds" umount $MNT Testing this change on box using a non-debug kernel (Debian's default kernel config) gave the following results: NUM_FILES=10000, before this patch: 27399 ms NUM_FILES=10000, after this patch: 9093 ms (-66.8%) NUM_FILES=5000, before this patch: 9241 ms NUM_FILES=5000, after this patch: 4642 ms (-49.8%) NUM_FILES=2000, before this patch: 2550 ms NUM_FILES=2000, after this patch: 1788 ms (-29.9%) NUM_FILES=1000, before this patch: 1088 ms NUM_FILES=1000, after this patch: 905 ms (-16.9%) Link: https://bugzilla.opensuse.org/show_bug.cgi?id=1193549 Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-01-20 14:00:08 +03:00
path = btrfs_alloc_path();
if (!path) {
btrfs: stop doing unnecessary log updates during a rename During a rename, we call __btrfs_unlink_inode(), which will call btrfs_del_inode_ref_in_log() and btrfs_del_dir_entries_in_log(), in order to remove an inode reference and a directory entry from the log. These are necessary when __btrfs_unlink_inode() is called from the unlink path, but not necessary when it's called from a rename context, because: 1) For the btrfs_del_inode_ref_in_log() call, it's pointless to delete the inode reference related to the old name, because later in the rename path we call btrfs_log_new_name(), which will drop all inode references from the log and copy all inode references from the subvolume tree to the log tree. So we are doing one unnecessary btree operation which adds additional latency and lock contention in case there are other tasks accessing the log tree; 2) For the btrfs_del_dir_entries_in_log() call, we are now doing the equivalent at btrfs_log_new_name() since the previous patch in the series, that has the subject "btrfs: avoid logging all directory changes during renames". In fact, having __btrfs_unlink_inode() call this function not only adds additional latency and lock contention due to the extra btree operation, but also can make btrfs_log_new_name() unnecessarily log a range item to track the deletion of the old name, since it has no way to known that the directory entry related to the old name was previously logged and already deleted by __btrfs_unlink_inode() through its call to btrfs_del_dir_entries_in_log(). So skip those calls at __btrfs_unlink_inode() when we are doing a rename. Skipping them also allows us now to reduce the duration of time we are pinning a log transaction during renames, which is always beneficial as it's not delaying so much other tasks trying to sync the log tree, in particular we end up not holding the log transaction pinned while adding the new name (adding inode ref, directory entry, etc). This change is part of a patchset comprised of the following patches: 1/5 btrfs: add helper to delete a dir entry from a log tree 2/5 btrfs: pass the dentry to btrfs_log_new_name() instead of the inode 3/5 btrfs: avoid logging all directory changes during renames 4/5 btrfs: stop doing unnecessary log updates during a rename 5/5 btrfs: avoid inode logging during rename and link when possible Just like the previous patch in the series, "btrfs: avoid logging all directory changes during renames", the following script mimics part of what a package installation/upgrade with zypper does, which is basically renaming a lot of files, in some directory under /usr, to a name with a suffix of "-RPMDELETE": $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=10000 mkfs.btrfs -f $DEV mount $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync # Do some change to testdir and fsync it. echo -n > $MNT/testdir/file_$((NUM_FILES + 1)) xfs_io -c "fsync" $MNT/testdir echo "Renaming $NUM_FILES files..." start=$(date +%s%N) for ((i = 1; i <= $NUM_FILES; i++)); do mv $MNT/testdir/file_$i $MNT/testdir/file_$i-RPMDELETE done end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "Renames took $dur milliseconds" umount $MNT Testing this change on box a using a non-debug kernel (Debian's default kernel config) gave the following results: NUM_FILES=10000, before patchset: 27399 ms NUM_FILES=10000, after patches 1/5 to 3/5 applied: 9093 ms (-66.8%) NUM_FILES=10000, after patches 1/5 to 4/5 applied: 9016 ms (-67.1%) NUM_FILES=5000, before patchset: 9241 ms NUM_FILES=5000, after patches 1/5 to 3/5 applied: 4642 ms (-49.8%) NUM_FILES=5000, after patches 1/5 to 4/5 applied: 4553 ms (-50.7%) NUM_FILES=2000, before patchset: 2550 ms NUM_FILES=2000, after patches 1/5 to 3/5 applied: 1788 ms (-29.9%) NUM_FILES=2000, after patches 1/5 to 4/5 applied: 1767 ms (-30.7%) NUM_FILES=1000, before patchset: 1088 ms NUM_FILES=1000, after patches 1/5 to 3/5 applied: 905 ms (-16.9%) NUM_FILES=1000, after patches 1/5 to 4/5 applied: 883 ms (-18.8%) The next patch in the series (5/5), also contains dbench results after applying to whole patchset. Link: https://bugzilla.opensuse.org/show_bug.cgi?id=1193549 Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-01-20 14:00:09 +03:00
ret = -ENOMEM;
goto out;
btrfs: avoid logging all directory changes during renames When doing a rename of a file, if the file or its old parent directory were logged before, we log the new name of the file and then make sure we log the old parent directory, to ensure that after a log replay the old name of the file is deleted and the new name added. The logging of the old parent directory can take some time, because it will scan all leaves modified in the current transaction, check which directory entries were already logged, copy the ones that were not logged before, etc. In this rename context all we need to do is make sure that the old name of the file is deleted on log replay, so instead of triggering a directory log operation, we can just delete the old directory entry from the log if it's there, or in case it isn't there, just log a range item to signal log replay that the old name must be deleted. So change btrfs_log_new_name() to do that. This scenario is actually not uncommon to trigger, and recently on a 5.15 kernel, an openSUSE Tumbleweed user reported package installations and upgrades, with the zypper tool, were often taking a long time to complete, much more than usual. With strace it could be observed that zypper was spending over 99% of its time on rename operations, and then with further analysis we checked that directory logging was happening too frequently and causing high latencies for the rename operations. Taking into account that installation/upgrade of some of these packages needed about a few thousand file renames, the slowdown was very noticeable for the user. The issue was caused indirectly due to an excessive number of inode evictions on a 5.15 kernel, about 100x more compared to a 5.13, 5.14 or a 5.16-rc8 kernel. After an inode eviction we can't tell for sure, in an efficient way, if an inode was previously logged in the current transaction, so we are pessimistic and assume it was, because in case it was we need to update the logged inode. More details on that in one of the patches in the same series (subject "btrfs: avoid inode logging during rename and link when possible"). Either way, in case the parent directory was logged before, we currently do more work then necessary during a rename, and this change minimizes that amount of work. The following script mimics part of what a package installation/upgrade with zypper does, which is basically renaming a lot of files, in some directory under /usr, to a name with a suffix of "-RPMDELETE": $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=10000 mkfs.btrfs -f $DEV mount $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync # Do some change to testdir and fsync it. echo -n > $MNT/testdir/file_$((NUM_FILES + 1)) xfs_io -c "fsync" $MNT/testdir echo "Renaming $NUM_FILES files..." start=$(date +%s%N) for ((i = 1; i <= $NUM_FILES; i++)); do mv $MNT/testdir/file_$i $MNT/testdir/file_$i-RPMDELETE done end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "Renames took $dur milliseconds" umount $MNT Testing this change on box using a non-debug kernel (Debian's default kernel config) gave the following results: NUM_FILES=10000, before this patch: 27399 ms NUM_FILES=10000, after this patch: 9093 ms (-66.8%) NUM_FILES=5000, before this patch: 9241 ms NUM_FILES=5000, after this patch: 4642 ms (-49.8%) NUM_FILES=2000, before this patch: 2550 ms NUM_FILES=2000, after this patch: 1788 ms (-29.9%) NUM_FILES=1000, before this patch: 1088 ms NUM_FILES=1000, after this patch: 905 ms (-16.9%) Link: https://bugzilla.opensuse.org/show_bug.cgi?id=1193549 Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-01-20 14:00:08 +03:00
}
/*
* Other concurrent task might be logging the old directory,
* as it can be triggered when logging other inode that had or
* still has a dentry in the old directory. So take the old
* directory's log_mutex to prevent getting an -EEXIST when
* logging a key to record the deletion, or having that other
* task logging the old directory get an -EEXIST if it attempts
* to log the same key after we just did it. In both cases that
* would result in falling back to a transaction commit.
*/
mutex_lock(&old_dir->log_mutex);
ret = del_logged_dentry(trans, log, path, btrfs_ino(old_dir),
old_dentry->d_name.name,
old_dentry->d_name.len, old_dir_index);
if (ret > 0) {
/*
* The dentry does not exist in the log, so record its
* deletion.
*/
btrfs_release_path(path);
ret = insert_dir_log_key(trans, log, path,
btrfs_ino(old_dir),
old_dir_index, old_dir_index);
}
mutex_unlock(&old_dir->log_mutex);
btrfs_free_path(path);
btrfs: stop doing unnecessary log updates during a rename During a rename, we call __btrfs_unlink_inode(), which will call btrfs_del_inode_ref_in_log() and btrfs_del_dir_entries_in_log(), in order to remove an inode reference and a directory entry from the log. These are necessary when __btrfs_unlink_inode() is called from the unlink path, but not necessary when it's called from a rename context, because: 1) For the btrfs_del_inode_ref_in_log() call, it's pointless to delete the inode reference related to the old name, because later in the rename path we call btrfs_log_new_name(), which will drop all inode references from the log and copy all inode references from the subvolume tree to the log tree. So we are doing one unnecessary btree operation which adds additional latency and lock contention in case there are other tasks accessing the log tree; 2) For the btrfs_del_dir_entries_in_log() call, we are now doing the equivalent at btrfs_log_new_name() since the previous patch in the series, that has the subject "btrfs: avoid logging all directory changes during renames". In fact, having __btrfs_unlink_inode() call this function not only adds additional latency and lock contention due to the extra btree operation, but also can make btrfs_log_new_name() unnecessarily log a range item to track the deletion of the old name, since it has no way to known that the directory entry related to the old name was previously logged and already deleted by __btrfs_unlink_inode() through its call to btrfs_del_dir_entries_in_log(). So skip those calls at __btrfs_unlink_inode() when we are doing a rename. Skipping them also allows us now to reduce the duration of time we are pinning a log transaction during renames, which is always beneficial as it's not delaying so much other tasks trying to sync the log tree, in particular we end up not holding the log transaction pinned while adding the new name (adding inode ref, directory entry, etc). This change is part of a patchset comprised of the following patches: 1/5 btrfs: add helper to delete a dir entry from a log tree 2/5 btrfs: pass the dentry to btrfs_log_new_name() instead of the inode 3/5 btrfs: avoid logging all directory changes during renames 4/5 btrfs: stop doing unnecessary log updates during a rename 5/5 btrfs: avoid inode logging during rename and link when possible Just like the previous patch in the series, "btrfs: avoid logging all directory changes during renames", the following script mimics part of what a package installation/upgrade with zypper does, which is basically renaming a lot of files, in some directory under /usr, to a name with a suffix of "-RPMDELETE": $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=10000 mkfs.btrfs -f $DEV mount $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync # Do some change to testdir and fsync it. echo -n > $MNT/testdir/file_$((NUM_FILES + 1)) xfs_io -c "fsync" $MNT/testdir echo "Renaming $NUM_FILES files..." start=$(date +%s%N) for ((i = 1; i <= $NUM_FILES; i++)); do mv $MNT/testdir/file_$i $MNT/testdir/file_$i-RPMDELETE done end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "Renames took $dur milliseconds" umount $MNT Testing this change on box a using a non-debug kernel (Debian's default kernel config) gave the following results: NUM_FILES=10000, before patchset: 27399 ms NUM_FILES=10000, after patches 1/5 to 3/5 applied: 9093 ms (-66.8%) NUM_FILES=10000, after patches 1/5 to 4/5 applied: 9016 ms (-67.1%) NUM_FILES=5000, before patchset: 9241 ms NUM_FILES=5000, after patches 1/5 to 3/5 applied: 4642 ms (-49.8%) NUM_FILES=5000, after patches 1/5 to 4/5 applied: 4553 ms (-50.7%) NUM_FILES=2000, before patchset: 2550 ms NUM_FILES=2000, after patches 1/5 to 3/5 applied: 1788 ms (-29.9%) NUM_FILES=2000, after patches 1/5 to 4/5 applied: 1767 ms (-30.7%) NUM_FILES=1000, before patchset: 1088 ms NUM_FILES=1000, after patches 1/5 to 3/5 applied: 905 ms (-16.9%) NUM_FILES=1000, after patches 1/5 to 4/5 applied: 883 ms (-18.8%) The next patch in the series (5/5), also contains dbench results after applying to whole patchset. Link: https://bugzilla.opensuse.org/show_bug.cgi?id=1193549 Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-01-20 14:00:09 +03:00
if (ret < 0)
goto out;
btrfs: avoid logging all directory changes during renames When doing a rename of a file, if the file or its old parent directory were logged before, we log the new name of the file and then make sure we log the old parent directory, to ensure that after a log replay the old name of the file is deleted and the new name added. The logging of the old parent directory can take some time, because it will scan all leaves modified in the current transaction, check which directory entries were already logged, copy the ones that were not logged before, etc. In this rename context all we need to do is make sure that the old name of the file is deleted on log replay, so instead of triggering a directory log operation, we can just delete the old directory entry from the log if it's there, or in case it isn't there, just log a range item to signal log replay that the old name must be deleted. So change btrfs_log_new_name() to do that. This scenario is actually not uncommon to trigger, and recently on a 5.15 kernel, an openSUSE Tumbleweed user reported package installations and upgrades, with the zypper tool, were often taking a long time to complete, much more than usual. With strace it could be observed that zypper was spending over 99% of its time on rename operations, and then with further analysis we checked that directory logging was happening too frequently and causing high latencies for the rename operations. Taking into account that installation/upgrade of some of these packages needed about a few thousand file renames, the slowdown was very noticeable for the user. The issue was caused indirectly due to an excessive number of inode evictions on a 5.15 kernel, about 100x more compared to a 5.13, 5.14 or a 5.16-rc8 kernel. After an inode eviction we can't tell for sure, in an efficient way, if an inode was previously logged in the current transaction, so we are pessimistic and assume it was, because in case it was we need to update the logged inode. More details on that in one of the patches in the same series (subject "btrfs: avoid inode logging during rename and link when possible"). Either way, in case the parent directory was logged before, we currently do more work then necessary during a rename, and this change minimizes that amount of work. The following script mimics part of what a package installation/upgrade with zypper does, which is basically renaming a lot of files, in some directory under /usr, to a name with a suffix of "-RPMDELETE": $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=10000 mkfs.btrfs -f $DEV mount $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync # Do some change to testdir and fsync it. echo -n > $MNT/testdir/file_$((NUM_FILES + 1)) xfs_io -c "fsync" $MNT/testdir echo "Renaming $NUM_FILES files..." start=$(date +%s%N) for ((i = 1; i <= $NUM_FILES; i++)); do mv $MNT/testdir/file_$i $MNT/testdir/file_$i-RPMDELETE done end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "Renames took $dur milliseconds" umount $MNT Testing this change on box using a non-debug kernel (Debian's default kernel config) gave the following results: NUM_FILES=10000, before this patch: 27399 ms NUM_FILES=10000, after this patch: 9093 ms (-66.8%) NUM_FILES=5000, before this patch: 9241 ms NUM_FILES=5000, after this patch: 4642 ms (-49.8%) NUM_FILES=2000, before this patch: 2550 ms NUM_FILES=2000, after this patch: 1788 ms (-29.9%) NUM_FILES=1000, before this patch: 1088 ms NUM_FILES=1000, after this patch: 905 ms (-16.9%) Link: https://bugzilla.opensuse.org/show_bug.cgi?id=1193549 Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-01-20 14:00:08 +03:00
}
btrfs: fix removed dentries still existing after log is synced When we move one inode from one directory to another and both the inode and its previous parent directory were logged before, we are not supposed to have the dentry for the old parent if we have a power failure after the log is synced. Only the new dentry is supposed to exist. Generally this works correctly, however there is a scenario where this is not currently working, because the old parent of the file/directory that was moved is not authoritative for a range that includes the dir index and dir item keys of the old dentry. This case is better explained with the following example and reproducer: # The test requires a very specific layout of keys and items in the # fs/subvolume btree to trigger the bug. So we want to make sure that # on whatever platform we are, we have the same leaf/node size. # # Currently in btrfs the node/leaf size can not be smaller than the page # size (but it can be greater than the page size). So use the largest # supported node/leaf size (64K). $ mkfs.btrfs -f -n 65536 /dev/sdc $ mount /dev/sdc /mnt # "testdir" is inode 257. $ mkdir /mnt/testdir $ chmod 755 /mnt/testdir # Create several empty files to have the directory "testdir" with its # items spread over several leaves (7 in this case). $ for ((i = 1; i <= 1200; i++)); do echo -n > /mnt/testdir/file$i done # Create our test directory "dira", inode number 1458, which gets all # its items in leaf 7. # # The BTRFS_DIR_ITEM_KEY item for inode 257 ("testdir") that points to # the entry named "dira" is in leaf 2, while the BTRFS_DIR_INDEX_KEY # item that points to that entry is in leaf 3. # # For this particular filesystem node size (64K), file count and file # names, we endup with the directory entry items from inode 257 in # leaves 2 and 3, as previously mentioned - what matters for triggering # the bug exercised by this test case is that those items are not placed # in leaf 1, they must be placed in a leaf different from the one # containing the inode item for inode 257. # # The corresponding BTRFS_DIR_ITEM_KEY and BTRFS_DIR_INDEX_KEY items for # the parent inode (257) are the following: # # item 460 key (257 DIR_ITEM 3724298081) itemoff 48344 itemsize 34 # location key (1458 INODE_ITEM 0) type DIR # transid 6 data_len 0 name_len 4 # name: dira # # and: # # item 771 key (257 DIR_INDEX 1202) itemoff 36673 itemsize 34 # location key (1458 INODE_ITEM 0) type DIR # transid 6 data_len 0 name_len 4 # name: dira $ mkdir /mnt/testdir/dira # Make sure everything done so far is durably persisted. $ sync # Now do a change to inode 257 ("testdir") that does not result in # COWing leaves 2 and 3 - the leaves that contain the directory items # pointing to inode 1458 (directory "dira"). # # Changing permissions, the owner/group, updating or adding a xattr, # etc, will not change (COW) leaves 2 and 3. So for the sake of # simplicity change the permissions of inode 257, which results in # updating its inode item and therefore change (COW) only leaf 1. $ chmod 700 /mnt/testdir # Now fsync directory inode 257. # # Since only the first leaf was changed/COWed, we log the inode item of # inode 257 and only the dentries found in the first leaf, all have a # key type of BTRFS_DIR_ITEM_KEY, and no keys of type # BTRFS_DIR_INDEX_KEY, because they sort after the former type and none # exist in the first leaf. # # We also log 3 items that represent ranges for dir items and dir # indexes for which the log is authoritative: # # 1) a key of type BTRFS_DIR_LOG_ITEM_KEY, which indicates the log is # authoritative for all BTRFS_DIR_ITEM_KEY keys that have an offset # in the range [0, 2285968570] (the offset here is the crc32c of the # dentry's name). The value 2285968570 corresponds to the offset of # the first key of leaf 2 (which is of type BTRFS_DIR_ITEM_KEY); # # 2) a key of type BTRFS_DIR_LOG_ITEM_KEY, which indicates the log is # authoritative for all BTRFS_DIR_ITEM_KEY keys that have an offset # in the range [4293818216, (u64)-1] (the offset here is the crc32c # of the dentry's name). The value 4293818216 corresponds to the # offset of the highest key of type BTRFS_DIR_ITEM_KEY plus 1 # (4293818215 + 1), which is located in leaf 2; # # 3) a key of type BTRFS_DIR_LOG_INDEX_KEY, with an offset of 1203, # which indicates the log is authoritative for all keys of type # BTRFS_DIR_INDEX_KEY that have an offset in the range # [1203, (u64)-1]. The value 1203 corresponds to the offset of the # last key of type BTRFS_DIR_INDEX_KEY plus 1 (1202 + 1), which is # located in leaf 3; # # Also, because "testdir" is a directory and inode 1458 ("dira") is a # child directory, we log inode 1458 too. $ xfs_io -c "fsync" /mnt/testdir # Now move "dira", inode 1458, to be a child of the root directory # (inode 256). # # Because this inode was previously logged, when "testdir" was fsynced, # the log is updated so that the old inode reference, referring to inode # 257 as the parent, is deleted and the new inode reference, referring # to inode 256 as the parent, is added to the log. $ mv /mnt/testdir/dira /mnt # Now change some file and fsync it. This guarantees the log changes # made by the previous move/rename operation are persisted. We do not # need to do any special modification to the file, just any change to # any file and sync the log. $ xfs_io -c "pwrite -S 0xab 0 64K" -c "fsync" /mnt/testdir/file1 # Simulate a power failure and then mount again the filesystem to # replay the log tree. We want to verify that we are able to mount the # filesystem, meaning log replay was successful, and that directory # inode 1458 ("dira") only has inode 256 (the filesystem's root) as # its parent (and no longer a child of inode 257). # # It used to happen that during log replay we would end up having # inode 1458 (directory "dira") with 2 hard links, being a child of # inode 257 ("testdir") and inode 256 (the filesystem's root). This # resulted in the tree checker detecting the issue and causing the # mount operation to fail (with -EIO). # # This happened because in the log we have the new name/parent for # inode 1458, which results in adding the new dentry with inode 256 # as the parent, but the previous dentry, under inode 257 was never # removed - this is because the ranges for dir items and dir indexes # of inode 257 for which the log is authoritative do not include the # old dir item and dir index for the dentry of inode 257 referring to # inode 1458: # # - for dir items, the log is authoritative for the ranges # [0, 2285968570] and [4293818216, (u64)-1]. The dir item at inode 257 # pointing to inode 1458 has a key of (257 DIR_ITEM 3724298081), as # previously mentioned, so the dir item is not deleted when the log # replay procedure processes the authoritative ranges, as 3724298081 # is outside both ranges; # # - for dir indexes, the log is authoritative for the range # [1203, (u64)-1], and the dir index item of inode 257 pointing to # inode 1458 has a key of (257 DIR_INDEX 1202), as previously # mentioned, so the dir index item is not deleted when the log # replay procedure processes the authoritative range. <power failure> $ mount /dev/sdc /mnt mount: /mnt: can't read superblock on /dev/sdc. $ dmesg (...) [87849.840509] BTRFS info (device sdc): start tree-log replay [87849.875719] BTRFS critical (device sdc): corrupt leaf: root=5 block=30539776 slot=554 ino=1458, invalid nlink: has 2 expect no more than 1 for dir [87849.878084] BTRFS info (device sdc): leaf 30539776 gen 7 total ptrs 557 free space 2092 owner 5 [87849.879516] BTRFS info (device sdc): refs 1 lock_owner 0 current 2099108 [87849.880613] item 0 key (1181 1 0) itemoff 65275 itemsize 160 [87849.881544] inode generation 6 size 0 mode 100644 [87849.882692] item 1 key (1181 12 257) itemoff 65258 itemsize 17 (...) [87850.562549] item 556 key (1458 12 257) itemoff 16017 itemsize 14 [87850.563349] BTRFS error (device dm-0): block=30539776 write time tree block corruption detected [87850.564386] ------------[ cut here ]------------ [87850.564920] WARNING: CPU: 3 PID: 2099108 at fs/btrfs/disk-io.c:465 csum_one_extent_buffer+0xed/0x100 [btrfs] [87850.566129] Modules linked in: btrfs dm_zero dm_snapshot (...) [87850.573789] CPU: 3 PID: 2099108 Comm: mount Not tainted 5.12.0-rc8-btrfs-next-86 #1 (...) [87850.587481] Call Trace: [87850.587768] btree_csum_one_bio+0x244/0x2b0 [btrfs] [87850.588354] ? btrfs_bio_fits_in_stripe+0xd8/0x110 [btrfs] [87850.589003] btrfs_submit_metadata_bio+0xb7/0x100 [btrfs] [87850.589654] submit_one_bio+0x61/0x70 [btrfs] [87850.590248] submit_extent_page+0x91/0x2f0 [btrfs] [87850.590842] write_one_eb+0x175/0x440 [btrfs] [87850.591370] ? find_extent_buffer_nolock+0x1c0/0x1c0 [btrfs] [87850.592036] btree_write_cache_pages+0x1e6/0x610 [btrfs] [87850.592665] ? free_debug_processing+0x1d5/0x240 [87850.593209] do_writepages+0x43/0xf0 [87850.593798] ? __filemap_fdatawrite_range+0xa4/0x100 [87850.594391] __filemap_fdatawrite_range+0xc5/0x100 [87850.595196] btrfs_write_marked_extents+0x68/0x160 [btrfs] [87850.596202] btrfs_write_and_wait_transaction.isra.0+0x4d/0xd0 [btrfs] [87850.597377] btrfs_commit_transaction+0x794/0xca0 [btrfs] [87850.598455] ? _raw_spin_unlock_irqrestore+0x32/0x60 [87850.599305] ? kmem_cache_free+0x15a/0x3d0 [87850.600029] btrfs_recover_log_trees+0x346/0x380 [btrfs] [87850.601021] ? replay_one_extent+0x7d0/0x7d0 [btrfs] [87850.601988] open_ctree+0x13c9/0x1698 [btrfs] [87850.602846] btrfs_mount_root.cold+0x13/0xed [btrfs] [87850.603771] ? kmem_cache_alloc_trace+0x7c9/0x930 [87850.604576] ? vfs_parse_fs_string+0x5d/0xb0 [87850.605293] ? kfree+0x276/0x3f0 [87850.605857] legacy_get_tree+0x30/0x50 [87850.606540] vfs_get_tree+0x28/0xc0 [87850.607163] fc_mount+0xe/0x40 [87850.607695] vfs_kern_mount.part.0+0x71/0x90 [87850.608440] btrfs_mount+0x13b/0x3e0 [btrfs] (...) [87850.629477] ---[ end trace 68802022b99a1ea0 ]--- [87850.630849] BTRFS: error (device sdc) in btrfs_commit_transaction:2381: errno=-5 IO failure (Error while writing out transaction) [87850.632422] BTRFS warning (device sdc): Skipping commit of aborted transaction. [87850.633416] BTRFS: error (device sdc) in cleanup_transaction:1978: errno=-5 IO failure [87850.634553] BTRFS: error (device sdc) in btrfs_replay_log:2431: errno=-5 IO failure (Failed to recover log tree) [87850.637529] BTRFS error (device sdc): open_ctree failed In this example the inode we moved was a directory, so it was easy to detect the problem because directories can only have one hard link and the tree checker immediately detects that. If the moved inode was a file, then the log replay would succeed and we would end up having both the new hard link (/mnt/foo) and the old hard link (/mnt/testdir/foo) present, but only the new one should be present. Fix this by forcing re-logging of the old parent directory when logging the new name during a rename operation. This ensures we end up with a log that is authoritative for a range covering the keys for the old dentry, therefore causing the old dentry do be deleted when replaying the log. A test case for fstests will follow up soon. Fixes: 64d6b281ba4db0 ("btrfs: remove unnecessary check_parent_dirs_for_sync()") CC: stable@vger.kernel.org # 5.12+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-05-12 18:27:16 +03:00
btrfs: do not commit logs and transactions during link and rename operations Since commit d4682ba03ef618 ("Btrfs: sync log after logging new name") we started to commit logs, and fallback to transaction commits when we failed to log the new names or commit the logs, after link and rename operations when the target inodes (or their parents) were previously logged in the current transaction. This was to avoid losing directories despite an explicit fsync on them when they are ancestors of some inode that got a new named logged, due to a link or rename operation. However that adds the cost of starting IO and waiting for it to complete, which can cause higher latencies for applications. Instead of doing that, just make sure that when we log a new name for an inode we don't mark any of its ancestors as logged, so that if any one does an fsync against any of them, without doing any other change on them, the fsync commits the log. This way we only pay the cost of a log commit (or a transaction commit if something goes wrong or a new block group was created) if the application explicitly asks to fsync any of the parent directories. Using dbench, which mixes several filesystems operations including renames, revealed some significant latency gains. The following script that uses dbench was used to test this: #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/btrfs MOUNT_OPTIONS="-o ssd -o space_cache=v2" MKFS_OPTIONS="-m single -d single" THREADS=16 echo "performance" | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -t 300 -D $MNT $THREADS umount $MNT The test was run on bare metal, no virtualization, on a box with 12 cores (Intel i7-8700), 64Gb of RAM and using a NVMe device, with a kernel configuration that is the default of typical distributions (debian in this case), without debug options enabled (kasan, kmemleak, slub debug, debug of page allocations, lock debugging, etc). Results before this patch: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 10750455 0.011 155.088 Close 7896674 0.001 0.243 Rename 455222 2.158 1101.947 Unlink 2171189 0.067 121.638 Deltree 256 2.425 7.816 Mkdir 128 0.002 0.003 Qpathinfo 9744323 0.006 21.370 Qfileinfo 1707092 0.001 0.146 Qfsinfo 1786756 0.001 11.228 Sfileinfo 875612 0.003 21.263 Find 3767281 0.025 9.617 WriteX 5356924 0.011 211.390 ReadX 16852694 0.003 9.442 LockX 35008 0.002 0.119 UnlockX 35008 0.001 0.138 Flush 753458 4.252 1102.249 Throughput 1128.35 MB/sec 16 clients 16 procs max_latency=1102.255 ms Results after this patch: 16 clients, after Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 11471098 0.012 448.281 Close 8426396 0.001 0.925 Rename 485746 0.123 267.183 Unlink 2316477 0.080 63.433 Deltree 288 2.830 11.144 Mkdir 144 0.003 0.010 Qpathinfo 10397420 0.006 10.288 Qfileinfo 1822039 0.001 0.169 Qfsinfo 1906497 0.002 14.039 Sfileinfo 934433 0.004 2.438 Find 4019879 0.026 10.200 WriteX 5718932 0.011 200.985 ReadX 17981671 0.003 10.036 LockX 37352 0.002 0.076 UnlockX 37352 0.001 0.109 Flush 804018 5.015 778.033 Throughput 1201.98 MB/sec 16 clients 16 procs max_latency=778.036 ms (+6.5% throughput, -29.4% max latency, -75.8% rename latency) Test case generic/498 from fstests tests the scenario that the previously mentioned commit fixed. Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-08-11 14:43:48 +03:00
btrfs_init_log_ctx(&ctx, &inode->vfs_inode);
ctx.logging_new_name = true;
/*
* We don't care about the return value. If we fail to log the new name
* then we know the next attempt to sync the log will fallback to a full
* transaction commit (due to a call to btrfs_set_log_full_commit()), so
* we don't need to worry about getting a log committed that has an
* inconsistent state after a rename operation.
*/
btrfs: make fast fsyncs wait only for writeback Currently regardless of a full or a fast fsync we always wait for ordered extents to complete, and then start logging the inode after that. However for fast fsyncs we can just wait for the writeback to complete, we don't need to wait for the ordered extents to complete since we use the list of modified extents maps to figure out which extents we must log and we can get their checksums directly from the ordered extents that are still in flight, otherwise look them up from the checksums tree. Until commit b5e6c3e170b770 ("btrfs: always wait on ordered extents at fsync time"), for fast fsyncs, we used to start logging without even waiting for the writeback to complete first, we would wait for it to complete after logging, while holding a transaction open, which lead to performance issues when using cgroups and probably for other cases too, as wait for IO while holding a transaction handle should be avoided as much as possible. After that, for fast fsyncs, we started to wait for ordered extents to complete before starting to log, which adds some latency to fsyncs and we even got at least one report about a performance drop which bisected to that particular change: https://lore.kernel.org/linux-btrfs/20181109215148.GF23260@techsingularity.net/ This change makes fast fsyncs only wait for writeback to finish before starting to log the inode, instead of waiting for both the writeback to finish and for the ordered extents to complete. This brings back part of the logic we had that extracts checksums from in flight ordered extents, which are not yet in the checksums tree, and making sure transaction commits wait for the completion of ordered extents previously logged (by far most of the time they have already completed by the time a transaction commit starts, resulting in no wait at all), to avoid any data loss if an ordered extent completes after the transaction used to log an inode is committed, followed by a power failure. When there are no other tasks accessing the checksums and the subvolume btrees, the ordered extent completion is pretty fast, typically taking 100 to 200 microseconds only in my observations. However when there are other tasks accessing these btrees, ordered extent completion can take a lot more time due to lock contention on nodes and leaves of these btrees. I've seen cases over 2 milliseconds, which starts to be significant. In particular when we do have concurrent fsyncs against different files there is a lot of contention on the checksums btree, since we have many tasks writing the checksums into the btree and other tasks that already started the logging phase are doing lookups for checksums in the btree. This change also turns all ranged fsyncs into full ranged fsyncs, which is something we already did when not using the NO_HOLES features or when doing a full fsync. This is to guarantee we never miss checksums due to writeback having been triggered only for a part of an extent, and we end up logging the full extent but only checksums for the written range, which results in missing checksums after log replay. Allowing ranged fsyncs to operate again only in the original range, when using the NO_HOLES feature and doing a fast fsync is doable but requires some non trivial changes to the writeback path, which can always be worked on later if needed, but I don't think they are a very common use case. Several tests were performed using fio for different numbers of concurrent jobs, each writing and fsyncing its own file, for both sequential and random file writes. The tests were run on bare metal, no virtualization, on a box with 12 cores (Intel i7-8700), 64Gb of RAM and a NVMe device, with a kernel configuration that is the default of typical distributions (debian in this case), without debug options enabled (kasan, kmemleak, slub debug, debug of page allocations, lock debugging, etc). The following script that calls fio was used: $ cat test-fsync.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/btrfs MOUNT_OPTIONS="-o ssd -o space_cache=v2" MKFS_OPTIONS="-d single -m single" if [ $# -ne 5 ]; then echo "Use $0 NUM_JOBS FILE_SIZE FSYNC_FREQ BLOCK_SIZE [write|randwrite]" exit 1 fi NUM_JOBS=$1 FILE_SIZE=$2 FSYNC_FREQ=$3 BLOCK_SIZE=$4 WRITE_MODE=$5 if [ "$WRITE_MODE" != "write" ] && [ "$WRITE_MODE" != "randwrite" ]; then echo "Invalid WRITE_MODE, must be 'write' or 'randwrite'" exit 1 fi cat <<EOF > /tmp/fio-job.ini [writers] rw=$WRITE_MODE fsync=$FSYNC_FREQ fallocate=none group_reporting=1 direct=0 bs=$BLOCK_SIZE ioengine=sync size=$FILE_SIZE directory=$MNT numjobs=$NUM_JOBS EOF echo "performance" | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor echo echo "Using config:" echo cat /tmp/fio-job.ini echo umount $MNT &> /dev/null mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT fio /tmp/fio-job.ini umount $MNT The results were the following: ************************* *** sequential writes *** ************************* ==== 1 job, 8GiB file, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=36.6MiB/s (38.4MB/s), 36.6MiB/s-36.6MiB/s (38.4MB/s-38.4MB/s), io=8192MiB (8590MB), run=223689-223689msec After patch: WRITE: bw=40.2MiB/s (42.1MB/s), 40.2MiB/s-40.2MiB/s (42.1MB/s-42.1MB/s), io=8192MiB (8590MB), run=203980-203980msec (+9.8%, -8.8% runtime) ==== 2 jobs, 4GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=35.8MiB/s (37.5MB/s), 35.8MiB/s-35.8MiB/s (37.5MB/s-37.5MB/s), io=8192MiB (8590MB), run=228950-228950msec After patch: WRITE: bw=43.5MiB/s (45.6MB/s), 43.5MiB/s-43.5MiB/s (45.6MB/s-45.6MB/s), io=8192MiB (8590MB), run=188272-188272msec (+21.5% throughput, -17.8% runtime) ==== 4 jobs, 2GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=50.1MiB/s (52.6MB/s), 50.1MiB/s-50.1MiB/s (52.6MB/s-52.6MB/s), io=8192MiB (8590MB), run=163446-163446msec After patch: WRITE: bw=64.5MiB/s (67.6MB/s), 64.5MiB/s-64.5MiB/s (67.6MB/s-67.6MB/s), io=8192MiB (8590MB), run=126987-126987msec (+28.7% throughput, -22.3% runtime) ==== 8 jobs, 1GiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=64.0MiB/s (68.1MB/s), 64.0MiB/s-64.0MiB/s (68.1MB/s-68.1MB/s), io=8192MiB (8590MB), run=126075-126075msec After patch: WRITE: bw=86.8MiB/s (91.0MB/s), 86.8MiB/s-86.8MiB/s (91.0MB/s-91.0MB/s), io=8192MiB (8590MB), run=94358-94358msec (+35.6% throughput, -25.2% runtime) ==== 16 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=79.8MiB/s (83.6MB/s), 79.8MiB/s-79.8MiB/s (83.6MB/s-83.6MB/s), io=8192MiB (8590MB), run=102694-102694msec After patch: WRITE: bw=107MiB/s (112MB/s), 107MiB/s-107MiB/s (112MB/s-112MB/s), io=8192MiB (8590MB), run=76446-76446msec (+34.1% throughput, -25.6% runtime) ==== 32 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=93.2MiB/s (97.7MB/s), 93.2MiB/s-93.2MiB/s (97.7MB/s-97.7MB/s), io=16.0GiB (17.2GB), run=175836-175836msec After patch: WRITE: bw=111MiB/s (117MB/s), 111MiB/s-111MiB/s (117MB/s-117MB/s), io=16.0GiB (17.2GB), run=147001-147001msec (+19.1% throughput, -16.4% runtime) ==== 64 jobs, 512MiB files, fsync frequency 1, block size 64KiB ==== Before patch: WRITE: bw=108MiB/s (114MB/s), 108MiB/s-108MiB/s (114MB/s-114MB/s), io=32.0GiB (34.4GB), run=302656-302656msec After patch: WRITE: bw=133MiB/s (140MB/s), 133MiB/s-133MiB/s (140MB/s-140MB/s), io=32.0GiB (34.4GB), run=246003-246003msec (+23.1% throughput, -18.7% runtime) ************************ *** random writes *** ************************ ==== 1 job, 8GiB file, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=11.5MiB/s (12.0MB/s), 11.5MiB/s-11.5MiB/s (12.0MB/s-12.0MB/s), io=8192MiB (8590MB), run=714281-714281msec After patch: WRITE: bw=11.6MiB/s (12.2MB/s), 11.6MiB/s-11.6MiB/s (12.2MB/s-12.2MB/s), io=8192MiB (8590MB), run=705959-705959msec (+0.9% throughput, -1.7% runtime) ==== 2 jobs, 4GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=12.8MiB/s (13.5MB/s), 12.8MiB/s-12.8MiB/s (13.5MB/s-13.5MB/s), io=8192MiB (8590MB), run=638101-638101msec After patch: WRITE: bw=13.1MiB/s (13.7MB/s), 13.1MiB/s-13.1MiB/s (13.7MB/s-13.7MB/s), io=8192MiB (8590MB), run=625374-625374msec (+2.3% throughput, -2.0% runtime) ==== 4 jobs, 2GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=15.4MiB/s (16.2MB/s), 15.4MiB/s-15.4MiB/s (16.2MB/s-16.2MB/s), io=8192MiB (8590MB), run=531146-531146msec After patch: WRITE: bw=17.8MiB/s (18.7MB/s), 17.8MiB/s-17.8MiB/s (18.7MB/s-18.7MB/s), io=8192MiB (8590MB), run=460431-460431msec (+15.6% throughput, -13.3% runtime) ==== 8 jobs, 1GiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=19.9MiB/s (20.8MB/s), 19.9MiB/s-19.9MiB/s (20.8MB/s-20.8MB/s), io=8192MiB (8590MB), run=412664-412664msec After patch: WRITE: bw=22.2MiB/s (23.3MB/s), 22.2MiB/s-22.2MiB/s (23.3MB/s-23.3MB/s), io=8192MiB (8590MB), run=368589-368589msec (+11.6% throughput, -10.7% runtime) ==== 16 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=29.3MiB/s (30.7MB/s), 29.3MiB/s-29.3MiB/s (30.7MB/s-30.7MB/s), io=8192MiB (8590MB), run=279924-279924msec After patch: WRITE: bw=30.4MiB/s (31.9MB/s), 30.4MiB/s-30.4MiB/s (31.9MB/s-31.9MB/s), io=8192MiB (8590MB), run=269258-269258msec (+3.8% throughput, -3.8% runtime) ==== 32 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=36.9MiB/s (38.7MB/s), 36.9MiB/s-36.9MiB/s (38.7MB/s-38.7MB/s), io=16.0GiB (17.2GB), run=443581-443581msec After patch: WRITE: bw=41.6MiB/s (43.6MB/s), 41.6MiB/s-41.6MiB/s (43.6MB/s-43.6MB/s), io=16.0GiB (17.2GB), run=394114-394114msec (+12.7% throughput, -11.2% runtime) ==== 64 jobs, 512MiB files, fsync frequency 16, block size 4KiB ==== Before patch: WRITE: bw=45.9MiB/s (48.1MB/s), 45.9MiB/s-45.9MiB/s (48.1MB/s-48.1MB/s), io=32.0GiB (34.4GB), run=714614-714614msec After patch: WRITE: bw=48.8MiB/s (51.1MB/s), 48.8MiB/s-48.8MiB/s (51.1MB/s-51.1MB/s), io=32.0GiB (34.4GB), run=672087-672087msec (+6.3% throughput, -6.0% runtime) Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2020-08-11 14:43:58 +03:00
btrfs_log_inode_parent(trans, inode, parent, LOG_INODE_EXISTS, &ctx);
btrfs: stop doing unnecessary log updates during a rename During a rename, we call __btrfs_unlink_inode(), which will call btrfs_del_inode_ref_in_log() and btrfs_del_dir_entries_in_log(), in order to remove an inode reference and a directory entry from the log. These are necessary when __btrfs_unlink_inode() is called from the unlink path, but not necessary when it's called from a rename context, because: 1) For the btrfs_del_inode_ref_in_log() call, it's pointless to delete the inode reference related to the old name, because later in the rename path we call btrfs_log_new_name(), which will drop all inode references from the log and copy all inode references from the subvolume tree to the log tree. So we are doing one unnecessary btree operation which adds additional latency and lock contention in case there are other tasks accessing the log tree; 2) For the btrfs_del_dir_entries_in_log() call, we are now doing the equivalent at btrfs_log_new_name() since the previous patch in the series, that has the subject "btrfs: avoid logging all directory changes during renames". In fact, having __btrfs_unlink_inode() call this function not only adds additional latency and lock contention due to the extra btree operation, but also can make btrfs_log_new_name() unnecessarily log a range item to track the deletion of the old name, since it has no way to known that the directory entry related to the old name was previously logged and already deleted by __btrfs_unlink_inode() through its call to btrfs_del_dir_entries_in_log(). So skip those calls at __btrfs_unlink_inode() when we are doing a rename. Skipping them also allows us now to reduce the duration of time we are pinning a log transaction during renames, which is always beneficial as it's not delaying so much other tasks trying to sync the log tree, in particular we end up not holding the log transaction pinned while adding the new name (adding inode ref, directory entry, etc). This change is part of a patchset comprised of the following patches: 1/5 btrfs: add helper to delete a dir entry from a log tree 2/5 btrfs: pass the dentry to btrfs_log_new_name() instead of the inode 3/5 btrfs: avoid logging all directory changes during renames 4/5 btrfs: stop doing unnecessary log updates during a rename 5/5 btrfs: avoid inode logging during rename and link when possible Just like the previous patch in the series, "btrfs: avoid logging all directory changes during renames", the following script mimics part of what a package installation/upgrade with zypper does, which is basically renaming a lot of files, in some directory under /usr, to a name with a suffix of "-RPMDELETE": $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=10000 mkfs.btrfs -f $DEV mount $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync # Do some change to testdir and fsync it. echo -n > $MNT/testdir/file_$((NUM_FILES + 1)) xfs_io -c "fsync" $MNT/testdir echo "Renaming $NUM_FILES files..." start=$(date +%s%N) for ((i = 1; i <= $NUM_FILES; i++)); do mv $MNT/testdir/file_$i $MNT/testdir/file_$i-RPMDELETE done end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "Renames took $dur milliseconds" umount $MNT Testing this change on box a using a non-debug kernel (Debian's default kernel config) gave the following results: NUM_FILES=10000, before patchset: 27399 ms NUM_FILES=10000, after patches 1/5 to 3/5 applied: 9093 ms (-66.8%) NUM_FILES=10000, after patches 1/5 to 4/5 applied: 9016 ms (-67.1%) NUM_FILES=5000, before patchset: 9241 ms NUM_FILES=5000, after patches 1/5 to 3/5 applied: 4642 ms (-49.8%) NUM_FILES=5000, after patches 1/5 to 4/5 applied: 4553 ms (-50.7%) NUM_FILES=2000, before patchset: 2550 ms NUM_FILES=2000, after patches 1/5 to 3/5 applied: 1788 ms (-29.9%) NUM_FILES=2000, after patches 1/5 to 4/5 applied: 1767 ms (-30.7%) NUM_FILES=1000, before patchset: 1088 ms NUM_FILES=1000, after patches 1/5 to 3/5 applied: 905 ms (-16.9%) NUM_FILES=1000, after patches 1/5 to 4/5 applied: 883 ms (-18.8%) The next patch in the series (5/5), also contains dbench results after applying to whole patchset. Link: https://bugzilla.opensuse.org/show_bug.cgi?id=1193549 Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-01-20 14:00:09 +03:00
out:
btrfs: avoid inode logging during rename and link when possible During a rename or link operation, we need to determine if an inode was previously logged or not, and if it was, do some update to the logged inode. We used to rely exclusively on the logged_trans field of struct btrfs_inode to determine that, but that was not reliable because the value of that field is not persisted in the inode item, so it's lost when an inode is evicted and loaded back again. That led to several issues in the past, such as not persisting deletions (such as the case fixed by commit 803f0f64d17769 ("Btrfs: fix fsync not persisting dentry deletions due to inode evictions")), or resulting in losing a file after an inode eviction followed by a rename (commit ecc64fab7d49c6 ("btrfs: fix lost inode on log replay after mix of fsync, rename and inode eviction")), besides other issues. So the inode_logged() helper was introduced and used to determine if an inode was possibly logged before in the current transaction, with the caveat that it could return false positives, in the sense that even if an inode was not logged before in the current transaction, it could still return true, but never to return false in case the inode was logged. >From a functional point of view that is fine, but from a performance perspective it can introduce significant latencies to rename and link operations, as they will end up doing inode logging even when it is not necessary. Recently on a 5.15 kernel, an openSUSE Tumbleweed user reported package installations and upgrades, with the zypper tool, were often taking a long time to complete. With strace it could be observed that zypper was spending about 99% of its time on rename operations, and then with further analysis we checked that directory logging was happening too frequently. Taking into account that installation/upgrade of some of the packages needed a few thousand file renames, the slowdown was very noticeable for the user. The issue was caused indirectly due to an excessive number of inode evictions on a 5.15 kernel, about 100x more compared to a 5.13, 5.14 or a 5.16-rc8 kernel. While triggering the inode evictions if something outside btrfs' control, btrfs could still behave better by eliminating the false positives from the inode_logged() helper. So change inode_logged() to actually eliminate such false positives caused by inode eviction and when an inode was never logged since the filesystem was mounted, as both cases relate to when the logged_trans field of struct btrfs_inode has a value of zero. When it can not determine if the inode was logged based only on the logged_trans value, lookup for the existence of the inode item in the log tree - if it's there then we known the inode was logged, if it's not there then it can not have been logged in the current transaction. Once we determine if the inode was logged, update the logged_trans value to avoid future calls to have to search in the log tree again. Alternatively, we could start storing logged_trans in the on disk inode item structure (struct btrfs_inode_item) in the unused space it still has, but that would be a bit odd because: 1) We only care about logged_trans since the filesystem was mounted, we don't care about its value from a previous mount. Having it persisted in the inode item structure would not make the best use of the precious unused space; 2) In order to get logged_trans persisted before inode eviction, we would have to update the delayed inode when we finish logging the inode and update its logged_trans in struct btrfs_inode, which makes it a bit cumbersome since we need to check if the delayed inode exists, if not create it and populate it and deal with any errors (-ENOMEM mostly). This change is part of a patchset comprised of the following patches: 1/5 btrfs: add helper to delete a dir entry from a log tree 2/5 btrfs: pass the dentry to btrfs_log_new_name() instead of the inode 3/5 btrfs: avoid logging all directory changes during renames 4/5 btrfs: stop doing unnecessary log updates during a rename 5/5 btrfs: avoid inode logging during rename and link when possible The following test script mimics part of what the zypper tool does during package installations/upgrades. It does not triggers inode evictions, but it's similar because it triggers false positives from the inode_logged() helper, because the inodes have a logged_trans of 0, there's a log tree due to a fsync of an unrelated file and the directory inode has its last_trans field set to the current transaction: $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=10000 mkfs.btrfs -f $DEV mount $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync # Now do some change to an unrelated file and fsync it. # This is just to create a log tree to make sure that inode_logged() # does not return false when called against "testdir". xfs_io -f -c "pwrite 0 4K" -c "fsync" $MNT/foo # Do some change to testdir. This is to make sure inode_logged() # will return true when called against "testdir", because its # logged_trans is 0, it was changed in the current transaction # and there's a log tree. echo -n > $MNT/testdir/file_$((NUM_FILES + 1)) echo "Renaming $NUM_FILES files..." start=$(date +%s%N) for ((i = 1; i <= $NUM_FILES; i++)); do mv $MNT/testdir/file_$i $MNT/testdir/file_$i-RPMDELETE done end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "Renames took $dur milliseconds" umount $MNT Testing this change on a box using a non-debug kernel (Debian's default kernel config) gave the following results: NUM_FILES=10000, before patchset: 27837 ms NUM_FILES=10000, after patches 1/5 to 4/5 applied: 9236 ms (-66.8%) NUM_FILES=10000, after whole patchset applied: 8902 ms (-68.0%) NUM_FILES=5000, before patchset: 9127 ms NUM_FILES=5000, after patches 1/5 to 4/5 applied: 4640 ms (-49.2%) NUM_FILES=5000, after whole patchset applied: 4441 ms (-51.3%) NUM_FILES=2000, before patchset: 2528 ms NUM_FILES=2000, after patches 1/5 to 4/5 applied: 1983 ms (-21.6%) NUM_FILES=2000, after whole patchset applied: 1747 ms (-30.9%) NUM_FILES=1000, before patchset: 1085 ms NUM_FILES=1000, after patches 1/5 to 4/5 applied: 893 ms (-17.7%) NUM_FILES=1000, after whole patchset applied: 867 ms (-20.1%) Running dbench on the same physical machine with the following script: $ cat run-dbench.sh #!/bin/bash NUM_JOBS=$(nproc --all) DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 MOUNT_OPTIONS="-o ssd" MKFS_OPTIONS="-O no-holes -R free-space-tree" echo "performance" | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor mkfs.btrfs -f $MKFS_OPTIONS $DEV mount $MOUNT_OPTIONS $DEV $MNT dbench -D $MNT -t 120 $NUM_JOBS umount $MNT Before patchset: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 3761352 0.032 143.843 Close 2762770 0.002 2.273 Rename 159304 0.291 67.037 Unlink 759784 0.207 143.998 Deltree 72 4.028 15.977 Mkdir 36 0.003 0.006 Qpathinfo 3409780 0.013 9.678 Qfileinfo 596772 0.001 0.878 Qfsinfo 625189 0.003 1.245 Sfileinfo 306443 0.006 1.840 Find 1318106 0.063 19.798 WriteX 1871137 0.021 8.532 ReadX 5897325 0.003 3.567 LockX 12252 0.003 0.258 UnlockX 12252 0.002 0.100 Flush 263666 3.327 155.632 Throughput 980.047 MB/sec 12 clients 12 procs max_latency=155.636 ms After whole patchset applied: Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 4195584 0.033 107.742 Close 3081932 0.002 1.935 Rename 177641 0.218 14.905 Unlink 847333 0.166 107.822 Deltree 118 5.315 15.247 Mkdir 59 0.004 0.048 Qpathinfo 3802612 0.014 10.302 Qfileinfo 666748 0.001 1.034 Qfsinfo 697329 0.003 0.944 Sfileinfo 341712 0.006 2.099 Find 1470365 0.065 9.359 WriteX 2093921 0.021 8.087 ReadX 6576234 0.003 3.407 LockX 13660 0.003 0.308 UnlockX 13660 0.002 0.114 Flush 294090 2.906 115.539 Throughput 1093.11 MB/sec 12 clients 12 procs max_latency=115.544 ms +11.5% throughput -25.8% max latency rename max latency -77.8% Link: https://bugzilla.opensuse.org/show_bug.cgi?id=1193549 Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-01-20 14:00:10 +03:00
/*
* If an error happened mark the log for a full commit because it's not
* consistent and up to date or we couldn't find out if one of the
* inodes was logged before in this transaction. Do it before unpinning
* the log, to avoid any races with someone else trying to commit it.
*/
if (ret < 0)
btrfs_set_log_full_commit(trans);
if (log_pinned)
btrfs: stop doing unnecessary log updates during a rename During a rename, we call __btrfs_unlink_inode(), which will call btrfs_del_inode_ref_in_log() and btrfs_del_dir_entries_in_log(), in order to remove an inode reference and a directory entry from the log. These are necessary when __btrfs_unlink_inode() is called from the unlink path, but not necessary when it's called from a rename context, because: 1) For the btrfs_del_inode_ref_in_log() call, it's pointless to delete the inode reference related to the old name, because later in the rename path we call btrfs_log_new_name(), which will drop all inode references from the log and copy all inode references from the subvolume tree to the log tree. So we are doing one unnecessary btree operation which adds additional latency and lock contention in case there are other tasks accessing the log tree; 2) For the btrfs_del_dir_entries_in_log() call, we are now doing the equivalent at btrfs_log_new_name() since the previous patch in the series, that has the subject "btrfs: avoid logging all directory changes during renames". In fact, having __btrfs_unlink_inode() call this function not only adds additional latency and lock contention due to the extra btree operation, but also can make btrfs_log_new_name() unnecessarily log a range item to track the deletion of the old name, since it has no way to known that the directory entry related to the old name was previously logged and already deleted by __btrfs_unlink_inode() through its call to btrfs_del_dir_entries_in_log(). So skip those calls at __btrfs_unlink_inode() when we are doing a rename. Skipping them also allows us now to reduce the duration of time we are pinning a log transaction during renames, which is always beneficial as it's not delaying so much other tasks trying to sync the log tree, in particular we end up not holding the log transaction pinned while adding the new name (adding inode ref, directory entry, etc). This change is part of a patchset comprised of the following patches: 1/5 btrfs: add helper to delete a dir entry from a log tree 2/5 btrfs: pass the dentry to btrfs_log_new_name() instead of the inode 3/5 btrfs: avoid logging all directory changes during renames 4/5 btrfs: stop doing unnecessary log updates during a rename 5/5 btrfs: avoid inode logging during rename and link when possible Just like the previous patch in the series, "btrfs: avoid logging all directory changes during renames", the following script mimics part of what a package installation/upgrade with zypper does, which is basically renaming a lot of files, in some directory under /usr, to a name with a suffix of "-RPMDELETE": $ cat test.sh #!/bin/bash DEV=/dev/nvme0n1 MNT=/mnt/nvme0n1 NUM_FILES=10000 mkfs.btrfs -f $DEV mount $DEV $MNT mkdir $MNT/testdir for ((i = 1; i <= $NUM_FILES; i++)); do echo -n > $MNT/testdir/file_$i done sync # Do some change to testdir and fsync it. echo -n > $MNT/testdir/file_$((NUM_FILES + 1)) xfs_io -c "fsync" $MNT/testdir echo "Renaming $NUM_FILES files..." start=$(date +%s%N) for ((i = 1; i <= $NUM_FILES; i++)); do mv $MNT/testdir/file_$i $MNT/testdir/file_$i-RPMDELETE done end=$(date +%s%N) dur=$(( (end - start) / 1000000 )) echo "Renames took $dur milliseconds" umount $MNT Testing this change on box a using a non-debug kernel (Debian's default kernel config) gave the following results: NUM_FILES=10000, before patchset: 27399 ms NUM_FILES=10000, after patches 1/5 to 3/5 applied: 9093 ms (-66.8%) NUM_FILES=10000, after patches 1/5 to 4/5 applied: 9016 ms (-67.1%) NUM_FILES=5000, before patchset: 9241 ms NUM_FILES=5000, after patches 1/5 to 3/5 applied: 4642 ms (-49.8%) NUM_FILES=5000, after patches 1/5 to 4/5 applied: 4553 ms (-50.7%) NUM_FILES=2000, before patchset: 2550 ms NUM_FILES=2000, after patches 1/5 to 3/5 applied: 1788 ms (-29.9%) NUM_FILES=2000, after patches 1/5 to 4/5 applied: 1767 ms (-30.7%) NUM_FILES=1000, before patchset: 1088 ms NUM_FILES=1000, after patches 1/5 to 3/5 applied: 905 ms (-16.9%) NUM_FILES=1000, after patches 1/5 to 4/5 applied: 883 ms (-18.8%) The next patch in the series (5/5), also contains dbench results after applying to whole patchset. Link: https://bugzilla.opensuse.org/show_bug.cgi?id=1193549 Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-01-20 14:00:09 +03:00
btrfs_end_log_trans(root);
Btrfs: tree logging unlink/rename fixes The tree logging code allows individual files or directories to be logged without including operations on other files and directories in the FS. It tries to commit the minimal set of changes to disk in order to fsync the single file or directory that was sent to fsync or O_SYNC. The tree logging code was allowing files and directories to be unlinked if they were part of a rename operation where only one directory in the rename was in the fsync log. This patch adds a few new rules to the tree logging. 1) on rename or unlink, if the inode being unlinked isn't in the fsync log, we must force a full commit before doing an fsync of the directory where the unlink was done. The commit isn't done during the unlink, but it is forced the next time we try to log the parent directory. Solution: record transid of last unlink/rename per directory when the directory wasn't already logged. For renames this is only done when renaming to a different directory. mkdir foo/some_dir normal commit rename foo/some_dir foo2/some_dir mkdir foo/some_dir fsync foo/some_dir/some_file The fsync above will unlink the original some_dir without recording it in its new location (foo2). After a crash, some_dir will be gone unless the fsync of some_file forces a full commit 2) we must log any new names for any file or dir that is in the fsync log. This way we make sure not to lose files that are unlinked during the same transaction. 2a) we must log any new names for any file or dir during rename when the directory they are being removed from was logged. 2a is actually the more important variant. Without the extra logging a crash might unlink the old name without recreating the new one 3) after a crash, we must go through any directories with a link count of zero and redo the rm -rf mkdir f1/foo normal commit rm -rf f1/foo fsync(f1) The directory f1 was fully removed from the FS, but fsync was never called on f1, only its parent dir. After a crash the rm -rf must be replayed. This must be able to recurse down the entire directory tree. The inode link count fixup code takes care of the ugly details. Signed-off-by: Chris Mason <chris.mason@oracle.com>
2009-03-24 17:24:20 +03:00
}