git/bulk-checkin.c

384 строки
10 KiB
C
Исходник Обычный вид История

/*
* Copyright (c) 2011, Google Inc.
*/
#include "git-compat-util.h"
#include "alloc.h"
#include "bulk-checkin.h"
#include "hex.h"
core.fsyncmethod: batched disk flushes for loose-objects When adding many objects to a repo with `core.fsync=loose-object`, the cost of fsync'ing each object file can become prohibitive. One major source of the cost of fsync is the implied flush of the hardware writeback cache within the disk drive. This commit introduces a new `core.fsyncMethod=batch` option that batches up hardware flushes. It hooks into the bulk-checkin odb-transaction functionality, takes advantage of tmp-objdir, and uses the writeout-only support code. When the new mode is enabled, we do the following for each new object: 1a. Create the object in a tmp-objdir. 2a. Issue a pagecache writeback request and wait for it to complete. At the end of the entire transaction when unplugging bulk checkin: 1b. Issue an fsync against a dummy file to flush the log and hardware writeback cache, which should by now have seen the tmp-objdir writes. 2b. Rename all of the tmp-objdir files to their final names. 3b. When updating the index and/or refs, we assume that Git will issue another fsync internal to that operation. This is not the default today, but the user now has the option of syncing the index and there is a separate patch series to implement syncing of refs. On a filesystem with a singular journal that is updated during name operations (e.g. create, link, rename, etc), such as NTFS, HFS+, or XFS we would expect the fsync to trigger a journal writeout so that this sequence is enough to ensure that the user's data is durable by the time the git command returns. This sequence also ensures that no object files appear in the main object store unless they are fsync-durable. Batch mode is only enabled if core.fsync includes loose-objects. If the legacy core.fsyncObjectFiles setting is enabled, but core.fsync does not include loose-objects, we will use file-by-file fsyncing. In step (1a) of the sequence, the tmp-objdir is created lazily to avoid work if no loose objects are ever added to the ODB. We use a tmp-objdir to maintain the invariant that no loose-objects are visible in the main ODB unless they are properly fsync-durable. This is important since future ODB operations that try to create an object with specific contents will silently drop the new data if an object with the target hash exists without checking that the loose-object contents match the hash. Only a full git-fsck would restore the ODB to a functional state where dataloss doesn't occur. In step (1b) of the sequence, we issue a fsync against a dummy file created specifically for the purpose. This method has a little higher cost than using one of the input object files, but makes adding new callers of this mechanism easier, since we don't need to figure out which object file is "last" or risk sharing violations by caching the fd of the last object file. _Performance numbers_: Linux - Hyper-V VM running Kernel 5.11 (Ubuntu 20.04) on a fast SSD. Mac - macOS 11.5.1 running on a Mac mini on a 1TB Apple SSD. Windows - Same host as Linux, a preview version of Windows 11. Adding 500 files to the repo with 'git add' Times reported in seconds. object file syncing | Linux | Mac | Windows --------------------|-------|-------|-------- disabled | 0.06 | 0.35 | 0.61 fsync | 1.88 | 11.18 | 2.47 batch | 0.15 | 0.41 | 1.53 Signed-off-by: Neeraj Singh <neerajsi@microsoft.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-04-05 08:20:09 +03:00
#include "lockfile.h"
#include "repository.h"
#include "csum-file.h"
#include "pack.h"
#include "strbuf.h"
core.fsyncmethod: batched disk flushes for loose-objects When adding many objects to a repo with `core.fsync=loose-object`, the cost of fsync'ing each object file can become prohibitive. One major source of the cost of fsync is the implied flush of the hardware writeback cache within the disk drive. This commit introduces a new `core.fsyncMethod=batch` option that batches up hardware flushes. It hooks into the bulk-checkin odb-transaction functionality, takes advantage of tmp-objdir, and uses the writeout-only support code. When the new mode is enabled, we do the following for each new object: 1a. Create the object in a tmp-objdir. 2a. Issue a pagecache writeback request and wait for it to complete. At the end of the entire transaction when unplugging bulk checkin: 1b. Issue an fsync against a dummy file to flush the log and hardware writeback cache, which should by now have seen the tmp-objdir writes. 2b. Rename all of the tmp-objdir files to their final names. 3b. When updating the index and/or refs, we assume that Git will issue another fsync internal to that operation. This is not the default today, but the user now has the option of syncing the index and there is a separate patch series to implement syncing of refs. On a filesystem with a singular journal that is updated during name operations (e.g. create, link, rename, etc), such as NTFS, HFS+, or XFS we would expect the fsync to trigger a journal writeout so that this sequence is enough to ensure that the user's data is durable by the time the git command returns. This sequence also ensures that no object files appear in the main object store unless they are fsync-durable. Batch mode is only enabled if core.fsync includes loose-objects. If the legacy core.fsyncObjectFiles setting is enabled, but core.fsync does not include loose-objects, we will use file-by-file fsyncing. In step (1a) of the sequence, the tmp-objdir is created lazily to avoid work if no loose objects are ever added to the ODB. We use a tmp-objdir to maintain the invariant that no loose-objects are visible in the main ODB unless they are properly fsync-durable. This is important since future ODB operations that try to create an object with specific contents will silently drop the new data if an object with the target hash exists without checking that the loose-object contents match the hash. Only a full git-fsck would restore the ODB to a functional state where dataloss doesn't occur. In step (1b) of the sequence, we issue a fsync against a dummy file created specifically for the purpose. This method has a little higher cost than using one of the input object files, but makes adding new callers of this mechanism easier, since we don't need to figure out which object file is "last" or risk sharing violations by caching the fd of the last object file. _Performance numbers_: Linux - Hyper-V VM running Kernel 5.11 (Ubuntu 20.04) on a fast SSD. Mac - macOS 11.5.1 running on a Mac mini on a 1TB Apple SSD. Windows - Same host as Linux, a preview version of Windows 11. Adding 500 files to the repo with 'git add' Times reported in seconds. object file syncing | Linux | Mac | Windows --------------------|-------|-------|-------- disabled | 0.06 | 0.35 | 0.61 fsync | 1.88 | 11.18 | 2.47 batch | 0.15 | 0.41 | 1.53 Signed-off-by: Neeraj Singh <neerajsi@microsoft.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-04-05 08:20:09 +03:00
#include "string-list.h"
#include "tmp-objdir.h"
#include "packfile.h"
#include "object-store.h"
static int odb_transaction_nesting;
core.fsyncmethod: batched disk flushes for loose-objects When adding many objects to a repo with `core.fsync=loose-object`, the cost of fsync'ing each object file can become prohibitive. One major source of the cost of fsync is the implied flush of the hardware writeback cache within the disk drive. This commit introduces a new `core.fsyncMethod=batch` option that batches up hardware flushes. It hooks into the bulk-checkin odb-transaction functionality, takes advantage of tmp-objdir, and uses the writeout-only support code. When the new mode is enabled, we do the following for each new object: 1a. Create the object in a tmp-objdir. 2a. Issue a pagecache writeback request and wait for it to complete. At the end of the entire transaction when unplugging bulk checkin: 1b. Issue an fsync against a dummy file to flush the log and hardware writeback cache, which should by now have seen the tmp-objdir writes. 2b. Rename all of the tmp-objdir files to their final names. 3b. When updating the index and/or refs, we assume that Git will issue another fsync internal to that operation. This is not the default today, but the user now has the option of syncing the index and there is a separate patch series to implement syncing of refs. On a filesystem with a singular journal that is updated during name operations (e.g. create, link, rename, etc), such as NTFS, HFS+, or XFS we would expect the fsync to trigger a journal writeout so that this sequence is enough to ensure that the user's data is durable by the time the git command returns. This sequence also ensures that no object files appear in the main object store unless they are fsync-durable. Batch mode is only enabled if core.fsync includes loose-objects. If the legacy core.fsyncObjectFiles setting is enabled, but core.fsync does not include loose-objects, we will use file-by-file fsyncing. In step (1a) of the sequence, the tmp-objdir is created lazily to avoid work if no loose objects are ever added to the ODB. We use a tmp-objdir to maintain the invariant that no loose-objects are visible in the main ODB unless they are properly fsync-durable. This is important since future ODB operations that try to create an object with specific contents will silently drop the new data if an object with the target hash exists without checking that the loose-object contents match the hash. Only a full git-fsck would restore the ODB to a functional state where dataloss doesn't occur. In step (1b) of the sequence, we issue a fsync against a dummy file created specifically for the purpose. This method has a little higher cost than using one of the input object files, but makes adding new callers of this mechanism easier, since we don't need to figure out which object file is "last" or risk sharing violations by caching the fd of the last object file. _Performance numbers_: Linux - Hyper-V VM running Kernel 5.11 (Ubuntu 20.04) on a fast SSD. Mac - macOS 11.5.1 running on a Mac mini on a 1TB Apple SSD. Windows - Same host as Linux, a preview version of Windows 11. Adding 500 files to the repo with 'git add' Times reported in seconds. object file syncing | Linux | Mac | Windows --------------------|-------|-------|-------- disabled | 0.06 | 0.35 | 0.61 fsync | 1.88 | 11.18 | 2.47 batch | 0.15 | 0.41 | 1.53 Signed-off-by: Neeraj Singh <neerajsi@microsoft.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-04-05 08:20:09 +03:00
static struct tmp_objdir *bulk_fsync_objdir;
bulk-checkin: rename 'state' variable and separate 'plugged' boolean This commit prepares for adding batch-fsync to the bulk-checkin infrastructure. The bulk-checkin infrastructure is currently used to batch up addition of large blobs to a packfile. When a blob is larger than big_file_threshold, we unconditionally add it to a pack. If bulk checkins are 'plugged', we allow multiple large blobs to be added to a single pack until we reach the packfile size limit; otherwise, we simply make a new packfile for each large blob. The 'unplug' call tells us when the series of blob additions is done so that we can finish the packfiles and make their objects available to subsequent operations. Stated another way, bulk-checkin allows callers to define a transaction that adds multiple objects to the object database, where the object database can optimize its internal operations within the transaction boundary. Batched fsync will fit into bulk-checkin by taking advantage of the plug/unplug functionality to determine the appropriate time to fsync and make newly-added objects available in the primary object database. * Rename 'state' variable to 'bulk_checkin_packfile', since we will later be adding 'bulk_fsync_objdir'. This also makes the variable easier to find in the debugger, since the name is more unique. * Rename finish_bulk_checkin to flush_bulk_checkin_packfile and call it unconditionally from unplug_bulk_checkin. Internally it will conditionally do a flush if there's any work to do. * Move the 'plugged' data member of 'bulk_checkin_state' into a separate static variable. Doing this avoids resetting the variable in finish_bulk_checkin when zeroing the 'bulk_checkin_state'. As-is, we seem to unintentionally disable the plugging functionality the first time a new packfile must be created due to packfile size limits. While disabling the plugging state only results in suboptimal behavior for the current code, it would be fatal for the bulk-fsync functionality later in this patch series. The net effect of these changes is to make a clear separation between the portion of the bulk-checkin infrastructure that is related to the packfile (nearly all of it at present) and the part that is related to other future optimizations of the ODB. Signed-off-by: Neeraj Singh <neerajsi@microsoft.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-04-05 08:20:07 +03:00
static struct bulk_checkin_packfile {
char *pack_tmp_name;
struct hashfile *f;
off_t offset;
struct pack_idx_option pack_idx_opts;
struct pack_idx_entry **written;
uint32_t alloc_written;
uint32_t nr_written;
bulk-checkin: rename 'state' variable and separate 'plugged' boolean This commit prepares for adding batch-fsync to the bulk-checkin infrastructure. The bulk-checkin infrastructure is currently used to batch up addition of large blobs to a packfile. When a blob is larger than big_file_threshold, we unconditionally add it to a pack. If bulk checkins are 'plugged', we allow multiple large blobs to be added to a single pack until we reach the packfile size limit; otherwise, we simply make a new packfile for each large blob. The 'unplug' call tells us when the series of blob additions is done so that we can finish the packfiles and make their objects available to subsequent operations. Stated another way, bulk-checkin allows callers to define a transaction that adds multiple objects to the object database, where the object database can optimize its internal operations within the transaction boundary. Batched fsync will fit into bulk-checkin by taking advantage of the plug/unplug functionality to determine the appropriate time to fsync and make newly-added objects available in the primary object database. * Rename 'state' variable to 'bulk_checkin_packfile', since we will later be adding 'bulk_fsync_objdir'. This also makes the variable easier to find in the debugger, since the name is more unique. * Rename finish_bulk_checkin to flush_bulk_checkin_packfile and call it unconditionally from unplug_bulk_checkin. Internally it will conditionally do a flush if there's any work to do. * Move the 'plugged' data member of 'bulk_checkin_state' into a separate static variable. Doing this avoids resetting the variable in finish_bulk_checkin when zeroing the 'bulk_checkin_state'. As-is, we seem to unintentionally disable the plugging functionality the first time a new packfile must be created due to packfile size limits. While disabling the plugging state only results in suboptimal behavior for the current code, it would be fatal for the bulk-fsync functionality later in this patch series. The net effect of these changes is to make a clear separation between the portion of the bulk-checkin infrastructure that is related to the packfile (nearly all of it at present) and the part that is related to other future optimizations of the ODB. Signed-off-by: Neeraj Singh <neerajsi@microsoft.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-04-05 08:20:07 +03:00
} bulk_checkin_packfile;
static void finish_tmp_packfile(struct strbuf *basename,
const char *pack_tmp_name,
struct pack_idx_entry **written_list,
uint32_t nr_written,
struct pack_idx_option *pack_idx_opts,
unsigned char hash[])
{
char *idx_tmp_name = NULL;
stage_tmp_packfiles(basename, pack_tmp_name, written_list, nr_written,
NULL, pack_idx_opts, hash, &idx_tmp_name);
rename_tmp_packfile_idx(basename, &idx_tmp_name);
free(idx_tmp_name);
}
bulk-checkin: rename 'state' variable and separate 'plugged' boolean This commit prepares for adding batch-fsync to the bulk-checkin infrastructure. The bulk-checkin infrastructure is currently used to batch up addition of large blobs to a packfile. When a blob is larger than big_file_threshold, we unconditionally add it to a pack. If bulk checkins are 'plugged', we allow multiple large blobs to be added to a single pack until we reach the packfile size limit; otherwise, we simply make a new packfile for each large blob. The 'unplug' call tells us when the series of blob additions is done so that we can finish the packfiles and make their objects available to subsequent operations. Stated another way, bulk-checkin allows callers to define a transaction that adds multiple objects to the object database, where the object database can optimize its internal operations within the transaction boundary. Batched fsync will fit into bulk-checkin by taking advantage of the plug/unplug functionality to determine the appropriate time to fsync and make newly-added objects available in the primary object database. * Rename 'state' variable to 'bulk_checkin_packfile', since we will later be adding 'bulk_fsync_objdir'. This also makes the variable easier to find in the debugger, since the name is more unique. * Rename finish_bulk_checkin to flush_bulk_checkin_packfile and call it unconditionally from unplug_bulk_checkin. Internally it will conditionally do a flush if there's any work to do. * Move the 'plugged' data member of 'bulk_checkin_state' into a separate static variable. Doing this avoids resetting the variable in finish_bulk_checkin when zeroing the 'bulk_checkin_state'. As-is, we seem to unintentionally disable the plugging functionality the first time a new packfile must be created due to packfile size limits. While disabling the plugging state only results in suboptimal behavior for the current code, it would be fatal for the bulk-fsync functionality later in this patch series. The net effect of these changes is to make a clear separation between the portion of the bulk-checkin infrastructure that is related to the packfile (nearly all of it at present) and the part that is related to other future optimizations of the ODB. Signed-off-by: Neeraj Singh <neerajsi@microsoft.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-04-05 08:20:07 +03:00
static void flush_bulk_checkin_packfile(struct bulk_checkin_packfile *state)
{
unsigned char hash[GIT_MAX_RAWSZ];
struct strbuf packname = STRBUF_INIT;
int i;
if (!state->f)
return;
if (state->nr_written == 0) {
close(state->f->fd);
unlink(state->pack_tmp_name);
goto clear_exit;
} else if (state->nr_written == 1) {
finalize_hashfile(state->f, hash, FSYNC_COMPONENT_PACK,
CSUM_HASH_IN_STREAM | CSUM_FSYNC | CSUM_CLOSE);
} else {
int fd = finalize_hashfile(state->f, hash, FSYNC_COMPONENT_PACK, 0);
fixup_pack_header_footer(fd, hash, state->pack_tmp_name,
state->nr_written, hash,
state->offset);
close(fd);
}
strbuf_addf(&packname, "%s/pack/pack-%s.", get_object_directory(),
hash_to_hex(hash));
finish_tmp_packfile(&packname, state->pack_tmp_name,
state->written, state->nr_written,
&state->pack_idx_opts, hash);
for (i = 0; i < state->nr_written; i++)
free(state->written[i]);
clear_exit:
free(state->written);
memset(state, 0, sizeof(*state));
strbuf_release(&packname);
/* Make objects we just wrote available to ourselves */
reprepare_packed_git(the_repository);
}
core.fsyncmethod: batched disk flushes for loose-objects When adding many objects to a repo with `core.fsync=loose-object`, the cost of fsync'ing each object file can become prohibitive. One major source of the cost of fsync is the implied flush of the hardware writeback cache within the disk drive. This commit introduces a new `core.fsyncMethod=batch` option that batches up hardware flushes. It hooks into the bulk-checkin odb-transaction functionality, takes advantage of tmp-objdir, and uses the writeout-only support code. When the new mode is enabled, we do the following for each new object: 1a. Create the object in a tmp-objdir. 2a. Issue a pagecache writeback request and wait for it to complete. At the end of the entire transaction when unplugging bulk checkin: 1b. Issue an fsync against a dummy file to flush the log and hardware writeback cache, which should by now have seen the tmp-objdir writes. 2b. Rename all of the tmp-objdir files to their final names. 3b. When updating the index and/or refs, we assume that Git will issue another fsync internal to that operation. This is not the default today, but the user now has the option of syncing the index and there is a separate patch series to implement syncing of refs. On a filesystem with a singular journal that is updated during name operations (e.g. create, link, rename, etc), such as NTFS, HFS+, or XFS we would expect the fsync to trigger a journal writeout so that this sequence is enough to ensure that the user's data is durable by the time the git command returns. This sequence also ensures that no object files appear in the main object store unless they are fsync-durable. Batch mode is only enabled if core.fsync includes loose-objects. If the legacy core.fsyncObjectFiles setting is enabled, but core.fsync does not include loose-objects, we will use file-by-file fsyncing. In step (1a) of the sequence, the tmp-objdir is created lazily to avoid work if no loose objects are ever added to the ODB. We use a tmp-objdir to maintain the invariant that no loose-objects are visible in the main ODB unless they are properly fsync-durable. This is important since future ODB operations that try to create an object with specific contents will silently drop the new data if an object with the target hash exists without checking that the loose-object contents match the hash. Only a full git-fsck would restore the ODB to a functional state where dataloss doesn't occur. In step (1b) of the sequence, we issue a fsync against a dummy file created specifically for the purpose. This method has a little higher cost than using one of the input object files, but makes adding new callers of this mechanism easier, since we don't need to figure out which object file is "last" or risk sharing violations by caching the fd of the last object file. _Performance numbers_: Linux - Hyper-V VM running Kernel 5.11 (Ubuntu 20.04) on a fast SSD. Mac - macOS 11.5.1 running on a Mac mini on a 1TB Apple SSD. Windows - Same host as Linux, a preview version of Windows 11. Adding 500 files to the repo with 'git add' Times reported in seconds. object file syncing | Linux | Mac | Windows --------------------|-------|-------|-------- disabled | 0.06 | 0.35 | 0.61 fsync | 1.88 | 11.18 | 2.47 batch | 0.15 | 0.41 | 1.53 Signed-off-by: Neeraj Singh <neerajsi@microsoft.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-04-05 08:20:09 +03:00
/*
* Cleanup after batch-mode fsync_object_files.
*/
static void flush_batch_fsync(void)
{
struct strbuf temp_path = STRBUF_INIT;
struct tempfile *temp;
if (!bulk_fsync_objdir)
return;
/*
* Issue a full hardware flush against a temporary file to ensure
* that all objects are durable before any renames occur. The code in
* fsync_loose_object_bulk_checkin has already issued a writeout
* request, but it has not flushed any writeback cache in the storage
* hardware or any filesystem logs. This fsync call acts as a barrier
* to ensure that the data in each new object file is durable before
* the final name is visible.
*/
strbuf_addf(&temp_path, "%s/bulk_fsync_XXXXXX", get_object_directory());
temp = xmks_tempfile(temp_path.buf);
fsync_or_die(get_tempfile_fd(temp), get_tempfile_path(temp));
delete_tempfile(&temp);
strbuf_release(&temp_path);
/*
* Make the object files visible in the primary ODB after their data is
* fully durable.
*/
tmp_objdir_migrate(bulk_fsync_objdir);
bulk_fsync_objdir = NULL;
}
bulk-checkin: rename 'state' variable and separate 'plugged' boolean This commit prepares for adding batch-fsync to the bulk-checkin infrastructure. The bulk-checkin infrastructure is currently used to batch up addition of large blobs to a packfile. When a blob is larger than big_file_threshold, we unconditionally add it to a pack. If bulk checkins are 'plugged', we allow multiple large blobs to be added to a single pack until we reach the packfile size limit; otherwise, we simply make a new packfile for each large blob. The 'unplug' call tells us when the series of blob additions is done so that we can finish the packfiles and make their objects available to subsequent operations. Stated another way, bulk-checkin allows callers to define a transaction that adds multiple objects to the object database, where the object database can optimize its internal operations within the transaction boundary. Batched fsync will fit into bulk-checkin by taking advantage of the plug/unplug functionality to determine the appropriate time to fsync and make newly-added objects available in the primary object database. * Rename 'state' variable to 'bulk_checkin_packfile', since we will later be adding 'bulk_fsync_objdir'. This also makes the variable easier to find in the debugger, since the name is more unique. * Rename finish_bulk_checkin to flush_bulk_checkin_packfile and call it unconditionally from unplug_bulk_checkin. Internally it will conditionally do a flush if there's any work to do. * Move the 'plugged' data member of 'bulk_checkin_state' into a separate static variable. Doing this avoids resetting the variable in finish_bulk_checkin when zeroing the 'bulk_checkin_state'. As-is, we seem to unintentionally disable the plugging functionality the first time a new packfile must be created due to packfile size limits. While disabling the plugging state only results in suboptimal behavior for the current code, it would be fatal for the bulk-fsync functionality later in this patch series. The net effect of these changes is to make a clear separation between the portion of the bulk-checkin infrastructure that is related to the packfile (nearly all of it at present) and the part that is related to other future optimizations of the ODB. Signed-off-by: Neeraj Singh <neerajsi@microsoft.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-04-05 08:20:07 +03:00
static int already_written(struct bulk_checkin_packfile *state, struct object_id *oid)
{
int i;
/* The object may already exist in the repository */
if (has_object_file(oid))
return 1;
/* Might want to keep the list sorted */
for (i = 0; i < state->nr_written; i++)
if (oideq(&state->written[i]->oid, oid))
return 1;
/* This is a new object we need to keep */
return 0;
}
/*
* Read the contents from fd for size bytes, streaming it to the
* packfile in state while updating the hash in ctx. Signal a failure
* by returning a negative value when the resulting pack would exceed
* the pack size limit and this is not the first object in the pack,
* so that the caller can discard what we wrote from the current pack
* by truncating it and opening a new one. The caller will then call
* us again after rewinding the input fd.
*
* The already_hashed_to pointer is kept untouched by the caller to
* make sure we do not hash the same byte when we are called
* again. This way, the caller does not have to checkpoint its hash
* status before calling us just in case we ask it to call us again
* with a new pack.
*/
bulk-checkin: rename 'state' variable and separate 'plugged' boolean This commit prepares for adding batch-fsync to the bulk-checkin infrastructure. The bulk-checkin infrastructure is currently used to batch up addition of large blobs to a packfile. When a blob is larger than big_file_threshold, we unconditionally add it to a pack. If bulk checkins are 'plugged', we allow multiple large blobs to be added to a single pack until we reach the packfile size limit; otherwise, we simply make a new packfile for each large blob. The 'unplug' call tells us when the series of blob additions is done so that we can finish the packfiles and make their objects available to subsequent operations. Stated another way, bulk-checkin allows callers to define a transaction that adds multiple objects to the object database, where the object database can optimize its internal operations within the transaction boundary. Batched fsync will fit into bulk-checkin by taking advantage of the plug/unplug functionality to determine the appropriate time to fsync and make newly-added objects available in the primary object database. * Rename 'state' variable to 'bulk_checkin_packfile', since we will later be adding 'bulk_fsync_objdir'. This also makes the variable easier to find in the debugger, since the name is more unique. * Rename finish_bulk_checkin to flush_bulk_checkin_packfile and call it unconditionally from unplug_bulk_checkin. Internally it will conditionally do a flush if there's any work to do. * Move the 'plugged' data member of 'bulk_checkin_state' into a separate static variable. Doing this avoids resetting the variable in finish_bulk_checkin when zeroing the 'bulk_checkin_state'. As-is, we seem to unintentionally disable the plugging functionality the first time a new packfile must be created due to packfile size limits. While disabling the plugging state only results in suboptimal behavior for the current code, it would be fatal for the bulk-fsync functionality later in this patch series. The net effect of these changes is to make a clear separation between the portion of the bulk-checkin infrastructure that is related to the packfile (nearly all of it at present) and the part that is related to other future optimizations of the ODB. Signed-off-by: Neeraj Singh <neerajsi@microsoft.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-04-05 08:20:07 +03:00
static int stream_to_pack(struct bulk_checkin_packfile *state,
git_hash_ctx *ctx, off_t *already_hashed_to,
int fd, size_t size, enum object_type type,
const char *path, unsigned flags)
{
git_zstream s;
bulk-checkin: make buffer reuse more obvious and safer ibuf can be reused for multiple iterations of the loop. Specifically: deflate() overwrites s.avail_in to show how much of the input buffer has not been processed yet - and sometimes leaves 'avail_in > 0', in which case ibuf will be processed again during the loop's subsequent iteration. But if we declare ibuf within the loop, then (in theory) we get a new (and uninitialised) buffer for every iteration. In practice, my compiler seems to resue the same buffer - meaning that this code does work - but it doesn't seem safe to rely on this behaviour. MSAN correctly catches this issue - as soon as we hit the 's.avail_in > 0' condition, we end up reading from what seems to be uninitialised memory. Therefore, we move ibuf out of the loop, making this reuse safe. See MSAN output from t1050-large below - the interesting part is the ibuf creation at the end, although there's a lot of indirection before we reach the read from unitialised memory: ==11294==WARNING: MemorySanitizer: use-of-uninitialized-value #0 0x7f75db58fb1c in crc32_little crc32.c:283:9 #1 0x7f75db58d5b3 in crc32_z crc32.c:220:20 #2 0x7f75db59668c in crc32 crc32.c:242:12 #3 0x8c94f8 in hashwrite csum-file.c:101:15 #4 0x825faf in stream_to_pack bulk-checkin.c:154:5 #5 0x82467b in deflate_to_pack bulk-checkin.c:225:8 #6 0x823ff1 in index_bulk_checkin bulk-checkin.c:264:15 #7 0xa7cff2 in index_stream object-file.c:2234:9 #8 0xa7bff7 in index_fd object-file.c:2256:9 #9 0xa7d22d in index_path object-file.c:2274:7 #10 0xb3c8c9 in add_to_index read-cache.c:802:7 #11 0xb3e039 in add_file_to_index read-cache.c:835:9 #12 0x4a99c3 in add_files add.c:458:7 #13 0x4a7276 in cmd_add add.c:670:18 #14 0x4a1e76 in run_builtin git.c:461:11 #15 0x49e1e7 in handle_builtin git.c:714:3 #16 0x4a0c08 in run_argv git.c:781:4 #17 0x49d5a8 in cmd_main git.c:912:19 #18 0x7974da in main common-main.c:52:11 #19 0x7f75da66f349 in __libc_start_main (/lib64/libc.so.6+0x24349) #20 0x421bd9 in _start start.S:120 Uninitialized value was stored to memory at #0 0x7f75db58fa6b in crc32_little crc32.c:283:9 #1 0x7f75db58d5b3 in crc32_z crc32.c:220:20 #2 0x7f75db59668c in crc32 crc32.c:242:12 #3 0x8c94f8 in hashwrite csum-file.c:101:15 #4 0x825faf in stream_to_pack bulk-checkin.c:154:5 #5 0x82467b in deflate_to_pack bulk-checkin.c:225:8 #6 0x823ff1 in index_bulk_checkin bulk-checkin.c:264:15 #7 0xa7cff2 in index_stream object-file.c:2234:9 #8 0xa7bff7 in index_fd object-file.c:2256:9 #9 0xa7d22d in index_path object-file.c:2274:7 #10 0xb3c8c9 in add_to_index read-cache.c:802:7 #11 0xb3e039 in add_file_to_index read-cache.c:835:9 #12 0x4a99c3 in add_files add.c:458:7 #13 0x4a7276 in cmd_add add.c:670:18 #14 0x4a1e76 in run_builtin git.c:461:11 #15 0x49e1e7 in handle_builtin git.c:714:3 #16 0x4a0c08 in run_argv git.c:781:4 #17 0x49d5a8 in cmd_main git.c:912:19 #18 0x7974da in main common-main.c:52:11 #19 0x7f75da66f349 in __libc_start_main (/lib64/libc.so.6+0x24349) Uninitialized value was stored to memory at #0 0x447eb9 in __msan_memcpy msan_interceptors.cpp:1558:3 #1 0x7f75db5c2011 in flush_pending deflate.c:746:5 #2 0x7f75db5cafa0 in deflate_stored deflate.c:1815:9 #3 0x7f75db5bb7d2 in deflate deflate.c:1005:34 #4 0xd80b7f in git_deflate zlib.c:244:12 #5 0x825dff in stream_to_pack bulk-checkin.c:140:12 #6 0x82467b in deflate_to_pack bulk-checkin.c:225:8 #7 0x823ff1 in index_bulk_checkin bulk-checkin.c:264:15 #8 0xa7cff2 in index_stream object-file.c:2234:9 #9 0xa7bff7 in index_fd object-file.c:2256:9 #10 0xa7d22d in index_path object-file.c:2274:7 #11 0xb3c8c9 in add_to_index read-cache.c:802:7 #12 0xb3e039 in add_file_to_index read-cache.c:835:9 #13 0x4a99c3 in add_files add.c:458:7 #14 0x4a7276 in cmd_add add.c:670:18 #15 0x4a1e76 in run_builtin git.c:461:11 #16 0x49e1e7 in handle_builtin git.c:714:3 #17 0x4a0c08 in run_argv git.c:781:4 #18 0x49d5a8 in cmd_main git.c:912:19 #19 0x7974da in main common-main.c:52:11 Uninitialized value was stored to memory at #0 0x447eb9 in __msan_memcpy msan_interceptors.cpp:1558:3 #1 0x7f75db644241 in _tr_stored_block trees.c:873:5 #2 0x7f75db5cad7c in deflate_stored deflate.c:1813:9 #3 0x7f75db5bb7d2 in deflate deflate.c:1005:34 #4 0xd80b7f in git_deflate zlib.c:244:12 #5 0x825dff in stream_to_pack bulk-checkin.c:140:12 #6 0x82467b in deflate_to_pack bulk-checkin.c:225:8 #7 0x823ff1 in index_bulk_checkin bulk-checkin.c:264:15 #8 0xa7cff2 in index_stream object-file.c:2234:9 #9 0xa7bff7 in index_fd object-file.c:2256:9 #10 0xa7d22d in index_path object-file.c:2274:7 #11 0xb3c8c9 in add_to_index read-cache.c:802:7 #12 0xb3e039 in add_file_to_index read-cache.c:835:9 #13 0x4a99c3 in add_files add.c:458:7 #14 0x4a7276 in cmd_add add.c:670:18 #15 0x4a1e76 in run_builtin git.c:461:11 #16 0x49e1e7 in handle_builtin git.c:714:3 #17 0x4a0c08 in run_argv git.c:781:4 #18 0x49d5a8 in cmd_main git.c:912:19 #19 0x7974da in main common-main.c:52:11 Uninitialized value was stored to memory at #0 0x447eb9 in __msan_memcpy msan_interceptors.cpp:1558:3 #1 0x7f75db5c8fcf in deflate_stored deflate.c:1783:9 #2 0x7f75db5bb7d2 in deflate deflate.c:1005:34 #3 0xd80b7f in git_deflate zlib.c:244:12 #4 0x825dff in stream_to_pack bulk-checkin.c:140:12 #5 0x82467b in deflate_to_pack bulk-checkin.c:225:8 #6 0x823ff1 in index_bulk_checkin bulk-checkin.c:264:15 #7 0xa7cff2 in index_stream object-file.c:2234:9 #8 0xa7bff7 in index_fd object-file.c:2256:9 #9 0xa7d22d in index_path object-file.c:2274:7 #10 0xb3c8c9 in add_to_index read-cache.c:802:7 #11 0xb3e039 in add_file_to_index read-cache.c:835:9 #12 0x4a99c3 in add_files add.c:458:7 #13 0x4a7276 in cmd_add add.c:670:18 #14 0x4a1e76 in run_builtin git.c:461:11 #15 0x49e1e7 in handle_builtin git.c:714:3 #16 0x4a0c08 in run_argv git.c:781:4 #17 0x49d5a8 in cmd_main git.c:912:19 #18 0x7974da in main common-main.c:52:11 #19 0x7f75da66f349 in __libc_start_main (/lib64/libc.so.6+0x24349) Uninitialized value was stored to memory at #0 0x447eb9 in __msan_memcpy msan_interceptors.cpp:1558:3 #1 0x7f75db5ea545 in read_buf deflate.c:1181:5 #2 0x7f75db5c97f7 in deflate_stored deflate.c:1791:9 #3 0x7f75db5bb7d2 in deflate deflate.c:1005:34 #4 0xd80b7f in git_deflate zlib.c:244:12 #5 0x825dff in stream_to_pack bulk-checkin.c:140:12 #6 0x82467b in deflate_to_pack bulk-checkin.c:225:8 #7 0x823ff1 in index_bulk_checkin bulk-checkin.c:264:15 #8 0xa7cff2 in index_stream object-file.c:2234:9 #9 0xa7bff7 in index_fd object-file.c:2256:9 #10 0xa7d22d in index_path object-file.c:2274:7 #11 0xb3c8c9 in add_to_index read-cache.c:802:7 #12 0xb3e039 in add_file_to_index read-cache.c:835:9 #13 0x4a99c3 in add_files add.c:458:7 #14 0x4a7276 in cmd_add add.c:670:18 #15 0x4a1e76 in run_builtin git.c:461:11 #16 0x49e1e7 in handle_builtin git.c:714:3 #17 0x4a0c08 in run_argv git.c:781:4 #18 0x49d5a8 in cmd_main git.c:912:19 #19 0x7974da in main common-main.c:52:11 Uninitialized value was created by an allocation of 'ibuf' in the stack frame of function 'stream_to_pack' #0 0x825710 in stream_to_pack bulk-checkin.c:101 SUMMARY: MemorySanitizer: use-of-uninitialized-value crc32.c:283:9 in crc32_little Exiting Signed-off-by: Andrzej Hunt <andrzej@ahunt.org> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-06-10 19:48:30 +03:00
unsigned char ibuf[16384];
unsigned char obuf[16384];
unsigned hdrlen;
int status = Z_OK;
int write_object = (flags & HASH_WRITE_OBJECT);
off_t offset = 0;
git_deflate_init(&s, pack_compression_level);
hdrlen = encode_in_pack_object_header(obuf, sizeof(obuf), type, size);
s.next_out = obuf + hdrlen;
s.avail_out = sizeof(obuf) - hdrlen;
while (status != Z_STREAM_END) {
if (size && !s.avail_in) {
ssize_t rsize = size < sizeof(ibuf) ? size : sizeof(ibuf);
ssize_t read_result = read_in_full(fd, ibuf, rsize);
if (read_result < 0)
die_errno("failed to read from '%s'", path);
if (read_result != rsize)
die("failed to read %d bytes from '%s'",
(int)rsize, path);
offset += rsize;
if (*already_hashed_to < offset) {
size_t hsize = offset - *already_hashed_to;
if (rsize < hsize)
hsize = rsize;
if (hsize)
the_hash_algo->update_fn(ctx, ibuf, hsize);
*already_hashed_to = offset;
}
s.next_in = ibuf;
s.avail_in = rsize;
size -= rsize;
}
status = git_deflate(&s, size ? 0 : Z_FINISH);
if (!s.avail_out || status == Z_STREAM_END) {
if (write_object) {
size_t written = s.next_out - obuf;
/* would we bust the size limit? */
if (state->nr_written &&
pack_size_limit_cfg &&
pack_size_limit_cfg < state->offset + written) {
git_deflate_abort(&s);
return -1;
}
hashwrite(state->f, obuf, written);
state->offset += written;
}
s.next_out = obuf;
s.avail_out = sizeof(obuf);
}
switch (status) {
case Z_OK:
case Z_BUF_ERROR:
case Z_STREAM_END:
continue;
default:
die("unexpected deflate failure: %d", status);
}
}
git_deflate_end(&s);
return 0;
}
/* Lazily create backing packfile for the state */
bulk-checkin: rename 'state' variable and separate 'plugged' boolean This commit prepares for adding batch-fsync to the bulk-checkin infrastructure. The bulk-checkin infrastructure is currently used to batch up addition of large blobs to a packfile. When a blob is larger than big_file_threshold, we unconditionally add it to a pack. If bulk checkins are 'plugged', we allow multiple large blobs to be added to a single pack until we reach the packfile size limit; otherwise, we simply make a new packfile for each large blob. The 'unplug' call tells us when the series of blob additions is done so that we can finish the packfiles and make their objects available to subsequent operations. Stated another way, bulk-checkin allows callers to define a transaction that adds multiple objects to the object database, where the object database can optimize its internal operations within the transaction boundary. Batched fsync will fit into bulk-checkin by taking advantage of the plug/unplug functionality to determine the appropriate time to fsync and make newly-added objects available in the primary object database. * Rename 'state' variable to 'bulk_checkin_packfile', since we will later be adding 'bulk_fsync_objdir'. This also makes the variable easier to find in the debugger, since the name is more unique. * Rename finish_bulk_checkin to flush_bulk_checkin_packfile and call it unconditionally from unplug_bulk_checkin. Internally it will conditionally do a flush if there's any work to do. * Move the 'plugged' data member of 'bulk_checkin_state' into a separate static variable. Doing this avoids resetting the variable in finish_bulk_checkin when zeroing the 'bulk_checkin_state'. As-is, we seem to unintentionally disable the plugging functionality the first time a new packfile must be created due to packfile size limits. While disabling the plugging state only results in suboptimal behavior for the current code, it would be fatal for the bulk-fsync functionality later in this patch series. The net effect of these changes is to make a clear separation between the portion of the bulk-checkin infrastructure that is related to the packfile (nearly all of it at present) and the part that is related to other future optimizations of the ODB. Signed-off-by: Neeraj Singh <neerajsi@microsoft.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-04-05 08:20:07 +03:00
static void prepare_to_stream(struct bulk_checkin_packfile *state,
unsigned flags)
{
if (!(flags & HASH_WRITE_OBJECT) || state->f)
return;
state->f = create_tmp_packfile(&state->pack_tmp_name);
reset_pack_idx_option(&state->pack_idx_opts);
/* Pretend we are going to write only one object */
state->offset = write_pack_header(state->f, 1);
if (!state->offset)
die_errno("unable to write pack header");
}
bulk-checkin: rename 'state' variable and separate 'plugged' boolean This commit prepares for adding batch-fsync to the bulk-checkin infrastructure. The bulk-checkin infrastructure is currently used to batch up addition of large blobs to a packfile. When a blob is larger than big_file_threshold, we unconditionally add it to a pack. If bulk checkins are 'plugged', we allow multiple large blobs to be added to a single pack until we reach the packfile size limit; otherwise, we simply make a new packfile for each large blob. The 'unplug' call tells us when the series of blob additions is done so that we can finish the packfiles and make their objects available to subsequent operations. Stated another way, bulk-checkin allows callers to define a transaction that adds multiple objects to the object database, where the object database can optimize its internal operations within the transaction boundary. Batched fsync will fit into bulk-checkin by taking advantage of the plug/unplug functionality to determine the appropriate time to fsync and make newly-added objects available in the primary object database. * Rename 'state' variable to 'bulk_checkin_packfile', since we will later be adding 'bulk_fsync_objdir'. This also makes the variable easier to find in the debugger, since the name is more unique. * Rename finish_bulk_checkin to flush_bulk_checkin_packfile and call it unconditionally from unplug_bulk_checkin. Internally it will conditionally do a flush if there's any work to do. * Move the 'plugged' data member of 'bulk_checkin_state' into a separate static variable. Doing this avoids resetting the variable in finish_bulk_checkin when zeroing the 'bulk_checkin_state'. As-is, we seem to unintentionally disable the plugging functionality the first time a new packfile must be created due to packfile size limits. While disabling the plugging state only results in suboptimal behavior for the current code, it would be fatal for the bulk-fsync functionality later in this patch series. The net effect of these changes is to make a clear separation between the portion of the bulk-checkin infrastructure that is related to the packfile (nearly all of it at present) and the part that is related to other future optimizations of the ODB. Signed-off-by: Neeraj Singh <neerajsi@microsoft.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-04-05 08:20:07 +03:00
static int deflate_to_pack(struct bulk_checkin_packfile *state,
struct object_id *result_oid,
int fd, size_t size,
enum object_type type, const char *path,
unsigned flags)
{
off_t seekback, already_hashed_to;
git_hash_ctx ctx;
unsigned char obuf[16384];
unsigned header_len;
struct hashfile_checkpoint checkpoint = {0};
struct pack_idx_entry *idx = NULL;
seekback = lseek(fd, 0, SEEK_CUR);
if (seekback == (off_t) -1)
return error("cannot find the current offset");
header_len = format_object_header((char *)obuf, sizeof(obuf),
type, size);
the_hash_algo->init_fn(&ctx);
the_hash_algo->update_fn(&ctx, obuf, header_len);
/* Note: idx is non-NULL when we are writing */
if ((flags & HASH_WRITE_OBJECT) != 0)
CALLOC_ARRAY(idx, 1);
already_hashed_to = 0;
while (1) {
prepare_to_stream(state, flags);
if (idx) {
hashfile_checkpoint(state->f, &checkpoint);
idx->offset = state->offset;
crc32_begin(state->f);
}
if (!stream_to_pack(state, &ctx, &already_hashed_to,
fd, size, type, path, flags))
break;
/*
* Writing this object to the current pack will make
* it too big; we need to truncate it, start a new
* pack, and write into it.
*/
if (!idx)
BUG("should not happen");
hashfile_truncate(state->f, &checkpoint);
state->offset = checkpoint.offset;
bulk-checkin: rename 'state' variable and separate 'plugged' boolean This commit prepares for adding batch-fsync to the bulk-checkin infrastructure. The bulk-checkin infrastructure is currently used to batch up addition of large blobs to a packfile. When a blob is larger than big_file_threshold, we unconditionally add it to a pack. If bulk checkins are 'plugged', we allow multiple large blobs to be added to a single pack until we reach the packfile size limit; otherwise, we simply make a new packfile for each large blob. The 'unplug' call tells us when the series of blob additions is done so that we can finish the packfiles and make their objects available to subsequent operations. Stated another way, bulk-checkin allows callers to define a transaction that adds multiple objects to the object database, where the object database can optimize its internal operations within the transaction boundary. Batched fsync will fit into bulk-checkin by taking advantage of the plug/unplug functionality to determine the appropriate time to fsync and make newly-added objects available in the primary object database. * Rename 'state' variable to 'bulk_checkin_packfile', since we will later be adding 'bulk_fsync_objdir'. This also makes the variable easier to find in the debugger, since the name is more unique. * Rename finish_bulk_checkin to flush_bulk_checkin_packfile and call it unconditionally from unplug_bulk_checkin. Internally it will conditionally do a flush if there's any work to do. * Move the 'plugged' data member of 'bulk_checkin_state' into a separate static variable. Doing this avoids resetting the variable in finish_bulk_checkin when zeroing the 'bulk_checkin_state'. As-is, we seem to unintentionally disable the plugging functionality the first time a new packfile must be created due to packfile size limits. While disabling the plugging state only results in suboptimal behavior for the current code, it would be fatal for the bulk-fsync functionality later in this patch series. The net effect of these changes is to make a clear separation between the portion of the bulk-checkin infrastructure that is related to the packfile (nearly all of it at present) and the part that is related to other future optimizations of the ODB. Signed-off-by: Neeraj Singh <neerajsi@microsoft.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-04-05 08:20:07 +03:00
flush_bulk_checkin_packfile(state);
if (lseek(fd, seekback, SEEK_SET) == (off_t) -1)
return error("cannot seek back");
}
the_hash_algo->final_oid_fn(result_oid, &ctx);
if (!idx)
return 0;
idx->crc32 = crc32_end(state->f);
if (already_written(state, result_oid)) {
hashfile_truncate(state->f, &checkpoint);
state->offset = checkpoint.offset;
free(idx);
} else {
oidcpy(&idx->oid, result_oid);
ALLOC_GROW(state->written,
state->nr_written + 1,
state->alloc_written);
state->written[state->nr_written++] = idx;
}
return 0;
}
core.fsyncmethod: batched disk flushes for loose-objects When adding many objects to a repo with `core.fsync=loose-object`, the cost of fsync'ing each object file can become prohibitive. One major source of the cost of fsync is the implied flush of the hardware writeback cache within the disk drive. This commit introduces a new `core.fsyncMethod=batch` option that batches up hardware flushes. It hooks into the bulk-checkin odb-transaction functionality, takes advantage of tmp-objdir, and uses the writeout-only support code. When the new mode is enabled, we do the following for each new object: 1a. Create the object in a tmp-objdir. 2a. Issue a pagecache writeback request and wait for it to complete. At the end of the entire transaction when unplugging bulk checkin: 1b. Issue an fsync against a dummy file to flush the log and hardware writeback cache, which should by now have seen the tmp-objdir writes. 2b. Rename all of the tmp-objdir files to their final names. 3b. When updating the index and/or refs, we assume that Git will issue another fsync internal to that operation. This is not the default today, but the user now has the option of syncing the index and there is a separate patch series to implement syncing of refs. On a filesystem with a singular journal that is updated during name operations (e.g. create, link, rename, etc), such as NTFS, HFS+, or XFS we would expect the fsync to trigger a journal writeout so that this sequence is enough to ensure that the user's data is durable by the time the git command returns. This sequence also ensures that no object files appear in the main object store unless they are fsync-durable. Batch mode is only enabled if core.fsync includes loose-objects. If the legacy core.fsyncObjectFiles setting is enabled, but core.fsync does not include loose-objects, we will use file-by-file fsyncing. In step (1a) of the sequence, the tmp-objdir is created lazily to avoid work if no loose objects are ever added to the ODB. We use a tmp-objdir to maintain the invariant that no loose-objects are visible in the main ODB unless they are properly fsync-durable. This is important since future ODB operations that try to create an object with specific contents will silently drop the new data if an object with the target hash exists without checking that the loose-object contents match the hash. Only a full git-fsck would restore the ODB to a functional state where dataloss doesn't occur. In step (1b) of the sequence, we issue a fsync against a dummy file created specifically for the purpose. This method has a little higher cost than using one of the input object files, but makes adding new callers of this mechanism easier, since we don't need to figure out which object file is "last" or risk sharing violations by caching the fd of the last object file. _Performance numbers_: Linux - Hyper-V VM running Kernel 5.11 (Ubuntu 20.04) on a fast SSD. Mac - macOS 11.5.1 running on a Mac mini on a 1TB Apple SSD. Windows - Same host as Linux, a preview version of Windows 11. Adding 500 files to the repo with 'git add' Times reported in seconds. object file syncing | Linux | Mac | Windows --------------------|-------|-------|-------- disabled | 0.06 | 0.35 | 0.61 fsync | 1.88 | 11.18 | 2.47 batch | 0.15 | 0.41 | 1.53 Signed-off-by: Neeraj Singh <neerajsi@microsoft.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-04-05 08:20:09 +03:00
void prepare_loose_object_bulk_checkin(void)
{
/*
* We lazily create the temporary object directory
* the first time an object might be added, since
* callers may not know whether any objects will be
* added at the time they call begin_odb_transaction.
*/
if (!odb_transaction_nesting || bulk_fsync_objdir)
return;
bulk_fsync_objdir = tmp_objdir_create("bulk-fsync");
if (bulk_fsync_objdir)
tmp_objdir_replace_primary_odb(bulk_fsync_objdir, 0);
}
void fsync_loose_object_bulk_checkin(int fd, const char *filename)
{
/*
* If we have an active ODB transaction, we issue a call that
* cleans the filesystem page cache but avoids a hardware flush
* command. Later on we will issue a single hardware flush
* before renaming the objects to their final names as part of
* flush_batch_fsync.
*/
if (!bulk_fsync_objdir ||
git_fsync(fd, FSYNC_WRITEOUT_ONLY) < 0) {
if (errno == ENOSYS)
warning(_("core.fsyncMethod = batch is unsupported on this platform"));
core.fsyncmethod: batched disk flushes for loose-objects When adding many objects to a repo with `core.fsync=loose-object`, the cost of fsync'ing each object file can become prohibitive. One major source of the cost of fsync is the implied flush of the hardware writeback cache within the disk drive. This commit introduces a new `core.fsyncMethod=batch` option that batches up hardware flushes. It hooks into the bulk-checkin odb-transaction functionality, takes advantage of tmp-objdir, and uses the writeout-only support code. When the new mode is enabled, we do the following for each new object: 1a. Create the object in a tmp-objdir. 2a. Issue a pagecache writeback request and wait for it to complete. At the end of the entire transaction when unplugging bulk checkin: 1b. Issue an fsync against a dummy file to flush the log and hardware writeback cache, which should by now have seen the tmp-objdir writes. 2b. Rename all of the tmp-objdir files to their final names. 3b. When updating the index and/or refs, we assume that Git will issue another fsync internal to that operation. This is not the default today, but the user now has the option of syncing the index and there is a separate patch series to implement syncing of refs. On a filesystem with a singular journal that is updated during name operations (e.g. create, link, rename, etc), such as NTFS, HFS+, or XFS we would expect the fsync to trigger a journal writeout so that this sequence is enough to ensure that the user's data is durable by the time the git command returns. This sequence also ensures that no object files appear in the main object store unless they are fsync-durable. Batch mode is only enabled if core.fsync includes loose-objects. If the legacy core.fsyncObjectFiles setting is enabled, but core.fsync does not include loose-objects, we will use file-by-file fsyncing. In step (1a) of the sequence, the tmp-objdir is created lazily to avoid work if no loose objects are ever added to the ODB. We use a tmp-objdir to maintain the invariant that no loose-objects are visible in the main ODB unless they are properly fsync-durable. This is important since future ODB operations that try to create an object with specific contents will silently drop the new data if an object with the target hash exists without checking that the loose-object contents match the hash. Only a full git-fsck would restore the ODB to a functional state where dataloss doesn't occur. In step (1b) of the sequence, we issue a fsync against a dummy file created specifically for the purpose. This method has a little higher cost than using one of the input object files, but makes adding new callers of this mechanism easier, since we don't need to figure out which object file is "last" or risk sharing violations by caching the fd of the last object file. _Performance numbers_: Linux - Hyper-V VM running Kernel 5.11 (Ubuntu 20.04) on a fast SSD. Mac - macOS 11.5.1 running on a Mac mini on a 1TB Apple SSD. Windows - Same host as Linux, a preview version of Windows 11. Adding 500 files to the repo with 'git add' Times reported in seconds. object file syncing | Linux | Mac | Windows --------------------|-------|-------|-------- disabled | 0.06 | 0.35 | 0.61 fsync | 1.88 | 11.18 | 2.47 batch | 0.15 | 0.41 | 1.53 Signed-off-by: Neeraj Singh <neerajsi@microsoft.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-04-05 08:20:09 +03:00
fsync_or_die(fd, filename);
}
}
int index_bulk_checkin(struct object_id *oid,
int fd, size_t size, enum object_type type,
const char *path, unsigned flags)
{
bulk-checkin: rename 'state' variable and separate 'plugged' boolean This commit prepares for adding batch-fsync to the bulk-checkin infrastructure. The bulk-checkin infrastructure is currently used to batch up addition of large blobs to a packfile. When a blob is larger than big_file_threshold, we unconditionally add it to a pack. If bulk checkins are 'plugged', we allow multiple large blobs to be added to a single pack until we reach the packfile size limit; otherwise, we simply make a new packfile for each large blob. The 'unplug' call tells us when the series of blob additions is done so that we can finish the packfiles and make their objects available to subsequent operations. Stated another way, bulk-checkin allows callers to define a transaction that adds multiple objects to the object database, where the object database can optimize its internal operations within the transaction boundary. Batched fsync will fit into bulk-checkin by taking advantage of the plug/unplug functionality to determine the appropriate time to fsync and make newly-added objects available in the primary object database. * Rename 'state' variable to 'bulk_checkin_packfile', since we will later be adding 'bulk_fsync_objdir'. This also makes the variable easier to find in the debugger, since the name is more unique. * Rename finish_bulk_checkin to flush_bulk_checkin_packfile and call it unconditionally from unplug_bulk_checkin. Internally it will conditionally do a flush if there's any work to do. * Move the 'plugged' data member of 'bulk_checkin_state' into a separate static variable. Doing this avoids resetting the variable in finish_bulk_checkin when zeroing the 'bulk_checkin_state'. As-is, we seem to unintentionally disable the plugging functionality the first time a new packfile must be created due to packfile size limits. While disabling the plugging state only results in suboptimal behavior for the current code, it would be fatal for the bulk-fsync functionality later in this patch series. The net effect of these changes is to make a clear separation between the portion of the bulk-checkin infrastructure that is related to the packfile (nearly all of it at present) and the part that is related to other future optimizations of the ODB. Signed-off-by: Neeraj Singh <neerajsi@microsoft.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-04-05 08:20:07 +03:00
int status = deflate_to_pack(&bulk_checkin_packfile, oid, fd, size, type,
path, flags);
if (!odb_transaction_nesting)
bulk-checkin: rename 'state' variable and separate 'plugged' boolean This commit prepares for adding batch-fsync to the bulk-checkin infrastructure. The bulk-checkin infrastructure is currently used to batch up addition of large blobs to a packfile. When a blob is larger than big_file_threshold, we unconditionally add it to a pack. If bulk checkins are 'plugged', we allow multiple large blobs to be added to a single pack until we reach the packfile size limit; otherwise, we simply make a new packfile for each large blob. The 'unplug' call tells us when the series of blob additions is done so that we can finish the packfiles and make their objects available to subsequent operations. Stated another way, bulk-checkin allows callers to define a transaction that adds multiple objects to the object database, where the object database can optimize its internal operations within the transaction boundary. Batched fsync will fit into bulk-checkin by taking advantage of the plug/unplug functionality to determine the appropriate time to fsync and make newly-added objects available in the primary object database. * Rename 'state' variable to 'bulk_checkin_packfile', since we will later be adding 'bulk_fsync_objdir'. This also makes the variable easier to find in the debugger, since the name is more unique. * Rename finish_bulk_checkin to flush_bulk_checkin_packfile and call it unconditionally from unplug_bulk_checkin. Internally it will conditionally do a flush if there's any work to do. * Move the 'plugged' data member of 'bulk_checkin_state' into a separate static variable. Doing this avoids resetting the variable in finish_bulk_checkin when zeroing the 'bulk_checkin_state'. As-is, we seem to unintentionally disable the plugging functionality the first time a new packfile must be created due to packfile size limits. While disabling the plugging state only results in suboptimal behavior for the current code, it would be fatal for the bulk-fsync functionality later in this patch series. The net effect of these changes is to make a clear separation between the portion of the bulk-checkin infrastructure that is related to the packfile (nearly all of it at present) and the part that is related to other future optimizations of the ODB. Signed-off-by: Neeraj Singh <neerajsi@microsoft.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-04-05 08:20:07 +03:00
flush_bulk_checkin_packfile(&bulk_checkin_packfile);
return status;
}
void begin_odb_transaction(void)
{
odb_transaction_nesting += 1;
}
void flush_odb_transaction(void)
{
core.fsyncmethod: batched disk flushes for loose-objects When adding many objects to a repo with `core.fsync=loose-object`, the cost of fsync'ing each object file can become prohibitive. One major source of the cost of fsync is the implied flush of the hardware writeback cache within the disk drive. This commit introduces a new `core.fsyncMethod=batch` option that batches up hardware flushes. It hooks into the bulk-checkin odb-transaction functionality, takes advantage of tmp-objdir, and uses the writeout-only support code. When the new mode is enabled, we do the following for each new object: 1a. Create the object in a tmp-objdir. 2a. Issue a pagecache writeback request and wait for it to complete. At the end of the entire transaction when unplugging bulk checkin: 1b. Issue an fsync against a dummy file to flush the log and hardware writeback cache, which should by now have seen the tmp-objdir writes. 2b. Rename all of the tmp-objdir files to their final names. 3b. When updating the index and/or refs, we assume that Git will issue another fsync internal to that operation. This is not the default today, but the user now has the option of syncing the index and there is a separate patch series to implement syncing of refs. On a filesystem with a singular journal that is updated during name operations (e.g. create, link, rename, etc), such as NTFS, HFS+, or XFS we would expect the fsync to trigger a journal writeout so that this sequence is enough to ensure that the user's data is durable by the time the git command returns. This sequence also ensures that no object files appear in the main object store unless they are fsync-durable. Batch mode is only enabled if core.fsync includes loose-objects. If the legacy core.fsyncObjectFiles setting is enabled, but core.fsync does not include loose-objects, we will use file-by-file fsyncing. In step (1a) of the sequence, the tmp-objdir is created lazily to avoid work if no loose objects are ever added to the ODB. We use a tmp-objdir to maintain the invariant that no loose-objects are visible in the main ODB unless they are properly fsync-durable. This is important since future ODB operations that try to create an object with specific contents will silently drop the new data if an object with the target hash exists without checking that the loose-object contents match the hash. Only a full git-fsck would restore the ODB to a functional state where dataloss doesn't occur. In step (1b) of the sequence, we issue a fsync against a dummy file created specifically for the purpose. This method has a little higher cost than using one of the input object files, but makes adding new callers of this mechanism easier, since we don't need to figure out which object file is "last" or risk sharing violations by caching the fd of the last object file. _Performance numbers_: Linux - Hyper-V VM running Kernel 5.11 (Ubuntu 20.04) on a fast SSD. Mac - macOS 11.5.1 running on a Mac mini on a 1TB Apple SSD. Windows - Same host as Linux, a preview version of Windows 11. Adding 500 files to the repo with 'git add' Times reported in seconds. object file syncing | Linux | Mac | Windows --------------------|-------|-------|-------- disabled | 0.06 | 0.35 | 0.61 fsync | 1.88 | 11.18 | 2.47 batch | 0.15 | 0.41 | 1.53 Signed-off-by: Neeraj Singh <neerajsi@microsoft.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-04-05 08:20:09 +03:00
flush_batch_fsync();
bulk-checkin: rename 'state' variable and separate 'plugged' boolean This commit prepares for adding batch-fsync to the bulk-checkin infrastructure. The bulk-checkin infrastructure is currently used to batch up addition of large blobs to a packfile. When a blob is larger than big_file_threshold, we unconditionally add it to a pack. If bulk checkins are 'plugged', we allow multiple large blobs to be added to a single pack until we reach the packfile size limit; otherwise, we simply make a new packfile for each large blob. The 'unplug' call tells us when the series of blob additions is done so that we can finish the packfiles and make their objects available to subsequent operations. Stated another way, bulk-checkin allows callers to define a transaction that adds multiple objects to the object database, where the object database can optimize its internal operations within the transaction boundary. Batched fsync will fit into bulk-checkin by taking advantage of the plug/unplug functionality to determine the appropriate time to fsync and make newly-added objects available in the primary object database. * Rename 'state' variable to 'bulk_checkin_packfile', since we will later be adding 'bulk_fsync_objdir'. This also makes the variable easier to find in the debugger, since the name is more unique. * Rename finish_bulk_checkin to flush_bulk_checkin_packfile and call it unconditionally from unplug_bulk_checkin. Internally it will conditionally do a flush if there's any work to do. * Move the 'plugged' data member of 'bulk_checkin_state' into a separate static variable. Doing this avoids resetting the variable in finish_bulk_checkin when zeroing the 'bulk_checkin_state'. As-is, we seem to unintentionally disable the plugging functionality the first time a new packfile must be created due to packfile size limits. While disabling the plugging state only results in suboptimal behavior for the current code, it would be fatal for the bulk-fsync functionality later in this patch series. The net effect of these changes is to make a clear separation between the portion of the bulk-checkin infrastructure that is related to the packfile (nearly all of it at present) and the part that is related to other future optimizations of the ODB. Signed-off-by: Neeraj Singh <neerajsi@microsoft.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-04-05 08:20:07 +03:00
flush_bulk_checkin_packfile(&bulk_checkin_packfile);
}
void end_odb_transaction(void)
{
odb_transaction_nesting -= 1;
if (odb_transaction_nesting < 0)
BUG("Unbalanced ODB transaction nesting");
if (odb_transaction_nesting)
return;
flush_odb_transaction();
}