Add a bunch of helper functions that calculate the sizes of various
btrees. These will be used to repair btrees and btree headers.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Since the transaction allocation helper is about to become more complex,
move it to common.c and remove the redundant parameters.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Strengthen the btree block header checks to detect the number of records
being less than the btree type's minimum record count. Certain blocks
are allowed to violate this constraint -- specifically any btree block
at the top of the tree can have fewer than minrecs records.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
All scrub code runs in transaction context, which means that memory
allocations are automatically run in PF_MEMALLOC_NOFS context. It's
therefore unnecessary to pass in KM_NOFS to allocation routines, so
clean them all out.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Refactor the quota scrubber to take the quotaofflock and grab the quota
inode in the setup function so that we can treat quota in the same
"scrub in the context of this inode" (i.e. sc->ip) manner as we treat
any other inode. We do have to drop the quota inode's ILOCK_EXCL to use
dqiterate, but since dquots have their own individual locks the ILOCK
wasn't helping us anyway.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Create a helper function to iterate all the dquots of a given type in
the system, and refactor the dquot scrub to use it. This will get more
use in the quota repair code.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
The function 'xfs_qm_dqiterate' doesn't iterate dquots at all, it
iterates all dquot blocks of a quota inode and clears the counters.
Therefore, change the name to something more descriptive so that we can
introduce a real dquot iterator later.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
DQALLOC is only ever used with xfs_qm_dqget*, and the only flag that the
_dqget family of functions cares about is DQALLOC. Therefore, change
it to a boolean 'can alloc?' flag for the dqget interfaces where that
makes sense.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
The quota initialization code needs an "uncached" variant of _dqget to
read in default quota limits and timers before the dquot cache is fully
set up. We've already split up _dqget into its component pieces so
create a fourth variant to address this need, and make dqread internal
to xfs_dquot.c again.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Separate the disk dquot read and allocation functionality into
two helper functions, then refactor dqread to call them directly.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Create two incore dquot initialization functions that will help us to
disentangle dqget and dqread.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Quotacheck only runs during mount, which means that there are no other
processes in the system that could be doing chown or chproj. Therefore
there's no potential for racing to attach dquots to the inode so we can
drop all the ILOCK and race detection bits from quotacheck.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
There are two uses of dqget here -- one is to return the dquot for a
given type and id, and the other is to return the dquot for a given type
and inode. Those are two separate things, so split them into two
smaller functions.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
The flags argument is always zero, get rid of it.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Move the dqget input checks to a separate function in preparation for
splitting up the dqget functionality.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Delegate the dquot cache handling (radix tree lookup and insertion) to
separate helper functions so that we can continue to simplify the body
of dqget.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
There's only one caller of DQNEXT and its semantics can be moved into a
separate function, so create the function and get rid of the flag.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
When dquot flush or purge fail there's no need to spam the logs, we've
already logged the IO error or fs shutdown that caused the flush
failures.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
In commit efa092f3d4 "[XFS] Fixes a bug in the quota code when
allocating a new dquot record", we allocate a new dquot block, grab a
buffer to initialize it, and return the locked initialized dquot buffer
to the caller for further in-core dquot initialization. Unfortunately,
if the _bmap_finish errored out, _qm_dqalloc would also error out
without bothering to free the (locked) buffer. Leaking a locked buffer
caused hangs in generic/388 when quotas are enabled.
Furthermore, the _bmap_finish -> _defer_finish conversion in
310a75a3c6 ("xfs: change xfs_bmap_{finish,cancel,init,free} ->
xfs_defer_*") failed to observe that the buffer was held going into
_defer_finish and therefore failed to notice that the buffer lock is
/not/ maintained afterwards. Now that we can bjoin a buffer to a
defer_ops, use this mechanism to ensure that the buffer stays locked
across the _defer_finish. Release the holds and locks on the buffer as
appropriate if we have to error out.
There is a subtlety here for the caller in that the buffer emerges
locked and held to the transaction, so if the _trans_commit fails we
have to release the buffer explicitly. This fixes the unmount hang
in generic/388 when quotas are enabled.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Unwritten extents by definition have not been written to until they
are converted to normal written extents. If unwritten extents are
freed from a file, it is therefore guaranteed that the blocks have
not been written to since allocation (note that zero range punches
and reallocates blocks).
To cut down on online discards generated from workloads that make
use of preallocation, skip discards of extents if they are in the
unwritten state when the extent is freed.
Note that this optimization does not apply to log recovery, during
which all freed extents are discarded if online discard is enabled.
Also note that it may be possible for a filesystem crash to occur
after write completion of an unwritten extent but before unwritten
conversion such that the extent remains unwritten after log
recovery. Since this pseudo-inconsistency may already be possible
after a crash (consider writing to recently allocated blocks where
the allocation transaction is lost after a crash), this change
shouldn't introduce any fundamental limitations that don't already
exist. In short, on storage stacks where discards are important,
it's good practice to run an occasional fstrim even with online
discard enabled in the filesystem, particularly after a crash.
Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
We've had reports of online discard operations being sent from XFS
on write-only workloads. These discards occur as a result of
eofblocks trims that can occur after a large file copy completes.
These discards are slightly confusing for users who might be paying
close attention to online discards (i.e., vdo) due to performance
sensitivity. They also happen to be spurious because freed post-eof
blocks by definition have not been written to during the current
allocation cycle.
Update xfs_free_eofblocks() to skip discards that are purely
attributed to eofblocks trims. This cuts down the number of spurious
discards that may occur on write-only workloads due to normal
preallocation activity.
Note that discards of post-eof extents can still occur from other
codepaths that do not isolate handling of post-eof blocks from those
within eof. For example, file unlinks and truncates may still cause
discards for any file blocks affected by the operation.
Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Freed extents are unconditionally discarded when online discard is
enabled. Define XFS_BMAPI_NODISCARD to allow callers to bypass
discards when unnecessary. For example, this will be useful for
eofblocks trimming.
This patch does not change behavior.
Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
It's just a connector between a transaction and a log item. There's
a 1:1 relationship between a log item descriptor and a log item,
and a 1:1 relationship between a log item descriptor and a
transaction. Both relationships are created and terminated at the
same time, so why do we even have the descriptor?
Replace it with a specific list_head in the log item and a new
log item dirtied flag to replace the XFS_LID_DIRTY flag.
Signed-Off-By: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
[darrick: fix up deferred agfl intent finish_item use of LID_DIRTY]
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Just to make sure the item isn't associated with another
transaction when we try to reuse it.
Signed-Off-By: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
xfs_reflink_clear_inode_flag double-joins an inode to a transaction,
which is not allowed. Fix that and document that the caller must have
already joined it.
Signed-Off-By: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
[darrick: edit out trace for nonexistent ASSERT]
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
xfs_reflink_cancel_cow_range joins an inode twice to the same
transaction. This is not allowed, so fix it and document that the
callers of xfs_reflink_cancel_cow_blocks() must have already joined the
inode to the permanent transaction passed in.
Signed-Off-By: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
[darrick: edited the commit log to remove trace for nonexistent ASSERT]
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
xfs_inactive_symlink_rmt() does something nasty - it joins an inode
into a transaction it is already joined to. This means the inode can
have multiple log item descriptors attached to the transaction for
it. This breaks teh 1:1 mapping that is supposed to exist
between the log item and log item descriptor.
This results in the log item being processed twice during
transaction commit and CIL formatting, and there are lots of other
potential issues tha arise from double processing of log items in
the transaction commit state machine.
In this case, the inode is already held by the rolling transaction
returned from xfs_defer_finish(), so there's no need to join it
again.
Signed-Off-By: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Been hitting AIL ordering assert failures recently, but been unable
to trace them down because the system immediately hangs up onteh
spinlock that was held when this assert fires:
XFS: Assertion failed: XFS_LSN_CMP(prev_lip->li_lsn, lip->li_lsn) <= 0, file: fs/xfs/xfs_trans_ail.c, line: 52
Move the assertions outside of the spinlock so the corpse can
be dissected. Thanks to Brian Foster for supplying a clean
way of doing this.
Signed-Off-By: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
So it's clear in the trace where they are being called from.
Signed-Off-By: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Because currently we have no idea what the transaction context we
are operating in is, and I need to know that information to track
down bugs in multiple log item joins to transactions.
Signed-Off-By: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
The log item flags contain a field that is protected by the AIL
lock - the XFS_LI_IN_AIL flag. We use non-atomic RMW operations to
set and clear these flags, but most of the updates and checks are
not done with the AIL lock held and so are susceptible to update
races.
Fix this by changing the log item flags to use atomic bitops rather
than be reliant on the AIL lock for update serialisation.
Signed-Off-By: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
xfs_rmap_lookup_le_range can return errors, so we need to check for
them and bail out.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Don't panic() the system if the bmap records are garbage, just call
ASSERT which gives us the same backtrace but enables developers to
control if the system goes down or not. This makes debugging with
generic/388 much easier because it won't reboot the machine midway
through a run just because btree_read_bufl returns EIO when the fs has
already shut down.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Directory operations can perform block allocations as entries are
added/removed from directories. Defer AGFL block frees from the
remaining directory operation transactions. This covers the hard
link, remove and rename operations.
Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Inode allocation can require block allocation for physical inode
chunk allocation, inode btree record insertion, and/or directory
block allocation for entry insertion. Any of these block allocation
requests can require AGFL fixups prior to the actual allocation.
Update the common file creation transacions to defer AGFL frees from
these contexts to avoid too much log reservation consumption
per-transaction.
Since these transactions are already passed down through the btree
cursors and da_args structure, this simply requires to attach dfops
to the transaction. Note that this covers tr_create, tr_mkdir and
tr_symlink. Other transactions such as tr_create_tmpfile do not
already make use of deferred operations and so are left alone for
the time being.
Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
XFS inode chunks are already freed via deferred operations (which
now also defer AGFL block frees), but inode btree blocks are freed
directly in the associated context. This has been known to lead to
log reservation overruns in particular workloads where an inobt
block free may require several AGFL block frees (and thus several
allocation btree modifications) before the inobt block itself is
actually freed.
To avoid this problem, defer the frees of any AGFL blocks before the
inobt block free takes place. This requires passing the dfops from
xfs_inactive_ifree() down through the inobt ->[alloc|free]_block()
callouts, which essentially only requires to attach the dfops to the
transaction since it is already carried all the way through to the
inobt update and allocation.
Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Now that AGFL block frees are deferred when dfops is set in the
transaction, start deferring AGFL block frees from contexts that are
known to push the limits of existing log reservations.
The first such context is deferred operation processing itself. This
primarily targets deferred extent frees (such as file extents and
inode chunks), but in doing so covers all allocation operations that
occur in deferred operation processing context.
Update xfs_defer_finish() to set and reset ->t_agfl_dfops across the
processing sequence. This means that any AGFL block frees due to
allocation events result in the addition of new EFIs to the dfops
rather than being processed immediately. xfs_defer_finish() rolls
the transaction at least once more to process the frees of the AGFL
blocks back to the allocation btrees and returns once the AGFL is
rectified.
Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
The AGFL fixup code executes before every block allocation/free and
rectifies the AGFL based on the current, dynamic allocation
requirements of the fs. The AGFL must hold a minimum number of
blocks to satisfy a worst case split of the free space btrees caused
by the impending allocation operation. The AGFL is also updated to
maintain the implicit requirement for a minimum number of free slots
to satisfy a worst case join of the free space btrees.
Since the AGFL caches individual blocks, AGFL reduction typically
involves multiple, single block frees. We've had reports of
transaction overrun problems during certain workloads that boil down
to AGFL reduction freeing multiple blocks and consuming more space
in the log than was reserved for the transaction.
Since the objective of freeing AGFL blocks is to ensure free AGFL
free slots are available for the upcoming allocation, one way to
address this problem is to release surplus blocks from the AGFL
immediately but defer the free of those blocks (similar to how
file-mapped blocks are unmapped from the file in one transaction and
freed via a deferred operation) until the transaction is rolled.
This turns AGFL reduction into an operation with predictable log
reservation consumption.
Add the capability to defer AGFL block frees when a deferred ops
list is available to the AGFL fixup code. Add a dfops pointer to the
transaction to carry dfops through various contexts to the allocator
context. Deferring AGFL frees is conditional behavior based on
whether the transaction pointer is populated. The long term
objective is to reuse the transaction pointer to clean up all
unrelated callchains that pass dfops on the stack along with a
transaction and in doing so, consistently defer AGFL blocks from the
allocator.
A bit of customization is required to handle deferred completion
processing because AGFL blocks are accounted against a per-ag
reservation pool and AGFL blocks are not inserted into the extent
busy list when freed (they are inserted when used and released back
to the AGFL). Reuse the majority of the existing deferred extent
free infrastructure and customize it appropriately to handle AGFL
blocks.
Note that this patch only adds infrastructure. It does not change
behavior because no callers have been updated to pass ->t_agfl_dfops
into the allocation code.
Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Refactor the AGFL block free code into a new helper such that it can
be invoked from deferred context. No functional changes.
Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Rather than printing the top of the buffer that held a corrupted dqblk,
restructure things to print out the specific one that failed by pushing
the calls to the verifier_error function down into the verifier which
iterates over the buffer and detects the error.
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Add an xfs_dqblk verifier so that it can check the uuid on V5 filesystems;
it calls the existing xfs_dquot_verify verifier to validate the
xfs_disk_dquot_t contained inside it. This lets us move the uuid
verification out of the crc verifier, which makes little sense.
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
It's a bit dicey to pass in the smaller xfs_disk_dquot and then cast it to
something larger; pass in the full xfs_dqblk so we know the caller has sent
us the right thing. Rename the function to xfs_dqblk_repair for
clarity.
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
During quotacheck we send in the quota type, so verify that as well.
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Long ago the flags argument was used to determine whether to issue warnings
about corruptions, but that's done elsewhere now and the flag is unused
here, so remove it.
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Rather than checking what kind of locking is needed in a helper
function and then jumping through hoops to do the locking in line,
move the locking to the helper function that does all the checks
and rename it to xfs_ilock_for_iomap().
This also allows us to hoist all the nonblocking checks up into the
locking helper, further simplifier the code flow in
xfs_file_iomap_begin() and making it easier to understand.
Signed-Off-By: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Carlos Maiolino <cmaiolino@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
The current logic that determines whether allocation should be done
has grown somewhat spaghetti like with the addition of IOMAP_NOWAIT
functionality. Separate out each of the different cases into single,
obvious checks to get rid most of the nested IOMAP_NOWAIT checks
in the allocation logic.
Signed-Off-By: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Carlos Maiolino <cmaiolino@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
If we are doing direct IO writes with datasync semantics, we often
have to flush metadata changes along with the data write. However,
if we are overwriting existing data, there are no metadata changes
that we need to flush. In this case, optimising the IO by using
FUA write makes sense.
We know from the IOMAP_F_DIRTY flag as to whether a specific inode
requires a metadata flush - this is currently used by DAX to ensure
extent modification as stable in page fault operations. For direct
IO writes, we can use it to determine if we need to flush metadata
or not once the data is on disk.
Hence if we have been returned a mapped extent that is not new and
the IO mapping is not dirty, then we can use a FUA write to provide
datasync semantics. This allows us to short-cut the
generic_write_sync() call in IO completion and hence avoid
unnecessary operations. This makes pure direct IO data write
behaviour identical to the way block devices use REQ_FUA to provide
datasync semantics.
On a FUA enabled device, a synchronous direct IO write workload
(sequential 4k overwrites in 32MB file) had the following results:
# xfs_io -fd -c "pwrite -V 1 -D 0 32m" /mnt/scratch/boo
kernel time write()s write iops Write b/w
------ ---- -------- ---------- ---------
(no dsync) 4s 2173/s 2173 8.5MB/s
vanilla 22s 370/s 750 1.4MB/s
patched 19s 420/s 420 1.6MB/s
The patched code clearly doesn't send cache flushes anymore, but
instead uses FUA (confirmed via blktrace), and performance improves
a bit as a result. However, the benefits will be higher on workloads
that mix O_DSYNC overwrites with other write IO as we won't be
flushing the entire device cache on every DSYNC overwrite IO
anymore.
Signed-Off-By: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Currently iomap_dio_rw() only handles (data)sync write completions
for AIO. This means we can't optimised non-AIO IO to minimise device
flushes as we can't tell the caller whether a flush is required or
not.
To solve this problem and enable further optimisations, make
iomap_dio_rw responsible for data sync behaviour for all IO, not
just AIO.
In doing so, the sync operation is now accounted as part of the DIO
IO by inode_dio_end(), hence post-IO data stability updates will no
long race against operations that serialise via inode_dio_wait()
such as truncate or hole punch.
Signed-Off-By: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
To prepare for iomap iinfrastructure based DSYNC optimisations.
While moving the code araound, move the XFS write bytes metric
update for direct IO into xfs_dio_write_end_io callback so that we
always capture the amount of data written via AIO+DIO. This fixes
the problem where queued AIO+DIO writes are not accounted to this
metric.
Signed-Off-By: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
When looking at an event trace recently, I noticed that non-blocking
buffer lookup attempts would fail on cached locked buffers and then
run the slow cache-miss path. This means we are doing an xfs_buf
allocation, lookup and free unnecessarily every time we avoid
blocking on a locked buffer.
Fix this by changing _xfs_buf_find() to return an error status to
the caller to indicate that we failed the lock attempt rather than
just returning a NULL. This allows the higher level code to
discriminate between a cache miss and an cache hit that we failed to
lock.
This also allows us to return a -EFSCORRUPTED state if we are asked
to look up a block number outside the range of the filesystem in
_xfs_buf_find(), which moves us one step closer to being able to
handle such errors in a more graceful manner at the higher levels.
Signed-Off-By: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Carlos Maiolino <cmaiolino@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Darrick J. Wong
Brian Foster
Dave Chinner
Eric Sandeen