Pull btrfs updates from Chris Mason:
"This pull is mostly cleanups and fixes:
- The raid5/6 cleanups from Zhao Lei fixup some long standing warts
in the code and add improvements on top of the scrubbing support
from 3.19.
- Josef has round one of our ENOSPC fixes coming from large btrfs
clusters here at FB.
- Dave Sterba continues a long series of cleanups (thanks Dave), and
Filipe continues hammering on corner cases in fsync and others
This all was held up a little trying to track down a use-after-free in
btrfs raid5/6. It's not clear yet if this is just made easier to
trigger with this pull or if its a new bug from the raid5/6 cleanups.
Dave Sterba is the only one to trigger it so far, but he has a
consistent way to reproduce, so we'll get it nailed shortly"
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/mason/linux-btrfs: (68 commits)
Btrfs: don't remove extents and xattrs when logging new names
Btrfs: fix fsync data loss after adding hard link to inode
Btrfs: fix BUG_ON in btrfs_orphan_add() when delete unused block group
Btrfs: account for large extents with enospc
Btrfs: don't set and clear delalloc for O_DIRECT writes
Btrfs: only adjust outstanding_extents when we do a short write
btrfs: Fix out-of-space bug
Btrfs: scrub, fix sleep in atomic context
Btrfs: fix scheduler warning when syncing log
Btrfs: Remove unnecessary placeholder in btrfs_err_code
btrfs: cleanup init for list in free-space-cache
btrfs: delete chunk allocation attemp when setting block group ro
btrfs: clear bio reference after submit_one_bio()
Btrfs: fix scrub race leading to use-after-free
Btrfs: add missing cleanup on sysfs init failure
Btrfs: fix race between transaction commit and empty block group removal
btrfs: add more checks to btrfs_read_sys_array
btrfs: cleanup, rename a few variables in btrfs_read_sys_array
btrfs: add checks for sys_chunk_array sizes
btrfs: more superblock checks, lower bounds on devices and sectorsize/nodesize
...
If we are recording in the tree log that an inode has new names (new hard
links were added), we would drop items, belonging to the inode, that we
shouldn't:
1) When the flag BTRFS_INODE_COPY_EVERYTHING is set in the inode's runtime
flags, we ended up dropping all the extent and xattr items that were
previously logged. This was done only in memory, since logging a new
name doesn't imply syncing the log;
2) When the flag BTRFS_INODE_COPY_EVERYTHING is set in the inode's runtime
flags, we ended up dropping all the xattr items that were previously
logged. Like the case before, this was done only in memory because
logging a new name doesn't imply syncing the log.
This led to some surprises in scenarios such as the following:
1) write some extents to an inode;
2) fsync the inode;
3) truncate the inode or delete/modify some of its xattrs
4) add a new hard link for that inode
5) fsync some other file, to force the log tree to be durably persisted
6) power failure happens
The next time the fs is mounted, the fsync log replay code is executed,
and the resulting file doesn't have the content it had when the last fsync
against it was performed, instead if has a content matching what it had
when the last transaction commit happened.
So change the behaviour such that when a new name is logged, only the inode
item and reference items are processed.
This is easy to reproduce with the test I just made for xfstests, whose
main body is:
_scratch_mkfs >> $seqres.full 2>&1
_init_flakey
_mount_flakey
# Create our test file with some data.
$XFS_IO_PROG -f -c "pwrite -S 0xaa -b 8K 0 8K" \
$SCRATCH_MNT/foo | _filter_xfs_io
# Make sure the file is durably persisted.
sync
# Append some data to our file, to increase its size.
$XFS_IO_PROG -f -c "pwrite -S 0xcc -b 4K 8K 4K" \
$SCRATCH_MNT/foo | _filter_xfs_io
# Fsync the file, so from this point on if a crash/power failure happens, our
# new data is guaranteed to be there next time the fs is mounted.
$XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foo
# Now shrink our file to 5000 bytes.
$XFS_IO_PROG -c "truncate 5000" $SCRATCH_MNT/foo
# Now do an expanding truncate to a size larger than what we had when we last
# fsync'ed our file. This is just to verify that after power failure and
# replaying the fsync log, our file matches what it was when we last fsync'ed
# it - 12Kb size, first 8Kb of data had a value of 0xaa and the last 4Kb of
# data had a value of 0xcc.
$XFS_IO_PROG -c "truncate 32K" $SCRATCH_MNT/foo
# Add one hard link to our file. This made btrfs drop all of our file's
# metadata from the fsync log, including the metadata relative to the
# extent we just wrote and fsync'ed. This change was made only to the fsync
# log in memory, so adding the hard link alone doesn't change the persisted
# fsync log. This happened because the previous truncates set the runtime
# flag BTRFS_INODE_NEEDS_FULL_SYNC in the btrfs inode structure.
ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link
# Now make sure the in memory fsync log is durably persisted.
# Creating and fsync'ing another file will do it.
# After this our persisted fsync log will no longer have metadata for our file
# foo that points to the extent we wrote and fsync'ed before.
touch $SCRATCH_MNT/bar
$XFS_IO_PROG -c "fsync" $SCRATCH_MNT/bar
# As expected, before the crash/power failure, we should be able to see a file
# with a size of 32Kb, with its first 5000 bytes having the value 0xaa and all
# the remaining bytes with value 0x00.
echo "File content before:"
od -t x1 $SCRATCH_MNT/foo
# Simulate a crash/power loss.
_load_flakey_table $FLAKEY_DROP_WRITES
_unmount_flakey
_load_flakey_table $FLAKEY_ALLOW_WRITES
_mount_flakey
# After mounting the fs again, the fsync log was replayed.
# The expected result is to see a file with a size of 12Kb, with its first 8Kb
# of data having the value 0xaa and its last 4Kb of data having a value of 0xcc.
# The btrfs bug used to leave the file as it used te be as of the last
# transaction commit - that is, with a size of 8Kb with all bytes having a
# value of 0xaa.
echo "File content after:"
od -t x1 $SCRATCH_MNT/foo
The test case for xfstests follows soon.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
We have a scenario where after the fsync log replay we can lose file data
that had been previously fsync'ed if we added an hard link for our inode
and after that we sync'ed the fsync log (for example by fsync'ing some
other file or directory).
This is because when adding an hard link we updated the inode item in the
log tree with an i_size value of 0. At that point the new inode item was
in memory only and a subsequent fsync log replay would not make us lose
the file data. However if after adding the hard link we sync the log tree
to disk, by fsync'ing some other file or directory for example, we ended
up losing the file data after log replay, because the inode item in the
persisted log tree had an an i_size of zero.
This is easy to reproduce, and the following excerpt from my test for
xfstests shows this:
_scratch_mkfs >> $seqres.full 2>&1
_init_flakey
_mount_flakey
# Create one file with data and fsync it.
# This made the btrfs fsync log persist the data and the inode metadata with
# a correct inode->i_size (4096 bytes).
$XFS_IO_PROG -f -c "pwrite -S 0xaa -b 4K 0 4K" -c "fsync" \
$SCRATCH_MNT/foo | _filter_xfs_io
# Now add one hard link to our file. This made the btrfs code update the fsync
# log, in memory only, with an inode metadata having a size of 0.
ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link
# Now force persistence of the fsync log to disk, for example, by fsyncing some
# other file.
touch $SCRATCH_MNT/bar
$XFS_IO_PROG -c "fsync" $SCRATCH_MNT/bar
# Before a power loss or crash, we could read the 4Kb of data from our file as
# expected.
echo "File content before:"
od -t x1 $SCRATCH_MNT/foo
# Simulate a crash/power loss.
_load_flakey_table $FLAKEY_DROP_WRITES
_unmount_flakey
_load_flakey_table $FLAKEY_ALLOW_WRITES
_mount_flakey
# After the fsync log replay, because the fsync log had a value of 0 for our
# inode's i_size, we couldn't read anymore the 4Kb of data that we previously
# wrote and fsync'ed. The size of the file became 0 after the fsync log replay.
echo "File content after:"
od -t x1 $SCRATCH_MNT/foo
Another alternative test, that doesn't need to fsync an inode in the same
transaction it was created, is:
_scratch_mkfs >> $seqres.full 2>&1
_init_flakey
_mount_flakey
# Create our test file with some data.
$XFS_IO_PROG -f -c "pwrite -S 0xaa -b 8K 0 8K" \
$SCRATCH_MNT/foo | _filter_xfs_io
# Make sure the file is durably persisted.
sync
# Append some data to our file, to increase its size.
$XFS_IO_PROG -f -c "pwrite -S 0xcc -b 4K 8K 4K" \
$SCRATCH_MNT/foo | _filter_xfs_io
# Fsync the file, so from this point on if a crash/power failure happens, our
# new data is guaranteed to be there next time the fs is mounted.
$XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foo
# Add one hard link to our file. This made btrfs write into the in memory fsync
# log a special inode with generation 0 and an i_size of 0 too. Note that this
# didn't update the inode in the fsync log on disk.
ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link
# Now make sure the in memory fsync log is durably persisted.
# Creating and fsync'ing another file will do it.
touch $SCRATCH_MNT/bar
$XFS_IO_PROG -c "fsync" $SCRATCH_MNT/bar
# As expected, before the crash/power failure, we should be able to read the
# 12Kb of file data.
echo "File content before:"
od -t x1 $SCRATCH_MNT/foo
# Simulate a crash/power loss.
_load_flakey_table $FLAKEY_DROP_WRITES
_unmount_flakey
_load_flakey_table $FLAKEY_ALLOW_WRITES
_mount_flakey
# After mounting the fs again, the fsync log was replayed.
# The btrfs fsync log replay code didn't update the i_size of the persisted
# inode because the inode item in the log had a special generation with a
# value of 0 (and it couldn't know the correct i_size, since that inode item
# had a 0 i_size too). This made the last 4Kb of file data inaccessible and
# effectively lost.
echo "File content after:"
od -t x1 $SCRATCH_MNT/foo
This isn't a new issue/regression. This problem has been around since the
log tree code was added in 2008:
Btrfs: Add a write ahead tree log to optimize synchronous operations
(commit e02119d5a7)
Test cases for xfstests follow soon.
CC: <stable@vger.kernel.org>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
Add missing blk_finish_plug in btrfs_sync_log()
Signed-off-by: Forrest Liu <forrestl@synology.com>
Reviewed-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Chris Mason <clm@fb.com>
If we have an inode with a large number of hard links, some of which may
be extrefs, turn a regular ref into an extref, fsync the inode and then
replay the fsync log (after a crash/reboot), we can endup with an fsync
log that makes the replay code always fail with -EOVERFLOW when processing
the inode's references.
This is easy to reproduce with the test case I made for xfstests. Its steps
are the following:
_scratch_mkfs "-O extref" >> $seqres.full 2>&1
_init_flakey
_mount_flakey
# Create a test file with 3001 hard links. This number is large enough to
# make btrfs start using extrefs at some point even if the fs has the maximum
# possible leaf/node size (64Kb).
echo "hello world" > $SCRATCH_MNT/foo
for i in `seq 1 3000`; do
ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link_`printf "%04d" $i`
done
# Make sure all metadata and data are durably persisted.
sync
# Now remove one link, add a new one with a new name, add another new one with
# the same name as the one we just removed and fsync the inode.
rm -f $SCRATCH_MNT/foo_link_0001
ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link_3001
ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link_0001
rm -f $SCRATCH_MNT/foo_link_0002
ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link_3002
ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link_3003
$XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foo
# Simulate a crash/power loss. This makes sure the next mount
# will see an fsync log and will replay that log.
_load_flakey_table $FLAKEY_DROP_WRITES
_unmount_flakey
_load_flakey_table $FLAKEY_ALLOW_WRITES
_mount_flakey
# Check that the number of hard links is correct, we are able to remove all
# the hard links and read the file's data. This is just to verify we don't
# get stale file handle errors (due to dangling directory index entries that
# point to inodes that no longer exist).
echo "Link count: $(stat --format=%h $SCRATCH_MNT/foo)"
[ -f $SCRATCH_MNT/foo ] || echo "Link foo is missing"
for ((i = 1; i <= 3003; i++)); do
name=foo_link_`printf "%04d" $i`
if [ $i -eq 2 ]; then
[ -f $SCRATCH_MNT/$name ] && echo "Link $name found"
else
[ -f $SCRATCH_MNT/$name ] || echo "Link $name is missing"
fi
done
rm -f $SCRATCH_MNT/foo_link_*
cat $SCRATCH_MNT/foo
rm -f $SCRATCH_MNT/foo
status=0
exit
The fix is simply to correct the overflow condition when overwriting a
reference item because it was wrong, trying to increase the item in the
fs/subvol tree by an impossible amount. Also ensure that we don't insert
one normal ref and one ext ref for the same dentry - this happened because
processing a dir index entry from the parent in the log happened when
the normal ref item was full, which made the logic insert an extref and
later when the normal ref had enough room, it would be inserted again
when processing the ref item from the child inode in the log.
This issue has been present since the introduction of the extrefs feature
(2012).
A test case for xfstests follows soon. This test only passes if the previous
patch titled "Btrfs: fix fsync when extend references are added to an inode"
is applied too.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
If we added an extended reference to an inode and fsync'ed it, the log
replay code would make our inode have an incorrect link count, which
was lower then the expected/correct count.
This resulted in stale directory index entries after deleting some of
the hard links, and any access to the dangling directory entries resulted
in -ESTALE errors because the entries pointed to inode items that don't
exist anymore.
This is easy to reproduce with the test case I made for xfstests, and
the bulk of that test is:
_scratch_mkfs "-O extref" >> $seqres.full 2>&1
_init_flakey
_mount_flakey
# Create a test file with 3001 hard links. This number is large enough to
# make btrfs start using extrefs at some point even if the fs has the maximum
# possible leaf/node size (64Kb).
echo "hello world" > $SCRATCH_MNT/foo
for i in `seq 1 3000`; do
ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link_`printf "%04d" $i`
done
# Make sure all metadata and data are durably persisted.
sync
# Add one more link to the inode that ends up being a btrfs extref and fsync
# the inode.
ln $SCRATCH_MNT/foo $SCRATCH_MNT/foo_link_3001
$XFS_IO_PROG -c "fsync" $SCRATCH_MNT/foo
# Simulate a crash/power loss. This makes sure the next mount
# will see an fsync log and will replay that log.
_load_flakey_table $FLAKEY_DROP_WRITES
_unmount_flakey
_load_flakey_table $FLAKEY_ALLOW_WRITES
_mount_flakey
# Now after the fsync log replay btrfs left our inode with a wrong link count N,
# which was smaller than the correct link count M (N < M).
# So after removing N hard links, the remaining M - N directory entries were
# still visible to user space but it was impossible to do anything with them
# because they pointed to an inode that didn't exist anymore. This resulted in
# stale file handle errors (-ESTALE) when accessing those dentries for example.
#
# So remove all hard links except the first one and then attempt to read the
# file, to verify we don't get an -ESTALE error when accessing the inodel
#
# The btrfs fsck tool also detected the incorrect inode link count and it
# reported an error message like the following:
#
# root 5 inode 257 errors 2001, no inode item, link count wrong
# unresolved ref dir 256 index 2978 namelen 13 name foo_link_2976 filetype 1 errors 4, no inode ref
#
# The fstests framework automatically calls fsck after a test is run, so we
# don't need to call fsck explicitly here.
rm -f $SCRATCH_MNT/foo_link_*
cat $SCRATCH_MNT/foo
status=0
exit
So make sure an fsync always flushes the delayed inode item, so that the
fsync log contains it (needed in order to trigger the link count fixup
code) and fix the extref counting function, which always return -ENOENT
to its caller (and made it assume there were always 0 extrefs).
This issue has been present since the introduction of the extrefs feature
(2012).
A test case for xfstests follows soon.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
If we have an inode (file) with a link count greater than 1, remove
one of its hard links, fsync the inode, power fail/crash and then
replay the fsync log on the next mount, we end up getting the parent
directory's metadata inconsistent - its i_size still reflects the
deleted hard link and has dangling index entries (with no matching
inode reference entries). This prevents the directory from ever being
deletable, as its i_size can never decrease to BTRFS_EMPTY_DIR_SIZE
even if all of its children inodes are deleted, and the dangling index
entries can never be removed (as they point to an inode that does not
exist anymore).
This is easy to reproduce with the following excerpt from the test case
for xfstests that I just made:
_scratch_mkfs >> $seqres.full 2>&1
_init_flakey
_mount_flakey
# Create a test file with 2 hard links in the same directory.
mkdir -p $SCRATCH_MNT/a/b
echo "hello world" > $SCRATCH_MNT/a/b/foo
ln $SCRATCH_MNT/a/b/foo $SCRATCH_MNT/a/b/bar
# Make sure all metadata and data are durably persisted.
sync
# Now remove one of the hard links and fsync the inode.
rm -f $SCRATCH_MNT/a/b/bar
$XFS_IO_PROG -c "fsync" $SCRATCH_MNT/a/b/foo
# Simulate a crash/power loss. This makes sure the next mount
# will see an fsync log and will replay that log.
_load_flakey_table $FLAKEY_DROP_WRITES
_unmount_flakey
_load_flakey_table $FLAKEY_ALLOW_WRITES
_mount_flakey
# Remove the last hard link of the file and attempt to remove its parent
# directory - this failed in btrfs because the fsync log and replay code
# didn't decrement the parent directory's i_size and left dangling directory
# index entries - this made the btrfs rmdir implementation always fail with
# the error -ENOTEMPTY.
#
# The dangling directory index entries were visible to user space, but it was
# impossible to do anything on them (unlink, open, read, write, stat, etc)
# because the inode they pointed to did not exist anymore.
#
# The parent directory's metadata inconsistency (stale index entries) was
# also detected by btrfs' fsck tool, which is run automatically by the fstests
# framework when the test finishes. The error message reported by fsck was:
#
# root 5 inode 259 errors 2001, no inode item, link count wrong
# unresolved ref dir 258 index 3 namelen 3 name bar filetype 1 errors 4, no inode ref
#
rm -f $SCRATCH_MNT/a/b/*
rmdir $SCRATCH_MNT/a/b
rmdir $SCRATCH_MNT/a
To fix this just make sure that after an unlink, if the inode is fsync'ed,
he parent inode is fully logged in the fsync log.
A test case for xfstests follows soon.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
We can search and add the orphan item in one go,
btrfs_insert_orphan_item will find out if the item already exists.
Signed-off-by: David Sterba <dsterba@suse.cz>
If btrfs_find_item is called with NULL path it allocates one locally but
does not free it. Affected paths are inserting an orphan item for a file
and for a subvol root.
Move the path allocation to the callers.
CC: <stable@vger.kernel.org> # 3.14+
Fixes: 3f870c2899 ("btrfs: expand btrfs_find_item() to include find_orphan_item functionality")
Signed-off-by: David Sterba <dsterba@suse.cz>
Finally it's clear that the requested blocksize is always equal to
nodesize, with one exception, the superblock.
Superblock has fixed size regardless of the metadata block size, but
uses the same helpers to initialize sys array/chunk tree and to work
with the chunk items. So it pretends to be an extent_buffer for a
moment, btrfs_read_sys_array is full of special cases, we're adding one
more.
Signed-off-by: David Sterba <dsterba@suse.cz>
When doing a fsync with a fast path we have a time window where we can miss
the fact that writeback of some file data failed, and therefore we endup
returning success (0) from fsync when we should return an error.
The steps that lead to this are the following:
1) We start all ordered extents by calling filemap_fdatawrite_range();
2) We do some other work like locking the inode's i_mutex, start a transaction,
start a log transaction, etc;
3) We enter btrfs_log_inode(), acquire the inode's log_mutex and collect all the
ordered extents from inode's ordered tree into a list;
4) But by the time we do ordered extent collection, some ordered extents we started
at step 1) might have already completed with an error, and therefore we didn't
found them in the ordered tree and had no idea they finished with an error. This
makes our fsync return success (0) to userspace, but has no bad effects on the log
like for example insertion of file extent items into the log that point to unwritten
extents, because the invalid extent maps were removed before the ordered extent
completed (in inode.c:btrfs_finish_ordered_io).
So after collecting the ordered extents just check if the inode's i_mapping has any
error flags set (AS_EIO or AS_ENOSPC) and leave with an error if it does. Whenever
writeback fails for a page of an ordered extent, we call mapping_set_error (done in
extent_io.c:end_extent_writepage, called by extent_io.c:end_bio_extent_writepage)
that sets one of those error flags in the inode's i_mapping flags.
This change also has the side effect of fixing the issue where for fast fsyncs we
never checked/cleared the error flags from the inode's i_mapping flags, which means
that a full fsync performed after a fast fsync could get such errors that belonged
to the fast fsync - because the full fsync calls btrfs_wait_ordered_range() which
calls filemap_fdatawait_range(), and the later checks for and clears those flags,
while for fast fsyncs we never call filemap_fdatawait_range() or anything else
that checks for and clears the error flags from the inode's i_mapping.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
Instead of collecting all ordered extents from the inode's ordered tree
and then wait for all of them to complete, just collect the ones that
overlap the fsync range.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
If an error happens during writeback of log btree extents, make sure the
error is returned to the caller (fsync), so that it takes proper action
(commit current transaction) instead of writing a superblock that points
to log btrees with all or some nodes that weren't durably persisted.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
Liu Bo pointed out that my previous fix would lose the generation update in the
scenario I described. It is actually much worse than that, we could lose the
entire extent if we lose power right after the transaction commits. Consider
the following
write extent 0-4k
log extent in log tree
commit transaction
< power fail happens here
ordered extent completes
We would lose the 0-4k extent because it hasn't updated the actual fs tree, and
the transaction commit will reset the log so it isn't replayed. If we lose
power before the transaction commit we are save, otherwise we are not.
Fix this by keeping track of all extents we logged in this transaction. Then
when we go to commit the transaction make sure we wait for all of those ordered
extents to complete before proceeding. This will make sure that if we lose
power after the transaction commit we still have our data. This also fixes the
problem of the improperly updated extent generation. Thanks,
cc: stable@vger.kernel.org
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>
If we have two fsync()'s race on different subvols one will do all of its work
to get into the log_tree, wait on it's outstanding IO, and then allow the
log_tree to finish it's commit. The problem is we were just free'ing that
subvols logged extents instead of waiting on them, so whoever lost the race
wouldn't really have their data on disk. Fix this by waiting properly instead
of freeing the logged extents. Thanks,
cc: stable@vger.kernel.org
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>
Pull btrfs fixes from Chris Mason:
"Filipe is nailing down some problems with our skinny extent variation,
and Dave's patch fixes endian problems in the new super block checks"
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/mason/linux-btrfs:
Btrfs: fix race that makes btrfs_lookup_extent_info miss skinny extent items
Btrfs: properly clean up btrfs_end_io_wq_cache
Btrfs: fix invalid leaf slot access in btrfs_lookup_extent()
btrfs: use macro accessors in superblock validation checks
If we couldn't find our extent item, we accessed the current slot
(path->slots[0]) to check if it corresponds to an equivalent skinny
metadata item. However this slot could be beyond our last item in the
leaf (i.e. path->slots[0] >= btrfs_header_nritems(leaf)), in which case
we shouldn't process it.
Since btrfs_lookup_extent() is only used to find extent items for data
extents, fix this by removing completely the logic that looks up for an
equivalent skinny metadata item, since it can not exist.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
Pull btrfs updates from Chris Mason:
"The largest set of changes here come from Miao Xie. He's cleaning up
and improving read recovery/repair for raid, and has a number of
related fixes.
I've merged another set of fsync fixes from Filipe, and he's also
improved the way we handle metadata write errors to make sure we force
the FS readonly if things go wrong.
Otherwise we have a collection of fixes and cleanups. Dave Sterba
gets a cookie for removing the most lines (thanks Dave)"
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/mason/linux-btrfs: (139 commits)
btrfs: Fix compile error when CONFIG_SECURITY is not set.
Btrfs: fix compiles when CONFIG_BTRFS_FS_RUN_SANITY_TESTS is off
btrfs: Make btrfs handle security mount options internally to avoid losing security label.
Btrfs: send, don't delay dir move if there's a new parent inode
btrfs: add more superblock checks
Btrfs: fix race in WAIT_SYNC ioctl
Btrfs: be aware of btree inode write errors to avoid fs corruption
Btrfs: remove redundant btrfs_verify_qgroup_counts declaration.
btrfs: fix shadow warning on cmp
Btrfs: fix compilation errors under DEBUG
Btrfs: fix crash of btrfs_release_extent_buffer_page
Btrfs: add missing end_page_writeback on submit_extent_page failure
btrfs: Fix the wrong condition judgment about subset extent map
Btrfs: fix build_backref_tree issue with multiple shared blocks
Btrfs: cleanup error handling in build_backref_tree
btrfs: move checks for DUMMY_ROOT into a helper
btrfs: new define for the inline extent data start
btrfs: kill extent_buffer_page helper
btrfs: drop constant param from btrfs_release_extent_buffer_page
btrfs: hide typecast to definition of BTRFS_SEND_TRANS_STUB
...
When we do a fast fsync, we start all ordered operations and then while
they're running in parallel we visit the list of modified extent maps
and construct their matching file extent items and write them to the
log btree. After that, in btrfs_sync_log() we wait for all the ordered
operations to finish (via btrfs_wait_logged_extents).
The problem with this is that we were completely ignoring errors that
can happen in the extent write path, such as -ENOSPC, a temporary -ENOMEM
or -EIO errors for example. When such error happens, it means we have parts
of the on disk extent that weren't written to, and so we end up logging
file extent items that point to these extents that contain garbage/random
data - so after a crash/reboot plus log replay, we get our inode's metadata
pointing to those extents.
This worked in contrast with the full (non-fast) fsync path, where we
start all ordered operations, wait for them to finish and then write
to the log btree. In this path, after each ordered operation completes
we check if it's flagged with an error (BTRFS_ORDERED_IOERR) and return
-EIO if so (via btrfs_wait_ordered_range).
So if an error happens with any ordered operation, just return a -EIO
error to userspace, so that it knows that not all of its previous writes
were durably persisted and the application can take proper action (like
redo the writes for e.g.) - and definitely not leave any file extent items
in the log refer to non fully written extents.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
When replaying a directory from the fsync log, if a directory entry
exists both in the fs/subvol tree and in the log, the directory's inode
got its i_size updated incorrectly, accounting for the dentry's name
twice.
Reproducer, from a test for xfstests:
_scratch_mkfs >> $seqres.full 2>&1
_init_flakey
_mount_flakey
touch $SCRATCH_MNT/foo
sync
touch $SCRATCH_MNT/bar
xfs_io -c "fsync" $SCRATCH_MNT
xfs_io -c "fsync" $SCRATCH_MNT/bar
_load_flakey_table $FLAKEY_DROP_WRITES
_unmount_flakey
_load_flakey_table $FLAKEY_ALLOW_WRITES
_mount_flakey
[ -f $SCRATCH_MNT/foo ] || echo "file foo is missing"
[ -f $SCRATCH_MNT/bar ] || echo "file bar is missing"
_unmount_flakey
_check_scratch_fs $FLAKEY_DEV
The filesystem check at the end failed with the message:
"root 5 root dir 256 error".
A test case for xfstests follows.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
None of the uses of btrfs_search_forward() need to have the path
nodes (level >= 1) read locked, only the leaf needs to be locked
while the caller processes it. Therefore make it return a path
with all nodes unlocked, except for the leaf.
This change is motivated by the observation that during a file
fsync we repeatdly call btrfs_search_forward() and process the
returned leaf while upper nodes of the returned path (level >= 1)
are read locked, which unnecessarily blocks other tasks that want
to write to the same fs/subvol btree.
Therefore instead of modifying the fsync code to unlock all nodes
with level >= 1 immediately after calling btrfs_search_forward(),
change btrfs_search_forward() to do it, so that it benefits all
callers.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
The nodesize and leafsize were never of different values. Unify the
usage and make nodesize the one. Cleanup the redundant checks and
helpers.
Shaves a few bytes from .text:
text data bss dec hex filename
852418 24560 23112 900090 dbbfa btrfs.ko.before
851074 24584 23112 898770 db6d2 btrfs.ko.after
Signed-off-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Chris Mason <clm@fb.com>
btrfs_set_key_type and btrfs_key_type are used inconsistently along with
open coded variants. Other members of btrfs_key are accessed directly
without any helpers anyway.
Signed-off-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Chris Mason <clm@fb.com>
When a ranged fsync finishes if there are still extent maps in the modified
list, still set the inode's logged_trans and last_log_commit. This is important
in case an inode is fsync'ed and unlinked in the same transaction, to ensure its
inode ref gets deleted from the log and the respective dentries in its parent
are deleted too from the log (if the parent directory was fsync'ed in the same
transaction).
Instead make btrfs_inode_in_log() return false if the list of modified extent
maps isn't empty.
This is an incremental on top of the v4 version of the patch:
"Btrfs: fix fsync data loss after a ranged fsync"
which was added to its v5, but didn't make it on time.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
While we're doing a full fsync (when the inode has the flag
BTRFS_INODE_NEEDS_FULL_SYNC set) that is ranged too (covers only a
portion of the file), we might have ordered operations that are started
before or while we're logging the inode and that fall outside the fsync
range.
Therefore when a full ranged fsync finishes don't remove every extent
map from the list of modified extent maps - as for some of them, that
fall outside our fsync range, their respective ordered operation hasn't
finished yet, meaning the corresponding file extent item wasn't inserted
into the fs/subvol tree yet and therefore we didn't log it, and we must
let the next fast fsync (one that checks only the modified list) see this
extent map and log a matching file extent item to the log btree and wait
for its ordered operation to finish (if it's still ongoing).
A test case for xfstests follows.
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
The file hole detection logic during a file fsync wasn't correct,
because it didn't look back (in a previous leaf) for the last file
extent item that can be in a leaf to the left of our leaf and that
has a generation lower than the current transaction id. This made it
assume that a hole exists when it really doesn't exist in the file.
Such false positive hole detection happens in the following scenario:
* We have a file that has many file extent items, covering 3 or more
btree leafs (the first leaf must contain non file extent items too).
* Two ranges of the file are modified, with their extent items being
located at 2 different leafs and those leafs aren't consecutive.
* When processing the second modified leaf, we weren't checking if
some file extent item exists that is located in some leaf that is
between our 2 modified leafs, and therefore assumed the range defined
between the last file extent item in the first leaf and the first file
extent item in the second leaf matched a hole.
Fortunately this didn't result in overriding the log with wrong data,
instead it made the last loop in copy_items() attempt to insert a
duplicated key (for a hole file extent item), which makes the file
fsync code return with -EEXIST to file.c:btrfs_sync_file() which in
turn ends up doing a full transaction commit, which is much more
expensive then writing only to the log tree and wait for it to be
durably persisted (as well as the file's modified extents/pages).
Therefore fix the hole detection logic, so that we don't pay the
cost of doing full transaction commits.
I could trigger this issue with the following test for xfstests (which
never fails, either without or with this patch). The last fsync call
results in a full transaction commit, due to the -EEXIST error mentioned
above. I could also observe this behaviour happening frequently when
running xfstests/generic/075 in a loop.
Test:
_cleanup()
{
_cleanup_flakey
rm -fr $tmp
}
# get standard environment, filters and checks
. ./common/rc
. ./common/filter
. ./common/dmflakey
# real QA test starts here
_supported_fs btrfs
_supported_os Linux
_require_scratch
_require_dm_flakey
_need_to_be_root
rm -f $seqres.full
# Create a file with many file extent items, each representing a 4Kb extent.
# These items span 3 btree leaves, of 16Kb each (default mkfs.btrfs leaf size
# as of btrfs-progs 3.12).
_scratch_mkfs -l 16384 >/dev/null 2>&1
_init_flakey
SAVE_MOUNT_OPTIONS="$MOUNT_OPTIONS"
MOUNT_OPTIONS="$MOUNT_OPTIONS -o commit=999"
_mount_flakey
# First fsync, inode has BTRFS_INODE_NEEDS_FULL_SYNC flag set.
$XFS_IO_PROG -f -c "pwrite -S 0x01 -b 4096 0 4096" -c "fsync" \
$SCRATCH_MNT/foo | _filter_xfs_io
# For any of the following fsync calls, inode doesn't have the flag
# BTRFS_INODE_NEEDS_FULL_SYNC set.
for ((i = 1; i <= 500; i++)); do
OFFSET=$((4096 * i))
LEN=4096
$XFS_IO_PROG -c "pwrite -S 0x01 $OFFSET $LEN" -c "fsync" \
$SCRATCH_MNT/foo | _filter_xfs_io
done
# Commit transaction and bump next transaction's id (to 7).
sync
# Truncate will set the BTRFS_INODE_NEEDS_FULL_SYNC flag in the btrfs's
# inode runtime flags.
$XFS_IO_PROG -c "truncate 2048000" $SCRATCH_MNT/foo
# Commit transaction and bump next transaction's id (to 8).
sync
# Touch 1 extent item from the first leaf and 1 from the last leaf. The leaf
# in the middle, containing only file extent items, isn't touched. So the
# next fsync, when calling btrfs_search_forward(), won't visit that middle
# leaf. First and 3rd leaf have now a generation with value 8, while the
# middle leaf remains with a generation with value 6.
$XFS_IO_PROG \
-c "pwrite -S 0xee -b 4096 0 4096" \
-c "pwrite -S 0xff -b 4096 2043904 4096" \
-c "fsync" \
$SCRATCH_MNT/foo | _filter_xfs_io
_load_flakey_table $FLAKEY_DROP_WRITES
md5sum $SCRATCH_MNT/foo | _filter_scratch
_unmount_flakey
_load_flakey_table $FLAKEY_ALLOW_WRITES
# During mount, we'll replay the log created by the fsync above, and the file's
# md5 digest should be the same we got before the unmount.
_mount_flakey
md5sum $SCRATCH_MNT/foo | _filter_scratch
_unmount_flakey
MOUNT_OPTIONS="$SAVE_MOUNT_OPTIONS"
status=0
exit
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
If the log sync fails, there is something wrong in the log tree, we
should not continue to join the log transaction and log the metadata.
What we should do is to do a full commit.
This patch fixes this problem by setting ->last_trans_log_full_commit
to the current transaction id, it will tell the tasks not to join
the log transaction, and do a full commit.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
We might commit the log sub-transaction which didn't contain the metadata we
logged. It was because we didn't record the log transid and just select
the current log sub-transaction to commit, but the right one might be
committed by the other task already. Actually, we needn't do anything
and it is safe that we go back directly in this case.
This patch improves the log sync by the above idea. We record the transid
of the log sub-transaction in which we log the metadata, and the transid
of the log sub-transaction we have committed. If the committed transid
is >= the transid we record when logging the metadata, we just go back.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
It is possible that many tasks sync the log tree at the same time, but
only one task can do the sync work, the others will wait for it. But those
wait tasks didn't get the result of the log sync, and returned 0 when they
ended the wait. It caused those tasks skipped the error handle, and the
serious problem was they told the users the file sync succeeded but in
fact they failed.
This patch fixes this problem by introducing a log context structure,
we insert it into the a global list. When the sync fails, we will set
the error number of every log context in the list, then the waiting tasks
get the error number of the log context and handle the error if need.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
The log trans id is initialized to be 0 every time we create a log tree,
and the log tree need be re-created after a new transaction is started,
it means the log trans id is unlikely to be a huge number, so we can use
signed integer instead of unsigned long integer to save a bit space.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
Mutex unlock implies certain memory barriers to make sure all the memory
operation completes before the unlock, and the next mutex lock implies memory
barriers to make sure the all the memory happens after the lock. So it is
a full memory barrier(smp_mb), we needn't add memory barriers. Remove them.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
The old code would start the log transaction even the log tree init
failed, it was unnecessary. Fix it.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
We may abort the wait earlier if ->last_trans_log_full_commit was set to
the current transaction id, at this case, we need commit the current
transaction instead of the log sub-transaction. But the current code
didn't tell the caller to do it (return 0, not -EAGAIN). Fix it.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
->last_trans_log_full_commit may be changed by the other tasks without lock,
so we need prevent the compiler from the optimize access just like
tmp = fs_info->last_trans_log_full_commit
if (tmp == ...)
...
<do something>
if (tmp == ...)
...
In fact, we need get the new value of ->last_trans_log_full_commit during
the second access. Fix it by ACCESS_ONCE().
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
There was a problem in the old code:
If we failed to log the csum, we would free all the ordered extents in the log list
including those ordered extents that were logged successfully, it would make the
log committer not to wait for the completion of the ordered extents.
This patch doesn't insert the ordered extents that is about to be logged into
a global list, instead, we insert them into a local list. If we log the ordered
extents successfully, we splice them with the global list, or we will throw them
away, then do full sync. It can also reduce the lock contention and the traverse
time of list.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
If we truncate an uncompressed inline item, ram_bytes isn't updated to reflect
the new size. The fixe uses the size directly from the item header when
reading uncompressed inlines, and also fixes truncate to update the
size as it goes.
Reported-by: Jens Axboe <axboe@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>
CC: stable@vger.kernel.org
The performance of fsync dropped down suddenly sometimes, the main reason
of this problem was that we might only flush part dirty pages in a ordered
extent, then got that ordered extent, wait for the csum calcucation. But if
no task flushed the left part, we would wait until the flusher flushed them,
sometimes we need wait for several seconds, it made the performance drop
down suddenly. (On my box, it drop down from 56MB/s to 4-10MB/s)
This patch improves the above problem by flushing left dirty pages aggressively.
Test Environment:
CPU: 2CPU * 2Cores
Memory: 4GB
Partition: 20GB(HDD)
Test Command:
# sysbench --num-threads=8 --test=fileio --file-num=1 \
> --file-total-size=8G --file-block-size=32768 \
> --file-io-mode=sync --file-fsync-freq=100 \
> --file-fsync-end=no --max-requests=10000 \
> --file-test-mode=rndwr run
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>
When writing to a file we drop existing file extent items that cover the
write range and then add a new file extent item that represents that write
range.
Before this change we were doing a tree lookup to remove the file extent
items, and then after we did another tree lookup to insert the new file
extent item.
Most of the time all the file extent items we need to drop are located
within a single leaf - this is the leaf where our new file extent item ends
up at. Therefore, in this common case just combine these 2 operations into
a single one.
By avoiding the second btree navigation for insertion of the new file extent
item, we reduce btree node/leaf lock acquisitions/releases, btree block/leaf
COW operations, CPU time on btree node/leaf key binary searches, etc.
Besides for file writes, this is an operation that happens for file fsync's
as well. However log btrees are much less likely to big as big as regular
fs btrees, therefore the impact of this change is smaller.
The following benchmark was performed against an SSD drive and a
HDD drive, both for random and sequential writes:
sysbench --test=fileio --file-num=4096 --file-total-size=8G \
--file-test-mode=[rndwr|seqwr] --num-threads=512 \
--file-block-size=8192 \ --max-requests=1000000 \
--file-fsync-freq=0 --file-io-mode=sync [prepare|run]
All results below are averages of 10 runs of the respective test.
** SSD sequential writes
Before this change: 225.88 Mb/sec
After this change: 277.26 Mb/sec
** SSD random writes
Before this change: 49.91 Mb/sec
After this change: 56.39 Mb/sec
** HDD sequential writes
Before this change: 68.53 Mb/sec
After this change: 69.87 Mb/sec
** HDD random writes
Before this change: 13.04 Mb/sec
After this change: 14.39 Mb/sec
Signed-off-by: Filipe David Borba Manana <fdmanana@gmail.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>
This is the third step in bootstrapping the btrfs_find_item interface.
The function find_orphan_item(), in orphan.c, is similar to the two
functions already replaced by the new interface. It uses two parameters,
which are already present in the interface, and is nearly identical to
the function brought in in the previous patch.
Replace the two calls to find_orphan_item() with calls to
btrfs_find_item(), with the defined objectid and type that was used
internally by find_orphan_item(), a null path, and a null key. Add a
test for a null path to btrfs_find_item, and if it passes, allocate and
free the path. Finally, remove find_orphan_item().
Signed-off-by: Kelley Nielsen <kelleynnn@gmail.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Chris Mason <clm@fb.com>
Btrfs has always had these filler extent data items for holes in inodes. This
has made somethings very easy, like logging hole punches and sending hole
punches. However for large holey files these extent data items are pure
overhead. So add an incompatible feature to no longer add hole extents to
reduce the amount of metadata used by these sort of files. This has a few
changes for logging and send obviously since they will need to detect holes and
log/send the holes if there are any. I've tested this thoroughly with xfstests
and it doesn't cause any issues with and without the incompat format set.
Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Chris Mason <clm@fb.com>
If we fsync, seek and write, rename and then fsync again we will lose the
modified hole extent because the rename will drop all of the modified extents
since we didn't do the fast search. We need to only drop the modified extents
if we didn't do the fast search and we were logging the entire inode as we don't
need them anymore, otherwise this is being premature. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
If we rename a file that is already in the log and we fsync again we will lose
the new name. This is because we just log the inode update and not the new ref.
To fix this we just need to check if we are logging the new name of the inode
and copy all the metadata instead of just updating the inode itself. With this
patch my testcase now passes. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
Use WARN_ON()'s return value in place of WARN_ON(1) for cleaner source
code that outputs a more descriptive warnings. Also fix the styling
warning of redundant braces that came up as a result of this fix.
Signed-off-by: Dulshani Gunawardhana <dulshani.gunawardhana89@gmail.com>
Reviewed-by: Zach Brown <zab@redhat.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>