generic/561 fails when testing xfs on next-20240506 kernel
Hi,
generic/561 fails when testing XFS on a next-20240506 kernel as shown below,
# ./check generic/561 FSTYP -- xfs (debug) PLATFORM -- Linux/x86_64 xfs-crc-rtdev-extsize-28k 6.9.0-rc7-next-20240506+ #1 SMP PREEMPT_DYNAMIC Mon May 6 07:53:46 GMT 2024 MKFS_OPTIONS -- -f -rrtdev=/dev/loop14 -f -m reflink=0,rmapbt=0, -d rtinherit=1 -r extsize=28k /dev/loop5 MOUNT_OPTIONS -- -o context=system_u:object_r:root_t:s0 -ortdev=/dev/loop14 /dev/loop5 /media/scratch
generic/561 - output mismatch (see /var/lib/xfstests/results/xfs-crc-rtdev-extsize-28k/6.9.0-rc7-next-20240506+/xfs_crc_rtdev_extsize_28k/generic/561.out.bad)
--- tests/generic/561.out 2024-05-06 08:18:09.681430366 +0000
+++ /var/lib/xfstests/results/xfs-crc-rtdev-extsize-28k/6.9.0-rc7-next-20240506+/xfs_crc_rtdev_extsize_28k/generic/561.out.bad 2024-05-08 09:14:24.908010133 +0000
@@ -1,2 +1,5 @@
QA output created by 561
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: FAILED
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: FAILED
+md5sum: WARNING: 2 computed checksums did NOT match
Silence is golden
...
(Run 'diff -u /var/lib/xfstests/tests/generic/561.out /var/lib/xfstests/results/xfs-crc-rtdev-extsize-28k/6.9.0-rc7-next-20240506+/xfs_crc_rtdev_extsize_28k/generic/561.out.bad' to see the entire diff)Ran: generic/561
Failures: generic/561
Failed 1 of 1 tests
The following was the fstest configuration used for the test run,
FSTYP=xfs
TEST_DIR=/media/test
SCRATCH_MNT=/media/scratch
TEST_DEV=/dev/loop16
TEST_LOGDEV=/dev/loop13
SCRATCH_DEV_POOL="/dev/loop5 /dev/loop6 /dev/loop7 /dev/loop8 /dev/loop9 /dev/loop10 /dev/loop11 /dev/loop12" MKFS_OPTIONS='-f -m crc=1,reflink=0,rmapbt=0, -i sparse=0 -lsize=1g'
TEST_FS_MOUNT_OPTS="-o logdev=/dev/loop13"
MOUNT_OPTIONS='-o usrquota,grpquota,prjquota'
TEST_FS_MOUNT_OPTS="$TEST_FS_MOUNT_OPTS -o usrquota,grpquota,prjquota" SCRATCH_LOGDEV=/dev/loop15
USE_EXTERNAL=yes
LOGWRITES_DEV=/dev/loop15
Git bisect produced the following as the first bad commit,
commit 943bc0882cebf482422640924062a7daac5a27ba
Author: Zhang Yi <[email protected]> Date: Wed Mar 20 19:05:45 2024 +0800
iomap: don't increase i_size if it's not a write operation
Increase i_size in iomap_zero_range() and iomap_unshare_iter() is not needed, the caller should handle it. Especially, when truncate partial block, we should not increase i_size beyond the new EOF here. It doesn't affect xfs and gfs2 now because they set the new file size after zero out, it doesn't matter that a transient increase in i_size, but it will affect ext4 because it set file size before truncate. So move the i_size updating logic to iomap_write_iter().
fs/iomap/buffered-io.c | 50 +++++++++++++++++++++++++------------------------- 1 file changed, 25 insertions(+), 25 deletions(-)
Signature · 1 line
Chandan
Quoted 21 lines · Chandan Babu R
Hi,
generic/561 fails when testing XFS on a next-20240506 kernel as shown below,
# ./check generic/561 FSTYP -- xfs (debug) PLATFORM -- Linux/x86_64 xfs-crc-rtdev-extsize-28k 6.9.0-rc7-next-20240506+ #1 SMP PREEMPT_DYNAMIC Mon May 6 07:53:46 GMT 2024 MKFS_OPTIONS -- -f -rrtdev=/dev/loop14 -f -m reflink=0,rmapbt=0, -d rtinherit=1 -r extsize=28k /dev/loop5 MOUNT_OPTIONS -- -o context=system_u:object_r:root_t:s0 -ortdev=/dev/loop14 /dev/loop5 /media/scratchgeneric/561 - output mismatch (see
/var/lib/xfstests/results/xfs-crc-rtdev-extsize-28k/6.9.0-rc7-next-20240506+/xfs_crc_rtdev_extsize_28k/generic/561.out.bad)--- tests/generic/561.out 2024-05-06 08:18:09.681430366 +0000 +++ /var/lib/xfstests/results/xfs-crc-rtdev-extsize-28k/6.9.0-rc7-next-20240506+/xfs_crc_rtdev_extsize_28k/generic/561.out.bad 2024-05-08 09:14:24.908010133 +0000 @@ -1,2 +1,5 @@ QA output created by 561 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: FAILED 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: FAILED +md5sum: WARNING: 2 computed checksums did NOT match Silence is golden ... (Run 'diff -u /var/lib/xfstests/tests/generic/561.out /var/lib/xfstests/results/xfs-crc-rtdev-extsize-28k/6.9.0-rc7-next-20240506+/xfs_crc_rtdev_extsize_28k/generic/561.out.bad' to see the entire diff)
Quoted 4 lines
Ran: generic/561
Failures: generic/561
Failed 1 of 1 tests
Sorry about this regression. After debuging and analyzing the code, I notice that this problem could only happens on xfs realtime inode. The real problem is about realtime extent alignment.
Please assume that if we have a file that contains a written extent [A, D). We unaligned truncate to the file to B, in the middle of this written extent.
A B D
+wwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwww
After the truncate, the i_size is set to B, but due to the sb_rextsize, xfs_itruncate_extents() truncate and aligned the written extent to C, so the data in [B, C) doesn't zeroed and becomes stale.
A B C
+wwwwwwwwwwwwwwSSSSSS
^
EOF
The if we write [E, F) beyond this written extent, xfs_file_write_checks()-> xfs_zero_range() would zero [B, C) in page cache, but since we don't increase i_size in iomap_zero_iter(), the writeback process doesn't write zero data to disk. After write, the data in [B, C) is still stale so once we clear the pagecache, this stale data is exposed.
A B C E F
+wwwwwwwwwwwwwwSSSSSS wwwwwwww
The reason this problem doesn't occur on normal inode is because normal inode doesn't have a post EOF written extent. For realtime inode, I guess it's not enough to just zero the EOF block (xfs_setattr_size()->xfs_truncate_page()), we should also zero the extra blocks that aligned to realtime extent size before updating i_size. Any suggestions?
Thanks,
Yi.
Quoted 42 lines
The following was the fstest configuration used for the test run,
FSTYP=xfs
TEST_DIR=/media/test
SCRATCH_MNT=/media/scratch
TEST_DEV=/dev/loop16
TEST_LOGDEV=/dev/loop13
SCRATCH_DEV_POOL="/dev/loop5/dev/loop6/dev/loop7/dev/loop8/dev/loop9/dev/loop10/dev/loop11/dev/loop12" MKFS_OPTIONS='-f -m crc=1,reflink=0,rmapbt=0, -i sparse=0 -lsize=1g'
TEST_FS_MOUNT_OPTS="-o logdev=/dev/loop13"
MOUNT_OPTIONS='-o usrquota,grpquota,prjquota'
TEST_FS_MOUNT_OPTS="$TEST_FS_MOUNT_OPTS -o usrquota,grpquota,prjquota" SCRATCH_LOGDEV=/dev/loop15
USE_EXTERNAL=yes
LOGWRITES_DEV=/dev/loop15Git bisect produced the following as the first bad commit,
commit
943bc0882cebf482422640924062a7daac5a27baAuthor: Zhang Yi <[email protected]> Date: Wed Mar 20 19:05:45 2024 +0800iomap: don't increase
i_sizeif it's not a write operationIncrease
i_sizeiniomap_zero_range()andiomap_unshare_iter()is not needed, the caller should handle it. Especially, when truncate partial block, we should not increasei_sizebeyond the new EOF here. It doesn't affect xfs and gfs2 now because they set the new file size after zero out, it doesn't matter that a transient increase ini_size, but it will affect ext4 because it set file size before truncate. So move thei_sizeupdating logic toiomap_write_iter().Signed-off-by Zhang Yi <[email protected]>Reviewed-by Christoph Hellwig <[email protected]>Reviewed-by Darrick J. Wong <[email protected]>Signed-off-by Christian Brauner <[email protected]>fs/iomap/buffered-io.c | 50 +++++++++++++++++++++++++------------------------- 1 file changed, 25 insertions(+), 25 deletions(-)
Quoted 37 lines · Zhang Yi
Quoted 21 lines · Chandan Babu R
Hi,
generic/561 fails when testing XFS on a next-20240506 kernel as shown below,
# ./check generic/561 FSTYP -- xfs (debug) PLATFORM -- Linux/x86_64 xfs-crc-rtdev-extsize-28k 6.9.0-rc7-next-20240506+ #1 SMP PREEMPT_DYNAMIC Mon May 6 07:53:46 GMT 2024 MKFS_OPTIONS -- -f -rrtdev=/dev/loop14 -f -m reflink=0,rmapbt=0, -d rtinherit=1 -r extsize=28k /dev/loop5 MOUNT_OPTIONS -- -o context=system_u:object_r:root_t:s0 -ortdev=/dev/loop14 /dev/loop5 /media/scratchgeneric/561 - output mismatch (see
/var/lib/xfstests/results/xfs-crc-rtdev-extsize-28k/6.9.0-rc7-next-20240506+/xfs_crc_rtdev_extsize_28k/generic/561.out.bad)--- tests/generic/561.out 2024-05-06 08:18:09.681430366 +0000 +++ /var/lib/xfstests/results/xfs-crc-rtdev-extsize-28k/6.9.0-rc7-next-20240506+/xfs_crc_rtdev_extsize_28k/generic/561.out.bad 2024-05-08 09:14:24.908010133 +0000 @@ -1,2 +1,5 @@ QA output created by 561 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: FAILED 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: FAILED +md5sum: WARNING: 2 computed checksums did NOT match Silence is golden ... (Run 'diff -u /var/lib/xfstests/tests/generic/561.out /var/lib/xfstests/results/xfs-crc-rtdev-extsize-28k/6.9.0-rc7-next-20240506+/xfs_crc_rtdev_extsize_28k/generic/561.out.bad' to see the entire diff)Quoted 4 lines
Ran: generic/561
Failures: generic/561
Failed 1 of 1 testsSorry about this regression. After debuging and analyzing the code, I notice that this problem could only happens on xfs realtime inode. The real problem is about realtime extent alignment.
Please assume that if we have a file that contains a written extent [A, D). We unaligned truncate to the file to B, in the middle of this written extent.
A B D
+wwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwww
Quoted 8 lines
After the truncate, the
i_sizeis set to B, but due to thesb_rextsize,xfs_itruncate_extents()truncate and aligned the written extent to C, so the data in [B, C) doesn't zeroed and becomes stale.A B C
+wwwwwwwwwwwwwwSSSSSS
^
EOF
This region must be zeroed on disk before we call xfs_itruncate_extents(). i.e completed xfs_setattr_size() via xfs_truncate_page() and flushed to disk before we start removing extents.
The problem is that iomap_truncate_page() only zeros the trailing portion of the i_blocksize() value, which is wrong for realtime devices with rtextsize != fs blocksize.
Further, xfs_setattr_size() then calls truncate_setsize(newsize) before the zeroing has been written back to disk, which means that the flush that occurs immediately after the truncate_setsize() call can not write blocks beyond the new EOF regardless of whether iomap_truncate_page() wrote zeroes to them or not.
Quoted 11 lines
The if we write [E, F) beyond this written extent, xfs_file_write_checks()->
xfs_zero_range()would zero [B, C) in page cache, but since we don't increasei_sizeiniomap_zero_iter(), the writeback process doesn't write zero data to disk. After write, the data in [B, C) is still stale so once we clear the pagecache, this stale data is exposed.A B C E F
+wwwwwwwwwwwwwwSSSSSS wwwwwwwwThe reason this problem doesn't occur on normal inode is because normal inode doesn't have a post EOF written extent.
That's incorrect - we can have post-eof written extents on normal files. The reason this doesn't get exposed for normal files is that the zeroing range used in iomap_truncate_page() covers the entire filesystem block and writeback can write the entire EOF page that covers that block containing the zeroes. Hence when we remove all the written extents beyond EOF later in the truncate, we don't leave any blocks beyond EOF that we haven't zeroed.
Quoted 4 lines
For realtime inode, I guess it's not
enough to just zero the EOF block (xfs_setattr_size()->xfs_truncate_page()), we should also zero the extra blocks that aligned to realtime extent size before updatingi_size. Any suggestions?
Right. xfs_setattr_size() needs fixing to flush the entire zeroed range *before* truncating the page cache and changing the inode size.
Of course, xfs_truncate_page() also needs fixing to zero the entire rtextsize range, not use iomap_truncate_page() which only zeroes to the end of the EOF filesystem block.
I note that dax_truncate_page() has the same problem with RT device block sizes as iomap_truncate_page(), so we need the same fix for both dax and non-dax paths here.
It might actually be easiest to pass the block size for zeroing into iomap_truncate_page() rather than relying on being able to extract the zeroing range from the inode via i_blocksize(). We can't use i_blocksize() for RT files, because inode->i_blkbits and hence i_blocksize() only supports power of 2 block sizes. Changing that is a *much* bigger job, so fixing xfs_truncate_page() is likely the best thing to do right now...
Cheers,
Dave.
Signature · 2 lines
Dave Chinner [email protected]
Quoted 98 lines · Dave Chinner
Quoted 37 lines · Zhang Yi
Quoted 21 lines · Chandan Babu R
Hi,
generic/561 fails when testing XFS on a next-20240506 kernel as shown below,
# ./check generic/561 FSTYP -- xfs (debug) PLATFORM -- Linux/x86_64 xfs-crc-rtdev-extsize-28k 6.9.0-rc7-next-20240506+ #1 SMP PREEMPT_DYNAMIC Mon May 6 07:53:46 GMT 2024 MKFS_OPTIONS -- -f -rrtdev=/dev/loop14 -f -m reflink=0,rmapbt=0, -d rtinherit=1 -r extsize=28k /dev/loop5 MOUNT_OPTIONS -- -o context=system_u:object_r:root_t:s0 -ortdev=/dev/loop14 /dev/loop5 /media/scratchgeneric/561 - output mismatch (see
/var/lib/xfstests/results/xfs-crc-rtdev-extsize-28k/6.9.0-rc7-next-20240506+/xfs_crc_rtdev_extsize_28k/generic/561.out.bad)--- tests/generic/561.out 2024-05-06 08:18:09.681430366 +0000 +++ /var/lib/xfstests/results/xfs-crc-rtdev-extsize-28k/6.9.0-rc7-next-20240506+/xfs_crc_rtdev_extsize_28k/generic/561.out.bad 2024-05-08 09:14:24.908010133 +0000 @@ -1,2 +1,5 @@ QA output created by 561 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: FAILED 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: FAILED +md5sum: WARNING: 2 computed checksums did NOT match Silence is golden ... (Run 'diff -u /var/lib/xfstests/tests/generic/561.out /var/lib/xfstests/results/xfs-crc-rtdev-extsize-28k/6.9.0-rc7-next-20240506+/xfs_crc_rtdev_extsize_28k/generic/561.out.bad' to see the entire diff)Quoted 4 lines
Ran: generic/561
Failures: generic/561
Failed 1 of 1 testsSorry about this regression. After debuging and analyzing the code, I notice that this problem could only happens on xfs realtime inode. The real problem is about realtime extent alignment.
Please assume that if we have a file that contains a written extent [A, D). We unaligned truncate to the file to B, in the middle of this written extent.
A B D
+wwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwQuoted 8 lines
After the truncate, the
i_sizeis set to B, but due to thesb_rextsize,xfs_itruncate_extents()truncate and aligned the written extent to C, so the data in [B, C) doesn't zeroed and becomes stale.A B C
+wwwwwwwwwwwwwwSSSSSS
^
EOFThis region must be zeroed on disk before we call
xfs_itruncate_extents(). i.e completedxfs_setattr_size()viaxfs_truncate_page()and flushed to disk before we start removing extents.The problem is that
iomap_truncate_page()only zeros the trailing portion of thei_blocksize()value, which is wrong for realtime devices with rtextsize != fs blocksize.Further,
xfs_setattr_size()then callstruncate_setsize(newsize)before the zeroing has been written back to disk, which means that the flush that occurs immediately after thetruncate_setsize()call can not write blocks beyond the new EOF regardless of whetheriomap_truncate_page()wrote zeroes to them or not.Quoted 11 lines
The if we write [E, F) beyond this written extent, xfs_file_write_checks()->
xfs_zero_range()would zero [B, C) in page cache, but since we don't increasei_sizeiniomap_zero_iter(), the writeback process doesn't write zero data to disk. After write, the data in [B, C) is still stale so once we clear the pagecache, this stale data is exposed.A B C E F
+wwwwwwwwwwwwwwSSSSSS wwwwwwwwThe reason this problem doesn't occur on normal inode is because normal inode doesn't have a post EOF written extent.
That's incorrect - we can have post-eof written extents on normal files. The reason this doesn't get exposed for normal files is that the zeroing range used in
iomap_truncate_page()covers the entire filesystem block and writeback can write the entire EOF page that covers that block containing the zeroes. Hence when we remove all the written extents beyond EOF later in the truncate, we don't leave any blocks beyond EOF that we haven't zeroed.Quoted 4 lines
For realtime inode, I guess it's not
enough to just zero the EOF block (xfs_setattr_size()->xfs_truncate_page()), we should also zero the extra blocks that aligned to realtime extent size before updatingi_size. Any suggestions?Right.
xfs_setattr_size()needs fixing to flush the entire zeroed range *before* truncating the page cache and changing the inode size.Of course,
xfs_truncate_page()also needs fixing to zero the entire rtextsize range, not useiomap_truncate_page()which only zeroes to the end of the EOF filesystem block.I note that
dax_truncate_page()has the same problem with RT device block sizes asiomap_truncate_page(), so we need the same fix for both dax and non-dax paths here.
Quoted 8 lines
It might actually be easiest to pass the block size for zeroing into
iomap_truncate_page()rather than relying on being able to extract the zeroing range from the inode viai_blocksize(). We can't usei_blocksize()for RT files, becauseinode->i_blkbitsand hencei_blocksize()only supports power of 2 block sizes. Changing that is a *much* bigger job, so fixingxfs_truncate_page()is likely the best thing to do right now...
Thanks for the explanation and suggestion, I agree with you. However, why do you recommend to pass block size for zeroing in to iomap_truncate_page()? It's looks like we could fix xfs_truncate_page() by using iomap_zero_range() and dax_zero_range() instead and don't use iomap_truncate_page() and dax_truncate_page().
Thanks,
Yi.
Quoted 12 lines · Zhang Yi
Quoted 10 lines · Dave Chinner
Quoted 1 line · Zhang Yi
On 2024/5/8 17:01, Chandan Babu R wrote:
It might actually be easiest to pass the block size for zeroing into
iomap_truncate_page()rather than relying on being able to extract the zeroing range from the inode viai_blocksize(). We can't usei_blocksize()for RT files, becauseinode->i_blkbitsand hencei_blocksize()only supports power of 2 block sizes. Changing that is a *much* bigger job, so fixingxfs_truncate_page()is likely the best thing to do right now...
Quoted 5 lines
Thanks for the explanation and suggestion, I agree with you. However, why do you recommend to pass block size for zeroing in to
iomap_truncate_page()? It's looks like we could fixxfs_truncate_page()by usingiomap_zero_range()anddax_zero_range()instead and don't useiomap_truncate_page()anddax_truncate_page().
Fair question. Yes, we could just fix it in XFS.
But then any other filesystem that uses iomap might have the same problem where the allocation block size (and hence the range that needs zeroing) is different to the filesystem block size (e.g. ext4 with bigalloc?). At that point, those filesystem developers need to:
a) be aware of the issue; and
b) implement their own iomap_zero_range() wrapper for the same function.
If the iomap infrastructure doesn't assume block sizes are always i_blocksize() (which is clearly a bad assumption!), then the API itself informs the filesystem developers of the fact they really have to care about using the correct allocation block sizes during EOF zeroing.
At the moment, only XFS uses iomap_truncate_page(), and only ext2 and XFS use dax_truncate_page(). It seems pretty simple to change the infrastructure and document why we don't use i_blocksize() at this point. Especially considering we have forced alignment stuff in XFS coming up which further decouples allocation unit (and hence zeroing range) from the filesystem block size....
-Dave.
Signature · 2 lines
Dave Chinner [email protected]
Quoted 36 lines · Dave Chinner
Quoted 12 lines · Zhang Yi
Quoted 10 lines · Dave Chinner
Quoted 1 line · Zhang Yi
On 2024/5/8 17:01, Chandan Babu R wrote:
It might actually be easiest to pass the block size for zeroing into
iomap_truncate_page()rather than relying on being able to extract the zeroing range from the inode viai_blocksize(). We can't usei_blocksize()for RT files, becauseinode->i_blkbitsand hencei_blocksize()only supports power of 2 block sizes. Changing that is a *much* bigger job, so fixingxfs_truncate_page()is likely the best thing to do right now...Quoted 5 lines
Thanks for the explanation and suggestion, I agree with you. However, why do you recommend to pass block size for zeroing in to
iomap_truncate_page()? It's looks like we could fixxfs_truncate_page()by usingiomap_zero_range()anddax_zero_range()instead and don't useiomap_truncate_page()anddax_truncate_page().Fair question. Yes, we could just fix it in XFS.
But then any other filesystem that uses iomap might have the same problem where the allocation block size (and hence the range that needs zeroing) is different to the filesystem block size (e.g. ext4 with bigalloc?). At that point, those filesystem developers need to:
a) be aware of the issue; and
b) implement their owniomap_zero_range()wrapper for the same function.If the iomap infrastructure doesn't assume block sizes are always
i_blocksize()(which is clearly a bad assumption!), then the API itself informs the filesystem developers of the fact they really have to care about using the correct allocation block sizes during EOF zeroing.
Quoted 7 lines
At the moment, only XFS uses
iomap_truncate_page(), and only ext2 and XFS usedax_truncate_page(). It seems pretty simple to change the infrastructure and document why we don't usei_blocksize()at this point. Especially considering we have forced alignment stuff in XFS coming up which further decouples allocation unit (and hence zeroing range) from the filesystem block size....
ext4 with bigalloc still use block size (not cluster size) to manage allocation extents, so it aligned to the block size when truncate and doesn't have this problem now. But it doesn't matter, your suggestion make sense to me, pass block size to iomap_truncate_page() and dax_truncate_page() could make them to have better compatibility.
Thanks,
Yi.
Quoted 3 lines · Chandan Babu R
Hi,
generic/561 fails when testing XFS on a next-20240506 kernel as shown below,
Hello,
I've send out the fix [1] based on the solution discussed with Dave, and I've tested it on generic/561 several times and -g auto with below 4 types of config, no new failures. I hope I've covered all the corner cases about zero range on xfs, please take a look at that series.
- Default config
- MKFS_OPTIONS="-f -m reflink=0"
- MOUNT_OPTIONS='-o dax'
- USE_EXTERNAL=yes
SCRATCH_RTDEV=/dev/nvme0n1
MKFS_OPTIONS="-f -m reflink=0,rmapbt=0, -d rtinherit=1 -r extsize=28k"
[1] https://lore.kernel.org/linux-xfs/[email protected]/
Thanks,
Yi.
Quoted 64 lines
# ./check generic/561 FSTYP -- xfs (debug) PLATFORM -- Linux/x86_64 xfs-crc-rtdev-extsize-28k 6.9.0-rc7-next-20240506+ #1 SMP PREEMPT_DYNAMIC Mon May 6 07:53:46 GMT 2024 MKFS_OPTIONS -- -f -rrtdev=/dev/loop14 -f -m reflink=0,rmapbt=0, -d rtinherit=1 -r extsize=28k /dev/loop5 MOUNT_OPTIONS -- -o context=system_u:object_r:root_t:s0 -ortdev=/dev/loop14 /dev/loop5 /media/scratchgeneric/561 - output mismatch (see
/var/lib/xfstests/results/xfs-crc-rtdev-extsize-28k/6.9.0-rc7-next-20240506+/xfs_crc_rtdev_extsize_28k/generic/561.out.bad)--- tests/generic/561.out 2024-05-06 08:18:09.681430366 +0000 +++ /var/lib/xfstests/results/xfs-crc-rtdev-extsize-28k/6.9.0-rc7-next-20240506+/xfs_crc_rtdev_extsize_28k/generic/561.out.bad 2024-05-08 09:14:24.908010133 +0000 @@ -1,2 +1,5 @@ QA output created by 561 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: FAILED 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: FAILED +md5sum: WARNING: 2 computed checksums did NOT match Silence is golden ... (Run 'diff -u /var/lib/xfstests/tests/generic/561.out /var/lib/xfstests/results/xfs-crc-rtdev-extsize-28k/6.9.0-rc7-next-20240506+/xfs_crc_rtdev_extsize_28k/generic/561.out.bad' to see the entire diff)Ran: generic/561
Failures: generic/561
Failed 1 of 1 testsThe following was the fstest configuration used for the test run,
FSTYP=xfs
TEST_DIR=/media/test
SCRATCH_MNT=/media/scratch
TEST_DEV=/dev/loop16
TEST_LOGDEV=/dev/loop13
SCRATCH_DEV_POOL="/dev/loop5/dev/loop6/dev/loop7/dev/loop8/dev/loop9/dev/loop10/dev/loop11/dev/loop12" MKFS_OPTIONS='-f -m crc=1,reflink=0,rmapbt=0, -i sparse=0 -lsize=1g'
TEST_FS_MOUNT_OPTS="-o logdev=/dev/loop13"
MOUNT_OPTIONS='-o usrquota,grpquota,prjquota'
TEST_FS_MOUNT_OPTS="$TEST_FS_MOUNT_OPTS -o usrquota,grpquota,prjquota" SCRATCH_LOGDEV=/dev/loop15
USE_EXTERNAL=yes
LOGWRITES_DEV=/dev/loop15Git bisect produced the following as the first bad commit,
commit
943bc0882cebf482422640924062a7daac5a27baAuthor: Zhang Yi <[email protected]> Date: Wed Mar 20 19:05:45 2024 +0800iomap: don't increase
i_sizeif it's not a write operationIncrease
i_sizeiniomap_zero_range()andiomap_unshare_iter()is not needed, the caller should handle it. Especially, when truncate partial block, we should not increasei_sizebeyond the new EOF here. It doesn't affect xfs and gfs2 now because they set the new file size after zero out, it doesn't matter that a transient increase ini_size, but it will affect ext4 because it set file size before truncate. So move thei_sizeupdating logic toiomap_write_iter().Signed-off-by Zhang Yi <[email protected]>Reviewed-by Christoph Hellwig <[email protected]>Reviewed-by Darrick J. Wong <[email protected]>Signed-off-by Christian Brauner <[email protected]>fs/iomap/buffered-io.c | 50 +++++++++++++++++++++++++------------------------- 1 file changed, 25 insertions(+), 25 deletions(-)