Re: [mm/readahead] a0b99df1aa: xfstests.xfs.421.fail

From: Dave Chinner
Date: Mon Feb 21 2022 - 16:44:01 EST

On Mon, Feb 21, 2022 at 09:10:20PM +0000, Matthew Wilcox wrote:
> On Tue, Feb 22, 2022 at 07:55:29AM +1100, Dave Chinner wrote:
> > On Mon, Feb 21, 2022 at 01:56:55PM +0000, Matthew Wilcox wrote:
> > > On Mon, Feb 21, 2022 at 04:02:18PM +0800, kernel test robot wrote:
> > > > commit: a0b99df1aa37d714eb80be5fb54efd56c88a3336 ("mm/readahead: Add large folio readahead")
> > >
> > > > xfs/420 - output mismatch (see /lkp/benchmarks/xfstests/results//xfs/420.out.bad)
> > > > --- tests/xfs/420.out 2022-02-17 11:55:00.000000000 +0000
> > > > +++ /lkp/benchmarks/xfstests/results//xfs/420.out.bad 2022-02-20 20:34:22.430378506 +0000
> > > > @@ -13,9 +13,7 @@
> > > > Seek holes and data in file2
> > > > Whence Result
> > > > DATA 0
> > > > -HOLE 131072
> > > > -DATA 196608
> > > > -HOLE 262144
> > > > +HOLE 524288
> > >
> > > Confirm this test now fails. I don't think it's actually a bug,
> > > though. I think the test is now using larger pages to cache the
> > > file, and it fails to report that there's a hole in the file.
> > > Maybe there actually isn't a hole in the file any more; using
> > > larger pages to cache the file means we'll now write more data
> > > than we used to.
> > >
> > > Adding XFS people for their thoughts.
> > >
> > > Complete output:
> > >
> > > $ diff -u ../ktest/tests/xfstests/tests/xfs/420.out ktest-out/xfstests/xfs/420.out.bad
> > > --- ../ktest/tests/xfstests/tests/xfs/420.out 2021-07-05 15:49:45.539887305 -0400
> > > +++ ktest-out/xfstests/xfs/420.out.bad 2022-02-21 08:14:40.000000000 -0500
> > > @@ -13,9 +13,7 @@
> > > Seek holes and data in file2
> > > Whence Result
> > > DATA 0
> > > -HOLE 131072
> > > -DATA 196608
> > > -HOLE 262144
> > > +HOLE 524288
> > > Compare files
> > > c2803804acc9936eef8aab42c119bfac SCRATCH_MNT/test-420/file1
> > > 017c08a9320aad844ce86aa9631afb98 SCRATCH_MNT/test-420/file2
> > > @@ -28,9 +26,7 @@
> > > Seek holes and data in file2
> > > Whence Result
> > > DATA 0
> > > -HOLE 131072
> > > -DATA 196608
> > > -HOLE 262144
> > > +HOLE 524288
> > > Compare files
> > > c2803804acc9936eef8aab42c119bfac SCRATCH_MNT/test-420/file1
> > > 017c08a9320aad844ce86aa9631afb98 SCRATCH_MNT/test-420/file2
> > >
> > > So the file checksums are right, which means I didn't break the COW
> > > functionality. But we're no longer reporting a hole at 128k.
> >
> > Can you post the contents of the 420.full output file so we can see
> > what the output of the various commands that are run are? e.g.
> > things like cowextsize that is configured, etc?
>
> Sure! It's short, so I've included it inline.

Ok, I'll cut this up so it makes sense...

>
>
> meta-data=/dev/sdc isize=512 agcount=4, agsize=3670016 blks
> = sectsz=512 attr=2, projid32bit=1
> = crc=1 finobt=1, sparse=1, rmapbt=1
> = reflink=1 bigtime=0
> data = bsize=1024 blocks=14680064, imaxpct=25
> = sunit=0 swidth=0 blks
> naming =version 2 bsize=4096 ascii-ci=0, ftype=1
> log =internal log bsize=1024 blocks=10240, version=2
> = sectsz=512 sunit=0 blks, lazy-count=1
> realtime =none extsz=4096 blocks=0, rtextents=0
> Discarding blocks...Done.
> [0] /mnt/scratch/test-420
> [524288] /mnt/scratch/test-420
> wrote 131072/131072 bytes at offset 0
> 128 KiB, 128 ops; 0.0000 sec (1.327 GiB/sec and 1391304.3478 ops/sec)
> wrote 524288/524288 bytes at offset 0
> 512 KiB, 512 ops; 0.0003 sec (1.341 GiB/sec and 1406593.4066 ops/sec)
> wrote 131072/131072 bytes at offset 0
> 128 KiB, 128 ops; 0.0000 sec (1.822 GiB/sec and 1910447.7612 ops/sec)
> CoW the shared part then write into the empty part
> [524288] /mnt/scratch/test-420/file1
> [524288] /mnt/scratch/test-420/file2
> wrote 65536/65536 bytes at offset 0
> 64 KiB, 64 ops; 0.0001 sec (416.667 MiB/sec and 426666.6667 ops/sec)
> wrote 65536/65536 bytes at offset 196608
> 64 KiB, 64 ops; 0.0000 sec (1.695 GiB/sec and 1777777.7778 ops/sec)
> wrote 65536/65536 bytes at offset 0
> 64 KiB, 64 ops; 0.0000 sec (1.387 GiB/sec and 1454545.4545 ops/sec)
> wrote 65536/65536 bytes at offset 196608
> 64 KiB, 64 ops; 0.0000 sec (1.526 GiB/sec and 1600000.0000 ops/sec)
> xfs_io: xfsctl(XFS_IOC_GETBMAPX) iflags=0x28 ["/mnt/scratch/test-420/file1"]: Invalid argument
> /mnt/scratch/test-420/file1:
> EXT: FILE-OFFSET BLOCK-RANGE AG AG-OFFSET TOTAL
> 0: [0..255]: 128..383 0 (128..383) 256 100000
> 1: [256..1023]: hole 768
> xfs_io: xfsctl(XFS_IOC_GETBMAPX) iflags=0x28 ["/mnt/scratch/test-420/file2"]: Invalid argument
> /mnt/scratch/test-420/file2:
> EXT: FILE-OFFSET BLOCK-RANGE AG AG-OFFSET TOTAL
> 0: [0..255]: 128..383 0 (128..383) 256 100000
> 1: [256..1023]: hole 768

So this is the extent list for file2 after the reflink. Note the
hole at 128-512kB. The flags tell us the 128kB data extent is
shared. There are no unwritten extents at all.

> > > Seek holes and data in file2
> > > Whence Result
> > > DATA 0
> > > -HOLE 131072
> > > -DATA 196608
> > > -HOLE 262144
> > > +HOLE 524288

indicates the file is completely full of data at this point. Based
on the extent list, and looking at the current
iomap_seek_hole/data() implementation, this looks wrong.

Technically speaking, however, it is valid because holes are allowed
to be reported as data but not vice versa. That said, in this case
there are no unwritten extents so iomap_seek_hole/data should be
reporting the exact on-disk extent mapping and not even be looking
at the page cache contents. The page cache is only only for
unwritten extents, and there are none of them so page cache
footprint should not change this result.

> xfs_io: xfsctl(XFS_IOC_GETBMAPX) iflags=0x28 ["/mnt/scratch/test-420/file3"]: Invalid argument
> /mnt/scratch/test-420/file3:
> EXT: FILE-OFFSET BLOCK-RANGE AG AG-OFFSET TOTAL
> 0: [0..1023]: 512..1535 0 (512..1535) 1024
> sync filesystem
> xfs_io: xfsctl(XFS_IOC_GETBMAPX) iflags=0x28 ["/mnt/scratch/test-420/file1"]: Invalid argument
> /mnt/scratch/test-420/file1:
> EXT: FILE-OFFSET BLOCK-RANGE AG AG-OFFSET TOTAL
> 0: [0..127]: 128..255 0 (128..255) 128
> 1: [128..255]: 256..383 0 (256..383) 128 100000
> 2: [256..1023]: hole 768
> xfs_io: xfsctl(XFS_IOC_GETBMAPX) iflags=0x28 ["/mnt/scratch/test-420/file2"]: Invalid argument
> /mnt/scratch/test-420/file2:
> EXT: FILE-OFFSET BLOCK-RANGE AG AG-OFFSET TOTAL
> 0: [0..127]: 1536..1663 0 (1536..1663) 128
> 1: [128..255]: 256..383 0 (256..383) 128 100000
> 2: [256..383]: hole 128
> 3: [384..511]: 1920..2047 0 (1920..2047) 128
> 4: [512..1023]: hole 512

And the extent list after we've overwritten the first 64kB and
another 64kB write at 192kB. Note that the 64kB hole at 128kB is
still there.

> > > @@ -13,9 +13,7 @@
> > > Seek holes and data in file2
> > > Whence Result
> > > DATA 0
> > > -HOLE 131072
> > > -DATA 196608
> > > -HOLE 262144
> > > +HOLE 524288

And this still looks wrong. Again, no unwritten extents, so we
should just be reporting the on-disk extent layout here.

> xfs_io: xfsctl(XFS_IOC_GETBMAPX) iflags=0x28 ["/mnt/scratch/test-420/file3"]: Invalid argument
> /mnt/scratch/test-420/file3:
> EXT: FILE-OFFSET BLOCK-RANGE AG AG-OFFSET TOTAL
> 0: [0..1023]: 512..1535 0 (512..1535) 1024
> Remount
> xfs_io: xfsctl(XFS_IOC_GETBMAPX) iflags=0x28 ["/mnt/scratch/test-420/file1"]: Invalid argument
> /mnt/scratch/test-420/file1:
> EXT: FILE-OFFSET BLOCK-RANGE AG AG-OFFSET TOTAL
> 0: [0..127]: 128..255 0 (128..255) 128
> 1: [128..255]: 256..383 0 (256..383) 128 100000
> 2: [256..1023]: hole 768
> xfs_io: xfsctl(XFS_IOC_GETBMAPX) iflags=0x28 ["/mnt/scratch/test-420/file2"]: Invalid argument
> /mnt/scratch/test-420/file2:
> EXT: FILE-OFFSET BLOCK-RANGE AG AG-OFFSET TOTAL
> 0: [0..127]: 1536..1663 0 (1536..1663) 128
> 1: [128..255]: 256..383 0 (256..383) 128 100000
> 2: [256..383]: hole 128
> 3: [384..511]: 1920..2047 0 (1920..2047) 128
> 4: [512..1023]: hole 512

This is after mount/unmount, where there is now no page cache over
the file. This reports the same layout as bmap does (i.e. matches
the golden output) and is obviously correct.

This smells of a bug in the large folio patchset as the on-disk
extent layouts are correct and unchanged. The question now is what
if the large folio patchset doing to iomap_seek_hole/data that makes
it behave differently to the current code?

Cheers,

Dave.
--
Dave Chinner
david@xxxxxxxxxxxxx