attr fork related fstests failures on for-next

attr fork related fstests failures on for-next

[Brian Foster]

Hi,

I'm seeing a couple different fstests failures on current for-next that
appear to be associated with e6a688c33238 ("xfs: initialise attr fork on
inode create"). The first is xfs_check complaining about sb versionnum
bits on various tests:

generic/003 16s ... _check_xfs_filesystem: filesystem on /dev/mapper/test-scratch is inconsistent (c)
(see /root/xfstests-dev/results//generic/003.full for details)
# cat results/generic/003.full
...
_check_xfs_filesystem: filesystem on /dev/mapper/test-scratch is inconsistent (c)
*** xfs_check output ***
sb versionnum missing attr bit 10
*** end xfs_check output
...
#

With xfs_check bypassed, repair eventually complains about some attr
forks. The first point I hit this variant is generic/117:

generic/117 9s ... _check_xfs_filesystem: filesystem on /dev/mapper/test-scratch is inconsistent (r)
(see /root/xfstests-dev/results//generic/117.full for details)
# cat results//generic/117.full
...
_check_xfs_filesystem: filesystem on /dev/mapper/test-scratch is inconsistent (r)
*** xfs_repair -n output ***
...
Phase 3 - for each AG...
        - scan (but don't clear) agi unlinked lists...
        - process known inodes and perform inode discovery...
        - agno = 0
bad attr fork offset 24 in dev inode 135, should be 1
would have cleared inode 135
bad attr fork offset 24 in dev inode 142, should be 1
would have cleared inode 142
...

Both problems disappear with e6a688c33238 reverted.

Brian

Re: attr fork related fstests failures on for-next

[Darrick J. Wong]

On Mon, Mar 29, 2021 at 02:16:04PM -0400, Brian Foster wrote:
> Hi,
> 
> I'm seeing a couple different fstests failures on current for-next that
> appear to be associated with e6a688c33238 ("xfs: initialise attr fork on
> inode create"). The first is xfs_check complaining about sb versionnum
> bits on various tests:
> 
> generic/003 16s ... _check_xfs_filesystem: filesystem on /dev/mapper/test-scratch is inconsistent (c)
> (see /root/xfstests-dev/results//generic/003.full for details)
> # cat results/generic/003.full
> ...
> _check_xfs_filesystem: filesystem on /dev/mapper/test-scratch is inconsistent (c)
> *** xfs_check output ***
> sb versionnum missing attr bit 10
> *** end xfs_check output

FWIW I think this because that commit sets up an attr fork without
setting ATTR and ATTR2 in sb_version like xfs_bmap_add_attrfork does...

> ...
> #
> 
> With xfs_check bypassed, repair eventually complains about some attr
> forks. The first point I hit this variant is generic/117:
> 
> generic/117 9s ... _check_xfs_filesystem: filesystem on /dev/mapper/test-scratch is inconsistent (r)
> (see /root/xfstests-dev/results//generic/117.full for details)
> # cat results//generic/117.full
> ...
> _check_xfs_filesystem: filesystem on /dev/mapper/test-scratch is inconsistent (r)
> *** xfs_repair -n output ***
> ...
> Phase 3 - for each AG...
>         - scan (but don't clear) agi unlinked lists...
>         - process known inodes and perform inode discovery...
>         - agno = 0
> bad attr fork offset 24 in dev inode 135, should be 1
> would have cleared inode 135
> bad attr fork offset 24 in dev inode 142, should be 1
> would have cleared inode 142

...and I think this is because xfs_default_attroffset doesn't set the
correct value for device files.  For those kinds of files, xfs_repair
requires the data fork to be exactly 8 bytes.

--D

> ...
> 
> Both problems disappear with e6a688c33238 reverted.
> 
> Brian
> 

Re: attr fork related fstests failures on for-next

[Dave Chinner]

On Mon, Mar 29, 2021 at 11:31:20AM -0700, Darrick J. Wong wrote:
> On Mon, Mar 29, 2021 at 02:16:04PM -0400, Brian Foster wrote:
> > Hi,
> > 
> > I'm seeing a couple different fstests failures on current for-next that
> > appear to be associated with e6a688c33238 ("xfs: initialise attr fork on
> > inode create"). The first is xfs_check complaining about sb versionnum
> > bits on various tests:
> > 
> > generic/003 16s ... _check_xfs_filesystem: filesystem on /dev/mapper/test-scratch is inconsistent (c)
> > (see /root/xfstests-dev/results//generic/003.full for details)
> > # cat results/generic/003.full
> > ...
> > _check_xfs_filesystem: filesystem on /dev/mapper/test-scratch is inconsistent (c)
> > *** xfs_check output ***
> > sb versionnum missing attr bit 10
> > *** end xfs_check output
> 
> FWIW I think this because that commit sets up an attr fork without
> setting ATTR and ATTR2 in sb_version like xfs_bmap_add_attrfork does...

Maybe, but mkfs.xfs sets ATTR2 by default and has for a long time.
I'm not seeing this fail on either v4 or v5 filesystems on for-next
with a current xfsprogs (5.11.0), so what environment is this test
failing in?

SECTION       -- xfs
FSTYP         -- xfs (debug)
PLATFORM      -- Linux/x86_64 test3 5.12.0-rc5-dgc+ #3074 SMP
PREEMPT Tue Mar 30 07:37:47 AEDT 2021
MKFS_OPTIONS  -- -f -m rmapbt=1,reflink=1 -i sparse=1 /dev/pmem1
MOUNT_OPTIONS -- /dev/pmem1 /mnt/scratch

generic/003 11s ...  11s
Passed all 1 tests
Xunit report: /home/dave/src/xfstests-dev/results//xfs/result.xml

SECTION       -- xfs_v4
FSTYP         -- xfs (debug)
PLATFORM      -- Linux/x86_64 test3 5.12.0-rc5-dgc+ #3074 SMP
PREEMPT Tue Mar 30 07:37:47 AEDT 2021
MKFS_OPTIONS  -- -f -m crc=0 /dev/pmem1
MOUNT_OPTIONS -- /dev/pmem1 /mnt/scratch

generic/003 11s ...  11s
Passed all 1 tests

> > With xfs_check bypassed, repair eventually complains about some attr
> > forks. The first point I hit this variant is generic/117:
> > 
> > generic/117 9s ... _check_xfs_filesystem: filesystem on /dev/mapper/test-scratch is inconsistent (r)
> > (see /root/xfstests-dev/results//generic/117.full for details)
> > # cat results//generic/117.full
> > ...
> > _check_xfs_filesystem: filesystem on /dev/mapper/test-scratch is inconsistent (r)
> > *** xfs_repair -n output ***
> > ...
> > Phase 3 - for each AG...
> >         - scan (but don't clear) agi unlinked lists...
> >         - process known inodes and perform inode discovery...
> >         - agno = 0
> > bad attr fork offset 24 in dev inode 135, should be 1
> > would have cleared inode 135
> > bad attr fork offset 24 in dev inode 142, should be 1
> > would have cleared inode 142
> 
> ...and I think this is because xfs_default_attroffset doesn't set the
> correct value for device files.  For those kinds of files, xfs_repair
> requires the data fork to be exactly 8 bytes.

Again, doesn't fail with xfsprogs 5.11.0 here for either v4 or v5
filesystems...

SECTION       -- xfs
FSTYP         -- xfs (debug)
PLATFORM      -- Linux/x86_64 test3 5.12.0-rc5-dgc+ #3074 SMP
PREEMPT Tue Mar 30 07:37:47 AEDT 2021
MKFS_OPTIONS  -- -f -m rmapbt=1,reflink=1 -i sparse=1 /dev/pmem1
MOUNT_OPTIONS -- /dev/pmem1 /mnt/scratch

generic/117 1s ...  2s
Passed all 1 tests
Xunit report: /home/dave/src/xfstests-dev/results//xfs/result.xml

SECTION       -- xfs_v4
FSTYP         -- xfs (debug)
PLATFORM      -- Linux/x86_64 test3 5.12.0-rc5-dgc+ #3074 SMP
PREEMPT Tue Mar 30 07:37:47 AEDT 2021
MKFS_OPTIONS  -- -f -m crc=0 /dev/pmem1
MOUNT_OPTIONS -- /dev/pmem1 /mnt/scratch

generic/117 2s ...  2s
Passed all 1 tests

I'm going to need more information on what environment these
failures are being generated in.

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com

Re: attr fork related fstests failures on for-next

[Dave Chinner]

On Tue, Mar 30, 2021 at 07:48:28AM +1100, Dave Chinner wrote:
> On Mon, Mar 29, 2021 at 11:31:20AM -0700, Darrick J. Wong wrote:
> > On Mon, Mar 29, 2021 at 02:16:04PM -0400, Brian Foster wrote:
> > > Hi,
> > > 
> > > I'm seeing a couple different fstests failures on current for-next that
> > > appear to be associated with e6a688c33238 ("xfs: initialise attr fork on
> > > inode create"). The first is xfs_check complaining about sb versionnum
> > > bits on various tests:
> > > 
> > > generic/003 16s ... _check_xfs_filesystem: filesystem on /dev/mapper/test-scratch is inconsistent (c)
> > > (see /root/xfstests-dev/results//generic/003.full for details)
> > > # cat results/generic/003.full
> > > ...
> > > _check_xfs_filesystem: filesystem on /dev/mapper/test-scratch is inconsistent (c)
> > > *** xfs_check output ***
> > > sb versionnum missing attr bit 10
> > > *** end xfs_check output
> > 
> > FWIW I think this because that commit sets up an attr fork without
> > setting ATTR and ATTR2 in sb_version like xfs_bmap_add_attrfork does...
> 
> Maybe, but mkfs.xfs sets ATTR2 by default and has for a long time.
> I'm not seeing this fail on either v4 or v5 filesystems on for-next
> with a current xfsprogs (5.11.0), so what environment is this test
> failing in?
> 
> SECTION       -- xfs
> FSTYP         -- xfs (debug)
> PLATFORM      -- Linux/x86_64 test3 5.12.0-rc5-dgc+ #3074 SMP
> PREEMPT Tue Mar 30 07:37:47 AEDT 2021
> MKFS_OPTIONS  -- -f -m rmapbt=1,reflink=1 -i sparse=1 /dev/pmem1
> MOUNT_OPTIONS -- /dev/pmem1 /mnt/scratch

Ok, this regression test VM had selinux set to permissive so it
should have been using selinux. But at some time in the past,
"selinux=0" had been added to the kernel CLI, hence turning it off
and so not actually testing this path.  I have a mix of selinux
enabled and disabled test VMs (because test matrix) and it looks
like this never made it to a VM that had selinux enabled...

Ok, I can reproduce it now, will fix.

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com

Re: attr fork related fstests failures on for-next

[Darrick J. Wong]

On Tue, Mar 30, 2021 at 07:48:28AM +1100, Dave Chinner wrote:
> On Mon, Mar 29, 2021 at 11:31:20AM -0700, Darrick J. Wong wrote:
> > On Mon, Mar 29, 2021 at 02:16:04PM -0400, Brian Foster wrote:
> > > Hi,
> > > 
> > > I'm seeing a couple different fstests failures on current for-next that
> > > appear to be associated with e6a688c33238 ("xfs: initialise attr fork on
> > > inode create"). The first is xfs_check complaining about sb versionnum
> > > bits on various tests:
> > > 
> > > generic/003 16s ... _check_xfs_filesystem: filesystem on /dev/mapper/test-scratch is inconsistent (c)
> > > (see /root/xfstests-dev/results//generic/003.full for details)
> > > # cat results/generic/003.full
> > > ...
> > > _check_xfs_filesystem: filesystem on /dev/mapper/test-scratch is inconsistent (c)
> > > *** xfs_check output ***
> > > sb versionnum missing attr bit 10
> > > *** end xfs_check output
> > 
> > FWIW I think this because that commit sets up an attr fork without
> > setting ATTR and ATTR2 in sb_version like xfs_bmap_add_attrfork does...
> 
> Maybe, but mkfs.xfs sets ATTR2 by default and has for a long time.

The xfs_check regression is a result of xfs_db being too stupid to
recognize ATTR2.

> I'm not seeing this fail on either v4 or v5 filesystems on for-next
> with a current xfsprogs (5.11.0), so what environment is this test
> failing in?

I /think/ any environment where xfs_create_need_xattr returns true is
enough to reproduce it; I triggered it by making that function reproduce
unconditionally and kicking off anything that runs mknod to create a
block device inode.

--D

> SECTION       -- xfs
> FSTYP         -- xfs (debug)
> PLATFORM      -- Linux/x86_64 test3 5.12.0-rc5-dgc+ #3074 SMP
> PREEMPT Tue Mar 30 07:37:47 AEDT 2021
> MKFS_OPTIONS  -- -f -m rmapbt=1,reflink=1 -i sparse=1 /dev/pmem1
> MOUNT_OPTIONS -- /dev/pmem1 /mnt/scratch
> 
> generic/003 11s ...  11s
> Passed all 1 tests
> Xunit report: /home/dave/src/xfstests-dev/results//xfs/result.xml
> 
> SECTION       -- xfs_v4
> FSTYP         -- xfs (debug)
> PLATFORM      -- Linux/x86_64 test3 5.12.0-rc5-dgc+ #3074 SMP
> PREEMPT Tue Mar 30 07:37:47 AEDT 2021
> MKFS_OPTIONS  -- -f -m crc=0 /dev/pmem1
> MOUNT_OPTIONS -- /dev/pmem1 /mnt/scratch
> 
> generic/003 11s ...  11s
> Passed all 1 tests
> 
> > > With xfs_check bypassed, repair eventually complains about some attr
> > > forks. The first point I hit this variant is generic/117:
> > > 
> > > generic/117 9s ... _check_xfs_filesystem: filesystem on /dev/mapper/test-scratch is inconsistent (r)
> > > (see /root/xfstests-dev/results//generic/117.full for details)
> > > # cat results//generic/117.full
> > > ...
> > > _check_xfs_filesystem: filesystem on /dev/mapper/test-scratch is inconsistent (r)
> > > *** xfs_repair -n output ***
> > > ...
> > > Phase 3 - for each AG...
> > >         - scan (but don't clear) agi unlinked lists...
> > >         - process known inodes and perform inode discovery...
> > >         - agno = 0
> > > bad attr fork offset 24 in dev inode 135, should be 1
> > > would have cleared inode 135
> > > bad attr fork offset 24 in dev inode 142, should be 1
> > > would have cleared inode 142
> > 
> > ...and I think this is because xfs_default_attroffset doesn't set the
> > correct value for device files.  For those kinds of files, xfs_repair
> > requires the data fork to be exactly 8 bytes.
> 
> Again, doesn't fail with xfsprogs 5.11.0 here for either v4 or v5
> filesystems...
> 
> SECTION       -- xfs
> FSTYP         -- xfs (debug)
> PLATFORM      -- Linux/x86_64 test3 5.12.0-rc5-dgc+ #3074 SMP
> PREEMPT Tue Mar 30 07:37:47 AEDT 2021
> MKFS_OPTIONS  -- -f -m rmapbt=1,reflink=1 -i sparse=1 /dev/pmem1
> MOUNT_OPTIONS -- /dev/pmem1 /mnt/scratch
> 
> generic/117 1s ...  2s
> Passed all 1 tests
> Xunit report: /home/dave/src/xfstests-dev/results//xfs/result.xml
> 
> SECTION       -- xfs_v4
> FSTYP         -- xfs (debug)
> PLATFORM      -- Linux/x86_64 test3 5.12.0-rc5-dgc+ #3074 SMP
> PREEMPT Tue Mar 30 07:37:47 AEDT 2021
> MKFS_OPTIONS  -- -f -m crc=0 /dev/pmem1
> MOUNT_OPTIONS -- /dev/pmem1 /mnt/scratch
> 
> generic/117 2s ...  2s
> Passed all 1 tests
> 
> I'm going to need more information on what environment these
> failures are being generated in.
> 
> Cheers,
> 
> Dave.
> -- 
> Dave Chinner
> david@fromorbit.com