On Tue, Aug 21, 2018 at 08:15:06AM +1000, Dave Chinner wrote:
> On Mon, Aug 20, 2018 at 11:59:18AM -0400, Brian Foster wrote:
> > On Mon, Aug 20, 2018 at 08:36:26AM -0700, Darrick J. Wong wrote:
> > > On Mon, Aug 20, 2018 at 10:27:42AM -0500, Eric Sandeen wrote:
> > > > On 8/20/18 10:06 AM, Brian Foster wrote:
> > > > > On Tue, Jul 24, 2018 at 09:43:46AM -0700, Darrick J. Wong wrote:
> > > > >> On Mon, Jul 23, 2018 at 11:39:53PM -0700, Eric Sandeen wrote:
> > > > >>> On 6/4/18 11:24 PM, Dave Chinner wrote:
> > > > >>>> From: Dave Chinner <dchinner@xxxxxxxxxx>
> > > > >>>>
> > > > >>>> There are rules for vald extent size hints. We enforce them when
> > > > >>>> applications set them, but fuzzers violate those rules and that
> > > > >>>> screws us over.
> > > > >>>>
> > > > >>>> This results in alignment assertion failures when setting up
> > > > >>>> allocations such as this in direct IO:
> > > > >>>>
> > > > >>>> XFS: Assertion failed: ap->length, file: fs/xfs/libxfs/xfs_bmap.c, line: 3432
> > > > >>>> ....
> > > > >>>> Call Trace:
> > > > >>>> xfs_bmap_btalloc+0x415/0x910
> > > > >>>> xfs_bmapi_write+0x71c/0x12e0
> > > > >>>> xfs_iomap_write_direct+0x2a9/0x420
> > > > >>>> xfs_file_iomap_begin+0x4dc/0xa70
> > > > >>>> iomap_apply+0x43/0x100
> > > > >>>> iomap_file_buffered_write+0x62/0x90
> > > > >>>> xfs_file_buffered_aio_write+0xba/0x300
> > > > >>>> __vfs_write+0xd5/0x150
> > > > >>>> vfs_write+0xb6/0x180
> > > > >>>> ksys_write+0x45/0xa0
> > > > >>>> do_syscall_64+0x5a/0x180
> > > > >>>> entry_SYSCALL_64_after_hwframe+0x49/0xbe
> > > > >>>>
> > > > >>>> And from xfs_db:
> > > > >>>>
> > > > >>>> core.extsize = 10380288
> > > > >>>>
> > > > >>>> Which is not an integer multiple of the block size, and so violates
> > > > >>>> Rule #7 for setting extent size hints. Validate extent size hint
> > > > >>>> rules in the inode verifier to catch this.
> > > > >>>
> > > > >>> So, I think that if I do:
> > > > >>>
> > > > >>> # mkfs.xfs -f -m crc=0 $TEST_DEV
> > > > >>> # ./check xfs/229
> > > > >>> # ./check xfs/229
> > > > >>>
> > > > >>> I trip the verifier, because I end up with freed inodes on disk with an
> > > > >>> extent size hints but zeroed flags.
> > > > >>>
> > > > >>> xfs_ifree sets di_flags = 0 but doesn't clear di_extsize; xfs_inode_validate_extsize
> > > > >>> says if extsize !=0 and the hint flag is set, it fails
> > > > >>>
> > > > >>> Anyone else see this?
> > > > >>
> > > > >> Yeah, I think I just hit this on the TEST_DEV in xfs/242.
> > > > >>
> > > > >> git blame says I lifted the code from the scrub code, and I probably
> > > > >> wrote the code having read the ioctl code (which clears the extsize
> > > > >> field if the iflag isn't set).
> > > > >>
> > > > >>> (crc=0 needed because that causes us to actually reread the inode chunks
> > > > >>> in xfs_iread vs. /* shortcut IO on inode allocation if possible */
> > > > >>
> > > > >> Hmmm, so a v5 fs mounted with ikeep will also read an inode chunk when
> > > > >> creating an inode. It looks like we do that (instead of zeroing the
> > > > >> incore inode and setting a random i_generation) to preserve the existing
> > > > >> generation number?
> > > > >>
> > > > >> In any case, it's pretty clear that kernels have been writing out freed
> > > > >> inode cores with di_mode == 0, di_flags == 0, and di_extsize == (some
> > > > >> number) so we clearly can't have that in the verifier. It looks like we
> > > > >> only examine di_extsize if either EXTSZ flag are set, so it's not
> > > > >> causing incorrect behavior. Maybe it can be a preening fix in
> > > > >> scrub/repair.
> > > > >>
> > > > >
> > > > > I just stumbled on this problem with xfs/229 that Eric reported. I'm
> > > > > confused by the comment above regarding this not causing incorrect
> > > > > behavior.
> > > >
> > > > I think Darrick meant that having a nonzero extent size hint on disk
> > > > won't cause incorrect behavior because "we only examine di_extsize if
> > > > either EXTSZ flag are set"
> > >
> > > Yeah, he probably did. :)
> > >
> >
> > Got it, thanks.
> >
> > > I think Brian's suggestion of
> > >
> > > if (i_mode != 0 && !hint && extsize != 0)
> > > barf_error();
> > >
> > > sounds reasonable (having not tested that at all).
> > >
> >
> > I'll run it through xfstests and get it posted if nothing else fails.
> >
> > BTW, do we have a similar issue with the cowextsize hint (assuming
> > v5+ikeep)? It looks like it's cleared similarly in xfs_ialloc(), but I'm
> > not sure if it's cleared somewhere else on free...
>
I should note for the list that we've since determined this was already
fixed in v4.18 [1]. The patch ended up in a common base branch between
what is used for upstream pull requests and XFS' for-next, being left
out of the latter just by accident.
[1] d4a34e1655 ("xfs: properly handle free inodes in extent hint
validators")
> We should clear them on free now, so that we can draw a line in the
> sand for when we can have verifiers check it. e.g. when the next
> feature bit gets introduced, filesystems with that feature bit set
> can also verify the extent size hints are zero on freed inodes
> because we know that kernels supporting that feature always zero
> them on free....
>
That seems fine (and harmless) to me if the goal is ultimately to have
this content clear on-disk. It keeps things consistent for verifiers,
scrub, repair, etc. to not have some bits with required initialized
values and others where we need to accommodate stale data.
Brian
> Cheers,
>
> Dave.
> --
> Dave Chinner
> david@xxxxxxxxxxxxx
