Re: XFS corruption on post 3.7 tree.

[Dave Chinner <david@fromorbit.com>]

On Thu, Dec 13, 2012 at 04:41:19PM -0600, Ben Myers wrote:
> Hi Dave,
> 
> On Thu, Dec 13, 2012 at 05:10:57PM -0500, Dave Jones wrote:
> > On Thu, Dec 13, 2012 at 03:55:22PM -0500, Dave Jones wrote:
> >  > Doing a kernel build while running on a 3.7+ tree from last night and I hit this...
> >  > 
> >  > 
> >  > [22637.787422] XFS: Internal error XFS_WANT_CORRUPTED_RETURN at line 163 of file fs/xfs/xfs_dir2_data.c.  Caller 0xffffffffa070086a
> 
> Looks like the dir v2 verifier found that a single block directory had a data
> entry without a corresponding leaf entry in the block.

Actually, a data entry with a corresponding name hash entry. i.e.
the data entry should contain XFS_DIR2_DATA_FREE_TAG, not contain a
dirent....

> > I unmounted, remounted, unmounted, and then ran xfs_repair on it, as prompted.
> > xfs_repair noted..
> > 
> > bad hash table for directory inode 201328949 (bad stale count): rebuilding

And that indicates that the header count of data and stale/free
entries does not add up. That is, it found more less free entries
than it shoul dhave, which means there was at least one entry that
didn't have a XFS_DIR2_DATA_FREE_TAG value when it should have.
That matches up precisely with the problem the write verifier
reported.

> Interesting!

Very!  The new metadata write verifiers appear to have exposed an
existing silent directory corruption within a day of going upstream.
:)

Now to try to find the needle in very complex haystack. :/

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com