On Mon, 14 Apr 2014, Peter Zijlstra wrote:
> On Mon, Apr 14, 2014 at 03:08:36PM -0400, Brian Foster wrote:
> > On Mon, Apr 14, 2014 at 12:43:14PM -0400, Brian Foster wrote:
> > > Hi all,
> > >
> > > This is a heads up that I'm seeing a blatant readdir hang on the current
> > > for-next with selinux enabled. To reproduce, I format a clean fs, mount
> > > and attempt an ls.
> > >
> > > The problem does not occur with selinux disabled, if I back out the
> > > following commit:
> > >
> > > 40194ecc6d78 xfs: reinstate the ilock in xfs_readdir
> > >
> > > ... or if I remove the locking around xfs_attr_get(), so I suspect this
> > > is another instance of a recursive deadlock. I'm getting no output
> > > whatsoever in order to confirm this and it also leads to a complete
> > > system lockup. It's also interesting that this hasn't been observed
> > > until now, given the above commit was introduced in 3.14. So the above
> > > commit doesn't appear to be the most recent change that triggers this.
> > >
> > > I reproduced on the latest linus tree and do not reproduce on 3.14, so
> > > I'm trying to do a bisect to find out what else might have changed to
> > > trigger this.
> > >
> > This bisected down to:
> > commit 6f008e72cd111a119b5d8de8c5438d892aae99eb
> > Author: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
> > Date: Wed Mar 12 13:24:42 2014 +0100
> > locking/mutex: Fix debug checks
> > ...
> > ... which suggests something down in the mutex debug code. Indeed, the
> > problem no longer occurs if I disable kernel debug in my .config. What
> > is also interesting is that it didn't return when I reenable
> > DEBUG_KERNEL and DEBUG_MUTEXES alone. It does return when I start to
> > enable some of the other lock debugging options. FWIW, I also cleared
> > out my tree and rebuilt from scratch just to be sure that I didn't have
> > anything stale/broken lying around.
> > Peter,
> > Any insight on this?
> That will make the kernel continue after the lockdep splat. I too see it
> on some of my XFS using machines.
It can happen on JFS, too, but my trusty "untar a system backup
until a splat happens" test barely worked for the merge-window
kernel. Therefore, I used xfstests generic/113 on XFS (kernel +
xfs-oss/for-next) to cause this situation. The patch above has been
through xfstests on both v4- and v5-superblock XFS, solving any new
lockdep issues down here on x86.
Sorry to not report it here, therefore costing you time in doing a
bisect. The second lockdep splat I got after kernel 3.14 wasn't XFS,
and so I treated it as a non-XFS issue.