XFS reclaim lock order bug

[Peter Zijlstra]

On Thu, 2010-11-25 at 06:25 -0500, Christoph Hellwig wrote:
> On Thu, Nov 25, 2010 at 09:29:40PM +1100, Dave Chinner wrote:
> > Yes, actually it is - see the XFS_IRECLAIMABLE case in
> > xfs_iget_cache_hit(). I guess we haven't seen the original lock
> > inversion false positives that this was supposed to fix because the
> > reclaim warnings trip first...
> > 
> > I think that means we also need to reinitialise the lock when we recycle
> > the inode out of the XFS_IRECLAIMABLE state.
> 
> I came up with the patch below when we had a previous report of the
> warning, but I couldn't convince myself that it really helps:
> 
> Index: linux-2.6/fs/xfs/xfs_iget.c
> ===================================================================
> --- linux-2.6.orig/fs/xfs/xfs_iget.c	2010-09-20 12:10:28.227444173 -0300
> +++ linux-2.6/fs/xfs/xfs_iget.c	2010-09-20 12:11:25.631444190 -0300
> @@ -207,6 +207,10 @@ xfs_iget_cache_hit(
>  
>  			ip->i_flags &= ~XFS_INEW;
>  			ip->i_flags |= XFS_IRECLAIMABLE;
> +
> +			ASSERT(!rwsem_is_locked(&ip->i_iolock.mr_lock));
> +			mrlock_init(&ip->i_iolock, MRLOCK_BARRIER, "xfsio", ip->i_ino);
> +
>  			__xfs_inode_set_reclaim_tag(pag, ip);
>  			trace_xfs_iget_reclaim_fail(ip);
>  			goto out_error;


That adds a 3rd class which should work, but doesn't validate that the
first -- xfs_inode_alloc() and this one are in fact similar.