Re: xfs deadlock on mm-unstable kernel?

[Dave Chinner <david@fromorbit.com>]

On Tue, Nov 12, 2024 at 06:14:28PM +0100, Sebastian Andrzej Siewior wrote:
> On 2024-07-08 20:14:44 [+1000], Dave Chinner wrote:
> > On Mon, Jul 08, 2024 at 04:36:08PM +0800, Alex Shi wrote:
> > >   372.297234][ T3001] ============================================
> > > [  372.297530][ T3001] WARNING: possible recursive locking detected
> > > [  372.297827][ T3001] 6.10.0-rc6-00453-g2be3de2b70e6 #64 Not tainted
> > > [  372.298137][ T3001] --------------------------------------------
> > > [  372.298436][ T3001] cc1/3001 is trying to acquire lock:
> > > [  372.298701][ T3001] ffff88802cb910d8 (&xfs_dir_ilock_class){++++}-{3:3}, at: xfs_reclaim_inode+0x59e/0x710
> > > [  372.299242][ T3001] 
> > > [  372.299242][ T3001] but task is already holding lock:
> > > [  372.299679][ T3001] ffff88800e145e58 (&xfs_dir_ilock_class){++++}-{3:3}, at: xfs_ilock_data_map_shared+0x4d/0x60
> > > [  372.300258][ T3001] 
> > > [  372.300258][ T3001] other info that might help us debug this:
> > > [  372.300650][ T3001]  Possible unsafe locking scenario:
> > > [  372.300650][ T3001] 
> > > [  372.301031][ T3001]        CPU0
> > > [  372.301231][ T3001]        ----
> > > [  372.301386][ T3001]   lock(&xfs_dir_ilock_class);
> > > [  372.301623][ T3001]   lock(&xfs_dir_ilock_class);
> > > [  372.301860][ T3001] 
> > > [  372.301860][ T3001]  *** DEADLOCK ***
> > > [  372.301860][ T3001] 
> > > [  372.302325][ T3001]  May be due to missing lock nesting notation
> > > [  372.302325][ T3001] 
> > > [  372.302723][ T3001] 3 locks held by cc1/3001:
> > > [  372.302944][ T3001]  #0: ffff88800e146078 (&inode->i_sb->s_type->i_mutex_dir_key){++++}-{3:3}, at: walk_component+0x2a5/0x500
> > > [  372.303554][ T3001]  #1: ffff88800e145e58 (&xfs_dir_ilock_class){++++}-{3:3}, at: xfs_ilock_data_map_shared+0x4d/0x60
> > > [  372.304183][ T3001]  #2: ffff8880040190e0 (&type->s_umount_key#48){++++}-{3:3}, at: super_cache_scan+0x82/0x4e0
> > 
> > False positive. Inodes above allocation must be actively referenced,
> > and inodes accees by xfs_reclaim_inode() must have no references and
> > been evicted and destroyed by the VFS. So there is no way that an
> > unreferenced inode being locked for reclaim in xfs_reclaim_inode()
> > can deadlock against the refrenced inode locked by the inode lookup
> > code.
> > 
> > Unfortunately, we don't have enough lockdep subclasses available to
> > annotate this correctly - we're already using all
> > MAX_LOCKDEP_SUBCLASSES to tell lockdep about all the ways we can
> > nest inode locks. That leaves us no space to add a "reclaim"
> > annotation for locking done from super_cache_scan() paths that would
> > avoid these false positives....
> 
> So the former inode (the one triggering the reclaim) is created and can
> not be the same as the one in reclaim list. Couldn't we assign it a
> different lock-class?

We've done that in the past. The problem with that is we lose lock
ordering verification across reclaim. i.e. inode lock ordering must
be the same both above and below reclaim, and changing the lock
class loses the ability to verify this.

This is important to us as some code (e.g. extent removal) can be
called from both above and below reclaim, and they require the same
transaction and inode lock contexts to be held regardless of where
they are called from...

-Dave.
-- 
Dave Chinner
david@fromorbit.com