Re: [PATCH V2 2/2] fs/super.c: don't fool lockdep in freeze_super() and thaw_super() paths
From: Oleg Nesterov
Date: Sun Oct 09 2016 - 12:18:26 EST
On 10/08, Dave Chinner wrote:
>
> On Fri, Oct 07, 2016 at 07:15:18PM +0200, Oleg Nesterov wrote:
> > > >
> > > > --- x/fs/xfs/xfs_trans.c
> > > > +++ x/fs/xfs/xfs_trans.c
> > > > @@ -245,7 +245,8 @@ xfs_trans_alloc(
> > > > atomic_inc(&mp->m_active_trans);
> > > >
> > > > tp = kmem_zone_zalloc(xfs_trans_zone,
> > > > - (flags & XFS_TRANS_NOFS) ? KM_NOFS : KM_SLEEP);
> > > > + (flags & (XFS_TRANS_NOFS | XFS_TRANS_NO_WRITECOUNT))
> > > > + ? KM_NOFS : KM_SLEEP);
> > > > tp->t_magic = XFS_TRANS_HEADER_MAGIC;
> > > > tp->t_flags = flags;
> > > > tp->t_mountp = mp;
> > >
> > > Brief examination says caller should set XFS_TRANS_NOFS, not change
> > > the implementation to make XFS_TRANS_NO_WRITECOUNT flag to also mean
> > > XFS_TRANS_NOFS.
> >
> > I didn't mean the change above can fix the problem, and I don't really
> > understand your suggestion.
>
> xfs_syncsb() does:
>
> tp = xfs_trans_alloc(... , XFS_TRANS_NO_WRITECOUNT, ....);
>
> but it's running in a GFP_NOFS context when a freeze is being
> finalised. SO, rather than changing what XFS_TRANS_NO_WRITECOUNT
> does in xfs_trans_alloc(), we should tell it to do a GFP_NOFS
> allocation. i.e.
>
> tp = xfs_trans_alloc(... , XFS_TRANS_NOFS | XFS_TRANS_NO_WRITECOUNT, ....);
Ah. This is clear but I am not sure it is enough,
> > Obviously any GFP_FS allocation in xfs_fs_freeze()
> > paths will trigger the same warning.
>
> Of which there should be none except for that xfs_trans_alloc()
> call.
Really? Again, I can be easily wrong, but when I look at xfs_freeze_fs()
paths I can see
xfs_fs_freeze()->xfs_quiesce_attr()->xfs_log_quiesce()->xfs_log_unmount_write()
->xfs_log_reserve()->xlog_ticket_alloc(KM_SLEEP)
at least. But I can test the change above, perhaps this call chain is
not possible...
> > I added this hack
> >
> > --- a/fs/xfs/xfs_super.c
> > +++ b/fs/xfs/xfs_super.c
> > @@ -1333,10 +1333,15 @@ xfs_fs_freeze(
> > struct super_block *sb)
> > {
> > struct xfs_mount *mp = XFS_M(sb);
> > + int ret;
> >
> > + current->flags |= PF_FSTRANS; // tell kmem_flags_convert() to remove GFP_FS
> > xfs_save_resvblks(mp);
> > xfs_quiesce_attr(mp);
> > - return xfs_sync_sb(mp, true);
> > + ret = xfs_sync_sb(mp, true);
> > + current->flags &= ~PF_FSTRANS;
> > +
> > + return ret;
> > }
>
> /me shudders
don't worry, this debugging change won't escape my testing machine!
> > just for testing purposes and after that I got another warning below. I didn't
> > read it carefully yet, but _at first glance_ it looks like the lock inversion
> > uncovered by 2/2, although I can be easily wrong. cancel_delayed_work_sync(l_work)
> > under sb_internal can hang if xfs_log_worker() waits for this rwsem?`
>
> Actually: I *can't read it*. I've got no fucking clue what lockdep
> is trying to say here. This /looks/ like a lockdep is getting
> confused
I can almost never understand what lockdep tells me, it is too clever for me.
But this time I think it is right.
Suppose that freeze_super() races with xfs_log_worker() callback.
freeze_super() takes sb_internal lock and xfs_log_quiesce() calls
cancel_delayed_work_sync(l_work). This will sleep until xfs_log_worker()
finishes.
xfs_log_worker() does a __GFP_FS alloc, triggers reclaim, and blocks
on the same sb_internal lock. Say, in xfs_do_writepage()->xfs_trans_alloc()
path.
Deadlock. The worker thread can't take sb_internal hold by freeze_super(),
cancel_delayed_work_sync() will sleep forever because xfs_log_worker()
can't finish.
So xfs_log_worker() should run in a GFP_NOFS context too. And perhaps
the change above in xfs_trans_alloc() or in xfs_sync_sb() can help if
it doesn't do other allocatiions, I dunno.
Oleg.