On Wed 05-10-16 18:44:32, Oleg Nesterov wrote:
> On 10/05, Dave Chinner wrote:
> >
> > On Tue, Oct 04, 2016 at 01:43:43PM +0200, Oleg Nesterov wrote:
> >
> > > plus the following warnings:
> > >
> > > [ 1894.500040] run fstests generic/070 at 2016-10-04 05:03:39
> > > [ 1895.076655] =================================
> > > [ 1895.077136] [ INFO: inconsistent lock state ]
> > > [ 1895.077574] 4.8.0 #1 Not tainted
> > > [ 1895.077900] ---------------------------------
> > > [ 1895.078330] inconsistent {IN-RECLAIM_FS-W} -> {RECLAIM_FS-ON-W}
> > > usage.
> > > [ 1895.078993] fsstress/18239 [HC0[0]:SC0[0]:HE1:SE1] takes:
> > > [ 1895.079522] (&xfs_nondir_ilock_class){++++?-}, at:
> > > [<ffffffffc049ad45>] xfs_ilock+0x165/0x210 [xfs]
> > > [ 1895.080529] {IN-RECLAIM_FS-W} state was registered at:
> >
> > And that is a bug in the lockdep annotations for memory allocation because
> > it
> > fails to take into account the current task flags that are set via
> > memalloc_noio_save() to prevent vmalloc from doing GFP_KERNEL allocations.
> > i.e.
> > in _xfs_buf_map_pages():
>
> OK, I see...
>
> I'll re-test with the following change:
>
> --- a/kernel/locking/lockdep.c
> +++ b/kernel/locking/lockdep.c
> @@ -2867,7 +2867,7 @@ static void __lockdep_trace_alloc(gfp_t gfp_mask,
> unsigned long flags)
> return;
>
> /* We're only interested __GFP_FS allocations for now */
> - if (!(gfp_mask & __GFP_FS))
> + if ((curr->flags & PF_MEMALLOC_NOIO) || !(gfp_mask & __GFP_FS))
> return;
>
>
> Hmm. This is off-topic and most probably I missed something... but at
> first glance we can simplify/improve the reclaim-fs lockdep annotations:
>
> 1. add the global "struct lockdep_map reclaim_fs_map"
>
> 2. change __lockdep_trace_alloc
>
> - mark_held_locks(curr, RECLAIM_FS);
> + lock_map_acquire(&reclaim_fs_map);
> + lock_map_release(&reclaim_fs_map);
>
> 3. turn lockdep_set/clear_current_reclaim_state() into
>
> void lockdep_set_current_reclaim_state(gfp_t gfp_mask)
> {
> if (gfp_mask & __GFP_FS)
> lock_map_acquire(&reclaim_fs_map);
> }
>
> void lockdep_clear_current_reclaim_state(gfp_t gfp_mask)
> {
> if (gfp_mask & __GFP_FS)
> lock_map_release(&reclaim_fs_map);
> }
>
> and now we can remove task_struct->lockdep_reclaim_gfp and all other
> RECLAIM_FS hacks in lockdep.c. Plus we can easily extend this logic to
> check more GFP_ flags.
Yeah, looks possible to me. I've added Peter to CC since he's most likely
to know.
Honza
--
Jan Kara <j...@suse.com>
SUSE Labs, CR
