Re: rcu_preempt detected stalls.

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Mon, Oct 27, 2014 at 04:44:25PM -0700, Paul E. McKenney wrote:
> On Mon, Oct 27, 2014 at 02:13:29PM -0700, Paul E. McKenney wrote:
> > On Fri, Oct 24, 2014 at 12:39:15PM -0400, Sasha Levin wrote:
> > > On 10/24/2014 12:13 PM, Paul E. McKenney wrote:
> > > > On Fri, Oct 24, 2014 at 08:28:40AM -0400, Sasha Levin wrote:
> > > >> > On 10/23/2014 03:58 PM, Paul E. McKenney wrote:
> > > >>> > > On Thu, Oct 23, 2014 at 02:55:43PM -0400, Sasha Levin wrote:
> > > >>>>> > >> > On 10/23/2014 02:39 PM, Paul E. McKenney wrote:
> > > >>>>>>> > >>> > > On Tue, Oct 14, 2014 at 10:35:10PM -0400, Sasha Levin wrote:
> > > >>>>>>>>> > >>>> > >> On 10/13/2014 01:35 PM, Dave Jones wrote:
> > > >>>>>>>>>>> > >>>>> > >>> oday in "rcu stall while fuzzing" news:
> > > >>>>>>>>>>> > >>>>> > >>>
> > > >>>>>>>>>>> > >>>>> > >>> INFO: rcu_preempt detected stalls on CPUs/tasks:
> > > >>>>>>>>>>> > >>>>> > >>> 	Tasks blocked on level-0 rcu_node (CPUs 0-3): P766 P646
> > > >>>>>>>>>>> > >>>>> > >>> 	Tasks blocked on level-0 rcu_node (CPUs 0-3): P766 P646
> > > >>>>>>>>>>> > >>>>> > >>> 	(detected by 0, t=6502 jiffies, g=75434, c=75433, q=0)
> > > >>>>>>>>> > >>>> > >>
> > > >>>>>>>>> > >>>> > >> I've complained about RCU stalls couple days ago (in a different context)
> > > >>>>>>>>> > >>>> > >> on -next. I guess whatever causing them made it into Linus's tree?
> > > >>>>>>>>> > >>>> > >>
> > > >>>>>>>>> > >>>> > >> https://lkml.org/lkml/2014/10/11/64
> > > >>>>>>> > >>> > > 
> > > >>>>>>> > >>> > > And on that one, I must confess that I don't see where the RCU read-side
> > > >>>>>>> > >>> > > critical section might be.
> > > >>>>>>> > >>> > > 
> > > >>>>>>> > >>> > > Hmmm...  Maybe someone forgot to put an rcu_read_unlock() somewhere.
> > > >>>>>>> > >>> > > Can you reproduce this with CONFIG_PROVE_RCU=y?
> > > >>>>> > >> > 
> > > >>>>> > >> > Paul, if that was directed to me - Yes, I see stalls with CONFIG_PROVE_RCU
> > > >>>>> > >> > set and nothing else is showing up before/after that.
> > > >>> > > Indeed it was directed to you.  ;-)
> > > >>> > > 
> > > >>> > > Does the following crude diagnostic patch turn up anything?
> > > >> > 
> > > >> > Nope, seeing stalls but not seeing that pr_err() you added.
> > > > OK, color me confused.  Could you please send me the full dmesg or a
> > > > pointer to it?
> > > 
> > > Attached.
> > 
> > Thank you!  I would complain about the FAULT_INJECTION messages, but
> > they don't appear to be happening all that frequently.
> > 
> > The stack dumps do look different here.  I suspect that this is a real
> > issue in the VM code.
> 
> And to that end...  The filemap_map_pages() function does have loop over
> a list of pages.  I wonder if the rcu_read_lock() should be moved into
> the radix_tree_for_each_slot() loop.  CCing linux-mm for their thoughts,
> though it looks to me like the current radix_tree_for_each_slot() wants
> to be under RCU protection.  But I am not seeing anything that requires
> all iterations of the loop to be under the same RCU read-side critical
> section.  Maybe something like the following patch?

Just following up, did the patch below help?

							Thanx, Paul

> ------------------------------------------------------------------------
> 
> mm: Attempted fix for RCU CPU stall warning
> 
> It appears that filemap_map_pages() can stay in a single RCU read-side
> critical section for a very long time if given a large area to map.
> This could result in RCU CPU stall warnings.  This commit therefore breaks
> the read-side critical section into per-iteration critical sections, taking
> care to make sure that the radix_tree_for_each_slot() call itself remains
> in an RCU read-side critical section, as required.
> 
> Reported-by: Sasha Levin <sasha.levin@oracle.com>
> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
> 
> diff --git a/mm/filemap.c b/mm/filemap.c
> index 14b4642279f1..f78f144fb41f 100644
> --- a/mm/filemap.c
> +++ b/mm/filemap.c
> @@ -2055,6 +2055,8 @@ skip:
>  next:
>  		if (iter.index == vmf->max_pgoff)
>  			break;
> +		rcu_read_unlock();
> +		rcu_read_lock();
>  	}
>  	rcu_read_unlock();
>  }

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>