Re: INFO: rcu_sched detected stalls on CPUs/tasks with `kswapd` and `mem_cgroup_shrink_node`

[Donald Buczek]

On 11/21/16 15:29, Paul E. McKenney wrote:
> On Mon, Nov 21, 2016 at 03:18:19PM +0100, Michal Hocko wrote:
> > On Mon 21-11-16 06:01:22, Paul E. McKenney wrote:
> > > On Mon, Nov 21, 2016 at 02:41:31PM +0100, Michal Hocko wrote:
> > [...]
> > > > To the patch. I cannot say I would like it. cond_resched_rcu_qs sounds
> > > > way too lowlevel for this usage. If anything cond_resched somewhere inside
> > > > mem_cgroup_iter would be more appropriate to me.
> > > Like this?
> > >
> > > 							Thanx, Paul
> > >
> > > ------------------------------------------------------------------------
> > >
> > > diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> > > index ae052b5e3315..81cb30d5b2fc 100644
> > > --- a/mm/memcontrol.c
> > > +++ b/mm/memcontrol.c
> > > @@ -867,6 +867,7 @@ struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
> > >   out:
> > >   	if (prev && prev != root)
> > >   		css_put(&prev->css);
> > > +	cond_resched_rcu_qs();
> > I still do not understand why should we play with _rcu_qs at all and a
> > regular cond_resched is not sufficient. Anyway I would have to double
> > check whether we can do cond_resched in the iterator. I do not remember
> > having users which are atomic but I might be easily wrong here. Before
> > we touch this code, though, I would really like to understand what is
> > actually going on here because as I've already pointed out we should
> > have some resched points in the reclaim path.
> If there is a tight loop in the kernel, cond_resched() will ensure that
> other tasks get a chance to run, but if there are no such tasks, it does
> nothing to give RCU the quiescent state that it needs from time to time.
> So if there is a possibility of a long-running in-kernel loop without
> preemption by some other task, cond_resched_rcu_qs() is required.
>
> I welcome your deeper investigation -- I am very much treating symptoms
> here, which might or might not have any relationship to fixing underlying
> problems.
>
> 							Thanx, Paul

Hello,

thanks a lot for looking into this!

Let me add some information from the reporting site:

* We've tried the patch from Paul E. McKenney (the one posted Wed, 16 
Nov 2016)  and it doesn't shut up the rcu stall warnings.

* Log file from a boot with the patch applied ( grep kernel 
/var/log/messages ) is here : 
http://owww.molgen.mpg.de/~buczek/321322/2016-11-21_syslog.txt

* This system is a backup server and walks over thousands of files 
sometimes with multiple parallel rsync processes.

* No rcu_* warnings on that machine with 4.7.2, but with 4.8.4 , 4.8.6 , 
4.8.8 and now 4.9.0-rc5+Pauls patch

* When the backups are actually happening there might be relevant memory 
pressure from inode cache and the rsync processes. We saw the oom-killer 
kick in on another machine with same hardware and similar (a bit higher) 
workload. This other machine also shows a lot of rcu stall warnings 
since 4.8.4.

* We see "rcu_sched detected stalls" also on some other machines since 
we switched to 4.8 but not as frequently as on the two backup servers. 
Usually there's "shrink_node" and "kswapd" on the top of the stack. 
Often "xfs_reclaim_inodes" variants on top of that.

Donald

--
Donald Buczek
buczek@xxxxxxxxxxxxx
Tel: +49 30 8413 1433