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

[Paul E. McKenney]

On Wed, Nov 30, 2016 at 01:31:37PM +0100, Paul Menzel wrote:
> On 11/30/16 12:54, Paul E. McKenney wrote:
> > On Wed, Nov 30, 2016 at 03:53:20AM -0800, Paul E. McKenney wrote:
> >> On Wed, Nov 30, 2016 at 12:09:44PM +0100, Michal Hocko wrote:
> >>> [CCing Paul]
> >>>
> >>> On Wed 30-11-16 11:28:34, Donald Buczek wrote:
> >>> [...]
> >>>> shrink_active_list gets and releases the spinlock and calls cond_resched().
> >>>> This should give other tasks a chance to run. Just as an experiment, I'm
> >>>> trying
> >>>>
> >>>> --- a/mm/vmscan.c
> >>>> +++ b/mm/vmscan.c
> >>>> @@ -1921,7 +1921,7 @@ static void shrink_active_list(unsigned long
> >>>> nr_to_scan,
> >>>>         spin_unlock_irq(&pgdat->lru_lock);
> >>>>
> >>>>         while (!list_empty(&l_hold)) {
> >>>> -               cond_resched();
> >>>> +               cond_resched_rcu_qs();
> >>>>                 page = lru_to_page(&l_hold);
> >>>>                 list_del(&page->lru);
> >>>>
> >>>> and didn't hit a rcu_sched warning for >21 hours uptime now. We'll see.
> >>>
> >>> This is really interesting! Is it possible that the RCU stall detector
> >>> is somehow confused?
> >>
> >> No, it is not confused.  Again, cond_resched() is not a quiescent
> >> state unless it does a context switch.  Therefore, if the task running
> >> in that loop was the only runnable task on its CPU, cond_resched()
> >> would -never- provide RCU with a quiescent state.
> >>
> >> In contrast, cond_resched_rcu_qs() unconditionally provides RCU
> >> with a quiescent state (hence the _rcu_qs in its name), regardless
> >> of whether or not a context switch happens.
> >>
> >> It is therefore expected behavior that this change might prevent
> >> RCU CPU stall warnings.
> > 
> > I should add...  This assumes that CONFIG_PREEMPT=n.  So what is
> > CONFIG_PREEMPT?
> 
> Itâs not selected.
> 
> ```
> # CONFIG_PREEMPT is not set
> ```

Thank you for the info!

As noted elsewhere in this thread, there are other ways to get stalls,
including the long irq-disabled execution that Michal suspects.

							Thanx, Paul

> >>>> Is preemption disabled for another reason?
> >>>
> >>> I do not think so. I will have to double check the code but this is a
> >>> standard sleepable context. Just wondering what is the PREEMPT
> >>> configuration here?
> 
> 
> Kind regards,
> 
> Paul
>