Re: [PATCH] sched: Fix nr_uninterruptible race causing increasing load average

[Phil Auld <pauld@redhat.com>]

Hi Peter,

On Thu, Jul 08, 2021 at 09:26:26AM +0200 Peter Zijlstra wrote:
> On Wed, Jul 07, 2021 at 03:04:57PM -0400, Phil Auld wrote:
> > On systems with weaker memory ordering (e.g. power) commit dbfb089d360b
> > ("sched: Fix loadavg accounting race") causes increasing values of load
> > average (via rq->calc_load_active and calc_load_tasks) due to the wakeup
> > CPU not always seeing the write to task->sched_contributes_to_load in
> > __schedule(). Missing that we fail to decrement nr_uninterruptible when
> > waking up a task which incremented nr_uninterruptible when it slept.
> > 
> > The rq->lock serialization is insufficient across different rq->locks.
> > 
> > Add smp_wmb() to schedule and smp_rmb() before the read in
> > ttwu_do_activate().
> 
> > diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> > index 4ca80df205ce..ced7074716eb 100644
> > --- a/kernel/sched/core.c
> > +++ b/kernel/sched/core.c
> > @@ -2992,6 +2992,8 @@ ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
> >  
> >  	lockdep_assert_held(&rq->lock);
> >  
> > +	/* Pairs with smp_wmb in __schedule() */
> > +	smp_rmb();
> >  	if (p->sched_contributes_to_load)
> >  		rq->nr_uninterruptible--;
> >  
> 
> Is this really needed ?! (this question is a big fat clue the comment is
> insufficient). AFAICT try_to_wake_up() has a LOAD-ACQUIRE on p->on_rq
> and hence the p->sched_contributed_to_load must already happen after.
>

Yes, it is needed.  We've got idle power systems with load average of 530.21.
Calc_load_tasks is 530, and the sum of both nr_uninterruptible and
calc_load_active across all the runqueues is 530. Basically monotonically
non-decreasing load average. With the patch this no longer happens.

We need the sched_contributed_to_load to "happen before" so that it's seen
on the other cpu on wakeup.

> > @@ -5084,6 +5086,11 @@ static void __sched notrace __schedule(bool preempt)
> >  				!(prev_state & TASK_NOLOAD) &&
> >  				!(prev->flags & PF_FROZEN);
> >  
> > +			/*
> > +			 * Make sure the previous write is ordered before p->on_rq etc so
> > +			 * that it is visible to other cpus in the wakeup path (ttwu_do_activate()).
> > +			 */
> > +			smp_wmb();
> >  			if (prev->sched_contributes_to_load)
> >  				rq->nr_uninterruptible++;
> 
> That comment is terrible, look at all the other barrier comments around
> there for clues; in effect you're worrying about:
> 
> 	p->sched_contributes_to_load = X	R1 = p->on_rq
> 	WMB					RMB
> 	p->on_rq = Y				R2 = p->sched_contributes_to_load
> 
> Right?

The only way I can see that decrememnt being missed is if the write to
sched_contributes_to_load is not being seen on the wakeup cpu. 

Before the previous patch the _state condition was checked again on the
wakeup cpu and that is ordered.

> 
> 
> Bah bah bah.. I so detest having to add barriers here for silly
> accounting. Let me think about this a little.
> 
> 

Thanks,
Phil

--