Re: [v4.8-rc1 Regression] sched/fair: Apply more PELT fixes

[Peter Zijlstra]

On Mon, Oct 17, 2016 at 12:49:55PM +0100, Dietmar Eggemann wrote:

> > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > index 8b03fb5..8926685 100644
> > --- a/kernel/sched/fair.c
> > +++ b/kernel/sched/fair.c
> > @@ -2902,7 +2902,8 @@ __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
> >   */
> >  static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
> >  {
> > -	long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
> > +	unsigned long load_avg = READ_ONCE(cfs_rq->avg.load_avg);
> > +	long delta = load_avg - cfs_rq->tg_load_avg_contrib;
> >  
> >  	/*
> >  	 * No need to update load_avg for root_task_group as it is not used.
> > @@ -2912,7 +2913,7 @@ static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
> >  
> >  	if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
> >  		atomic_long_add(delta, &cfs_rq->tg->load_avg);
> > -		cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
> > +		cfs_rq->tg_load_avg_contrib = load_avg;
> >  	}
> >  }
> 
> Tested it on an Ubuntu 16.10 Server (on top of the default 4.8.0-22-generic
> kernel) on a Lenovo T430 and it didn't help.

Right, I don't think that race exists, we update cfs_rq->avg.load_avg
with rq->lock held and have it held here, so it cannot change under us.

This might do with a few lockdep_assert_held() instances to clarify this
though.

> What seems to cure it is to get rid of this snippet (part of the commit
> mentioned earlier in this thread: 3d30544f0212):
> 
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index 039de34f1521..16c692049fbf 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -726,7 +726,6 @@ void post_init_entity_util_avg(struct sched_entity *se)
>         struct sched_avg *sa = &se->avg;
>         long cap = (long)(SCHED_CAPACITY_SCALE - cfs_rq->avg.util_avg) / 2;
>         u64 now = cfs_rq_clock_task(cfs_rq);
> -       int tg_update;
>  
>         if (cap > 0) {
>                 if (cfs_rq->avg.util_avg != 0) {
> @@ -759,10 +758,8 @@ void post_init_entity_util_avg(struct sched_entity *se)
>                 }
>         }
>  
> -       tg_update = update_cfs_rq_load_avg(now, cfs_rq, false);
> +       update_cfs_rq_load_avg(now, cfs_rq, false);
>         attach_entity_load_avg(cfs_rq, se);
> -       if (tg_update)
> -               update_tg_load_avg(cfs_rq, false);
>  }
>  
>  #else /* !CONFIG_SMP */
> 
> BTW, I guess we can reach .tg_load_avg up to ~300000-400000 on such a system
> initially because systemd will create all ~100 services (and therefore the
> corresponding 2. level tg's) at once. In my previous example, there was 500ms
> between the creation of 2 tg's so there was a lot of decaying going on in between.

Cute... on current kernels that translates to simply removing the call
to update_tg_load_avg(), lets see if we can figure out what goes
sideways first though, because it _should_ decay back out. And if that
can fail here, I'm not seeing why that wouldn't fail elsewhere either.

I'll see if I can reproduce this with a script creating heaps of cgroups
in a hurry, I have a total lack of system-disease on all my machines.


/me goes prod..