Re: [RFC PATCH v3 06/10] sched: Add over-utilization/tipping point indicator

[Dietmar Eggemann]

On 06/19/2018 09:01 AM, Pavan Kondeti wrote:
> On Mon, May 21, 2018 at 03:25:01PM +0100, Quentin Perret wrote:

[...]

> > @@ -8152,6 +8176,9 @@ static inline void update_sg_lb_stats(struct lb_env *env,
> >   		if (nr_running > 1)
> >   			*overload = true;
> >   
> > +		if (cpu_overutilized(i))
> > +			*overutilized = 1;
> > +
>
> There is no need to check if every CPU is overutilized or not once
> *overutilized is marked as true, right?

True, so you want to check *overutilized before calling 
cpu_overutilized() to save a little bit on compute?

[...]

> > @@ -8586,6 +8621,10 @@ static struct sched_group *find_busiest_group(struct lb_env *env)
> >   	 * this level.
> >   	 */
> >   	update_sd_lb_stats(env, &sds);
> > +
> > +	if (sched_energy_enabled() && !READ_ONCE(env->dst_rq->rd->overutilized))
> > +		goto out_balanced;
> > +
>
> Is there any reason for sending no-hz idle kicks but bailing out here when
> system is not overutilized?

Even if a system is not-overutilized, we want to update stale cpu 
blocked load and utilization so NOHZ_STATS_KICK have to get through.

So calling find_busiest_group() -> update_sd_lb_stats() -> 
update_sg_lb_stats() to possibly execute update_nohz_stats() is IMHO the 
right thing to do.