Re: [PATCH v4] sched/freq: move call to cpufreq_update_util

[Peter Zijlstra <peterz@infradead.org>]

On Thu, Nov 14, 2019 at 06:07:31PM +0100, Vincent Guittot wrote:

> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index 69a81a5..3be44e1 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -3504,9 +3504,6 @@ update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
>  	cfs_rq->load_last_update_time_copy = sa->last_update_time;
>  #endif
>  
> -	if (decayed)
> -		cfs_rq_util_change(cfs_rq, 0);
> -
>  	return decayed;
>  }

This removes the call from the for_each_leaf_cfs_rq_safe() loop.

> @@ -7543,18 +7544,19 @@ static void update_blocked_averages(int cpu)
>  	const struct sched_class *curr_class;
>  	struct rq_flags rf;
>  	bool done = true;
> +	int decayed;
>  
>  	rq_lock_irqsave(rq, &rf);
>  	update_rq_clock(rq);
>  
>  	/*
> -	 * update_cfs_rq_load_avg() can call cpufreq_update_util(). Make sure
> -	 * that RT, DL and IRQ signals have been updated before updating CFS.
> +	 * update_load_avg() can call cpufreq_update_util(). Make sure that RT,
> +	 * DL and IRQ signals have been updated before updating CFS.
>  	 */
>  	curr_class = rq->curr->sched_class;
> -	update_rt_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &rt_sched_class);
> -	update_dl_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &dl_sched_class);
> -	update_irq_load_avg(rq, 0);
> +	decayed = update_rt_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &rt_sched_class);
> +	decayed |= update_dl_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &dl_sched_class);
> +	decayed |= update_irq_load_avg(rq, 0);

Should not all 3 have their windows aligned and thus alway return the
exact same value?

>  
>  	/* Don't need periodic decay once load/util_avg are null */
>  	if (others_have_blocked(rq))
> @@ -7567,9 +7569,13 @@ static void update_blocked_averages(int cpu)
>  	for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
>  		struct sched_entity *se;
>  
> -		if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq))
> +		if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq)) {
>  			update_tg_load_avg(cfs_rq, 0);
>  
> +			if (cfs_rq == &rq->cfs)
> +				decayed = 1;

And that restores it.

But should not also rq->cfs's window be aligned with the above 3?
Meaning that this one, with exception of the list_del, covers all 4.

> +		}
> +
>  		/* Propagate pending load changes to the parent, if any: */
>  		se = cfs_rq->tg->se[cpu];
>  		if (se && !skip_blocked_update(se))
> @@ -7588,6 +7594,9 @@ static void update_blocked_averages(int cpu)
>  	}
>  
>  	update_blocked_load_status(rq, !done);
> +
> +	if (decayed)
> +		cpufreq_update_util(rq, 0);
>  	rq_unlock_irqrestore(rq, &rf);
>  }
>  
> @@ -7644,22 +7653,22 @@ static inline void update_blocked_averages(int cpu)
>  	struct cfs_rq *cfs_rq = &rq->cfs;
>  	const struct sched_class *curr_class;
>  	struct rq_flags rf;
> +	int decayed;
>  
>  	rq_lock_irqsave(rq, &rf);
>  	update_rq_clock(rq);
>  
> -	/*
> -	 * update_cfs_rq_load_avg() can call cpufreq_update_util(). Make sure
> -	 * that RT, DL and IRQ signals have been updated before updating CFS.
> -	 */
>  	curr_class = rq->curr->sched_class;
> -	update_rt_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &rt_sched_class);
> -	update_dl_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &dl_sched_class);
> -	update_irq_load_avg(rq, 0);
> +	decayed = update_rt_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &rt_sched_class);
> +	decayed |= update_dl_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &dl_sched_class);
> +	decayed |= update_irq_load_avg(rq, 0);
>  
> -	update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq);
> +	decayed |= update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq);

And that thus this one makes all 3 above redundant.

>  
>  	update_blocked_load_status(rq, cfs_rq_has_blocked(cfs_rq) || others_have_blocked(rq));
> +
> +	if (decayed)
> +		cpufreq_update_util(rq, 0);
>  	rq_unlock_irqrestore(rq, &rf);
>  }

That is, I'm almost tempted to prefer a variant of your initial hack,
that refuses to remove rq->cfs from the list.

That avoids having to care about the rt,dl,irq decays (their windows
align with rq->cfs) and makes smp/up similar.


I still don't actually understand how any of this makes intel_pstate
happy though.