Re: [PATCH]sched: stop hrtick timer if running task is switching from fair scheduling class to another

[Oleg Nesterov <oleg@redhat.com>]

On 12/23, Kirill Tkhai wrote:
>
> [PATCH]sched: stop hrtick timer if running task is switching from fair
> scheduling class to another
>
> We have to stop hrtick timer to avoid excess interrupt. Not-fair tasks
> are not interested in fair's hrtick. RT class uses its own fixed
> timeslice (in case of RR), which doesn't depend on current value of hrtick
> timer.

Well. I shouldn't try to comment this patch, I do not really understand
this code.

But since nobody else replies...

> @@ -5271,6 +5271,13 @@ static void switched_from_fair(struct rq *rq,
> struct task_struct *p)
>  		place_entity(cfs_rq, se, 0);
>  		se->vruntime -= cfs_rq->min_vruntime;
>  	}
> +
> +	/*
> +	 * Other scheduling classes are not interested in fair's hrtick timer.
> +	 */
> +	if (task_current(rq, p) && sched_feat(HRTICK))
> +		hrtick_clear(rq);
> +
>  }

May be... but in this case, perhaps instead we should teach
dequeue_task_fair() or put_prev_task_fair() to do this. Then we can
probably remove __schedule()->hrtick_clear().

I simply can't understand the current hrtick logic. For example,
why dequeue_task_fair() does hrtick_update() ? OK, probably
because "nr_running < sched_nr_latency" can become true. But at
least this doesn't make sense to me when p == rq->curr, say,
__schedule() path.

Hmm. In any case, how it is possible to do hrtick_start() with
rq->lock held? hrtimer_restart() may want to wakeup_softirqd().

Peter, Ingo, does this code really work? SCHED_FEAT(HRTICK) == 0
by default, and afaics you can't change it without CONFIG_SCHED_DEBUG.

Confused...

Oleg.