Re: [PATCH 4/4] sched: Idle task shortcut optimization

On 01/17/2014 03:26 PM, Peter Zijlstra wrote:

On Fri, Jan 17, 2014 at 10:04:04AM +0100, Daniel Lezcano wrote:

...

@@ -2679,11 +2715,8 @@ need_resched:

pre_schedule(rq, prev);

• if (unlikely(!rq->nr_running))

• rq->idle_stamp = idle_balance(rq) ? 0 : rq_clock(rq);
  

• put_prev_task(rq, prev);
• next = pick_next_task(rq);

• next = pick_next_task_or_idle(rq); clear_tsk_need_resched(prev); clear_preempt_need_resched(); rq->skip_clock_update = 0;

I have vague memories that we need to have idle_balance() before put_prev_task(), but I can't recollect why this would be so.

That said, if I resurrect these patches:

https://lkml.org/lkml/2012/6/14/271

I suppose we could write something like:

struct task_struct *pick_next_task(struct rq *rq, struct task_struct *prev) { const struct sched_class *class; struct task_struct *p;

again: if (likely(rq->nr_running)) {

if (likely(rq->nr_running == rq->cfs.h_nr_running))
	return fair_sched_class.pick_next_task(rq, prev);

for_each_class(class) {
	p = class->pick_next_task(rq, prev);
	if (p)
		return p;
}

}

if (idle_balance(rq)) goto again;

rq->idle_stamp = rq_clock(rq);

return idle_sched_class.pick_next_task(rq, prev); }

Which keeps idle_balance() before put_prev_task(), and by using idle_sched_clas.pick_next_task() doesn't rape the idle class interface like you did :-)

But put_prev_task is called before pick_next_task, so idle_balance() is called after now, no ?