Re: [tip:sched/core] sched: Push down pre_schedule() and idle_balance ()

[Peter Zijlstra <peterz@infradead.org>]

On Wed, Feb 12, 2014 at 02:26:25PM +0800, Michael wang wrote:
> Hi, Peter
> 
> On 02/11/2014 08:17 PM, tip-bot for Peter Zijlstra wrote:
> [snip]
> > +
> > +idle:
> > +#ifdef CONFIG_SMP
> > +	idle_enter_fair(rq);
> > +	/*
> > +	 * We must set idle_stamp _before_ calling idle_balance(), such that we
> > +	 * measure the duration of idle_balance() as idle time.
> > +	 */
> > +	rq->idle_stamp = rq_clock(rq);
> > +	if (idle_balance(rq)) { /* drops rq->lock */
> 
> Since idle_balance() will release the rq lock, will it happen that some
> rt or dl tasks was waken up and enqueued before it hold the lock again?
> 
> Should we recheck 'rq->nr_running == rq->cfs.h_nr_running' here before
> goto pick fair entity to make sure the priority?
> 
> May be like:
> 
> if (idle_balance(rq) &&
> 	rq->nr_running == rq->cfs.h_nr_running)

Yes I think there might be a problem here because of how we re-arranged
things. Let me brew of pot of tea and try to actually wake up.

I suspect we might be good if we clear the need_resched flags before
calling pick_next_task. Then any RT/DL task that gets moved here will
set need resched, and we'll retry the pick_next_task loop.