Re: [patch 1/2] sched: Change the nohz ilb logic from pull to push model

[Peter Zijlstra <peterz@infradead.org>]

On Mon, 2009-12-21 at 13:13 +0100, Peter Zijlstra wrote:
> 
> > +     ret = atomic_cmpxchg(&nohz.first_pick_cpu, -1, cpu);
> > +     if (ret == -1 || ret == cpu) {
> > +             atomic_cmpxchg(&nohz.second_pick_cpu, cpu, -1);
> > +             if (rq->nr_running > 1)
> > +                     return 1;
> > +     } else {
> > +             ret = atomic_cmpxchg(&nohz.second_pick_cpu, -1, cpu);
> > +             if (ret == -1 || ret == cpu) {
> > +                     if (rq->nr_running)
> > +                             return 1;
> >               }
> >       }
> 
> Looked very funny, and took a while to understand why you're doing that,
> but yeah, I can't see a better way of doing it either. 

That is, the sanest way to write that is to do something like:

 weight(~nohz & online) == 1 && nr_running == 1

except that with the recent cpumask blowout that's a very expensive op.