Re: [PATCH] sched: keep quiescent cpu out of idle balance loop

[Peter Zijlstra <peterz@infradead.org>]

On Thu, Feb 20, 2014 at 10:42:51AM +0800, Lei Wen wrote:
> >> -     int ilb = cpumask_first(nohz.idle_cpus_mask);
> >> +     int ilb;
> >> +     int cpu = smp_processor_id();
> >> +     struct sched_domain *tmp;
> >>
> >> -     if (ilb < nr_cpu_ids && idle_cpu(ilb))
> >> -             return ilb;
> >> +     for_each_domain(cpu, tmp) {
> >> +             ilb = cpumask_first_and(nohz.idle_cpus_mask,
> >> +                             sched_domain_span(tmp));
> >> +             if (ilb < nr_cpu_ids && idle_cpu(ilb))
> >> +                     return ilb;
> >> +     }
> >
> > The ILB code is bad; but you just made it horrible. Don't add pointless
> > for_each_domain() iterations.
> >
> > I'm thinking something like:
> >
> >   ilb = cpumask_first_and(nohz.idle_cpus_mask, this_rq()->rd.span);
> >
> > Should work just fine, no?
> 
> Yes, it has the same result as my previous patch did.
> 
> >
> > Better still would be to maybe not participate in the ILB in the first
> > place and leave this selection loop alone.
> 
> Not quitely get your point here...
> Do you mean that you want idle cpu selection be put in earlier place
> than current find_new_ilb is?

I meant that if you stop an idle CPU setting its bit in
nohz.idle_cpus_mask, you don't have to mask it out either.