lkml.org 
[lkml]   [2013]   [Jun]   [4]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH] sched: fix clear NOHZ_BALANCE_KICK
On Tue, Jun 04, 2013 at 01:15:10PM +0200, Peter Zijlstra wrote:
> On Tue, Jun 04, 2013 at 12:26:22PM +0200, Frederic Weisbecker wrote:
> > @@ -1393,8 +1392,12 @@ static void sched_ttwu_pending(void)
> >
> > void scheduler_ipi(void)
> > {
> > - if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick()
> > - && !tick_nohz_full_cpu(smp_processor_id()))
> > + int cpu = smp_processor_id();
> > + bool idle_kick = got_nohz_idle_kick(cpu);
>
> This puts an unconditional atomic instruction in the IPI path.
> if (test) clear();
> is lots cheaper, esp. since most IPIs won't have this flag set.

Agreed but I'm a bit worried about ordering:

CPU 0 CPU 1

test_and_set_bit(nohz_kick, CPU 1) scheduler_ipi
smp_send_reschedule(CPU 1) if (test_and_clear_bit(nohz_kick))
do_something

I'm not sure what base guarantee we have with ordering against raw IPIs such as the
the scheduler ipi. But unless both IPI trigger and IPI receive imply a full barrier
(or just IPI receive implies read barrier, it seems that's all we need), we need
test_and_set_bit() or smp_rmb()/smp_mb__before_clear_bit() && smp_mb__after_clear_bit().

>
> > +
> > + if (!(idle_kick && idle_cpu(cpu))
> > + && llist_empty(&this_rq()->wake_list)
> > + && !tick_nohz_full_cpu(cpu)
>
> What's with this weird operator first split style?

Yeah ugly, I'll fix.

>
> > return;
> >
> > /*
>
> > +enum idle_balance_type {
> > + IDLE_BALANCE = 1,
> > + IDLE_NOHZ_BALANCE = 2,
> > +};
>
> You might want to update the rq->idle_balance assignment in
> scheduler_tick() to make sure it uses the right value (it does now, but
> there's nothing stopping people from changing the values).

Agreed!

Thanks.


\
 
 \ /
  Last update: 2013-06-04 14:21    [W:0.423 / U:0.096 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site