Re: [PATCH] sched: fix clear NOHZ_BALANCE_KICK

[Frederic Weisbecker <fweisbec@gmail.com>]

On Tue, Jun 04, 2013 at 01:48:47PM +0200, Vincent Guittot wrote:
> On 4 June 2013 13:19, Frederic Weisbecker <fweisbec@gmail.com> wrote:
> > On Tue, Jun 04, 2013 at 01:11:47PM +0200, Vincent Guittot wrote:
> >> On 4 June 2013 12:26, Frederic Weisbecker <fweisbec@gmail.com> wrote:
> >> > On Tue, Jun 04, 2013 at 11:36:11AM +0200, Peter Zijlstra wrote:
> >> >>
> >> >> The best I can seem to come up with is something like the below; but I think
> >> >> its ghastly. Surely we can do something saner with that bit.
> >> >>
> >> >> Having to clear it at 3 different places is just wrong.
> >> >
> >> > We could clear the flag early in scheduler_ipi() and set some
> >> > specific value in rq->idle_balance that tells we want nohz idle
> >> > balancing from the softirq, something like this untested:
> >>
> >> I'm not sure that we can have less than 2 places to clear it: cancel
> >> place or acknowledge place otherwise we can face a situation where
> >> idle load balance will be triggered 2 consecutive times because
> >> NOHZ_BALANCE_KICK will be cleared before the idle load balance has
> >> been done and had a chance to migrate tasks.
> >
> > I guess it depends what is the minimum value of rq->next_balance, it seems
> > to be large enough to avoid this kind of incident. Although I don't
> > know well the whole logic with rq->next_balance and ilb trigger so I must
> > defer to you.
> 
> In the trace that was showing the issue, i can see that both CPU0 and
> CPU1 were trying to trig ILB almost simultaneously and the
> test_and_set NOHZ_BALANCE_KICK filters one request so i would say that
> clearing the bit before the end of the idle load balance sequence can
> generate such sequence

I see.

> 
> In the sequence below, i have minimized the clear of NOHZ_BALANCE_KICK
> in 2 places : acknowledge and cancel. I have reused part of the
> proposal from peter which clears the bit if the condition doesn't
> match but i have reordered the tests to done that only if all other
> condition are matching
> 
>  static inline bool got_nohz_idle_kick(void)
>  {
> - int cpu = smp_processor_id();
> - return idle_cpu(cpu) && test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
> + bool nohz_kick = test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
> +
> +       if (!nohz_kick)
> +               return false;
> +
> +       if (idle_cpu(cpu) && !need_resched())
> +               return true;
> +
> +       clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
> +       return false;
>  }
> 
>  #else /* CONFIG_NO_HZ_COMMON */
> @@ -1393,8 +1401,9 @@ 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()))
> + if (llist_empty(&this_rq()->wake_list)
> + && !tick_nohz_full_cpu(smp_processor_id())
> + && !got_nohz_idle_kick())
>   return;

But we still need got_nohz_idle_kick() to be the first check, don't we? Otherwise
if we run an "idle -> quick task slice -> idle" sequence we may keep the flag
but lose the notifying IPI in between.