Re: [PATCH 3/7] perf: Migrate perf to use new tick dependency mask model

[Frederic Weisbecker <fweisbec@gmail.com>]

On Tue, Nov 24, 2015 at 11:19:33AM -0500, Chris Metcalf wrote:
> On 11/13/2015 09:22 AM, Frederic Weisbecker wrote:
> >Instead of providing asynchronous checks for the nohz subsystem to verify
> >perf event tick dependency, migrate perf to the new mask.
> >
> >Perf needs the tick for two situations:
> >
> >1) Freq events. We could set the tick dependency when those are
> >installed on a CPU context. But setting a global dependency on top of
> >the global freq events accounting is much easier. If people want that
> >to be optimized, we can still refine that on the per-CPU tick dependency
> >level. This patch dooesn't change the current behaviour anyway.
> >
> >2) Throttled events: this is a per-cpu dependency.
> >
> >
> >@@ -3540,8 +3530,10 @@ static void unaccount_event(struct perf_event *event)
> >  		atomic_dec(&nr_comm_events);
> >  	if (event->attr.task)
> >  		atomic_dec(&nr_task_events);
> >-	if (event->attr.freq)
> >-		atomic_dec(&nr_freq_events);
> >+	if (event->attr.freq) {
> >+		if (atomic_dec_and_test(&nr_freq_events))
> >+			tick_nohz_clear_dep(TICK_PERF_EVENTS_BIT);
> >+	}
> >  	if (event->attr.context_switch) {
> >  		static_key_slow_dec_deferred(&perf_sched_events);
> >  		atomic_dec(&nr_switch_events);
> >
> >@@ -7695,7 +7687,7 @@ static void account_event(struct perf_event *event)
> >  		atomic_inc(&nr_task_events);
> >  	if (event->attr.freq) {
> >  		if (atomic_inc_return(&nr_freq_events) == 1)
> >-			tick_nohz_full_kick_all();
> >+			tick_nohz_set_dep(TICK_PERF_EVENTS_BIT);
> >  	}
> >  	if (event->attr.context_switch) {
> >  		atomic_inc(&nr_switch_events);
> 
> It would be helpful to have a comment explaining why these two
> can't race with each other, e.g. this race:
> 
> [cpu 1]  atomic_dec_and_test
> [cpu 2] atomic_inc_return
> [cpu 2] tick_nohz_set_dep()
> [cpu 1] tick_nohz_clear_dep()
> 
> Or perhaps this is a true race condition possibility?
> 
> I think we're OK for the sched cases since they're protected under
> the rq lock, I think.  I'm not sure about the POSIX cpu timers.

Hmm, how did I miss that...

So in the case of perf, either we need locking, in which case we may want
to use something like tick_nohz_add_dep() which takes care of counting.
But perf would be the only user.

Another possibility is to rather set/clear the tick mask on the task level
in event_sched_in/event_sched_out using ctx->nr_freq which is protected by
ctx->lock. I think I should rather do that.

Concerning the others:

_ sched: we are under the rq lock, like you noticed, we are fine.

_ posix timers: we are under sighand lock, so we are fine too.

_ sched_clock_stable: that one is more obscure. It seems that set_sched_clock_stable()
  and clear_sched_clock_stable() can race on static keys if running concurrently, and
  that would concern tick mask as well.

Thanks.

> 
> -- 
> Chris Metcalf, EZChip Semiconductor
> http://www.ezchip.com
>