Re: kernel/timer: avoid spurious ksoftirqd wakeups

[Marcelo Tosatti <mtosatti@redhat.com>]

On Thu, Apr 02, 2015 at 04:59:40PM +0200, Frederic Weisbecker wrote:
> On Wed, Apr 01, 2015 at 10:44:55PM -0300, Marcelo Tosatti wrote:
> > 
> > It is only necessary to raise timer softirq
> > in case there are active timers or irq work 
> > to do.
> > 
> > Limit the ksoftirqd wakeup to those cases.
> > 
> > Fixes a latency spike with isolated CPUs and 
> > nohz full mode.
> > 
> > Reported-and-tested-by: Luiz Capitulino <lcapitulino@redhat.com>
> > Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
> > 
> > diff --git a/include/linux/timer.h b/include/linux/timer.h
> > index 8c5a197..0c065f9 100644
> > --- a/include/linux/timer.h
> > +++ b/include/linux/timer.h
> > @@ -192,7 +192,7 @@ extern void set_timer_slack(struct timer_list *time, int slack_hz);
> >   * locks the timer base and does the comparison against the given
> >   * jiffie.
> >   */
> > -extern unsigned long get_next_timer_interrupt(unsigned long now);
> > +extern unsigned long get_next_timer_interrupt(unsigned long now, bool *raise_softirq);
> >  
> >  /*
> >   * Timer-statistics info:
> > diff --git a/kernel/time/tick-sched.c b/kernel/time/tick-sched.c
> > index a4c4eda..615e276 100644
> > --- a/kernel/time/tick-sched.c
> > +++ b/kernel/time/tick-sched.c
> > @@ -568,6 +568,7 @@ static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts,
> >  	unsigned long rcu_delta_jiffies;
> >  	struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
> >  	u64 time_delta;
> > +	bool raise_softirq;
> >  
> >  	time_delta = timekeeping_max_deferment();
> >  
> > @@ -582,9 +583,11 @@ static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts,
> >  	    arch_needs_cpu() || irq_work_needs_cpu()) {
> >  		next_jiffies = last_jiffies + 1;
> >  		delta_jiffies = 1;
> > +		raise_softirq = true;
> 
> I believe that irq_work doesn't need the softirq. 

Can drop that, right.

> It needs a tick only in order to call
> irq_work_tick(). And I believe this is the same for RCU which needs a call to
> rcu_check_callbacks(), but it might need something else that the softirq does
> (but this is the timer softirq, not the rcu one). 
> 
> >  	} else {
> >  		/* Get the next timer wheel timer */
> > -		next_jiffies = get_next_timer_interrupt(last_jiffies);
> > +		next_jiffies = get_next_timer_interrupt(last_jiffies,
> > +							&raise_softirq);
> >  		delta_jiffies = next_jiffies - last_jiffies;
> >  		if (rcu_delta_jiffies < delta_jiffies) {
> >  			next_jiffies = last_jiffies + rcu_delta_jiffies;
> > @@ -703,7 +706,8 @@ static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts,
> >  		 */
> >  		tick_do_update_jiffies64(ktime_get());
> >  	}
> > -	raise_softirq_irqoff(TIMER_SOFTIRQ);
> > +	if (raise_softirq)
> > +		raise_softirq_irqoff(TIMER_SOFTIRQ);
> >  out:
> >  	ts->next_jiffies = next_jiffies;
> >  	ts->last_jiffies = last_jiffies;
> 
> Lets look at the things outside the pending timer list that can end up failing
> to program the timer because it is in the past already:

Is this an attempt to find possible regressions introduced 
by this change ?

> _ timekeeping_max_deferment(): I doubt, the value is pretty high
> _ scheduler_tick_max_deferment(); it's one second long, way enough to never be in
>   the past by the time we program the clock
> _ RCU, irq_work, arch: may be, if the last jiffies update is far enough. But apparently
>   the problem is elsewhere since you keep the softirq for these and your patch solves your
>   problem.
> _ In case hrtimer runs in low-res mode and the next hrtimer is very close, or even in the past
>   already, you may run into such issue. And hrtimer doesn't need the timer softirq, at least not
>   to run the callbacks. It needs it only if hrtimer is switching to high-res mode, I think it's
>   a rare event.
> 
> Now it would be nice to identify the issue we are facing here. Are you running in hrtimer low-res
> mode?

The issue is a latency spike due to ksoftirqd waking up to 
process pending timers, processing two deferred timers, 
but no non-deferred timers.

hrtimer is not in low-res mode.

The issue is ksoftirqd waking up in the first place.