Re: kernel/timer: avoid spurious ksoftirqd wakeups (v2)

[Luiz Capitulino <lcapitulino@redhat.com>]

On Fri, 10 Apr 2015 15:09:07 -0300
Marcelo Tosatti <mtosatti@redhat.com> wrote:

> On Wed, Apr 08, 2015 at 12:12:45AM +0200, Frederic Weisbecker wrote:
> > On Tue, Apr 07, 2015 at 11:10:49PM +0200, Thomas Gleixner wrote:
> > > On Mon, 6 Apr 2015, Marcelo Tosatti wrote:
> > > > It is only necessary to raise timer softirq
> > > > in case there are active timers.
> > > 
> > > Depends. See below.
> > >  
> > > > Limit the ksoftirqd wakeup to that case.
> > > > 
> > > > Fixes a latency spike with isolated CPUs and
> > > > nohz full mode.
> > > 
> > > This lacks a proper explanation of the observed issue.
> > > 
> > > > --- 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 = false;
> > > 
> > > This shadows the function name raise_softirq(). Not pretty.
> > >   
> > > >  	time_delta = timekeeping_max_deferment();
> > > >  
> > > > @@ -584,7 +585,8 @@ static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts,
> > > >  		delta_jiffies = 1;
> > > >  	} 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 +705,8 @@ 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);
> > > 
> > > This breaks when high resolution timers are disabled (compile or
> > > runtime) because then the hrtimer queues are run from the timer
> > > softirq.
> > > 
> > > Now assume the following situation:
> > > 
> > >   Tick is stopped completely with no timers and no hrtimers pending.
> > > 
> > >   Interrupt happens and schedules a hrtimer.
> > > 
> > >   nohz_stop_sched_tick()
> > >     get_next_timer_interrupt(..., &raise_softirq);
> > > 
> > >       ---> base->active_timers = 0, so raise_softirq is false
> > > 
> > >     tick_program_event(expires)
> > >       clockevents_program_event(expires)
> > >       
> > >       ---> Assume expires is already in the past
> > > 
> > >         if (expires <= ktime_get())
> > > 	   return -ETIME;
> > > 
> > >     if (raise_softirq)
> > >        raise_softirq_irqoff(TIMER_SOFTIRQ);
> > > 
> > > So because the tick device was not armed you wont get a tick
> > > interrupt up to the point where tick_nohz_stop_sched_tick() is called
> > > again which might be far off.
> > > 
> > > I can see that the unconditional raise_softirq_irqoff() is suboptimal,
> > > but it was a rather simple solution to get stuff rolling again because
> > > it forces the cpu out of the inner idle loop which in turn restarts
> > > the tick.
> > 
> > Doh, that's the kind of side effect I was worried about, thanks for the
> > explanation. The necessary exit out of the idle loop implied by this
> > softirq when the timer fails to be programmed really deserves a comment.
> > 
> > And note how it relies on the magic !in_interrupt() in this piece of
> > hardirq code, otherwise that would be softirq from hardirq without
> > reschedule() and thus no exit from idle loop, and thus no tick
> > reprogramming.
> > 
> > Let's see if I can come up with some solution to clean this up, if
> > Marcelo doesn't beat me at it.
> 
> The problem is the following from -RT:
> 
> #ifdef CONFIG_PREEMPT_RT_BASE
>        if (!hrtimer_rt_defer(timer))
>                return -ETIME;
> #endif

Just to clarify, Marcelo and I have found that this is the code
that fails in clockevents_program_event() returning -ETIME:

    delta = ktime_to_ns(ktime_sub(expires, ktime_get()));
    if (delta <= 0)
        return force ? clockevents_program_min_delta(dev) : -ETIME;

It fails in this call trace:

tick_nohz_stop_sched_tick()
  hrtimer_start()
    __hrtimer_start_range_ns()
      hrtimer_enqueue_reprogram()
        hrtimer_reprogram() /* timer expires in 38259845000000 */
          tick_program_event(38259845000000, 0) /* returns -ETIME */
            tick_program_event()
              clockevents_program_event() /* returns -ETIME */

> It seems a valid solution for this interrupt is to program 
> sched_timer to the nearest future possible. 

What about calling the timer function right there, like
hrtimer_interrupt() does?

	if (!hrtimer_rt_defer(timer))
		__run_hrtimer(timer, &basenow);

> 
> 	if (expires < now)
> 		expires = now + safe_margin;
> 
> 	program_timer(expires);
> 
> (perhaps a for loop increasing safe_margin if program_timer fails...)
> 
> Is that what you mean by clean up, Frederic?
>