Re: [PATCH v2] cpuidle: menu: Handle stopped tick more aggressively

[leo . yan]

On Sun, Aug 12, 2018 at 12:07:45PM +0200, Rafael J. Wysocki wrote:

[...]

> > > > > --- linux-pm.orig/drivers/cpuidle/governors/menu.c
> > > > > +++ linux-pm/drivers/cpuidle/governors/menu.c
> > > > > @@ -285,9 +285,8 @@ static int menu_select(struct cpuidle_dr
> > > > >  {
> > > > >       struct menu_device *data = this_cpu_ptr(&menu_devices);
> > > > >       int latency_req = cpuidle_governor_latency_req(dev->cpu);
> > > > > -     int i;
> > > > > -     int first_idx;
> > > > > -     int idx;
> > > > > +     int first_idx = 0;
> > > > > +     int idx, i;
> > > > >       unsigned int interactivity_req;
> > > > >       unsigned int expected_interval;
> > > > >       unsigned long nr_iowaiters, cpu_load;
> > > > > @@ -307,6 +306,18 @@ static int menu_select(struct cpuidle_dr
> > > > >       /* determine the expected residency time, round up */
> > > > >       data->next_timer_us = ktime_to_us(tick_nohz_get_sleep_length(&delta_next));
> > > > >
> > > > > +     /*
> > > > > +      * If the tick is already stopped, the cost of possible short idle
> > > > > +      * duration misprediction is much higher, because the CPU may be stuck
> > > > > +      * in a shallow idle state for a long time as a result of it.  In that
> > > > > +      * case say we might mispredict and use the known time till the closest
> > > > > +      * timer event for the idle state selection.
> > > > > +      */
> > > > > +     if (tick_nohz_tick_stopped()) {
> > > > > +             data->predicted_us = ktime_to_us(delta_next);
> > > > > +             goto select;
> > > > > +     }
> > > > > +
> > > >
> > > > This introduce two potential issues:
> > > >
> > > > - This will totally ignore the typical pattern in idle loop; I
> > > >   observed on the mmc driver can trigger multiple times (> 10 times)
> > > >   with consistent interval;
> > >
> > > I'm not sure what you mean by "ignore".
> >
> > You could see after move code from blow to this position, the typical
> > pattern interval will not be accounted; so if in the middle of idles
> > there have a bunch of interrupts with fix pattern, the upper code
> > cannot detect this pattern anymore.
> 
> I'm not really following you here.
> 
> The part of the code skipped for tick_nohz_tick_stopped() doesn't
> update the data at all AFAICS.  It only computes some values that
> would be discarded later anyway, so I'm not sure what the point of
> running that computation is.

Sorry I don't explain clearly, so try to rephrase:

With your patch for the tick stopped case, it directly uses tick delta
value as prediction and goto 'select' tag.  So it skips below code
pieces, these codes have minor improvement for typical pattern which
can be applied in the middle of idles, for example, the mmc driver
triggers 16 interrupts with ~1500us interval, these interrupts are all
handled within the idle loop, so the typical pattern can detect the mmc
interrupts pattern and it will help idle governor to select a shallower
idle state so can avoid to break the residency.

You mentioned these computed values would be discarded later, this is
true for most cases, but it isn't always true actually.  Without your
patch, the governor will discard the computed values only when
'data->predicted_us < TICK_USEC', otherwise the interval pattern is
still be applied in the prediction.

	expected_interval = get_typical_interval(data);
	expected_interval = min(expected_interval, data->next_timer_us);

	[...]

	/*
	 * Use the lowest expected idle interval to pick the idle state.
	 */
	data->predicted_us = min(data->predicted_us, expected_interval);

> The statistics are updated by menu_update() and that still runs and it
> will take the actual wakeup events into account, won't it?

Yes.

> > [...]
> >
> > > > > -     if ((drv->states[idx].flags & CPUIDLE_FLAG_POLLING) ||
> > > > > -         expected_interval < TICK_USEC) {
> > > > > +     if (((drv->states[idx].flags & CPUIDLE_FLAG_POLLING) ||
> > > > > +         expected_interval < TICK_USEC) && !tick_nohz_tick_stopped()) {
> > > >
> > > > I am not sure this logic is right... Why not use below checking, so
> > > > for POLLING state we will never ask to stop the tick?
> > > >
> > > >         if (drv->states[idx].flags & CPUIDLE_FLAG_POLLING ||
> > > >             (expected_interval < TICK_USEC && !tick_nohz_tick_stopped())) {
> > > >
> > >
> > > The only effect of it would be setting stop_tick to false, but why
> > > would that matter?
> >
> > Please consider below situation, not sure if this case is existed or
> > not:
> >
> >   step1: first time: enter one idle state with stopping tick;
> >   step2: second time: select POLLING state and tick_nohz_tick_stopped()
> >   is true;
> >
> > So in step2, it cannot set stop_tick to false with below sentence.
> >
> > > > >               unsigned int delta_next_us = ktime_to_us(delta_next);
> > > > >
> > > > >               *stop_tick = false;
> 
> But setting *stop_tick has no effect as far as the current code is
> concerned (up to the bug fixed by the other patch).

Yeah.

> Also the POLLING state can only be selected if there are no other
> states matching delta_next available in that case which means that
> there will be a timer to break the polling loop soon enough (and BTW
> the polling has a built-in timeout too), so I don't really see a
> problem here.

Ah, now I understand the logic and I misunderstand the POLLING mode
before; now agree with this.  Sorry for noise.

Thanks,
Leo Yan