Re: [PATCH v3 3/4] cpufreq: schedutil: Simplify iowait boosting

["Rafael J. Wysocki" <rafael@kernel.org>]

On Thu, Mar 28, 2019 at 10:48 AM Quentin Perret <quentin.perret@arm.com> wrote:
>
> Hi Rafael,
>
> On Tuesday 26 Mar 2019 at 12:18:00 (+0100), Rafael J. Wysocki wrote:
> > @@ -13,6 +13,8 @@
> >  #include <linux/sched/cpufreq.h>
> >  #include <trace/events/power.h>
> >
> > +#define IOWAIT_BOOST_MIN     (SCHED_CAPACITY_SCALE / 8)
> > +
> >  struct sugov_tunables {
> >       struct gov_attr_set     attr_set;
> >       unsigned int            rate_limit_us;
> > @@ -51,7 +53,6 @@ struct sugov_cpu {
> >       u64                     last_update;
> >
> >       unsigned long           bw_dl;
> > -     unsigned long           min;
> >       unsigned long           max;
> >
> >       /* The field below is for single-CPU policies only: */
> > @@ -303,7 +304,7 @@ static bool sugov_iowait_reset(struct sua
>
> The comment above this function needs updating I think.
>
> >       if (delta_ns <= TICK_NSEC)
> >               return false;
> >
> > -     sg_cpu->iowait_boost = set_iowait_boost ? sg_cpu->min : 0;
> > +     sg_cpu->iowait_boost = set_iowait_boost ? IOWAIT_BOOST_MIN : 0;
> >       sg_cpu->iowait_boost_pending = set_iowait_boost;
> >
> >       return true;
> > @@ -349,7 +350,7 @@ static void sugov_iowait_boost(struct su
>
> Ditto.

I overlooked these two, thanks for pointing that out!

Will send an update momentarily.

> >       }
> >
> >       /* First wakeup after IO: start with minimum boost */
> > -     sg_cpu->iowait_boost = sg_cpu->min;
> > +     sg_cpu->iowait_boost = IOWAIT_BOOST_MIN;
> >  }
> >
> >  /**
> > @@ -389,7 +390,7 @@ static unsigned long sugov_iowait_apply(
> >                * No boost pending; reduce the boost value.
> >                */
> >               sg_cpu->iowait_boost >>= 1;
> > -             if (sg_cpu->iowait_boost < sg_cpu->min) {
> > +             if (sg_cpu->iowait_boost < IOWAIT_BOOST_MIN) {
> >                       sg_cpu->iowait_boost = 0;
> >                       return util;
> >               }
> > @@ -826,9 +827,6 @@ static int sugov_start(struct cpufreq_po
> >               memset(sg_cpu, 0, sizeof(*sg_cpu));
> >               sg_cpu->cpu                     = cpu;
> >               sg_cpu->sg_policy               = sg_policy;
> > -             sg_cpu->min                     =
> > -                     (SCHED_CAPACITY_SCALE * policy->cpuinfo.min_freq) /
> > -                     policy->cpuinfo.max_freq;
> >       }
> >
> >       for_each_cpu(cpu, policy->cpus) {
> >
>
> Other than that, I tried a backport of this on a Pixel 3 with Snapdragon
> 845 (which is relevant because it has tons of OPPs, so starting at 128
> makes it ramp up faster) to check the impact on power, but the only
> differences appeared to be in the noise margin, so it's all good :)

Cool. :-)

> Full test results available at [1]. Note that I did enable the iowait
> boost feature for these tests -- it is disabled by default on P3 ...

Thanks!

> ---
> [1] https://nbviewer.jupyter.org/gist/qperret/69c9bde13aad2d783689e78c9ba2d9bc