Re: çå: [PATCH V4] sched/cpufreq: initialize iowait_boost_max and iowait_boost with cpu capacity

[Vincent Guittot]

On Mon, 4 Mar 2019 at 17:48, Quentin Perret <quentin.perret@xxxxxxx> wrote:
>
> On Monday 04 Mar 2019 at 16:26:16 (+0100), Peter Zijlstra wrote:
> > On Mon, Mar 04, 2019 at 01:58:12PM +0000, Quentin Perret wrote:
> > > You could also update the values in sugov_get_util() at the cost of a
> > > small overhead to compute 'min'. I'm not sure what's preferable since
> > > we wanted to avoid that kind of overhead in the first place ...
> >
> > Or,... we could actually make things simpler.
> >
> > How's the below? I have a feq questions wrt min, mostly:
> >
> >  - what's the difference between policy->min and
> >    policy->cpuinfo.min_freq; it used to be the former, the below uses
> >    the latter.
>
> As mentioned on IRC, IIRC policy->min is something that can be written
> from userspace (for example) to cap the min freq. OTOH, cpuinfo.min_freq
> is read-only and just reports the lowest OPP.
>
> Rafael is this correct ?
>
> >  - should we have a min_freq based value, instead of a constant; the
> >    difference being that with this the actual boost speed depends in the
> >    gap between min/max.
>
> If the above is correct, then I agree. Looking at min_freq simplifies
> things quite a bit since it doesn't need to be updated all the time,
> and the whole policy->min stuff is dealt with at the CPUFreq core level
> so it's not obvious sugov should care.
>
> > Anyway; the below converts iowait_boost to capacity scale (from kHz), it
> > side-steps the whole issue you guys are bickering about by limiting it
> > to SCHED_CAPACITY_SCALE (aka. 1024) on the boost path, and then limiting
> > it to @max by the time we figured out we ought to use it.
> >
> > And since that means we never change @max anymore; we can simplify whole
> > return value thing.
>
> [...]
>
> > @@ -837,7 +818,9 @@ static int sugov_start(struct cpufreq_policy *policy)
> >               memset(sg_cpu, 0, sizeof(*sg_cpu));
> >               sg_cpu->cpu                     = cpu;
> >               sg_cpu->sg_policy               = sg_policy;
> > -             sg_cpu->iowait_boost_max        = policy->cpuinfo.max_freq;
> > +             sg_cpu->min                     =
> > +                     (SCHED_CAPACITY_SCALE * policy->cpuinfo.min_freq) /
> > +                     policy->cpuinfo.max_freq;
>
> The 'funny' thing is that on big little this 'min' can end up being
> larger than 'max' ...

yes arch_scale_cpu_capacity() should be used instead of
SCHED_CAPACITY_SCALE to be compared with sg_cpu->max and sg_cpu->util

>
> On juno r0 for example, min_freq and max_freq for little CPUs are
> respectively 450MHz and 850MHz. So you get sg_cpu->min=542, but
> sg_cpu->max=446 ... So you'll max out after the first IO wakeup.
> And since iowait_boost is reset whenever it is smaller than sg_cpu->min,
> you end up with something that can either force max freq or apply no
> boost at all ...
>
> Perhaps you could keep the 'util' and 'max' pointers in
> sugov_iowait_apply() and overwrite them like before, but in the
> SCHED_CAPACITY_SCALE scale as you suggest ?
>
> Thanks,
> Quentin