Re: [PATCH v3 3/3] [RFC] CPUFreq: Add support for cpu-perf-dependencies

From: Viresh Kumar
Date: Thu Nov 19 2020 - 01:40:33 EST


On 18-11-20, 13:00, Rafael J. Wysocki wrote:
> On Wed, Nov 18, 2020 at 5:42 AM Viresh Kumar <viresh.kumar@xxxxxxxxxx> wrote:
> >
> > On 17-11-20, 14:06, Rafael J. Wysocki wrote:
> > > Is this really a cpufreq thing, though, or is it arch stuff? I think
> > > the latter, because it is not necessary for anything in cpufreq.
> > >
> > > Yes, acpi-cpufreq happens to know this information, because it uses
> > > processor_perflib, but the latter may as well be used by the arch
> > > enumeration of CPUs and the freqdomain_cpus mask may be populated from
> > > there.
> > >
> > > As far as cpufreq is concerned, if the interface to the hardware is
> > > per-CPU, there is one CPU per policy and cpufreq has no business
> > > knowing anything about the underlying hardware coordination.
> >
> > It won't be used by cpufreq for now at least and yes I understand your
> > concern. I opted for this because we already have a cpufreq
> > implementation for the same thing and it is usually better to reuse
> > this kind of stuff instead of inventing it over.
>
> Do you mean related_cpus and real_cpus?

Sorry about the confusion, I meant freqdomain_cpus only.

> That's the granularity of the interface to the hardware I'm talking about.
>
> Strictly speaking, it means "these CPUs share a HW interface for perf
> control" and it need not mean "these CPUs are in the same
> clock/voltage domain". Specifically, it need not mean "these CPUs are
> the only CPUs in the given clock/voltage domain". That's what it
> means when the control is exercised by manipulating OPPs directly, but
> not in general.
>
> In the ACPI case, for example, what the firmware tells you need not
> reflect the HW topology in principle. It only tells you whether or
> not it wants you to coordinate a given group of CPUs and in what way,
> but this may not be the whole picture from the HW perspective. If you
> need the latter, you need more information in general (at least you
> need to assume that what the firmware tells you actually does reflect
> the HW topology on the given SoC).
>
> So yes, in the particular case of OPP-based perf control, cpufreq
> happens to have the same information that is needed by the other
> subsystems, but otherwise it may not and what I'm saying is that it
> generally is a mistake to expect cpufreq to have that information or
> to be able to obtain it without the help of the arch/platform code.
> Hence, it would be a mistake to design an interface based on that
> expectation.
>
> Or looking at it from a different angle, today a cpufreq driver is
> only required to specify the granularity of the HW interface for perf
> control via related_cpus. It is not required to obtain extra
> information beyond that. If a new mask to be populated by it is
> added, the driver may need to do more work which is not necessary from
> the perf control perspective. That doesn't look particularly clean to
> me.
>
> Moreover, adding such a mask to cpufreq_policy would make the users of
> it depend on cpufreq sort of artificially, which need not be useful
> even.
>
> IMO, the information needed by all of the subsystems in question
> should be obtained and made available at the arch/platform level and
> everyone who needs it should be able to access it from there,
> including the cpufreq driver for the given platform if that's what it
> needs to do.
>
> BTW, cpuidle may need the information in question too, so why should
> it be provided via cpufreq rather than via cpuidle?

Right.

--
viresh