Re: [PATCH] ACPI: processor: Add QoS requests for all CPUs

[Rafael J. Wysocki]

On Fri, Oct 25, 2019 at 10:17 AM Rafael J. Wysocki <rafael@xxxxxxxxxx> wrote:
>
> On Fri, Oct 25, 2019 at 4:53 AM Viresh Kumar <viresh.kumar@xxxxxxxxxx> wrote:
> >
> > On 25-10-19, 02:41, Rafael J. Wysocki wrote:
> > > From: Rafael J. Wysocki <rafael.j.wysocki@xxxxxxxxx>
> > >
> > > The _PPC change notifications from the platform firmware are per-CPU,
> > > so acpi_processor_ppc_init() needs to add a frequency QoS request
> > > for each CPU covered by a cpufreq policy to take all of them into
> > > account.
> > >
> > > Even though ACPI thermal control of CPUs sets frequency limits
> > > per processor package, it also needs a frequency QoS request for each
> > > CPU in a cpufreq policy in case some of them are taken offline and
> > > the frequency limit needs to be set through the remaining online
> > > ones (this is slightly excessive, because all CPUs covered by one
> > > cpufreq policy will set the same frequency limit through their QoS
> > > requests, but it is not incorrect).
> > >
> > > Modify the code in accordance with the above observations.
> >
> > I am not sure if I understood everything you just said, but I don't
> > see how things can break with the current code we have.
> >
> > Both acpi_thermal_cpufreq_init() and acpi_processor_ppc_init() are
> > called from acpi_processor_notifier() which is registered as a policy
> > notifier and is called when a policy is created or removed. Even if
> > some CPUs of a policy go offline, it won't matter as the request for
> > the policy stays and it will be dropped only when all the CPUs of a
> > policy go offline.
> >
> > What am I missing ?
>
> The way the request is used.
>
> Say there are two CPUs, A and B, in the same policy.  A is
> policy->cpu, so acpi_processor_ppc_init() adds a QoS request for A
> only (note that the B's QoS request, B->perflib_req, remains
> inactive).
>
> Now, some time later, the platform firmware notifies the OS of a _PPC
> change for B.  That means acpi_processor_notify() is called and it
> calls acpi_processor_ppc_has_changed(B) and that invokes
> acpi_processor_get_platform_limit(B), which in turn looks at the B's
> QoS request (B->perflib_req) and sees that it is inactive, so 0 is
> returned without doing anything.  However, *some* QoS request should
> be updated then.
>
> Would it be correct to update the A's QoS request in that case?  No,
> because the _PPC limit for A may be different that the _PPC limit for
> B in principle.
>
> The thermal case is not completely analogous, because
> cpufreq_set_cur_state() finds online CPUs in the same package as the
> target one and tries to update the QoS request for each of them, which
> will include the original policy->cpu, whose QoS request has been
> registered by acpi_thermal_cpufreq_init(), as long as it is online.
> If it is offline, it will be skipped and there is no easy way to find
> a "previous policy->cpu".  It is possible to do that, but IMO it is
> more straightforward to have a request for each CPU added.

BTW, IMO processor_thremal can be changed to use one frequency QoS
request per policy on top of this, but I'd rather take one step at a
time. :-)