Re: [PATCH V7 5/7] cpufreq: Register notifiers with the PM QoS framework

From: Viresh Kumar <viresh.kumar@linaro.org>
To: "Rafael J. Wysocki" <rafael@kernel.org>
Cc: Dmitry Osipenko <digetx@gmail.com>,
	Rafael Wysocki <rjw@rjwysocki.net>,
	Linux PM <linux-pm@vger.kernel.org>,
	Vincent Guittot <vincent.guittot@linaro.org>,
	Matthias Kaehlcke <mka@chromium.org>,
	Ulf Hansson <ulf.hansson@linaro.org>,
	Stephen Rothwell <sfr@canb.auug.org.au>,
	Pavel Machek <pavel@ucw.cz>,
	"Rafael J . Wysocki" <rafael.j.wysocki@intel.com>,
	Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
	linux-tegra <linux-tegra@vger.kernel.org>
Subject: Re: [PATCH V7 5/7] cpufreq: Register notifiers with the PM QoS framework
Date: Wed, 16 Oct 2019 13:57:42 +0530	[thread overview]
Message-ID: <20191016082742.nttzuofes6uds4pu@vireshk-i7> (raw)
In-Reply-To: <CAJZ5v0hxsy3ZKFvtWULHAVog4=3rYQfd3-61A9dNaKeUbiDtrg@mail.gmail.com>

On 15-10-19, 23:50, Rafael J. Wysocki wrote:
> On Tue, Oct 15, 2019 at 5:53 PM Rafael J. Wysocki <rafael@kernel.org> wrote:

> > > - Update QoS framework with the knowledge of related CPUs, this has been pending
> > >   until now from my side. And this is the thing we really need to do. Eventually
> > >   we shall have only a single notifier list for all CPUs of a policy, at least
> > >   for MIN/MAX frequencies.
> >
> > - Move the PM QoS requests and notifiers to the new policy CPU on all
> > changes of that.  That is, when cpufreq_offline() nominates the new
> > "leader", all of the QoS stuff for the policy needs to go to this one.
> 
> Alas, that still will not work, because things like
> acpi_processor_ppc_init() only work accidentally for one-CPU policies.

I am not sure what problem you see here ? Can you please explain a bit more.

> Generally, adding such a PM QoS request to a non-policy CPU simply has
> no effect until it becomes a policy CPU which may be never.

I was thinking maybe we can read the constraints for all CPUs in the
policy->cpus mask in cpufreq_set_policy() and so this part of the problem will
just go away. The only part that would be left is to remove the QoS constraints
properly.

> It looks like using device PM QoS for cpufreq is a mistake in general
> and what is needed is a struct pm_qos_constraints member in struct
> cpufreq_policy and something like
> 
> struct freq_pm_qos_request {
>         enum freq_pm_qos_req_type type; /* min or max */
>         struct plist_node pnode;
>         struct cpufreq_policy *policy;
> };
> 
> Then, pm_qos_update_target() can be used for adding, updating and
> removing requests.

-- 
viresh