Re: [PATCH v1] Revert "cpufreq: schedutil: Move max CPU capacity to sugov_policy"

From: Vincent Guittot <vincent.guittot@linaro.org>
To: Lukasz Luba <lukasz.luba@arm.com>
Cc: "Rafael J. Wysocki" <rafael@kernel.org>,
	Sam Wu <wusamuel@google.com>,
	Saravana Kannan <saravanak@google.com>,
	Viresh Kumar <viresh.kumar@linaro.org>,
	Ingo Molnar <mingo@redhat.com>,
	Peter Zijlstra <peterz@infradead.org>,
	Juri Lelli <juri.lelli@redhat.com>,
	Dietmar Eggemann <dietmar.eggemann@arm.com>,
	Steven Rostedt <rostedt@goodmis.org>,
	Ben Segall <bsegall@google.com>, Mel Gorman <mgorman@suse.de>,
	Daniel Bristot de Oliveira <bristot@redhat.com>,
	Valentin Schneider <vschneid@redhat.com>,
	"Isaac J . Manjarres" <isaacmanjarres@google.com>,
	kernel-team@android.com,
	"Rafael J. Wysocki" <rafael.j.wysocki@intel.com>,
	linux-pm@vger.kernel.org, linux-kernel@vger.kernel.org
Subject: Re: [PATCH v1] Revert "cpufreq: schedutil: Move max CPU capacity to sugov_policy"
Date: Wed, 30 Nov 2022 11:42:20 +0100	[thread overview]
Message-ID: <CAKfTPtAPniqQyDzh=Yu8Z9R9+H2PzBKkHT0SJgHZiUOdNdw3Mg@mail.gmail.com> (raw)
In-Reply-To: <97af1300-541d-a79c-404c-92886f10b220@arm.com>

Hi All

Just for the log and because it took me a while to figure out the root
cause of the problem: This patch also creates a regression for
snapdragon845 based systems and probably on any QC chipsets that use a
LUT to update the OPP table at boot. The behavior is the same as
described by Sam with a staled value in sugov_policy.max field.

Regards,
Vincent

On Tue, 22 Nov 2022 at 09:58, Lukasz Luba <lukasz.luba@arm.com> wrote:
>
> Hi Rafael and Sam
>
> On 11/21/22 19:18, Rafael J. Wysocki wrote:
> > On Fri, Nov 18, 2022 at 2:00 AM Sam Wu <wusamuel@google.com> wrote:
> >>
> >> On Wed, Nov 16, 2022 at 3:35 AM Lukasz Luba <lukasz.luba@arm.com> wrote:
> >>> Which mainline kernel version you use in pixel6?
> >> I am using kernel version 6.1-rc5.
> >>>
> >>> Could you elaborate a bit how is it possible?
> >>> Do you have the sg_policy setup properly (and at right time)?
> >>> Do you have the cpu capacity from arch_scale_cpu_capacity()
> >>> set correctly and at the right time during this cpufreq
> >>> governor setup?
> >>>
> >>> IIRC in Android there is a different code for setting up the
> >>> cpufreq sched governor clones. In mainline we don't have to do
> >>> those tricks, so this might be the main difference.
> >> This behavior is seen on the mainline kernel. There isn't any vendor code
> >> modifying the behavior, and the schedutil governor is being used.
> >>>
> >>> Could you trace the value that is read from
> >>> arch_scale_cpu_capacity() and share it with us?
> >>> I suspect this value changes in time in your kernel.
> >> There's an additional CPU capacity normalization step during
> >> init_cpu_capacity_callback() that does not happen until all the CPUs come
> >> online. However, the sugov_start() function can be called for a subset of
> >> CPUs before all the CPUs are brought up and before the normalization of
> >> the CPU capacity values, so there could be a stale value stored
> >> in sugov_policy.max field.
> >
> > OK, the revert has been applied as 6.1-rc material, thanks!
>
> I was on a business trip last week so couldn't check this.
> Now I'm back and I've checked the booting sequence.
> Yes, there is some race condition and the mechanism
> using blocking_notifier_call_chain() in the cpufreq_online()
> doesn't help while we are registering that schedutil
> new policy.
>
> I will have to go through those mechanisms and check them.
> I agree, for now the best option is to revert the patch.
>
> My apologies for introducing this issues.
> Thanks Sam for capturing it.
>
> Regards,
> Lukasz