[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <0f486d81-ec0f-22f0-a7cd-942da3288d6e@arm.com>
Date: Fri, 22 Sep 2023 00:04:29 +0200
From: Dietmar Eggemann <dietmar.eggemann@....com>
To: Shrikanth Hegde <sshegde@...ux.vnet.ibm.com>,
Pierre Gondois <pierre.gondois@....com>
Cc: vschneid@...hat.com, "Rafael J. Wysocki" <rafael@...nel.org>,
Viresh Kumar <viresh.kumar@...aro.org>,
linux-pm@...r.kernel.org, linux-kernel@...r.kernel.org
Subject: Re: [PATCH] cpufreq: Rebuild sched-domains when removing cpufreq
driver
On 20/09/2023 07:29, Shrikanth Hegde wrote:
>
> On 9/19/23 1:19 PM, Pierre Gondois wrote:
>>
>> On 9/19/23 01:03, Dietmar Eggemann wrote:
>>> On 18/09/2023 13:29, Pierre Gondois wrote:
[...]
>>> Rebuilding SDs when inserting the driver is already covered by
>>>
>>> cpufreq_online()
>>> cpufreq_set_policy()
>>> sched_cpufreq_governor_change()
>>> if (old or new gov eq. schedutil)
>>> schedule_work(&rebuild_sd_work)
>>>
>>> So what's missing is only a sched_cpufreq_governor_change() call when
>>> removing the driver, right?
>>
>> Yes exact, removing a cpufreq driver (e.g. `rmmod cppc_cpufreq.ko`) goes
>> through:
>> cpufreq_remove_dev()
>> \-__cpufreq_offline()
>>
>> so the path you mentioned is not used in this case.
But IMHO the name of sched_cpufreq_governor_change() is now misleading
when called from the driver removal path. It's not the governor which
has changed. We want to disable EAS when unloading the driver in (3):
...
if (!arch_scale_freq_invariant()) {
if (sched_debug()) {
pr_warn("rd %*pbl: Disabling EAS: frequency-invariant load tracking
not yet supported",
cpumask_pr_args(cpu_map));
}
goto free;
}
...
> One Doubt, while looking through code. Not well versed with this area.
>
> cpuhp_cpufreq_offline is being registered with CPU hotplug. That ends up
> calling cpufreq_offline. This may cause non desired issues.
> 1. rebuild of sched domains twice instead, once by CPU hotplug and once by this.
That's possible when you CPU hp out the last CPU of a policy (or Perf
Domain (PD).
Otherwise `cpuhp_cpufreq_offline -> cpufreq_offline() ->
__cpufreq_offline()` returns early in `if (!policy_is_inactive(policy))`
condition.
partition_sched_domains_locked() does:
(1) detach_destroy_domains()
(2) build_sched_domains()
(3) build_perf_domains()
But only the first workqueue event (from cpuset_hotplug_workfn or
rebuild_sd_workfn) will rebuild Sched Domains and PDs (1)-(3), the
second one will only build PDs (3) again.
That's not nice but AFAICS it is not functional incorrect.
> 2. offline/online of CPU (non-SMT) may not disabling EAS.
Can't see this issue right now. When we offline the last CPU of a PD on
a 2 PD system, EAS should be stopped since the root domain does not have
any asymmetric CPU capacities left and when we online it again, EAS
should be started (sched_energy_set()).
Powered by blists - more mailing lists