[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <2cfbc633-1e94-d741-2337-e1b0cf48b81b@nvidia.com>
Date: Fri, 8 Dec 2023 21:04:32 +0530
From: Sumit Gupta <sumitg@...dia.com>
To: Ionela Voinescu <ionela.voinescu@....com>
CC: Beata Michalska <beata.michalska@....com>,
<linux-kernel@...r.kernel.org>,
<linux-arm-kernel@...ts.infradead.org>, <linux-pm@...r.kernel.org>,
<sudeep.holla@....covm>, <will@...nel.org>,
<catalin.marinas@....com>, <viresh.kumar@...aro.org>,
<rafael@...nel.org>, <yang@...amperecomputing.com>,
<linux-tegra@...r.kernel.org>, Sumit Gupta <sumitg@...dia.com>
Subject: Re: [PATCH v2 2/2] cpufreq: Wire-up arch-flavored freq info into
cpufreq_verify_current_freq
>>>>>> --- a/drivers/cpufreq/cpufreq.c
>>>>>> +++ b/drivers/cpufreq/cpufreq.c
>>>>>> @@ -1756,7 +1756,8 @@ static unsigned int cpufreq_verify_current_freq(struct cpufreq_policy *policy, b
>>>>>> {
>>>>>> unsigned int new_freq;
>>>>>>
>>>>>> - new_freq = cpufreq_driver->get(policy->cpu);
>>>>>> + new_freq = arch_freq_get_on_cpu(policy->cpu);
>>>>>> + new_freq = new_freq ?: cpufreq_driver->get(policy->cpu);
>>>>>
>>>>> Given that arch_freq_get_on_cpu() is an average frequency, it does not
>>>>> seem right to me to trigger the sync & update process of
>>>>> cpufreq_verify_current_freq() based on it.
>>>>>
>>>>> cpufreq_verify_current_freq() will at least modify the internal state of
>>>>> the policy and send PRE and POST notifications, if not do a full frequency
>>>>> update, based on this average frequency, which is likely different from
>>>>> the current frequency, even beyond the 1MHz threshold.
>>>>>
>>>>> While I believe it's okay to return this average frequency in
>>>>> cpuinfo_cur_freq, I don't think it should be used as an indication of
>>>>> an accurate current frequency, which is what
>>>>> cpufreq_verify_current_freq() expects.
>>>>>
>>>>> Sumit, can you give more details on the issue at [1] and why this change
>>>>> fixes it?
>>>>>
>>>>> [1] https://lore.kernel.org/lkml/6a5710f6-bfbb-5dfd-11cd-0cd02220cee7@nvidia.com/
>>>>>
>>>>> Thank you,
>>>>> Ionela.
>>>>>
>>>> cpufreq_verify_current_freq() also updates 'policy->cur' in POST
>>>> notification if the frequency from hardware has more delta (out of sync).
>>>>
>>>> As the value from 'cpufreq_driver->get()' is not reliable due to [1],
>>>> calling the 'get' hook can update the 'policy->cur' with a wrong value when
>>>> governor starts in cpufreq_start_governor().
>>>> And if the frequency is never changed after the governor starts during
>>>> boot e.g. when performance governor is set as default, then
>>>> 'scaling_cur_freq' always returns wrong value.
>>>>
>>>> Instead, the arch_freq_get_on_cpu() API updates 'policy->cur' with a more
>>>> stable freq value.
>>>>
>>>> [1] https://lore.kernel.org/lkml/20230418113459.12860-7-sumitg@nvidia.com/
>>>
>>> Got it, many thanks!
>>>
>>> As the code is right now in v2, arch_freq_get_on_cpu() is called on
>>> show_scaling_cur_freq(), so the problem you describe would not show up.
>>> policy->cur would still be incorrect, but 'scaling_cur_freq' would
>>> return the value from arch_freq_get_on_cpu().
>>>
>>> Would it be enough if arch_freq_get_on_cpu() gets also called from
>>> show_cpuinfo_cur_freq() instead of cpufreq_verify_current_freq()?
>>>
>>> Thanks,
>>> Ionela.
>>>
>>
>> Yes.
>> I am not sure if making both the nodes 'scaling_cur_freq' and
>> 'cpuinfo_cur_freq' same is fine?
>
> That would happen anyway if arch_freq_get_on_cpu() is called from
> cpufreq_verify_current_freq().
>
Yes, that will happen in both the cases.
> In principle, according to [1], it would be correct to use it for
> 'cpuinfo_cur_freq' and not 'scaling_cur_freq'. But the call from
> show_scaling_cur_freq() is already there before these patches,
> introduced a long time ago for x86.
>
> The topic was discussed at [2] and the agreement so far was that it
> would be best to keep the behaviour the same for both x86 and arm.
>
Looking at the previous discussion in [2], seems to be fine.
Best Regards,
Sumit Gupta
> I don't like going against the user-guide, but these patches don't
> actually go against the user-guide. The old call to
> arch_freq_get_on_cpu() from show_scaling_cur_freq() goes against it.
> But I agree that's something necessary to keep, as legacy for x86.
> Additionally, you also mentioned that you'd prefer to have a more
> accurate frequency returned for 'scaling_cur_freq'.
>
> [1] https://www.kernel.org/doc/Documentation/cpu-freq/user-guide.txt
> [2] https://lore.kernel.org/lkml/20230609043922.eyyqutbwlofqaddz@vireshk-i7/
>
> Thanks,
> Ionela.
>
>>
>> Best Regards,
>> Sumit Gupta
Powered by blists - more mailing lists