On 6/16/21 6:47 AM, Dmitry Osipenko wrote:
> 16.06.2021 05:50, Thara Gopinath пишет:
> ...
>>
>> Hi,
>>
>> Thermal pressure is letting scheduler know that the max capacity
>> available for a cpu to schedule tasks is reduced due to a thermal event.
>> So you cannot have a h/w thermal pressure and s/w thermal pressure.
>> There is eventually only one capping applied at h/w level and the
>> frequency corresponding to this capping should be used for thermal
>> pressure.
>>
>> Ideally you should not be having both s/w and h/w trying to throttle at
>> the same time. Why is this a scenario and what prevents you from
>> disabling s/w throttling when h/w throttling is enabled. Now if there
>> has to a aggregation for whatever reason this should be done at the
>> thermal driver level and passed to scheduler.
>
> Hello,
>
> The h/w mitigation is much more reactive than software, in the same time
> it's much less flexible than software. It should provide additional
> protection in a cases where software isn't doing a good job. Ideally h/w
> mitigation should stay inactive all the time, nevertheless it should be
> modeled properly by the driver.

Ok. This is kind of opposite to what I am doing on the Qcom platform I
am working on. The h/w throttling is the default since like you
mentioned it is more reactive. And s/w does only critical trip management.

>
>>>>
>>>> That is a good question. IMO, first step would be to call
>>>> cpufreq_update_limits().
>>>
>>> Right
>>>
>>>> [ Cc Thara who implemented the thermal pressure ]
>>>>
>>>> May be Thara has an idea about how to aggregate both? There is another
>>>> series floating around with hardware limiter [1] and the same
>>>> problematic.
>>>>
>>>> [1] https://lkml.org/lkml/2021/6/8/1791
>>>
>>> Thanks, it indeed looks similar.
>>>
>>> I guess the common thermal pressure update code could be moved out into
>>> a new special cpufreq thermal QoS handler (policy->thermal_constraints),
>>> where handler will select the frequency constraint and set up the
>>> pressure accordingly. So there won't be any races in the code.
>>>
>> It was a conscious decision to keep thermal pressure update out of qos
>> max freq update because there are platforms that don't use the qos
>> framework. For eg acpi uses cpufreq_update_policy.
>> But you are right. We have two platforms now applying h/w throttling and
>> cpufreq_cooling applying s/w throttling. So it does make sense to have
>> one api doing all the computation to update thermal pressure. I am not
>> sure how exactly/where exactly this will reside.
>
> The generic cpufreq_cooling already uses QoS for limiting the CPU
> frequency. It could be okay to use QoS for the OF drivers, this needs a
> closer look.
>
> We have the case where CPU frequency is changed by the thermal event and
> the thermal pressure equation is the same for both s/w cpufreq_cooling
> and h/w thermal driver. The pressure is calculated based on the QoS
> cpufreq constraint that is already aggregated.
>
> Hence what we may need to do on the thermal event is:
>
> 1. Update the QoS request
> 2. Update the thermal pressure
> 3. Ensure that updates are not racing

Yes. So the first two steps you mentioned is exactly what
cpufreq_cooling.c also does except for the fact that it is a s/w
mitigation. Now if you have two sources that is updating the max
frequency via qos, I think you can do either of the following before
calculating thermal pressure
1. Read the throttled frequency from h/w if your h/w supports this feature.
or
2. Use freq_qos_read_value to get the max frequency value.

Either way only the correct throttled capacity should be passed to
scheduler.

--
Warm Regards
Thara (She/Her/Hers)
>
>> So for starters, I think you should replicate the update of thermal
>> pressure in your h/w driver when you know that h/w is
>> throttling/throttled the frequency. You can refer to cpufreq_cooling.c
>> to see how it is done.
>>
>> Moving to a common api can be done as a separate patch series.
>>
>
> Thank you for the clarification and suggestion.
>