Re: [PATCH v4 2/3] sched/fair: Take thermal pressure into account while estimating energy

From: Lukasz Luba
Date: Wed Jun 16 2021 - 14:31:28 EST

On 6/16/21 6:24 PM, Dietmar Eggemann wrote:
On 15/06/2021 18:09, Lukasz Luba wrote:

On 6/15/21 4:31 PM, Dietmar Eggemann wrote:
On 14/06/2021 21:11, Lukasz Luba wrote:


It's important to highlight that this will only fix this issue between
schedutil and EAS when it's due to `thermal pressure` (today only via
CPU cooling). There are other places which could restrict policy->max
via freq_qos_update_request() and EAS will be unaware of it.

True, but for this I have some other plans.

As long as people are aware of the fact that this was developed to be
beneficial for `EAS - IPA` integration, I'm fine with this.

Good. I had in mind that I will have to do some re-work on this
thermal pressure code in the cpufreq cooling, to satisfy our roadmap


IMHO, this means that this is catered for the IPA governor then. I'm not
sure if this would be beneficial when another thermal governor is used?

Yes, it will be, the cpufreq_set_cur_state() is called by
thermal exported function:
      cdev->ops->set_cur_state(cdev, target)

So it can be called not only by IPA. All governors call it, because
that's the default mechanism.

True, but I'm still not convinced that it is useful outside `EAS - IPA`.

It is. So in mainline thermal there is another governor: fair_share [1],
which uses 'weights' to split the cooling effort across cooling devices
in the thermal zone. That governor would manage CPUs and GPU and
set throttling like IPA.

The mechanical side of the code would allow for such benefits, I just
don't know if their CPU cooling device + thermal zone setups would cater
for this?

Yes, it's possible. Even for custom vendor governors (modified clones
of IPA)

Let's stick to mainline here ;-) It's complicated enough ...

I agree, so there isn't only IPA in mainline.


Maybe shorter?

         struct cpumask *pd_mask = perf_domain_span(pd);
-       unsigned long cpu_cap =
+       int cpu = cpumask_first(pd_mask);
+       unsigned long cpu_cap = arch_scale_cpu_capacity(cpu);
+       unsigned long _cpu_cap = cpu_cap -
         unsigned long max_util = 0, sum_util = 0;
-       unsigned long _cpu_cap = cpu_cap;
-       int cpu;
-       _cpu_cap -= arch_scale_thermal_pressure(cpumask_first(pd_mask));

Could be, but still, the definitions should be sorted from longest on
top, to shortest at the bottom. I wanted to avoid modifying too many
lines with this simple patch.

Only if there are no dependencies, but here we have already `cpu_cap ->
pd_mask`. OK, not a big deal.

True, those dependencies are tricky to sort them properly, so I coded
it this way.


I see what you mean, but this might cause some issues in the design
(per-cpu scmi cpu perf control). Let's use this EM pointer gently ;)

OK, with the requirement that clients see the EM as ro:

Reviewed-by: Dietmar Eggemann <dietmar.eggemann@xxxxxxx>

Thank you Dietmar for the review!