Re: [Patch v8 4/7] sched/fair: Enable periodic update of average thermal pressure

[Dietmar Eggemann]

On 17/01/2020 16:39, Vincent Guittot wrote:
> On Fri, 17 Jan 2020 at 15:55, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>>
>> On Fri, Jan 17, 2020 at 02:22:51PM +0100, Vincent Guittot wrote:
>>> On Thu, 16 Jan 2020 at 16:15, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>>
>>>>
>>>> That there indentation trainwreck is a reason to rename the function.
>>>>
>>>>         decayed = update_rt_rq_load_avg(now, rq, curr_class == &rt_sched_class) |
>>>>                   update_dl_rq_load_avg(now, rq, curr_class == &dl_sched_class) |
>>>>                   update_thermal_load_avg(rq_clock_task(rq), rq, thermal_pressure) |
>>>>                   update_irq_load_avg(rq, 0);
>>>>
>>>> Is much better.
>>>>
>>>> But now that you made me look at that, I noticed it's using a different
>>>> clock -- it is _NOT_ using now/rq_clock_pelt(), which means it'll not be
>>>> in sync with the other averages.
>>>>
>>>> Is there a good reason for that?
>>>
>>> We don't need to apply frequency and cpu capacity invariance on the
>>> thermal capping signal which is  what rq_clock_pelt does
>>
>> Hmm, I suppose that is true, and that really could've done with a
>> comment. Now clock_pelt is sort-of in sync with clock_task, but won't it
>> still give weird artifacts by having it on a slightly different basis?
> 
> No we should not. Weird artifacts happens when we
> add/subtract/propagate signals between each other and then apply pelt
> algorithm on the results. In the case of thermal signal, we only add
> it to others to update cpu_capacity but pelt algo is then not applied
> on it. The error because of some signals being at segment boundaries
> whereas others are not, is limited to 2% and doesn't accumulate over
> time.
> 
>>
>> Anyway, looking at this, would it make sense to remove the @now argument
>> from update_*_load_avg()? All those functions already take @rq, and
>> rq_clock_*() are fairly trivial inlines.
> 
> TBH I was thinking of doing the opposite for update_irq_load_avg which
> hides the clock that is used for irq_avg. This helps to easily
> identify which signals use the exact same clock and can be mixed to
> create a new pelt signal and which can't

The 'now' argument is one thing but why not:

-int update_thermal_load_avg(u64 now, struct rq *rq, u64 capacity)
+int update_thermal_load_avg(u64 now, struct rq *rq)
 {
+       u64 capacity = arch_cpu_thermal_pressure(cpu_of(rq));
+
        if (___update_load_sum(now, &rq->avg_thermal,

This would make the call-sites __update_blocked_others() and
task_tick(_fair)() cleaner.

I guess the argument is not to pollute pelt.c. But it already contains
arch_scale_[freq|cpu]_capacity() for irq.