Re: [PATCH] sched/fair: Prevent cpu_busy_time from exceeding actual_cpu_capacity

[Dietmar Eggemann <dietmar.eggemann@arm.com>]

On 07/06/2024 12:37, Xuewen Yan wrote:
> On Fri, Jun 7, 2024 at 6:30 PM Dietmar Eggemann
> <dietmar.eggemann@arm.com> wrote:
>>
>> On 07/06/2024 10:20, Xuewen Yan wrote:
>>> Hi Dietmar
>>>
>>> On Fri, Jun 7, 2024 at 3:19 PM Dietmar Eggemann
>>> <dietmar.eggemann@arm.com> wrote:
>>>>
>>>> On 06/06/2024 09:06, Xuewen Yan wrote:
>>>>> Because the effective_cpu_util() would return a util which
>>>>> maybe bigger than the actual_cpu_capacity, this could cause
>>>>> the pd_busy_time calculation errors.
>>>>
>>>> Doesn't return effective_cpu_util() either scale or min(scale, util)
>>>> with scale = arch_scale_cpu_capacity(cpu)? So the util sum over the PD
>>>> cannot exceed eenv->cpu_cap?
>>>
>>> In effective_cpu_util, the scale = arch_scale_cpu_capacity(cpu);
>>>  Although there is the clamp of eenv->pd_cap, but let us consider the
>>> following simple scenario:
>>> The pd cpus are 4-7, and the arch_scale_capacity is 1024, and because
>>> of cpufreq-limit,
>>
>> Ah, this is due to:
>>
>> find_energy_efficient_cpu()
>>
>>    ...
>>    for (; pd; pd = pd->next)
>>        ...
>>        cpu_actual_cap = get_actual_cpu_capacity(cpu)
>>
>>        for_each_cpu(cpu, cpus)
>>            ...
>>            eenv.pd_cap += cpu_actual_cap
>>
>> and:
>>
>> get_actual_cpu_capacity()
>>
>>    ...
>>    capacity = arch_scale_cpu_capacity(cpu)
>>
>>    capacity -= max(hw_load_avg(cpu_rq(cpu)), cpufreq_get_pressure(cpu))
>>
>> which got introduced by f1f8d0a22422 ("sched/cpufreq: Take cpufreq
>> feedback into account").
> 
> I don't think it was introduced by f1f8d0a22422, because f1f8d0a22422
> just replaced the cpu_thermal_cap with get_actual_cpu_capacity(cpu).
> The eenv.cpu_cap was  introduced by 3e8c6c9aac42 ("sched/fair: Remove
> task_util from effective utilization in feec()").

Yes, you're right. 3e8c6c9aac42 changed it from per-CPU to per-PD
capping.

In case we want to go back to per-CPU then we should remove the
eenv->pd_cap capping in eenv_pd_busy_time().

-->8--

@@ -7864,16 +7864,15 @@ static inline void eenv_pd_busy_time(struct energy_env *eenv,
                                     struct cpumask *pd_cpus,
                                     struct task_struct *p)
 {
-       unsigned long busy_time = 0;
        int cpu;
 
        for_each_cpu(cpu, pd_cpus) {
                unsigned long util = cpu_util(cpu, p, -1, 0);
 
-               busy_time += effective_cpu_util(cpu, util, NULL, NULL);
+               util = effective_cpu_util(cpu, util, NULL, NULL);
+               util = min(util, eenv->cpu_cap);
+               eenv->pd_busy_time += util;
        }
-
-       eenv->pd_busy_time = min(eenv->pd_cap, busy_time);
 }



I'm wondering whether we would need the:

if (dst_cpu >= 0)
    busy_time = min(eenv->pd_cap, busy_time + eenv->task_busy_time);

in compute_energy() anymore since we only get a candidate CPU in feec()
after checking with util_fits_cpu() if cpu can accommodate p :

feec()

    ...

    for_each_cpu()

        util = cpu_util(cpu, p, cpu, ...)
        cpu_cap = capacity_of()

        ...

        fits = util_fits_cpu(util, ..., cpu);
        if (!fits)
            continue

        /* check if candidate CPU */