Re: [PATCH] sched/fair: Remove the duplicate check from group_has_capacity()

[Valentin Schneider]

On 11/08/20 12:44, Qi Zheng wrote:
> On 2020/8/11 下午6:38, Valentin Schneider wrote:
>>
>> On 11/08/20 04:39, Qi Zheng wrote:
>>> On 2020/8/11 上午2:33, Valentin Schneider wrote:
>>>>
>>>> On 10/08/20 02:00, Qi Zheng wrote:
>>>>> 1. The group_has_capacity() function is only called in
>>>>>      group_classify().
>>>>> 2. The following inequality has already been checked in
>>>>>      group_is_overloaded() which was also called in
>>>>>      group_classify().
>>>>>
>>>>>         (sgs->group_capacity * imbalance_pct) <
>>>>>                           (sgs->group_runnable * 100)
>>>>>
>>>>
>>>> Consider group_is_overloaded() returns false because of the first
>>>> condition:
>>>>
>>>>           if (sgs->sum_nr_running <= sgs->group_weight)
>>>>                   return false;
>>>>
>>>> then group_has_capacity() would be the first place where the group_runnable
>>>> vs group_capacity comparison would be done.
>>>>
>>>> Now in that specific case we'll actually only check it if
>>>>
>>>>     sgs->sum_nr_running == sgs->group_weight
>>>>
>>>> and the only case where the runnable vs capacity check can fail here is if
>>>> there's significant capacity pressure going on. TBH this capacity pressure
>>>> could be happening even when there are fewer tasks than CPUs, so I'm not
>>>> sure how intentional that corner case is.
>>>
>>> Maybe some cpus in sg->cpumask are no longer active at the == case,
>>> which causes the significant capacity pressure?
>>>
>>
>> That can only happen in that short window between deactivating a CPU and
>> not having rebuilt the sched_domains yet, which sounds quite elusive.
>>
>
> In fact, at the beginning, I added unlikely() here to hint the compiler:
>
> -	if ((sgs->group_capacity * imbalance_pct) <
> -			(sgs->group_runnable * 100))
> +	if (unlikely((sgs->group_capacity * imbalance_pct) <
> +			(sgs->group_runnable * 100)))
>
> The corresponding patch is as follows:
>
>       [PATCH]sched/core: add unlikely in group_has_capacity()
>
> Do you think it is necessary?

The "unlikely" approach has the benefit of keeping all corner cases in
place. I was tempted to say it could still make sense to get rid of the
extra check entirely, given that it has an impact only when:

- sum_nr_running == group_weight
- group capacity has been noticeably reduced

If sum_nr_running < group_weight, we won't evaluate it.
If sum_nr_running > group_weight, we either won't call into
  group_has_capacity() or we'll have checked it already in
  group_overloaded().

That said, it does make very much sense to check it in that ==
case. Vincent might have a different take on this, but right now I'd say
the unlikely approach is the safest one of the two.