Re: [PATCH 4/4] sched,fair: Fix PELT integrity for new tasks

[Dietmar Eggemann]

On 20/06/16 13:35, Vincent Guittot wrote:
> On 20 June 2016 at 13:35, Dietmar Eggemann <dietmar.eggemann@xxxxxxx> wrote:
>>
>>
>> On 17/06/16 17:18, Peter Zijlstra wrote:
>>> On Fri, Jun 17, 2016 at 06:02:39PM +0200, Peter Zijlstra wrote:
>>>> So yes, ho-humm, how to go about doing that bestest. Lemme have a play.
>>>
>>> This is what I came up with, not entirely pretty, but I suppose it'll
>>> have to do.
>>>
>>> ---
>>> --- a/kernel/sched/fair.c
>>> +++ b/kernel/sched/fair.c
>>> @@ -724,6 +724,7 @@ void post_init_entity_util_avg(struct sc
>>>       struct cfs_rq *cfs_rq = cfs_rq_of(se);
>>>       struct sched_avg *sa = &se->avg;
>>>       long cap = (long)(SCHED_CAPACITY_SCALE - cfs_rq->avg.util_avg) / 2;
>>> +     u64 now = cfs_rq_clock_task(cfs_rq);
>>>
>>>       if (cap > 0) {
>>>               if (cfs_rq->avg.util_avg != 0) {
>>> @@ -738,7 +739,20 @@ void post_init_entity_util_avg(struct sc
>>>               sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
>>>       }
>>>
>>> -     update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, false);
>>> +     if (entity_is_task(se)) {
>>> +             struct task_struct *p = task_of(se);
>>> +             if (p->sched_class != &fair_sched_class) {
>>> +                     /*
>>> +                      * For !fair tasks do attach_entity_load_avg()
>>> +                      * followed by detach_entity_load_avg() as per
>>> +                      * switched_from_fair().
>>> +                      */
>>> +                     se->avg.last_update_time = now;
>>> +                     return;
>>> +             }
>>> +     }
>>> +
>>> +     update_cfs_rq_load_avg(now, cfs_rq, false);
>>>       attach_entity_load_avg(cfs_rq, se);
>>>  }
>>>
>>>
>>
>> Doesn't a sleeping !fair_sched_class task which switches to fair uses
>> try_to_wake_up() rather than wake_up_new_task() so it won't go through
>> post_init_entity_util_avg()?
> 
> It will go through wake_up_new_task and post_init_entity_util_avg
> during its fork which is enough to set last_update_time. Then, it will
> use the switched_to_fair if the task becomes a fair one

Oh I see. We want to make sure that every task (even when forked as
!fair) has a last_update_time value != 0, when becoming fair one day.