Re: [PATCH v2] sched/fair: Sync se's load_avg with cfs_rq in reweight_task

[Chengming Zhou]

On 2024/7/22 13:17, K Prateek Nayak wrote:
> (+ Qais)
>
> Hello Chuyi,
>
> On 7/20/2024 10:42 AM, Chuyi Zhou wrote:
> > In reweight_task(), there are two situations:
> >
> > 1. The task was on_rq, then the task's load_avg is accurate because we
> > synchronized it with cfs_rq through update_load_avg() in dequeue_task().
> >
> > 2. The task is sleeping, its load_avg might not have been updated for 
> > some
> > time, which can result in inaccurate dequeue_load_avg() in
> > reweight_entity().
> >
> > This patch solves this by using update_load_avg() to synchronize the
> > load_avg of se with cfs_rq. For tasks were on_rq, since we already update
> > load_avg to accurate values in dequeue_task(), this change will not have
> > other effects due to the short time interval between the two updates.
> >
> > Signed-off-by: Chuyi Zhou <zhouchuyi@xxxxxxxxxxxxx>
> > ---
> > Changes in v2:
> > - change the description in commit log.
> > - use update_load_avg() in reweight_task() rather than in reweight_entity
> > suggested by chengming.
> > - Link to v1: 
> > https://lore.kernel.org/lkml/20240716150840.23061-1-zhouchuyi@xxxxxxxxxxxxx/
> > ---
> >   kernel/sched/fair.c | 3 +++
> >   1 file changed, 3 insertions(+)
> >
> > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > index 9057584ec06d..b1e07ce90284 100644
> > --- a/kernel/sched/fair.c
> > +++ b/kernel/sched/fair.c
> > @@ -3835,12 +3835,15 @@ static void reweight_entity(struct cfs_rq 
> > *cfs_rq, struct sched_entity *se,
> >       }
> >   }
> > +static inline void update_load_avg(struct cfs_rq *cfs_rq, struct 
> > sched_entity *se, int flags);
> > +
> >   void reweight_task(struct task_struct *p, const struct load_weight *lw)
> >   {
> >       struct sched_entity *se = &p->se;
> >       struct cfs_rq *cfs_rq = cfs_rq_of(se);
> >       struct load_weight *load = &se->load;
> > +    update_load_avg(cfs_rq, se, 0);
>
> Seems to be necessary when we reach here from __setscheduler_params() or
> set_user_nice() for a sleeping task. Please feel free to add:
>
> Reviewed-by: K Prateek Nayak <kprateek.nayak@xxxxxxx>
>
> But since we are on the subject of accurate PELT accounting, one question
> I have here is whether a reweight_task() for a sleeping task race with
> its wakeup? Something like the following scenario:

These two paths are impossible to race, since we have the task->pi_lock.

Thanks.

> CPU0                                        CPU1
> ====                                        ====
> /* No rq locks held until ttwu_queue() */
> try_to_wake_up(p) {
>      ...
>      /* Migrating task */
>      set_task_cpu(p, cpu) {
>          /* p->sched_class->migrate_task_rq(p, new_cpu); */        /* 
> Called with task_cpu(p)'s rq lock held */
>          migrate_task_rq_fair() {                    reweight_task(p) {
>              /* p is still sleeping */                    ...
>              if (!task_on_rq_migrating(p)) {                    
> dequeue_load_avg(cfs_rq, se);
>                          remove_entity_load_avg(se);                
> update_load_set(&se->load, weight);
>                  ...                            enqueue_load_avg(cfs_rq, 
> se);
>              }                                ...
>          }                                }
>          /* task_cpu() is updated here */
>          __set_task_cpu(p, new_cpu);
>      }
>      ttwu_queue();
> }
>
> In theory, the remove_entity_load_avg() could record stale value of
> "load_avg" that gets removed at the next dequeue if I'm not mistaken?
> But I believe these small inaccuracies are tolerable since they'll decay
> in a while anyways?
>
> >       reweight_entity(cfs_rq, se, lw->weight);
> >       load->inv_weight = lw->inv_weight;
> >   }