Messages in this thread | | | From | Vincent Guittot <> | Date | Tue, 19 Jul 2022 12:29:31 +0200 | Subject | Re: [External] Re: [PATCH v2 07/10] sched/fair: use update_load_avg() to attach/detach entity load_avg |
| |
On Fri, 15 Jul 2022 at 18:21, Chengming Zhou <zhouchengming@bytedance.com> wrote: > > On 2022/7/15 19:18, Dietmar Eggemann wrote: > > On 13/07/2022 06:04, Chengming Zhou wrote: > >> Since update_load_avg() support DO_ATTACH and DO_DETACH now, we can > >> use update_load_avg() to implement attach/detach entity load_avg. > >> > >> Another advantage of using update_load_avg() is that it will check > >> last_update_time before attach or detach, instead of unconditional > >> attach/detach in the current code. > >> > >> This way can avoid some corner problematic cases of load tracking, > >> like twice attach problem, detach unattached NEW task problem. > > > > This explanation is somewhat hard to follow for me. Since both issues > > have been fixed already (you mention this further below) you're saying > > that with you change you don't reintroduce them? > > Sorry for this not very clear explanation. > > Yes, both issues have been fixed already, what I want to say is that bugfix > brings its own problem and limitation mentioned below.
As Dietmar said, the commit message is misleading because someone can think you fix these bugs whereas it's not the case
> > So I want to use another way to solve these problems better. > > > > >> 1. switch to fair class (twice attach problem) > >> > >> p->sched_class = fair_class; --> p.se->avg.last_update_time = 0 > >> if (queued) > >> enqueue_task(p); > >> ... > >> enqueue_entity() > >> update_load_avg(UPDATE_TG | DO_ATTACH) > >> if (!se->avg.last_update_time && (flags & DO_ATTACH)) --> true > >> attach_entity_load_avg() --> attached, will set last_update_time > >> check_class_changed() > >> switched_from() (!fair) > >> switched_to() (fair) > >> switched_to_fair() > >> attach_entity_load_avg() --> unconditional attach again! > >> > >> 2. change cgroup of NEW task (detach unattached task problem) > >> > >> sched_move_group(p) > >> if (queued) > >> dequeue_task() > >> task_move_group_fair() > >> detach_task_cfs_rq() > >> detach_entity_load_avg() --> detach unattached NEW task > >> set_task_rq() > >> attach_task_cfs_rq() > >> attach_entity_load_avg() > >> if (queued) > >> enqueue_task() > >> > >> These problems have been fixed in commit 7dc603c9028e > >> ("sched/fair: Fix PELT integrity for new tasks"), which also > >> bring its own problems. > >> > >> First, it add a new task state TASK_NEW and an unnessary limitation > >> that we would fail when change the cgroup of TASK_NEW tasks. > > This is the limitation that bugfix has brought. > > We can't change cgroup or switch to fair for task with last_update_time=0 > if we don't have conditional detach/attach. > > So we have to: > > 1. !fair task also need to set last_update_time. > 2. cpu_cgroup_can_attach() have to wait for TASK_NEW to fully attached. > > >> > >> Second, it attach entity load_avg in post_init_entity_util_avg(), > >> in which we only set sched_avg last_update_time for !fair tasks, > >> will cause PELT integrity problem when switched_to_fair(). > > > > I guess those PELT integrity problems are less severe since we have the > > enqueue_task_fair() before the switched_to_fair() for enqueued tasks. So > > we always decay the time the task spend outside fair. > > enqueue_task_fair() > enqueue_entity() > update_load_avg() > if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD)) --> true > __update_load_avg_se(now, cfs_rq, se); --> (1) > > We can see above, for queued !fair task, (1) will deay the delta time > (now - se.avg.last_update_time) even for a NEW !fair task. > > > > > Looks to me that you want to replace this by your `freeze PELT when > > outside fair` model. > > Yes, want to freeze PELT for two !fair cases: > > 1. !fair task hasn't been fair before: will still have its init load_avg > when switch to fair.
But I'm not sure it makes sense to keep these init values. As an example, the util_avg is set according to the cpu utilization at the time of the task creation. I would tend to decay them as these init values become less and less relevant.
so we should return early in post_init_entity_util_avg() and don't set util_avg if sched class is not cfs
> > 2. !fair task has been switched_from_fair(): will still keep its lastest > load_avg when switch to fair. > > > > >> This patch make update_load_avg() the only location of attach/detach, > >> and can handle these corner cases like change cgroup of NEW tasks, > >> by checking last_update_time before attach/detach. > > > > [...] > > > >> @@ -11527,9 +11522,7 @@ static void detach_entity_cfs_rq(struct sched_entity *se) > >> struct cfs_rq *cfs_rq = cfs_rq_of(se); > >> > >> /* Catch up with the cfs_rq and remove our load when we leave */ > >> - update_load_avg(cfs_rq, se, 0); > >> - detach_entity_load_avg(cfs_rq, se); > >> - update_tg_load_avg(cfs_rq); > >> + update_load_avg(cfs_rq, se, UPDATE_TG | DO_DETACH); > > > > IMHO, the DO_[DE|AT]TACH comments in update_load_avg() would have to be > > updated in this case. > > Correct, will do. > > Thanks. > > > > > [...]
| |