Re: [PATCH v2] sched/fair: sanitize vruntime of entity being migrated

From: Vincent Guittot
Date: Wed Mar 15 2023 - 03:19:02 EST


On Tue, 14 Mar 2023 at 18:16, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>
> On Tue, Mar 14, 2023 at 02:24:37PM +0100, Vincent Guittot wrote:
>
> > > @@ -7632,11 +7646,8 @@ static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
> > > * min_vruntime -- the latter is done by enqueue_entity() when placing
> > > * the task on the new runqueue.
> > > */
> > > - if (READ_ONCE(p->__state) == TASK_WAKING) {
> > > - struct cfs_rq *cfs_rq = cfs_rq_of(se);
> > > -
> > > + if (READ_ONCE(p->__state) == TASK_WAKING || reset_vruntime(cfs_rq, se))
> >
> > That's somehow what was proposed in one of the previous proposals but
> > we can't call rq_clock_task(rq_of(cfs_rq)) because rq lock might not
> > be hold and rq task clock has not been updated before being used
>
> Argh indeed. I spend a lot of time ensuring we didn't take the old rq
> lock on wakeup -- and then a lot of time cursing about how we don't :-)
>
> Now, if we could rely on the rq-clock being no more than 1 tick behind
> current, this would still be entirely sufficient to catch the long sleep
> case.

We should also take care when loading rq_clock_task that we are not
racing with an update especially for a 32bits system like pelt
last_update_time

>
> Except I suppose that NOHZ can bite us here. If the old CPU is idle, the
> timestamps can be arbitrarily old. Mooo :/

That should not be a real problem because if the cpu is idle and the
rq clock is not updated, the min_vruntime will not move forward so we
are "safe" in regard to the overflow.

That's what was done in the v2 and v3 of this patch


>
>