Re: [PATCH 08/20 v3] sched/fair: Add push task mechanism for fair
From: Vincent Guittot
Date: Thu Nov 12 2026 - 02:06:00 EST
On Wed, 11 Nov 2026 at 09:59, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>
> On Mon, Nov 09, 2026 at 05:07:31PM +0100, Vincent Guittot wrote:
> > On Mon, 9 Nov 2026 at 15:43, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> > >
> > > On Fri, Nov 06, 2026 at 04:19:35PM +0100, Vincent Guittot wrote:
> > > > +static struct task_struct *pick_next_pushable_fair_task(struct rq *rq)
> > > > +{
> > > > + struct task_struct *p;
> > > > +
> > > > + if (!has_pushable_tasks(rq))
> > > > + return NULL;
> > > > +
> > > > + p = plist_first_entry(&rq->cfs.pushable_tasks,
> > > > + struct task_struct, pushable_tasks);
> > > > +
> > > > + WARN_ON_ONCE(rq->cpu != task_cpu(p));
> > > > + WARN_ON_ONCE(task_current(rq, p));
> > > > + WARN_ON_ONCE(p->nr_cpus_allowed <= 1);
> > > > + WARN_ON_ONCE(!task_on_rq_queued(p));
> > > > +
> > > > + /*
> > > > + * Remove task from the pushable list as we try only once after that
> > > > + * the task has been put back in enqueued list.
> > > > + */
> > > > + plist_del(&p->pushable_tasks, &rq->cfs.pushable_tasks);
> > > > +
> > > > + return p;
> > > > +}
> > >
> > > This means a task will only be pushed once? After which it needs to be
> > > re-enqueued on the runqueue?
> >
> > Yes, my goal is to keep the list small to avoid ending up with many
> > tasks to push. So I only check if prev could be pushed somewhere else
> >
> > >
> > > So if it was found during migrate_disable(), its thereafter stuck? --
> > > this seems a little weird.
> >
> > If a task is migrate disabled while we are about to look for a
> > possibly better CPU, it will have to wait until after the next time it
> > becomes current. This is a simple way to pace the number of tasks to
> > push in one callback
>
> There is still a worst case where a lot of tasks get woken, they all get
> enqueued and then they all get on this list and you need to iterate them
> all.
Yes but this case still limits the number of tasks added to the list
because of the case which involves short slice preemption and
simultaneous enqueue on an idle CPU
>
> If you want to bound this thing, you might as well do like
> detach_one_task(). Only ever consider the first task from the end that
> is suitable.
>