Re: [PATCH 4/7] sched/fair: Take care of group/affinity/sched_class change for throttled task

[Aaron Lu <ziqianlu@bytedance.com>]

On Fri, May 23, 2025 at 04:59:42PM +0200, Peter Zijlstra wrote:
> On Thu, May 22, 2025 at 08:49:43PM +0800, Aaron Lu wrote:
> > On Thu, May 22, 2025 at 02:03:36PM +0200, Peter Zijlstra wrote:
> 
> > > This is asymmetric -- dequeue removes it from that throttle list, but
> > > the corresponding enqueue will not add it back, what gives?
> > > 
> > > Because now we have:
> > > 
> > >  p->on_rq=1
> > >  p->throttle_node on list
> > > 
> > > move_queued_task()
> > >   deactivate_task()
> > >     dequeue_task_fair()
> > >       list_del_init(throttle_node)
> > >     p->on_rq = 2
> > > 
> > >   activate_task()
> > >     enqueue_task_fair()
> > >       // nothing special, makes the thing runnable
> > >     p->on_rq = 1;
> > > 
> > > and we exit with a task that is on-rq and not throttled ?!?
> > >
> > > Why is this? Are we relying on pick_task_fair() to dequeue it again and
> > > fix up our inconsistencies? If so, that had better have a comment on.
> > 
> > Correct.
> 
> But would it not be better to have enqueue bail when we're trying to
> enqueue an already throttled task into a throttled cfs_rq?
> 
> It seems a waste to do the actual enqueue, pick, dequeue when we
> could've just avoided all that.
>

The original idea is to keep code simple but surely this can be
optimized. I'm working on it and will paste diff here once I get it
work.

Thanks,
Aaron

> The immediate problem seems to be that you destroy the
> task_is_throttled() state on dequeue, but surely that is trivially
> fixable by not keeping that state in the list.