Re: [PATCH] sched: Fix PI handling vs sched_setscheduler()

From: Juri Lelli
Date: Wed Jan 20 2016 - 05:33:58 EST


On 20/01/16 11:24, Peter Zijlstra wrote:
> On Wed, Jan 20, 2016 at 10:08:39AM +0000, Juri Lelli wrote:
> > Hi,
> >
> > On 19/01/16 12:18, Peter Zijlstra wrote:
> >
> > [...]
> >
> > > @@ -4097,15 +4099,14 @@ static int __sched_setscheduler(struct t
> > > if (running)
> > > p->sched_class->set_curr_task(rq);
> > > if (queued) {
> > > - int enqueue_flags = ENQUEUE_RESTORE;
> > > /*
> > > * We enqueue to tail when the priority of a task is
> > > * increased (user space view).
> > > */
> > > - if (oldprio <= p->prio)
> > > - enqueue_flags |= ENQUEUE_HEAD;
> > > + if (oldprio < p->prio)
> > > + queue_flags |= ENQUEUE_HEAD;
> >
> > Was this condition broken before or it needs to be changed now with this
> > patch?
>
> I could not see how we could get there before, seeing how the == case is
> handled above.
>
> > > --- a/kernel/sched/sched.h
> > > +++ b/kernel/sched/sched.h
> > > @@ -1130,18 +1130,35 @@ static inline void finish_lock_switch(st
> > > extern const int sched_prio_to_weight[40];
> > > extern const u32 sched_prio_to_wmult[40];
> > >
> > > +/*
> > > + * {de,en}queue flags:
> > > + *
> > > + * SAVE/RESTORE - an otherwise spurious dequeue/enqueue, done to ensure tasks
> > > + * are in a known state which allows modification. Such pairs
> > > + * should preserve as much state as possible.
> > > + *
> > > + * MOVE - paired with SAVE/RESTORE, explicitly does not preserve the location
> > > + * in the runqueue.
> > > + *
> > > + * ENQUEUE_HEAD - place at front of runqueue (tail if not specificed)
> > > + *
> > > + */
> >
> > Do we want to document all the flags while we are at it? :)
>
> Probably.. does the below work? I think I bailed on replenish because I
> could not come up with a coherent short description, does the below make
> any sense?
>
>
> --- a/kernel/sched/sched.h
> +++ b/kernel/sched/sched.h
> @@ -1140,7 +1140,9 @@ extern const u32 sched_prio_to_wmult[40]
> * MOVE - paired with SAVE/RESTORE, explicitly does not preserve the location
> * in the runqueue.
> *
> - * ENQUEUE_HEAD - place at front of runqueue (tail if not specificed)
> + * ENQUEUE_HEAD - place at front of runqueue (tail if not specificed)
> + * ENQUEUE_REPLENISH - push the CBS slot forward


* ENQUEUE_REPLENISH - CBS (replenish runtime and postpone deadline)

Maybe? :)

> + * ENQUEUE_WAKING - sched_class::task_waking was called
> *
> */
>
>