Re: [RFC][PATCH 3/3] sched/eevdf: Fix sched idle task group preemption
From: Josh Don
Date: Fri Nov 13 2026 - 15:38:48 EST
On Thu, Nov 12, 2026 at 4:38 AM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>
> On Thu, Nov 12, 2026 at 01:31:19PM +0100, Peter Zijlstra wrote:
> > On Tue, Nov 10, 2026 at 09:51:21AM -0800, Josh Don wrote:
> > > On Tue, Nov 10, 2026 at 6:20 AM Kayra Cizmeci <kayracizmeci@xxxxxxxxx> wrote:
> > > >
> > > > Hi Josh!
> > > >
> > > > On Mon, Nov 09, 2026 at 04:10:30PM -0800, Josh Don wrote:
> > > >
> > > > > > /*
> > > > > > - * Preempt the current task with a newly woken task if needed:
> > > > > > + * Preempt an idle entity in favor of a non-idle entity (and don't preempt
> > > > > > + * in the inverse case).
> > > > > > + */
> > > > > > +static inline enum preempt_wakeup_action
> > > > > > +wakeup_preempt_sched_idle(struct sched_entity *cse, struct sched_entity *pse)
> > > > > > +{
> > > > > > + int cse_is_idle, pse_is_idle;
> > > > > > +
> > > > > > + cse_is_idle = se_is_idle(cse);
> > > > > > + pse_is_idle = se_is_idle(pse);
> > > > > > +
> > > > > > + if (cse_is_idle || pse_is_idle) {
> > > > > > + find_matching_se(&cse, &pse);
> > > > > > + WARN_ON_ONCE(!pse);
> > > > > > +
> > > > > > + cse_is_idle = se_is_idle(cse);
> > > > > > + pse_is_idle = se_is_idle(pse);
> > > > >
> > > > > now that se_is_idle() is checking for any ancestors with idle, this
> > > > > doesn't appear to be correct. Some motivaing examples:
> > > >
> > > > So we want to keep the idle difference between groups with idle
> > > > ancestors? FWIW, I asked about this but then thought it was knowingly
> > > > and that was what we wanted. I'm still not sure what we want...
> > >
> > > cgroup properties are cgroup hierarchy position relative, this just
> > > maintins that behavior for cpu.idle as well. There was a more
> > > extensive discussion on this in the previous RFC, which is how we
> > > landed back on the find_matching_se() restoration.
> >
> > Well, I'm still not convinced we need this.
> >
> > Realistically idle means: "don't care, run whenever you have a minute".
> > To then preempt such a task for another such a task just because its
> > somewhere else in the hierarchy makes no sense, you don't care about
> > either of them, we shouldn't be preempting.
> >
> > Anyway, I've pushed out the first two patches to queue/sched/urgent
> > since they address the worst of it. We can chew on all this a bit more I
> > suppose. But I cannot convince myself that there is any situation in
> > which we (should) actually care about the timeliness of anything in a
> > cgroup::idle container.
>
> Perhaps even to the point that we should have idle (either the task or
> the cgroup property) imply an extra long slice length, much like how it
> implies a very low weight.
Does the example I gave earlier with separate users (jobs) that can
subdivide their own cpu slice into regular and "idle" usage not feel
convincing to you? If a user is entitled to 50% of the cpu, and they
have both latency sensitive and background batch work to do, they
still want their batch work entitled to forward progress under the
job's overall 50% system cpu share.