Re: [PATCH] Sched/fair: Block nohz tick_stop when cfs bandwidth in use

From: Phil Auld
Date: Fri Jun 23 2023 - 09:09:53 EST


On Thu, Jun 22, 2023 at 05:37:30PM -0400 Phil Auld wrote:
> On Thu, Jun 22, 2023 at 01:49:52PM -0700 Benjamin Segall wrote:
> > Phil Auld <pauld@xxxxxxxxxx> writes:
> >
> > > CFS bandwidth limits and NOHZ full don't play well together. Tasks
> > > can easily run well past their quotas before a remote tick does
> > > accounting. This leads to long, multi-period stalls before such
> > > tasks can run again. Currentlyi, when presented with these conflicting
> > > requirements the scheduler is favoring nohz_full and letting the tick
> > > be stopped. However, nohz tick stopping is already best-effort, there
> > > are a number of conditions that can prevent it, whereas cfs runtime
> > > bandwidth is expected to be enforced.
> > >
> > > Make the scheduler favor bandwidth over stopping the tick by setting
> > > TICK_DEP_BIT_SCHED when the only running task is a cfs task with
> > > runtime limit enabled.
> > >
> > > Add sched_feat HZ_BW (off by default) to control this behavior.
> > >
> > > Signed-off-by: Phil Auld <pauld@xxxxxxxxxx>
> > > Cc: Ingo Molnar <mingo@xxxxxxxxxx>
> > > Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
> > > Cc: Vincent Guittot <vincent.guittot@xxxxxxxxxx>
> > > Cc: Juri Lelli <juri.lelli@xxxxxxxxxx>
> > > Cc: Dietmar Eggemann <dietmar.eggemann@xxxxxxx>
> > > Cc: Valentin Schneider <vschneid@xxxxxxxxxx>
> > > Cc: Ben Segall <bsegall@xxxxxxxxxx>
> > > ---
> > > kernel/sched/fair.c | 33 ++++++++++++++++++++++++++++++++-
> > > kernel/sched/features.h | 2 ++
> > > 2 files changed, 34 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > > index 373ff5f55884..880eadfac330 100644
> > > --- a/kernel/sched/fair.c
> > > +++ b/kernel/sched/fair.c
> > > @@ -6139,6 +6139,33 @@ static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
> > > rcu_read_unlock();
> > > }
> > >
> > > +#ifdef CONFIG_NO_HZ_FULL
> > > +/* called from pick_next_task_fair() */
> > > +static void sched_fair_update_stop_tick(struct rq *rq, struct task_struct *p)
> > > +{
> > > + struct cfs_rq *cfs_rq = task_cfs_rq(p);
> > > + int cpu = cpu_of(rq);
> > > +
> > > + if (!sched_feat(HZ_BW) || !cfs_bandwidth_used())
> > > + return;
> > > +
> > > + if (!tick_nohz_full_cpu(cpu))
> > > + return;
> > > +
> > > + if (rq->nr_running != 1 || !sched_can_stop_tick(rq))
> > > + return;
> > > +
> > > + /*
> > > + * We know there is only one task runnable and we've just picked it. The
> > > + * normal enqueue path will have cleared TICK_DEP_BIT_SCHED if we will
> > > + * be otherwise able to stop the tick. Just need to check if we are using
> > > + * bandwidth control.
> > > + */
> > > + if (cfs_rq->runtime_enabled)
> > > + tick_nohz_dep_set_cpu(cpu, TICK_DEP_BIT_SCHED);
> > > +}
> > > +#endif
> >
> > So from a CFS_BANDWIDTH pov runtime_enabled && nr_running == 1 seems
> > fine. But working around sched_can_stop_tick instead of with it seems
> > sketchy in general, and in an edge case like "migrate a task onto the
> > cpu and then off again" you'd get sched_update_tick_dependency resetting
> > the TICK_DEP_BIT and then not call PNT (ie a task wakes up onto this cpu
> > without preempting, and then another cpu goes idle and pulls it, causing
> > this cpu to go into nohz_full).
> >
>
> The information to make these tests is not available in sched_can_stop_tick.
> I did start there. When that is called, and we are likely to go nohz_full,
> curr is null so it's hard to find the right cfs_rq to make that
> runtime_enabled test against. We could, maybe, plumb the task being enqueued
> in but it would not be valid for the dequeue path and would be a bit messy.
>

Sorry, mispoke... rq->curr == rq-idle not null. But still we don't have
access to the task and its cfs_rq which will have runtime_enabled set.

> But yes, I suppose you could end up in a state that is just as bad as today.
>
> Maybe I could add a redundant check in sched_can_stop_tick for when
> nr_running == 1 and curr is not null and make sure the bit does not get
> cleared. I'll look into that.
>
>
> Thanks,
> Phil
>
> --
>

--