Re: [PATCH v2] Sched/fair: Block nohz tick_stop when cfs bandwidth in use

[Benjamin Segall]

Phil Auld <pauld@xxxxxxxxxx> writes:

> On Wed, Jun 28, 2023 at 02:42:16PM -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>
>> > ---
>> >
>> > v2:  Ben pointed out that the bit could get cleared in the dequeue path
>> > if we migrate a newly enqueued task without preempting curr. Added a 
>> > check for that edge case to sched_can_stop_tick. Removed the call to 
>> > sched_can_stop_tick from sched_fair_update_stop_tick since it was 
>> > redundant.
>> >
>> >  kernel/sched/core.c     | 12 +++++++++++
>> >  kernel/sched/fair.c     | 45 +++++++++++++++++++++++++++++++++++++++++
>> >  kernel/sched/features.h |  2 ++
>> >  3 files changed, 59 insertions(+)
>> >
>> > diff --git a/kernel/sched/core.c b/kernel/sched/core.c
>> > index a68d1276bab0..646f60bfc7e7 100644
>> > --- a/kernel/sched/core.c
>> > +++ b/kernel/sched/core.c
>> > @@ -1194,6 +1194,8 @@ static void nohz_csd_func(void *info)
>> >  #endif /* CONFIG_NO_HZ_COMMON */
>> >  
>> >  #ifdef CONFIG_NO_HZ_FULL
>> > +extern bool sched_cfs_bandwidth_active(struct cfs_rq *cfs_rq);
>> > +
>> >  bool sched_can_stop_tick(struct rq *rq)
>> >  {
>> >  	int fifo_nr_running;
>> > @@ -1229,6 +1231,16 @@ bool sched_can_stop_tick(struct rq *rq)
>> >  	if (rq->nr_running > 1)
>> >  		return false;
>> >  
>> > +	/*
>> > +	 * If there is one task and it has CFS runtime bandwidth constraints
>> > +	 * and it's on the cpu now we don't want to stop the tick.
>> > +	 */
>> > +	if (sched_feat(HZ_BW) && rq->nr_running == 1 && rq->curr
>> > +	    && rq->curr->sched_class == &fair_sched_class && task_on_rq_queued(rq->curr)) {
>> > +		if (sched_cfs_bandwidth_active(task_cfs_rq(rq->curr)))
>> 
>> Actually, something I should have noticed earlier is that this should
>> probably be hierarchical, right? You need to check every ancestor
>> cfs_rq, not just the immediate parent. And at that point it probably
>> makes sense to have sched_cfs_bandwidth_active take a task_struct.
>> 
>
> Are you saying a child cfs_rq with a parent that has runtime_enabled could
> itself not have runtime_enabled?   I may be missing something but I don't
> see how that works.

Correct.

>
> account_cfs_rq_runtime() for example just looks at the immediate cfs_rq of
> curr and bails if it does not have runtime_enabled. How could that task get
> throttled if it exceeds some parent's limit?

account_cfs_rq_runtime() is called (primarily) from update_curr(), which
is called by enqueue_entity/dequeue_entity/entity_tick/etc, which are
called at each level of the hierarchy.

The worse cache behavior of doing a separate walk in sched_can_stop_tick
aka add/sub_nr_running could I guess be avoided by having some
runtime_enabled flag on the task struct or rq that is up to date for
rq->curr only. That would only be a little annoying to keep accurate,
and there's the dual arguments of "task_struct/rq is already too
cluttered"/"well they're already so cluttered a little more won't hurt".