Re: [PATCH RESEND] sched/nohz: Add HRTICK_BW for using cfs bandwidth with nohz_full

[Benjamin Segall <bsegall@google.com>]

Phil Auld <pauld@redhat.com> 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.  Use the hrtick mechanism to set a sched
> tick to fire at remaining_runtime in the future if we are on
> a nohz full cpu, if the task has quota and if we are likely to
> disable the tick (nr_running == 1).  This allows for bandwidth
> accounting before tasks go too far over quota.
>
> A number of container workloads use a dynamic number of real
> nohz tasks but also have other work that is limited which ends
> up running on the "spare" nohz cpus.  This is an artifact of
> having to specify nohz_full cpus at boot. Adding this hrtick
> resolves the issue of long stalls on these tasks.
>
> Add the sched_feat HRTICK_BW off by default to allow users to
> enable this only when needed.
>
> Signed-off-by: Phil Auld <pauld@redhat.com>
> Suggested-by: Juri Lelli <jlelli@redhat.com>
> Cc: Ingo Molnar <mingo@redhat.com>
> Cc: Peter Zijlstra <peterz@infradead.org>
> Cc: Vincent Guittot <vincent.guittot@linaro.org>
> Cc: Juri Lelli <juri.lelli@redhat.com>
> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com>
> Cc: Valentin Schneider <vschneid@redhat.com>
> Cc: Ben Segall <bsegall@google.com>
> ---
>
> Resend with LKML address instead of rh list...
>
>  kernel/sched/core.c     |  2 +-
>  kernel/sched/fair.c     | 20 ++++++++++++++++++++
>  kernel/sched/features.h |  1 +
>  kernel/sched/sched.h    | 12 ++++++++++++
>  4 files changed, 34 insertions(+), 1 deletion(-)
>
> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> index a68d1276bab0..76425c377245 100644
> --- a/kernel/sched/core.c
> +++ b/kernel/sched/core.c
> @@ -6562,7 +6562,7 @@ static void __sched notrace __schedule(unsigned int sched_mode)
>  
>  	schedule_debug(prev, !!sched_mode);
>  
> -	if (sched_feat(HRTICK) || sched_feat(HRTICK_DL))
> +	if (sched_feat(HRTICK) || sched_feat(HRTICK_DL) || sched_feat(HRTICK_BW))
>  		hrtick_clear(rq);
>  
>  	local_irq_disable();
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index 373ff5f55884..0dd1f6a874bc 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -5309,6 +5309,22 @@ static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
>  	return ret;
>  }
>  
> +#if defined(CONFIG_SCHED_HRTICK) && defined(CONFIG_NO_HZ_FULL)
> +static void start_hrtick_cfs_bw(struct rq *rq, struct cfs_rq *cfs_rq)
> +{
> +	if (!tick_nohz_full_cpu(rq->cpu) || !cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
> +		return;
> +
> +	/* runtime_remaining should never be negative here but just in case */
> +	if (rq->nr_running == 1 && cfs_rq->runtime_remaining > 0)
> +		hrtick_start(rq, cfs_rq->runtime_remaining);
> +}
> +#else
> +static void start_hrtick_cfs_bw(struct rq *rq, struct cfs_rq *cfs_rq)
> +{
> +}
> +#endif
> +
>  static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
>  {
>  	/* dock delta_exec before expiring quota (as it could span periods) */
> @@ -5481,6 +5497,7 @@ static bool throttle_cfs_rq(struct cfs_rq *cfs_rq)
>  	 */
>  	cfs_rq->throttled = 1;
>  	cfs_rq->throttled_clock = rq_clock(rq);
> +
>  	return true;
>  }
>  
> @@ -8096,6 +8113,9 @@ done: __maybe_unused;
>  	if (hrtick_enabled_fair(rq))
>  		hrtick_start_fair(rq, p);
>  
> +	if (hrtick_enabled_bw(rq))
> +		start_hrtick_cfs_bw(rq, task_cfs_rq(p));
> +
>  	update_misfit_status(p, rq);

Implementation-wise this winds up with a tick of
sysctl_sched_cfs_bandwidth_slice, which I suppose the admin could _also_
configure depending on how much NOHZ benefit vs cfsb issues they want.

It's not great that this implementation winds up going all the way
through schedule() for each 5ms-default tick, though.