Re: [PATCH v2] sched_ext: Count cap-rejected local DSQ inserts in SCX_EV_SUB_REJECT

[Andrea Righi <arighi@nvidia.com>]

Hi Liang,

On Tue, Sep 29, 2026 at 11:03:56AM +0800, luoliang@kylinos.cn wrote:
> From: Liang Luo <luoliang@kylinos.cn>
> 
> When a task is enqueued on a local DSQ without the caps its enqueue flags
> require, __scx_resolve_local_dsq() has three outcomes: admit the task
> anyway because the rq is draining offline, the task is migration-disabled
> or a migration is pending (counted in SCX_EV_SUB_FORCED_ADMIT), divert it
> to the rescue path when the enqueue requested rescue (counted in
> SCX_EV_SUB_RESCUE), or divert it to the reject DSQ to be re-enqueued so
> that the BPF scheduler can re-decide.
> 
> Unlike the other two, which require a draining rq, a task that cannot
> migrate anywhere else, or an explicit SCX_ENQ_RESCUE opt-in, the reject
> diversion has no preconditions: every other cap-missing insert lands
> there. Its traffic is currently not visible in any counter -
> SCX_EV_REENQ_REPEAT only counts it once the same task fails its
> placement again, and nr_rejected counts tasks refused by ops.init_task()
> through task->scx.disallow.
> 
> Add SCX_EV_SUB_REJECT and count the reject diversion so that the events
> cover all three outcomes of a cap-missing local insert.

I still don't see a need for this event. The BPF scheduler receives
SCX_TASK_REENQ_CAP on cap-related re-enqueues and can count those itself, as
scx_qmap already does. That count isn't identical to the proposed diversion
count, but the commit message doesn't explain why the distinction matters.

Is there a concrete case where this new counter would help diagnose a problem
that the scheduler-side count cannot?

Thanks,
-Andrea

> 
> Signed-off-by: Liang Luo <luoliang@kylinos.cn>
> 
> ---
> 
> v2: Rework the rationale in the commit message and drop the Fixes: tag
>     per Tejun. No code changes.
> ---
>  kernel/sched/ext/internal.h | 10 +++++++++-
>  kernel/sched/ext/sub.c      |  2 ++
>  2 files changed, 11 insertions(+), 1 deletion(-)
> 
> diff --git a/kernel/sched/ext/internal.h b/kernel/sched/ext/internal.h
> index 0bcceab612de..36da1a6560db 100644
> --- a/kernel/sched/ext/internal.h
> +++ b/kernel/sched/ext/internal.h
> @@ -1308,6 +1308,13 @@ struct scx_event_stats {
>  	 * caps for its cid and the task entered the rescue path.
>  	 */
>  	s64		SCX_EV_SUB_RESCUE;
> +
> +	/*
> +	 * The number of times an insert lacked the caps for its cid and was
> +	 * diverted to the reject DSQ to be re-enqueued so that the BPF
> +	 * scheduler can re-decide.
> +	 */
> +	s64		SCX_EV_SUB_REJECT;
>  };
>  
>  #define SCX_EVENTS_LIST(SCX_EVENT)					\
> @@ -1331,7 +1338,8 @@ struct scx_event_stats {
>  	SCX_EVENT(SCX_EV_SUB_KICK_DENIED);				\
>  	SCX_EVENT(SCX_EV_SUB_REENQ_DENIED);				\
>  	SCX_EVENT(SCX_EV_SUB_CIDPERF_DENIED);				\
> -	SCX_EVENT(SCX_EV_SUB_RESCUE)
> +	SCX_EVENT(SCX_EV_SUB_RESCUE);					\
> +	SCX_EVENT(SCX_EV_SUB_REJECT)
>  
>  struct scx_sched;
>  
> diff --git a/kernel/sched/ext/sub.c b/kernel/sched/ext/sub.c
> index 10567196be96..3429b84f319a 100644
> --- a/kernel/sched/ext/sub.c
> +++ b/kernel/sched/ext/sub.c
> @@ -748,6 +748,8 @@ struct scx_dispatch_q *__scx_resolve_local_dsq(struct scx_sched *sch, struct rq
>  		return &rq->scx.rescue.dsq;
>  	}
>  
> +	__scx_add_event(sch, SCX_EV_SUB_REJECT, 1);
> +
>  	p->scx.reenq_reason_caps = missing;
>  	p->scx.reenq_reason_cid = cid;
>  
> -- 
> 2.43.0
> 
>