Re: [PATCHSET v14 sched_ext/for-7.4] sched: Make proxy execution compatible with sched_ext

[Andrea Righi <arighi@nvidia.com>]

Hi Peter,

On Thu, Sep 24, 2026 at 09:52:37AM +0200, Peter Zijlstra wrote:
> On Tue, Sep 22, 2026 at 06:51:39PM +0200, Andrea Righi wrote:
> 
> > Andrea Righi (16):
> >       sched/core: Drop mutex locks before proxy rescheduling
> >       sched/core: Dequeue waking proxy donors before reset
> >       sched/core: Mark wakeups completed through ttwu_runnable()
> >       sched: Add helper to block retained proxy donors
> >       sched: Add sched_ext hooks for proxy execution
> 
> Right, so these add:
> 
> WF_TTWU_RQ:
> 
>   Used like ENQUEUE_DELAYED; could be fixed by generalizing that to cover all
>   of p->is_blocked.
> 
> scx_allow_proxy_exec():
> 
>   Hook to kill proxy exec for scx
> 
> scx_proxy_reenqueue_retry():
> 
>   Like put_prev_task(), but for current. Ensures current gets put back on a DSQ
>   once its done running.
> 
> scx_proxy_donor_start():
> 
>   Delayed set_next_task(), confirms donor will be used.
> 
> sched_proxy_block_task():
> 
>   Almost like switching_to_scx(), except it needs to change ctx->queued in case
>   of p->is_blocked. Hence a new callback ran before sched_change_begin().

Yes, that matches the intent, with one small clarification:
scx_proxy_reenqueue_retry() doesn't directly put current back on a DSQ, a task
that couldn't be reenqueued remains on the reject DSQ and the hook schedules a
deferred retry after proxy resolution.

> 
> 
> 
> Now, I have:
> 
>   https://patch.msgid.link/20260917-sched-fair-hrtick-restart-v4-1-4dd1414da81a@gentwo.org,
> 
> pending, would something like the below on top of both this work?
> 
> (although I'm not convinced SC_CONFIRM is actually making it better)

Yes, this works.

I applied Shubhang's v4 hrtick patch as a preliminary commit, then added the
SNT_CONFIRM callback and reworked the proxy-exec series to use it. sched_ext now
confirms the donor in set_next_task_scx() and scx_proxy_donor_start() is gone.

I tested the updated series and it passed all my scx proxy exec tests. The
branch is available here:

 git://git.kernel.org/pub/scm/linux/kernel/git/arighi/linux.git scx-proxy-exec-next

It's not an obvious simplification, but if we want to go this way, we can
express both the provisional pick and donor confirmation through the sched class
callback.

Thanks,
-Andrea

> 
> --- a/kernel/sched/core.c
> +++ b/kernel/sched/core.c
> @@ -1876,7 +1876,7 @@ static inline void uclamp_rq_inc(struct
>  	if (!uclamp_is_used())
>  		return;
>  
> -	if (unlikely(!p->sched_class->uclamp_enabled))
> +	if (unlikely(!(p->sched_class->flags & SC_UCLAMP)))
>  		return;
>  
>  	/* Only inc the delayed task which being woken up. */
> @@ -1904,7 +1904,7 @@ static inline void uclamp_rq_dec(struct
>  	if (!uclamp_is_used())
>  		return;
>  
> -	if (unlikely(!p->sched_class->uclamp_enabled))
> +	if (unlikely(!(p->sched_class->flags & SC_UCLAMP)))
>  		return;
>  
>  	if (p->se.sched_delayed)
> @@ -7268,9 +7268,10 @@ static void __sched notrace __schedule(i
>  	rq->next_class = next->sched_class;
>  	if (sched_proxy_exec()) {
>  		struct task_struct *prev_donor = rq->donor;
> +		struct task_struct *donor = next;
>  
> -		rq_set_donor(rq, next);
> -		next->blocked_donor = NULL;
> +		rq_set_donor(rq, donor);
> +		donor->blocked_donor = NULL;
>  		if (unlikely(next->is_blocked)) {
>  			next = find_proxy_task(rq, next, &rf);
>  			if (!next) {
> @@ -7283,8 +7284,7 @@ static void __sched notrace __schedule(i
>  				goto keep_resched;
>  			}
>  		}
> -		if (rq->donor == prev_donor && prev != next) {
> -			struct task_struct *donor = rq->donor;
> +		if (donor == prev_donor && prev != next) {
>  			/*
>  			 * When transitioning like:
>  			 *
> @@ -7298,9 +7298,10 @@ static void __sched notrace __schedule(i
>  			 * on_cpu.
>  			 */
>  			donor->sched_class->put_prev_task(rq, donor, donor);
> -			donor->sched_class->set_next_task(rq, donor, true);
> +			donor->sched_class->set_next_task(rq, donor, SNT_PICK);
>  		}
> -		scx_proxy_donor_start(rq);
> +		if (donor->sched_class->flags & SC_CONFIRM)
> +			donor->sched_class->set_next_task(rq, donor, SNT_CONFIRM);
>  		scx_proxy_reenqueue_retry(rq, next);
>  	} else {
>  		rq_set_donor(rq, next);
> --- a/kernel/sched/ext/ext.c
> +++ b/kernel/sched/ext/ext.c
> @@ -5069,8 +5069,12 @@ DEFINE_SCHED_CLASS(ext) = {
>  
>  	.update_curr		= update_curr_scx,
>  
> +	.flags			= 0
>  #ifdef CONFIG_UCLAMP_TASK
> -	.uclamp_enabled		= 1,
> +				| SC_UCLAMP
> +#endif
> +#ifdef CONFIG_SCHED_PROXY_EXEC
> +				| SC_CONFIRM
>  #endif
>  };
>  
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -15887,7 +15887,7 @@ DEFINE_SCHED_CLASS(fair) = {
>  #endif
>  
>  #ifdef CONFIG_UCLAMP_TASK
> -	.uclamp_enabled		= 1,
> +	.flags			= SC_UCLAMP,
>  #endif
>  };
>  
> --- a/kernel/sched/rt.c
> +++ b/kernel/sched/rt.c
> @@ -2636,7 +2636,7 @@ DEFINE_SCHED_CLASS(rt) = {
>  #endif
>  
>  #ifdef CONFIG_UCLAMP_TASK
> -	.uclamp_enabled		= 1,
> +	.flags			= SC_UCLAMP,
>  #endif
>  };
>  
> --- a/kernel/sched/sched.h
> +++ b/kernel/sched/sched.h
> @@ -2641,13 +2641,17 @@ enum snt_e {
>  	SNT_NORMAL,	/* set_next_task() */
>  	SNT_PICK,	/* put_prev_set_next_task(): prev != next */
>  	SNT_REPICK,	/* put_prev_set_next_task(): prev == next */
> +	SNT_CONFIRM,	/* confirm pick */
> +};
> +
> +enum {
> +	SC_UCLAMP	= 1<<0,
> +	SC_CONFIRM	= 1<<1,
>  };
>  
>  struct sched_class {
>  
> -#ifdef CONFIG_UCLAMP_TASK
> -	int uclamp_enabled;
> -#endif
> +	unsigned int flags;
>  
>  	/*
>  	 * move_queued_task/activate_task/enqueue_task: rq->lock