Re: [PATCH v3 1/3] sched_ext: Fix pnt_seq calculation

['Tejun Heo' <tj@kernel.org>]

Hello,

Sorry for another delay. I'm finallyback from a long vacation and should be
more responsive from now on.

On Sun, Jul 20, 2025 at 09:36:22AM +0000, liuwenfang wrote:
> Fix pnt_seq calculation for all transitions.

This needs a lot more explanation about the bug it fixes and how.

> +void scx_put_prev_set_next(struct rq *rq, struct task_struct *prev,
> +			   struct task_struct *next)
> +{
> +#ifdef CONFIG_SMP
> +	/*
> +	 * Pairs with the smp_load_acquire() issued by a CPU in
> +	 * kick_cpus_irq_workfn() who is waiting for this CPU to perform a
> +	 * resched.
> +	 */
> +	smp_store_release(&rq->scx.pnt_seq, rq->scx.pnt_seq + 1);
> +#endif
> +}

Let's use a more specific name - e.g. something like scx_bump_sched_seq().
Note that pnt_seq is a bit of misnomer at this point. We probablys should
rename it to sched_seq in a separate patch.

> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index 0fb9bf995..50d757e92 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -8887,6 +8887,9 @@ pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf
>  
>  	__put_prev_set_next_dl_server(rq, prev, p);
>  
> +	if (scx_enabled())
> +		scx_put_prev_set_next(rq, prev, p);
> +
>  	/*
>  	 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
>  	 * likely that a next task is from the same cgroup as the current.
> diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h
> index 47972f34e..bcb7f175c 100644
> @@ -2465,6 +2470,9 @@ static inline void put_prev_set_next_task(struct rq *rq,
>  
>  	__put_prev_set_next_dl_server(rq, prev, next);
>  
> +	if (scx_enabled())
> +		scx_put_prev_set_next(rq, prev, next);
> +
>  	if (next == prev)
>  		return;

I'm not sure these are the best spots to call this function. How about
putting it in the CONFIG_SCHED_CLASS_EXT section in prev_balance()? The goal
of the seq counter is to wait for scheduler path to be entered, so that's
good enough a spot and there already is scx specific section, so it doesn't
add too much noise.

Thanks.

-- 
tejun