Re: [PATCH v4 6/7] sched/deadline: Deferrable dl server

[Daniel Bristot de Oliveira <bristot@kernel.org>]

On 9/7/23 10:07, Peter Zijlstra wrote:
> On Wed, Sep 06, 2023 at 04:58:11PM +0200, Daniel Bristot de Oliveira wrote:
> 
>>> Yeah, it's a wee hack to move it to the zero-laxity point. I was
>>> considering if it makes sense to push that down and make it available
>>> for all DL tasks, but I'm not sure..
>>
>> It might be useful in the future, like when DL dominates all other schedulers, so
>> we can have a way to schedule a deferred work, like kworkers... :-) But it might be
>> too early for that..
> 
> So... that scheme I was pushing where we unconditionally decrement
> fair_server.dl_runtime from update_curr_fair(), that relies on it being
> a proper zero-laxity scheduler, and doesn't work with the proposed defer
> hack.
> 
> That is, it relies on dl_runtime > 0 during throttle, and you explicitly
> set it 0.
> 
> Now, I've not looked at all this code in detail in a minute, but would
> not something like the below work?
> 
> AFAICT the regular dl_task_timer() callback works to make it go, because
> replenish will see positive runtime (or not, when already consumed) and
> DTRT.
> 
> 
> Index: linux-2.6/include/linux/sched.h
> ===================================================================
> --- linux-2.6.orig/include/linux/sched.h
> +++ linux-2.6/include/linux/sched.h
> @@ -657,6 +657,7 @@ struct sched_dl_entity {
>  	unsigned int			dl_non_contending : 1;
>  	unsigned int			dl_overrun	  : 1;
>  	unsigned int			dl_server         : 1;
> +	unsigned int			dl_zerolax        : 1;
>  
>  	/*
>  	 * Bandwidth enforcement timer. Each -deadline task has its
> Index: linux-2.6/kernel/sched/deadline.c
> ===================================================================
> --- linux-2.6.orig/kernel/sched/deadline.c
> +++ linux-2.6/kernel/sched/deadline.c
> @@ -895,6 +895,16 @@ static void replenish_dl_entity(struct s
>  		dl_se->dl_yielded = 0;
>  	if (dl_se->dl_throttled)
>  		dl_se->dl_throttled = 0;
> +
> +	/*
> +	 * If this is a zero-laxity task, and we're before the zero-laxity
> +	 * point, throttle it.
> +	 */
> +	if (dl_se->dl_zerolax &&
> +	    dl_time_before(dl_se->deadline - dl_se->runtime, rq_clock(rq))) {
> +		if (!is_dl_boosted(dl_se) && start_dl_timer(dl_se))
> +			dl_se->dl_throttled = 1;
> +	}
>  }
>  
>  /*
> @@ -1078,7 +1088,12 @@ static int start_dl_timer(struct sched_d
>  	 * that it is actually coming from rq->clock and not from
>  	 * hrtimer's time base reading.
>  	 */
> -	act = ns_to_ktime(dl_next_period(dl_se));
> +	if (dl_se->dl_zerolax && !dl_se->dl_throttled) {
> +		act = ns_to_ktime(dl_se->deadline - dl_se->runtime);
> +	} else {
> +		act = ns_to_ktime(dl_next_period(dl_se));
> +	}
> +
>  	now = hrtimer_cb_get_time(timer);
>  	delta = ktime_to_ns(now) - rq_clock(rq);
>  	act = ktime_add_ns(act, delta);
> @@ -1794,6 +1809,13 @@ enqueue_dl_entity(struct sched_dl_entity
>  		setup_new_dl_entity(dl_se);
>  	}
>  
> +	/*
> +	 * If we are still throttled, eg. we got replenished but are a
> +	 * zero-laxity task and still got to wait, don't enqueue.
> +	 */
> +	if (dl_se->dl_throttled)
> +		return;
> +
>  	__enqueue_dl_entity(dl_se);
>  }

Let me see if I got it:

	- Always start the server, but throttled with full runtime...
	- Unconditionally decrement fair_server.dl_runtime from update_curr_fair()
		(check if it is not decremented twice as it runs)
	- When the dl timer fire, replenish or throttle for the next period?

is that the base for it?

-- Daniel