Re: [RFC v5 5/9] sched/deadline: do not reclaim the whole CPU bandwidth

[Peter Zijlstra <peterz@infradead.org>]

On Fri, Mar 24, 2017 at 04:52:58AM +0100, luca abeni wrote:

> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> index 20c62e7..efa88eb 100644
> --- a/kernel/sched/core.c
> +++ b/kernel/sched/core.c
> @@ -6716,6 +6716,12 @@ static void sched_dl_do_global(void)
>  		raw_spin_unlock_irqrestore(&dl_b->lock, flags);
>  
>  		rcu_read_unlock_sched();
> +		if (dl_b->bw == -1)
> +			cpu_rq(cpu)->dl.deadline_bw_inv = 1 << 8;
> +		else
> +			cpu_rq(cpu)->dl.deadline_bw_inv =
> +				to_ratio(global_rt_runtime(),
> +					 global_rt_period()) >> 12;

Coding style requires braces here (on both legs of the condition)..

Also, I find deadline_bw_inv an awkward name; would something like
bw_ratio or so be more accurate?

> +	if (global_rt_runtime() == RUNTIME_INF)
> +		dl_rq->deadline_bw_inv = 1 << 8;
> +	else
> +		dl_rq->deadline_bw_inv =
> +			to_ratio(global_rt_runtime(), global_rt_period()) >> 12;

That's almost the same code; do we want a helper function?

>  
>  u64 grub_reclaim(u64 delta, struct rq *rq)
>  {
> +	return (delta * rq->dl.running_bw * rq->dl.deadline_bw_inv) >> 20 >> 8;
>  }

At which point we might want a note about how this doesn't overflow I
suppose.

Also:

	delta *= rq->dl.running_bw;
	delta *= rq->dl.bw_ratio;
	delta >>= 20 + 8;

	return delta;

Might be more readable ?

Alternatively:

	delta = (delta * rq->dl.running_bw) >> 8;
	delta = (delta * rq->dl.bw_ratio) >> 20;

	return delta;

But I doubt we care about those extra 8 bit of space; delta should not
be over 36 bits (~64 seconds) anyway I suppose.