Re: [PATCH v5 3/4] sched/deadline: Add support for SD_PREFER_SIBLING on find_later_rq()

[Byungchul Park]

On Thu, Aug 03, 2017 at 02:03:34PM +0200, Peter Zijlstra wrote:
> This one I'm not sure on..  at the very least we should exclude all of
> the prefer sibling domain when we do the next domain, and if there are
> multiple prefer sibling levels, we should only pick the first
> fallback_cpu -- there is no point is overriding it with a possible CPU
> further away.

I agree.

> I implemented that below -- although the find_cpu() function is really
> rather horrible.
> 
> But still this isn't quite right, because when we consider this for SMT
> (as was the intent here) we'll happily occupy a full sibling core over
> finding an empty one.
> 
> Now, the problem is that actually doing the right thing quickly ends up
> very expensive, we'd have to scan the entire cache domain at least once,
> so maybe this is good enough.. no idea :/
> 
> 
> ---
> --- a/kernel/sched/deadline.c
> +++ b/kernel/sched/deadline.c
> @@ -1793,12 +1793,35 @@ static struct task_struct *pick_earliest
>  
>  static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask_dl);
>  
> +/*
> + * Find the first cpu in: mask & sd & ~prefer
                                          ^
                                Yes, I missed it.

> + */
> +static int find_cpu(const struct cpumask *mask,
> +		    const struct sched_domain *sd,
> +		    const struct sched_domain *prefer)
> +{
> +	const struct cpumask *sds = sched_domain_span(sd);
> +	const struct cpumask *ps  = prefer ? sched_domain_span(prefer) : NULL;
> +	int cpu = -1;
> +
> +	while ((cpu = cpumask_next(cpu, mask)) < nr_cpu_ids) {
> +		if (!cpumask_test_cpu(cpu, sds))
> +			continue;
> +		if (ps && cpumask_test_cpu(cpu, ps))
> +			continue;
> +		break;
> +	}
> +
> +	return cpu;
> +}
> +
>  static int find_later_rq(struct task_struct *task)
>  {
> -	struct sched_domain *sd;
> +	struct sched_domain *sd, *prefer = NULL;
>  	struct cpumask *later_mask = this_cpu_cpumask_var_ptr(local_cpu_mask_dl);
>  	int this_cpu = smp_processor_id();
>  	int cpu = task_cpu(task);
> +	int fallback_cpu = -1;
>  
>  	/* Make sure the mask is initialized first */
>  	if (unlikely(!later_mask))
> @@ -1850,8 +1873,7 @@ static int find_later_rq(struct task_str
>  				return this_cpu;
>  			}
>  
> -			best_cpu = cpumask_first_and(later_mask,
> -							sched_domain_span(sd));
> +			best_cpu = find_cpu(later_mask, sd, prefer);
>  			/*
>  			 * Last chance: if a cpu being in both later_mask
>  			 * and current sd span is valid, that becomes our
> @@ -1859,6 +1881,17 @@ static int find_later_rq(struct task_str
>  			 * already under consideration through later_mask.
>  			 */
>  			if (best_cpu < nr_cpu_ids) {
> +				/*
> +				 * If current domain is SD_PREFER_SIBLING
> +				 * flaged, we have to get more chances to
> +				 * check other siblings.
> +				 */
> +				if (sd->flags & SD_PREFER_SIBLING) {
> +					prefer = sd;
> +					if (fallback_cpu == -1)
                                        ^
                                     I like the 'if' statement.
                                     I should have done this.

> +						fallback_cpu = best_cpu;
> +					continue;
> +				}
>  				rcu_read_unlock();
>  				return best_cpu;
>  			}

Thank you.