Re: [PATCH v9 1/2] sched/deadline: Add support for SD_PREFER_SIBLING on find_later_rq()

[Juri Lelli]

Hi,

On 22/08/17 17:30, Byungchul Park wrote:
> It would be better to try to check other siblings first if
> SD_PREFER_SIBLING is flaged when pushing tasks - migration.
> 
> Signed-off-by: Byungchul Park <byungchul.park@xxxxxxx>
> Suggested-by: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
       ^
This has to come before your SoB.

> ---
>  kernel/sched/deadline.c | 80 ++++++++++++++++++++++++++++++++++++++++++++-----
>  1 file changed, 73 insertions(+), 7 deletions(-)
> 
> diff --git a/kernel/sched/deadline.c b/kernel/sched/deadline.c
> index 0223694..b6b3855 100644
> --- a/kernel/sched/deadline.c
> +++ b/kernel/sched/deadline.c
> @@ -1319,12 +1319,35 @@ static struct task_struct *pick_earliest_pushable_dl_task(struct rq *rq, int cpu
>  
>  static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask_dl);
>  
> +/*
> + * Find the first cpu in: mask & sd & ~prefer
> + */
> +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;
> +
> +	for_each_cpu(cpu, mask) {
> +		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))
> @@ -1376,15 +1399,35 @@ static int find_later_rq(struct task_struct *task)
>  				return this_cpu;
>  			}
>  
> -			best_cpu = cpumask_first_and(later_mask,
> -							sched_domain_span(sd));
>  			/*
> -			 * Last chance: if a cpu being in both later_mask
> -			 * and current sd span is valid, that becomes our
> -			 * choice. Of course, the latest possible cpu is
> -			 * already under consideration through later_mask.
> +			 * If a cpu being in later_mask & current sd &
> +			 * ~prefer sd is valid, that becomes our choice.
> +			 * Of course, the latest possible cpu is already
> +			 * under consideration through later_mask.
>  			 */
> +			best_cpu = find_cpu(later_mask, sd, prefer);
> +
>  			if (best_cpu < nr_cpu_ids) {
> +				/*
> +				 * If current domain is SD_PREFER_SIBLING
> +				 * flaged, we have to try to check other
> +				 * siblings first.
> +				 */
> +				if (sd->flags & SD_PREFER_SIBLING) {
> +					prefer = sd;
> +
> +					/*
> +					 * fallback_cpu should be one
> +					 * in the closest domain among
> +					 * SD_PREFER_SIBLING domains,
> +					 * in case that more than one
> +					 * SD_PREFER_SIBLING domains
> +					 * exist in the hierachy.
> +					 */
> +					if (fallback_cpu == -1)
> +						fallback_cpu = best_cpu;
> +					continue;
> +				}
>  				rcu_read_unlock();
>  				return best_cpu;
>  			}
> @@ -1393,6 +1436,29 @@ static int find_later_rq(struct task_struct *task)
>  	rcu_read_unlock();
>  
>  	/*
> +	 * If fallback_cpu is valid, all our guesses failed *except* for
> +	 * SD_PREFER_SIBLING domain. Now, we can return the fallback cpu.
> +	 *
> +	 * XXX: Consider the following example, 4 cores SMT2 system:
> +	 *
> +	 *    LLC [0       -        7]
> +	 *    SMT [0 1][2 3][4 5][6 7]
> +	 *         o x  o x  x x  x x
> +	 *
> +	 *    where 'o': occupied and 'x': empty.
> +	 *
> +	 * A wakeup on cpu0 will exclude cpu1 and choose cpu3, since
> +	 * cpu1 is in a SD_PREFER_SIBLING sd and cpu3 is not. However,
> +	 * in this case, we have to choose cpu4 for better work, instead

... in this case cpu4 would have been a better choice, since cpu3 is a
(SMT) thread of an already loaded core.

> +	 * of cpu3 that is fully loaded.
> +	 *
> +	 * We have to do the best if possible, but choose the second
> +	 * best here since that is too expensive to adopt.
> +	 */

I'd also modify this last paragraph with something like:

"Doing it 'right' is difficult and expensive. The current solution is an
acceptable approximation."

Apart from these minor points, patch looks ok to me.

Acked-by: Juri Lelli <juri.lelli@xxxxxxx>

Best,

- Juri