On 10/05/20 16:57, Peter Zijlstra wrote:
> Replace a bunch of cpumask_any*() instances with
> cpumask_any*_distribute(), by injecting this little bit of random in
> cpu selection, we reduce the chance two competing balance operations
> working off the same lowest_mask pick the same CPU.
> 
> Signed-off-by: Peter Zijlstra (Intel) <[email protected]>
> ---
>  include/linux/cpumask.h |    6 ++++++
>  kernel/sched/cpupri.c   |    4 ++--
>  kernel/sched/deadline.c |    2 +-
>  kernel/sched/rt.c       |    6 +++---
>  lib/cpumask.c           |   18 ++++++++++++++++++
>  5 files changed, 30 insertions(+), 6 deletions(-)
> 

[...]

> --- a/kernel/sched/rt.c
> +++ b/kernel/sched/rt.c
> @@ -1752,8 +1752,8 @@ static int find_lowest_rq(struct task_st
>                               return this_cpu;
>                       }
>  
> -                     best_cpu = cpumask_first_and(lowest_mask,
> -                                                  sched_domain_span(sd));
> +                     best_cpu = cpumask_any_and_distribute(lowest_mask,
> +                                                           
> sched_domain_span(sd));

I guess I should have done this 6 months ago and just got done with it :)

        [email protected]

>                       if (best_cpu < nr_cpu_ids) {
>                               rcu_read_unlock();
>                               return best_cpu;
> @@ -1770,7 +1770,7 @@ static int find_lowest_rq(struct task_st
>       if (this_cpu != -1)
>               return this_cpu;
>  
> -     cpu = cpumask_any(lowest_mask);
> +     cpu = cpumask_any_distribute(lowest_mask);
>       if (cpu < nr_cpu_ids)
>               return cpu;
>  
> --- a/lib/cpumask.c
> +++ b/lib/cpumask.c
> @@ -267,3 +267,21 @@ int cpumask_any_and_distribute(const str
>       return next;
>  }
>  EXPORT_SYMBOL(cpumask_any_and_distribute);
> +
> +int cpumask_any_distribute(const struct cpumask *srcp)
> +{
> +     int next, prev;
> +
> +     /* NOTE: our first selection will skip 0. */
> +     prev = __this_cpu_read(distribute_cpu_mask_prev);

We had a discussion then that __this_cpu*() variant assumes preemption being
disabled and it's safer to use this_cpu*() variant instead. Still holds true
here?

Thanks

--
Qais Yousef

> +
> +     next = cpumask_next(prev, srcp);
> +     if (next >= nr_cpu_ids)
> +             next = cpumask_first(srcp);
> +
> +     if (next < nr_cpu_ids)
> +             __this_cpu_write(distribute_cpu_mask_prev, next);
> +
> +     return next;
> +}
> +EXPORT_SYMBOL(cpumask_any_distribute);
> 
> 

Reply via email to