Re: [PATCH v2] sched/fair: static cpumasks for load balance

[huangbing@kylinos]

On 6/1/22 16:45, Dietmar Eggemann wrote:
> On 31/05/2022 05:12, Bing Huang wrote:
> > The both cpu mask load_balance_mask and select_idle_mask just only used
> > in fair.c, but allocation in core.c in CONFIG_CPUMASK_OFFSTACK=y case,
> > and global via declare per cpu variations. More or less, it looks wired.
> >
> > Signed-off-by: Bing Huang <huangbing@xxxxxxxxxx>
> > ---
> >
> >   v2: move load_balance_mask and select_idle_mask allocation from
> > sched_init() to init_sched_fair_class()
> This would align CFS with RT (local_cpu_mask) and DL
> (local_cpu_mask_dl).
>
> [...]
>
> > @@ -11841,6 +11841,16 @@ void show_numa_stats(struct task_struct *p, struct seq_file *m)
> >   __init void init_sched_fair_class(void)
> >   {
> >   #ifdef CONFIG_SMP
>      `int i` missing for DEBUG_PER_CPU_MAPS/CONFIG_CPUMASK_OFFSTACK case.
sorry,my fault.
> > +
> > +#ifdef CONFIG_CPUMASK_OFFSTACK
> > +	for_each_possible_cpu(i) {
> > +		per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
> > +			cpumask_size(), GFP_KERNEL, cpu_to_node(i));
> > +		per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
> > +			cpumask_size(), GFP_KERNEL, cpu_to_node(i));
> > +	}
> > +#endif
> > +
> What about:
>
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index 98319b654788..9ef5133c72d6 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -11811,15 +11811,14 @@ void show_numa_stats(struct task_struct *p, struct seq_file *m)
>   __init void init_sched_fair_class(void)
>   {
>   #ifdef CONFIG_SMP
> +       int i;
>   
> -#ifdef CONFIG_CPUMASK_OFFSTACK
>          for_each_possible_cpu(i) {
> -               per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
> -                       cpumask_size(), GFP_KERNEL, cpu_to_node(i));
> -               per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
> -                       cpumask_size(), GFP_KERNEL, cpu_to_node(i));
> +               zalloc_cpumask_var_node(&per_cpu(load_balance_mask, i),
> +                                       GFP_KERNEL, cpu_to_node(i));
> +               zalloc_cpumask_var_node(&per_cpu(select_idle_mask, i),
> +                                       GFP_KERNEL, cpu_to_node(i));
>          }
> -#endif
>
> to get rid of the #ifdef ? We do the same for RT (local_cpu_mask) and DL
> (local_cpu_mask_dl).
good suggestion, thanks very much. happy to do that.
> [...]