Re: [PATCH v1 1/3] SMP: kill redundant call_function_data->cpumask_ipi field

[Wang YanQing <udknight@gmail.com>]

On Sun, Sep 08, 2013 at 11:22:23PM +0800, Jiang Liu wrote:
> From: Jiang Liu <jiang.liu@huawei.com>
> 
> Commit f44310b98ddb7 "smp: Fix SMP function call empty cpu mask race"
> introduced field call_function_data->cpumask_ipi to resolve a race
> condition in smp_call_function_many().
> 
> Later commit 9a46ad6d6df3 "smp: make smp_call_function_many() use logic
> similar to smp_call_function_single()" fixed the same issue in another
> way when optimizing smp_call_function_many(), which then obsoletes
> changes introduced by commit f44310b98ddb7. So revert it.
> 

Yes, you are right, after commit 9a46ad6d6df3, we can revert f44310b98ddb7.
Maybe use "Revert smp: Fix SMP function call empty cpu mask race" is a better
subject.

> We may also keep call_function_data->cpumask_ipi field and use it to
> optimize smp_call_function_many() as below:
> diff --git a/kernel/smp.c b/kernel/smp.c
> index fe9f773..dd852fb 100644
> --- a/kernel/smp.c
> +++ b/kernel/smp.c
> @@ -428,6 +428,8 @@ void smp_call_function_many(const struct cpumask *mask,
>                 csd->info = info;
> 
>                 raw_spin_lock_irqsave(&dst->lock, flags);
> +               if (list_empty(&dst->list))
> +                       cpumask_clear_cpu(cpu, cfd->cpumask_ipi);
>                 list_add_tail(&csd->list, &dst->list);
>                 raw_spin_unlock_irqrestore(&dst->lock, flags);
>         }
> 

Your optimization don't need to keep cpumask_ipi, just clear cfd->cpumask,
and test whether cfd->cpumask is empty after "for_each_cpu(cpu, cfd->cpumask)".


Acked-by: Wang YanQing <udknight@gmail.com>

Thanks.