Re: [RFC PATCH] sched: avoid unnecessary atomic_read when sync_core_before_usermode() is empty

[Tianchen Ding]

We've run schbench and found wakeup latency on some arm64 machines worse 
than others. Perf shows there's a hotspot on 
atomic_read(&mm->membarrier_state);

We're still working for the real reason behind it (maybe cache or sth 
hardware related), and we do see remove this function can help improve 
performance.

Thanks.

On 2022/4/2 11:08, Tianchen Ding wrote:
> On archs except x86, CONFIG_ARCH_HAS_SYNC_CORE_BEFORE_USERMODE is not
> defined. We found membarrier_mm_sync_core_before_usermode() looks like
> this when compiled by gcc10:
>
> 	if (current->mm != mm)
> 		return;
> 	atomic_read(&mm->membarrier_state);
>
> This memory access is unnecessary. Remove it to improve performance.
>
> Signed-off-by: Tianchen Ding <dtcccc@xxxxxxxxxxxxxxxxx>
> ---
>   include/linux/sched/mm.h | 6 ++++++
>   1 file changed, 6 insertions(+)
>
> diff --git a/include/linux/sched/mm.h b/include/linux/sched/mm.h
> index a80356e9dc69..3ded68d9f913 100644
> --- a/include/linux/sched/mm.h
> +++ b/include/linux/sched/mm.h
> @@ -401,6 +401,7 @@ enum {
>   #include <asm/membarrier.h>
>   #endif
>   
> +#ifdef CONFIG_ARCH_HAS_SYNC_CORE_BEFORE_USERMODE
>   static inline void membarrier_mm_sync_core_before_usermode(struct mm_struct *mm)
>   {
>   	if (current->mm != mm)
> @@ -410,6 +411,11 @@ static inline void membarrier_mm_sync_core_before_usermode(struct mm_struct *mm)
>   		return;
>   	sync_core_before_usermode();
>   }
> +#else
> +static inline void membarrier_mm_sync_core_before_usermode(struct mm_struct *mm)
> +{
> +}
> +#endif
>   
>   extern void membarrier_exec_mmap(struct mm_struct *mm);
>