Re: [PATCH 1/1] mm/page_alloc: merge two get_pfnblock_migratetype() into one

[Matthew Wilcox]

On Wed, Nov 23, 2022 at 09:40:43PM +0800, wonder_rock@xxxxxxx wrote:
> From: Deyan Wang <wonder_rock@xxxxxxx>
> 
> It is safe to get_pfnblock_migratetype() in critical section,
> so just call get_pfnblock_migratetype() once.

No, you don't understand.  zone->lock is HOT and we want to avoid
calling get_pfnblock_migratetype() under the lock, if we can avoid it.
Which we usually can.

> Signed-off-by: Deyan Wang <wonder_rock@xxxxxxx>
> ---
>  mm/page_alloc.c | 7 +------
>  1 file changed, 1 insertion(+), 6 deletions(-)
> 
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index 6652df1f5cc8..c7ba9d65d1c0 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -1704,13 +1704,8 @@ static void __free_pages_ok(struct page *page, unsigned int order,
>  	if (!free_pages_prepare(page, order, true, fpi_flags))
>  		return;
>  
> -	migratetype = get_pfnblock_migratetype(page, pfn);
> -
>  	spin_lock_irqsave(&zone->lock, flags);
> -	if (unlikely(has_isolate_pageblock(zone) ||
> -		is_migrate_isolate(migratetype))) {
> -		migratetype = get_pfnblock_migratetype(page, pfn);
> -	}
> +	migratetype = get_pfnblock_migratetype(page, pfn);
>  	__free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
>  	spin_unlock_irqrestore(&zone->lock, flags);
>  
> -- 
> 2.25.1
> 
>