[PATCH v3] mm/page_alloc: speeding up the iteration of max_order

[PATCH v3] mm/page_alloc: speeding up the iteration of max_order

[Muchun Song]

When we free a page whose order is very close to MAX_ORDER and greater
than pageblock_order, it wastes some CPU cycles to increase max_order
to MAX_ORDER one by one and check the pageblock migratetype of that page
repeatedly especially when MAX_ORDER is much larger than pageblock_order.

We also should not be checking migratetype of buddy when "order ==
MAX_ORDER - 1" as the buddy pfn may be invalid, so adjust the condition.
With the new check, we don't need the max_order check anymore, so we
replace it.

Also adjust max_order initialization so that it's lower by one than
previously, which makes the code hopefully more clear.

Fixes: d9dddbf55667 ("mm/page_alloc: prevent merging between isolated and other pageblocks")
Signed-off-by: Muchun Song <songmuchun@bytedance.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
---
Changes in v3:
 - Update commit log.

Changes in v2:
 - Rework the code suggested by Vlastimil. Thanks.

 mm/page_alloc.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index f91df593bf71..56e603eea1dd 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -1002,7 +1002,7 @@ static inline void __free_one_page(struct page *page,
 	struct page *buddy;
 	bool to_tail;
 
-	max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
+	max_order = min_t(unsigned int, MAX_ORDER - 1, pageblock_order);
 
 	VM_BUG_ON(!zone_is_initialized(zone));
 	VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
@@ -1015,7 +1015,7 @@ static inline void __free_one_page(struct page *page,
 	VM_BUG_ON_PAGE(bad_range(zone, page), page);
 
 continue_merging:
-	while (order < max_order - 1) {
+	while (order < max_order) {
 		if (compaction_capture(capc, page, order, migratetype)) {
 			__mod_zone_freepage_state(zone, -(1 << order),
 								migratetype);
@@ -1041,7 +1041,7 @@ static inline void __free_one_page(struct page *page,
 		pfn = combined_pfn;
 		order++;
 	}
-	if (max_order < MAX_ORDER) {
+	if (order < MAX_ORDER - 1) {
 		/* If we are here, it means order is >= pageblock_order.
 		 * We want to prevent merge between freepages on isolate
 		 * pageblock and normal pageblock. Without this, pageblock
@@ -1062,7 +1062,7 @@ static inline void __free_one_page(struct page *page,
 						is_migrate_isolate(buddy_mt)))
 				goto done_merging;
 		}
-		max_order++;
+		max_order = order + 1;
 		goto continue_merging;
 	}
 
-- 
2.11.0

Re: [PATCH v3] mm/page_alloc: speeding up the iteration of max_order

[David Hildenbrand]

On 04.12.20 16:51, Muchun Song wrote:
> When we free a page whose order is very close to MAX_ORDER and greater
> than pageblock_order, it wastes some CPU cycles to increase max_order
> to MAX_ORDER one by one and check the pageblock migratetype of that page
> repeatedly especially when MAX_ORDER is much larger than pageblock_order.
> 
> We also should not be checking migratetype of buddy when "order ==
> MAX_ORDER - 1" as the buddy pfn may be invalid, so adjust the condition.
> With the new check, we don't need the max_order check anymore, so we
> replace it.
> 
> Also adjust max_order initialization so that it's lower by one than
> previously, which makes the code hopefully more clear.
> 
> Fixes: d9dddbf55667 ("mm/page_alloc: prevent merging between isolated and other pageblocks")
> Signed-off-by: Muchun Song <songmuchun@bytedance.com>
> Acked-by: Vlastimil Babka <vbabka@suse.cz>
> ---
> Changes in v3:
>  - Update commit log.
> 
> Changes in v2:
>  - Rework the code suggested by Vlastimil. Thanks.
> 
>  mm/page_alloc.c | 8 ++++----
>  1 file changed, 4 insertions(+), 4 deletions(-)
> 
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index f91df593bf71..56e603eea1dd 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -1002,7 +1002,7 @@ static inline void __free_one_page(struct page *page,
>  	struct page *buddy;
>  	bool to_tail;
>  
> -	max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
> +	max_order = min_t(unsigned int, MAX_ORDER - 1, pageblock_order);
>  
>  	VM_BUG_ON(!zone_is_initialized(zone));
>  	VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
> @@ -1015,7 +1015,7 @@ static inline void __free_one_page(struct page *page,
>  	VM_BUG_ON_PAGE(bad_range(zone, page), page);
>  
>  continue_merging:
> -	while (order < max_order - 1) {
> +	while (order < max_order) {
>  		if (compaction_capture(capc, page, order, migratetype)) {
>  			__mod_zone_freepage_state(zone, -(1 << order),
>  								migratetype);
> @@ -1041,7 +1041,7 @@ static inline void __free_one_page(struct page *page,
>  		pfn = combined_pfn;
>  		order++;
>  	}
> -	if (max_order < MAX_ORDER) {
> +	if (order < MAX_ORDER - 1) {
>  		/* If we are here, it means order is >= pageblock_order.
>  		 * We want to prevent merge between freepages on isolate
>  		 * pageblock and normal pageblock. Without this, pageblock
> @@ -1062,7 +1062,7 @@ static inline void __free_one_page(struct page *page,
>  						is_migrate_isolate(buddy_mt)))
>  				goto done_merging;
>  		}
> -		max_order++;
> +		max_order = order + 1;
>  		goto continue_merging;
>  	}
>  
> 


LGTM

Reviewed-by: David Hildenbrand <david@redhat.com>

-- 
Thanks,

David / dhildenb

Re: [PATCH v3] mm/page_alloc: speeding up the iteration of max_order

[Oscar Salvador]

On Fri, Dec 04, 2020 at 11:51:09PM +0800, Muchun Song wrote:
> When we free a page whose order is very close to MAX_ORDER and greater
> than pageblock_order, it wastes some CPU cycles to increase max_order
> to MAX_ORDER one by one and check the pageblock migratetype of that page
> repeatedly especially when MAX_ORDER is much larger than pageblock_order.
> 
> We also should not be checking migratetype of buddy when "order ==
> MAX_ORDER - 1" as the buddy pfn may be invalid, so adjust the condition.
> With the new check, we don't need the max_order check anymore, so we
> replace it.
> 
> Also adjust max_order initialization so that it's lower by one than
> previously, which makes the code hopefully more clear.
> 
> Fixes: d9dddbf55667 ("mm/page_alloc: prevent merging between isolated and other pageblocks")
> Signed-off-by: Muchun Song <songmuchun@bytedance.com>
> Acked-by: Vlastimil Babka <vbabka@suse.cz>

Reviewed-by: Oscar Salvador <osalvador@suse.de>

-- 
Oscar Salvador
SUSE L3