2007-12-13 05:20:41

by Shaohua Li

Subject: [RFC] move free page back to MIGRATE_MOVABLE?

pages are MIGRATE_MOVABLE initially, and fallback to other types. If the
pages are freed, I wonder why not move them back to MIGRATE_MOVABLE.
This, for example, can help memory remove.

Thanks,
Shaohua

Index: linux/mm/page_alloc.c
===================================================================
--- linux.orig/mm/page_alloc.c 2007-12-13 11:44:36.000000000 +0800
+++ linux/mm/page_alloc.c 2007-12-13 12:04:59.000000000 +0800
@@ -386,6 +386,9 @@ static inline int page_is_buddy(struct p
return 0;
}

+int move_freepages(struct zone *zone,
+ struct page *start_page, struct page *end_page,
+ int migratetype);
/*
* Freeing function for a buddy system allocator.
*
@@ -446,6 +449,17 @@ static inline void __free_one_page(struc
list_add(&page->lru,
&zone->free_area[order].free_list[migratetype]);
zone->free_area[order].nr_free++;
+
+ if (order >= pageblock_order && migratetype != MIGRATE_MOVABLE) {
+ struct page *tmp = page;
+
+ move_freepages(zone, page, page + (1 << order) - 1, MIGRATE_MOVABLE);
+
+ while (tmp < page + (1 << order)) {
+ set_pageblock_migratetype(tmp, MIGRATE_MOVABLE);
+ tmp += pageblock_nr_pages;
+ }
+ }
}

static inline int free_pages_check(struct page *page)

2007-12-17 10:45:06

by Mel Gorman

Subject: Re: [RFC] move free page back to MIGRATE_MOVABLE?

On (13/12/07 13:19), Shaohua Li didst pronounce:
> pages are MIGRATE_MOVABLE initially, and fallback to other types. If the
> pages are freed, I wonder why not move them back to MIGRATE_MOVABLE.

Because there is no need, if a MIGRATE_MOVABLE allocation needs more pages,
they will be acquired during fallback which always steals the largest
block. Lets say that a MIGRATE_UNMOVABLE block is freed and it is larger
than pageblock_order. The next MIGRATE_MOVABLE allocation will find it during
fallback if necessary.

> This, for example, can help memory remove.
>

I don't see how. When hot-removing, MIGRATE_MOVABLE blocks are considered
possible to remove but they are also considered movable if the pages are free.

> Thanks,
> Shaohua
>
> Index: linux/mm/page_alloc.c
> ===================================================================
> --- linux.orig/mm/page_alloc.c 2007-12-13 11:44:36.000000000 +0800
> +++ linux/mm/page_alloc.c 2007-12-13 12:04:59.000000000 +0800
> @@ -386,6 +386,9 @@ static inline int page_is_buddy(struct p
> return 0;
> }
>
> +int move_freepages(struct zone *zone,
> + struct page *start_page, struct page *end_page,
> + int migratetype);
> /*
> * Freeing function for a buddy system allocator.
> *
> @@ -446,6 +449,17 @@ static inline void __free_one_page(struc
> list_add(&page->lru,
> &zone->free_area[order].free_list[migratetype]);
> zone->free_area[order].nr_free++;
> +
> + if (order >= pageblock_order && migratetype != MIGRATE_MOVABLE) {
> + struct page *tmp = page;
> +
> + move_freepages(zone, page, page + (1 << order) - 1, MIGRATE_MOVABLE);
> +

This is unnecessary and affects a commonly used path. If a MOVABLE allocation
needs more pages later, it will find the large free block during fallback.

> + while (tmp < page + (1 << order)) {
> + set_pageblock_migratetype(tmp, MIGRATE_MOVABLE);
> + tmp += pageblock_nr_pages;
> + }
> + }
> }
>
> static inline int free_pages_check(struct page *page)
>

--
Mel Gorman
Part-time Phd Student Linux Technology Center
University of Limerick IBM Dublin Software Lab