2014-12-09 07:53:55

by Weijie Yang

[permalink] [raw]
Subject: [PATCH 3/3] mm: page_alloc: remove redundant set_freepage_migratetype() calls

The freepage_migratetype is a temporary cached value which represents
the free page's pageblock migratetype. Now we use it in two scenarios:

1. Use it as a cached value in page freeing path. This cached value
is temporary and non-100% update, which help us decide which pcp
freelist and buddy freelist the page should go rather than using
get_pfnblock_migratetype() to save some instructions.
When there is race between page isolation and free path, we need use
additional method to get a accurate value to put the free pages to
the correct freelist and get a precise free pages statistics.

2. Use it in page alloc path to update NR_FREE_CMA_PAGES statistics.

This patch aims at the scenario 1 and removes two redundant
set_freepage_migratetype() calls, which will make sense in the hot path.

Signed-off-by: Weijie Yang <[email protected]>
---
mm/page_alloc.c | 2 --
1 file changed, 2 deletions(-)

diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 616a2c9..99af01a 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -775,7 +775,6 @@ static void __free_pages_ok(struct page *page, unsigned int order)
migratetype = get_pfnblock_migratetype(page, pfn);
local_irq_save(flags);
__count_vm_events(PGFREE, 1 << order);
- set_freepage_migratetype(page, migratetype);
free_one_page(page_zone(page), page, pfn, order, migratetype);
local_irq_restore(flags);
}
@@ -1024,7 +1023,6 @@ int move_freepages(struct zone *zone,
order = page_order(page);
list_move(&page->lru,
&zone->free_area[order].free_list[migratetype]);
- set_freepage_migratetype(page, migratetype);
page += 1 << order;
pages_moved += 1 << order;
}
--
1.7.10.4


2014-12-09 09:49:10

by Vlastimil Babka

[permalink] [raw]
Subject: Re: [PATCH 3/3] mm: page_alloc: remove redundant set_freepage_migratetype() calls

On 12/09/2014 08:51 AM, Weijie Yang wrote:
> The freepage_migratetype is a temporary cached value which represents
> the free page's pageblock migratetype. Now we use it in two scenarios:
>
> 1. Use it as a cached value in page freeing path. This cached value
> is temporary and non-100% update, which help us decide which pcp
> freelist and buddy freelist the page should go rather than using
> get_pfnblock_migratetype() to save some instructions.
> When there is race between page isolation and free path, we need use
> additional method to get a accurate value to put the free pages to
> the correct freelist and get a precise free pages statistics.
>
> 2. Use it in page alloc path to update NR_FREE_CMA_PAGES statistics.

Maybe add that in this case, the value is only valid between being set
by __rmqueue_smallest/__rmqueue_fallback and being consumed by
rmqueue_bulk or buffered_rmqueue for the purposes of statistics.
Oh, except that in rmqueue_bulk, we are placing it on pcplists, so it's
case 1. Tricky.

Anyway, the comments for get/set_freepage_migratetype() say:

/* It's valid only if the page is free path or free_list */

And that's not really true. So should it instead say something like "The
value is only valid when the page is on pcp list, for determining on
which free list the page should go if the pcp list is flushed. It is
also temporarily valid during allocation from free list."

> This patch aims at the scenario 1 and removes two redundant
> set_freepage_migratetype() calls, which will make sense in the hot path.
>
> Signed-off-by: Weijie Yang <[email protected]>
> ---
> mm/page_alloc.c | 2 --
> 1 file changed, 2 deletions(-)
>
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index 616a2c9..99af01a 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -775,7 +775,6 @@ static void __free_pages_ok(struct page *page, unsigned int order)
> migratetype = get_pfnblock_migratetype(page, pfn);
> local_irq_save(flags);
> __count_vm_events(PGFREE, 1 << order);
> - set_freepage_migratetype(page, migratetype);
> free_one_page(page_zone(page), page, pfn, order, migratetype);
> local_irq_restore(flags);
> }
> @@ -1024,7 +1023,6 @@ int move_freepages(struct zone *zone,
> order = page_order(page);
> list_move(&page->lru,
> &zone->free_area[order].free_list[migratetype]);
> - set_freepage_migratetype(page, migratetype);
> page += 1 << order;
> pages_moved += 1 << order;
> }
>

2014-12-10 14:44:15

by Weijie Yang

[permalink] [raw]
Subject: Re: [PATCH 3/3] mm: page_alloc: remove redundant set_freepage_migratetype() calls

On Tue, Dec 9, 2014 at 5:49 PM, Vlastimil Babka <[email protected]> wrote:
> On 12/09/2014 08:51 AM, Weijie Yang wrote:
>>
>> The freepage_migratetype is a temporary cached value which represents
>> the free page's pageblock migratetype. Now we use it in two scenarios:
>>
>> 1. Use it as a cached value in page freeing path. This cached value
>> is temporary and non-100% update, which help us decide which pcp
>> freelist and buddy freelist the page should go rather than using
>> get_pfnblock_migratetype() to save some instructions.
>> When there is race between page isolation and free path, we need use
>> additional method to get a accurate value to put the free pages to
>> the correct freelist and get a precise free pages statistics.
>>
>> 2. Use it in page alloc path to update NR_FREE_CMA_PAGES statistics.
>
>
> Maybe add that in this case, the value is only valid between being set by
> __rmqueue_smallest/__rmqueue_fallback and being consumed by rmqueue_bulk or
> buffered_rmqueue for the purposes of statistics.
> Oh, except that in rmqueue_bulk, we are placing it on pcplists, so it's case
> 1. Tricky.

I will add more description and comments in the next version.
Thanks.

> Anyway, the comments for get/set_freepage_migratetype() say:
>
> /* It's valid only if the page is free path or free_list */
>
> And that's not really true. So should it instead say something like "The
> value is only valid when the page is on pcp list, for determining on which
> free list the page should go if the pcp list is flushed. It is also
> temporarily valid during allocation from free list."

I will update the comments. Thanks

>> This patch aims at the scenario 1 and removes two redundant
>> set_freepage_migratetype() calls, which will make sense in the hot path.
>>
>> Signed-off-by: Weijie Yang <[email protected]>
>> ---
>> mm/page_alloc.c | 2 --
>> 1 file changed, 2 deletions(-)
>>
>> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
>> index 616a2c9..99af01a 100644
>> --- a/mm/page_alloc.c
>> +++ b/mm/page_alloc.c
>> @@ -775,7 +775,6 @@ static void __free_pages_ok(struct page *page,
>> unsigned int order)
>> migratetype = get_pfnblock_migratetype(page, pfn);
>> local_irq_save(flags);
>> __count_vm_events(PGFREE, 1 << order);
>> - set_freepage_migratetype(page, migratetype);
>> free_one_page(page_zone(page), page, pfn, order, migratetype);
>> local_irq_restore(flags);
>> }
>> @@ -1024,7 +1023,6 @@ int move_freepages(struct zone *zone,
>> order = page_order(page);
>> list_move(&page->lru,
>> &zone->free_area[order].free_list[migratetype]);
>> - set_freepage_migratetype(page, migratetype);
>> page += 1 << order;
>> pages_moved += 1 << order;
>> }
>>
>