[lkml]   [2012]   [Jan]   [31]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [RFCv1 3/6] PASR: mm: Integrate PASR in Buddy allocator
Hello Mel,
On 01/30/2012 05:52 PM, Maxime Coquelin wrote:
> On 01/30/2012 04:22 PM, Mel Gorman wrote:
>> You may be able to use the existing arch_alloc_page() hook and
>> call PASR on architectures that support it if and only if PASR is
>> present and enabled by the administrator but even this is likely to be
>> unpopular as it'll have a measurable performance impact on platforms
>> with PASR (not to mention the PASR lock will be even heavier as it'll
>> now be also used for per-cpu page allocations). To get the hook you
>> want, you'd need to show significant benefit before they were happy with
>> the hook.
> Your proposal sounds good.
> AFAIK, per-cpu allocation maximum size is 32KB. Please correct me if
> I'm wrong.
> Since pasr_kget/kput() calls the PASR framework only on MAX_ORDER
> allocations, we wouldn't add any locking risks nor contention compared
> to current patch.
> I will update the patch set using arch_alloc/free_page().
I just had a deeper look at when arch_alloc_page() is called. I think it
does not fit with PASR framework needs.
pasr_kget() calls pasr_get() only for max order pages (same for
pasr_kput()) to avoid overhead.

In current patch set, pasr_kget() is called when pages are removed from
the free lists, and pasr_kput() when pages are inserted in the free lists.
So, pasr_get() is called in case of :
- allocation of a max order page
- split of a max order page into lower order pages to fulfill
allocation of pages smaller than max order
And pasr_put() is called in case of:
- release of a max order page
- coalescence of two "max order -1" pages when smaller pages are

If we call the PASR framework in arch_alloc_page(), we have two
1) using pasr_kget(): the PASR framework will only be notified of
max order allocations, so the coalesce/split of free pages case will not
be taken into account.
2) using pasr_get(): the PASR framework will be called for every
orders of page allocation/release. The induced overhead is not acceptable.

To avoid calling pasr_kget/kput() directly in page_alloc.c, do you think
adding some arch specific hooks when a page is inserted or removed from
the free lists could be acceptable?
Something like arch_insert_freepage(struct page *page, int order) and
arch_remove_freepage(struct page *page, int order).


 \ /
  Last update: 2012-01-31 13:19    [W:0.052 / U:23.736 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site