lkml.org 
[lkml]   [2012]   [Aug]   [16]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [RFC 1/2] cma: remove __reclaim_pages
On Tue, Aug 14, 2012 at 05:57:06PM +0900, Minchan Kim wrote:
> Now cma reclaims too many pages by __reclaim_pages which says
> following as
>
> * Reclaim enough pages to make sure that contiguous allocation
> * will not starve the system.
>
> Starve? What does it starve the system? The function which allocate
> free page for migration target would wake up kswapd and do direct reclaim
> if needed during migration so system doesn't starve.
>

I thought this patch was overkill at the time it was introduced but
didn't have a concrete reason to reject it when I commented on it
https://lkml.org/lkml/2012/1/30/136 . Marek did want this and followed
up with "contiguous allocations should have higher priority than others"
which I took to mean that he was also ok with excessive reclaim.

> Let remove __reclaim_pages and related function and fields.
>

That should be one patch and I don't object to it being removed as such
but it's Marek's call.

> I modified split_free_page slightly because I removed __reclaim_pages,
> isolate_freepages_range can fail by split_free_page's watermark check.
> It's very critical in CMA because it ends up failing alloc_contig_range.
>

This is a big change and should have been in a patch on its
own. split_free_page checks watermarks because if the watermarks are
not obeyed a zone can become fully allocated. This can cause a system to
livelock under certain circumstances if a page cannot be allocated and a
free page is required before other pages can be freed.

> I think we don't need the check in case of CMA because CMA allocates
> free pages by alloc_pages, not isolate_freepages_block in migrate_pages
> so watermark is already checked in alloc_pages.

It uses alloc_pages when migrating pages out of the CMA area but note
that it uses isolate_freepages_block when allocating the CMA buffer when
alloc_contig_range calls isolate_freepages_range

isolate_freepages_range(unsigned long start_pfn, unsigned long end_pfn)
{
for (pfn = start_pfn; pfn < end_pfn; pfn += isolated) {
isolated = isolate_freepages_block(pfn, block_end_pfn,
&freelist, true);
}
map_pages(&freelist);
}

so the actual CMA allocation itself is not using alloc_pages. By removing
the watermark check you allow the CMA to breach watermarks and puts the
system at risk of livelock.

I'm not keen on the split_free_page() change at all.

--
Mel Gorman
SUSE Labs


\
 
 \ /
  Last update: 2012-08-17 06:21    [W:0.122 / U:1.276 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site