Re: [PATCH 1/2] Align the node_mem_map endpoints to a MAX_ORDER boundary

[Andy Whitcroft <apw@shadowen.org>]

Andrew Morton wrote:
> Mel Gorman <mel@csn.ul.ie> wrote:
> 
>>Andy added code to buddy allocator which does not require the zone's
>>endpoints to be aligned to MAX_ORDER. An issue is that the buddy
>>allocator requires the node_mem_map's endpoints to be MAX_ORDER aligned.
>>Otherwise __page_find_buddy could compute a buddy not in node_mem_map for
>>partial MAX_ORDER regions at zone's endpoints. page_is_buddy will detect
>>that these pages at endpoints are not PG_buddy (they were zeroed out by
>>bootmem allocator and not part of zone). Of course the negative here is
>>we could waste a little memory but the positive is eliminating all the
>>old checks for zone boundary conditions.
>>
>>SPARSEMEM won't encounter this issue because of MAX_ORDER size constraint
>>when SPARSEMEM is configured. ia64 VIRTUAL_MEM_MAP doesn't need the
>>logic either because the holes and endpoints are handled differently.
>>This leaves checking alloc_remap and other arches which privately allocate
>>for node_mem_map.
> 
> 
> Do we think we need this in 2.6.17?

I would say yes, it is a very low risk patch in my view and provides a
very large part of the protections we require.  i386 as our largest
userbase should be safe from zone/node alignment issues with just this
change.  Others need slightly more (the page_zone_idx check) which is
being discussed in another thread.

-apw