Appended is a patch which stops using the zone->zone_mem_map to calculate
the buddy and combined page pointers.  It uses the fact that the mem_map
array is guaranteed to be contigious for the surrounding (1 << MAX_ORDER)
pages.  The relative positions of the pages in the physical address space
to provide the alignement; which conicidentally fixes the issue where zones
are not aligned at MAX_ORDER.  There is a very comprehensive comment in the
new code explaining the mathematical relationship between a page and its
buddy so I won't reproduce it here.

This kind of approach is required for CONFIG_NONLINEAR systems where the
mem_map is not contiguous within a zone, and the zone->zone_mem_map is not
used at all.

This patch has been boot-tested on a large variety of systems and
architectures: my P4 laptop, 16-way NUMAQ, 16-way Summit, 4-way x86 SMP,
ppc64 LPAR, x86_64, and several ia64 configurations.

It has been performance-tested on a 16-way NUMAQ.  SDET shows a very slight
(within margin of error) performance gain.  Kernbench shows an
approximately ~1% decrease in system time with this patch applied.  So, it
has a likely positive performance impact.

However, the patch has the potential to have a negative performance impact
on systems with an expensive page_to_pfn() implementation.  But, I think
the NUMAQ has one of the more expensive ones around, and it doesn't seem
mind too much.
Signed-off-by: Andy Whitcroft <apw@shadowen.org>
Signed-off-by: Dave Hansen <haveblue@us.ibm.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>