From Martin Bligh.

This patch remaps the lmem_map (struct page) arrays for each node onto
their own nodes.  This is non-trivial, since all of ZONE_NORMAL, and
hence permanently mapped KVA resides on node 0.

Very early in the boot sequence, it calculates the size of the lmem_map
arrays (rounding up to the nearest large page size), and reserves a
suitable amount of permanent KVA by shifting down max_low_pfn to create
a gap between max_low_pfn and highstart_pfn (both of which are normally
about 896Mb).

It then uses the new set_pmd_pfn function to set up the pmds correctly
so that the large pages point at the physical addresses reserved from
the remote nodes.

Tested on NUMA-Q and some ratty old i386 PC kicking around under my
desk (on 2.5.36-mm1).  Was good for a 20% improvement in system time on
kernel compile when I initially benchmarked it against 2.5.32 or
something - due to a reduction in inter-node traffic, better
interconnect cache usage and locality.  Should have no effect on any
system other than i386 NUMA systems.