We have a 64GB NUMA machine that doesn't run very well on the stock
kernel due to lack of ZONE_NORMAL.  I think it has ~150MB of free lowmem
immediately after boot, due to consumption of mem_map.

The attached patch is effectively mem= designed for NUMA.  It restricts
the total amount of memory on each NUMA node to the specified amount. 
We need this because a plain mem= on the machine will make the memory
get distributed unevenly which is very bad.

node | plain | mem=32G | memnode=8G
-----+-------+---------+-----------
   0 |   16G |     16G |        8G
   1 |   16G |     16G |        8G
   2 |   16G |      0G |        8G
   3 |   16G |      0G |        8G

Tested on 4-node 32GB NUMA-Q and 4-node 64GB x440.  

In playing with this, I also noticed that we should clean up some of
they variable names in srat.c.  They're long and if you reference 3 of
them you take up an entire 80 column line.
-- 
Dave Hansen
haveblue@us.ibm.com