So, this patch started out with me trying to keep from passing
contiguous, node-specific mem_map into free_area_init_node() and
cousins.  Instead, I relied on some calls to pfn_to_page().

This works fine and dandy when all you need is the pgdat->node_mem_map
to do pfn_to_page().  However, the non-NUMA/DISCONTIG architectures use
the real, global mem_map[] instead of a node_mem_map in the
pfn_to_page() calculation.  So, I ended up effectively trying to
initialize mem_map from itself, when it was NULL.  That was bad, and
caused some very pretty colors on someone's screen when he tested it.

So, I had to make sure to initialize the global mem_map[] before calling
into free_area_init_node().  Then, I realized how many architectures do
this on their own, and have comments like this:

        /* XXX: MRB-remove - this doesn't seem sane, should this be done somewhere else ?*/
        mem_map = NODE_DATA(0)->node_mem_map;

Also, ppc64 has some interesting hacks^Wcode in this area to make up for
any empty-memory NUMA nodes, which I think can go away now.

The patch magically removes more code than it adds.  It could be
smaller, but I shamelessly added some comments.

Applies on top of 2.6.8.1-mm4.  It touches enough architectures that it
probably needs plenty of cooling-off time in -mm.  

-- Dave