Instead of statically allocating the NUMA node structures, do it at runtime
with node local memory where possible.  With NR_CPUS=128 we had over 800kB
of memory allocated before this patch, now we allocate only as many nodes
as the machine requires.

We have to do some fancy footwork in do_init_bootmem since we use both the
LMB allocator and the bootmem allocator at the same time.  The problem has
always been there although I only recently found a test case that hit it.
Wrap that logic in careful_allocation which uses the LMB allocator then
falls back to the bootmem allocator if required.

Now alloc_bootmem on a node with no memory doesnt panic, we dont need the
hack in paging_init so remove it.
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>