Allow restricted per node allocations via __GFP_THISNODE

This implements a kernel facility to allocate on a particular node
without fallback and without obeying either memory policies or
cpuset constraints.

We add a new gfp flag __GFP_THISNODE that disables
cpusets, memory policies and fallback for an allocation. If
there is no memory on a node then the allocation will return
NULL.

This is necessary for certain kernel subsystems that manage list of
pages per node. Without this flag the page allocator may return
pages from other nodes or the allocations may be redirected
by memory policies or cpuset constraints.

These redirections have shown to be a problem because they cause
longer latencies during object draining in the slab allocator (but that is
probably just the first visible symptom of this fundamental problem).

Similar issues as for the slab allocator exist for the uncached
allocator.

__GFP_THISNODE can also be passed to vmalloc_* function to
insure placement on a certain node. This may be necessary
for certain driver allocations and allocation in the network
stack to avoid redirection there.

The patchset consists of 7 patches:

Subject: [RFC 0/7] cpuset / mempolicy interference with kernel allocs fix
Subject: [RFC 1/7] Cleanup get_page_from_freelist a bit
Subject: [RFC 2/7] Add __GFP_THISNODE to control off node fallback
Subject: [RFC 3/7] Enforce node allocation in sys_move_pages()
Subject: [RFC 4/7] Genericize slab_node and huge_zonelist
Subject: [RFC 5/7] Enforce node allocation in the uncached allocator
Subject: [RFC 6/7] Enforce node allocation for profiling
Subject: [RFC 7/7] slab: Keep nodelists pure and implement fall back