The VM allocates pte_chains with GFP_ATOMIC, under deep locking.  If that
allocation fails, we oops.

My approach to solving this is to require that the caller of page_add_rmap()
pass in a pte_chain structure for page_add_rmap() to use.  Then, callers can
arrange to allocate that structure outside locks with GFP_KERNEL.

This patch provides the base infrastructure.

A common case is that page_add_rmap() will in fact not consume the pte_chain,
because an empty slot was found within one of the page's existing pte_chain
structures.  So this patch provides for a special one-deep per-cpu pte_chain
cache to optimise this case of taking just one pte_chain and then immediately
putting it back.

We end up adding maybe 20-30 instructions to the pagefault path to handle the
eventuality of pte_chain allocation failures.

Lots of other design ideas were considered.  This is the best I could come up
with.