The atomicicty fix for move_one_page() was not quite right.

We only do the page_table_present() test if CONFIG_HIGHPTE=y.  Which is
fine, but even with CONFIG_HIGHPTE=n, the pte mapping functions still
do an inc_preempt_count() due to their unconditional kmap_atomic().  So
we get a might_sleep() warning.

The warning is actually bogus, because those pte's are always in
direct-mapped memory.

So hm.  Three fixes suggest themselves:

1: Run the page_table_present() test if CONFIG_HIGHMEM.

   Rejected: penalises non-pte_highmem setups

2: Make kmap_atomic() not do inc_preempt_count() is the page was
   direct mapped.

   Rejected: I don't think we want kmap_atomic side effects to be
   varying according to the page which was passed.

3: Change the pte mapping functions so they don't run kmap_atomic at
   all if CONFIG_HIGHPTE=n

   This is what I did.  And guess what?  For CONFIG_HIGHMEM=y,
   CONFIG_HIGHPTE=n this patch shrinks the kernel by 5 kbytes.  Because
   kmap_atomic is inlined.

   The lesson: we do way too much damn inlining.