hugetlb: fix race condition in hugetlb_fault()

The race is as follows:

Suppose a multi-threaded task forks a new process (on cpu A), thus
bumping up the ref count on all the pages.  While the fork is occurring
(and thus we have marked all the PTEs as read-only), another thread in
the original process (on cpu B) tries to write to a huge page, taking an
access violation from the write-protect and calling hugetlb_cow().  Now,
suppose the fork() fails.  It will undo the COW and decrement the ref
count on the pages, so the ref count on the huge page drops back to 1.
Meanwhile hugetlb_cow() also decrements the ref count by one on the
original page, since the original address space doesn't need it any
more, having copied a new page to replace the original page.  This
leaves the ref count at zero, and when we call unlock_page(), we panic.

	fork on CPU A				fault on CPU B
	=============				==============
	...
	down_write(&parent->mmap_sem);
	down_write_nested(&child->mmap_sem);
	...
	while duplicating vmas
		if ...