thp: fix MADV_DONTNEED vs. MADV_FREE race

Both MADV_DONTNEED and MADV_FREE handled with down_read(mmap_sem).

It's critical to not clear pmd intermittently while handling MADV_FREE
to avoid race with MADV_DONTNEED:

	CPU0:				CPU1:
				madvise_free_huge_pmd()
				 pmdp_huge_get_and_clear_full()
madvise_dontneed()
 zap_pmd_range()
  pmd_trans_huge(*pmd) == 0 (without ptl)
  // skip the pmd
				 set_pmd_at();
				 // pmd is re-established

It results in MADV_DONTNEED skipping the pmd, leaving it not cleared.
It violates MADV_DONTNEED interface and can result is userspace
misbehaviour.

Basically it's the same race as with numa balancing in
change_huge_pmd(), but a bit simpler to mitigate: we don't need to
preserve dirty/young flags here due to MADV_FREE functionality.

[kirill.shutemov@linux.intel.com: Urgh... Power is special again]
  Link: http://lkml.kernel.org/r/20170303102636.bhd2zhtpds4mt62a@black.fi.intel.com
Link: http://lkml.kernel.org/r/20170302151034.27829-4-kirill.shutemov@linux.intel.com
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Acked-by: Minchan Kim <minchan@kernel.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Hillf Danton <hillf.zj@alibaba-inc.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

Author: kiryl

mm/huge_memory.c

@@ -1568,8 +1568,7 @@ bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
 		deactivate_page(page);
 
 	if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
-		orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
-			tlb->fullmm);
+		pmdp_invalidate(vma, addr, pmd);
 		orig_pmd = pmd_mkold(orig_pmd);
 		orig_pmd = pmd_mkclean(orig_pmd);