mm: Fix NULL pointer dereference in madvise(MADV_WILLNEED) support

Sasha Levin found a NULL pointer dereference that is due to a missing
page table lock, which in turn is due to the pmd entry in question being
a transparent huge-table entry.

The code - introduced in commit 1998cc0489 ("mm: make
madvise(MADV_WILLNEED) support swap file prefetch") - correctly checks
for this situation using pmd_none_or_trans_huge_or_clear_bad(), but it
turns out that that function doesn't work correctly.

pmd_none_or_trans_huge_or_clear_bad() expected that pmd_bad() would
trigger if the transparent hugepage bit was set, but it doesn't do that
if pmd_numa() is also set. Note that the NUMA bit only gets set on real
NUMA machines, so people trying to reproduce this on most normal
development systems would never actually trigger this.

Fix it by removing the very subtle (and subtly incorrect) expectation,
and instead just checking pmd_trans_huge() explicitly.

Reported-by: Sasha Levin <sasha.levin@oracle.com>
Acked-by: Andrea Arcangeli <aarcange@redhat.com>
[ Additionally remove the now stale test for pmd_trans_huge() inside the
  pmd_bad() case - Linus ]
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

Change-Id: I3f3763f236ef102de735297cd175cf514d40d28f
This commit is contained in:
Kirill A. Shutemov 2013-12-20 15:10:03 +02:00 committed by Zhao Wei Liew
parent ba0d2ae061
commit 397b074320
1 changed files with 2 additions and 3 deletions

View File

@ -471,11 +471,10 @@ static inline int pmd_none_or_trans_huge_or_clear_bad(pmd_t *pmd)
#ifdef CONFIG_TRANSPARENT_HUGEPAGE
barrier();
#endif
if (pmd_none(pmdval))
if (pmd_none(pmdval) || pmd_trans_huge(pmdval))
return 1;
if (unlikely(pmd_bad(pmdval))) {
if (!pmd_trans_huge(pmdval))
pmd_clear_bad(pmd);
pmd_clear_bad(pmd);
return 1;
}
return 0;