mm: thp: fix SMP race condition between THP page fault and MADV_DONTNEED

pmd_trans_unstable()/pmd_none_or_trans_huge_or_clear_bad() were
introduced to locklessy (but atomically) detect when a pmd is a regular
(stable) pmd or when the pmd is unstable and can infinitely transition
from pmd_none() and pmd_trans_huge() from under us, while only holding
the mmap_sem for reading (for writing not).

While holding the mmap_sem only for reading, MADV_DONTNEED can run from
under us and so before we can assume the pmd to be a regular stable pmd
we need to compare it against pmd_none() and pmd_trans_huge() in an
atomic way, with pmd_trans_unstable().  The old pmd_trans_huge() left a
tiny window for a race.

Useful applications are unlikely to notice the difference as doing
MADV_DONTNEED concurrently with a page fault would lead to undefined
behavior.

[akpm@linux-foundation.org: tidy up comment grammar/layout]
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Reported-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

diff --git a/mm/memory.c b/mm/memory.c
index 635451abc8f7c9b84663535678a26a9b0848c89d..8132787ae4d509d475ed6a77705076ddfee30630 100644 (file)
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -3404,8 +3404,18 @@ static int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
        if (unlikely(pmd_none(*pmd)) &&
            unlikely(__pte_alloc(mm, vma, pmd, address)))
                return VM_FAULT_OOM;
-       /* if an huge pmd materialized from under us just retry later */
-       if (unlikely(pmd_trans_huge(*pmd) || pmd_devmap(*pmd)))
+       /*
+        * If a huge pmd materialized under us just retry later.  Use
+        * pmd_trans_unstable() instead of pmd_trans_huge() to ensure the pmd
+        * didn't become pmd_trans_huge under us and then back to pmd_none, as
+        * a result of MADV_DONTNEED running immediately after a huge pmd fault
+        * in a different thread of this mm, in turn leading to a misleading
+        * pmd_trans_huge() retval.  All we have to ensure is that it is a
+        * regular pmd that we can walk with pte_offset_map() and we can do that
+        * through an atomic read in C, which is what pmd_trans_unstable()
+        * provides.
+        */
+       if (unlikely(pmd_trans_unstable(pmd) || pmd_devmap(*pmd)))
                return 0;
        /*
         * A regular pmd is established and it can't morph into a huge pmd