thp: close race between split and zap huge pages
authorKirill A. Shutemov <kirill.shutemov@linux.intel.com>
Fri, 18 Apr 2014 22:07:25 +0000 (15:07 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Fri, 18 Apr 2014 23:40:09 +0000 (16:40 -0700)
Sasha Levin has reported two THP BUGs[1][2].  I believe both of them
have the same root cause.  Let's look to them one by one.

The first bug[1] is "kernel BUG at mm/huge_memory.c:1829!".  It's
BUG_ON(mapcount != page_mapcount(page)) in __split_huge_page().  From my
testing I see that page_mapcount() is higher than mapcount here.

I think it happens due to race between zap_huge_pmd() and
page_check_address_pmd().  page_check_address_pmd() misses PMD which is
under zap:

CPU0 CPU1
zap_huge_pmd()
  pmdp_get_and_clear()
__split_huge_page()
  anon_vma_interval_tree_foreach()
    __split_huge_page_splitting()
      page_check_address_pmd()
        mm_find_pmd()
  /*
   * We check if PMD present without taking ptl: no
   * serialization against zap_huge_pmd(). We miss this PMD,
   * it's not accounted to 'mapcount' in __split_huge_page().
   */
  pmd_present(pmd) == 0

  BUG_ON(mapcount != page_mapcount(page)) // CRASH!!!

  page_remove_rmap(page)
    atomic_add_negative(-1, &page->_mapcount)

The second bug[2] is "kernel BUG at mm/huge_memory.c:1371!".
It's VM_BUG_ON_PAGE(!PageHead(page), page) in zap_huge_pmd().

This happens in similar way:

CPU0 CPU1
zap_huge_pmd()
  pmdp_get_and_clear()
  page_remove_rmap(page)
    atomic_add_negative(-1, &page->_mapcount)
__split_huge_page()
  anon_vma_interval_tree_foreach()
    __split_huge_page_splitting()
      page_check_address_pmd()
        mm_find_pmd()
  pmd_present(pmd) == 0 /* The same comment as above */
  /*
   * No crash this time since we already decremented page->_mapcount in
   * zap_huge_pmd().
   */
  BUG_ON(mapcount != page_mapcount(page))

  /*
   * We split the compound page here into small pages without
   * serialization against zap_huge_pmd()
   */
  __split_huge_page_refcount()
VM_BUG_ON_PAGE(!PageHead(page), page); // CRASH!!!

So my understanding the problem is pmd_present() check in mm_find_pmd()
without taking page table lock.

The bug was introduced by me commit with commit 117b0791ac42. Sorry for
that. :(

Let's open code mm_find_pmd() in page_check_address_pmd() and do the
check under page table lock.

Note that __page_check_address() does the same for PTE entires
if sync != 0.

I've stress tested split and zap code paths for 36+ hours by now and
don't see crashes with the patch applied. Before it took <20 min to
trigger the first bug and few hours for second one (if we ignore
first).

[1] https://lkml.kernel.org/g/<53440991.9090001@oracle.com>
[2] https://lkml.kernel.org/g/<5310C56C.60709@oracle.com>

Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Reported-by: Sasha Levin <sasha.levin@oracle.com>
Tested-by: Sasha Levin <sasha.levin@oracle.com>
Cc: Bob Liu <lliubbo@gmail.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Michel Lespinasse <walken@google.com>
Cc: Dave Jones <davej@redhat.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: <stable@vger.kernel.org> [3.13+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/huge_memory.c

index 64635f5278ff2d41525fe1ca3ebcf5314c95e3e4..b4b1feba64724234dee1b66a482c79a0cd3c0f95 100644 (file)
@@ -1536,16 +1536,23 @@ pmd_t *page_check_address_pmd(struct page *page,
                              enum page_check_address_pmd_flag flag,
                              spinlock_t **ptl)
 {
+       pgd_t *pgd;
+       pud_t *pud;
        pmd_t *pmd;
 
        if (address & ~HPAGE_PMD_MASK)
                return NULL;
 
-       pmd = mm_find_pmd(mm, address);
-       if (!pmd)
+       pgd = pgd_offset(mm, address);
+       if (!pgd_present(*pgd))
                return NULL;
+       pud = pud_offset(pgd, address);
+       if (!pud_present(*pud))
+               return NULL;
+       pmd = pmd_offset(pud, address);
+
        *ptl = pmd_lock(mm, pmd);
-       if (pmd_none(*pmd))
+       if (!pmd_present(*pmd))
                goto unlock;
        if (pmd_page(*pmd) != page)
                goto unlock;