In some cases it may happen that pmd_none_or_clear_bad() is called with
the mmap_sem hold in read mode.  In those cases the huge page faults can
allocate hugepmds under pmd_none_or_clear_bad() and that can trigger a
false positive from pmd_bad() that will not like to see a pmd
materializing as trans huge.

It's not khugepaged causing the problem, khugepaged holds the mmap_sem
in write mode (and all those sites must hold the mmap_sem in read mode
to prevent pagetables to go away from under them, during code review it
seems vm86 mode on 32bit kernels requires that too unless it's
restricted to 1 thread per process or UP builds).  The race is only with
the huge pagefaults that can convert a pmd_none() into a
pmd_trans_huge().

Effectively all these pmd_none_or_clear_bad() sites running with
mmap_sem in read mode are somewhat speculative with the page faults, and
the result is always undefined when they run simultaneously.  This is
probably why it wasn't common to run into this.  For example if the
madvise(MADV_DONTNEED) runs zap_page_range() shortly before the page
fault, the hugepage will not be zapped, if the page fault runs first it
will be zapped.

Altering pmd_bad() not to error out if it finds hugepmds won't be enough
to fix this, because zap_pmd_range would then proceed to call
zap_pte_range (which would be incorrect if the pmd become a
pmd_trans_huge()).

The simplest way to fix this is to read the pmd in the local stack
(regardless of what we read, no need of actual CPU barriers, only
compiler barrier needed), and be sure it is not changing under the code
that computes its value.  Even if the real pmd is changing under the
value we hold on the stack, we don't care.  If we actually end up in
zap_pte_range it means the pmd was not none already and it was not huge,
and it can't become huge from under us (khugepaged locking explained
above).

All we need is to enforce that there is no way anymore that in a code
path like below, pmd_trans_huge can be false, but pmd_none_or_clear_bad
can run into a hugepmd.  The overhead of a barrier() is just a compiler
tweak and should not be measurable (I only added it for THP builds).  I
don't exclude different compiler versions may have prevented the race
too by caching the value of *pmd on the stack (that hasn't been
verified, but it wouldn't be impossible considering
pmd_none_or_clear_bad, pmd_bad, pmd_trans_huge, pmd_none are all inlines
and there's no external function called in between pmd_trans_huge and
pmd_none_or_clear_bad).

		if (pmd_trans_huge(*pmd)) {
			if (next-addr != HPAGE_PMD_SIZE) {
				VM_BUG_ON(!rwsem_is_locked(&tlb->mm->mmap_sem));
				split_huge_page_pmd(vma->vm_mm, pmd);
			} else if (zap_huge_pmd(tlb, vma, pmd, addr))
				continue;
			/* fall through */
		}
		if (pmd_none_or_clear_bad(pmd))

Because this race condition could be exercised without special
privileges this was reported in CVE-2012-1179.

The race was identified and fully explained by Ulrich who debugged it.
I'm quoting his accurate explanation below, for reference.

====== start quote =======
      mapcount 0 page_mapcount 1
      kernel BUG at mm/huge_memory.c:1384!

    At some point prior to the panic, a "bad pmd ..." message similar to the
    following is logged on the console:

      mm/memory.c:145: bad pmd ffff8800376e1f98(80000000314000e7).

    The "bad pmd ..." message is logged by pmd_clear_bad() before it clears
    the page's PMD table entry.

        143 void pmd_clear_bad(pmd_t *pmd)
        144 {
    ->  145         pmd_ERROR(*pmd);
        146         pmd_clear(pmd);
        147 }

    After the PMD table entry has been cleared, there is an inconsistency
    between the actual number of PMD table entries that are mapping the page
    and the page's map count (_mapcount field in struct page). When the page
    is subsequently reclaimed, __split_huge_page() detects this inconsistency.

       1381         if (mapcount != page_mapcount(page))
       1382                 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
       1383                        mapcount, page_mapcount(page));
    -> 1384         BUG_ON(mapcount != page_mapcount(page));

    The root cause of the problem is a race of two threads in a multithreaded
    process. Thread B incurs a page fault on a virtual address that has never
    been accessed (PMD entry is zero) while Thread A is executing an madvise()
    system call on a virtual address within the same 2 MB (huge page) range.

               virtual address space
              .---------------------.
              |                     |
              |                     |
            .-|---------------------|
            | |                     |
            | |                     |<-- B(fault)
            | |                     |
      2 MB  | |/////////////////////|-.
      huge <  |/////////////////////|  > A(range)
      page  | |/////////////////////|-'
            | |                     |
            | |                     |
            '-|---------------------|
              |                     |
              |                     |
              '---------------------'

    - Thread A is executing an madvise(..., MADV_DONTNEED) system call
      on the virtual address range "A(range)" shown in the picture.

    sys_madvise
      // Acquire the semaphore in shared mode.
      down_read(&current->mm->mmap_sem)
      ...
      madvise_vma
        switch (behavior)
        case MADV_DONTNEED:
             madvise_dontneed
               zap_page_range
                 unmap_vmas
                   unmap_page_range
                     zap_pud_range
                       zap_pmd_range
                         //
                         // Assume that this huge page has never been accessed.
                         // I.e. content of the PMD entry is zero (not mapped).
                         //
                         if (pmd_trans_huge(*pmd)) {
                             // We don't get here due to the above assumption.
                         }
                         //
                         // Assume that Thread B incurred a page fault and
             .---------> // sneaks in here as shown below.
             |           //
             |           if (pmd_none_or_clear_bad(pmd))
             |               {
             |                 if (unlikely(pmd_bad(*pmd)))
             |                     pmd_clear_bad
             |                     {
             |                       pmd_ERROR
             |                         // Log "bad pmd ..." message here.
             |                       pmd_clear
             |                         // Clear the page's PMD entry.
             |                         // Thread B incremented the map count
             |                         // in page_add_new_anon_rmap(), but
             |                         // now the page is no longer mapped
             |                         // by a PMD entry (-> inconsistency).
             |                     }
             |               }
             |
             v
    - Thread B is handling a page fault on virtual address "B(fault)" shown
      in the picture.

    ...
    do_page_fault
      __do_page_fault
        // Acquire the semaphore in shared mode.
        down_read_trylock(&mm->mmap_sem)
        ...
        handle_mm_fault
          if (pmd_none(*pmd) && transparent_hugepage_enabled(vma))
              // We get here due to the above assumption (PMD entry is zero).
              do_huge_pmd_anonymous_page
                alloc_hugepage_vma
                  // Allocate a new transparent huge page here.
                ...
                __do_huge_pmd_anonymous_page
                  ...
                  spin_lock(&mm->page_table_lock)
                  ...
                  page_add_new_anon_rmap
                    // Here we increment the page's map count (starts at -1).
                    atomic_set(&page->_mapcount, 0)
                  set_pmd_at
                    // Here we set the page's PMD entry which will be cleared
                    // when Thread A calls pmd_clear_bad().
                  ...
                  spin_unlock(&mm->page_table_lock)

    The mmap_sem does not prevent the race because both threads are acquiring
    it in shared mode (down_read).  Thread B holds the page_table_lock while
    the page's map count and PMD table entry are updated.  However, Thread A
    does not synchronize on that lock.

====== end quote =======

[akpm@linux-foundation.org: checkpatch fixes]
Reported-by: Ulrich Obergfell <uobergfe@redhat.com>
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Hugh Dickins <hughd@google.com>
Cc: Dave Jones <davej@redhat.com>
Acked-by: Larry Woodman <lwoodman@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Cc: <stable@vger.kernel.org>		[2.6.38+]
Cc: Mark Salter <msalter@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>