mm: numa: recheck for transhuge pages under lock during protection changes
authorMel Gorman <mgorman@suse.de>
Mon, 7 Apr 2014 22:36:56 +0000 (15:36 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Mon, 7 Apr 2014 23:35:50 +0000 (16:35 -0700)
Sasha reported the following bug using trinity

  kernel BUG at mm/mprotect.c:149!
  invalid opcode: 0000 [#1] PREEMPT SMP DEBUG_PAGEALLOC
  Dumping ftrace buffer:
     (ftrace buffer empty)
  Modules linked in:
  CPU: 20 PID: 26219 Comm: trinity-c216 Tainted: G        W    3.14.0-rc5-next-20140305-sasha-00011-ge06f5f3-dirty #105
  task: ffff8800b6c80000 ti: ffff880228436000 task.ti: ffff880228436000
  RIP: change_protection_range+0x3b3/0x500
  Call Trace:
    change_protection+0x25/0x30
    change_prot_numa+0x1b/0x30
    task_numa_work+0x279/0x360
    task_work_run+0xae/0xf0
    do_notify_resume+0x8e/0xe0
    retint_signal+0x4d/0x92

The VM_BUG_ON was added in -mm by the patch "mm,numa: reorganize
change_pmd_range".  The race existed without the patch but was just
harder to hit.

The problem is that a transhuge check is made without holding the PTL.
It's possible at the time of the check that a parallel fault clears the
pmd and inserts a new one which then triggers the VM_BUG_ON check.  This
patch removes the VM_BUG_ON but fixes the race by rechecking transhuge
under the PTL when marking page tables for NUMA hinting and bailing if a
race occurred.  It is not a problem for calls to mprotect() as they hold
mmap_sem for write.

Signed-off-by: Mel Gorman <mgorman@suse.de>
Reported-by: Sasha Levin <sasha.levin@oracle.com>
Reviewed-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/mprotect.c

index 79cb51866e02f217ddb7d98d03b7c7dd2dc4b0f7..2c51c79c8a698189c53380352f9770d234d0772c 100644 (file)
@@ -36,6 +36,34 @@ static inline pgprot_t pgprot_modify(pgprot_t oldprot, pgprot_t newprot)
 }
 #endif
 
+/*
+ * For a prot_numa update we only hold mmap_sem for read so there is a
+ * potential race with faulting where a pmd was temporarily none. This
+ * function checks for a transhuge pmd under the appropriate lock. It
+ * returns a pte if it was successfully locked or NULL if it raced with
+ * a transhuge insertion.
+ */
+static pte_t *lock_pte_protection(struct vm_area_struct *vma, pmd_t *pmd,
+                       unsigned long addr, int prot_numa, spinlock_t **ptl)
+{
+       pte_t *pte;
+       spinlock_t *pmdl;
+
+       /* !prot_numa is protected by mmap_sem held for write */
+       if (!prot_numa)
+               return pte_offset_map_lock(vma->vm_mm, pmd, addr, ptl);
+
+       pmdl = pmd_lock(vma->vm_mm, pmd);
+       if (unlikely(pmd_trans_huge(*pmd) || pmd_none(*pmd))) {
+               spin_unlock(pmdl);
+               return NULL;
+       }
+
+       pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, ptl);
+       spin_unlock(pmdl);
+       return pte;
+}
+
 static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
                unsigned long addr, unsigned long end, pgprot_t newprot,
                int dirty_accountable, int prot_numa)
@@ -45,7 +73,10 @@ static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
        spinlock_t *ptl;
        unsigned long pages = 0;
 
-       pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
+       pte = lock_pte_protection(vma, pmd, addr, prot_numa, &ptl);
+       if (!pte)
+               return 0;
+
        arch_enter_lazy_mmu_mode();
        do {
                oldpte = *pte;
@@ -132,12 +163,13 @@ static inline unsigned long change_pmd_range(struct vm_area_struct *vma,
                                                pages += HPAGE_PMD_NR;
                                                nr_huge_updates++;
                                        }
+
+                                       /* huge pmd was handled */
                                        continue;
                                }
                        }
                        /* fall through, the trans huge pmd just split */
                }
-               VM_BUG_ON(pmd_trans_huge(*pmd));
                this_pages = change_pte_range(vma, pmd, addr, next, newprot,
                                 dirty_accountable, prot_numa);
                pages += this_pages;