mm: vma_merge: fix vm_page_prot SMP race condition against rmap_walk

The rmap_walk can access vm_page_prot (and potentially vm_flags in the
pte/pmd manipulations).  So it's not safe to wait the caller to update
the vm_page_prot/vm_flags after vma_merge returned potentially removing
the "next" vma and extending the "current" vma over the
next->vm_start,vm_end range, but still with the "current" vma
vm_page_prot, after releasing the rmap locks.

The vm_page_prot/vm_flags must be transferred from the "next" vma to the
current vma while vma_merge still holds the rmap locks.

The side effect of this race condition is pte corruption during migrate
as remove_migration_ptes when run on a address of the "next" vma that
got removed, used the vm_page_prot of the current vma.

  migrate   	      	        mprotect
  ------------			-------------
  migrating in "next" vma
				vma_merge() # removes "next" vma and
			        	    # extends "current" vma
					    # current vma is not with
					    # vm_page_prot updated
  remove_migration_ptes
  read vm_page_prot of current "vma"
  establish pte with wrong permissions
				vm_set_page_prot(vma) # too late!
				change_protection in the old vma range
				only, next range is not updated

This caused segmentation faults and potentially memory corruption in
heavy mprotect loads with some light page migration caused by compaction
in the background.

Hugh Dickins pointed out the comment about the Odd case 8 in vma_merge
which confirms the case 8 is only buggy one where the race can trigger,
in all other vma_merge cases the above cannot happen.

This fix removes the oddness factor from case 8 and it converts it from:

      AAAA
  PPPPNNNNXXXX -> PPPPNNNNNNNN

to:

      AAAA
  PPPPNNNNXXXX -> PPPPXXXXXXXX

XXXX has the right vma properties for the whole merged vma returned by
vma_adjust, so it solves the problem fully.  It has the added benefits
that the callers could stop updating vma properties when vma_merge
succeeds however the callers are not updated by this patch (there are
bits like VM_SOFTDIRTY that still need special care for the whole range,
as the vma merging ignores them, but as long as they're not processed by
rmap walks and instead they're accessed with the mmap_sem at least for
reading, they are fine not to be updated within vma_adjust before
releasing the rmap_locks).

Link: http://lkml.kernel.org/r/1474309513-20313-1-git-send-email-aarcange@redhat.com
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Reported-by: Aditya Mandaleeka <adityam@microsoft.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Jan Vorlicek <janvorli@microsoft.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

1 files changed, 8 insertions, 2 deletions

diff --git a/include/linux/mm.h b/include/linux/mm.h
index 040a04a88996..2c8ed8a894c8 100644
--- a/include/linux/mm.h
+++ b/include/linux/mm.h
@@ -1968,8 +1968,14 @@ void anon_vma_interval_tree_verify(struct anon_vma_chain *node);
 
 /* mmap.c */
 extern int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin);
-extern int vma_adjust(struct vm_area_struct *vma, unsigned long start,
-	unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert);
+extern int __vma_adjust(struct vm_area_struct *vma, unsigned long start,
+	unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert,
+	struct vm_area_struct *expand);
+static inline int vma_adjust(struct vm_area_struct *vma, unsigned long start,
+	unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
+{
+	return __vma_adjust(vma, start, end, pgoff, insert, NULL);
+}
 extern struct vm_area_struct *vma_merge(struct mm_struct *,
 	struct vm_area_struct *prev, unsigned long addr, unsigned long end,
 	unsigned long vm_flags, struct anon_vma *, struct file *, pgoff_t,