mm/filemap_xip.c: fix race condition in xip_file_fault()
Carsten Otte [Fri, 3 Feb 2012 23:37:14 +0000 (15:37 -0800)]
Fix a race condition that shows in conjunction with xip_file_fault() when
two threads of the same user process fault on the same memory page.

In this case, the race winner will install the page table entry and the
unlucky loser will cause an oops: xip_file_fault calls vm_insert_pfn (via
vm_insert_mixed) which drops out at this check:

retval = -EBUSY;
if (!pte_none(*pte))
goto out_unlock;

The resulting -EBUSY return value will trigger a BUG_ON() in
xip_file_fault.

This fix simply considers the fault as fixed in this case, because the
race winner has successfully installed the pte.

[akpm@linux-foundation.org: use conventional (and consistent) comment layout]
Reported-by: David Sadler <dsadler@us.ibm.com>
Signed-off-by: Carsten Otte <cotte@de.ibm.com>
Reported-by: Louis Alex Eisner <leisner@cs.ucsd.edu>
Cc: Hugh Dickins <hughd@google.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm/filemap_xip.c

index f91b2f6..a4eb311 100644 (file)
@@ -263,7 +263,12 @@ found:
                                                        xip_pfn);
                if (err == -ENOMEM)
                        return VM_FAULT_OOM;
-               BUG_ON(err);
+               /*
+                * err == -EBUSY is fine, we've raced against another thread
+                * that faulted-in the same page
+                */
+               if (err != -EBUSY)
+                       BUG_ON(err);
                return VM_FAULT_NOPAGE;
        } else {
                int err, ret = VM_FAULT_OOM;