aio: check for multiplication overflow in do_io_submit
authorJeff Moyer <jmoyer@redhat.com>
Fri, 10 Sep 2010 21:16:00 +0000 (14:16 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Wed, 15 Sep 2010 00:02:37 +0000 (17:02 -0700)
Tavis Ormandy pointed out that do_io_submit does not do proper bounds
checking on the passed-in iocb array:

       if (unlikely(nr < 0))
               return -EINVAL;

       if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(iocbpp)))))
               return -EFAULT;                      ^^^^^^^^^^^^^^^^^^

The attached patch checks for overflow, and if it is detected, the
number of iocbs submitted is scaled down to a number that will fit in
the long.  This is an ok thing to do, as sys_io_submit is documented as
returning the number of iocbs submitted, so callers should handle a
return value of less than the 'nr' argument passed in.

Reported-by: Tavis Ormandy <taviso@cmpxchg8b.com>
Signed-off-by: Jeff Moyer <jmoyer@redhat.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
fs/aio.c

index 3006b5bc33d697a2f773b2d81016fda57ee5ff68..1320b2a05fb2aaed4b261965683d814855abefe6 100644 (file)
--- a/fs/aio.c
+++ b/fs/aio.c
@@ -1659,6 +1659,9 @@ long do_io_submit(aio_context_t ctx_id, long nr,
        if (unlikely(nr < 0))
                return -EINVAL;
 
+       if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
+               nr = LONG_MAX/sizeof(*iocbpp);
+
        if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
                return -EFAULT;