Messages in this thread | | | From | Roland Dreier <> | Subject | Re: bug list: range checking issues 2.6.34-rc1 | Date | Tue, 16 Mar 2010 15:22:33 -0700 |
| |
> > drivers/infiniband/core/user_mad.c +646 ib_umad_reg_agent() 'umm' 4 <= 6 > 641 u32 *umm = (u32 *) ureq.method_mask; > 642 int i; > 643 > 644 for (i = 0; i < BITS_TO_LONGS(IB_MGMT_MAX_METHODS); ++i) > 645 req.method_mask[i] = > 646 umm[i * 2] | ((u64) umm[i * 2 + 1] << 32); > "umm" points to a array with 4 elements. > i can be 0 to 3, so "i * 2" goes up to 6 > And 4 <= 6 so it's a problem. > Smatch also complained about "i * 2 + 1" but I didn't include that.
It's a bit tricky, but I believe this is a false positive. The code in question is compatibility handling for 32-bit userspace on a 64-bit kernel. In that case the range of i will be 0 to 1 (IB_MGMT_MAX_METHODS is 128, so BITS_TO_LONGS on that is 2), and so we will only access elements 0, 1, 2, and 3 of umm[], which is OK.
(Not sure how easily a static checker could find this; the code in question is guarded by test of compat_method_mask, which can only be 1 if ib_umad_reg_agent() is called from ib_umad_compat_ioctl(), which will only be built with CONFIG_COMPAT set, which can only happen on a 64-bit architecture -- but it seems a bit hard for a checker to follow all that) -- Roland Dreier <rolandd@cisco.com> For corporate legal information go to: http://www.cisco.com/web/about/doing_business/legal/cri/index.html
| |