Messages in this thread | | | From | Andrew Lutomirski <> | Date | Mon, 27 Feb 2012 10:35:55 -0800 | Subject | Re: [PATCH v11 07/12] seccomp: add SECCOMP_RET_ERRNO |
| |
On Mon, Feb 27, 2012 at 10:14 AM, Oleg Nesterov <oleg@redhat.com> wrote: > On 02/27, Kees Cook wrote: >> >> On Mon, Feb 27, 2012 at 9:11 AM, Oleg Nesterov <oleg@redhat.com> wrote: >> > On 02/24, Will Drewry wrote: >> >> >> >> static u32 seccomp_run_filters(int syscall) >> >> { >> >> struct seccomp_filter *f; >> >> - u32 ret = SECCOMP_RET_KILL; >> >> static const struct bpf_load_fn fns = { >> >> bpf_load, >> >> sizeof(struct seccomp_data), >> >> }; >> >> + u32 ret = SECCOMP_RET_ALLOW; >> >> const void *sc_ptr = (const void *)(uintptr_t)syscall; >> >> >> >> + /* Ensure unexpected behavior doesn't result in failing open. */ >> >> + if (unlikely(current->seccomp.filter == NULL)) >> >> + ret = SECCOMP_RET_KILL; >> > >> > Is "seccomp.filter == NULL" really possible? >> >> It should not be, but I'm much more comfortable with this failing >> closed. I think it's important to be as defensive as possible with >> this code given its intended use. > > Can't resists... Sorry, I know I am troll but personally I think > in this case the most defensive code is BUG_ON(->filter == NULL) > or at least WARN_ON().
Linus will probably object because he objected (correctly) to a very similar problem in my old vsyscall emulation series. A userspace security feature shouldn't have a failure mode in which it confuses the kernel and results in an oops, unless the situation is really unrecoverable. So WARN_ON plus do_exit would be okay but BUG_ON would not.
--Andy -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
| |