On Mon, 2016-10-31 at 21:45 +0100, Florian Weimer wrote:
> * Jann Horn:
> 
> > On Mon, Oct 31, 2016 at 09:04:02AM -0700, Kees Cook wrote:
> > > On Mon, Oct 31, 2016 at 7:04 AM, Jann Horn <jann@thejh.net> wrote:
> > > > On machines with sizeof(unsigned long)==8, this ensures that the
> > > > more
> > > > significant 32 bits of stack_canary are random, too.
> > > > stack_canary is defined as unsigned long, all the architectures
> > > > with stack
> > > > protector support already pick the stack_canary of init as a
> > > > random
> > > > unsigned long, and get_random_long() should be as fast as
> > > > get_random_int(),
> > > > so there seems to be no good reason against this.
> > > > 
> > > > This should help if someone tries to guess a stack canary with
> > > > brute force.
> > > > 
> > > > (This change has been made in PaX already, with a different
> > > > RNG.)
> > > > 
> > > > Signed-off-by: Jann Horn <jann@thejh.net>
> > > 
> > > Acked-by: Kees Cook <keescook@chromium.org>
> > > 
> > > (A separate change might be to make sure that the leading byte is
> > > zeroed. Entropy of the value, I think, is less important than
> > > blocking
> > > canary exposures from unbounded str* functions. Brute forcing
> > > kernel
> > > stack canaries isn't like it bruting them in userspace...)
> > 
> > Yeah, makes sense. Especially on 64bit, 56 bits of entropy ought to
> > be
> > enough anyway.
> 
> So you two approve of the way glibc does this currently?  (See the
> other thread.)
> 
> I was under the impression that the kernel performs far less
> null-terminated string processing the average user space application,
> especially on the stack.  (A lot of userspace code assumes large
> stacks and puts essentially arbitrarily long strings into VLAs.)

It makes a lot of sense on x86_64 where it means the canary is still 56
bits. Also, you want -fstack-check for protecting again stack overflows
rather than stack *buffer* overflow. SSP won't really help you in that
regard. Sadly, while -fstack-check now works well in GCC 6 with little
performance cost, it's not really a complete feature (and Clang impls it
as a no-op!).