On Mon, Jul 6, 2020 at 4:31 AM Kees Cook <keescook@xxxxxxxxxxxx> wrote:
>
> On Sun, Jul 05, 2020 at 10:16:14PM +0800, Guo Ren wrote:
> > On Sun, Jul 5, 2020 at 2:53 PM Kees Cook <keescook@xxxxxxxxxxxx> wrote:
> > > On Sun, Jul 05, 2020 at 06:24:15AM +0000, guoren@xxxxxxxxxx wrote:
> > > > +static __always_inline void boot_init_stack_canary(void)
> > > > +{
> > > > + unsigned long canary;
> > > > +
> > > > + /* Try to get a semi random initial value. */
> > > > + get_random_bytes(&canary, sizeof(canary));
> > > > + canary ^= LINUX_VERSION_CODE;
> > > > + canary &= CANARY_MASK;
> > >
> > > Does riscv have any kind of instruction counters or other trivial timers
> > > that could be mixed in here? (e.g. x86's TSC)
> > Do you mean:
> > get_random_bytes(&canary, sizeof(canary));
> > + canary += get_cycles64() + (get_cycles64() << 32UL);
> > canary ^= LINUX_VERSION_CODE;
> > canary &= CANARY_MASK;
> >
> > Ok ?
>
> Sure -- I assume get_cycles64() is architecturally "simple"? (i.e. it
> doesn't require that the entire time-keeping subsystem has started?)
Yes, it's just a csr read. But it's necessary? get_random_bytes should enough.
>
> > >
> > > > +
> > > > + current->stack_canary = canary;
> > > > + __stack_chk_guard = current->stack_canary;
> > >
> > > What's needed for riscv to support a per-task canary? (e.g. x86's TLS or
> > > arm64's register-specific methods)
> > Some archs change __stack_chk_guard in _switch_to of entry.S, but it
> > depends on !CONFIG_SMP.
>
> Oh, funny. I hadn't actually noticed that logic for the !CONFIG_SMP
> cases. I see to problem with that, but the more important case, I think
> is the per-task canaries.
Maybe some race condition problems. When canary changed the in the
switch to, but other CPUs still get that value concurrently.
>
> > #if defined(CONFIG_STACKPROTECTOR) && !defined(CONFIG_SMP)
> > get value from next_task->stack_canary
> > store value to __stack_chk_guard
> > #endif
> >
> > It's a so limitation solution for per-task canary, so I didn't copy it
> > into riscv?
>
> Right -- it's a limited solution. On the other had, is !CONFIG_SMP
> expected to be a common config for riscv? If so, it's worth adding. If
> not, I'd say skip it. (Though it looks very simple to do...)
CONFIG_SMP is mostly default for me and let's talk about arm64/x86
per-task solution. That is the right way.
--
Best Regards
Guo Ren
ML: https://lore.kernel.org/linux-csky/
