Re: [PATCH v11 01/12] add support for Clang's Shadow Call Stack (SCS)

[Kees Cook]

On Wed, Apr 22, 2020 at 04:51:34PM -0700, Sami Tolvanen wrote:
> On Wed, Apr 22, 2020 at 06:39:47PM +0100, Will Deacon wrote:
> > On Mon, Apr 20, 2020 at 02:18:30PM -0700, Sami Tolvanen wrote:
> > > On Mon, Apr 20, 2020 at 06:17:28PM +0100, Will Deacon wrote:
> > > > > +	 * The shadow call stack is aligned to SCS_SIZE, and grows
> > > > > +	 * upwards, so we can mask out the low bits to extract the base
> > > > > +	 * when the task is not running.
> > > > > +	 */
> > > > > +	return (void *)((unsigned long)task_scs(tsk) & ~(SCS_SIZE - 1));
> > > > 
> > > > Could we avoid forcing this alignment it we stored the SCS pointer as a
> > > > (base,offset) pair instead? That might be friendlier on the allocations
> > > > later on.
> > > 
> > > The idea is to avoid storing the current task's shadow stack address in
> > > memory, which is why I would rather not store the base address either.
> > 
> > What I mean is that, instead of storing the current shadow stack pointer,
> > we instead store a base and an offset. We can still clear the base, as you
> > do with the pointer today, and I don't see that the offset is useful to
> > an attacker on its own.
> 
> I see what you mean. However, even if we store the base address +
> the offset, we still need aligned allocation if we want to clear
> the address. This would basically just move __scs_base() logic to
> cpu_switch_to() / scs_save().

Okay, so, I feel like this has gotten off into the weeds, or I'm really
dense (or both). :) Going back to the original comment:

> > > > Could we avoid forcing this alignment it we stored the SCS
> > > > pointer as a (base,offset) pair instead? That might be friendlier
> > > > on the allocations later on.

I think there was some confusion about mixing the "we want to be able to
wipe the value" combined with the masking in __scs_base(). These are
unrelated, as was correctly observed with "We can still clear the base".

What I don't understand here is the suggestion to store two values:

Why is two better than storing one? With one, we only need a single access.

Why would storing the base be "friendlier on the allocations later on"?
This is coming out of a single kmem cache, in 1K chunks. They will be
naturally aligned to 1K (unless redzoing has been turned on for some
slab debugging reason). The base masking is a way to avoid needing to
store two values, and only happens at task death.

Storing two values eats memory for all tasks for seemingly no meaningful
common benefit. What am I missing here?

-- 
Kees Cook