Re: [PATCH v2 05/17] add support for Clang's Shadow Call Stack (SCS)

From: Sami Tolvanen
Date: Tue Oct 29 2019 - 14:06:47 EST

On Mon, Oct 28, 2019 at 12:57 PM Kees Cook <keescook@xxxxxxxxxxxx> wrote:
> On Mon, Oct 28, 2019 at 04:35:33PM +0000, Mark Rutland wrote:
> > On Fri, Oct 25, 2019 at 01:49:21PM -0700, Sami Tolvanen wrote:
> > > To keep the address of the currently active shadow stack out of
> > > memory, the arm64 implementation clears this field when it loads x18
> > > and saves the current value before a context switch. The generic code
> > > doesn't expect the arch code to necessarily do so, but does allow it.
> > > This requires us to use __scs_base() when accessing the base pointer
> > > and to reset it in idle tasks before they're reused, hence
> > > scs_task_reset().
> >
> > Ok. That'd be worth a comment somewhere, since it adds a number of
> > things which would otherwise be unnecessary.
> >
> > IIUC this assumes an adversary who knows the address of a task's
> > thread_info, and has an arbitrary-read (to extract the SCS base from
> > thead_info) and an arbitrary-write (to modify the SCS area).
> >
> > Assuming that's the case, I don't think this buys much. If said
> > adversary controls two userspace threads A and B, they only need to wait
> > until A is context-switched out or in userspace, and read A's SCS base
> > using B.
> >
> > Given that, I'd rather always store the SCS base in the thread_info, and
> > simplify the rest of the code manipulating it.
> I'd like to keep this as-is since it provides a temporal protection.
> Having arbitrary kernel read and write at arbitrary time is a very
> powerful attack primitive, and is, IMO, not very common. Many attacks
> tend to be chains of bugs that give attackers narrow visibility in to the
> kernel at specific moments. I would say this design is more about stopping
> "current" from dumping thread_info (as there are many more opportunities
> for current to see its own thread_info compared to arbitrary addresses
> or another task's thread_info). As such, I think it's a reasonable
> precaution to take.

I'm not sure if always storing the base address in thread_info would
simplify the code that much. We could remove __scs_base() and
scs_task_reset(), which are both trivial, and drop a few instructions
in the arch-specific code that clear the field. I do agree that a
comment or two would help understand what's going on here though.