Re: [PATCH v13 00/12] add support for Clang's Shadow Call Stack

[Ard Biesheuvel]

On Mon, 27 Apr 2020 at 19:39, Ard Biesheuvel <ardb@xxxxxxxxxx> wrote:
>
> On Mon, 27 Apr 2020 at 18:00, Sami Tolvanen <samitolvanen@xxxxxxxxxx> wrote:
> >
> > This patch series adds support for Clang's Shadow Call Stack
> > (SCS) mitigation, which uses a separately allocated shadow stack
> > to protect against return address overwrites. More information
> > can be found here:
> >
> >   https://clang.llvm.org/docs/ShadowCallStack.html
> >
> > SCS provides better protection against traditional buffer
> > overflows than CONFIG_STACKPROTECTOR_*, but it should be noted
> > that SCS security guarantees in the kernel differ from the ones
> > documented for user space. The kernel must store addresses of
> > shadow stacks in memory, which means an attacker capable of
> > reading and writing arbitrary memory may be able to locate them
> > and hijack control flow by modifying the shadow stacks.
> >
> > SCS is currently supported only on arm64, where the compiler
> > requires the x18 register to be reserved for holding the current
> > task's shadow stack pointer.
> >
> > With -fsanitize=shadow-call-stack, the compiler injects
> > instructions to all non-leaf C functions to store the return
> > address to the shadow stack, and unconditionally load it again
> > before returning. As a result, SCS is incompatible with features
> > that rely on modifying function return addresses in the kernel
> > stack to alter control flow. A copy of the return address is
> > still kept in the kernel stack for compatibility with stack
> > unwinding, for example.
> >
> > SCS has a minimal performance overhead, but allocating
> > shadow stacks increases kernel memory usage. The feature is
> > therefore mostly useful on hardware that lacks support for PAC
> > instructions.
> >
> > Changes in v13:
> >  - Changed thread_info::shadow_call_stack to a base address and
> >    an offset instead, and removed the now unneeded __scs_base()
> >    and scs_save().
> >  - Removed alignment from the kmem_cache and static allocations.
> >  - Removed the task_set_scs() helper function.
> >  - Moved the assembly code for loading and storing the offset in
> >    thread_info to scs_load/save macros.
> >  - Added offset checking to scs_corrupted().
> >  - Switched to cmpxchg_relaxed() in scs_check_usage().
> >
>
> OK, so one thing that came up in an offline discussion about SCS is
> the way it interacts with the vmap'ed stack.
>
> The vmap'ed stack is great for robustness, but it only works if things
> don't explode for other reasons in the mean time. This means the
> ordinary-to-shadow-call-stack size ratio should be chosen such that it
> is *really* unlikely you could ever overflow the shadow call stack and
> corrupt another task's call stack before hitting the vmap stack's
> guard region.
>
> Alternatively, I wonder if there is a way we could let the SCS and
> ordinary stack share the [bottom of] the vmap'ed region. That would
> give rather nasty results if the ordinary stack overflows into the
> SCS, but for cases where we really recurse out of control, we could
> catch this occurrence on either stack, whichever one occurs first. And
> the nastiness -when it does occur- will not corrupt any state beyond
> the stack of the current task.

Hmm, I guess that would make it quite hard to keep the SCS address
secret though :-(