Re: [RFC] Retpoline: Binary mitigation for branch-target-injection (aka "Spectre")

[Paul Turner]

On Thu, Jan 04, 2018 at 08:18:57AM -0800, Andy Lutomirski wrote:
> On Thu, Jan 4, 2018 at 1:30 AM, Woodhouse, David <dwmw@xxxxxxxxxxxx> wrote:
> > On Thu, 2018-01-04 at 01:10 -0800, Paul Turner wrote:
> >> Apologies for the discombobulation around today's disclosure.  Obviously the
> >> original goal was to communicate this a little more coherently, but the
> >> unscheduled advances in the disclosure disrupted the efforts to pull this
> >> together more cleanly.
> >>
> >> I wanted to open discussion the "retpoline" approach and and define its
> >> requirements so that we can separate the core
> >> details from questions regarding any particular implementation thereof.
> >>
> >> As a starting point, a full write-up describing the approach is available at:
> >>   https://support.google.com/faqs/answer/7625886
> >
> > Note that (ab)using 'ret' in this way is incompatible with CET on
> > upcoming processors. HJ added a -mno-indirect-branch-register option to
> > the latest round of GCC patches, which puts the branch target in a
> > register instead of on the stack. My kernel patches (which I'm about to
> > reconcile with Andi's tweaks and post) do the same.
> >
> > That means that in the cases where at runtime we want to ALTERNATIVE
> > out the retpoline, it just turns back into a bare 'jmp *\reg'.
> >
> >
> 
> I hate to say this, but I think Intel should postpone CET until the
> dust settles.  Intel should also consider a hardware-protected stack
> that is only accessible with PUSH, POP, CALL, RET, and a new MOVSTACK
> instruction.  That, by itself, would give considerable protection.
> But we still need JMP_NO_SPECULATE.  Or, better yet, get the CPU to
> stop leaking data during speculative execution.

Echoing Andy's thoughts, but from a slightly different angle:

1) BTI is worse than the current classes of return attack.  Given this,
   considered as a binary choice, it's equivalent to the current state of the
   world (e.g. no CET).
2) CET will not be "free".  I suspect in its initial revisions it will be more
   valuable for protecting end-users then enterprise workloads (cost is not
   observable for interactive workloads because there's tons of headroom in the
   first place).

While the potential incompatibility is unfortunate; I'm not sure it makes a
significant adoption to the adoption rate of CET.