Re: [tip:x86/pti] x86/retpoline: Fill return stack buffer on vmexit

[Josh Poimboeuf]

On Thu, Jan 25, 2018 at 09:05:37AM -0800, Andy Lutomirski wrote:
> On Thu, Jan 25, 2018 at 9:00 AM, David Woodhouse <dwmw2@xxxxxxxxxxxxx> wrote:
> > On Thu, 2018-01-25 at 10:56 -0600, Josh Poimboeuf wrote:
> >> On Thu, Jan 25, 2018 at 04:03:18PM +0000, David Woodhouse wrote:
> >> > On Thu, 2018-01-25 at 16:51 +0100, Borislav Petkov wrote:
> >> > >
> >> > > > And the seg fault is objtool's way of telling you you need a
> >> > > > ANNOTATE_NOSPEC_ALTERNATIVE above the alternative ;-)
> >> > >
> >> > > Except that it blew up when I did this which doesn't have ALTERNATIVE
> >> > > (it's the diff I saved :-))
> >> >
> >> > Yeah, ANNOTATE_NOSPEC_ALTERNATIVE just tells objtool "don't look at the
> >> > alternative; you're not going to like it".
> >> >
> >> > If you start putting a __fill_rsb() function out of line somewhere and
> >> > only *calling* it from alternatives, then objtool is going to shit
> >> > itself when it sees that function, regardless.
> >>
> >> Right, if you *really* want it always inline, the short term solution is
> >> to just patch it in with X86_FEATURE_ALWAYS.
> >
> > And the whole problem here is that patching it in with alternatives is
> > painful on kernels < 4.1 because back then, we didn't cope with
> > oldinstr and altinstr being different lengths.
> >
> > And they don't want to fix *that* because kABI...
> >
> > I just stopped caring.
> 
> Screw kABI.

There are *many* real world users who depend on kABI, so it's a fact of
life with no better known solution for the given constraints.

If you have a better idea then I suggest you build your own enterprise
distro.

> Distros that use retpolines need their driver vendors to recompile no
> matter what.  Distros that use IBRS and refuse to use retpolines
> should get put on a list of "didn't actually adequately mitigate
> spectre".

Retpolines don't need to break kABI.  We can just detect when they're
missing and report it to the user.

And I don't think there's anybody *refusing* to use retpolines, is
there?  But they're still not fully baked, especially for Skylake+.

IBRS was just easier to implement out of the gate.  And it mitigates
Spectre just fine, but it's too slow to stick with long term.

-- 
Josh