Re: [PATCH v2.1] x86/retpoline: Fill return stack buffer on vmexit

[Josh Poimboeuf]

On Thu, Jan 11, 2018 at 02:53:57PM +0000, David Woodhouse wrote:
> On Thu, 2018-01-11 at 15:32 +0100, Peter Zijlstra wrote:
> > On Thu, Jan 11, 2018 at 02:28:32PM +0000, David Woodhouse wrote:
> > > 
> > > On Thu, 2018-01-11 at 08:20 -0600, Josh Poimboeuf wrote:
> > > > 
> > > > 
> > > > This seems weird. I liked v1 a lot better. What's the problem with
> > > > patching in the whole thing?
> > > > 
> > > > Also, if you go back to v1, it should be an easy objtool fix, just add
> > > > ANNOTATE_NOSPEC_ALTERNATIVE in front of it.
> > > The objection was that I was patching in a fairly long set of
> > > instructions. I confess I don't actually know why that's a problem,
> >
> > You get a giant string of NOPs, a single jmp should be faster.
> 
> How about this one then (with ANNOTATE_NOSPEC_ALTERNATIVE):
> 
> -ÂÂÂÂÂÂÂasm volatile (ALTERNATIVE("",
> +ÂÂÂÂÂÂÂasm volatile (ALTERNATIVE("jmp " alt_end_marker "f",
> ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ__stringify(__FILL_RETURN_BUFFER(%0, %1, _%=)),
> ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂX86_FEATURE_RETPOLINE)
> ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ: "=r" (dummy), ASM_CALL_CONSTRAINT : : "memory" );

Looks good to me.

Another IRC discussion was that Boris may eventually add a feature to
the alternatives code to automatically insert such a jump if there are a
lot of nops.

-- 
Josh