On Tue, 2018-01-09 at 16:39 -0800, Linus Torvalds wrote:
> On Tue, Jan 9, 2018 at 4:31 PM, Andi Kleen <andi@firstfloor.org>
> wrote:
> > 
> > 
> > The following patch fixes it for me. Something doesn't
> > seem to work with ALTERNATIVE_2. It adds only a few bytes
> > more code, so seems acceptable.
> Ugh. It's kind of stupid, though.
> 
> Why is the code sequence not simply:
> 
>   ALTERNATIVE "", "lfence", X86_FEATURE_RETPOLINE_AMD
>   ALTERNATIVE __stringify(jmp *\reg), __stringify(RETPOLINE_JMP
> \reg),
> X86_FEATURE_RETPOLINE
> 
> ie make that X86_FEATURE_RETPOLINE_AMD _only_ emit the "lfence", and
> simply fall through to what will be the "jmp *\reg" of the
> non-RETPOLINE version.
> 
> Then just make sure X86_FEATURE_RETPOLINE_AMD disables
> X86_FEATURE_RETPOLINE.
> 
> That is both simpler and smaller, no?

Not smaller, as the lfence *could* have gone in all the space left by
turning the whole retpoline thing into a single 'jmp'. But I'll
certainly give you simpler. I'll retest and merge Andi's latest patches
for that; thanks.

I'd really like to know what went wrong though. Did we merge Borislav's
attempt to peek at jumps inside alternatives, perchance? Will take a
look...