On Thu, 2018-01-25 at 10:34 +0100, Peter Zijlstra wrote:
> On Wed, Jan 24, 2018 at 11:43:05AM +0100, Paolo Bonzini wrote:
> > 
> > On 24/01/2018 11:35, Peter Zijlstra wrote:
> > > 
> > > On Tue, Jan 23, 2018 at 08:48:13PM +0000, David Woodhouse wrote:
> > > > 
> > > > On Tue, 2018-01-23 at 21:28 +0100, Borislav Petkov wrote:
> > > > > 
> > > > > 
> > > > > > 
> > > > > >        flags = (flags & EFLAGS_MASK) | X86_EFLAGS_IF;
> > > > > > -     asm("push %[flags]; popf; call *%[fastop]"
> > > > > > -         : "=a"(rc) : [fastop]"r"(fop), [flags]"r"(flags));
> > > > > > +     asm("push %[flags]; popf; " CALL_NOSPEC
> > > > > > +         : "=a"(rc) : [thunk_target]"r"(fop), [flags]"r"(flags));
> > > > > Oh, "thunk_target" is magical.
> > > > You can use THUNK_TARGET(fop), which will be "rm" on 32-bit and avoids
> > > > register starvation in some cases (I don't think the hyperv calls
> > > > worked until I did that).
> > > The reason I didn't use THUNK_TARGET() was exactly because it used "rm"
> > > and the current code did "r" only. I'm happy to change if people can
> > > agree on something ;-)
> > In practice, "fastop" is going to be in a register because of how it's
> > computed, but "rm" is okay.
> OK, so the other occurence in that file uses "+S", which is the SI
> register. That cannot use THUNK_TARGET(), right?
> 
> So do you want one THUNK_TARGET and one open coded, or keep the patch as
> is (both open coded) ?

As long as it builds for i386, you might as well keep them both open-
coded. The "rm" was there mostly for the hyperv call, which ran out of
registers completely when it was just "r".

This patch *really* wants to be going to Linus urgently as a retpoline
fix, and not buried as patch 6/24 though, right? Likewise 7/24? 

Both of them,

Reviewed-by: David Woodhouse <dwmw@amazon.co.uk>

... unless you want me to send them on, but it makes most sense for
Thomas to round them I suspect?