Re: [PATCH 1/2][4.15?] x86/shadow: suppress "fast fault path" optimization when running virtualized

[Ian Jackson <iwj@xxxxxxxxxxxxxx>]

Andrew Cooper writes ("Re: [PATCH 1/2][4.15?] x86/shadow: suppress "fast fault 
path" optimization when running virtualized"):
> On 05/03/2021 16:40, Ian Jackson wrote:
> > Andrew Cooper writes ("Re: [PATCH 1/2][4.15?] x86/shadow: suppress "fast 
> > fault path" optimization when running virtualized"):
> >> This wants backporting to stable releases, so I would recommend for 4.15
> >> even at this point.
> > Can someone explain to me the implications of not taking these patch,
> > and the risks of taking them ?
> >
> > AFIACT the implications of not taking 1/ are that we would misbehave
> > in a security relevant way, sometimes, when we are running under
> > another hypervisor ?
> 
> Correct.  Specifically if you've got a migration pool containing an
> IceLake server and something older.
> 
> > As to the risks, 1/ looks obviously correct even to me.
> 
> I agree, although Tim has the deciding maintainer vote.

Right, well, for patch 1 then

Release-Acked-by: Ian Jackson <iwj@xxxxxxxxxxxxxx>

> > And the implications of not taking 2/ is a performance problem ?
> 
> Correct (I believe).
> 
> > 2/ seems complex.  What would go wrong if there were a misplaced ) or
> > confused bit-twiddling or something ?
> 
> The bit twiddling can be independency checked by disassembling the binary.
> 
> However, I have some concerns with the patch as-is, in relation to L1TF
> / XSA-273.

I'm going to hold off on this for now.  I think to give it a
release-ack I would want someone to argue the case.  Concerns would
include Andy's comments (which I saw earlier but do not fully
understand) and me wanting to to know (i) how bad is the perf impact
without it (ii) how has this bit-twiddling been checked.

I hope that makes sense.

Thanks,
Ian.