Re: [PATCH 1/3] Revert "x86/msr: drop compatibility #GP handling in guest_{rd,wr}msr()"

[Roger Pau Monné <roger.pau@xxxxxxxxxx>]

On Wed, Mar 17, 2021 at 02:46:20PM +0000, Ian Jackson wrote:
> Andrew Cooper writes ("Re: [PATCH 1/3] Revert "x86/msr: drop compatibility 
> #GP handling in guest_{rd,wr}msr()""):
> > On 17/03/2021 13:37, Ian Jackson wrote:
> > > AFAICT there is no explanation for why patches 1/ and 2/ deserve to go
> > > into 4.15.
> 
> I see now, rereading the thread, that there was a sentence about this
> in each patch betwen the commit message and the diff.  Sorry for
> missing that.  (Although TBH at least one of those messages could
> usefully have gone into the commit message, as useful motivational
> background.)
> 
> > >   We are late in the freeze now, so I would ideally be
> > > looking for a clear and compelling argument.  I'd also like to
> > > understand what the risks are of taking these.  Can someone please
> > > enlighten me ?
> > 
> > To make the code in 4.15 match 4.14, so patch 3 can be written in the
> > first place.
> > 
> > Also, as a side benefit, patches 1 and 2 reduce the quantity of logspew
> > from the impacted MSRs.
> > 
> > We cannot simply take patch 3 as-is, and say "4.14 and earlier" for
> > backport, because that still forces end users to specify msr_relaxed to
> > unbreak their Solaris guests, which is usability regression vs 4.14
> 
> This is plausible and going in the right direction but I still feel
> uncertain.
> 
> Jan, what is your summary opinion about patch 3 ?
> 
> Roger, can I get your opinion about the possible downside risks of
> this patch ?

For patches 1 and 2 the risk is low I think. This is already the same
handling that we do in pre-4.15, so it's unlikely to cause issues.
>From a guests PoV they don't change the result of trying to access any
of the modified MSRs, accessing them will still result in a #GP being
injected to the guest.

The main risk for patch 3 would be that reporting 0 for
MSR_RAPL_POWER_UNIT would result in some kind of issue on certain
guests, or that it triggers the poking of other MSRs in the
expectation that they would be available. I think those are quite
unlikely, and the patch fixes a real issue with Solaris guests.

Roger.