Re: [REPOST PATCH] arm/arm64: KVM: Add PSCI version selection API

[Christoffer Dall]

On Mon, Apr 09, 2018 at 01:47:50PM +0100, Marc Zyngier wrote:
> +Drew, who's look at the whole save/restore thing extensively
> 
> On 09/04/18 13:30, Christoffer Dall wrote:
> > On Thu, Mar 15, 2018 at 07:26:48PM +0000, Marc Zyngier wrote:
> >> On 15/03/18 19:13, Peter Maydell wrote:
> >>> On 15 March 2018 at 19:00, Marc Zyngier <marc.zyngier@xxxxxxx> wrote:
> >>>> On 06/03/18 09:21, Andrew Jones wrote:
> >>>>> On Mon, Mar 05, 2018 at 04:47:55PM +0000, Peter Maydell wrote:
> >>>>>> On 2 March 2018 at 11:11, Marc Zyngier <marc.zyngier@xxxxxxx> wrote:
> >>>>>>> On Fri, 02 Mar 2018 10:44:48 +0000,
> >>>>>>> Auger Eric wrote:
> >>>>>>>> I understand the get/set is called as part of the migration process.
> >>>>>>>> So my understanding is the benefit of this series is migration fails in
> >>>>>>>> those cases:
> >>>>>>>>
> >>>>>>>>> =0.2 source -> 0.1 destination
> >>>>>>>> 0.1 source -> >=0.2 destination
> >>>>>>>
> >>>>>>> It also fails in the case where you migrate a 1.0 guest to something
> >>>>>>> that cannot support it.
> >>>>>>
> >>>>>> I think it would be useful if we could write out the various
> >>>>>> combinations of source, destination and what we expect/want to
> >>>>>> have happen. My gut feeling here is that we're sacrificing
> >>>>>> exact migration compatibility in favour of having the guest
> >>>>>> automatically get the variant-2 mitigations, but it's not clear
> >>>>>> to me exactly which migration combinations that's intended to
> >>>>>> happen for. Marc?
> >>>>>>
> >>>>>> If this wasn't a mitigation issue the desired behaviour would be
> >>>>>> straightforward:
> >>>>>>  * kernel should default to 0.2 on the basis that
> >>>>>>    that's what it did before
> >>>>>>  * new QEMU version should enable 1.0 by default for virt-2.12
> >>>>>>    and 0.2 for virt-2.11 and earlier
> >>>>>>  * PSCI version info shouldn't appear in migration stream unless
> >>>>>>    it's something other than 0.2
> >>>>>> But that would leave some setups (which?) unnecessarily without the
> >>>>>> mitigation, so we're not doing that. The question is, exactly
> >>>>>> what *are* we aiming for?
> >>>>>
> >>>>> The reason Marc dropped this patch from the series it was first introduced
> >>>>> in was because we didn't have the aim 100% understood. We want the
> >>>>> mitigation by default, but also to have the least chance of migration
> >>>>> failure, and when we must fail (because we're not doing the
> >>>>> straightforward approach listed above, which would prevent failures), then
> >>>>> we want to fail with the least amount of damage to the user.
> >>>>>
> >>>>> I experimented with a couple different approaches and provided tables[1]
> >>>>> with my results. I even recommended an approach, but I may have changed
> >>>>> my mind after reading Marc's follow-up[2]. The thread continues from
> >>>>> there as well with follow-ups from Christoffer, Marc, and myself. Anyway,
> >>>>> Marc did this repost for us to debate it and work out the best approach
> >>>>> here.
> >>>> It doesn't look like we've made much progress on this, which makes me
> >>>> think that we probably don't need anything of the like.
> >>>
> >>> I was waiting for a better explanation from you of what we're trying to
> >>> achieve. If you want to take the "do nothing" approach then a list
> >>> also of what migrations succeed/fail/break in that case would also
> >>> be useful.
> >>>
> >>> (I am somewhat lazily trying to avoid having to spend time reverse
> >>> engineering the "what are we trying to do and what effects are
> >>> we accepting" parts from the patch and the code that's already gone
> >>> into the kernel.)
> >>
> >> OK, let me (re)state the problem:
> >>
> >> For a guest that requests PSCI 0.2 (i.e. all guests from the past 4 or 5
> >> years), we now silently upgrade the PSCI version to 1.0 allowing the new
> >> SMCCC to be discovered, and the ARCH_WORKAROUND_1 service to be called.
> >>
> >> Things get funny, specially with migration (and the way QEMU works).
> >>
> >> If we "do nothing":
> >>
> >> (1) A guest migrating from an "old" host to a "new" host will silently
> >> see its PSCI version upgraded. Not a big deal in my opinion, as 1.0 is a
> >> strict superset of 0.2 (apart from the version number...).
> >>
> >> (2) A guest migrating from a "new" host to an "old" host will silently
> >> loose its Spectre v2 mitigation. That's quite a big deal.
> >>
> >> (3, not related to migration) A guest having a hardcoded knowledge of
> >> PSCI 0.2 will se that we've changed something, and may decide to catch
> >> fire. Oh well.
> >>
> >> If we take this patch:
> >>
> >> (1) still exists
> > 
> > No problem, IMHO.
> > 
> >>
> >> (2) will now fail to migrate. I see this as a feature.
> > 
> > Yes, I agree.  This is actually the most important reason for doing
> > anything beyond what's already merged.
> 
> Indeed, and that's the reason I wrote this patch the first place.
> 
> > 
> >>
> >> (3) can be worked around by setting the "PSCI version pseudo register"
> >> to 0.2.
> > 
> > Nice to have, but we're probably not expecting this to be of major
> > concern.  I initially thought it was a nice debugging feature as well,
> > but that may be a ridiculous point.
> > 
> >>
> >> These are the main things I can think of at the moment.
> > 
> > So I think we we should merge this patch.
> > 
> > If userspace then wants to support "migrate from explicitly set v0.2 new
> > kernel to old kernel", then it must add specific support to filter out
> > the register from the register list; not that I think anyone will need
> > that or bother to implement it.
> > 
> > In other words, I think you should merge this:
> > 
> > Reviewed-by: Christoffer Dall <cdall@xxxxxxxxxx>
> > 
> 
> Thanks. One issue is that we've now missed the 4.16 train, and that this
> effectively is an ABI change (a fairly minor one, but still). Would we
> consider slapping this as a retrospective fix to 4.16-stable, or keep it
> as a 4.17 feature?

Given that it fixes a potentially dangerous migration, and it's a fairly
simple patch, I think it's reasonable to apply as a fix to the next 4.16
release.  Would we be violating any hard-set rules in doing so?

Thanks,
-Christoffer