KVM/arm64: Guest ABI changes do not appear rollback-safe

[Andrew Jones]

On Wed, Aug 25, 2021 at 11:14:59AM -0700, Oliver Upton wrote:
> On Wed, Aug 25, 2021 at 8:07 AM Andrew Jones <drjones at redhat.com> wrote:
> >
> > On Wed, Aug 25, 2021 at 11:39:34AM +0100, Marc Zyngier wrote:
> > > On Wed, 25 Aug 2021 11:02:28 +0100,
> > > Oliver Upton <oupton at google.com> wrote:
> > > >
> > > > On Wed, Aug 25, 2021 at 2:27 AM Marc Zyngier <maz at kernel.org> wrote:
> > > > > > Exposing new hypercalls to guests in this manner seems very unsafe to
> > > > > > me. Suppose an operator is trying to upgrade from kernel N to kernel
> > > > > > N+1, which brings in the new 'widget' hypercall. Guests are live
> > > > > > migrated onto the N+1 kernel, but the operator finds a defect that
> > > > > > warrants a kernel rollback. VMs are then migrated from kernel N+1 -> N.
> > > > > > Any guests that discovered the 'widget' hypercall are likely going to
> > > > > > get fussy _very_ quickly on the old kernel.
> > > > >
> > > > > This goes against what we decided to support for the *only* publicly
> > > > > available VMM that cares about save/restore, which is that we only
> > > > > move forward and don't rollback.
> > > >
> > > > Ah, I was definitely missing this context. Current behavior makes much
> > > > more sense then.
> > > >
> > > > > Hypercalls are the least of your
> > > > > worries, and there is a whole range of other architectural features
> > > > > that will have also appeared/disappeared (your own CNTPOFF series is a
> > > > > glaring example of this).
> > > >
> > > > Isn't that a tad bit different though? I'll admit, I'm just as guilty
> > > > with my own series forgetting to add a KVM_CAP (oops), but it is in my
> > > > queue to kick out with the fix for nVHE/ptimer. Nonetheless, if a user
> > > > takes up a new KVM UAPI, it is up to the user to run on a new kernel.
> > >
> > > The two are linked. Exposing a new register to userspace and/or guest
> > > result in the same thing: you can't rollback. That's specially true in
> > > the QEMU case, which *learns* from the kernel what registers are
> > > available, and doesn't maintain a fixed list.
> > >
> > > > My concerns are explicitly with the 'under the nose' changes, where
> > > > KVM modifies the guest feature set without userspace opting in. Based
> > > > on your comment, though, it would appear that other parts of KVM are
> > > > affected too.
> > >
> > > Any new system register that is exposed by a new kernel feature breaks
> > > rollback. And so far, we only consider it a bug if the set of exposed
> > > registers reduces. Anything can be added safely (as checked by one of
> > > the selftests added by Drew).
> > >
> > > < It doesn't have to be rollback safety, either. There may
> > > > simply be a hypercall which an operator doesn't want to give its
> > > > guests, and it needs a way to tell KVM to hide it.
> > >
> > > Fair enough. But this has to be done in a scalable way, which
> > > individual capability cannot provide.
> > >
> > > > > > Have I missed something blatantly obvious, or do others see this as an
> > > > > > issue as well? I'll reply with an example of adding opt-out for PTP.
> > > > > > I'm sure other hypercalls could be handled similarly.
> > > > >
> > > > > Why do we need this? For future hypercalls, we could have some buy-in
> > > > > capabilities. For existing ones, it is too late, and negative features
> > > > > are just too horrible.
> > > >
> > > > Oh, agreed on the nastiness. Lazy hack to realize the intended
> > > > functional change..
> > >
> > > Well, you definitely achieved your goal of attracting my attention :).
> > >
> > > > > For KVM-specific hypercalls, we could get the VMM to save/restore the
> > > > > bitmap of supported functions. That would be "less horrible". This
> > > > > could be implemented using extra "firmware pseudo-registers" such as
> > > > > the ones described in Documentation/virt/kvm/arm/psci.rst.
> > > >
> > > > This seems more reasonable, especially since we do this for migrating
> > > > the guest's PSCI version.
> > > >
> > > > Alternatively, I had thought about using a VM attribute, given the
> > > > fact that it is non-architectural information and we avoid ABI issues
> > > > in KVM_GET_REG_LIST without buy-in through a KVM_CAP.
> > >
> > > The whole point is that these settings get exposed by
> > > KVM_GET_REG_LIST, as this is QEMU's way to dump a VM state. Given that
> > > we already have this for things like the spectre management state, we
> > > can just as well expose the bitmaps that deal with the KVM-specific
> > > hypercalls. After all, this falls into the realm of "KVM as VM
> > > firmware".
> > >
> > > For ARM-architected hypercalls (TRNG, stolen time), we may need a
> > > similar extension.
> > >
> >
> > Thanks for including me Marc. I think you've mentioned all the examples
> > of why we don't generally expect N+1 -> N migrations to work that I
> > can think of. While some of the examples like get-reg-list could
> > eventually be eliminated if we had CPU models to tighten our machine type
> > state, I think N+1 -> N migrations will always be best effort at most.
> >
> > I agree with giving userspace control over the exposer of the hypercalls
> > though. Using pseudo-registers for that purpose rather than a pile of
> > CAPs also seems reasonable to me.
> >
> > And, while I don't think this patch is going to proceed, I thought I'd
> > point out that the opt-out approach doesn't help much with expanding
> > our migration support unless we require the VMM to be upgraded first.
> >
> > And, even then, the (N_kern, N+1_vmm) -> (N+1_kern, N_vmm) case won't
> > work as expected, since the source enforce opt-out, but the destination
> > won't.
> 
> Right, there's going to need to be a fence in both kernel and VMM
> versions. Before the fence, you can't rollback with either component.
> Once on the other side of the fence, the user may freely migrate
> between kernel + VMM combinations.
> 
> > Also, since the VMM doesn't key off the kernel version, for the
> > most part N+1 VMMs won't know when they're supposed to opt-out or not,
> > leaving it to the user to ensure they consider everything. opt-in
> > usually only needs the user to consider what machine type they want to
> > launch.
> 
> Going the register route will implicitly require opt-out for all old
> hypercalls. We exposed them unconditionally to the guest before, and
> we must uphold that behavior. The default value for the bitmap will
> have those features set. Any hypercalls added after that register
> interface will then require explicit opt-in from userspace.
> 
> With regards to the pseudoregister interface, how would a VMM discover
> new bits? From my perspective, you need to have two bitmaps that the
> VMM can get at: the set of supported feature bits and the active
> bitmap of features for a running guest.
>

I think we should model the pseudo-register approach off of x86's
CPUID approach. x86 has specific get/set ioctls for CPUIDs
(KVM_GET/SET_CPUID2), but I think we should get by with just
get/set-one-reg. However, it might be nice/necessary to have something
like x86's KVM_GET_SUPPORTED_CPUID which returns all the registers
at once as a bitmap and the set bits would inform userspace of what's
supported by the hardware and KVM.

So a new ioctl similar to KVM_GET_SUPPORTED_CPUID would be your
first bitmap that shows what's supported and then userpace can
determine what it wants to change and calculate the appropriate
pseudo-registers to set/clear bits on with set-one-reg, as x86
would do with kvm-set-cpuid2.

Thanks,
drew