[PATCH] KVM: arm64: Prevent FWD_TO_USER of SMCCC to pKVM

[Pierre-Clément Tosi]

Hi Marc,

Thanks for your quick review!

On Mon, Dec 01, 2025 at 06:48:48PM +0000, Marc Zyngier wrote:
> On Mon, 01 Dec 2025 18:19:52 +0000,
> "=?utf-8?q?Pierre-Cl=C3=A9ment_Tosi?=" <ptosi at google.com> wrote:
> > 
> > With protected VMs, forwarding guest HVC/SMCs happens at two interfaces:
> > 
> >      pKVM [EL2] <--> KVM [EL1] <--> VMM [EL0]
> > 
> > so it might be possible for EL0 to successfully register with EL1 to
> > handle guest SMCCC calls but never see the KVM_EXIT_HYPERCALL, even if
> > the calls are properly issued by the guest, due to EL2 handling them so
> > that (host) EL1 never gets a chance to exit to EL0.
> > 
> > Instead, avoid that confusing situation and make userspace fail early by
> > disallowing KVM_ARM_VM_SMCCC_FILTER-ing calls from protected guests in
> > the KVM FID range (which pKVM re-uses).
> > 
> > DEN0028 defines 65536 "Vendor Specific Hypervisor Service Calls":
> > 
> > - the first ARM_SMCCC_KVM_NUM_FUNCS (128) can be custom-defined
> > - the following 3 are currently standardized
> > - the rest is "reserved for future expansion"
> > 
> > so reserve them all, like commit 821d935c87bc ("KVM: arm64: Introduce
> > support for userspace SMCCC filtering") with the Arm Architecture Calls.
> 
> I don't think preventing all hypercalls from reaching userspace is
> acceptable from an API perspective. For example, it is highly expected
> that the hypercall that exposes the various MIDR/REVIDR/AIDR that the
> guest can be expected to run on is handled in userspace.
> 
> Given that this hypercall is critical to the correct behaviour of a
> guest in an asymmetric system, you can't really forbid it. If you
> don't want it, that's fine -- don't implement it in your VMM.
> 
> But I fully expect pKVM to inherit the existing APIs by virtue of
> being a KVM backend.
> 
> > Alternatively, we could have only reserved the ARM_SMCCC_KVM_NUM_FUNCS
> > (or even a subset of it) and the "Call UID Query" but that would have
> > risked future conflicts between that uAPI and an extension of the SMCCC
> > or of the pKVM ABI.
> 
> I disagree. The only ones you can legitimately block are the ones that
> are earmarked for pKVM itself (2-63), and only these. Everything else
> should make it to userspace if the guest and the VMM agree to do so.

Sounds good, I'll limit v2 to these FIDs and the "Call UID Query" (unavoidable).

> 
> This is part of the KVM ABI, and pKVM should be fixed.

I agree. In particular on not restricting {S,G}ET_ONE_REG to a point where the
KVM uAPI for SMCCC filtering can't be supported, as some experimental Linux
forks might [1] have done! I couldn't find the corresponding patches on LKML to
contribute this feedback to, so something to keep in mind when they come in :)

[1]: https://r.android.com/q/I1c7bc93ebe0bfb597cca5c4284ceb7fd53e4713c

> 
> Thanks,
> 
> 	M.
> 
> -- 
> Without deviation from the norm, progress is not possible.

Thanks!

-- 
Pierre