Re: [PATCH 1/3] KVM: arm64: Restrict FF-A host version renegotiation

[Marc Zyngier <maz@kernel.org>]

On Tue, 06 May 2025 10:29:41 +0100,
Per Larsen <perl@immunant.com> wrote:
> 
> From: Per Larsen <perlarsen@google.com>
> 
> On Fri, May 02, 2025 at 09:47:45AM +0100, Marc Zyngier wrote:
> > On Fri, 02 May 2025 04:52:39 +0100,
> > Per Larsen <perl@immunant.com> wrote:
> > > 
> > > FF-A implementations with the same major version must interoperate with
> > > earlier minor versions per DEN0077A 1.2 REL0 13.2.1 but FF-A version 1.1
> > > broke the ABI on several structures and 1.2 relies on SMCCC 1.2 is not
> > > backwards compatible with SMCCC 1.2 (see DEN0028 1.6 G BET0 Appendix F).
> > > 
> > > If we return the negotiated hypervisor version when the host requests a
> > > lesser minor version, the host will rely on the FF-A interoperability
> > > rules. Since the hypervisor does not currently have the necessary
> > > compatibility paths (e.g. to handle breaking changes to the SMC calling
> > > convention), return NOT_SUPPORTED.
> > > 
> > > Signed-off-by: Per Larsen <perlarsen@google.com>
> > > Signed-off-by: Per Larsen <perl@immunant.com>
> > > ---
> > >  arch/arm64/kvm/hyp/nvhe/ffa.c | 19 ++++++++++++++++++-
> > >  1 file changed, 18 insertions(+), 1 deletion(-)
> > > 
> > > diff --git a/arch/arm64/kvm/hyp/nvhe/ffa.c b/arch/arm64/kvm/hyp/nvhe/ffa.c
> > > index 3369dd0c4009..10e88207b78e 100644
> > > --- a/arch/arm64/kvm/hyp/nvhe/ffa.c
> > > +++ b/arch/arm64/kvm/hyp/nvhe/ffa.c
> > > @@ -712,7 +712,24 @@ static void do_ffa_version(struct arm_smccc_res *res,
> > > 
> > >   hyp_spin_lock(&version_lock);
> > >   if (has_version_negotiated) {
> > > - res->a0 = hyp_ffa_version;
> > > + /*
> > > + * FF-A implementations with the same major version must
> > > + * interoperate with earlier minor versions per DEN0077A 1.2
> > > + * REL0 13.2.1 but FF-A version 1.1 broke the ABI on several
> > > + * structures and 1.2 relies on SMCCC 1.2 is not backwards
> > > + * compatible with SMCCC 1.2 (see DEN0028 1.6 G BET0 Appendix F).
> > 
> > I can't parse this sentence. Missing words?
> > 
> 
> Yes, I will fix this in v2.
> 
> > > + *
> > > + * If we return the negotiated hypervisor version when the host
> > > + * requests a lesser minor version, the host will rely on the
> > > + * aforementioned FF-A interoperability rules. Since the
> > > + * hypervisor does not currently have the necessary compatibility
> > > + * paths (e.g. to paper over the above-mentioned calling
> > > + * convention changes), return NOT_SUPPORTED.
> > > + */
> > > + if (FFA_MINOR_VERSION(ffa_req_version) < FFA_MINOR_VERSION(hyp_ffa_version))
> > > + res->a0 = FFA_RET_NOT_SUPPORTED;
> > > + else
> > > + res->a0 = hyp_ffa_version;
> > >   goto unlock;
> > >   }
> > > 
> > 
> > Something has gone seriously wrong with your email, and the patches
> > are badly mangled and unusable. They are also sent as individual
> > patches and not as a thread, which is a sign that you didn't send them
> > using git. Please fix this for your next posting.
> > 
> 
> Yes, my apologies. I will use git send-email to post v2.
> 
> > More to the meat of the patches: why should the hypervisor paper over
> > anything if the spec is broken? Why can't the host just as well decide
> > for itself what to do?
> > 
> 
> Asssuming we drop this patch from the series and apply the rest, the
> hypervisor and host can negotiate FF-A 1.2. If the host then calls
> FFA_VERSION a second time to request FF-A 1.1, the hypervisor would
> return version 1.2 (without this patch).

Why would it do that? Once a particular version has been negotiated, I
expect to be immutable.

> Per the spec, that means the
> host is can use the compatibility rules (DEN0077A Sec 13.2.1) to go
> ahead and use FF-A 1.1 (every function in 1.A must work in a compatible
> way in 1.B if B>A).

I don't interpret this as "you can switch between versions" after the
initial negotiation.

> However, the hypervisor negotiated version stays at 1.2 so it will use
> SMCCC 1.2 for 64-bit interfaces. The host has no way of knowing this and
> might as well assume that the hypervisor was implemented to fall back to
> SMCCC 1.1 in this particular case. 
> 
> I don't even know that the host will ever try to renegotiate as it is
> explicitly not allowed by the FF-A spec. There is no way for the
> hypervisor to say, "stay at the negotiated version" so we must return
> NOT_SUPPORTED. 

If it is not allowed, why should we do *anything*? And if the host is
broken, let's fix the host rather than adding pointless validation
code to EL2.

Thanks,

	M.

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