Re: Hang with nVHE mode and SME

[Marc Zyngier <maz@kernel.org>]

+ Peter

On Wed, 26 Oct 2022 16:07:21 +0100,
Mark Brown <broonie@kernel.org> wrote:
> 
> On Wed, Oct 26, 2022 at 03:29:44PM +0100, Vincent Donnefort wrote:
> 
> >   static void __activate_traps(struct kvm_vcpu *vcpu)
> >   {
> >       ...
> >       if (cpus_have_final_cap(ARM64_SME)) {
> >            // HANG !
> 
> You're not entirely specific on where the hang is - I assume you mean
> the second SME block and that it's hanging on the first sysreg read in
> there:
> 
> 	if (cpus_have_final_cap(ARM64_SME)) {
> 		u64 val;
> 
> 		val = read_sysreg_s(SYS_HFGRTR_EL2);
> 
> rather than on the if statement?  A brief grep around the qemu source
> suggests that this register and fine grained traps in general are not
> implemented which is invalid with SME since this is required for SME in
> nVHE.  SME is a v9.2 feature, and v9.2 includes all the requirements of
> v8.7.  FEAT_FGT has been mandatory since v8.6 (where it was added).

It is very unfortunate that the SME spec doesn't call out this
explicitly, while it is calling out the dependency on FEAT_HCX (OK,
one is mandatory and the other isn't).

> I did originally have a check in there for either/both this or FEAT_HCX
> but it got taken out during review due to the architecturual
> requirement.

So the question is whether we work around this in the kernel (not
enabling either KVM or SME if FEAT_FGT is not present), or leave it as
is with the hope that QEMU gets updated.

I'm inclined to do the latter. Thoughts anyone?

Thanks,

	M.

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