Hi Oliver,
On Fri, May 26, 2023 at 10:16:51PM +0000, Oliver Upton wrote:
> The 6.4 kernel picks up support for a generalized SMCCC filter, allowing
> userspace to select hypercall ranges that should be forwarded to
> userspace. This is a shameless attempt of making future SMCCC interfaces
> the responsibility of userspace :)
>
> As a starting point, let's move PSCI up into userspace. KVM already
> leans on userspace for handling calls that have a system-wide effect.
>
> Tested on linux-next with a 64 vCPU VM. Additionally, I took a stab at
> running kvm-unit-test's psci test, which passes.
I applied these patches to kvmtool and tested 6.4, however they don't work
with an SVE enabled kernel/device.
# ./lkvm run -c 2 -m 4 -k psci.flat
INFO: psci: PSCI version 1.0
PASS: psci: invalid-function
PASS: psci: affinity-info-on
PASS: psci: affinity-info-off
Error: KVM_ARM_VCPU_FINALIZE: Operation not permitted
Fatal: Unable to configure requested vcpu features
`kvm_cpu__configure_features` in kvmtool is failing because Linux returns an
error if SVE was already finalised (arch/arm64/kvm/reset.c):
```
int kvm_arm_vcpu_finalize(struct kvm_vcpu *vcpu, int feature)
{
switch (feature) {
case KVM_ARM_VCPU_SVE:
if (!vcpu_has_sve(vcpu))
return -EINVAL;
if (kvm_arm_vcpu_sve_finalized(vcpu))
return -EPERM; // <---- returns here
return kvm_vcpu_finalize_sve(vcpu);
}
return -EINVAL;
}
```
It's not immediately obvious to me why finalising SVE twice is an error.
Changing that to `return 0;` gets the test passing, but not sure if there
are other implications.
I also booted with `arm64.nosve` and the test passed.
Thanks,
Joey
