Re: [PATCH v18 20/23] KVM: arm64: CCA: Don't expose unsupported capabilities for realm guests

[Fuad Tabba]

Hi Suzuki,

On Tue, 15 Sep 2026 17:01:38 +0100, Suzuki K Poulose
<suzuki.poulose@xxxxxxx> wrote:
[...]
> diff --git a/arch/arm64/kvm/arm.c b/arch/arm64/kvm/arm.c
[...]
> +static inline bool kvm_realm_ext_allowed(long ext)
> +{
> +        switch (ext) {
> +        case KVM_CAP_ARM_PSCI:
> +        case KVM_CAP_ARM_PSCI_0_2:
> +        case KVM_CAP_NR_VCPUS:
> +        case KVM_CAP_MAX_VCPUS:
> +        case KVM_CAP_MAX_VCPU_ID:
> +        case KVM_CAP_MSI_DEVID:
> +        case KVM_CAP_ARM_VM_IPA_SIZE:
> +        case KVM_CAP_ARM_SVE:
> +        case KVM_CAP_ONE_REG:
> +        case KVM_CAP_ARM_PTRAUTH_ADDRESS:
> +        case KVM_CAP_ARM_PTRAUTH_GENERIC:
> +        case KVM_CAP_SYNC_MMU:
> +                return true;
> +        }
> +        return false;
> +}

I'd keep KVM_CAP_IRQCHIP in this list. api.rst has it advertising
KVM_IRQ_LINE as well as KVM_CREATE_IRQCHIP, so the cap means in-kernel
irqchip rather than GICv2, and the GICv2 case is already rejected by
patch 18 in kvm_vgic_create(). It also keeps the two protected lists
the same where the support is the same. And could this live in
kvm_rmi.h, next to where kvm_pkvm_ext_allowed() lives in kvm_pkvm.h,
rather than in arm.c?

[...]
> +static inline bool kvm_arch_vm_ioctl_allowed(struct kvm *kvm, unsigned int ioctl)
> +{
> +        long ext;
> +        int r;
> +
> +        r = kvm_get_cap_for_kvm_ioctl(ioctl, &ext);
> +        if (WARN_ON_ONCE(r < 0))
> +                return false;

You told Sashiko my fix covers this one [1], but that fix changes
kvm_pkvm_ioctl_allowed(), which this patch deletes, so the `if (r <
0)` has to go into this copy too. The guard you posted for the second
finding still leaves the WARN reachable on pKVM and Realm hosts, which
is the case the fix removed.

nit: kvm_arch_ is the prefix generic KVM uses for the hooks it calls
into the arch. For two static helpers in arm.c, kvm_vm_ext_allowed()
and kvm_vm_ioctl_allowed() would avoid the collision, and the same for
kvm_arch_vm_ext_allowed() above.

Cheers,
/fuad

[1] https://lore.kernel.org/all/20260914093838.1082637-1-fuad.tabba@xxxxxxxxx/