Failing no-vgic-v3 test

[Sascha Bischoff]

On Mon, 2025-10-27 at 12:37 +0100, Sebastian Ott wrote:
> Hey,
> 
> on an ampere altra I've got a sad selftest:
> 
> [root at virtlab-arm11 kvm]# ./arm64/no-vgic-v3
> Random seed: 0x6b8b4567
> ==== Test Assertion Failure ====
>    arm64/no-vgic-v3.c:66: handled
>    pid=3793 tid=3793 errno=4 - Interrupted system call
>       1  0x0000000000402feb: test_run_vcpu at no-vgic-v3.c:128
>       2  0x000000000040214f: test_guest_no_gicv3 at no-vgic-v3.c:155
> (discriminator 17)
>       3   (inlined by) main at no-vgic-v3.c:174 (discriminator 17)
>       4  0x0000ffff873eb587: ?? ??:0
>       5  0x0000ffff873eb65f: ?? ??:0
>       6  0x00000000004022af: _start at ??:?
>    ICC_PMR_EL1 no read trap
> 
> This is a guest without VGICv3 on GICv3 HW. The test expects UNDEF on
> reg
> access - which is not happening since:
> 
> 3193287ddffb KVM: arm64: gic-v3: Only set ICH_HCR traps for v2-on-v3
> or v3 guests
> 
> As a local fix I've done:
> diff --git a/arch/arm64/kvm/vgic/vgic-v3.c
> b/arch/arm64/kvm/vgic/vgic-v3.c
> index 6fbb4b099855..1fe53a021926 100644
> --- a/arch/arm64/kvm/vgic/vgic-v3.c
> +++ b/arch/arm64/kvm/vgic/vgic-v3.c
> @@ -297,11 +297,15 @@ void vcpu_set_ich_hcr(struct kvm_vcpu *vcpu)
>   {
>   	struct vgic_v3_cpu_if *vgic_v3 = &vcpu-
> >arch.vgic_cpu.vgic_v3;
> 
>   	if (!vgic_is_v3(vcpu->kvm))
>   		return;
> 
>   	/* Hide GICv3 sysreg if necessary */
> -	if (vcpu->kvm->arch.vgic.vgic_model ==
> KVM_DEV_TYPE_ARM_VGIC_V2) {
> +	if (!kvm_has_gicv3(vcpu->kvm) ||
> +	    vcpu->kvm->arch.vgic.vgic_model ==
> KVM_DEV_TYPE_ARM_VGIC_V2) {
>   		vgic_v3->vgic_hcr |= (ICH_HCR_EL2_TALL0 |
> ICH_HCR_EL2_TALL1 |
>   				      ICH_HCR_EL2_TC);
>   		return;
> 
> 
> but following the intention of the patch above maybe we should do
> smth
> like:
> diff --git a/arch/arm64/kvm/vgic/vgic-v3.c
> b/arch/arm64/kvm/vgic/vgic-v3.c
> index 6fbb4b099855..1fe53a021926 100644
> --- a/arch/arm64/kvm/vgic/vgic-v3.c
> +++ b/arch/arm64/kvm/vgic/vgic-v3.c
> @@ -297,11 +297,15 @@ void vcpu_set_ich_hcr(struct kvm_vcpu *vcpu)
>   {
>   	struct vgic_v3_cpu_if *vgic_v3 = &vcpu-
> >arch.vgic_cpu.vgic_v3;
> 
>   	if (!vgic_is_v3(vcpu->kvm))
>   		return;
> 
>   	/* Hide GICv3 sysreg if necessary */
> -	if (vcpu->kvm->arch.vgic.vgic_model ==
> KVM_DEV_TYPE_ARM_VGIC_V2) {
> +	if (kvm_has_feat(vcpu->kvm, ID_AA64PFR0_EL1, GIC, NI) ||
> +	    vcpu->kvm->arch.vgic.vgic_model ==
> KVM_DEV_TYPE_ARM_VGIC_V2) {
>   		vgic_v3->vgic_hcr |= (ICH_HCR_EL2_TALL0 |
> ICH_HCR_EL2_TALL1 |
>   				      ICH_HCR_EL2_TC);
>   		return;
> 
> Thoughts?
> Sebastian
> 

Hi Sebastian.

Apologies - I broke that test by failing to consider that it is
possible to not create an in-kernel vgic. We need to trap all accesses
on a GICv3-compatible host if either we're running a GICv2 guest, or if
we don't have an in-kernel vgic at all (the case I'd missed).

I posted a fix for that the other day:

https://lore.kernel.org/all/20251021094358.1963807-1-sascha.bischoff@arm.com/

Could you please try the fix and confirm if it works for you?

Thanks,
Sascha