It doesn't make much sense as a realm guest wouldn't want to trust the
host. It will also need some extra work to ensure that KVM will only
attempt to write into a shared memory region. So for now just disable
it.

Reviewed-by: Suzuki K Poulose <suzuki.poulose@arm.com>
Signed-off-by: Steven Price <steven.price@arm.com>
---
 arch/arm64/kvm/arm.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/arch/arm64/kvm/arm.c b/arch/arm64/kvm/arm.c
index eff1a4ec892b..134acb4ee26f 100644
--- a/arch/arm64/kvm/arm.c
+++ b/arch/arm64/kvm/arm.c
@@ -432,7 +432,10 @@ int kvm_vm_ioctl_check_extension(struct kvm *kvm, long ext)
 		r = system_supports_mte();
 		break;
 	case KVM_CAP_STEAL_TIME:
-		r = kvm_arm_pvtime_supported();
+		if (kvm_is_realm(kvm))
+			r = 0;
+		else
+			r = kvm_arm_pvtime_supported();
 		break;
 	case KVM_CAP_ARM_EL1_32BIT:
 		r = cpus_have_final_cap(ARM64_HAS_32BIT_EL1);
-- 
2.43.0

On 12/13/24 1:55 AM, Steven Price wrote:
> It doesn't make much sense as a realm guest wouldn't want to trust the
> host. It will also need some extra work to ensure that KVM will only
> attempt to write into a shared memory region. So for now just disable
> it.
> 
> Reviewed-by: Suzuki K Poulose <suzuki.poulose@arm.com>
> Signed-off-by: Steven Price <steven.price@arm.com>
> ---
>   arch/arm64/kvm/arm.c | 5 ++++-
>   1 file changed, 4 insertions(+), 1 deletion(-)
> 
> diff --git a/arch/arm64/kvm/arm.c b/arch/arm64/kvm/arm.c
> index eff1a4ec892b..134acb4ee26f 100644
> --- a/arch/arm64/kvm/arm.c
> +++ b/arch/arm64/kvm/arm.c
> @@ -432,7 +432,10 @@ int kvm_vm_ioctl_check_extension(struct kvm *kvm, long ext)
>   		r = system_supports_mte();
>   		break;
>   	case KVM_CAP_STEAL_TIME:
> -		r = kvm_arm_pvtime_supported();
> +		if (kvm_is_realm(kvm))
> +			r = 0;
> +		else
> +			r = kvm_arm_pvtime_supported();
>   		break;

kvm_vm_ioctl_check_extension() can be called on the file descriptor of "/dev/kvm".
'kvm' is NULL and kvm_is_realm() returns false, which is the missed corner case.

>   	case KVM_CAP_ARM_EL1_32BIT:
>   		r = cpus_have_final_cap(ARM64_HAS_32BIT_EL1);

Thanks,
Gavin

On 02/02/2025 02:15, Gavin Shan wrote:
> On 12/13/24 1:55 AM, Steven Price wrote:
>> It doesn't make much sense as a realm guest wouldn't want to trust the
>> host. It will also need some extra work to ensure that KVM will only
>> attempt to write into a shared memory region. So for now just disable
>> it.
>>
>> Reviewed-by: Suzuki K Poulose <suzuki.poulose@arm.com>
>> Signed-off-by: Steven Price <steven.price@arm.com>
>> ---
>>   arch/arm64/kvm/arm.c | 5 ++++-
>>   1 file changed, 4 insertions(+), 1 deletion(-)
>>
>> diff --git a/arch/arm64/kvm/arm.c b/arch/arm64/kvm/arm.c
>> index eff1a4ec892b..134acb4ee26f 100644
>> --- a/arch/arm64/kvm/arm.c
>> +++ b/arch/arm64/kvm/arm.c
>> @@ -432,7 +432,10 @@ int kvm_vm_ioctl_check_extension(struct kvm *kvm,
>> long ext)
>>           r = system_supports_mte();
>>           break;
>>       case KVM_CAP_STEAL_TIME:
>> -        r = kvm_arm_pvtime_supported();
>> +        if (kvm_is_realm(kvm))
>> +            r = 0;
>> +        else
>> +            r = kvm_arm_pvtime_supported();
>>           break;
> 
> kvm_vm_ioctl_check_extension() can be called on the file descriptor of
> "/dev/kvm".
> 'kvm' is NULL and kvm_is_realm() returns false, which is the missed
> corner case.

This is a general problem with checking extensions on the /dev/kvm file
descriptor. In this case the kernel does support stolen time, but it
doesn't support it with a realm guest. Because there's no context to
know whether the query is about realm guests or not there's little the
kernel can do other than report support.

This is the same situation with other extensions that are keyed off
kvm_is_realm(). I'm not sure what we can do other than say that VMMs
shouldn't do that.

Steve

>>       case KVM_CAP_ARM_EL1_32BIT:
>>           r = cpus_have_final_cap(ARM64_HAS_32BIT_EL1);
> 
> Thanks,
> Gavin
>