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

[Suzuki K Poulose <suzuki.poulose@arm.com>]

On 15/09/2026 18:55, Suzuki K Poulose wrote:
> On 15/09/2026 17:43, sashiko-bot@kernel.org wrote:
>> Thank you for your contribution! Sashiko AI review found 1 potential 
>> issue(s) to consider:
>> - [High] kvm: host Denial of Service via WARN_ON_ONCE in 
>> kvm_arch_vm_ioctl_allowed()
>> -- 
>>
>> commit f9fdb12a2363dfa41ace289596ec2d2bd95cde48
>> Author: Suzuki K Poulose <suzuki.poulose@arm.com>
>> KVM: arm64: CCA: Don't expose unsupported capabilities for realm guests
>>
>> This commit limits the capabilities allowed for Realm VMs and blocks
>> vm_ioctls backed by unsupported capabilities. It introduces generic
>> capability and ioctl filtering helpers kvm_arch_vm_ext_allowed() and
>> kvm_arch_vm_ioctl_allowed() to support both pKVM and Realm guests.
>>
>>> diff --git a/arch/arm64/kvm/arm.c b/arch/arm64/kvm/arm.c
>>> --- a/arch/arm64/kvm/arm.c
>>> +++ b/arch/arm64/kvm/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;
>>
>> [Severity: High]
>> Could this allow an unprivileged host userspace process to crash the 
>> host?
> 
> This is addressed by a patch from Fuad, hence I didn't try to fix that
> here.
> 
> https://lore.kernel.org/all/20260914093838.1082637-1-fuad.tabba@linux.dev
> 
> 
>>
>> When kvm_arch_vm_ioctl() receives an unrecognized or garbage ioctl,
>> kvm_get_cap_for_kvm_ioctl() returns a negative value. This evaluates 
>> to true
>> and triggers the WARN_ON_ONCE() here. On systems configured with 
>> panic_on_warn
>> (which is common in production and Android environments), this results 
>> in a
>> kernel panic and a complete host denial of service.
>>
>> [ ... ]
>>> @@ -2150,7 +2202,7 @@ int kvm_arch_vm_ioctl(struct file *filp, 
>>> unsigned int ioctl, unsigned long arg)
>>>       void __user *argp = (void __user *)arg;
>>>       struct kvm_device_attr attr;
>>> -    if (is_protected_kvm_enabled() && !kvm_pkvm_ioctl_allowed(kvm, 
>>> ioctl))
>>> +    if (!kvm_arch_vm_ioctl_allowed(kvm, ioctl))
>>>           return -EINVAL;
>>
>> [Severity: High]
>> Does removing the is_protected_kvm_enabled() check expose the 
>> WARN_ON_ONCE()
>> above to all ARM64 KVM deployments?
> 
> Thats right, I will add the following check in the 
> kvm_arch_vm_ioctl_allowed()
> 
> if (!kvm || !kvm_vm_is_hyp_distrusting(kvm))
>      return true;

Thats not correct. It should be :

if (!is_protected_kvm_enabled() &&
     (!kvm || !kvm_is_realm(kvm)))
	return true;

Cheers
Suzuki