[PATCH v18 01/23] KVM: arm64: protected VM: Handle set_one_reg CNTVCT_EL0/CNTPCT_EL0

Suzuki K Poulose suzuki.poulose at arm.com
Tue Sep 15 14:20:36 PDT 2026


On 15/09/2026 18:48, Suzuki K Poulose wrote:
> On 15/09/2026 17:46, Marc Zyngier wrote:
>> On Tue, 15 Sep 2026 17:01:19 +0100,
>> Suzuki K Poulose <suzuki.poulose at arm.com> wrote:
>>>
>>> Protected VMs doesn't allow setting offsets for virtual and phyiscal
>>> counters, as the offset is always fixed to 0. The VM ioclt is filtered
>>> out based on the cap. However we don't prevent the userspace from trying
>>> to write to the CNTVCT/CNTPCT registers. This would lead to KVM 
>>> triggering
>>> a WARN() in timer_set_offset() as the vm_offset pointer is set to NULL.
>>>
>>> Fix this by always "fixing" the timer offsets to 0 and marking that the
>>> timer offset is set in the kvm->arch.flags at pKVM init time. The
>>> userspace cannot use the KVM_ARM_SET_COUNTER_OFFSET, as it is blocked 
>>> for a
>>> protected VM.
>>>
>>> A userspace writing to the SYS_CNT*CT would observe success, without
>>> any real effect. This was chosen over preventing the writes to these
>>> registers and returning -EPERM.
>>>
>>> With that, we always have a valid vm_offset pointer, remove the 
>>> checks for
>>> vm_offset == NULL.
>>>
>>> Reported by Sashiko here
>>> https://lore.kernel.org/all/20260908164641.416911F00A3A@smtp.kernel.org
>>>
>>> Fixes: f7d05ee84a6a ("KVM: arm64: Prevent host from managing timer 
>>> offsets for protected VMs")
>>> Suggested-by: Marc Zyngier <maz at krenel.org>
>>> Signed-off-by: Suzuki K Poulose <suzuki.poulose at arm.com>
>>> ---
>>>   arch/arm64/kvm/arch_timer.c    | 15 +++++----------
>>>   arch/arm64/kvm/arm.c           | 15 +++++++++++++++
>>>   arch/arm64/kvm/hyp/nvhe/pkvm.c | 26 +++++++++++++-------------
>>>   include/kvm/arm_arch_timer.h   |  3 +--
>>>   4 files changed, 34 insertions(+), 25 deletions(-)
>>>
>>> diff --git a/arch/arm64/kvm/arch_timer.c b/arch/arm64/kvm/arch_timer.c
>>> index 6ac3321f4c575..dda020da4c9c7 100644
>>> --- a/arch/arm64/kvm/arch_timer.c
>>> +++ b/arch/arm64/kvm/arch_timer.c
>>> @@ -1079,14 +1079,10 @@ static void timer_context_init(struct 
>>> kvm_vcpu *vcpu, int timerid)
>>>       ctxt->timer_id = timerid;
>>> -    if (!kvm_vm_is_protected(vcpu->kvm)) {
>>> -        if (timerid == TIMER_VTIMER)
>>> -            ctxt->offset.vm_offset = &kvm->arch.timer_data.voffset;
>>> -        else
>>> -            ctxt->offset.vm_offset = &kvm->arch.timer_data.poffset;
>>> -    } else {
>>> -        ctxt->offset.vm_offset = NULL;
>>> -    }
>>> +    if (timerid == TIMER_VTIMER)
>>> +        ctxt->offset.vm_offset = &kvm->arch.timer_data.voffset;
>>> +    else
>>> +        ctxt->offset.vm_offset = &kvm->arch.timer_data.poffset;
>>>       hrtimer_setup(&ctxt->hrtimer, kvm_hrtimer_expire, 
>>> CLOCK_MONOTONIC, HRTIMER_MODE_ABS_HARD);
>>> @@ -1110,8 +1106,7 @@ void kvm_timer_vcpu_init(struct kvm_vcpu *vcpu)
>>>           timer_context_init(vcpu, i);
>>>       /* Synchronize offsets across timers of a VM if not already 
>>> provided */
>>> -    if (!vcpu_is_protected(vcpu) &&
>>> -        !test_bit(KVM_ARCH_FLAG_VM_COUNTER_OFFSET, &vcpu->kvm- 
>>> >arch.flags)) {
>>> +    if (!test_bit(KVM_ARCH_FLAG_VM_COUNTER_OFFSET, &vcpu->kvm- 
>>> >arch.flags)) {
>>>           timer_set_offset(vcpu_vtimer(vcpu), kvm_phys_timer_read());
>>>           timer_set_offset(vcpu_ptimer(vcpu), 0);
>>>       }
>>> diff --git a/arch/arm64/kvm/arm.c b/arch/arm64/kvm/arm.c
>>> index 8b080804bc90b..7c88508cac8a1 100644
>>> --- a/arch/arm64/kvm/arm.c
>>> +++ b/arch/arm64/kvm/arm.c
>>> @@ -214,6 +214,20 @@ static int kvm_arm_default_max_vcpus(void)
>>>       return vgic_present ? kvm_vgic_get_max_vcpus() : KVM_MAX_VCPUS;
>>>   }
>>> +/*
>>> + * Fix the counter offset to 0 for Protected VMs and mark the
>>> + * offset flag. The user can't set the offset via 
>>> KVM_ARM_SET_COUNTER_OFFSET.
>>> + */
>>> +static void kvm_arch_fix_timer_offsets(struct kvm *kvm)
>>> +{
>>> +    if (!kvm_vm_is_protected(kvm))
>>> +        return;
>>> +
>>> +    /* Fix the counter offset to 0 and mark the offset initialised */
>>> +    kvm->arch.timer_data.poffset = kvm->arch.timer_data.voffset = 0;
>>> +    set_bit(KVM_ARCH_FLAG_VM_COUNTER_OFFSET, &kvm->arch.flags);
>>> +}
>>> +
>>>   /**
>>>    * kvm_arch_init_vm - initializes a VM data structure
>>>    * @kvm:    pointer to the KVM struct
>>> @@ -267,6 +281,7 @@ int kvm_arch_init_vm(struct kvm *kvm, unsigned 
>>> long type)
>>>       kvm_vgic_early_init(kvm);
>>> +    kvm_arch_fix_timer_offsets(kvm);
>>>       kvm_timer_init_vm(kvm);
>>
>> This should all be moved to the timer code.
>>
>>>       /* The maximum number of VCPUs is limited by the host's GIC 
>>> model */
>>> diff --git a/arch/arm64/kvm/hyp/nvhe/pkvm.c b/arch/arm64/kvm/hyp/ 
>>> nvhe/pkvm.c
>>> index 459bd9eb7e4bc..e7b38eff63bd1 100644
>>> --- a/arch/arm64/kvm/hyp/nvhe/pkvm.c
>>> +++ b/arch/arm64/kvm/hyp/nvhe/pkvm.c
>>> @@ -528,19 +528,19 @@ static int init_pkvm_hyp_vcpu(struct 
>>> pkvm_hyp_vcpu *hyp_vcpu,
>>>       hyp_vcpu->vcpu.arch.cflags = READ_ONCE(host_vcpu->arch.cflags);
>>>       hyp_vcpu->vcpu.arch.mp_state.mp_state = KVM_MP_STATE_STOPPED;
>>> -    if (!pkvm_hyp_vcpu_is_protected(hyp_vcpu)) {
>>> -        /*
>>> -         * Timer offsets are pointing to the untrusted KVM copy,
>>> -         * which is pinned in __pkvm_init_vm() for the VM life time.
>>> -         * It is worth noting that hyp_vm->host_kvm points to an EL2
>>> -         * linear map address and timer_get_offset() will use
>>> -         * kern_hyp_va() which is safe as it is idempotent.
>>> -         */
>>> -        vcpu_vtimer(&hyp_vcpu->vcpu)->offset.vm_offset =
>>> -            &hyp_vm->host_kvm->arch.timer_data.voffset;
>>> -        vcpu_ptimer(&hyp_vcpu->vcpu)->offset.vm_offset =
>>> -            &hyp_vm->host_kvm->arch.timer_data.poffset;
>>> -    }
>>> +    /*
>>> +     * Timer offsets are pointing to the untrusted KVM copy,
>>> +     * which is pinned in __pkvm_init_vm() for the VM life time.
>>> +     * It is worth noting that hyp_vm->host_kvm points to an EL2
>>> +     * linear map address and timer_get_offset() will use
>>> +     * kern_hyp_va() which is safe as it is idempotent.
>>> +     * Also for protected VMs the offset is fixed to 0 and is prevented
>>> +     * from changing.
>>> +     */
>>> +    vcpu_vtimer(&hyp_vcpu->vcpu)->offset.vm_offset =
>>> +        &hyp_vm->host_kvm->arch.timer_data.voffset;
>>> +    vcpu_ptimer(&hyp_vcpu->vcpu)->offset.vm_offset =
>>> +        &hyp_vm->host_kvm->arch.timer_data.poffset;
>>
>> I don't think this is right. Protected guests have no offset, and this
>> needs to be ensured by the hypervisor. Here, the host can change the
>> offset any time it wants, and that's not acceptable.
> 
> Ah, you're right. :facepalm:
> 
>>
>>>       ret = pkvm_vcpu_init_sysregs(hyp_vcpu);
>>>       if (ret)
>>> diff --git a/include/kvm/arm_arch_timer.h b/include/kvm/arm_arch_timer.h
>>> index bc6f2fdd7ad33..4f0aa3bb69f45 100644
>>> --- a/include/kvm/arm_arch_timer.h
>>> +++ b/include/kvm/arm_arch_timer.h
>>> @@ -176,8 +176,7 @@ static inline bool has_cntpoff(void)
>>>           if (__ctxt) {                        \
>>>               struct arch_timer_offset *ato = &__ctxt->offset;\
>>>                                       \
>>> -            if (ato->vm_offset)                \
>>> -                off += *KERN_HYP_VA(ato->vm_offset);    \
>>> +            off += *KERN_HYP_VA(ato->vm_offset);        \
>>>               if (ato->vcpu_offset)                \
>>>                   off += *KERN_HYP_VA(ato->vcpu_offset);    \
>>>           }                            \
>>
>> And as you drop the previous hunk, this also needs to be restored to
>> its original state.
> 
> Ack.
> 
> Suzuki


This is how it looks like now, reworked.


diff --git a/arch/arm64/kvm/arch_timer.c b/arch/arm64/kvm/arch_timer.c
index 6ac3321f4c575..1c8fac222a643 100644
--- a/arch/arm64/kvm/arch_timer.c
+++ b/arch/arm64/kvm/arch_timer.c
@@ -1110,8 +1110,7 @@ void kvm_timer_vcpu_init(struct kvm_vcpu *vcpu)
                 timer_context_init(vcpu, i);

         /* Synchronize offsets across timers of a VM if not already 
provided */
-       if (!vcpu_is_protected(vcpu) &&
-           !test_bit(KVM_ARCH_FLAG_VM_COUNTER_OFFSET, 
&vcpu->kvm->arch.flags)) {
+       if (!test_bit(KVM_ARCH_FLAG_VM_COUNTER_OFFSET, 
&vcpu->kvm->arch.flags)) {
                 timer_set_offset(vcpu_vtimer(vcpu), kvm_phys_timer_read());
                 timer_set_offset(vcpu_ptimer(vcpu), 0);
         }
@@ -1133,6 +1132,15 @@ void kvm_timer_init_vm(struct kvm *kvm)
          */
         for (int i = 0; i < NR_KVM_TIMERS; i++)
                 kvm->arch.timer_data.ppi[i] = get_vgic_ppi(kvm, 
default_ppi[i]);
+
+       /*
+        * For the protected VMs, fix the counter offset to 0 by marking the
+        * KVM_ARCH_FLAG_VM_COUNTER_OFFSET. This would also prevent 
userspace
+        * writes to CNTVCT_ELO & CNTPCT_EL0.
+        * The user can't set the offset via vm ioctl 
KVM_ARM_SET_COUNTER_OFFSET.
+        */
+       if (kvm_vm_is_protected(kvm))
+               set_bit(KVM_ARCH_FLAG_VM_COUNTER_OFFSET, &kvm->arch.flags);
  }

  void kvm_timer_cpu_up(void)





More information about the linux-arm-kernel mailing list