[PATCH] KVM: nVMX: pass valid guest linear-address to the L1

[PATCH] KVM: nVMX: pass valid guest linear-address to the L1

[Jan Dakinevich]

If EPT support is exposed to L1 hypervisor, guest linear-address field
of VMCS should contain GVA of L2, the access to which caused EPT violation.

Signed-off-by: Jan Dakinevich <jan.dakinevich@gmail.com>
---
 arch/x86/kvm/vmx.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/arch/x86/kvm/vmx.c b/arch/x86/kvm/vmx.c
index 5cede40..a4bb2bd 100644
--- a/arch/x86/kvm/vmx.c
+++ b/arch/x86/kvm/vmx.c
@@ -10500,6 +10500,9 @@ static void prepare_vmcs12(struct kvm_vcpu *vcpu, struct vmcs12 *vmcs12,
 		vmcs12->guest_pdptr3 = vmcs_read64(GUEST_PDPTR3);
 	}
 
+	if (nested_cpu_has_ept(vmcs12))
+		vmcs12->guest_linear_address = vmcs_readl(GUEST_LINEAR_ADDRESS);
+
 	if (nested_cpu_has_vid(vmcs12))
 		vmcs12->guest_intr_status = vmcs_read16(GUEST_INTR_STATUS);
 
-- 
1.9.1

Re: [PATCH] KVM: nVMX: pass valid guest linear-address to the L1

[Wanpeng Li]

2016-09-05 2:22 GMT+08:00 Jan Dakinevich <jan.dakinevich@gmail.com>:
> If EPT support is exposed to L1 hypervisor, guest linear-address field
> of VMCS should contain GVA of L2, the access to which caused EPT violation.
>
> Signed-off-by: Jan Dakinevich <jan.dakinevich@gmail.com>
> ---
>  arch/x86/kvm/vmx.c | 3 +++
>  1 file changed, 3 insertions(+)
>
> diff --git a/arch/x86/kvm/vmx.c b/arch/x86/kvm/vmx.c
> index 5cede40..a4bb2bd 100644
> --- a/arch/x86/kvm/vmx.c
> +++ b/arch/x86/kvm/vmx.c
> @@ -10500,6 +10500,9 @@ static void prepare_vmcs12(struct kvm_vcpu *vcpu, struct vmcs12 *vmcs12,
>                 vmcs12->guest_pdptr3 = vmcs_read64(GUEST_PDPTR3);
>         }
>
> +       if (nested_cpu_has_ept(vmcs12))
> +               vmcs12->guest_linear_address = vmcs_readl(GUEST_LINEAR_ADDRESS);
> +

No, nested_ept_inject_page_fault() will set
vmcs12->guest_linear_address after L0 walks L1's EPT page table and
finds that the mapping is invalid if nested EPT is enabled.
prepare_vmcs12() just copies the vmcs field that could have changed by
the L2 guest or the exit-information etc instead of all fields since
other fields are modified by L1 with VMWRITE, which already writes to
vmcs12 directly.

Regards,
Wanpeng Li

Re: [PATCH] KVM: nVMX: pass valid guest linear-address to the L1

[Jan Dakinevich]

On 09/05/2016 03:49 AM, Wanpeng Li wrote:
> 2016-09-05 2:22 GMT+08:00 Jan Dakinevich <jan.dakinevich@gmail.com>:
>> If EPT support is exposed to L1 hypervisor, guest linear-address field
>> of VMCS should contain GVA of L2, the access to which caused EPT violation.
>>
>> Signed-off-by: Jan Dakinevich <jan.dakinevich@gmail.com>
>> ---
>>  arch/x86/kvm/vmx.c | 3 +++
>>  1 file changed, 3 insertions(+)
>>
>> diff --git a/arch/x86/kvm/vmx.c b/arch/x86/kvm/vmx.c
>> index 5cede40..a4bb2bd 100644
>> --- a/arch/x86/kvm/vmx.c
>> +++ b/arch/x86/kvm/vmx.c
>> @@ -10500,6 +10500,9 @@ static void prepare_vmcs12(struct kvm_vcpu *vcpu, struct vmcs12 *vmcs12,
>>                 vmcs12->guest_pdptr3 = vmcs_read64(GUEST_PDPTR3);
>>         }
>>
>> +       if (nested_cpu_has_ept(vmcs12))
>> +               vmcs12->guest_linear_address = vmcs_readl(GUEST_LINEAR_ADDRESS);
>> +
> 
> No, nested_ept_inject_page_fault() will set
> vmcs12->guest_linear_address after L0 walks L1's EPT page table and
> finds that the mapping is invalid if nested EPT is enabled.

Acctually, nested_ept_inject_page_fault() doesn't do that, the routine
sets only vmcs12->guest_physical_address, but
vmcs12->guest_linear_address remains untouched. As result, after EPT
fault from L2, vmcs_readl(GUEST_LINEAR_ADDRESS) in L1 always returns 0.

> prepare_vmcs12() just copies the vmcs field that could have changed by
> the L2 guest or the exit-information etc instead of all fields since
> other fields are modified by L1 with VMWRITE, which already writes to
> vmcs12 directly.

Yes, and guest linear-address considered as a part of exit information,
provided by hardware.

> 
> Regards,
> Wanpeng Li
> 

-- 
Best regards
Jan Dakinevich

Re: [PATCH] KVM: nVMX: pass valid guest linear-address to the L1

[Paolo Bonzini]

On 05/09/2016 15:02, Jan Dakinevich wrote:
> > No, nested_ept_inject_page_fault() will set
> > vmcs12->guest_linear_address after L0 walks L1's EPT page table and
> > finds that the mapping is invalid if nested EPT is enabled.
>
> Acctually, nested_ept_inject_page_fault() doesn't do that, the routine
> sets only vmcs12->guest_physical_address, but
> vmcs12->guest_linear_address remains untouched. As result, after EPT
> fault from L2, vmcs_readl(GUEST_LINEAR_ADDRESS) in L1 always returns 0.
> 
> > prepare_vmcs12() just copies the vmcs field that could have changed by
> > the L2 guest or the exit-information etc instead of all fields since
> > other fields are modified by L1 with VMWRITE, which already writes to
> > vmcs12 directly.
> 
> Yes, and guest linear-address considered as a part of exit information,
> provided by hardware.

I agree.  The patch seems okay to me, but I'll leave a couple days more
to Wanpeng.

Paolo

Re: [PATCH] KVM: nVMX: pass valid guest linear-address to the L1

[Wanpeng Li]

2016-09-05 21:02 GMT+08:00 Jan Dakinevich <jan.dakinevich@gmail.com>:
>
>
> On 09/05/2016 03:49 AM, Wanpeng Li wrote:
>> 2016-09-05 2:22 GMT+08:00 Jan Dakinevich <jan.dakinevich@gmail.com>:
>>> If EPT support is exposed to L1 hypervisor, guest linear-address field
>>> of VMCS should contain GVA of L2, the access to which caused EPT violation.
>>>
>>> Signed-off-by: Jan Dakinevich <jan.dakinevich@gmail.com>
>>> ---
>>>  arch/x86/kvm/vmx.c | 3 +++
>>>  1 file changed, 3 insertions(+)
>>>
>>> diff --git a/arch/x86/kvm/vmx.c b/arch/x86/kvm/vmx.c
>>> index 5cede40..a4bb2bd 100644
>>> --- a/arch/x86/kvm/vmx.c
>>> +++ b/arch/x86/kvm/vmx.c
>>> @@ -10500,6 +10500,9 @@ static void prepare_vmcs12(struct kvm_vcpu *vcpu, struct vmcs12 *vmcs12,
>>>                 vmcs12->guest_pdptr3 = vmcs_read64(GUEST_PDPTR3);
>>>         }
>>>
>>> +       if (nested_cpu_has_ept(vmcs12))
>>> +               vmcs12->guest_linear_address = vmcs_readl(GUEST_LINEAR_ADDRESS);
>>> +
>>
>> No, nested_ept_inject_page_fault() will set
>> vmcs12->guest_linear_address after L0 walks L1's EPT page table and
>> finds that the mapping is invalid if nested EPT is enabled.
>
> Acctually, nested_ept_inject_page_fault() doesn't do that, the routine
> sets only vmcs12->guest_physical_address, but
> vmcs12->guest_linear_address remains untouched. As result, after EPT
> fault from L2, vmcs_readl(GUEST_LINEAR_ADDRESS) in L1 always returns 0.

Agreed, I misread guest_linear_address as guest_physical_address.

>
>> prepare_vmcs12() just copies the vmcs field that could have changed by
>> the L2 guest or the exit-information etc instead of all fields since
>> other fields are modified by L1 with VMWRITE, which already writes to
>> vmcs12 directly.
>
> Yes, and guest linear-address considered as a part of exit information,
> provided by hardware.

Thanks for the patch.

Reviewed-by: Wanpeng Li <wanpeng.li@hotmail.com>