Re: [PATCH v3 2/2] KVM: X86: Introduce KVM_HC_PAGE_ENC_STATUS hypercall

[Paolo Bonzini <pbonzini@redhat.com>]

On 30/04/21 22:10, Sean Christopherson wrote:
> On Thu, Apr 29, 2021, Paolo Bonzini wrote:
>> diff --git a/Documentation/virt/kvm/msr.rst b/Documentation/virt/kvm/msr.rst
>> index 57fc4090031a..cf1b0b2099b0 100644
>> --- a/Documentation/virt/kvm/msr.rst
>> +++ b/Documentation/virt/kvm/msr.rst
>> @@ -383,5 +383,10 @@ MSR_KVM_MIGRATION_CONTROL:
>>   data:
>>           This MSR is available if KVM_FEATURE_MIGRATION_CONTROL is present in
>>           CPUID.  Bit 0 represents whether live migration of the guest is allowed.
>> +
>>           When a guest is started, bit 0 will be 1 if the guest has encrypted
>> -        memory and 0 if the guest does not have encrypted memory.
>> +        memory and 0 if the guest does not have encrypted memory.  If the
>> +        guest is communicating page encryption status to the host using the
>> +        ``KVM_HC_PAGE_ENC_STATUS`` hypercall, it can set bit 0 in this MSR to
>> +        allow live migration of the guest.  The MSR is read-only if
>> +        ``KVM_FEATURE_HC_PAGE_STATUS`` is not advertised to the guest.
> 
> I still don't get the desire to tie MSR_KVM_MIGRATION_CONTROL to PAGE_ENC_STATUS
> in any way shape or form.  I can understand making it read-only or dropping
> writes if it's not intercepted by userspace, but making it read-only for
> non-encrypted guests makes it useful only for encrypted guests, which defeats
> the purpose of genericizing the MSR.

Yeah, I see your point.  On the other hand by making it unconditionally 
writable we must implement the writability in KVM, because a read-only 
implementation would not comply with the spec.

>> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
>> index e9c40be9235c..0c2524bbaa84 100644
>> --- a/arch/x86/kvm/x86.c
>> +++ b/arch/x86/kvm/x86.c
>> @@ -3279,6 +3279,12 @@ int kvm_set_msr_common(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
>>   		if (!guest_pv_has(vcpu, KVM_FEATURE_MIGRATION_CONTROL))
>>   			return 1;
>>   
>> +		/*
>> +		 * This implementation is only good if userspace has *not*
>> +		 * enabled KVM_FEATURE_HC_PAGE_ENC_STATUS.  If userspace
>> +		 * enables KVM_FEATURE_HC_PAGE_ENC_STATUS it must set up an
>> +		 * MSR filter in order to accept writes that change bit 0.
>> +		 */
>>   		if (data != !static_call(kvm_x86_has_encrypted_memory)(vcpu->kvm))
>>   			return 1;
> 
> This behavior doesn't match the documentation.
> 
>    a. The MSR is not read-only for legacy guests since they can write '0'.
>    b. The MSR is not read-only if KVM_FEATURE_HC_PAGE_STATUS isn't advertised,
>       a guest with encrypted memory can write '1' regardless of whether userspace
>       has enabled KVM_FEATURE_HC_PAGE_STATUS.

Right, I should have said "not changeable" rather than "read-only".

>    c. The MSR is never fully writable, e.g. a guest with encrypted memory can set
>       bit 0, but not clear it.  This doesn't seem intentional?

It is intentional, clearing it would mean preserving the value in the 
kernel so that userspace can read it.

So... I don't know, all in all having both the separate CPUID and the 
userspace implementation reeks of overengineering.  It should be either 
of these:

- separate CPUID bit, MSR unconditionally writable and implemented in 
KVM.  Userspace is expected to ignore the MSR value for encrypted guests 
unless KVM_FEATURE_HC_PAGE_STATUS is exposed.  Userspace should respect 
it even for unencrypted guests (not a migration-DoS vector, because 
userspace can just not expose the feature).

- make it completely independent from migration, i.e. it's just a facet 
of MSR_KVM_PAGE_ENC_STATUS saying whether the bitmap is up-to-date.  It 
would use CPUID bit as the encryption status bitmap and have no code at 
all in KVM (userspace needs to set up the filter and implement everything).

At this point I very much prefer the latter, which is basically Ashish's 
earlier patch.

Paolo

> Why not simply drop writes?  E.g.
> 
> 		if (data & ~KVM_MIGRATION_READY)
> 			return 1;
> 		break;
> 
> And then do "msr->data = 0;" in the read path.  That's just as effective as
> making the MSR read-only to force userspace to intercept the MSR if it wants to
> do anything useful with the information, and it's easy to document.