Re: [PATCH] KVM: nSVM: check for EFER.SVME=1 before entering guest

[Paolo Bonzini <pbonzini@redhat.com>]

On 18/03/20 19:40, Krish Sadhukhan wrote:
> 
> On 3/18/20 5:41 AM, Paolo Bonzini wrote:
>> EFER is set for L2 using svm_set_efer, which hardcodes EFER_SVME to 1
>> and hides
>> an incorrect value for EFER.SVME in the L1 VMCB.  Perform the check
>> manually
>> to detect invalid guest state.
>>
>> Reported-by: Krish Sadhukhan <krish.sadhukhan@oracle.com>
>> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
>> ---
>>   arch/x86/kvm/svm.c | 3 +++
>>   1 file changed, 3 insertions(+)
>>
>> diff --git a/arch/x86/kvm/svm.c b/arch/x86/kvm/svm.c
>> index 08568ae9f7a1..2125c6ae5951 100644
>> --- a/arch/x86/kvm/svm.c
>> +++ b/arch/x86/kvm/svm.c
>> @@ -3558,6 +3558,9 @@ static bool nested_svm_vmrun_msrpm(struct
>> vcpu_svm *svm)
>>     static bool nested_vmcb_checks(struct vmcb *vmcb)
>>   {
>> +    if ((vmcb->save.efer & EFER_SVME) == 0)
>> +        return false;
>> +
>>       if ((vmcb->control.intercept & (1ULL << INTERCEPT_VMRUN)) == 0)
>>           return false;
>>   
> 
> Ah! This now tells me that I forgot the KVM fix that was supposed to
> accompany my patchset.

Heh, indeed.  I was puzzled for a second after applying it, then decided
I would just fix it myself. :)

> Do we need this check in software ? I wasn't checking the bit in KVM and
> instead I was just making sure that L0 sets that bit based on the
> setting in nested vmcb:

The only effect of the function below over svm_set_efer is to guarantee
a vmrun error to happen.  Doing the test in nested_vmcb_checks is more
consistent with other must-be-one bits such as the VMRUN intercept, and
it's also a smaller patch.

Paolo

> 
> +static void nested_svm_set_efer(struct kvm_vcpu *vcpu, u64
> nested_vmcb_efer)
> +{
> +       svm_set_efer(vcpu, nested_vmcb_efer);
> +
> +       if (!(nested_vmcb_efer & EFER_SVME))
> +               to_svm(vcpu)->vmcb->save.efer &= ~EFER_SVME;
> +}
> +
>  static int is_external_interrupt(u32 info)
>  {
>         info &= SVM_EVTINJ_TYPE_MASK | SVM_EVTINJ_VALID;
> @@ -3554,7 +3562,7 @@ static void enter_svm_guest_mode(struct vcpu_svm
> *svm, u64
>         svm->vmcb->save.gdtr = nested_vmcb->save.gdtr;
>         svm->vmcb->save.idtr = nested_vmcb->save.idtr;
>         kvm_set_rflags(&svm->vcpu, nested_vmcb->save.rflags);
> -       svm_set_efer(&svm->vcpu, nested_vmcb->save.efer);
> +       nested_svm_set_efer(&svm->vcpu, nested_vmcb->save.efer);
>         svm_set_cr0(&svm->vcpu, nested_vmcb->save.cr0);
>         svm_set_cr4(&svm->vcpu, nested_vmcb->save.cr4);
>         if (npt_enabled) {
>