Re: [PATCH v2 1/2] KVM: SVM: Move asid to vcpu_svm

[Paolo Bonzini]

On 13/10/20 03:29, Sean Christopherson wrote:
> > @@ -3446,6 +3447,7 @@ static __no_kcsan fastpath_t svm_vcpu_run(struct kvm_vcpu *vcpu)
> >   
> >   	sync_lapic_to_cr8(vcpu);
> >   
> > +	svm->vmcb->control.asid = svm->asid;
>
> Related to the above, handling this in vcpu_run() feels wrong.  There really
> shouldn't be a need to track the ASID.  vmcb01 will always exist if vmcb02
> exits, e.g. the ASID can be copied and marked dirty when loading vmcb02.
> For new_asid(), it can unconditionally update vmcb01 and conditionally update
> vmcb02.

Yeah, it is a bit ugly and it is only needed on processors without 
flush-by-ASID.  On those processors the ASID is used by KVM as a sort of 
TLB flush generation count.  A TLB flush should affect both VMCB01 and 
VMCB02, and that's the reason to have a global ASID in struct vcpu_svm.

On processors with flush-by-ASID, currently we bump the ASID on every 
physical CPU change.  However even that is not needed, in principle 
svm->asid should never change and we could also have different ASID01 
and ASID02, similar to VMX.  So: long term, svm->asid should only be 
used only on processors without flush-by-ASID.  kvm-amd however is not 
quite ready for that.

Paolo