Re: [PATCH v3 25/27] KVM: nVMX: Add FRED VMCS fields

[Sean Christopherson <seanjc@google.com>]

On Mon, Oct 28, 2024, Chao Gao wrote:
> On Fri, Oct 25, 2024 at 12:25:45AM -0700, Xin Li wrote:
> >> > static void nested_vmx_setup_cr_fixed(struct nested_vmx_msrs *msrs)
> >> > diff --git a/arch/x86/kvm/vmx/nested.h b/arch/x86/kvm/vmx/nested.h
> >> > index 2c296b6abb8c..5272f617fcef 100644
> >> > --- a/arch/x86/kvm/vmx/nested.h
> >> > +++ b/arch/x86/kvm/vmx/nested.h
> >> > @@ -251,6 +251,14 @@ static inline bool nested_cpu_has_encls_exit(struct vmcs12 *vmcs12)
> >> > 	return nested_cpu_has2(vmcs12, SECONDARY_EXEC_ENCLS_EXITING);
> >> > }
> >> > 
> >> > +static inline bool nested_cpu_has_fred(struct vmcs12 *vmcs12)
> >> > +{
> >> > +	return vmcs12->vm_entry_controls & VM_ENTRY_LOAD_IA32_FRED &&
> >> > +	       vmcs12->vm_exit_controls & VM_EXIT_ACTIVATE_SECONDARY_CONTROLS &&
> >> > +	       vmcs12->secondary_vm_exit_controls & SECONDARY_VM_EXIT_SAVE_IA32_FRED &&
> >> > +	       vmcs12->secondary_vm_exit_controls & SECONDARY_VM_EXIT_LOAD_IA32_FRED;
> >> 
> >> Is it a requirement in the SDM that the VMM should enable all FRED controls or
> >> none? If not, the VMM is allowed to enable only one or two of them. This means
> >> KVM would need to emulate FRED controls for the L1 VMM as three separate
> >> features.
> >
> >The SDM doesn't say that.  But FRED states are used during and
> >immediately after VM entry and exit, I don't see a good reason for a VMM
> >to enable only one or two of the 3 save/load configs.

Not KVM's concern.

> >Say if VM_ENTRY_LOAD_IA32_FRED is not set, it means a VMM needs to
> >switch to guest FRED states before it does a VM entry, which is
> >absolutely a big mess.

Again, not KVM's concern.

> If the VMM doesn't enable FRED, it's fine to load guest FRED states before VM
> entry, right?

Yep.  Or if L1 is simply broken and elects to manually load FRED state before
VM-Enter instead of using VM_ENTRY_LOAD_IA32_FRED, then any badness that happens
is 100% L1's problem to deal with.  KVM's responsiblity is to emulate the
architectural behavior, what L1 may or may not do is irrelevant.

> The key is to emulate hardware behavior accurately without making assumptions
> about guests.

+1000

> If some combinations of controls cannot be emulated properly, KVM
> should report internal errors at some point.