Re: [PATCH v9 07/10] VMX: support MSR-IMM

[Andrew Cooper <andrew.cooper3@xxxxxxxxxx>]

On 24/11/2025 3:00 pm, Jan Beulich wrote:
> Hook up the new VM exit codes and handle guest uses of the insns.
>
> Signed-off-by: Jan Beulich <jbeulich@xxxxxxxx>
> ---
> v9: New.
> ---
> The lack of an enable bit is concerning; at least for the nested case
> that's a security issue afaict (when L0 isn't aware of the insns, or more
> specifically the exit codes).

This is why we need support statements of new CPUs.

Intel say that unknown VMExits turning into #UD is the expected course
of action.  That covers all of these cases which don't have an explicit
enable.

This is better than our current behaviour, which is non-architectural
for supervisor code and practically the most unhelpful course of action
going.

Obviously, logic turning on a new feature is expected to handle all the
VMExit cases it can produce.

The corollary for nested virt is that L0 must never make a Virtual
VMExit with case that isn't enabled.  Combined with #UD in the unknown
case, that covers things reasonably well.

> --- a/xen/arch/x86/domain.c
> +++ b/xen/arch/x86/domain.c
> @@ -453,7 +453,7 @@ void domain_cpu_policy_changed(struct do
>      }
>  
>      /* Nested doesn't have the necessary processing, yet. */
> -    if ( nestedhvm_enabled(d) && p->feat.user_msr )
> +    if ( nestedhvm_enabled(d) && (p->feat.user_msr || p->feat.msr_imm) )
>          return /* -EINVAL */;

What processing is missing?  (Aside from correcting the unknown case.)

>  
>      for_each_vcpu ( d, v )
> --- a/xen/arch/x86/hvm/vmx/vmx.c
> +++ b/xen/arch/x86/hvm/vmx/vmx.c
> @@ -4762,6 +4762,7 @@ void asmlinkage vmx_vmexit_handler(struc
>          break;
>  
>      case EXIT_REASON_URDMSR:
> +    case EXIT_REASON_RDMSR_IMM:

Instructions which aren't enumerated in CPUID have reserved behaviour.

The exit handler needs to check cp->feat.msr_imm and inject #UD.

It's not perfect; un-intercepted MSRs will happen to execute correctly
on capable hardware, but most MSRs are not intercepted and it's far
closer to adequate behaviour than omitting the #UD check.

~Andrew