Re: [Xen-devel] [PATCH 23/27] x86/cpuid: Move all leaf 7 handling into guest_cpuid()

["Jan Beulich" <JBeulich@xxxxxxxx>]

>>> On 04.01.17 at 13:39, <andrew.cooper3@xxxxxxxxxx> wrote:
> @@ -380,14 +385,42 @@ void guest_cpuid(const struct vcpu *v, unsigned int 
> leaf,
>      case 0x80000000 ... 0x80000000 + CPUID_GUEST_NR_EXTD - 1:
>          if ( leaf > p->extd.max_leaf )
>              return;
> -        break;
> +        goto legacy;
>  
>      default:
>          return;
>      }
>  
> +    /* Skip dynamic adjustments if we are in the wrong context. */
> +    if ( v != curr )
> +        return;
> +
> +    /*
> +     * Second pass:
> +     * - Dynamic adjustments
> +     */
> +    switch ( leaf )
> +    {
> +    case 0x7:
> +        switch ( subleaf )
> +        {
> +        case 0:
> +            /* OSPKE clear in policy.  Fast-forward CR4 back in. */
> +            if ( (is_pv_vcpu(v)
> +                  ? v->arch.pv_vcpu.ctrlreg[4]
> +                  : v->arch.hvm_vcpu.guest_cr[4]) & X86_CR4_PKE )
> +                res->c |= cpufeat_mask(X86_FEATURE_OSPKE);

What's wrong with doing this adjustment when v != curr? By
the time the caller looks at the result, the state of guest
software controlled bits can't be relied upon anyway. Which
then raises the question whether a second switch() statement
for the a second pass is all that useful in the first place (I
realize this may depend on future plans of yours).

Jan


_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
https://lists.xen.org/xen-devel