Re: [PATCH 4/6] x86: KVM: Enable AVX-VNNI-INT8 CPUID and expose it to guest

[Jiaxi Chen]

On 10/19/2022 11:09 PM, Sean Christopherson wrote:
> On Wed, Oct 19, 2022, Sean Christopherson wrote:
>> On Wed, Oct 19, 2022, Borislav Petkov wrote:
>>> On Wed, Oct 19, 2022 at 04:47:32PM +0800, Jiaxi Chen wrote:
>>>> AVX-VNNI-INT8 is a new set of instructions in the latest Intel platform
>>>> Sierra Forest. It multiplies the individual bytes of two unsigned or
>>>> unsigned source operands, then add and accumulate the results into the
>>>> destination dword element size operand. This instruction allows for the
>>>> platform to have superior AI capabilities.
>>>>
>>>> The bit definition:
>>>> CPUID.(EAX=7,ECX=1):EDX[bit 4]
>>>
>>> For this particular one, use scattered.c instead of adding a new leaf.
>>
>> Unless the kernel wants to use X86_FEATURE_AVX_VNNI_INT8, which seems unlikely,
>> there's no need to create a scattered entry.  This can be handled in KVM by adding
>> a KVM-only leaf entry (which will be needed no matter what), plus a #define for
>> X86_FEATURE_AVX_VNNI_INT8 to direct it to the KVM entry.
>>
>> E.g. 
>>
>> diff --git a/arch/x86/kvm/reverse_cpuid.h b/arch/x86/kvm/reverse_cpuid.h
>> index a19d473d0184..25e7bfc61607 100644
>> --- a/arch/x86/kvm/reverse_cpuid.h
>> +++ b/arch/x86/kvm/reverse_cpuid.h
>> @@ -13,6 +13,7 @@
>>   */
>>  enum kvm_only_cpuid_leafs {
>>         CPUID_12_EAX     = NCAPINTS,
>> +       CPUID_7_1_EDX,
>>         NR_KVM_CPU_CAPS,
>>  
>>         NKVMCAPINTS = NR_KVM_CPU_CAPS - NCAPINTS,
>> @@ -24,6 +25,16 @@ enum kvm_only_cpuid_leafs {
>>  #define KVM_X86_FEATURE_SGX1           KVM_X86_FEATURE(CPUID_12_EAX, 0)
>>  #define KVM_X86_FEATURE_SGX2           KVM_X86_FEATURE(CPUID_12_EAX, 1)
>>  
>> +#define KVM_X86_FEATURE_AVX_VNNI_INT8  KVM_X86_FEATURE(CPUID_7_1_EDX, 4)
>> +
>> +/*
>> + * Alias X86_FEATURE_* to the KVM variant for features in KVM-only leafs that
>> + * aren't scattered by cpufeatures.h so that X86_FEATURE_* can be used in KVM,
>> + * e.g. to query guest CPUID.  As a bonus, no translation is needed for these
>> + * features in __feature_translate().
>> + */
>> +#define X86_FEATURE_AVX_VNNI_INT8      KVM_X86_FEATURE_AVX_VNNI_INT8
> 
> Actually, there's no need for KVM_X86_FEATURE_AVX_VNNI_INT8 in this case, just
> #define X86_FEATURE_AVX_VNNI_INT8 directly.  The KVM_ prefixed macro exists purely
> to redirect the non-KVM_ version, but that's unnecessary in this case.
> 
> diff --git a/arch/x86/kvm/reverse_cpuid.h b/arch/x86/kvm/reverse_cpuid.h
> index a19d473d0184..38adafb03490 100644
> --- a/arch/x86/kvm/reverse_cpuid.h
> +++ b/arch/x86/kvm/reverse_cpuid.h
> @@ -13,6 +13,7 @@
>   */
>  enum kvm_only_cpuid_leafs {
>         CPUID_12_EAX     = NCAPINTS,
> +       CPUID_7_1_EDX,
>         NR_KVM_CPU_CAPS,
>  
>         NKVMCAPINTS = NR_KVM_CPU_CAPS - NCAPINTS,
> @@ -24,6 +25,13 @@ enum kvm_only_cpuid_leafs {
>  #define KVM_X86_FEATURE_SGX1           KVM_X86_FEATURE(CPUID_12_EAX, 0)
>  #define KVM_X86_FEATURE_SGX2           KVM_X86_FEATURE(CPUID_12_EAX, 1)
>  
> +/*
> + * Omit the KVM_ prefix for features in KVM-only leafs that aren't scattered by
> + * cpufeatures.h so that X86_FEATURE_* can be used in KVM,* e.g. to query guest
> + * CPUID.  As a bonus, no handling in __feature_translate() is needed.
> + */
> +#define X86_FEATURE_AVX_VNNI_INT8      KVM_X86_FEATURE(CPUID_7_1_EDX, 4)
> +
>  struct cpuid_reg {
>         u32 function;
>         u32 index;
> @@ -48,6 +56,7 @@ static const struct cpuid_reg reverse_cpuid[] = {
>         [CPUID_7_1_EAX]       = {         7, 1, CPUID_EAX},
>         [CPUID_12_EAX]        = {0x00000012, 0, CPUID_EAX},
>         [CPUID_8000_001F_EAX] = {0x8000001f, 0, CPUID_EAX},
> +       [CPUID_7_1_EDX]       = {         7, 1, CPUID_EDX},
>  };

Use KVM-only leafs is better for edx in this case. Will follow this suggestion in v2. 
Thanks Sean very much.
-- 
Regards,
Jiaxi