On Fri, Aug 16, 2024 at 07:51:48AM -0700, Sean Christopherson wrote:
> On Mon, Jun 03, 2024, Tao Su wrote:
> > Advertise AVX10.1 related CPUIDs, i.e. report AVX10 support bit via
> > CPUID.(EAX=07H, ECX=01H):EDX[bit 19] and new CPUID leaf 0x24H so that
> > guest OS and applications can query the AVX10.1 CPUIDs directly. Intel
> > AVX10 represents the first major new vector ISA since the introduction of
> > Intel AVX512, which will establish a common, converged vector instruction
> > set across all Intel architectures[1].
> >
> > AVX10.1 is an early version of AVX10, that enumerates the Intel AVX512
> > instruction set at 128, 256, and 512 bits which is enabled on
> > Granite Rapids. I.e., AVX10.1 is only a new CPUID enumeration with no
> > VMX capability, Embedded rounding and Suppress All Exceptions (SAE),
> > which will be introduced in AVX10.2.
> >
> > Advertising AVX10.1 is safe because kernel doesn't enable AVX10.1 which is
>
> I thought there is nothing to enable for AVX10.1? I.e. it's purely a new way to
> enumerate support, thus there will never be anything for the kernel to enable.
>
Yes, AVX10.1 is just a new enumeration way.
> > on KVM-only leaf now, just the CPUID checking is changed when using AVX512
> > related instructions, e.g. if using one AVX512 instruction needs to check
> > (AVX512 AND AVX512DQ), it can check ((AVX512 AND AVX512DQ) OR AVX10.1)
> > after checking XCR0[7:5].
> >
> > The versions of AVX10 are expected to be inclusive, e.g. version N+1 is
> > a superset of version N. Per the spec, the version can never be 0, just
> > advertise AVX10.1 if it's supported in hardware.
>
> I think it's also worth calling out that advertising AVX10_{128,256,512} needs
> to land in the same patch (this patch) as AVX10 (and thus AVX10.1), because
> otherwise KVM would advertise an impossible CPU model, e.g. with AVX512 but not
> AVX10.1/512, which per "Feature Differences Between Intel® AVX-512 and Intel® AVX10"
> should be impossible.
>
Indeed, that's the reason why I only used one patch, will do in v3, thanks!
> > As more and more AVX related CPUIDs are added (it would have resulted in
> > around 40-50 CPUID flags when developing AVX10), the versioning approach
> > is introduced. But incrementing version numbers are bad for virtualization.
> > E.g. if AVX10.2 has a feature that shouldn't be enumerated to guests for
> > whatever reason, then KVM can't enumerate any "later" features either,
> > because the only way to hide the problematic AVX10.2 feature is to set the
> > version to AVX10.1 or lower[2]. But most AVX features are just passed
> > through and don’t have virtualization controls, so AVX10 should not be
> > problematic in practice.
> >
> > [1] https://cdrdv2.intel.com/v1/dl/getContent/784267
> > [2] https://lore.kernel.org/all/Zkz5Ak0PQlAN8DxK@xxxxxxxxxx/
> >
> > Signed-off-by: Tao Su <tao1.su@xxxxxxxxxxxxxxx>
> > ---
> > Changelog:
> > v1 -> v2:
> > - Directly advertise version 1 because version can never be 0.
> > - Add and advertise feature bits for the supported vector sizes.
> >
> > v1: https://lore.kernel.org/all/20240520022002.1494056-1-tao1.su@xxxxxxxxxxxxxxx/
> > ---
> > arch/x86/include/asm/cpuid.h | 1 +
> > arch/x86/kvm/cpuid.c | 21 +++++++++++++++++++--
> > arch/x86/kvm/reverse_cpuid.h | 8 ++++++++
> > 3 files changed, 28 insertions(+), 2 deletions(-)
> >
> > diff --git a/arch/x86/include/asm/cpuid.h b/arch/x86/include/asm/cpuid.h
> > index 6b122a31da06..aa21c105eef1 100644
> > --- a/arch/x86/include/asm/cpuid.h
> > +++ b/arch/x86/include/asm/cpuid.h
> > @@ -179,6 +179,7 @@ static __always_inline bool cpuid_function_is_indexed(u32 function)
> > case 0x1d:
> > case 0x1e:
> > case 0x1f:
> > + case 0x24:
> > case 0x8000001d:
> > return true;
> > }
> > diff --git a/arch/x86/kvm/cpuid.c b/arch/x86/kvm/cpuid.c
> > index f2f2be5d1141..6717a5b7d9cd 100644
> > --- a/arch/x86/kvm/cpuid.c
> > +++ b/arch/x86/kvm/cpuid.c
> > @@ -693,7 +693,7 @@ void kvm_set_cpu_caps(void)
> >
> > kvm_cpu_cap_init_kvm_defined(CPUID_7_1_EDX,
> > F(AVX_VNNI_INT8) | F(AVX_NE_CONVERT) | F(PREFETCHITI) |
> > - F(AMX_COMPLEX)
> > + F(AMX_COMPLEX) | F(AVX10)
> > );
> >
> > kvm_cpu_cap_init_kvm_defined(CPUID_7_2_EDX,
> > @@ -709,6 +709,10 @@ void kvm_set_cpu_caps(void)
> > SF(SGX1) | SF(SGX2) | SF(SGX_EDECCSSA)
> > );
> >
> > + kvm_cpu_cap_init_kvm_defined(CPUID_24_0_EBX,
> > + F(AVX10_128) | F(AVX10_256) | F(AVX10_512)
> > + );
> > +
> > kvm_cpu_cap_mask(CPUID_8000_0001_ECX,
> > F(LAHF_LM) | F(CMP_LEGACY) | 0 /*SVM*/ | 0 /* ExtApicSpace */ |
> > F(CR8_LEGACY) | F(ABM) | F(SSE4A) | F(MISALIGNSSE) |
> > @@ -937,7 +941,7 @@ static inline int __do_cpuid_func(struct kvm_cpuid_array *array, u32 function)
> > switch (function) {
> > case 0:
> > /* Limited to the highest leaf implemented in KVM. */
> > - entry->eax = min(entry->eax, 0x1fU);
> > + entry->eax = min(entry->eax, 0x24U);
> > break;
> > case 1:
> > cpuid_entry_override(entry, CPUID_1_EDX);
> > @@ -1162,6 +1166,19 @@ static inline int __do_cpuid_func(struct kvm_cpuid_array *array, u32 function)
> > break;
> > }
> > break;
> > + case 0x24: {
>
> No need for the curly braces on the case. But, my suggestion below will change
> that ;-)
>
I got it :-)
> > + if (!kvm_cpu_cap_has(X86_FEATURE_AVX10)) {
> > + entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
> > + break;
> > + }
> > + entry->eax = 0;
> > + cpuid_entry_override(entry, CPUID_24_0_EBX);
> > + /* EBX[7:0] hold the AVX10 version; KVM supports version '1'. */
> > + entry->ebx |= 1;
>
> Ah, rather than hardcode this to '1', I think we should do:
>
> u8 avx10_version;
>
> if (!kvm_cpu_cap_has(X86_FEATURE_AVX10)) {
> entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
> break;
> }
>
> /*
> * The AVX10 version is encoded in EBX[7:0]. Note, the version
> * is guaranteed to be >=1 if AVX10 is supported. Note #2, the
> * version needs to be captured before overriding EBX features!
> */
> avx10_version = min_t(u8, entry->ebx & 0xff, 1);
>
> cpuid_entry_override(entry, CPUID_24_0_EBX);
> entry->ebx |= avx10_version;
>
> I.e. use the same approach as limiting the max leaf, which does:
>
> entry->eax = min(entry->eax, 0x1fU);
>
> Unless I'm misunderstanding how all of this is expected to play out, we're going
> to need the min_t() code for AVX10.2 anyways, might as well implement it now.
Yes, that's better for the upcoming AVX10.2, thanks for the suggestions!
