On 11/3/21 5:46 AM, Maxim Levitsky wrote:
> Fix two semi-theoretical issues that are present.
>
> 1. AVX is assumed to be present when AVX2 is present.
> That can be false in a VM.
> This can be considered a hypervisor bug,
> but the kernel should not crash in this case if this is possible.
The kernel shouldn't crash in this case. We've got a software
dependency which should disable AVX2 if AVX is off:
static const struct cpuid_dep cpuid_deps[] = {
...
{ X86_FEATURE_AVX2, X86_FEATURE_AVX, },
> 2. YMM state can be soft disabled in XCR0.
>
> Fix both issues by using 'cpu_has_xfeatures(XFEATURE_MASK_YMM')
> to check for usable AVX support.
There's another table to ensure that this doesn't happen:
> static unsigned short xsave_cpuid_features[] __initdata = {
> [XFEATURE_FP] = X86_FEATURE_FPU,
> [XFEATURE_SSE] = X86_FEATURE_XMM,
> [XFEATURE_YMM] = X86_FEATURE_AVX,
So, if XFEATURE_YMM isn't supported, X86_FEATURE_AVX should be cleared.
But, that's all how it's _supposed_ to work. It's quite possible we've
got bugs somewhere, so if you're hitting an issue in practice please let
us know.
If this did end up confusing you and Paulo, that's not great either.
Any patches that make these dependency tables easier to find or grok
would be appreciated too.
