Re: [PATCH v2 1/2] x86/fpu: Add a helper to prepare AMX state for low-power CPU idle

[Dave Hansen]

On 3/9/22 15:12, Chang S. Bae wrote:
> On 3/9/2022 2:46 PM, Dave Hansen wrote:
>> On 3/9/22 14:34, Chang S. Bae wrote:
>>> +/*
>>> + * Initialize register state that may prevent from entering
>>> low-power idle.
>>> + * This function will be invoked from the cpuidle driver only when
>>> needed.
>>> + */
>>> +void fpu_idle_fpregs(void)
>>> +{
>>> +    if (!fpu_state_size_dynamic())
>>> +        return;
>>
>> Is this check just an optimization?  
> 
> No. 0day reported the splat [3] with the earlier code in v1 [1]:
> 
> if (fpu_state_size_dynamic() && (xfeatures_in_use() &
> XFEATURE_MASK_XTILE)) { ... }
> 
> It looks like GCC-9 reordered to hit XGETBV without checking
> fpu_state_size_dynamic(). So this line was separated to avoid that.

I assume that splat is because 0day found a CPU which doesn't support
XGETBV1.  Since fpu_state_size_dynamic() only ever returns true on
XGETBV1 systems so it works as a proxy for checking XGETBV1 support.

Right?

If so, then fpu_state_size_dynamic() is a *bit* of an oblique way to
check for XGETBV1 support.

Why don't we do a good old:

	cpu_feature_enabled(X86_FEATURE_XGETBV1)

check?

Also, did we get the asm constraints wrong on xgetbv()?  Surely we
shouldn't be allowing the compiler to reorder it.  Do we need a "memory"
constraint?