Re: [PATCH v2 1/2] x86/fpu: Extend kernel_fpu_begin_mask() to initialize AMX state

[Dave Hansen]

On 5/9/24 11:41, Chang S. Bae wrote:
> On 5/9/2024 10:36 AM, Dave Hansen wrote:
>>
>> I'd probably just do this:
>>
>> +    kernel_fpu_begin();
>> +    // AMX *MUST* be in the init state for the wrmsr() to work.
>> +    // But, the more in the init state, the less state the test
>> +    // has to save and restore.  Just zap everything.
>> +    restore_fpregs_from_fpstate(&init_fpstate,   
>> +                    fpu_user_cfg.max_features);
>> +
> 
> I assume that this snippet goes to the IFS driver side. Then, we need
> to introduce and export a new wrapper for this. 
> restore_fpregs_from_fpstate() and its arguments are not accessible as
> of now.

Yes, a new wrapper to initialize all the user FPU state is fine.

> Also, I think we should encapsulate them. If we follow this style, we
> could have invoked tilerelease() directly from the idle driver, right?

You could have...  But I think the point there truly was to do a minimal
amount of work because you're going idle and need to race to get there.
The path to idle is super hot and you don't want to do _any_ additional
work.  It's worth writing highly AMX-specific code because generic code
would be slow there.

The IFS path is a super duper slow one.  It takes *FOREVER*.  I like the
idea of being simple, dumb and slow.