Re: [RFC PATCH v2] x86/fpu: make kernel-mode FPU reliably usable in softirqs

[Dave Hansen]

On 3/5/25 09:37, Ingo Molnar wrote:
> 
> * Dave Hansen <dave.hansen@xxxxxxxxx> wrote:
> 
>> On 3/5/25 01:07, Ingo Molnar wrote:>> Alternatives considered:
>>>> - Make kernel-mode FPU sections fully preemptible.  This would require
>>>>   growing task_struct by another struct fpstate which is more than 2K.
>>>
>>> So that's something that will probably happen once the kernel is built 
>>> using APX anyway?
>>
>> I was expecting that building the kernel with APX would be very 
>> different than a kernel_fpu_begin(). We don't just need *one* more 
>> save area for APX registers: we need a stack, just like normal GPRs.
> 
> Yes - but my point is: with any APX build we'd probably be saving 
> FPU(-ish) registers at entry points, into a separate context area. If 
> that includes FPU registers then we'd not have to do 
> kernel_fpu_begin()/end().

That's true. But wouldn't it be a bit silly to include _all_ FPU
registers? If the kernel isn't using AVX512, why bother saving and
restoring AVX512?

> In other words, we'd be doing something close to 'growing task_struct 
> by another struct fpstate', or so - regardless of whether it's in 
> task_struct or some sort of extended pt_regs. The kernel would also be 
> close to 'FPU-safe', i.e. there likely wouldn't be a need for 
> kernel_fpu_begin()/end().
The new APX registers are 128 bytes, total. 'struct fpstate' is ~3k on
most CPUs these days and >11k with AMX.

I was thinking that growing things (say pt_regs) by 128b would be
acceptable, given some nice performance gains from AMX. Growing by 3k
would be cause some real headaches. Growing by 11k would be a non-starter.

I'm pretty sure the torches and pitchforks would come out if our syscall
latency included another 1k of save/restore much less 3k or 11k.