Re: Lazy FPU restoration / moving kernel_fpu_end() to context switch

[Brian Gerst]

On Fri, Jun 15, 2018 at 12:25 PM, Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:
> On Fri, 15 Jun 2018, Jason A. Donenfeld wrote:
>> In a loop this looks like:
>>
>> for (thing) {
>>   kernel_fpu_begin();
>>   encrypt(thing);
>>   kernel_fpu_end();
>> }
>>
>> This is obviously very bad, because begin() and end() are slow, so
>> WireGuard does the obvious:
>>
>> kernel_fpu_begin();
>> for (thing)
>>   encrypt(thing);
>> kernel_fpu_end();
>>
>> This is fine and well, and the crypto API I'm working on will enable
>
> It might be fine crypto performance wise, but it's a total nightmare
> latency wise because kernel_fpu_begin() disables preemption. We've seen
> latencies in the larger millisecond range due to processing large data sets
> with kernel FPU.
>
> If you want to go there then we really need a better approach which allows
> kernel FPU usage in preemptible context and in case of preemption a way to
> stash the preempted FPU context and restore it when the task gets scheduled
> in again. Just using the existing FPU stuff and moving the loops inside the
> begin/end section and keeping preemption disabled for arbitrary time spans
> is not going to fly.

One optimization that can be done is to delay restoring the user FPU
state until we exit to userspace.  That way the FPU is saved and
restored only once no matter how many times
kernel_fpu_begin()/kernel_fpu_end() are called.

--
Brian Gerst