On 18.10.2020 10:02, Eric Dumazet wrote:
> On Sun, Oct 18, 2020 at 1:29 AM Jakub Kicinski <k...@kernel.org> wrote:
>>
>> On Sat, 17 Oct 2020 15:45:57 +0200 Heiner Kallweit wrote:
>>> When __napi_schedule_irqoff was added with bc9ad166e38a
>>> ("net: introduce napi_schedule_irqoff()") the commit message stated:
>>> "Many NIC drivers can use it from their hard IRQ handler instead of
>>> generic variant."
>>
>> Eric, do you think it still matters? Does it matter on x86?
>>
>>> It turned out that this most of the time isn't safe in certain
>>> configurations:
>>> - if CONFIG_PREEMPT_RT is set
>>> - if command line parameter threadirqs is set
>>>
>>> Having said that drivers are being switched back to __napi_schedule(),
>>> see e.g. patch in [0] and related discussion. I thought about a
>>> __napi_schedule version checking dynamically whether interrupts are
>>> disabled. But checking e.g. variable force_irqthreads also comes at
>>> a cost, so that we may not see a benefit compared to calling
>>> local_irq_save/local_irq_restore.
>>>
>>> If more or less all users have to switch back, then the question is
>>> whether we should remove __napi_schedule_irqoff.
>>> Instead of touching all users we could make __napi_schedule_irqoff
>>> an alias for __napi_schedule for now.
>>>
>>> [0] https://lkml.org/lkml/2020/10/8/706
>>
>> We're effectively calling raise_softirq_irqoff() from IRQ handlers,
>> with force_irqthreads == true that's no longer legal.
>>
>> Thomas - is the expectation that IRQ handlers never assume they have
>> IRQs disabled going forward? We don't have any performance numbers
>> but if I'm reading Agner's tables right POPF is 18 cycles on Broadwell.
>> Is PUSHF/POPF too cheap to bother?
>>
>> Otherwise a non-solution could be to make IRQ_FORCED_THREADING
>> configurable.
>
> I have to say I do not understand why we want to defer to a thread the
> hard IRQ that we use in NAPI model.
>
Seems like the current forced threading comes with the big hammer and
thread-ifies all hard irq's. To avoid this all NAPI network drivers
would have to request the interrupt with IRQF_NO_THREAD.
> Whole point of NAPI was to keep hard irq handler very short.
>
> We should focus on transferring the NAPI work (potentially disrupting
> ) to a thread context, instead of the very minor hard irq trigger.
>
