Re: [RFC PATCH 2/2] arm64: support HAVE_IRQ_EXIT_ON_IRQ_STACK

[Qi Zheng <zhengqi.arch@bytedance.com>]

On 2022/7/8 16:52, Arnd Bergmann wrote:
> On Fri, Jul 8, 2022 at 5:13 AM Qi Zheng <zhengqi.arch@bytedance.com> wrote:
>> On 2022/7/8 04:55, Arnd Bergmann wrote:
>>> On Thu, Jul 7, 2022 at 5:00 PM Qi Zheng <zhengqi.arch@bytedance.com> wrote:
>>>> On 2022/7/7 22:41, Arnd Bergmann wrote:
>>>>> On Thu, Jul 7, 2022 at 3:38 PM Qi Zheng <zhengqi.arch@bytedance.com> wrote:
>>>>>> On 2022/7/7 20:49, Arnd Bergmann wrote:
>>>>>
>>>>> -asmlinkage void noinstr el1h_64_fiq_handler(struct pt_regs *regs)
>>>>> +asmlinkage void noinstr el1h_64_irq_handler(struct pt_regs *regs)
>>>>> +{
>>>>> +       if (on_thread_stack())
>>>>> +               call_on_irq_stack(regs, el1_irq);
>>>>
>>>> IMO, this can't work. Because el1_interrupt() will invoke
>>>> arm64_preempt_schedule_irq(), which will cause scheduling on the
>>>> IRQ stack.
>>>
>>> Ah, too bad. I spent some more time looking for a simpler approach,
>>> but couldn't find one I'm happy with. One idea might be to have
>>> callback functions for each combinations of irq/fiq with irq/pnmi
>>> to avoid the nested callback pointers. Not sure if that helps.
>>
>> Maybe nested callback pointers are not always a wild beast. ;)
>> This method does not change much, and we can also conveniently stuff
>> all kinds of things in do_handler() that we want to run on the IRQ
>> stack in addition to the handler().
> 
> Right, your approach is probably the one that changes the existing
> code the least. I see that x86 handles this by having call_on_irq_stack()
> in an inline asm, but this in turn complicates the asm implementation,
> which is also worth keeping simple.

Yes, and I see that the commit f2c5092190f2 ("arch/*: Disable softirq
stacks on PREEMPT_RT.") has been merged into next-20220707, so I will
rebase to the next-20220707 and send the next version.

Thank you very much :)

> 
>           Arnd

-- 
Thanks,
Qi