Re: napi_busy_poll

[Hao Xu <haoxu@linux.alibaba.com>]

On 2/13/22 03:51, Olivier Langlois wrote:
> On Wed, 2022-02-09 at 11:34 +0800, Hao Xu wrote:
>> 在 2022/2/9 上午1:05, Jens Axboe 写道:
>>> On 2/8/22 7:58 AM, Olivier Langlois wrote:
>>>> Hi,
>>>>
>>>> I was wondering if integrating the NAPI busy poll for socket fds
>>>> into
>>>> io_uring like how select/poll/epoll are doing has ever been
>>>> considered?
>>>>
>>>> It seems to me that it could be an awesome feature when used
>>>> along with
>>>> a io_qpoll thread and something not too difficult to add...
>>> Should be totally doable and it's been brought up before, just
>>> needs
>>> someone to actually do it... Would love to see it.
>>>
>> We've done some investigation before, would like to have a try.
>>
> Hao,
>
> Let me know if I can help you with coding or testing. I have done very
> preliminary investigation too. It doesn't seem like it would be very
> hard to implement but I get confused with small details.
>
> For instance, the epoll implementation, unless there is something that
> I don't understand, appears to have a serious limitation. It seems like
> it would not work correctly if there are sockets associated to more
> than 1 NAPI device in the fd set. As far as I am concerned, that

Yes, it seems that epoll_wait only does busy polling for 1 NAPI.

I think it is because the busy polling there is just an optimization

(doing some polling before trapping into sleep) not a must have,

so it's kind of trade-off between polling and reacting to other events

I guess. Not very sure about this too..

The iouring implementation I'm thinking of in my mind is polling for every

NAPI involved.


Regards,

Hao

> limitation would be ok since in my setup I only use 1 device but if it
> was necessary to be better than the epoll implementation, I am not sure
> at all how this could be addressed. I do not have enough kernel dev
> experience to find easy solutions to those type of issues...
>
> Worse case scenario, I guess that I could give it a shot creating a
> good enough implementation for my needs and show it to the list to get
> feedback...
>
> Greetings,
> Olivier