Re: [PATCH liburing 1/2] io_uring_enter: add timeout support

[Jens Axboe <axboe@kernel.dk>]

On 7/30/20 8:12 PM, Jiufei Xue wrote:
> 
> 
> On 2020/7/30 下午11:28, Jens Axboe wrote:
>> On 7/29/20 8:32 PM, Jiufei Xue wrote:
>>> Hi Jens,
>>>
>>> On 2020/7/30 上午1:51, Jens Axboe wrote:
>>>> On 7/29/20 4:10 AM, Jiufei Xue wrote:
>>>>> Kernel can handle timeout when feature IORING_FEAT_GETEVENTS_TIMEOUT
>>>>> supported. Add two new interfaces: io_uring_wait_cqes2(),
>>>>> io_uring_wait_cqe_timeout2() for applications to use this feature.
>>>>
>>>> Why add new new interfaces, when the old ones already pass in the
>>>> timeout? Surely they could just use this new feature, instead of the
>>>> internal timeout, if it's available?
>>>>
>>> Applications use the old one may not call io_uring_submit() because
>>> io_uring_wait_cqes() will do it. So I considered to add a new one.
>>
>> Not sure I see how that's a problem - previously, you could not do that
>> either, if you were doing separate submit/complete threads. So this
>> doesn't really add any new restrictions. The app can check for the
>> feature flag to see if it's safe to do so now.
>> Yes, new applications can check for the feature flag. What about the existing
>
> apps? The existing applications which do not separate submit/complete
> threads may use io_uring_wait_cqes() or io_uring_wait_cqe_timeout() without
> submiting the requests. No one will do that now when the feature is supported.

Right, and I feel like I'm missing something here, what's the issue with
that? As far as the application is concerned, a different mechanism may be
used to achieve the timeout, but it should work in the same way.

So please explain this as clearly as you can, as I'm probably missing
something...

>>> Oh, yes, I haven't considering that before. So could I add this feature
>>> bit to io_uring.flags. Any suggestion?
>>
>> Either that, or we add this (and add pad that we can use later) and just
>> say that for the next release you have to re-compile against the lib.
>> That will break existing applications, unless they are recompiled... But
>> it might not be a bad idea to do so, just so we can pad io_uring out a
>> little bit to provide for future flexibility.
>>
> Agree about that. So should we increase the major version after adding the
> feature flag and some pad?

I think so, also a good time to think about other cases where that might be
useful.

But I think we need to flesh out the API first, as that might impact things.

-- 
Jens Axboe