Re: [PATCH liburing v3 0/2] add support for new timeout feature

[Jiufei Xue <jiufei.xue@linux.alibaba.com>]

On 2020/8/5 上午11:28, Carter Li 李通洲 wrote:
>> diff --git a/src/include/liburing.h b/src/include/liburing.h
>> index 0505a4f..82c2980 100644
>> --- a/src/include/liburing.h
>> +++ b/src/include/liburing.h
>> @@ -56,6 +56,9 @@ struct io_uring {
>>  	struct io_uring_sq sq;
>>  	struct io_uring_cq cq;
>>  	unsigned flags;
>> +	unsigned flags_internal;
>> +	unsigned features;
>> +	unsigned pad[4];
>>  	int ring_fd;
>>  };
> 
> Won't it break existing code runs on newer kernel?

io_uring is a structure that used in userspace. It breaks the API
with existing compiled application. So I have changed the soname
to 2.0.7.

And for syscall io_uring_enter(), I have added a new feature bit
IORING_FEAT_GETEVENTS_TIMEOUT and io_uring_enter() flag
IORING_ENTER_GETEVENTS_TIMEOUT. Here are 3 cases below:

1) old liburing <-> new kernel: old liburing can not pass the flag
   IORING_ENTER_GETEVENTS_TIMEOUT, so new kernel will parse the arguments
   the original way.

2) new liburing <-> old kernel: feature IORING_FEAT_GETEVENTS_TIMEOUT
   not supported, liburing will do things like before.

3) new liburing <-> new kernel: feature IORING_FEAT_GETEVENTS_TIMEOUT
   supported, liburing pass the new arguments with the flag
   IORING_ENTER_GETEVENTS_TIMEOUT which helps kernel parse the arguments
   correctly.

Thanks,
Jiufei.

> Won't it break code compiled with new liburing but runs on older kernel?
> 

> IMO In this case, a new syscall `io_uring_setup2` is required at least.
> 
> Regards,
> Carter
>