Question about how liburing works

[tickox]

Hello,

My questions may have been asked before but I can't find an answer that suits me.

1)
If I create 1 ring on 1 thread, how many cores are used ? 1 or 2 ?
If I create a new ring on another thread, do I use 2 or 4 cores ?
(to be sure of the number of cores I will have left without having context switch)
If I understood correctly, there will be 2 cores used by ring (one for me and the other for the kernel).

2)
A thread waits via io_uring_wait_cqe on a ring.
A second thread requests a sqe and publishes it on the same ring as thread 1.

Is this a bug risk? (only thread 1 gets the cqe but one or more other threads can publish sqe)
If yes, what is the approach often used for this kind of scenario?

3)
Using blocking/non-blocking sockets changes something or not?
I assume that you should use non-blocking sockets.

Thanks

[isilence]

1) If I create 1 ring on 1 thread, how many cores are used ? 1 or 2 ? If I create a new ring on another thread, do I use 2 or 4 cores ? (to be sure of the number of cores I will have left without having context switch) If I understood correctly, there will be 2 cores used by ring (one for me and the other for the kernel).

In case no SQPOLL is specified, and there is a good chance you don't need it, there will be no additional threads. IOW, all io_uring syscalls will work in a context of the task it was called from.

If you use SQPOLL, it'll create one additional in-kernel thread doing polling and eating CPU per ring. However, you can share SQPOLL backend between rings (see IORING_SETUP_ATTACH_WQ), in that case you will get only 1 additional thread per N rings.

2) A thread waits via io_uring_wait_cqe on a ring. A second thread requests a sqe and publishes it on the same ring as thread 1.

Is this a bug risk? (only thread 1 gets the cqe but one or more other threads can publish sqe) If yes, what is the approach often used for this kind of scenario?

From the io_uring perspective you can safely have one thread doing completions and one submissions. Note, there was a catch with older kernels using any flavour of *wait_cqe_timeout(), see IORING_FEAT_EXT_ARG.

3) Using blocking/non-blocking sockets changes something or not? I assume that you should use non-blocking sockets.

It does, if there is no data it's not going to wait but would just return -EAGAIN. That behaviour may be wanted in some cases, e.g. if you do polling (including with poll requests).

[tickox]

Thanks you !