Is there a limit to the io uring queue size?

[ioquatix]

What if there are way more tasks produced than consumed?

[ioquatix]

The uring backend has a fixed sized submission and completion queue, defined here:

https://github.com/socketry/event/blob/ee7f6bfa0b4a1df20af91639a73a23a241238a2c/ext/event/selector/uring.c#L34

In order to submit a system call, we must first get a submission queue entry to fill in with the details:

https://github.com/socketry/event/blob/ee7f6bfa0b4a1df20af91639a73a23a241238a2c/ext/event/selector/uring.c#L211-L222

If we couldn't get a sqe because the submission queue was overflowing, we call io_uring_submit_now(data):

https://github.com/socketry/event/blob/ee7f6bfa0b4a1df20af91639a73a23a241238a2c/ext/event/selector/uring.c#L188-L204

This can in fact yield the current fiber to implement a form of internal queue load balancing, i.e. if you have too many fibers trying to submit system calls, they will start to yield and try again later (on the next iteration of the run loop):

https://github.com/socketry/event/blob/ee7f6bfa0b4a1df20af91639a73a23a241238a2c/ext/event/selector/selector.h#L110-L114

Followed by:

https://github.com/socketry/event/blob/master/ext/event/selector/selector.c#L172-L193

This code puts the current fiber into the event selector's ready list, and on the next iteration of the event loop it will be resumed, which is done by Event_Selector_queue_flush:

https://github.com/socketry/event/blob/master/ext/event/selector/selector.c#L257-L278

This ensures that the interface will correctly handle any arbitrary number of tasks without overloading the io_uring queues.