Using ring per thread with multi-thread scheduler

[altanozlu]

To minimize impact on normal tokio application code base, tokio-uring should also support something other than thread per core model (or pinning futures to specific threads), such as work stealing scheduler.
I would like to share my approach to make this happen

• Each Tokio worker thread will have its own ring, accessed via thread-local storage.
• When an operation is submitted to the ring, that ring will handle its completion and task notification(waker).
• Futures can be safely shared across threads as it's used to be.
• Worker threads will park using io_uring_enter with min_complete=1 (can be changed since there is no need to call kernel for CQ if SQ is submitted)
• Since it will be modified by different threads (submitter and poller) operation information must be held thread safe.
• Operation cancellation will be asynchronous, potentially using IORING_OP_MSG_RING or an MPSC channel.
• This approach eliminates the need for epoll.

This design enables horizontal scalability by using a dedicated ring per thread and thread-local rings will make coding this implementation relatively straightforward.

I would like to get opinions about this approach.