High-performance I/O using io_uring via osmo_io

Andreas Eversberg and Harald Welte (sysmocom)

Playlists: 'osmodevcon2024' videos starting here / audio

Traditional socket I/O via `read/write/recvfrom/sendto/recvmsg/sendmsg` and friends creates a very high system call load. A highly-loaded osmo-bsc spends most of its time in syscall entry and syscall exit.

`io_uring` is a modern Linux kernel mechanism to avoid this syscall overhead. We have introduced the `osmo_io`API to libosmocore as a generic back-end for non-blocking/asynchronous I/O and a back-end for our classic `osmo_fd` / `poll` approach as well as a new backend for `io_uring`.

The talk will cover
* a very basic io_uring introduction
* a description of the osmo_io API
* the difficulties porting from osmo_fd to osmo_io
* status of porting various sub-systems over to osmo_io