kevy-uring

Pure-Rust io_uring bindings against the Linux kernel ABI. A completion-based engine: submit reads / writes / accepts into a shared submission queue (SQ), then reap their results from the completion queue (CQ) — batching many operations into one io_uring_enter syscall.

• io_uring_setup / io_uring_enter / io_uring_register issued as raw syscalls. No liburing C dependency, no libc crate, no third-party Rust dependency.
• SQ / CQ / SQE regions mmap'd and driven through the documented head/tail cursors, with the appropriate Acquire / Release fences against the kernel-side updates.
• Multishot recv + provided-buffer ring (kernel 5.19+) supported, the same primitives a thread-per-core reactor needs.
• Linux-only: on every other target the crate compiles to an empty module that any caller can cfg-gate.

use std::io::Write;
use std::net::TcpListener;
use std::os::fd::AsRawFd;
use kevy_uring::IoUring;

let listener = TcpListener::bind("127.0.0.1:0")?;
let mut ring = IoUring::new(64)?;
assert!(ring.prep_accept(listener.as_raw_fd(), /* user_data */ 1));
ring.submit_and_wait(1)?;
ring.for_each_completion(|c| println!("accepted fd {}", c.res));
# Ok::<(), std::io::Error>(())

Why a separate crate

Carved out of kevy-sys so the engine can be reused independently of kevy's network internals (sockets, readiness pollers). The bindings are generic Linux infrastructure; nothing here is specific to kevy's command surface.

Part of the kevy key–value server.

Safety

The shared ring cursors are accessed as AtomicU32 over the mmap'd memory (the kernel is the other party): the producer publishes the SQ tail with Release and reads the SQ head with Acquire; the consumer reads the CQ tail with Acquire and publishes the CQ head with Release. IoUring owns its ring fd and three mappings, freed on drop.

unsafe is confined to the FFI module's extern "C" declarations and the wrappers that read / write through the mmap'd ring memory. Every unsafe block carries a SAFETY: comment naming the kernel invariant it relies on.

License

Licensed under either of MIT or Apache-2.0 at your option.