ringzero

Pure C io_uring TCP server using liburing. Built for benchmark comparison against zerg (C# io_uring).

Architecture

• 1 acceptor thread — multishot accept, round-robin to N reactors via lock-free SPSC queue
• N reactor threads — each with its own io_uring (SINGLE_ISSUER | DEFER_TASKRUN), provided buffer ring, and multishot recv
• Synchronous handler — runs inline on the reactor thread, zero cross-thread signaling

All liburing static inline functions are fully inlined by gcc at -O2.

Building

Requires liburing (liburing-dev on Debian/Ubuntu).

make

Produces:

• rgzero — server binary (linked against libringzero.so with $ORIGIN rpath)
• libringzero.so — shared library
• libringzero.a — static library

Running

./rgzero [reactor_count]    # default: 12 reactors

Listens on 0.0.0.0:8080. Stop with Ctrl+C or SIGTERM.

Project Structure

include/
  constants.h    — UD packing macros, ring/buffer tunables
  queue.h        — lock-free SPSC queue (acceptor → reactor fd handoff)
  listener.h     — TCP listen socket setup
  connection.h   — per-connection state, write buffering
  reactor.h      — reactor struct, event loop, provided buffer ring
  acceptor.h     — multishot accept loop
  engine.h       — thread orchestration (acceptor + N reactors)
  ringzero.h     — umbrella header

src/lib/
  engine.c       — spawn/join acceptor and reactor threads
  acceptor.c     — multishot accept, round-robin fd distribution
  reactor.c      — io_uring event loop (recv/send/cancel)
  connection.c   — write slab management, flush logic
  listener.c     — SO_REUSEADDR/PORT, nonblocking listen socket

src/app/
  main.c         — CLI, signal handling, plaintext HTTP handler

Key io_uring patterns

• Single syscall loop — io_uring_submit_and_wait_timeout submits + harvests in one kernel entry
• Zero-copy recv — provided buffer rings (io_uring_setup_buf_ring), kernel writes directly into pre-registered memory
• Multishot recv — one SQE arms continuous receive, no re-arming per event
• Batch CQE processing — io_uring_peek_batch_cqe harvests up to 4096 completions at once
• User-data packing — PACK_UD(kind, fd) encodes operation type + fd in 64 bits