Application                    Kernel
    │                            │
    ├── mmap SQ ring ───────────►│ submission queue (SQE)
    ├── mmap CQ ring ◄──────────│ completion queue (CQE)
    ├── io_uring_submit() ──────►│ processes SQEs
    └── io_uring_wait_cqe() ◄──│ posts CQEs

#include <liburing.h>
#include <fcntl.h>
#include <unistd.h>
#include <stdio.h>

int main(void) {
    struct io_uring ring;
    io_uring_queue_init(32, &ring, 0);

    int fd = open("test.txt", O_RDONLY);
    char buf[4096];

    struct io_uring_sqe *sqe = io_uring_get_sqe(&ring);
    io_uring_prep_read(sqe, fd, buf, sizeof(buf), 0);
    io_uring_sqe_set_data(sqe, (void *)1);

    io_uring_submit(&ring);

    struct io_uring_cqe *cqe;
    io_uring_wait_cqe(&ring, &cqe);
    if (cqe->res >= 0)
        printf("read %d bytes\n", cqe->res);
    else
        perror("read");

    io_uring_cqe_seen(&ring, cqe);
    io_uring_queue_exit(&ring);
    close(fd);
    return 0;
}

gcc -o uring_read uring_read.c -luring
./uring_read

// Batch multiple SQEs before single submit
for (int i = 0; i < n; i++) {
    sqe = io_uring_get_sqe(&ring);
    io_uring_prep_read(sqe, fds[i], bufs[i], sizes[i], 0);
    io_uring_sqe_set_data(sqe, (void *)(intptr_t)i);
}
io_uring_submit(&ring);

// Multi-shot accept — one SQE handles many connections
sqe = io_uring_get_sqe(&ring);
io_uring_prep_multishot_accept(sqe, listen_fd, NULL, NULL, 0);
io_uring_sqe_set_data(sqe, (void *)ACCEPT_COOKIE);

// Loop on CQEs until CQE_FLAG_MORE is clear
while (1) {
    io_uring_wait_cqe(&ring, &cqe);
    if (cqe->user_data == ACCEPT_COOKIE) {
        int client_fd = cqe->res;
        if (client_fd >= 0)
            handle_client(client_fd);
        if (!(cqe->flags & IORING_CQE_F_MORE))
            break;  // re-arm accept SQE
    }
    io_uring_cqe_seen(&ring, cqe);
}

#include <liburing.h>

#define BUF_GROUP 0
#define BUF_SIZE  4096
#define BUF_COUNT 64

struct io_uring_buf_ring *buf_ring;
char bufs[BUF_COUNT][BUF_SIZE];

// Setup buffer ring
buf_ring = io_uring_setup_buf_ring(&ring, BUF_COUNT, BUF_GROUP, 0, &ret);
for (int i = 0; i < BUF_COUNT; i++)
    io_uring_buf_ring_add(buf_ring, bufs[i], BUF_SIZE, i, BUF_COUNT - 1 - i, 0);
io_uring_buf_ring_advance(buf_ring, BUF_COUNT);

// Recv with provided buffers
sqe = io_uring_get_sqe(&ring);
io_uring_prep_recv_multishot(sqe, sock_fd, NULL, 0, 0);
sqe->buf_group = BUF_GROUP;

// Register files — avoids per-op fd table lookup
int fds[64];
io_uring_register_files(&ring, fds, 64);
io_uring_prep_read(sqe, 0, buf, size, offset);  // index 0, not fd

// Registered buffers — pinned, DMA-friendly
struct iovec iov = { .iov_base = buf, .iov_len = size };
io_uring_register_buffers(&ring, &iov, 1);
io_uring_prep_read_fixed(sqe, fd, buf, size, offset, 0);

sqe = io_uring_get_sqe(&ring);
io_uring_prep_send_zc(sqe, sock_fd, data, len, 0, 0);
// May complete with IORING_CQE_F_MORE — wait for notification CQE

use tokio_uring::fs::File;

#[tokio_uring::main]
async fn main() -> std::io::Result<()> {
    let file = File::open("test.txt").await?;
    let buf = vec![0u8; 4096];
    let (res, buf) = file.read_at(buf, 0).await;
    println!("read {} bytes", res?);
    Ok(())
}

cargo add tokio-uring

Choose io_uring when
├── Syscall overhead is measurable bottleneck
├── High connection count with multi-shot
└── Kernel 6.x+ with needed op support

Stay on epoll when
├── Simple app, portability matters
└── Team familiarity outweighs gains

# Restrict unprivileged io_uring (some distros disable by default)
sysctl kernel.io_uring_disabled   # 0=enabled, 1=disabled, 2=allow registered
sysctl kernel.io_uring_group      # gid allowed when disabled=2