Linux io_uring 核心机制剖析：从双 Ring Buffer 原理到高性能服务端改造实践

+-------------------+                      +-------------------+
|  User Space App   |                      |    Kernel Core    |
+---------+---------+                      +---------+---------+
          |                                          |
          |  [1] 填充 SQE (Submission Queue Entry)    |
          v                                          |
    +-----------+      Shared Memory (mmap)     +----+------+
    |    SQ     | ============================> | SQ Kernel | (读取并派发任务)
    +-----------+                               +-----------+
                                                      |
    +-----------+      Shared Memory (mmap)     +----+------+
    |    CQ     | <============================ | CQ Kernel | (完成写回 CQE)
    +-----------+                               +-----------+
          ^                                          |
          |  [2] 消费 CQE (Completion Queue Entry)    |
          |                                          |

$ strace -c -p $(pidof epoll_server)
% time     seconds  usecs/call     calls    errors syscall
------ ----------- ----------- --------- --------- ----------------
 42.18    0.852100           3    284033           read
 35.12    0.709320           2    284030           write
 22.70    0.458440           5     91688           epoll_wait
------ ----------- ----------- --------- --------- ----------------
100.00    2.019860                659751           total

$ strace -c -p $(pidof io_uring_server)
% time     seconds  usecs/call     calls    errors syscall
------ ----------- ----------- --------- --------- ----------------
------ ----------- ----------- --------- --------- ----------------
100.00    0.000000                     0           total

# epoll 模式下：
382,109  context-switches          #    0.038 M/sec

# io_uring SQPOLL 模式下：
    124  context-switches          #    0.000 M/sec

#include <stdio.h>
#include <stdlib.h>
#include <netinet/in.h>
#include <liburing.h>
#include <string.h>
#include <unistd.h>

#define QUEUE_DEPTH 1024
#define BUF_SIZE 1024

enum { EVENT_ACCEPT, EVENT_READ, EVENT_WRITE };

struct conn_info {
    int fd;
    int event_type;
    char buffer[BUF_SIZE];
    int bytes;
};

int main() {
    struct io_uring ring;
    // 1. 初始化 io_uring，队列深度 1024
    io_uring_queue_init(QUEUE_DEPTH, &ring, 0);

    int listen_fd = socket(AF_INET, SOCK_STREAM, 0);
    struct sockaddr_in serv_addr;
    memset(&serv_addr, 0, sizeof(serv_addr));
    serv_addr.sin_family = AF_INET;
    serv_addr.sin_port = htons(8080);
    serv_addr.sin_addr.s_addr = INADDR_ANY;

    bind(listen_fd, (struct sockaddr *)&serv_addr, sizeof(serv_addr));
    listen(listen_fd, 512);

    // 提交第一个 Accept 任务
    struct io_uring_sqe *sqe = io_uring_get_sqe(&ring);
    struct conn_info accept_conn = { .fd = listen_fd, .event_type = EVENT_ACCEPT };
    io_uring_prep_accept(sqe, listen_fd, NULL, NULL, 0);
    io_uring_sqe_set_data(sqe, &accept_conn);
    io_uring_submit(&ring);

    printf("Server listening on port 8080...\n");

    while (1) {
        struct io_uring_cqe *cqe;
        // 等待 CQ 产生完成事件
        int ret = io_uring_wait_cqe(&ring, &cqe);
        if (ret < 0) break;

        struct conn_info *conn = (struct conn_info *)io_uring_cqe_get_data(cqe);
        int res = cqe->res; // 返回值（如 read 到的字节数，或 accept 的新 fd）

        if (conn->event_type == EVENT_ACCEPT) {
            int client_fd = res;
            // 为新连接提交异步 Read 请求
            struct conn_info *read_conn = malloc(sizeof(struct conn_info));
            read_conn->fd = client_fd;
            read_conn->event_type = EVENT_READ;

            sqe = io_uring_get_sqe(&ring);
            io_uring_prep_read(sqe, client_fd, read_conn->buffer, BUF_SIZE, 0);
            io_uring_sqe_set_data(sqe, read_conn);

            // 重新挂起 Accept 监听
            sqe = io_uring_get_sqe(&ring);
            io_uring_prep_accept(sqe, listen_fd, NULL, NULL, 0);
            io_uring_sqe_set_data(sqe, &accept_conn);

            io_uring_submit(&ring);
        } 
        else if (conn->event_type == EVENT_READ) {
            if (res <= 0) {
                close(conn->fd);
                free(conn);
            } else {
                // Echo 回写数据
                conn->bytes = res;
                conn->event_type = EVENT_WRITE;
                sqe = io_uring_get_sqe(&ring);
                io_uring_prep_write(sqe, conn->fd, conn->buffer, conn->bytes, 0);
                io_uring_sqe_set_data(sqe, conn);
                io_uring_submit(&ring);
            }
        } 
        else if (conn->event_type == EVENT_WRITE) {
            // 写完后继续读
            conn->event_type = EVENT_READ;
            sqe = io_uring_get_sqe(&ring);
            io_uring_prep_read(sqe, conn->fd, conn->buffer, BUF_SIZE, 0);
            io_uring_sqe_set_data(sqe, conn);
            io_uring_submit(&ring);
        }

        io_uring_cqe_seen(&ring, cqe); // 标记 CQE 已经被消费
    }

    io_uring_queue_exit(&ring);
    return 0;
}

int files[2] = { client_fd, server_fd };
io_uring_register_files(&ring, files, 2);
// 在 sqe 中标记 IOSQE_FIXED_FILE
sqe->flags |= IOSQE_FIXED_FILE;