C++ 被 Linux 系统和异步线程上的 epoll 和 socket fd 混淆

Question

I am trying to understand how to use the epoll library (Linux) in combination with socket file descriptors. There is limited information about epoll available online as far as I can tell. So far the most useful resource I have found is this:

https://www.suchprogramming.com/epoll-in-3-easy-steps/

Which gives a complete working example of epoll however it is used with the stdin file descriptor. (Elsewhere I read that epoll can not be used with stdin , stdout and stderr . I assume that information was wrong.)

I have tried to build on this with a MWE.

#include <string>
#include <sstream>
#include <iostream>
#include <fstream>
#include <string.h>
#include <unistd.h>
#include <future>
#include <sys/socket.h>
#include <netinet/in.h>
#include <sys/epoll.h>


int thread_function(int accept_sock_fd)
{
    char *read_buffer = new char[1024];

    unsigned int read_length = 0;
    std::string read_string = "";
    for(;;)
    {
        unsigned int t_read_length = read(accept_sock_fd, read_buffer, 1024 - 1);
        //std::cout << "t_read_length=" << t_read_length << std::endl;
        //std::cout << read_buffer << std::endl;
        read_string += std::string(read_buffer);
        read_length += t_read_length;

        // dodgy ?
        // tries to detect end of data using new lines
        // but what if client does not send new line char?
        if(read_buffer[t_read_length - 1] == '\n')
        {
            break;
        }
    }
    std::cout << read_string << std::endl;
    //std::cout << read_length << std::endl;
    //std::cout << read_buffer << std::endl;

    std::string status_line = "HTTP/1.0 200 OK\r\n\r\n";
    std::string html_lines = "Hello World\r\n\r\n";
    std::string write_string = status_line + html_lines;
    //std::cout << write_string << std::endl;
    unsigned int write_length = write(accept_sock_fd, write_string.data(), write_string.length());
    //std::cout << "write_length=" << write_length << std::endl;
    close(accept_sock_fd);

    delete [] read_buffer;

    return 0;
}


int main()
{

    // read configuration file
    //Config config;
    //config.ReadFromFile("config.txt");
    //std::cout << config << std::endl;

    unsigned int num_threads = 4; //config.GetNumThreads();


    ///////////////////////////////////////////////////////////////////////////
    // init epoll
    //
    //

    int epoll_fd = epoll_create(num_threads);
    if(epoll_fd < 0)
    {
        std::cout << "epoll error" << std::endl;
    }

    struct epoll_event event;
    struct epoll_event *events = new struct epoll_event[num_threads * sizeof(struct epoll_event)];
    


    ///////////////////////////////////////////////////////////////////////////
    // create socket fd
    //
    //

    int sock_fd = socket(AF_INET, SOCK_STREAM, 0);
    if(sock_fd == 0)
    {
        std::cout << "Error creating sock_fd" << std::endl;
    }
//    int setsockopt(sockfd, 

    struct sockaddr_in server_address;
    memset(&server_address, 0, sizeof(struct sockaddr_in));
    server_address.sin_family = AF_INET;
    server_address.sin_addr.s_addr = INADDR_ANY;
    server_address.sin_port = htons(55555 /*config.GetPort()*/);

    bind(sock_fd, (struct sockaddr *)&server_address, sizeof(server_address));
    if(bind < 0)
    {
        std::cout << "Error: bind" << std::endl;
    }

    if(listen(sock_fd, 10) < 0)
    {
        std::cout << "Error: listen" << std::endl;
    }


    ///////////////////////////////////////////////////////////////////////////
    // continue epoll setup
    //
    //

    event.events = EPOLLIN; // | EPOLLPRI | EPOLLERR | EPOLLHUP;
    event.data.fd = 0; //client_sock;
    if(epoll_ctl(epoll_fd, EPOLL_CTL_ADD, sock_fd, &event)) // not sure about this
    {
        std::cout << "Failed to add epoll" << std::endl;
        return 1;
    }


    for(;;)
    {
        int n_fd = epoll_wait(epoll_fd, events, num_threads, 0);
        if(n_fd < 0)
        {
            std::cout << "die" << std::endl;
            break;
        }

        for(int i = 0; i < n_fd; ++ i)
        {
            int fd = events[i].data.fd;
            
            // accpt connection
            int accept_sock_fd = 0;
            struct sockaddr client_address;
            int addrlen = sizeof(client_address);
            accept_sock_fd = accept(fd, NULL, NULL); // not sure what goes here
            accept_sock_fd = accept(sock_fd, NULL, NULL); // not sure what goes here
            if(accept_sock_fd < 0)
            {
                std::cout << "Error: accept" << std::endl;
            }

            // do something to handle data on accept_sock_fd
            std::future<int> thread_future = std::async(thread_function, accept_sock_fd);
            int ret = thread_future.get();
            std::cout << ret << std::endl;

            // close connection
            close(accept_sock_fd);
            
        }
    }


    delete [] events;

    close(sock_fd);


    if(close(epoll_fd))
    {
        std::cout << "error: failed to close epoll fd" << std::endl;
    }


    return 0;
}

I have searched pretty extensively for further information on how to combine epoll with sockets, but there isn't much information about this. I suppose it is quite a niche area.

Ideally if anyone knows of any good resources (even a book) where I can find more information about this and thus solve the problem myself that would be great, otherwise some advice on how to proceed would be appreciated also.

I am mostly confused as to how I watch for events on the socket fd and also how I then accept events and handle the new fd returned by accept .

See these lines in the above code:

// correct fd
epoll_ctl(epoll_fd, EPOLL_CTL_ADD, /*the_server_socket_fd*/, &event);
...
// not sure about this one, particularly the waiting time of 0
for(;;)
    epoll_wait(epoll_fd, events, /*num_threads*/, 0);
    
    for(i ...)
        fd = events[i].data.fd; // what fd is this? what does it mean?
        accept(/*something here, but what*/ fd OR the_server_socket_fd ?, NULL, NULL);
        // launch thread with retuned value from accept, probably

Answer 1

Using epoll is not so much different with using poll . The tricky thing is that keep track which client owns file descriptor returned by epoll (if there are a lot of clients).

---

How to handle socket file asynchronously with epoll (in this case as TCP socket server).

1. Open an epoll file descriptor with epoll_create(2) .
2. Create a TCP socket with socket(2) , bind(2) and listen(2) .
3. Add the main TCP socket file descriptor to epoll with epoll_ctl + EPOLL_CTL_ADD .
4. Call epoll_wait inside a loop, the program will sleep on epoll_wait , the kernel will wake the program up when there is event coming from monitored file descriptors or when timeout is reached .
5. If epoll_wait returns a value greater than zero, then you need to decide, what file descriptor returned by epoll.

---

5.1. If it is main TCP file descriptor, then you need to accept(2) . And then add the client file descriptor returned by accept(2) to epoll with epoll_ctl + EPOLL_CTL_ADD . Other

5.2. If it is client file descriptor, then you need to call recv(2) and do any action you want with that client.

At step 5.2, if you see EPOLLHUP in the events[i].events , then the client has closed its connection and you need to call epoll_ctl + EPOLL_CTL_DEL and close the client file descriptor (it may be safe not to call epoll_ctl + EPOLL_CTL_DEL and just close the client file descriptor, but I prefer to delete it from epoll first).

---

6. Goto step 4.

Detailed mechanism about how to determine which client owns file descriptor returned by epoll can be seen from server.c code below.

---

Flowchart

Flowchart can make it more clear.

---

Working example you can try at home

server.c

/*
 * https://stackoverflow.com/questions/66916835/c-confused-by-epoll-and-socket-fd-on-linux-systems-and-async-threads
 */
#include <stdio.h>
#include <errno.h>
#include <unistd.h>
#include <string.h>
#include <stdbool.h>
#include <sys/epoll.h>
#include <arpa/inet.h>
#include <netinet/tcp.h>

#define PRERF "(errno=%d) %s\n"
#define PREAR(NUM) NUM, strerror(NUM)
#define EPOLL_MAP_TO_NOP (0u)
#define EPOLL_MAP_SHIFT  (1u) /* Shift to cover reserved value MAP_TO_NOP */

struct client_slot {
    bool                is_used;
    int                 client_fd;
    char                src_ip[sizeof("xxx.xxx.xxx.xxx")];
    uint16_t            src_port;
    uint16_t            my_index;
};

struct tcp_state {
    bool                stop;
    int                 tcp_fd;
    int                 epoll_fd;
    uint16_t            client_c;
    struct client_slot  clients[10];

    /*
     * Map the file descriptor to client_slot array index
     * Note: We assume there is no file descriptor greater than 10000.
     *
     * You must adjust this in production.
     */
    uint32_t            client_map[10000];
};


static int my_epoll_add(int epoll_fd, int fd, uint32_t events)
{
    int err;
    struct epoll_event event;

    /* Shut the valgrind up! */
    memset(&event, 0, sizeof(struct epoll_event));

    event.events  = events;
    event.data.fd = fd;
    if (epoll_ctl(epoll_fd, EPOLL_CTL_ADD, fd, &event) < 0) {
        err = errno;
        printf("epoll_ctl(EPOLL_CTL_ADD): " PRERF, PREAR(err));
        return -1;
    }
    return 0;
}



static int my_epoll_delete(int epoll_fd, int fd)
{
    int err;

    if (epoll_ctl(epoll_fd, EPOLL_CTL_DEL, fd, NULL) < 0) {
        err = errno;
        printf("epoll_ctl(EPOLL_CTL_DEL): " PRERF, PREAR(err));
        return -1;
    }
    return 0;
}


static const char *convert_addr_ntop(struct sockaddr_in *addr, char *src_ip_buf)
{
    int err;
    const char *ret;
    in_addr_t saddr = addr->sin_addr.s_addr;

    ret = inet_ntop(AF_INET, &saddr, src_ip_buf, sizeof("xxx.xxx.xxx.xxx"));
    if (ret == NULL) {
        err = errno;
        err = err ? err : EINVAL;
        printf("inet_ntop(): " PRERF, PREAR(err));
        return NULL;
    }

    return ret;
}


static int accept_new_client(int tcp_fd, struct tcp_state *state)
{
    int err;
    int client_fd;
    struct sockaddr_in addr;
    socklen_t addr_len = sizeof(addr);
    uint16_t src_port;
    const char *src_ip;
    char src_ip_buf[sizeof("xxx.xxx.xxx.xxx")];
    const size_t client_slot_num = sizeof(state->clients) / sizeof(*state->clients);


    memset(&addr, 0, sizeof(addr));
    client_fd = accept(tcp_fd, (struct sockaddr *)&addr, &addr_len);
    if (client_fd < 0) {
        err = errno;
        if (err == EAGAIN)
            return 0;

        /* Error */
        printf("accept(): " PRERF, PREAR(err));
        return -1;
    }

    src_port = ntohs(addr.sin_port);
    src_ip   = convert_addr_ntop(&addr, src_ip_buf);
    if (!src_ip) {
        printf("Cannot parse source address\n");
        goto out_close;
    }


    /*
     * Find unused client slot.
     *
     * In real world application, you don't want to iterate
     * the whole array, instead you can use stack data structure
     * to retrieve unused index in O(1).
     *
     */
    for (size_t i = 0; i < client_slot_num; i++) {
        struct client_slot *client = &state->clients[i];

        if (!client->is_used) {
            /*
             * We found unused slot.
             */

            client->client_fd = client_fd;
            memcpy(client->src_ip, src_ip_buf, sizeof(src_ip_buf));
            client->src_port = src_port;
            client->is_used = true;
            client->my_index = i;

            /*
             * We map the client_fd to client array index that we accept
             * here.
             */
            state->client_map[client_fd] = client->my_index + EPOLL_MAP_SHIFT;

            /*
             * Let's tell to `epoll` to monitor this client file descriptor.
             */
            my_epoll_add(state->epoll_fd, client_fd, EPOLLIN | EPOLLPRI);

            printf("Client %s:%u has been accepted!\n", src_ip, src_port);
            return 0;
        }
    }
    printf("Sorry, can't accept more client at the moment, slot is full\n");


out_close:
    close(client_fd);
    return 0;
}


static void handle_client_event(int client_fd, uint32_t revents,
                                struct tcp_state *state)
{
    int err;
    ssize_t recv_ret;
    char buffer[1024];
    const uint32_t err_mask = EPOLLERR | EPOLLHUP;
    /*
     * Read the mapped value to get client index.
     */
    uint32_t index = state->client_map[client_fd] - EPOLL_MAP_SHIFT;
    struct client_slot *client = &state->clients[index];

    if (revents & err_mask)
        goto close_conn;

    recv_ret = recv(client_fd, buffer, sizeof(buffer), 0);
    if (recv_ret == 0)
        goto close_conn;

    if (recv_ret < 0) {
        err = errno;
        if (err == EAGAIN)
            return;

        /* Error */
        printf("recv(): " PRERF, PREAR(err));
        goto close_conn;
    }


    /*
     * Safe printing
     */
    buffer[recv_ret] = '\0';
    if (buffer[recv_ret - 1] == '\n') {
        buffer[recv_ret - 1] = '\0';
    }

    printf("Client %s:%u sends: \"%s\"\n", client->src_ip, client->src_port,
           buffer);
    return;


close_conn:
    printf("Client %s:%u has closed its connection\n", client->src_ip,
           client->src_port);
    my_epoll_delete(state->epoll_fd, client_fd);
    close(client_fd);
    client->is_used = false;
    return;
}


static int event_loop(struct tcp_state *state)
{
    int err;
    int ret = 0;
    int timeout = 3000; /* in milliseconds */
    int maxevents = 32;
    int epoll_ret;
    int epoll_fd = state->epoll_fd;
    struct epoll_event events[32];

    printf("Entering event loop...\n");

    while (!state->stop) {

        /*
         * I sleep on `epoll_wait` and the kernel will wake me up
         * when event comes to my monitored file descriptors, or
         * when the timeout reached.
         */
        epoll_ret = epoll_wait(epoll_fd, events, maxevents, timeout);


        if (epoll_ret == 0) {
            /*
             *`epoll_wait` reached its timeout
             */
            printf("I don't see any event within %d milliseconds\n", timeout);
            continue;
        }


        if (epoll_ret == -1) {
            err = errno;
            if (err == EINTR) {
                printf("Something interrupted me!\n");
                continue;
            }

            /* Error */
            ret = -1;
            printf("epoll_wait(): " PRERF, PREAR(err));
            break;
        }


        for (int i = 0; i < epoll_ret; i++) {
            int fd = events[i].data.fd;

            if (fd == state->tcp_fd) {
                /*
                 * A new client is connecting to us...
                 */
                if (accept_new_client(fd, state) < 0) {
                    ret = -1;
                    goto out;
                }
                continue;
            }

            /*
             * We have event(s) from client, let's call `recv()` to read it.
             */
            handle_client_event(fd, events[i].events, state);
        }
    }

out:
    return ret;
}


static int init_epoll(struct tcp_state *state)
{
    int err;
    int epoll_fd;

    printf("Initializing epoll_fd...\n");

    /* The epoll_create argument is ignored on modern Linux */
    epoll_fd = epoll_create(255);
    if (epoll_fd < 0) {
        err = errno;
        printf("epoll_create(): " PRERF, PREAR(err));
        return -1;
    }

    state->epoll_fd = epoll_fd;
    return 0;
}


static int init_socket(struct tcp_state *state)
{
    int ret;
    int err;
    int tcp_fd = -1;
    struct sockaddr_in addr;
    socklen_t addr_len = sizeof(addr);
    const char *bind_addr = "0.0.0.0";
    uint16_t bind_port = 1234;

    printf("Creating TCP socket...\n");
    tcp_fd = socket(AF_INET, SOCK_STREAM | SOCK_NONBLOCK, IPPROTO_TCP);
    if (tcp_fd < 0) {
        err = errno;
        printf("socket(): " PRERF, PREAR(err));
        return -1;
    }

    memset(&addr, 0, sizeof(addr));
    addr.sin_family = AF_INET;
    addr.sin_port = htons(bind_port);
    addr.sin_addr.s_addr = inet_addr(bind_addr);

    ret = bind(tcp_fd, (struct sockaddr *)&addr, addr_len);
    if (ret < 0) {
        ret = -1;
        err = errno;
        printf("bind(): " PRERF, PREAR(err));
        goto out;
    }

    ret = listen(tcp_fd, 10);
    if (ret < 0) {
        ret = -1;
        err = errno;
        printf("listen(): " PRERF, PREAR(err));
        goto out;
    }

    /*
     * Add `tcp_fd` to epoll monitoring.
     *
     * If epoll returned tcp_fd in `events` then a client is
     * trying to connect to us.
     */
    ret = my_epoll_add(state->epoll_fd, tcp_fd, EPOLLIN | EPOLLPRI);
    if (ret < 0) {
        ret = -1;
        goto out;
    }

    printf("Listening on %s:%u...\n", bind_addr, bind_port);
    state->tcp_fd = tcp_fd;
    return 0;
out:
    close(tcp_fd);
    return ret;
}


static void init_state(struct tcp_state *state)
{
    const size_t client_slot_num = sizeof(state->clients) / sizeof(*state->clients);
    const uint16_t client_map_num = sizeof(state->client_map) / sizeof(*state->client_map);

    for (size_t i = 0; i < client_slot_num; i++) {
        state->clients[i].is_used = false;
        state->clients[i].client_fd = -1;
    }

    for (uint16_t i = 0; i < client_map_num; i++) {
        state->client_map[i] = EPOLL_MAP_TO_NOP;
    }
}


int main(void)
{
    int ret;
    struct tcp_state state;

    init_state(&state);

    ret = init_epoll(&state);
    if (ret != 0)
        goto out;


    ret = init_socket(&state);
    if (ret != 0)
        goto out;


    state.stop = false;

    ret = event_loop(&state);

out:
    /*
     * You should write a cleaner here.
     *
     * Close all client file descriptors and release
     * some resources you may have.
     *
     * You may also want to set interrupt handler
     * before the event_loop.
     *
     * For example, if you get SIGINT or SIGTERM
     * set `state->stop` to true, so that it exits
     * gracefully.
     */
    return ret;
}

---

test.php (PHP script to simulate client)

I am too lazy to write a client TCP socket in C. So I use PHP.

<?php

function main(): int
{
    $sock = socket_create(AF_INET, SOCK_STREAM, 0);
    $conn = socket_connect($sock, "127.0.0.1", 1234);
    if (!$conn)
        return 1;
    socket_write($sock, "AAAAAAAA\n", 9);
    sleep(1);
    socket_write($sock, "BBBBBBBB\n", 9);
    sleep(1);
    socket_write($sock, "CCCCCCCC\n", 9);
    sleep(1);
    socket_close($sock);
    return 0;
}

exit(main());

---

Compile and run the server

ammarfaizi2@integral:~/ex$ gcc -Wall -Wextra -ggdb3 server.c -o server
ammarfaizi2@integral:~/ex$ ./server
Initializing epoll_fd...
Creating TCP socket...
Listening on 0.0.0.0:1234...
Entering event loop...

---

Simulate multiple clients with test.php

ammarfaizi2@integral:~$ for i in {1..10}; do php test.php & done;
[1] 14214
[2] 14215
[3] 14216
[4] 14217
[5] 14218
[6] 14219
[7] 14220
[8] 14221
[9] 14222
[10] 14223
ammarfaizi2@integral:~$ 

---

Server output when clients connect

ammarfaizi2@integral:~/ex$ ./server
Initializing epoll_fd...
Creating TCP socket...
Listening on 0.0.0.0:1234...
Entering event loop...
Client 127.0.0.1:60866 has been accepted!
Client 127.0.0.1:60866 sends: "AAAAAAAA"
Client 127.0.0.1:60868 has been accepted!
Client 127.0.0.1:60868 sends: "AAAAAAAA"
Client 127.0.0.1:60870 has been accepted!
Client 127.0.0.1:60872 has been accepted!
Client 127.0.0.1:60870 sends: "AAAAAAAA"
Client 127.0.0.1:60872 sends: "AAAAAAAA"
Client 127.0.0.1:60874 has been accepted!
Client 127.0.0.1:60874 sends: "AAAAAAAA"
Client 127.0.0.1:60876 has been accepted!
Client 127.0.0.1:60878 has been accepted!
Client 127.0.0.1:60878 sends: "AAAAAAAA"
Client 127.0.0.1:60876 sends: "AAAAAAAA"
Client 127.0.0.1:60880 has been accepted!
Client 127.0.0.1:60880 sends: "AAAAAAAA"
Client 127.0.0.1:60882 has been accepted!
Client 127.0.0.1:60882 sends: "AAAAAAAA"
Client 127.0.0.1:60884 has been accepted!
Client 127.0.0.1:60884 sends: "AAAAAAAA"
Client 127.0.0.1:60866 sends: "BBBBBBBB"
Client 127.0.0.1:60868 sends: "BBBBBBBB"
Client 127.0.0.1:60870 sends: "BBBBBBBB"
Client 127.0.0.1:60872 sends: "BBBBBBBB"
Client 127.0.0.1:60874 sends: "BBBBBBBB"
Client 127.0.0.1:60878 sends: "BBBBBBBB"
Client 127.0.0.1:60876 sends: "BBBBBBBB"
Client 127.0.0.1:60880 sends: "BBBBBBBB"
Client 127.0.0.1:60882 sends: "BBBBBBBB"
Client 127.0.0.1:60884 sends: "BBBBBBBB"
Client 127.0.0.1:60866 sends: "CCCCCCCC"
Client 127.0.0.1:60868 sends: "CCCCCCCC"
Client 127.0.0.1:60870 sends: "CCCCCCCC"
Client 127.0.0.1:60872 sends: "CCCCCCCC"
Client 127.0.0.1:60874 sends: "CCCCCCCC"
Client 127.0.0.1:60878 sends: "CCCCCCCC"
Client 127.0.0.1:60876 sends: "CCCCCCCC"
Client 127.0.0.1:60880 sends: "CCCCCCCC"
Client 127.0.0.1:60882 sends: "CCCCCCCC"
Client 127.0.0.1:60884 sends: "CCCCCCCC"
Client 127.0.0.1:60866 has closed its connection
Client 127.0.0.1:60868 has closed its connection
Client 127.0.0.1:60870 has closed its connection
Client 127.0.0.1:60872 has closed its connection
Client 127.0.0.1:60874 has closed its connection
Client 127.0.0.1:60878 has closed its connection
Client 127.0.0.1:60876 has closed its connection
Client 127.0.0.1:60880 has closed its connection
Client 127.0.0.1:60882 has closed its connection
Client 127.0.0.1:60884 has closed its connection

Answer 2

Assuming you have already open server socket you can use code below. It is based on man epoll accessible here .

// assume you have opened listen_sock properly
// int listen_sock;
// 
int epollfd = epoll_create1(0);
if (epollfd == -1) {
  exit(1);
}

struct epoll_event ev;
ev.events = EPOLLIN;
ev.data.fd = listen_sock;
if (epoll_ctl(epollfd, EPOLL_CTL_ADD, listen_sock, &ev) == -1) {
  exit(2);
}

for (;;) {
  #define MAX_EVENTS 64
  struct epoll_event events[MAX_EVENTS];
  int events_count = epoll_wait(epollfd, events, MAX_EVENTS, -1);
  if (n == -1) {
    exit(3);
  }

  for (int n = 0; n < events_count; ++ n) {
    if (events[n].data.fd == listen_sock) {
      struct sockaddr_un addr;
      socklen_t addrlen;
      int socket = accept(listen_sock, (struct sockaddr *) &addr, &addrlen);
    }
  }
}