nonblocking socket recv problem while using it with epoll

Question

hi there. i've got a problem: sometimes (not regularly) recv returns -1 and errno == EAGAIN while using epoll in edge-triggered mode. piece of code:

server_sock = startup(&port);

if ( (epollfd = epoll_create(4096)) < 0) {
    perror("epoll_create error");
    exit(EXIT_FAILURE);
}

ev.events = EPOLLIN | EPOLLET;
ev.data.fd = server_sock;
if (epoll_ctl(epollfd, EPOLL_CTL_ADD, server_sock, &ev) == -1) {
    perror("epoll_ctl: server_sock");
    exit(EXIT_FAILURE);
}

while (1) {
    int nfds = epoll_wait(epollfd, events, 4096, -1);
    if (nfds == -1) {
        perror("epoll_wait");
        exit(EXIT_FAILURE);
    }

    for (int i = 0; i < nfds; i++) {
        if (events[i].data.fd == server_sock) {
            client_sock = accept(server_sock,
                         (struct sockaddr *)&client_name,
                         (socklen_t *)(&client_name_len));

        if (client_sock == -1) //server overloaded
            continue;

        if (events[i].events & EPOLLIN) {
            std::cout << "EPOLLIN on " << client_sock << std::endl;
        }

        Arch::set_nonblocking(client_sock);
        ev.events = EPOLLIN | EPOLLRDHUP | EPOLLET; //input data and connection closing
        ev.data.fd = client_sock;

        if (epoll_ctl(epollfd, EPOLL_CTL_ADD, client_sock, &ev) == -1) {
            perror("epoll_ctl: client_socket");
            exit(EXIT_FAILURE);
        }

        accept_request(client_sock);

        } else {
            if (events[i].events & EPOLLRDHUP) {
                epoll_ctl(epollfd, EPOLL_CTL_DEL, events[i].data.fd, &ev);
            }
        }
    }
}

startup(&port) creates nonblocking socket, binding with port and so on. my script sends following data: GET /connect?id=1&secret=1 HTTP/1.0\r\n\r\n but sometimes recv return -1 in this function (calling inside of accept_request) :

/**********************************************************************/
/* Get a line from a socket, whether the line ends in a newline,
 * carriage return, or a CRLF combination.  Terminates the string read
 * with a null character.  If no newline indicator is found before the
 * end of the buffer, the string is terminated with a null.  If any of
 * the above three line terminators is read, the last character of the
 * string will be a linefeed and the string will be terminated with a
 * null character.
 * Parameters: the socket descriptor
 *             the buffer to save the data in
 *             the size of the buffer
 * Returns: the number of bytes stored (excluding null) */
/**********************************************************************/
int get_line(int sock, char *buf, int size) {
    int i = 0;
    char c = '\0';
    int n;

    while ((i < size - 1) && (c != '\n')) {
        n = recv(sock, &c, 1, 0);
        //debug
        std::cout << "n = " << n << std::endl;
        if (n > 0) {
            if (c == '\r') {
                n = recv(sock, &c, 1, MSG_PEEK);
                if ((n > 0) && (c == '\n'))
                    recv(sock, &c, 1, 0);
                else
                    c = '\n';
            }
            buf[i] = c;
            i++;
        } else {
            //debug
            if (errno == EWOULDBLOCK)
                std::cout << "EWOULDBLOCK" << std::endl;
            c = '\n';
        }
    }
    buf[i] = '\0';

    return(i);
}

as epoll man page wrote i have to read/write until i get EAGAIN, but i got it already! and i sure the buffer don't be empty. what i do wrong?

UPD: i've found out an interesting thing: when such situation is happened i use in my code sleep(1) and recc(...) again and i get data that i expect! it's a dirty trick. is there any more graceful approach to solve this problem?

Answer 1

It's completely normal for the first recv() in that case to return EAGAIN. epoll() never told you if it was readable yet or not.

Every single recv() should be prepared to handle EAGAIN if you are using non-blocking sockets. Spurious wakeups are possible, so whenever an API like select(), poll() or epoll() tells you that a socket is readable, it's only saying "it might be readable - give it a try".