Linux server socket - Bad file descriptor

Question

I have a problem with a server socket under Linux. For some reason unknown to me the server socket vanishes and I get a Bad file descriptor error in the select call that waits for an incomming connection. This problem always occurs when I close an unrelated socket connection in a different thread. This happens on an embedded Linux with 2.6.36 Kernel.

Does anyone know why this would happen? Is it normal that a server socket can simply vanish resulting in Bad file descriptor?

edit: The other socket code implements a VNC Server and runs in a completely different thread. The only thing special in that other code is the use of setjmp/longjmp but that should not be a problem.

The code that create the server socket is the following:

int server_socket = socket(PF_INET, SOCK_STREAM, IPPROTO_TCP);

struct sockaddr_in saddr;
memset(&saddr, 0, sizeof(saddr));
saddr.sin_family = AF_INET;
saddr.sin_addr.s_addr = htonl(INADDR_ANY);
saddr.sin_port = htons(1234);

const int optionval = 1;
setsockopt(server_socket, SOL_SOCKET, SO_REUSEADDR, &optionval, sizeof(optionval));

if (bind(server_socket, (struct sockaddr *) &saddr, sizeof(saddr)) < 0) {
    perror("bind");
    return 0;
}

if (listen(server_socket, 1) < 0) {
    perror("listen");
    return 0;
}

I wait for an incomming connection using the code below:

static int WaitForConnection(int server_socket, struct timeval *timeout)
{
    fd_set read_fds;

    FD_ZERO(&read_fds);
    int max_sd = server_socket;
    FD_SET(server_socket, &read_fds);

    // This select will result in 'EBADFD' in the error case.
    // Even though the server socket was not closed with 'close'.
    int res = select(max_sd + 1, &read_fds, NULL, NULL, timeout);
    if (res > 0) {
        struct sockaddr_in caddr;
        socklen_t clen = sizeof(caddr);
        return accept(server_socket, (struct sockaddr *) &caddr, &clen);
    }

    return -1;
}

edit: When the problem case happens i currently simply restart the server but I don't understand why the server socket id should suddenly become an invalid file descriptor:

int error = 0;
socklen_t len = sizeof (error);
int retval = getsockopt (server_socket, SOL_SOCKET, SO_ERROR, &error, &len );
if (retval < 0) {
    close(server_socket);
    goto server_start;
}

Answer 1

Sockets (file descriptors) usually suffer from the same management issues as raw pointers in C. Whenever you close a socket, do not forget to assign -1 to the variable that keeps the descriptor value:

close(socket);
socket = -1;

As you would do to C pointer

free(buffer);
buffer = NULL;

If you forget to do this yo can later close socket twice, as you would free() memory twice if it was a pointer.

The other issue might be related to the fact that people usually forget: file descriptors in UNIX environment start from 0. If somewhere in the code you have

struct FooData {
    int foo;
    int socket;
    ...
}

// Either
FooData my_data_1 = {0};
// Or
FooData my_data_2;
memset(&my_data_2, 0, sizeof(my_data_2));

In both cases my_data_1 and my_data_2 have a valid descriptor (socket) value. And later, some piece of code, responsible for freeing FooData structure may blindly close() this descriptor, that happens to be you server's listening socket (0).