socket connect timeout in c using poll

Question

I have an existing multi-threaded application which uses blocking connect() call.

However, I want to introduce connect timeout for application where if server does not respond to our query in x milliseconds, application will stop trying and give error.

However, I am not able to figure out how to do that using poll.

@caf 's non blocking connect using select has been of great help. But I have read that select in slow compared to poll, hence I want to use poll. Could you please tell me if that is true?

I am pasting his code from the post here

int main(int argc, char **argv) {
u_short port;                /* user specified port number */
char *addr;                  /* will be a pointer to the address */
struct sockaddr_in address;  /* the libc network address data structure */
short int sock = -1;         /* file descriptor for the network socket */
fd_set fdset;
struct timeval tv;

if (argc != 3) {
    fprintf(stderr, "Usage %s <port_num> <address>\n", argv[0]);
    return EXIT_FAILURE;
}

port = atoi(argv[1]);
addr = argv[2];

address.sin_family = AF_INET;
address.sin_addr.s_addr = inet_addr(addr); /* assign the address */
address.sin_port = htons(port);            /* translate int2port num */

sock = socket(AF_INET, SOCK_STREAM, 0);
fcntl(sock, F_SETFL, O_NONBLOCK);

connect(sock, (struct sockaddr *)&address, sizeof(address));

FD_ZERO(&fdset);
FD_SET(sock, &fdset);
tv.tv_sec = 10;             /* 10 second timeout */
tv.tv_usec = 0;

if (select(sock + 1, NULL, &fdset, NULL, &tv) == 1)
{
    int so_error;
    socklen_t len = sizeof so_error;

    getsockopt(sock, SOL_SOCKET, SO_ERROR, &so_error, &len);

    if (so_error == 0) {
        printf("%s:%d is open\n", addr, port);
    }
}

close(sock);
return 0;

}

Could you please help me to write similar functionality using poll.

I am on RHEL and using gcc 4.5.x version.

Update: For current code, how can change socket to blocking mode once app creates connections to server. I am not able to find a way to unset this O_NONBLOCK. Update 2: fcntl(sockfd, F_SETFL, fcntl(sockfd, F_GETFL) & ~O_NONBLOCK); One post has pointed that we can do this with above command. Didn't get the login though.

I think both poll and select are pretty slow. [Check here.](http://daniel.haxx.se/docs/poll-vs-select.html) — Darcy Rayner, Sep 22 '11 at 11:59
Actually, I need to build both solaris & linux complienet code. Hence I can not go for epoll(). Need to live with poll or select. — Amey Jah, Sep 22 '11 at 13:09
@AmeyJah select is slower then epoll, not poll and that also only if you're dealing with a large number of sockets at once - just use select. it is perfectly fine. — gby, Sep 22 '11 at 13:15
@gby Could you please tell what would be the result for large number of sockets at once. Will select go slower than poll? — Amey Jah, Sep 22 '11 at 15:14
@AmeyJah select will not be slower then poll. They both will be slower then epoll, for a large number of sockets — gby, Sep 22 '11 at 17:23
@AmeyJah you probably want to read this: http://www.kegel.com/c10k.html — gby, Sep 26 '11 at 14:39

score 5 · Accepted Answer · answered Sep 22 '11 at 15:18

The performance of select() and poll() relative to other system-dependent polling APIs is equivalent if there are only a few file descriptors that are being polled. They both scale poorly, but can perform just fine when the program is only listening to a few fds.

An advantage of other system-specific polling APIs tends to be that they push associated state into the kernel so that each polling syscall does not require a copy of state from userspace to kernelspace. With only a few events, this copy and scan overhead is insignificant, and this benefit does not exist.

Just stick with select() unless you need to scale to handling hundreds or more fds.

If you really will be handling many fds, and you must support both Linux and Solaris, consider using a library like libevent or libev to abstract away the kernel-specific efficient wait APIs.

score -1 · Answer 2 · answered Sep 22 '11 at 14:38

-1

select won't be slower than poll. But even if it were, it's unlikely that the minimal difference would matter in your application.

For doing a nonblocking timeout connect you can copy from existing applications that provide such functionality. (see for example here, for an idea on how to do it).

answered Sep 22 '11 at 14:38

rumpel

7,870
2
38
39

oh sorry, I saw to late that you already have code for nonblocking connect with select. – rumpel Sep 22 '11 at 14:40

socket connect timeout in c using poll

2 Answers2