SELECT_TUT(T) Linux Programmer's Manual SELECT_TUT(T) NAME select, pselect, FD_CLR, FD_ISSET, FD_SET, FD_ZERO - syn- chronous I/O multiplexing SYNOPSIS #include <sys/time.h> #include <sys/types.h> #include <unistd.h> int select(int n, fd_set *readfds, fd_set *writefds, fd_set *exceptfds, struct timeval *utimeout); int pselect(int n, fd_set *readfds, fd_set *writefds, fd_set *exceptfds, const struct timespec *ntimeout, sigset_t *sigmask); FD_CLR(int fd, fd_set *set); FD_ISSET(int fd, fd_set *set); FD_SET(int fd, fd_set *set); FD_ZERO(fd_set *set); DESCRIPTION select (or pselect) is the pivot function of most C pro- grams that handle more than one simultaneous file descrip- tor (or socket handle) in an efficient manner. Its princi- pal arguments are three arrays of file descriptors: readfds, writefds, and exceptfds. The way that select is usually used is to block while waiting for a "change of status" on one or more of the file descriptors. A "change of status" is when more characters become available from the file descriptor; or when space becomes available within the kernel's internal buffers for more to be writ- ten to the file descriptor, or when a file descriptor goes into error (in the case of a socket or pipe this is when the other end of the connection is closed). In summary, select just watches multiple file descriptors, and is the standard Unix call to do so. The arrays of file descriptors are called file descriptor sets. Each set is declared as type fd_set, and its con- tents can be altered with the macros FD_CLR, FD_ISSET, FD_SET, and FD_ZERO. FD_ZERO is usually the first func- tion to be used on a newly declared set. Thereafter, the individual file descriptors that you are interested in can be added one by one with FD_SET. select modifies the con- tents of the sets according to the rules described below; after calling select you can test if your file descriptor is still present in the set with the FD_ISSET macro. FD_ISSET returns zero if the descriptor is present and non-zero if it is not. FD_CLR removes a file descriptor from the set although I can't see the use for it in a clean program. ARGUMENTS readfds This set is watched to see if data is available for reading from any of its file descriptors. After select has returned, readfds will be cleared of all file descriptors except for those file descriptors that are immediately available for reading with a recv() (for sockets) or read() (for pipes, files, and sockets) call. writefds This set is watched to see if there is space to write data to any of its file descriptor. After select has returned, writefds will be cleared of all file descriptors except for those file descrip- tors that are immediately available for writing with a send() (for sockets) or write() (for pipes, files, and sockets) call. exceptfds This set is watched for exceptions or errors on any of the file descriptors. However, that is actually just a rumor. How you use exceptfds is to watch for Out of Bounds (OOB) data. OOB data is data sent on a socket using the MSG_OOB flag, and hence exceptfds only really applies to sockets. See recv(v) and send(d) about this. After select has returned, exceptfds will be cleared of all file descriptors except for those descriptors that are available for reading OOB data. You can only ever read one byte of OOB data though (which is done with recv()), and writing OOB data (done with send) can be done at any time and will not block. Hence there is no need for a fourth set to check if a socket is available for writing OOB data. n This is an integer one more than the maximum of any file descriptor in any of the sets. In other words, while you are busy adding file descriptors to your sets, you must calculate the maximum integer value of all of them, then increment this value by one, and then pass this as n to select. utimeout This is the longest time select must wait before returning, even if nothing interesting happened. If this value is passed as NULL, then select blocks indefinitely waiting for an event. utimeout can be set to zero seconds, which causes select to return immediately. The structure struct timeval is defined as, struct timeval { long tv_sec; /* seconds */ long tv_usec; /* microseconds */ }; ntimeout This argument has the meaning as utimeout but struct timespec has nanosecond precision as fol- lows, struct timespec { long tv_sec; /* seconds */ long tv_nsec; /* nanoseconds */ }; sigmask This argument holds a set of signals to allow while performing a pselect call (see sigaddset(t) and sigprocmask(k)). It can be passed as NULL, in which case it does not modify the set of allowed signals on entry and exit to the function. It will then behave just like select. COMBINING SIGNAL AND DATA EVENTS pselect must be used if you are waiting for a signal as well as data from a file descriptor. Programs that receive signals as events normally use the signal handler only to raise a global flag. The global flag will indicate that the event must be processed in the main loop of the pro- gram. A signal will cause the select (or pselect) call to return with errno set to EINTR. This behavior is essential so that signals can be processed in the main loop of the program, otherwise select would block indefinitely. Now, somewhere in the main loop will be a conditional to check the global flag. So we must ask: what if a signal arrives after conditional, but before the select call? The answer is that select would block indefinitely, even though an event is actually pending. This race condition is solved by the pselect call. This call can be used to mask out signals that are not to be received except within the pse- lect call. For instance, let us say that the event in question was the exit of a child process. Before the start of the main loop, we would block SIGCHLD using sigproc- mask. Our pselect call would enable SIGCHLD by using the virgin signal mask. Our program would look like: int child_events = 0; void child_sig_handler (int x) { child_events++; signal (SIGCHLD, child_sig_handler); } int main (int argc, char **argv) { sigset_t sigmask, orig_sigmask; sigemptyset (&sigmask); sigaddset (&sigmask, SIGCHLD); sigprocmask (SIG_BLOCK, &sigmask, &orig_sigmask); signal (SIGCHLD, child_sig_handler); for (;;) { /* main loop */ for (; child_events > 0; child_events--) { /* do event work here */ } r = pselect (n, &rd, &wr, &er, 0, &orig_sigmask); /* main body of program */ } } Note that the above pselect call can be replaced with: sigprocmask (SIG_BLOCK, &orig_sigmask, 0); r = select (n, &rd, &wr, &er, 0); sigprocmask (SIG_BLOCK, &sigmask, 0); but then there is still the possibility that a signal could arrive after the first sigprocmask and before the select. If you do do this, it is prudent to at least put a finite timeout so that the process does not block. At pre- sent glibc probably works this way. The Linux kernel does not have a native pselect system call as yet so this is all probably much of a mute point. PRACTICAL So what is the point of select? Can't I just read and write to my descriptors whenever I want? The point of select is that it watches multiple descriptors at the same time and properly puts the process to sleep if there is no activity. It does this while enabling you to handle multi- ple simultaneous pipes and sockets. Unix programmers often find themselves in a position where they have to handle IO from more than one file descriptor where the data flow may be intermittent. If you were to merely create a sequence of read and write calls, you would find that one of your calls may block waiting for data from/to a file descrip- tor, while another file descriptor is unused though avail- able for data. select efficiently copes with this situa- tion. A classic example of select comes from the select man page: #include <stdio.h> #include <sys/time.h> #include <sys/types.h> #include <unistd.h> int main(n) { fd_set rfds; struct timeval tv; int retval; /* Watch stdin (fd 0) to see when it has input. */ FD_ZERO(&rfds); FD_SET(0, &rfds); /* Wait up to five seconds. */ tv.tv_sec = 5; tv.tv_usec = 0; retval = select(1, &rfds, NULL, NULL, &tv); /* Don't rely on the value of tv now! */ if (retval) printf("Data is available now.\n"); /* FD_ISSET(0, &rfds) will be true. */ else printf("No data within five seconds.\n"); exit(t); } PORT FORWARDING EXAMPLE Here is an example that better demonstrates the true util- ity of select. The listing below a TCP forwarding program that forwards from one TCP port to another. #include <stdlib.h> #include <stdio.h> #include <unistd.h> #include <sys/time.h> #include <sys/types.h> #include <string.h> #include <signal.h> #include <sys/socket.h> #include <netinet/in.h> #include <arpa/inet.h> #include <errno.h> static int forward_port; #undef max #define max(x,y) ((x) > (y) ? (x) : (y)) static int listen_socket (int listen_port) { struct sockaddr_in a; int s; int yes; if ((s = socket (AF_INET, SOCK_STREAM, 0)) < 0) { perror ("socket"); return -1; } yes = 1; if (setsockopt (s, SOL_SOCKET, SO_REUSEADDR, (char *) &yes, sizeof (yes)) < 0) { perror ("setsockopt"); close (s); return -1; } memset (&a, 0, sizeof (a)); a.sin_port = htons (listen_port); a.sin_family = AF_INET; if (bind (s, (struct sockaddr *) &a, sizeof (a)) < 0) { perror ("bind"); close (s); return -1; } printf ("accepting connections on port %d\n", (int) listen_port); listen (s, 10); return s; } static int connect_socket (int connect_port, char *address) { struct sockaddr_in a; int s; if ((s = socket (AF_INET, SOCK_STREAM, 0)) < 0) { perror ("socket"); close (s); return -1; } memset (&a, 0, sizeof (a)); a.sin_port = htons (connect_port); a.sin_family = AF_INET; if (!inet_aton (address, (struct in_addr *) &a.sin_addr.s_addr)) { perror ("bad IP address format"); close (s); return -1; } if (connect (s, (struct sockaddr *) &a, sizeof (a)) < 0) { perror ("connect()"); shutdown (s, SHUT_RDWR); close (s); return -1; } return s; } #define SHUT_FD1 { \ if (fd1 >= 0) { \ shutdown (fd1, SHUT_RDWR); \ close (fd1); \ fd1 = -1; \ } \ } #define SHUT_FD2 { \ if (fd2 >= 0) { \ shutdown (fd2, SHUT_RDWR); \ close (fd2); \ fd2 = -1; \ } \ } #define BUF_SIZE 1024 int main (int argc, char **argv) { int h; int fd1 = -1, fd2 = -1; char buf1[BUF_SIZE], buf2[BUF_SIZE]; int buf1_avail, buf1_written; int buf2_avail, buf2_written; if (argc != 4) { fprintf (stderr, "Usage\n\tfwd <listen-port> \ <forward-to-port> <forward-to-ip-address>\n"); exit (1); } signal (SIGPIPE, SIG_IGN); forward_port = atoi (argv[2]); h = listen_socket (atoi (argv[1])); if (h < 0) exit (1); for (;;) { int r, n = 0; fd_set rd, wr, er; FD_ZERO (&rd); FD_ZERO (&wr); FD_ZERO (&er); FD_SET (h, &rd); n = max (n, h); if (fd1 > 0 && buf1_avail < BUF_SIZE) { FD_SET (fd1, &rd); n = max (n, fd1); } if (fd2 > 0 && buf2_avail < BUF_SIZE) { FD_SET (fd2, &rd); n = max (n, fd2); } if (fd1 > 0 && buf2_avail - buf2_written > 0) { FD_SET (fd1, &wr); n = max (n, fd1); } if (fd2 > 0 && buf1_avail - buf1_written > 0) { FD_SET (fd2, &wr); n = max (n, fd2); } if (fd1 > 0) { FD_SET (fd1, &er); n = max (n, fd1); } if (fd2 > 0) { FD_SET (fd2, &er); n = max (n, fd2); } r = select (n + 1, &rd, &wr, &er, NULL); if (r == -1 && errno == EINTR) continue; if (r < 0) { perror ("select()"); exit (1); } if (FD_ISSET (h, &rd)) { unsigned int l; struct sockaddr_in client_address; memset (&client_address, 0, l = sizeof (client_address)); r = accept (h, (struct sockaddr *) &client_address, &l); if (r < 0) { perror ("accept()"); } else { SHUT_FD1; SHUT_FD2; buf1_avail = buf1_written = 0; buf2_avail = buf2_written = 0; fd1 = r; fd2 = connect_socket (forward_port, argv[3]); if (fd2 < 0) { SHUT_FD1; } else printf ("connect from %s\n", inet_ntoa (client_address.sin_addr)); } } /* NB: read oob data before normal reads */ if (fd1 > 0) if (FD_ISSET (fd1, &er)) { char c; errno = 0; r = recv (fd1, &c, 1, MSG_OOB); if (r < 1) { SHUT_FD1; } else send (fd2, &c, 1, MSG_OOB); } if (fd2 > 0) if (FD_ISSET (fd2, &er)) { char c; errno = 0; r = recv (fd2, &c, 1, MSG_OOB); if (r < 1) { SHUT_FD1; } else send (fd1, &c, 1, MSG_OOB); } if (fd1 > 0) if (FD_ISSET (fd1, &rd)) { r = read (fd1, buf1 + buf1_avail, BUF_SIZE - buf1_avail); if (r < 1) { SHUT_FD1; } else buf1_avail += r; } if (fd2 > 0) if (FD_ISSET (fd2, &rd)) { r = read (fd2, buf2 + buf2_avail, BUF_SIZE - buf2_avail); if (r < 1) { SHUT_FD2; } else buf2_avail += r; } if (fd1 > 0) if (FD_ISSET (fd1, &wr)) { r = write (fd1, buf2 + buf2_written, buf2_avail - buf2_written); if (r < 1) { SHUT_FD1; } else buf2_written += r; } if (fd2 > 0) if (FD_ISSET (fd2, &wr)) { r = write (fd2, buf1 + buf1_written, buf1_avail - buf1_written); if (r < 1) { SHUT_FD2; } else buf1_written += r; } /* check if write data has caught read data */ if (buf1_written == buf1_avail) buf1_written = buf1_avail = 0; if (buf2_written == buf2_avail) buf2_written = buf2_avail = 0; /* one side has closed the connection, keep writing to the other side until empty */ if (fd1 < 0 && buf1_avail - buf1_written == 0) { SHUT_FD2; } if (fd2 < 0 && buf2_avail - buf2_written == 0) { SHUT_FD1; } } return 0; } The above program properly forwards most kinds of TCP con- nections including OOB signal data transmitted by telnet servers. It handles the tricky problem of having data flow in both directions simultaneously. You might think it more efficient to use a fork() call and devote a thread to each stream. This becomes more tricky than you might suspect. Another idea is to set non-blocking IO using an ioctl() call. This also has its problems because you end up having to have inefficient timeouts. The program does not handle more than one simultaneous connection at a time, although it could easily be extended to do this with a linked list of buffers - one for each connection. At the moment, new connections cause the cur- rent connection to be dropped. SELECT LAW Many people who try to use select come across behavior that is difficult to understand and produces non-portable or borderline results. For instance, the above program is carefully written not to block at any point, even though it does not set its file descriptors to non-blocking mode at all (see ioctl(l)). It is easy to introduce subtle errors that will remove the advantage of using select, hence I will present a list of essentials to watch for when using the select call. 1. You should always try use select without a timeout. Your program should have nothing to do if there is no data available. Code that depends on timeouts is not usually portable and difficult to debug. 2. The value n must be properly calculated for effi- ciency as explained above. 3. No file descriptor must be added to any set if you do not intend to check its result after the select call, and respond appropriately. See next rule. 4. After select returns, all file descriptors in all sets must be checked. Any file descriptor that is available for writing must be written to, and any file descriptor available for reading must be read, etc. 5. The functions read(), recv(), write(), and send() do not necessarily read/write the full amount of data that you have requested. If they do read/write the full amount, its because you have a low traffic load and a fast stream. This is not always going to be the case. You should cope with the case of your functions only managing to send or receive a single byte. 6. Never read/write only in single bytes at a time unless your are really sure that you have a small amount of data to process. It is extremely ineffi- cient not to read/write as much data as you can buffer each time. The buffers in the example above are 1024 bytes although they could easily be made as large as the maximum possible packet size on your local network. 7. The functions read(), recv(), write(), and send() as well as the select() call can return -1 with an errno of EINTR or EAGAIN (EWOULDBLOCK) which are not errors. These results must be properly managed (not done properly above). If your program is not going to receive any signals then it is unlikely you will get EINTR. If your program does not set non-blocking IO, you will not get EAGAIN. Nonethe- less you should still cope with these errors for completeness. 8. Except as indicated in 7., the functions read(), recv(), write(), and send() never have a return value less than 1 except if an error has occurred. For instance, a read() on a pipe where the other end has died returns zero (so does and end-of-file error), but only returns zero once. Should any of these functions return 0 or -1, you should not use that descriptor again. In the above example, I close the descriptor immediately, and then set it to -1 to prevent it being included in a set. 9. The timeout value must be initialized with each new call to select, since some operating systems modify the structure. pselect however does not modify its timeout structure. 10. I have heard that the Windows socket layer does not cope with OOB data properly. It also does not cope with select calls when no file descriptors are set at all. Having no file descriptors set is a useful way to sleep the process with sub-second precision by using the timeout. (See further on.) USLEEP EMULATION On systems that do not have a usleep function, you can call select with a finite timeout and no file descriptors as follows: struct timeval tv; tv.tv_sec = 0; tv.tv_usec = 200000; /* 0.2 seconds */ select (0, NULL, NULL, NULL, &tv); This is only guarenteed to work on Unix systems, however. RETURN VALUE On success, select returns the total number of file descriptors still present in the file descriptor sets. If select timed out, then the file descriptors sets should be all empty (but may not be on some systems). However the return value will definitely be zero. A return value of -1 indicates an error, with errno being set appropriately. In the case of an error, the returned sets and the timeout struct contents are undefined and should not be used. pselect however never modifies ntime- out. ERRORS EBADF A set contained an invalid file descriptor. This error often occurs when you add a file descriptor to a set that you have already issued a close on, or when that file descriptor has experienced some kind of error. Hence you should cease adding to sets any file descriptor that returns an error on reading or writing. EINTR An interrupting signal was caught like SIGINT or SIGCHLD etc. In this case you should rebuild your file descriptor sets and retry. EINVAL Occurs if n is negative. ENOMEM Internal memory allocation failure. NOTES Generally speaking, all operating systems that support sockets, also support select. Some people consider select to be an esoteric and rarely used function. Indeed, many types of programs become extremely complicated without it. select can be used to solve many problems in a portable and efficient way that naive programmers try to solve with threads, forking, IPCs, signals, memory sharing and other dirty methods. pselect is a newer function that is less commonly used. CONFORMING TO 4.4BSD (the select function first appeared in 4.2BSD). Generally portable to/from non-BSD systems supporting clones of the BSD socket layer (including System V vari- ants). However, note that the System V variant typically sets the timeout variable before exit, but the BSD variant does not. The pselect function is defined in IEEE Std 1003.1g-2000 (POSIX.1g). It is found in glibc2.1 and later. Glibc2.0 has a function with this name, that however does not take a sigmask parameter. SEE ALSO accept(t), connect(t), ioctl(l), poll(l), read(d), recv(v), select(t), send(d), sigaddset(t), sigdelset(t), sigemptyset(t), sigfillset(t), sigismember(r), sigproc- mask(k), write(e) AUTHORS This man page was written by Paul Sheer. Linux 2.4 October 21, 2001 SELECT_TUT(T)