1 .\" This manpage is copyright (C) 2001 Paul Sheer.
3 .\" SPDX-License-Identifier: Linux-man-pages-copyleft
5 .\" very minor changes, aeb
7 .\" Modified 5 June 2002, Michael Kerrisk <mtk.manpages@gmail.com>
8 .\" 2006-05-13, mtk, removed much material that is redundant with select.2
9 .\" various other changes
10 .\" 2008-01-26, mtk, substantial changes and rewrites
12 .TH SELECT_TUT 2 2021-03-22 "Linux" "Linux Programmer's Manual"
14 select, pselect \- synchronous I/O multiplexing
17 .RI ( libc ", " \-lc )
26 system calls are used to efficiently monitor multiple file descriptors,
27 to see if any of them is, or becomes, "ready";
28 that is, to see whether I/O becomes possible,
29 or an "exceptional condition" has occurred on any of the file descriptors.
31 This page provides background and tutorial information
32 on the use of these system calls.
33 For details of the arguments and semantics of
40 .SS Combining signal and data events
42 is useful if you are waiting for a signal as well as
43 for file descriptor(s) to become ready for I/O.
44 Programs that receive signals
45 normally use the signal handler only to raise a global flag.
46 The global flag will indicate that the event must be processed
47 in the main loop of the program.
48 A signal will cause the
52 call to return with \fIerrno\fP set to \fBEINTR\fP.
53 This behavior is essential so that signals can be processed
54 in the main loop of the program, otherwise
56 would block indefinitely.
59 in the main loop will be a conditional to check the global flag.
61 what if a signal arrives after the conditional, but before the
66 would block indefinitely, even though an event is actually pending.
67 This race condition is solved by the
70 This call can be used to set the signal mask to a set of signals
71 that are to be received only within the
74 For instance, let us say that the event in question
75 was the exit of a child process.
76 Before the start of the main loop, we
77 would block \fBSIGCHLD\fP using
83 by using an empty signal mask.
84 Our program would look like:
87 static volatile sig_atomic_t got_SIGCHLD = 0;
90 child_sig_handler(int sig)
96 main(int argc, char *argv[])
98 sigset_t sigmask, empty_mask;
100 fd_set readfds, writefds, exceptfds;
103 sigemptyset(&sigmask);
104 sigaddset(&sigmask, SIGCHLD);
105 if (sigprocmask(SIG_BLOCK, &sigmask, NULL) == \-1) {
106 perror("sigprocmask");
111 sa.sa_handler = child_sig_handler;
112 sigemptyset(&sa.sa_mask);
113 if (sigaction(SIGCHLD, &sa, NULL) == \-1) {
118 sigemptyset(&empty_mask);
120 for (;;) { /* main loop */
121 /* Initialize readfds, writefds, and exceptfds
122 before the pselect() call. (Code omitted.) */
124 r = pselect(nfds, &readfds, &writefds, &exceptfds,
126 if (r == \-1 && errno != EINTR) {
133 /* Handle signalled event here; e.g., wait() for all
134 terminated children. (Code omitted.) */
137 /* main body of program */
142 So what is the point of
144 Can't I just read and write to my file descriptors whenever I want?
148 multiple descriptors at the same time and properly puts the process to
149 sleep if there is no activity.
150 UNIX programmers often find
151 themselves in a position where they have to handle I/O from more than one
152 file descriptor where the data flow may be intermittent.
153 If you were to merely create a sequence of
158 find that one of your calls may block waiting for data from/to a file
159 descriptor, while another file descriptor is unused though ready for I/O.
161 efficiently copes with this situation.
163 Many people who try to use
165 come across behavior that is
166 difficult to understand and produces nonportable or borderline results.
167 For instance, the above program is carefully written not to
168 block at any point, even though it does not set its file descriptors to
170 It is easy to introduce
171 subtle errors that will remove the advantage of using
173 so here is a list of essentials to watch for when using
177 You should always try to use
181 should have nothing to do if there is no data available.
183 depends on timeouts is not usually portable and is difficult to debug.
186 The value \fInfds\fP must be properly calculated for efficiency as
190 No file descriptor must be added to any set if you do not intend
191 to check its result after the
193 call, and respond appropriately.
199 returns, all file descriptors in all sets
200 should be checked to see if they are ready.
209 do \fInot\fP necessarily read/write the full amount of data
210 that you have requested.
211 If they do read/write the full amount, it's
212 because you have a low traffic load and a fast stream.
213 This is not always going to be the case.
214 You should cope with the case of your
215 functions managing to send or receive only a single byte.
218 Never read/write only in single bytes at a time unless you are really
219 sure that you have a small amount of data to process.
221 inefficient not to read/write as much data as you can buffer each time.
222 The buffers in the example below are 1024 bytes although they could
223 easily be made larger.
233 can fail with the error
243 set to \fBEAGAIN\fP (\fBEWOULDBLOCK\fP).
244 These results must be properly managed (not done properly above).
245 If your program is not going to receive any signals, then
246 it is unlikely you will get \fBEINTR\fP.
247 If your program does not set nonblocking I/O,
248 you will not get \fBEAGAIN\fP.
249 .\" Nonetheless, you should still cope with these errors for completeness.
258 with a buffer length of zero.
267 fail with errors other than those listed in \fB7.\fP,
268 or one of the input functions returns 0, indicating end of file,
269 then you should \fInot\fP pass that file descriptor to
272 In the example below,
273 I close the file descriptor immediately, and then set it to \-1
274 to prevent it being included in a set.
277 The timeout value must be initialized with each new call to
279 since some operating systems modify the structure.
281 however does not modify its timeout structure.
286 modifies its file descriptor sets,
287 if the call is being used in a loop,
288 then the sets must be reinitialized before each call.
289 .\" "I have heard" does not fill me with confidence, and doesn't
290 .\" belong in a man page, so I've commented this point out.
293 .\" I have heard that the Windows socket layer does not cope with OOB data
295 .\" It also does not cope with
297 .\" calls when no file descriptors are set at all.
298 .\" Having no file descriptors set is a useful
299 .\" way to sleep the process with subsecond precision by using the timeout.
300 .\" (See further on.)
306 all operating systems that support sockets also support
310 many problems in a portable and efficient way that naive programmers try
311 to solve in a more complicated manner using
312 threads, forking, IPCs, signals, memory sharing, and so on.
316 system call has the same functionality as
318 and is somewhat more efficient when monitoring sparse
319 file descriptor sets.
320 It is nowadays widely available, but historically was less portable than
325 API provides an interface that is more efficient than
329 when monitoring large numbers of file descriptors.
331 Here is an example that better demonstrates the true utility of
333 The listing below is a TCP forwarding program that forwards
334 from one TCP port to another.
336 .\" SRC BEGIN (select.c)
338 #include <arpa/inet.h>
340 #include <netinet/in.h>
345 #include <sys/select.h>
346 #include <sys/socket.h>
349 static int forward_port;
352 #define max(x, y) ((x) > (y) ? (x) : (y))
355 listen_socket(int listen_port)
357 struct sockaddr_in addr;
361 lfd = socket(AF_INET, SOCK_STREAM, 0);
368 if (setsockopt(lfd, SOL_SOCKET, SO_REUSEADDR,
369 &yes, sizeof(yes)) == \-1)
371 perror("setsockopt");
376 memset(&addr, 0, sizeof(addr));
377 addr.sin_port = htons(listen_port);
378 addr.sin_family = AF_INET;
379 if (bind(lfd, (struct sockaddr *) &addr, sizeof(addr)) == \-1) {
385 printf("accepting connections on port %d\en", listen_port);
391 connect_socket(int connect_port, char *address)
393 struct sockaddr_in addr;
396 cfd = socket(AF_INET, SOCK_STREAM, 0);
402 memset(&addr, 0, sizeof(addr));
403 addr.sin_port = htons(connect_port);
404 addr.sin_family = AF_INET;
406 if (!inet_aton(address, (struct in_addr *) &addr.sin_addr.s_addr)) {
407 fprintf(stderr, "inet_aton(): bad IP address format\en");
412 if (connect(cfd, (struct sockaddr *) &addr, sizeof(addr)) == \-1) {
414 shutdown(cfd, SHUT_RDWR);
421 #define SHUT_FD1 do { \e
423 shutdown(fd1, SHUT_RDWR); \e
429 #define SHUT_FD2 do { \e
431 shutdown(fd2, SHUT_RDWR); \e
437 #define BUF_SIZE 1024
440 main(int argc, char *argv[])
443 int fd1 = \-1, fd2 = \-1;
444 char buf1[BUF_SIZE], buf2[BUF_SIZE];
445 int buf1_avail = 0, buf1_written = 0;
446 int buf2_avail = 0, buf2_written = 0;
449 fprintf(stderr, "Usage\en\etfwd <listen\-port> "
450 "<forward\-to\-port> <forward\-to\-ip\-address>\en");
454 signal(SIGPIPE, SIG_IGN);
456 forward_port = atoi(argv[2]);
458 h = listen_socket(atoi(argv[1]));
465 fd_set readfds, writefds, exceptfds;
473 if (fd1 > 0 && buf1_avail < BUF_SIZE)
474 FD_SET(fd1, &readfds);
475 /* Note: nfds is updated below, when fd1 is added to
477 if (fd2 > 0 && buf2_avail < BUF_SIZE)
478 FD_SET(fd2, &readfds);
480 if (fd1 > 0 && buf2_avail \- buf2_written > 0)
481 FD_SET(fd1, &writefds);
482 if (fd2 > 0 && buf1_avail \- buf1_written > 0)
483 FD_SET(fd2, &writefds);
486 FD_SET(fd1, &exceptfds);
487 nfds = max(nfds, fd1);
490 FD_SET(fd2, &exceptfds);
491 nfds = max(nfds, fd2);
494 ready = select(nfds + 1, &readfds, &writefds, &exceptfds, NULL);
496 if (ready == \-1 && errno == EINTR)
504 if (FD_ISSET(h, &readfds)) {
506 struct sockaddr_in client_addr;
509 addrlen = sizeof(client_addr);
510 memset(&client_addr, 0, addrlen);
511 fd = accept(h, (struct sockaddr *) &client_addr, &addrlen);
517 buf1_avail = buf1_written = 0;
518 buf2_avail = buf2_written = 0;
520 fd2 = connect_socket(forward_port, argv[3]);
524 printf("connect from %s\en",
525 inet_ntoa(client_addr.sin_addr));
527 /* Skip any events on the old, closed file
534 /* NB: read OOB data before normal reads. */
536 if (fd1 > 0 && FD_ISSET(fd1, &exceptfds)) {
539 nbytes = recv(fd1, &c, 1, MSG_OOB);
543 send(fd2, &c, 1, MSG_OOB);
545 if (fd2 > 0 && FD_ISSET(fd2, &exceptfds)) {
548 nbytes = recv(fd2, &c, 1, MSG_OOB);
552 send(fd1, &c, 1, MSG_OOB);
554 if (fd1 > 0 && FD_ISSET(fd1, &readfds)) {
555 nbytes = read(fd1, buf1 + buf1_avail,
556 BUF_SIZE \- buf1_avail);
560 buf1_avail += nbytes;
562 if (fd2 > 0 && FD_ISSET(fd2, &readfds)) {
563 nbytes = read(fd2, buf2 + buf2_avail,
564 BUF_SIZE \- buf2_avail);
568 buf2_avail += nbytes;
570 if (fd1 > 0 && FD_ISSET(fd1, &writefds) && buf2_avail > 0) {
571 nbytes = write(fd1, buf2 + buf2_written,
572 buf2_avail \- buf2_written);
576 buf2_written += nbytes;
578 if (fd2 > 0 && FD_ISSET(fd2, &writefds) && buf1_avail > 0) {
579 nbytes = write(fd2, buf1 + buf1_written,
580 buf1_avail \- buf1_written);
584 buf1_written += nbytes;
587 /* Check if write data has caught read data. */
589 if (buf1_written == buf1_avail)
590 buf1_written = buf1_avail = 0;
591 if (buf2_written == buf2_avail)
592 buf2_written = buf2_avail = 0;
594 /* One side has closed the connection, keep
595 writing to the other side until empty. */
597 if (fd1 < 0 && buf1_avail \- buf1_written == 0)
599 if (fd2 < 0 && buf2_avail \- buf2_written == 0)
607 The above program properly forwards most kinds of TCP connections
608 including OOB signal data transmitted by \fBtelnet\fP servers.
609 It handles the tricky problem of having data flow in both directions
611 You might think it more efficient to use a
613 call and devote a thread to each stream.
614 This becomes more tricky than you might suspect.
615 Another idea is to set nonblocking I/O using
617 This also has its problems because you end up using
618 inefficient timeouts.
620 The program does not handle more than one simultaneous connection at a
621 time, although it could easily be extended to do this with a linked list
622 of buffers\(emone for each connection.
624 connections cause the current connection to be dropped.
637 .\" This man page was written by Paul Sheer.