1SELECT_TUT(2) Linux Programmer's Manual SELECT_TUT(2)
2
6 select, pselect, FD_CLR, FD_ISSET, FD_SET, FD_ZERO - synchronous I/O
7 multiplexing
8
10 #include <sys/time.h>
11 #include <sys/types.h>
12 #include <unistd.h>
13 int select(int nfds, fd_set *readfds, fd_set *writefds, fd_set
15 *exceptfds, struct timeval *utimeout);
16 int pselect(int nfds, fd_set *readfds, fd_set *writefds, fd_set
18 *exceptfds, const struct timespec *ntimeout, sigset_t *sigmask);
19 FD_CLR(int fd, fd_set *set);
21 FD_ISSET(int fd, fd_set *set);
22 FD_SET(int fd, fd_set *set);
23 FD_ZERO(fd_set *set);
24
26 select() (or pselect()) is the pivot function of most C programs that
27 handle more than one simultaneous file descriptor (or socket handle) in
28 an efficient manner. Its principal arguments are three arrays of file
29 descriptors: readfds, writefds, and exceptfds. The way that select() is
30 usually used is to block while waiting for a "change of status" on one
31 or more of the file descriptors. A "change of status" is when more
32 characters become available from the file descriptor, or when space
33 becomes available within the kernel's internal buffers for more to be
34 written to the file descriptor, or when a file descriptor goes into
35 error (in the case of a socket or pipe this is when the other end of
36 the connection is closed).
37 In summary, select() just watches multiple file descriptors, and is the
39 standard Unix call to do so.
40 The arrays of file descriptors are called file descriptor sets. Each
42 set is declared as type fd_set, and its contents can be altered with
43 the macros FD_CLR(), FD_ISSET(), FD_SET(), and FD_ZERO(). FD_ZERO() is
44 usually the first function to be used on a newly declared set. There‐
45 after, the individual file descriptors that you are interested in can
46 be added one by one with FD_SET(). select() modifies the contents of
47 the sets according to the rules described below; after calling select()
48 you can test if your file descriptor is still present in the set with
49 the FD_ISSET() macro. FD_ISSET() returns non-zero if the descriptor is
50 present and zero if it is not. FD_CLR() removes a file descriptor from
51 the set.
52
54 readfds
55 This set is watched to see if data is available for reading from
56 any of its file descriptors. After select() has returned,
57 readfds will be cleared of all file descriptors except for those
58 file descriptors that are immediately available for reading with
59 a recv() (for sockets) or read() (for pipes, files, and sockets)
60 call.
61 writefds
63 This set is watched to see if there is space to write data to
64 any of its file descriptors. After select() has returned,
65 writefds will be cleared of all file descriptors except for
66 those file descriptors that are immediately available for writ‐
67 ing with a send() (for sockets) or write() (for pipes, files,
68 and sockets) call.
69 exceptfds
71 This set is watched for exceptions or errors on any of the file
72 descriptors. However, that is actually just a rumor. How you use
73 exceptfds is to watch for out-of-band (OOB) data. OOB data is
74 data sent on a socket using the MSG_OOB flag, and hence
75 exceptfds only really applies to sockets. See recv(2) and
76 send(2) about this. After select() has returned, exceptfds will
77 be cleared of all file descriptors except for those descriptors
78 that are available for reading OOB data. You can only ever read
79 one byte of OOB data though (which is done with recv()), and
80 writing OOB data (done with send()) can be done at any time and
81 will not block. Hence there is no need for a fourth set to check
82 if a socket is available for writing OOB data.
83 nfds This is an integer one more than the maximum of any file
85 descriptor in any of the sets. In other words, while you are
86 busy adding file descriptors to your sets, you must calculate
87 the maximum integer value of all of them, then increment this
88 value by one, and then pass this as nfds to select().
89 utimeout
91 This is the longest time select() must wait before returning,
92 even if nothing interesting happened. If this value is passed as
93 NULL, then select() blocks indefinitely waiting for an event.
94 utimeout can be set to zero seconds, which causes select() to
95 return immediately. The structure struct timeval is defined as,
96 struct timeval {
98 time_t tv_sec; /* seconds */
99 long tv_usec; /* microseconds */
100 };
101 ntimeout
103 This argument has the same meaning as utimeout but struct time‐
104 spec has nanosecond precision as follows,
105 struct timespec {
107 long tv_sec; /* seconds */
108 long tv_nsec; /* nanoseconds */
109 };
110 sigmask
112 This argument holds a set of signals to allow while performing a
113 pselect() call (see sigaddset(3) and sigprocmask(2)). It can be
114 passed as NULL, in which case it does not modify the set of
115 allowed signals on entry and exit to the function. It will then
116 behave just like select().
117
119 pselect() must be used if you are waiting for a signal as well as data
120 from a file descriptor. Programs that receive signals as events nor‐
121 mally use the signal handler only to raise a global flag. The global
122 flag will indicate that the event must be processed in the main loop of
123 the program. A signal will cause the select() (or pselect()) call to
124 return with errno set to EINTR. This behavior is essential so that sig‐
125 nals can be processed in the main loop of the program, otherwise
126 select() would block indefinitely. Now, somewhere in the main loop will
127 be a conditional to check the global flag. So we must ask: what if a
128 signal arrives after the conditional, but before the select() call? The
129 answer is that select() would block indefinitely, even though an event
130 is actually pending. This race condition is solved by the pselect()
131 call. This call can be used to mask out signals that are not to be
132 received except within the pselect() call. For instance, let us say
133 that the event in question was the exit of a child process. Before the
134 start of the main loop, we would block SIGCHLD using sigprocmask(). Our
135 pselect() call would enable SIGCHLD by using the virgin signal mask.
136 Our program would look like:
137 int child_events = 0;
139 void child_sig_handler (int x) {
141 child_events++;
142 signal (SIGCHLD, child_sig_handler);
143 }
144 int main (int argc, char **argv) {
146 sigset_t sigmask, orig_sigmask;
147 sigemptyset (&sigmask);
149 sigaddset (&sigmask, SIGCHLD);
150 sigprocmask (SIG_BLOCK, &sigmask,
151 &orig_sigmask);
152 signal (SIGCHLD, child_sig_handler);
154 for (;;) { /* main loop */
156 for (; child_events > 0; child_events--) {
157 /* do event work here */
158 }
159 r = pselect (nfds, &rd, &wr, &er, 0, &orig_sigmask);
160 /* main body of program */
162 }
163 }
164
166 So what is the point of select()? Can't I just read and write to my
167 descriptors whenever I want? The point of select() is that it watches
168 multiple descriptors at the same time and properly puts the process to
169 sleep if there is no activity. It does this while enabling you to han‐
170 dle multiple simultaneous pipes and sockets. Unix programmers often
171 find themselves in a position where they have to handle I/O from more
172 than one file descriptor where the data flow may be intermittent. If
173 you were to merely create a sequence of read() and write() calls, you
174 would find that one of your calls may block waiting for data from/to a
175 file descriptor, while another file descriptor is unused though avail‐
176 able for data. select() efficiently copes with this situation.
177 A simple example of the use of select() can be found in the select(2)
179 manual page.
180
182 Here is an example that better demonstrates the true utility of
183 select(). The listing below is a TCP forwarding program that forwards
184 from one TCP port to another.
185 #include <stdlib.h>
187 #include <stdio.h>
188 #include <unistd.h>
189 #include <sys/time.h>
190 #include <sys/types.h>
191 #include <string.h>
192 #include <signal.h>
193 #include <sys/socket.h>
194 #include <netinet/in.h>
195 #include <arpa/inet.h>
196 #include <errno.h>
197 static int forward_port;
199 #undef max
201 #define max(x,y) ((x) > (y) ? (x) : (y))
202 static int listen_socket (int listen_port) {
204 struct sockaddr_in a;
205 int s;
206 int yes;
207 if ((s = socket (AF_INET, SOCK_STREAM, 0)) < 0) {
208 perror ("socket");
209 return -1;
210 }
211 yes = 1;
212 if (setsockopt
213 (s, SOL_SOCKET, SO_REUSEADDR,
214 (char *) &yes, sizeof (yes)) < 0) {
215 perror ("setsockopt");
216 close (s);
217 return -1;
218 }
219 memset (&a, 0, sizeof (a));
220 a.sin_port = htons (listen_port);
221 a.sin_family = AF_INET;
222 if (bind
223 (s, (struct sockaddr *) &a, sizeof (a)) < 0) {
224 perror ("bind");
225 close (s);
226 return -1;
227 }
228 printf ("accepting connections on port %d\n",
229 (int) listen_port);
230 listen (s, 10);
231 return s;
232 }
233 static int connect_socket (int connect_port,
235 char *address) {
236 struct sockaddr_in a;
237 int s;
238 if ((s = socket (AF_INET, SOCK_STREAM, 0)) < 0) {
239 perror ("socket");
240 close (s);
241 return -1;
242 }
243 memset (&a, 0, sizeof (a));
245 a.sin_port = htons (connect_port);
246 a.sin_family = AF_INET;
247 if (!inet_aton
249 (address,
250 (struct in_addr *) &a.sin_addr.s_addr)) {
251 perror ("bad IP address format");
252 close (s);
253 return -1;
254 }
255 if (connect
257 (s, (struct sockaddr *) &a,
258 sizeof (a)) < 0) {
259 perror ("connect()");
260 shutdown (s, SHUT_RDWR);
261 close (s);
262 return -1;
263 }
264 return s;
265 }
266 #define SHUT_FD1 { \
268 if (fd1 >= 0) { \
269 shutdown (fd1, SHUT_RDWR); \
270 close (fd1); \
271 fd1 = -1; \
272 } \
273 }
274 #define SHUT_FD2 { \
276 if (fd2 >= 0) { \
277 shutdown (fd2, SHUT_RDWR); \
278 close (fd2); \
279 fd2 = -1; \
280 } \
281 }
282 #define BUF_SIZE 1024
284 int main (int argc, char **argv) {
286 int h;
287 int fd1 = -1, fd2 = -1;
288 char buf1[BUF_SIZE], buf2[BUF_SIZE];
289 int buf1_avail, buf1_written;
290 int buf2_avail, buf2_written;
291 if (argc != 4) {
293 fprintf (stderr,
294 "Usage\n\tfwd <listen-port> \
295 <forward-to-port> <forward-to-ip-address>\n");
296 exit (1);
297 }
298 signal (SIGPIPE, SIG_IGN);
300 forward_port = atoi (argv[2]);
302 h = listen_socket (atoi (argv[1]));
304 if (h < 0)
305 exit (1);
306 for (;;) {
308 int r, nfds = 0;
309 fd_set rd, wr, er;
310 FD_ZERO (&rd);
311 FD_ZERO (&wr);
312 FD_ZERO (&er);
313 FD_SET (h, &rd);
314 nfds = max (nfds, h);
315 if (fd1 > 0 && buf1_avail < BUF_SIZE) {
316 FD_SET (fd1, &rd);
317 nfds = max (nfds, fd1);
318 }
319 if (fd2 > 0 && buf2_avail < BUF_SIZE) {
320 FD_SET (fd2, &rd);
321 nfds = max (nfds, fd2);
322 }
323 if (fd1 > 0
324 && buf2_avail - buf2_written > 0) {
325 FD_SET (fd1, &wr);
326 nfds = max (nfds, fd1);
327 }
328 if (fd2 > 0
329 && buf1_avail - buf1_written > 0) {
330 FD_SET (fd2, &wr);
331 nfds = max (nfds, fd2);
332 }
333 if (fd1 > 0) {
334 FD_SET (fd1, &er);
335 nfds = max (nfds, fd1);
336 }
337 if (fd2 > 0) {
338 FD_SET (fd2, &er);
339 nfds = max (nfds, fd2);
340 }
341 r = select (nfds + 1, &rd, &wr, &er, NULL);
343 if (r == -1 && errno == EINTR)
345 continue;
346 if (r < 0) {
347 perror ("select()");
348 exit (1);
349 }
350 if (FD_ISSET (h, &rd)) {
351 unsigned int l;
352 struct sockaddr_in client_address;
353 memset (&client_address, 0, l =
354 sizeof (client_address));
355 r = accept (h, (struct sockaddr *)
356 &client_address, &l);
357 if (r < 0) {
358 perror ("accept()");
359 } else {
360 SHUT_FD1;
361 SHUT_FD2;
362 buf1_avail = buf1_written = 0;
363 buf2_avail = buf2_written = 0;
364 fd1 = r;
365 fd2 =
366 connect_socket (forward_port,
367 argv[3]);
368 if (fd2 < 0) {
369 SHUT_FD1;
370 } else
371 printf ("connect from %s\n",
372 inet_ntoa
373 (client_address.sin_addr));
374 }
375 }
376 /* NB: read oob data before normal reads */
377 if (fd1 > 0)
378 if (FD_ISSET (fd1, &er)) {
379 char c;
380 errno = 0;
381 r = recv (fd1, &c, 1, MSG_OOB);
382 if (r < 1) {
383 SHUT_FD1;
384 } else
385 send (fd2, &c, 1, MSG_OOB);
386 }
387 if (fd2 > 0)
388 if (FD_ISSET (fd2, &er)) {
389 char c;
390 errno = 0;
391 r = recv (fd2, &c, 1, MSG_OOB);
392 if (r < 1) {
393 SHUT_FD1;
394 } else
395 send (fd1, &c, 1, MSG_OOB);
396 }
397 if (fd1 > 0)
398 if (FD_ISSET (fd1, &rd)) {
399 r =
400 read (fd1, buf1 + buf1_avail,
401 BUF_SIZE - buf1_avail);
402 if (r < 1) {
403 SHUT_FD1;
404 } else
405 buf1_avail += r;
406 }
407 if (fd2 > 0)
408 if (FD_ISSET (fd2, &rd)) {
409 r =
410 read (fd2, buf2 + buf2_avail,
411 BUF_SIZE - buf2_avail);
412 if (r < 1) {
413 SHUT_FD2;
414 } else
415 buf2_avail += r;
416 }
417 if (fd1 > 0)
418 if (FD_ISSET (fd1, &wr)) {
419 r =
420 write (fd1,
421 buf2 + buf2_written,
422 buf2_avail -
423 buf2_written);
424 if (r < 1) {
425 SHUT_FD1;
426 } else
427 buf2_written += r;
428 }
429 if (fd2 > 0)
430 if (FD_ISSET (fd2, &wr)) {
431 r =
432 write (fd2,
433 buf1 + buf1_written,
434 buf1_avail -
435 buf1_written);
436 if (r < 1) {
437 SHUT_FD2;
438 } else
439 buf1_written += r;
440 }
441 /* check if write data has caught read data */
442 if (buf1_written == buf1_avail)
443 buf1_written = buf1_avail = 0;
444 if (buf2_written == buf2_avail)
445 buf2_written = buf2_avail = 0;
446 /* one side has closed the connection, keep
447 writing to the other side until empty */
448 if (fd1 < 0
449 && buf1_avail - buf1_written == 0) {
450 SHUT_FD2;
451 }
452 if (fd2 < 0
453 && buf2_avail - buf2_written == 0) {
454 SHUT_FD1;
455 }
456 }
457 return 0;
458 }
459 The above program properly forwards most kinds of TCP connections
461 including OOB signal data transmitted by telnet servers. It handles the
462 tricky problem of having data flow in both directions simultaneously.
463 You might think it more efficient to use a fork() call and devote a
464 thread to each stream. This becomes more tricky than you might suspect.
465 Another idea is to set non-blocking I/O using an ioctl() call. This
466 also has its problems because you end up having to have inefficient
467 timeouts.
468 The program does not handle more than one simultaneous connection at a
470 time, although it could easily be extended to do this with a linked
471 list of buffers — one for each connection. At the moment, new connec‐
472 tions cause the current connection to be dropped.
473
475 Many people who try to use select() come across behavior that is diffi‐
476 cult to understand and produces non-portable or borderline results. For
477 instance, the above program is carefully written not to block at any
478 point, even though it does not set its file descriptors to non-blocking
479 mode at all (see ioctl(2)). It is easy to introduce subtle errors that
480 will remove the advantage of using select(), hence I will present a
481 list of essentials to watch for when using the select() call.
482 1. You should always try to use select() without a timeout. Your
484 program should have nothing to do if there is no data available.
485 Code that depends on timeouts is not usually portable and is
486 difficult to debug.
487 2. The value nfds must be properly calculated for efficiency as
489 explained above.
490 3. No file descriptor must be added to any set if you do not intend
492 to check its result after the select() call, and respond appro‐
493 priately. See next rule.
494 4. After select() returns, all file descriptors in all sets should
496 be checked to see if they are ready.
497 5. The functions read(), recv(), write(), and send() do not neces‐
499 sarily read/write the full amount of data that you have
500 requested. If they do read/write the full amount, its because
501 you have a low traffic load and a fast stream. This is not
502 always going to be the case. You should cope with the case of
503 your functions only managing to send or receive a single byte.
504 6. Never read/write only in single bytes at a time unless your are
506 really sure that you have a small amount of data to process. It
507 is extremely inefficient not to read/write as much data as you
508 can buffer each time. The buffers in the example above are 1024
509 bytes although they could easily be made larger.
510 7. The functions read(), recv(), write(), and send() as well as the
512 select() call can return -1 with errno set to EINTR, or with
513 errno set to EAGAIN (EWOULDBLOCK). These results must be prop‐
514 erly managed (not done properly above). If your program is not
515 going to receive any signals then it is unlikely you will get
516 EINTR. If your program does not set non-blocking I/O, you will
517 not get EAGAIN. Nonetheless you should still cope with these
518 errors for completeness.
519 8. Never call read(), recv(), write(), or send() with a buffer
521 length of zero.
522 9. If the functions read(), recv(), write(), and send() fail with
524 errors other than those listed in 7., or one of the input func‐
525 tions returns 0, indicating end of file, then you should not
526 pass that descriptor to select() again. In the above example, I
527 close the descriptor immediately, and then set it to -1 to pre‐
528 vent it being included in a set.
529 10. The timeout value must be initialized with each new call to
531 select(), since some operating systems modify the structure.
532 pselect() however does not modify its timeout structure.
533 11. I have heard that the Windows socket layer does not cope with
535 OOB data properly. It also does not cope with select() calls
536 when no file descriptors are set at all. Having no file descrip‐
537 tors set is a useful way to sleep the process with sub-second
538 precision by using the timeout. (See further on.)
539
541 On systems that do not have a usleep() function, you can call select()
542 with a finite timeout and no file descriptors as follows:
543 struct timeval tv;
545 tv.tv_sec = 0;
546 tv.tv_usec = 200000; /* 0.2 seconds */
547 select (0, NULL, NULL, NULL, &tv);
548 This is only guaranteed to work on Unix systems, however.
550
552 On success, select() returns the total number of file descriptors still
553 present in the file descriptor sets.
554 If select() timed out, then the return value will be zero. The file
556 descriptors set should be all empty (but may not be on some systems).
557 A return value of -1 indicates an error, with errno being set appropri‐
559 ately. In the case of an error, the returned sets and the timeout
560 struct contents are undefined and should not be used. pselect() how‐
561 ever never modifies ntimeout.
562
564 Generally speaking, all operating systems that support sockets, also
565 support select(). Many types of programs become extremely complicated
566 without the use of select(). select() can be used to solve many prob‐
567 lems in a portable and efficient way that naive programmers try to
568 solve in a more complicated manner using threads, forking, IPCs, sig‐
569 nals, memory sharing, and so on.
570 The poll(2) system call has the same functionality as select(), and is
572 somewhat more efficient when monitoring sparse file descriptor sets.
573 It is nowadays widely available, but historically was less portable
574 than select().
575 The Linux-specific epoll(7) API provides an interface that is more
577 efficient than select(2) and poll(2) when monitoring large numbers of
578 file descriptors.
579
581 accept(2), connect(2), ioctl(2), poll(2), read(2), recv(2), select(2),
582 send(2), sigprocmask(2), write(2), sigaddset(3), sigdelset(3), sigemp‐
583 tyset(3), sigfillset(3), sigismember(3), epoll(7)
584Linux 2006-05-13 SELECT_TUT(2)