Home Man
Today's Posts

Linux & Unix Commands - Search Man Pages
Man Page or Keyword Search:
Select Section of Man Page:
Select Man Page Repository:

Linux 2.6 - man page for select_tut (linux section 2)

SELECT_TUT(2)			    Linux Programmer's Manual			    SELECT_TUT(2)

       select, pselect, FD_CLR, FD_ISSET, FD_SET, FD_ZERO - synchronous I/O multiplexing

       /* According to POSIX.1-2001 */
       #include <sys/select.h>

       /* According to earlier standards */
       #include <sys/time.h>
       #include <sys/types.h>
       #include <unistd.h>

       int select(int nfds, fd_set *readfds, fd_set *writefds,
		  fd_set *exceptfds, struct timeval *utimeout);

       void FD_CLR(int fd, fd_set *set);
       int  FD_ISSET(int fd, fd_set *set);
       void FD_SET(int fd, fd_set *set);
       void FD_ZERO(fd_set *set);

       #include <sys/select.h>

       int pselect(int nfds, fd_set *readfds, fd_set *writefds,
		   fd_set *exceptfds, const struct timespec *ntimeout,
		   const sigset_t *sigmask);

   Feature Test Macro Requirements for glibc (see feature_test_macros(7)):

       pselect(): _POSIX_C_SOURCE >= 200112L || _XOPEN_SOURCE >= 600

       select()  (or  pselect()) is used to efficiently monitor multiple file descriptors, to see
       if any of them is, or becomes, "ready"; that is, to see whether I/O becomes  possible,  or
       an "exceptional condition" has occurred on any of the descriptors.

       Its  principal  arguments  are  three  "sets"  of file descriptors: readfds, writefds, and
       exceptfds.  Each set is declared as type fd_set, and its contents can be manipulated  with
       the  macros  FD_CLR(),  FD_ISSET(),  FD_SET(), and FD_ZERO().  A newly declared set should
       first be cleared using FD_ZERO().  select() modifies the contents of the sets according to
       the  rules  described  below;  after calling select() you can test if a file descriptor is
       still present in a set with the FD_ISSET() macro.  FD_ISSET() returns nonzero if a  speci-
       fied  file  descriptor is present in a set and zero if it is not.  FD_CLR() removes a file
       descriptor from a set.

	      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 that are immediately available for reading.

	      This set is watched to see if there is space to write  data  to  any  of	its  file
	      descriptors.   After  select()  has  returned, writefds will be cleared of all file
	      descriptors except for those that are immediately available for writing.

	      This set is watched for "exceptional  conditions".   In  practice,  only	one  such
	      exceptional  condition  is  common:  the availability of out-of-band (OOB) data for
	      reading from a TCP socket.  See recv(2), send(2), and tcp(7) for more details about
	      OOB  data.   (One  other	less common case where select(2) indicates an exceptional
	      condition occurs with pseudoterminals in packet  mode;  see  tty_ioctl(4).)   After
	      select() has returned, exceptfds will be cleared of all file descriptors except for
	      those for which an exceptional condition has occurred.

       nfds   This is an integer one more than the maximum of any file descriptor in any  of  the
	      sets.   In other words, while adding file descriptors to each of the sets, you must
	      calculate the maximum integer value of all of them, then increment  this	value  by
	      one, and then pass this as nfds.

	      This is the longest time select() may wait before returning, even if nothing inter-
	      esting happened.	If this value is passed as NULL,  then	select()  blocks  indefi-
	      nitely  waiting for a file descriptor to become ready.  utimeout can be set to zero
	      seconds, which causes select() to return immediately, with  information  about  the
	      readiness  of  file  descriptors	at  the  time  of the call.  The structure struct
	      timeval is defined as:

		  struct timeval {
		      time_t tv_sec;	/* seconds */
		      long tv_usec;	/* microseconds */

	      This argument for pselect() has the same meaning as utimeout, but  struct  timespec
	      has nanosecond precision as follows:

		  struct timespec {
		      long tv_sec;    /* seconds */
		      long tv_nsec;   /* nanoseconds */

	      This  argument  holds a set of signals that the kernel should unblock (i.e., remove
	      from the signal mask of the calling thread), while the caller is blocked inside the
	      pselect()  call  (see  sigaddset(3)  and sigprocmask(2)).  It may be NULL, in which
	      case the call does not modify the signal mask on entry and exit  to  the	function.
	      In this case, pselect() will then behave just like select().

   Combining signal and data events
       pselect()  is  useful if you are waiting for a signal as well as for file descriptor(s) to
       become ready for I/O.  Programs that receive signals 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 program.  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 pro-
       cessed 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 the conditional, but before the  select()  call?   The
       answer  is  that select() would block indefinitely, even though an event is actually pend-
       ing.  This race condition is solved by the pselect() call.  This call can be used  to  set
       the  signal  mask  to  a  set of signals that are only to be received within the pselect()
       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 sigprocmask(2).
       Our pselect() call would enable SIGCHLD by using an empty signal mask.  Our program  would
       look like:

       static volatile sig_atomic_t got_SIGCHLD = 0;

       static void
       child_sig_handler(int sig)
	   got_SIGCHLD = 1;

       main(int argc, char *argv[])
	   sigset_t sigmask, empty_mask;
	   struct sigaction sa;
	   fd_set readfds, writefds, exceptfds;
	   int r;

	   sigaddset(&sigmask, SIGCHLD);
	   if (sigprocmask(SIG_BLOCK, &sigmask, NULL) == -1) {

	   sa.sa_flags = 0;
	   sa.sa_handler = child_sig_handler;
	   if (sigaction(SIGCHLD, &sa, NULL) == -1) {


	   for (;;) {	       /* main loop */
	       /* Initialize readfds, writefds, and exceptfds
		  before the pselect() call. (Code omitted.) */

	       r = pselect(nfds, &readfds, &writefds, &exceptfds,
			   NULL, &empty_mask);
	       if (r == -1 && errno != EINTR) {
		   /* Handle error */

	       if (got_SIGCHLD) {
		   got_SIGCHLD = 0;

		   /* Handle signalled event here; e.g., wait() for all
		      terminated children. (Code omitted.) */

	       /* main body of program */

       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.  UNIX programmers often
       find themselves in a position where they have to  handle  I/O  from  more  than	one  file
       descriptor  where  the  data  flow  may	be  intermittent.  If you were to merely create a
       sequence of read(2) and write(2) calls, you would find that one of your	calls  may  block
       waiting for data from/to a file descriptor, while another file descriptor is unused though
       ready for I/O.  select() efficiently copes with this situation.

   Select law
       Many people who try to use select() come across behavior that is difficult  to  understand
       and  produces nonportable or borderline results.  For instance, the above program is care-
       fully written not to block at any point, even though it does not set its file  descriptors
       to nonblocking mode.  It is easy to introduce subtle errors that will remove the advantage
       of using select(), so here is a list of essentials to watch for when using select().

       1.  You should always try to 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 is difficult to debug.

       2.  The value nfds must be properly calculated for efficiency 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 should be checked to see if
	   they are ready.

       5.  The functions read(2), recv(2), write(2), and send(2) do  not  necessarily  read/write
	   the	full  amount  of  data	that  you have requested.  If they do read/write the full
	   amount, it's 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 managing
	   to send or receive only a single byte.

       6.  Never read/write only in single bytes at a time unless you are really  sure	that  you
	   have a small amount of data to process.  It is extremely inefficient not to read/write
	   as much data as you can buffer each time.  The buffers in the example below	are  1024
	   bytes although they could easily be made larger.

       7.  The functions read(2), recv(2), write(2), and send(2) as well as the select() call can
	   return -1 with errno set to EINTR, or with errno set to EAGAIN  (EWOULDBLOCK).   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 nonblocking I/O, you will not get EAGAIN.

       8.  Never call read(2), recv(2), write(2), or send(2) with a buffer length of zero.

       9.  If  the  functions read(2), recv(2), write(2), and send(2) fail with errors other than
	   those listed in 7., or one of the input functions returns 0, indicating end	of  file,
	   then  you  should not pass that descriptor to select() again.  In the example below, I
	   close the descriptor immediately, and then set it to -1 to prevent it  being  included
	   in a set.

       10. The timeout value must be initialized with each new call to select(), since some oper-
	   ating systems modify the structure.	pselect() however does	not  modify  its  timeout

       11. Since select() modifies its file descriptor sets, if the call is being used in a loop,
	   then the sets must be reinitialized before each call.

   Usleep emulation
       On systems that do not have a usleep(3) 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 guaranteed to work only on UNIX systems, however.

       On  success,  select()  returns	the total number of file descriptors still present in the
       file descriptor sets.

       If select() timed out, then the return value will  be  zero.   The  file  descriptors  set
       should be all empty (but may not be on some systems).

       A  return value of -1 indicates an error, with errno being set appropriately.  In the case
       of an error, the contents of the returned sets and the struct timeout contents  are  unde-
       fined and should not be used.  pselect() however never modifies ntimeout.

       Generally  speaking,  all  operating  systems  that support sockets also support select().
       select() can be used to solve many problems in a portable and  efficient  way  that  naive
       programmers  try  to solve in a more complicated manner using threads, forking, IPCs, sig-
       nals, memory sharing, and so on.

       The poll(2) system call has the same functionality as select(), and is somewhat more effi-
       cient  when  monitoring sparse file descriptor sets.  It is nowadays widely available, but
       historically was less portable than select().

       The Linux-specific epoll(7)  API  provides  an  interface  that	is  more  efficient  than
       select(2) and poll(2) when monitoring large numbers of file descriptors.

       Here  is  an  example  that better demonstrates the true utility of select().  The listing
       below is 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)) == -1) {
	       return -1;
	   yes = 1;
	   if (setsockopt(s, SOL_SOCKET, SO_REUSEADDR,
		   &yes, sizeof(yes)) == -1) {
	       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)) == -1) {
	       return -1;
	   printf("accepting connections on port %d\n", 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)) == -1) {
	       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");
	       return -1;

	   if (connect(s, (struct sockaddr *) &a, sizeof(a)) == -1) {
	       shutdown(s, SHUT_RDWR);
	       return -1;
	   return s;

       #define SHUT_FD1 do {				    \
			    if (fd1 >= 0) {		    \
				shutdown(fd1, SHUT_RDWR);   \
				close(fd1);		    \
				fd1 = -1;		    \
			    }				    \
			} while (0)

       #define SHUT_FD2 do {				    \
			    if (fd2 >= 0) {		    \
				shutdown(fd2, SHUT_RDWR);   \
				close(fd2);		    \
				fd2 = -1;		    \
			    }				    \
			} while (0)

       #define BUF_SIZE 1024

       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");

	   signal(SIGPIPE, SIG_IGN);

	   forward_port = atoi(argv[2]);

	   h = listen_socket(atoi(argv[1]));
	   if (h == -1)

	   for (;;) {
	       int r, nfds = 0;
	       fd_set rd, wr, er;

	       FD_SET(h, &rd);
	       nfds = max(nfds, h);
	       if (fd1 > 0 && buf1_avail < BUF_SIZE) {
		   FD_SET(fd1, &rd);
		   nfds = max(nfds, fd1);
	       if (fd2 > 0 && buf2_avail < BUF_SIZE) {
		   FD_SET(fd2, &rd);
		   nfds = max(nfds, fd2);
	       if (fd1 > 0 && buf2_avail - buf2_written > 0) {
		   FD_SET(fd1, &wr);
		   nfds = max(nfds, fd1);
	       if (fd2 > 0 && buf1_avail - buf1_written > 0) {
		   FD_SET(fd2, &wr);
		   nfds = max(nfds, fd2);
	       if (fd1 > 0) {
		   FD_SET(fd1, &er);
		   nfds = max(nfds, fd1);
	       if (fd2 > 0) {
		   FD_SET(fd2, &er);
		   nfds = max(nfds, fd2);

	       r = select(nfds + 1, &rd, &wr, &er, NULL);

	       if (r == -1 && errno == EINTR)

	       if (r == -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 == -1) {
		   } else {
		       buf1_avail = buf1_written = 0;
		       buf2_avail = buf2_written = 0;
		       fd1 = r;
		       fd2 = connect_socket(forward_port, argv[3]);
		       if (fd2 == -1)
			   printf("connect from %s\n",

	       /* NB: read oob data before normal reads */

	       if (fd1 > 0)
		   if (FD_ISSET(fd1, &er)) {
		       char c;

		       r = recv(fd1, &c, 1, MSG_OOB);
		       if (r < 1)
			   send(fd2, &c, 1, MSG_OOB);
	       if (fd2 > 0)
		   if (FD_ISSET(fd2, &er)) {
		       char c;

		       r = recv(fd2, &c, 1, MSG_OOB);
		       if (r < 1)
			   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)
			   buf1_avail += r;
	       if (fd2 > 0)
		   if (FD_ISSET(fd2, &rd)) {
		       r = read(fd2, buf2 + buf2_avail,
				 BUF_SIZE - buf2_avail);
		       if (r < 1)
			   buf2_avail += r;
	       if (fd1 > 0)
		   if (FD_ISSET(fd1, &wr)) {
		       r = write(fd1, buf2 + buf2_written,
				  buf2_avail - buf2_written);
		       if (r < 1)
			   buf2_written += r;
	       if (fd2 > 0)
		   if (FD_ISSET(fd2, &wr)) {
		       r = write(fd2, buf1 + buf1_written,
				  buf1_avail - buf1_written);
		       if (r < 1)
			   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)
	       if (fd2 < 0 && buf2_avail - buf2_written == 0)

       The above program properly forwards most kinds of TCP  connections  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(2)  call
       and  devote  a  thread  to  each stream.  This becomes more tricky than you might suspect.
       Another idea is to set nonblocking I/O using fcntl(2).  This also has its problems because
       you end up using 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  connec-
       tion.  At the moment, new connections cause the current connection to be dropped.

       accept(2),  connect(2),	ioctl(2), poll(2), read(2), recv(2), select(2), send(2), sigproc-
       mask(2), write(2), sigaddset(3), sigdelset(3),  sigemptyset(3),	sigfillset(3),	sigismem-
       ber(3), epoll(7)

       This  page  is  part of release 3.55 of the Linux man-pages project.  A description of the
       project,    and	  information	 about	  reporting    bugs,	can    be    found     at

Linux					    2012-08-03				    SELECT_TUT(2)

All times are GMT -4. The time now is 08:43 PM.

Unix & Linux Forums Content Copyrightę1993-2018. All Rights Reserved.
Show Password