1 /* Establishing and handling network connections.
2 Copyright (C) 1995, 1996, 1997, 2001, 2002 Free Software Foundation, Inc.
4 This file is part of GNU Wget.
6 GNU Wget is free software; you can redistribute it and/or modify
7 it under the terms of the GNU General Public License as published by
8 the Free Software Foundation; either version 2 of the License, or
9 (at your option) any later version.
11 GNU Wget is distributed in the hope that it will be useful,
12 but WITHOUT ANY WARRANTY; without even the implied warranty of
13 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 GNU General Public License for more details.
16 You should have received a copy of the GNU General Public License
17 along with Wget; if not, write to the Free Software
18 Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
20 In addition, as a special exception, the Free Software Foundation
21 gives permission to link the code of its release of Wget with the
22 OpenSSL project's "OpenSSL" library (or with modified versions of it
23 that use the same license as the "OpenSSL" library), and distribute
24 the linked executables. You must obey the GNU General Public License
25 in all respects for all of the code used other than "OpenSSL". If you
26 modify this file, you may extend this exception to your version of the
27 file, but you are not obligated to do so. If you do not wish to do
28 so, delete this exception statement from your version. */
40 # include <sys/socket.h>
42 # include <netinet/in.h>
44 # include <arpa/inet.h>
46 #endif /* not WINDOWS */
50 #ifdef HAVE_SYS_SELECT_H
51 # include <sys/select.h>
52 #endif /* HAVE_SYS_SELECT_H */
60 /* Define sockaddr_storage where unavailable (presumably on IPv4-only
64 # ifndef HAVE_STRUCT_SOCKADDR_STORAGE
65 # define sockaddr_storage sockaddr_in
67 #endif /* ENABLE_IPV6 */
69 /* Fill SA as per the data in IP and PORT. SA shoult point to struct
70 sockaddr_storage if ENABLE_IPV6 is defined, to struct sockaddr_in
74 sockaddr_set_data (struct sockaddr *sa, const ip_address *ip, int port)
80 struct sockaddr_in *sin = (struct sockaddr_in *)sa;
82 sin->sin_family = AF_INET;
83 sin->sin_port = htons (port);
84 sin->sin_addr = ADDRESS_IPV4_IN_ADDR (ip);
90 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)sa;
92 sin6->sin6_family = AF_INET6;
93 sin6->sin6_port = htons (port);
94 sin6->sin6_addr = ADDRESS_IPV6_IN6_ADDR (ip);
95 #ifdef HAVE_SOCKADDR_IN6_SCOPE_ID
96 sin6->sin6_scope_id = ADDRESS_IPV6_SCOPE (ip);
100 #endif /* ENABLE_IPV6 */
106 /* Get the data of SA, specifically the IP address and the port. If
107 you're not interested in one or the other information, pass NULL as
111 sockaddr_get_data (const struct sockaddr *sa, ip_address *ip, int *port)
113 switch (sa->sa_family)
117 struct sockaddr_in *sin = (struct sockaddr_in *)sa;
120 ip->type = IPV4_ADDRESS;
121 ADDRESS_IPV4_IN_ADDR (ip) = sin->sin_addr;
124 *port = ntohs (sin->sin_port);
130 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)sa;
133 ip->type = IPV6_ADDRESS;
134 ADDRESS_IPV6_IN6_ADDR (ip) = sin6->sin6_addr;
135 #ifdef HAVE_SOCKADDR_IN6_SCOPE_ID
136 ADDRESS_IPV6_SCOPE (ip) = sin6->sin6_scope_id;
140 *port = ntohs (sin6->sin6_port);
149 /* Return the size of the sockaddr structure depending on its
153 sockaddr_size (const struct sockaddr *sa)
155 switch (sa->sa_family)
158 return sizeof (struct sockaddr_in);
161 return sizeof (struct sockaddr_in6);
169 resolve_bind_address (struct sockaddr *sa)
171 struct address_list *al;
173 /* Make sure this is called only once. opt.bind_address doesn't
174 change during a Wget run. */
175 static int called, should_bind;
176 static ip_address ip;
180 sockaddr_set_data (sa, &ip, 0);
185 al = lookup_host (opt.bind_address, LH_BIND | LH_SILENT);
188 /* #### We should be able to print the error message here. */
189 logprintf (LOG_NOTQUIET,
190 _("%s: unable to resolve bind address `%s'; disabling bind.\n"),
191 exec_name, opt.bind_address);
196 /* Pick the first address in the list and use it as bind address.
197 Perhaps we should try multiple addresses in succession, but I
198 don't think that's necessary in practice. */
199 ip = *address_list_address_at (al, 0);
200 address_list_release (al);
202 sockaddr_set_data (sa, &ip, 0);
209 const struct sockaddr *addr;
215 connect_with_timeout_callback (void *arg)
217 struct cwt_context *ctx = (struct cwt_context *)arg;
218 ctx->result = connect (ctx->fd, ctx->addr, ctx->addrlen);
221 /* Like connect, but specifies a timeout. If connecting takes longer
222 than TIMEOUT seconds, -1 is returned and errno is set to
226 connect_with_timeout (int fd, const struct sockaddr *addr, socklen_t addrlen,
229 struct cwt_context ctx;
232 ctx.addrlen = addrlen;
234 if (run_with_timeout (timeout, connect_with_timeout_callback, &ctx))
239 if (ctx.result == -1 && errno == EINTR)
244 /* Connect via TCP to the specified address and port.
246 If PRINT is non-NULL, it is the host name to print that we're
250 connect_to_ip (const ip_address *ip, int port, const char *print)
252 struct sockaddr_storage ss;
253 struct sockaddr *sa = (struct sockaddr *)&ss;
256 /* If PRINT is non-NULL, print the "Connecting to..." line, with
257 PRINT being the host name we're connecting to. */
260 const char *txt_addr = pretty_print_address (ip);
261 if (print && 0 != strcmp (print, txt_addr))
262 logprintf (LOG_VERBOSE, _("Connecting to %s|%s|:%d... "),
263 escnonprint (print), txt_addr, port);
265 logprintf (LOG_VERBOSE, _("Connecting to %s:%d... "), txt_addr, port);
268 /* Store the sockaddr info to SA. */
269 sockaddr_set_data (sa, ip, port);
271 /* Create the socket of the family appropriate for the address. */
272 sock = socket (sa->sa_family, SOCK_STREAM, 0);
276 #if defined(ENABLE_IPV6) && defined(IPV6_V6ONLY)
279 /* In case of error, we will go on anyway... */
280 int err = setsockopt (sock, IPPROTO_IPV6, IPV6_V6ONLY, &on, sizeof (on));
283 DEBUGP (("Failed setting IPV6_V6ONLY: %s", strerror (errno)));
288 /* For very small rate limits, set the buffer size (and hence,
289 hopefully, the kernel's TCP window size) to the per-second limit.
290 That way we should never have to sleep for more than 1s between
292 if (opt.limit_rate && opt.limit_rate < 8192)
294 int bufsize = opt.limit_rate;
296 bufsize = 512; /* avoid pathologically small values */
298 setsockopt (sock, SOL_SOCKET, SO_RCVBUF,
299 (void *)&bufsize, (socklen_t)sizeof (bufsize));
301 /* When we add limit_rate support for writing, which is useful
302 for POST, we should also set SO_SNDBUF here. */
305 if (opt.bind_address)
307 /* Bind the client side of the socket to the requested
309 struct sockaddr_storage bind_ss;
310 struct sockaddr *bind_sa = (struct sockaddr *)&bind_ss;
311 if (resolve_bind_address (bind_sa))
313 if (bind (sock, bind_sa, sockaddr_size (bind_sa)) < 0)
318 /* Connect the socket to the remote endpoint. */
319 if (connect_with_timeout (sock, sa, sockaddr_size (sa),
320 opt.connect_timeout) < 0)
326 logprintf (LOG_VERBOSE, _("connected.\n"));
327 DEBUGP (("Created socket %d.\n", sock));
332 /* Protect errno from possible modifications by close and
334 int save_errno = errno;
338 logprintf (LOG_VERBOSE, _("failed: %s.\n"), strerror (errno));
344 /* Connect via TCP to a remote host on the specified port.
346 HOST is resolved as an Internet host name. If HOST resolves to
347 more than one IP address, they are tried in the order returned by
348 DNS until connecting to one of them succeeds. */
351 connect_to_host (const char *host, int port)
356 struct address_list *al = lookup_host (host, 0);
362 address_list_get_bounds (al, &start, &end);
363 for (i = start; i < end; i++)
365 const ip_address *ip = address_list_address_at (al, i);
366 sock = connect_to_ip (ip, port, host);
370 address_list_set_connected (al);
371 address_list_release (al);
375 /* The attempt to connect has failed. Continue with the loop
376 and try next address. */
378 address_list_set_faulty (al, i);
381 /* Failed to connect to any of the addresses in AL. */
383 if (address_list_connected_p (al))
385 /* We connected to AL before, but cannot do so now. That might
386 indicate that our DNS cache entry for HOST has expired. */
387 address_list_release (al);
388 al = lookup_host (host, LH_REFRESH);
391 address_list_release (al);
396 /* Create a socket, bind it to local interface BIND_ADDRESS on port
397 *PORT, set up a listen backlog, and return the resulting socket, or
400 BIND_ADDRESS is the address of the interface to bind to. If it is
401 NULL, the socket is bound to the default address. PORT should
402 point to the port number that will be used for the binding. If
403 that number is 0, the system will choose a suitable port, and the
404 chosen value will be written to *PORT.
406 Calling accept() on such a socket waits for and accepts incoming
410 bind_local (const ip_address *bind_address, int *port)
413 int family = AF_INET;
414 struct sockaddr_storage ss;
415 struct sockaddr *sa = (struct sockaddr *)&ss;
417 /* For setting options with setsockopt. */
419 void *setopt_ptr = (void *)&setopt_val;
420 socklen_t setopt_size = sizeof (setopt_val);
423 if (bind_address->type == IPV6_ADDRESS)
427 sock = socket (family, SOCK_STREAM, 0);
432 setsockopt (sock, SOL_SOCKET, SO_REUSEADDR, setopt_ptr, setopt_size);
436 sockaddr_set_data (sa, bind_address, *port);
437 if (bind (sock, sa, sockaddr_size (sa)) < 0)
442 DEBUGP (("Local socket fd %d bound.\n", sock));
444 /* If *PORT is 0, find out which port we've bound to. */
447 socklen_t addrlen = sockaddr_size (sa);
448 if (getsockname (sock, sa, &addrlen) < 0)
450 /* If we can't find out the socket's local address ("name"),
451 something is seriously wrong with the socket, and it's
452 unusable for us anyway because we must know the chosen
457 sockaddr_get_data (sa, NULL, port);
458 DEBUGP (("binding to address %s using port %i.\n",
459 pretty_print_address (bind_address), *port));
461 if (listen (sock, 1) < 0)
469 /* Like a call to accept(), but with the added check for timeout.
471 In other words, accept a client connection on LOCAL_SOCK, and
472 return the new socket used for communication with the client.
473 LOCAL_SOCK should have been bound, e.g. using bind_local().
475 The caller is blocked until a connection is established. If no
476 connection is established for opt.connect_timeout seconds, the
477 function exits with an error status. */
480 accept_connection (int local_sock)
484 /* We don't need the values provided by accept, but accept
485 apparently requires them to be present. */
486 struct sockaddr_storage ss;
487 struct sockaddr *sa = (struct sockaddr *)&ss;
488 socklen_t addrlen = sizeof (ss);
490 if (opt.connect_timeout)
492 int test = select_fd (local_sock, opt.connect_timeout, WAIT_FOR_READ);
498 sock = accept (local_sock, sa, &addrlen);
499 DEBUGP (("Accepted client at socket %d.\n", sock));
503 /* Get the IP address associated with the connection on FD and store
504 it to IP. Return 1 on success, 0 otherwise.
506 If ENDPOINT is ENDPOINT_LOCAL, it returns the address of the local
507 (client) side of the socket. Else if ENDPOINT is ENDPOINT_PEER, it
508 returns the address of the remote (peer's) side of the socket. */
511 socket_ip_address (int sock, ip_address *ip, int endpoint)
513 struct sockaddr_storage storage;
514 struct sockaddr *sockaddr = (struct sockaddr *)&storage;
515 socklen_t addrlen = sizeof (storage);
518 if (endpoint == ENDPOINT_LOCAL)
519 ret = getsockname (sock, sockaddr, &addrlen);
520 else if (endpoint == ENDPOINT_PEER)
521 ret = getpeername (sock, sockaddr, &addrlen);
527 switch (sockaddr->sa_family)
532 struct sockaddr_in6 *sa6 = (struct sockaddr_in6 *)&storage;
533 ip->type = IPV6_ADDRESS;
534 ADDRESS_IPV6_IN6_ADDR (ip) = sa6->sin6_addr;
535 #ifdef HAVE_SOCKADDR_IN6_SCOPE_ID
536 ADDRESS_IPV6_SCOPE (ip) = sa6->sin6_scope_id;
538 DEBUGP (("conaddr is: %s\n", pretty_print_address (ip)));
544 struct sockaddr_in *sa = (struct sockaddr_in *)&storage;
545 ip->type = IPV4_ADDRESS;
546 ADDRESS_IPV4_IN_ADDR (ip) = sa->sin_addr;
547 DEBUGP (("conaddr is: %s\n", pretty_print_address (ip)));
555 /* Return non-zero if the error from the connect code can be
556 considered retryable. Wget normally retries after errors, but the
557 exception are the "unsupported protocol" type errors (possible on
558 IPv4/IPv6 dual family systems) and "connection refused". */
561 retryable_socket_connect_error (int err)
563 /* Have to guard against some of these values not being defined.
564 Cannot use a switch statement because some of the values might be
568 || err == EAFNOSUPPORT
571 || err == EPFNOSUPPORT
573 #ifdef ESOCKTNOSUPPORT /* no, "sockt" is not a typo! */
574 || err == ESOCKTNOSUPPORT
576 #ifdef EPROTONOSUPPORT
577 || err == EPROTONOSUPPORT
580 || err == ENOPROTOOPT
582 /* Apparently, older versions of Linux and BSD used EINVAL
583 instead of EAFNOSUPPORT and such. */
588 if (!opt.retry_connrefused)
589 if (err == ECONNREFUSED
591 || err == ENETUNREACH /* network is unreachable */
594 || err == EHOSTUNREACH /* host is unreachable */
602 /* Wait for a single descriptor to become available, timing out after
603 MAXTIME seconds. Returns 1 if FD is available, 0 for timeout and
604 -1 for error. The argument WAIT_FOR can be a combination of
605 WAIT_FOR_READ and WAIT_FOR_WRITE.
607 This is a mere convenience wrapper around the select call, and
608 should be taken as such (for example, it doesn't implement Wget's
609 0-timeout-means-no-timeout semantics.) */
612 select_fd (int fd, double maxtime, int wait_for)
615 fd_set *rd = NULL, *wr = NULL;
616 struct timeval tmout;
621 if (wait_for & WAIT_FOR_READ)
623 if (wait_for & WAIT_FOR_WRITE)
626 tmout.tv_sec = (long) maxtime;
627 tmout.tv_usec = 1000000 * (maxtime - (long) maxtime);
630 result = select (fd + 1, rd, wr, NULL, &tmout);
631 while (result < 0 && errno == EINTR);
637 test_socket_open (int sock)
642 /* Check if we still have a valid (non-EOF) connection. From Andrew
643 * Maholski's code in the Unix Socket FAQ. */
645 FD_ZERO (&check_set);
646 FD_SET (sock, &check_set);
648 /* Wait one microsecond */
652 /* If we get a timeout, then that means still connected */
653 if (select (sock + 1, &check_set, NULL, NULL, &to) == 0)
655 /* Connection is valid (not EOF), so continue */
662 /* Basic socket operations, mostly EINTR wrappers. */
665 # define read(fd, buf, cnt) recv (fd, buf, cnt, 0)
666 # define write(fd, buf, cnt) send (fd, buf, cnt, 0)
667 # define close(fd) closesocket (fd)
671 # define read(fd, buf, cnt) recv (fd, buf, cnt, 0)
672 # define write(fd, buf, cnt) send (fd, buf, cnt, 0)
676 sock_read (int fd, char *buf, int bufsize)
680 res = read (fd, buf, bufsize);
681 while (res == -1 && errno == EINTR);
686 sock_write (int fd, char *buf, int bufsize)
690 res = write (fd, buf, bufsize);
691 while (res == -1 && errno == EINTR);
696 sock_poll (int fd, double timeout, int wait_for)
698 return select_fd (fd, timeout, wait_for);
702 sock_peek (int fd, char *buf, int bufsize)
706 res = recv (fd, buf, bufsize, MSG_PEEK);
707 while (res == -1 && errno == EINTR);
715 DEBUGP (("Closed fd %d\n", fd));
721 /* Reading and writing from the network. We build around the socket
722 (file descriptor) API, but support "extended" operations for things
723 that are not mere file descriptors under the hood, such as SSL
726 That way the user code can call fd_read(fd, ...) and we'll run read
727 or SSL_read or whatever is necessary. */
729 static struct hash_table *transport_map;
730 static int transport_map_modified_tick;
732 struct transport_info {
741 /* Register the transport layer operations that will be used when
742 reading, writing, and polling FD.
744 This should be used for transport layers like SSL that piggyback on
745 sockets. FD should otherwise be a real socket, on which you can
746 call getpeername, etc. */
749 fd_register_transport (int fd, fd_reader_t reader, fd_writer_t writer,
750 fd_poller_t poller, fd_peeker_t peeker,
751 fd_closer_t closer, void *ctx)
753 struct transport_info *info;
755 /* The file descriptor must be non-negative to be registered.
756 Negative values are ignored by fd_close(), and -1 cannot be used as
760 info = xnew (struct transport_info);
761 info->reader = reader;
762 info->writer = writer;
763 info->poller = poller;
764 info->peeker = peeker;
765 info->closer = closer;
768 transport_map = hash_table_new (0, NULL, NULL);
769 hash_table_put (transport_map, (void *) fd, info);
770 ++transport_map_modified_tick;
773 /* Return context of the transport registered with
774 fd_register_transport. This assumes fd_register_transport was
775 previously called on FD. */
778 fd_transport_context (int fd)
780 struct transport_info *info = hash_table_get (transport_map, (void *) fd);
784 /* When fd_read/fd_write are called multiple times in a loop, they should
785 remember the INFO pointer instead of fetching it every time. It is
786 not enough to compare FD to LAST_FD because FD might have been
787 closed and reopened. modified_tick ensures that changes to
788 transport_map will not be unnoticed.
790 This is a macro because we want the static storage variables to be
793 #define LAZY_RETRIEVE_INFO(info) do { \
794 static struct transport_info *last_info; \
795 static int last_fd = -1, last_tick; \
796 if (!transport_map) \
798 else if (last_fd == fd && last_tick == transport_map_modified_tick) \
802 info = hash_table_get (transport_map, (void *) fd); \
805 last_tick = transport_map_modified_tick; \
810 poll_internal (int fd, struct transport_info *info, int wf, double timeout)
813 timeout = opt.read_timeout;
817 if (info && info->poller)
818 test = info->poller (fd, timeout, wf, info->ctx);
820 test = sock_poll (fd, timeout, wf);
829 /* Read no more than BUFSIZE bytes of data from FD, storing them to
830 BUF. If TIMEOUT is non-zero, the operation aborts if no data is
831 received after that many seconds. If TIMEOUT is -1, the value of
832 opt.timeout is used for TIMEOUT. */
835 fd_read (int fd, char *buf, int bufsize, double timeout)
837 struct transport_info *info;
838 LAZY_RETRIEVE_INFO (info);
839 if (!poll_internal (fd, info, WAIT_FOR_READ, timeout))
841 if (info && info->reader)
842 return info->reader (fd, buf, bufsize, info->ctx);
844 return sock_read (fd, buf, bufsize);
847 /* Like fd_read, except it provides a "preview" of the data that will
848 be read by subsequent calls to fd_read. Specifically, it copies no
849 more than BUFSIZE bytes of the currently available data to BUF and
850 returns the number of bytes copied. Return values and timeout
851 semantics are the same as those of fd_read.
853 CAVEAT: Do not assume that the first subsequent call to fd_read
854 will retrieve the same amount of data. Reading can return more or
855 less data, depending on the TCP implementation and other
856 circumstances. However, barring an error, it can be expected that
857 all the peeked data will eventually be read by fd_read. */
860 fd_peek (int fd, char *buf, int bufsize, double timeout)
862 struct transport_info *info;
863 LAZY_RETRIEVE_INFO (info);
864 if (!poll_internal (fd, info, WAIT_FOR_READ, timeout))
866 if (info && info->peeker)
867 return info->peeker (fd, buf, bufsize, info->ctx);
869 return sock_peek (fd, buf, bufsize);
872 /* Write the entire contents of BUF to FD. If TIMEOUT is non-zero,
873 the operation aborts if no data is received after that many
874 seconds. If TIMEOUT is -1, the value of opt.timeout is used for
878 fd_write (int fd, char *buf, int bufsize, double timeout)
881 struct transport_info *info;
882 LAZY_RETRIEVE_INFO (info);
884 /* `write' may write less than LEN bytes, thus the loop keeps trying
885 it until all was written, or an error occurred. */
889 if (!poll_internal (fd, info, WAIT_FOR_WRITE, timeout))
891 if (info && info->writer)
892 res = info->writer (fd, buf, bufsize, info->ctx);
894 res = sock_write (fd, buf, bufsize);
903 /* Close the file descriptor FD. */
908 struct transport_info *info;
912 /* Don't use LAZY_RETRIEVE_INFO because fd_close() is only called once
913 per socket, so that particular optimization wouldn't work. */
916 info = hash_table_get (transport_map, (void *) fd);
918 if (info && info->closer)
919 info->closer (fd, info->ctx);
925 hash_table_remove (transport_map, (void *) fd);
927 ++transport_map_modified_tick;