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)));
287 /* For very small rate limits, set the buffer size (and hence,
288 hopefully, the kernel's TCP window size) to the per-second limit.
289 That way we should never have to sleep for more than 1s between
291 if (opt.limit_rate && opt.limit_rate < 8192)
293 int bufsize = opt.limit_rate;
295 bufsize = 512; /* avoid pathologically small values */
297 setsockopt (sock, SOL_SOCKET, SO_RCVBUF,
298 (void *)&bufsize, (socklen_t)sizeof (bufsize));
300 /* When we add limit_rate support for writing, which is useful
301 for POST, we should also set SO_SNDBUF here. */
304 if (opt.bind_address)
306 /* Bind the client side of the socket to the requested
308 struct sockaddr_storage bind_ss;
309 struct sockaddr *bind_sa = (struct sockaddr *)&bind_ss;
310 if (resolve_bind_address (bind_sa))
312 if (bind (sock, bind_sa, sockaddr_size (bind_sa)) < 0)
317 /* Connect the socket to the remote endpoint. */
318 if (connect_with_timeout (sock, sa, sockaddr_size (sa),
319 opt.connect_timeout) < 0)
325 logprintf (LOG_VERBOSE, _("connected.\n"));
326 DEBUGP (("Created socket %d.\n", sock));
331 /* Protect errno from possible modifications by close and
333 int save_errno = errno;
337 logprintf (LOG_VERBOSE, _("failed: %s.\n"), strerror (errno));
343 /* Connect via TCP to a remote host on the specified port.
345 HOST is resolved as an Internet host name. If HOST resolves to
346 more than one IP address, they are tried in the order returned by
347 DNS until connecting to one of them succeeds. */
350 connect_to_host (const char *host, int port)
355 struct address_list *al = lookup_host (host, 0);
361 address_list_get_bounds (al, &start, &end);
362 for (i = start; i < end; i++)
364 const ip_address *ip = address_list_address_at (al, i);
365 sock = connect_to_ip (ip, port, host);
369 address_list_set_connected (al);
370 address_list_release (al);
374 /* The attempt to connect has failed. Continue with the loop
375 and try next address. */
377 address_list_set_faulty (al, i);
380 /* Failed to connect to any of the addresses in AL. */
382 if (address_list_connected_p (al))
384 /* We connected to AL before, but cannot do so now. That might
385 indicate that our DNS cache entry for HOST has expired. */
386 address_list_release (al);
387 al = lookup_host (host, LH_REFRESH);
390 address_list_release (al);
395 /* Create a socket, bind it to local interface BIND_ADDRESS on port
396 *PORT, set up a listen backlog, and return the resulting socket, or
399 BIND_ADDRESS is the address of the interface to bind to. If it is
400 NULL, the socket is bound to the default address. PORT should
401 point to the port number that will be used for the binding. If
402 that number is 0, the system will choose a suitable port, and the
403 chosen value will be written to *PORT.
405 Calling accept() on such a socket waits for and accepts incoming
409 bind_local (const ip_address *bind_address, int *port)
412 int family = AF_INET;
413 struct sockaddr_storage ss;
414 struct sockaddr *sa = (struct sockaddr *)&ss;
416 /* For setting options with setsockopt. */
418 void *setopt_ptr = (void *)&setopt_val;
419 socklen_t setopt_size = sizeof (setopt_val);
422 if (bind_address->type == IPV6_ADDRESS)
426 sock = socket (family, SOCK_STREAM, 0);
431 setsockopt (sock, SOL_SOCKET, SO_REUSEADDR, setopt_ptr, setopt_size);
435 sockaddr_set_data (sa, bind_address, *port);
436 if (bind (sock, sa, sockaddr_size (sa)) < 0)
441 DEBUGP (("Local socket fd %d bound.\n", sock));
443 /* If *PORT is 0, find out which port we've bound to. */
446 socklen_t addrlen = sockaddr_size (sa);
447 if (getsockname (sock, sa, &addrlen) < 0)
449 /* If we can't find out the socket's local address ("name"),
450 something is seriously wrong with the socket, and it's
451 unusable for us anyway because we must know the chosen
456 sockaddr_get_data (sa, NULL, port);
457 DEBUGP (("binding to address %s using port %i.\n",
458 pretty_print_address (bind_address), *port));
460 if (listen (sock, 1) < 0)
468 /* Like a call to accept(), but with the added check for timeout.
470 In other words, accept a client connection on LOCAL_SOCK, and
471 return the new socket used for communication with the client.
472 LOCAL_SOCK should have been bound, e.g. using bind_local().
474 The caller is blocked until a connection is established. If no
475 connection is established for opt.connect_timeout seconds, the
476 function exits with an error status. */
479 accept_connection (int local_sock)
483 /* We don't need the values provided by accept, but accept
484 apparently requires them to be present. */
485 struct sockaddr_storage ss;
486 struct sockaddr *sa = (struct sockaddr *)&ss;
487 socklen_t addrlen = sizeof (ss);
489 if (opt.connect_timeout)
491 int test = select_fd (local_sock, opt.connect_timeout, WAIT_FOR_READ);
497 sock = accept (local_sock, sa, &addrlen);
498 DEBUGP (("Accepted client at socket %d.\n", sock));
502 /* Get the IP address associated with the connection on FD and store
503 it to IP. Return 1 on success, 0 otherwise.
505 If ENDPOINT is ENDPOINT_LOCAL, it returns the address of the local
506 (client) side of the socket. Else if ENDPOINT is ENDPOINT_PEER, it
507 returns the address of the remote (peer's) side of the socket. */
510 socket_ip_address (int sock, ip_address *ip, int endpoint)
512 struct sockaddr_storage storage;
513 struct sockaddr *sockaddr = (struct sockaddr *)&storage;
514 socklen_t addrlen = sizeof (storage);
517 if (endpoint == ENDPOINT_LOCAL)
518 ret = getsockname (sock, sockaddr, &addrlen);
519 else if (endpoint == ENDPOINT_PEER)
520 ret = getpeername (sock, sockaddr, &addrlen);
526 switch (sockaddr->sa_family)
531 struct sockaddr_in6 *sa6 = (struct sockaddr_in6 *)&storage;
532 ip->type = IPV6_ADDRESS;
533 ADDRESS_IPV6_IN6_ADDR (ip) = sa6->sin6_addr;
534 #ifdef HAVE_SOCKADDR_IN6_SCOPE_ID
535 ADDRESS_IPV6_SCOPE (ip) = sa6->sin6_scope_id;
537 DEBUGP (("conaddr is: %s\n", pretty_print_address (ip)));
543 struct sockaddr_in *sa = (struct sockaddr_in *)&storage;
544 ip->type = IPV4_ADDRESS;
545 ADDRESS_IPV4_IN_ADDR (ip) = sa->sin_addr;
546 DEBUGP (("conaddr is: %s\n", pretty_print_address (ip)));
554 /* Return non-zero if the error from the connect code can be
555 considered retryable. Wget normally retries after errors, but the
556 exception are the "unsupported protocol" type errors (possible on
557 IPv4/IPv6 dual family systems) and "connection refused". */
560 retryable_socket_connect_error (int err)
562 /* Have to guard against some of these values not being defined.
563 Cannot use a switch statement because some of the values might be
567 || err == EAFNOSUPPORT
570 || err == EPFNOSUPPORT
572 #ifdef ESOCKTNOSUPPORT /* no, "sockt" is not a typo! */
573 || err == ESOCKTNOSUPPORT
575 #ifdef EPROTONOSUPPORT
576 || err == EPROTONOSUPPORT
579 || err == ENOPROTOOPT
581 /* Apparently, older versions of Linux and BSD used EINVAL
582 instead of EAFNOSUPPORT and such. */
587 if (!opt.retry_connrefused)
588 if (err == ECONNREFUSED
590 || err == ENETUNREACH /* network is unreachable */
593 || err == EHOSTUNREACH /* host is unreachable */
601 /* Wait for a single descriptor to become available, timing out after
602 MAXTIME seconds. Returns 1 if FD is available, 0 for timeout and
603 -1 for error. The argument WAIT_FOR can be a combination of
604 WAIT_FOR_READ and WAIT_FOR_WRITE.
606 This is a mere convenience wrapper around the select call, and
607 should be taken as such (for example, it doesn't implement Wget's
608 0-timeout-means-no-timeout semantics.) */
611 select_fd (int fd, double maxtime, int wait_for)
614 fd_set *rd = NULL, *wr = NULL;
615 struct timeval tmout;
620 if (wait_for & WAIT_FOR_READ)
622 if (wait_for & WAIT_FOR_WRITE)
625 tmout.tv_sec = (long) maxtime;
626 tmout.tv_usec = 1000000 * (maxtime - (long) maxtime);
629 result = select (fd + 1, rd, wr, NULL, &tmout);
630 while (result < 0 && errno == EINTR);
636 test_socket_open (int sock)
641 /* Check if we still have a valid (non-EOF) connection. From Andrew
642 * Maholski's code in the Unix Socket FAQ. */
644 FD_ZERO (&check_set);
645 FD_SET (sock, &check_set);
647 /* Wait one microsecond */
651 /* If we get a timeout, then that means still connected */
652 if (select (sock + 1, &check_set, NULL, NULL, &to) == 0)
654 /* Connection is valid (not EOF), so continue */
661 /* Basic socket operations, mostly EINTR wrappers. */
664 # define read(fd, buf, cnt) recv (fd, buf, cnt, 0)
665 # define write(fd, buf, cnt) send (fd, buf, cnt, 0)
666 # define close(fd) closesocket (fd)
670 # define read(fd, buf, cnt) recv (fd, buf, cnt, 0)
671 # define write(fd, buf, cnt) send (fd, buf, cnt, 0)
675 sock_read (int fd, char *buf, int bufsize)
679 res = read (fd, buf, bufsize);
680 while (res == -1 && errno == EINTR);
685 sock_write (int fd, char *buf, int bufsize)
689 res = write (fd, buf, bufsize);
690 while (res == -1 && errno == EINTR);
695 sock_poll (int fd, double timeout, int wait_for)
697 return select_fd (fd, timeout, wait_for);
701 sock_peek (int fd, char *buf, int bufsize)
705 res = recv (fd, buf, bufsize, MSG_PEEK);
706 while (res == -1 && errno == EINTR);
714 DEBUGP (("Closed fd %d\n", fd));
720 /* Reading and writing from the network. We build around the socket
721 (file descriptor) API, but support "extended" operations for things
722 that are not mere file descriptors under the hood, such as SSL
725 That way the user code can call fd_read(fd, ...) and we'll run read
726 or SSL_read or whatever is necessary. */
728 static struct hash_table *transport_map;
729 static int transport_map_modified_tick;
731 struct transport_info {
740 /* Register the transport layer operations that will be used when
741 reading, writing, and polling FD.
743 This should be used for transport layers like SSL that piggyback on
744 sockets. FD should otherwise be a real socket, on which you can
745 call getpeername, etc. */
748 fd_register_transport (int fd, fd_reader_t reader, fd_writer_t writer,
749 fd_poller_t poller, fd_peeker_t peeker,
750 fd_closer_t closer, void *ctx)
752 struct transport_info *info;
754 /* The file descriptor must be non-negative to be registered.
755 Negative values are ignored by fd_close(), and -1 cannot be used as
759 info = xnew (struct transport_info);
760 info->reader = reader;
761 info->writer = writer;
762 info->poller = poller;
763 info->peeker = peeker;
764 info->closer = closer;
767 transport_map = hash_table_new (0, NULL, NULL);
768 hash_table_put (transport_map, (void *) fd, info);
769 ++transport_map_modified_tick;
772 /* Return context of the transport registered with
773 fd_register_transport. This assumes fd_register_transport was
774 previously called on FD. */
777 fd_transport_context (int fd)
779 struct transport_info *info = hash_table_get (transport_map, (void *) fd);
783 /* When fd_read/fd_write are called multiple times in a loop, they should
784 remember the INFO pointer instead of fetching it every time. It is
785 not enough to compare FD to LAST_FD because FD might have been
786 closed and reopened. modified_tick ensures that changes to
787 transport_map will not be unnoticed.
789 This is a macro because we want the static storage variables to be
792 #define LAZY_RETRIEVE_INFO(info) do { \
793 static struct transport_info *last_info; \
794 static int last_fd = -1, last_tick; \
795 if (!transport_map) \
797 else if (last_fd == fd && last_tick == transport_map_modified_tick) \
801 info = hash_table_get (transport_map, (void *) fd); \
804 last_tick = transport_map_modified_tick; \
809 poll_internal (int fd, struct transport_info *info, int wf, double timeout)
812 timeout = opt.read_timeout;
816 if (info && info->poller)
817 test = info->poller (fd, timeout, wf, info->ctx);
819 test = sock_poll (fd, timeout, wf);
828 /* Read no more than BUFSIZE bytes of data from FD, storing them to
829 BUF. If TIMEOUT is non-zero, the operation aborts if no data is
830 received after that many seconds. If TIMEOUT is -1, the value of
831 opt.timeout is used for TIMEOUT. */
834 fd_read (int fd, char *buf, int bufsize, double timeout)
836 struct transport_info *info;
837 LAZY_RETRIEVE_INFO (info);
838 if (!poll_internal (fd, info, WAIT_FOR_READ, timeout))
840 if (info && info->reader)
841 return info->reader (fd, buf, bufsize, info->ctx);
843 return sock_read (fd, buf, bufsize);
846 /* Like fd_read, except it provides a "preview" of the data that will
847 be read by subsequent calls to fd_read. Specifically, it copies no
848 more than BUFSIZE bytes of the currently available data to BUF and
849 returns the number of bytes copied. Return values and timeout
850 semantics are the same as those of fd_read.
852 CAVEAT: Do not assume that the first subsequent call to fd_read
853 will retrieve the same amount of data. Reading can return more or
854 less data, depending on the TCP implementation and other
855 circumstances. However, barring an error, it can be expected that
856 all the peeked data will eventually be read by fd_read. */
859 fd_peek (int fd, char *buf, int bufsize, double timeout)
861 struct transport_info *info;
862 LAZY_RETRIEVE_INFO (info);
863 if (!poll_internal (fd, info, WAIT_FOR_READ, timeout))
865 if (info && info->peeker)
866 return info->peeker (fd, buf, bufsize, info->ctx);
868 return sock_peek (fd, buf, bufsize);
871 /* Write the entire contents of BUF to FD. If TIMEOUT is non-zero,
872 the operation aborts if no data is received after that many
873 seconds. If TIMEOUT is -1, the value of opt.timeout is used for
877 fd_write (int fd, char *buf, int bufsize, double timeout)
880 struct transport_info *info;
881 LAZY_RETRIEVE_INFO (info);
883 /* `write' may write less than LEN bytes, thus the loop keeps trying
884 it until all was written, or an error occurred. */
888 if (!poll_internal (fd, info, WAIT_FOR_WRITE, timeout))
890 if (info && info->writer)
891 res = info->writer (fd, buf, bufsize, info->ctx);
893 res = sock_write (fd, buf, bufsize);
902 /* Close the file descriptor FD. */
907 struct transport_info *info;
911 /* Don't use LAZY_RETRIEVE_INFO because fd_close() is only called once
912 per socket, so that particular optimization wouldn't work. */
915 info = hash_table_get (transport_map, (void *) fd);
917 if (info && info->closer)
918 info->closer (fd, info->ctx);
924 hash_table_remove (transport_map, (void *) fd);
926 ++transport_map_modified_tick;