1 /* Establishing and handling network connections.
2 Copyright (C) 1995, 1996, 1997, 1998, 1999, 2000, 2001, 2002, 2003,
3 2004, 2005, 2006, 2007, 2008, 2009, 2010, 2011 Free Software
6 This file is part of GNU Wget.
8 GNU Wget is free software; you can redistribute it and/or modify
9 it under the terms of the GNU General Public License as published by
10 the Free Software Foundation; either version 3 of the License, or
11 (at your option) any later version.
13 GNU Wget is distributed in the hope that it will be useful,
14 but WITHOUT ANY WARRANTY; without even the implied warranty of
15 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 GNU General Public License for more details.
18 You should have received a copy of the GNU General Public License
19 along with Wget. If not, see <http://www.gnu.org/licenses/>.
21 Additional permission under GNU GPL version 3 section 7
23 If you modify this program, or any covered work, by linking or
24 combining it with the OpenSSL project's OpenSSL library (or a
25 modified version of that library), containing parts covered by the
26 terms of the OpenSSL or SSLeay licenses, the Free Software Foundation
27 grants you additional permission to convey the resulting work.
28 Corresponding Source for a non-source form of such a combination
29 shall include the source code for the parts of OpenSSL used as well
30 as that of the covered work. */
39 #ifdef HAVE_SYS_SOCKET_H
40 # include <sys/socket.h>
41 #endif /* def HAVE_SYS_SOCKET_H */
43 #ifdef HAVE_SYS_SELECT_H
44 # include <sys/select.h>
45 #endif /* def HAVE_SYS_SELECT_H */
50 # else /* def __VMS */
52 # endif /* def __VMS [else] */
53 # include <netinet/in.h>
55 # include <arpa/inet.h>
57 #endif /* not WINDOWS */
61 #ifdef HAVE_SYS_TIME_H
62 # include <sys/time.h>
69 /* Apparently needed for Interix: */
74 /* Define sockaddr_storage where unavailable (presumably on IPv4-only
78 # ifndef HAVE_STRUCT_SOCKADDR_STORAGE
79 # define sockaddr_storage sockaddr_in
81 #endif /* ENABLE_IPV6 */
83 /* Fill SA as per the data in IP and PORT. SA shoult point to struct
84 sockaddr_storage if ENABLE_IPV6 is defined, to struct sockaddr_in
88 sockaddr_set_data (struct sockaddr *sa, const ip_address *ip, int port)
94 struct sockaddr_in *sin = (struct sockaddr_in *)sa;
96 sin->sin_family = AF_INET;
97 sin->sin_port = htons (port);
98 sin->sin_addr = ip->data.d4;
104 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)sa;
106 sin6->sin6_family = AF_INET6;
107 sin6->sin6_port = htons (port);
108 sin6->sin6_addr = ip->data.d6;
109 #ifdef HAVE_SOCKADDR_IN6_SCOPE_ID
110 sin6->sin6_scope_id = ip->ipv6_scope;
114 #endif /* ENABLE_IPV6 */
120 /* Get the data of SA, specifically the IP address and the port. If
121 you're not interested in one or the other information, pass NULL as
125 sockaddr_get_data (const struct sockaddr *sa, ip_address *ip, int *port)
127 switch (sa->sa_family)
131 struct sockaddr_in *sin = (struct sockaddr_in *)sa;
134 ip->family = AF_INET;
135 ip->data.d4 = sin->sin_addr;
138 *port = ntohs (sin->sin_port);
144 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)sa;
147 ip->family = AF_INET6;
148 ip->data.d6 = sin6->sin6_addr;
149 #ifdef HAVE_SOCKADDR_IN6_SCOPE_ID
150 ip->ipv6_scope = sin6->sin6_scope_id;
154 *port = ntohs (sin6->sin6_port);
163 /* Return the size of the sockaddr structure depending on its
167 sockaddr_size (const struct sockaddr *sa)
169 switch (sa->sa_family)
172 return sizeof (struct sockaddr_in);
175 return sizeof (struct sockaddr_in6);
182 /* Resolve the bind address specified via --bind-address and store it
183 to SA. The resolved value is stored in a static variable and
184 reused after the first invocation of this function.
186 Returns true on success, false on failure. */
189 resolve_bind_address (struct sockaddr *sa)
191 struct address_list *al;
193 /* Make sure this is called only once. opt.bind_address doesn't
194 change during a Wget run. */
195 static bool called, should_bind;
196 static ip_address ip;
200 sockaddr_set_data (sa, &ip, 0);
205 al = lookup_host (opt.bind_address, LH_BIND | LH_SILENT);
208 /* #### We should be able to print the error message here. */
209 logprintf (LOG_NOTQUIET,
210 _("%s: unable to resolve bind address %s; disabling bind.\n"),
211 exec_name, quote (opt.bind_address));
216 /* Pick the first address in the list and use it as bind address.
217 Perhaps we should try multiple addresses in succession, but I
218 don't think that's necessary in practice. */
219 ip = *address_list_address_at (al, 0);
220 address_list_release (al);
222 sockaddr_set_data (sa, &ip, 0);
229 const struct sockaddr *addr;
235 connect_with_timeout_callback (void *arg)
237 struct cwt_context *ctx = (struct cwt_context *)arg;
238 ctx->result = connect (ctx->fd, ctx->addr, ctx->addrlen);
241 /* Like connect, but specifies a timeout. If connecting takes longer
242 than TIMEOUT seconds, -1 is returned and errno is set to
246 connect_with_timeout (int fd, const struct sockaddr *addr, socklen_t addrlen,
249 struct cwt_context ctx;
252 ctx.addrlen = addrlen;
254 if (run_with_timeout (timeout, connect_with_timeout_callback, &ctx))
259 if (ctx.result == -1 && errno == EINTR)
264 /* Connect via TCP to the specified address and port.
266 If PRINT is non-NULL, it is the host name to print that we're
270 connect_to_ip (const ip_address *ip, int port, const char *print)
272 struct sockaddr_storage ss;
273 struct sockaddr *sa = (struct sockaddr *)&ss;
276 /* If PRINT is non-NULL, print the "Connecting to..." line, with
277 PRINT being the host name we're connecting to. */
280 const char *txt_addr = print_address (ip);
281 if (0 != strcmp (print, txt_addr))
283 char *str = NULL, *name;
285 if (opt.enable_iri && (name = idn_decode ((char *) print)) != NULL)
287 int len = strlen (print) + strlen (name) + 4;
289 snprintf (str, len, "%s (%s)", name, print);
294 logprintf (LOG_VERBOSE, _("Connecting to %s|%s|:%d... "),
295 str ? str : escnonprint_uri (print), txt_addr, port);
301 logprintf (LOG_VERBOSE, _("Connecting to %s:%d... "), txt_addr, port);
304 /* Store the sockaddr info to SA. */
305 sockaddr_set_data (sa, ip, port);
307 /* Create the socket of the family appropriate for the address. */
308 sock = socket (sa->sa_family, SOCK_STREAM, 0);
312 #if defined(ENABLE_IPV6) && defined(IPV6_V6ONLY)
315 /* In case of error, we will go on anyway... */
316 int err = setsockopt (sock, IPPROTO_IPV6, IPV6_V6ONLY, &on, sizeof (on));
319 DEBUGP (("Failed setting IPV6_V6ONLY: %s", strerror (errno)));
323 /* For very small rate limits, set the buffer size (and hence,
324 hopefully, the kernel's TCP window size) to the per-second limit.
325 That way we should never have to sleep for more than 1s between
327 if (opt.limit_rate && opt.limit_rate < 8192)
329 int bufsize = opt.limit_rate;
331 bufsize = 512; /* avoid pathologically small values */
333 setsockopt (sock, SOL_SOCKET, SO_RCVBUF,
334 (void *)&bufsize, (socklen_t)sizeof (bufsize));
336 /* When we add limit_rate support for writing, which is useful
337 for POST, we should also set SO_SNDBUF here. */
340 if (opt.bind_address)
342 /* Bind the client side of the socket to the requested
344 struct sockaddr_storage bind_ss;
345 struct sockaddr *bind_sa = (struct sockaddr *)&bind_ss;
346 if (resolve_bind_address (bind_sa))
348 if (bind (sock, bind_sa, sockaddr_size (bind_sa)) < 0)
353 /* Connect the socket to the remote endpoint. */
354 if (connect_with_timeout (sock, sa, sockaddr_size (sa),
355 opt.connect_timeout) < 0)
361 logprintf (LOG_VERBOSE, _("connected.\n"));
362 DEBUGP (("Created socket %d.\n", sock));
367 /* Protect errno from possible modifications by close and
369 int save_errno = errno;
373 logprintf (LOG_VERBOSE, _("failed: %s.\n"), strerror (errno));
379 /* Connect via TCP to a remote host on the specified port.
381 HOST is resolved as an Internet host name. If HOST resolves to
382 more than one IP address, they are tried in the order returned by
383 DNS until connecting to one of them succeeds. */
386 connect_to_host (const char *host, int port)
391 struct address_list *al = lookup_host (host, 0);
396 logprintf (LOG_NOTQUIET,
397 _("%s: unable to resolve host address %s\n"),
398 exec_name, quote (host));
402 address_list_get_bounds (al, &start, &end);
403 for (i = start; i < end; i++)
405 const ip_address *ip = address_list_address_at (al, i);
406 sock = connect_to_ip (ip, port, host);
410 address_list_set_connected (al);
411 address_list_release (al);
415 /* The attempt to connect has failed. Continue with the loop
416 and try next address. */
418 address_list_set_faulty (al, i);
421 /* Failed to connect to any of the addresses in AL. */
423 if (address_list_connected_p (al))
425 /* We connected to AL before, but cannot do so now. That might
426 indicate that our DNS cache entry for HOST has expired. */
427 address_list_release (al);
428 al = lookup_host (host, LH_REFRESH);
431 address_list_release (al);
436 /* Create a socket, bind it to local interface BIND_ADDRESS on port
437 *PORT, set up a listen backlog, and return the resulting socket, or
440 BIND_ADDRESS is the address of the interface to bind to. If it is
441 NULL, the socket is bound to the default address. PORT should
442 point to the port number that will be used for the binding. If
443 that number is 0, the system will choose a suitable port, and the
444 chosen value will be written to *PORT.
446 Calling accept() on such a socket waits for and accepts incoming
450 bind_local (const ip_address *bind_address, int *port)
453 struct sockaddr_storage ss;
454 struct sockaddr *sa = (struct sockaddr *)&ss;
456 /* For setting options with setsockopt. */
458 void *setopt_ptr = (void *)&setopt_val;
459 socklen_t setopt_size = sizeof (setopt_val);
461 sock = socket (bind_address->family, SOCK_STREAM, 0);
466 setsockopt (sock, SOL_SOCKET, SO_REUSEADDR, setopt_ptr, setopt_size);
470 sockaddr_set_data (sa, bind_address, *port);
471 if (bind (sock, sa, sockaddr_size (sa)) < 0)
476 DEBUGP (("Local socket fd %d bound.\n", sock));
478 /* If *PORT is 0, find out which port we've bound to. */
481 socklen_t addrlen = sockaddr_size (sa);
482 if (getsockname (sock, sa, &addrlen) < 0)
484 /* If we can't find out the socket's local address ("name"),
485 something is seriously wrong with the socket, and it's
486 unusable for us anyway because we must know the chosen
491 sockaddr_get_data (sa, NULL, port);
492 DEBUGP (("binding to address %s using port %i.\n",
493 print_address (bind_address), *port));
495 if (listen (sock, 1) < 0)
503 /* Like a call to accept(), but with the added check for timeout.
505 In other words, accept a client connection on LOCAL_SOCK, and
506 return the new socket used for communication with the client.
507 LOCAL_SOCK should have been bound, e.g. using bind_local().
509 The caller is blocked until a connection is established. If no
510 connection is established for opt.connect_timeout seconds, the
511 function exits with an error status. */
514 accept_connection (int local_sock)
518 /* We don't need the values provided by accept, but accept
519 apparently requires them to be present. */
520 struct sockaddr_storage ss;
521 struct sockaddr *sa = (struct sockaddr *)&ss;
522 socklen_t addrlen = sizeof (ss);
524 if (opt.connect_timeout)
526 int test = select_fd (local_sock, opt.connect_timeout, WAIT_FOR_READ);
532 sock = accept (local_sock, sa, &addrlen);
533 DEBUGP (("Accepted client at socket %d.\n", sock));
537 /* Get the IP address associated with the connection on FD and store
538 it to IP. Return true on success, false otherwise.
540 If ENDPOINT is ENDPOINT_LOCAL, it returns the address of the local
541 (client) side of the socket. Else if ENDPOINT is ENDPOINT_PEER, it
542 returns the address of the remote (peer's) side of the socket. */
545 socket_ip_address (int sock, ip_address *ip, int endpoint)
547 struct sockaddr_storage storage;
548 struct sockaddr *sockaddr = (struct sockaddr *) &storage;
549 socklen_t addrlen = sizeof (storage);
552 memset (sockaddr, 0, addrlen);
553 if (endpoint == ENDPOINT_LOCAL)
554 ret = getsockname (sock, sockaddr, &addrlen);
555 else if (endpoint == ENDPOINT_PEER)
556 ret = getpeername (sock, sockaddr, &addrlen);
562 ip->family = sockaddr->sa_family;
563 switch (sockaddr->sa_family)
568 struct sockaddr_in6 *sa6 = (struct sockaddr_in6 *)&storage;
569 ip->data.d6 = sa6->sin6_addr;
570 #ifdef HAVE_SOCKADDR_IN6_SCOPE_ID
571 ip->ipv6_scope = sa6->sin6_scope_id;
573 DEBUGP (("conaddr is: %s\n", print_address (ip)));
579 struct sockaddr_in *sa = (struct sockaddr_in *)&storage;
580 ip->data.d4 = sa->sin_addr;
581 DEBUGP (("conaddr is: %s\n", print_address (ip)));
589 /* Return true if the error from the connect code can be considered
590 retryable. Wget normally retries after errors, but the exception
591 are the "unsupported protocol" type errors (possible on IPv4/IPv6
592 dual family systems) and "connection refused". */
595 retryable_socket_connect_error (int err)
597 /* Have to guard against some of these values not being defined.
598 Cannot use a switch statement because some of the values might be
602 || err == EAFNOSUPPORT
605 || err == EPFNOSUPPORT
607 #ifdef ESOCKTNOSUPPORT /* no, "sockt" is not a typo! */
608 || err == ESOCKTNOSUPPORT
610 #ifdef EPROTONOSUPPORT
611 || err == EPROTONOSUPPORT
614 || err == ENOPROTOOPT
616 /* Apparently, older versions of Linux and BSD used EINVAL
617 instead of EAFNOSUPPORT and such. */
622 if (!opt.retry_connrefused)
623 if (err == ECONNREFUSED
625 || err == ENETUNREACH /* network is unreachable */
628 || err == EHOSTUNREACH /* host is unreachable */
636 /* Wait for a single descriptor to become available, timing out after
637 MAXTIME seconds. Returns 1 if FD is available, 0 for timeout and
638 -1 for error. The argument WAIT_FOR can be a combination of
639 WAIT_FOR_READ and WAIT_FOR_WRITE.
641 This is a mere convenience wrapper around the select call, and
642 should be taken as such (for example, it doesn't implement Wget's
643 0-timeout-means-no-timeout semantics.) */
646 select_fd (int fd, double maxtime, int wait_for)
649 fd_set *rd = NULL, *wr = NULL;
650 struct timeval tmout;
655 if (wait_for & WAIT_FOR_READ)
657 if (wait_for & WAIT_FOR_WRITE)
660 tmout.tv_sec = (long) maxtime;
661 tmout.tv_usec = 1000000 * (maxtime - (long) maxtime);
665 result = select (fd + 1, rd, wr, NULL, &tmout);
667 /* gnulib select() converts blocking sockets to nonblocking in windows.
668 wget uses blocking sockets so we must convert them back to blocking. */
669 set_windows_fd_as_blocking_socket (fd);
672 while (result < 0 && errno == EINTR);
677 /* Return true iff the connection to the remote site established
678 through SOCK is still open.
680 Specifically, this function returns true if SOCK is not ready for
681 reading. This is because, when the connection closes, the socket
682 is ready for reading because EOF is about to be delivered. A side
683 effect of this method is that sockets that have pending data are
684 considered non-open. This is actually a good thing for callers of
685 this function, where such pending data can only be unwanted
686 leftover from a previous request. */
689 test_socket_open (int sock)
695 /* Check if we still have a valid (non-EOF) connection. From Andrew
696 * Maholski's code in the Unix Socket FAQ. */
698 FD_ZERO (&check_set);
699 FD_SET (sock, &check_set);
701 /* Wait one microsecond */
705 ret = select (sock + 1, &check_set, NULL, NULL, &to);
707 /* gnulib select() converts blocking sockets to nonblocking in windows.
708 wget uses blocking sockets so we must convert them back to blocking
710 set_windows_fd_as_blocking_socket ( sock );
714 /* We got a timeout, it means we're still connected. */
717 /* Read now would not wait, it means we have either pending data
722 /* Basic socket operations, mostly EINTR wrappers. */
725 sock_read (int fd, char *buf, int bufsize)
729 res = read (fd, buf, bufsize);
730 while (res == -1 && errno == EINTR);
735 sock_write (int fd, char *buf, int bufsize)
739 res = write (fd, buf, bufsize);
740 while (res == -1 && errno == EINTR);
745 sock_poll (int fd, double timeout, int wait_for)
747 return select_fd (fd, timeout, wait_for);
751 sock_peek (int fd, char *buf, int bufsize)
755 res = recv (fd, buf, bufsize, MSG_PEEK);
756 while (res == -1 && errno == EINTR);
764 DEBUGP (("Closed fd %d\n", fd));
770 /* Reading and writing from the network. We build around the socket
771 (file descriptor) API, but support "extended" operations for things
772 that are not mere file descriptors under the hood, such as SSL
775 That way the user code can call fd_read(fd, ...) and we'll run read
776 or SSL_read or whatever is necessary. */
778 static struct hash_table *transport_map;
779 static unsigned int transport_map_modified_tick;
781 struct transport_info {
782 struct transport_implementation *imp;
786 /* Register the transport layer operations that will be used when
787 reading, writing, and polling FD.
789 This should be used for transport layers like SSL that piggyback on
790 sockets. FD should otherwise be a real socket, on which you can
791 call getpeername, etc. */
794 fd_register_transport (int fd, struct transport_implementation *imp, void *ctx)
796 struct transport_info *info;
798 /* The file descriptor must be non-negative to be registered.
799 Negative values are ignored by fd_close(), and -1 cannot be used as
803 info = xnew (struct transport_info);
807 transport_map = hash_table_new (0, NULL, NULL);
808 hash_table_put (transport_map, (void *)(intptr_t) fd, info);
809 ++transport_map_modified_tick;
812 /* Return context of the transport registered with
813 fd_register_transport. This assumes fd_register_transport was
814 previously called on FD. */
817 fd_transport_context (int fd)
819 struct transport_info *info = hash_table_get (transport_map, (void *)(intptr_t) fd);
823 /* When fd_read/fd_write are called multiple times in a loop, they should
824 remember the INFO pointer instead of fetching it every time. It is
825 not enough to compare FD to LAST_FD because FD might have been
826 closed and reopened. modified_tick ensures that changes to
827 transport_map will not be unnoticed.
829 This is a macro because we want the static storage variables to be
832 #define LAZY_RETRIEVE_INFO(info) do { \
833 static struct transport_info *last_info; \
834 static int last_fd = -1; \
835 static unsigned int last_tick; \
836 if (!transport_map) \
838 else if (last_fd == fd && last_tick == transport_map_modified_tick) \
842 info = hash_table_get (transport_map, (void *)(intptr_t) fd); \
845 last_tick = transport_map_modified_tick; \
850 poll_internal (int fd, struct transport_info *info, int wf, double timeout)
853 timeout = opt.read_timeout;
857 if (info && info->imp->poller)
858 test = info->imp->poller (fd, timeout, wf, info->ctx);
860 test = sock_poll (fd, timeout, wf);
869 /* Read no more than BUFSIZE bytes of data from FD, storing them to
870 BUF. If TIMEOUT is non-zero, the operation aborts if no data is
871 received after that many seconds. If TIMEOUT is -1, the value of
872 opt.timeout is used for TIMEOUT. */
875 fd_read (int fd, char *buf, int bufsize, double timeout)
877 struct transport_info *info;
878 LAZY_RETRIEVE_INFO (info);
879 if (!poll_internal (fd, info, WAIT_FOR_READ, timeout))
881 if (info && info->imp->reader)
882 return info->imp->reader (fd, buf, bufsize, info->ctx);
884 return sock_read (fd, buf, bufsize);
887 /* Like fd_read, except it provides a "preview" of the data that will
888 be read by subsequent calls to fd_read. Specifically, it copies no
889 more than BUFSIZE bytes of the currently available data to BUF and
890 returns the number of bytes copied. Return values and timeout
891 semantics are the same as those of fd_read.
893 CAVEAT: Do not assume that the first subsequent call to fd_read
894 will retrieve the same amount of data. Reading can return more or
895 less data, depending on the TCP implementation and other
896 circumstances. However, barring an error, it can be expected that
897 all the peeked data will eventually be read by fd_read. */
900 fd_peek (int fd, char *buf, int bufsize, double timeout)
902 struct transport_info *info;
903 LAZY_RETRIEVE_INFO (info);
904 if (!poll_internal (fd, info, WAIT_FOR_READ, timeout))
906 if (info && info->imp->peeker)
907 return info->imp->peeker (fd, buf, bufsize, info->ctx);
909 return sock_peek (fd, buf, bufsize);
912 /* Write the entire contents of BUF to FD. If TIMEOUT is non-zero,
913 the operation aborts if no data is received after that many
914 seconds. If TIMEOUT is -1, the value of opt.timeout is used for
918 fd_write (int fd, char *buf, int bufsize, double timeout)
921 struct transport_info *info;
922 LAZY_RETRIEVE_INFO (info);
924 /* `write' may write less than LEN bytes, thus the loop keeps trying
925 it until all was written, or an error occurred. */
929 if (!poll_internal (fd, info, WAIT_FOR_WRITE, timeout))
931 if (info && info->imp->writer)
932 res = info->imp->writer (fd, buf, bufsize, info->ctx);
934 res = sock_write (fd, buf, bufsize);
943 /* Report the most recent error(s) on FD. This should only be called
944 after fd_* functions, such as fd_read and fd_write, and only if
945 they return a negative result. For errors coming from other calls
946 such as setsockopt or fopen, strerror should continue to be
949 If the transport doesn't support error messages or doesn't supply
950 one, strerror(errno) is returned. The returned error message
951 should not be used after fd_close has been called. */
956 /* Don't bother with LAZY_RETRIEVE_INFO, as this will only be called
957 in case of error, never in a tight loop. */
958 struct transport_info *info = NULL;
960 info = hash_table_get (transport_map, (void *)(intptr_t) fd);
962 if (info && info->imp->errstr)
964 const char *err = info->imp->errstr (fd, info->ctx);
967 /* else, fall through and print the system error. */
969 return strerror (errno);
972 /* Close the file descriptor FD. */
977 struct transport_info *info;
981 /* Don't use LAZY_RETRIEVE_INFO because fd_close() is only called once
982 per socket, so that particular optimization wouldn't work. */
985 info = hash_table_get (transport_map, (void *)(intptr_t) fd);
987 if (info && info->imp->closer)
988 info->imp->closer (fd, info->ctx);
994 hash_table_remove (transport_map, (void *)(intptr_t) fd);
996 ++transport_map_modified_tick;