1 /*************************************************
2 * Exim - an Internet mail transport agent *
3 *************************************************/
5 /* Copyright (c) University of Cambridge 1995 - 2018 */
6 /* See the file NOTICE for conditions of use and distribution. */
8 /* Functions for doing things with sockets. With the advent of IPv6 this has
9 got messier, so that it's worth pulling out the code into separate functions
10 that other parts of Exim can call, especially as there are now several
11 different places in the code where sockets are used. */
17 #if defined(TCP_FASTOPEN)
18 # if defined(MSG_FASTOPEN) || defined(EXIM_TFO_CONNECTX) || defined(EXIM_TFO_FREEBSD)
19 # define EXIM_SUPPORT_TFO
23 /*************************************************
25 *************************************************/
27 /* Socket creation happens in a number of places so it's packaged here for
31 type SOCK_DGRAM or SOCK_STREAM
32 af AF_INET or AF_INET6
34 Returns: socket number or -1 on failure
38 ip_socket(int type, int af)
40 int sock = socket(af, type, 0);
42 log_write(0, LOG_MAIN, "IPv%c socket creation failed: %s",
43 (af == AF_INET6)? '6':'4', strerror(errno));
51 /*************************************************
52 * Convert printing address to numeric *
53 *************************************************/
55 /* This function converts the textual form of an IP address into a numeric form
56 in an appropriate structure in an IPv6 environment. The getaddrinfo() function
57 can (apparently) handle more complicated addresses (e.g. those containing
58 scopes) than inet_pton() in some environments. We use hints to tell it that the
59 input must be a numeric address.
61 However, apparently some operating systems (or libraries) don't support
62 getaddrinfo(), so there is a build-time option to revert to inet_pton() (which
63 does not support scopes).
66 address textual form of the address
67 addr where to copy back the answer
69 Returns: nothing - failure provokes a panic-die
73 ip_addrinfo(const uschar *address, struct sockaddr_in6 *saddr)
75 #ifdef IPV6_USE_INET_PTON
77 if (inet_pton(AF_INET6, CCS address, &saddr->sin6_addr) != 1)
78 log_write(0, LOG_MAIN|LOG_PANIC_DIE, "unable to parse \"%s\" as an "
79 "IP address", address);
80 saddr->sin6_family = AF_INET6;
85 struct addrinfo hints, *res;
86 memset(&hints, 0, sizeof(hints));
87 hints.ai_family = AF_INET6;
88 hints.ai_socktype = SOCK_STREAM;
89 hints.ai_flags = AI_NUMERICHOST;
90 if ((rc = getaddrinfo(CCS address, NULL, &hints, &res)) != 0 || res == NULL)
91 log_write(0, LOG_MAIN|LOG_PANIC_DIE, "unable to parse \"%s\" as an "
92 "IP address: %s", address,
93 (rc == 0)? "NULL result returned" : gai_strerror(rc));
94 memcpy(saddr, res->ai_addr, res->ai_addrlen);
99 #endif /* HAVE_IPV6 */
102 /*************************************************
103 * Bind socket to interface and port *
104 *************************************************/
107 ip_addr(void * sin_, int af, const uschar * address, int port)
109 union sockaddr_46 * sin = sin_;
110 memset(sin, 0, sizeof(*sin));
112 /* Setup code when using an IPv6 socket. The wildcard address is ":", to
113 ensure an IPv6 socket is used. */
118 if (address[0] == ':' && address[1] == 0)
120 sin->v6.sin6_family = AF_INET6;
121 sin->v6.sin6_addr = in6addr_any;
124 ip_addrinfo(address, &sin->v6); /* Panic-dies on error */
125 sin->v6.sin6_port = htons(port);
126 return sizeof(sin->v6);
129 #else /* HAVE_IPv6 */
130 af = af; /* Avoid compiler warning */
131 #endif /* HAVE_IPV6 */
133 /* Setup code when using IPv4 socket. The wildcard address is "". */
136 sin->v4.sin_family = AF_INET;
137 sin->v4.sin_port = htons(port);
138 sin->v4.sin_addr.s_addr = address[0] == 0
139 ? (S_ADDR_TYPE)INADDR_ANY
140 : (S_ADDR_TYPE)inet_addr(CS address);
141 return sizeof(sin->v4);
147 /* This function binds a socket to a local interface address and port. For a
148 wildcard IPv6 bind, the address is ":".
152 af AF_INET or AF_INET6 - the socket type
153 address the IP address, in text form
154 port the IP port (host order)
156 Returns: the result of bind()
160 ip_bind(int sock, int af, uschar *address, int port)
162 union sockaddr_46 sin;
163 int s_len = ip_addr(&sin, af, address, port);
164 return bind(sock, (struct sockaddr *)&sin, s_len);
169 /*************************************************
170 * Connect socket to remote host *
171 *************************************************/
173 /* This function connects a socket to a remote address and port. The socket may
174 or may not have previously been bound to a local interface. The socket is not
175 closed, even in cases of error. It is expected that the calling function, which
176 created the socket, will be the one that closes it.
180 af AF_INET6 or AF_INET for the socket type
181 address the remote address, in text form
183 timeout a timeout (zero for indefinite timeout)
184 fastopen_blob non-null iff TCP_FASTOPEN can be used; may indicate early-data to
185 be sent in SYN segment. Any such data must be idempotent.
187 Returns: 0 on success; -1 on failure, with errno set
191 ip_connect(int sock, int af, const uschar *address, int port, int timeout,
192 const blob * fastopen_blob)
194 struct sockaddr_in s_in4;
195 struct sockaddr *s_ptr;
196 int s_len, rc, save_errno;
198 /* For an IPv6 address, use an IPv6 sockaddr structure. */
201 struct sockaddr_in6 s_in6;
204 memset(&s_in6, 0, sizeof(s_in6));
205 ip_addrinfo(address, &s_in6); /* Panic-dies on error */
206 s_in6.sin6_port = htons(port);
207 s_ptr = (struct sockaddr *)&s_in6;
208 s_len = sizeof(s_in6);
211 #else /* HAVE_IPV6 */
212 af = af; /* Avoid compiler warning */
213 #endif /* HAVE_IPV6 */
215 /* For an IPv4 address, use an IPv4 sockaddr structure, even on a system with
219 memset(&s_in4, 0, sizeof(s_in4));
220 s_in4.sin_family = AF_INET;
221 s_in4.sin_port = htons(port);
222 s_in4.sin_addr.s_addr = (S_ADDR_TYPE)inet_addr(CCS address);
223 s_ptr = (struct sockaddr *)&s_in4;
224 s_len = sizeof(s_in4);
227 /* If no connection timeout is set, just call connect() without setting a
228 timer, thereby allowing the inbuilt OS timeout to operate. */
230 callout_address = string_sprintf("[%s]:%d", address, port);
231 sigalrm_seen = FALSE;
232 if (timeout > 0) ALARM(timeout);
234 #ifdef EXIM_SUPPORT_TFO
235 /* TCP Fast Open, if the system has a cookie from a previous call to
236 this peer, can send data in the SYN packet. The peer can send data
237 before it gets our ACK of its SYN,ACK - the latter is useful for
238 the SMTP banner. Other (than SMTP) cases of TCP connections can
239 possibly use the data-on-syn, so support that too. */
241 if (fastopen_blob && f.tcp_fastopen_ok)
244 /* This is a Linux implementation. */
246 if ((rc = sendto(sock, fastopen_blob->data, fastopen_blob->len,
247 MSG_FASTOPEN | MSG_DONTWAIT, s_ptr, s_len)) >= 0)
248 /* seen for with-data, experimental TFO option, with-cookie case */
249 /* seen for with-data, proper TFO opt, with-cookie case */
251 DEBUG(D_transport|D_v)
252 debug_printf(" TFO mode connection attempt to %s, %lu data\n",
253 address, (unsigned long)fastopen_blob->len);
254 /*XXX also seen on successful TFO, sigh */
255 tcp_out_fastopen = fastopen_blob->len > 0 ? TFO_ATTEMPTED_DATA : TFO_ATTEMPTED_NODATA;
257 else if (errno == EINPROGRESS) /* expected if we had no cookie for peer */
258 /* seen for no-data, proper TFO option, both cookie-request and with-cookie cases */
259 /* apparently no visibility of the diffference at this point */
260 /* seen for with-data, proper TFO opt, cookie-req */
261 /* with netwk delay, post-conn tcp_info sees unacked 1 for R, 2 for C; code in smtp_out.c */
262 /* ? older Experimental TFO option behaviour ? */
263 { /* queue unsent data */
264 DEBUG(D_transport|D_v) debug_printf(" TFO mode sendto, %s data: EINPROGRESS\n",
265 fastopen_blob->len > 0 ? "with" : "no");
266 if (!fastopen_blob->data)
268 tcp_out_fastopen = TFO_ATTEMPTED_NODATA; /* we tried; unknown if useful yet */
272 rc = send(sock, fastopen_blob->data, fastopen_blob->len, 0);
274 else if(errno == EOPNOTSUPP)
277 debug_printf("Tried TCP Fast Open but apparently not enabled by sysctl\n");
281 # elif defined(EXIM_TFO_FREEBSD)
282 /* Re: https://people.freebsd.org/~pkelsey/tfo-tools/tfo-client.c */
284 if (setsockopt(sock, IPPROTO_TCP, TCP_FASTOPEN, &on, sizeof(on)) < 0)
287 debug_printf("Tried TCP Fast Open but apparently not enabled by sysctl\n");
290 if ((rc = sendto(sock, fastopen_blob->data, fastopen_blob->len, 0,
293 DEBUG(D_transport|D_v)
294 debug_printf(" TFO mode connection attempt to %s, %lu data\n",
295 address, (unsigned long)fastopen_blob->len);
296 tcp_out_fastopen = fastopen_blob->len > 0 ? TFO_ATTEMPTED_DATA : TFO_ATTEMPTED_NODATA;
299 # elif defined(EXIM_TFO_CONNECTX)
301 sa_endpoints_t ends = {
302 .sae_srcif = 0, .sae_srcaddr = NULL, .sae_srcaddrlen = 0,
303 .sae_dstaddr = s_ptr, .sae_dstaddrlen = s_len };
305 .iov_base = fastopen_blob->data, .iov_len = fastopen_blob->len };
308 if ((rc = connectx(sock, &ends, SAE_ASSOCID_ANY,
309 CONNECT_DATA_IDEMPOTENT, &iov, 1, &len, NULL)) == 0)
311 DEBUG(D_transport|D_v)
312 debug_printf(" TFO mode connection attempt to %s, %lu data\n",
313 address, (unsigned long)fastopen_blob->len);
314 tcp_out_fastopen = fastopen_blob->len > 0 ? TFO_ATTEMPTED_DATA : TFO_ATTEMPTED_NODATA;
316 if (len != fastopen_blob->len)
317 DEBUG(D_transport|D_v)
318 debug_printf(" only queued %lu data!\n", (unsigned long)len);
320 else if (errno == EINPROGRESS)
322 DEBUG(D_transport|D_v) debug_printf(" TFO mode connectx, %s data: EINPROGRESS\n",
323 fastopen_blob->len > 0 ? "with" : "no");
324 if (!fastopen_blob->data)
326 tcp_out_fastopen = TFO_ATTEMPTED_NODATA; /* we tried; unknown if useful yet */
329 else /* assume that no data was queued; block in send */
330 rc = send(sock, fastopen_blob->data, fastopen_blob->len, 0);
335 #endif /*EXIM_SUPPORT_TFO*/
337 #if defined(EXIM_SUPPORT_TFO) && !defined(EXIM_TFO_CONNECTX)
341 DEBUG(D_transport|D_v) if (fastopen_blob)
342 debug_printf(" non-TFO mode connection attempt to %s, %lu data\n",
343 address, (unsigned long)fastopen_blob->len);
344 if ((rc = connect(sock, s_ptr, s_len)) >= 0)
345 if ( fastopen_blob && fastopen_blob->data && fastopen_blob->len
346 && send(sock, fastopen_blob->data, fastopen_blob->len, 0) < 0)
353 /* There is a testing facility for simulating a connection timeout, as I
354 can't think of any other way of doing this. It converts a connection refused
355 into a timeout if the timeout is set to 999999. */
357 if (f.running_in_test_harness && save_errno == ECONNREFUSED && timeout == 999999)
369 /* A failure whose error code is "Interrupted system call" is in fact
370 an externally applied timeout if the signal handler has been run. */
372 errno = save_errno == EINTR && sigalrm_seen ? ETIMEDOUT : save_errno;
378 /*************************************************
379 * Create connected socket to remote host *
380 *************************************************/
382 /* Create a socket and connect to host (name or number, ipv6 ok)
383 at one of port-range.
386 type SOCK_DGRAM or SOCK_STREAM
387 af AF_INET6 or AF_INET for the socket type
388 hostname host name, or ip address (as text)
389 portlo,porthi the remote port range
391 connhost if not NULL, host_item to be filled in with connection details
392 errstr pointer for allocated string on error
393 fastopen_blob with SOCK_STREAM, if non-null, request TCP Fast Open.
394 Additionally, optional idempotent early-data to send
397 socket fd, or -1 on failure (having allocated an error string)
400 ip_connectedsocket(int type, const uschar * hostname, int portlo, int porthi,
401 int timeout, host_item * connhost, uschar ** errstr, const blob * fastopen_blob)
405 int af = 0, fd, fd4 = -1, fd6 = -1;
408 shost.address = NULL;
412 namelen = Ustrlen(hostname);
414 /* Anything enclosed in [] must be an IP address. */
416 if (hostname[0] == '[' &&
417 hostname[namelen - 1] == ']')
419 uschar * host = string_copyn(hostname+1, namelen-2);
420 if (string_is_ip_address(host, NULL) == 0)
422 *errstr = string_sprintf("malformed IP address \"%s\"", hostname);
425 shost.name = shost.address = host;
428 /* Otherwise check for an unadorned IP address */
430 else if (string_is_ip_address(hostname, NULL) != 0)
431 shost.name = shost.address = string_copyn(hostname, namelen);
433 /* Otherwise lookup IP address(es) from the name */
437 shost.name = string_copyn(hostname, namelen);
438 if (host_find_byname(&shost, NULL, HOST_FIND_QUALIFY_SINGLE,
439 NULL, FALSE) != HOST_FOUND)
441 *errstr = string_sprintf("no IP address found for host %s", shost.name);
446 /* Try to connect to the server - test each IP till one works */
448 for (host_item * h = &shost; h; h = h->next)
450 fd = Ustrchr(h->address, ':') != 0
451 ? fd6 < 0 ? (fd6 = ip_socket(type, af = AF_INET6)) : fd6
452 : fd4 < 0 ? (fd4 = ip_socket(type, af = AF_INET )) : fd4;
456 *errstr = string_sprintf("failed to create socket: %s", strerror(errno));
460 for (int port = portlo; port <= porthi; port++)
461 if (ip_connect(fd, af, h->address, port, timeout, fastopen_blob) == 0)
463 if (fd != fd6) close(fd6);
464 if (fd != fd4) close(fd4);
469 connhost->next = NULL;
475 *errstr = string_sprintf("failed to connect to any address for %s: %s",
476 hostname, strerror(errno));
479 close(fd4); close(fd6); return -1;
485 ip_tcpsocket(const uschar * hostport, uschar ** errstr, int tmo)
488 uschar hostname[256];
489 unsigned int portlow, porthigh;
491 /* extract host and port part */
492 scan = sscanf(CS hostport, "%255s %u-%u", hostname, &portlow, &porthigh);
497 *errstr = string_sprintf("invalid socket '%s'", hostport);
503 return ip_connectedsocket(SOCK_STREAM, hostname, portlow, porthigh,
504 tmo, NULL, errstr, NULL);
508 ip_unixsocket(const uschar * path, uschar ** errstr)
511 struct sockaddr_un server;
513 if ((sock = socket(AF_UNIX, SOCK_STREAM, 0)) < 0)
515 *errstr = US"can't open UNIX socket.";
519 callout_address = string_copy(path);
520 server.sun_family = AF_UNIX;
521 Ustrncpy(US server.sun_path, path, sizeof(server.sun_path)-1);
522 server.sun_path[sizeof(server.sun_path)-1] = '\0';
523 if (connect(sock, (struct sockaddr *) &server, sizeof(server)) < 0)
527 *errstr = string_sprintf("unable to connect to UNIX socket (%s): %s",
528 path, strerror(err));
535 ip_streamsocket(const uschar * spec, uschar ** errstr, int tmo)
538 ? ip_unixsocket(spec, errstr) : ip_tcpsocket(spec, errstr, tmo);
541 /*************************************************
542 * Set keepalive on a socket *
543 *************************************************/
545 /* Can be called for both incoming and outgoing sockets.
549 address the remote host address, for failure logging
550 torf true for outgoing connection, false for incoming
556 ip_keepalive(int sock, const uschar *address, BOOL torf)
559 if (setsockopt(sock, SOL_SOCKET, SO_KEEPALIVE,
560 US (&fodder), sizeof(fodder)) != 0)
561 log_write(0, LOG_MAIN, "setsockopt(SO_KEEPALIVE) on connection %s %s "
562 "failed: %s", torf? "to":"from", address, strerror(errno));
567 /*************************************************
568 * Receive from a socket with timeout *
569 *************************************************/
573 fd the file descriptor
574 timelimit the timeout endpoint, seconds-since-epoch
575 Returns: TRUE => ready for i/o
576 FALSE => timed out, or other error
579 fd_ready(int fd, time_t timelimit)
582 int time_left = timelimit - time(NULL);
590 /* Wait until the socket is ready */
594 struct timeval tv = { .tv_sec = time_left, .tv_usec = 0 };
595 FD_ZERO (&select_inset);
596 FD_SET (fd, &select_inset);
598 /*DEBUG(D_transport) debug_printf("waiting for data on fd\n");*/
599 rc = select(fd + 1, (SELECT_ARG2_TYPE *)&select_inset, NULL, NULL, &tv);
601 /* If some interrupt arrived, just retry. We presume this to be rare,
602 but it can happen (e.g. the SIGUSR1 signal sent by exiwhat causes
605 Aug 2004: Somebody set up a cron job that ran exiwhat every 2 minutes, making
606 the interrupt not at all rare. Since the timeout is typically more than 2
607 minutes, the effect was to block the timeout completely. To prevent this
608 happening again, we do an explicit time test and adjust the timeout
611 if (rc < 0 && errno == EINTR)
613 DEBUG(D_transport) debug_printf("EINTR while waiting for socket data\n");
615 /* Watch out, 'continue' jumps to the condition, not to the loops top */
616 if ((time_left = timelimit - time(NULL)) > 0) continue;
625 /* Checking the FD_ISSET is not enough, if we're interrupted, the
626 select_inset may still contain the 'input'. */
628 while (rc < 0 || !FD_ISSET(fd, &select_inset));
632 /* The timeout is implemented using select(), and we loop to cover select()
633 getting interrupted, and the possibility of select() returning with a positive
634 result but no ready descriptor. Is this in fact possible?
637 cctx the connection context (socket fd, possibly TLS context)
639 bufsize the buffer size
640 timelimit the timeout endpoint, seconds-since-epoch
642 Returns: > 0 => that much data read
643 <= 0 on error or EOF; errno set - zero for EOF
647 ip_recv(client_conn_ctx * cctx, uschar * buffer, int buffsize, time_t timelimit)
651 if (!fd_ready(cctx->sock, timelimit))
654 /* The socket is ready, read from it (via TLS if it's active). On EOF (i.e.
655 close down of the connection), set errno to zero; otherwise leave it alone. */
658 if (cctx->tls_ctx) /* client TLS */
659 rc = tls_read(cctx->tls_ctx, buffer, buffsize);
660 else if (tls_in.active.sock == cctx->sock) /* server TLS */
661 rc = tls_read(NULL, buffer, buffsize);
664 rc = recv(cctx->sock, buffer, buffsize, 0);
666 if (rc > 0) return rc;
667 if (rc == 0) errno = 0;
674 /*************************************************
675 * Lookup address family of potential socket *
676 *************************************************/
678 /* Given a file-descriptor, check to see if it's a socket and, if so,
679 return the address family; detects IPv4 vs IPv6. If not a socket then
682 The value 0 is typically AF_UNSPEC, which should not be seen on a connected
683 fd. If the return is -1, the errno will be from getsockname(); probably
684 ENOTSOCK or ECONNRESET.
686 Arguments: socket-or-not fd
687 Returns: address family or -1
691 ip_get_address_family(int fd)
693 struct sockaddr_storage ss;
694 socklen_t sslen = sizeof(ss);
696 if (getsockname(fd, (struct sockaddr *) &ss, &sslen) < 0)
699 return (int) ss.ss_family;
705 /*************************************************
706 * Lookup DSCP settings for a socket *
707 *************************************************/
709 struct dscp_name_tableentry {
713 /* Keep both of these tables sorted! */
714 static struct dscp_name_tableentry dscp_table[] = {
715 #ifdef IPTOS_DSCP_AF11
716 { CUS"af11", IPTOS_DSCP_AF11 },
717 { CUS"af12", IPTOS_DSCP_AF12 },
718 { CUS"af13", IPTOS_DSCP_AF13 },
719 { CUS"af21", IPTOS_DSCP_AF21 },
720 { CUS"af22", IPTOS_DSCP_AF22 },
721 { CUS"af23", IPTOS_DSCP_AF23 },
722 { CUS"af31", IPTOS_DSCP_AF31 },
723 { CUS"af32", IPTOS_DSCP_AF32 },
724 { CUS"af33", IPTOS_DSCP_AF33 },
725 { CUS"af41", IPTOS_DSCP_AF41 },
726 { CUS"af42", IPTOS_DSCP_AF42 },
727 { CUS"af43", IPTOS_DSCP_AF43 },
728 { CUS"ef", IPTOS_DSCP_EF },
731 { CUS"lowcost", IPTOS_LOWCOST },
733 { CUS"lowdelay", IPTOS_LOWDELAY },
735 { CUS"mincost", IPTOS_MINCOST },
737 { CUS"reliability", IPTOS_RELIABILITY },
738 { CUS"throughput", IPTOS_THROUGHPUT }
740 static int dscp_table_size =
741 sizeof(dscp_table) / sizeof(struct dscp_name_tableentry);
743 /* DSCP values change by protocol family, and so do the options used for
744 setsockopt(); this utility does all the lookups. It takes an unexpanded
745 option string, expands it, strips off affix whitespace, then checks if it's
746 a number. If all of what's left is a number, then that's how the option will
747 be parsed and success/failure is a range check. If it's not all a number,
748 then it must be a supported keyword.
751 dscp_name a string, so far unvalidated
752 af address_family in use
753 level setsockopt level to use
754 optname setsockopt name to use
755 dscp_value value for dscp_name
757 Returns: TRUE if okay to setsockopt(), else FALSE
759 *level and *optname may be set even if FALSE is returned
763 dscp_lookup(const uschar *dscp_name, int af,
764 int *level, int *optname, int *dscp_value)
766 uschar *dscp_lookup, *p;
775 #if HAVE_IPV6 && defined(IPV6_TCLASS)
776 else if (af == AF_INET6)
778 *level = IPPROTO_IPV6;
779 *optname = IPV6_TCLASS;
785 debug_printf("Unhandled address family %d in dscp_lookup()\n", af);
791 debug_printf("[empty DSCP]\n");
794 dscp_lookup = expand_string(US dscp_name);
795 if (dscp_lookup == NULL || *dscp_lookup == '\0')
798 p = dscp_lookup + Ustrlen(dscp_lookup) - 1;
799 while (isspace(*p)) *p-- = '\0';
800 while (isspace(*dscp_lookup) && dscp_lookup < p) dscp_lookup++;
801 if (*dscp_lookup == '\0')
804 rawlong = Ustrtol(dscp_lookup, &p, 0);
805 if (p != dscp_lookup && *p == '\0')
807 /* We have six bits available, which will end up shifted to fit in 0xFC mask.
808 RFC 2597 defines the values unshifted. */
809 if (rawlong < 0 || rawlong > 0x3F)
812 debug_printf("DSCP value %ld out of range, ignored.\n", rawlong);
815 *dscp_value = rawlong << 2;
820 last = dscp_table_size;
823 int middle = (first + last)/2;
824 int c = Ustrcmp(dscp_lookup, dscp_table[middle].name);
827 *dscp_value = dscp_table[middle].value;
839 dscp_list_to_stream(FILE *stream)
841 for (int i = 0; i < dscp_table_size; ++i)
842 fprintf(stream, "%s\n", dscp_table[i].name);