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 /*************************************************
19 *************************************************/
21 /* Socket creation happens in a number of places so it's packaged here for
25 type SOCK_DGRAM or SOCK_STREAM
26 af AF_INET or AF_INET6
28 Returns: socket number or -1 on failure
32 ip_socket(int type, int af)
34 int sock = socket(af, type, 0);
36 log_write(0, LOG_MAIN, "IPv%c socket creation failed: %s",
37 (af == AF_INET6)? '6':'4', strerror(errno));
45 /*************************************************
46 * Convert printing address to numeric *
47 *************************************************/
49 /* This function converts the textual form of an IP address into a numeric form
50 in an appropriate structure in an IPv6 environment. The getaddrinfo() function
51 can (apparently) handle more complicated addresses (e.g. those containing
52 scopes) than inet_pton() in some environments. We use hints to tell it that the
53 input must be a numeric address.
55 However, apparently some operating systems (or libraries) don't support
56 getaddrinfo(), so there is a build-time option to revert to inet_pton() (which
57 does not support scopes).
60 address textual form of the address
61 addr where to copy back the answer
63 Returns: nothing - failure provokes a panic-die
67 ip_addrinfo(const uschar *address, struct sockaddr_in6 *saddr)
69 #ifdef IPV6_USE_INET_PTON
71 if (inet_pton(AF_INET6, CCS address, &saddr->sin6_addr) != 1)
72 log_write(0, LOG_MAIN|LOG_PANIC_DIE, "unable to parse \"%s\" as an "
73 "IP address", address);
74 saddr->sin6_family = AF_INET6;
79 struct addrinfo hints, *res;
80 memset(&hints, 0, sizeof(hints));
81 hints.ai_family = AF_INET6;
82 hints.ai_socktype = SOCK_STREAM;
83 hints.ai_flags = AI_NUMERICHOST;
84 if ((rc = getaddrinfo(CCS address, NULL, &hints, &res)) != 0 || res == NULL)
85 log_write(0, LOG_MAIN|LOG_PANIC_DIE, "unable to parse \"%s\" as an "
86 "IP address: %s", address,
87 (rc == 0)? "NULL result returned" : gai_strerror(rc));
88 memcpy(saddr, res->ai_addr, res->ai_addrlen);
93 #endif /* HAVE_IPV6 */
96 /*************************************************
97 * Bind socket to interface and port *
98 *************************************************/
101 ip_addr(void * sin_, int af, const uschar * address, int port)
103 union sockaddr_46 * sin = sin_;
104 memset(sin, 0, sizeof(*sin));
106 /* Setup code when using an IPv6 socket. The wildcard address is ":", to
107 ensure an IPv6 socket is used. */
112 if (address[0] == ':' && address[1] == 0)
114 sin->v6.sin6_family = AF_INET6;
115 sin->v6.sin6_addr = in6addr_any;
118 ip_addrinfo(address, &sin->v6); /* Panic-dies on error */
119 sin->v6.sin6_port = htons(port);
120 return sizeof(sin->v6);
123 #else /* HAVE_IPv6 */
124 af = af; /* Avoid compiler warning */
125 #endif /* HAVE_IPV6 */
127 /* Setup code when using IPv4 socket. The wildcard address is "". */
130 sin->v4.sin_family = AF_INET;
131 sin->v4.sin_port = htons(port);
132 sin->v4.sin_addr.s_addr = address[0] == 0
133 ? (S_ADDR_TYPE)INADDR_ANY
134 : (S_ADDR_TYPE)inet_addr(CS address);
135 return sizeof(sin->v4);
141 /* This function binds a socket to a local interface address and port. For a
142 wildcard IPv6 bind, the address is ":".
146 af AF_INET or AF_INET6 - the socket type
147 address the IP address, in text form
148 port the IP port (host order)
150 Returns: the result of bind()
154 ip_bind(int sock, int af, uschar *address, int port)
156 union sockaddr_46 sin;
157 int s_len = ip_addr(&sin, af, address, port);
158 return bind(sock, (struct sockaddr *)&sin, s_len);
163 /*************************************************
164 *************************************************/
166 #ifdef EXIM_TFO_PROBE
171 int sock, backlog = 5;
173 if ( (sock = socket(SOCK_STREAM, AF_INET, 0)) < 0
174 && setsockopt(sock, IPPROTO_TCP, TCP_FASTOPEN, &backlog, sizeof(backlog))
176 f.tcp_fastopen_ok = TRUE;
183 /*************************************************
184 * Connect socket to remote host *
185 *************************************************/
187 /* This function connects a socket to a remote address and port. The socket may
188 or may not have previously been bound to a local interface. The socket is not
189 closed, even in cases of error. It is expected that the calling function, which
190 created the socket, will be the one that closes it.
194 af AF_INET6 or AF_INET for the socket type
195 address the remote address, in text form
197 timeout a timeout (zero for indefinite timeout)
198 fastopen_blob non-null iff TCP_FASTOPEN can be used; may indicate early-data to
199 be sent in SYN segment. Any such data must be idempotent.
201 Returns: 0 on success; -1 on failure, with errno set
205 ip_connect(int sock, int af, const uschar *address, int port, int timeout,
206 const blob * fastopen_blob)
208 struct sockaddr_in s_in4;
209 struct sockaddr *s_ptr;
210 int s_len, rc, save_errno;
212 /* For an IPv6 address, use an IPv6 sockaddr structure. */
215 struct sockaddr_in6 s_in6;
218 memset(&s_in6, 0, sizeof(s_in6));
219 ip_addrinfo(address, &s_in6); /* Panic-dies on error */
220 s_in6.sin6_port = htons(port);
221 s_ptr = (struct sockaddr *)&s_in6;
222 s_len = sizeof(s_in6);
225 #else /* HAVE_IPV6 */
226 af = af; /* Avoid compiler warning */
227 #endif /* HAVE_IPV6 */
229 /* For an IPv4 address, use an IPv4 sockaddr structure, even on a system with
233 memset(&s_in4, 0, sizeof(s_in4));
234 s_in4.sin_family = AF_INET;
235 s_in4.sin_port = htons(port);
236 s_in4.sin_addr.s_addr = (S_ADDR_TYPE)inet_addr(CCS address);
237 s_ptr = (struct sockaddr *)&s_in4;
238 s_len = sizeof(s_in4);
241 /* If no connection timeout is set, just call connect() without setting a
242 timer, thereby allowing the inbuilt OS timeout to operate. */
244 callout_address = string_sprintf("[%s]:%d", address, port);
245 sigalrm_seen = FALSE;
246 if (timeout > 0) ALARM(timeout);
249 /* TCP Fast Open, if the system has a cookie from a previous call to
250 this peer, can send data in the SYN packet. The peer can send data
251 before it gets our ACK of its SYN,ACK - the latter is useful for
252 the SMTP banner. Other (than SMTP) cases of TCP connections can
253 possibly use the data-on-syn, so support that too. */
255 if (fastopen_blob && f.tcp_fastopen_ok)
258 /* This is a Linux implementation. It might be useable on FreeBSD; I have
261 if ((rc = sendto(sock, fastopen_blob->data, fastopen_blob->len,
262 MSG_FASTOPEN | MSG_DONTWAIT, s_ptr, s_len)) >= 0)
263 /* seen for with-data, experimental TFO option, with-cookie case */
264 /* seen for with-data, proper TFO opt, with-cookie case */
266 DEBUG(D_transport|D_v)
267 debug_printf("TFO mode connection attempt to %s, %lu data\n",
268 address, (unsigned long)fastopen_blob->len);
269 /*XXX also seen on successful TFO, sigh */
270 tcp_out_fastopen = fastopen_blob->len > 0 ? TFO_ATTEMPTED_DATA : TFO_ATTEMPTED_NODATA;
272 else if (errno == EINPROGRESS) /* expected if we had no cookie for peer */
273 /* seen for no-data, proper TFO option, both cookie-request and with-cookie cases */
274 /* apparently no visibility of the diffference at this point */
275 /* seen for with-data, proper TFO opt, cookie-req */
276 /* with netwk delay, post-conn tcp_info sees unacked 1 for R, 2 for C; code in smtp_out.c */
277 /* ? older Experimental TFO option behaviour ? */
278 { /* queue unsent data */
279 DEBUG(D_transport|D_v) debug_printf("TFO mode sendto, %s data: EINPROGRESS\n",
280 fastopen_blob->len > 0 ? "with" : "no");
281 if (!fastopen_blob->data)
283 tcp_out_fastopen = TFO_ATTEMPTED_NODATA; /* we tried; unknown if useful yet */
287 rc = send(sock, fastopen_blob->data, fastopen_blob->len, 0);
289 else if(errno == EOPNOTSUPP)
292 debug_printf("Tried TCP Fast Open but apparently not enabled by sysctl\n");
296 # ifdef EXIM_TFO_CONNECTX
298 sa_endpoints_t ends = {
299 .sae_srcif = 0, .sae_srcaddr = NULL, .sae_srcaddrlen = 0,
300 .sae_dstaddr = s_ptr, .sae_dstaddrlen = s_len };
302 .iov_base = fastopen_blob->data, .iov_len = fastopen_blob->len };
305 if ((rc = connectx(sock, &ends, SAE_ASSOCID_ANY,
306 CONNECT_DATA_IDEMPOTENT, &iov, 1, &len, NULL)) == 0)
308 DEBUG(D_transport|D_v)
309 debug_printf("TFO mode connection attempt to %s, %lu data\n",
310 address, (unsigned long)fastopen_blob->len);
311 tcp_out_fastopen = fastopen_blob->len > 0 ? TFO_ATTEMPTED_DATA : TFO_ATTEMPTED_NODATA;
313 if (len != fastopen_blob->len)
314 DEBUG(D_transport|D_v)
315 debug_printf(" only queued %lu data!\n", (unsigned long)len);
317 else if (errno == EINPROGRESS)
319 DEBUG(D_transport|D_v) debug_printf("TFO mode sendto, %s data: EINPROGRESS\n",
320 fastopen_blob->len > 0 ? "with" : "no");
321 if (!fastopen_blob->data)
323 tcp_out_fastopen = TFO_ATTEMPTED_NODATA; /* we tried; unknown if useful yet */
326 else /* assume that no data was queued; block in send */
327 rc = send(sock, fastopen_blob->data, fastopen_blob->len, 0);
332 #endif /*TCP_FASTOPEN*/
334 #if defined(TCP_FASTOPEN) && defined(MSG_FASTOPEN)
338 DEBUG(D_transport|D_v) if (fastopen_blob)
339 debug_printf("non-TFO mode connection attempt to %s, %lu data\n",
340 address, (unsigned long)fastopen_blob->len);
341 if ((rc = connect(sock, s_ptr, s_len)) >= 0)
342 if ( fastopen_blob && fastopen_blob->data && fastopen_blob->len
343 && send(sock, fastopen_blob->data, fastopen_blob->len, 0) < 0)
350 /* There is a testing facility for simulating a connection timeout, as I
351 can't think of any other way of doing this. It converts a connection refused
352 into a timeout if the timeout is set to 999999. */
354 if (f.running_in_test_harness && save_errno == ECONNREFUSED && timeout == 999999)
366 /* A failure whose error code is "Interrupted system call" is in fact
367 an externally applied timeout if the signal handler has been run. */
369 errno = save_errno == EINTR && sigalrm_seen ? ETIMEDOUT : save_errno;
375 /*************************************************
376 * Create connected socket to remote host *
377 *************************************************/
379 /* Create a socket and connect to host (name or number, ipv6 ok)
380 at one of port-range.
383 type SOCK_DGRAM or SOCK_STREAM
384 af AF_INET6 or AF_INET for the socket type
385 hostname host name, or ip address (as text)
386 portlo,porthi the remote port range
388 connhost if not NULL, host_item to be filled in with connection details
389 errstr pointer for allocated string on error
390 fastopen_blob with SOCK_STREAM, if non-null, request TCP Fast Open.
391 Additionally, optional idempotent early-data to send
394 socket fd, or -1 on failure (having allocated an error string)
397 ip_connectedsocket(int type, const uschar * hostname, int portlo, int porthi,
398 int timeout, host_item * connhost, uschar ** errstr, const blob * fastopen_blob)
402 int af = 0, fd, fd4 = -1, fd6 = -1;
405 shost.address = NULL;
409 namelen = Ustrlen(hostname);
411 /* Anything enclosed in [] must be an IP address. */
413 if (hostname[0] == '[' &&
414 hostname[namelen - 1] == ']')
416 uschar * host = string_copyn(hostname+1, namelen-2);
417 if (string_is_ip_address(host, NULL) == 0)
419 *errstr = string_sprintf("malformed IP address \"%s\"", hostname);
422 shost.name = shost.address = host;
425 /* Otherwise check for an unadorned IP address */
427 else if (string_is_ip_address(hostname, NULL) != 0)
428 shost.name = shost.address = string_copyn(hostname, namelen);
430 /* Otherwise lookup IP address(es) from the name */
434 shost.name = string_copyn(hostname, namelen);
435 if (host_find_byname(&shost, NULL, HOST_FIND_QUALIFY_SINGLE,
436 NULL, FALSE) != HOST_FOUND)
438 *errstr = string_sprintf("no IP address found for host %s", shost.name);
443 /* Try to connect to the server - test each IP till one works */
445 for (host_item * h = &shost; h; h = h->next)
447 fd = Ustrchr(h->address, ':') != 0
448 ? fd6 < 0 ? (fd6 = ip_socket(type, af = AF_INET6)) : fd6
449 : fd4 < 0 ? (fd4 = ip_socket(type, af = AF_INET )) : fd4;
453 *errstr = string_sprintf("failed to create socket: %s", strerror(errno));
457 for (int port = portlo; port <= porthi; port++)
458 if (ip_connect(fd, af, h->address, port, timeout, fastopen_blob) == 0)
460 if (fd != fd6) close(fd6);
461 if (fd != fd4) close(fd4);
466 connhost->next = NULL;
472 *errstr = string_sprintf("failed to connect to any address for %s: %s",
473 hostname, strerror(errno));
476 close(fd4); close(fd6); return -1;
482 ip_tcpsocket(const uschar * hostport, uschar ** errstr, int tmo)
485 uschar hostname[256];
486 unsigned int portlow, porthigh;
488 /* extract host and port part */
489 scan = sscanf(CS hostport, "%255s %u-%u", hostname, &portlow, &porthigh);
494 *errstr = string_sprintf("invalid socket '%s'", hostport);
500 return ip_connectedsocket(SOCK_STREAM, hostname, portlow, porthigh,
501 tmo, NULL, errstr, NULL);
505 ip_unixsocket(const uschar * path, uschar ** errstr)
508 struct sockaddr_un server;
510 if ((sock = socket(AF_UNIX, SOCK_STREAM, 0)) < 0)
512 *errstr = US"can't open UNIX socket.";
516 callout_address = string_copy(path);
517 server.sun_family = AF_UNIX;
518 Ustrncpy(server.sun_path, path, sizeof(server.sun_path)-1);
519 server.sun_path[sizeof(server.sun_path)-1] = '\0';
520 if (connect(sock, (struct sockaddr *) &server, sizeof(server)) < 0)
524 *errstr = string_sprintf("unable to connect to UNIX socket (%s): %s",
525 path, strerror(err));
532 ip_streamsocket(const uschar * spec, uschar ** errstr, int tmo)
535 ? ip_unixsocket(spec, errstr) : ip_tcpsocket(spec, errstr, tmo);
538 /*************************************************
539 * Set keepalive on a socket *
540 *************************************************/
542 /* Can be called for both incoming and outgoing sockets.
546 address the remote host address, for failure logging
547 torf true for outgoing connection, false for incoming
553 ip_keepalive(int sock, const uschar *address, BOOL torf)
556 if (setsockopt(sock, SOL_SOCKET, SO_KEEPALIVE,
557 US (&fodder), sizeof(fodder)) != 0)
558 log_write(0, LOG_MAIN, "setsockopt(SO_KEEPALIVE) on connection %s %s "
559 "failed: %s", torf? "to":"from", address, strerror(errno));
564 /*************************************************
565 * Receive from a socket with timeout *
566 *************************************************/
570 fd the file descriptor
571 timeout the timeout, seconds
572 Returns: TRUE => ready for i/o
573 FALSE => timed out, or other error
576 fd_ready(int fd, int timeout)
579 time_t start_recv = time(NULL);
580 int time_left = timeout;
588 /* Wait until the socket is ready */
592 struct timeval tv = { .tv_sec = time_left, .tv_usec = 0 };
593 FD_ZERO (&select_inset);
594 FD_SET (fd, &select_inset);
596 /*DEBUG(D_transport) debug_printf("waiting for data on fd\n");*/
597 rc = select(fd + 1, (SELECT_ARG2_TYPE *)&select_inset, NULL, NULL, &tv);
599 /* If some interrupt arrived, just retry. We presume this to be rare,
600 but it can happen (e.g. the SIGUSR1 signal sent by exiwhat causes
603 Aug 2004: Somebody set up a cron job that ran exiwhat every 2 minutes, making
604 the interrupt not at all rare. Since the timeout is typically more than 2
605 minutes, the effect was to block the timeout completely. To prevent this
606 happening again, we do an explicit time test and adjust the timeout
609 if (rc < 0 && errno == EINTR)
611 DEBUG(D_transport) debug_printf("EINTR while waiting for socket data\n");
613 /* Watch out, 'continue' jumps to the condition, not to the loops top */
614 time_left = timeout - (time(NULL) - start_recv);
615 if (time_left > 0) continue;
624 /* Checking the FD_ISSET is not enough, if we're interrupted, the
625 select_inset may still contain the 'input'. */
627 while (rc < 0 || !FD_ISSET(fd, &select_inset));
631 /* The timeout is implemented using select(), and we loop to cover select()
632 getting interrupted, and the possibility of select() returning with a positive
633 result but no ready descriptor. Is this in fact possible?
636 cctx the connection context (socket fd, possibly TLS context)
638 bufsize the buffer size
641 Returns: > 0 => that much data read
642 <= 0 on error or EOF; errno set - zero for EOF
646 ip_recv(client_conn_ctx * cctx, uschar * buffer, int buffsize, int timeout)
650 if (!fd_ready(cctx->sock, timeout))
653 /* The socket is ready, read from it (via TLS if it's active). On EOF (i.e.
654 close down of the connection), set errno to zero; otherwise leave it alone. */
657 if (cctx->tls_ctx) /* client TLS */
658 rc = tls_read(cctx->tls_ctx, buffer, buffsize);
659 else if (tls_in.active.sock == cctx->sock) /* server TLS */
660 rc = tls_read(NULL, buffer, buffsize);
663 rc = recv(cctx->sock, buffer, buffsize, 0);
665 if (rc > 0) return rc;
666 if (rc == 0) errno = 0;
673 /*************************************************
674 * Lookup address family of potential socket *
675 *************************************************/
677 /* Given a file-descriptor, check to see if it's a socket and, if so,
678 return the address family; detects IPv4 vs IPv6. If not a socket then
681 The value 0 is typically AF_UNSPEC, which should not be seen on a connected
682 fd. If the return is -1, the errno will be from getsockname(); probably
683 ENOTSOCK or ECONNRESET.
685 Arguments: socket-or-not fd
686 Returns: address family or -1
690 ip_get_address_family(int fd)
692 struct sockaddr_storage ss;
693 socklen_t sslen = sizeof(ss);
695 if (getsockname(fd, (struct sockaddr *) &ss, &sslen) < 0)
698 return (int) ss.ss_family;
704 /*************************************************
705 * Lookup DSCP settings for a socket *
706 *************************************************/
708 struct dscp_name_tableentry {
712 /* Keep both of these tables sorted! */
713 static struct dscp_name_tableentry dscp_table[] = {
714 #ifdef IPTOS_DSCP_AF11
715 { CUS"af11", IPTOS_DSCP_AF11 },
716 { CUS"af12", IPTOS_DSCP_AF12 },
717 { CUS"af13", IPTOS_DSCP_AF13 },
718 { CUS"af21", IPTOS_DSCP_AF21 },
719 { CUS"af22", IPTOS_DSCP_AF22 },
720 { CUS"af23", IPTOS_DSCP_AF23 },
721 { CUS"af31", IPTOS_DSCP_AF31 },
722 { CUS"af32", IPTOS_DSCP_AF32 },
723 { CUS"af33", IPTOS_DSCP_AF33 },
724 { CUS"af41", IPTOS_DSCP_AF41 },
725 { CUS"af42", IPTOS_DSCP_AF42 },
726 { CUS"af43", IPTOS_DSCP_AF43 },
727 { CUS"ef", IPTOS_DSCP_EF },
730 { CUS"lowcost", IPTOS_LOWCOST },
732 { CUS"lowdelay", IPTOS_LOWDELAY },
734 { CUS"mincost", IPTOS_MINCOST },
736 { CUS"reliability", IPTOS_RELIABILITY },
737 { CUS"throughput", IPTOS_THROUGHPUT }
739 static int dscp_table_size =
740 sizeof(dscp_table) / sizeof(struct dscp_name_tableentry);
742 /* DSCP values change by protocol family, and so do the options used for
743 setsockopt(); this utility does all the lookups. It takes an unexpanded
744 option string, expands it, strips off affix whitespace, then checks if it's
745 a number. If all of what's left is a number, then that's how the option will
746 be parsed and success/failure is a range check. If it's not all a number,
747 then it must be a supported keyword.
750 dscp_name a string, so far unvalidated
751 af address_family in use
752 level setsockopt level to use
753 optname setsockopt name to use
754 dscp_value value for dscp_name
756 Returns: TRUE if okay to setsockopt(), else FALSE
758 *level and *optname may be set even if FALSE is returned
762 dscp_lookup(const uschar *dscp_name, int af,
763 int *level, int *optname, int *dscp_value)
765 uschar *dscp_lookup, *p;
774 #if HAVE_IPV6 && defined(IPV6_TCLASS)
775 else if (af == AF_INET6)
777 *level = IPPROTO_IPV6;
778 *optname = IPV6_TCLASS;
784 debug_printf("Unhandled address family %d in dscp_lookup()\n", af);
790 debug_printf("[empty DSCP]\n");
793 dscp_lookup = expand_string(US dscp_name);
794 if (dscp_lookup == NULL || *dscp_lookup == '\0')
797 p = dscp_lookup + Ustrlen(dscp_lookup) - 1;
798 while (isspace(*p)) *p-- = '\0';
799 while (isspace(*dscp_lookup) && dscp_lookup < p) dscp_lookup++;
800 if (*dscp_lookup == '\0')
803 rawlong = Ustrtol(dscp_lookup, &p, 0);
804 if (p != dscp_lookup && *p == '\0')
806 /* We have six bits available, which will end up shifted to fit in 0xFC mask.
807 RFC 2597 defines the values unshifted. */
808 if (rawlong < 0 || rawlong > 0x3F)
811 debug_printf("DSCP value %ld out of range, ignored.\n", rawlong);
814 *dscp_value = rawlong << 2;
819 last = dscp_table_size;
822 int middle = (first + last)/2;
823 int c = Ustrcmp(dscp_lookup, dscp_table[middle].name);
826 *dscp_value = dscp_table[middle].value;
838 dscp_list_to_stream(FILE *stream)
840 for (int i = 0; i < dscp_table_size; ++i)
841 fprintf(stream, "%s\n", dscp_table[i].name);