1 /*************************************************
2 * Exim - an Internet mail transport agent *
3 *************************************************/
5 /* Copyright (c) University of Cambridge 1995 - 2018 */
6 /* Copyright (c) The Exim Maintainers 2020 */
7 /* See the file NOTICE for conditions of use and distribution. */
9 /* Functions for doing things with sockets. With the advent of IPv6 this has
10 got messier, so that it's worth pulling out the code into separate functions
11 that other parts of Exim can call, especially as there are now several
12 different places in the code where sockets are used. */
18 #if defined(TCP_FASTOPEN)
19 # if defined(MSG_FASTOPEN) || defined(EXIM_TFO_CONNECTX) || defined(EXIM_TFO_FREEBSD)
20 # define EXIM_SUPPORT_TFO
24 /*************************************************
26 *************************************************/
28 /* Socket creation happens in a number of places so it's packaged here for
32 type SOCK_DGRAM or SOCK_STREAM
33 af AF_INET or AF_INET6
35 Returns: socket number or -1 on failure
39 ip_socket(int type, int af)
41 int sock = socket(af, type, 0);
43 log_write(0, LOG_MAIN, "IPv%c socket creation failed: %s",
44 (af == AF_INET6)? '6':'4', strerror(errno));
52 /*************************************************
53 * Convert printing address to numeric *
54 *************************************************/
56 /* This function converts the textual form of an IP address into a numeric form
57 in an appropriate structure in an IPv6 environment. The getaddrinfo() function
58 can (apparently) handle more complicated addresses (e.g. those containing
59 scopes) than inet_pton() in some environments. We use hints to tell it that the
60 input must be a numeric address.
62 However, apparently some operating systems (or libraries) don't support
63 getaddrinfo(), so there is a build-time option to revert to inet_pton() (which
64 does not support scopes).
67 address textual form of the address
68 addr where to copy back the answer
70 Returns: nothing - failure provokes a panic-die
74 ip_addrinfo(const uschar *address, struct sockaddr_in6 *saddr)
76 #ifdef IPV6_USE_INET_PTON
78 if (inet_pton(AF_INET6, CCS address, &saddr->sin6_addr) != 1)
79 log_write(0, LOG_MAIN|LOG_PANIC_DIE, "unable to parse \"%s\" as an "
80 "IP address", address);
81 saddr->sin6_family = AF_INET6;
86 struct addrinfo hints, *res;
87 memset(&hints, 0, sizeof(hints));
88 hints.ai_family = AF_INET6;
89 hints.ai_socktype = SOCK_STREAM;
90 hints.ai_flags = AI_NUMERICHOST;
91 if ((rc = getaddrinfo(CCS address, NULL, &hints, &res)) != 0 || res == NULL)
92 log_write(0, LOG_MAIN|LOG_PANIC_DIE, "unable to parse \"%s\" as an "
93 "IP address: %s", address,
94 (rc == 0)? "NULL result returned" : gai_strerror(rc));
95 memcpy(saddr, res->ai_addr, res->ai_addrlen);
100 #endif /* HAVE_IPV6 */
103 /*************************************************
104 * Bind socket to interface and port *
105 *************************************************/
108 ip_addr(void * sin_, int af, const uschar * address, int port)
110 union sockaddr_46 * sin = sin_;
111 memset(sin, 0, sizeof(*sin));
113 /* Setup code when using an IPv6 socket. The wildcard address is ":", to
114 ensure an IPv6 socket is used. */
119 if (address[0] == ':' && address[1] == 0)
121 sin->v6.sin6_family = AF_INET6;
122 sin->v6.sin6_addr = in6addr_any;
125 ip_addrinfo(address, &sin->v6); /* Panic-dies on error */
126 sin->v6.sin6_port = htons(port);
127 return sizeof(sin->v6);
130 #else /* HAVE_IPv6 */
131 af = af; /* Avoid compiler warning */
132 #endif /* HAVE_IPV6 */
134 /* Setup code when using IPv4 socket. The wildcard address is "". */
137 sin->v4.sin_family = AF_INET;
138 sin->v4.sin_port = htons(port);
139 sin->v4.sin_addr.s_addr = address[0] == 0
140 ? (S_ADDR_TYPE)INADDR_ANY
141 : (S_ADDR_TYPE)inet_addr(CS address);
142 return sizeof(sin->v4);
148 /* This function binds a socket to a local interface address and port. For a
149 wildcard IPv6 bind, the address is ":".
153 af AF_INET or AF_INET6 - the socket type
154 address the IP address, in text form
155 port the IP port (host order)
157 Returns: the result of bind()
161 ip_bind(int sock, int af, uschar *address, int port)
163 union sockaddr_46 sin;
164 int s_len = ip_addr(&sin, af, address, port);
165 return bind(sock, (struct sockaddr *)&sin, s_len);
170 /*************************************************
171 * Connect socket to remote host *
172 *************************************************/
174 /* This function connects a socket to a remote address and port. The socket may
175 or may not have previously been bound to a local interface. The socket is not
176 closed, even in cases of error. It is expected that the calling function, which
177 created the socket, will be the one that closes it.
181 af AF_INET6 or AF_INET for the socket type
182 address the remote address, in text form
184 timeout a timeout (zero for indefinite timeout)
185 fastopen_blob non-null iff TCP_FASTOPEN can be used; may indicate early-data to
186 be sent in SYN segment. Any such data must be idempotent.
188 Returns: 0 on success; -1 on failure, with errno set
192 ip_connect(int sock, int af, const uschar *address, int port, int timeout,
193 const blob * fastopen_blob)
195 struct sockaddr_in s_in4;
196 struct sockaddr *s_ptr;
197 int s_len, rc, save_errno;
199 /* For an IPv6 address, use an IPv6 sockaddr structure. */
202 struct sockaddr_in6 s_in6;
205 memset(&s_in6, 0, sizeof(s_in6));
206 ip_addrinfo(address, &s_in6); /* Panic-dies on error */
207 s_in6.sin6_port = htons(port);
208 s_ptr = (struct sockaddr *)&s_in6;
209 s_len = sizeof(s_in6);
212 #else /* HAVE_IPV6 */
213 af = af; /* Avoid compiler warning */
214 #endif /* HAVE_IPV6 */
216 /* For an IPv4 address, use an IPv4 sockaddr structure, even on a system with
220 memset(&s_in4, 0, sizeof(s_in4));
221 s_in4.sin_family = AF_INET;
222 s_in4.sin_port = htons(port);
223 s_in4.sin_addr.s_addr = (S_ADDR_TYPE)inet_addr(CCS address);
224 s_ptr = (struct sockaddr *)&s_in4;
225 s_len = sizeof(s_in4);
228 /* If no connection timeout is set, just call connect() without setting a
229 timer, thereby allowing the inbuilt OS timeout to operate. */
231 callout_address = string_sprintf("[%s]:%d", address, port);
232 sigalrm_seen = FALSE;
233 if (timeout > 0) ALARM(timeout);
235 #ifdef EXIM_SUPPORT_TFO
236 /* TCP Fast Open, if the system has a cookie from a previous call to
237 this peer, can send data in the SYN packet. The peer can send data
238 before it gets our ACK of its SYN,ACK - the latter is useful for
239 the SMTP banner. Other (than SMTP) cases of TCP connections can
240 possibly use the data-on-syn, so support that too. */
242 if (fastopen_blob && f.tcp_fastopen_ok)
245 /* This is a Linux implementation. */
247 if ((rc = sendto(sock, fastopen_blob->data, fastopen_blob->len,
248 MSG_FASTOPEN | MSG_DONTWAIT, s_ptr, s_len)) >= 0)
249 /* seen for with-data, experimental TFO option, with-cookie case */
250 /* seen for with-data, proper TFO opt, with-cookie case */
252 DEBUG(D_transport|D_v)
253 debug_printf(" TFO mode connection attempt to %s, %lu data\n",
254 address, (unsigned long)fastopen_blob->len);
255 /*XXX also seen on successful TFO, sigh */
256 tcp_out_fastopen = fastopen_blob->len > 0 ? TFO_ATTEMPTED_DATA : TFO_ATTEMPTED_NODATA;
260 case EINPROGRESS: /* expected if we had no cookie for peer */
261 /* seen for no-data, proper TFO option, both cookie-request and with-cookie cases */
262 /* apparently no visibility of the diffference at this point */
263 /* seen for with-data, proper TFO opt, cookie-req */
264 /* with netwk delay, post-conn tcp_info sees unacked 1 for R, 2 for C; code in smtp_out.c */
265 /* ? older Experimental TFO option behaviour ? */
266 DEBUG(D_transport|D_v) debug_printf(" TFO mode sendto, %s data: EINPROGRESS\n",
267 fastopen_blob->len > 0 ? "with" : "no");
268 if (!fastopen_blob->data)
270 tcp_out_fastopen = TFO_ATTEMPTED_NODATA; /* we tried; unknown if useful yet */
273 else /* queue unsent data */
274 rc = send(sock, fastopen_blob->data, fastopen_blob->len, 0);
279 debug_printf("Tried TCP Fast Open but apparently not enabled by sysctl\n");
284 debug_printf("Tried TCP Fast Open but kernel too old to support it\n");
288 # elif defined(EXIM_TFO_FREEBSD)
289 /* Re: https://people.freebsd.org/~pkelsey/tfo-tools/tfo-client.c */
291 if (setsockopt(sock, IPPROTO_TCP, TCP_FASTOPEN, &on, sizeof(on)) < 0)
294 debug_printf("Tried TCP Fast Open but apparently not enabled by sysctl\n");
297 if ((rc = sendto(sock, fastopen_blob->data, fastopen_blob->len, 0,
300 DEBUG(D_transport|D_v)
301 debug_printf(" TFO mode connection attempt to %s, %lu data\n",
302 address, (unsigned long)fastopen_blob->len);
303 tcp_out_fastopen = fastopen_blob->len > 0 ? TFO_ATTEMPTED_DATA : TFO_ATTEMPTED_NODATA;
306 # elif defined(EXIM_TFO_CONNECTX)
308 sa_endpoints_t ends = {
309 .sae_srcif = 0, .sae_srcaddr = NULL, .sae_srcaddrlen = 0,
310 .sae_dstaddr = s_ptr, .sae_dstaddrlen = s_len };
312 .iov_base = fastopen_blob->data, .iov_len = fastopen_blob->len };
315 if ((rc = connectx(sock, &ends, SAE_ASSOCID_ANY,
316 CONNECT_DATA_IDEMPOTENT, &iov, 1, &len, NULL)) == 0)
318 DEBUG(D_transport|D_v)
319 debug_printf(" TFO mode connection attempt to %s, %lu data\n",
320 address, (unsigned long)fastopen_blob->len);
321 tcp_out_fastopen = fastopen_blob->len > 0 ? TFO_ATTEMPTED_DATA : TFO_ATTEMPTED_NODATA;
323 if (len != fastopen_blob->len)
324 DEBUG(D_transport|D_v)
325 debug_printf(" only queued %lu data!\n", (unsigned long)len);
327 else if (errno == EINPROGRESS)
329 DEBUG(D_transport|D_v) debug_printf(" TFO mode connectx, %s data: EINPROGRESS\n",
330 fastopen_blob->len > 0 ? "with" : "no");
331 if (!fastopen_blob->data)
333 tcp_out_fastopen = TFO_ATTEMPTED_NODATA; /* we tried; unknown if useful yet */
336 else /* assume that no data was queued; block in send */
337 rc = send(sock, fastopen_blob->data, fastopen_blob->len, 0);
342 #endif /*EXIM_SUPPORT_TFO*/
344 #if defined(EXIM_SUPPORT_TFO) && !defined(EXIM_TFO_CONNECTX)
348 DEBUG(D_transport|D_v) if (fastopen_blob)
349 debug_printf(" non-TFO mode connection attempt to %s, %lu data\n",
350 address, (unsigned long)fastopen_blob->len);
351 if ((rc = connect(sock, s_ptr, s_len)) >= 0)
352 if ( fastopen_blob && fastopen_blob->data && fastopen_blob->len
353 && send(sock, fastopen_blob->data, fastopen_blob->len, 0) < 0)
360 /* There is a testing facility for simulating a connection timeout, as I
361 can't think of any other way of doing this. It converts a connection refused
362 into a timeout if the timeout is set to 999999. */
364 if (f.running_in_test_harness && save_errno == ECONNREFUSED && timeout == 999999)
376 /* A failure whose error code is "Interrupted system call" is in fact
377 an externally applied timeout if the signal handler has been run. */
379 errno = save_errno == EINTR && sigalrm_seen ? ETIMEDOUT : save_errno;
385 /*************************************************
386 * Create connected socket to remote host *
387 *************************************************/
389 /* Create a socket and connect to host (name or number, ipv6 ok)
390 at one of port-range.
393 type SOCK_DGRAM or SOCK_STREAM
394 af AF_INET6 or AF_INET for the socket type
395 hostname host name, or ip address (as text)
396 portlo,porthi the remote port range
398 connhost if not NULL, host_item to be filled in with connection details
399 errstr pointer for allocated string on error
400 fastopen_blob with SOCK_STREAM, if non-null, request TCP Fast Open.
401 Additionally, optional idempotent early-data to send
404 socket fd, or -1 on failure (having allocated an error string)
407 ip_connectedsocket(int type, const uschar * hostname, int portlo, int porthi,
408 int timeout, host_item * connhost, uschar ** errstr, const blob * fastopen_blob)
412 int af = 0, fd, fd4 = -1, fd6 = -1;
415 shost.address = NULL;
419 namelen = Ustrlen(hostname);
421 /* Anything enclosed in [] must be an IP address. */
423 if (hostname[0] == '[' &&
424 hostname[namelen - 1] == ']')
426 uschar * host = string_copyn(hostname+1, namelen-2);
427 if (string_is_ip_address(host, NULL) == 0)
429 *errstr = string_sprintf("malformed IP address \"%s\"", hostname);
432 shost.name = shost.address = host;
435 /* Otherwise check for an unadorned IP address */
437 else if (string_is_ip_address(hostname, NULL) != 0)
438 shost.name = shost.address = string_copyn(hostname, namelen);
440 /* Otherwise lookup IP address(es) from the name */
444 shost.name = string_copyn(hostname, namelen);
445 if (host_find_byname(&shost, NULL, HOST_FIND_QUALIFY_SINGLE,
446 NULL, FALSE) != HOST_FOUND)
448 *errstr = string_sprintf("no IP address found for host %s", shost.name);
453 /* Try to connect to the server - test each IP till one works */
455 for (host_item * h = &shost; h; h = h->next)
457 fd = Ustrchr(h->address, ':') != 0
458 ? fd6 < 0 ? (fd6 = ip_socket(type, af = AF_INET6)) : fd6
459 : fd4 < 0 ? (fd4 = ip_socket(type, af = AF_INET )) : fd4;
463 *errstr = string_sprintf("failed to create socket: %s", strerror(errno));
467 for (int port = portlo; port <= porthi; port++)
468 if (ip_connect(fd, af, h->address, port, timeout, fastopen_blob) == 0)
470 if (fd != fd6) close(fd6);
471 if (fd != fd4) close(fd4);
476 connhost->next = NULL;
482 *errstr = string_sprintf("failed to connect to any address for %s: %s",
483 hostname, strerror(errno));
486 close(fd4); close(fd6); return -1;
492 ip_tcpsocket(const uschar * hostport, uschar ** errstr, int tmo,
493 host_item * connhost)
496 uschar hostname[256];
497 unsigned int portlow, porthigh;
499 /* extract host and port part */
500 scan = sscanf(CS hostport, "%255s %u-%u", hostname, &portlow, &porthigh);
505 *errstr = string_sprintf("invalid socket '%s'", hostport);
511 return ip_connectedsocket(SOCK_STREAM, hostname, portlow, porthigh,
512 tmo, connhost, errstr, NULL);
516 ip_unixsocket(const uschar * path, uschar ** errstr)
519 struct sockaddr_un server;
521 if ((sock = socket(AF_UNIX, SOCK_STREAM, 0)) < 0)
523 *errstr = US"can't open UNIX socket.";
527 callout_address = string_copy(path);
528 server.sun_family = AF_UNIX;
529 Ustrncpy(US server.sun_path, path, sizeof(server.sun_path)-1);
530 server.sun_path[sizeof(server.sun_path)-1] = '\0';
531 if (connect(sock, (struct sockaddr *) &server, sizeof(server)) < 0)
535 *errstr = string_sprintf("unable to connect to UNIX socket (%s): %s",
536 path, strerror(err));
542 /* spec is either an absolute path (with a leading /), or
543 a host (name or IP) and port (whitespace-separated).
544 The port can be a range, dash-separated, or a single number.
546 For a TCP socket, optionally fill in a host_item.
549 ip_streamsocket(const uschar * spec, uschar ** errstr, int tmo,
550 host_item * connhost)
553 ? ip_unixsocket(spec, errstr) : ip_tcpsocket(spec, errstr, tmo, connhost);
556 /*************************************************
557 * Set keepalive on a socket *
558 *************************************************/
560 /* Can be called for both incoming and outgoing sockets.
564 address the remote host address, for failure logging
565 torf true for outgoing connection, false for incoming
571 ip_keepalive(int sock, const uschar *address, BOOL torf)
574 if (setsockopt(sock, SOL_SOCKET, SO_KEEPALIVE,
575 US (&fodder), sizeof(fodder)) != 0)
576 log_write(0, LOG_MAIN, "setsockopt(SO_KEEPALIVE) on connection %s %s "
577 "failed: %s", torf? "to":"from", address, strerror(errno));
582 /*************************************************
583 * Receive from a socket with timeout *
584 *************************************************/
588 fd the file descriptor
589 timelimit the timeout endpoint, seconds-since-epoch
590 Returns: TRUE => ready for i/o
591 FALSE => timed out, or other error
594 fd_ready(int fd, time_t timelimit)
597 int time_left = timelimit - time(NULL);
605 /* Wait until the socket is ready */
609 struct timeval tv = { .tv_sec = time_left, .tv_usec = 0 };
610 FD_ZERO (&select_inset);
611 FD_SET (fd, &select_inset);
613 /*DEBUG(D_transport) debug_printf("waiting for data on fd\n");*/
614 rc = select(fd + 1, (SELECT_ARG2_TYPE *)&select_inset, NULL, NULL, &tv);
616 /* If some interrupt arrived, just retry. We presume this to be rare,
617 but it can happen (e.g. the SIGUSR1 signal sent by exiwhat causes
620 Aug 2004: Somebody set up a cron job that ran exiwhat every 2 minutes, making
621 the interrupt not at all rare. Since the timeout is typically more than 2
622 minutes, the effect was to block the timeout completely. To prevent this
623 happening again, we do an explicit time test and adjust the timeout
626 if (rc < 0 && errno == EINTR)
628 DEBUG(D_transport) debug_printf("EINTR while waiting for socket data\n");
630 /* Watch out, 'continue' jumps to the condition, not to the loops top */
631 if ((time_left = timelimit - time(NULL)) > 0) continue;
640 /* Checking the FD_ISSET is not enough, if we're interrupted, the
641 select_inset may still contain the 'input'. */
643 while (rc < 0 || !FD_ISSET(fd, &select_inset));
647 /* The timeout is implemented using select(), and we loop to cover select()
648 getting interrupted, and the possibility of select() returning with a positive
649 result but no ready descriptor. Is this in fact possible?
652 cctx the connection context (socket fd, possibly TLS context)
654 bufsize the buffer size
655 timelimit the timeout endpoint, seconds-since-epoch
657 Returns: > 0 => that much data read
658 <= 0 on error or EOF; errno set - zero for EOF
662 ip_recv(client_conn_ctx * cctx, uschar * buffer, int buffsize, time_t timelimit)
666 if (!fd_ready(cctx->sock, timelimit))
669 /* The socket is ready, read from it (via TLS if it's active). On EOF (i.e.
670 close down of the connection), set errno to zero; otherwise leave it alone. */
673 if (cctx->tls_ctx) /* client TLS */
674 rc = tls_read(cctx->tls_ctx, buffer, buffsize);
675 else if (tls_in.active.sock == cctx->sock) /* server TLS */
676 rc = tls_read(NULL, buffer, buffsize);
679 rc = recv(cctx->sock, buffer, buffsize, 0);
681 if (rc > 0) return rc;
682 if (rc == 0) errno = 0;
689 /*************************************************
690 * Lookup address family of potential socket *
691 *************************************************/
693 /* Given a file-descriptor, check to see if it's a socket and, if so,
694 return the address family; detects IPv4 vs IPv6. If not a socket then
697 The value 0 is typically AF_UNSPEC, which should not be seen on a connected
698 fd. If the return is -1, the errno will be from getsockname(); probably
699 ENOTSOCK or ECONNRESET.
701 Arguments: socket-or-not fd
702 Returns: address family or -1
706 ip_get_address_family(int fd)
708 struct sockaddr_storage ss;
709 socklen_t sslen = sizeof(ss);
711 if (getsockname(fd, (struct sockaddr *) &ss, &sslen) < 0)
714 return (int) ss.ss_family;
720 /*************************************************
721 * Lookup DSCP settings for a socket *
722 *************************************************/
724 struct dscp_name_tableentry {
728 /* Keep both of these tables sorted! */
729 static struct dscp_name_tableentry dscp_table[] = {
730 #ifdef IPTOS_DSCP_AF11
731 { CUS"af11", IPTOS_DSCP_AF11 },
732 { CUS"af12", IPTOS_DSCP_AF12 },
733 { CUS"af13", IPTOS_DSCP_AF13 },
734 { CUS"af21", IPTOS_DSCP_AF21 },
735 { CUS"af22", IPTOS_DSCP_AF22 },
736 { CUS"af23", IPTOS_DSCP_AF23 },
737 { CUS"af31", IPTOS_DSCP_AF31 },
738 { CUS"af32", IPTOS_DSCP_AF32 },
739 { CUS"af33", IPTOS_DSCP_AF33 },
740 { CUS"af41", IPTOS_DSCP_AF41 },
741 { CUS"af42", IPTOS_DSCP_AF42 },
742 { CUS"af43", IPTOS_DSCP_AF43 },
743 { CUS"ef", IPTOS_DSCP_EF },
746 { CUS"lowcost", IPTOS_LOWCOST },
748 { CUS"lowdelay", IPTOS_LOWDELAY },
750 { CUS"mincost", IPTOS_MINCOST },
752 { CUS"reliability", IPTOS_RELIABILITY },
753 { CUS"throughput", IPTOS_THROUGHPUT }
755 static int dscp_table_size =
756 sizeof(dscp_table) / sizeof(struct dscp_name_tableentry);
758 /* DSCP values change by protocol family, and so do the options used for
759 setsockopt(); this utility does all the lookups. It takes an unexpanded
760 option string, expands it, strips off affix whitespace, then checks if it's
761 a number. If all of what's left is a number, then that's how the option will
762 be parsed and success/failure is a range check. If it's not all a number,
763 then it must be a supported keyword.
766 dscp_name a string, so far unvalidated
767 af address_family in use
768 level setsockopt level to use
769 optname setsockopt name to use
770 dscp_value value for dscp_name
772 Returns: TRUE if okay to setsockopt(), else FALSE
774 *level and *optname may be set even if FALSE is returned
778 dscp_lookup(const uschar *dscp_name, int af,
779 int *level, int *optname, int *dscp_value)
781 uschar *dscp_lookup, *p;
790 #if HAVE_IPV6 && defined(IPV6_TCLASS)
791 else if (af == AF_INET6)
793 *level = IPPROTO_IPV6;
794 *optname = IPV6_TCLASS;
800 debug_printf("Unhandled address family %d in dscp_lookup()\n", af);
806 debug_printf("[empty DSCP]\n");
809 dscp_lookup = expand_string(US dscp_name);
810 if (dscp_lookup == NULL || *dscp_lookup == '\0')
813 p = dscp_lookup + Ustrlen(dscp_lookup) - 1;
814 while (isspace(*p)) *p-- = '\0';
815 while (isspace(*dscp_lookup) && dscp_lookup < p) dscp_lookup++;
816 if (*dscp_lookup == '\0')
819 rawlong = Ustrtol(dscp_lookup, &p, 0);
820 if (p != dscp_lookup && *p == '\0')
822 /* We have six bits available, which will end up shifted to fit in 0xFC mask.
823 RFC 2597 defines the values unshifted. */
824 if (rawlong < 0 || rawlong > 0x3F)
827 debug_printf("DSCP value %ld out of range, ignored.\n", rawlong);
830 *dscp_value = rawlong << 2;
835 last = dscp_table_size;
838 int middle = (first + last)/2;
839 int c = Ustrcmp(dscp_lookup, dscp_table[middle].name);
842 *dscp_value = dscp_table[middle].value;
854 dscp_list_to_stream(FILE *stream)
856 for (int i = 0; i < dscp_table_size; ++i)
857 fprintf(stream, "%s\n", dscp_table[i].name);