4c6c90878a2dee901624b6416e0973f85104ca31
[exim.git] / src / src / daemon.c
1 /* $Cambridge: exim/src/src/daemon.c,v 1.22 2007/01/23 14:34:02 ph10 Exp $ */
2
3 /*************************************************
4 * Exim - an Internet mail transport agent *
5 *************************************************/
6
7 /* Copyright (c) University of Cambridge 1995 - 2007 */
8 /* See the file NOTICE for conditions of use and distribution. */
9
10 /* Functions concerned with running Exim as a daemon */
11
12
13 #include "exim.h"
14
15
16 /* Structure for holding data for each SMTP connection */
17
18 typedef struct smtp_slot {
19 pid_t pid; /* pid of the spawned reception process */
20 uschar *host_address; /* address of the client host */
21 } smtp_slot;
22
23 /* An empty slot for initializing (Standard C does not allow constructor
24 expressions in assigments except as initializers in declarations). */
25
26 static smtp_slot empty_smtp_slot = { 0, NULL };
27
28
29
30 /*************************************************
31 * Local static variables *
32 *************************************************/
33
34 static volatile BOOL sigchld_seen;
35 static volatile BOOL sighup_seen;
36
37 static int accept_retry_count = 0;
38 static int accept_retry_errno;
39 static BOOL accept_retry_select_failed;
40
41 static int queue_run_count = 0;
42 static pid_t *queue_pid_slots = NULL;
43 static smtp_slot *smtp_slots = NULL;
44
45 static BOOL write_pid = TRUE;
46
47
48
49 /*************************************************
50 * SIGHUP Handler *
51 *************************************************/
52
53 /* All this handler does is to set a flag and re-enable the signal.
54
55 Argument: the signal number
56 Returns: nothing
57 */
58
59 static void
60 sighup_handler(int sig)
61 {
62 sig = sig; /* Keep picky compilers happy */
63 sighup_seen = TRUE;
64 signal(SIGHUP, sighup_handler);
65 }
66
67
68
69 /*************************************************
70 * SIGCHLD handler for main daemon process *
71 *************************************************/
72
73 /* Don't re-enable the handler here, since we aren't doing the
74 waiting here. If the signal is re-enabled, there will just be an
75 infinite sequence of calls to this handler. The SIGCHLD signal is
76 used just as a means of waking up the daemon so that it notices
77 terminated subprocesses as soon as possible.
78
79 Argument: the signal number
80 Returns: nothing
81 */
82
83 static void
84 main_sigchld_handler(int sig)
85 {
86 sig = sig; /* Keep picky compilers happy */
87 os_non_restarting_signal(SIGCHLD, SIG_DFL);
88 sigchld_seen = TRUE;
89 }
90
91
92
93
94 /*************************************************
95 * Unexpected errors in SMTP calls *
96 *************************************************/
97
98 /* This function just saves a bit of repetitious coding.
99
100 Arguments:
101 log_msg Text of message to be logged
102 smtp_msg Text of SMTP error message
103 was_errno The failing errno
104
105 Returns: nothing
106 */
107
108 static void
109 never_error(uschar *log_msg, uschar *smtp_msg, int was_errno)
110 {
111 uschar *emsg = (was_errno <= 0)? US"" :
112 string_sprintf(": %s", strerror(was_errno));
113 log_write(0, LOG_MAIN|LOG_PANIC, "%s%s", log_msg, emsg);
114 if (smtp_out != NULL) smtp_printf("421 %s\r\n", smtp_msg);
115 }
116
117
118
119
120 /*************************************************
121 * Handle a connected SMTP call *
122 *************************************************/
123
124 /* This function is called when an SMTP connection has been accepted.
125 If there are too many, give an error message and close down. Otherwise
126 spin off a sub-process to handle the call. The list of listening sockets
127 is required so that they can be closed in the sub-process. Take care not to
128 leak store in this process - reset the stacking pool at the end.
129
130 Arguments:
131 listen_sockets sockets which are listening for incoming calls
132 listen_socket_count count of listening sockets
133 accept_socket socket of the current accepted call
134 accepted socket information about the current call
135
136 Returns: nothing
137 */
138
139 static void
140 handle_smtp_call(int *listen_sockets, int listen_socket_count,
141 int accept_socket, struct sockaddr *accepted)
142 {
143 pid_t pid;
144 union sockaddr_46 interface_sockaddr;
145 EXIM_SOCKLEN_T ifsize = sizeof(interface_sockaddr);
146 int dup_accept_socket = -1;
147 int max_for_this_host = 0;
148 int wfsize = 0;
149 int wfptr = 0;
150 int use_log_write_selector = log_write_selector;
151 uschar *whofrom = NULL;
152
153 void *reset_point = store_get(0);
154
155 /* Make the address available in ASCII representation, and also fish out
156 the remote port. */
157
158 sender_host_address = host_ntoa(-1, accepted, NULL, &sender_host_port);
159 DEBUG(D_any) debug_printf("Connection request from %s port %d\n",
160 sender_host_address, sender_host_port);
161
162 /* Set up the output stream, check the socket has duplicated, and set up the
163 input stream. These operations fail only the exceptional circumstances. Note
164 that never_error() won't use smtp_out if it is NULL. */
165
166 smtp_out = fdopen(accept_socket, "wb");
167 if (smtp_out == NULL)
168 {
169 never_error(US"daemon: fdopen() for smtp_out failed", US"", errno);
170 goto ERROR_RETURN;
171 }
172
173 dup_accept_socket = dup(accept_socket);
174 if (dup_accept_socket < 0)
175 {
176 never_error(US"daemon: couldn't dup socket descriptor",
177 US"Connection setup failed", errno);
178 goto ERROR_RETURN;
179 }
180
181 smtp_in = fdopen(dup_accept_socket, "rb");
182 if (smtp_in == NULL)
183 {
184 never_error(US"daemon: fdopen() for smtp_in failed",
185 US"Connection setup failed", errno);
186 goto ERROR_RETURN;
187 }
188
189 /* Get the data for the local interface address. Panic for most errors, but
190 "connection reset by peer" just means the connection went away. */
191
192 if (getsockname(accept_socket, (struct sockaddr *)(&interface_sockaddr),
193 &ifsize) < 0)
194 {
195 log_write(0, LOG_MAIN | ((errno == ECONNRESET)? 0 : LOG_PANIC),
196 "getsockname() failed: %s", strerror(errno));
197 smtp_printf("421 Local problem: getsockname() failed; please try again later\r\n");
198 goto ERROR_RETURN;
199 }
200
201 interface_address = host_ntoa(-1, &interface_sockaddr, NULL, &interface_port);
202 DEBUG(D_interface) debug_printf("interface address=%s port=%d\n",
203 interface_address, interface_port);
204
205 /* Build a string identifying the remote host and, if requested, the port and
206 the local interface data. This is for logging; at the end of this function the
207 memory is reclaimed. */
208
209 whofrom = string_append(whofrom, &wfsize, &wfptr, 3, "[", sender_host_address, "]");
210
211 if ((log_extra_selector & LX_incoming_port) != 0)
212 whofrom = string_append(whofrom, &wfsize, &wfptr, 2, ":", string_sprintf("%d",
213 sender_host_port));
214
215 if ((log_extra_selector & LX_incoming_interface) != 0)
216 whofrom = string_append(whofrom, &wfsize, &wfptr, 4, " I=[",
217 interface_address, "]:", string_sprintf("%d", interface_port));
218
219 whofrom[wfptr] = 0; /* Terminate the newly-built string */
220
221 /* Check maximum number of connections. We do not check for reserved
222 connections or unacceptable hosts here. That is done in the subprocess because
223 it might take some time. */
224
225 if (smtp_accept_max > 0 && smtp_accept_count >= smtp_accept_max)
226 {
227 DEBUG(D_any) debug_printf("rejecting SMTP connection: count=%d max=%d\n",
228 smtp_accept_count, smtp_accept_max);
229 smtp_printf("421 Too many concurrent SMTP connections; "
230 "please try again later.\r\n");
231 log_write(L_connection_reject,
232 LOG_MAIN, "Connection from %s refused: too many connections",
233 whofrom);
234 goto ERROR_RETURN;
235 }
236
237 /* If a load limit above which only reserved hosts are acceptable is defined,
238 get the load average here, and if there are in fact no reserved hosts, do
239 the test right away (saves a fork). If there are hosts, do the check in the
240 subprocess because it might take time. */
241
242 if (smtp_load_reserve >= 0)
243 {
244 load_average = os_getloadavg();
245 if (smtp_reserve_hosts == NULL && load_average > smtp_load_reserve)
246 {
247 DEBUG(D_any) debug_printf("rejecting SMTP connection: load average = %.2f\n",
248 (double)load_average/1000.0);
249 smtp_printf("421 Too much load; please try again later.\r\n");
250 log_write(L_connection_reject,
251 LOG_MAIN, "Connection from %s refused: load average = %.2f",
252 whofrom, (double)load_average/1000.0);
253 goto ERROR_RETURN;
254 }
255 }
256
257 /* Check that one specific host (strictly, IP address) is not hogging
258 resources. This is done here to prevent a denial of service attack by someone
259 forcing you to fork lots of times before denying service. The value of
260 smtp_accept_max_per_host is a string which is expanded. This makes it possible
261 to provide host-specific limits according to $sender_host address, but because
262 this is in the daemon mainline, only fast expansions (such as inline address
263 checks) should be used. The documentation is full of warnings. */
264
265 if (smtp_accept_max_per_host != NULL)
266 {
267 uschar *expanded = expand_string(smtp_accept_max_per_host);
268 if (expanded == NULL)
269 {
270 if (!expand_string_forcedfail)
271 log_write(0, LOG_MAIN|LOG_PANIC, "expansion of smtp_accept_max_per_host "
272 "failed for %s: %s", whofrom, expand_string_message);
273 }
274 /* For speed, interpret a decimal number inline here */
275 else
276 {
277 uschar *s = expanded;
278 while (isdigit(*s))
279 max_for_this_host = max_for_this_host * 10 + *s++ - '0';
280 if (*s != 0)
281 log_write(0, LOG_MAIN|LOG_PANIC, "expansion of smtp_accept_max_per_host "
282 "for %s contains non-digit: %s", whofrom, expanded);
283 }
284 }
285
286 /* If we have fewer connections than max_for_this_host, we can skip the tedious
287 per host_address checks. Note that at this stage smtp_accept_count contains the
288 count of *other* connections, not including this one. */
289
290 if ((max_for_this_host > 0) &&
291 (smtp_accept_count >= max_for_this_host))
292 {
293 int i;
294 int host_accept_count = 0;
295 int other_host_count = 0; /* keep a count of non matches to optimise */
296
297 for (i = 0; i < smtp_accept_max; ++i)
298 {
299 if (smtp_slots[i].host_address != NULL)
300 {
301 if (Ustrcmp(sender_host_address, smtp_slots[i].host_address) == 0)
302 host_accept_count++;
303 else
304 other_host_count++;
305
306 /* Testing all these strings is expensive - see if we can drop out
307 early, either by hitting the target, or finding there are not enough
308 connections left to make the target. */
309
310 if ((host_accept_count >= max_for_this_host) ||
311 ((smtp_accept_count - other_host_count) < max_for_this_host))
312 break;
313 }
314 }
315
316 if (host_accept_count >= max_for_this_host)
317 {
318 DEBUG(D_any) debug_printf("rejecting SMTP connection: too many from this "
319 "IP address: count=%d max=%d\n",
320 host_accept_count, max_for_this_host);
321 smtp_printf("421 Too many concurrent SMTP connections "
322 "from this IP address; please try again later.\r\n");
323 log_write(L_connection_reject,
324 LOG_MAIN, "Connection from %s refused: too many connections "
325 "from that IP address", whofrom);
326 goto ERROR_RETURN;
327 }
328 }
329
330 /* OK, the connection count checks have been passed. Before we can fork the
331 accepting process, we must first log the connection if requested. This logging
332 used to happen in the subprocess, but doing that means that the value of
333 smtp_accept_count can be out of step by the time it is logged. So we have to do
334 the logging here and accept the performance cost. Note that smtp_accept_count
335 hasn't yet been incremented to take account of this connection.
336
337 In order to minimize the cost (because this is going to happen for every
338 connection), do a preliminary selector test here. This saves ploughing through
339 the generalized logging code each time when the selector is false. If the
340 selector is set, check whether the host is on the list for logging. If not,
341 arrange to unset the selector in the subprocess. */
342
343 if ((log_write_selector & L_smtp_connection) != 0)
344 {
345 uschar *list = hosts_connection_nolog;
346 if (list != NULL && verify_check_host(&list) == OK)
347 use_log_write_selector &= ~L_smtp_connection;
348 else
349 log_write(L_smtp_connection, LOG_MAIN, "SMTP connection from %s "
350 "(TCP/IP connection count = %d)", whofrom, smtp_accept_count + 1);
351 }
352
353 /* Now we can fork the accepting process; do a lookup tidy, just in case any
354 expansion above did a lookup. */
355
356 search_tidyup();
357 pid = fork();
358
359 /* Handle the child process */
360
361 if (pid == 0)
362 {
363 int i;
364 int queue_only_reason = 0;
365 int old_pool = store_pool;
366 int save_debug_selector = debug_selector;
367 BOOL local_queue_only;
368 #ifdef SA_NOCLDWAIT
369 struct sigaction act;
370 #endif
371
372 smtp_accept_count++; /* So that it includes this process */
373
374 /* May have been modified for the subprocess */
375
376 log_write_selector = use_log_write_selector;
377
378 /* Get the local interface address into permanent store */
379
380 store_pool = POOL_PERM;
381 interface_address = string_copy(interface_address);
382 store_pool = old_pool;
383
384 /* Check for a tls-on-connect port */
385
386 if (host_is_tls_on_connect_port(interface_port)) tls_on_connect = TRUE;
387
388 /* Expand smtp_active_hostname if required. We do not do this any earlier,
389 because it may depend on the local interface address (indeed, that is most
390 likely what it depends on.) */
391
392 smtp_active_hostname = primary_hostname;
393 if (raw_active_hostname != NULL)
394 {
395 uschar *nah = expand_string(raw_active_hostname);
396 if (nah == NULL)
397 {
398 if (!expand_string_forcedfail)
399 {
400 log_write(0, LOG_MAIN|LOG_PANIC, "failed to expand \"%s\" "
401 "(smtp_active_hostname): %s", raw_active_hostname,
402 expand_string_message);
403 smtp_printf("421 Local configuration error; "
404 "please try again later.\r\n");
405 mac_smtp_fflush();
406 search_tidyup();
407 _exit(EXIT_FAILURE);
408 }
409 }
410 else if (nah[0] != 0) smtp_active_hostname = nah;
411 }
412
413 /* Initialize the queueing flags */
414
415 queue_check_only();
416 local_queue_only = queue_only;
417
418 /* Close the listening sockets, and set the SIGCHLD handler to SIG_IGN.
419 We also attempt to set things up so that children are automatically reaped,
420 but just in case this isn't available, there's a paranoid waitpid() in the
421 loop too (except for systems where we are sure it isn't needed). See the more
422 extensive comment before the reception loop in exim.c for a fuller
423 explanation of this logic. */
424
425 for (i = 0; i < listen_socket_count; i++) (void)close(listen_sockets[i]);
426
427 #ifdef SA_NOCLDWAIT
428 act.sa_handler = SIG_IGN;
429 sigemptyset(&(act.sa_mask));
430 act.sa_flags = SA_NOCLDWAIT;
431 sigaction(SIGCHLD, &act, NULL);
432 #else
433 signal(SIGCHLD, SIG_IGN);
434 #endif
435
436 /* Attempt to get an id from the sending machine via the RFC 1413
437 protocol. We do this in the sub-process in order not to hold up the
438 main process if there is any delay. Then set up the fullhost information
439 in case there is no HELO/EHLO.
440
441 If debugging is enabled only for the daemon, we must turn if off while
442 finding the id, but turn it on again afterwards so that information about the
443 incoming connection is output. */
444
445 if (debug_daemon) debug_selector = 0;
446 verify_get_ident(IDENT_PORT);
447 host_build_sender_fullhost();
448 debug_selector = save_debug_selector;
449
450 DEBUG(D_any)
451 debug_printf("Process %d is handling incoming connection from %s\n",
452 (int)getpid(), sender_fullhost);
453
454 /* Now disable debugging permanently if it's required only for the daemon
455 process. */
456
457 if (debug_daemon) debug_selector = 0;
458
459 /* If there are too many child processes for immediate delivery,
460 set the local_queue_only flag, which is initialized from the
461 configured value and may therefore already be TRUE. Leave logging
462 till later so it will have a message id attached. */
463
464 if (smtp_accept_queue > 0 && smtp_accept_count > smtp_accept_queue)
465 {
466 local_queue_only = TRUE;
467 queue_only_reason = 1;
468 }
469
470 /* Handle the start of the SMTP session, then loop, accepting incoming
471 messages from the SMTP connection. The end will come at the QUIT command,
472 when smtp_setup_msg() returns 0. A break in the connection causes the
473 process to die (see accept.c).
474
475 NOTE: We do *not* call smtp_log_no_mail() if smtp_start_session() fails,
476 because a log line has already been written for all its failure exists
477 (usually "connection refused: <reason>") and writing another one is
478 unnecessary clutter. */
479
480 if (!smtp_start_session())
481 {
482 mac_smtp_fflush();
483 search_tidyup();
484 _exit(EXIT_SUCCESS);
485 }
486
487 for (;;)
488 {
489 int rc;
490 message_id[0] = 0; /* Clear out any previous message_id */
491 reset_point = store_get(0); /* Save current store high water point */
492
493 DEBUG(D_any)
494 debug_printf("Process %d is ready for new message\n", (int)getpid());
495
496 /* Smtp_setup_msg() returns 0 on QUIT or if the call is from an
497 unacceptable host or if an ACL "drop" command was triggered, -1 on
498 connection lost, and +1 on validly reaching DATA. Receive_msg() almost
499 always returns TRUE when smtp_input is true; just retry if no message was
500 accepted (can happen for invalid message parameters). However, it can yield
501 FALSE if the connection was forcibly dropped by the DATA ACL. */
502
503 if ((rc = smtp_setup_msg()) > 0)
504 {
505 BOOL ok = receive_msg(FALSE);
506 search_tidyup(); /* Close cached databases */
507 if (!ok) /* Connection was dropped */
508 {
509 mac_smtp_fflush();
510 smtp_log_no_mail(); /* Log no mail if configured */
511 _exit(EXIT_SUCCESS);
512 }
513 if (message_id[0] == 0) continue; /* No message was accepted */
514 }
515 else
516 {
517 mac_smtp_fflush();
518 search_tidyup();
519 smtp_log_no_mail(); /* Log no mail if configured */
520 _exit((rc == 0)? EXIT_SUCCESS : EXIT_FAILURE);
521 }
522
523 /* Show the recipients when debugging */
524
525 DEBUG(D_receive)
526 {
527 int i;
528 if (sender_address != NULL)
529 debug_printf("Sender: %s\n", sender_address);
530 if (recipients_list != NULL)
531 {
532 debug_printf("Recipients:\n");
533 for (i = 0; i < recipients_count; i++)
534 debug_printf(" %s\n", recipients_list[i].address);
535 }
536 }
537
538 /* A message has been accepted. Clean up any previous delivery processes
539 that have completed and are defunct, on systems where they don't go away
540 by themselves (see comments when setting SIG_IGN above). On such systems
541 (if any) these delivery processes hang around after termination until
542 the next message is received. */
543
544 #ifndef SIG_IGN_WORKS
545 while (waitpid(-1, NULL, WNOHANG) > 0);
546 #endif
547
548 /* Reclaim up the store used in accepting this message */
549
550 store_reset(reset_point);
551
552 /* If queue_only is set or if there are too many incoming connections in
553 existence, local_queue_only will be TRUE. If it is not, check whether we
554 have received too many messages in this session for immediate delivery. If
555 not, and queue_only_load is set, check that the load average is below it.
556 Note that, once set, local_queue_only remains set for any subsequent
557 messages on the same SMTP connection. This is a deliberate choice; even
558 though the load average may fall, it doesn't seem right to deliver later
559 messages on the same call when not delivering earlier ones. */
560
561 if (!local_queue_only)
562 {
563 if (smtp_accept_queue_per_connection > 0 &&
564 receive_messagecount > smtp_accept_queue_per_connection)
565 {
566 local_queue_only = TRUE;
567 queue_only_reason = 2;
568 }
569 else if (queue_only_load >= 0)
570 {
571 local_queue_only = (load_average = os_getloadavg()) > queue_only_load;
572 if (local_queue_only) queue_only_reason = 3;
573 }
574 }
575
576 /* Log the queueing here, when it will get a message id attached, but
577 not if queue_only is set (case 0). */
578
579 if (local_queue_only) switch(queue_only_reason)
580 {
581 case 1:
582 log_write(L_delay_delivery,
583 LOG_MAIN, "no immediate delivery: too many connections "
584 "(%d, max %d)", smtp_accept_count, smtp_accept_queue);
585 break;
586
587 case 2:
588 log_write(L_delay_delivery,
589 LOG_MAIN, "no immediate delivery: more than %d messages "
590 "received in one connection", smtp_accept_queue_per_connection);
591 break;
592
593 case 3:
594 log_write(L_delay_delivery,
595 LOG_MAIN, "no immediate delivery: load average %.2f",
596 (double)load_average/1000.0);
597 break;
598 }
599
600 /* If a delivery attempt is required, spin off a new process to handle it.
601 If we are not root, we have to re-exec exim unless deliveries are being
602 done unprivileged. */
603
604 else if (!queue_only_policy && !deliver_freeze)
605 {
606 pid_t dpid;
607
608 /* Before forking, ensure that the C output buffer is flushed. Otherwise
609 anything that it in it will get duplicated, leading to duplicate copies
610 of the pending output. */
611
612 mac_smtp_fflush();
613
614 if ((dpid = fork()) == 0)
615 {
616 (void)fclose(smtp_in);
617 (void)fclose(smtp_out);
618
619 /* Don't ever molest the parent's SSL connection, but do clean up
620 the data structures if necessary. */
621
622 #ifdef SUPPORT_TLS
623 tls_close(FALSE);
624 #endif
625
626 /* Reset SIGHUP and SIGCHLD in the child in both cases. */
627
628 signal(SIGHUP, SIG_DFL);
629 signal(SIGCHLD, SIG_DFL);
630
631 if (geteuid() != root_uid && !deliver_drop_privilege)
632 {
633 signal(SIGALRM, SIG_DFL);
634 (void)child_exec_exim(CEE_EXEC_PANIC, FALSE, NULL, FALSE, 2, US"-Mc",
635 message_id);
636 /* Control does not return here. */
637 }
638
639 /* No need to re-exec; SIGALRM remains set to the default handler */
640
641 (void)deliver_message(message_id, FALSE, FALSE);
642 search_tidyup();
643 _exit(EXIT_SUCCESS);
644 }
645
646 if (dpid > 0)
647 {
648 DEBUG(D_any) debug_printf("forked delivery process %d\n", (int)dpid);
649 }
650 else
651 {
652 log_write(0, LOG_MAIN|LOG_PANIC, "daemon: delivery process fork "
653 "failed: %s", strerror(errno));
654 }
655 }
656 }
657 }
658
659
660 /* Carrying on in the parent daemon process... Can't do much if the fork
661 failed. Otherwise, keep count of the number of accepting processes and
662 remember the pid for ticking off when the child completes. */
663
664 if (pid < 0)
665 {
666 never_error(US"daemon: accept process fork failed", US"Fork failed", errno);
667 }
668 else
669 {
670 int i;
671 for (i = 0; i < smtp_accept_max; ++i)
672 {
673 if (smtp_slots[i].pid <= 0)
674 {
675 smtp_slots[i].pid = pid;
676 if (smtp_accept_max_per_host != NULL)
677 smtp_slots[i].host_address = string_copy_malloc(sender_host_address);
678 smtp_accept_count++;
679 break;
680 }
681 }
682 DEBUG(D_any) debug_printf("%d SMTP accept process%s running\n",
683 smtp_accept_count, (smtp_accept_count == 1)? "" : "es");
684 }
685
686 /* Get here via goto in error cases */
687
688 ERROR_RETURN:
689
690 /* Close the streams associated with the socket which will also close the
691 socket fds in this process. We can't do anything if fclose() fails, but
692 logging brings it to someone's attention. However, "connection reset by peer"
693 isn't really a problem, so skip that one. On Solaris, a dropped connection can
694 manifest itself as a broken pipe, so drop that one too. If the streams don't
695 exist, something went wrong while setting things up. Make sure the socket
696 descriptors are closed, in order to drop the connection. */
697
698 if (smtp_out != NULL)
699 {
700 if (fclose(smtp_out) != 0 && errno != ECONNRESET && errno != EPIPE)
701 log_write(0, LOG_MAIN|LOG_PANIC, "daemon: fclose(smtp_out) failed: %s",
702 strerror(errno));
703 smtp_out = NULL;
704 }
705 else (void)close(accept_socket);
706
707 if (smtp_in != NULL)
708 {
709 if (fclose(smtp_in) != 0 && errno != ECONNRESET && errno != EPIPE)
710 log_write(0, LOG_MAIN|LOG_PANIC, "daemon: fclose(smtp_in) failed: %s",
711 strerror(errno));
712 smtp_in = NULL;
713 }
714 else (void)close(dup_accept_socket);
715
716 /* Release any store used in this process, including the store used for holding
717 the incoming host address and an expanded active_hostname. */
718
719 store_reset(reset_point);
720 sender_host_address = NULL;
721 }
722
723
724
725
726 /*************************************************
727 * Check wildcard listen special cases *
728 *************************************************/
729
730 /* This function is used when binding and listening on lists of addresses and
731 ports. It tests for special cases of wildcard listening, when IPv4 and IPv6
732 sockets may interact in different ways in different operating systems. It is
733 passed an error number, the list of listening addresses, and the current
734 address. Two checks are available: for a previous wildcard IPv6 address, or for
735 a following wildcard IPv4 address, in both cases on the same port.
736
737 In practice, pairs of wildcard addresses should be adjacent in the address list
738 because they are sorted that way below.
739
740 Arguments:
741 eno the error number
742 addresses the list of addresses
743 ipa the current IP address
744 back if TRUE, check for previous wildcard IPv6 address
745 if FALSE, check for a following wildcard IPv4 address
746
747 Returns: TRUE or FALSE
748 */
749
750 static BOOL
751 check_special_case(int eno, ip_address_item *addresses, ip_address_item *ipa,
752 BOOL back)
753 {
754 ip_address_item *ipa2;
755
756 /* For the "back" case, if the failure was "address in use" for a wildcard IPv4
757 address, seek a previous IPv6 wildcard address on the same port. As it is
758 previous, it must have been successfully bound and be listening. Flag it as a
759 "6 including 4" listener. */
760
761 if (back)
762 {
763 if (eno != EADDRINUSE || ipa->address[0] != 0) return FALSE;
764 for (ipa2 = addresses; ipa2 != ipa; ipa2 = ipa2->next)
765 {
766 if (ipa2->address[1] == 0 && ipa2->port == ipa->port)
767 {
768 ipa2->v6_include_v4 = TRUE;
769 return TRUE;
770 }
771 }
772 }
773
774 /* For the "forward" case, if the current address is a wildcard IPv6 address,
775 we seek a following wildcard IPv4 address on the same port. */
776
777 else
778 {
779 if (ipa->address[0] != ':' || ipa->address[1] != 0) return FALSE;
780 for (ipa2 = ipa->next; ipa2 != NULL; ipa2 = ipa2->next)
781 if (ipa2->address[0] == 0 && ipa->port == ipa2->port) return TRUE;
782 }
783
784 return FALSE;
785 }
786
787
788
789
790 /*************************************************
791 * Handle terminating subprocesses *
792 *************************************************/
793
794 /* Handle the termination of child processes. Theoretically, this need be done
795 only when sigchld_seen is TRUE, but rumour has it that some systems lose
796 SIGCHLD signals at busy times, so to be on the safe side, this function is
797 called each time round. It shouldn't be too expensive.
798
799 Arguments: none
800 Returns: nothing
801 */
802
803 static void
804 handle_ending_processes(void)
805 {
806 int status;
807 pid_t pid;
808
809 while ((pid = waitpid(-1, &status, WNOHANG)) > 0)
810 {
811 int i;
812 DEBUG(D_any) debug_printf("child %d ended: status=0x%x\n", (int)pid,
813 status);
814
815 /* If it's a listening daemon for which we are keeping track of individual
816 subprocesses, deal with an accepting process that has terminated. */
817
818 if (smtp_slots != NULL)
819 {
820 for (i = 0; i < smtp_accept_max; i++)
821 {
822 if (smtp_slots[i].pid == pid)
823 {
824 if (smtp_slots[i].host_address != NULL)
825 store_free(smtp_slots[i].host_address);
826 smtp_slots[i] = empty_smtp_slot;
827 if (--smtp_accept_count < 0) smtp_accept_count = 0;
828 DEBUG(D_any) debug_printf("%d SMTP accept process%s now running\n",
829 smtp_accept_count, (smtp_accept_count == 1)? "" : "es");
830 break;
831 }
832 }
833 if (i < smtp_accept_max) continue; /* Found an accepting process */
834 }
835
836 /* If it wasn't an accepting process, see if it was a queue-runner
837 process that we are tracking. */
838
839 if (queue_pid_slots != NULL)
840 {
841 for (i = 0; i < queue_run_max; i++)
842 {
843 if (queue_pid_slots[i] == pid)
844 {
845 queue_pid_slots[i] = 0;
846 if (--queue_run_count < 0) queue_run_count = 0;
847 DEBUG(D_any) debug_printf("%d queue-runner process%s now running\n",
848 queue_run_count, (queue_run_count == 1)? "" : "es");
849 break;
850 }
851 }
852 }
853 }
854 }
855
856
857
858 /*************************************************
859 * Exim Daemon Mainline *
860 *************************************************/
861
862 /* The daemon can do two jobs, either of which is optional:
863
864 (1) Listens for incoming SMTP calls and spawns off a sub-process to handle
865 each one. This is requested by the -bd option, with -oX specifying the SMTP
866 port on which to listen (for testing).
867
868 (2) Spawns a queue-running process every so often. This is controlled by the
869 -q option with a an interval time. (If no time is given, a single queue run
870 is done from the main function, and control doesn't get here.)
871
872 Root privilege is required in order to attach to port 25. Some systems require
873 it when calling socket() rather than bind(). To cope with all cases, we run as
874 root for both socket() and bind(). Some systems also require root in order to
875 write to the pid file directory. This function must therefore be called as root
876 if it is to work properly in all circumstances. Once the socket is bound and
877 the pid file written, root privilege is given up if there is an exim uid.
878
879 There are no arguments to this function, and it never returns. */
880
881 void
882 daemon_go(void)
883 {
884 int *listen_sockets = NULL;
885 int listen_socket_count = 0;
886 ip_address_item *addresses = NULL;
887
888 /* If any debugging options are set, turn on the D_pid bit so that all
889 debugging lines get the pid added. */
890
891 DEBUG(D_any|D_v) debug_selector |= D_pid;
892
893
894 /* Do the preparation for setting up a listener on one or more interfaces, and
895 possible on various ports. This is controlled by the combination of
896 local_interfaces (which can set IP addresses and ports) and daemon_smtp_port
897 (which is a list of default ports to use for those items in local_interfaces
898 that do not specify a port). The -oX command line option can be used to
899 override one or both of these options.
900
901 If local_interfaces is not set, the default is to listen on all interfaces.
902 When it is set, it can include "all IPvx interfaces" as an item. This is useful
903 when different ports are in use.
904
905 It turns out that listening on all interfaces is messy in an IPv6 world,
906 because several different implementation approaches have been taken. This code
907 is now supposed to work with all of them. The point of difference is whether an
908 IPv6 socket that is listening on all interfaces will receive incoming IPv4
909 calls or not. We also have to cope with the case when IPv6 libraries exist, but
910 there is no IPv6 support in the kernel.
911
912 . On Solaris, an IPv6 socket will accept IPv4 calls, and give them as mapped
913 addresses. However, if an IPv4 socket is also listening on all interfaces,
914 calls are directed to the appropriate socket.
915
916 . On (some versions of) Linux, an IPv6 socket will accept IPv4 calls, and
917 give them as mapped addresses, but an attempt also to listen on an IPv4
918 socket on all interfaces causes an error.
919
920 . On OpenBSD, an IPv6 socket will not accept IPv4 calls. You have to set up
921 two sockets if you want to accept both kinds of call.
922
923 . FreeBSD is like OpenBSD, but it has the IPV6_V6ONLY socket option, which
924 can be turned off, to make it behave like the versions of Linux described
925 above.
926
927 . I heard a report that the USAGI IPv6 stack for Linux has implemented
928 IPV6_V6ONLY.
929
930 So, what we do when IPv6 is supported is as follows:
931
932 (1) After it is set up, the list of interfaces is scanned for wildcard
933 addresses. If an IPv6 and an IPv4 wildcard are both found for the same
934 port, the list is re-arranged so that they are together, with the IPv6
935 wildcard first.
936
937 (2) If the creation of a wildcard IPv6 socket fails, we just log the error and
938 carry on if an IPv4 wildcard socket for the same port follows later in the
939 list. This allows Exim to carry on in the case when the kernel has no IPv6
940 support.
941
942 (3) Having created an IPv6 wildcard socket, we try to set IPV6_V6ONLY if that
943 option is defined. However, if setting fails, carry on regardless (but log
944 the incident).
945
946 (4) If binding or listening on an IPv6 wildcard socket fails, it is a serious
947 error.
948
949 (5) If binding or listening on an IPv4 wildcard socket fails with the error
950 EADDRINUSE, and a previous interface was an IPv6 wildcard for the same
951 port (which must have succeeded or we wouldn't have got this far), we
952 assume we are in the situation where just a single socket is permitted,
953 and ignore the error.
954
955 Phew!
956
957 The preparation code decodes options and sets up the relevant data. We do this
958 first, so that we can return non-zero if there are any syntax errors, and also
959 write to stderr. */
960
961 if (daemon_listen)
962 {
963 int *default_smtp_port;
964 int sep;
965 int pct = 0;
966 uschar *s;
967 uschar *list;
968 uschar *local_iface_source = US"local_interfaces";
969 ip_address_item *ipa;
970 ip_address_item **pipa;
971
972 /* If any option requiring a load average to be available during the
973 reception of a message is set, call os_getloadavg() while we are root
974 for those OS for which this is necessary the first time it is called (in
975 order to perform an "open" on the kernel memory file). */
976
977 #ifdef LOAD_AVG_NEEDS_ROOT
978 if (queue_only_load >= 0 || smtp_load_reserve >= 0 ||
979 (deliver_queue_load_max >= 0 && deliver_drop_privilege))
980 (void)os_getloadavg();
981 #endif
982
983 /* If -oX was used, disable the writing of a pid file unless -oP was
984 explicitly used to force it. Then scan the string given to -oX. Any items
985 that contain neither a dot nor a colon are used to override daemon_smtp_port.
986 Any other items are used to override local_interfaces. */
987
988 if (override_local_interfaces != NULL)
989 {
990 uschar *new_smtp_port = NULL;
991 uschar *new_local_interfaces = NULL;
992 int portsize = 0;
993 int portptr = 0;
994 int ifacesize = 0;
995 int ifaceptr = 0;
996
997 if (override_pid_file_path == NULL) write_pid = FALSE;
998
999 list = override_local_interfaces;
1000 sep = 0;
1001 while ((s = string_nextinlist(&list,&sep,big_buffer,big_buffer_size))
1002 != NULL)
1003 {
1004 uschar joinstr[4];
1005 uschar **ptr;
1006 int *sizeptr;
1007 int *ptrptr;
1008
1009 if (Ustrpbrk(s, ".:") == NULL)
1010 {
1011 ptr = &new_smtp_port;
1012 sizeptr = &portsize;
1013 ptrptr = &portptr;
1014 }
1015 else
1016 {
1017 ptr = &new_local_interfaces;
1018 sizeptr = &ifacesize;
1019 ptrptr = &ifaceptr;
1020 }
1021
1022 if (*ptr == NULL)
1023 {
1024 joinstr[0] = sep;
1025 joinstr[1] = ' ';
1026 *ptr = string_cat(*ptr, sizeptr, ptrptr, US"<", 1);
1027 }
1028
1029 *ptr = string_cat(*ptr, sizeptr, ptrptr, joinstr, 2);
1030 *ptr = string_cat(*ptr, sizeptr, ptrptr, s, Ustrlen(s));
1031 }
1032
1033 if (new_smtp_port != NULL)
1034 {
1035 new_smtp_port[portptr] = 0;
1036 daemon_smtp_port = new_smtp_port;
1037 DEBUG(D_any) debug_printf("daemon_smtp_port overridden by -oX:\n %s\n",
1038 daemon_smtp_port);
1039 }
1040
1041 if (new_local_interfaces != NULL)
1042 {
1043 new_local_interfaces[ifaceptr] = 0;
1044 local_interfaces = new_local_interfaces;
1045 local_iface_source = US"-oX data";
1046 DEBUG(D_any) debug_printf("local_interfaces overridden by -oX:\n %s\n",
1047 local_interfaces);
1048 }
1049 }
1050
1051 /* Create a list of default SMTP ports, to be used if local_interfaces
1052 contains entries without explict ports. First count the number of ports, then
1053 build a translated list in a vector. */
1054
1055 list = daemon_smtp_port;
1056 sep = 0;
1057 while ((s = string_nextinlist(&list,&sep,big_buffer,big_buffer_size)) != NULL)
1058 pct++;
1059 default_smtp_port = store_get((pct+1) * sizeof(int));
1060 list = daemon_smtp_port;
1061 sep = 0;
1062 for (pct = 0;
1063 (s = string_nextinlist(&list,&sep,big_buffer,big_buffer_size)) != NULL;
1064 pct++)
1065 {
1066 if (isdigit(*s))
1067 {
1068 uschar *end;
1069 default_smtp_port[pct] = Ustrtol(s, &end, 0);
1070 if (end != s + Ustrlen(s))
1071 log_write(0, LOG_PANIC_DIE|LOG_CONFIG, "invalid SMTP port: %s", s);
1072 }
1073 else
1074 {
1075 struct servent *smtp_service = getservbyname(CS s, "tcp");
1076 if (smtp_service == NULL)
1077 log_write(0, LOG_PANIC_DIE|LOG_CONFIG, "TCP port \"%s\" not found", s);
1078 default_smtp_port[pct] = ntohs(smtp_service->s_port);
1079 }
1080 }
1081 default_smtp_port[pct] = 0;
1082
1083 /* Create the list of local interfaces, possibly with ports included. This
1084 list may contain references to 0.0.0.0 and ::0 as wildcards. These special
1085 values are converted below. */
1086
1087 addresses = host_build_ifacelist(local_interfaces, local_iface_source);
1088
1089 /* In the list of IP addresses, convert 0.0.0.0 into an empty string, and ::0
1090 into the string ":". We use these to recognize wildcards in IPv4 and IPv6. In
1091 fact, many IP stacks recognize 0.0.0.0 and ::0 and handle them as wildcards
1092 anyway, but we need to know which are the wildcard addresses, and the shorter
1093 strings are neater.
1094
1095 In the same scan, fill in missing port numbers from the default list. When
1096 there is more than one item in the list, extra items are created. */
1097
1098 for (ipa = addresses; ipa != NULL; ipa = ipa->next)
1099 {
1100 int i;
1101
1102 if (Ustrcmp(ipa->address, "0.0.0.0") == 0) ipa->address[0] = 0;
1103 else if (Ustrcmp(ipa->address, "::0") == 0)
1104 {
1105 ipa->address[0] = ':';
1106 ipa->address[1] = 0;
1107 }
1108
1109 if (ipa->port > 0) continue;
1110
1111 if (daemon_smtp_port[0] <= 0)
1112 log_write(0, LOG_MAIN|LOG_PANIC_DIE, "no port specified for interface "
1113 "%s and daemon_smtp_port is unset; cannot start daemon",
1114 (ipa->address[0] == 0)? US"\"all IPv4\"" :
1115 (ipa->address[1] == 0)? US"\"all IPv6\"" : ipa->address);
1116 ipa->port = default_smtp_port[0];
1117 for (i = 1; default_smtp_port[i] > 0; i++)
1118 {
1119 ip_address_item *new = store_get(sizeof(ip_address_item));
1120 memcpy(new->address, ipa->address, Ustrlen(ipa->address) + 1);
1121 new->port = default_smtp_port[i];
1122 new->next = ipa->next;
1123 ipa->next = new;
1124 ipa = new;
1125 }
1126 }
1127
1128 /* Scan the list of addresses for wildcards. If we find an IPv4 and an IPv6
1129 wildcard for the same port, ensure that (a) they are together and (b) the
1130 IPv6 address comes first. This makes handling the messy features easier, and
1131 also simplifies the construction of the "daemon started" log line. */
1132
1133 pipa = &addresses;
1134 for (ipa = addresses; ipa != NULL; pipa = &(ipa->next), ipa = ipa->next)
1135 {
1136 ip_address_item *ipa2;
1137
1138 /* Handle an IPv4 wildcard */
1139
1140 if (ipa->address[0] == 0)
1141 {
1142 for (ipa2 = ipa; ipa2->next != NULL; ipa2 = ipa2->next)
1143 {
1144 ip_address_item *ipa3 = ipa2->next;
1145 if (ipa3->address[0] == ':' &&
1146 ipa3->address[1] == 0 &&
1147 ipa3->port == ipa->port)
1148 {
1149 ipa2->next = ipa3->next;
1150 ipa3->next = ipa;
1151 *pipa = ipa3;
1152 break;
1153 }
1154 }
1155 }
1156
1157 /* Handle an IPv6 wildcard. */
1158
1159 else if (ipa->address[0] == ':' && ipa->address[1] == 0)
1160 {
1161 for (ipa2 = ipa; ipa2->next != NULL; ipa2 = ipa2->next)
1162 {
1163 ip_address_item *ipa3 = ipa2->next;
1164 if (ipa3->address[0] == 0 && ipa3->port == ipa->port)
1165 {
1166 ipa2->next = ipa3->next;
1167 ipa3->next = ipa->next;
1168 ipa->next = ipa3;
1169 ipa = ipa3;
1170 break;
1171 }
1172 }
1173 }
1174 }
1175
1176 /* Get a vector to remember all the sockets in */
1177
1178 for (ipa = addresses; ipa != NULL; ipa = ipa->next)
1179 listen_socket_count++;
1180 listen_sockets = store_get(sizeof(int *) * listen_socket_count);
1181
1182 /* Do a sanity check on the max connects value just to save us from getting
1183 a huge amount of store. */
1184
1185 if (smtp_accept_max > 4095) smtp_accept_max = 4096;
1186
1187 /* There's no point setting smtp_accept_queue unless it is less than the max
1188 connects limit. The configuration reader ensures that the max is set if the
1189 queue-only option is set. */
1190
1191 if (smtp_accept_queue > smtp_accept_max) smtp_accept_queue = 0;
1192
1193 /* Get somewhere to keep the list of SMTP accepting pids if we are keeping
1194 track of them for total number and queue/host limits. */
1195
1196 if (smtp_accept_max > 0)
1197 {
1198 int i;
1199 smtp_slots = store_get(smtp_accept_max * sizeof(smtp_slot));
1200 for (i = 0; i < smtp_accept_max; i++) smtp_slots[i] = empty_smtp_slot;
1201 }
1202 }
1203
1204 /* The variable background_daemon is always false when debugging, but
1205 can also be forced false in order to keep a non-debugging daemon in the
1206 foreground. If background_daemon is true, close all open file descriptors that
1207 we know about, but then re-open stdin, stdout, and stderr to /dev/null.
1208
1209 This is protection against any called functions (in libraries, or in
1210 Perl, or whatever) that think they can write to stderr (or stdout). Before this
1211 was added, it was quite likely that an SMTP connection would use one of these
1212 file descriptors, in which case writing random stuff to it caused chaos.
1213
1214 Then disconnect from the controlling terminal, Most modern Unixes seem to have
1215 setsid() for getting rid of the controlling terminal. For any OS that doesn't,
1216 setsid() can be #defined as a no-op, or as something else. */
1217
1218 if (background_daemon)
1219 {
1220 log_close_all(); /* Just in case anything was logged earlier */
1221 search_tidyup(); /* Just in case any were used in reading the config. */
1222 (void)close(0); /* Get rid of stdin/stdout/stderr */
1223 (void)close(1);
1224 (void)close(2);
1225 exim_nullstd(); /* Connect stdin/stdout/stderr to /dev/null */
1226 log_stderr = NULL; /* So no attempt to copy paniclog output */
1227
1228 /* If the parent process of this one has pid == 1, we are re-initializing the
1229 daemon as the result of a SIGHUP. In this case, there is no need to do
1230 anything, because the controlling terminal has long gone. Otherwise, fork, in
1231 case current process is a process group leader (see 'man setsid' for an
1232 explanation) before calling setsid(). */
1233
1234 if (getppid() != 1)
1235 {
1236 pid_t pid = fork();
1237 if (pid < 0) log_write(0, LOG_MAIN|LOG_PANIC_DIE,
1238 "fork() failed when starting daemon: %s", strerror(errno));
1239 if (pid > 0) exit(EXIT_SUCCESS); /* in parent process, just exit */
1240 (void)setsid(); /* release controlling terminal */
1241 }
1242 }
1243
1244 /* We are now in the disconnected, daemon process (unless debugging). Set up
1245 the listening sockets if required. */
1246
1247 if (daemon_listen)
1248 {
1249 int sk;
1250 int on = 1;
1251 ip_address_item *ipa;
1252
1253 /* For each IP address, create a socket, bind it to the appropriate port, and
1254 start listening. See comments above about IPv6 sockets that may or may not
1255 accept IPv4 calls when listening on all interfaces. We also have to cope with
1256 the case of a system with IPv6 libraries, but no IPv6 support in the kernel.
1257 listening, provided a wildcard IPv4 socket for the same port follows. */
1258
1259 for (ipa = addresses, sk = 0; sk < listen_socket_count; ipa = ipa->next, sk++)
1260 {
1261 BOOL wildcard;
1262 ip_address_item *ipa2;
1263 int af;
1264
1265 if (Ustrchr(ipa->address, ':') != NULL)
1266 {
1267 af = AF_INET6;
1268 wildcard = ipa->address[1] == 0;
1269 }
1270 else
1271 {
1272 af = AF_INET;
1273 wildcard = ipa->address[0] == 0;
1274 }
1275
1276 listen_sockets[sk] = ip_socket(SOCK_STREAM, af);
1277 if (listen_sockets[sk] < 0)
1278 {
1279 if (check_special_case(0, addresses, ipa, FALSE))
1280 {
1281 log_write(0, LOG_MAIN, "Failed to create IPv6 socket for wildcard "
1282 "listening (%s): will use IPv4", strerror(errno));
1283 goto SKIP_SOCKET;
1284 }
1285 log_write(0, LOG_PANIC_DIE, "IPv%c socket creation failed: %s",
1286 (af == AF_INET6)? '6' : '4', strerror(errno));
1287 }
1288
1289 /* If this is an IPv6 wildcard socket, set IPV6_V6ONLY if that option is
1290 available. Just log failure (can get protocol not available, just like
1291 socket creation can). */
1292
1293 #ifdef IPV6_V6ONLY
1294 if (af == AF_INET6 && wildcard &&
1295 setsockopt(listen_sockets[sk], IPPROTO_IPV6, IPV6_V6ONLY, (char *)(&on),
1296 sizeof(on)) < 0)
1297 log_write(0, LOG_MAIN, "Setting IPV6_V6ONLY on daemon's IPv6 wildcard "
1298 "socket failed (%s): carrying on without it", strerror(errno));
1299 #endif /* IPV6_V6ONLY */
1300
1301 /* Set SO_REUSEADDR so that the daemon can be restarted while a connection
1302 is being handled. Without this, a connection will prevent reuse of the
1303 smtp port for listening. */
1304
1305 if (setsockopt(listen_sockets[sk], SOL_SOCKET, SO_REUSEADDR,
1306 (uschar *)(&on), sizeof(on)) < 0)
1307 log_write(0, LOG_MAIN|LOG_PANIC_DIE, "setting SO_REUSEADDR on socket "
1308 "failed when starting daemon: %s", strerror(errno));
1309
1310 /* Set TCP_NODELAY; Exim does its own buffering. There is a switch to
1311 disable this because it breaks some broken clients. */
1312
1313 if (tcp_nodelay) setsockopt(listen_sockets[sk], IPPROTO_TCP, TCP_NODELAY,
1314 (uschar *)(&on), sizeof(on));
1315
1316 /* Now bind the socket to the required port; if Exim is being restarted
1317 it may not always be possible to bind immediately, even with SO_REUSEADDR
1318 set, so try 10 times, waiting between each try. After 10 failures, we give
1319 up. In an IPv6 environment, if bind () fails with the error EADDRINUSE and
1320 we are doing wildcard IPv4 listening and there was a previous IPv6 wildcard
1321 address for the same port, ignore the error on the grounds that we must be
1322 in a system where the IPv6 socket accepts both kinds of call. This is
1323 necessary for (some release of) USAGI Linux; other IP stacks fail at the
1324 listen() stage instead. */
1325
1326 for(;;)
1327 {
1328 uschar *msg, *addr;
1329 if (ip_bind(listen_sockets[sk], af, ipa->address, ipa->port) >= 0) break;
1330 if (check_special_case(errno, addresses, ipa, TRUE))
1331 {
1332 DEBUG(D_any) debug_printf("wildcard IPv4 bind() failed after IPv6 "
1333 "listen() success; EADDRINUSE ignored\n");
1334 (void)close(listen_sockets[sk]);
1335 goto SKIP_SOCKET;
1336 }
1337 msg = US strerror(errno);
1338 addr = wildcard? ((af == AF_INET6)? US"(any IPv6)" : US"(any IPv4)") :
1339 ipa->address;
1340 if (daemon_startup_retries <= 0)
1341 log_write(0, LOG_MAIN|LOG_PANIC_DIE,
1342 "socket bind() to port %d for address %s failed: %s: "
1343 "daemon abandoned", ipa->port, addr, msg);
1344 log_write(0, LOG_MAIN, "socket bind() to port %d for address %s "
1345 "failed: %s: waiting %s before trying again (%d more %s)",
1346 ipa->port, addr, msg, readconf_printtime(daemon_startup_sleep),
1347 daemon_startup_retries, (daemon_startup_retries > 1)? "tries" : "try");
1348 daemon_startup_retries--;
1349 sleep(daemon_startup_sleep);
1350 }
1351
1352 DEBUG(D_any)
1353 {
1354 if (wildcard)
1355 debug_printf("listening on all interfaces (IPv%c) port %d\n",
1356 (af == AF_INET6)? '6' : '4', ipa->port);
1357 else
1358 debug_printf("listening on %s port %d\n", ipa->address, ipa->port);
1359 }
1360
1361 /* Start listening on the bound socket, establishing the maximum backlog of
1362 connections that is allowed. On success, continue to the next address. */
1363
1364 if (listen(listen_sockets[sk], smtp_connect_backlog) >= 0) continue;
1365
1366 /* Listening has failed. In an IPv6 environment, as for bind(), if listen()
1367 fails with the error EADDRINUSE and we are doing IPv4 wildcard listening
1368 and there was a previous successful IPv6 wildcard listen on the same port,
1369 we want to ignore the error on the grounds that we must be in a system
1370 where the IPv6 socket accepts both kinds of call. */
1371
1372 if (!check_special_case(errno, addresses, ipa, TRUE))
1373 log_write(0, LOG_PANIC_DIE, "listen() failed on interface %s: %s",
1374 wildcard? ((af == AF_INET6)? US"(any IPv6)" : US"(any IPv4)") :
1375 ipa->address,
1376 strerror(errno));
1377
1378 DEBUG(D_any) debug_printf("wildcard IPv4 listen() failed after IPv6 "
1379 "listen() success; EADDRINUSE ignored\n");
1380 (void)close(listen_sockets[sk]);
1381
1382 /* Come here if there has been a problem with the socket which we
1383 are going to ignore. We remove the address from the chain, and back up the
1384 counts. */
1385
1386 SKIP_SOCKET:
1387 sk--; /* Back up the count */
1388 listen_socket_count--; /* Reduce the total */
1389 if (ipa == addresses) addresses = ipa->next; else
1390 {
1391 for (ipa2 = addresses; ipa2->next != ipa; ipa2 = ipa2->next);
1392 ipa2->next = ipa->next;
1393 ipa = ipa2;
1394 }
1395 } /* End of bind/listen loop for each address */
1396 } /* End of setup for listening */
1397
1398
1399 /* If we are not listening, we want to write a pid file only if -oP was
1400 explicitly given. */
1401
1402 else if (override_pid_file_path == NULL) write_pid = FALSE;
1403
1404 /* Write the pid to a known file for assistance in identification, if required.
1405 We do this before giving up root privilege, because on some systems it is
1406 necessary to be root in order to write into the pid file directory. There's
1407 nothing to stop multiple daemons running, as long as no more than one listens
1408 on a given TCP/IP port on the same interface(s). However, in these
1409 circumstances it gets far too complicated to mess with pid file names
1410 automatically. Consequently, Exim 4 writes a pid file only
1411
1412 (a) When running in the test harness, or
1413 (b) When -bd is used and -oX is not used, or
1414 (c) When -oP is used to supply a path.
1415
1416 The variable daemon_write_pid is used to control this. */
1417
1418 if (running_in_test_harness || write_pid)
1419 {
1420 FILE *f;
1421
1422 if (override_pid_file_path != NULL)
1423 pid_file_path = override_pid_file_path;
1424
1425 if (pid_file_path[0] == 0)
1426 pid_file_path = string_sprintf("%s/exim-daemon.pid", spool_directory);
1427
1428 f = modefopen(pid_file_path, "wb", 0644);
1429 if (f != NULL)
1430 {
1431 (void)fprintf(f, "%d\n", (int)getpid());
1432 (void)fclose(f);
1433 DEBUG(D_any) debug_printf("pid written to %s\n", pid_file_path);
1434 }
1435 else
1436 {
1437 DEBUG(D_any)
1438 debug_printf("%s\n", string_open_failed(errno, "pid file %s",
1439 pid_file_path));
1440 }
1441 }
1442
1443 /* Set up the handler for SIGHUP, which causes a restart of the daemon. */
1444
1445 sighup_seen = FALSE;
1446 signal(SIGHUP, sighup_handler);
1447
1448 /* Give up root privilege at this point (assuming that exim_uid and exim_gid
1449 are not root). The third argument controls the running of initgroups().
1450 Normally we do this, in order to set up the groups for the Exim user. However,
1451 if we are not root at this time - some odd installations run that way - we
1452 cannot do this. */
1453
1454 exim_setugid(exim_uid, exim_gid, geteuid()==root_uid, US"running as a daemon");
1455
1456 /* Get somewhere to keep the list of queue-runner pids if we are keeping track
1457 of them (and also if we are doing queue runs). */
1458
1459 if (queue_interval > 0 && queue_run_max > 0)
1460 {
1461 int i;
1462 queue_pid_slots = store_get(queue_run_max * sizeof(pid_t));
1463 for (i = 0; i < queue_run_max; i++) queue_pid_slots[i] = 0;
1464 }
1465
1466 /* Set up the handler for termination of child processes. */
1467
1468 sigchld_seen = FALSE;
1469 os_non_restarting_signal(SIGCHLD, main_sigchld_handler);
1470
1471 /* If we are to run the queue periodically, pretend the alarm has just gone
1472 off. This will cause the first queue-runner to get kicked off straight away. */
1473
1474 sigalrm_seen = (queue_interval > 0);
1475
1476 /* Log the start up of a daemon - at least one of listening or queue running
1477 must be set up. */
1478
1479 if (daemon_listen)
1480 {
1481 int i, j;
1482 int smtp_ports = 0;
1483 int smtps_ports = 0;
1484 ip_address_item *ipa;
1485 uschar *p = big_buffer;
1486 uschar *qinfo = (queue_interval > 0)?
1487 string_sprintf("-q%s", readconf_printtime(queue_interval))
1488 :
1489 US"no queue runs";
1490
1491 /* Build a list of listening addresses in big_buffer, but limit it to 10
1492 items. The style is for backwards compatibility.
1493
1494 It is now possible to have some ports listening for SMTPS (the old,
1495 deprecated protocol that starts TLS without using STARTTLS), and others
1496 listening for standard SMTP. Keep their listings separate. */
1497
1498 for (j = 0; j < 2; j++)
1499 {
1500 for (i = 0, ipa = addresses; i < 10 && ipa != NULL; i++, ipa = ipa->next)
1501 {
1502 /* First time round, look for SMTP ports; second time round, look for
1503 SMTPS ports. For the first one of each, insert leading text. */
1504
1505 if (host_is_tls_on_connect_port(ipa->port) == (j > 0))
1506 {
1507 if (j == 0)
1508 {
1509 if (smtp_ports++ == 0)
1510 {
1511 memcpy(p, "SMTP on", 8);
1512 p += 7;
1513 }
1514 }
1515 else
1516 {
1517 if (smtps_ports++ == 0)
1518 {
1519 (void)sprintf(CS p, "%sSMTPS on",
1520 (smtp_ports == 0)? "":" and for ");
1521 while (*p != 0) p++;
1522 }
1523 }
1524
1525 /* Now the information about the port (and sometimes interface) */
1526
1527 if (ipa->address[0] == ':' && ipa->address[1] == 0)
1528 {
1529 if (ipa->next != NULL && ipa->next->address[0] == 0 &&
1530 ipa->next->port == ipa->port)
1531 {
1532 (void)sprintf(CS p, " port %d (IPv6 and IPv4)", ipa->port);
1533 ipa = ipa->next;
1534 }
1535 else if (ipa->v6_include_v4)
1536 (void)sprintf(CS p, " port %d (IPv6 with IPv4)", ipa->port);
1537 else
1538 (void)sprintf(CS p, " port %d (IPv6)", ipa->port);
1539 }
1540 else if (ipa->address[0] == 0)
1541 (void)sprintf(CS p, " port %d (IPv4)", ipa->port);
1542 else
1543 (void)sprintf(CS p, " [%s]:%d", ipa->address, ipa->port);
1544 while (*p != 0) p++;
1545 }
1546 }
1547
1548 if (ipa != NULL)
1549 {
1550 memcpy(p, " ...", 5);
1551 p += 4;
1552 }
1553 }
1554
1555 log_write(0, LOG_MAIN,
1556 "exim %s daemon started: pid=%d, %s, listening for %s",
1557 version_string, getpid(), qinfo, big_buffer);
1558 set_process_info("daemon: %s, listening for %s", qinfo, big_buffer);
1559 }
1560
1561 else
1562 {
1563 log_write(0, LOG_MAIN,
1564 "exim %s daemon started: pid=%d, -q%s, not listening for SMTP",
1565 version_string, getpid(), readconf_printtime(queue_interval));
1566 set_process_info("daemon: -q%s, not listening",
1567 readconf_printtime(queue_interval));
1568 }
1569
1570
1571 /* Close the log so it can be renamed and moved. In the few cases below where
1572 this long-running process writes to the log (always exceptional conditions), it
1573 closes the log afterwards, for the same reason. */
1574
1575 log_close_all();
1576
1577 DEBUG(D_any) debug_print_ids(US"daemon running with");
1578
1579 /* Any messages accepted via this route are going to be SMTP. */
1580
1581 smtp_input = TRUE;
1582
1583 /* Enter the never-ending loop... */
1584
1585 for (;;)
1586 {
1587 #if HAVE_IPV6
1588 struct sockaddr_in6 accepted;
1589 #else
1590 struct sockaddr_in accepted;
1591 #endif
1592
1593 EXIM_SOCKLEN_T len = sizeof(accepted);
1594 pid_t pid;
1595
1596 /* This code is placed first in the loop, so that it gets obeyed at the
1597 start, before the first wait. This causes the first queue-runner to be
1598 started immediately. */
1599
1600 if (sigalrm_seen)
1601 {
1602 DEBUG(D_any) debug_printf("SIGALRM received\n");
1603
1604 /* Do a full queue run in a child process, if required, unless we already
1605 have enough queue runners on the go. If we are not running as root, a
1606 re-exec is required. */
1607
1608 if (queue_interval > 0 &&
1609 (queue_run_max <= 0 || queue_run_count < queue_run_max))
1610 {
1611 if ((pid = fork()) == 0)
1612 {
1613 int sk;
1614
1615 DEBUG(D_any) debug_printf("Starting queue-runner: pid %d\n",
1616 (int)getpid());
1617
1618 /* Disable debugging if it's required only for the daemon process. We
1619 leave the above message, because it ties up with the "child ended"
1620 debugging messages. */
1621
1622 if (debug_daemon) debug_selector = 0;
1623
1624 /* Close any open listening sockets in the child */
1625
1626 for (sk = 0; sk < listen_socket_count; sk++)
1627 (void)close(listen_sockets[sk]);
1628
1629 /* Reset SIGHUP and SIGCHLD in the child in both cases. */
1630
1631 signal(SIGHUP, SIG_DFL);
1632 signal(SIGCHLD, SIG_DFL);
1633
1634 /* Re-exec if privilege has been given up, unless deliver_drop_
1635 privilege is set. Reset SIGALRM before exec(). */
1636
1637 if (geteuid() != root_uid && !deliver_drop_privilege)
1638 {
1639 uschar opt[8];
1640 uschar *p = opt;
1641 uschar *extra[4];
1642 int extracount = 1;
1643
1644 signal(SIGALRM, SIG_DFL);
1645 *p++ = '-';
1646 *p++ = 'q';
1647 if (queue_2stage) *p++ = 'q';
1648 if (queue_run_first_delivery) *p++ = 'i';
1649 if (queue_run_force) *p++ = 'f';
1650 if (deliver_force_thaw) *p++ = 'f';
1651 if (queue_run_local) *p++ = 'l';
1652 *p = 0;
1653 extra[0] = opt;
1654
1655 /* If -R or -S were on the original command line, ensure they get
1656 passed on. */
1657
1658 if (deliver_selectstring != NULL)
1659 {
1660 extra[extracount++] = deliver_selectstring_regex? US"-Rr" : US"-R";
1661 extra[extracount++] = deliver_selectstring;
1662 }
1663
1664 if (deliver_selectstring_sender != NULL)
1665 {
1666 extra[extracount++] = deliver_selectstring_sender_regex?
1667 US"-Sr" : US"-S";
1668 extra[extracount++] = deliver_selectstring_sender;
1669 }
1670
1671 /* Overlay this process with a new execution. */
1672
1673 (void)child_exec_exim(CEE_EXEC_PANIC, FALSE, NULL, TRUE, extracount,
1674 extra[0], extra[1], extra[2], extra[3], extra[4]);
1675
1676 /* Control never returns here. */
1677 }
1678
1679 /* No need to re-exec; SIGALRM remains set to the default handler */
1680
1681 queue_run(NULL, NULL, FALSE);
1682 _exit(EXIT_SUCCESS);
1683 }
1684
1685 if (pid < 0)
1686 {
1687 log_write(0, LOG_MAIN|LOG_PANIC, "daemon: fork of queue-runner "
1688 "process failed: %s", strerror(errno));
1689 log_close_all();
1690 }
1691 else
1692 {
1693 int i;
1694 for (i = 0; i < queue_run_max; ++i)
1695 {
1696 if (queue_pid_slots[i] <= 0)
1697 {
1698 queue_pid_slots[i] = pid;
1699 queue_run_count++;
1700 break;
1701 }
1702 }
1703 DEBUG(D_any) debug_printf("%d queue-runner process%s running\n",
1704 queue_run_count, (queue_run_count == 1)? "" : "es");
1705 }
1706 }
1707
1708 /* Reset the alarm clock */
1709
1710 sigalrm_seen = FALSE;
1711 alarm(queue_interval);
1712 }
1713
1714
1715 /* Sleep till a connection happens if listening, and handle the connection if
1716 that is why we woke up. The FreeBSD operating system requires the use of
1717 select() before accept() because the latter function is not interrupted by
1718 a signal, and we want to wake up for SIGCHLD and SIGALRM signals. Some other
1719 OS do notice signals in accept() but it does no harm to have the select()
1720 in for all of them - and it won't then be a lurking problem for ports to
1721 new OS. In fact, the later addition of listening on specific interfaces only
1722 requires this way of working anyway. */
1723
1724 if (daemon_listen)
1725 {
1726 int sk, lcount, select_errno;
1727 int max_socket = 0;
1728 BOOL select_failed = FALSE;
1729 fd_set select_listen;
1730
1731 FD_ZERO(&select_listen);
1732 for (sk = 0; sk < listen_socket_count; sk++)
1733 {
1734 FD_SET(listen_sockets[sk], &select_listen);
1735 if (listen_sockets[sk] > max_socket) max_socket = listen_sockets[sk];
1736 }
1737
1738 DEBUG(D_any) debug_printf("Listening...\n");
1739
1740 /* In rare cases we may have had a SIGCHLD signal in the time between
1741 setting the handler (below) and getting back here. If so, pretend that the
1742 select() was interrupted so that we reap the child. This might still leave
1743 a small window when a SIGCHLD could get lost. However, since we use SIGCHLD
1744 only to do the reaping more quickly, it shouldn't result in anything other
1745 than a delay until something else causes a wake-up. */
1746
1747 if (sigchld_seen)
1748 {
1749 lcount = -1;
1750 errno = EINTR;
1751 }
1752 else
1753 {
1754 lcount = select(max_socket + 1, (SELECT_ARG2_TYPE *)&select_listen,
1755 NULL, NULL, NULL);
1756 }
1757
1758 if (lcount < 0)
1759 {
1760 select_failed = TRUE;
1761 lcount = 1;
1762 }
1763
1764 /* Clean up any subprocesses that may have terminated. We need to do this
1765 here so that smtp_accept_max_per_host works when a connection to that host
1766 has completed, and we are about to accept a new one. When this code was
1767 later in the sequence, a new connection could be rejected, even though an
1768 old one had just finished. Preserve the errno from any select() failure for
1769 the use of the common select/accept error processing below. */
1770
1771 select_errno = errno;
1772 handle_ending_processes();
1773 errno = select_errno;
1774
1775 /* Loop for all the sockets that are currently ready to go. If select
1776 actually failed, we have set the count to 1 and select_failed=TRUE, so as
1777 to use the common error code for select/accept below. */
1778
1779 while (lcount-- > 0)
1780 {
1781 int accept_socket = -1;
1782 if (!select_failed)
1783 {
1784 for (sk = 0; sk < listen_socket_count; sk++)
1785 {
1786 if (FD_ISSET(listen_sockets[sk], &select_listen))
1787 {
1788 accept_socket = accept(listen_sockets[sk],
1789 (struct sockaddr *)&accepted, &len);
1790 FD_CLR(listen_sockets[sk], &select_listen);
1791 break;
1792 }
1793 }
1794 }
1795
1796 /* If select or accept has failed and this was not caused by an
1797 interruption, log the incident and try again. With asymmetric TCP/IP
1798 routing errors such as "No route to network" have been seen here. Also
1799 "connection reset by peer" has been seen. These cannot be classed as
1800 disastrous errors, but they could fill up a lot of log. The code in smail
1801 crashes the daemon after 10 successive failures of accept, on the grounds
1802 that some OS fail continuously. Exim originally followed suit, but this
1803 appears to have caused problems. Now it just keeps going, but instead of
1804 logging each error, it batches them up when they are continuous. */
1805
1806 if (accept_socket < 0 && errno != EINTR)
1807 {
1808 if (accept_retry_count == 0)
1809 {
1810 accept_retry_errno = errno;
1811 accept_retry_select_failed = select_failed;
1812 }
1813 else
1814 {
1815 if (errno != accept_retry_errno ||
1816 select_failed != accept_retry_select_failed ||
1817 accept_retry_count >= 50)
1818 {
1819 log_write(0, LOG_MAIN | ((accept_retry_count >= 50)? LOG_PANIC : 0),
1820 "%d %s() failure%s: %s",
1821 accept_retry_count,
1822 accept_retry_select_failed? "select" : "accept",
1823 (accept_retry_count == 1)? "" : "s",
1824 strerror(accept_retry_errno));
1825 log_close_all();
1826 accept_retry_count = 0;
1827 accept_retry_errno = errno;
1828 accept_retry_select_failed = select_failed;
1829 }
1830 }
1831 accept_retry_count++;
1832 }
1833
1834 else
1835 {
1836 if (accept_retry_count > 0)
1837 {
1838 log_write(0, LOG_MAIN, "%d %s() failure%s: %s",
1839 accept_retry_count,
1840 accept_retry_select_failed? "select" : "accept",
1841 (accept_retry_count == 1)? "" : "s",
1842 strerror(accept_retry_errno));
1843 log_close_all();
1844 accept_retry_count = 0;
1845 }
1846 }
1847
1848 /* If select/accept succeeded, deal with the connection. */
1849
1850 if (accept_socket >= 0)
1851 handle_smtp_call(listen_sockets, listen_socket_count, accept_socket,
1852 (struct sockaddr *)&accepted);
1853 }
1854 }
1855
1856 /* If not listening, then just sleep for the queue interval. If we woke
1857 up early the last time for some other signal, it won't matter because
1858 the alarm signal will wake at the right time. This code originally used
1859 sleep() but it turns out that on the FreeBSD system, sleep() is not inter-
1860 rupted by signals, so it wasn't waking up for SIGALRM or SIGCHLD. Luckily
1861 select() can be used as an interruptible sleep() on all versions of Unix. */
1862
1863 else
1864 {
1865 struct timeval tv;
1866 tv.tv_sec = queue_interval;
1867 tv.tv_usec = 0;
1868 select(0, NULL, NULL, NULL, &tv);
1869 handle_ending_processes();
1870 }
1871
1872 /* Re-enable the SIGCHLD handler if it has been run. It can't do it
1873 for itself, because it isn't doing the waiting itself. */
1874
1875 if (sigchld_seen)
1876 {
1877 sigchld_seen = FALSE;
1878 os_non_restarting_signal(SIGCHLD, main_sigchld_handler);
1879 }
1880
1881 /* Handle being woken by SIGHUP. We know at this point that the result
1882 of accept() has been dealt with, so we can re-exec exim safely, first
1883 closing the listening sockets so that they can be reused. Cancel any pending
1884 alarm in case it is just about to go off, and set SIGHUP to be ignored so
1885 that another HUP in quick succession doesn't clobber the new daemon before it
1886 gets going. All log files get closed by the close-on-exec flag; however, if
1887 the exec fails, we need to close the logs. */
1888
1889 if (sighup_seen)
1890 {
1891 int sk;
1892 log_write(0, LOG_MAIN, "pid %d: SIGHUP received: re-exec daemon",
1893 getpid());
1894 for (sk = 0; sk < listen_socket_count; sk++)
1895 (void)close(listen_sockets[sk]);
1896 alarm(0);
1897 signal(SIGHUP, SIG_IGN);
1898 sighup_argv[0] = exim_path;
1899 exim_nullstd();
1900 execv(CS exim_path, (char *const *)sighup_argv);
1901 log_write(0, LOG_MAIN|LOG_PANIC_DIE, "pid %d: exec of %s failed: %s",
1902 getpid(), exim_path, strerror(errno));
1903 log_close_all();
1904 }
1905
1906 } /* End of main loop */
1907
1908 /* Control never reaches here */
1909 }
1910
1911 /* End of exim_daemon.c */
1912