Start
[exim.git] / src / src / daemon.c
1 /* $Cambridge: exim/src/src/daemon.c,v 1.1 2004/10/07 10:39:01 ph10 Exp $ */
2
3 /*************************************************
4 * Exim - an Internet mail transport agent *
5 *************************************************/
6
7 /* Copyright (c) University of Cambridge 1995 - 2004 */
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;
43 static smtp_slot *smtp_slots;
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 sigchld_seen = TRUE;
88 signal(SIGCHLD, SIG_DFL);
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 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. */
190
191 if (getsockname(accept_socket, (struct sockaddr *)(&interface_sockaddr),
192 &ifsize) < 0)
193 {
194 log_write(0, LOG_MAIN|LOG_PANIC, "getsockname() failed: %s",
195 strerror(errno));
196 smtp_printf("421 Local problem: getsockname() failed; please try again later\r\n");
197 goto ERROR_RETURN;
198 }
199
200 interface_address = host_ntoa(-1, &interface_sockaddr, NULL, &interface_port);
201 DEBUG(D_interface) debug_printf("interface address=%s port=%d\n",
202 interface_address, interface_port);
203
204 /* Build a string identifying the remote host and, if requested, the port and
205 the local interface data. This is for logging; at the end of this function the
206 memory is reclaimed. */
207
208 whofrom = string_append(whofrom, &wfsize, &wfptr, 3, "[", sender_host_address, "]");
209
210 if ((log_extra_selector & LX_incoming_port) != 0)
211 whofrom = string_append(whofrom, &wfsize, &wfptr, 2, ":", string_sprintf("%d",
212 sender_host_port));
213
214 if ((log_extra_selector & LX_incoming_interface) != 0)
215 whofrom = string_append(whofrom, &wfsize, &wfptr, 4, " I=[",
216 interface_address, "]:", string_sprintf("%d", interface_port));
217
218 whofrom[wfptr] = 0; /* Terminate the newly-built string */
219
220 /* Check maximum number of connections. We do not check for reserved
221 connections or unacceptable hosts here. That is done in the subprocess because
222 it might take some time. */
223
224 if (smtp_accept_max > 0 && smtp_accept_count >= smtp_accept_max)
225 {
226 DEBUG(D_any) debug_printf("rejecting SMTP connection: count=%d max=%d\n",
227 smtp_accept_count, smtp_accept_max);
228 smtp_printf("421 Too many concurrent SMTP connections; "
229 "please try again later.\r\n");
230 log_write(L_connection_reject,
231 LOG_MAIN, "Connection from %s refused: too many connections",
232 whofrom);
233 goto ERROR_RETURN;
234 }
235
236 /* If a load limit above which only reserved hosts are acceptable is defined,
237 get the load average here, and if there are in fact no reserved hosts, do
238 the test right away (saves a fork). If there are hosts, do the check in the
239 subprocess because it might take time. */
240
241 if (smtp_load_reserve >= 0)
242 {
243 load_average = os_getloadavg();
244 if (smtp_reserve_hosts == NULL && load_average > smtp_load_reserve)
245 {
246 DEBUG(D_any) debug_printf("rejecting SMTP connection: load average = %.2f\n",
247 (double)load_average/1000.0);
248 smtp_printf("421 Too much load; please try again later.\r\n");
249 log_write(L_connection_reject,
250 LOG_MAIN, "Connection from %s refused: load average = %.2f",
251 whofrom, (double)load_average/1000.0);
252 goto ERROR_RETURN;
253 }
254 }
255
256 /* Check that one specific host (strictly, IP address) is not hogging
257 resources. This is done here to prevent a denial of service attack by someone
258 forcing you to fork lots of times before denying service. The value of
259 smtp_accept_max_per_host is a string which is expanded. This makes it possible
260 to provide host-specific limits according to $sender_host address, but because
261 this is in the daemon mainline, only fast expansions (such as inline address
262 checks) should be used. The documentation is full of warnings. */
263
264 if (smtp_accept_max_per_host != NULL)
265 {
266 uschar *expanded = expand_string(smtp_accept_max_per_host);
267 if (expanded == NULL)
268 {
269 if (!expand_string_forcedfail)
270 log_write(0, LOG_MAIN|LOG_PANIC, "expansion of smtp_accept_max_per_host "
271 "failed for %s: %s", whofrom, expand_string_message);
272 }
273 /* For speed, interpret a decimal number inline here */
274 else
275 {
276 uschar *s = expanded;
277 while (isdigit(*s))
278 max_for_this_host = max_for_this_host * 10 + *s++ - '0';
279 if (*s != 0)
280 log_write(0, LOG_MAIN|LOG_PANIC, "expansion of smtp_accept_max_per_host "
281 "for %s contains non-digit: %s", whofrom, expanded);
282 }
283 }
284
285 /* If we have fewer connections than max_for_this_host, we can skip the tedious
286 per host_address checks. Note that at this stage smtp_accept_count contains the
287 count of *other* connections, not including this one. */
288
289 if ((max_for_this_host > 0) &&
290 (smtp_accept_count >= max_for_this_host))
291 {
292 int i;
293 int host_accept_count = 0;
294 int other_host_count = 0; /* keep a count of non matches to optimise */
295
296 for (i = 0; i < smtp_accept_max; ++i)
297 {
298 if (smtp_slots[i].host_address != NULL)
299 {
300 if (Ustrcmp(sender_host_address, smtp_slots[i].host_address) == 0)
301 host_accept_count++;
302 else
303 other_host_count++;
304
305 /* Testing all these strings is expensive - see if we can drop out
306 early, either by hitting the target, or finding there are not enough
307 connections left to make the target. */
308
309 if ((host_accept_count >= max_for_this_host) ||
310 ((smtp_accept_count - other_host_count) < max_for_this_host))
311 break;
312 }
313 }
314
315 if (host_accept_count >= max_for_this_host)
316 {
317 DEBUG(D_any) debug_printf("rejecting SMTP connection: too many from this "
318 "IP address: count=%d max=%d\n",
319 host_accept_count, max_for_this_host);
320 smtp_printf("421 Too many concurrent SMTP connections "
321 "from this IP address; please try again later.\r\n");
322 log_write(L_connection_reject,
323 LOG_MAIN, "Connection from %s refused: too many connections "
324 "from that IP address", whofrom);
325 goto ERROR_RETURN;
326 }
327 }
328
329 /* OK, the connection count checks have been passed. Before we can fork the
330 accepting process, we must first log the connection if requested. This logging
331 used to happen in the subprocess, but doing that means that the value of
332 smtp_accept_count can be out of step by the time it is logged. So we have to do
333 the logging here and accept the performance cost. Note that smtp_accept_count
334 hasn't yet been incremented to take account of this connection.
335
336 In order to minimize the cost (because this is going to happen for every
337 connection), do a preliminary selector test here. This saves ploughing through
338 the generalized logging code each time when the selector is false. If the
339 selector is set, check whether the host is on the list for logging. If not,
340 arrange to unset the selector in the subprocess. */
341
342 if ((log_write_selector & L_smtp_connection) != 0)
343 {
344 uschar *list = hosts_connection_nolog;
345 if (list != NULL && verify_check_host(&list) == OK)
346 use_log_write_selector &= ~L_smtp_connection;
347 else
348 log_write(L_smtp_connection, LOG_MAIN, "SMTP connection from %s "
349 "(TCP/IP connection count = %d)", whofrom, smtp_accept_count + 1);
350 }
351
352 /* Now we can fork the accepting process; do a lookup tidy, just in case any
353 expansion above did a lookup. */
354
355 search_tidyup();
356 pid = fork();
357
358 /* Handle the child process */
359
360 if (pid == 0)
361 {
362 int i;
363 int queue_only_reason = 0;
364 int old_pool = store_pool;
365 BOOL local_queue_only;
366 #ifdef SA_NOCLDWAIT
367 struct sigaction act;
368 #endif
369
370 /* May have been modified for the subprocess */
371
372 log_write_selector = use_log_write_selector;
373
374 /* Get the local interface address into permanent store */
375
376 store_pool = POOL_PERM;
377 interface_address = string_copy(interface_address);
378 store_pool = old_pool;
379
380 /* Check for a tls-on-connect port */
381
382 if (host_is_tls_on_connect_port(interface_port)) tls_on_connect = TRUE;
383
384 /* Expand smtp_active_hostname if required. We do not do this any earlier,
385 because it may depend on the local interface address (indeed, that is most
386 likely what it depends on.) */
387
388 smtp_active_hostname = primary_hostname;
389 if (raw_active_hostname != NULL)
390 {
391 uschar *nah = expand_string(raw_active_hostname);
392 if (nah == NULL)
393 {
394 if (!expand_string_forcedfail)
395 {
396 log_write(0, LOG_MAIN|LOG_PANIC, "failed to expand \"%s\" "
397 "(smtp_active_hostname): %s", raw_active_hostname,
398 expand_string_message);
399 smtp_printf("421 Local configuration error; "
400 "please try again later.\r\n");
401 mac_smtp_fflush();
402 search_tidyup();
403 _exit(EXIT_FAILURE);
404 }
405 }
406 else if (nah[0] != 0) smtp_active_hostname = nah;
407 }
408
409 /* Initialize the queueing flags */
410
411 queue_check_only();
412 local_queue_only = queue_only;
413
414 /* Close the listening sockets, and set the SIGCHLD handler to SIG_IGN.
415 We also attempt to set things up so that children are automatically reaped,
416 but just in case this isn't available, there's a paranoid waitpid() in the
417 loop too (except for systems where we are sure it isn't needed). See the more
418 extensive comment before the reception loop in exim.c for a fuller
419 explanation of this logic. */
420
421 for (i = 0; i < listen_socket_count; i++) close(listen_sockets[i]);
422
423 #ifdef SA_NOCLDWAIT
424 act.sa_handler = SIG_IGN;
425 sigemptyset(&(act.sa_mask));
426 act.sa_flags = SA_NOCLDWAIT;
427 sigaction(SIGCHLD, &act, NULL);
428 #else
429 signal(SIGCHLD, SIG_IGN);
430 #endif
431
432 /* Attempt to get an id from the sending machine via the RFC 1413
433 protocol. We do this in the sub-process in order not to hold up the
434 main process if there is any delay. Then set up the fullhost information
435 in case there is no HELO/EHLO. */
436
437 verify_get_ident(IDENT_PORT);
438 host_build_sender_fullhost();
439
440 DEBUG(D_any)
441 debug_printf("Process %d is handling incoming connection from %s\n",
442 (int)getpid(), sender_fullhost);
443
444 /* If there are too many child processes for immediate delivery,
445 set the local_queue_only flag, which is initialized from the
446 configured value and may therefore already be TRUE. Leave logging
447 till later so it will have a message id attached. */
448
449 if (smtp_accept_queue > 0 && smtp_accept_count >= smtp_accept_queue)
450 {
451 local_queue_only = TRUE;
452 queue_only_reason = 1;
453 }
454
455 /* Handle the start of the SMTP session, then loop, accepting incoming
456 messages from the SMTP connection. The end will come at the QUIT command,
457 when smtp_setup_msg() returns 0. A break in the connection causes the
458 process to die (see accept.c). */
459
460 if (!smtp_start_session())
461 {
462 mac_smtp_fflush();
463 search_tidyup();
464 _exit(EXIT_SUCCESS);
465 }
466
467 for (;;)
468 {
469 int rc;
470 message_id[0] = 0; /* Clear out any previous message_id */
471 reset_point = store_get(0); /* Save current store high water point */
472
473 DEBUG(D_any)
474 debug_printf("Process %d is ready for new message\n", (int)getpid());
475
476 /* Smtp_setup_msg() returns 0 on QUIT or if the call is from an
477 unacceptable host or if an ACL "drop" command was triggered, -1 on
478 connection lost, and +1 on validly reaching DATA. Receive_msg() almost
479 always returns TRUE when smtp_input is true; just retry if no message was
480 accepted (can happen for invalid message parameters). However, it can yield
481 FALSE if the connection was forcibly dropped by the DATA ACL. */
482
483 if ((rc = smtp_setup_msg()) > 0)
484 {
485 BOOL ok = receive_msg(FALSE);
486 search_tidyup(); /* Close cached databases */
487 if (!ok) /* Connection was dropped */
488 {
489 mac_smtp_fflush();
490 _exit(EXIT_SUCCESS);
491 }
492 if (message_id[0] == 0) continue; /* No message was accepted */
493 }
494 else
495 {
496 mac_smtp_fflush();
497 search_tidyup();
498 _exit((rc == 0)? EXIT_SUCCESS : EXIT_FAILURE);
499 }
500
501 /* Show the recipients when debugging */
502
503 DEBUG(D_receive)
504 {
505 int i;
506 if (sender_address != NULL)
507 debug_printf("Sender: %s\n", sender_address);
508 if (recipients_list != NULL)
509 {
510 debug_printf("Recipients:\n");
511 for (i = 0; i < recipients_count; i++)
512 debug_printf(" %s\n", recipients_list[i].address);
513 }
514 }
515
516 /* A message has been accepted. Clean up any previous delivery processes
517 that have completed and are defunct, on systems where they don't go away
518 by themselves (see comments when setting SIG_IGN above). On such systems
519 (if any) these delivery processes hang around after termination until
520 the next message is received. */
521
522 #ifndef SIG_IGN_WORKS
523 while (waitpid(-1, NULL, WNOHANG) > 0);
524 #endif
525
526 /* Reclaim up the store used in accepting this message */
527
528 store_reset(reset_point);
529
530 /* If queue_only is set or if there are too many incoming connections in
531 existence, local_queue_only will be TRUE. If it is not, check whether we
532 have received too many messages in this session for immediate delivery. If
533 not, and queue_only_load is set, check that the load average is below it.
534 Note that, once set, local_queue_only remains set for any subsequent
535 messages on the same SMTP connection. This is a deliberate choice; even
536 though the load average may fall, it doesn't seem right to deliver later
537 messages on the same call when not delivering earlier ones. */
538
539 if (!local_queue_only)
540 {
541 if (smtp_accept_queue_per_connection > 0 &&
542 receive_messagecount > smtp_accept_queue_per_connection)
543 {
544 local_queue_only = TRUE;
545 queue_only_reason = 2;
546 }
547 else if (queue_only_load >= 0)
548 {
549 local_queue_only = (load_average = os_getloadavg()) > queue_only_load;
550 if (local_queue_only) queue_only_reason = 3;
551 }
552 }
553
554 /* Log the queueing here, when it will get a message id attached, but
555 not if queue_only is set (case 0). */
556
557 if (local_queue_only) switch(queue_only_reason)
558 {
559 case 1:
560 log_write(L_delay_delivery,
561 LOG_MAIN, "no immediate delivery: too many connections "
562 "(%d, max %d)", smtp_accept_count, smtp_accept_queue);
563 break;
564
565 case 2:
566 log_write(L_delay_delivery,
567 LOG_MAIN, "no immediate delivery: more than %d messages "
568 "received in one connection", smtp_accept_queue_per_connection);
569 break;
570
571 case 3:
572 log_write(L_delay_delivery,
573 LOG_MAIN, "no immediate delivery: load average %.2f",
574 (double)load_average/1000.0);
575 break;
576 }
577
578 /* If a delivery attempt is required, spin off a new process to handle it.
579 If we are not root, we have to re-exec exim unless deliveries are being
580 done unprivileged. */
581
582 else if (!queue_only_policy && !deliver_freeze)
583 {
584 pid_t dpid;
585
586 /* Before forking, ensure that the C output buffer is flushed. Otherwise
587 anything that it in it will get duplicated, leading to duplicate copies
588 of the pending output. */
589
590 mac_smtp_fflush();
591
592 if ((dpid = fork()) == 0)
593 {
594 fclose(smtp_in);
595 fclose(smtp_out);
596
597 /* Don't ever molest the parent's SSL connection, but do clean up
598 the data structures if necessary. */
599
600 #ifdef SUPPORT_TLS
601 tls_close(FALSE);
602 #endif
603
604 /* Reset SIGHUP and SIGCHLD in the child in both cases. */
605
606 signal(SIGHUP, SIG_DFL);
607 signal(SIGCHLD, SIG_DFL);
608
609 if (geteuid() != root_uid && !deliver_drop_privilege)
610 {
611 signal(SIGALRM, SIG_DFL);
612 (void)child_exec_exim(CEE_EXEC_PANIC, FALSE, NULL, FALSE, 2, US"-Mc",
613 message_id);
614 /* Control does not return here. */
615 }
616
617 /* No need to re-exec; SIGALRM remains set to the default handler */
618
619 (void)deliver_message(message_id, FALSE, FALSE);
620 search_tidyup();
621 _exit(EXIT_SUCCESS);
622 }
623
624 if (dpid > 0)
625 {
626 DEBUG(D_any) debug_printf("forked delivery process %d\n", (int)dpid);
627 }
628 else
629 {
630 log_write(0, LOG_MAIN|LOG_PANIC, "daemon: delivery process fork "
631 "failed: %s", strerror(errno));
632 }
633 }
634 }
635 }
636
637
638 /* Carrying on in the parent daemon process... Can't do much if the fork
639 failed. Otherwise, keep count of the number of accepting processes and
640 remember the pid for ticking off when the child completes. */
641
642 if (pid < 0)
643 {
644 never_error(US"daemon: accept process fork failed", US"Fork failed", errno);
645 }
646 else
647 {
648 int i;
649 for (i = 0; i < smtp_accept_max; ++i)
650 {
651 if (smtp_slots[i].pid <= 0)
652 {
653 smtp_slots[i].pid = pid;
654 if (smtp_accept_max_per_host != NULL)
655 smtp_slots[i].host_address = string_copy_malloc(sender_host_address);
656 smtp_accept_count++;
657 break;
658 }
659 }
660 DEBUG(D_any) debug_printf("%d SMTP accept process%s running\n",
661 smtp_accept_count, (smtp_accept_count == 1)? "" : "es");
662 }
663
664 /* Get here via goto in error cases */
665
666 ERROR_RETURN:
667
668 /* Close the streams associated with the socket which will also close the
669 socket fds in this process. We can't do anything if fclose() fails, but
670 logging brings it to someone's attention. However, "connection reset by peer"
671 isn't really a problem, so skip that one. If the streams don't exist, something
672 went wrong while setting things up. Make sure the socket descriptors are
673 closed, in order to drop the connection. */
674
675 if (smtp_out != NULL)
676 {
677 if (fclose(smtp_out) != 0 && errno != ECONNRESET)
678 log_write(0, LOG_MAIN|LOG_PANIC, "daemon: fclose(smtp_out) failed: %s",
679 strerror(errno));
680 smtp_out = NULL;
681 }
682 else close(accept_socket);
683
684 if (smtp_in != NULL)
685 {
686 if (fclose(smtp_in) != 0 && errno != ECONNRESET)
687 log_write(0, LOG_MAIN|LOG_PANIC, "daemon: fclose(smtp_in) failed: %s",
688 strerror(errno));
689 smtp_in = NULL;
690 }
691 else close(dup_accept_socket);
692
693 /* Release any store used in this process, including the store used for holding
694 the incoming host address and an expanded active_hostname. */
695
696 store_reset(reset_point);
697 sender_host_address = NULL;
698 }
699
700
701
702
703 /*************************************************
704 * Check wildcard listen special cases *
705 *************************************************/
706
707 /* This function is used when binding and listening on lists of addresses and
708 ports. It tests for special cases of wildcard listening, when IPv4 and IPv6
709 sockets may interact in different ways in different operating systems. It is
710 passed an error number, the list of listening addresses, and the current
711 address. Two checks are available: for a previous wildcard IPv6 address, or for
712 a following wildcard IPv4 address, in both cases on the same port.
713
714 In practice, pairs of wildcard addresses should be adjacent in the address list
715 because they are sorted that way below.
716
717 Arguments:
718 eno the error number
719 addresses the list of addresses
720 ipa the current IP address
721 back if TRUE, check for previous wildcard IPv6 address
722 if FALSE, check for a following wildcard IPv4 address
723
724 Returns: TRUE or FALSE
725 */
726
727 static BOOL
728 check_special_case(int eno, ip_address_item *addresses, ip_address_item *ipa,
729 BOOL back)
730 {
731 ip_address_item *ipa2;
732
733 /* For the "back" case, if the failure was "address in use" for a wildcard IPv4
734 address, seek a previous IPv6 wildcard address on the same port. As it is
735 previous, it must have been successfully bound and be listening. Flag it as a
736 "6 including 4" listener. */
737
738 if (back)
739 {
740 if (eno != EADDRINUSE || ipa->address[0] != 0) return FALSE;
741 for (ipa2 = addresses; ipa2 != ipa; ipa2 = ipa2->next)
742 {
743 if (ipa2->address[1] == 0 && ipa2->port == ipa->port)
744 {
745 ipa2->v6_include_v4 = TRUE;
746 return TRUE;
747 }
748 }
749 }
750
751 /* For the "forward" case, if the current address is a wildcard IPv6 address,
752 we seek a following wildcard IPv4 address on the same port. */
753
754 else
755 {
756 if (ipa->address[0] != ':' || ipa->address[1] != 0) return FALSE;
757 for (ipa2 = ipa->next; ipa2 != NULL; ipa2 = ipa2->next)
758 if (ipa2->address[0] == 0 && ipa->port == ipa2->port) return TRUE;
759 }
760
761 return FALSE;
762 }
763
764
765
766
767
768
769 /*************************************************
770 * Exim Daemon Mainline *
771 *************************************************/
772
773 /* The daemon can do two jobs, either of which is optional:
774
775 (1) Listens for incoming SMTP calls and spawns off a sub-process to handle
776 each one. This is requested by the -bd option, with -oX specifying the SMTP
777 port on which to listen (for testing).
778
779 (2) Spawns a queue-running process every so often. This is controlled by the
780 -q option with a an interval time. (If no time is given, a single queue run
781 is done from the main function, and control doesn't get here.)
782
783 Root privilege is required in order to attach to port 25. Some systems require
784 it when calling socket() rather than bind(). To cope with all cases, we run as
785 root for both socket() and bind(). Some systems also require root in order to
786 write to the pid file directory. This function must therefore be called as root
787 if it is to work properly in all circumstances. Once the socket is bound and
788 the pid file written, root privilege is given up if there is an exim uid.
789
790 There are no arguments to this function, and it never returns. */
791
792 void
793 daemon_go(void)
794 {
795 int *listen_sockets = NULL;
796 int listen_socket_count = 0;
797 ip_address_item *addresses = NULL;
798
799 /* If any debugging options are set, turn on the D_pid bit so that all
800 debugging lines get the pid added. */
801
802 DEBUG(D_any|D_v) debug_selector |= D_pid;
803
804
805 /* Do the preparation for setting up a listener on one or more interfaces, and
806 possible on various ports. This is controlled by the combination of
807 local_interfaces (which can set IP addresses and ports) and daemon_smtp_port
808 (which is a list of default ports to use for those items in local_interfaces
809 that do not specify a port). The -oX command line option can be used to
810 override one or both of these options.
811
812 If local_interfaces is not set, the default is to listen on all interfaces.
813 When it is set, it can include "all IPvx interfaces" as an item. This is useful
814 when different ports are in use.
815
816 It turns out that listening on all interfaces is messy in an IPv6 world,
817 because several different implementation approaches have been taken. This code
818 is now supposed to work with all of them. The point of difference is whether an
819 IPv6 socket that is listening on all interfaces will receive incoming IPv4
820 calls or not. We also have to cope with the case when IPv6 libraries exist, but
821 there is no IPv6 support in the kernel.
822
823 . On Solaris, an IPv6 socket will accept IPv4 calls, and give them as mapped
824 addresses. However, if an IPv4 socket is also listening on all interfaces,
825 calls are directed to the appropriate socket.
826
827 . On (some versions of) Linux, an IPv6 socket will accept IPv4 calls, and
828 give them as mapped addresses, but an attempt also to listen on an IPv4
829 socket on all interfaces causes an error.
830
831 . On OpenBSD, an IPv6 socket will not accept IPv4 calls. You have to set up
832 two sockets if you want to accept both kinds of call.
833
834 . FreeBSD is like OpenBSD, but it has the IPV6_V6ONLY socket option, which
835 can be turned off, to make it behave like the versions of Linux described
836 above.
837
838 . I heard a report that the USAGI IPv6 stack for Linux has implemented
839 IPV6_V6ONLY.
840
841 So, what we do when IPv6 is supported is as follows:
842
843 (1) After it is set up, the list of interfaces is scanned for wildcard
844 addresses. If an IPv6 and an IPv4 wildcard are both found for the same
845 port, the list is re-arranged so that they are together, with the IPv6
846 wildcard first.
847
848 (2) If the creation of a wildcard IPv6 socket fails, we just log the error and
849 carry on if an IPv4 wildcard socket for the same port follows later in the
850 list. This allows Exim to carry on in the case when the kernel has no IPv6
851 support.
852
853 (3) Having created an IPv6 wildcard socket, we try to set IPV6_V6ONLY if that
854 option is defined. However, if setting fails, carry on regardless (but log
855 the incident).
856
857 (4) If binding or listening on an IPv6 wildcard socket fails, it is a serious
858 error.
859
860 (5) If binding or listening on an IPv4 wildcard socket fails with the error
861 EADDRINUSE, and a previous interface was an IPv6 wildcard for the same
862 port (which must have succeeded or we wouldn't have got this far), we
863 assume we are in the situation where just a single socket is permitted,
864 and ignore the error.
865
866 Phew!
867
868 The preparation code decodes options and sets up the relevant data. We do this
869 first, so that we can return non-zero if there are any syntax errors, and also
870 write to stderr. */
871
872 if (daemon_listen)
873 {
874 int *default_smtp_port;
875 int sep;
876 int pct = 0;
877 uschar *s;
878 uschar *list;
879 uschar *local_iface_source = US"local_interfaces";
880 ip_address_item *ipa;
881 ip_address_item **pipa;
882
883 /* If any option requiring a load average to be available during the
884 reception of a message is set, call os_getloadavg() while we are root
885 for those OS for which this is necessary the first time it is called (in
886 order to perform an "open" on the kernel memory file). */
887
888 #ifdef LOAD_AVG_NEEDS_ROOT
889 if (queue_only_load >= 0 || smtp_load_reserve >= 0) (void)os_getloadavg();
890 #endif
891
892 /* If -oX was used, disable the writing of a pid file unless -oP was
893 explicitly used to force it. Then scan the string given to -oX. Any items
894 that contain neither a dot nor a colon are used to override daemon_smtp_port.
895 Any other items are used to override local_interfaces. */
896
897 if (override_local_interfaces != NULL)
898 {
899 uschar *new_smtp_port = NULL;
900 uschar *new_local_interfaces = NULL;
901 int portsize = 0;
902 int portptr = 0;
903 int ifacesize = 0;
904 int ifaceptr = 0;
905
906 if (override_pid_file_path == NULL) write_pid = FALSE;
907
908 list = override_local_interfaces;
909 sep = 0;
910 while ((s = string_nextinlist(&list,&sep,big_buffer,big_buffer_size))
911 != NULL)
912 {
913 uschar joinstr[4];
914 uschar **ptr;
915 int *sizeptr;
916 int *ptrptr;
917
918 if (Ustrpbrk(s, ".:") == NULL)
919 {
920 ptr = &new_smtp_port;
921 sizeptr = &portsize;
922 ptrptr = &portptr;
923 }
924 else
925 {
926 ptr = &new_local_interfaces;
927 sizeptr = &ifacesize;
928 ptrptr = &ifaceptr;
929 }
930
931 if (*ptr == NULL)
932 {
933 joinstr[0] = sep;
934 joinstr[1] = ' ';
935 *ptr = string_cat(*ptr, sizeptr, ptrptr, US"<", 1);
936 }
937
938 *ptr = string_cat(*ptr, sizeptr, ptrptr, joinstr, 2);
939 *ptr = string_cat(*ptr, sizeptr, ptrptr, s, Ustrlen(s));
940 }
941
942 if (new_smtp_port != NULL)
943 {
944 new_smtp_port[portptr] = 0;
945 daemon_smtp_port = new_smtp_port;
946 DEBUG(D_any) debug_printf("daemon_smtp_port overridden by -oX:\n %s\n",
947 daemon_smtp_port);
948 }
949
950 if (new_local_interfaces != NULL)
951 {
952 new_local_interfaces[ifaceptr] = 0;
953 local_interfaces = new_local_interfaces;
954 local_iface_source = US"-oX data";
955 DEBUG(D_any) debug_printf("local_interfaces overridden by -oX:\n %s\n",
956 local_interfaces);
957 }
958 }
959
960 /* Create a list of default SMTP ports, to be used if local_interfaces
961 contains entries without explict ports. First count the number of ports, then
962 build a translated list in a vector. */
963
964 list = daemon_smtp_port;
965 sep = 0;
966 while ((s = string_nextinlist(&list,&sep,big_buffer,big_buffer_size)) != NULL)
967 pct++;
968 default_smtp_port = store_get((pct+1) * sizeof(int));
969 list = daemon_smtp_port;
970 sep = 0;
971 for (pct = 0;
972 (s = string_nextinlist(&list,&sep,big_buffer,big_buffer_size)) != NULL;
973 pct++)
974 {
975 if (isdigit(*s))
976 {
977 uschar *end;
978 default_smtp_port[pct] = Ustrtol(s, &end, 0);
979 if (end != s + Ustrlen(s))
980 log_write(0, LOG_PANIC_DIE|LOG_CONFIG, "invalid SMTP port: %s", s);
981 }
982 else
983 {
984 struct servent *smtp_service = getservbyname(CS s, "tcp");
985 if (smtp_service == NULL)
986 log_write(0, LOG_PANIC_DIE|LOG_CONFIG, "TCP port \"%s\" not found", s);
987 default_smtp_port[pct] = ntohs(smtp_service->s_port);
988 }
989 }
990 default_smtp_port[pct] = 0;
991
992 /* Create the list of local interfaces, possibly with ports included. This
993 list may contain references to 0.0.0.0 and ::0 as wildcards. These special
994 values are converted below. */
995
996 addresses = host_build_ifacelist(local_interfaces, local_iface_source);
997
998 /* In the list of IP addresses, convert 0.0.0.0 into an empty string, and ::0
999 into the string ":". We use these to recognize wildcards in IPv4 and IPv6. In
1000 fact, many IP stacks recognize 0.0.0.0 and ::0 and handle them as wildcards
1001 anyway, but we need to know which are the wildcard addresses, and the shorter
1002 strings are neater.
1003
1004 In the same scan, fill in missing port numbers from the default list. When
1005 there is more than one item in the list, extra items are created. */
1006
1007 for (ipa = addresses; ipa != NULL; ipa = ipa->next)
1008 {
1009 int i;
1010
1011 if (Ustrcmp(ipa->address, "0.0.0.0") == 0) ipa->address[0] = 0;
1012 else if (Ustrcmp(ipa->address, "::0") == 0)
1013 {
1014 ipa->address[0] = ':';
1015 ipa->address[1] = 0;
1016 }
1017
1018 if (ipa->port > 0) continue;
1019
1020 if (daemon_smtp_port[0] <= 0)
1021 log_write(0, LOG_MAIN|LOG_PANIC_DIE, "no port specified for interface "
1022 "%s and daemon_smtp_port is unset; cannot start daemon",
1023 (ipa->address[0] == 0)? US"\"all IPv4\"" :
1024 (ipa->address[1] == 0)? US"\"all IPv6\"" : ipa->address);
1025 ipa->port = default_smtp_port[0];
1026 for (i = 1; default_smtp_port[i] > 0; i++)
1027 {
1028 ip_address_item *new = store_get(sizeof(ip_address_item));
1029 memcpy(new->address, ipa->address, Ustrlen(ipa->address) + 1);
1030 new->port = default_smtp_port[i];
1031 new->next = ipa->next;
1032 ipa->next = new;
1033 ipa = new;
1034 }
1035 }
1036
1037 /* Scan the list of addresses for wildcards. If we find an IPv4 and an IPv6
1038 wildcard for the same port, ensure that (a) they are together and (b) the
1039 IPv6 address comes first. This makes handling the messy features easier, and
1040 also simplifies the construction of the "daemon started" log line. */
1041
1042 pipa = &addresses;
1043 for (ipa = addresses; ipa != NULL; pipa = &(ipa->next), ipa = ipa->next)
1044 {
1045 ip_address_item *ipa2;
1046
1047 /* Handle an IPv4 wildcard */
1048
1049 if (ipa->address[0] == 0)
1050 {
1051 for (ipa2 = ipa; ipa2->next != NULL; ipa2 = ipa2->next)
1052 {
1053 ip_address_item *ipa3 = ipa2->next;
1054 if (ipa3->address[0] == ':' &&
1055 ipa3->address[1] == 0 &&
1056 ipa3->port == ipa->port)
1057 {
1058 ipa2->next = ipa3->next;
1059 ipa3->next = ipa;
1060 *pipa = ipa3;
1061 break;
1062 }
1063 }
1064 }
1065
1066 /* Handle an IPv6 wildcard. */
1067
1068 else if (ipa->address[0] == ':' && ipa->address[1] == 0)
1069 {
1070 for (ipa2 = ipa; ipa2->next != NULL; ipa2 = ipa2->next)
1071 {
1072 ip_address_item *ipa3 = ipa2->next;
1073 if (ipa3->address[0] == 0 && ipa3->port == ipa->port)
1074 {
1075 ipa2->next = ipa3->next;
1076 ipa3->next = ipa->next;
1077 ipa->next = ipa3;
1078 ipa = ipa3;
1079 break;
1080 }
1081 }
1082 }
1083 }
1084
1085 /* Get a vector to remember all the sockets in */
1086
1087 for (ipa = addresses; ipa != NULL; ipa = ipa->next)
1088 listen_socket_count++;
1089 listen_sockets = store_get(sizeof(int *) * listen_socket_count);
1090
1091 /* Do a sanity check on the max connects value just to save us from getting
1092 a huge amount of store. */
1093
1094 if (smtp_accept_max > 4095) smtp_accept_max = 4096;
1095
1096 /* There's no point setting smtp_accept_queue unless it is less than the max
1097 connects limit. The configuration reader ensures that the max is set if the
1098 queue-only option is set. */
1099
1100 if (smtp_accept_queue > smtp_accept_max) smtp_accept_queue = 0;
1101
1102 /* Get somewhere to keep the list of SMTP accepting pids if we are keeping
1103 track of them for total number and queue/host limits. */
1104
1105 if (smtp_accept_max > 0)
1106 {
1107 int i;
1108 smtp_slots = store_get(smtp_accept_max * sizeof(smtp_slot));
1109 for (i = 0; i < smtp_accept_max; i++) smtp_slots[i] = empty_smtp_slot;
1110 }
1111 }
1112
1113 /* We now close all open file descriptors that we know about, and disconnect
1114 from the controlling terminal, unless background_daemon is unset. This is
1115 always unset when debugging, but can also be forced. Most modern Unixes seem to
1116 have setsid() for getting rid of the controlling terminal. For any OS that
1117 doesn't, setsid() can be #defined as a no-op, or as something else. */
1118
1119 if (background_daemon)
1120 {
1121 log_close_all(); /* Just in case anything was logged earlier */
1122 search_tidyup(); /* Just in case any were used in reading the config. */
1123 close(0); /* Get rid of stdin/stdout/stderr */
1124 close(1);
1125 close(2);
1126 log_stderr = NULL; /* So no attempt to copy paniclog output */
1127
1128 /* If the parent process of this one has pid == 1, we are re-initializing the
1129 daemon as the result of a SIGHUP. In this case, there is no need to do any
1130 forking, because the controlling terminal has long gone. Otherwise, fork,
1131 in case current process is a process group leader (see 'man setsid' for an
1132 explanation). */
1133
1134 if (getppid() != 1)
1135 {
1136 pid_t pid = fork();
1137 if (pid < 0) log_write(0, LOG_MAIN|LOG_PANIC_DIE,
1138 "fork() failed when starting daemon: %s", strerror(errno));
1139 if (pid > 0) exit(EXIT_SUCCESS); /* in parent process, just exit */
1140 (void)setsid(); /* release controlling terminal */
1141 }
1142 }
1143
1144 /* We are now in the disconnected, daemon process (unless debugging). Set up
1145 the listening sockets if required. */
1146
1147 if (daemon_listen)
1148 {
1149 int sk;
1150 int on = 1;
1151 ip_address_item *ipa;
1152
1153 /* For each IP address, create a socket, bind it to the appropriate port, and
1154 start listening. See comments above about IPv6 sockets that may or may not
1155 accept IPv4 calls when listening on all interfaces. We also have to cope with
1156 the case of a system with IPv6 libraries, but no IPv6 support in the kernel.
1157 listening, provided a wildcard IPv4 socket for the same port follows. */
1158
1159 for (ipa = addresses, sk = 0; sk < listen_socket_count; ipa = ipa->next, sk++)
1160 {
1161 BOOL wildcard;
1162 ip_address_item *ipa2;
1163 int retries = 9;
1164 int af;
1165
1166 if (Ustrchr(ipa->address, ':') != NULL)
1167 {
1168 af = AF_INET6;
1169 wildcard = ipa->address[1] == 0;
1170 }
1171 else
1172 {
1173 af = AF_INET;
1174 wildcard = ipa->address[0] == 0;
1175 }
1176
1177 listen_sockets[sk] = ip_socket(SOCK_STREAM, af);
1178 if (listen_sockets[sk] < 0)
1179 {
1180 if (check_special_case(0, addresses, ipa, FALSE))
1181 {
1182 log_write(0, LOG_MAIN, "Failed to create IPv6 socket for wildcard "
1183 "listening (%s): will use IPv4", strerror(errno));
1184 goto SKIP_SOCKET;
1185 }
1186 log_write(0, LOG_PANIC_DIE, "IPv%c socket creation failed: %s",
1187 (af == AF_INET6)? '6' : '4', strerror(errno));
1188 }
1189
1190 /* If this is an IPv6 wildcard socket, set IPV6_V6ONLY if that option is
1191 available. Just log failure (can get protocol not available, just like
1192 socket creation can). */
1193
1194 #ifdef IPV6_V6ONLY
1195 if (af == AF_INET6 && wildcard &&
1196 setsockopt(listen_sockets[sk], IPPROTO_IPV6, IPV6_V6ONLY, (char *)(&on),
1197 sizeof(on)) < 0)
1198 log_write(0, LOG_MAIN, "Setting IPV6_V6ONLY on daemon's IPv6 wildcard "
1199 "socket failed (%s): carrying on without it", strerror(errno));
1200 #endif /* IPV6_V6ONLY */
1201
1202 /* Set SO_REUSEADDR so that the daemon can be restarted while a connection
1203 is being handled. Without this, a connection will prevent reuse of the
1204 smtp port for listening. */
1205
1206 if (setsockopt(listen_sockets[sk], SOL_SOCKET, SO_REUSEADDR,
1207 (uschar *)(&on), sizeof(on)) < 0)
1208 log_write(0, LOG_MAIN|LOG_PANIC_DIE, "setting SO_REUSEADDR on socket "
1209 "failed when starting daemon: %s", strerror(errno));
1210
1211 /* Set TCP_NODELAY; Exim does its own buffering. There is a switch to
1212 disable this because it breaks some broken clients. */
1213
1214 if (tcp_nodelay) setsockopt(listen_sockets[sk], IPPROTO_TCP, TCP_NODELAY,
1215 (uschar *)(&on), sizeof(on));
1216
1217 /* Now bind the socket to the required port; if Exim is being restarted
1218 it may not always be possible to bind immediately, even with SO_REUSEADDR
1219 set, so try 10 times, waiting between each try. After 10 failures, we give
1220 up. In an IPv6 environment, if bind () fails with the error EADDRINUSE and
1221 we are doing wildcard IPv4 listening and there was a previous IPv6 wildcard
1222 address for the same port, ignore the error on the grounds that we must be
1223 in a system where the IPv6 socket accepts both kinds of call. This is
1224 necessary for (some release of) USAGI Linux; other IP stacks fail at the
1225 listen() stage instead. */
1226
1227 for(;;)
1228 {
1229 uschar *msg, *addr;
1230 if (ip_bind(listen_sockets[sk], af, ipa->address, ipa->port) >= 0) break;
1231 if (check_special_case(errno, addresses, ipa, TRUE))
1232 {
1233 DEBUG(D_any) debug_printf("wildcard IPv4 bind() failed after IPv6 "
1234 "listen() success; EADDRINUSE ignored\n");
1235 close(listen_sockets[sk]);
1236 goto SKIP_SOCKET;
1237 }
1238 msg = US strerror(errno);
1239 addr = wildcard? ((af == AF_INET6)? US"(any IPv6)" : US"(any IPv4)") :
1240 ipa->address;
1241 if (retries-- <= 0)
1242 log_write(0, LOG_MAIN|LOG_PANIC_DIE,
1243 "socket bind() to port %d for address %s failed: %s: "
1244 "daemon abandoned", ipa->port, addr, msg);
1245 log_write(0, LOG_MAIN, "socket bind() to port %d for address %s "
1246 "failed: %s: waiting before trying again", ipa->port, addr, msg);
1247 sleep(30);
1248 }
1249
1250 DEBUG(D_any)
1251 {
1252 if (wildcard)
1253 debug_printf("listening on all interfaces (IPv%c) port %d\n",
1254 (af == AF_INET6)? '6' : '4', ipa->port);
1255 else
1256 debug_printf("listening on %s port %d\n", ipa->address, ipa->port);
1257 }
1258
1259 /* Start listening on the bound socket, establishing the maximum backlog of
1260 connections that is allowed. On success, continue to the next address. */
1261
1262 if (listen(listen_sockets[sk], smtp_connect_backlog) >= 0) continue;
1263
1264 /* Listening has failed. In an IPv6 environment, as for bind(), if listen()
1265 fails with the error EADDRINUSE and we are doing IPv4 wildcard listening
1266 and there was a previous successful IPv6 wildcard listen on the same port,
1267 we want to ignore the error on the grounds that we must be in a system
1268 where the IPv6 socket accepts both kinds of call. */
1269
1270 if (!check_special_case(errno, addresses, ipa, TRUE))
1271 log_write(0, LOG_PANIC_DIE, "listen() failed on interface %s: %s",
1272 wildcard? ((af == AF_INET6)? US"(any IPv6)" : US"(any IPv4)") :
1273 ipa->address,
1274 strerror(errno));
1275
1276 DEBUG(D_any) debug_printf("wildcard IPv4 listen() failed after IPv6 "
1277 "listen() success; EADDRINUSE ignored\n");
1278 close(listen_sockets[sk]);
1279
1280 /* Come here if there has been a problem with the socket which we
1281 are going to ignore. We remove the address from the chain, and back up the
1282 counts. */
1283
1284 SKIP_SOCKET:
1285 sk--; /* Back up the count */
1286 listen_socket_count--; /* Reduce the total */
1287 if (ipa == addresses) addresses = ipa->next; else
1288 {
1289 for (ipa2 = addresses; ipa2->next != ipa; ipa2 = ipa2->next);
1290 ipa2->next = ipa->next;
1291 ipa = ipa2;
1292 }
1293 } /* End of bind/listen loop for each address */
1294 } /* End of setup for listening */
1295
1296
1297 /* If we are not listening, we want to write a pid file only if -oP was
1298 explicitly given. */
1299
1300 else if (override_pid_file_path == NULL) write_pid = FALSE;
1301
1302 /* Write the pid to a known file for assistance in identification, if required.
1303 We do this before giving up root privilege, because on some systems it is
1304 necessary to be root in order to write into the pid file directory. There's
1305 nothing to stop multiple daemons running, as long as no more than one listens
1306 on a given TCP/IP port on the same interface(s). However, in these
1307 circumstances it gets far too complicated to mess with pid file names
1308 automatically. Consequently, Exim 4 writes a pid file only
1309
1310 (a) When running in the test harness, or
1311 (b) When -bd is used and -oX is not used, or
1312 (c) When -oP is used to supply a path.
1313
1314 The variable daemon_write_pid is used to control this. */
1315
1316 if (running_in_test_harness || write_pid)
1317 {
1318 FILE *f;
1319
1320 if (override_pid_file_path != NULL)
1321 pid_file_path = override_pid_file_path;
1322
1323 if (pid_file_path[0] == 0)
1324 pid_file_path = string_sprintf("%s/exim-daemon.pid", spool_directory);
1325
1326 f = Ufopen(pid_file_path, "wb");
1327 if (f != NULL)
1328 {
1329 fprintf(f, "%d\n", (int)getpid());
1330 fchmod(fileno(f), 0644);
1331 fclose(f);
1332 DEBUG(D_any) debug_printf("pid written to %s\n", pid_file_path);
1333 }
1334 else
1335 {
1336 DEBUG(D_any)
1337 debug_printf("%s\n", string_open_failed(errno, "pid file %s",
1338 pid_file_path));
1339 }
1340 }
1341
1342 /* Set up the handler for SIGHUP, which causes a restart of the daemon. */
1343
1344 sighup_seen = FALSE;
1345 signal(SIGHUP, sighup_handler);
1346
1347 /* Give up root privilege at this point (assuming that exim_uid and exim_gid
1348 are not root). The third argument controls the running of initgroups().
1349 Normally we do this, in order to set up the groups for the Exim user. However,
1350 if we are not root at this time - some odd installations run that way - we
1351 cannot do this. */
1352
1353 exim_setugid(exim_uid, exim_gid, geteuid()==root_uid, US"running as a daemon");
1354
1355 /* Get somewhere to keep the list of queue-runner pids if we are keeping track
1356 of them (and also if we are doing queue runs). */
1357
1358 if (queue_interval > 0 && queue_run_max > 0)
1359 {
1360 int i;
1361 queue_pid_slots = store_get(queue_run_max * sizeof(pid_t));
1362 for (i = 0; i < queue_run_max; i++) queue_pid_slots[i] = 0;
1363 }
1364
1365 /* Set up the handler for termination of child processes. */
1366
1367 sigchld_seen = FALSE;
1368 signal(SIGCHLD, main_sigchld_handler);
1369
1370 /* If we are to run the queue periodically, pretend the alarm has just gone
1371 off. This will cause the first queue-runner to get kicked off straight away. */
1372
1373 sigalrm_seen = (queue_interval > 0);
1374
1375 /* Log the start up of a daemon - at least one of listening or queue running
1376 must be set up. */
1377
1378 if (daemon_listen)
1379 {
1380 int i, j;
1381 int smtp_ports = 0;
1382 int smtps_ports = 0;
1383 ip_address_item *ipa;
1384 uschar *p = big_buffer;
1385 uschar *qinfo = (queue_interval > 0)?
1386 string_sprintf("-q%s", readconf_printtime(queue_interval))
1387 :
1388 US"no queue runs";
1389
1390 /* Build a list of listening addresses in big_buffer, but limit it to 10
1391 items. The style is for backwards compatibility.
1392
1393 It is now possible to have some ports listening for SMTPS (the old,
1394 deprecated protocol that starts TLS without using STARTTLS), and others
1395 listening for standard SMTP. Keep their listings separate. */
1396
1397 for (j = 0; j < 2; j++)
1398 {
1399 for (i = 0, ipa = addresses; i < 10 && ipa != NULL; i++, ipa = ipa->next)
1400 {
1401 /* First time round, look for SMTP ports; second time round, look for
1402 SMTPS ports. For the first one of each, insert leading text. */
1403
1404 if (host_is_tls_on_connect_port(ipa->port) == (j > 0))
1405 {
1406 if (j == 0)
1407 {
1408 if (smtp_ports++ == 0)
1409 {
1410 memcpy(p, "SMTP on", 8);
1411 p += 7;
1412 }
1413 }
1414 else
1415 {
1416 if (smtps_ports++ == 0)
1417 {
1418 (void)sprintf(CS p, "%sSMTPS on",
1419 (smtp_ports == 0)? "":" and for ");
1420 while (*p != 0) p++;
1421 }
1422 }
1423
1424 /* Now the information about the port (and sometimes interface) */
1425
1426 if (ipa->address[0] == ':' && ipa->address[1] == 0)
1427 {
1428 if (ipa->next != NULL && ipa->next->address[0] == 0 &&
1429 ipa->next->port == ipa->port)
1430 {
1431 (void)sprintf(CS p, " port %d (IPv6 and IPv4)", ipa->port);
1432 ipa = ipa->next;
1433 }
1434 else if (ipa->v6_include_v4)
1435 (void)sprintf(CS p, " port %d (IPv6 with IPv4)", ipa->port);
1436 else
1437 (void)sprintf(CS p, " port %d (IPv6)", ipa->port);
1438 }
1439 else if (ipa->address[0] == 0)
1440 (void)sprintf(CS p, " port %d (IPv4)", ipa->port);
1441 else
1442 (void)sprintf(CS p, " [%s]:%d", ipa->address, ipa->port);
1443 while (*p != 0) p++;
1444 }
1445 }
1446
1447 if (ipa != NULL)
1448 {
1449 memcpy(p, " ...", 5);
1450 p += 4;
1451 }
1452 }
1453
1454 log_write(0, LOG_MAIN,
1455 "exim %s daemon started: pid=%d, %s, listening for %s",
1456 version_string, getpid(), qinfo, big_buffer);
1457 set_process_info("daemon: %s, listening for %s", qinfo, big_buffer);
1458 }
1459
1460 else
1461 {
1462 log_write(0, LOG_MAIN,
1463 "exim %s daemon started: pid=%d, -q%s, not listening for SMTP",
1464 version_string, getpid(), readconf_printtime(queue_interval));
1465 set_process_info("daemon: -q%s, not listening",
1466 readconf_printtime(queue_interval));
1467 }
1468
1469
1470 /* Close the log so it can be renamed and moved. In the few cases below where
1471 this long-running process writes to the log (always exceptional conditions), it
1472 closes the log afterwards, for the same reason. */
1473
1474 log_close_all();
1475
1476 DEBUG(D_any) debug_print_ids(US"daemon running with");
1477
1478 /* Any messages accepted via this route are going to be SMTP. */
1479
1480 smtp_input = TRUE;
1481
1482 /* Enter the never-ending loop... */
1483
1484 for (;;)
1485 {
1486 #if HAVE_IPV6
1487 struct sockaddr_in6 accepted;
1488 #else
1489 struct sockaddr_in accepted;
1490 #endif
1491
1492 SOCKLEN_T len = sizeof(accepted);
1493 int status;
1494 pid_t pid;
1495
1496 /* This code is placed first in the loop, so that it gets obeyed at the
1497 start, before the first wait. This causes the first queue-runner to be
1498 started immediately. */
1499
1500 if (sigalrm_seen)
1501 {
1502 DEBUG(D_any) debug_printf("SIGALRM received\n");
1503
1504 /* Do a full queue run in a child process, if required, unless we already
1505 have enough queue runners on the go. If we are not running as root, a
1506 re-exec is required. */
1507
1508 if (queue_interval > 0 &&
1509 (queue_run_max <= 0 || queue_run_count < queue_run_max))
1510 {
1511 if ((pid = fork()) == 0)
1512 {
1513 int sk;
1514 DEBUG(D_any) debug_printf("Starting queue-runner: pid %d\n",
1515 (int)getpid());
1516
1517 /* Close any open listening sockets in the child */
1518
1519 for (sk = 0; sk < listen_socket_count; sk++) close(listen_sockets[sk]);
1520
1521 /* Reset SIGHUP and SIGCHLD in the child in both cases. */
1522
1523 signal(SIGHUP, SIG_DFL);
1524 signal(SIGCHLD, SIG_DFL);
1525
1526 /* Re-exec if privilege has been given up, unless deliver_drop_
1527 privilege is set. Reset SIGALRM before exec(). */
1528
1529 if (geteuid() != root_uid && !deliver_drop_privilege)
1530 {
1531 uschar opt[8];
1532 uschar *p = opt;
1533
1534 signal(SIGALRM, SIG_DFL);
1535 *p++ = '-';
1536 *p++ = 'q';
1537 if (queue_2stage) *p++ = 'q';
1538 if (queue_run_first_delivery) *p++ = 'i';
1539 if (queue_run_force) *p++ = 'f';
1540 if (deliver_force_thaw) *p++ = 'f';
1541 if (queue_run_local) *p++ = 'l';
1542 *p = 0;
1543
1544 (void)child_exec_exim(CEE_EXEC_PANIC, FALSE, NULL, TRUE, 1, opt);
1545 /* Control never returns here. */
1546 }
1547
1548 /* No need to re-exec; SIGALRM remains set to the default handler */
1549
1550 queue_run(NULL, NULL, FALSE);
1551 _exit(EXIT_SUCCESS);
1552 }
1553
1554 if (pid < 0)
1555 {
1556 log_write(0, LOG_MAIN|LOG_PANIC, "daemon: fork of queue-runner "
1557 "process failed: %s", strerror(errno));
1558 log_close_all();
1559 }
1560 else
1561 {
1562 int i;
1563 for (i = 0; i < queue_run_max; ++i)
1564 {
1565 if (queue_pid_slots[i] <= 0)
1566 {
1567 queue_pid_slots[i] = pid;
1568 queue_run_count++;
1569 break;
1570 }
1571 }
1572 DEBUG(D_any) debug_printf("%d queue-runner process%s running\n",
1573 queue_run_count, (queue_run_count == 1)? "" : "es");
1574 }
1575 }
1576
1577 /* Reset the alarm clock */
1578
1579 sigalrm_seen = FALSE;
1580 alarm(queue_interval);
1581 }
1582
1583
1584 /* Sleep till a connection happens if listening, and handle the connection if
1585 that is why we woke up. The FreeBSD operating system requires the use of
1586 select() before accept() because the latter function is not interrupted by
1587 a signal, and we want to wake up for SIGCHLD and SIGALRM signals. Some other
1588 OS do notice signals in accept() but it does no harm to have the select()
1589 in for all of them - and it won't then be a lurking problem for ports to
1590 new OS. In fact, the later addition of listening on specific interfaces only
1591 requires this way of working anyway. */
1592
1593 if (daemon_listen)
1594 {
1595 int sk, lcount;
1596 int max_socket = 0;
1597 BOOL select_failed = FALSE;
1598 fd_set select_listen;
1599
1600 FD_ZERO(&select_listen);
1601 for (sk = 0; sk < listen_socket_count; sk++)
1602 {
1603 FD_SET(listen_sockets[sk], &select_listen);
1604 if (listen_sockets[sk] > max_socket) max_socket = listen_sockets[sk];
1605 }
1606
1607 DEBUG(D_any) debug_printf("Listening...\n");
1608
1609 if ((lcount = select(max_socket + 1, (SELECT_ARG2_TYPE *)&select_listen,
1610 NULL, NULL, NULL)) < 0)
1611 {
1612 select_failed = TRUE;
1613 lcount = 1;
1614 }
1615
1616 /* Loop for all the sockets that are currently ready to go. If select
1617 actually failed, we have set the count to 1 and a flag, so as to use the
1618 common error code for select/accept below. */
1619
1620 while (lcount-- > 0)
1621 {
1622 int accept_socket = -1;
1623 if (!select_failed)
1624 {
1625 for (sk = 0; sk < listen_socket_count; sk++)
1626 {
1627 if (FD_ISSET(listen_sockets[sk], &select_listen))
1628 {
1629 accept_socket = accept(listen_sockets[sk],
1630 (struct sockaddr *)&accepted, &len);
1631 FD_CLR(listen_sockets[sk], &select_listen);
1632 break;
1633 }
1634 }
1635 }
1636
1637 /* If select or accept has failed and this was not caused by an
1638 interruption, log the incident and try again. With asymmetric TCP/IP
1639 routing errors such as "No route to network" have been seen here. Also
1640 "connection reset by peer" has been seen. These cannot be classed as
1641 disastrous errors, but they could fill up a lot of log. The code in smail
1642 crashes the daemon after 10 successive failures of accept, on the grounds
1643 that some OS fail continuously. Exim originally followed suit, but this
1644 appears to have caused problems. Now it just keeps going, but instead of
1645 logging each error, it batches them up when they are continuous. */
1646
1647 if (accept_socket < 0 && errno != EINTR)
1648 {
1649 if (accept_retry_count == 0)
1650 {
1651 accept_retry_errno = errno;
1652 accept_retry_select_failed = select_failed;
1653 }
1654 else
1655 {
1656 if (errno != accept_retry_errno ||
1657 select_failed != accept_retry_select_failed ||
1658 accept_retry_count >= 50)
1659 {
1660 log_write(0, LOG_MAIN | ((accept_retry_count >= 50)? LOG_PANIC : 0),
1661 "%d %s() failure%s: %s",
1662 accept_retry_count,
1663 accept_retry_select_failed? "select" : "accept",
1664 (accept_retry_count == 1)? "" : "s",
1665 strerror(accept_retry_errno));
1666 log_close_all();
1667 accept_retry_count = 0;
1668 accept_retry_errno = errno;
1669 accept_retry_select_failed = select_failed;
1670 }
1671 }
1672 accept_retry_count++;
1673 }
1674
1675 else
1676 {
1677 if (accept_retry_count > 0)
1678 {
1679 log_write(0, LOG_MAIN, "%d %s() failure%s: %s",
1680 accept_retry_count,
1681 accept_retry_select_failed? "select" : "accept",
1682 (accept_retry_count == 1)? "" : "s",
1683 strerror(accept_retry_errno));
1684 log_close_all();
1685 accept_retry_count = 0;
1686 }
1687 }
1688
1689 /* If select/accept succeeded, deal with the connection. */
1690
1691 if (accept_socket >= 0)
1692 handle_smtp_call(listen_sockets, listen_socket_count, accept_socket,
1693 (struct sockaddr *)&accepted);
1694 }
1695 }
1696
1697 /* If not listening, then just sleep for the queue interval. If we woke
1698 up early the last time for some other signal, it won't matter because
1699 the alarm signal will wake at the right time. This code originally used
1700 sleep() but it turns out that on the FreeBSD system, sleep() is not inter-
1701 rupted by signals, so it wasn't waking up for SIGALRM or SIGCHLD. Luckily
1702 select() can be used as an interruptible sleep() on all versions of Unix. */
1703
1704 else
1705 {
1706 struct timeval tv;
1707 tv.tv_sec = queue_interval;
1708 tv.tv_usec = 0;
1709 select(0, NULL, NULL, NULL, &tv);
1710 }
1711
1712 /* Handle the termination of a child process. Theoretically, this need
1713 be done only when sigchld_seen is TRUE, but rumour has it that some systems
1714 lose SIGCHLD signals at busy times, so to be on the safe side, just
1715 do it each time round. It shouldn't be too expensive. */
1716
1717 while ((pid = waitpid(-1, &status, WNOHANG)) > 0)
1718 {
1719 int i;
1720 DEBUG(D_any) debug_printf("child %d ended: status=0x%x\n", (int)pid,
1721 status);
1722
1723 /* If it's a listening daemon, deal with an accepting process that has
1724 terminated. */
1725
1726 if (daemon_listen)
1727 {
1728 for (i = 0; i < smtp_accept_max; i++)
1729 {
1730 if (smtp_slots[i].pid == pid)
1731 {
1732 if (smtp_slots[i].host_address != NULL)
1733 store_free(smtp_slots[i].host_address);
1734 smtp_slots[i] = empty_smtp_slot;
1735 if (--smtp_accept_count < 0) smtp_accept_count = 0;
1736 DEBUG(D_any) debug_printf("%d SMTP accept process%s now running\n",
1737 smtp_accept_count, (smtp_accept_count == 1)? "" : "es");
1738 break;
1739 }
1740 }
1741 if (i < smtp_accept_max) continue; /* Found an accepting process */
1742 }
1743
1744 /* If it wasn't an accepting process, see if it was a queue-runner
1745 process, if we are keeping track of them. */
1746
1747 if (queue_interval > 0)
1748 {
1749 for (i = 0; i < queue_run_max; i++)
1750 {
1751 if (queue_pid_slots[i] == pid)
1752 {
1753 queue_pid_slots[i] = 0;
1754 if (--queue_run_count < 0) queue_run_count = 0;
1755 DEBUG(D_any) debug_printf("%d queue-runner process%s now running\n",
1756 queue_run_count, (queue_run_count == 1)? "" : "es");
1757 break;
1758 }
1759 }
1760 }
1761 }
1762
1763 /* Re-enable the SIGCHLD handler if it has been run. It can't do it
1764 for itself, because it isn't doing the waiting itself. */
1765
1766 if (sigchld_seen)
1767 {
1768 sigchld_seen = FALSE;
1769 signal(SIGCHLD, main_sigchld_handler);
1770 }
1771
1772 /* Handle being woken by SIGHUP. We know at this point that the result
1773 of accept() has been dealt with, so we can re-exec exim safely, first
1774 closing the listening sockets so that they can be reused. Cancel any pending
1775 alarm in case it is just about to go off, and set SIGHUP to be ignored so
1776 that another HUP in quick succession doesn't clobber the new daemon before it
1777 gets going. All log files get closed by the close-on-exec flag; however, if
1778 the exec fails, we need to close the logs. */
1779
1780 if (sighup_seen)
1781 {
1782 int sk;
1783 log_write(0, LOG_MAIN, "pid %d: SIGHUP received: re-exec daemon",
1784 getpid());
1785 for (sk = 0; sk < listen_socket_count; sk++) close(listen_sockets[sk]);
1786 alarm(0);
1787 signal(SIGHUP, SIG_IGN);
1788 sighup_argv[0] = exim_path;
1789 exim_nullstd();
1790 execv(CS exim_path, (char *const *)sighup_argv);
1791 log_write(0, LOG_MAIN|LOG_PANIC_DIE, "pid %d: exec of %s failed: %s",
1792 getpid(), exim_path, strerror(errno));
1793 log_close_all();
1794 }
1795
1796 } /* End of main loop */
1797
1798 /* Control never reaches here */
1799 }
1800
1801 /* End of exim_daemon.c */
1802