27b4cb26567045add35870477d9f3f30e21a3714
[exim.git] / src / src / daemon.c
1 /*************************************************
2 * Exim - an Internet mail transport agent *
3 *************************************************/
4
5 /* Copyright (c) University of Cambridge 1995 - 2009 */
6 /* See the file NOTICE for conditions of use and distribution. */
7
8 /* Functions concerned with running Exim as a daemon */
9
10
11 #include "exim.h"
12
13
14 /* Structure for holding data for each SMTP connection */
15
16 typedef struct smtp_slot {
17 pid_t pid; /* pid of the spawned reception process */
18 uschar *host_address; /* address of the client host */
19 } smtp_slot;
20
21 /* An empty slot for initializing (Standard C does not allow constructor
22 expressions in assigments except as initializers in declarations). */
23
24 static smtp_slot empty_smtp_slot = { 0, NULL };
25
26
27
28 /*************************************************
29 * Local static variables *
30 *************************************************/
31
32 static SIGNAL_BOOL sigchld_seen;
33 static SIGNAL_BOOL sighup_seen;
34
35 static int accept_retry_count = 0;
36 static int accept_retry_errno;
37 static BOOL accept_retry_select_failed;
38
39 static int queue_run_count = 0;
40 static pid_t *queue_pid_slots = NULL;
41 static smtp_slot *smtp_slots = NULL;
42
43 static BOOL write_pid = TRUE;
44
45
46
47 /*************************************************
48 * SIGHUP Handler *
49 *************************************************/
50
51 /* All this handler does is to set a flag and re-enable the signal.
52
53 Argument: the signal number
54 Returns: nothing
55 */
56
57 static void
58 sighup_handler(int sig)
59 {
60 sig = sig; /* Keep picky compilers happy */
61 sighup_seen = TRUE;
62 signal(SIGHUP, sighup_handler);
63 }
64
65
66
67 /*************************************************
68 * SIGCHLD handler for main daemon process *
69 *************************************************/
70
71 /* Don't re-enable the handler here, since we aren't doing the
72 waiting here. If the signal is re-enabled, there will just be an
73 infinite sequence of calls to this handler. The SIGCHLD signal is
74 used just as a means of waking up the daemon so that it notices
75 terminated subprocesses as soon as possible.
76
77 Argument: the signal number
78 Returns: nothing
79 */
80
81 static void
82 main_sigchld_handler(int sig)
83 {
84 sig = sig; /* Keep picky compilers happy */
85 os_non_restarting_signal(SIGCHLD, SIG_DFL);
86 sigchld_seen = TRUE;
87 }
88
89
90
91
92 /*************************************************
93 * Unexpected errors in SMTP calls *
94 *************************************************/
95
96 /* This function just saves a bit of repetitious coding.
97
98 Arguments:
99 log_msg Text of message to be logged
100 smtp_msg Text of SMTP error message
101 was_errno The failing errno
102
103 Returns: nothing
104 */
105
106 static void
107 never_error(uschar *log_msg, uschar *smtp_msg, int was_errno)
108 {
109 uschar *emsg = (was_errno <= 0)? US"" :
110 string_sprintf(": %s", strerror(was_errno));
111 log_write(0, LOG_MAIN|LOG_PANIC, "%s%s", log_msg, emsg);
112 if (smtp_out != NULL) smtp_printf("421 %s\r\n", smtp_msg);
113 }
114
115
116
117
118 /*************************************************
119 * Handle a connected SMTP call *
120 *************************************************/
121
122 /* This function is called when an SMTP connection has been accepted.
123 If there are too many, give an error message and close down. Otherwise
124 spin off a sub-process to handle the call. The list of listening sockets
125 is required so that they can be closed in the sub-process. Take care not to
126 leak store in this process - reset the stacking pool at the end.
127
128 Arguments:
129 listen_sockets sockets which are listening for incoming calls
130 listen_socket_count count of listening sockets
131 accept_socket socket of the current accepted call
132 accepted socket information about the current call
133
134 Returns: nothing
135 */
136
137 static void
138 handle_smtp_call(int *listen_sockets, int listen_socket_count,
139 int accept_socket, struct sockaddr *accepted)
140 {
141 pid_t pid;
142 union sockaddr_46 interface_sockaddr;
143 EXIM_SOCKLEN_T ifsize = sizeof(interface_sockaddr);
144 int dup_accept_socket = -1;
145 int max_for_this_host = 0;
146 int wfsize = 0;
147 int wfptr = 0;
148 int use_log_write_selector = log_write_selector;
149 uschar *whofrom = NULL;
150
151 void *reset_point = store_get(0);
152
153 /* Make the address available in ASCII representation, and also fish out
154 the remote port. */
155
156 sender_host_address = host_ntoa(-1, accepted, NULL, &sender_host_port);
157 DEBUG(D_any) debug_printf("Connection request from %s port %d\n",
158 sender_host_address, sender_host_port);
159
160 /* Set up the output stream, check the socket has duplicated, and set up the
161 input stream. These operations fail only the exceptional circumstances. Note
162 that never_error() won't use smtp_out if it is NULL. */
163
164 smtp_out = fdopen(accept_socket, "wb");
165 if (smtp_out == NULL)
166 {
167 never_error(US"daemon: fdopen() for smtp_out failed", US"", errno);
168 goto ERROR_RETURN;
169 }
170
171 dup_accept_socket = dup(accept_socket);
172 if (dup_accept_socket < 0)
173 {
174 never_error(US"daemon: couldn't dup socket descriptor",
175 US"Connection setup failed", errno);
176 goto ERROR_RETURN;
177 }
178
179 smtp_in = fdopen(dup_accept_socket, "rb");
180 if (smtp_in == NULL)
181 {
182 never_error(US"daemon: fdopen() for smtp_in failed",
183 US"Connection setup failed", errno);
184 goto ERROR_RETURN;
185 }
186
187 /* Get the data for the local interface address. Panic for most errors, but
188 "connection reset by peer" just means the connection went away. */
189
190 if (getsockname(accept_socket, (struct sockaddr *)(&interface_sockaddr),
191 &ifsize) < 0)
192 {
193 log_write(0, LOG_MAIN | ((errno == ECONNRESET)? 0 : LOG_PANIC),
194 "getsockname() failed: %s", strerror(errno));
195 smtp_printf("421 Local problem: getsockname() failed; please try again later\r\n");
196 goto ERROR_RETURN;
197 }
198
199 interface_address = host_ntoa(-1, &interface_sockaddr, NULL, &interface_port);
200 DEBUG(D_interface) debug_printf("interface address=%s port=%d\n",
201 interface_address, interface_port);
202
203 /* Build a string identifying the remote host and, if requested, the port and
204 the local interface data. This is for logging; at the end of this function the
205 memory is reclaimed. */
206
207 whofrom = string_append(whofrom, &wfsize, &wfptr, 3, "[", sender_host_address, "]");
208
209 if ((log_extra_selector & LX_incoming_port) != 0)
210 whofrom = string_append(whofrom, &wfsize, &wfptr, 2, ":", string_sprintf("%d",
211 sender_host_port));
212
213 if ((log_extra_selector & LX_incoming_interface) != 0)
214 whofrom = string_append(whofrom, &wfsize, &wfptr, 4, " I=[",
215 interface_address, "]:", string_sprintf("%d", interface_port));
216
217 whofrom[wfptr] = 0; /* Terminate the newly-built string */
218
219 /* Check maximum number of connections. We do not check for reserved
220 connections or unacceptable hosts here. That is done in the subprocess because
221 it might take some time. */
222
223 if (smtp_accept_max > 0 && smtp_accept_count >= smtp_accept_max)
224 {
225 DEBUG(D_any) debug_printf("rejecting SMTP connection: count=%d max=%d\n",
226 smtp_accept_count, smtp_accept_max);
227 smtp_printf("421 Too many concurrent SMTP connections; "
228 "please try again later.\r\n");
229 log_write(L_connection_reject,
230 LOG_MAIN, "Connection from %s refused: too many connections",
231 whofrom);
232 goto ERROR_RETURN;
233 }
234
235 /* If a load limit above which only reserved hosts are acceptable is defined,
236 get the load average here, and if there are in fact no reserved hosts, do
237 the test right away (saves a fork). If there are hosts, do the check in the
238 subprocess because it might take time. */
239
240 if (smtp_load_reserve >= 0)
241 {
242 load_average = OS_GETLOADAVG();
243 if (smtp_reserve_hosts == NULL && load_average > smtp_load_reserve)
244 {
245 DEBUG(D_any) debug_printf("rejecting SMTP connection: load average = %.2f\n",
246 (double)load_average/1000.0);
247 smtp_printf("421 Too much load; please try again later.\r\n");
248 log_write(L_connection_reject,
249 LOG_MAIN, "Connection from %s refused: load average = %.2f",
250 whofrom, (double)load_average/1000.0);
251 goto ERROR_RETURN;
252 }
253 }
254
255 /* Check that one specific host (strictly, IP address) is not hogging
256 resources. This is done here to prevent a denial of service attack by someone
257 forcing you to fork lots of times before denying service. The value of
258 smtp_accept_max_per_host is a string which is expanded. This makes it possible
259 to provide host-specific limits according to $sender_host address, but because
260 this is in the daemon mainline, only fast expansions (such as inline address
261 checks) should be used. The documentation is full of warnings. */
262
263 if (smtp_accept_max_per_host != NULL)
264 {
265 uschar *expanded = expand_string(smtp_accept_max_per_host);
266 if (expanded == NULL)
267 {
268 if (!expand_string_forcedfail)
269 log_write(0, LOG_MAIN|LOG_PANIC, "expansion of smtp_accept_max_per_host "
270 "failed for %s: %s", whofrom, expand_string_message);
271 }
272 /* For speed, interpret a decimal number inline here */
273 else
274 {
275 uschar *s = expanded;
276 while (isdigit(*s))
277 max_for_this_host = max_for_this_host * 10 + *s++ - '0';
278 if (*s != 0)
279 log_write(0, LOG_MAIN|LOG_PANIC, "expansion of smtp_accept_max_per_host "
280 "for %s contains non-digit: %s", whofrom, expanded);
281 }
282 }
283
284 /* If we have fewer connections than max_for_this_host, we can skip the tedious
285 per host_address checks. Note that at this stage smtp_accept_count contains the
286 count of *other* connections, not including this one. */
287
288 if ((max_for_this_host > 0) &&
289 (smtp_accept_count >= max_for_this_host))
290 {
291 int i;
292 int host_accept_count = 0;
293 int other_host_count = 0; /* keep a count of non matches to optimise */
294
295 for (i = 0; i < smtp_accept_max; ++i)
296 {
297 if (smtp_slots[i].host_address != NULL)
298 {
299 if (Ustrcmp(sender_host_address, smtp_slots[i].host_address) == 0)
300 host_accept_count++;
301 else
302 other_host_count++;
303
304 /* Testing all these strings is expensive - see if we can drop out
305 early, either by hitting the target, or finding there are not enough
306 connections left to make the target. */
307
308 if ((host_accept_count >= max_for_this_host) ||
309 ((smtp_accept_count - other_host_count) < max_for_this_host))
310 break;
311 }
312 }
313
314 if (host_accept_count >= max_for_this_host)
315 {
316 DEBUG(D_any) debug_printf("rejecting SMTP connection: too many from this "
317 "IP address: count=%d max=%d\n",
318 host_accept_count, max_for_this_host);
319 smtp_printf("421 Too many concurrent SMTP connections "
320 "from this IP address; please try again later.\r\n");
321 log_write(L_connection_reject,
322 LOG_MAIN, "Connection from %s refused: too many connections "
323 "from that IP address", whofrom);
324 goto ERROR_RETURN;
325 }
326 }
327
328 /* OK, the connection count checks have been passed. Before we can fork the
329 accepting process, we must first log the connection if requested. This logging
330 used to happen in the subprocess, but doing that means that the value of
331 smtp_accept_count can be out of step by the time it is logged. So we have to do
332 the logging here and accept the performance cost. Note that smtp_accept_count
333 hasn't yet been incremented to take account of this connection.
334
335 In order to minimize the cost (because this is going to happen for every
336 connection), do a preliminary selector test here. This saves ploughing through
337 the generalized logging code each time when the selector is false. If the
338 selector is set, check whether the host is on the list for logging. If not,
339 arrange to unset the selector in the subprocess. */
340
341 if ((log_write_selector & L_smtp_connection) != 0)
342 {
343 uschar *list = hosts_connection_nolog;
344 if (list != NULL && verify_check_host(&list) == OK)
345 use_log_write_selector &= ~L_smtp_connection;
346 else
347 log_write(L_smtp_connection, LOG_MAIN, "SMTP connection from %s "
348 "(TCP/IP connection count = %d)", whofrom, smtp_accept_count + 1);
349 }
350
351 /* Now we can fork the accepting process; do a lookup tidy, just in case any
352 expansion above did a lookup. */
353
354 search_tidyup();
355 pid = fork();
356
357 /* Handle the child process */
358
359 if (pid == 0)
360 {
361 int i;
362 int queue_only_reason = 0;
363 int old_pool = store_pool;
364 int save_debug_selector = debug_selector;
365 BOOL local_queue_only;
366 BOOL session_local_queue_only;
367 #ifdef SA_NOCLDWAIT
368 struct sigaction act;
369 #endif
370
371 smtp_accept_count++; /* So that it includes this process */
372
373 /* May have been modified for the subprocess */
374
375 log_write_selector = use_log_write_selector;
376
377 /* Get the local interface address into permanent store */
378
379 store_pool = POOL_PERM;
380 interface_address = string_copy(interface_address);
381 store_pool = old_pool;
382
383 /* Check for a tls-on-connect port */
384
385 if (host_is_tls_on_connect_port(interface_port)) tls_on_connect = TRUE;
386
387 /* Expand smtp_active_hostname if required. We do not do this any earlier,
388 because it may depend on the local interface address (indeed, that is most
389 likely what it depends on.) */
390
391 smtp_active_hostname = primary_hostname;
392 if (raw_active_hostname != NULL)
393 {
394 uschar *nah = expand_string(raw_active_hostname);
395 if (nah == NULL)
396 {
397 if (!expand_string_forcedfail)
398 {
399 log_write(0, LOG_MAIN|LOG_PANIC, "failed to expand \"%s\" "
400 "(smtp_active_hostname): %s", raw_active_hostname,
401 expand_string_message);
402 smtp_printf("421 Local configuration error; "
403 "please try again later.\r\n");
404 mac_smtp_fflush();
405 search_tidyup();
406 _exit(EXIT_FAILURE);
407 }
408 }
409 else if (nah[0] != 0) smtp_active_hostname = nah;
410 }
411
412 /* Initialize the queueing flags */
413
414 queue_check_only();
415 session_local_queue_only = queue_only;
416
417 /* Close the listening sockets, and set the SIGCHLD handler to SIG_IGN.
418 We also attempt to set things up so that children are automatically reaped,
419 but just in case this isn't available, there's a paranoid waitpid() in the
420 loop too (except for systems where we are sure it isn't needed). See the more
421 extensive comment before the reception loop in exim.c for a fuller
422 explanation of this logic. */
423
424 for (i = 0; i < listen_socket_count; i++) (void)close(listen_sockets[i]);
425
426 /* Set FD_CLOEXEC on the SMTP socket. We don't want any rogue child processes
427 to be able to communicate with them, under any circumstances. */
428 (void)fcntl(accept_socket, F_SETFD,
429 fcntl(accept_socket, F_GETFD) | FD_CLOEXEC);
430 (void)fcntl(dup_accept_socket, F_SETFD,
431 fcntl(dup_accept_socket, F_GETFD) | FD_CLOEXEC);
432
433 #ifdef SA_NOCLDWAIT
434 act.sa_handler = SIG_IGN;
435 sigemptyset(&(act.sa_mask));
436 act.sa_flags = SA_NOCLDWAIT;
437 sigaction(SIGCHLD, &act, NULL);
438 #else
439 signal(SIGCHLD, SIG_IGN);
440 #endif
441
442 /* Attempt to get an id from the sending machine via the RFC 1413
443 protocol. We do this in the sub-process in order not to hold up the
444 main process if there is any delay. Then set up the fullhost information
445 in case there is no HELO/EHLO.
446
447 If debugging is enabled only for the daemon, we must turn if off while
448 finding the id, but turn it on again afterwards so that information about the
449 incoming connection is output. */
450
451 if (debug_daemon) debug_selector = 0;
452 verify_get_ident(IDENT_PORT);
453 host_build_sender_fullhost();
454 debug_selector = save_debug_selector;
455
456 DEBUG(D_any)
457 debug_printf("Process %d is handling incoming connection from %s\n",
458 (int)getpid(), sender_fullhost);
459
460 /* Now disable debugging permanently if it's required only for the daemon
461 process. */
462
463 if (debug_daemon) debug_selector = 0;
464
465 /* If there are too many child processes for immediate delivery,
466 set the session_local_queue_only flag, which is initialized from the
467 configured value and may therefore already be TRUE. Leave logging
468 till later so it will have a message id attached. Note that there is no
469 possibility of re-calculating this per-message, because the value of
470 smtp_accept_count does not change in this subprocess. */
471
472 if (smtp_accept_queue > 0 && smtp_accept_count > smtp_accept_queue)
473 {
474 session_local_queue_only = TRUE;
475 queue_only_reason = 1;
476 }
477
478 /* Handle the start of the SMTP session, then loop, accepting incoming
479 messages from the SMTP connection. The end will come at the QUIT command,
480 when smtp_setup_msg() returns 0. A break in the connection causes the
481 process to die (see accept.c).
482
483 NOTE: We do *not* call smtp_log_no_mail() if smtp_start_session() fails,
484 because a log line has already been written for all its failure exists
485 (usually "connection refused: <reason>") and writing another one is
486 unnecessary clutter. */
487
488 if (!smtp_start_session())
489 {
490 mac_smtp_fflush();
491 search_tidyup();
492 _exit(EXIT_SUCCESS);
493 }
494
495 for (;;)
496 {
497 int rc;
498 message_id[0] = 0; /* Clear out any previous message_id */
499 reset_point = store_get(0); /* Save current store high water point */
500
501 DEBUG(D_any)
502 debug_printf("Process %d is ready for new message\n", (int)getpid());
503
504 /* Smtp_setup_msg() returns 0 on QUIT or if the call is from an
505 unacceptable host or if an ACL "drop" command was triggered, -1 on
506 connection lost, and +1 on validly reaching DATA. Receive_msg() almost
507 always returns TRUE when smtp_input is true; just retry if no message was
508 accepted (can happen for invalid message parameters). However, it can yield
509 FALSE if the connection was forcibly dropped by the DATA ACL. */
510
511 if ((rc = smtp_setup_msg()) > 0)
512 {
513 BOOL ok = receive_msg(FALSE);
514 search_tidyup(); /* Close cached databases */
515 if (!ok) /* Connection was dropped */
516 {
517 mac_smtp_fflush();
518 smtp_log_no_mail(); /* Log no mail if configured */
519 _exit(EXIT_SUCCESS);
520 }
521 if (message_id[0] == 0) continue; /* No message was accepted */
522 }
523 else
524 {
525 mac_smtp_fflush();
526 search_tidyup();
527 smtp_log_no_mail(); /* Log no mail if configured */
528 _exit((rc == 0)? EXIT_SUCCESS : EXIT_FAILURE);
529 }
530
531 /* Show the recipients when debugging */
532
533 DEBUG(D_receive)
534 {
535 int i;
536 if (sender_address != NULL)
537 debug_printf("Sender: %s\n", sender_address);
538 if (recipients_list != NULL)
539 {
540 debug_printf("Recipients:\n");
541 for (i = 0; i < recipients_count; i++)
542 debug_printf(" %s\n", recipients_list[i].address);
543 }
544 }
545
546 /* A message has been accepted. Clean up any previous delivery processes
547 that have completed and are defunct, on systems where they don't go away
548 by themselves (see comments when setting SIG_IGN above). On such systems
549 (if any) these delivery processes hang around after termination until
550 the next message is received. */
551
552 #ifndef SIG_IGN_WORKS
553 while (waitpid(-1, NULL, WNOHANG) > 0);
554 #endif
555
556 /* Reclaim up the store used in accepting this message */
557
558 store_reset(reset_point);
559
560 /* If queue_only is set or if there are too many incoming connections in
561 existence, session_local_queue_only will be TRUE. If it is not, check
562 whether we have received too many messages in this session for immediate
563 delivery. */
564
565 if (!session_local_queue_only &&
566 smtp_accept_queue_per_connection > 0 &&
567 receive_messagecount > smtp_accept_queue_per_connection)
568 {
569 session_local_queue_only = TRUE;
570 queue_only_reason = 2;
571 }
572
573 /* Initialize local_queue_only from session_local_queue_only. If it is not
574 true, and queue_only_load is set, check that the load average is below it.
575 If local_queue_only is set by this means, we also set if for the session if
576 queue_only_load_latch is true (the default). This means that, once set,
577 local_queue_only remains set for any subsequent messages on the same SMTP
578 connection. This is a deliberate choice; even though the load average may
579 fall, it doesn't seem right to deliver later messages on the same call when
580 not delivering earlier ones. However, the are special circumstances such as
581 very long-lived connections from scanning appliances where this is not the
582 best strategy. In such cases, queue_only_load_latch should be set false. */
583
584 local_queue_only = session_local_queue_only;
585 if (!local_queue_only && queue_only_load >= 0)
586 {
587 local_queue_only = (load_average = OS_GETLOADAVG()) > queue_only_load;
588 if (local_queue_only)
589 {
590 queue_only_reason = 3;
591 if (queue_only_load_latch) session_local_queue_only = TRUE;
592 }
593 }
594
595 /* Log the queueing here, when it will get a message id attached, but
596 not if queue_only is set (case 0). */
597
598 if (local_queue_only) switch(queue_only_reason)
599 {
600 case 1:
601 log_write(L_delay_delivery,
602 LOG_MAIN, "no immediate delivery: too many connections "
603 "(%d, max %d)", smtp_accept_count, smtp_accept_queue);
604 break;
605
606 case 2:
607 log_write(L_delay_delivery,
608 LOG_MAIN, "no immediate delivery: more than %d messages "
609 "received in one connection", smtp_accept_queue_per_connection);
610 break;
611
612 case 3:
613 log_write(L_delay_delivery,
614 LOG_MAIN, "no immediate delivery: load average %.2f",
615 (double)load_average/1000.0);
616 break;
617 }
618
619 /* If a delivery attempt is required, spin off a new process to handle it.
620 If we are not root, we have to re-exec exim unless deliveries are being
621 done unprivileged. */
622
623 else if (!queue_only_policy && !deliver_freeze)
624 {
625 pid_t dpid;
626
627 /* Before forking, ensure that the C output buffer is flushed. Otherwise
628 anything that it in it will get duplicated, leading to duplicate copies
629 of the pending output. */
630
631 mac_smtp_fflush();
632
633 if ((dpid = fork()) == 0)
634 {
635 (void)fclose(smtp_in);
636 (void)fclose(smtp_out);
637
638 /* Don't ever molest the parent's SSL connection, but do clean up
639 the data structures if necessary. */
640
641 #ifdef SUPPORT_TLS
642 tls_close(FALSE);
643 #endif
644
645 /* Reset SIGHUP and SIGCHLD in the child in both cases. */
646
647 signal(SIGHUP, SIG_DFL);
648 signal(SIGCHLD, SIG_DFL);
649
650 if (geteuid() != root_uid && !deliver_drop_privilege)
651 {
652 signal(SIGALRM, SIG_DFL);
653 (void)child_exec_exim(CEE_EXEC_PANIC, FALSE, NULL, FALSE, 2, US"-Mc",
654 message_id);
655 /* Control does not return here. */
656 }
657
658 /* No need to re-exec; SIGALRM remains set to the default handler */
659
660 (void)deliver_message(message_id, FALSE, FALSE);
661 search_tidyup();
662 _exit(EXIT_SUCCESS);
663 }
664
665 if (dpid > 0)
666 {
667 DEBUG(D_any) debug_printf("forked delivery process %d\n", (int)dpid);
668 }
669 else
670 {
671 log_write(0, LOG_MAIN|LOG_PANIC, "daemon: delivery process fork "
672 "failed: %s", strerror(errno));
673 }
674 }
675 }
676 }
677
678
679 /* Carrying on in the parent daemon process... Can't do much if the fork
680 failed. Otherwise, keep count of the number of accepting processes and
681 remember the pid for ticking off when the child completes. */
682
683 if (pid < 0)
684 {
685 never_error(US"daemon: accept process fork failed", US"Fork failed", errno);
686 }
687 else
688 {
689 int i;
690 for (i = 0; i < smtp_accept_max; ++i)
691 {
692 if (smtp_slots[i].pid <= 0)
693 {
694 smtp_slots[i].pid = pid;
695 if (smtp_accept_max_per_host != NULL)
696 smtp_slots[i].host_address = string_copy_malloc(sender_host_address);
697 smtp_accept_count++;
698 break;
699 }
700 }
701 DEBUG(D_any) debug_printf("%d SMTP accept process%s running\n",
702 smtp_accept_count, (smtp_accept_count == 1)? "" : "es");
703 }
704
705 /* Get here via goto in error cases */
706
707 ERROR_RETURN:
708
709 /* Close the streams associated with the socket which will also close the
710 socket fds in this process. We can't do anything if fclose() fails, but
711 logging brings it to someone's attention. However, "connection reset by peer"
712 isn't really a problem, so skip that one. On Solaris, a dropped connection can
713 manifest itself as a broken pipe, so drop that one too. If the streams don't
714 exist, something went wrong while setting things up. Make sure the socket
715 descriptors are closed, in order to drop the connection. */
716
717 if (smtp_out != NULL)
718 {
719 if (fclose(smtp_out) != 0 && errno != ECONNRESET && errno != EPIPE)
720 log_write(0, LOG_MAIN|LOG_PANIC, "daemon: fclose(smtp_out) failed: %s",
721 strerror(errno));
722 smtp_out = NULL;
723 }
724 else (void)close(accept_socket);
725
726 if (smtp_in != NULL)
727 {
728 if (fclose(smtp_in) != 0 && errno != ECONNRESET && errno != EPIPE)
729 log_write(0, LOG_MAIN|LOG_PANIC, "daemon: fclose(smtp_in) failed: %s",
730 strerror(errno));
731 smtp_in = NULL;
732 }
733 else (void)close(dup_accept_socket);
734
735 /* Release any store used in this process, including the store used for holding
736 the incoming host address and an expanded active_hostname. */
737
738 store_reset(reset_point);
739 sender_host_address = NULL;
740 }
741
742
743
744
745 /*************************************************
746 * Check wildcard listen special cases *
747 *************************************************/
748
749 /* This function is used when binding and listening on lists of addresses and
750 ports. It tests for special cases of wildcard listening, when IPv4 and IPv6
751 sockets may interact in different ways in different operating systems. It is
752 passed an error number, the list of listening addresses, and the current
753 address. Two checks are available: for a previous wildcard IPv6 address, or for
754 a following wildcard IPv4 address, in both cases on the same port.
755
756 In practice, pairs of wildcard addresses should be adjacent in the address list
757 because they are sorted that way below.
758
759 Arguments:
760 eno the error number
761 addresses the list of addresses
762 ipa the current IP address
763 back if TRUE, check for previous wildcard IPv6 address
764 if FALSE, check for a following wildcard IPv4 address
765
766 Returns: TRUE or FALSE
767 */
768
769 static BOOL
770 check_special_case(int eno, ip_address_item *addresses, ip_address_item *ipa,
771 BOOL back)
772 {
773 ip_address_item *ipa2;
774
775 /* For the "back" case, if the failure was "address in use" for a wildcard IPv4
776 address, seek a previous IPv6 wildcard address on the same port. As it is
777 previous, it must have been successfully bound and be listening. Flag it as a
778 "6 including 4" listener. */
779
780 if (back)
781 {
782 if (eno != EADDRINUSE || ipa->address[0] != 0) return FALSE;
783 for (ipa2 = addresses; ipa2 != ipa; ipa2 = ipa2->next)
784 {
785 if (ipa2->address[1] == 0 && ipa2->port == ipa->port)
786 {
787 ipa2->v6_include_v4 = TRUE;
788 return TRUE;
789 }
790 }
791 }
792
793 /* For the "forward" case, if the current address is a wildcard IPv6 address,
794 we seek a following wildcard IPv4 address on the same port. */
795
796 else
797 {
798 if (ipa->address[0] != ':' || ipa->address[1] != 0) return FALSE;
799 for (ipa2 = ipa->next; ipa2 != NULL; ipa2 = ipa2->next)
800 if (ipa2->address[0] == 0 && ipa->port == ipa2->port) return TRUE;
801 }
802
803 return FALSE;
804 }
805
806
807
808
809 /*************************************************
810 * Handle terminating subprocesses *
811 *************************************************/
812
813 /* Handle the termination of child processes. Theoretically, this need be done
814 only when sigchld_seen is TRUE, but rumour has it that some systems lose
815 SIGCHLD signals at busy times, so to be on the safe side, this function is
816 called each time round. It shouldn't be too expensive.
817
818 Arguments: none
819 Returns: nothing
820 */
821
822 static void
823 handle_ending_processes(void)
824 {
825 int status;
826 pid_t pid;
827
828 while ((pid = waitpid(-1, &status, WNOHANG)) > 0)
829 {
830 int i;
831 DEBUG(D_any)
832 {
833 debug_printf("child %d ended: status=0x%x\n", (int)pid, status);
834 #ifdef WCOREDUMP
835 if (WIFEXITED(status))
836 debug_printf(" normal exit, %d\n", WEXITSTATUS(status));
837 else if (WIFSIGNALED(status))
838 debug_printf(" signal exit, signal %d%s\n", WTERMSIG(status),
839 WCOREDUMP(status) ? " (core dumped)" : "");
840 #endif
841 }
842
843 /* If it's a listening daemon for which we are keeping track of individual
844 subprocesses, deal with an accepting process that has terminated. */
845
846 if (smtp_slots != NULL)
847 {
848 for (i = 0; i < smtp_accept_max; i++)
849 {
850 if (smtp_slots[i].pid == pid)
851 {
852 if (smtp_slots[i].host_address != NULL)
853 store_free(smtp_slots[i].host_address);
854 smtp_slots[i] = empty_smtp_slot;
855 if (--smtp_accept_count < 0) smtp_accept_count = 0;
856 DEBUG(D_any) debug_printf("%d SMTP accept process%s now running\n",
857 smtp_accept_count, (smtp_accept_count == 1)? "" : "es");
858 break;
859 }
860 }
861 if (i < smtp_accept_max) continue; /* Found an accepting process */
862 }
863
864 /* If it wasn't an accepting process, see if it was a queue-runner
865 process that we are tracking. */
866
867 if (queue_pid_slots != NULL)
868 {
869 for (i = 0; i < queue_run_max; i++)
870 {
871 if (queue_pid_slots[i] == pid)
872 {
873 queue_pid_slots[i] = 0;
874 if (--queue_run_count < 0) queue_run_count = 0;
875 DEBUG(D_any) debug_printf("%d queue-runner process%s now running\n",
876 queue_run_count, (queue_run_count == 1)? "" : "es");
877 break;
878 }
879 }
880 }
881 }
882 }
883
884
885
886 /*************************************************
887 * Exim Daemon Mainline *
888 *************************************************/
889
890 /* The daemon can do two jobs, either of which is optional:
891
892 (1) Listens for incoming SMTP calls and spawns off a sub-process to handle
893 each one. This is requested by the -bd option, with -oX specifying the SMTP
894 port on which to listen (for testing).
895
896 (2) Spawns a queue-running process every so often. This is controlled by the
897 -q option with a an interval time. (If no time is given, a single queue run
898 is done from the main function, and control doesn't get here.)
899
900 Root privilege is required in order to attach to port 25. Some systems require
901 it when calling socket() rather than bind(). To cope with all cases, we run as
902 root for both socket() and bind(). Some systems also require root in order to
903 write to the pid file directory. This function must therefore be called as root
904 if it is to work properly in all circumstances. Once the socket is bound and
905 the pid file written, root privilege is given up if there is an exim uid.
906
907 There are no arguments to this function, and it never returns. */
908
909 void
910 daemon_go(void)
911 {
912 struct passwd *pw;
913 int *listen_sockets = NULL;
914 int listen_socket_count = 0;
915 ip_address_item *addresses = NULL;
916
917 /* If any debugging options are set, turn on the D_pid bit so that all
918 debugging lines get the pid added. */
919
920 DEBUG(D_any|D_v) debug_selector |= D_pid;
921
922
923 /* Do the preparation for setting up a listener on one or more interfaces, and
924 possible on various ports. This is controlled by the combination of
925 local_interfaces (which can set IP addresses and ports) and daemon_smtp_port
926 (which is a list of default ports to use for those items in local_interfaces
927 that do not specify a port). The -oX command line option can be used to
928 override one or both of these options.
929
930 If local_interfaces is not set, the default is to listen on all interfaces.
931 When it is set, it can include "all IPvx interfaces" as an item. This is useful
932 when different ports are in use.
933
934 It turns out that listening on all interfaces is messy in an IPv6 world,
935 because several different implementation approaches have been taken. This code
936 is now supposed to work with all of them. The point of difference is whether an
937 IPv6 socket that is listening on all interfaces will receive incoming IPv4
938 calls or not. We also have to cope with the case when IPv6 libraries exist, but
939 there is no IPv6 support in the kernel.
940
941 . On Solaris, an IPv6 socket will accept IPv4 calls, and give them as mapped
942 addresses. However, if an IPv4 socket is also listening on all interfaces,
943 calls are directed to the appropriate socket.
944
945 . On (some versions of) Linux, an IPv6 socket will accept IPv4 calls, and
946 give them as mapped addresses, but an attempt also to listen on an IPv4
947 socket on all interfaces causes an error.
948
949 . On OpenBSD, an IPv6 socket will not accept IPv4 calls. You have to set up
950 two sockets if you want to accept both kinds of call.
951
952 . FreeBSD is like OpenBSD, but it has the IPV6_V6ONLY socket option, which
953 can be turned off, to make it behave like the versions of Linux described
954 above.
955
956 . I heard a report that the USAGI IPv6 stack for Linux has implemented
957 IPV6_V6ONLY.
958
959 So, what we do when IPv6 is supported is as follows:
960
961 (1) After it is set up, the list of interfaces is scanned for wildcard
962 addresses. If an IPv6 and an IPv4 wildcard are both found for the same
963 port, the list is re-arranged so that they are together, with the IPv6
964 wildcard first.
965
966 (2) If the creation of a wildcard IPv6 socket fails, we just log the error and
967 carry on if an IPv4 wildcard socket for the same port follows later in the
968 list. This allows Exim to carry on in the case when the kernel has no IPv6
969 support.
970
971 (3) Having created an IPv6 wildcard socket, we try to set IPV6_V6ONLY if that
972 option is defined. However, if setting fails, carry on regardless (but log
973 the incident).
974
975 (4) If binding or listening on an IPv6 wildcard socket fails, it is a serious
976 error.
977
978 (5) If binding or listening on an IPv4 wildcard socket fails with the error
979 EADDRINUSE, and a previous interface was an IPv6 wildcard for the same
980 port (which must have succeeded or we wouldn't have got this far), we
981 assume we are in the situation where just a single socket is permitted,
982 and ignore the error.
983
984 Phew!
985
986 The preparation code decodes options and sets up the relevant data. We do this
987 first, so that we can return non-zero if there are any syntax errors, and also
988 write to stderr. */
989
990 if (daemon_listen)
991 {
992 int *default_smtp_port;
993 int sep;
994 int pct = 0;
995 uschar *s;
996 uschar *list;
997 uschar *local_iface_source = US"local_interfaces";
998 ip_address_item *ipa;
999 ip_address_item **pipa;
1000
1001 /* If any option requiring a load average to be available during the
1002 reception of a message is set, call os_getloadavg() while we are root
1003 for those OS for which this is necessary the first time it is called (in
1004 order to perform an "open" on the kernel memory file). */
1005
1006 #ifdef LOAD_AVG_NEEDS_ROOT
1007 if (queue_only_load >= 0 || smtp_load_reserve >= 0 ||
1008 (deliver_queue_load_max >= 0 && deliver_drop_privilege))
1009 (void)os_getloadavg();
1010 #endif
1011
1012 /* If -oX was used, disable the writing of a pid file unless -oP was
1013 explicitly used to force it. Then scan the string given to -oX. Any items
1014 that contain neither a dot nor a colon are used to override daemon_smtp_port.
1015 Any other items are used to override local_interfaces. */
1016
1017 if (override_local_interfaces != NULL)
1018 {
1019 uschar *new_smtp_port = NULL;
1020 uschar *new_local_interfaces = NULL;
1021 int portsize = 0;
1022 int portptr = 0;
1023 int ifacesize = 0;
1024 int ifaceptr = 0;
1025
1026 if (override_pid_file_path == NULL) write_pid = FALSE;
1027
1028 list = override_local_interfaces;
1029 sep = 0;
1030 while ((s = string_nextinlist(&list,&sep,big_buffer,big_buffer_size))
1031 != NULL)
1032 {
1033 uschar joinstr[4];
1034 uschar **ptr;
1035 int *sizeptr;
1036 int *ptrptr;
1037
1038 if (Ustrpbrk(s, ".:") == NULL)
1039 {
1040 ptr = &new_smtp_port;
1041 sizeptr = &portsize;
1042 ptrptr = &portptr;
1043 }
1044 else
1045 {
1046 ptr = &new_local_interfaces;
1047 sizeptr = &ifacesize;
1048 ptrptr = &ifaceptr;
1049 }
1050
1051 if (*ptr == NULL)
1052 {
1053 joinstr[0] = sep;
1054 joinstr[1] = ' ';
1055 *ptr = string_cat(*ptr, sizeptr, ptrptr, US"<", 1);
1056 }
1057
1058 *ptr = string_cat(*ptr, sizeptr, ptrptr, joinstr, 2);
1059 *ptr = string_cat(*ptr, sizeptr, ptrptr, s, Ustrlen(s));
1060 }
1061
1062 if (new_smtp_port != NULL)
1063 {
1064 new_smtp_port[portptr] = 0;
1065 daemon_smtp_port = new_smtp_port;
1066 DEBUG(D_any) debug_printf("daemon_smtp_port overridden by -oX:\n %s\n",
1067 daemon_smtp_port);
1068 }
1069
1070 if (new_local_interfaces != NULL)
1071 {
1072 new_local_interfaces[ifaceptr] = 0;
1073 local_interfaces = new_local_interfaces;
1074 local_iface_source = US"-oX data";
1075 DEBUG(D_any) debug_printf("local_interfaces overridden by -oX:\n %s\n",
1076 local_interfaces);
1077 }
1078 }
1079
1080 /* Create a list of default SMTP ports, to be used if local_interfaces
1081 contains entries without explict ports. First count the number of ports, then
1082 build a translated list in a vector. */
1083
1084 list = daemon_smtp_port;
1085 sep = 0;
1086 while ((s = string_nextinlist(&list,&sep,big_buffer,big_buffer_size)) != NULL)
1087 pct++;
1088 default_smtp_port = store_get((pct+1) * sizeof(int));
1089 list = daemon_smtp_port;
1090 sep = 0;
1091 for (pct = 0;
1092 (s = string_nextinlist(&list,&sep,big_buffer,big_buffer_size)) != NULL;
1093 pct++)
1094 {
1095 if (isdigit(*s))
1096 {
1097 uschar *end;
1098 default_smtp_port[pct] = Ustrtol(s, &end, 0);
1099 if (end != s + Ustrlen(s))
1100 log_write(0, LOG_PANIC_DIE|LOG_CONFIG, "invalid SMTP port: %s", s);
1101 }
1102 else
1103 {
1104 struct servent *smtp_service = getservbyname(CS s, "tcp");
1105 if (smtp_service == NULL)
1106 log_write(0, LOG_PANIC_DIE|LOG_CONFIG, "TCP port \"%s\" not found", s);
1107 default_smtp_port[pct] = ntohs(smtp_service->s_port);
1108 }
1109 }
1110 default_smtp_port[pct] = 0;
1111
1112 /* Create the list of local interfaces, possibly with ports included. This
1113 list may contain references to 0.0.0.0 and ::0 as wildcards. These special
1114 values are converted below. */
1115
1116 addresses = host_build_ifacelist(local_interfaces, local_iface_source);
1117
1118 /* In the list of IP addresses, convert 0.0.0.0 into an empty string, and ::0
1119 into the string ":". We use these to recognize wildcards in IPv4 and IPv6. In
1120 fact, many IP stacks recognize 0.0.0.0 and ::0 and handle them as wildcards
1121 anyway, but we need to know which are the wildcard addresses, and the shorter
1122 strings are neater.
1123
1124 In the same scan, fill in missing port numbers from the default list. When
1125 there is more than one item in the list, extra items are created. */
1126
1127 for (ipa = addresses; ipa != NULL; ipa = ipa->next)
1128 {
1129 int i;
1130
1131 if (Ustrcmp(ipa->address, "0.0.0.0") == 0) ipa->address[0] = 0;
1132 else if (Ustrcmp(ipa->address, "::0") == 0)
1133 {
1134 ipa->address[0] = ':';
1135 ipa->address[1] = 0;
1136 }
1137
1138 if (ipa->port > 0) continue;
1139
1140 if (daemon_smtp_port[0] <= 0)
1141 log_write(0, LOG_MAIN|LOG_PANIC_DIE, "no port specified for interface "
1142 "%s and daemon_smtp_port is unset; cannot start daemon",
1143 (ipa->address[0] == 0)? US"\"all IPv4\"" :
1144 (ipa->address[1] == 0)? US"\"all IPv6\"" : ipa->address);
1145 ipa->port = default_smtp_port[0];
1146 for (i = 1; default_smtp_port[i] > 0; i++)
1147 {
1148 ip_address_item *new = store_get(sizeof(ip_address_item));
1149 memcpy(new->address, ipa->address, Ustrlen(ipa->address) + 1);
1150 new->port = default_smtp_port[i];
1151 new->next = ipa->next;
1152 ipa->next = new;
1153 ipa = new;
1154 }
1155 }
1156
1157 /* Scan the list of addresses for wildcards. If we find an IPv4 and an IPv6
1158 wildcard for the same port, ensure that (a) they are together and (b) the
1159 IPv6 address comes first. This makes handling the messy features easier, and
1160 also simplifies the construction of the "daemon started" log line. */
1161
1162 pipa = &addresses;
1163 for (ipa = addresses; ipa != NULL; pipa = &(ipa->next), ipa = ipa->next)
1164 {
1165 ip_address_item *ipa2;
1166
1167 /* Handle an IPv4 wildcard */
1168
1169 if (ipa->address[0] == 0)
1170 {
1171 for (ipa2 = ipa; ipa2->next != NULL; ipa2 = ipa2->next)
1172 {
1173 ip_address_item *ipa3 = ipa2->next;
1174 if (ipa3->address[0] == ':' &&
1175 ipa3->address[1] == 0 &&
1176 ipa3->port == ipa->port)
1177 {
1178 ipa2->next = ipa3->next;
1179 ipa3->next = ipa;
1180 *pipa = ipa3;
1181 break;
1182 }
1183 }
1184 }
1185
1186 /* Handle an IPv6 wildcard. */
1187
1188 else if (ipa->address[0] == ':' && ipa->address[1] == 0)
1189 {
1190 for (ipa2 = ipa; ipa2->next != NULL; ipa2 = ipa2->next)
1191 {
1192 ip_address_item *ipa3 = ipa2->next;
1193 if (ipa3->address[0] == 0 && ipa3->port == ipa->port)
1194 {
1195 ipa2->next = ipa3->next;
1196 ipa3->next = ipa->next;
1197 ipa->next = ipa3;
1198 ipa = ipa3;
1199 break;
1200 }
1201 }
1202 }
1203 }
1204
1205 /* Get a vector to remember all the sockets in */
1206
1207 for (ipa = addresses; ipa != NULL; ipa = ipa->next)
1208 listen_socket_count++;
1209 listen_sockets = store_get(sizeof(int *) * listen_socket_count);
1210
1211 /* Do a sanity check on the max connects value just to save us from getting
1212 a huge amount of store. */
1213
1214 if (smtp_accept_max > 4095) smtp_accept_max = 4096;
1215
1216 /* There's no point setting smtp_accept_queue unless it is less than the max
1217 connects limit. The configuration reader ensures that the max is set if the
1218 queue-only option is set. */
1219
1220 if (smtp_accept_queue > smtp_accept_max) smtp_accept_queue = 0;
1221
1222 /* Get somewhere to keep the list of SMTP accepting pids if we are keeping
1223 track of them for total number and queue/host limits. */
1224
1225 if (smtp_accept_max > 0)
1226 {
1227 int i;
1228 smtp_slots = store_get(smtp_accept_max * sizeof(smtp_slot));
1229 for (i = 0; i < smtp_accept_max; i++) smtp_slots[i] = empty_smtp_slot;
1230 }
1231 }
1232
1233 /* The variable background_daemon is always false when debugging, but
1234 can also be forced false in order to keep a non-debugging daemon in the
1235 foreground. If background_daemon is true, close all open file descriptors that
1236 we know about, but then re-open stdin, stdout, and stderr to /dev/null.
1237
1238 This is protection against any called functions (in libraries, or in
1239 Perl, or whatever) that think they can write to stderr (or stdout). Before this
1240 was added, it was quite likely that an SMTP connection would use one of these
1241 file descriptors, in which case writing random stuff to it caused chaos.
1242
1243 Then disconnect from the controlling terminal, Most modern Unixes seem to have
1244 setsid() for getting rid of the controlling terminal. For any OS that doesn't,
1245 setsid() can be #defined as a no-op, or as something else. */
1246
1247 if (background_daemon)
1248 {
1249 log_close_all(); /* Just in case anything was logged earlier */
1250 search_tidyup(); /* Just in case any were used in reading the config. */
1251 (void)close(0); /* Get rid of stdin/stdout/stderr */
1252 (void)close(1);
1253 (void)close(2);
1254 exim_nullstd(); /* Connect stdin/stdout/stderr to /dev/null */
1255 log_stderr = NULL; /* So no attempt to copy paniclog output */
1256
1257 /* If the parent process of this one has pid == 1, we are re-initializing the
1258 daemon as the result of a SIGHUP. In this case, there is no need to do
1259 anything, because the controlling terminal has long gone. Otherwise, fork, in
1260 case current process is a process group leader (see 'man setsid' for an
1261 explanation) before calling setsid(). */
1262
1263 if (getppid() != 1)
1264 {
1265 pid_t pid = fork();
1266 if (pid < 0) log_write(0, LOG_MAIN|LOG_PANIC_DIE,
1267 "fork() failed when starting daemon: %s", strerror(errno));
1268 if (pid > 0) exit(EXIT_SUCCESS); /* in parent process, just exit */
1269 (void)setsid(); /* release controlling terminal */
1270 }
1271 }
1272
1273 /* We are now in the disconnected, daemon process (unless debugging). Set up
1274 the listening sockets if required. */
1275
1276 if (daemon_listen)
1277 {
1278 int sk;
1279 int on = 1;
1280 ip_address_item *ipa;
1281
1282 /* For each IP address, create a socket, bind it to the appropriate port, and
1283 start listening. See comments above about IPv6 sockets that may or may not
1284 accept IPv4 calls when listening on all interfaces. We also have to cope with
1285 the case of a system with IPv6 libraries, but no IPv6 support in the kernel.
1286 listening, provided a wildcard IPv4 socket for the same port follows. */
1287
1288 for (ipa = addresses, sk = 0; sk < listen_socket_count; ipa = ipa->next, sk++)
1289 {
1290 BOOL wildcard;
1291 ip_address_item *ipa2;
1292 int af;
1293
1294 if (Ustrchr(ipa->address, ':') != NULL)
1295 {
1296 af = AF_INET6;
1297 wildcard = ipa->address[1] == 0;
1298 }
1299 else
1300 {
1301 af = AF_INET;
1302 wildcard = ipa->address[0] == 0;
1303 }
1304
1305 listen_sockets[sk] = ip_socket(SOCK_STREAM, af);
1306 if (listen_sockets[sk] < 0)
1307 {
1308 if (check_special_case(0, addresses, ipa, FALSE))
1309 {
1310 log_write(0, LOG_MAIN, "Failed to create IPv6 socket for wildcard "
1311 "listening (%s): will use IPv4", strerror(errno));
1312 goto SKIP_SOCKET;
1313 }
1314 log_write(0, LOG_PANIC_DIE, "IPv%c socket creation failed: %s",
1315 (af == AF_INET6)? '6' : '4', strerror(errno));
1316 }
1317
1318 /* If this is an IPv6 wildcard socket, set IPV6_V6ONLY if that option is
1319 available. Just log failure (can get protocol not available, just like
1320 socket creation can). */
1321
1322 #ifdef IPV6_V6ONLY
1323 if (af == AF_INET6 && wildcard &&
1324 setsockopt(listen_sockets[sk], IPPROTO_IPV6, IPV6_V6ONLY, (char *)(&on),
1325 sizeof(on)) < 0)
1326 log_write(0, LOG_MAIN, "Setting IPV6_V6ONLY on daemon's IPv6 wildcard "
1327 "socket failed (%s): carrying on without it", strerror(errno));
1328 #endif /* IPV6_V6ONLY */
1329
1330 /* Set SO_REUSEADDR so that the daemon can be restarted while a connection
1331 is being handled. Without this, a connection will prevent reuse of the
1332 smtp port for listening. */
1333
1334 if (setsockopt(listen_sockets[sk], SOL_SOCKET, SO_REUSEADDR,
1335 (uschar *)(&on), sizeof(on)) < 0)
1336 log_write(0, LOG_MAIN|LOG_PANIC_DIE, "setting SO_REUSEADDR on socket "
1337 "failed when starting daemon: %s", strerror(errno));
1338
1339 /* Set TCP_NODELAY; Exim does its own buffering. There is a switch to
1340 disable this because it breaks some broken clients. */
1341
1342 if (tcp_nodelay) setsockopt(listen_sockets[sk], IPPROTO_TCP, TCP_NODELAY,
1343 (uschar *)(&on), sizeof(on));
1344
1345 /* Now bind the socket to the required port; if Exim is being restarted
1346 it may not always be possible to bind immediately, even with SO_REUSEADDR
1347 set, so try 10 times, waiting between each try. After 10 failures, we give
1348 up. In an IPv6 environment, if bind () fails with the error EADDRINUSE and
1349 we are doing wildcard IPv4 listening and there was a previous IPv6 wildcard
1350 address for the same port, ignore the error on the grounds that we must be
1351 in a system where the IPv6 socket accepts both kinds of call. This is
1352 necessary for (some release of) USAGI Linux; other IP stacks fail at the
1353 listen() stage instead. */
1354
1355 for(;;)
1356 {
1357 uschar *msg, *addr;
1358 if (ip_bind(listen_sockets[sk], af, ipa->address, ipa->port) >= 0) break;
1359 if (check_special_case(errno, addresses, ipa, TRUE))
1360 {
1361 DEBUG(D_any) debug_printf("wildcard IPv4 bind() failed after IPv6 "
1362 "listen() success; EADDRINUSE ignored\n");
1363 (void)close(listen_sockets[sk]);
1364 goto SKIP_SOCKET;
1365 }
1366 msg = US strerror(errno);
1367 addr = wildcard? ((af == AF_INET6)? US"(any IPv6)" : US"(any IPv4)") :
1368 ipa->address;
1369 if (daemon_startup_retries <= 0)
1370 log_write(0, LOG_MAIN|LOG_PANIC_DIE,
1371 "socket bind() to port %d for address %s failed: %s: "
1372 "daemon abandoned", ipa->port, addr, msg);
1373 log_write(0, LOG_MAIN, "socket bind() to port %d for address %s "
1374 "failed: %s: waiting %s before trying again (%d more %s)",
1375 ipa->port, addr, msg, readconf_printtime(daemon_startup_sleep),
1376 daemon_startup_retries, (daemon_startup_retries > 1)? "tries" : "try");
1377 daemon_startup_retries--;
1378 sleep(daemon_startup_sleep);
1379 }
1380
1381 DEBUG(D_any)
1382 {
1383 if (wildcard)
1384 debug_printf("listening on all interfaces (IPv%c) port %d\n",
1385 (af == AF_INET6)? '6' : '4', ipa->port);
1386 else
1387 debug_printf("listening on %s port %d\n", ipa->address, ipa->port);
1388 }
1389
1390 /* Start listening on the bound socket, establishing the maximum backlog of
1391 connections that is allowed. On success, continue to the next address. */
1392
1393 if (listen(listen_sockets[sk], smtp_connect_backlog) >= 0) continue;
1394
1395 /* Listening has failed. In an IPv6 environment, as for bind(), if listen()
1396 fails with the error EADDRINUSE and we are doing IPv4 wildcard listening
1397 and there was a previous successful IPv6 wildcard listen on the same port,
1398 we want to ignore the error on the grounds that we must be in a system
1399 where the IPv6 socket accepts both kinds of call. */
1400
1401 if (!check_special_case(errno, addresses, ipa, TRUE))
1402 log_write(0, LOG_PANIC_DIE, "listen() failed on interface %s: %s",
1403 wildcard? ((af == AF_INET6)? US"(any IPv6)" : US"(any IPv4)") :
1404 ipa->address,
1405 strerror(errno));
1406
1407 DEBUG(D_any) debug_printf("wildcard IPv4 listen() failed after IPv6 "
1408 "listen() success; EADDRINUSE ignored\n");
1409 (void)close(listen_sockets[sk]);
1410
1411 /* Come here if there has been a problem with the socket which we
1412 are going to ignore. We remove the address from the chain, and back up the
1413 counts. */
1414
1415 SKIP_SOCKET:
1416 sk--; /* Back up the count */
1417 listen_socket_count--; /* Reduce the total */
1418 if (ipa == addresses) addresses = ipa->next; else
1419 {
1420 for (ipa2 = addresses; ipa2->next != ipa; ipa2 = ipa2->next);
1421 ipa2->next = ipa->next;
1422 ipa = ipa2;
1423 }
1424 } /* End of bind/listen loop for each address */
1425 } /* End of setup for listening */
1426
1427
1428 /* If we are not listening, we want to write a pid file only if -oP was
1429 explicitly given. */
1430
1431 else if (override_pid_file_path == NULL) write_pid = FALSE;
1432
1433 /* Write the pid to a known file for assistance in identification, if required.
1434 We do this before giving up root privilege, because on some systems it is
1435 necessary to be root in order to write into the pid file directory. There's
1436 nothing to stop multiple daemons running, as long as no more than one listens
1437 on a given TCP/IP port on the same interface(s). However, in these
1438 circumstances it gets far too complicated to mess with pid file names
1439 automatically. Consequently, Exim 4 writes a pid file only
1440
1441 (a) When running in the test harness, or
1442 (b) When -bd is used and -oX is not used, or
1443 (c) When -oP is used to supply a path.
1444
1445 The variable daemon_write_pid is used to control this. */
1446
1447 if (running_in_test_harness || write_pid)
1448 {
1449 FILE *f;
1450
1451 if (override_pid_file_path != NULL)
1452 pid_file_path = override_pid_file_path;
1453
1454 if (pid_file_path[0] == 0)
1455 pid_file_path = string_sprintf("%s/exim-daemon.pid", spool_directory);
1456
1457 f = modefopen(pid_file_path, "wb", 0644);
1458 if (f != NULL)
1459 {
1460 (void)fprintf(f, "%d\n", (int)getpid());
1461 (void)fclose(f);
1462 DEBUG(D_any) debug_printf("pid written to %s\n", pid_file_path);
1463 }
1464 else
1465 {
1466 DEBUG(D_any)
1467 debug_printf("%s\n", string_open_failed(errno, "pid file %s",
1468 pid_file_path));
1469 }
1470 }
1471
1472 /* Set up the handler for SIGHUP, which causes a restart of the daemon. */
1473
1474 sighup_seen = FALSE;
1475 signal(SIGHUP, sighup_handler);
1476
1477 /* Give up root privilege at this point (assuming that exim_uid and exim_gid
1478 are not root). The third argument controls the running of initgroups().
1479 Normally we do this, in order to set up the groups for the Exim user. However,
1480 if we are not root at this time - some odd installations run that way - we
1481 cannot do this. */
1482
1483 exim_setugid(exim_uid, exim_gid, geteuid()==root_uid, US"running as a daemon");
1484
1485 /* Update the originator_xxx fields so that received messages as listed as
1486 coming from Exim, not whoever started the daemon. */
1487
1488 originator_uid = exim_uid;
1489 originator_gid = exim_gid;
1490 originator_login = ((pw = getpwuid(exim_uid)) != NULL)?
1491 string_copy_malloc(US pw->pw_name) : US"exim";
1492
1493 /* Get somewhere to keep the list of queue-runner pids if we are keeping track
1494 of them (and also if we are doing queue runs). */
1495
1496 if (queue_interval > 0 && queue_run_max > 0)
1497 {
1498 int i;
1499 queue_pid_slots = store_get(queue_run_max * sizeof(pid_t));
1500 for (i = 0; i < queue_run_max; i++) queue_pid_slots[i] = 0;
1501 }
1502
1503 /* Set up the handler for termination of child processes. */
1504
1505 sigchld_seen = FALSE;
1506 os_non_restarting_signal(SIGCHLD, main_sigchld_handler);
1507
1508 /* If we are to run the queue periodically, pretend the alarm has just gone
1509 off. This will cause the first queue-runner to get kicked off straight away. */
1510
1511 sigalrm_seen = (queue_interval > 0);
1512
1513 /* Log the start up of a daemon - at least one of listening or queue running
1514 must be set up. */
1515
1516 if (daemon_listen)
1517 {
1518 int i, j;
1519 int smtp_ports = 0;
1520 int smtps_ports = 0;
1521 ip_address_item *ipa;
1522 uschar *p = big_buffer;
1523 uschar *qinfo = (queue_interval > 0)?
1524 string_sprintf("-q%s", readconf_printtime(queue_interval))
1525 :
1526 US"no queue runs";
1527
1528 /* Build a list of listening addresses in big_buffer, but limit it to 10
1529 items. The style is for backwards compatibility.
1530
1531 It is now possible to have some ports listening for SMTPS (the old,
1532 deprecated protocol that starts TLS without using STARTTLS), and others
1533 listening for standard SMTP. Keep their listings separate. */
1534
1535 for (j = 0; j < 2; j++)
1536 {
1537 for (i = 0, ipa = addresses; i < 10 && ipa != NULL; i++, ipa = ipa->next)
1538 {
1539 /* First time round, look for SMTP ports; second time round, look for
1540 SMTPS ports. For the first one of each, insert leading text. */
1541
1542 if (host_is_tls_on_connect_port(ipa->port) == (j > 0))
1543 {
1544 if (j == 0)
1545 {
1546 if (smtp_ports++ == 0)
1547 {
1548 memcpy(p, "SMTP on", 8);
1549 p += 7;
1550 }
1551 }
1552 else
1553 {
1554 if (smtps_ports++ == 0)
1555 {
1556 (void)sprintf(CS p, "%sSMTPS on",
1557 (smtp_ports == 0)? "":" and for ");
1558 while (*p != 0) p++;
1559 }
1560 }
1561
1562 /* Now the information about the port (and sometimes interface) */
1563
1564 if (ipa->address[0] == ':' && ipa->address[1] == 0)
1565 {
1566 if (ipa->next != NULL && ipa->next->address[0] == 0 &&
1567 ipa->next->port == ipa->port)
1568 {
1569 (void)sprintf(CS p, " port %d (IPv6 and IPv4)", ipa->port);
1570 ipa = ipa->next;
1571 }
1572 else if (ipa->v6_include_v4)
1573 (void)sprintf(CS p, " port %d (IPv6 with IPv4)", ipa->port);
1574 else
1575 (void)sprintf(CS p, " port %d (IPv6)", ipa->port);
1576 }
1577 else if (ipa->address[0] == 0)
1578 (void)sprintf(CS p, " port %d (IPv4)", ipa->port);
1579 else
1580 (void)sprintf(CS p, " [%s]:%d", ipa->address, ipa->port);
1581 while (*p != 0) p++;
1582 }
1583 }
1584
1585 if (ipa != NULL)
1586 {
1587 memcpy(p, " ...", 5);
1588 p += 4;
1589 }
1590 }
1591
1592 log_write(0, LOG_MAIN,
1593 "exim %s daemon started: pid=%d, %s, listening for %s",
1594 version_string, getpid(), qinfo, big_buffer);
1595 set_process_info("daemon: %s, listening for %s", qinfo, big_buffer);
1596 }
1597
1598 else
1599 {
1600 log_write(0, LOG_MAIN,
1601 "exim %s daemon started: pid=%d, -q%s, not listening for SMTP",
1602 version_string, getpid(), readconf_printtime(queue_interval));
1603 set_process_info("daemon: -q%s, not listening",
1604 readconf_printtime(queue_interval));
1605 }
1606
1607
1608 /* Close the log so it can be renamed and moved. In the few cases below where
1609 this long-running process writes to the log (always exceptional conditions), it
1610 closes the log afterwards, for the same reason. */
1611
1612 log_close_all();
1613
1614 DEBUG(D_any) debug_print_ids(US"daemon running with");
1615
1616 /* Any messages accepted via this route are going to be SMTP. */
1617
1618 smtp_input = TRUE;
1619
1620 /* Enter the never-ending loop... */
1621
1622 for (;;)
1623 {
1624 #if HAVE_IPV6
1625 struct sockaddr_in6 accepted;
1626 #else
1627 struct sockaddr_in accepted;
1628 #endif
1629
1630 EXIM_SOCKLEN_T len;
1631 pid_t pid;
1632
1633 /* This code is placed first in the loop, so that it gets obeyed at the
1634 start, before the first wait. This causes the first queue-runner to be
1635 started immediately. */
1636
1637 if (sigalrm_seen)
1638 {
1639 DEBUG(D_any) debug_printf("SIGALRM received\n");
1640
1641 /* Do a full queue run in a child process, if required, unless we already
1642 have enough queue runners on the go. If we are not running as root, a
1643 re-exec is required. */
1644
1645 if (queue_interval > 0 &&
1646 (queue_run_max <= 0 || queue_run_count < queue_run_max))
1647 {
1648 if ((pid = fork()) == 0)
1649 {
1650 int sk;
1651
1652 DEBUG(D_any) debug_printf("Starting queue-runner: pid %d\n",
1653 (int)getpid());
1654
1655 /* Disable debugging if it's required only for the daemon process. We
1656 leave the above message, because it ties up with the "child ended"
1657 debugging messages. */
1658
1659 if (debug_daemon) debug_selector = 0;
1660
1661 /* Close any open listening sockets in the child */
1662
1663 for (sk = 0; sk < listen_socket_count; sk++)
1664 (void)close(listen_sockets[sk]);
1665
1666 /* Reset SIGHUP and SIGCHLD in the child in both cases. */
1667
1668 signal(SIGHUP, SIG_DFL);
1669 signal(SIGCHLD, SIG_DFL);
1670
1671 /* Re-exec if privilege has been given up, unless deliver_drop_
1672 privilege is set. Reset SIGALRM before exec(). */
1673
1674 if (geteuid() != root_uid && !deliver_drop_privilege)
1675 {
1676 uschar opt[8];
1677 uschar *p = opt;
1678 uschar *extra[5];
1679 int extracount = 1;
1680
1681 signal(SIGALRM, SIG_DFL);
1682 *p++ = '-';
1683 *p++ = 'q';
1684 if (queue_2stage) *p++ = 'q';
1685 if (queue_run_first_delivery) *p++ = 'i';
1686 if (queue_run_force) *p++ = 'f';
1687 if (deliver_force_thaw) *p++ = 'f';
1688 if (queue_run_local) *p++ = 'l';
1689 *p = 0;
1690 extra[0] = opt;
1691
1692 /* If -R or -S were on the original command line, ensure they get
1693 passed on. */
1694
1695 if (deliver_selectstring != NULL)
1696 {
1697 extra[extracount++] = deliver_selectstring_regex? US"-Rr" : US"-R";
1698 extra[extracount++] = deliver_selectstring;
1699 }
1700
1701 if (deliver_selectstring_sender != NULL)
1702 {
1703 extra[extracount++] = deliver_selectstring_sender_regex?
1704 US"-Sr" : US"-S";
1705 extra[extracount++] = deliver_selectstring_sender;
1706 }
1707
1708 /* Overlay this process with a new execution. */
1709
1710 (void)child_exec_exim(CEE_EXEC_PANIC, FALSE, NULL, TRUE, extracount,
1711 extra[0], extra[1], extra[2], extra[3], extra[4]);
1712
1713 /* Control never returns here. */
1714 }
1715
1716 /* No need to re-exec; SIGALRM remains set to the default handler */
1717
1718 queue_run(NULL, NULL, FALSE);
1719 _exit(EXIT_SUCCESS);
1720 }
1721
1722 if (pid < 0)
1723 {
1724 log_write(0, LOG_MAIN|LOG_PANIC, "daemon: fork of queue-runner "
1725 "process failed: %s", strerror(errno));
1726 log_close_all();
1727 }
1728 else
1729 {
1730 int i;
1731 for (i = 0; i < queue_run_max; ++i)
1732 {
1733 if (queue_pid_slots[i] <= 0)
1734 {
1735 queue_pid_slots[i] = pid;
1736 queue_run_count++;
1737 break;
1738 }
1739 }
1740 DEBUG(D_any) debug_printf("%d queue-runner process%s running\n",
1741 queue_run_count, (queue_run_count == 1)? "" : "es");
1742 }
1743 }
1744
1745 /* Reset the alarm clock */
1746
1747 sigalrm_seen = FALSE;
1748 alarm(queue_interval);
1749 }
1750
1751
1752 /* Sleep till a connection happens if listening, and handle the connection if
1753 that is why we woke up. The FreeBSD operating system requires the use of
1754 select() before accept() because the latter function is not interrupted by
1755 a signal, and we want to wake up for SIGCHLD and SIGALRM signals. Some other
1756 OS do notice signals in accept() but it does no harm to have the select()
1757 in for all of them - and it won't then be a lurking problem for ports to
1758 new OS. In fact, the later addition of listening on specific interfaces only
1759 requires this way of working anyway. */
1760
1761 if (daemon_listen)
1762 {
1763 int sk, lcount, select_errno;
1764 int max_socket = 0;
1765 BOOL select_failed = FALSE;
1766 fd_set select_listen;
1767
1768 FD_ZERO(&select_listen);
1769 for (sk = 0; sk < listen_socket_count; sk++)
1770 {
1771 FD_SET(listen_sockets[sk], &select_listen);
1772 if (listen_sockets[sk] > max_socket) max_socket = listen_sockets[sk];
1773 }
1774
1775 DEBUG(D_any) debug_printf("Listening...\n");
1776
1777 /* In rare cases we may have had a SIGCHLD signal in the time between
1778 setting the handler (below) and getting back here. If so, pretend that the
1779 select() was interrupted so that we reap the child. This might still leave
1780 a small window when a SIGCHLD could get lost. However, since we use SIGCHLD
1781 only to do the reaping more quickly, it shouldn't result in anything other
1782 than a delay until something else causes a wake-up. */
1783
1784 if (sigchld_seen)
1785 {
1786 lcount = -1;
1787 errno = EINTR;
1788 }
1789 else
1790 {
1791 lcount = select(max_socket + 1, (SELECT_ARG2_TYPE *)&select_listen,
1792 NULL, NULL, NULL);
1793 }
1794
1795 if (lcount < 0)
1796 {
1797 select_failed = TRUE;
1798 lcount = 1;
1799 }
1800
1801 /* Clean up any subprocesses that may have terminated. We need to do this
1802 here so that smtp_accept_max_per_host works when a connection to that host
1803 has completed, and we are about to accept a new one. When this code was
1804 later in the sequence, a new connection could be rejected, even though an
1805 old one had just finished. Preserve the errno from any select() failure for
1806 the use of the common select/accept error processing below. */
1807
1808 select_errno = errno;
1809 handle_ending_processes();
1810 errno = select_errno;
1811
1812 /* Loop for all the sockets that are currently ready to go. If select
1813 actually failed, we have set the count to 1 and select_failed=TRUE, so as
1814 to use the common error code for select/accept below. */
1815
1816 while (lcount-- > 0)
1817 {
1818 int accept_socket = -1;
1819 if (!select_failed)
1820 {
1821 for (sk = 0; sk < listen_socket_count; sk++)
1822 {
1823 if (FD_ISSET(listen_sockets[sk], &select_listen))
1824 {
1825 len = sizeof(accepted);
1826 accept_socket = accept(listen_sockets[sk],
1827 (struct sockaddr *)&accepted, &len);
1828 FD_CLR(listen_sockets[sk], &select_listen);
1829 break;
1830 }
1831 }
1832 }
1833
1834 /* If select or accept has failed and this was not caused by an
1835 interruption, log the incident and try again. With asymmetric TCP/IP
1836 routing errors such as "No route to network" have been seen here. Also
1837 "connection reset by peer" has been seen. These cannot be classed as
1838 disastrous errors, but they could fill up a lot of log. The code in smail
1839 crashes the daemon after 10 successive failures of accept, on the grounds
1840 that some OS fail continuously. Exim originally followed suit, but this
1841 appears to have caused problems. Now it just keeps going, but instead of
1842 logging each error, it batches them up when they are continuous. */
1843
1844 if (accept_socket < 0 && errno != EINTR)
1845 {
1846 if (accept_retry_count == 0)
1847 {
1848 accept_retry_errno = errno;
1849 accept_retry_select_failed = select_failed;
1850 }
1851 else
1852 {
1853 if (errno != accept_retry_errno ||
1854 select_failed != accept_retry_select_failed ||
1855 accept_retry_count >= 50)
1856 {
1857 log_write(0, LOG_MAIN | ((accept_retry_count >= 50)? LOG_PANIC : 0),
1858 "%d %s() failure%s: %s",
1859 accept_retry_count,
1860 accept_retry_select_failed? "select" : "accept",
1861 (accept_retry_count == 1)? "" : "s",
1862 strerror(accept_retry_errno));
1863 log_close_all();
1864 accept_retry_count = 0;
1865 accept_retry_errno = errno;
1866 accept_retry_select_failed = select_failed;
1867 }
1868 }
1869 accept_retry_count++;
1870 }
1871
1872 else
1873 {
1874 if (accept_retry_count > 0)
1875 {
1876 log_write(0, LOG_MAIN, "%d %s() failure%s: %s",
1877 accept_retry_count,
1878 accept_retry_select_failed? "select" : "accept",
1879 (accept_retry_count == 1)? "" : "s",
1880 strerror(accept_retry_errno));
1881 log_close_all();
1882 accept_retry_count = 0;
1883 }
1884 }
1885
1886 /* If select/accept succeeded, deal with the connection. */
1887
1888 if (accept_socket >= 0)
1889 handle_smtp_call(listen_sockets, listen_socket_count, accept_socket,
1890 (struct sockaddr *)&accepted);
1891 }
1892 }
1893
1894 /* If not listening, then just sleep for the queue interval. If we woke
1895 up early the last time for some other signal, it won't matter because
1896 the alarm signal will wake at the right time. This code originally used
1897 sleep() but it turns out that on the FreeBSD system, sleep() is not inter-
1898 rupted by signals, so it wasn't waking up for SIGALRM or SIGCHLD. Luckily
1899 select() can be used as an interruptible sleep() on all versions of Unix. */
1900
1901 else
1902 {
1903 struct timeval tv;
1904 tv.tv_sec = queue_interval;
1905 tv.tv_usec = 0;
1906 select(0, NULL, NULL, NULL, &tv);
1907 handle_ending_processes();
1908 }
1909
1910 /* Re-enable the SIGCHLD handler if it has been run. It can't do it
1911 for itself, because it isn't doing the waiting itself. */
1912
1913 if (sigchld_seen)
1914 {
1915 sigchld_seen = FALSE;
1916 os_non_restarting_signal(SIGCHLD, main_sigchld_handler);
1917 }
1918
1919 /* Handle being woken by SIGHUP. We know at this point that the result
1920 of accept() has been dealt with, so we can re-exec exim safely, first
1921 closing the listening sockets so that they can be reused. Cancel any pending
1922 alarm in case it is just about to go off, and set SIGHUP to be ignored so
1923 that another HUP in quick succession doesn't clobber the new daemon before it
1924 gets going. All log files get closed by the close-on-exec flag; however, if
1925 the exec fails, we need to close the logs. */
1926
1927 if (sighup_seen)
1928 {
1929 int sk;
1930 log_write(0, LOG_MAIN, "pid %d: SIGHUP received: re-exec daemon",
1931 getpid());
1932 for (sk = 0; sk < listen_socket_count; sk++)
1933 (void)close(listen_sockets[sk]);
1934 alarm(0);
1935 signal(SIGHUP, SIG_IGN);
1936 sighup_argv[0] = exim_path;
1937 exim_nullstd();
1938 execv(CS exim_path, (char *const *)sighup_argv);
1939 log_write(0, LOG_MAIN|LOG_PANIC_DIE, "pid %d: exec of %s failed: %s",
1940 getpid(), exim_path, strerror(errno));
1941 log_close_all();
1942 }
1943
1944 } /* End of main loop */
1945
1946 /* Control never reaches here */
1947 }
1948
1949 /* End of exim_daemon.c */
1950