Add comments as to why smtp_log_no_mail() is not called when
[exim.git] / src / src / daemon.c
1 /* $Cambridge: exim/src/src/daemon.c,v 1.21 2007/01/17 11:29:39 ph10 Exp $ */
2
3 /*************************************************
4 * Exim - an Internet mail transport agent *
5 *************************************************/
6
7 /* Copyright (c) University of Cambridge 1995 - 2007 */
8 /* See the file NOTICE for conditions of use and distribution. */
9
10 /* Functions concerned with running Exim as a daemon */
11
12
13 #include "exim.h"
14
15
16 /* Structure for holding data for each SMTP connection */
17
18 typedef struct smtp_slot {
19 pid_t pid; /* pid of the spawned reception process */
20 uschar *host_address; /* address of the client host */
21 } smtp_slot;
22
23 /* An empty slot for initializing (Standard C does not allow constructor
24 expressions in assigments except as initializers in declarations). */
25
26 static smtp_slot empty_smtp_slot = { 0, NULL };
27
28
29
30 /*************************************************
31 * Local static variables *
32 *************************************************/
33
34 static volatile BOOL sigchld_seen;
35 static volatile BOOL sighup_seen;
36
37 static int accept_retry_count = 0;
38 static int accept_retry_errno;
39 static BOOL accept_retry_select_failed;
40
41 static int queue_run_count = 0;
42 static pid_t *queue_pid_slots = NULL;
43 static smtp_slot *smtp_slots = NULL;
44
45 static BOOL write_pid = TRUE;
46
47
48
49 /*************************************************
50 * SIGHUP Handler *
51 *************************************************/
52
53 /* All this handler does is to set a flag and re-enable the signal.
54
55 Argument: the signal number
56 Returns: nothing
57 */
58
59 static void
60 sighup_handler(int sig)
61 {
62 sig = sig; /* Keep picky compilers happy */
63 sighup_seen = TRUE;
64 signal(SIGHUP, sighup_handler);
65 }
66
67
68
69 /*************************************************
70 * SIGCHLD handler for main daemon process *
71 *************************************************/
72
73 /* Don't re-enable the handler here, since we aren't doing the
74 waiting here. If the signal is re-enabled, there will just be an
75 infinite sequence of calls to this handler. The SIGCHLD signal is
76 used just as a means of waking up the daemon so that it notices
77 terminated subprocesses as soon as possible.
78
79 Argument: the signal number
80 Returns: nothing
81 */
82
83 static void
84 main_sigchld_handler(int sig)
85 {
86 sig = sig; /* Keep picky compilers happy */
87 os_non_restarting_signal(SIGCHLD, SIG_DFL);
88 sigchld_seen = TRUE;
89 }
90
91
92
93
94 /*************************************************
95 * Unexpected errors in SMTP calls *
96 *************************************************/
97
98 /* This function just saves a bit of repetitious coding.
99
100 Arguments:
101 log_msg Text of message to be logged
102 smtp_msg Text of SMTP error message
103 was_errno The failing errno
104
105 Returns: nothing
106 */
107
108 static void
109 never_error(uschar *log_msg, uschar *smtp_msg, int was_errno)
110 {
111 uschar *emsg = (was_errno <= 0)? US"" :
112 string_sprintf(": %s", strerror(was_errno));
113 log_write(0, LOG_MAIN|LOG_PANIC, "%s%s", log_msg, emsg);
114 if (smtp_out != NULL) smtp_printf("421 %s\r\n", smtp_msg);
115 }
116
117
118
119
120 /*************************************************
121 * Handle a connected SMTP call *
122 *************************************************/
123
124 /* This function is called when an SMTP connection has been accepted.
125 If there are too many, give an error message and close down. Otherwise
126 spin off a sub-process to handle the call. The list of listening sockets
127 is required so that they can be closed in the sub-process. Take care not to
128 leak store in this process - reset the stacking pool at the end.
129
130 Arguments:
131 listen_sockets sockets which are listening for incoming calls
132 listen_socket_count count of listening sockets
133 accept_socket socket of the current accepted call
134 accepted socket information about the current call
135
136 Returns: nothing
137 */
138
139 static void
140 handle_smtp_call(int *listen_sockets, int listen_socket_count,
141 int accept_socket, struct sockaddr *accepted)
142 {
143 pid_t pid;
144 union sockaddr_46 interface_sockaddr;
145 EXIM_SOCKLEN_T ifsize = sizeof(interface_sockaddr);
146 int dup_accept_socket = -1;
147 int max_for_this_host = 0;
148 int wfsize = 0;
149 int wfptr = 0;
150 int use_log_write_selector = log_write_selector;
151 uschar *whofrom = NULL;
152
153 void *reset_point = store_get(0);
154
155 /* Make the address available in ASCII representation, and also fish out
156 the remote port. */
157
158 sender_host_address = host_ntoa(-1, accepted, NULL, &sender_host_port);
159 DEBUG(D_any) debug_printf("Connection request from %s port %d\n",
160 sender_host_address, sender_host_port);
161
162 /* Set up the output stream, check the socket has duplicated, and set up the
163 input stream. These operations fail only the exceptional circumstances. Note
164 that never_error() won't use smtp_out if it is NULL. */
165
166 smtp_out = fdopen(accept_socket, "wb");
167 if (smtp_out == NULL)
168 {
169 never_error(US"daemon: fdopen() for smtp_out failed", US"", errno);
170 goto ERROR_RETURN;
171 }
172
173 dup_accept_socket = dup(accept_socket);
174 if (dup_accept_socket < 0)
175 {
176 never_error(US"daemon: couldn't dup socket descriptor",
177 US"Connection setup failed", errno);
178 goto ERROR_RETURN;
179 }
180
181 smtp_in = fdopen(dup_accept_socket, "rb");
182 if (smtp_in == NULL)
183 {
184 never_error(US"daemon: fdopen() for smtp_in failed",
185 US"Connection setup failed", errno);
186 goto ERROR_RETURN;
187 }
188
189 /* Get the data for the local interface address. Panic for most errors, but
190 "connection reset by peer" just means the connection went away. */
191
192 if (getsockname(accept_socket, (struct sockaddr *)(&interface_sockaddr),
193 &ifsize) < 0)
194 {
195 log_write(0, LOG_MAIN | ((errno == ECONNRESET)? 0 : LOG_PANIC),
196 "getsockname() failed: %s", strerror(errno));
197 smtp_printf("421 Local problem: getsockname() failed; please try again later\r\n");
198 goto ERROR_RETURN;
199 }
200
201 interface_address = host_ntoa(-1, &interface_sockaddr, NULL, &interface_port);
202 DEBUG(D_interface) debug_printf("interface address=%s port=%d\n",
203 interface_address, interface_port);
204
205 /* Build a string identifying the remote host and, if requested, the port and
206 the local interface data. This is for logging; at the end of this function the
207 memory is reclaimed. */
208
209 whofrom = string_append(whofrom, &wfsize, &wfptr, 3, "[", sender_host_address, "]");
210
211 if ((log_extra_selector & LX_incoming_port) != 0)
212 whofrom = string_append(whofrom, &wfsize, &wfptr, 2, ":", string_sprintf("%d",
213 sender_host_port));
214
215 if ((log_extra_selector & LX_incoming_interface) != 0)
216 whofrom = string_append(whofrom, &wfsize, &wfptr, 4, " I=[",
217 interface_address, "]:", string_sprintf("%d", interface_port));
218
219 whofrom[wfptr] = 0; /* Terminate the newly-built string */
220
221 /* Check maximum number of connections. We do not check for reserved
222 connections or unacceptable hosts here. That is done in the subprocess because
223 it might take some time. */
224
225 if (smtp_accept_max > 0 && smtp_accept_count >= smtp_accept_max)
226 {
227 DEBUG(D_any) debug_printf("rejecting SMTP connection: count=%d max=%d\n",
228 smtp_accept_count, smtp_accept_max);
229 smtp_printf("421 Too many concurrent SMTP connections; "
230 "please try again later.\r\n");
231 log_write(L_connection_reject,
232 LOG_MAIN, "Connection from %s refused: too many connections",
233 whofrom);
234 goto ERROR_RETURN;
235 }
236
237 /* If a load limit above which only reserved hosts are acceptable is defined,
238 get the load average here, and if there are in fact no reserved hosts, do
239 the test right away (saves a fork). If there are hosts, do the check in the
240 subprocess because it might take time. */
241
242 if (smtp_load_reserve >= 0)
243 {
244 load_average = os_getloadavg();
245 if (smtp_reserve_hosts == NULL && load_average > smtp_load_reserve)
246 {
247 DEBUG(D_any) debug_printf("rejecting SMTP connection: load average = %.2f\n",
248 (double)load_average/1000.0);
249 smtp_printf("421 Too much load; please try again later.\r\n");
250 log_write(L_connection_reject,
251 LOG_MAIN, "Connection from %s refused: load average = %.2f",
252 whofrom, (double)load_average/1000.0);
253 goto ERROR_RETURN;
254 }
255 }
256
257 /* Check that one specific host (strictly, IP address) is not hogging
258 resources. This is done here to prevent a denial of service attack by someone
259 forcing you to fork lots of times before denying service. The value of
260 smtp_accept_max_per_host is a string which is expanded. This makes it possible
261 to provide host-specific limits according to $sender_host address, but because
262 this is in the daemon mainline, only fast expansions (such as inline address
263 checks) should be used. The documentation is full of warnings. */
264
265 if (smtp_accept_max_per_host != NULL)
266 {
267 uschar *expanded = expand_string(smtp_accept_max_per_host);
268 if (expanded == NULL)
269 {
270 if (!expand_string_forcedfail)
271 log_write(0, LOG_MAIN|LOG_PANIC, "expansion of smtp_accept_max_per_host "
272 "failed for %s: %s", whofrom, expand_string_message);
273 }
274 /* For speed, interpret a decimal number inline here */
275 else
276 {
277 uschar *s = expanded;
278 while (isdigit(*s))
279 max_for_this_host = max_for_this_host * 10 + *s++ - '0';
280 if (*s != 0)
281 log_write(0, LOG_MAIN|LOG_PANIC, "expansion of smtp_accept_max_per_host "
282 "for %s contains non-digit: %s", whofrom, expanded);
283 }
284 }
285
286 /* If we have fewer connections than max_for_this_host, we can skip the tedious
287 per host_address checks. Note that at this stage smtp_accept_count contains the
288 count of *other* connections, not including this one. */
289
290 if ((max_for_this_host > 0) &&
291 (smtp_accept_count >= max_for_this_host))
292 {
293 int i;
294 int host_accept_count = 0;
295 int other_host_count = 0; /* keep a count of non matches to optimise */
296
297 for (i = 0; i < smtp_accept_max; ++i)
298 {
299 if (smtp_slots[i].host_address != NULL)
300 {
301 if (Ustrcmp(sender_host_address, smtp_slots[i].host_address) == 0)
302 host_accept_count++;
303 else
304 other_host_count++;
305
306 /* Testing all these strings is expensive - see if we can drop out
307 early, either by hitting the target, or finding there are not enough
308 connections left to make the target. */
309
310 if ((host_accept_count >= max_for_this_host) ||
311 ((smtp_accept_count - other_host_count) < max_for_this_host))
312 break;
313 }
314 }
315
316 if (host_accept_count >= max_for_this_host)
317 {
318 DEBUG(D_any) debug_printf("rejecting SMTP connection: too many from this "
319 "IP address: count=%d max=%d\n",
320 host_accept_count, max_for_this_host);
321 smtp_printf("421 Too many concurrent SMTP connections "
322 "from this IP address; please try again later.\r\n");
323 log_write(L_connection_reject,
324 LOG_MAIN, "Connection from %s refused: too many connections "
325 "from that IP address", whofrom);
326 goto ERROR_RETURN;
327 }
328 }
329
330 /* OK, the connection count checks have been passed. Before we can fork the
331 accepting process, we must first log the connection if requested. This logging
332 used to happen in the subprocess, but doing that means that the value of
333 smtp_accept_count can be out of step by the time it is logged. So we have to do
334 the logging here and accept the performance cost. Note that smtp_accept_count
335 hasn't yet been incremented to take account of this connection.
336
337 In order to minimize the cost (because this is going to happen for every
338 connection), do a preliminary selector test here. This saves ploughing through
339 the generalized logging code each time when the selector is false. If the
340 selector is set, check whether the host is on the list for logging. If not,
341 arrange to unset the selector in the subprocess. */
342
343 if ((log_write_selector & L_smtp_connection) != 0)
344 {
345 uschar *list = hosts_connection_nolog;
346 if (list != NULL && verify_check_host(&list) == OK)
347 use_log_write_selector &= ~L_smtp_connection;
348 else
349 log_write(L_smtp_connection, LOG_MAIN, "SMTP connection from %s "
350 "(TCP/IP connection count = %d)", whofrom, smtp_accept_count + 1);
351 }
352
353 /* Now we can fork the accepting process; do a lookup tidy, just in case any
354 expansion above did a lookup. */
355
356 search_tidyup();
357 pid = fork();
358
359 /* Handle the child process */
360
361 if (pid == 0)
362 {
363 int i;
364 int queue_only_reason = 0;
365 int old_pool = store_pool;
366 int save_debug_selector = debug_selector;
367 BOOL local_queue_only;
368 #ifdef SA_NOCLDWAIT
369 struct sigaction act;
370 #endif
371
372 /* May have been modified for the subprocess */
373
374 log_write_selector = use_log_write_selector;
375
376 /* Get the local interface address into permanent store */
377
378 store_pool = POOL_PERM;
379 interface_address = string_copy(interface_address);
380 store_pool = old_pool;
381
382 /* Check for a tls-on-connect port */
383
384 if (host_is_tls_on_connect_port(interface_port)) tls_on_connect = TRUE;
385
386 /* Expand smtp_active_hostname if required. We do not do this any earlier,
387 because it may depend on the local interface address (indeed, that is most
388 likely what it depends on.) */
389
390 smtp_active_hostname = primary_hostname;
391 if (raw_active_hostname != NULL)
392 {
393 uschar *nah = expand_string(raw_active_hostname);
394 if (nah == NULL)
395 {
396 if (!expand_string_forcedfail)
397 {
398 log_write(0, LOG_MAIN|LOG_PANIC, "failed to expand \"%s\" "
399 "(smtp_active_hostname): %s", raw_active_hostname,
400 expand_string_message);
401 smtp_printf("421 Local configuration error; "
402 "please try again later.\r\n");
403 mac_smtp_fflush();
404 search_tidyup();
405 _exit(EXIT_FAILURE);
406 }
407 }
408 else if (nah[0] != 0) smtp_active_hostname = nah;
409 }
410
411 /* Initialize the queueing flags */
412
413 queue_check_only();
414 local_queue_only = queue_only;
415
416 /* Close the listening sockets, and set the SIGCHLD handler to SIG_IGN.
417 We also attempt to set things up so that children are automatically reaped,
418 but just in case this isn't available, there's a paranoid waitpid() in the
419 loop too (except for systems where we are sure it isn't needed). See the more
420 extensive comment before the reception loop in exim.c for a fuller
421 explanation of this logic. */
422
423 for (i = 0; i < listen_socket_count; i++) (void)close(listen_sockets[i]);
424
425 #ifdef SA_NOCLDWAIT
426 act.sa_handler = SIG_IGN;
427 sigemptyset(&(act.sa_mask));
428 act.sa_flags = SA_NOCLDWAIT;
429 sigaction(SIGCHLD, &act, NULL);
430 #else
431 signal(SIGCHLD, SIG_IGN);
432 #endif
433
434 /* Attempt to get an id from the sending machine via the RFC 1413
435 protocol. We do this in the sub-process in order not to hold up the
436 main process if there is any delay. Then set up the fullhost information
437 in case there is no HELO/EHLO.
438
439 If debugging is enabled only for the daemon, we must turn if off while
440 finding the id, but turn it on again afterwards so that information about the
441 incoming connection is output. */
442
443 if (debug_daemon) debug_selector = 0;
444 verify_get_ident(IDENT_PORT);
445 host_build_sender_fullhost();
446 debug_selector = save_debug_selector;
447
448 DEBUG(D_any)
449 debug_printf("Process %d is handling incoming connection from %s\n",
450 (int)getpid(), sender_fullhost);
451
452 /* Now disable debugging permanently if it's required only for the daemon
453 process. */
454
455 if (debug_daemon) debug_selector = 0;
456
457 /* If there are too many child processes for immediate delivery,
458 set the local_queue_only flag, which is initialized from the
459 configured value and may therefore already be TRUE. Leave logging
460 till later so it will have a message id attached. */
461
462 if (smtp_accept_queue > 0 && smtp_accept_count >= smtp_accept_queue)
463 {
464 local_queue_only = TRUE;
465 queue_only_reason = 1;
466 }
467
468 /* Handle the start of the SMTP session, then loop, accepting incoming
469 messages from the SMTP connection. The end will come at the QUIT command,
470 when smtp_setup_msg() returns 0. A break in the connection causes the
471 process to die (see accept.c).
472
473 NOTE: We do *not* call smtp_log_no_mail() if smtp_start_session() fails,
474 because a log line has already been written for all its failure exists
475 (usually "connection refused: <reason>") and writing another one is
476 unnecessary clutter. */
477
478 if (!smtp_start_session())
479 {
480 mac_smtp_fflush();
481 search_tidyup();
482 _exit(EXIT_SUCCESS);
483 }
484
485 for (;;)
486 {
487 int rc;
488 message_id[0] = 0; /* Clear out any previous message_id */
489 reset_point = store_get(0); /* Save current store high water point */
490
491 DEBUG(D_any)
492 debug_printf("Process %d is ready for new message\n", (int)getpid());
493
494 /* Smtp_setup_msg() returns 0 on QUIT or if the call is from an
495 unacceptable host or if an ACL "drop" command was triggered, -1 on
496 connection lost, and +1 on validly reaching DATA. Receive_msg() almost
497 always returns TRUE when smtp_input is true; just retry if no message was
498 accepted (can happen for invalid message parameters). However, it can yield
499 FALSE if the connection was forcibly dropped by the DATA ACL. */
500
501 if ((rc = smtp_setup_msg()) > 0)
502 {
503 BOOL ok = receive_msg(FALSE);
504 search_tidyup(); /* Close cached databases */
505 if (!ok) /* Connection was dropped */
506 {
507 mac_smtp_fflush();
508 smtp_log_no_mail(); /* Log no mail if configured */
509 _exit(EXIT_SUCCESS);
510 }
511 if (message_id[0] == 0) continue; /* No message was accepted */
512 }
513 else
514 {
515 mac_smtp_fflush();
516 search_tidyup();
517 smtp_log_no_mail(); /* Log no mail if configured */
518 _exit((rc == 0)? EXIT_SUCCESS : EXIT_FAILURE);
519 }
520
521 /* Show the recipients when debugging */
522
523 DEBUG(D_receive)
524 {
525 int i;
526 if (sender_address != NULL)
527 debug_printf("Sender: %s\n", sender_address);
528 if (recipients_list != NULL)
529 {
530 debug_printf("Recipients:\n");
531 for (i = 0; i < recipients_count; i++)
532 debug_printf(" %s\n", recipients_list[i].address);
533 }
534 }
535
536 /* A message has been accepted. Clean up any previous delivery processes
537 that have completed and are defunct, on systems where they don't go away
538 by themselves (see comments when setting SIG_IGN above). On such systems
539 (if any) these delivery processes hang around after termination until
540 the next message is received. */
541
542 #ifndef SIG_IGN_WORKS
543 while (waitpid(-1, NULL, WNOHANG) > 0);
544 #endif
545
546 /* Reclaim up the store used in accepting this message */
547
548 store_reset(reset_point);
549
550 /* If queue_only is set or if there are too many incoming connections in
551 existence, local_queue_only will be TRUE. If it is not, check whether we
552 have received too many messages in this session for immediate delivery. If
553 not, and queue_only_load is set, check that the load average is below it.
554 Note that, once set, local_queue_only remains set for any subsequent
555 messages on the same SMTP connection. This is a deliberate choice; even
556 though the load average may fall, it doesn't seem right to deliver later
557 messages on the same call when not delivering earlier ones. */
558
559 if (!local_queue_only)
560 {
561 if (smtp_accept_queue_per_connection > 0 &&
562 receive_messagecount > smtp_accept_queue_per_connection)
563 {
564 local_queue_only = TRUE;
565 queue_only_reason = 2;
566 }
567 else if (queue_only_load >= 0)
568 {
569 local_queue_only = (load_average = os_getloadavg()) > queue_only_load;
570 if (local_queue_only) queue_only_reason = 3;
571 }
572 }
573
574 /* Log the queueing here, when it will get a message id attached, but
575 not if queue_only is set (case 0). */
576
577 if (local_queue_only) switch(queue_only_reason)
578 {
579 case 1:
580 log_write(L_delay_delivery,
581 LOG_MAIN, "no immediate delivery: too many connections "
582 "(%d, max %d)", smtp_accept_count, smtp_accept_queue);
583 break;
584
585 case 2:
586 log_write(L_delay_delivery,
587 LOG_MAIN, "no immediate delivery: more than %d messages "
588 "received in one connection", smtp_accept_queue_per_connection);
589 break;
590
591 case 3:
592 log_write(L_delay_delivery,
593 LOG_MAIN, "no immediate delivery: load average %.2f",
594 (double)load_average/1000.0);
595 break;
596 }
597
598 /* If a delivery attempt is required, spin off a new process to handle it.
599 If we are not root, we have to re-exec exim unless deliveries are being
600 done unprivileged. */
601
602 else if (!queue_only_policy && !deliver_freeze)
603 {
604 pid_t dpid;
605
606 /* Before forking, ensure that the C output buffer is flushed. Otherwise
607 anything that it in it will get duplicated, leading to duplicate copies
608 of the pending output. */
609
610 mac_smtp_fflush();
611
612 if ((dpid = fork()) == 0)
613 {
614 (void)fclose(smtp_in);
615 (void)fclose(smtp_out);
616
617 /* Don't ever molest the parent's SSL connection, but do clean up
618 the data structures if necessary. */
619
620 #ifdef SUPPORT_TLS
621 tls_close(FALSE);
622 #endif
623
624 /* Reset SIGHUP and SIGCHLD in the child in both cases. */
625
626 signal(SIGHUP, SIG_DFL);
627 signal(SIGCHLD, SIG_DFL);
628
629 if (geteuid() != root_uid && !deliver_drop_privilege)
630 {
631 signal(SIGALRM, SIG_DFL);
632 (void)child_exec_exim(CEE_EXEC_PANIC, FALSE, NULL, FALSE, 2, US"-Mc",
633 message_id);
634 /* Control does not return here. */
635 }
636
637 /* No need to re-exec; SIGALRM remains set to the default handler */
638
639 (void)deliver_message(message_id, FALSE, FALSE);
640 search_tidyup();
641 _exit(EXIT_SUCCESS);
642 }
643
644 if (dpid > 0)
645 {
646 DEBUG(D_any) debug_printf("forked delivery process %d\n", (int)dpid);
647 }
648 else
649 {
650 log_write(0, LOG_MAIN|LOG_PANIC, "daemon: delivery process fork "
651 "failed: %s", strerror(errno));
652 }
653 }
654 }
655 }
656
657
658 /* Carrying on in the parent daemon process... Can't do much if the fork
659 failed. Otherwise, keep count of the number of accepting processes and
660 remember the pid for ticking off when the child completes. */
661
662 if (pid < 0)
663 {
664 never_error(US"daemon: accept process fork failed", US"Fork failed", errno);
665 }
666 else
667 {
668 int i;
669 for (i = 0; i < smtp_accept_max; ++i)
670 {
671 if (smtp_slots[i].pid <= 0)
672 {
673 smtp_slots[i].pid = pid;
674 if (smtp_accept_max_per_host != NULL)
675 smtp_slots[i].host_address = string_copy_malloc(sender_host_address);
676 smtp_accept_count++;
677 break;
678 }
679 }
680 DEBUG(D_any) debug_printf("%d SMTP accept process%s running\n",
681 smtp_accept_count, (smtp_accept_count == 1)? "" : "es");
682 }
683
684 /* Get here via goto in error cases */
685
686 ERROR_RETURN:
687
688 /* Close the streams associated with the socket which will also close the
689 socket fds in this process. We can't do anything if fclose() fails, but
690 logging brings it to someone's attention. However, "connection reset by peer"
691 isn't really a problem, so skip that one. On Solaris, a dropped connection can
692 manifest itself as a broken pipe, so drop that one too. If the streams don't
693 exist, something went wrong while setting things up. Make sure the socket
694 descriptors are closed, in order to drop the connection. */
695
696 if (smtp_out != NULL)
697 {
698 if (fclose(smtp_out) != 0 && errno != ECONNRESET && errno != EPIPE)
699 log_write(0, LOG_MAIN|LOG_PANIC, "daemon: fclose(smtp_out) failed: %s",
700 strerror(errno));
701 smtp_out = NULL;
702 }
703 else (void)close(accept_socket);
704
705 if (smtp_in != NULL)
706 {
707 if (fclose(smtp_in) != 0 && errno != ECONNRESET && errno != EPIPE)
708 log_write(0, LOG_MAIN|LOG_PANIC, "daemon: fclose(smtp_in) failed: %s",
709 strerror(errno));
710 smtp_in = NULL;
711 }
712 else (void)close(dup_accept_socket);
713
714 /* Release any store used in this process, including the store used for holding
715 the incoming host address and an expanded active_hostname. */
716
717 store_reset(reset_point);
718 sender_host_address = NULL;
719 }
720
721
722
723
724 /*************************************************
725 * Check wildcard listen special cases *
726 *************************************************/
727
728 /* This function is used when binding and listening on lists of addresses and
729 ports. It tests for special cases of wildcard listening, when IPv4 and IPv6
730 sockets may interact in different ways in different operating systems. It is
731 passed an error number, the list of listening addresses, and the current
732 address. Two checks are available: for a previous wildcard IPv6 address, or for
733 a following wildcard IPv4 address, in both cases on the same port.
734
735 In practice, pairs of wildcard addresses should be adjacent in the address list
736 because they are sorted that way below.
737
738 Arguments:
739 eno the error number
740 addresses the list of addresses
741 ipa the current IP address
742 back if TRUE, check for previous wildcard IPv6 address
743 if FALSE, check for a following wildcard IPv4 address
744
745 Returns: TRUE or FALSE
746 */
747
748 static BOOL
749 check_special_case(int eno, ip_address_item *addresses, ip_address_item *ipa,
750 BOOL back)
751 {
752 ip_address_item *ipa2;
753
754 /* For the "back" case, if the failure was "address in use" for a wildcard IPv4
755 address, seek a previous IPv6 wildcard address on the same port. As it is
756 previous, it must have been successfully bound and be listening. Flag it as a
757 "6 including 4" listener. */
758
759 if (back)
760 {
761 if (eno != EADDRINUSE || ipa->address[0] != 0) return FALSE;
762 for (ipa2 = addresses; ipa2 != ipa; ipa2 = ipa2->next)
763 {
764 if (ipa2->address[1] == 0 && ipa2->port == ipa->port)
765 {
766 ipa2->v6_include_v4 = TRUE;
767 return TRUE;
768 }
769 }
770 }
771
772 /* For the "forward" case, if the current address is a wildcard IPv6 address,
773 we seek a following wildcard IPv4 address on the same port. */
774
775 else
776 {
777 if (ipa->address[0] != ':' || ipa->address[1] != 0) return FALSE;
778 for (ipa2 = ipa->next; ipa2 != NULL; ipa2 = ipa2->next)
779 if (ipa2->address[0] == 0 && ipa->port == ipa2->port) return TRUE;
780 }
781
782 return FALSE;
783 }
784
785
786
787
788 /*************************************************
789 * Handle terminating subprocesses *
790 *************************************************/
791
792 /* Handle the termination of child processes. Theoretically, this need be done
793 only when sigchld_seen is TRUE, but rumour has it that some systems lose
794 SIGCHLD signals at busy times, so to be on the safe side, this function is
795 called each time round. It shouldn't be too expensive.
796
797 Arguments: none
798 Returns: nothing
799 */
800
801 static void
802 handle_ending_processes(void)
803 {
804 int status;
805 pid_t pid;
806
807 while ((pid = waitpid(-1, &status, WNOHANG)) > 0)
808 {
809 int i;
810 DEBUG(D_any) debug_printf("child %d ended: status=0x%x\n", (int)pid,
811 status);
812
813 /* If it's a listening daemon for which we are keeping track of individual
814 subprocesses, deal with an accepting process that has terminated. */
815
816 if (smtp_slots != NULL)
817 {
818 for (i = 0; i < smtp_accept_max; i++)
819 {
820 if (smtp_slots[i].pid == pid)
821 {
822 if (smtp_slots[i].host_address != NULL)
823 store_free(smtp_slots[i].host_address);
824 smtp_slots[i] = empty_smtp_slot;
825 if (--smtp_accept_count < 0) smtp_accept_count = 0;
826 DEBUG(D_any) debug_printf("%d SMTP accept process%s now running\n",
827 smtp_accept_count, (smtp_accept_count == 1)? "" : "es");
828 break;
829 }
830 }
831 if (i < smtp_accept_max) continue; /* Found an accepting process */
832 }
833
834 /* If it wasn't an accepting process, see if it was a queue-runner
835 process that we are tracking. */
836
837 if (queue_pid_slots != NULL)
838 {
839 for (i = 0; i < queue_run_max; i++)
840 {
841 if (queue_pid_slots[i] == pid)
842 {
843 queue_pid_slots[i] = 0;
844 if (--queue_run_count < 0) queue_run_count = 0;
845 DEBUG(D_any) debug_printf("%d queue-runner process%s now running\n",
846 queue_run_count, (queue_run_count == 1)? "" : "es");
847 break;
848 }
849 }
850 }
851 }
852 }
853
854
855
856 /*************************************************
857 * Exim Daemon Mainline *
858 *************************************************/
859
860 /* The daemon can do two jobs, either of which is optional:
861
862 (1) Listens for incoming SMTP calls and spawns off a sub-process to handle
863 each one. This is requested by the -bd option, with -oX specifying the SMTP
864 port on which to listen (for testing).
865
866 (2) Spawns a queue-running process every so often. This is controlled by the
867 -q option with a an interval time. (If no time is given, a single queue run
868 is done from the main function, and control doesn't get here.)
869
870 Root privilege is required in order to attach to port 25. Some systems require
871 it when calling socket() rather than bind(). To cope with all cases, we run as
872 root for both socket() and bind(). Some systems also require root in order to
873 write to the pid file directory. This function must therefore be called as root
874 if it is to work properly in all circumstances. Once the socket is bound and
875 the pid file written, root privilege is given up if there is an exim uid.
876
877 There are no arguments to this function, and it never returns. */
878
879 void
880 daemon_go(void)
881 {
882 int *listen_sockets = NULL;
883 int listen_socket_count = 0;
884 ip_address_item *addresses = NULL;
885
886 /* If any debugging options are set, turn on the D_pid bit so that all
887 debugging lines get the pid added. */
888
889 DEBUG(D_any|D_v) debug_selector |= D_pid;
890
891
892 /* Do the preparation for setting up a listener on one or more interfaces, and
893 possible on various ports. This is controlled by the combination of
894 local_interfaces (which can set IP addresses and ports) and daemon_smtp_port
895 (which is a list of default ports to use for those items in local_interfaces
896 that do not specify a port). The -oX command line option can be used to
897 override one or both of these options.
898
899 If local_interfaces is not set, the default is to listen on all interfaces.
900 When it is set, it can include "all IPvx interfaces" as an item. This is useful
901 when different ports are in use.
902
903 It turns out that listening on all interfaces is messy in an IPv6 world,
904 because several different implementation approaches have been taken. This code
905 is now supposed to work with all of them. The point of difference is whether an
906 IPv6 socket that is listening on all interfaces will receive incoming IPv4
907 calls or not. We also have to cope with the case when IPv6 libraries exist, but
908 there is no IPv6 support in the kernel.
909
910 . On Solaris, an IPv6 socket will accept IPv4 calls, and give them as mapped
911 addresses. However, if an IPv4 socket is also listening on all interfaces,
912 calls are directed to the appropriate socket.
913
914 . On (some versions of) Linux, an IPv6 socket will accept IPv4 calls, and
915 give them as mapped addresses, but an attempt also to listen on an IPv4
916 socket on all interfaces causes an error.
917
918 . On OpenBSD, an IPv6 socket will not accept IPv4 calls. You have to set up
919 two sockets if you want to accept both kinds of call.
920
921 . FreeBSD is like OpenBSD, but it has the IPV6_V6ONLY socket option, which
922 can be turned off, to make it behave like the versions of Linux described
923 above.
924
925 . I heard a report that the USAGI IPv6 stack for Linux has implemented
926 IPV6_V6ONLY.
927
928 So, what we do when IPv6 is supported is as follows:
929
930 (1) After it is set up, the list of interfaces is scanned for wildcard
931 addresses. If an IPv6 and an IPv4 wildcard are both found for the same
932 port, the list is re-arranged so that they are together, with the IPv6
933 wildcard first.
934
935 (2) If the creation of a wildcard IPv6 socket fails, we just log the error and
936 carry on if an IPv4 wildcard socket for the same port follows later in the
937 list. This allows Exim to carry on in the case when the kernel has no IPv6
938 support.
939
940 (3) Having created an IPv6 wildcard socket, we try to set IPV6_V6ONLY if that
941 option is defined. However, if setting fails, carry on regardless (but log
942 the incident).
943
944 (4) If binding or listening on an IPv6 wildcard socket fails, it is a serious
945 error.
946
947 (5) If binding or listening on an IPv4 wildcard socket fails with the error
948 EADDRINUSE, and a previous interface was an IPv6 wildcard for the same
949 port (which must have succeeded or we wouldn't have got this far), we
950 assume we are in the situation where just a single socket is permitted,
951 and ignore the error.
952
953 Phew!
954
955 The preparation code decodes options and sets up the relevant data. We do this
956 first, so that we can return non-zero if there are any syntax errors, and also
957 write to stderr. */
958
959 if (daemon_listen)
960 {
961 int *default_smtp_port;
962 int sep;
963 int pct = 0;
964 uschar *s;
965 uschar *list;
966 uschar *local_iface_source = US"local_interfaces";
967 ip_address_item *ipa;
968 ip_address_item **pipa;
969
970 /* If any option requiring a load average to be available during the
971 reception of a message is set, call os_getloadavg() while we are root
972 for those OS for which this is necessary the first time it is called (in
973 order to perform an "open" on the kernel memory file). */
974
975 #ifdef LOAD_AVG_NEEDS_ROOT
976 if (queue_only_load >= 0 || smtp_load_reserve >= 0 ||
977 (deliver_queue_load_max >= 0 && deliver_drop_privilege))
978 (void)os_getloadavg();
979 #endif
980
981 /* If -oX was used, disable the writing of a pid file unless -oP was
982 explicitly used to force it. Then scan the string given to -oX. Any items
983 that contain neither a dot nor a colon are used to override daemon_smtp_port.
984 Any other items are used to override local_interfaces. */
985
986 if (override_local_interfaces != NULL)
987 {
988 uschar *new_smtp_port = NULL;
989 uschar *new_local_interfaces = NULL;
990 int portsize = 0;
991 int portptr = 0;
992 int ifacesize = 0;
993 int ifaceptr = 0;
994
995 if (override_pid_file_path == NULL) write_pid = FALSE;
996
997 list = override_local_interfaces;
998 sep = 0;
999 while ((s = string_nextinlist(&list,&sep,big_buffer,big_buffer_size))
1000 != NULL)
1001 {
1002 uschar joinstr[4];
1003 uschar **ptr;
1004 int *sizeptr;
1005 int *ptrptr;
1006
1007 if (Ustrpbrk(s, ".:") == NULL)
1008 {
1009 ptr = &new_smtp_port;
1010 sizeptr = &portsize;
1011 ptrptr = &portptr;
1012 }
1013 else
1014 {
1015 ptr = &new_local_interfaces;
1016 sizeptr = &ifacesize;
1017 ptrptr = &ifaceptr;
1018 }
1019
1020 if (*ptr == NULL)
1021 {
1022 joinstr[0] = sep;
1023 joinstr[1] = ' ';
1024 *ptr = string_cat(*ptr, sizeptr, ptrptr, US"<", 1);
1025 }
1026
1027 *ptr = string_cat(*ptr, sizeptr, ptrptr, joinstr, 2);
1028 *ptr = string_cat(*ptr, sizeptr, ptrptr, s, Ustrlen(s));
1029 }
1030
1031 if (new_smtp_port != NULL)
1032 {
1033 new_smtp_port[portptr] = 0;
1034 daemon_smtp_port = new_smtp_port;
1035 DEBUG(D_any) debug_printf("daemon_smtp_port overridden by -oX:\n %s\n",
1036 daemon_smtp_port);
1037 }
1038
1039 if (new_local_interfaces != NULL)
1040 {
1041 new_local_interfaces[ifaceptr] = 0;
1042 local_interfaces = new_local_interfaces;
1043 local_iface_source = US"-oX data";
1044 DEBUG(D_any) debug_printf("local_interfaces overridden by -oX:\n %s\n",
1045 local_interfaces);
1046 }
1047 }
1048
1049 /* Create a list of default SMTP ports, to be used if local_interfaces
1050 contains entries without explict ports. First count the number of ports, then
1051 build a translated list in a vector. */
1052
1053 list = daemon_smtp_port;
1054 sep = 0;
1055 while ((s = string_nextinlist(&list,&sep,big_buffer,big_buffer_size)) != NULL)
1056 pct++;
1057 default_smtp_port = store_get((pct+1) * sizeof(int));
1058 list = daemon_smtp_port;
1059 sep = 0;
1060 for (pct = 0;
1061 (s = string_nextinlist(&list,&sep,big_buffer,big_buffer_size)) != NULL;
1062 pct++)
1063 {
1064 if (isdigit(*s))
1065 {
1066 uschar *end;
1067 default_smtp_port[pct] = Ustrtol(s, &end, 0);
1068 if (end != s + Ustrlen(s))
1069 log_write(0, LOG_PANIC_DIE|LOG_CONFIG, "invalid SMTP port: %s", s);
1070 }
1071 else
1072 {
1073 struct servent *smtp_service = getservbyname(CS s, "tcp");
1074 if (smtp_service == NULL)
1075 log_write(0, LOG_PANIC_DIE|LOG_CONFIG, "TCP port \"%s\" not found", s);
1076 default_smtp_port[pct] = ntohs(smtp_service->s_port);
1077 }
1078 }
1079 default_smtp_port[pct] = 0;
1080
1081 /* Create the list of local interfaces, possibly with ports included. This
1082 list may contain references to 0.0.0.0 and ::0 as wildcards. These special
1083 values are converted below. */
1084
1085 addresses = host_build_ifacelist(local_interfaces, local_iface_source);
1086
1087 /* In the list of IP addresses, convert 0.0.0.0 into an empty string, and ::0
1088 into the string ":". We use these to recognize wildcards in IPv4 and IPv6. In
1089 fact, many IP stacks recognize 0.0.0.0 and ::0 and handle them as wildcards
1090 anyway, but we need to know which are the wildcard addresses, and the shorter
1091 strings are neater.
1092
1093 In the same scan, fill in missing port numbers from the default list. When
1094 there is more than one item in the list, extra items are created. */
1095
1096 for (ipa = addresses; ipa != NULL; ipa = ipa->next)
1097 {
1098 int i;
1099
1100 if (Ustrcmp(ipa->address, "0.0.0.0") == 0) ipa->address[0] = 0;
1101 else if (Ustrcmp(ipa->address, "::0") == 0)
1102 {
1103 ipa->address[0] = ':';
1104 ipa->address[1] = 0;
1105 }
1106
1107 if (ipa->port > 0) continue;
1108
1109 if (daemon_smtp_port[0] <= 0)
1110 log_write(0, LOG_MAIN|LOG_PANIC_DIE, "no port specified for interface "
1111 "%s and daemon_smtp_port is unset; cannot start daemon",
1112 (ipa->address[0] == 0)? US"\"all IPv4\"" :
1113 (ipa->address[1] == 0)? US"\"all IPv6\"" : ipa->address);
1114 ipa->port = default_smtp_port[0];
1115 for (i = 1; default_smtp_port[i] > 0; i++)
1116 {
1117 ip_address_item *new = store_get(sizeof(ip_address_item));
1118 memcpy(new->address, ipa->address, Ustrlen(ipa->address) + 1);
1119 new->port = default_smtp_port[i];
1120 new->next = ipa->next;
1121 ipa->next = new;
1122 ipa = new;
1123 }
1124 }
1125
1126 /* Scan the list of addresses for wildcards. If we find an IPv4 and an IPv6
1127 wildcard for the same port, ensure that (a) they are together and (b) the
1128 IPv6 address comes first. This makes handling the messy features easier, and
1129 also simplifies the construction of the "daemon started" log line. */
1130
1131 pipa = &addresses;
1132 for (ipa = addresses; ipa != NULL; pipa = &(ipa->next), ipa = ipa->next)
1133 {
1134 ip_address_item *ipa2;
1135
1136 /* Handle an IPv4 wildcard */
1137
1138 if (ipa->address[0] == 0)
1139 {
1140 for (ipa2 = ipa; ipa2->next != NULL; ipa2 = ipa2->next)
1141 {
1142 ip_address_item *ipa3 = ipa2->next;
1143 if (ipa3->address[0] == ':' &&
1144 ipa3->address[1] == 0 &&
1145 ipa3->port == ipa->port)
1146 {
1147 ipa2->next = ipa3->next;
1148 ipa3->next = ipa;
1149 *pipa = ipa3;
1150 break;
1151 }
1152 }
1153 }
1154
1155 /* Handle an IPv6 wildcard. */
1156
1157 else if (ipa->address[0] == ':' && ipa->address[1] == 0)
1158 {
1159 for (ipa2 = ipa; ipa2->next != NULL; ipa2 = ipa2->next)
1160 {
1161 ip_address_item *ipa3 = ipa2->next;
1162 if (ipa3->address[0] == 0 && ipa3->port == ipa->port)
1163 {
1164 ipa2->next = ipa3->next;
1165 ipa3->next = ipa->next;
1166 ipa->next = ipa3;
1167 ipa = ipa3;
1168 break;
1169 }
1170 }
1171 }
1172 }
1173
1174 /* Get a vector to remember all the sockets in */
1175
1176 for (ipa = addresses; ipa != NULL; ipa = ipa->next)
1177 listen_socket_count++;
1178 listen_sockets = store_get(sizeof(int *) * listen_socket_count);
1179
1180 /* Do a sanity check on the max connects value just to save us from getting
1181 a huge amount of store. */
1182
1183 if (smtp_accept_max > 4095) smtp_accept_max = 4096;
1184
1185 /* There's no point setting smtp_accept_queue unless it is less than the max
1186 connects limit. The configuration reader ensures that the max is set if the
1187 queue-only option is set. */
1188
1189 if (smtp_accept_queue > smtp_accept_max) smtp_accept_queue = 0;
1190
1191 /* Get somewhere to keep the list of SMTP accepting pids if we are keeping
1192 track of them for total number and queue/host limits. */
1193
1194 if (smtp_accept_max > 0)
1195 {
1196 int i;
1197 smtp_slots = store_get(smtp_accept_max * sizeof(smtp_slot));
1198 for (i = 0; i < smtp_accept_max; i++) smtp_slots[i] = empty_smtp_slot;
1199 }
1200 }
1201
1202 /* The variable background_daemon is always false when debugging, but
1203 can also be forced false in order to keep a non-debugging daemon in the
1204 foreground. If background_daemon is true, close all open file descriptors that
1205 we know about, but then re-open stdin, stdout, and stderr to /dev/null.
1206
1207 This is protection against any called functions (in libraries, or in
1208 Perl, or whatever) that think they can write to stderr (or stdout). Before this
1209 was added, it was quite likely that an SMTP connection would use one of these
1210 file descriptors, in which case writing random stuff to it caused chaos.
1211
1212 Then disconnect from the controlling terminal, Most modern Unixes seem to have
1213 setsid() for getting rid of the controlling terminal. For any OS that doesn't,
1214 setsid() can be #defined as a no-op, or as something else. */
1215
1216 if (background_daemon)
1217 {
1218 log_close_all(); /* Just in case anything was logged earlier */
1219 search_tidyup(); /* Just in case any were used in reading the config. */
1220 (void)close(0); /* Get rid of stdin/stdout/stderr */
1221 (void)close(1);
1222 (void)close(2);
1223 exim_nullstd(); /* Connect stdin/stdout/stderr to /dev/null */
1224 log_stderr = NULL; /* So no attempt to copy paniclog output */
1225
1226 /* If the parent process of this one has pid == 1, we are re-initializing the
1227 daemon as the result of a SIGHUP. In this case, there is no need to do
1228 anything, because the controlling terminal has long gone. Otherwise, fork, in
1229 case current process is a process group leader (see 'man setsid' for an
1230 explanation) before calling setsid(). */
1231
1232 if (getppid() != 1)
1233 {
1234 pid_t pid = fork();
1235 if (pid < 0) log_write(0, LOG_MAIN|LOG_PANIC_DIE,
1236 "fork() failed when starting daemon: %s", strerror(errno));
1237 if (pid > 0) exit(EXIT_SUCCESS); /* in parent process, just exit */
1238 (void)setsid(); /* release controlling terminal */
1239 }
1240 }
1241
1242 /* We are now in the disconnected, daemon process (unless debugging). Set up
1243 the listening sockets if required. */
1244
1245 if (daemon_listen)
1246 {
1247 int sk;
1248 int on = 1;
1249 ip_address_item *ipa;
1250
1251 /* For each IP address, create a socket, bind it to the appropriate port, and
1252 start listening. See comments above about IPv6 sockets that may or may not
1253 accept IPv4 calls when listening on all interfaces. We also have to cope with
1254 the case of a system with IPv6 libraries, but no IPv6 support in the kernel.
1255 listening, provided a wildcard IPv4 socket for the same port follows. */
1256
1257 for (ipa = addresses, sk = 0; sk < listen_socket_count; ipa = ipa->next, sk++)
1258 {
1259 BOOL wildcard;
1260 ip_address_item *ipa2;
1261 int af;
1262
1263 if (Ustrchr(ipa->address, ':') != NULL)
1264 {
1265 af = AF_INET6;
1266 wildcard = ipa->address[1] == 0;
1267 }
1268 else
1269 {
1270 af = AF_INET;
1271 wildcard = ipa->address[0] == 0;
1272 }
1273
1274 listen_sockets[sk] = ip_socket(SOCK_STREAM, af);
1275 if (listen_sockets[sk] < 0)
1276 {
1277 if (check_special_case(0, addresses, ipa, FALSE))
1278 {
1279 log_write(0, LOG_MAIN, "Failed to create IPv6 socket for wildcard "
1280 "listening (%s): will use IPv4", strerror(errno));
1281 goto SKIP_SOCKET;
1282 }
1283 log_write(0, LOG_PANIC_DIE, "IPv%c socket creation failed: %s",
1284 (af == AF_INET6)? '6' : '4', strerror(errno));
1285 }
1286
1287 /* If this is an IPv6 wildcard socket, set IPV6_V6ONLY if that option is
1288 available. Just log failure (can get protocol not available, just like
1289 socket creation can). */
1290
1291 #ifdef IPV6_V6ONLY
1292 if (af == AF_INET6 && wildcard &&
1293 setsockopt(listen_sockets[sk], IPPROTO_IPV6, IPV6_V6ONLY, (char *)(&on),
1294 sizeof(on)) < 0)
1295 log_write(0, LOG_MAIN, "Setting IPV6_V6ONLY on daemon's IPv6 wildcard "
1296 "socket failed (%s): carrying on without it", strerror(errno));
1297 #endif /* IPV6_V6ONLY */
1298
1299 /* Set SO_REUSEADDR so that the daemon can be restarted while a connection
1300 is being handled. Without this, a connection will prevent reuse of the
1301 smtp port for listening. */
1302
1303 if (setsockopt(listen_sockets[sk], SOL_SOCKET, SO_REUSEADDR,
1304 (uschar *)(&on), sizeof(on)) < 0)
1305 log_write(0, LOG_MAIN|LOG_PANIC_DIE, "setting SO_REUSEADDR on socket "
1306 "failed when starting daemon: %s", strerror(errno));
1307
1308 /* Set TCP_NODELAY; Exim does its own buffering. There is a switch to
1309 disable this because it breaks some broken clients. */
1310
1311 if (tcp_nodelay) setsockopt(listen_sockets[sk], IPPROTO_TCP, TCP_NODELAY,
1312 (uschar *)(&on), sizeof(on));
1313
1314 /* Now bind the socket to the required port; if Exim is being restarted
1315 it may not always be possible to bind immediately, even with SO_REUSEADDR
1316 set, so try 10 times, waiting between each try. After 10 failures, we give
1317 up. In an IPv6 environment, if bind () fails with the error EADDRINUSE and
1318 we are doing wildcard IPv4 listening and there was a previous IPv6 wildcard
1319 address for the same port, ignore the error on the grounds that we must be
1320 in a system where the IPv6 socket accepts both kinds of call. This is
1321 necessary for (some release of) USAGI Linux; other IP stacks fail at the
1322 listen() stage instead. */
1323
1324 for(;;)
1325 {
1326 uschar *msg, *addr;
1327 if (ip_bind(listen_sockets[sk], af, ipa->address, ipa->port) >= 0) break;
1328 if (check_special_case(errno, addresses, ipa, TRUE))
1329 {
1330 DEBUG(D_any) debug_printf("wildcard IPv4 bind() failed after IPv6 "
1331 "listen() success; EADDRINUSE ignored\n");
1332 (void)close(listen_sockets[sk]);
1333 goto SKIP_SOCKET;
1334 }
1335 msg = US strerror(errno);
1336 addr = wildcard? ((af == AF_INET6)? US"(any IPv6)" : US"(any IPv4)") :
1337 ipa->address;
1338 if (daemon_startup_retries <= 0)
1339 log_write(0, LOG_MAIN|LOG_PANIC_DIE,
1340 "socket bind() to port %d for address %s failed: %s: "
1341 "daemon abandoned", ipa->port, addr, msg);
1342 log_write(0, LOG_MAIN, "socket bind() to port %d for address %s "
1343 "failed: %s: waiting %s before trying again (%d more %s)",
1344 ipa->port, addr, msg, readconf_printtime(daemon_startup_sleep),
1345 daemon_startup_retries, (daemon_startup_retries > 1)? "tries" : "try");
1346 daemon_startup_retries--;
1347 sleep(daemon_startup_sleep);
1348 }
1349
1350 DEBUG(D_any)
1351 {
1352 if (wildcard)
1353 debug_printf("listening on all interfaces (IPv%c) port %d\n",
1354 (af == AF_INET6)? '6' : '4', ipa->port);
1355 else
1356 debug_printf("listening on %s port %d\n", ipa->address, ipa->port);
1357 }
1358
1359 /* Start listening on the bound socket, establishing the maximum backlog of
1360 connections that is allowed. On success, continue to the next address. */
1361
1362 if (listen(listen_sockets[sk], smtp_connect_backlog) >= 0) continue;
1363
1364 /* Listening has failed. In an IPv6 environment, as for bind(), if listen()
1365 fails with the error EADDRINUSE and we are doing IPv4 wildcard listening
1366 and there was a previous successful IPv6 wildcard listen on the same port,
1367 we want to ignore the error on the grounds that we must be in a system
1368 where the IPv6 socket accepts both kinds of call. */
1369
1370 if (!check_special_case(errno, addresses, ipa, TRUE))
1371 log_write(0, LOG_PANIC_DIE, "listen() failed on interface %s: %s",
1372 wildcard? ((af == AF_INET6)? US"(any IPv6)" : US"(any IPv4)") :
1373 ipa->address,
1374 strerror(errno));
1375
1376 DEBUG(D_any) debug_printf("wildcard IPv4 listen() failed after IPv6 "
1377 "listen() success; EADDRINUSE ignored\n");
1378 (void)close(listen_sockets[sk]);
1379
1380 /* Come here if there has been a problem with the socket which we
1381 are going to ignore. We remove the address from the chain, and back up the
1382 counts. */
1383
1384 SKIP_SOCKET:
1385 sk--; /* Back up the count */
1386 listen_socket_count--; /* Reduce the total */
1387 if (ipa == addresses) addresses = ipa->next; else
1388 {
1389 for (ipa2 = addresses; ipa2->next != ipa; ipa2 = ipa2->next);
1390 ipa2->next = ipa->next;
1391 ipa = ipa2;
1392 }
1393 } /* End of bind/listen loop for each address */
1394 } /* End of setup for listening */
1395
1396
1397 /* If we are not listening, we want to write a pid file only if -oP was
1398 explicitly given. */
1399
1400 else if (override_pid_file_path == NULL) write_pid = FALSE;
1401
1402 /* Write the pid to a known file for assistance in identification, if required.
1403 We do this before giving up root privilege, because on some systems it is
1404 necessary to be root in order to write into the pid file directory. There's
1405 nothing to stop multiple daemons running, as long as no more than one listens
1406 on a given TCP/IP port on the same interface(s). However, in these
1407 circumstances it gets far too complicated to mess with pid file names
1408 automatically. Consequently, Exim 4 writes a pid file only
1409
1410 (a) When running in the test harness, or
1411 (b) When -bd is used and -oX is not used, or
1412 (c) When -oP is used to supply a path.
1413
1414 The variable daemon_write_pid is used to control this. */
1415
1416 if (running_in_test_harness || write_pid)
1417 {
1418 FILE *f;
1419
1420 if (override_pid_file_path != NULL)
1421 pid_file_path = override_pid_file_path;
1422
1423 if (pid_file_path[0] == 0)
1424 pid_file_path = string_sprintf("%s/exim-daemon.pid", spool_directory);
1425
1426 f = modefopen(pid_file_path, "wb", 0644);
1427 if (f != NULL)
1428 {
1429 (void)fprintf(f, "%d\n", (int)getpid());
1430 (void)fclose(f);
1431 DEBUG(D_any) debug_printf("pid written to %s\n", pid_file_path);
1432 }
1433 else
1434 {
1435 DEBUG(D_any)
1436 debug_printf("%s\n", string_open_failed(errno, "pid file %s",
1437 pid_file_path));
1438 }
1439 }
1440
1441 /* Set up the handler for SIGHUP, which causes a restart of the daemon. */
1442
1443 sighup_seen = FALSE;
1444 signal(SIGHUP, sighup_handler);
1445
1446 /* Give up root privilege at this point (assuming that exim_uid and exim_gid
1447 are not root). The third argument controls the running of initgroups().
1448 Normally we do this, in order to set up the groups for the Exim user. However,
1449 if we are not root at this time - some odd installations run that way - we
1450 cannot do this. */
1451
1452 exim_setugid(exim_uid, exim_gid, geteuid()==root_uid, US"running as a daemon");
1453
1454 /* Get somewhere to keep the list of queue-runner pids if we are keeping track
1455 of them (and also if we are doing queue runs). */
1456
1457 if (queue_interval > 0 && queue_run_max > 0)
1458 {
1459 int i;
1460 queue_pid_slots = store_get(queue_run_max * sizeof(pid_t));
1461 for (i = 0; i < queue_run_max; i++) queue_pid_slots[i] = 0;
1462 }
1463
1464 /* Set up the handler for termination of child processes. */
1465
1466 sigchld_seen = FALSE;
1467 os_non_restarting_signal(SIGCHLD, main_sigchld_handler);
1468
1469 /* If we are to run the queue periodically, pretend the alarm has just gone
1470 off. This will cause the first queue-runner to get kicked off straight away. */
1471
1472 sigalrm_seen = (queue_interval > 0);
1473
1474 /* Log the start up of a daemon - at least one of listening or queue running
1475 must be set up. */
1476
1477 if (daemon_listen)
1478 {
1479 int i, j;
1480 int smtp_ports = 0;
1481 int smtps_ports = 0;
1482 ip_address_item *ipa;
1483 uschar *p = big_buffer;
1484 uschar *qinfo = (queue_interval > 0)?
1485 string_sprintf("-q%s", readconf_printtime(queue_interval))
1486 :
1487 US"no queue runs";
1488
1489 /* Build a list of listening addresses in big_buffer, but limit it to 10
1490 items. The style is for backwards compatibility.
1491
1492 It is now possible to have some ports listening for SMTPS (the old,
1493 deprecated protocol that starts TLS without using STARTTLS), and others
1494 listening for standard SMTP. Keep their listings separate. */
1495
1496 for (j = 0; j < 2; j++)
1497 {
1498 for (i = 0, ipa = addresses; i < 10 && ipa != NULL; i++, ipa = ipa->next)
1499 {
1500 /* First time round, look for SMTP ports; second time round, look for
1501 SMTPS ports. For the first one of each, insert leading text. */
1502
1503 if (host_is_tls_on_connect_port(ipa->port) == (j > 0))
1504 {
1505 if (j == 0)
1506 {
1507 if (smtp_ports++ == 0)
1508 {
1509 memcpy(p, "SMTP on", 8);
1510 p += 7;
1511 }
1512 }
1513 else
1514 {
1515 if (smtps_ports++ == 0)
1516 {
1517 (void)sprintf(CS p, "%sSMTPS on",
1518 (smtp_ports == 0)? "":" and for ");
1519 while (*p != 0) p++;
1520 }
1521 }
1522
1523 /* Now the information about the port (and sometimes interface) */
1524
1525 if (ipa->address[0] == ':' && ipa->address[1] == 0)
1526 {
1527 if (ipa->next != NULL && ipa->next->address[0] == 0 &&
1528 ipa->next->port == ipa->port)
1529 {
1530 (void)sprintf(CS p, " port %d (IPv6 and IPv4)", ipa->port);
1531 ipa = ipa->next;
1532 }
1533 else if (ipa->v6_include_v4)
1534 (void)sprintf(CS p, " port %d (IPv6 with IPv4)", ipa->port);
1535 else
1536 (void)sprintf(CS p, " port %d (IPv6)", ipa->port);
1537 }
1538 else if (ipa->address[0] == 0)
1539 (void)sprintf(CS p, " port %d (IPv4)", ipa->port);
1540 else
1541 (void)sprintf(CS p, " [%s]:%d", ipa->address, ipa->port);
1542 while (*p != 0) p++;
1543 }
1544 }
1545
1546 if (ipa != NULL)
1547 {
1548 memcpy(p, " ...", 5);
1549 p += 4;
1550 }
1551 }
1552
1553 log_write(0, LOG_MAIN,
1554 "exim %s daemon started: pid=%d, %s, listening for %s",
1555 version_string, getpid(), qinfo, big_buffer);
1556 set_process_info("daemon: %s, listening for %s", qinfo, big_buffer);
1557 }
1558
1559 else
1560 {
1561 log_write(0, LOG_MAIN,
1562 "exim %s daemon started: pid=%d, -q%s, not listening for SMTP",
1563 version_string, getpid(), readconf_printtime(queue_interval));
1564 set_process_info("daemon: -q%s, not listening",
1565 readconf_printtime(queue_interval));
1566 }
1567
1568
1569 /* Close the log so it can be renamed and moved. In the few cases below where
1570 this long-running process writes to the log (always exceptional conditions), it
1571 closes the log afterwards, for the same reason. */
1572
1573 log_close_all();
1574
1575 DEBUG(D_any) debug_print_ids(US"daemon running with");
1576
1577 /* Any messages accepted via this route are going to be SMTP. */
1578
1579 smtp_input = TRUE;
1580
1581 /* Enter the never-ending loop... */
1582
1583 for (;;)
1584 {
1585 #if HAVE_IPV6
1586 struct sockaddr_in6 accepted;
1587 #else
1588 struct sockaddr_in accepted;
1589 #endif
1590
1591 EXIM_SOCKLEN_T len = sizeof(accepted);
1592 pid_t pid;
1593
1594 /* This code is placed first in the loop, so that it gets obeyed at the
1595 start, before the first wait. This causes the first queue-runner to be
1596 started immediately. */
1597
1598 if (sigalrm_seen)
1599 {
1600 DEBUG(D_any) debug_printf("SIGALRM received\n");
1601
1602 /* Do a full queue run in a child process, if required, unless we already
1603 have enough queue runners on the go. If we are not running as root, a
1604 re-exec is required. */
1605
1606 if (queue_interval > 0 &&
1607 (queue_run_max <= 0 || queue_run_count < queue_run_max))
1608 {
1609 if ((pid = fork()) == 0)
1610 {
1611 int sk;
1612
1613 DEBUG(D_any) debug_printf("Starting queue-runner: pid %d\n",
1614 (int)getpid());
1615
1616 /* Disable debugging if it's required only for the daemon process. We
1617 leave the above message, because it ties up with the "child ended"
1618 debugging messages. */
1619
1620 if (debug_daemon) debug_selector = 0;
1621
1622 /* Close any open listening sockets in the child */
1623
1624 for (sk = 0; sk < listen_socket_count; sk++)
1625 (void)close(listen_sockets[sk]);
1626
1627 /* Reset SIGHUP and SIGCHLD in the child in both cases. */
1628
1629 signal(SIGHUP, SIG_DFL);
1630 signal(SIGCHLD, SIG_DFL);
1631
1632 /* Re-exec if privilege has been given up, unless deliver_drop_
1633 privilege is set. Reset SIGALRM before exec(). */
1634
1635 if (geteuid() != root_uid && !deliver_drop_privilege)
1636 {
1637 uschar opt[8];
1638 uschar *p = opt;
1639 uschar *extra[4];
1640 int extracount = 1;
1641
1642 signal(SIGALRM, SIG_DFL);
1643 *p++ = '-';
1644 *p++ = 'q';
1645 if (queue_2stage) *p++ = 'q';
1646 if (queue_run_first_delivery) *p++ = 'i';
1647 if (queue_run_force) *p++ = 'f';
1648 if (deliver_force_thaw) *p++ = 'f';
1649 if (queue_run_local) *p++ = 'l';
1650 *p = 0;
1651 extra[0] = opt;
1652
1653 /* If -R or -S were on the original command line, ensure they get
1654 passed on. */
1655
1656 if (deliver_selectstring != NULL)
1657 {
1658 extra[extracount++] = deliver_selectstring_regex? US"-Rr" : US"-R";
1659 extra[extracount++] = deliver_selectstring;
1660 }
1661
1662 if (deliver_selectstring_sender != NULL)
1663 {
1664 extra[extracount++] = deliver_selectstring_sender_regex?
1665 US"-Sr" : US"-S";
1666 extra[extracount++] = deliver_selectstring_sender;
1667 }
1668
1669 /* Overlay this process with a new execution. */
1670
1671 (void)child_exec_exim(CEE_EXEC_PANIC, FALSE, NULL, TRUE, extracount,
1672 extra[0], extra[1], extra[2], extra[3], extra[4]);
1673
1674 /* Control never returns here. */
1675 }
1676
1677 /* No need to re-exec; SIGALRM remains set to the default handler */
1678
1679 queue_run(NULL, NULL, FALSE);
1680 _exit(EXIT_SUCCESS);
1681 }
1682
1683 if (pid < 0)
1684 {
1685 log_write(0, LOG_MAIN|LOG_PANIC, "daemon: fork of queue-runner "
1686 "process failed: %s", strerror(errno));
1687 log_close_all();
1688 }
1689 else
1690 {
1691 int i;
1692 for (i = 0; i < queue_run_max; ++i)
1693 {
1694 if (queue_pid_slots[i] <= 0)
1695 {
1696 queue_pid_slots[i] = pid;
1697 queue_run_count++;
1698 break;
1699 }
1700 }
1701 DEBUG(D_any) debug_printf("%d queue-runner process%s running\n",
1702 queue_run_count, (queue_run_count == 1)? "" : "es");
1703 }
1704 }
1705
1706 /* Reset the alarm clock */
1707
1708 sigalrm_seen = FALSE;
1709 alarm(queue_interval);
1710 }
1711
1712
1713 /* Sleep till a connection happens if listening, and handle the connection if
1714 that is why we woke up. The FreeBSD operating system requires the use of
1715 select() before accept() because the latter function is not interrupted by
1716 a signal, and we want to wake up for SIGCHLD and SIGALRM signals. Some other
1717 OS do notice signals in accept() but it does no harm to have the select()
1718 in for all of them - and it won't then be a lurking problem for ports to
1719 new OS. In fact, the later addition of listening on specific interfaces only
1720 requires this way of working anyway. */
1721
1722 if (daemon_listen)
1723 {
1724 int sk, lcount, select_errno;
1725 int max_socket = 0;
1726 BOOL select_failed = FALSE;
1727 fd_set select_listen;
1728
1729 FD_ZERO(&select_listen);
1730 for (sk = 0; sk < listen_socket_count; sk++)
1731 {
1732 FD_SET(listen_sockets[sk], &select_listen);
1733 if (listen_sockets[sk] > max_socket) max_socket = listen_sockets[sk];
1734 }
1735
1736 DEBUG(D_any) debug_printf("Listening...\n");
1737
1738 /* In rare cases we may have had a SIGCHLD signal in the time between
1739 setting the handler (below) and getting back here. If so, pretend that the
1740 select() was interrupted so that we reap the child. This might still leave
1741 a small window when a SIGCHLD could get lost. However, since we use SIGCHLD
1742 only to do the reaping more quickly, it shouldn't result in anything other
1743 than a delay until something else causes a wake-up. */
1744
1745 if (sigchld_seen)
1746 {
1747 lcount = -1;
1748 errno = EINTR;
1749 }
1750 else
1751 {
1752 lcount = select(max_socket + 1, (SELECT_ARG2_TYPE *)&select_listen,
1753 NULL, NULL, NULL);
1754 }
1755
1756 if (lcount < 0)
1757 {
1758 select_failed = TRUE;
1759 lcount = 1;
1760 }
1761
1762 /* Clean up any subprocesses that may have terminated. We need to do this
1763 here so that smtp_accept_max_per_host works when a connection to that host
1764 has completed, and we are about to accept a new one. When this code was
1765 later in the sequence, a new connection could be rejected, even though an
1766 old one had just finished. Preserve the errno from any select() failure for
1767 the use of the common select/accept error processing below. */
1768
1769 select_errno = errno;
1770 handle_ending_processes();
1771 errno = select_errno;
1772
1773 /* Loop for all the sockets that are currently ready to go. If select
1774 actually failed, we have set the count to 1 and select_failed=TRUE, so as
1775 to use the common error code for select/accept below. */
1776
1777 while (lcount-- > 0)
1778 {
1779 int accept_socket = -1;
1780 if (!select_failed)
1781 {
1782 for (sk = 0; sk < listen_socket_count; sk++)
1783 {
1784 if (FD_ISSET(listen_sockets[sk], &select_listen))
1785 {
1786 accept_socket = accept(listen_sockets[sk],
1787 (struct sockaddr *)&accepted, &len);
1788 FD_CLR(listen_sockets[sk], &select_listen);
1789 break;
1790 }
1791 }
1792 }
1793
1794 /* If select or accept has failed and this was not caused by an
1795 interruption, log the incident and try again. With asymmetric TCP/IP
1796 routing errors such as "No route to network" have been seen here. Also
1797 "connection reset by peer" has been seen. These cannot be classed as
1798 disastrous errors, but they could fill up a lot of log. The code in smail
1799 crashes the daemon after 10 successive failures of accept, on the grounds
1800 that some OS fail continuously. Exim originally followed suit, but this
1801 appears to have caused problems. Now it just keeps going, but instead of
1802 logging each error, it batches them up when they are continuous. */
1803
1804 if (accept_socket < 0 && errno != EINTR)
1805 {
1806 if (accept_retry_count == 0)
1807 {
1808 accept_retry_errno = errno;
1809 accept_retry_select_failed = select_failed;
1810 }
1811 else
1812 {
1813 if (errno != accept_retry_errno ||
1814 select_failed != accept_retry_select_failed ||
1815 accept_retry_count >= 50)
1816 {
1817 log_write(0, LOG_MAIN | ((accept_retry_count >= 50)? LOG_PANIC : 0),
1818 "%d %s() failure%s: %s",
1819 accept_retry_count,
1820 accept_retry_select_failed? "select" : "accept",
1821 (accept_retry_count == 1)? "" : "s",
1822 strerror(accept_retry_errno));
1823 log_close_all();
1824 accept_retry_count = 0;
1825 accept_retry_errno = errno;
1826 accept_retry_select_failed = select_failed;
1827 }
1828 }
1829 accept_retry_count++;
1830 }
1831
1832 else
1833 {
1834 if (accept_retry_count > 0)
1835 {
1836 log_write(0, LOG_MAIN, "%d %s() failure%s: %s",
1837 accept_retry_count,
1838 accept_retry_select_failed? "select" : "accept",
1839 (accept_retry_count == 1)? "" : "s",
1840 strerror(accept_retry_errno));
1841 log_close_all();
1842 accept_retry_count = 0;
1843 }
1844 }
1845
1846 /* If select/accept succeeded, deal with the connection. */
1847
1848 if (accept_socket >= 0)
1849 handle_smtp_call(listen_sockets, listen_socket_count, accept_socket,
1850 (struct sockaddr *)&accepted);
1851 }
1852 }
1853
1854 /* If not listening, then just sleep for the queue interval. If we woke
1855 up early the last time for some other signal, it won't matter because
1856 the alarm signal will wake at the right time. This code originally used
1857 sleep() but it turns out that on the FreeBSD system, sleep() is not inter-
1858 rupted by signals, so it wasn't waking up for SIGALRM or SIGCHLD. Luckily
1859 select() can be used as an interruptible sleep() on all versions of Unix. */
1860
1861 else
1862 {
1863 struct timeval tv;
1864 tv.tv_sec = queue_interval;
1865 tv.tv_usec = 0;
1866 select(0, NULL, NULL, NULL, &tv);
1867 handle_ending_processes();
1868 }
1869
1870 /* Re-enable the SIGCHLD handler if it has been run. It can't do it
1871 for itself, because it isn't doing the waiting itself. */
1872
1873 if (sigchld_seen)
1874 {
1875 sigchld_seen = FALSE;
1876 os_non_restarting_signal(SIGCHLD, main_sigchld_handler);
1877 }
1878
1879 /* Handle being woken by SIGHUP. We know at this point that the result
1880 of accept() has been dealt with, so we can re-exec exim safely, first
1881 closing the listening sockets so that they can be reused. Cancel any pending
1882 alarm in case it is just about to go off, and set SIGHUP to be ignored so
1883 that another HUP in quick succession doesn't clobber the new daemon before it
1884 gets going. All log files get closed by the close-on-exec flag; however, if
1885 the exec fails, we need to close the logs. */
1886
1887 if (sighup_seen)
1888 {
1889 int sk;
1890 log_write(0, LOG_MAIN, "pid %d: SIGHUP received: re-exec daemon",
1891 getpid());
1892 for (sk = 0; sk < listen_socket_count; sk++)
1893 (void)close(listen_sockets[sk]);
1894 alarm(0);
1895 signal(SIGHUP, SIG_IGN);
1896 sighup_argv[0] = exim_path;
1897 exim_nullstd();
1898 execv(CS exim_path, (char *const *)sighup_argv);
1899 log_write(0, LOG_MAIN|LOG_PANIC_DIE, "pid %d: exec of %s failed: %s",
1900 getpid(), exim_path, strerror(errno));
1901 log_close_all();
1902 }
1903
1904 } /* End of main loop */
1905
1906 /* Control never reaches here */
1907 }
1908
1909 /* End of exim_daemon.c */
1910