No paniclog for ECONNRESET when getsockname() fails.
[exim.git] / src / src / daemon.c
1 /* $Cambridge: exim/src/src/daemon.c,v 1.17 2006/11/06 11:27:54 ph10 Exp $ */
2
3 /*************************************************
4 * Exim - an Internet mail transport agent *
5 *************************************************/
6
7 /* Copyright (c) University of Cambridge 1995 - 2006 */
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 if (!smtp_start_session())
474 {
475 mac_smtp_fflush();
476 search_tidyup();
477 _exit(EXIT_SUCCESS);
478 }
479
480 for (;;)
481 {
482 int rc;
483 message_id[0] = 0; /* Clear out any previous message_id */
484 reset_point = store_get(0); /* Save current store high water point */
485
486 DEBUG(D_any)
487 debug_printf("Process %d is ready for new message\n", (int)getpid());
488
489 /* Smtp_setup_msg() returns 0 on QUIT or if the call is from an
490 unacceptable host or if an ACL "drop" command was triggered, -1 on
491 connection lost, and +1 on validly reaching DATA. Receive_msg() almost
492 always returns TRUE when smtp_input is true; just retry if no message was
493 accepted (can happen for invalid message parameters). However, it can yield
494 FALSE if the connection was forcibly dropped by the DATA ACL. */
495
496 if ((rc = smtp_setup_msg()) > 0)
497 {
498 BOOL ok = receive_msg(FALSE);
499 search_tidyup(); /* Close cached databases */
500 if (!ok) /* Connection was dropped */
501 {
502 mac_smtp_fflush();
503 _exit(EXIT_SUCCESS);
504 }
505 if (message_id[0] == 0) continue; /* No message was accepted */
506 }
507 else
508 {
509 mac_smtp_fflush();
510 search_tidyup();
511 _exit((rc == 0)? EXIT_SUCCESS : EXIT_FAILURE);
512 }
513
514 /* Show the recipients when debugging */
515
516 DEBUG(D_receive)
517 {
518 int i;
519 if (sender_address != NULL)
520 debug_printf("Sender: %s\n", sender_address);
521 if (recipients_list != NULL)
522 {
523 debug_printf("Recipients:\n");
524 for (i = 0; i < recipients_count; i++)
525 debug_printf(" %s\n", recipients_list[i].address);
526 }
527 }
528
529 /* A message has been accepted. Clean up any previous delivery processes
530 that have completed and are defunct, on systems where they don't go away
531 by themselves (see comments when setting SIG_IGN above). On such systems
532 (if any) these delivery processes hang around after termination until
533 the next message is received. */
534
535 #ifndef SIG_IGN_WORKS
536 while (waitpid(-1, NULL, WNOHANG) > 0);
537 #endif
538
539 /* Reclaim up the store used in accepting this message */
540
541 store_reset(reset_point);
542
543 /* If queue_only is set or if there are too many incoming connections in
544 existence, local_queue_only will be TRUE. If it is not, check whether we
545 have received too many messages in this session for immediate delivery. If
546 not, and queue_only_load is set, check that the load average is below it.
547 Note that, once set, local_queue_only remains set for any subsequent
548 messages on the same SMTP connection. This is a deliberate choice; even
549 though the load average may fall, it doesn't seem right to deliver later
550 messages on the same call when not delivering earlier ones. */
551
552 if (!local_queue_only)
553 {
554 if (smtp_accept_queue_per_connection > 0 &&
555 receive_messagecount > smtp_accept_queue_per_connection)
556 {
557 local_queue_only = TRUE;
558 queue_only_reason = 2;
559 }
560 else if (queue_only_load >= 0)
561 {
562 local_queue_only = (load_average = os_getloadavg()) > queue_only_load;
563 if (local_queue_only) queue_only_reason = 3;
564 }
565 }
566
567 /* Log the queueing here, when it will get a message id attached, but
568 not if queue_only is set (case 0). */
569
570 if (local_queue_only) switch(queue_only_reason)
571 {
572 case 1:
573 log_write(L_delay_delivery,
574 LOG_MAIN, "no immediate delivery: too many connections "
575 "(%d, max %d)", smtp_accept_count, smtp_accept_queue);
576 break;
577
578 case 2:
579 log_write(L_delay_delivery,
580 LOG_MAIN, "no immediate delivery: more than %d messages "
581 "received in one connection", smtp_accept_queue_per_connection);
582 break;
583
584 case 3:
585 log_write(L_delay_delivery,
586 LOG_MAIN, "no immediate delivery: load average %.2f",
587 (double)load_average/1000.0);
588 break;
589 }
590
591 /* If a delivery attempt is required, spin off a new process to handle it.
592 If we are not root, we have to re-exec exim unless deliveries are being
593 done unprivileged. */
594
595 else if (!queue_only_policy && !deliver_freeze)
596 {
597 pid_t dpid;
598
599 /* Before forking, ensure that the C output buffer is flushed. Otherwise
600 anything that it in it will get duplicated, leading to duplicate copies
601 of the pending output. */
602
603 mac_smtp_fflush();
604
605 if ((dpid = fork()) == 0)
606 {
607 (void)fclose(smtp_in);
608 (void)fclose(smtp_out);
609
610 /* Don't ever molest the parent's SSL connection, but do clean up
611 the data structures if necessary. */
612
613 #ifdef SUPPORT_TLS
614 tls_close(FALSE);
615 #endif
616
617 /* Reset SIGHUP and SIGCHLD in the child in both cases. */
618
619 signal(SIGHUP, SIG_DFL);
620 signal(SIGCHLD, SIG_DFL);
621
622 if (geteuid() != root_uid && !deliver_drop_privilege)
623 {
624 signal(SIGALRM, SIG_DFL);
625 (void)child_exec_exim(CEE_EXEC_PANIC, FALSE, NULL, FALSE, 2, US"-Mc",
626 message_id);
627 /* Control does not return here. */
628 }
629
630 /* No need to re-exec; SIGALRM remains set to the default handler */
631
632 (void)deliver_message(message_id, FALSE, FALSE);
633 search_tidyup();
634 _exit(EXIT_SUCCESS);
635 }
636
637 if (dpid > 0)
638 {
639 DEBUG(D_any) debug_printf("forked delivery process %d\n", (int)dpid);
640 }
641 else
642 {
643 log_write(0, LOG_MAIN|LOG_PANIC, "daemon: delivery process fork "
644 "failed: %s", strerror(errno));
645 }
646 }
647 }
648 }
649
650
651 /* Carrying on in the parent daemon process... Can't do much if the fork
652 failed. Otherwise, keep count of the number of accepting processes and
653 remember the pid for ticking off when the child completes. */
654
655 if (pid < 0)
656 {
657 never_error(US"daemon: accept process fork failed", US"Fork failed", errno);
658 }
659 else
660 {
661 int i;
662 for (i = 0; i < smtp_accept_max; ++i)
663 {
664 if (smtp_slots[i].pid <= 0)
665 {
666 smtp_slots[i].pid = pid;
667 if (smtp_accept_max_per_host != NULL)
668 smtp_slots[i].host_address = string_copy_malloc(sender_host_address);
669 smtp_accept_count++;
670 break;
671 }
672 }
673 DEBUG(D_any) debug_printf("%d SMTP accept process%s running\n",
674 smtp_accept_count, (smtp_accept_count == 1)? "" : "es");
675 }
676
677 /* Get here via goto in error cases */
678
679 ERROR_RETURN:
680
681 /* Close the streams associated with the socket which will also close the
682 socket fds in this process. We can't do anything if fclose() fails, but
683 logging brings it to someone's attention. However, "connection reset by peer"
684 isn't really a problem, so skip that one. On Solaris, a dropped connection can
685 manifest itself as a broken pipe, so drop that one too. If the streams don't
686 exist, something went wrong while setting things up. Make sure the socket
687 descriptors are closed, in order to drop the connection. */
688
689 if (smtp_out != NULL)
690 {
691 if (fclose(smtp_out) != 0 && errno != ECONNRESET && errno != EPIPE)
692 log_write(0, LOG_MAIN|LOG_PANIC, "daemon: fclose(smtp_out) failed: %s",
693 strerror(errno));
694 smtp_out = NULL;
695 }
696 else (void)close(accept_socket);
697
698 if (smtp_in != NULL)
699 {
700 if (fclose(smtp_in) != 0 && errno != ECONNRESET && errno != EPIPE)
701 log_write(0, LOG_MAIN|LOG_PANIC, "daemon: fclose(smtp_in) failed: %s",
702 strerror(errno));
703 smtp_in = NULL;
704 }
705 else (void)close(dup_accept_socket);
706
707 /* Release any store used in this process, including the store used for holding
708 the incoming host address and an expanded active_hostname. */
709
710 store_reset(reset_point);
711 sender_host_address = NULL;
712 }
713
714
715
716
717 /*************************************************
718 * Check wildcard listen special cases *
719 *************************************************/
720
721 /* This function is used when binding and listening on lists of addresses and
722 ports. It tests for special cases of wildcard listening, when IPv4 and IPv6
723 sockets may interact in different ways in different operating systems. It is
724 passed an error number, the list of listening addresses, and the current
725 address. Two checks are available: for a previous wildcard IPv6 address, or for
726 a following wildcard IPv4 address, in both cases on the same port.
727
728 In practice, pairs of wildcard addresses should be adjacent in the address list
729 because they are sorted that way below.
730
731 Arguments:
732 eno the error number
733 addresses the list of addresses
734 ipa the current IP address
735 back if TRUE, check for previous wildcard IPv6 address
736 if FALSE, check for a following wildcard IPv4 address
737
738 Returns: TRUE or FALSE
739 */
740
741 static BOOL
742 check_special_case(int eno, ip_address_item *addresses, ip_address_item *ipa,
743 BOOL back)
744 {
745 ip_address_item *ipa2;
746
747 /* For the "back" case, if the failure was "address in use" for a wildcard IPv4
748 address, seek a previous IPv6 wildcard address on the same port. As it is
749 previous, it must have been successfully bound and be listening. Flag it as a
750 "6 including 4" listener. */
751
752 if (back)
753 {
754 if (eno != EADDRINUSE || ipa->address[0] != 0) return FALSE;
755 for (ipa2 = addresses; ipa2 != ipa; ipa2 = ipa2->next)
756 {
757 if (ipa2->address[1] == 0 && ipa2->port == ipa->port)
758 {
759 ipa2->v6_include_v4 = TRUE;
760 return TRUE;
761 }
762 }
763 }
764
765 /* For the "forward" case, if the current address is a wildcard IPv6 address,
766 we seek a following wildcard IPv4 address on the same port. */
767
768 else
769 {
770 if (ipa->address[0] != ':' || ipa->address[1] != 0) return FALSE;
771 for (ipa2 = ipa->next; ipa2 != NULL; ipa2 = ipa2->next)
772 if (ipa2->address[0] == 0 && ipa->port == ipa2->port) return TRUE;
773 }
774
775 return FALSE;
776 }
777
778
779
780
781 /*************************************************
782 * Handle terminating subprocesses *
783 *************************************************/
784
785 /* Handle the termination of child processes. Theoretically, this need be done
786 only when sigchld_seen is TRUE, but rumour has it that some systems lose
787 SIGCHLD signals at busy times, so to be on the safe side, this function is
788 called each time round. It shouldn't be too expensive.
789
790 Arguments: none
791 Returns: nothing
792 */
793
794 static void
795 handle_ending_processes(void)
796 {
797 int status;
798 pid_t pid;
799
800 while ((pid = waitpid(-1, &status, WNOHANG)) > 0)
801 {
802 int i;
803 DEBUG(D_any) debug_printf("child %d ended: status=0x%x\n", (int)pid,
804 status);
805
806 /* If it's a listening daemon for which we are keeping track of individual
807 subprocesses, deal with an accepting process that has terminated. */
808
809 if (smtp_slots != NULL)
810 {
811 for (i = 0; i < smtp_accept_max; i++)
812 {
813 if (smtp_slots[i].pid == pid)
814 {
815 if (smtp_slots[i].host_address != NULL)
816 store_free(smtp_slots[i].host_address);
817 smtp_slots[i] = empty_smtp_slot;
818 if (--smtp_accept_count < 0) smtp_accept_count = 0;
819 DEBUG(D_any) debug_printf("%d SMTP accept process%s now running\n",
820 smtp_accept_count, (smtp_accept_count == 1)? "" : "es");
821 break;
822 }
823 }
824 if (i < smtp_accept_max) continue; /* Found an accepting process */
825 }
826
827 /* If it wasn't an accepting process, see if it was a queue-runner
828 process that we are tracking. */
829
830 if (queue_pid_slots != NULL)
831 {
832 for (i = 0; i < queue_run_max; i++)
833 {
834 if (queue_pid_slots[i] == pid)
835 {
836 queue_pid_slots[i] = 0;
837 if (--queue_run_count < 0) queue_run_count = 0;
838 DEBUG(D_any) debug_printf("%d queue-runner process%s now running\n",
839 queue_run_count, (queue_run_count == 1)? "" : "es");
840 break;
841 }
842 }
843 }
844 }
845 }
846
847
848
849 /*************************************************
850 * Exim Daemon Mainline *
851 *************************************************/
852
853 /* The daemon can do two jobs, either of which is optional:
854
855 (1) Listens for incoming SMTP calls and spawns off a sub-process to handle
856 each one. This is requested by the -bd option, with -oX specifying the SMTP
857 port on which to listen (for testing).
858
859 (2) Spawns a queue-running process every so often. This is controlled by the
860 -q option with a an interval time. (If no time is given, a single queue run
861 is done from the main function, and control doesn't get here.)
862
863 Root privilege is required in order to attach to port 25. Some systems require
864 it when calling socket() rather than bind(). To cope with all cases, we run as
865 root for both socket() and bind(). Some systems also require root in order to
866 write to the pid file directory. This function must therefore be called as root
867 if it is to work properly in all circumstances. Once the socket is bound and
868 the pid file written, root privilege is given up if there is an exim uid.
869
870 There are no arguments to this function, and it never returns. */
871
872 void
873 daemon_go(void)
874 {
875 int *listen_sockets = NULL;
876 int listen_socket_count = 0;
877 ip_address_item *addresses = NULL;
878
879 /* If any debugging options are set, turn on the D_pid bit so that all
880 debugging lines get the pid added. */
881
882 DEBUG(D_any|D_v) debug_selector |= D_pid;
883
884
885 /* Do the preparation for setting up a listener on one or more interfaces, and
886 possible on various ports. This is controlled by the combination of
887 local_interfaces (which can set IP addresses and ports) and daemon_smtp_port
888 (which is a list of default ports to use for those items in local_interfaces
889 that do not specify a port). The -oX command line option can be used to
890 override one or both of these options.
891
892 If local_interfaces is not set, the default is to listen on all interfaces.
893 When it is set, it can include "all IPvx interfaces" as an item. This is useful
894 when different ports are in use.
895
896 It turns out that listening on all interfaces is messy in an IPv6 world,
897 because several different implementation approaches have been taken. This code
898 is now supposed to work with all of them. The point of difference is whether an
899 IPv6 socket that is listening on all interfaces will receive incoming IPv4
900 calls or not. We also have to cope with the case when IPv6 libraries exist, but
901 there is no IPv6 support in the kernel.
902
903 . On Solaris, an IPv6 socket will accept IPv4 calls, and give them as mapped
904 addresses. However, if an IPv4 socket is also listening on all interfaces,
905 calls are directed to the appropriate socket.
906
907 . On (some versions of) Linux, an IPv6 socket will accept IPv4 calls, and
908 give them as mapped addresses, but an attempt also to listen on an IPv4
909 socket on all interfaces causes an error.
910
911 . On OpenBSD, an IPv6 socket will not accept IPv4 calls. You have to set up
912 two sockets if you want to accept both kinds of call.
913
914 . FreeBSD is like OpenBSD, but it has the IPV6_V6ONLY socket option, which
915 can be turned off, to make it behave like the versions of Linux described
916 above.
917
918 . I heard a report that the USAGI IPv6 stack for Linux has implemented
919 IPV6_V6ONLY.
920
921 So, what we do when IPv6 is supported is as follows:
922
923 (1) After it is set up, the list of interfaces is scanned for wildcard
924 addresses. If an IPv6 and an IPv4 wildcard are both found for the same
925 port, the list is re-arranged so that they are together, with the IPv6
926 wildcard first.
927
928 (2) If the creation of a wildcard IPv6 socket fails, we just log the error and
929 carry on if an IPv4 wildcard socket for the same port follows later in the
930 list. This allows Exim to carry on in the case when the kernel has no IPv6
931 support.
932
933 (3) Having created an IPv6 wildcard socket, we try to set IPV6_V6ONLY if that
934 option is defined. However, if setting fails, carry on regardless (but log
935 the incident).
936
937 (4) If binding or listening on an IPv6 wildcard socket fails, it is a serious
938 error.
939
940 (5) If binding or listening on an IPv4 wildcard socket fails with the error
941 EADDRINUSE, and a previous interface was an IPv6 wildcard for the same
942 port (which must have succeeded or we wouldn't have got this far), we
943 assume we are in the situation where just a single socket is permitted,
944 and ignore the error.
945
946 Phew!
947
948 The preparation code decodes options and sets up the relevant data. We do this
949 first, so that we can return non-zero if there are any syntax errors, and also
950 write to stderr. */
951
952 if (daemon_listen)
953 {
954 int *default_smtp_port;
955 int sep;
956 int pct = 0;
957 uschar *s;
958 uschar *list;
959 uschar *local_iface_source = US"local_interfaces";
960 ip_address_item *ipa;
961 ip_address_item **pipa;
962
963 /* If any option requiring a load average to be available during the
964 reception of a message is set, call os_getloadavg() while we are root
965 for those OS for which this is necessary the first time it is called (in
966 order to perform an "open" on the kernel memory file). */
967
968 #ifdef LOAD_AVG_NEEDS_ROOT
969 if (queue_only_load >= 0 || smtp_load_reserve >= 0 ||
970 (deliver_queue_load_max >= 0 && deliver_drop_privilege))
971 (void)os_getloadavg();
972 #endif
973
974 /* If -oX was used, disable the writing of a pid file unless -oP was
975 explicitly used to force it. Then scan the string given to -oX. Any items
976 that contain neither a dot nor a colon are used to override daemon_smtp_port.
977 Any other items are used to override local_interfaces. */
978
979 if (override_local_interfaces != NULL)
980 {
981 uschar *new_smtp_port = NULL;
982 uschar *new_local_interfaces = NULL;
983 int portsize = 0;
984 int portptr = 0;
985 int ifacesize = 0;
986 int ifaceptr = 0;
987
988 if (override_pid_file_path == NULL) write_pid = FALSE;
989
990 list = override_local_interfaces;
991 sep = 0;
992 while ((s = string_nextinlist(&list,&sep,big_buffer,big_buffer_size))
993 != NULL)
994 {
995 uschar joinstr[4];
996 uschar **ptr;
997 int *sizeptr;
998 int *ptrptr;
999
1000 if (Ustrpbrk(s, ".:") == NULL)
1001 {
1002 ptr = &new_smtp_port;
1003 sizeptr = &portsize;
1004 ptrptr = &portptr;
1005 }
1006 else
1007 {
1008 ptr = &new_local_interfaces;
1009 sizeptr = &ifacesize;
1010 ptrptr = &ifaceptr;
1011 }
1012
1013 if (*ptr == NULL)
1014 {
1015 joinstr[0] = sep;
1016 joinstr[1] = ' ';
1017 *ptr = string_cat(*ptr, sizeptr, ptrptr, US"<", 1);
1018 }
1019
1020 *ptr = string_cat(*ptr, sizeptr, ptrptr, joinstr, 2);
1021 *ptr = string_cat(*ptr, sizeptr, ptrptr, s, Ustrlen(s));
1022 }
1023
1024 if (new_smtp_port != NULL)
1025 {
1026 new_smtp_port[portptr] = 0;
1027 daemon_smtp_port = new_smtp_port;
1028 DEBUG(D_any) debug_printf("daemon_smtp_port overridden by -oX:\n %s\n",
1029 daemon_smtp_port);
1030 }
1031
1032 if (new_local_interfaces != NULL)
1033 {
1034 new_local_interfaces[ifaceptr] = 0;
1035 local_interfaces = new_local_interfaces;
1036 local_iface_source = US"-oX data";
1037 DEBUG(D_any) debug_printf("local_interfaces overridden by -oX:\n %s\n",
1038 local_interfaces);
1039 }
1040 }
1041
1042 /* Create a list of default SMTP ports, to be used if local_interfaces
1043 contains entries without explict ports. First count the number of ports, then
1044 build a translated list in a vector. */
1045
1046 list = daemon_smtp_port;
1047 sep = 0;
1048 while ((s = string_nextinlist(&list,&sep,big_buffer,big_buffer_size)) != NULL)
1049 pct++;
1050 default_smtp_port = store_get((pct+1) * sizeof(int));
1051 list = daemon_smtp_port;
1052 sep = 0;
1053 for (pct = 0;
1054 (s = string_nextinlist(&list,&sep,big_buffer,big_buffer_size)) != NULL;
1055 pct++)
1056 {
1057 if (isdigit(*s))
1058 {
1059 uschar *end;
1060 default_smtp_port[pct] = Ustrtol(s, &end, 0);
1061 if (end != s + Ustrlen(s))
1062 log_write(0, LOG_PANIC_DIE|LOG_CONFIG, "invalid SMTP port: %s", s);
1063 }
1064 else
1065 {
1066 struct servent *smtp_service = getservbyname(CS s, "tcp");
1067 if (smtp_service == NULL)
1068 log_write(0, LOG_PANIC_DIE|LOG_CONFIG, "TCP port \"%s\" not found", s);
1069 default_smtp_port[pct] = ntohs(smtp_service->s_port);
1070 }
1071 }
1072 default_smtp_port[pct] = 0;
1073
1074 /* Create the list of local interfaces, possibly with ports included. This
1075 list may contain references to 0.0.0.0 and ::0 as wildcards. These special
1076 values are converted below. */
1077
1078 addresses = host_build_ifacelist(local_interfaces, local_iface_source);
1079
1080 /* In the list of IP addresses, convert 0.0.0.0 into an empty string, and ::0
1081 into the string ":". We use these to recognize wildcards in IPv4 and IPv6. In
1082 fact, many IP stacks recognize 0.0.0.0 and ::0 and handle them as wildcards
1083 anyway, but we need to know which are the wildcard addresses, and the shorter
1084 strings are neater.
1085
1086 In the same scan, fill in missing port numbers from the default list. When
1087 there is more than one item in the list, extra items are created. */
1088
1089 for (ipa = addresses; ipa != NULL; ipa = ipa->next)
1090 {
1091 int i;
1092
1093 if (Ustrcmp(ipa->address, "0.0.0.0") == 0) ipa->address[0] = 0;
1094 else if (Ustrcmp(ipa->address, "::0") == 0)
1095 {
1096 ipa->address[0] = ':';
1097 ipa->address[1] = 0;
1098 }
1099
1100 if (ipa->port > 0) continue;
1101
1102 if (daemon_smtp_port[0] <= 0)
1103 log_write(0, LOG_MAIN|LOG_PANIC_DIE, "no port specified for interface "
1104 "%s and daemon_smtp_port is unset; cannot start daemon",
1105 (ipa->address[0] == 0)? US"\"all IPv4\"" :
1106 (ipa->address[1] == 0)? US"\"all IPv6\"" : ipa->address);
1107 ipa->port = default_smtp_port[0];
1108 for (i = 1; default_smtp_port[i] > 0; i++)
1109 {
1110 ip_address_item *new = store_get(sizeof(ip_address_item));
1111 memcpy(new->address, ipa->address, Ustrlen(ipa->address) + 1);
1112 new->port = default_smtp_port[i];
1113 new->next = ipa->next;
1114 ipa->next = new;
1115 ipa = new;
1116 }
1117 }
1118
1119 /* Scan the list of addresses for wildcards. If we find an IPv4 and an IPv6
1120 wildcard for the same port, ensure that (a) they are together and (b) the
1121 IPv6 address comes first. This makes handling the messy features easier, and
1122 also simplifies the construction of the "daemon started" log line. */
1123
1124 pipa = &addresses;
1125 for (ipa = addresses; ipa != NULL; pipa = &(ipa->next), ipa = ipa->next)
1126 {
1127 ip_address_item *ipa2;
1128
1129 /* Handle an IPv4 wildcard */
1130
1131 if (ipa->address[0] == 0)
1132 {
1133 for (ipa2 = ipa; ipa2->next != NULL; ipa2 = ipa2->next)
1134 {
1135 ip_address_item *ipa3 = ipa2->next;
1136 if (ipa3->address[0] == ':' &&
1137 ipa3->address[1] == 0 &&
1138 ipa3->port == ipa->port)
1139 {
1140 ipa2->next = ipa3->next;
1141 ipa3->next = ipa;
1142 *pipa = ipa3;
1143 break;
1144 }
1145 }
1146 }
1147
1148 /* Handle an IPv6 wildcard. */
1149
1150 else if (ipa->address[0] == ':' && ipa->address[1] == 0)
1151 {
1152 for (ipa2 = ipa; ipa2->next != NULL; ipa2 = ipa2->next)
1153 {
1154 ip_address_item *ipa3 = ipa2->next;
1155 if (ipa3->address[0] == 0 && ipa3->port == ipa->port)
1156 {
1157 ipa2->next = ipa3->next;
1158 ipa3->next = ipa->next;
1159 ipa->next = ipa3;
1160 ipa = ipa3;
1161 break;
1162 }
1163 }
1164 }
1165 }
1166
1167 /* Get a vector to remember all the sockets in */
1168
1169 for (ipa = addresses; ipa != NULL; ipa = ipa->next)
1170 listen_socket_count++;
1171 listen_sockets = store_get(sizeof(int *) * listen_socket_count);
1172
1173 /* Do a sanity check on the max connects value just to save us from getting
1174 a huge amount of store. */
1175
1176 if (smtp_accept_max > 4095) smtp_accept_max = 4096;
1177
1178 /* There's no point setting smtp_accept_queue unless it is less than the max
1179 connects limit. The configuration reader ensures that the max is set if the
1180 queue-only option is set. */
1181
1182 if (smtp_accept_queue > smtp_accept_max) smtp_accept_queue = 0;
1183
1184 /* Get somewhere to keep the list of SMTP accepting pids if we are keeping
1185 track of them for total number and queue/host limits. */
1186
1187 if (smtp_accept_max > 0)
1188 {
1189 int i;
1190 smtp_slots = store_get(smtp_accept_max * sizeof(smtp_slot));
1191 for (i = 0; i < smtp_accept_max; i++) smtp_slots[i] = empty_smtp_slot;
1192 }
1193 }
1194
1195 /* The variable background_daemon is always false when debugging, but
1196 can also be forced false in order to keep a non-debugging daemon in the
1197 foreground. If background_daemon is true, close all open file descriptors that
1198 we know about, but then re-open stdin, stdout, and stderr to /dev/null.
1199
1200 This is protection against any called functions (in libraries, or in
1201 Perl, or whatever) that think they can write to stderr (or stdout). Before this
1202 was added, it was quite likely that an SMTP connection would use one of these
1203 file descriptors, in which case writing random stuff to it caused chaos.
1204
1205 Then disconnect from the controlling terminal, Most modern Unixes seem to have
1206 setsid() for getting rid of the controlling terminal. For any OS that doesn't,
1207 setsid() can be #defined as a no-op, or as something else. */
1208
1209 if (background_daemon)
1210 {
1211 log_close_all(); /* Just in case anything was logged earlier */
1212 search_tidyup(); /* Just in case any were used in reading the config. */
1213 (void)close(0); /* Get rid of stdin/stdout/stderr */
1214 (void)close(1);
1215 (void)close(2);
1216 exim_nullstd(); /* Connect stdin/stdout/stderr to /dev/null */
1217 log_stderr = NULL; /* So no attempt to copy paniclog output */
1218
1219 /* If the parent process of this one has pid == 1, we are re-initializing the
1220 daemon as the result of a SIGHUP. In this case, there is no need to do
1221 anything, because the controlling terminal has long gone. Otherwise, fork, in
1222 case current process is a process group leader (see 'man setsid' for an
1223 explanation) before calling setsid(). */
1224
1225 if (getppid() != 1)
1226 {
1227 pid_t pid = fork();
1228 if (pid < 0) log_write(0, LOG_MAIN|LOG_PANIC_DIE,
1229 "fork() failed when starting daemon: %s", strerror(errno));
1230 if (pid > 0) exit(EXIT_SUCCESS); /* in parent process, just exit */
1231 (void)setsid(); /* release controlling terminal */
1232 }
1233 }
1234
1235 /* We are now in the disconnected, daemon process (unless debugging). Set up
1236 the listening sockets if required. */
1237
1238 if (daemon_listen)
1239 {
1240 int sk;
1241 int on = 1;
1242 ip_address_item *ipa;
1243
1244 /* For each IP address, create a socket, bind it to the appropriate port, and
1245 start listening. See comments above about IPv6 sockets that may or may not
1246 accept IPv4 calls when listening on all interfaces. We also have to cope with
1247 the case of a system with IPv6 libraries, but no IPv6 support in the kernel.
1248 listening, provided a wildcard IPv4 socket for the same port follows. */
1249
1250 for (ipa = addresses, sk = 0; sk < listen_socket_count; ipa = ipa->next, sk++)
1251 {
1252 BOOL wildcard;
1253 ip_address_item *ipa2;
1254 int af;
1255
1256 if (Ustrchr(ipa->address, ':') != NULL)
1257 {
1258 af = AF_INET6;
1259 wildcard = ipa->address[1] == 0;
1260 }
1261 else
1262 {
1263 af = AF_INET;
1264 wildcard = ipa->address[0] == 0;
1265 }
1266
1267 listen_sockets[sk] = ip_socket(SOCK_STREAM, af);
1268 if (listen_sockets[sk] < 0)
1269 {
1270 if (check_special_case(0, addresses, ipa, FALSE))
1271 {
1272 log_write(0, LOG_MAIN, "Failed to create IPv6 socket for wildcard "
1273 "listening (%s): will use IPv4", strerror(errno));
1274 goto SKIP_SOCKET;
1275 }
1276 log_write(0, LOG_PANIC_DIE, "IPv%c socket creation failed: %s",
1277 (af == AF_INET6)? '6' : '4', strerror(errno));
1278 }
1279
1280 /* If this is an IPv6 wildcard socket, set IPV6_V6ONLY if that option is
1281 available. Just log failure (can get protocol not available, just like
1282 socket creation can). */
1283
1284 #ifdef IPV6_V6ONLY
1285 if (af == AF_INET6 && wildcard &&
1286 setsockopt(listen_sockets[sk], IPPROTO_IPV6, IPV6_V6ONLY, (char *)(&on),
1287 sizeof(on)) < 0)
1288 log_write(0, LOG_MAIN, "Setting IPV6_V6ONLY on daemon's IPv6 wildcard "
1289 "socket failed (%s): carrying on without it", strerror(errno));
1290 #endif /* IPV6_V6ONLY */
1291
1292 /* Set SO_REUSEADDR so that the daemon can be restarted while a connection
1293 is being handled. Without this, a connection will prevent reuse of the
1294 smtp port for listening. */
1295
1296 if (setsockopt(listen_sockets[sk], SOL_SOCKET, SO_REUSEADDR,
1297 (uschar *)(&on), sizeof(on)) < 0)
1298 log_write(0, LOG_MAIN|LOG_PANIC_DIE, "setting SO_REUSEADDR on socket "
1299 "failed when starting daemon: %s", strerror(errno));
1300
1301 /* Set TCP_NODELAY; Exim does its own buffering. There is a switch to
1302 disable this because it breaks some broken clients. */
1303
1304 if (tcp_nodelay) setsockopt(listen_sockets[sk], IPPROTO_TCP, TCP_NODELAY,
1305 (uschar *)(&on), sizeof(on));
1306
1307 /* Now bind the socket to the required port; if Exim is being restarted
1308 it may not always be possible to bind immediately, even with SO_REUSEADDR
1309 set, so try 10 times, waiting between each try. After 10 failures, we give
1310 up. In an IPv6 environment, if bind () fails with the error EADDRINUSE and
1311 we are doing wildcard IPv4 listening and there was a previous IPv6 wildcard
1312 address for the same port, ignore the error on the grounds that we must be
1313 in a system where the IPv6 socket accepts both kinds of call. This is
1314 necessary for (some release of) USAGI Linux; other IP stacks fail at the
1315 listen() stage instead. */
1316
1317 for(;;)
1318 {
1319 uschar *msg, *addr;
1320 if (ip_bind(listen_sockets[sk], af, ipa->address, ipa->port) >= 0) break;
1321 if (check_special_case(errno, addresses, ipa, TRUE))
1322 {
1323 DEBUG(D_any) debug_printf("wildcard IPv4 bind() failed after IPv6 "
1324 "listen() success; EADDRINUSE ignored\n");
1325 (void)close(listen_sockets[sk]);
1326 goto SKIP_SOCKET;
1327 }
1328 msg = US strerror(errno);
1329 addr = wildcard? ((af == AF_INET6)? US"(any IPv6)" : US"(any IPv4)") :
1330 ipa->address;
1331 if (daemon_startup_retries <= 0)
1332 log_write(0, LOG_MAIN|LOG_PANIC_DIE,
1333 "socket bind() to port %d for address %s failed: %s: "
1334 "daemon abandoned", ipa->port, addr, msg);
1335 log_write(0, LOG_MAIN, "socket bind() to port %d for address %s "
1336 "failed: %s: waiting %s before trying again (%d more %s)",
1337 ipa->port, addr, msg, readconf_printtime(daemon_startup_sleep),
1338 daemon_startup_retries, (daemon_startup_retries > 1)? "tries" : "try");
1339 daemon_startup_retries--;
1340 sleep(daemon_startup_sleep);
1341 }
1342
1343 DEBUG(D_any)
1344 {
1345 if (wildcard)
1346 debug_printf("listening on all interfaces (IPv%c) port %d\n",
1347 (af == AF_INET6)? '6' : '4', ipa->port);
1348 else
1349 debug_printf("listening on %s port %d\n", ipa->address, ipa->port);
1350 }
1351
1352 /* Start listening on the bound socket, establishing the maximum backlog of
1353 connections that is allowed. On success, continue to the next address. */
1354
1355 if (listen(listen_sockets[sk], smtp_connect_backlog) >= 0) continue;
1356
1357 /* Listening has failed. In an IPv6 environment, as for bind(), if listen()
1358 fails with the error EADDRINUSE and we are doing IPv4 wildcard listening
1359 and there was a previous successful IPv6 wildcard listen on the same port,
1360 we want to ignore the error on the grounds that we must be in a system
1361 where the IPv6 socket accepts both kinds of call. */
1362
1363 if (!check_special_case(errno, addresses, ipa, TRUE))
1364 log_write(0, LOG_PANIC_DIE, "listen() failed on interface %s: %s",
1365 wildcard? ((af == AF_INET6)? US"(any IPv6)" : US"(any IPv4)") :
1366 ipa->address,
1367 strerror(errno));
1368
1369 DEBUG(D_any) debug_printf("wildcard IPv4 listen() failed after IPv6 "
1370 "listen() success; EADDRINUSE ignored\n");
1371 (void)close(listen_sockets[sk]);
1372
1373 /* Come here if there has been a problem with the socket which we
1374 are going to ignore. We remove the address from the chain, and back up the
1375 counts. */
1376
1377 SKIP_SOCKET:
1378 sk--; /* Back up the count */
1379 listen_socket_count--; /* Reduce the total */
1380 if (ipa == addresses) addresses = ipa->next; else
1381 {
1382 for (ipa2 = addresses; ipa2->next != ipa; ipa2 = ipa2->next);
1383 ipa2->next = ipa->next;
1384 ipa = ipa2;
1385 }
1386 } /* End of bind/listen loop for each address */
1387 } /* End of setup for listening */
1388
1389
1390 /* If we are not listening, we want to write a pid file only if -oP was
1391 explicitly given. */
1392
1393 else if (override_pid_file_path == NULL) write_pid = FALSE;
1394
1395 /* Write the pid to a known file for assistance in identification, if required.
1396 We do this before giving up root privilege, because on some systems it is
1397 necessary to be root in order to write into the pid file directory. There's
1398 nothing to stop multiple daemons running, as long as no more than one listens
1399 on a given TCP/IP port on the same interface(s). However, in these
1400 circumstances it gets far too complicated to mess with pid file names
1401 automatically. Consequently, Exim 4 writes a pid file only
1402
1403 (a) When running in the test harness, or
1404 (b) When -bd is used and -oX is not used, or
1405 (c) When -oP is used to supply a path.
1406
1407 The variable daemon_write_pid is used to control this. */
1408
1409 if (running_in_test_harness || write_pid)
1410 {
1411 FILE *f;
1412
1413 if (override_pid_file_path != NULL)
1414 pid_file_path = override_pid_file_path;
1415
1416 if (pid_file_path[0] == 0)
1417 pid_file_path = string_sprintf("%s/exim-daemon.pid", spool_directory);
1418
1419 f = modefopen(pid_file_path, "wb", 0644);
1420 if (f != NULL)
1421 {
1422 (void)fprintf(f, "%d\n", (int)getpid());
1423 (void)fclose(f);
1424 DEBUG(D_any) debug_printf("pid written to %s\n", pid_file_path);
1425 }
1426 else
1427 {
1428 DEBUG(D_any)
1429 debug_printf("%s\n", string_open_failed(errno, "pid file %s",
1430 pid_file_path));
1431 }
1432 }
1433
1434 /* Set up the handler for SIGHUP, which causes a restart of the daemon. */
1435
1436 sighup_seen = FALSE;
1437 signal(SIGHUP, sighup_handler);
1438
1439 /* Give up root privilege at this point (assuming that exim_uid and exim_gid
1440 are not root). The third argument controls the running of initgroups().
1441 Normally we do this, in order to set up the groups for the Exim user. However,
1442 if we are not root at this time - some odd installations run that way - we
1443 cannot do this. */
1444
1445 exim_setugid(exim_uid, exim_gid, geteuid()==root_uid, US"running as a daemon");
1446
1447 /* Get somewhere to keep the list of queue-runner pids if we are keeping track
1448 of them (and also if we are doing queue runs). */
1449
1450 if (queue_interval > 0 && queue_run_max > 0)
1451 {
1452 int i;
1453 queue_pid_slots = store_get(queue_run_max * sizeof(pid_t));
1454 for (i = 0; i < queue_run_max; i++) queue_pid_slots[i] = 0;
1455 }
1456
1457 /* Set up the handler for termination of child processes. */
1458
1459 sigchld_seen = FALSE;
1460 os_non_restarting_signal(SIGCHLD, main_sigchld_handler);
1461
1462 /* If we are to run the queue periodically, pretend the alarm has just gone
1463 off. This will cause the first queue-runner to get kicked off straight away. */
1464
1465 sigalrm_seen = (queue_interval > 0);
1466
1467 /* Log the start up of a daemon - at least one of listening or queue running
1468 must be set up. */
1469
1470 if (daemon_listen)
1471 {
1472 int i, j;
1473 int smtp_ports = 0;
1474 int smtps_ports = 0;
1475 ip_address_item *ipa;
1476 uschar *p = big_buffer;
1477 uschar *qinfo = (queue_interval > 0)?
1478 string_sprintf("-q%s", readconf_printtime(queue_interval))
1479 :
1480 US"no queue runs";
1481
1482 /* Build a list of listening addresses in big_buffer, but limit it to 10
1483 items. The style is for backwards compatibility.
1484
1485 It is now possible to have some ports listening for SMTPS (the old,
1486 deprecated protocol that starts TLS without using STARTTLS), and others
1487 listening for standard SMTP. Keep their listings separate. */
1488
1489 for (j = 0; j < 2; j++)
1490 {
1491 for (i = 0, ipa = addresses; i < 10 && ipa != NULL; i++, ipa = ipa->next)
1492 {
1493 /* First time round, look for SMTP ports; second time round, look for
1494 SMTPS ports. For the first one of each, insert leading text. */
1495
1496 if (host_is_tls_on_connect_port(ipa->port) == (j > 0))
1497 {
1498 if (j == 0)
1499 {
1500 if (smtp_ports++ == 0)
1501 {
1502 memcpy(p, "SMTP on", 8);
1503 p += 7;
1504 }
1505 }
1506 else
1507 {
1508 if (smtps_ports++ == 0)
1509 {
1510 (void)sprintf(CS p, "%sSMTPS on",
1511 (smtp_ports == 0)? "":" and for ");
1512 while (*p != 0) p++;
1513 }
1514 }
1515
1516 /* Now the information about the port (and sometimes interface) */
1517
1518 if (ipa->address[0] == ':' && ipa->address[1] == 0)
1519 {
1520 if (ipa->next != NULL && ipa->next->address[0] == 0 &&
1521 ipa->next->port == ipa->port)
1522 {
1523 (void)sprintf(CS p, " port %d (IPv6 and IPv4)", ipa->port);
1524 ipa = ipa->next;
1525 }
1526 else if (ipa->v6_include_v4)
1527 (void)sprintf(CS p, " port %d (IPv6 with IPv4)", ipa->port);
1528 else
1529 (void)sprintf(CS p, " port %d (IPv6)", ipa->port);
1530 }
1531 else if (ipa->address[0] == 0)
1532 (void)sprintf(CS p, " port %d (IPv4)", ipa->port);
1533 else
1534 (void)sprintf(CS p, " [%s]:%d", ipa->address, ipa->port);
1535 while (*p != 0) p++;
1536 }
1537 }
1538
1539 if (ipa != NULL)
1540 {
1541 memcpy(p, " ...", 5);
1542 p += 4;
1543 }
1544 }
1545
1546 log_write(0, LOG_MAIN,
1547 "exim %s daemon started: pid=%d, %s, listening for %s",
1548 version_string, getpid(), qinfo, big_buffer);
1549 set_process_info("daemon: %s, listening for %s", qinfo, big_buffer);
1550 }
1551
1552 else
1553 {
1554 log_write(0, LOG_MAIN,
1555 "exim %s daemon started: pid=%d, -q%s, not listening for SMTP",
1556 version_string, getpid(), readconf_printtime(queue_interval));
1557 set_process_info("daemon: -q%s, not listening",
1558 readconf_printtime(queue_interval));
1559 }
1560
1561
1562 /* Close the log so it can be renamed and moved. In the few cases below where
1563 this long-running process writes to the log (always exceptional conditions), it
1564 closes the log afterwards, for the same reason. */
1565
1566 log_close_all();
1567
1568 DEBUG(D_any) debug_print_ids(US"daemon running with");
1569
1570 /* Any messages accepted via this route are going to be SMTP. */
1571
1572 smtp_input = TRUE;
1573
1574 /* Enter the never-ending loop... */
1575
1576 for (;;)
1577 {
1578 #if HAVE_IPV6
1579 struct sockaddr_in6 accepted;
1580 #else
1581 struct sockaddr_in accepted;
1582 #endif
1583
1584 EXIM_SOCKLEN_T len = sizeof(accepted);
1585 pid_t pid;
1586
1587 /* This code is placed first in the loop, so that it gets obeyed at the
1588 start, before the first wait. This causes the first queue-runner to be
1589 started immediately. */
1590
1591 if (sigalrm_seen)
1592 {
1593 DEBUG(D_any) debug_printf("SIGALRM received\n");
1594
1595 /* Do a full queue run in a child process, if required, unless we already
1596 have enough queue runners on the go. If we are not running as root, a
1597 re-exec is required. */
1598
1599 if (queue_interval > 0 &&
1600 (queue_run_max <= 0 || queue_run_count < queue_run_max))
1601 {
1602 if ((pid = fork()) == 0)
1603 {
1604 int sk;
1605
1606 DEBUG(D_any) debug_printf("Starting queue-runner: pid %d\n",
1607 (int)getpid());
1608
1609 /* Disable debugging if it's required only for the daemon process. We
1610 leave the above message, because it ties up with the "child ended"
1611 debugging messages. */
1612
1613 if (debug_daemon) debug_selector = 0;
1614
1615 /* Close any open listening sockets in the child */
1616
1617 for (sk = 0; sk < listen_socket_count; sk++)
1618 (void)close(listen_sockets[sk]);
1619
1620 /* Reset SIGHUP and SIGCHLD in the child in both cases. */
1621
1622 signal(SIGHUP, SIG_DFL);
1623 signal(SIGCHLD, SIG_DFL);
1624
1625 /* Re-exec if privilege has been given up, unless deliver_drop_
1626 privilege is set. Reset SIGALRM before exec(). */
1627
1628 if (geteuid() != root_uid && !deliver_drop_privilege)
1629 {
1630 uschar opt[8];
1631 uschar *p = opt;
1632
1633 signal(SIGALRM, SIG_DFL);
1634 *p++ = '-';
1635 *p++ = 'q';
1636 if (queue_2stage) *p++ = 'q';
1637 if (queue_run_first_delivery) *p++ = 'i';
1638 if (queue_run_force) *p++ = 'f';
1639 if (deliver_force_thaw) *p++ = 'f';
1640 if (queue_run_local) *p++ = 'l';
1641 *p = 0;
1642
1643 (void)child_exec_exim(CEE_EXEC_PANIC, FALSE, NULL, TRUE, 1, opt);
1644 /* Control never returns here. */
1645 }
1646
1647 /* No need to re-exec; SIGALRM remains set to the default handler */
1648
1649 queue_run(NULL, NULL, FALSE);
1650 _exit(EXIT_SUCCESS);
1651 }
1652
1653 if (pid < 0)
1654 {
1655 log_write(0, LOG_MAIN|LOG_PANIC, "daemon: fork of queue-runner "
1656 "process failed: %s", strerror(errno));
1657 log_close_all();
1658 }
1659 else
1660 {
1661 int i;
1662 for (i = 0; i < queue_run_max; ++i)
1663 {
1664 if (queue_pid_slots[i] <= 0)
1665 {
1666 queue_pid_slots[i] = pid;
1667 queue_run_count++;
1668 break;
1669 }
1670 }
1671 DEBUG(D_any) debug_printf("%d queue-runner process%s running\n",
1672 queue_run_count, (queue_run_count == 1)? "" : "es");
1673 }
1674 }
1675
1676 /* Reset the alarm clock */
1677
1678 sigalrm_seen = FALSE;
1679 alarm(queue_interval);
1680 }
1681
1682
1683 /* Sleep till a connection happens if listening, and handle the connection if
1684 that is why we woke up. The FreeBSD operating system requires the use of
1685 select() before accept() because the latter function is not interrupted by
1686 a signal, and we want to wake up for SIGCHLD and SIGALRM signals. Some other
1687 OS do notice signals in accept() but it does no harm to have the select()
1688 in for all of them - and it won't then be a lurking problem for ports to
1689 new OS. In fact, the later addition of listening on specific interfaces only
1690 requires this way of working anyway. */
1691
1692 if (daemon_listen)
1693 {
1694 int sk, lcount, select_errno;
1695 int max_socket = 0;
1696 BOOL select_failed = FALSE;
1697 fd_set select_listen;
1698
1699 FD_ZERO(&select_listen);
1700 for (sk = 0; sk < listen_socket_count; sk++)
1701 {
1702 FD_SET(listen_sockets[sk], &select_listen);
1703 if (listen_sockets[sk] > max_socket) max_socket = listen_sockets[sk];
1704 }
1705
1706 DEBUG(D_any) debug_printf("Listening...\n");
1707
1708 /* In rare cases we may have had a SIGCHLD signal in the time between
1709 setting the handler (below) and getting back here. If so, pretend that the
1710 select() was interrupted so that we reap the child. This might still leave
1711 a small window when a SIGCHLD could get lost. However, since we use SIGCHLD
1712 only to do the reaping more quickly, it shouldn't result in anything other
1713 than a delay until something else causes a wake-up. */
1714
1715 if (sigchld_seen)
1716 {
1717 lcount = -1;
1718 errno = EINTR;
1719 }
1720 else
1721 {
1722 lcount = select(max_socket + 1, (SELECT_ARG2_TYPE *)&select_listen,
1723 NULL, NULL, NULL);
1724 }
1725
1726 if (lcount < 0)
1727 {
1728 select_failed = TRUE;
1729 lcount = 1;
1730 }
1731
1732 /* Clean up any subprocesses that may have terminated. We need to do this
1733 here so that smtp_accept_max_per_host works when a connection to that host
1734 has completed, and we are about to accept a new one. When this code was
1735 later in the sequence, a new connection could be rejected, even though an
1736 old one had just finished. Preserve the errno from any select() failure for
1737 the use of the common select/accept error processing below. */
1738
1739 select_errno = errno;
1740 handle_ending_processes();
1741 errno = select_errno;
1742
1743 /* Loop for all the sockets that are currently ready to go. If select
1744 actually failed, we have set the count to 1 and select_failed=TRUE, so as
1745 to use the common error code for select/accept below. */
1746
1747 while (lcount-- > 0)
1748 {
1749 int accept_socket = -1;
1750 if (!select_failed)
1751 {
1752 for (sk = 0; sk < listen_socket_count; sk++)
1753 {
1754 if (FD_ISSET(listen_sockets[sk], &select_listen))
1755 {
1756 accept_socket = accept(listen_sockets[sk],
1757 (struct sockaddr *)&accepted, &len);
1758 FD_CLR(listen_sockets[sk], &select_listen);
1759 break;
1760 }
1761 }
1762 }
1763
1764 /* If select or accept has failed and this was not caused by an
1765 interruption, log the incident and try again. With asymmetric TCP/IP
1766 routing errors such as "No route to network" have been seen here. Also
1767 "connection reset by peer" has been seen. These cannot be classed as
1768 disastrous errors, but they could fill up a lot of log. The code in smail
1769 crashes the daemon after 10 successive failures of accept, on the grounds
1770 that some OS fail continuously. Exim originally followed suit, but this
1771 appears to have caused problems. Now it just keeps going, but instead of
1772 logging each error, it batches them up when they are continuous. */
1773
1774 if (accept_socket < 0 && errno != EINTR)
1775 {
1776 if (accept_retry_count == 0)
1777 {
1778 accept_retry_errno = errno;
1779 accept_retry_select_failed = select_failed;
1780 }
1781 else
1782 {
1783 if (errno != accept_retry_errno ||
1784 select_failed != accept_retry_select_failed ||
1785 accept_retry_count >= 50)
1786 {
1787 log_write(0, LOG_MAIN | ((accept_retry_count >= 50)? LOG_PANIC : 0),
1788 "%d %s() failure%s: %s",
1789 accept_retry_count,
1790 accept_retry_select_failed? "select" : "accept",
1791 (accept_retry_count == 1)? "" : "s",
1792 strerror(accept_retry_errno));
1793 log_close_all();
1794 accept_retry_count = 0;
1795 accept_retry_errno = errno;
1796 accept_retry_select_failed = select_failed;
1797 }
1798 }
1799 accept_retry_count++;
1800 }
1801
1802 else
1803 {
1804 if (accept_retry_count > 0)
1805 {
1806 log_write(0, LOG_MAIN, "%d %s() failure%s: %s",
1807 accept_retry_count,
1808 accept_retry_select_failed? "select" : "accept",
1809 (accept_retry_count == 1)? "" : "s",
1810 strerror(accept_retry_errno));
1811 log_close_all();
1812 accept_retry_count = 0;
1813 }
1814 }
1815
1816 /* If select/accept succeeded, deal with the connection. */
1817
1818 if (accept_socket >= 0)
1819 handle_smtp_call(listen_sockets, listen_socket_count, accept_socket,
1820 (struct sockaddr *)&accepted);
1821 }
1822 }
1823
1824 /* If not listening, then just sleep for the queue interval. If we woke
1825 up early the last time for some other signal, it won't matter because
1826 the alarm signal will wake at the right time. This code originally used
1827 sleep() but it turns out that on the FreeBSD system, sleep() is not inter-
1828 rupted by signals, so it wasn't waking up for SIGALRM or SIGCHLD. Luckily
1829 select() can be used as an interruptible sleep() on all versions of Unix. */
1830
1831 else
1832 {
1833 struct timeval tv;
1834 tv.tv_sec = queue_interval;
1835 tv.tv_usec = 0;
1836 select(0, NULL, NULL, NULL, &tv);
1837 handle_ending_processes();
1838 }
1839
1840 /* Re-enable the SIGCHLD handler if it has been run. It can't do it
1841 for itself, because it isn't doing the waiting itself. */
1842
1843 if (sigchld_seen)
1844 {
1845 sigchld_seen = FALSE;
1846 os_non_restarting_signal(SIGCHLD, main_sigchld_handler);
1847 }
1848
1849 /* Handle being woken by SIGHUP. We know at this point that the result
1850 of accept() has been dealt with, so we can re-exec exim safely, first
1851 closing the listening sockets so that they can be reused. Cancel any pending
1852 alarm in case it is just about to go off, and set SIGHUP to be ignored so
1853 that another HUP in quick succession doesn't clobber the new daemon before it
1854 gets going. All log files get closed by the close-on-exec flag; however, if
1855 the exec fails, we need to close the logs. */
1856
1857 if (sighup_seen)
1858 {
1859 int sk;
1860 log_write(0, LOG_MAIN, "pid %d: SIGHUP received: re-exec daemon",
1861 getpid());
1862 for (sk = 0; sk < listen_socket_count; sk++)
1863 (void)close(listen_sockets[sk]);
1864 alarm(0);
1865 signal(SIGHUP, SIG_IGN);
1866 sighup_argv[0] = exim_path;
1867 exim_nullstd();
1868 execv(CS exim_path, (char *const *)sighup_argv);
1869 log_write(0, LOG_MAIN|LOG_PANIC_DIE, "pid %d: exec of %s failed: %s",
1870 getpid(), exim_path, strerror(errno));
1871 log_close_all();
1872 }
1873
1874 } /* End of main loop */
1875
1876 /* Control never reaches here */
1877 }
1878
1879 /* End of exim_daemon.c */
1880