Documentation for randint. Better randomness defaults. Fixes: #722
[exim.git] / src / src / daemon.c
1 /* $Cambridge: exim/src/src/daemon.c,v 1.26 2008/12/12 14:44:25 nm4 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 BOOL session_local_queue_only;
369 #ifdef SA_NOCLDWAIT
370 struct sigaction act;
371 #endif
372
373 smtp_accept_count++; /* So that it includes this process */
374
375 /* May have been modified for the subprocess */
376
377 log_write_selector = use_log_write_selector;
378
379 /* Get the local interface address into permanent store */
380
381 store_pool = POOL_PERM;
382 interface_address = string_copy(interface_address);
383 store_pool = old_pool;
384
385 /* Check for a tls-on-connect port */
386
387 if (host_is_tls_on_connect_port(interface_port)) tls_on_connect = TRUE;
388
389 /* Expand smtp_active_hostname if required. We do not do this any earlier,
390 because it may depend on the local interface address (indeed, that is most
391 likely what it depends on.) */
392
393 smtp_active_hostname = primary_hostname;
394 if (raw_active_hostname != NULL)
395 {
396 uschar *nah = expand_string(raw_active_hostname);
397 if (nah == NULL)
398 {
399 if (!expand_string_forcedfail)
400 {
401 log_write(0, LOG_MAIN|LOG_PANIC, "failed to expand \"%s\" "
402 "(smtp_active_hostname): %s", raw_active_hostname,
403 expand_string_message);
404 smtp_printf("421 Local configuration error; "
405 "please try again later.\r\n");
406 mac_smtp_fflush();
407 search_tidyup();
408 _exit(EXIT_FAILURE);
409 }
410 }
411 else if (nah[0] != 0) smtp_active_hostname = nah;
412 }
413
414 /* Initialize the queueing flags */
415
416 queue_check_only();
417 session_local_queue_only = queue_only;
418
419 /* Close the listening sockets, and set the SIGCHLD handler to SIG_IGN.
420 We also attempt to set things up so that children are automatically reaped,
421 but just in case this isn't available, there's a paranoid waitpid() in the
422 loop too (except for systems where we are sure it isn't needed). See the more
423 extensive comment before the reception loop in exim.c for a fuller
424 explanation of this logic. */
425
426 for (i = 0; i < listen_socket_count; i++) (void)close(listen_sockets[i]);
427
428 #ifdef SA_NOCLDWAIT
429 act.sa_handler = SIG_IGN;
430 sigemptyset(&(act.sa_mask));
431 act.sa_flags = SA_NOCLDWAIT;
432 sigaction(SIGCHLD, &act, NULL);
433 #else
434 signal(SIGCHLD, SIG_IGN);
435 #endif
436
437 /* Attempt to get an id from the sending machine via the RFC 1413
438 protocol. We do this in the sub-process in order not to hold up the
439 main process if there is any delay. Then set up the fullhost information
440 in case there is no HELO/EHLO.
441
442 If debugging is enabled only for the daemon, we must turn if off while
443 finding the id, but turn it on again afterwards so that information about the
444 incoming connection is output. */
445
446 if (debug_daemon) debug_selector = 0;
447 verify_get_ident(IDENT_PORT);
448 host_build_sender_fullhost();
449 debug_selector = save_debug_selector;
450
451 DEBUG(D_any)
452 debug_printf("Process %d is handling incoming connection from %s\n",
453 (int)getpid(), sender_fullhost);
454
455 /* Now disable debugging permanently if it's required only for the daemon
456 process. */
457
458 if (debug_daemon) debug_selector = 0;
459
460 /* If there are too many child processes for immediate delivery,
461 set the session_local_queue_only flag, which is initialized from the
462 configured value and may therefore already be TRUE. Leave logging
463 till later so it will have a message id attached. Note that there is no
464 possibility of re-calculating this per-message, because the value of
465 smtp_accept_count does not change in this subprocess. */
466
467 if (smtp_accept_queue > 0 && smtp_accept_count > smtp_accept_queue)
468 {
469 session_local_queue_only = TRUE;
470 queue_only_reason = 1;
471 }
472
473 /* Handle the start of the SMTP session, then loop, accepting incoming
474 messages from the SMTP connection. The end will come at the QUIT command,
475 when smtp_setup_msg() returns 0. A break in the connection causes the
476 process to die (see accept.c).
477
478 NOTE: We do *not* call smtp_log_no_mail() if smtp_start_session() fails,
479 because a log line has already been written for all its failure exists
480 (usually "connection refused: <reason>") and writing another one is
481 unnecessary clutter. */
482
483 if (!smtp_start_session())
484 {
485 mac_smtp_fflush();
486 search_tidyup();
487 _exit(EXIT_SUCCESS);
488 }
489
490 for (;;)
491 {
492 int rc;
493 message_id[0] = 0; /* Clear out any previous message_id */
494 reset_point = store_get(0); /* Save current store high water point */
495
496 DEBUG(D_any)
497 debug_printf("Process %d is ready for new message\n", (int)getpid());
498
499 /* Smtp_setup_msg() returns 0 on QUIT or if the call is from an
500 unacceptable host or if an ACL "drop" command was triggered, -1 on
501 connection lost, and +1 on validly reaching DATA. Receive_msg() almost
502 always returns TRUE when smtp_input is true; just retry if no message was
503 accepted (can happen for invalid message parameters). However, it can yield
504 FALSE if the connection was forcibly dropped by the DATA ACL. */
505
506 if ((rc = smtp_setup_msg()) > 0)
507 {
508 BOOL ok = receive_msg(FALSE);
509 search_tidyup(); /* Close cached databases */
510 if (!ok) /* Connection was dropped */
511 {
512 mac_smtp_fflush();
513 smtp_log_no_mail(); /* Log no mail if configured */
514 _exit(EXIT_SUCCESS);
515 }
516 if (message_id[0] == 0) continue; /* No message was accepted */
517 }
518 else
519 {
520 mac_smtp_fflush();
521 search_tidyup();
522 smtp_log_no_mail(); /* Log no mail if configured */
523 _exit((rc == 0)? EXIT_SUCCESS : EXIT_FAILURE);
524 }
525
526 /* Show the recipients when debugging */
527
528 DEBUG(D_receive)
529 {
530 int i;
531 if (sender_address != NULL)
532 debug_printf("Sender: %s\n", sender_address);
533 if (recipients_list != NULL)
534 {
535 debug_printf("Recipients:\n");
536 for (i = 0; i < recipients_count; i++)
537 debug_printf(" %s\n", recipients_list[i].address);
538 }
539 }
540
541 /* A message has been accepted. Clean up any previous delivery processes
542 that have completed and are defunct, on systems where they don't go away
543 by themselves (see comments when setting SIG_IGN above). On such systems
544 (if any) these delivery processes hang around after termination until
545 the next message is received. */
546
547 #ifndef SIG_IGN_WORKS
548 while (waitpid(-1, NULL, WNOHANG) > 0);
549 #endif
550
551 /* Reclaim up the store used in accepting this message */
552
553 store_reset(reset_point);
554
555 /* If queue_only is set or if there are too many incoming connections in
556 existence, session_local_queue_only will be TRUE. If it is not, check
557 whether we have received too many messages in this session for immediate
558 delivery. */
559
560 if (!session_local_queue_only &&
561 smtp_accept_queue_per_connection > 0 &&
562 receive_messagecount > smtp_accept_queue_per_connection)
563 {
564 session_local_queue_only = TRUE;
565 queue_only_reason = 2;
566 }
567
568 /* Initialize local_queue_only from session_local_queue_only. If it is not
569 true, and queue_only_load is set, check that the load average is below it.
570 If local_queue_only is set by this means, we also set if for the session if
571 queue_only_load_latch is true (the default). This means that, once set,
572 local_queue_only remains set for any subsequent messages on the same SMTP
573 connection. This is a deliberate choice; even though the load average may
574 fall, it doesn't seem right to deliver later messages on the same call when
575 not delivering earlier ones. However, the are special circumstances such as
576 very long-lived connections from scanning appliances where this is not the
577 best strategy. In such cases, queue_only_load_latch should be set false. */
578
579 local_queue_only = session_local_queue_only;
580 if (!local_queue_only && queue_only_load >= 0)
581 {
582 local_queue_only = (load_average = OS_GETLOADAVG()) > queue_only_load;
583 if (local_queue_only)
584 {
585 queue_only_reason = 3;
586 if (queue_only_load_latch) session_local_queue_only = TRUE;
587 }
588 }
589
590 /* Log the queueing here, when it will get a message id attached, but
591 not if queue_only is set (case 0). */
592
593 if (local_queue_only) switch(queue_only_reason)
594 {
595 case 1:
596 log_write(L_delay_delivery,
597 LOG_MAIN, "no immediate delivery: too many connections "
598 "(%d, max %d)", smtp_accept_count, smtp_accept_queue);
599 break;
600
601 case 2:
602 log_write(L_delay_delivery,
603 LOG_MAIN, "no immediate delivery: more than %d messages "
604 "received in one connection", smtp_accept_queue_per_connection);
605 break;
606
607 case 3:
608 log_write(L_delay_delivery,
609 LOG_MAIN, "no immediate delivery: load average %.2f",
610 (double)load_average/1000.0);
611 break;
612 }
613
614 /* If a delivery attempt is required, spin off a new process to handle it.
615 If we are not root, we have to re-exec exim unless deliveries are being
616 done unprivileged. */
617
618 else if (!queue_only_policy && !deliver_freeze)
619 {
620 pid_t dpid;
621
622 /* Before forking, ensure that the C output buffer is flushed. Otherwise
623 anything that it in it will get duplicated, leading to duplicate copies
624 of the pending output. */
625
626 mac_smtp_fflush();
627
628 if ((dpid = fork()) == 0)
629 {
630 (void)fclose(smtp_in);
631 (void)fclose(smtp_out);
632
633 /* Don't ever molest the parent's SSL connection, but do clean up
634 the data structures if necessary. */
635
636 #ifdef SUPPORT_TLS
637 tls_close(FALSE);
638 #endif
639
640 /* Reset SIGHUP and SIGCHLD in the child in both cases. */
641
642 signal(SIGHUP, SIG_DFL);
643 signal(SIGCHLD, SIG_DFL);
644
645 if (geteuid() != root_uid && !deliver_drop_privilege)
646 {
647 signal(SIGALRM, SIG_DFL);
648 (void)child_exec_exim(CEE_EXEC_PANIC, FALSE, NULL, FALSE, 2, US"-Mc",
649 message_id);
650 /* Control does not return here. */
651 }
652
653 /* No need to re-exec; SIGALRM remains set to the default handler */
654
655 (void)deliver_message(message_id, FALSE, FALSE);
656 search_tidyup();
657 _exit(EXIT_SUCCESS);
658 }
659
660 if (dpid > 0)
661 {
662 DEBUG(D_any) debug_printf("forked delivery process %d\n", (int)dpid);
663 }
664 else
665 {
666 log_write(0, LOG_MAIN|LOG_PANIC, "daemon: delivery process fork "
667 "failed: %s", strerror(errno));
668 }
669 }
670 }
671 }
672
673
674 /* Carrying on in the parent daemon process... Can't do much if the fork
675 failed. Otherwise, keep count of the number of accepting processes and
676 remember the pid for ticking off when the child completes. */
677
678 if (pid < 0)
679 {
680 never_error(US"daemon: accept process fork failed", US"Fork failed", errno);
681 }
682 else
683 {
684 int i;
685 for (i = 0; i < smtp_accept_max; ++i)
686 {
687 if (smtp_slots[i].pid <= 0)
688 {
689 smtp_slots[i].pid = pid;
690 if (smtp_accept_max_per_host != NULL)
691 smtp_slots[i].host_address = string_copy_malloc(sender_host_address);
692 smtp_accept_count++;
693 break;
694 }
695 }
696 DEBUG(D_any) debug_printf("%d SMTP accept process%s running\n",
697 smtp_accept_count, (smtp_accept_count == 1)? "" : "es");
698 }
699
700 /* Get here via goto in error cases */
701
702 ERROR_RETURN:
703
704 /* Close the streams associated with the socket which will also close the
705 socket fds in this process. We can't do anything if fclose() fails, but
706 logging brings it to someone's attention. However, "connection reset by peer"
707 isn't really a problem, so skip that one. On Solaris, a dropped connection can
708 manifest itself as a broken pipe, so drop that one too. If the streams don't
709 exist, something went wrong while setting things up. Make sure the socket
710 descriptors are closed, in order to drop the connection. */
711
712 if (smtp_out != NULL)
713 {
714 if (fclose(smtp_out) != 0 && errno != ECONNRESET && errno != EPIPE)
715 log_write(0, LOG_MAIN|LOG_PANIC, "daemon: fclose(smtp_out) failed: %s",
716 strerror(errno));
717 smtp_out = NULL;
718 }
719 else (void)close(accept_socket);
720
721 if (smtp_in != NULL)
722 {
723 if (fclose(smtp_in) != 0 && errno != ECONNRESET && errno != EPIPE)
724 log_write(0, LOG_MAIN|LOG_PANIC, "daemon: fclose(smtp_in) failed: %s",
725 strerror(errno));
726 smtp_in = NULL;
727 }
728 else (void)close(dup_accept_socket);
729
730 /* Release any store used in this process, including the store used for holding
731 the incoming host address and an expanded active_hostname. */
732
733 store_reset(reset_point);
734 sender_host_address = NULL;
735 }
736
737
738
739
740 /*************************************************
741 * Check wildcard listen special cases *
742 *************************************************/
743
744 /* This function is used when binding and listening on lists of addresses and
745 ports. It tests for special cases of wildcard listening, when IPv4 and IPv6
746 sockets may interact in different ways in different operating systems. It is
747 passed an error number, the list of listening addresses, and the current
748 address. Two checks are available: for a previous wildcard IPv6 address, or for
749 a following wildcard IPv4 address, in both cases on the same port.
750
751 In practice, pairs of wildcard addresses should be adjacent in the address list
752 because they are sorted that way below.
753
754 Arguments:
755 eno the error number
756 addresses the list of addresses
757 ipa the current IP address
758 back if TRUE, check for previous wildcard IPv6 address
759 if FALSE, check for a following wildcard IPv4 address
760
761 Returns: TRUE or FALSE
762 */
763
764 static BOOL
765 check_special_case(int eno, ip_address_item *addresses, ip_address_item *ipa,
766 BOOL back)
767 {
768 ip_address_item *ipa2;
769
770 /* For the "back" case, if the failure was "address in use" for a wildcard IPv4
771 address, seek a previous IPv6 wildcard address on the same port. As it is
772 previous, it must have been successfully bound and be listening. Flag it as a
773 "6 including 4" listener. */
774
775 if (back)
776 {
777 if (eno != EADDRINUSE || ipa->address[0] != 0) return FALSE;
778 for (ipa2 = addresses; ipa2 != ipa; ipa2 = ipa2->next)
779 {
780 if (ipa2->address[1] == 0 && ipa2->port == ipa->port)
781 {
782 ipa2->v6_include_v4 = TRUE;
783 return TRUE;
784 }
785 }
786 }
787
788 /* For the "forward" case, if the current address is a wildcard IPv6 address,
789 we seek a following wildcard IPv4 address on the same port. */
790
791 else
792 {
793 if (ipa->address[0] != ':' || ipa->address[1] != 0) return FALSE;
794 for (ipa2 = ipa->next; ipa2 != NULL; ipa2 = ipa2->next)
795 if (ipa2->address[0] == 0 && ipa->port == ipa2->port) return TRUE;
796 }
797
798 return FALSE;
799 }
800
801
802
803
804 /*************************************************
805 * Handle terminating subprocesses *
806 *************************************************/
807
808 /* Handle the termination of child processes. Theoretically, this need be done
809 only when sigchld_seen is TRUE, but rumour has it that some systems lose
810 SIGCHLD signals at busy times, so to be on the safe side, this function is
811 called each time round. It shouldn't be too expensive.
812
813 Arguments: none
814 Returns: nothing
815 */
816
817 static void
818 handle_ending_processes(void)
819 {
820 int status;
821 pid_t pid;
822
823 while ((pid = waitpid(-1, &status, WNOHANG)) > 0)
824 {
825 int i;
826 DEBUG(D_any) debug_printf("child %d ended: status=0x%x\n", (int)pid,
827 status);
828
829 /* If it's a listening daemon for which we are keeping track of individual
830 subprocesses, deal with an accepting process that has terminated. */
831
832 if (smtp_slots != NULL)
833 {
834 for (i = 0; i < smtp_accept_max; i++)
835 {
836 if (smtp_slots[i].pid == pid)
837 {
838 if (smtp_slots[i].host_address != NULL)
839 store_free(smtp_slots[i].host_address);
840 smtp_slots[i] = empty_smtp_slot;
841 if (--smtp_accept_count < 0) smtp_accept_count = 0;
842 DEBUG(D_any) debug_printf("%d SMTP accept process%s now running\n",
843 smtp_accept_count, (smtp_accept_count == 1)? "" : "es");
844 break;
845 }
846 }
847 if (i < smtp_accept_max) continue; /* Found an accepting process */
848 }
849
850 /* If it wasn't an accepting process, see if it was a queue-runner
851 process that we are tracking. */
852
853 if (queue_pid_slots != NULL)
854 {
855 for (i = 0; i < queue_run_max; i++)
856 {
857 if (queue_pid_slots[i] == pid)
858 {
859 queue_pid_slots[i] = 0;
860 if (--queue_run_count < 0) queue_run_count = 0;
861 DEBUG(D_any) debug_printf("%d queue-runner process%s now running\n",
862 queue_run_count, (queue_run_count == 1)? "" : "es");
863 break;
864 }
865 }
866 }
867 }
868 }
869
870
871
872 /*************************************************
873 * Exim Daemon Mainline *
874 *************************************************/
875
876 /* The daemon can do two jobs, either of which is optional:
877
878 (1) Listens for incoming SMTP calls and spawns off a sub-process to handle
879 each one. This is requested by the -bd option, with -oX specifying the SMTP
880 port on which to listen (for testing).
881
882 (2) Spawns a queue-running process every so often. This is controlled by the
883 -q option with a an interval time. (If no time is given, a single queue run
884 is done from the main function, and control doesn't get here.)
885
886 Root privilege is required in order to attach to port 25. Some systems require
887 it when calling socket() rather than bind(). To cope with all cases, we run as
888 root for both socket() and bind(). Some systems also require root in order to
889 write to the pid file directory. This function must therefore be called as root
890 if it is to work properly in all circumstances. Once the socket is bound and
891 the pid file written, root privilege is given up if there is an exim uid.
892
893 There are no arguments to this function, and it never returns. */
894
895 void
896 daemon_go(void)
897 {
898 struct passwd *pw;
899 int *listen_sockets = NULL;
900 int listen_socket_count = 0;
901 ip_address_item *addresses = NULL;
902
903 /* If any debugging options are set, turn on the D_pid bit so that all
904 debugging lines get the pid added. */
905
906 DEBUG(D_any|D_v) debug_selector |= D_pid;
907
908
909 /* Do the preparation for setting up a listener on one or more interfaces, and
910 possible on various ports. This is controlled by the combination of
911 local_interfaces (which can set IP addresses and ports) and daemon_smtp_port
912 (which is a list of default ports to use for those items in local_interfaces
913 that do not specify a port). The -oX command line option can be used to
914 override one or both of these options.
915
916 If local_interfaces is not set, the default is to listen on all interfaces.
917 When it is set, it can include "all IPvx interfaces" as an item. This is useful
918 when different ports are in use.
919
920 It turns out that listening on all interfaces is messy in an IPv6 world,
921 because several different implementation approaches have been taken. This code
922 is now supposed to work with all of them. The point of difference is whether an
923 IPv6 socket that is listening on all interfaces will receive incoming IPv4
924 calls or not. We also have to cope with the case when IPv6 libraries exist, but
925 there is no IPv6 support in the kernel.
926
927 . On Solaris, an IPv6 socket will accept IPv4 calls, and give them as mapped
928 addresses. However, if an IPv4 socket is also listening on all interfaces,
929 calls are directed to the appropriate socket.
930
931 . On (some versions of) Linux, an IPv6 socket will accept IPv4 calls, and
932 give them as mapped addresses, but an attempt also to listen on an IPv4
933 socket on all interfaces causes an error.
934
935 . On OpenBSD, an IPv6 socket will not accept IPv4 calls. You have to set up
936 two sockets if you want to accept both kinds of call.
937
938 . FreeBSD is like OpenBSD, but it has the IPV6_V6ONLY socket option, which
939 can be turned off, to make it behave like the versions of Linux described
940 above.
941
942 . I heard a report that the USAGI IPv6 stack for Linux has implemented
943 IPV6_V6ONLY.
944
945 So, what we do when IPv6 is supported is as follows:
946
947 (1) After it is set up, the list of interfaces is scanned for wildcard
948 addresses. If an IPv6 and an IPv4 wildcard are both found for the same
949 port, the list is re-arranged so that they are together, with the IPv6
950 wildcard first.
951
952 (2) If the creation of a wildcard IPv6 socket fails, we just log the error and
953 carry on if an IPv4 wildcard socket for the same port follows later in the
954 list. This allows Exim to carry on in the case when the kernel has no IPv6
955 support.
956
957 (3) Having created an IPv6 wildcard socket, we try to set IPV6_V6ONLY if that
958 option is defined. However, if setting fails, carry on regardless (but log
959 the incident).
960
961 (4) If binding or listening on an IPv6 wildcard socket fails, it is a serious
962 error.
963
964 (5) If binding or listening on an IPv4 wildcard socket fails with the error
965 EADDRINUSE, and a previous interface was an IPv6 wildcard for the same
966 port (which must have succeeded or we wouldn't have got this far), we
967 assume we are in the situation where just a single socket is permitted,
968 and ignore the error.
969
970 Phew!
971
972 The preparation code decodes options and sets up the relevant data. We do this
973 first, so that we can return non-zero if there are any syntax errors, and also
974 write to stderr. */
975
976 if (daemon_listen)
977 {
978 int *default_smtp_port;
979 int sep;
980 int pct = 0;
981 uschar *s;
982 uschar *list;
983 uschar *local_iface_source = US"local_interfaces";
984 ip_address_item *ipa;
985 ip_address_item **pipa;
986
987 /* If any option requiring a load average to be available during the
988 reception of a message is set, call os_getloadavg() while we are root
989 for those OS for which this is necessary the first time it is called (in
990 order to perform an "open" on the kernel memory file). */
991
992 #ifdef LOAD_AVG_NEEDS_ROOT
993 if (queue_only_load >= 0 || smtp_load_reserve >= 0 ||
994 (deliver_queue_load_max >= 0 && deliver_drop_privilege))
995 (void)os_getloadavg();
996 #endif
997
998 /* If -oX was used, disable the writing of a pid file unless -oP was
999 explicitly used to force it. Then scan the string given to -oX. Any items
1000 that contain neither a dot nor a colon are used to override daemon_smtp_port.
1001 Any other items are used to override local_interfaces. */
1002
1003 if (override_local_interfaces != NULL)
1004 {
1005 uschar *new_smtp_port = NULL;
1006 uschar *new_local_interfaces = NULL;
1007 int portsize = 0;
1008 int portptr = 0;
1009 int ifacesize = 0;
1010 int ifaceptr = 0;
1011
1012 if (override_pid_file_path == NULL) write_pid = FALSE;
1013
1014 list = override_local_interfaces;
1015 sep = 0;
1016 while ((s = string_nextinlist(&list,&sep,big_buffer,big_buffer_size))
1017 != NULL)
1018 {
1019 uschar joinstr[4];
1020 uschar **ptr;
1021 int *sizeptr;
1022 int *ptrptr;
1023
1024 if (Ustrpbrk(s, ".:") == NULL)
1025 {
1026 ptr = &new_smtp_port;
1027 sizeptr = &portsize;
1028 ptrptr = &portptr;
1029 }
1030 else
1031 {
1032 ptr = &new_local_interfaces;
1033 sizeptr = &ifacesize;
1034 ptrptr = &ifaceptr;
1035 }
1036
1037 if (*ptr == NULL)
1038 {
1039 joinstr[0] = sep;
1040 joinstr[1] = ' ';
1041 *ptr = string_cat(*ptr, sizeptr, ptrptr, US"<", 1);
1042 }
1043
1044 *ptr = string_cat(*ptr, sizeptr, ptrptr, joinstr, 2);
1045 *ptr = string_cat(*ptr, sizeptr, ptrptr, s, Ustrlen(s));
1046 }
1047
1048 if (new_smtp_port != NULL)
1049 {
1050 new_smtp_port[portptr] = 0;
1051 daemon_smtp_port = new_smtp_port;
1052 DEBUG(D_any) debug_printf("daemon_smtp_port overridden by -oX:\n %s\n",
1053 daemon_smtp_port);
1054 }
1055
1056 if (new_local_interfaces != NULL)
1057 {
1058 new_local_interfaces[ifaceptr] = 0;
1059 local_interfaces = new_local_interfaces;
1060 local_iface_source = US"-oX data";
1061 DEBUG(D_any) debug_printf("local_interfaces overridden by -oX:\n %s\n",
1062 local_interfaces);
1063 }
1064 }
1065
1066 /* Create a list of default SMTP ports, to be used if local_interfaces
1067 contains entries without explict ports. First count the number of ports, then
1068 build a translated list in a vector. */
1069
1070 list = daemon_smtp_port;
1071 sep = 0;
1072 while ((s = string_nextinlist(&list,&sep,big_buffer,big_buffer_size)) != NULL)
1073 pct++;
1074 default_smtp_port = store_get((pct+1) * sizeof(int));
1075 list = daemon_smtp_port;
1076 sep = 0;
1077 for (pct = 0;
1078 (s = string_nextinlist(&list,&sep,big_buffer,big_buffer_size)) != NULL;
1079 pct++)
1080 {
1081 if (isdigit(*s))
1082 {
1083 uschar *end;
1084 default_smtp_port[pct] = Ustrtol(s, &end, 0);
1085 if (end != s + Ustrlen(s))
1086 log_write(0, LOG_PANIC_DIE|LOG_CONFIG, "invalid SMTP port: %s", s);
1087 }
1088 else
1089 {
1090 struct servent *smtp_service = getservbyname(CS s, "tcp");
1091 if (smtp_service == NULL)
1092 log_write(0, LOG_PANIC_DIE|LOG_CONFIG, "TCP port \"%s\" not found", s);
1093 default_smtp_port[pct] = ntohs(smtp_service->s_port);
1094 }
1095 }
1096 default_smtp_port[pct] = 0;
1097
1098 /* Create the list of local interfaces, possibly with ports included. This
1099 list may contain references to 0.0.0.0 and ::0 as wildcards. These special
1100 values are converted below. */
1101
1102 addresses = host_build_ifacelist(local_interfaces, local_iface_source);
1103
1104 /* In the list of IP addresses, convert 0.0.0.0 into an empty string, and ::0
1105 into the string ":". We use these to recognize wildcards in IPv4 and IPv6. In
1106 fact, many IP stacks recognize 0.0.0.0 and ::0 and handle them as wildcards
1107 anyway, but we need to know which are the wildcard addresses, and the shorter
1108 strings are neater.
1109
1110 In the same scan, fill in missing port numbers from the default list. When
1111 there is more than one item in the list, extra items are created. */
1112
1113 for (ipa = addresses; ipa != NULL; ipa = ipa->next)
1114 {
1115 int i;
1116
1117 if (Ustrcmp(ipa->address, "0.0.0.0") == 0) ipa->address[0] = 0;
1118 else if (Ustrcmp(ipa->address, "::0") == 0)
1119 {
1120 ipa->address[0] = ':';
1121 ipa->address[1] = 0;
1122 }
1123
1124 if (ipa->port > 0) continue;
1125
1126 if (daemon_smtp_port[0] <= 0)
1127 log_write(0, LOG_MAIN|LOG_PANIC_DIE, "no port specified for interface "
1128 "%s and daemon_smtp_port is unset; cannot start daemon",
1129 (ipa->address[0] == 0)? US"\"all IPv4\"" :
1130 (ipa->address[1] == 0)? US"\"all IPv6\"" : ipa->address);
1131 ipa->port = default_smtp_port[0];
1132 for (i = 1; default_smtp_port[i] > 0; i++)
1133 {
1134 ip_address_item *new = store_get(sizeof(ip_address_item));
1135 memcpy(new->address, ipa->address, Ustrlen(ipa->address) + 1);
1136 new->port = default_smtp_port[i];
1137 new->next = ipa->next;
1138 ipa->next = new;
1139 ipa = new;
1140 }
1141 }
1142
1143 /* Scan the list of addresses for wildcards. If we find an IPv4 and an IPv6
1144 wildcard for the same port, ensure that (a) they are together and (b) the
1145 IPv6 address comes first. This makes handling the messy features easier, and
1146 also simplifies the construction of the "daemon started" log line. */
1147
1148 pipa = &addresses;
1149 for (ipa = addresses; ipa != NULL; pipa = &(ipa->next), ipa = ipa->next)
1150 {
1151 ip_address_item *ipa2;
1152
1153 /* Handle an IPv4 wildcard */
1154
1155 if (ipa->address[0] == 0)
1156 {
1157 for (ipa2 = ipa; ipa2->next != NULL; ipa2 = ipa2->next)
1158 {
1159 ip_address_item *ipa3 = ipa2->next;
1160 if (ipa3->address[0] == ':' &&
1161 ipa3->address[1] == 0 &&
1162 ipa3->port == ipa->port)
1163 {
1164 ipa2->next = ipa3->next;
1165 ipa3->next = ipa;
1166 *pipa = ipa3;
1167 break;
1168 }
1169 }
1170 }
1171
1172 /* Handle an IPv6 wildcard. */
1173
1174 else if (ipa->address[0] == ':' && ipa->address[1] == 0)
1175 {
1176 for (ipa2 = ipa; ipa2->next != NULL; ipa2 = ipa2->next)
1177 {
1178 ip_address_item *ipa3 = ipa2->next;
1179 if (ipa3->address[0] == 0 && ipa3->port == ipa->port)
1180 {
1181 ipa2->next = ipa3->next;
1182 ipa3->next = ipa->next;
1183 ipa->next = ipa3;
1184 ipa = ipa3;
1185 break;
1186 }
1187 }
1188 }
1189 }
1190
1191 /* Get a vector to remember all the sockets in */
1192
1193 for (ipa = addresses; ipa != NULL; ipa = ipa->next)
1194 listen_socket_count++;
1195 listen_sockets = store_get(sizeof(int *) * listen_socket_count);
1196
1197 /* Do a sanity check on the max connects value just to save us from getting
1198 a huge amount of store. */
1199
1200 if (smtp_accept_max > 4095) smtp_accept_max = 4096;
1201
1202 /* There's no point setting smtp_accept_queue unless it is less than the max
1203 connects limit. The configuration reader ensures that the max is set if the
1204 queue-only option is set. */
1205
1206 if (smtp_accept_queue > smtp_accept_max) smtp_accept_queue = 0;
1207
1208 /* Get somewhere to keep the list of SMTP accepting pids if we are keeping
1209 track of them for total number and queue/host limits. */
1210
1211 if (smtp_accept_max > 0)
1212 {
1213 int i;
1214 smtp_slots = store_get(smtp_accept_max * sizeof(smtp_slot));
1215 for (i = 0; i < smtp_accept_max; i++) smtp_slots[i] = empty_smtp_slot;
1216 }
1217 }
1218
1219 /* The variable background_daemon is always false when debugging, but
1220 can also be forced false in order to keep a non-debugging daemon in the
1221 foreground. If background_daemon is true, close all open file descriptors that
1222 we know about, but then re-open stdin, stdout, and stderr to /dev/null.
1223
1224 This is protection against any called functions (in libraries, or in
1225 Perl, or whatever) that think they can write to stderr (or stdout). Before this
1226 was added, it was quite likely that an SMTP connection would use one of these
1227 file descriptors, in which case writing random stuff to it caused chaos.
1228
1229 Then disconnect from the controlling terminal, Most modern Unixes seem to have
1230 setsid() for getting rid of the controlling terminal. For any OS that doesn't,
1231 setsid() can be #defined as a no-op, or as something else. */
1232
1233 if (background_daemon)
1234 {
1235 log_close_all(); /* Just in case anything was logged earlier */
1236 search_tidyup(); /* Just in case any were used in reading the config. */
1237 (void)close(0); /* Get rid of stdin/stdout/stderr */
1238 (void)close(1);
1239 (void)close(2);
1240 exim_nullstd(); /* Connect stdin/stdout/stderr to /dev/null */
1241 log_stderr = NULL; /* So no attempt to copy paniclog output */
1242
1243 /* If the parent process of this one has pid == 1, we are re-initializing the
1244 daemon as the result of a SIGHUP. In this case, there is no need to do
1245 anything, because the controlling terminal has long gone. Otherwise, fork, in
1246 case current process is a process group leader (see 'man setsid' for an
1247 explanation) before calling setsid(). */
1248
1249 if (getppid() != 1)
1250 {
1251 pid_t pid = fork();
1252 if (pid < 0) log_write(0, LOG_MAIN|LOG_PANIC_DIE,
1253 "fork() failed when starting daemon: %s", strerror(errno));
1254 if (pid > 0) exit(EXIT_SUCCESS); /* in parent process, just exit */
1255 (void)setsid(); /* release controlling terminal */
1256 }
1257 }
1258
1259 /* We are now in the disconnected, daemon process (unless debugging). Set up
1260 the listening sockets if required. */
1261
1262 if (daemon_listen)
1263 {
1264 int sk;
1265 int on = 1;
1266 ip_address_item *ipa;
1267
1268 /* For each IP address, create a socket, bind it to the appropriate port, and
1269 start listening. See comments above about IPv6 sockets that may or may not
1270 accept IPv4 calls when listening on all interfaces. We also have to cope with
1271 the case of a system with IPv6 libraries, but no IPv6 support in the kernel.
1272 listening, provided a wildcard IPv4 socket for the same port follows. */
1273
1274 for (ipa = addresses, sk = 0; sk < listen_socket_count; ipa = ipa->next, sk++)
1275 {
1276 BOOL wildcard;
1277 ip_address_item *ipa2;
1278 int af;
1279
1280 if (Ustrchr(ipa->address, ':') != NULL)
1281 {
1282 af = AF_INET6;
1283 wildcard = ipa->address[1] == 0;
1284 }
1285 else
1286 {
1287 af = AF_INET;
1288 wildcard = ipa->address[0] == 0;
1289 }
1290
1291 listen_sockets[sk] = ip_socket(SOCK_STREAM, af);
1292 if (listen_sockets[sk] < 0)
1293 {
1294 if (check_special_case(0, addresses, ipa, FALSE))
1295 {
1296 log_write(0, LOG_MAIN, "Failed to create IPv6 socket for wildcard "
1297 "listening (%s): will use IPv4", strerror(errno));
1298 goto SKIP_SOCKET;
1299 }
1300 log_write(0, LOG_PANIC_DIE, "IPv%c socket creation failed: %s",
1301 (af == AF_INET6)? '6' : '4', strerror(errno));
1302 }
1303
1304 /* If this is an IPv6 wildcard socket, set IPV6_V6ONLY if that option is
1305 available. Just log failure (can get protocol not available, just like
1306 socket creation can). */
1307
1308 #ifdef IPV6_V6ONLY
1309 if (af == AF_INET6 && wildcard &&
1310 setsockopt(listen_sockets[sk], IPPROTO_IPV6, IPV6_V6ONLY, (char *)(&on),
1311 sizeof(on)) < 0)
1312 log_write(0, LOG_MAIN, "Setting IPV6_V6ONLY on daemon's IPv6 wildcard "
1313 "socket failed (%s): carrying on without it", strerror(errno));
1314 #endif /* IPV6_V6ONLY */
1315
1316 /* Set SO_REUSEADDR so that the daemon can be restarted while a connection
1317 is being handled. Without this, a connection will prevent reuse of the
1318 smtp port for listening. */
1319
1320 if (setsockopt(listen_sockets[sk], SOL_SOCKET, SO_REUSEADDR,
1321 (uschar *)(&on), sizeof(on)) < 0)
1322 log_write(0, LOG_MAIN|LOG_PANIC_DIE, "setting SO_REUSEADDR on socket "
1323 "failed when starting daemon: %s", strerror(errno));
1324
1325 /* Set TCP_NODELAY; Exim does its own buffering. There is a switch to
1326 disable this because it breaks some broken clients. */
1327
1328 if (tcp_nodelay) setsockopt(listen_sockets[sk], IPPROTO_TCP, TCP_NODELAY,
1329 (uschar *)(&on), sizeof(on));
1330
1331 /* Now bind the socket to the required port; if Exim is being restarted
1332 it may not always be possible to bind immediately, even with SO_REUSEADDR
1333 set, so try 10 times, waiting between each try. After 10 failures, we give
1334 up. In an IPv6 environment, if bind () fails with the error EADDRINUSE and
1335 we are doing wildcard IPv4 listening and there was a previous IPv6 wildcard
1336 address for the same port, ignore the error on the grounds that we must be
1337 in a system where the IPv6 socket accepts both kinds of call. This is
1338 necessary for (some release of) USAGI Linux; other IP stacks fail at the
1339 listen() stage instead. */
1340
1341 for(;;)
1342 {
1343 uschar *msg, *addr;
1344 if (ip_bind(listen_sockets[sk], af, ipa->address, ipa->port) >= 0) break;
1345 if (check_special_case(errno, addresses, ipa, TRUE))
1346 {
1347 DEBUG(D_any) debug_printf("wildcard IPv4 bind() failed after IPv6 "
1348 "listen() success; EADDRINUSE ignored\n");
1349 (void)close(listen_sockets[sk]);
1350 goto SKIP_SOCKET;
1351 }
1352 msg = US strerror(errno);
1353 addr = wildcard? ((af == AF_INET6)? US"(any IPv6)" : US"(any IPv4)") :
1354 ipa->address;
1355 if (daemon_startup_retries <= 0)
1356 log_write(0, LOG_MAIN|LOG_PANIC_DIE,
1357 "socket bind() to port %d for address %s failed: %s: "
1358 "daemon abandoned", ipa->port, addr, msg);
1359 log_write(0, LOG_MAIN, "socket bind() to port %d for address %s "
1360 "failed: %s: waiting %s before trying again (%d more %s)",
1361 ipa->port, addr, msg, readconf_printtime(daemon_startup_sleep),
1362 daemon_startup_retries, (daemon_startup_retries > 1)? "tries" : "try");
1363 daemon_startup_retries--;
1364 sleep(daemon_startup_sleep);
1365 }
1366
1367 DEBUG(D_any)
1368 {
1369 if (wildcard)
1370 debug_printf("listening on all interfaces (IPv%c) port %d\n",
1371 (af == AF_INET6)? '6' : '4', ipa->port);
1372 else
1373 debug_printf("listening on %s port %d\n", ipa->address, ipa->port);
1374 }
1375
1376 /* Start listening on the bound socket, establishing the maximum backlog of
1377 connections that is allowed. On success, continue to the next address. */
1378
1379 if (listen(listen_sockets[sk], smtp_connect_backlog) >= 0) continue;
1380
1381 /* Listening has failed. In an IPv6 environment, as for bind(), if listen()
1382 fails with the error EADDRINUSE and we are doing IPv4 wildcard listening
1383 and there was a previous successful IPv6 wildcard listen on the same port,
1384 we want to ignore the error on the grounds that we must be in a system
1385 where the IPv6 socket accepts both kinds of call. */
1386
1387 if (!check_special_case(errno, addresses, ipa, TRUE))
1388 log_write(0, LOG_PANIC_DIE, "listen() failed on interface %s: %s",
1389 wildcard? ((af == AF_INET6)? US"(any IPv6)" : US"(any IPv4)") :
1390 ipa->address,
1391 strerror(errno));
1392
1393 DEBUG(D_any) debug_printf("wildcard IPv4 listen() failed after IPv6 "
1394 "listen() success; EADDRINUSE ignored\n");
1395 (void)close(listen_sockets[sk]);
1396
1397 /* Come here if there has been a problem with the socket which we
1398 are going to ignore. We remove the address from the chain, and back up the
1399 counts. */
1400
1401 SKIP_SOCKET:
1402 sk--; /* Back up the count */
1403 listen_socket_count--; /* Reduce the total */
1404 if (ipa == addresses) addresses = ipa->next; else
1405 {
1406 for (ipa2 = addresses; ipa2->next != ipa; ipa2 = ipa2->next);
1407 ipa2->next = ipa->next;
1408 ipa = ipa2;
1409 }
1410 } /* End of bind/listen loop for each address */
1411 } /* End of setup for listening */
1412
1413
1414 /* If we are not listening, we want to write a pid file only if -oP was
1415 explicitly given. */
1416
1417 else if (override_pid_file_path == NULL) write_pid = FALSE;
1418
1419 /* Write the pid to a known file for assistance in identification, if required.
1420 We do this before giving up root privilege, because on some systems it is
1421 necessary to be root in order to write into the pid file directory. There's
1422 nothing to stop multiple daemons running, as long as no more than one listens
1423 on a given TCP/IP port on the same interface(s). However, in these
1424 circumstances it gets far too complicated to mess with pid file names
1425 automatically. Consequently, Exim 4 writes a pid file only
1426
1427 (a) When running in the test harness, or
1428 (b) When -bd is used and -oX is not used, or
1429 (c) When -oP is used to supply a path.
1430
1431 The variable daemon_write_pid is used to control this. */
1432
1433 if (running_in_test_harness || write_pid)
1434 {
1435 FILE *f;
1436
1437 if (override_pid_file_path != NULL)
1438 pid_file_path = override_pid_file_path;
1439
1440 if (pid_file_path[0] == 0)
1441 pid_file_path = string_sprintf("%s/exim-daemon.pid", spool_directory);
1442
1443 f = modefopen(pid_file_path, "wb", 0644);
1444 if (f != NULL)
1445 {
1446 (void)fprintf(f, "%d\n", (int)getpid());
1447 (void)fclose(f);
1448 DEBUG(D_any) debug_printf("pid written to %s\n", pid_file_path);
1449 }
1450 else
1451 {
1452 DEBUG(D_any)
1453 debug_printf("%s\n", string_open_failed(errno, "pid file %s",
1454 pid_file_path));
1455 }
1456 }
1457
1458 /* Set up the handler for SIGHUP, which causes a restart of the daemon. */
1459
1460 sighup_seen = FALSE;
1461 signal(SIGHUP, sighup_handler);
1462
1463 /* Give up root privilege at this point (assuming that exim_uid and exim_gid
1464 are not root). The third argument controls the running of initgroups().
1465 Normally we do this, in order to set up the groups for the Exim user. However,
1466 if we are not root at this time - some odd installations run that way - we
1467 cannot do this. */
1468
1469 exim_setugid(exim_uid, exim_gid, geteuid()==root_uid, US"running as a daemon");
1470
1471 /* Update the originator_xxx fields so that received messages as listed as
1472 coming from Exim, not whoever started the daemon. */
1473
1474 originator_uid = exim_uid;
1475 originator_gid = exim_gid;
1476 originator_login = ((pw = getpwuid(exim_uid)) != NULL)?
1477 string_copy_malloc(US pw->pw_name) : US"exim";
1478
1479 /* Get somewhere to keep the list of queue-runner pids if we are keeping track
1480 of them (and also if we are doing queue runs). */
1481
1482 if (queue_interval > 0 && queue_run_max > 0)
1483 {
1484 int i;
1485 queue_pid_slots = store_get(queue_run_max * sizeof(pid_t));
1486 for (i = 0; i < queue_run_max; i++) queue_pid_slots[i] = 0;
1487 }
1488
1489 /* Set up the handler for termination of child processes. */
1490
1491 sigchld_seen = FALSE;
1492 os_non_restarting_signal(SIGCHLD, main_sigchld_handler);
1493
1494 /* If we are to run the queue periodically, pretend the alarm has just gone
1495 off. This will cause the first queue-runner to get kicked off straight away. */
1496
1497 sigalrm_seen = (queue_interval > 0);
1498
1499 /* Log the start up of a daemon - at least one of listening or queue running
1500 must be set up. */
1501
1502 if (daemon_listen)
1503 {
1504 int i, j;
1505 int smtp_ports = 0;
1506 int smtps_ports = 0;
1507 ip_address_item *ipa;
1508 uschar *p = big_buffer;
1509 uschar *qinfo = (queue_interval > 0)?
1510 string_sprintf("-q%s", readconf_printtime(queue_interval))
1511 :
1512 US"no queue runs";
1513
1514 /* Build a list of listening addresses in big_buffer, but limit it to 10
1515 items. The style is for backwards compatibility.
1516
1517 It is now possible to have some ports listening for SMTPS (the old,
1518 deprecated protocol that starts TLS without using STARTTLS), and others
1519 listening for standard SMTP. Keep their listings separate. */
1520
1521 for (j = 0; j < 2; j++)
1522 {
1523 for (i = 0, ipa = addresses; i < 10 && ipa != NULL; i++, ipa = ipa->next)
1524 {
1525 /* First time round, look for SMTP ports; second time round, look for
1526 SMTPS ports. For the first one of each, insert leading text. */
1527
1528 if (host_is_tls_on_connect_port(ipa->port) == (j > 0))
1529 {
1530 if (j == 0)
1531 {
1532 if (smtp_ports++ == 0)
1533 {
1534 memcpy(p, "SMTP on", 8);
1535 p += 7;
1536 }
1537 }
1538 else
1539 {
1540 if (smtps_ports++ == 0)
1541 {
1542 (void)sprintf(CS p, "%sSMTPS on",
1543 (smtp_ports == 0)? "":" and for ");
1544 while (*p != 0) p++;
1545 }
1546 }
1547
1548 /* Now the information about the port (and sometimes interface) */
1549
1550 if (ipa->address[0] == ':' && ipa->address[1] == 0)
1551 {
1552 if (ipa->next != NULL && ipa->next->address[0] == 0 &&
1553 ipa->next->port == ipa->port)
1554 {
1555 (void)sprintf(CS p, " port %d (IPv6 and IPv4)", ipa->port);
1556 ipa = ipa->next;
1557 }
1558 else if (ipa->v6_include_v4)
1559 (void)sprintf(CS p, " port %d (IPv6 with IPv4)", ipa->port);
1560 else
1561 (void)sprintf(CS p, " port %d (IPv6)", ipa->port);
1562 }
1563 else if (ipa->address[0] == 0)
1564 (void)sprintf(CS p, " port %d (IPv4)", ipa->port);
1565 else
1566 (void)sprintf(CS p, " [%s]:%d", ipa->address, ipa->port);
1567 while (*p != 0) p++;
1568 }
1569 }
1570
1571 if (ipa != NULL)
1572 {
1573 memcpy(p, " ...", 5);
1574 p += 4;
1575 }
1576 }
1577
1578 log_write(0, LOG_MAIN,
1579 "exim %s daemon started: pid=%d, %s, listening for %s",
1580 version_string, getpid(), qinfo, big_buffer);
1581 set_process_info("daemon: %s, listening for %s", qinfo, big_buffer);
1582 }
1583
1584 else
1585 {
1586 log_write(0, LOG_MAIN,
1587 "exim %s daemon started: pid=%d, -q%s, not listening for SMTP",
1588 version_string, getpid(), readconf_printtime(queue_interval));
1589 set_process_info("daemon: -q%s, not listening",
1590 readconf_printtime(queue_interval));
1591 }
1592
1593
1594 /* Close the log so it can be renamed and moved. In the few cases below where
1595 this long-running process writes to the log (always exceptional conditions), it
1596 closes the log afterwards, for the same reason. */
1597
1598 log_close_all();
1599
1600 DEBUG(D_any) debug_print_ids(US"daemon running with");
1601
1602 /* Any messages accepted via this route are going to be SMTP. */
1603
1604 smtp_input = TRUE;
1605
1606 /* Enter the never-ending loop... */
1607
1608 for (;;)
1609 {
1610 #if HAVE_IPV6
1611 struct sockaddr_in6 accepted;
1612 #else
1613 struct sockaddr_in accepted;
1614 #endif
1615
1616 EXIM_SOCKLEN_T len;
1617 pid_t pid;
1618
1619 /* This code is placed first in the loop, so that it gets obeyed at the
1620 start, before the first wait. This causes the first queue-runner to be
1621 started immediately. */
1622
1623 if (sigalrm_seen)
1624 {
1625 DEBUG(D_any) debug_printf("SIGALRM received\n");
1626
1627 /* Do a full queue run in a child process, if required, unless we already
1628 have enough queue runners on the go. If we are not running as root, a
1629 re-exec is required. */
1630
1631 if (queue_interval > 0 &&
1632 (queue_run_max <= 0 || queue_run_count < queue_run_max))
1633 {
1634 if ((pid = fork()) == 0)
1635 {
1636 int sk;
1637
1638 DEBUG(D_any) debug_printf("Starting queue-runner: pid %d\n",
1639 (int)getpid());
1640
1641 /* Disable debugging if it's required only for the daemon process. We
1642 leave the above message, because it ties up with the "child ended"
1643 debugging messages. */
1644
1645 if (debug_daemon) debug_selector = 0;
1646
1647 /* Close any open listening sockets in the child */
1648
1649 for (sk = 0; sk < listen_socket_count; sk++)
1650 (void)close(listen_sockets[sk]);
1651
1652 /* Reset SIGHUP and SIGCHLD in the child in both cases. */
1653
1654 signal(SIGHUP, SIG_DFL);
1655 signal(SIGCHLD, SIG_DFL);
1656
1657 /* Re-exec if privilege has been given up, unless deliver_drop_
1658 privilege is set. Reset SIGALRM before exec(). */
1659
1660 if (geteuid() != root_uid && !deliver_drop_privilege)
1661 {
1662 uschar opt[8];
1663 uschar *p = opt;
1664 uschar *extra[5];
1665 int extracount = 1;
1666
1667 signal(SIGALRM, SIG_DFL);
1668 *p++ = '-';
1669 *p++ = 'q';
1670 if (queue_2stage) *p++ = 'q';
1671 if (queue_run_first_delivery) *p++ = 'i';
1672 if (queue_run_force) *p++ = 'f';
1673 if (deliver_force_thaw) *p++ = 'f';
1674 if (queue_run_local) *p++ = 'l';
1675 *p = 0;
1676 extra[0] = opt;
1677
1678 /* If -R or -S were on the original command line, ensure they get
1679 passed on. */
1680
1681 if (deliver_selectstring != NULL)
1682 {
1683 extra[extracount++] = deliver_selectstring_regex? US"-Rr" : US"-R";
1684 extra[extracount++] = deliver_selectstring;
1685 }
1686
1687 if (deliver_selectstring_sender != NULL)
1688 {
1689 extra[extracount++] = deliver_selectstring_sender_regex?
1690 US"-Sr" : US"-S";
1691 extra[extracount++] = deliver_selectstring_sender;
1692 }
1693
1694 /* Overlay this process with a new execution. */
1695
1696 (void)child_exec_exim(CEE_EXEC_PANIC, FALSE, NULL, TRUE, extracount,
1697 extra[0], extra[1], extra[2], extra[3], extra[4]);
1698
1699 /* Control never returns here. */
1700 }
1701
1702 /* No need to re-exec; SIGALRM remains set to the default handler */
1703
1704 queue_run(NULL, NULL, FALSE);
1705 _exit(EXIT_SUCCESS);
1706 }
1707
1708 if (pid < 0)
1709 {
1710 log_write(0, LOG_MAIN|LOG_PANIC, "daemon: fork of queue-runner "
1711 "process failed: %s", strerror(errno));
1712 log_close_all();
1713 }
1714 else
1715 {
1716 int i;
1717 for (i = 0; i < queue_run_max; ++i)
1718 {
1719 if (queue_pid_slots[i] <= 0)
1720 {
1721 queue_pid_slots[i] = pid;
1722 queue_run_count++;
1723 break;
1724 }
1725 }
1726 DEBUG(D_any) debug_printf("%d queue-runner process%s running\n",
1727 queue_run_count, (queue_run_count == 1)? "" : "es");
1728 }
1729 }
1730
1731 /* Reset the alarm clock */
1732
1733 sigalrm_seen = FALSE;
1734 alarm(queue_interval);
1735 }
1736
1737
1738 /* Sleep till a connection happens if listening, and handle the connection if
1739 that is why we woke up. The FreeBSD operating system requires the use of
1740 select() before accept() because the latter function is not interrupted by
1741 a signal, and we want to wake up for SIGCHLD and SIGALRM signals. Some other
1742 OS do notice signals in accept() but it does no harm to have the select()
1743 in for all of them - and it won't then be a lurking problem for ports to
1744 new OS. In fact, the later addition of listening on specific interfaces only
1745 requires this way of working anyway. */
1746
1747 if (daemon_listen)
1748 {
1749 int sk, lcount, select_errno;
1750 int max_socket = 0;
1751 BOOL select_failed = FALSE;
1752 fd_set select_listen;
1753
1754 FD_ZERO(&select_listen);
1755 for (sk = 0; sk < listen_socket_count; sk++)
1756 {
1757 FD_SET(listen_sockets[sk], &select_listen);
1758 if (listen_sockets[sk] > max_socket) max_socket = listen_sockets[sk];
1759 }
1760
1761 DEBUG(D_any) debug_printf("Listening...\n");
1762
1763 /* In rare cases we may have had a SIGCHLD signal in the time between
1764 setting the handler (below) and getting back here. If so, pretend that the
1765 select() was interrupted so that we reap the child. This might still leave
1766 a small window when a SIGCHLD could get lost. However, since we use SIGCHLD
1767 only to do the reaping more quickly, it shouldn't result in anything other
1768 than a delay until something else causes a wake-up. */
1769
1770 if (sigchld_seen)
1771 {
1772 lcount = -1;
1773 errno = EINTR;
1774 }
1775 else
1776 {
1777 lcount = select(max_socket + 1, (SELECT_ARG2_TYPE *)&select_listen,
1778 NULL, NULL, NULL);
1779 }
1780
1781 if (lcount < 0)
1782 {
1783 select_failed = TRUE;
1784 lcount = 1;
1785 }
1786
1787 /* Clean up any subprocesses that may have terminated. We need to do this
1788 here so that smtp_accept_max_per_host works when a connection to that host
1789 has completed, and we are about to accept a new one. When this code was
1790 later in the sequence, a new connection could be rejected, even though an
1791 old one had just finished. Preserve the errno from any select() failure for
1792 the use of the common select/accept error processing below. */
1793
1794 select_errno = errno;
1795 handle_ending_processes();
1796 errno = select_errno;
1797
1798 /* Loop for all the sockets that are currently ready to go. If select
1799 actually failed, we have set the count to 1 and select_failed=TRUE, so as
1800 to use the common error code for select/accept below. */
1801
1802 while (lcount-- > 0)
1803 {
1804 int accept_socket = -1;
1805 if (!select_failed)
1806 {
1807 for (sk = 0; sk < listen_socket_count; sk++)
1808 {
1809 if (FD_ISSET(listen_sockets[sk], &select_listen))
1810 {
1811 len = sizeof(accepted);
1812 accept_socket = accept(listen_sockets[sk],
1813 (struct sockaddr *)&accepted, &len);
1814 FD_CLR(listen_sockets[sk], &select_listen);
1815 break;
1816 }
1817 }
1818 }
1819
1820 /* If select or accept has failed and this was not caused by an
1821 interruption, log the incident and try again. With asymmetric TCP/IP
1822 routing errors such as "No route to network" have been seen here. Also
1823 "connection reset by peer" has been seen. These cannot be classed as
1824 disastrous errors, but they could fill up a lot of log. The code in smail
1825 crashes the daemon after 10 successive failures of accept, on the grounds
1826 that some OS fail continuously. Exim originally followed suit, but this
1827 appears to have caused problems. Now it just keeps going, but instead of
1828 logging each error, it batches them up when they are continuous. */
1829
1830 if (accept_socket < 0 && errno != EINTR)
1831 {
1832 if (accept_retry_count == 0)
1833 {
1834 accept_retry_errno = errno;
1835 accept_retry_select_failed = select_failed;
1836 }
1837 else
1838 {
1839 if (errno != accept_retry_errno ||
1840 select_failed != accept_retry_select_failed ||
1841 accept_retry_count >= 50)
1842 {
1843 log_write(0, LOG_MAIN | ((accept_retry_count >= 50)? LOG_PANIC : 0),
1844 "%d %s() failure%s: %s",
1845 accept_retry_count,
1846 accept_retry_select_failed? "select" : "accept",
1847 (accept_retry_count == 1)? "" : "s",
1848 strerror(accept_retry_errno));
1849 log_close_all();
1850 accept_retry_count = 0;
1851 accept_retry_errno = errno;
1852 accept_retry_select_failed = select_failed;
1853 }
1854 }
1855 accept_retry_count++;
1856 }
1857
1858 else
1859 {
1860 if (accept_retry_count > 0)
1861 {
1862 log_write(0, LOG_MAIN, "%d %s() failure%s: %s",
1863 accept_retry_count,
1864 accept_retry_select_failed? "select" : "accept",
1865 (accept_retry_count == 1)? "" : "s",
1866 strerror(accept_retry_errno));
1867 log_close_all();
1868 accept_retry_count = 0;
1869 }
1870 }
1871
1872 /* If select/accept succeeded, deal with the connection. */
1873
1874 if (accept_socket >= 0)
1875 handle_smtp_call(listen_sockets, listen_socket_count, accept_socket,
1876 (struct sockaddr *)&accepted);
1877 }
1878 }
1879
1880 /* If not listening, then just sleep for the queue interval. If we woke
1881 up early the last time for some other signal, it won't matter because
1882 the alarm signal will wake at the right time. This code originally used
1883 sleep() but it turns out that on the FreeBSD system, sleep() is not inter-
1884 rupted by signals, so it wasn't waking up for SIGALRM or SIGCHLD. Luckily
1885 select() can be used as an interruptible sleep() on all versions of Unix. */
1886
1887 else
1888 {
1889 struct timeval tv;
1890 tv.tv_sec = queue_interval;
1891 tv.tv_usec = 0;
1892 select(0, NULL, NULL, NULL, &tv);
1893 handle_ending_processes();
1894 }
1895
1896 /* Re-enable the SIGCHLD handler if it has been run. It can't do it
1897 for itself, because it isn't doing the waiting itself. */
1898
1899 if (sigchld_seen)
1900 {
1901 sigchld_seen = FALSE;
1902 os_non_restarting_signal(SIGCHLD, main_sigchld_handler);
1903 }
1904
1905 /* Handle being woken by SIGHUP. We know at this point that the result
1906 of accept() has been dealt with, so we can re-exec exim safely, first
1907 closing the listening sockets so that they can be reused. Cancel any pending
1908 alarm in case it is just about to go off, and set SIGHUP to be ignored so
1909 that another HUP in quick succession doesn't clobber the new daemon before it
1910 gets going. All log files get closed by the close-on-exec flag; however, if
1911 the exec fails, we need to close the logs. */
1912
1913 if (sighup_seen)
1914 {
1915 int sk;
1916 log_write(0, LOG_MAIN, "pid %d: SIGHUP received: re-exec daemon",
1917 getpid());
1918 for (sk = 0; sk < listen_socket_count; sk++)
1919 (void)close(listen_sockets[sk]);
1920 alarm(0);
1921 signal(SIGHUP, SIG_IGN);
1922 sighup_argv[0] = exim_path;
1923 exim_nullstd();
1924 execv(CS exim_path, (char *const *)sighup_argv);
1925 log_write(0, LOG_MAIN|LOG_PANIC_DIE, "pid %d: exec of %s failed: %s",
1926 getpid(), exim_path, strerror(errno));
1927 log_close_all();
1928 }
1929
1930 } /* End of main loop */
1931
1932 /* Control never reaches here */
1933 }
1934
1935 /* End of exim_daemon.c */
1936