Fix problem with smtp_max_per_host; test for completed processes before
[exim.git] / src / src / daemon.c
CommitLineData
bb6e88ff 1/* $Cambridge: exim/src/src/daemon.c,v 1.8 2005/02/16 15:24:21 ph10 Exp $ */
059ec3d9
PH
2
3/*************************************************
4* Exim - an Internet mail transport agent *
5*************************************************/
6
c988f1f4 7/* Copyright (c) University of Cambridge 1995 - 2005 */
059ec3d9
PH
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
18typedef 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
24expressions in assigments except as initializers in declarations). */
25
26static smtp_slot empty_smtp_slot = { 0, NULL };
27
28
29
30/*************************************************
31* Local static variables *
32*************************************************/
33
34static volatile BOOL sigchld_seen;
35static volatile BOOL sighup_seen;
36
37static int accept_retry_count = 0;
38static int accept_retry_errno;
39static BOOL accept_retry_select_failed;
40
41static int queue_run_count = 0;
bb6e88ff
PH
42static pid_t *queue_pid_slots = NULL;
43static smtp_slot *smtp_slots = NULL;
059ec3d9
PH
44
45static 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
55Argument: the signal number
56Returns: nothing
57*/
58
59static void
60sighup_handler(int sig)
61{
62sig = sig; /* Keep picky compilers happy */
63sighup_seen = TRUE;
64signal(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
74waiting here. If the signal is re-enabled, there will just be an
75infinite sequence of calls to this handler. The SIGCHLD signal is
76used just as a means of waking up the daemon so that it notices
77terminated subprocesses as soon as possible.
78
79Argument: the signal number
80Returns: nothing
81*/
82
83static void
84main_sigchld_handler(int sig)
85{
86sig = sig; /* Keep picky compilers happy */
7d468ab8 87os_non_restarting_signal(SIGCHLD, SIG_DFL);
059ec3d9 88sigchld_seen = TRUE;
059ec3d9
PH
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
100Arguments:
101 log_msg Text of message to be logged
102 smtp_msg Text of SMTP error message
103 was_errno The failing errno
104
105Returns: nothing
106*/
107
108static void
109never_error(uschar *log_msg, uschar *smtp_msg, int was_errno)
110{
111uschar *emsg = (was_errno <= 0)? US"" :
112 string_sprintf(": %s", strerror(was_errno));
113log_write(0, LOG_MAIN|LOG_PANIC, "%s%s", log_msg, emsg);
114if (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.
125If there are too many, give an error message and close down. Otherwise
126spin off a sub-process to handle the call. The list of listening sockets
127is required so that they can be closed in the sub-process. Take care not to
128leak store in this process - reset the stacking pool at the end.
129
130Arguments:
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
136Returns: nothing
137*/
138
139static void
140handle_smtp_call(int *listen_sockets, int listen_socket_count,
141 int accept_socket, struct sockaddr *accepted)
142{
143pid_t pid;
144union sockaddr_46 interface_sockaddr;
145SOCKLEN_T ifsize = sizeof(interface_sockaddr);
146int dup_accept_socket = -1;
147int max_for_this_host = 0;
148int wfsize = 0;
149int wfptr = 0;
150int use_log_write_selector = log_write_selector;
151uschar *whofrom = NULL;
152
153void *reset_point = store_get(0);
154
155/* Make the address available in ASCII representation, and also fish out
156the remote port. */
157
158sender_host_address = host_ntoa(-1, accepted, NULL, &sender_host_port);
159DEBUG(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
163input stream. These operations fail only the exceptional circumstances. Note
164that never_error() won't use smtp_out if it is NULL. */
165
166smtp_out = fdopen(accept_socket, "wb");
167if (smtp_out == NULL)
168 {
169 never_error(US"daemon: fdopen() for smtp_out failed", US"", errno);
170 goto ERROR_RETURN;
171 }
172
173dup_accept_socket = dup(accept_socket);
174if (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
181smtp_in = fdopen(dup_accept_socket, "rb");
182if (smtp_in == NULL)
183 {
184 never_error(US"daemon: fdopen() for smtp_in failed",
185 US"Connection setup failed", errno);
186 goto ERROR_RETURN;
187 }
188
189/* Get the data for the local interface address. */
190
191if (getsockname(accept_socket, (struct sockaddr *)(&interface_sockaddr),
192 &ifsize) < 0)
193 {
194 log_write(0, LOG_MAIN|LOG_PANIC, "getsockname() failed: %s",
195 strerror(errno));
196 smtp_printf("421 Local problem: getsockname() failed; please try again later\r\n");
197 goto ERROR_RETURN;
198 }
199
200interface_address = host_ntoa(-1, &interface_sockaddr, NULL, &interface_port);
201DEBUG(D_interface) debug_printf("interface address=%s port=%d\n",
202 interface_address, interface_port);
203
204/* Build a string identifying the remote host and, if requested, the port and
205the local interface data. This is for logging; at the end of this function the
206memory is reclaimed. */
207
208whofrom = string_append(whofrom, &wfsize, &wfptr, 3, "[", sender_host_address, "]");
209
210if ((log_extra_selector & LX_incoming_port) != 0)
211 whofrom = string_append(whofrom, &wfsize, &wfptr, 2, ":", string_sprintf("%d",
212 sender_host_port));
213
214if ((log_extra_selector & LX_incoming_interface) != 0)
215 whofrom = string_append(whofrom, &wfsize, &wfptr, 4, " I=[",
216 interface_address, "]:", string_sprintf("%d", interface_port));
217
218whofrom[wfptr] = 0; /* Terminate the newly-built string */
219
220/* Check maximum number of connections. We do not check for reserved
221connections or unacceptable hosts here. That is done in the subprocess because
222it might take some time. */
223
224if (smtp_accept_max > 0 && smtp_accept_count >= smtp_accept_max)
225 {
226 DEBUG(D_any) debug_printf("rejecting SMTP connection: count=%d max=%d\n",
227 smtp_accept_count, smtp_accept_max);
228 smtp_printf("421 Too many concurrent SMTP connections; "
229 "please try again later.\r\n");
230 log_write(L_connection_reject,
231 LOG_MAIN, "Connection from %s refused: too many connections",
232 whofrom);
233 goto ERROR_RETURN;
234 }
235
236/* If a load limit above which only reserved hosts are acceptable is defined,
237get the load average here, and if there are in fact no reserved hosts, do
238the test right away (saves a fork). If there are hosts, do the check in the
239subprocess because it might take time. */
240
241if (smtp_load_reserve >= 0)
242 {
243 load_average = os_getloadavg();
244 if (smtp_reserve_hosts == NULL && load_average > smtp_load_reserve)
245 {
246 DEBUG(D_any) debug_printf("rejecting SMTP connection: load average = %.2f\n",
247 (double)load_average/1000.0);
248 smtp_printf("421 Too much load; please try again later.\r\n");
249 log_write(L_connection_reject,
250 LOG_MAIN, "Connection from %s refused: load average = %.2f",
251 whofrom, (double)load_average/1000.0);
252 goto ERROR_RETURN;
253 }
254 }
255
256/* Check that one specific host (strictly, IP address) is not hogging
257resources. This is done here to prevent a denial of service attack by someone
258forcing you to fork lots of times before denying service. The value of
259smtp_accept_max_per_host is a string which is expanded. This makes it possible
260to provide host-specific limits according to $sender_host address, but because
261this is in the daemon mainline, only fast expansions (such as inline address
262checks) should be used. The documentation is full of warnings. */
263
264if (smtp_accept_max_per_host != NULL)
265 {
266 uschar *expanded = expand_string(smtp_accept_max_per_host);
267 if (expanded == NULL)
268 {
269 if (!expand_string_forcedfail)
270 log_write(0, LOG_MAIN|LOG_PANIC, "expansion of smtp_accept_max_per_host "
271 "failed for %s: %s", whofrom, expand_string_message);
272 }
273 /* For speed, interpret a decimal number inline here */
274 else
275 {
276 uschar *s = expanded;
277 while (isdigit(*s))
278 max_for_this_host = max_for_this_host * 10 + *s++ - '0';
279 if (*s != 0)
280 log_write(0, LOG_MAIN|LOG_PANIC, "expansion of smtp_accept_max_per_host "
281 "for %s contains non-digit: %s", whofrom, expanded);
282 }
283 }
284
285/* If we have fewer connections than max_for_this_host, we can skip the tedious
286per host_address checks. Note that at this stage smtp_accept_count contains the
287count of *other* connections, not including this one. */
288
289if ((max_for_this_host > 0) &&
290 (smtp_accept_count >= max_for_this_host))
291 {
292 int i;
293 int host_accept_count = 0;
294 int other_host_count = 0; /* keep a count of non matches to optimise */
295
296 for (i = 0; i < smtp_accept_max; ++i)
297 {
298 if (smtp_slots[i].host_address != NULL)
299 {
300 if (Ustrcmp(sender_host_address, smtp_slots[i].host_address) == 0)
301 host_accept_count++;
302 else
303 other_host_count++;
304
305 /* Testing all these strings is expensive - see if we can drop out
306 early, either by hitting the target, or finding there are not enough
307 connections left to make the target. */
308
309 if ((host_accept_count >= max_for_this_host) ||
310 ((smtp_accept_count - other_host_count) < max_for_this_host))
311 break;
312 }
313 }
314
315 if (host_accept_count >= max_for_this_host)
316 {
317 DEBUG(D_any) debug_printf("rejecting SMTP connection: too many from this "
318 "IP address: count=%d max=%d\n",
319 host_accept_count, max_for_this_host);
320 smtp_printf("421 Too many concurrent SMTP connections "
321 "from this IP address; please try again later.\r\n");
322 log_write(L_connection_reject,
323 LOG_MAIN, "Connection from %s refused: too many connections "
324 "from that IP address", whofrom);
325 goto ERROR_RETURN;
326 }
327 }
328
329/* OK, the connection count checks have been passed. Before we can fork the
330accepting process, we must first log the connection if requested. This logging
331used to happen in the subprocess, but doing that means that the value of
332smtp_accept_count can be out of step by the time it is logged. So we have to do
333the logging here and accept the performance cost. Note that smtp_accept_count
334hasn't yet been incremented to take account of this connection.
335
336In order to minimize the cost (because this is going to happen for every
337connection), do a preliminary selector test here. This saves ploughing through
338the generalized logging code each time when the selector is false. If the
339selector is set, check whether the host is on the list for logging. If not,
340arrange to unset the selector in the subprocess. */
341
342if ((log_write_selector & L_smtp_connection) != 0)
343 {
344 uschar *list = hosts_connection_nolog;
345 if (list != NULL && verify_check_host(&list) == OK)
346 use_log_write_selector &= ~L_smtp_connection;
347 else
348 log_write(L_smtp_connection, LOG_MAIN, "SMTP connection from %s "
349 "(TCP/IP connection count = %d)", whofrom, smtp_accept_count + 1);
350 }
351
352/* Now we can fork the accepting process; do a lookup tidy, just in case any
353expansion above did a lookup. */
354
355search_tidyup();
356pid = fork();
357
358/* Handle the child process */
359
360if (pid == 0)
361 {
362 int i;
363 int queue_only_reason = 0;
364 int old_pool = store_pool;
3d235903 365 int save_debug_selector = debug_selector;
059ec3d9
PH
366 BOOL local_queue_only;
367 #ifdef SA_NOCLDWAIT
368 struct sigaction act;
369 #endif
370
371 /* May have been modified for the subprocess */
372
373 log_write_selector = use_log_write_selector;
374
375 /* Get the local interface address into permanent store */
376
377 store_pool = POOL_PERM;
378 interface_address = string_copy(interface_address);
379 store_pool = old_pool;
380
381 /* Check for a tls-on-connect port */
382
383 if (host_is_tls_on_connect_port(interface_port)) tls_on_connect = TRUE;
384
385 /* Expand smtp_active_hostname if required. We do not do this any earlier,
386 because it may depend on the local interface address (indeed, that is most
387 likely what it depends on.) */
388
389 smtp_active_hostname = primary_hostname;
390 if (raw_active_hostname != NULL)
391 {
392 uschar *nah = expand_string(raw_active_hostname);
393 if (nah == NULL)
394 {
395 if (!expand_string_forcedfail)
396 {
397 log_write(0, LOG_MAIN|LOG_PANIC, "failed to expand \"%s\" "
398 "(smtp_active_hostname): %s", raw_active_hostname,
399 expand_string_message);
400 smtp_printf("421 Local configuration error; "
401 "please try again later.\r\n");
402 mac_smtp_fflush();
403 search_tidyup();
404 _exit(EXIT_FAILURE);
405 }
406 }
407 else if (nah[0] != 0) smtp_active_hostname = nah;
408 }
409
410 /* Initialize the queueing flags */
411
412 queue_check_only();
413 local_queue_only = queue_only;
414
415 /* Close the listening sockets, and set the SIGCHLD handler to SIG_IGN.
416 We also attempt to set things up so that children are automatically reaped,
417 but just in case this isn't available, there's a paranoid waitpid() in the
418 loop too (except for systems where we are sure it isn't needed). See the more
419 extensive comment before the reception loop in exim.c for a fuller
420 explanation of this logic. */
421
422 for (i = 0; i < listen_socket_count; i++) close(listen_sockets[i]);
423
424 #ifdef SA_NOCLDWAIT
425 act.sa_handler = SIG_IGN;
426 sigemptyset(&(act.sa_mask));
427 act.sa_flags = SA_NOCLDWAIT;
428 sigaction(SIGCHLD, &act, NULL);
429 #else
430 signal(SIGCHLD, SIG_IGN);
431 #endif
432
433 /* Attempt to get an id from the sending machine via the RFC 1413
434 protocol. We do this in the sub-process in order not to hold up the
435 main process if there is any delay. Then set up the fullhost information
3d235903
PH
436 in case there is no HELO/EHLO.
437
438 If debugging is enabled only for the daemon, we must turn if off while
439 finding the id, but turn it on again afterwards so that information about the
440 incoming connection is output. */
441
442 if (debug_daemon) debug_selector = 0;
059ec3d9
PH
443 verify_get_ident(IDENT_PORT);
444 host_build_sender_fullhost();
3d235903 445 debug_selector = save_debug_selector;
059ec3d9
PH
446
447 DEBUG(D_any)
448 debug_printf("Process %d is handling incoming connection from %s\n",
449 (int)getpid(), sender_fullhost);
450
3d235903
PH
451 /* Now disable debugging permanently if it's required only for the daemon
452 process. */
453
454 if (debug_daemon) debug_selector = 0;
455
059ec3d9
PH
456 /* If there are too many child processes for immediate delivery,
457 set the local_queue_only flag, which is initialized from the
458 configured value and may therefore already be TRUE. Leave logging
459 till later so it will have a message id attached. */
460
461 if (smtp_accept_queue > 0 && smtp_accept_count >= smtp_accept_queue)
462 {
463 local_queue_only = TRUE;
464 queue_only_reason = 1;
465 }
466
467 /* Handle the start of the SMTP session, then loop, accepting incoming
468 messages from the SMTP connection. The end will come at the QUIT command,
469 when smtp_setup_msg() returns 0. A break in the connection causes the
470 process to die (see accept.c). */
471
472 if (!smtp_start_session())
473 {
474 mac_smtp_fflush();
475 search_tidyup();
476 _exit(EXIT_SUCCESS);
477 }
478
479 for (;;)
480 {
481 int rc;
482 message_id[0] = 0; /* Clear out any previous message_id */
483 reset_point = store_get(0); /* Save current store high water point */
484
485 DEBUG(D_any)
486 debug_printf("Process %d is ready for new message\n", (int)getpid());
487
488 /* Smtp_setup_msg() returns 0 on QUIT or if the call is from an
489 unacceptable host or if an ACL "drop" command was triggered, -1 on
490 connection lost, and +1 on validly reaching DATA. Receive_msg() almost
491 always returns TRUE when smtp_input is true; just retry if no message was
492 accepted (can happen for invalid message parameters). However, it can yield
493 FALSE if the connection was forcibly dropped by the DATA ACL. */
494
495 if ((rc = smtp_setup_msg()) > 0)
496 {
497 BOOL ok = receive_msg(FALSE);
498 search_tidyup(); /* Close cached databases */
499 if (!ok) /* Connection was dropped */
500 {
501 mac_smtp_fflush();
502 _exit(EXIT_SUCCESS);
503 }
504 if (message_id[0] == 0) continue; /* No message was accepted */
505 }
506 else
507 {
508 mac_smtp_fflush();
509 search_tidyup();
510 _exit((rc == 0)? EXIT_SUCCESS : EXIT_FAILURE);
511 }
512
513 /* Show the recipients when debugging */
514
515 DEBUG(D_receive)
516 {
517 int i;
518 if (sender_address != NULL)
519 debug_printf("Sender: %s\n", sender_address);
520 if (recipients_list != NULL)
521 {
522 debug_printf("Recipients:\n");
523 for (i = 0; i < recipients_count; i++)
524 debug_printf(" %s\n", recipients_list[i].address);
525 }
526 }
527
528 /* A message has been accepted. Clean up any previous delivery processes
529 that have completed and are defunct, on systems where they don't go away
530 by themselves (see comments when setting SIG_IGN above). On such systems
531 (if any) these delivery processes hang around after termination until
532 the next message is received. */
533
534 #ifndef SIG_IGN_WORKS
535 while (waitpid(-1, NULL, WNOHANG) > 0);
536 #endif
537
538 /* Reclaim up the store used in accepting this message */
539
540 store_reset(reset_point);
541
542 /* If queue_only is set or if there are too many incoming connections in
543 existence, local_queue_only will be TRUE. If it is not, check whether we
544 have received too many messages in this session for immediate delivery. If
545 not, and queue_only_load is set, check that the load average is below it.
546 Note that, once set, local_queue_only remains set for any subsequent
547 messages on the same SMTP connection. This is a deliberate choice; even
548 though the load average may fall, it doesn't seem right to deliver later
549 messages on the same call when not delivering earlier ones. */
550
551 if (!local_queue_only)
552 {
553 if (smtp_accept_queue_per_connection > 0 &&
554 receive_messagecount > smtp_accept_queue_per_connection)
555 {
556 local_queue_only = TRUE;
557 queue_only_reason = 2;
558 }
559 else if (queue_only_load >= 0)
560 {
561 local_queue_only = (load_average = os_getloadavg()) > queue_only_load;
562 if (local_queue_only) queue_only_reason = 3;
563 }
564 }
565
566 /* Log the queueing here, when it will get a message id attached, but
567 not if queue_only is set (case 0). */
568
569 if (local_queue_only) switch(queue_only_reason)
570 {
571 case 1:
572 log_write(L_delay_delivery,
573 LOG_MAIN, "no immediate delivery: too many connections "
574 "(%d, max %d)", smtp_accept_count, smtp_accept_queue);
575 break;
576
577 case 2:
578 log_write(L_delay_delivery,
579 LOG_MAIN, "no immediate delivery: more than %d messages "
580 "received in one connection", smtp_accept_queue_per_connection);
581 break;
582
583 case 3:
584 log_write(L_delay_delivery,
585 LOG_MAIN, "no immediate delivery: load average %.2f",
586 (double)load_average/1000.0);
587 break;
588 }
589
590 /* If a delivery attempt is required, spin off a new process to handle it.
591 If we are not root, we have to re-exec exim unless deliveries are being
592 done unprivileged. */
593
594 else if (!queue_only_policy && !deliver_freeze)
595 {
596 pid_t dpid;
597
598 /* Before forking, ensure that the C output buffer is flushed. Otherwise
599 anything that it in it will get duplicated, leading to duplicate copies
600 of the pending output. */
601
602 mac_smtp_fflush();
603
604 if ((dpid = fork()) == 0)
605 {
606 fclose(smtp_in);
607 fclose(smtp_out);
608
609 /* Don't ever molest the parent's SSL connection, but do clean up
610 the data structures if necessary. */
611
612 #ifdef SUPPORT_TLS
613 tls_close(FALSE);
614 #endif
615
616 /* Reset SIGHUP and SIGCHLD in the child in both cases. */
617
618 signal(SIGHUP, SIG_DFL);
619 signal(SIGCHLD, SIG_DFL);
620
621 if (geteuid() != root_uid && !deliver_drop_privilege)
622 {
623 signal(SIGALRM, SIG_DFL);
624 (void)child_exec_exim(CEE_EXEC_PANIC, FALSE, NULL, FALSE, 2, US"-Mc",
625 message_id);
626 /* Control does not return here. */
627 }
628
629 /* No need to re-exec; SIGALRM remains set to the default handler */
630
631 (void)deliver_message(message_id, FALSE, FALSE);
632 search_tidyup();
633 _exit(EXIT_SUCCESS);
634 }
635
636 if (dpid > 0)
637 {
638 DEBUG(D_any) debug_printf("forked delivery process %d\n", (int)dpid);
639 }
640 else
641 {
642 log_write(0, LOG_MAIN|LOG_PANIC, "daemon: delivery process fork "
643 "failed: %s", strerror(errno));
644 }
645 }
646 }
647 }
648
649
650/* Carrying on in the parent daemon process... Can't do much if the fork
651failed. Otherwise, keep count of the number of accepting processes and
652remember the pid for ticking off when the child completes. */
653
654if (pid < 0)
655 {
656 never_error(US"daemon: accept process fork failed", US"Fork failed", errno);
657 }
658else
659 {
660 int i;
661 for (i = 0; i < smtp_accept_max; ++i)
662 {
663 if (smtp_slots[i].pid <= 0)
664 {
665 smtp_slots[i].pid = pid;
666 if (smtp_accept_max_per_host != NULL)
667 smtp_slots[i].host_address = string_copy_malloc(sender_host_address);
668 smtp_accept_count++;
669 break;
670 }
671 }
672 DEBUG(D_any) debug_printf("%d SMTP accept process%s running\n",
673 smtp_accept_count, (smtp_accept_count == 1)? "" : "es");
674 }
675
676/* Get here via goto in error cases */
677
678ERROR_RETURN:
679
680/* Close the streams associated with the socket which will also close the
681socket fds in this process. We can't do anything if fclose() fails, but
682logging brings it to someone's attention. However, "connection reset by peer"
683isn't really a problem, so skip that one. If the streams don't exist, something
684went wrong while setting things up. Make sure the socket descriptors are
685closed, in order to drop the connection. */
686
687if (smtp_out != NULL)
688 {
689 if (fclose(smtp_out) != 0 && errno != ECONNRESET)
690 log_write(0, LOG_MAIN|LOG_PANIC, "daemon: fclose(smtp_out) failed: %s",
691 strerror(errno));
692 smtp_out = NULL;
693 }
694else close(accept_socket);
695
696if (smtp_in != NULL)
697 {
698 if (fclose(smtp_in) != 0 && errno != ECONNRESET)
699 log_write(0, LOG_MAIN|LOG_PANIC, "daemon: fclose(smtp_in) failed: %s",
700 strerror(errno));
701 smtp_in = NULL;
702 }
703else close(dup_accept_socket);
704
705/* Release any store used in this process, including the store used for holding
706the incoming host address and an expanded active_hostname. */
707
708store_reset(reset_point);
709sender_host_address = NULL;
710}
711
712
713
714
715/*************************************************
716* Check wildcard listen special cases *
717*************************************************/
718
719/* This function is used when binding and listening on lists of addresses and
720ports. It tests for special cases of wildcard listening, when IPv4 and IPv6
721sockets may interact in different ways in different operating systems. It is
722passed an error number, the list of listening addresses, and the current
723address. Two checks are available: for a previous wildcard IPv6 address, or for
724a following wildcard IPv4 address, in both cases on the same port.
725
726In practice, pairs of wildcard addresses should be adjacent in the address list
727because they are sorted that way below.
728
729Arguments:
730 eno the error number
731 addresses the list of addresses
732 ipa the current IP address
733 back if TRUE, check for previous wildcard IPv6 address
734 if FALSE, check for a following wildcard IPv4 address
735
736Returns: TRUE or FALSE
737*/
738
739static BOOL
740check_special_case(int eno, ip_address_item *addresses, ip_address_item *ipa,
741 BOOL back)
742{
743ip_address_item *ipa2;
744
745/* For the "back" case, if the failure was "address in use" for a wildcard IPv4
746address, seek a previous IPv6 wildcard address on the same port. As it is
747previous, it must have been successfully bound and be listening. Flag it as a
748"6 including 4" listener. */
749
750if (back)
751 {
752 if (eno != EADDRINUSE || ipa->address[0] != 0) return FALSE;
753 for (ipa2 = addresses; ipa2 != ipa; ipa2 = ipa2->next)
754 {
755 if (ipa2->address[1] == 0 && ipa2->port == ipa->port)
756 {
757 ipa2->v6_include_v4 = TRUE;
758 return TRUE;
759 }
760 }
761 }
762
763/* For the "forward" case, if the current address is a wildcard IPv6 address,
764we seek a following wildcard IPv4 address on the same port. */
765
766else
767 {
768 if (ipa->address[0] != ':' || ipa->address[1] != 0) return FALSE;
769 for (ipa2 = ipa->next; ipa2 != NULL; ipa2 = ipa2->next)
770 if (ipa2->address[0] == 0 && ipa->port == ipa2->port) return TRUE;
771 }
772
773return FALSE;
774}
775
776
777
778
bb6e88ff
PH
779/*************************************************
780* Handle terminating subprocesses *
781*************************************************/
782
783/* Handle the termination of child processes. Theoretically, this need be done
784only when sigchld_seen is TRUE, but rumour has it that some systems lose
785SIGCHLD signals at busy times, so to be on the safe side, this function is
786called each time round. It shouldn't be too expensive.
787
788Arguments: none
789Returns: nothing
790*/
791
792static void
793handle_ending_processes(void)
794{
795int status;
796pid_t pid;
797
798while ((pid = waitpid(-1, &status, WNOHANG)) > 0)
799 {
800 int i;
801 DEBUG(D_any) debug_printf("child %d ended: status=0x%x\n", (int)pid,
802 status);
803
804 /* If it's a listening daemon for which we are keeping track of individual
805 subprocesses, deal with an accepting process that has terminated. */
806
807 if (smtp_slots != NULL)
808 {
809 for (i = 0; i < smtp_accept_max; i++)
810 {
811 if (smtp_slots[i].pid == pid)
812 {
813 if (smtp_slots[i].host_address != NULL)
814 store_free(smtp_slots[i].host_address);
815 smtp_slots[i] = empty_smtp_slot;
816 if (--smtp_accept_count < 0) smtp_accept_count = 0;
817 DEBUG(D_any) debug_printf("%d SMTP accept process%s now running\n",
818 smtp_accept_count, (smtp_accept_count == 1)? "" : "es");
819 break;
820 }
821 }
822 if (i < smtp_accept_max) continue; /* Found an accepting process */
823 }
824
825 /* If it wasn't an accepting process, see if it was a queue-runner
826 process that we are tracking. */
827
828 if (queue_pid_slots != NULL)
829 {
830 for (i = 0; i < queue_run_max; i++)
831 {
832 if (queue_pid_slots[i] == pid)
833 {
834 queue_pid_slots[i] = 0;
835 if (--queue_run_count < 0) queue_run_count = 0;
836 DEBUG(D_any) debug_printf("%d queue-runner process%s now running\n",
837 queue_run_count, (queue_run_count == 1)? "" : "es");
838 break;
839 }
840 }
841 }
842 }
843}
844
059ec3d9
PH
845
846
847/*************************************************
848* Exim Daemon Mainline *
849*************************************************/
850
851/* The daemon can do two jobs, either of which is optional:
852
853(1) Listens for incoming SMTP calls and spawns off a sub-process to handle
854each one. This is requested by the -bd option, with -oX specifying the SMTP
855port on which to listen (for testing).
856
857(2) Spawns a queue-running process every so often. This is controlled by the
858-q option with a an interval time. (If no time is given, a single queue run
859is done from the main function, and control doesn't get here.)
860
861Root privilege is required in order to attach to port 25. Some systems require
862it when calling socket() rather than bind(). To cope with all cases, we run as
863root for both socket() and bind(). Some systems also require root in order to
864write to the pid file directory. This function must therefore be called as root
865if it is to work properly in all circumstances. Once the socket is bound and
866the pid file written, root privilege is given up if there is an exim uid.
867
868There are no arguments to this function, and it never returns. */
869
870void
871daemon_go(void)
872{
873int *listen_sockets = NULL;
874int listen_socket_count = 0;
875ip_address_item *addresses = NULL;
876
877/* If any debugging options are set, turn on the D_pid bit so that all
878debugging lines get the pid added. */
879
880DEBUG(D_any|D_v) debug_selector |= D_pid;
881
882
883/* Do the preparation for setting up a listener on one or more interfaces, and
884possible on various ports. This is controlled by the combination of
885local_interfaces (which can set IP addresses and ports) and daemon_smtp_port
886(which is a list of default ports to use for those items in local_interfaces
887that do not specify a port). The -oX command line option can be used to
888override one or both of these options.
889
890If local_interfaces is not set, the default is to listen on all interfaces.
891When it is set, it can include "all IPvx interfaces" as an item. This is useful
892when different ports are in use.
893
894It turns out that listening on all interfaces is messy in an IPv6 world,
895because several different implementation approaches have been taken. This code
896is now supposed to work with all of them. The point of difference is whether an
897IPv6 socket that is listening on all interfaces will receive incoming IPv4
898calls or not. We also have to cope with the case when IPv6 libraries exist, but
899there is no IPv6 support in the kernel.
900
901. On Solaris, an IPv6 socket will accept IPv4 calls, and give them as mapped
902 addresses. However, if an IPv4 socket is also listening on all interfaces,
903 calls are directed to the appropriate socket.
904
905. On (some versions of) Linux, an IPv6 socket will accept IPv4 calls, and
906 give them as mapped addresses, but an attempt also to listen on an IPv4
907 socket on all interfaces causes an error.
908
909. On OpenBSD, an IPv6 socket will not accept IPv4 calls. You have to set up
910 two sockets if you want to accept both kinds of call.
911
912. FreeBSD is like OpenBSD, but it has the IPV6_V6ONLY socket option, which
913 can be turned off, to make it behave like the versions of Linux described
914 above.
915
916. I heard a report that the USAGI IPv6 stack for Linux has implemented
917 IPV6_V6ONLY.
918
919So, what we do when IPv6 is supported is as follows:
920
921 (1) After it is set up, the list of interfaces is scanned for wildcard
922 addresses. If an IPv6 and an IPv4 wildcard are both found for the same
923 port, the list is re-arranged so that they are together, with the IPv6
924 wildcard first.
925
926 (2) If the creation of a wildcard IPv6 socket fails, we just log the error and
927 carry on if an IPv4 wildcard socket for the same port follows later in the
928 list. This allows Exim to carry on in the case when the kernel has no IPv6
929 support.
930
931 (3) Having created an IPv6 wildcard socket, we try to set IPV6_V6ONLY if that
932 option is defined. However, if setting fails, carry on regardless (but log
933 the incident).
934
935 (4) If binding or listening on an IPv6 wildcard socket fails, it is a serious
936 error.
937
938 (5) If binding or listening on an IPv4 wildcard socket fails with the error
939 EADDRINUSE, and a previous interface was an IPv6 wildcard for the same
940 port (which must have succeeded or we wouldn't have got this far), we
941 assume we are in the situation where just a single socket is permitted,
942 and ignore the error.
943
944Phew!
945
946The preparation code decodes options and sets up the relevant data. We do this
947first, so that we can return non-zero if there are any syntax errors, and also
948write to stderr. */
949
950if (daemon_listen)
951 {
952 int *default_smtp_port;
953 int sep;
954 int pct = 0;
955 uschar *s;
956 uschar *list;
957 uschar *local_iface_source = US"local_interfaces";
958 ip_address_item *ipa;
959 ip_address_item **pipa;
960
961 /* If any option requiring a load average to be available during the
962 reception of a message is set, call os_getloadavg() while we are root
963 for those OS for which this is necessary the first time it is called (in
964 order to perform an "open" on the kernel memory file). */
965
966 #ifdef LOAD_AVG_NEEDS_ROOT
d95f9fdb
PH
967 if (queue_only_load >= 0 || smtp_load_reserve >= 0 ||
968 (deliver_queue_load_max >= 0 && deliver_drop_privilege))
969 (void)os_getloadavg();
059ec3d9
PH
970 #endif
971
972 /* If -oX was used, disable the writing of a pid file unless -oP was
973 explicitly used to force it. Then scan the string given to -oX. Any items
974 that contain neither a dot nor a colon are used to override daemon_smtp_port.
975 Any other items are used to override local_interfaces. */
976
977 if (override_local_interfaces != NULL)
978 {
979 uschar *new_smtp_port = NULL;
980 uschar *new_local_interfaces = NULL;
981 int portsize = 0;
982 int portptr = 0;
983 int ifacesize = 0;
984 int ifaceptr = 0;
985
986 if (override_pid_file_path == NULL) write_pid = FALSE;
987
988 list = override_local_interfaces;
989 sep = 0;
990 while ((s = string_nextinlist(&list,&sep,big_buffer,big_buffer_size))
991 != NULL)
992 {
993 uschar joinstr[4];
994 uschar **ptr;
995 int *sizeptr;
996 int *ptrptr;
997
998 if (Ustrpbrk(s, ".:") == NULL)
999 {
1000 ptr = &new_smtp_port;
1001 sizeptr = &portsize;
1002 ptrptr = &portptr;
1003 }
1004 else
1005 {
1006 ptr = &new_local_interfaces;
1007 sizeptr = &ifacesize;
1008 ptrptr = &ifaceptr;
1009 }
1010
1011 if (*ptr == NULL)
1012 {
1013 joinstr[0] = sep;
1014 joinstr[1] = ' ';
1015 *ptr = string_cat(*ptr, sizeptr, ptrptr, US"<", 1);
1016 }
1017
1018 *ptr = string_cat(*ptr, sizeptr, ptrptr, joinstr, 2);
1019 *ptr = string_cat(*ptr, sizeptr, ptrptr, s, Ustrlen(s));
1020 }
1021
1022 if (new_smtp_port != NULL)
1023 {
1024 new_smtp_port[portptr] = 0;
1025 daemon_smtp_port = new_smtp_port;
1026 DEBUG(D_any) debug_printf("daemon_smtp_port overridden by -oX:\n %s\n",
1027 daemon_smtp_port);
1028 }
1029
1030 if (new_local_interfaces != NULL)
1031 {
1032 new_local_interfaces[ifaceptr] = 0;
1033 local_interfaces = new_local_interfaces;
1034 local_iface_source = US"-oX data";
1035 DEBUG(D_any) debug_printf("local_interfaces overridden by -oX:\n %s\n",
1036 local_interfaces);
1037 }
1038 }
1039
1040 /* Create a list of default SMTP ports, to be used if local_interfaces
1041 contains entries without explict ports. First count the number of ports, then
1042 build a translated list in a vector. */
1043
1044 list = daemon_smtp_port;
1045 sep = 0;
1046 while ((s = string_nextinlist(&list,&sep,big_buffer,big_buffer_size)) != NULL)
1047 pct++;
1048 default_smtp_port = store_get((pct+1) * sizeof(int));
1049 list = daemon_smtp_port;
1050 sep = 0;
1051 for (pct = 0;
1052 (s = string_nextinlist(&list,&sep,big_buffer,big_buffer_size)) != NULL;
1053 pct++)
1054 {
1055 if (isdigit(*s))
1056 {
1057 uschar *end;
1058 default_smtp_port[pct] = Ustrtol(s, &end, 0);
1059 if (end != s + Ustrlen(s))
1060 log_write(0, LOG_PANIC_DIE|LOG_CONFIG, "invalid SMTP port: %s", s);
1061 }
1062 else
1063 {
1064 struct servent *smtp_service = getservbyname(CS s, "tcp");
1065 if (smtp_service == NULL)
1066 log_write(0, LOG_PANIC_DIE|LOG_CONFIG, "TCP port \"%s\" not found", s);
1067 default_smtp_port[pct] = ntohs(smtp_service->s_port);
1068 }
1069 }
1070 default_smtp_port[pct] = 0;
1071
1072 /* Create the list of local interfaces, possibly with ports included. This
1073 list may contain references to 0.0.0.0 and ::0 as wildcards. These special
1074 values are converted below. */
1075
1076 addresses = host_build_ifacelist(local_interfaces, local_iface_source);
1077
1078 /* In the list of IP addresses, convert 0.0.0.0 into an empty string, and ::0
1079 into the string ":". We use these to recognize wildcards in IPv4 and IPv6. In
1080 fact, many IP stacks recognize 0.0.0.0 and ::0 and handle them as wildcards
1081 anyway, but we need to know which are the wildcard addresses, and the shorter
1082 strings are neater.
1083
1084 In the same scan, fill in missing port numbers from the default list. When
1085 there is more than one item in the list, extra items are created. */
1086
1087 for (ipa = addresses; ipa != NULL; ipa = ipa->next)
1088 {
1089 int i;
1090
1091 if (Ustrcmp(ipa->address, "0.0.0.0") == 0) ipa->address[0] = 0;
1092 else if (Ustrcmp(ipa->address, "::0") == 0)
1093 {
1094 ipa->address[0] = ':';
1095 ipa->address[1] = 0;
1096 }
1097
1098 if (ipa->port > 0) continue;
1099
1100 if (daemon_smtp_port[0] <= 0)
1101 log_write(0, LOG_MAIN|LOG_PANIC_DIE, "no port specified for interface "
1102 "%s and daemon_smtp_port is unset; cannot start daemon",
1103 (ipa->address[0] == 0)? US"\"all IPv4\"" :
1104 (ipa->address[1] == 0)? US"\"all IPv6\"" : ipa->address);
1105 ipa->port = default_smtp_port[0];
1106 for (i = 1; default_smtp_port[i] > 0; i++)
1107 {
1108 ip_address_item *new = store_get(sizeof(ip_address_item));
1109 memcpy(new->address, ipa->address, Ustrlen(ipa->address) + 1);
1110 new->port = default_smtp_port[i];
1111 new->next = ipa->next;
1112 ipa->next = new;
1113 ipa = new;
1114 }
1115 }
1116
1117 /* Scan the list of addresses for wildcards. If we find an IPv4 and an IPv6
1118 wildcard for the same port, ensure that (a) they are together and (b) the
1119 IPv6 address comes first. This makes handling the messy features easier, and
1120 also simplifies the construction of the "daemon started" log line. */
1121
1122 pipa = &addresses;
1123 for (ipa = addresses; ipa != NULL; pipa = &(ipa->next), ipa = ipa->next)
1124 {
1125 ip_address_item *ipa2;
1126
1127 /* Handle an IPv4 wildcard */
1128
1129 if (ipa->address[0] == 0)
1130 {
1131 for (ipa2 = ipa; ipa2->next != NULL; ipa2 = ipa2->next)
1132 {
1133 ip_address_item *ipa3 = ipa2->next;
1134 if (ipa3->address[0] == ':' &&
1135 ipa3->address[1] == 0 &&
1136 ipa3->port == ipa->port)
1137 {
1138 ipa2->next = ipa3->next;
1139 ipa3->next = ipa;
1140 *pipa = ipa3;
1141 break;
1142 }
1143 }
1144 }
1145
1146 /* Handle an IPv6 wildcard. */
1147
1148 else if (ipa->address[0] == ':' && ipa->address[1] == 0)
1149 {
1150 for (ipa2 = ipa; ipa2->next != NULL; ipa2 = ipa2->next)
1151 {
1152 ip_address_item *ipa3 = ipa2->next;
1153 if (ipa3->address[0] == 0 && ipa3->port == ipa->port)
1154 {
1155 ipa2->next = ipa3->next;
1156 ipa3->next = ipa->next;
1157 ipa->next = ipa3;
1158 ipa = ipa3;
1159 break;
1160 }
1161 }
1162 }
1163 }
1164
1165 /* Get a vector to remember all the sockets in */
1166
1167 for (ipa = addresses; ipa != NULL; ipa = ipa->next)
1168 listen_socket_count++;
1169 listen_sockets = store_get(sizeof(int *) * listen_socket_count);
1170
1171 /* Do a sanity check on the max connects value just to save us from getting
1172 a huge amount of store. */
1173
1174 if (smtp_accept_max > 4095) smtp_accept_max = 4096;
1175
1176 /* There's no point setting smtp_accept_queue unless it is less than the max
1177 connects limit. The configuration reader ensures that the max is set if the
1178 queue-only option is set. */
1179
1180 if (smtp_accept_queue > smtp_accept_max) smtp_accept_queue = 0;
1181
1182 /* Get somewhere to keep the list of SMTP accepting pids if we are keeping
1183 track of them for total number and queue/host limits. */
1184
1185 if (smtp_accept_max > 0)
1186 {
1187 int i;
1188 smtp_slots = store_get(smtp_accept_max * sizeof(smtp_slot));
1189 for (i = 0; i < smtp_accept_max; i++) smtp_slots[i] = empty_smtp_slot;
1190 }
1191 }
1192
76a2d7ba
PH
1193/* The variable background_daemon is always false when debugging, but
1194can also be forced false in order to keep a non-debugging daemon in the
1195foreground. If background_daemon is true, close all open file descriptors that
1196we know about, but then re-open stdin, stdout, and stderr to /dev/null.
1197
1198This is protection against any called functions (in libraries, or in
1199Perl, or whatever) that think they can write to stderr (or stdout). Before this
1200was added, it was quite likely that an SMTP connection would use one of these
1201file descriptors, in which case writing random stuff to it caused chaos.
1202
1203Then disconnect from the controlling terminal, Most modern Unixes seem to have
1204setsid() for getting rid of the controlling terminal. For any OS that doesn't,
1205setsid() can be #defined as a no-op, or as something else. */
059ec3d9
PH
1206
1207if (background_daemon)
1208 {
76a2d7ba
PH
1209 log_close_all(); /* Just in case anything was logged earlier */
1210 search_tidyup(); /* Just in case any were used in reading the config. */
1211 close(0); /* Get rid of stdin/stdout/stderr */
059ec3d9
PH
1212 close(1);
1213 close(2);
76a2d7ba 1214 exim_nullstd(); /* Connect stdin/stdout/stderr to /dev/null */
059ec3d9
PH
1215 log_stderr = NULL; /* So no attempt to copy paniclog output */
1216
1217 /* If the parent process of this one has pid == 1, we are re-initializing the
76a2d7ba
PH
1218 daemon as the result of a SIGHUP. In this case, there is no need to do
1219 anything, because the controlling terminal has long gone. Otherwise, fork, in
1220 case current process is a process group leader (see 'man setsid' for an
1221 explanation) before calling setsid(). */
059ec3d9
PH
1222
1223 if (getppid() != 1)
1224 {
1225 pid_t pid = fork();
1226 if (pid < 0) log_write(0, LOG_MAIN|LOG_PANIC_DIE,
1227 "fork() failed when starting daemon: %s", strerror(errno));
1228 if (pid > 0) exit(EXIT_SUCCESS); /* in parent process, just exit */
1229 (void)setsid(); /* release controlling terminal */
1230 }
1231 }
1232
1233/* We are now in the disconnected, daemon process (unless debugging). Set up
1234the listening sockets if required. */
1235
1236if (daemon_listen)
1237 {
1238 int sk;
1239 int on = 1;
1240 ip_address_item *ipa;
1241
1242 /* For each IP address, create a socket, bind it to the appropriate port, and
1243 start listening. See comments above about IPv6 sockets that may or may not
1244 accept IPv4 calls when listening on all interfaces. We also have to cope with
1245 the case of a system with IPv6 libraries, but no IPv6 support in the kernel.
1246 listening, provided a wildcard IPv4 socket for the same port follows. */
1247
1248 for (ipa = addresses, sk = 0; sk < listen_socket_count; ipa = ipa->next, sk++)
1249 {
1250 BOOL wildcard;
1251 ip_address_item *ipa2;
1252 int retries = 9;
1253 int af;
1254
1255 if (Ustrchr(ipa->address, ':') != NULL)
1256 {
1257 af = AF_INET6;
1258 wildcard = ipa->address[1] == 0;
1259 }
1260 else
1261 {
1262 af = AF_INET;
1263 wildcard = ipa->address[0] == 0;
1264 }
1265
1266 listen_sockets[sk] = ip_socket(SOCK_STREAM, af);
1267 if (listen_sockets[sk] < 0)
1268 {
1269 if (check_special_case(0, addresses, ipa, FALSE))
1270 {
1271 log_write(0, LOG_MAIN, "Failed to create IPv6 socket for wildcard "
1272 "listening (%s): will use IPv4", strerror(errno));
1273 goto SKIP_SOCKET;
1274 }
1275 log_write(0, LOG_PANIC_DIE, "IPv%c socket creation failed: %s",
1276 (af == AF_INET6)? '6' : '4', strerror(errno));
1277 }
1278
1279 /* If this is an IPv6 wildcard socket, set IPV6_V6ONLY if that option is
1280 available. Just log failure (can get protocol not available, just like
1281 socket creation can). */
1282
1283 #ifdef IPV6_V6ONLY
1284 if (af == AF_INET6 && wildcard &&
1285 setsockopt(listen_sockets[sk], IPPROTO_IPV6, IPV6_V6ONLY, (char *)(&on),
1286 sizeof(on)) < 0)
1287 log_write(0, LOG_MAIN, "Setting IPV6_V6ONLY on daemon's IPv6 wildcard "
1288 "socket failed (%s): carrying on without it", strerror(errno));
1289 #endif /* IPV6_V6ONLY */
1290
1291 /* Set SO_REUSEADDR so that the daemon can be restarted while a connection
1292 is being handled. Without this, a connection will prevent reuse of the
1293 smtp port for listening. */
1294
1295 if (setsockopt(listen_sockets[sk], SOL_SOCKET, SO_REUSEADDR,
1296 (uschar *)(&on), sizeof(on)) < 0)
1297 log_write(0, LOG_MAIN|LOG_PANIC_DIE, "setting SO_REUSEADDR on socket "
1298 "failed when starting daemon: %s", strerror(errno));
1299
1300 /* Set TCP_NODELAY; Exim does its own buffering. There is a switch to
1301 disable this because it breaks some broken clients. */
1302
1303 if (tcp_nodelay) setsockopt(listen_sockets[sk], IPPROTO_TCP, TCP_NODELAY,
1304 (uschar *)(&on), sizeof(on));
1305
1306 /* Now bind the socket to the required port; if Exim is being restarted
1307 it may not always be possible to bind immediately, even with SO_REUSEADDR
1308 set, so try 10 times, waiting between each try. After 10 failures, we give
1309 up. In an IPv6 environment, if bind () fails with the error EADDRINUSE and
1310 we are doing wildcard IPv4 listening and there was a previous IPv6 wildcard
1311 address for the same port, ignore the error on the grounds that we must be
1312 in a system where the IPv6 socket accepts both kinds of call. This is
1313 necessary for (some release of) USAGI Linux; other IP stacks fail at the
1314 listen() stage instead. */
1315
1316 for(;;)
1317 {
1318 uschar *msg, *addr;
1319 if (ip_bind(listen_sockets[sk], af, ipa->address, ipa->port) >= 0) break;
1320 if (check_special_case(errno, addresses, ipa, TRUE))
1321 {
1322 DEBUG(D_any) debug_printf("wildcard IPv4 bind() failed after IPv6 "
1323 "listen() success; EADDRINUSE ignored\n");
1324 close(listen_sockets[sk]);
1325 goto SKIP_SOCKET;
1326 }
1327 msg = US strerror(errno);
1328 addr = wildcard? ((af == AF_INET6)? US"(any IPv6)" : US"(any IPv4)") :
1329 ipa->address;
1330 if (retries-- <= 0)
1331 log_write(0, LOG_MAIN|LOG_PANIC_DIE,
1332 "socket bind() to port %d for address %s failed: %s: "
1333 "daemon abandoned", ipa->port, addr, msg);
1334 log_write(0, LOG_MAIN, "socket bind() to port %d for address %s "
1335 "failed: %s: waiting before trying again", ipa->port, addr, msg);
1336 sleep(30);
1337 }
1338
1339 DEBUG(D_any)
1340 {
1341 if (wildcard)
1342 debug_printf("listening on all interfaces (IPv%c) port %d\n",
1343 (af == AF_INET6)? '6' : '4', ipa->port);
1344 else
1345 debug_printf("listening on %s port %d\n", ipa->address, ipa->port);
1346 }
1347
1348 /* Start listening on the bound socket, establishing the maximum backlog of
1349 connections that is allowed. On success, continue to the next address. */
1350
1351 if (listen(listen_sockets[sk], smtp_connect_backlog) >= 0) continue;
1352
1353 /* Listening has failed. In an IPv6 environment, as for bind(), if listen()
1354 fails with the error EADDRINUSE and we are doing IPv4 wildcard listening
1355 and there was a previous successful IPv6 wildcard listen on the same port,
1356 we want to ignore the error on the grounds that we must be in a system
1357 where the IPv6 socket accepts both kinds of call. */
1358
1359 if (!check_special_case(errno, addresses, ipa, TRUE))
1360 log_write(0, LOG_PANIC_DIE, "listen() failed on interface %s: %s",
1361 wildcard? ((af == AF_INET6)? US"(any IPv6)" : US"(any IPv4)") :
1362 ipa->address,
1363 strerror(errno));
1364
1365 DEBUG(D_any) debug_printf("wildcard IPv4 listen() failed after IPv6 "
1366 "listen() success; EADDRINUSE ignored\n");
1367 close(listen_sockets[sk]);
1368
1369 /* Come here if there has been a problem with the socket which we
1370 are going to ignore. We remove the address from the chain, and back up the
1371 counts. */
1372
1373 SKIP_SOCKET:
1374 sk--; /* Back up the count */
1375 listen_socket_count--; /* Reduce the total */
1376 if (ipa == addresses) addresses = ipa->next; else
1377 {
1378 for (ipa2 = addresses; ipa2->next != ipa; ipa2 = ipa2->next);
1379 ipa2->next = ipa->next;
1380 ipa = ipa2;
1381 }
1382 } /* End of bind/listen loop for each address */
1383 } /* End of setup for listening */
1384
1385
1386/* If we are not listening, we want to write a pid file only if -oP was
1387explicitly given. */
1388
1389else if (override_pid_file_path == NULL) write_pid = FALSE;
1390
1391/* Write the pid to a known file for assistance in identification, if required.
1392We do this before giving up root privilege, because on some systems it is
1393necessary to be root in order to write into the pid file directory. There's
1394nothing to stop multiple daemons running, as long as no more than one listens
1395on a given TCP/IP port on the same interface(s). However, in these
1396circumstances it gets far too complicated to mess with pid file names
1397automatically. Consequently, Exim 4 writes a pid file only
1398
1399 (a) When running in the test harness, or
1400 (b) When -bd is used and -oX is not used, or
1401 (c) When -oP is used to supply a path.
1402
1403The variable daemon_write_pid is used to control this. */
1404
1405if (running_in_test_harness || write_pid)
1406 {
1407 FILE *f;
1408
1409 if (override_pid_file_path != NULL)
1410 pid_file_path = override_pid_file_path;
1411
1412 if (pid_file_path[0] == 0)
1413 pid_file_path = string_sprintf("%s/exim-daemon.pid", spool_directory);
1414
1415 f = Ufopen(pid_file_path, "wb");
1416 if (f != NULL)
1417 {
1418 fprintf(f, "%d\n", (int)getpid());
1419 fchmod(fileno(f), 0644);
1420 fclose(f);
1421 DEBUG(D_any) debug_printf("pid written to %s\n", pid_file_path);
1422 }
1423 else
1424 {
1425 DEBUG(D_any)
1426 debug_printf("%s\n", string_open_failed(errno, "pid file %s",
1427 pid_file_path));
1428 }
1429 }
1430
1431/* Set up the handler for SIGHUP, which causes a restart of the daemon. */
1432
1433sighup_seen = FALSE;
1434signal(SIGHUP, sighup_handler);
1435
1436/* Give up root privilege at this point (assuming that exim_uid and exim_gid
1437are not root). The third argument controls the running of initgroups().
1438Normally we do this, in order to set up the groups for the Exim user. However,
1439if we are not root at this time - some odd installations run that way - we
1440cannot do this. */
1441
1442exim_setugid(exim_uid, exim_gid, geteuid()==root_uid, US"running as a daemon");
1443
1444/* Get somewhere to keep the list of queue-runner pids if we are keeping track
1445of them (and also if we are doing queue runs). */
1446
1447if (queue_interval > 0 && queue_run_max > 0)
1448 {
1449 int i;
1450 queue_pid_slots = store_get(queue_run_max * sizeof(pid_t));
1451 for (i = 0; i < queue_run_max; i++) queue_pid_slots[i] = 0;
1452 }
1453
1454/* Set up the handler for termination of child processes. */
1455
1456sigchld_seen = FALSE;
7d468ab8 1457os_non_restarting_signal(SIGCHLD, main_sigchld_handler);
059ec3d9
PH
1458
1459/* If we are to run the queue periodically, pretend the alarm has just gone
1460off. This will cause the first queue-runner to get kicked off straight away. */
1461
1462sigalrm_seen = (queue_interval > 0);
1463
1464/* Log the start up of a daemon - at least one of listening or queue running
1465must be set up. */
1466
1467if (daemon_listen)
1468 {
1469 int i, j;
1470 int smtp_ports = 0;
1471 int smtps_ports = 0;
1472 ip_address_item *ipa;
1473 uschar *p = big_buffer;
1474 uschar *qinfo = (queue_interval > 0)?
1475 string_sprintf("-q%s", readconf_printtime(queue_interval))
1476 :
1477 US"no queue runs";
1478
1479 /* Build a list of listening addresses in big_buffer, but limit it to 10
1480 items. The style is for backwards compatibility.
1481
1482 It is now possible to have some ports listening for SMTPS (the old,
1483 deprecated protocol that starts TLS without using STARTTLS), and others
1484 listening for standard SMTP. Keep their listings separate. */
1485
1486 for (j = 0; j < 2; j++)
1487 {
1488 for (i = 0, ipa = addresses; i < 10 && ipa != NULL; i++, ipa = ipa->next)
1489 {
1490 /* First time round, look for SMTP ports; second time round, look for
1491 SMTPS ports. For the first one of each, insert leading text. */
1492
1493 if (host_is_tls_on_connect_port(ipa->port) == (j > 0))
1494 {
1495 if (j == 0)
1496 {
1497 if (smtp_ports++ == 0)
1498 {
1499 memcpy(p, "SMTP on", 8);
1500 p += 7;
1501 }
1502 }
1503 else
1504 {
1505 if (smtps_ports++ == 0)
1506 {
1507 (void)sprintf(CS p, "%sSMTPS on",
1508 (smtp_ports == 0)? "":" and for ");
1509 while (*p != 0) p++;
1510 }
1511 }
1512
1513 /* Now the information about the port (and sometimes interface) */
1514
1515 if (ipa->address[0] == ':' && ipa->address[1] == 0)
1516 {
1517 if (ipa->next != NULL && ipa->next->address[0] == 0 &&
1518 ipa->next->port == ipa->port)
1519 {
1520 (void)sprintf(CS p, " port %d (IPv6 and IPv4)", ipa->port);
1521 ipa = ipa->next;
1522 }
1523 else if (ipa->v6_include_v4)
1524 (void)sprintf(CS p, " port %d (IPv6 with IPv4)", ipa->port);
1525 else
1526 (void)sprintf(CS p, " port %d (IPv6)", ipa->port);
1527 }
1528 else if (ipa->address[0] == 0)
1529 (void)sprintf(CS p, " port %d (IPv4)", ipa->port);
1530 else
1531 (void)sprintf(CS p, " [%s]:%d", ipa->address, ipa->port);
1532 while (*p != 0) p++;
1533 }
1534 }
1535
1536 if (ipa != NULL)
1537 {
1538 memcpy(p, " ...", 5);
1539 p += 4;
1540 }
1541 }
1542
1543 log_write(0, LOG_MAIN,
1544 "exim %s daemon started: pid=%d, %s, listening for %s",
1545 version_string, getpid(), qinfo, big_buffer);
1546 set_process_info("daemon: %s, listening for %s", qinfo, big_buffer);
1547 }
1548
1549else
1550 {
1551 log_write(0, LOG_MAIN,
1552 "exim %s daemon started: pid=%d, -q%s, not listening for SMTP",
1553 version_string, getpid(), readconf_printtime(queue_interval));
1554 set_process_info("daemon: -q%s, not listening",
1555 readconf_printtime(queue_interval));
1556 }
1557
1558
1559/* Close the log so it can be renamed and moved. In the few cases below where
1560this long-running process writes to the log (always exceptional conditions), it
1561closes the log afterwards, for the same reason. */
1562
1563log_close_all();
1564
1565DEBUG(D_any) debug_print_ids(US"daemon running with");
1566
1567/* Any messages accepted via this route are going to be SMTP. */
1568
1569smtp_input = TRUE;
1570
1571/* Enter the never-ending loop... */
1572
1573for (;;)
1574 {
1575 #if HAVE_IPV6
1576 struct sockaddr_in6 accepted;
1577 #else
1578 struct sockaddr_in accepted;
1579 #endif
1580
1581 SOCKLEN_T len = sizeof(accepted);
059ec3d9
PH
1582 pid_t pid;
1583
1584 /* This code is placed first in the loop, so that it gets obeyed at the
1585 start, before the first wait. This causes the first queue-runner to be
1586 started immediately. */
1587
1588 if (sigalrm_seen)
1589 {
1590 DEBUG(D_any) debug_printf("SIGALRM received\n");
1591
1592 /* Do a full queue run in a child process, if required, unless we already
1593 have enough queue runners on the go. If we are not running as root, a
1594 re-exec is required. */
1595
1596 if (queue_interval > 0 &&
1597 (queue_run_max <= 0 || queue_run_count < queue_run_max))
1598 {
1599 if ((pid = fork()) == 0)
1600 {
1601 int sk;
3d235903 1602
059ec3d9
PH
1603 DEBUG(D_any) debug_printf("Starting queue-runner: pid %d\n",
1604 (int)getpid());
1605
3d235903
PH
1606 /* Disable debugging if it's required only for the daemon process. We
1607 leave the above message, because it ties up with the "child ended"
1608 debugging messages. */
1609
1610 if (debug_daemon) debug_selector = 0;
1611
059ec3d9
PH
1612 /* Close any open listening sockets in the child */
1613
1614 for (sk = 0; sk < listen_socket_count; sk++) close(listen_sockets[sk]);
1615
1616 /* Reset SIGHUP and SIGCHLD in the child in both cases. */
1617
1618 signal(SIGHUP, SIG_DFL);
1619 signal(SIGCHLD, SIG_DFL);
1620
1621 /* Re-exec if privilege has been given up, unless deliver_drop_
1622 privilege is set. Reset SIGALRM before exec(). */
1623
1624 if (geteuid() != root_uid && !deliver_drop_privilege)
1625 {
1626 uschar opt[8];
1627 uschar *p = opt;
1628
1629 signal(SIGALRM, SIG_DFL);
1630 *p++ = '-';
1631 *p++ = 'q';
1632 if (queue_2stage) *p++ = 'q';
1633 if (queue_run_first_delivery) *p++ = 'i';
1634 if (queue_run_force) *p++ = 'f';
1635 if (deliver_force_thaw) *p++ = 'f';
1636 if (queue_run_local) *p++ = 'l';
1637 *p = 0;
1638
1639 (void)child_exec_exim(CEE_EXEC_PANIC, FALSE, NULL, TRUE, 1, opt);
1640 /* Control never returns here. */
1641 }
1642
1643 /* No need to re-exec; SIGALRM remains set to the default handler */
1644
1645 queue_run(NULL, NULL, FALSE);
1646 _exit(EXIT_SUCCESS);
1647 }
1648
1649 if (pid < 0)
1650 {
1651 log_write(0, LOG_MAIN|LOG_PANIC, "daemon: fork of queue-runner "
1652 "process failed: %s", strerror(errno));
1653 log_close_all();
1654 }
1655 else
1656 {
1657 int i;
1658 for (i = 0; i < queue_run_max; ++i)
1659 {
1660 if (queue_pid_slots[i] <= 0)
1661 {
1662 queue_pid_slots[i] = pid;
1663 queue_run_count++;
1664 break;
1665 }
1666 }
1667 DEBUG(D_any) debug_printf("%d queue-runner process%s running\n",
1668 queue_run_count, (queue_run_count == 1)? "" : "es");
1669 }
1670 }
1671
1672 /* Reset the alarm clock */
1673
1674 sigalrm_seen = FALSE;
1675 alarm(queue_interval);
1676 }
1677
1678
1679 /* Sleep till a connection happens if listening, and handle the connection if
1680 that is why we woke up. The FreeBSD operating system requires the use of
1681 select() before accept() because the latter function is not interrupted by
1682 a signal, and we want to wake up for SIGCHLD and SIGALRM signals. Some other
1683 OS do notice signals in accept() but it does no harm to have the select()
1684 in for all of them - and it won't then be a lurking problem for ports to
1685 new OS. In fact, the later addition of listening on specific interfaces only
1686 requires this way of working anyway. */
1687
1688 if (daemon_listen)
1689 {
bb6e88ff 1690 int sk, lcount, select_errno;
059ec3d9
PH
1691 int max_socket = 0;
1692 BOOL select_failed = FALSE;
1693 fd_set select_listen;
1694
1695 FD_ZERO(&select_listen);
1696 for (sk = 0; sk < listen_socket_count; sk++)
1697 {
1698 FD_SET(listen_sockets[sk], &select_listen);
1699 if (listen_sockets[sk] > max_socket) max_socket = listen_sockets[sk];
1700 }
1701
1702 DEBUG(D_any) debug_printf("Listening...\n");
7d468ab8
PH
1703
1704 /* In rare cases we may have had a SIGCHLD signal in the time between
1705 setting the handler (below) and getting back here. If so, pretend that the
1706 select() was interrupted so that we reap the child. This might still leave
1707 a small window when a SIGCHLD could get lost. However, since we use SIGCHLD
1708 only to do the reaping more quickly, it shouldn't result in anything other
1709 than a delay until something else causes a wake-up. */
1710
1711 if (sigchld_seen)
1712 {
1713 lcount = -1;
1714 errno = EINTR;
1715 }
1716 else
1717 {
1718 lcount = select(max_socket + 1, (SELECT_ARG2_TYPE *)&select_listen,
1719 NULL, NULL, NULL);
1720 }
059ec3d9 1721
7d468ab8 1722 if (lcount < 0)
059ec3d9
PH
1723 {
1724 select_failed = TRUE;
1725 lcount = 1;
1726 }
7d468ab8 1727
bb6e88ff
PH
1728 /* Clean up any subprocesses that may have terminated. We need to do this
1729 here so that smtp_accept_max_per_host works when a connection to that host
1730 has completed, and we are about to accept a new one. When this code was
1731 later in the sequence, a new connection could be rejected, even though an
1732 old one had just finished. Preserve the errno from any select() failure for
1733 the use of the common select/accept error processing below. */
1734
1735 select_errno = errno;
1736 handle_ending_processes();
1737 errno = select_errno;
1738
059ec3d9 1739 /* Loop for all the sockets that are currently ready to go. If select
7d468ab8
PH
1740 actually failed, we have set the count to 1 and select_failed=TRUE, so as
1741 to use the common error code for select/accept below. */
059ec3d9
PH
1742
1743 while (lcount-- > 0)
1744 {
1745 int accept_socket = -1;
1746 if (!select_failed)
1747 {
1748 for (sk = 0; sk < listen_socket_count; sk++)
1749 {
1750 if (FD_ISSET(listen_sockets[sk], &select_listen))
1751 {
1752 accept_socket = accept(listen_sockets[sk],
1753 (struct sockaddr *)&accepted, &len);
1754 FD_CLR(listen_sockets[sk], &select_listen);
1755 break;
1756 }
1757 }
1758 }
1759
1760 /* If select or accept has failed and this was not caused by an
1761 interruption, log the incident and try again. With asymmetric TCP/IP
1762 routing errors such as "No route to network" have been seen here. Also
1763 "connection reset by peer" has been seen. These cannot be classed as
1764 disastrous errors, but they could fill up a lot of log. The code in smail
1765 crashes the daemon after 10 successive failures of accept, on the grounds
1766 that some OS fail continuously. Exim originally followed suit, but this
1767 appears to have caused problems. Now it just keeps going, but instead of
1768 logging each error, it batches them up when they are continuous. */
1769
1770 if (accept_socket < 0 && errno != EINTR)
1771 {
1772 if (accept_retry_count == 0)
1773 {
1774 accept_retry_errno = errno;
1775 accept_retry_select_failed = select_failed;
1776 }
1777 else
1778 {
1779 if (errno != accept_retry_errno ||
1780 select_failed != accept_retry_select_failed ||
1781 accept_retry_count >= 50)
1782 {
1783 log_write(0, LOG_MAIN | ((accept_retry_count >= 50)? LOG_PANIC : 0),
1784 "%d %s() failure%s: %s",
1785 accept_retry_count,
1786 accept_retry_select_failed? "select" : "accept",
1787 (accept_retry_count == 1)? "" : "s",
1788 strerror(accept_retry_errno));
1789 log_close_all();
1790 accept_retry_count = 0;
1791 accept_retry_errno = errno;
1792 accept_retry_select_failed = select_failed;
1793 }
1794 }
1795 accept_retry_count++;
1796 }
1797
1798 else
1799 {
1800 if (accept_retry_count > 0)
1801 {
1802 log_write(0, LOG_MAIN, "%d %s() failure%s: %s",
1803 accept_retry_count,
1804 accept_retry_select_failed? "select" : "accept",
1805 (accept_retry_count == 1)? "" : "s",
1806 strerror(accept_retry_errno));
1807 log_close_all();
1808 accept_retry_count = 0;
1809 }
1810 }
1811
1812 /* If select/accept succeeded, deal with the connection. */
1813
1814 if (accept_socket >= 0)
1815 handle_smtp_call(listen_sockets, listen_socket_count, accept_socket,
1816 (struct sockaddr *)&accepted);
1817 }
1818 }
1819
1820 /* If not listening, then just sleep for the queue interval. If we woke
1821 up early the last time for some other signal, it won't matter because
1822 the alarm signal will wake at the right time. This code originally used
1823 sleep() but it turns out that on the FreeBSD system, sleep() is not inter-
1824 rupted by signals, so it wasn't waking up for SIGALRM or SIGCHLD. Luckily
1825 select() can be used as an interruptible sleep() on all versions of Unix. */
1826
1827 else
1828 {
1829 struct timeval tv;
1830 tv.tv_sec = queue_interval;
1831 tv.tv_usec = 0;
1832 select(0, NULL, NULL, NULL, &tv);
bb6e88ff 1833 handle_ending_processes();
059ec3d9
PH
1834 }
1835
1836 /* Re-enable the SIGCHLD handler if it has been run. It can't do it
1837 for itself, because it isn't doing the waiting itself. */
1838
1839 if (sigchld_seen)
1840 {
1841 sigchld_seen = FALSE;
7d468ab8 1842 os_non_restarting_signal(SIGCHLD, main_sigchld_handler);
059ec3d9
PH
1843 }
1844
1845 /* Handle being woken by SIGHUP. We know at this point that the result
1846 of accept() has been dealt with, so we can re-exec exim safely, first
1847 closing the listening sockets so that they can be reused. Cancel any pending
1848 alarm in case it is just about to go off, and set SIGHUP to be ignored so
1849 that another HUP in quick succession doesn't clobber the new daemon before it
1850 gets going. All log files get closed by the close-on-exec flag; however, if
1851 the exec fails, we need to close the logs. */
1852
1853 if (sighup_seen)
1854 {
1855 int sk;
1856 log_write(0, LOG_MAIN, "pid %d: SIGHUP received: re-exec daemon",
1857 getpid());
1858 for (sk = 0; sk < listen_socket_count; sk++) close(listen_sockets[sk]);
1859 alarm(0);
1860 signal(SIGHUP, SIG_IGN);
1861 sighup_argv[0] = exim_path;
1862 exim_nullstd();
1863 execv(CS exim_path, (char *const *)sighup_argv);
1864 log_write(0, LOG_MAIN|LOG_PANIC_DIE, "pid %d: exec of %s failed: %s",
1865 getpid(), exim_path, strerror(errno));
1866 log_close_all();
1867 }
1868
1869 } /* End of main loop */
1870
1871/* Control never reaches here */
1872}
1873
1874/* End of exim_daemon.c */
1875