1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21
22 /*
23 * Copyright (c) 2004, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Copyright (c) 2013, Joyent, Inc. All rights reserved.
25 * Copyright 2017 RackTop Systems.
26 */
27
28 /*
29 * restarter.c - service manipulation
30 *
31 * This component manages services whose restarter is svc.startd, the standard
32 * restarter. It translates restarter protocol events from the graph engine
33 * into actions on processes, as a delegated restarter would do.
34 *
35 * The master restarter manages a number of always-running threads:
36 * - restarter event thread: events from the graph engine
37 * - timeout thread: thread to fire queued timeouts
38 * - contract thread: thread to handle contract events
39 * - wait thread: thread to handle wait-based services
40 *
41 * The other threads are created as-needed:
42 * - per-instance method threads
43 * - per-instance event processing threads
44 *
45 * The interaction of all threads must result in the following conditions
46 * being satisfied (on a per-instance basis):
47 * - restarter events must be processed in order
48 * - method execution must be serialized
49 * - instance delete must be held until outstanding methods are complete
50 * - contract events shouldn't be processed while a method is running
51 * - timeouts should fire even when a method is running
52 *
53 * Service instances are represented by restarter_inst_t's and are kept in the
54 * instance_list list.
55 *
56 * Service States
57 * The current state of a service instance is kept in
58 * restarter_inst_t->ri_i.i_state. If transition to a new state could take
59 * some time, then before we effect the transition we set
60 * restarter_inst_t->ri_i.i_next_state to the target state, and afterwards we
61 * rotate i_next_state to i_state and set i_next_state to
62 * RESTARTER_STATE_NONE. So usually i_next_state is _NONE when ri_lock is not
63 * held. The exception is when we launch methods, which are done with
64 * a separate thread. To keep any other threads from grabbing ri_lock before
65 * method_thread() does, we set ri_method_thread to the thread id of the
66 * method thread, and when it is nonzero any thread with a different thread id
67 * waits on ri_method_cv.
68 *
69 * Method execution is serialized by blocking on ri_method_cv in
70 * inst_lookup_by_id() and waiting for a 0 value of ri_method_thread. This
71 * also prevents the instance structure from being deleted until all
72 * outstanding operations such as method_thread() have finished.
73 *
74 * Lock ordering:
75 *
76 * dgraph_lock [can be held when taking:]
77 * utmpx_lock
78 * dictionary->dict_lock
79 * st->st_load_lock
80 * wait_info_lock
81 * ru->restarter_update_lock
82 * restarter_queue->rpeq_lock
83 * instance_list.ril_lock
84 * inst->ri_lock
85 * st->st_configd_live_lock
86 *
87 * instance_list.ril_lock
88 * graph_queue->gpeq_lock
89 * gu->gu_lock
90 * st->st_configd_live_lock
91 * dictionary->dict_lock
92 * inst->ri_lock
93 * graph_queue->gpeq_lock
94 * gu->gu_lock
95 * tu->tu_lock
96 * tq->tq_lock
97 * inst->ri_queue_lock
98 * wait_info_lock
99 * bp->cb_lock
100 * utmpx_lock
101 *
102 * single_user_thread_lock
103 * wait_info_lock
104 * utmpx_lock
105 *
106 * gu_freeze_lock
107 *
108 * logbuf_mutex nests inside pretty much everything.
109 */
110
111 #include <sys/contract/process.h>
112 #include <sys/ctfs.h>
113 #include <sys/stat.h>
114 #include <sys/time.h>
115 #include <sys/types.h>
116 #include <sys/uio.h>
117 #include <sys/wait.h>
118 #include <assert.h>
119 #include <errno.h>
120 #include <fcntl.h>
121 #include <libcontract.h>
122 #include <libcontract_priv.h>
123 #include <libintl.h>
124 #include <librestart.h>
125 #include <librestart_priv.h>
126 #include <libuutil.h>
127 #include <limits.h>
128 #include <poll.h>
129 #include <port.h>
130 #include <pthread.h>
131 #include <stdarg.h>
132 #include <stdio.h>
133 #include <strings.h>
134 #include <unistd.h>
135
136 #include "startd.h"
137 #include "protocol.h"
138
139 static uu_list_pool_t *restarter_instance_pool;
140 static restarter_instance_list_t instance_list;
141
142 static uu_list_pool_t *restarter_queue_pool;
143
144 #define WT_SVC_ERR_THROTTLE 1 /* 1 sec delay for erroring wait svc */
145
146 /*
147 * Function used to reset the restart times for an instance, when
148 * an administrative task comes along and essentially makes the times
149 * in this array ineffective.
150 */
151 static void
152 reset_start_times(restarter_inst_t *inst)
153 {
154 inst->ri_start_index = 0;
155 bzero(inst->ri_start_time, sizeof (inst->ri_start_time));
156 }
157
158 /*ARGSUSED*/
159 static int
160 restarter_instance_compare(const void *lc_arg, const void *rc_arg,
161 void *private)
162 {
163 int lc_id = ((const restarter_inst_t *)lc_arg)->ri_id;
164 int rc_id = *(int *)rc_arg;
165
166 if (lc_id > rc_id)
167 return (1);
168 if (lc_id < rc_id)
169 return (-1);
170 return (0);
171 }
172
173 static restarter_inst_t *
174 inst_lookup_by_name(const char *name)
175 {
176 int id;
177
178 id = dict_lookup_byname(name);
179 if (id == -1)
180 return (NULL);
181
182 return (inst_lookup_by_id(id));
183 }
184
185 restarter_inst_t *
186 inst_lookup_by_id(int id)
187 {
188 restarter_inst_t *inst;
189
190 MUTEX_LOCK(&instance_list.ril_lock);
191 inst = uu_list_find(instance_list.ril_instance_list, &id, NULL, NULL);
192 if (inst != NULL)
193 MUTEX_LOCK(&inst->ri_lock);
194 MUTEX_UNLOCK(&instance_list.ril_lock);
195
196 if (inst != NULL) {
197 while (inst->ri_method_thread != 0 &&
198 !pthread_equal(inst->ri_method_thread, pthread_self())) {
199 ++inst->ri_method_waiters;
200 (void) pthread_cond_wait(&inst->ri_method_cv,
201 &inst->ri_lock);
202 assert(inst->ri_method_waiters > 0);
203 --inst->ri_method_waiters;
204 }
205 }
206
207 return (inst);
208 }
209
210 static restarter_inst_t *
211 inst_lookup_queue(const char *name)
212 {
213 int id;
214 restarter_inst_t *inst;
215
216 id = dict_lookup_byname(name);
217 if (id == -1)
218 return (NULL);
219
220 MUTEX_LOCK(&instance_list.ril_lock);
221 inst = uu_list_find(instance_list.ril_instance_list, &id, NULL, NULL);
222 if (inst != NULL)
223 MUTEX_LOCK(&inst->ri_queue_lock);
224 MUTEX_UNLOCK(&instance_list.ril_lock);
225
226 return (inst);
227 }
228
229 const char *
230 service_style(int flags)
231 {
232 switch (flags & RINST_STYLE_MASK) {
233 case RINST_CONTRACT: return ("contract");
234 case RINST_TRANSIENT: return ("transient");
235 case RINST_WAIT: return ("wait");
236
237 default:
238 #ifndef NDEBUG
239 uu_warn("%s:%d: Bad flags 0x%x.\n", __FILE__, __LINE__, flags);
240 #endif
241 abort();
242 /* NOTREACHED */
243 }
244 }
245
246 /*
247 * Fails with ECONNABORTED or ECANCELED.
248 */
249 static int
250 check_contract(restarter_inst_t *inst, boolean_t primary,
251 scf_instance_t *scf_inst)
252 {
253 ctid_t *ctidp;
254 int fd, r;
255
256 ctidp = primary ? &inst->ri_i.i_primary_ctid :
257 &inst->ri_i.i_transient_ctid;
258
259 assert(*ctidp >= 1);
260
261 fd = contract_open(*ctidp, NULL, "status", O_RDONLY);
262 if (fd >= 0) {
263 r = close(fd);
264 assert(r == 0);
265 return (0);
266 }
267
268 r = restarter_remove_contract(scf_inst, *ctidp, primary ?
269 RESTARTER_CONTRACT_PRIMARY : RESTARTER_CONTRACT_TRANSIENT);
270 switch (r) {
271 case 0:
272 case ECONNABORTED:
273 case ECANCELED:
274 *ctidp = 0;
275 return (r);
276
277 case ENOMEM:
278 uu_die("Out of memory\n");
279 /* NOTREACHED */
280
281 case EPERM:
282 uu_die("Insufficient privilege.\n");
283 /* NOTREACHED */
284
285 case EACCES:
286 uu_die("Repository backend access denied.\n");
287 /* NOTREACHED */
288
289 case EROFS:
290 log_error(LOG_INFO, "Could not remove unusable contract id %ld "
291 "for %s from repository.\n", *ctidp, inst->ri_i.i_fmri);
292 return (0);
293
294 case EINVAL:
295 case EBADF:
296 default:
297 assert(0);
298 abort();
299 /* NOTREACHED */
300 }
301 }
302
303 static int stop_instance(scf_handle_t *, restarter_inst_t *, stop_cause_t);
304
305 /*
306 * int restarter_insert_inst(scf_handle_t *, char *)
307 * If the inst is already in the restarter list, return its id. If the inst
308 * is not in the restarter list, initialize a restarter_inst_t, initialize its
309 * states, insert it into the list, and return 0.
310 *
311 * Fails with
312 * ENOENT - name is not in the repository
313 */
314 static int
315 restarter_insert_inst(scf_handle_t *h, const char *name)
316 {
317 int id, r;
318 restarter_inst_t *inst;
319 uu_list_index_t idx;
320 scf_service_t *scf_svc;
321 scf_instance_t *scf_inst;
322 scf_snapshot_t *snap = NULL;
323 scf_propertygroup_t *pg;
324 char *svc_name, *inst_name;
325 char logfilebuf[PATH_MAX];
326 char *c;
327 boolean_t do_commit_states;
328 restarter_instance_state_t state, next_state;
329 protocol_states_t *ps;
330 pid_t start_pid;
331 restarter_str_t reason = restarter_str_insert_in_graph;
332
333 MUTEX_LOCK(&instance_list.ril_lock);
334
335 /*
336 * We don't use inst_lookup_by_name() here because we want the lookup
337 * & insert to be atomic.
338 */
339 id = dict_lookup_byname(name);
340 if (id != -1) {
341 inst = uu_list_find(instance_list.ril_instance_list, &id, NULL,
342 &idx);
343 if (inst != NULL) {
344 MUTEX_UNLOCK(&instance_list.ril_lock);
345 return (0);
346 }
347 }
348
349 /* Allocate an instance */
350 inst = startd_zalloc(sizeof (restarter_inst_t));
351 inst->ri_utmpx_prefix = startd_alloc(max_scf_value_size);
352 inst->ri_utmpx_prefix[0] = '\0';
353
354 inst->ri_i.i_fmri = startd_alloc(strlen(name) + 1);
355 (void) strcpy((char *)inst->ri_i.i_fmri, name);
356
357 inst->ri_queue = startd_list_create(restarter_queue_pool, inst, 0);
358
359 /*
360 * id shouldn't be -1 since we use the same dictionary as graph.c, but
361 * just in case.
362 */
363 inst->ri_id = (id != -1 ? id : dict_insert(name));
364
365 special_online_hooks_get(name, &inst->ri_pre_online_hook,
366 &inst->ri_post_online_hook, &inst->ri_post_offline_hook);
367
368 scf_svc = safe_scf_service_create(h);
369 scf_inst = safe_scf_instance_create(h);
370 pg = safe_scf_pg_create(h);
371 svc_name = startd_alloc(max_scf_name_size);
372 inst_name = startd_alloc(max_scf_name_size);
373
374 rep_retry:
375 if (snap != NULL)
376 scf_snapshot_destroy(snap);
377 if (inst->ri_logstem != NULL)
378 startd_free(inst->ri_logstem, PATH_MAX);
379 if (inst->ri_common_name != NULL)
380 free(inst->ri_common_name);
381 if (inst->ri_C_common_name != NULL)
382 free(inst->ri_C_common_name);
383 snap = NULL;
384 inst->ri_logstem = NULL;
385 inst->ri_common_name = NULL;
386 inst->ri_C_common_name = NULL;
387
388 if (scf_handle_decode_fmri(h, name, NULL, scf_svc, scf_inst, NULL,
389 NULL, SCF_DECODE_FMRI_EXACT) != 0) {
390 switch (scf_error()) {
391 case SCF_ERROR_CONNECTION_BROKEN:
392 libscf_handle_rebind(h);
393 goto rep_retry;
394
395 case SCF_ERROR_NOT_FOUND:
396 goto deleted;
397 }
398
399 uu_die("Can't decode FMRI %s: %s\n", name,
400 scf_strerror(scf_error()));
401 }
402
403 /*
404 * If there's no running snapshot, then we execute using the editing
405 * snapshot. Pending snapshots will be taken later.
406 */
407 snap = libscf_get_running_snapshot(scf_inst);
408
409 if ((scf_service_get_name(scf_svc, svc_name, max_scf_name_size) < 0) ||
410 (scf_instance_get_name(scf_inst, inst_name, max_scf_name_size) <
411 0)) {
412 switch (scf_error()) {
413 case SCF_ERROR_NOT_SET:
414 break;
415
416 case SCF_ERROR_CONNECTION_BROKEN:
417 libscf_handle_rebind(h);
418 goto rep_retry;
419
420 default:
421 assert(0);
422 abort();
423 }
424
425 goto deleted;
426 }
427
428 (void) snprintf(logfilebuf, PATH_MAX, "%s:%s", svc_name, inst_name);
429 for (c = logfilebuf; *c != '\0'; c++)
430 if (*c == '/')
431 *c = '-';
432
433 inst->ri_logstem = startd_alloc(PATH_MAX);
434 (void) snprintf(inst->ri_logstem, PATH_MAX, "%s%s", logfilebuf,
435 LOG_SUFFIX);
436
437 /*
438 * If the restarter group is missing, use uninit/none. Otherwise,
439 * we're probably being restarted & don't want to mess up the states
440 * that are there.
441 */
442 state = RESTARTER_STATE_UNINIT;
443 next_state = RESTARTER_STATE_NONE;
444
445 r = scf_instance_get_pg(scf_inst, SCF_PG_RESTARTER, pg);
446 if (r != 0) {
447 switch (scf_error()) {
448 case SCF_ERROR_CONNECTION_BROKEN:
449 libscf_handle_rebind(h);
450 goto rep_retry;
451
452 case SCF_ERROR_NOT_SET:
453 goto deleted;
454
455 case SCF_ERROR_NOT_FOUND:
456 /*
457 * This shouldn't happen since the graph engine should
458 * have initialized the state to uninitialized/none if
459 * there was no restarter pg. In case somebody
460 * deleted it, though....
461 */
462 do_commit_states = B_TRUE;
463 break;
464
465 default:
466 assert(0);
467 abort();
468 }
469 } else {
470 r = libscf_read_states(pg, &state, &next_state);
471 if (r != 0) {
472 do_commit_states = B_TRUE;
473 } else {
474 if (next_state != RESTARTER_STATE_NONE) {
475 /*
476 * Force next_state to _NONE since we
477 * don't look for method processes.
478 */
479 next_state = RESTARTER_STATE_NONE;
480 do_commit_states = B_TRUE;
481 } else {
482 /*
483 * The reason for transition will depend on
484 * state.
485 */
486 if (st->st_initial == 0)
487 reason = restarter_str_startd_restart;
488 else if (state == RESTARTER_STATE_MAINT)
489 reason = restarter_str_bad_repo_state;
490 /*
491 * Inform the restarter of our state without
492 * changing the STIME in the repository.
493 */
494 ps = startd_alloc(sizeof (*ps));
495 inst->ri_i.i_state = ps->ps_state = state;
496 inst->ri_i.i_next_state = ps->ps_state_next =
497 next_state;
498 ps->ps_reason = reason;
499
500 graph_protocol_send_event(inst->ri_i.i_fmri,
501 GRAPH_UPDATE_STATE_CHANGE, ps);
502
503 do_commit_states = B_FALSE;
504 }
505 }
506 }
507
508 switch (libscf_get_startd_properties(scf_inst, snap, &inst->ri_flags,
509 &inst->ri_utmpx_prefix)) {
510 case 0:
511 break;
512
513 case ECONNABORTED:
514 libscf_handle_rebind(h);
515 goto rep_retry;
516
517 case ECANCELED:
518 goto deleted;
519
520 case ENOENT:
521 /*
522 * This is odd, because the graph engine should have required
523 * the general property group. So we'll just use default
524 * flags in anticipation of the graph engine sending us
525 * REMOVE_INSTANCE when it finds out that the general property
526 * group has been deleted.
527 */
528 inst->ri_flags = RINST_CONTRACT;
529 break;
530
531 default:
532 assert(0);
533 abort();
534 }
535
536 r = libscf_get_template_values(scf_inst, snap,
537 &inst->ri_common_name, &inst->ri_C_common_name);
538
539 /*
540 * Copy our names to smaller buffers to reduce our memory footprint.
541 */
542 if (inst->ri_common_name != NULL) {
543 char *tmp = safe_strdup(inst->ri_common_name);
544 startd_free(inst->ri_common_name, max_scf_value_size);
545 inst->ri_common_name = tmp;
546 }
547
548 if (inst->ri_C_common_name != NULL) {
549 char *tmp = safe_strdup(inst->ri_C_common_name);
550 startd_free(inst->ri_C_common_name, max_scf_value_size);
551 inst->ri_C_common_name = tmp;
552 }
553
554 switch (r) {
555 case 0:
556 break;
557
558 case ECONNABORTED:
559 libscf_handle_rebind(h);
560 goto rep_retry;
561
562 case ECANCELED:
563 goto deleted;
564
565 case ECHILD:
566 case ENOENT:
567 break;
568
569 default:
570 assert(0);
571 abort();
572 }
573
574 switch (libscf_read_method_ids(h, scf_inst, inst->ri_i.i_fmri,
575 &inst->ri_i.i_primary_ctid, &inst->ri_i.i_transient_ctid,
576 &start_pid)) {
577 case 0:
578 break;
579
580 case ECONNABORTED:
581 libscf_handle_rebind(h);
582 goto rep_retry;
583
584 case ECANCELED:
585 goto deleted;
586
587 default:
588 assert(0);
589 abort();
590 }
591
592 if (inst->ri_i.i_primary_ctid >= 1) {
593 contract_hash_store(inst->ri_i.i_primary_ctid, inst->ri_id);
594
595 switch (check_contract(inst, B_TRUE, scf_inst)) {
596 case 0:
597 break;
598
599 case ECONNABORTED:
600 libscf_handle_rebind(h);
601 goto rep_retry;
602
603 case ECANCELED:
604 goto deleted;
605
606 default:
607 assert(0);
608 abort();
609 }
610 }
611
612 if (inst->ri_i.i_transient_ctid >= 1) {
613 switch (check_contract(inst, B_FALSE, scf_inst)) {
614 case 0:
615 break;
616
617 case ECONNABORTED:
618 libscf_handle_rebind(h);
619 goto rep_retry;
620
621 case ECANCELED:
622 goto deleted;
623
624 default:
625 assert(0);
626 abort();
627 }
628 }
629
630 /* No more failures we live through, so add it to the list. */
631 (void) pthread_mutex_init(&inst->ri_lock, &mutex_attrs);
632 (void) pthread_mutex_init(&inst->ri_queue_lock, &mutex_attrs);
633 MUTEX_LOCK(&inst->ri_lock);
634 MUTEX_LOCK(&inst->ri_queue_lock);
635
636 (void) pthread_cond_init(&inst->ri_method_cv, NULL);
637
638 uu_list_node_init(inst, &inst->ri_link, restarter_instance_pool);
639 uu_list_insert(instance_list.ril_instance_list, inst, idx);
640 MUTEX_UNLOCK(&instance_list.ril_lock);
641
642 if (start_pid != -1 &&
643 (inst->ri_flags & RINST_STYLE_MASK) == RINST_WAIT) {
644 int ret;
645 ret = wait_register(start_pid, inst->ri_i.i_fmri, 0, 1);
646 if (ret == -1) {
647 /*
648 * Implication: if we can't reregister the
649 * instance, we will start another one. Two
650 * instances may or may not result in a resource
651 * conflict.
652 */
653 log_error(LOG_WARNING,
654 "%s: couldn't reregister %ld for wait\n",
655 inst->ri_i.i_fmri, start_pid);
656 } else if (ret == 1) {
657 /*
658 * Leading PID has exited.
659 */
660 (void) stop_instance(h, inst, RSTOP_EXIT);
661 }
662 }
663
664
665 scf_pg_destroy(pg);
666
667 if (do_commit_states)
668 (void) restarter_instance_update_states(h, inst, state,
669 next_state, RERR_NONE, reason);
670
671 log_framework(LOG_DEBUG, "%s is a %s-style service\n", name,
672 service_style(inst->ri_flags));
673
674 MUTEX_UNLOCK(&inst->ri_queue_lock);
675 MUTEX_UNLOCK(&inst->ri_lock);
676
677 startd_free(svc_name, max_scf_name_size);
678 startd_free(inst_name, max_scf_name_size);
679 scf_snapshot_destroy(snap);
680 scf_instance_destroy(scf_inst);
681 scf_service_destroy(scf_svc);
682
683 log_framework(LOG_DEBUG, "%s: inserted instance into restarter list\n",
684 name);
685
686 return (0);
687
688 deleted:
689 MUTEX_UNLOCK(&instance_list.ril_lock);
690 startd_free(inst_name, max_scf_name_size);
691 startd_free(svc_name, max_scf_name_size);
692 if (snap != NULL)
693 scf_snapshot_destroy(snap);
694 scf_pg_destroy(pg);
695 scf_instance_destroy(scf_inst);
696 scf_service_destroy(scf_svc);
697 startd_free((void *)inst->ri_i.i_fmri, strlen(inst->ri_i.i_fmri) + 1);
698 uu_list_destroy(inst->ri_queue);
699 if (inst->ri_logstem != NULL)
700 startd_free(inst->ri_logstem, PATH_MAX);
701 if (inst->ri_common_name != NULL)
702 free(inst->ri_common_name);
703 if (inst->ri_C_common_name != NULL)
704 free(inst->ri_C_common_name);
705 startd_free(inst->ri_utmpx_prefix, max_scf_value_size);
706 startd_free(inst, sizeof (restarter_inst_t));
707 return (ENOENT);
708 }
709
710 static void
711 restarter_delete_inst(restarter_inst_t *ri)
712 {
713 int id;
714 restarter_inst_t *rip;
715 void *cookie = NULL;
716 restarter_instance_qentry_t *e;
717
718 assert(MUTEX_HELD(&ri->ri_lock));
719
720 /*
721 * Must drop the instance lock so we can pick up the instance_list
722 * lock & remove the instance.
723 */
724 id = ri->ri_id;
725 MUTEX_UNLOCK(&ri->ri_lock);
726
727 MUTEX_LOCK(&instance_list.ril_lock);
728
729 rip = uu_list_find(instance_list.ril_instance_list, &id, NULL, NULL);
730 if (rip == NULL) {
731 MUTEX_UNLOCK(&instance_list.ril_lock);
732 return;
733 }
734
735 assert(ri == rip);
736
737 uu_list_remove(instance_list.ril_instance_list, ri);
738
739 log_framework(LOG_DEBUG, "%s: deleted instance from restarter list\n",
740 ri->ri_i.i_fmri);
741
742 MUTEX_UNLOCK(&instance_list.ril_lock);
743
744 /*
745 * We can lock the instance without holding the instance_list lock
746 * since we removed the instance from the list.
747 */
748 MUTEX_LOCK(&ri->ri_lock);
749 MUTEX_LOCK(&ri->ri_queue_lock);
750
751 if (ri->ri_i.i_primary_ctid >= 1)
752 contract_hash_remove(ri->ri_i.i_primary_ctid);
753
754 while (ri->ri_method_thread != 0 || ri->ri_method_waiters > 0)
755 (void) pthread_cond_wait(&ri->ri_method_cv, &ri->ri_lock);
756
757 while ((e = uu_list_teardown(ri->ri_queue, &cookie)) != NULL)
758 startd_free(e, sizeof (*e));
759 uu_list_destroy(ri->ri_queue);
760
761 startd_free((void *)ri->ri_i.i_fmri, strlen(ri->ri_i.i_fmri) + 1);
762 startd_free(ri->ri_logstem, PATH_MAX);
763 if (ri->ri_common_name != NULL)
764 free(ri->ri_common_name);
765 if (ri->ri_C_common_name != NULL)
766 free(ri->ri_C_common_name);
767 startd_free(ri->ri_utmpx_prefix, max_scf_value_size);
768 (void) pthread_mutex_destroy(&ri->ri_lock);
769 (void) pthread_mutex_destroy(&ri->ri_queue_lock);
770 startd_free(ri, sizeof (restarter_inst_t));
771 }
772
773 /*
774 * instance_is_wait_style()
775 *
776 * Returns 1 if the given instance is a "wait-style" service instance.
777 */
778 int
779 instance_is_wait_style(restarter_inst_t *inst)
780 {
781 assert(MUTEX_HELD(&inst->ri_lock));
782 return ((inst->ri_flags & RINST_STYLE_MASK) == RINST_WAIT);
783 }
784
785 /*
786 * instance_is_transient_style()
787 *
788 * Returns 1 if the given instance is a transient service instance.
789 */
790 int
791 instance_is_transient_style(restarter_inst_t *inst)
792 {
793 assert(MUTEX_HELD(&inst->ri_lock));
794 return ((inst->ri_flags & RINST_STYLE_MASK) == RINST_TRANSIENT);
795 }
796
797 /*
798 * instance_in_transition()
799 * Returns 1 if instance is in transition, 0 if not
800 */
801 int
802 instance_in_transition(restarter_inst_t *inst)
803 {
804 assert(MUTEX_HELD(&inst->ri_lock));
805 if (inst->ri_i.i_next_state == RESTARTER_STATE_NONE)
806 return (0);
807 return (1);
808 }
809
810 /*
811 * returns 1 if instance is already started, 0 if not
812 */
813 static int
814 instance_started(restarter_inst_t *inst)
815 {
816 int ret;
817
818 assert(MUTEX_HELD(&inst->ri_lock));
819
820 if (inst->ri_i.i_state == RESTARTER_STATE_ONLINE ||
821 inst->ri_i.i_state == RESTARTER_STATE_DEGRADED)
822 ret = 1;
823 else
824 ret = 0;
825
826 return (ret);
827 }
828
829 /*
830 * Returns
831 * 0 - success
832 * ECONNRESET - success, but h was rebound
833 */
834 int
835 restarter_instance_update_states(scf_handle_t *h, restarter_inst_t *ri,
836 restarter_instance_state_t new_state,
837 restarter_instance_state_t new_state_next, restarter_error_t err,
838 restarter_str_t reason)
839 {
840 protocol_states_t *states;
841 int e;
842 uint_t retry_count = 0, msecs = ALLOC_DELAY;
843 boolean_t rebound = B_FALSE;
844 int prev_state_online;
845 int state_online;
846
847 assert(MUTEX_HELD(&ri->ri_lock));
848
849 prev_state_online = instance_started(ri);
850
851 retry:
852 e = _restarter_commit_states(h, &ri->ri_i, new_state, new_state_next,
853 restarter_get_str_short(reason));
854 switch (e) {
855 case 0:
856 break;
857
858 case ENOMEM:
859 ++retry_count;
860 if (retry_count < ALLOC_RETRY) {
861 (void) poll(NULL, 0, msecs);
862 msecs *= ALLOC_DELAY_MULT;
863 goto retry;
864 }
865
866 /* Like startd_alloc(). */
867 uu_die("Insufficient memory.\n");
868 /* NOTREACHED */
869
870 case ECONNABORTED:
871 libscf_handle_rebind(h);
872 rebound = B_TRUE;
873 goto retry;
874
875 case EPERM:
876 case EACCES:
877 case EROFS:
878 log_error(LOG_NOTICE, "Could not commit state change for %s "
879 "to repository: %s.\n", ri->ri_i.i_fmri, strerror(e));
880 /* FALLTHROUGH */
881
882 case ENOENT:
883 ri->ri_i.i_state = new_state;
884 ri->ri_i.i_next_state = new_state_next;
885 break;
886
887 case EINVAL:
888 default:
889 bad_error("_restarter_commit_states", e);
890 }
891
892 states = startd_alloc(sizeof (protocol_states_t));
893 states->ps_state = new_state;
894 states->ps_state_next = new_state_next;
895 states->ps_err = err;
896 states->ps_reason = reason;
897 graph_protocol_send_event(ri->ri_i.i_fmri, GRAPH_UPDATE_STATE_CHANGE,
898 (void *)states);
899
900 state_online = instance_started(ri);
901
902 if (prev_state_online && !state_online)
903 ri->ri_post_offline_hook();
904 else if (!prev_state_online && state_online)
905 ri->ri_post_online_hook();
906
907 return (rebound ? ECONNRESET : 0);
908 }
909
910 void
911 restarter_mark_pending_snapshot(const char *fmri, uint_t flag)
912 {
913 restarter_inst_t *inst;
914
915 assert(flag == RINST_RETAKE_RUNNING || flag == RINST_RETAKE_START);
916
917 inst = inst_lookup_by_name(fmri);
918 if (inst == NULL)
919 return;
920
921 inst->ri_flags |= flag;
922
923 MUTEX_UNLOCK(&inst->ri_lock);
924 }
925
926 static void
927 restarter_take_pending_snapshots(scf_handle_t *h)
928 {
929 restarter_inst_t *inst;
930 int r;
931
932 MUTEX_LOCK(&instance_list.ril_lock);
933
934 for (inst = uu_list_first(instance_list.ril_instance_list);
935 inst != NULL;
936 inst = uu_list_next(instance_list.ril_instance_list, inst)) {
937 const char *fmri;
938 scf_instance_t *sinst = NULL;
939
940 MUTEX_LOCK(&inst->ri_lock);
941
942 /*
943 * This is where we'd check inst->ri_method_thread and if it
944 * were nonzero we'd wait in anticipation of another thread
945 * executing a method for inst. Doing so with the instance_list
946 * locked, though, leads to deadlock. Since taking a snapshot
947 * during that window won't hurt anything, we'll just continue.
948 */
949
950 fmri = inst->ri_i.i_fmri;
951
952 if (inst->ri_flags & RINST_RETAKE_RUNNING) {
953 scf_snapshot_t *rsnap;
954
955 (void) libscf_fmri_get_instance(h, fmri, &sinst);
956
957 rsnap = libscf_get_or_make_running_snapshot(sinst,
958 fmri, B_FALSE);
959
960 scf_instance_destroy(sinst);
961
962 if (rsnap != NULL)
963 inst->ri_flags &= ~RINST_RETAKE_RUNNING;
964
965 scf_snapshot_destroy(rsnap);
966 }
967
968 if (inst->ri_flags & RINST_RETAKE_START) {
969 switch (r = libscf_snapshots_poststart(h, fmri,
970 B_FALSE)) {
971 case 0:
972 case ENOENT:
973 inst->ri_flags &= ~RINST_RETAKE_START;
974 break;
975
976 case ECONNABORTED:
977 break;
978
979 case EACCES:
980 default:
981 bad_error("libscf_snapshots_poststart", r);
982 }
983 }
984
985 MUTEX_UNLOCK(&inst->ri_lock);
986 }
987
988 MUTEX_UNLOCK(&instance_list.ril_lock);
989 }
990
991 /* ARGSUSED */
992 void *
993 restarter_post_fsminimal_thread(void *unused)
994 {
995 scf_handle_t *h;
996 int r;
997
998 h = libscf_handle_create_bound_loop();
999
1000 for (;;) {
1001 r = libscf_create_self(h);
1002 if (r == 0)
1003 break;
1004
1005 assert(r == ECONNABORTED);
1006 libscf_handle_rebind(h);
1007 }
1008
1009 restarter_take_pending_snapshots(h);
1010
1011 (void) scf_handle_unbind(h);
1012 scf_handle_destroy(h);
1013
1014 return (NULL);
1015 }
1016
1017 /*
1018 * int stop_instance()
1019 *
1020 * Stop the instance identified by the instance given as the second argument,
1021 * for the cause stated.
1022 *
1023 * Returns
1024 * 0 - success
1025 * -1 - inst is in transition
1026 */
1027 static int
1028 stop_instance(scf_handle_t *local_handle, restarter_inst_t *inst,
1029 stop_cause_t cause)
1030 {
1031 fork_info_t *info;
1032 const char *cp;
1033 int err;
1034 restarter_error_t re;
1035 restarter_str_t reason;
1036 restarter_instance_state_t new_state;
1037
1038 assert(MUTEX_HELD(&inst->ri_lock));
1039 assert(inst->ri_method_thread == 0);
1040
1041 switch (cause) {
1042 case RSTOP_EXIT:
1043 re = RERR_RESTART;
1044 reason = restarter_str_ct_ev_exit;
1045 cp = "all processes in service exited";
1046 break;
1047 case RSTOP_ERR_CFG:
1048 re = RERR_FAULT;
1049 reason = restarter_str_method_failed;
1050 cp = "service exited with a configuration error";
1051 break;
1052 case RSTOP_ERR_EXIT:
1053 re = RERR_RESTART;
1054 reason = restarter_str_ct_ev_exit;
1055 cp = "service exited with an error";
1056 break;
1057 case RSTOP_CORE:
1058 re = RERR_FAULT;
1059 reason = restarter_str_ct_ev_core;
1060 cp = "process dumped core";
1061 break;
1062 case RSTOP_SIGNAL:
1063 re = RERR_FAULT;
1064 reason = restarter_str_ct_ev_signal;
1065 cp = "process received fatal signal from outside the service";
1066 break;
1067 case RSTOP_HWERR:
1068 re = RERR_FAULT;
1069 reason = restarter_str_ct_ev_hwerr;
1070 cp = "process killed due to uncorrectable hardware error";
1071 break;
1072 case RSTOP_DEPENDENCY:
1073 re = RERR_RESTART;
1074 reason = restarter_str_dependency_activity;
1075 cp = "dependency activity requires stop";
1076 break;
1077 case RSTOP_DISABLE:
1078 re = RERR_RESTART;
1079 reason = restarter_str_disable_request;
1080 cp = "service disabled";
1081 break;
1082 case RSTOP_RESTART:
1083 re = RERR_RESTART;
1084 reason = restarter_str_restart_request;
1085 cp = "service restarting";
1086 break;
1087 default:
1088 #ifndef NDEBUG
1089 (void) fprintf(stderr, "Unknown cause %d at %s:%d.\n",
1090 cause, __FILE__, __LINE__);
1091 #endif
1092 abort();
1093 }
1094
1095 /* Services in the disabled and maintenance state are ignored */
1096 if (inst->ri_i.i_state == RESTARTER_STATE_MAINT ||
1097 inst->ri_i.i_state == RESTARTER_STATE_DISABLED) {
1098 log_framework(LOG_DEBUG,
1099 "%s: stop_instance -> is maint/disabled\n",
1100 inst->ri_i.i_fmri);
1101 return (0);
1102 }
1103
1104 /* Already stopped instances are left alone */
1105 if (instance_started(inst) == 0) {
1106 log_framework(LOG_DEBUG, "Restarter: %s is already stopped.\n",
1107 inst->ri_i.i_fmri);
1108 return (0);
1109 }
1110
1111 if (instance_in_transition(inst)) {
1112 /* requeue event by returning -1 */
1113 log_framework(LOG_DEBUG,
1114 "Restarter: Not stopping %s, in transition.\n",
1115 inst->ri_i.i_fmri);
1116 return (-1);
1117 }
1118
1119 log_instance(inst, B_TRUE, "Stopping because %s.", cp);
1120
1121 log_framework(re == RERR_FAULT ? LOG_INFO : LOG_DEBUG,
1122 "%s: Instance stopping because %s.\n", inst->ri_i.i_fmri, cp);
1123
1124 if (instance_is_wait_style(inst) &&
1125 (cause == RSTOP_EXIT ||
1126 cause == RSTOP_ERR_CFG ||
1127 cause == RSTOP_ERR_EXIT)) {
1128 /*
1129 * No need to stop instance, as child has exited; remove
1130 * contract and move the instance to the offline state.
1131 */
1132 switch (err = restarter_instance_update_states(local_handle,
1133 inst, inst->ri_i.i_state, RESTARTER_STATE_OFFLINE, re,
1134 reason)) {
1135 case 0:
1136 case ECONNRESET:
1137 break;
1138
1139 default:
1140 bad_error("restarter_instance_update_states", err);
1141 }
1142
1143 if (cause == RSTOP_ERR_EXIT) {
1144 /*
1145 * The RSTOP_ERR_EXIT cause is set via the
1146 * wait_thread -> wait_remove code path when we have
1147 * a "wait" style svc that exited with an error. If
1148 * the svc is failing too quickly, we throttle it so
1149 * that we don't restart it more than once/second.
1150 * Since we know we're running in the wait thread its
1151 * ok to throttle it right here.
1152 */
1153 (void) update_fault_count(inst, FAULT_COUNT_INCR);
1154 if (method_rate_critical(inst)) {
1155 log_instance(inst, B_TRUE, "Failing too "
1156 "quickly, throttling.");
1157 (void) sleep(WT_SVC_ERR_THROTTLE);
1158 }
1159 } else {
1160 (void) update_fault_count(inst, FAULT_COUNT_RESET);
1161 reset_start_times(inst);
1162 }
1163
1164 if (inst->ri_i.i_primary_ctid != 0) {
1165 inst->ri_m_inst =
1166 safe_scf_instance_create(local_handle);
1167 inst->ri_mi_deleted = B_FALSE;
1168
1169 libscf_reget_instance(inst);
1170 method_remove_contract(inst, B_TRUE, B_TRUE);
1171
1172 scf_instance_destroy(inst->ri_m_inst);
1173 inst->ri_m_inst = NULL;
1174 }
1175
1176 switch (err = restarter_instance_update_states(local_handle,
1177 inst, inst->ri_i.i_next_state, RESTARTER_STATE_NONE, re,
1178 reason)) {
1179 case 0:
1180 case ECONNRESET:
1181 break;
1182
1183 default:
1184 bad_error("restarter_instance_update_states", err);
1185 }
1186
1187 if (cause != RSTOP_ERR_CFG)
1188 return (0);
1189 } else if (instance_is_wait_style(inst) && re == RERR_RESTART) {
1190 /*
1191 * Stopping a wait service through means other than the pid
1192 * exiting should keep wait_thread() from restarting the
1193 * service, by removing it from the wait list.
1194 * We cannot remove it right now otherwise the process will
1195 * end up <defunct> so mark it to be ignored.
1196 */
1197 wait_ignore_by_fmri(inst->ri_i.i_fmri);
1198 }
1199
1200 /*
1201 * There are some configuration errors which we cannot detect until we
1202 * try to run the method. For example, see exec_method() where the
1203 * restarter_set_method_context() call can return SMF_EXIT_ERR_CONFIG
1204 * in several cases. If this happens for a "wait-style" svc,
1205 * wait_remove() sets the cause as RSTOP_ERR_CFG so that we can detect
1206 * the configuration error and go into maintenance, even though it is
1207 * a "wait-style" svc.
1208 */
1209 if (cause == RSTOP_ERR_CFG)
1210 new_state = RESTARTER_STATE_MAINT;
1211 else
1212 new_state = inst->ri_i.i_enabled ?
1213 RESTARTER_STATE_OFFLINE : RESTARTER_STATE_DISABLED;
1214
1215 switch (err = restarter_instance_update_states(local_handle, inst,
1216 inst->ri_i.i_state, new_state, RERR_NONE, reason)) {
1217 case 0:
1218 case ECONNRESET:
1219 break;
1220
1221 default:
1222 bad_error("restarter_instance_update_states", err);
1223 }
1224
1225 info = startd_zalloc(sizeof (fork_info_t));
1226
1227 info->sf_id = inst->ri_id;
1228 info->sf_method_type = METHOD_STOP;
1229 info->sf_event_type = re;
1230 info->sf_reason = reason;
1231 inst->ri_method_thread = startd_thread_create(method_thread, info);
1232
1233 return (0);
1234 }
1235
1236 /*
1237 * Returns
1238 * ENOENT - fmri is not in instance_list
1239 * 0 - success
1240 * ECONNRESET - success, though handle was rebound
1241 * -1 - instance is in transition
1242 */
1243 int
1244 stop_instance_fmri(scf_handle_t *h, const char *fmri, uint_t flags)
1245 {
1246 restarter_inst_t *rip;
1247 int r;
1248
1249 rip = inst_lookup_by_name(fmri);
1250 if (rip == NULL)
1251 return (ENOENT);
1252
1253 r = stop_instance(h, rip, flags);
1254
1255 MUTEX_UNLOCK(&rip->ri_lock);
1256
1257 return (r);
1258 }
1259
1260 static void
1261 unmaintain_instance(scf_handle_t *h, restarter_inst_t *rip,
1262 unmaint_cause_t cause)
1263 {
1264 ctid_t ctid;
1265 scf_instance_t *inst;
1266 int r;
1267 uint_t tries = 0, msecs = ALLOC_DELAY;
1268 const char *cp;
1269 restarter_str_t reason;
1270
1271 assert(MUTEX_HELD(&rip->ri_lock));
1272
1273 if (rip->ri_i.i_state != RESTARTER_STATE_MAINT) {
1274 log_error(LOG_DEBUG, "Restarter: "
1275 "Ignoring maintenance off command because %s is not in the "
1276 "maintenance state.\n", rip->ri_i.i_fmri);
1277 return;
1278 }
1279
1280 switch (cause) {
1281 case RUNMAINT_CLEAR:
1282 cp = "clear requested";
1283 reason = restarter_str_clear_request;
1284 break;
1285 case RUNMAINT_DISABLE:
1286 cp = "disable requested";
1287 reason = restarter_str_disable_request;
1288 break;
1289 default:
1290 #ifndef NDEBUG
1291 (void) fprintf(stderr, "Uncaught case for %d at %s:%d.\n",
1292 cause, __FILE__, __LINE__);
1293 #endif
1294 abort();
1295 }
1296
1297 log_instance(rip, B_TRUE, "Leaving maintenance because %s.",
1298 cp);
1299 log_framework(LOG_DEBUG, "%s: Instance leaving maintenance because "
1300 "%s.\n", rip->ri_i.i_fmri, cp);
1301
1302 (void) restarter_instance_update_states(h, rip, RESTARTER_STATE_UNINIT,
1303 RESTARTER_STATE_NONE, RERR_RESTART, reason);
1304
1305 /*
1306 * If we did ADMIN_MAINT_ON_IMMEDIATE, then there might still be
1307 * a primary contract.
1308 */
1309 if (rip->ri_i.i_primary_ctid == 0)
1310 return;
1311
1312 ctid = rip->ri_i.i_primary_ctid;
1313 contract_abandon(ctid);
1314 rip->ri_i.i_primary_ctid = 0;
1315
1316 rep_retry:
1317 switch (r = libscf_fmri_get_instance(h, rip->ri_i.i_fmri, &inst)) {
1318 case 0:
1319 break;
1320
1321 case ECONNABORTED:
1322 libscf_handle_rebind(h);
1323 goto rep_retry;
1324
1325 case ENOENT:
1326 /* Must have been deleted. */
1327 return;
1328
1329 case EINVAL:
1330 case ENOTSUP:
1331 default:
1332 bad_error("libscf_handle_rebind", r);
1333 }
1334
1335 again:
1336 r = restarter_remove_contract(inst, ctid, RESTARTER_CONTRACT_PRIMARY);
1337 switch (r) {
1338 case 0:
1339 break;
1340
1341 case ENOMEM:
1342 ++tries;
1343 if (tries < ALLOC_RETRY) {
1344 (void) poll(NULL, 0, msecs);
1345 msecs *= ALLOC_DELAY_MULT;
1346 goto again;
1347 }
1348
1349 uu_die("Insufficient memory.\n");
1350 /* NOTREACHED */
1351
1352 case ECONNABORTED:
1353 scf_instance_destroy(inst);
1354 libscf_handle_rebind(h);
1355 goto rep_retry;
1356
1357 case ECANCELED:
1358 break;
1359
1360 case EPERM:
1361 case EACCES:
1362 case EROFS:
1363 log_error(LOG_INFO,
1364 "Could not remove contract id %lu for %s (%s).\n", ctid,
1365 rip->ri_i.i_fmri, strerror(r));
1366 break;
1367
1368 case EINVAL:
1369 case EBADF:
1370 default:
1371 bad_error("restarter_remove_contract", r);
1372 }
1373
1374 scf_instance_destroy(inst);
1375 }
1376
1377 /*
1378 * enable_inst()
1379 * Set inst->ri_i.i_enabled. Expects 'e' to be _ENABLE, _DISABLE, or
1380 * _ADMIN_DISABLE. If the event is _ENABLE and inst is uninitialized or
1381 * disabled, move it to offline. If the event is _DISABLE or
1382 * _ADMIN_DISABLE, make sure inst will move to disabled.
1383 *
1384 * Returns
1385 * 0 - success
1386 * ECONNRESET - h was rebound
1387 */
1388 static int
1389 enable_inst(scf_handle_t *h, restarter_inst_t *inst,
1390 restarter_instance_qentry_t *riq)
1391 {
1392 restarter_instance_state_t state;
1393 restarter_event_type_t e = riq->riq_type;
1394 restarter_str_t reason = restarter_str_per_configuration;
1395 int r;
1396
1397 assert(MUTEX_HELD(&inst->ri_lock));
1398 assert(e == RESTARTER_EVENT_TYPE_ADMIN_DISABLE ||
1399 e == RESTARTER_EVENT_TYPE_DISABLE ||
1400 e == RESTARTER_EVENT_TYPE_ENABLE);
1401 assert(instance_in_transition(inst) == 0);
1402
1403 state = inst->ri_i.i_state;
1404
1405 if (e == RESTARTER_EVENT_TYPE_ENABLE) {
1406 inst->ri_i.i_enabled = 1;
1407
1408 if (state == RESTARTER_STATE_UNINIT ||
1409 state == RESTARTER_STATE_DISABLED) {
1410 /*
1411 * B_FALSE: Don't log an error if the log_instance()
1412 * fails because it will fail on the miniroot before
1413 * install-discovery runs.
1414 */
1415 log_instance(inst, B_FALSE, "Enabled.");
1416 log_framework(LOG_DEBUG, "%s: Instance enabled.\n",
1417 inst->ri_i.i_fmri);
1418
1419 /*
1420 * If we are coming from DISABLED, it was obviously an
1421 * enable request. If we are coming from UNINIT, it may
1422 * have been a sevice in MAINT that was cleared.
1423 */
1424 if (riq->riq_reason == restarter_str_clear_request)
1425 reason = restarter_str_clear_request;
1426 else if (state == RESTARTER_STATE_DISABLED)
1427 reason = restarter_str_enable_request;
1428 (void) restarter_instance_update_states(h, inst,
1429 RESTARTER_STATE_OFFLINE, RESTARTER_STATE_NONE,
1430 RERR_NONE, reason);
1431 } else {
1432 log_framework(LOG_DEBUG, "Restarter: "
1433 "Not changing state of %s for enable command.\n",
1434 inst->ri_i.i_fmri);
1435 }
1436 } else {
1437 inst->ri_i.i_enabled = 0;
1438
1439 switch (state) {
1440 case RESTARTER_STATE_ONLINE:
1441 case RESTARTER_STATE_DEGRADED:
1442 r = stop_instance(h, inst, RSTOP_DISABLE);
1443 return (r == ECONNRESET ? 0 : r);
1444
1445 case RESTARTER_STATE_OFFLINE:
1446 case RESTARTER_STATE_UNINIT:
1447 if (inst->ri_i.i_primary_ctid != 0) {
1448 inst->ri_m_inst = safe_scf_instance_create(h);
1449 inst->ri_mi_deleted = B_FALSE;
1450
1451 libscf_reget_instance(inst);
1452 method_remove_contract(inst, B_TRUE, B_TRUE);
1453
1454 scf_instance_destroy(inst->ri_m_inst);
1455 }
1456 /* B_FALSE: See log_instance(..., "Enabled."); above */
1457 log_instance(inst, B_FALSE, "Disabled.");
1458 log_framework(LOG_DEBUG, "%s: Instance disabled.\n",
1459 inst->ri_i.i_fmri);
1460
1461 /*
1462 * If we are coming from OFFLINE, it was obviously a
1463 * disable request. But if we are coming from
1464 * UNINIT, it may have been a disable request for a
1465 * service in MAINT.
1466 */
1467 if (riq->riq_reason == restarter_str_disable_request ||
1468 state == RESTARTER_STATE_OFFLINE)
1469 reason = restarter_str_disable_request;
1470 (void) restarter_instance_update_states(h, inst,
1471 RESTARTER_STATE_DISABLED, RESTARTER_STATE_NONE,
1472 RERR_RESTART, reason);
1473 return (0);
1474
1475 case RESTARTER_STATE_DISABLED:
1476 break;
1477
1478 case RESTARTER_STATE_MAINT:
1479 /*
1480 * We only want to pull the instance out of maintenance
1481 * if the disable is on adminstrative request. The
1482 * graph engine sends _DISABLE events whenever a
1483 * service isn't in the disabled state, and we don't
1484 * want to pull the service out of maintenance if,
1485 * for example, it is there due to a dependency cycle.
1486 */
1487 if (e == RESTARTER_EVENT_TYPE_ADMIN_DISABLE)
1488 unmaintain_instance(h, inst, RUNMAINT_DISABLE);
1489 break;
1490
1491 default:
1492 #ifndef NDEBUG
1493 (void) fprintf(stderr, "Restarter instance %s has "
1494 "unknown state %d.\n", inst->ri_i.i_fmri, state);
1495 #endif
1496 abort();
1497 }
1498 }
1499
1500 return (0);
1501 }
1502
1503 static void
1504 start_instance(scf_handle_t *local_handle, restarter_inst_t *inst,
1505 int32_t reason)
1506 {
1507 fork_info_t *info;
1508 restarter_str_t new_reason;
1509
1510 assert(MUTEX_HELD(&inst->ri_lock));
1511 assert(instance_in_transition(inst) == 0);
1512 assert(inst->ri_method_thread == 0);
1513
1514 log_framework(LOG_DEBUG, "%s: trying to start instance\n",
1515 inst->ri_i.i_fmri);
1516
1517 /*
1518 * We want to keep the original reason for restarts and clear actions
1519 */
1520 switch (reason) {
1521 case restarter_str_restart_request:
1522 case restarter_str_clear_request:
1523 new_reason = reason;
1524 break;
1525 default:
1526 new_reason = restarter_str_dependencies_satisfied;
1527 }
1528
1529 /* Services in the disabled and maintenance state are ignored */
1530 if (inst->ri_i.i_state == RESTARTER_STATE_MAINT ||
1531 inst->ri_i.i_state == RESTARTER_STATE_DISABLED ||
1532 inst->ri_i.i_enabled == 0) {
1533 log_framework(LOG_DEBUG,
1534 "%s: start_instance -> is maint/disabled\n",
1535 inst->ri_i.i_fmri);
1536 return;
1537 }
1538
1539 /* Already started instances are left alone */
1540 if (instance_started(inst) == 1) {
1541 log_framework(LOG_DEBUG,
1542 "%s: start_instance -> is already started\n",
1543 inst->ri_i.i_fmri);
1544 return;
1545 }
1546
1547 log_framework(LOG_DEBUG, "%s: starting instance.\n", inst->ri_i.i_fmri);
1548
1549 (void) restarter_instance_update_states(local_handle, inst,
1550 inst->ri_i.i_state, RESTARTER_STATE_ONLINE, RERR_NONE, new_reason);
1551
1552 info = startd_zalloc(sizeof (fork_info_t));
1553
1554 info->sf_id = inst->ri_id;
1555 info->sf_method_type = METHOD_START;
1556 info->sf_event_type = RERR_NONE;
1557 info->sf_reason = new_reason;
1558 inst->ri_method_thread = startd_thread_create(method_thread, info);
1559 }
1560
1561 static int
1562 event_from_tty(scf_handle_t *h, restarter_inst_t *rip)
1563 {
1564 scf_instance_t *inst;
1565 int ret = 0;
1566
1567 if (libscf_fmri_get_instance(h, rip->ri_i.i_fmri, &inst))
1568 return (-1);
1569
1570 ret = restarter_inst_ractions_from_tty(inst);
1571
1572 scf_instance_destroy(inst);
1573 return (ret);
1574 }
1575
1576 static boolean_t
1577 restart_dump(scf_handle_t *h, restarter_inst_t *rip)
1578 {
1579 scf_instance_t *inst;
1580 boolean_t ret = B_FALSE;
1581
1582 if (libscf_fmri_get_instance(h, rip->ri_i.i_fmri, &inst))
1583 return (-1);
1584
1585 if (restarter_inst_dump(inst) == 1)
1586 ret = B_TRUE;
1587
1588 scf_instance_destroy(inst);
1589 return (ret);
1590 }
1591
1592 static void
1593 maintain_instance(scf_handle_t *h, restarter_inst_t *rip, int immediate,
1594 restarter_str_t reason)
1595 {
1596 fork_info_t *info;
1597 scf_instance_t *scf_inst = NULL;
1598
1599 assert(MUTEX_HELD(&rip->ri_lock));
1600 assert(reason != restarter_str_none);
1601 assert(rip->ri_method_thread == 0);
1602
1603 log_instance(rip, B_TRUE, "Stopping for maintenance due to %s.",
1604 restarter_get_str_short(reason));
1605 log_framework(LOG_DEBUG, "%s: stopping for maintenance due to %s.\n",
1606 rip->ri_i.i_fmri, restarter_get_str_short(reason));
1607
1608 /* Services in the maintenance state are ignored */
1609 if (rip->ri_i.i_state == RESTARTER_STATE_MAINT) {
1610 log_framework(LOG_DEBUG,
1611 "%s: maintain_instance -> is already in maintenance\n",
1612 rip->ri_i.i_fmri);
1613 return;
1614 }
1615
1616 /*
1617 * If reason state is restarter_str_service_request and
1618 * restarter_actions/auxiliary_fmri property is set with a valid fmri,
1619 * copy the fmri to restarter/auxiliary_fmri so svcs -x can use.
1620 */
1621 if (reason == restarter_str_service_request &&
1622 libscf_fmri_get_instance(h, rip->ri_i.i_fmri, &scf_inst) == 0) {
1623 if (restarter_inst_validate_ractions_aux_fmri(scf_inst) == 0) {
1624 if (restarter_inst_set_aux_fmri(scf_inst))
1625 log_framework(LOG_DEBUG, "%s: "
1626 "restarter_inst_set_aux_fmri failed: ",
1627 rip->ri_i.i_fmri);
1628 } else {
1629 log_framework(LOG_DEBUG, "%s: "
1630 "restarter_inst_validate_ractions_aux_fmri "
1631 "failed: ", rip->ri_i.i_fmri);
1632
1633 if (restarter_inst_reset_aux_fmri(scf_inst))
1634 log_framework(LOG_DEBUG, "%s: "
1635 "restarter_inst_reset_aux_fmri failed: ",
1636 rip->ri_i.i_fmri);
1637 }
1638 scf_instance_destroy(scf_inst);
1639 }
1640
1641 if (immediate || !instance_started(rip)) {
1642 if (rip->ri_i.i_primary_ctid != 0) {
1643 rip->ri_m_inst = safe_scf_instance_create(h);
1644 rip->ri_mi_deleted = B_FALSE;
1645
1646 libscf_reget_instance(rip);
1647 method_remove_contract(rip, B_TRUE, B_TRUE);
1648
1649 scf_instance_destroy(rip->ri_m_inst);
1650 }
1651
1652 (void) restarter_instance_update_states(h, rip,
1653 RESTARTER_STATE_MAINT, RESTARTER_STATE_NONE, RERR_RESTART,
1654 reason);
1655 return;
1656 }
1657
1658 (void) restarter_instance_update_states(h, rip, rip->ri_i.i_state,
1659 RESTARTER_STATE_MAINT, RERR_NONE, reason);
1660
1661 log_transition(rip, MAINT_REQUESTED);
1662
1663 info = startd_zalloc(sizeof (*info));
1664 info->sf_id = rip->ri_id;
1665 info->sf_method_type = METHOD_STOP;
1666 info->sf_event_type = RERR_RESTART;
1667 info->sf_reason = reason;
1668 rip->ri_method_thread = startd_thread_create(method_thread, info);
1669 }
1670
1671 static void
1672 refresh_instance(scf_handle_t *h, restarter_inst_t *rip)
1673 {
1674 scf_instance_t *inst;
1675 scf_snapshot_t *snap;
1676 fork_info_t *info;
1677 int r;
1678
1679 assert(MUTEX_HELD(&rip->ri_lock));
1680
1681 log_instance(rip, B_TRUE, "Rereading configuration.");
1682 log_framework(LOG_DEBUG, "%s: rereading configuration.\n",
1683 rip->ri_i.i_fmri);
1684
1685 rep_retry:
1686 r = libscf_fmri_get_instance(h, rip->ri_i.i_fmri, &inst);
1687 switch (r) {
1688 case 0:
1689 break;
1690
1691 case ECONNABORTED:
1692 libscf_handle_rebind(h);
1693 goto rep_retry;
1694
1695 case ENOENT:
1696 /* Must have been deleted. */
1697 return;
1698
1699 case EINVAL:
1700 case ENOTSUP:
1701 default:
1702 bad_error("libscf_fmri_get_instance", r);
1703 }
1704
1705 snap = libscf_get_running_snapshot(inst);
1706
1707 r = libscf_get_startd_properties(inst, snap, &rip->ri_flags,
1708 &rip->ri_utmpx_prefix);
1709 switch (r) {
1710 case 0:
1711 log_framework(LOG_DEBUG, "%s is a %s-style service\n",
1712 rip->ri_i.i_fmri, service_style(rip->ri_flags));
1713 break;
1714
1715 case ECONNABORTED:
1716 scf_instance_destroy(inst);
1717 scf_snapshot_destroy(snap);
1718 libscf_handle_rebind(h);
1719 goto rep_retry;
1720
1721 case ECANCELED:
1722 case ENOENT:
1723 /* Succeed in anticipation of REMOVE_INSTANCE. */
1724 break;
1725
1726 default:
1727 bad_error("libscf_get_startd_properties", r);
1728 }
1729
1730 if (instance_started(rip)) {
1731 /* Refresh does not change the state. */
1732 (void) restarter_instance_update_states(h, rip,
1733 rip->ri_i.i_state, rip->ri_i.i_state, RERR_NONE,
1734 restarter_str_refresh);
1735
1736 info = startd_zalloc(sizeof (*info));
1737 info->sf_id = rip->ri_id;
1738 info->sf_method_type = METHOD_REFRESH;
1739 info->sf_event_type = RERR_REFRESH;
1740 info->sf_reason = NULL;
1741
1742 assert(rip->ri_method_thread == 0);
1743 rip->ri_method_thread =
1744 startd_thread_create(method_thread, info);
1745 }
1746
1747 scf_snapshot_destroy(snap);
1748 scf_instance_destroy(inst);
1749 }
1750
1751 static void
1752 degrade_instance(scf_handle_t *h, restarter_inst_t *rip, restarter_str_t reason)
1753 {
1754 scf_instance_t *scf_inst = NULL;
1755
1756 assert(MUTEX_HELD(&rip->ri_lock));
1757
1758 log_instance(rip, B_TRUE, "Marking degraded due to %s.",
1759 restarter_get_str_short(reason));
1760 log_framework(LOG_DEBUG, "%s: marking degraded due to %s.\n",
1761 rip->ri_i.i_fmri, restarter_get_str_short(reason));
1762
1763 /* Services that aren't online are ignored */
1764 if (rip->ri_i.i_state != RESTARTER_STATE_ONLINE) {
1765 log_framework(LOG_DEBUG,
1766 "%s: degrade_instance -> is not online\n",
1767 rip->ri_i.i_fmri);
1768 return;
1769 }
1770
1771 /*
1772 * If reason state is restarter_str_service_request and
1773 * restarter_actions/auxiliary_fmri property is set with a valid fmri,
1774 * copy the fmri to restarter/auxiliary_fmri so svcs -x can use.
1775 */
1776 if (reason == restarter_str_service_request &&
1777 libscf_fmri_get_instance(h, rip->ri_i.i_fmri, &scf_inst) == 0) {
1778 if (restarter_inst_validate_ractions_aux_fmri(scf_inst) == 0) {
1779 if (restarter_inst_set_aux_fmri(scf_inst))
1780 log_framework(LOG_DEBUG, "%s: "
1781 "restarter_inst_set_aux_fmri failed: ",
1782 rip->ri_i.i_fmri);
1783 } else {
1784 log_framework(LOG_DEBUG, "%s: "
1785 "restarter_inst_validate_ractions_aux_fmri "
1786 "failed: ", rip->ri_i.i_fmri);
1787
1788 if (restarter_inst_reset_aux_fmri(scf_inst))
1789 log_framework(LOG_DEBUG, "%s: "
1790 "restarter_inst_reset_aux_fmri failed: ",
1791 rip->ri_i.i_fmri);
1792 }
1793 scf_instance_destroy(scf_inst);
1794 }
1795
1796 (void) restarter_instance_update_states(h, rip,
1797 RESTARTER_STATE_DEGRADED, RESTARTER_STATE_NONE, RERR_NONE, reason);
1798
1799 log_transition(rip, DEGRADE_REQUESTED);
1800 }
1801
1802 const char *event_names[] = { "INVALID", "ADD_INSTANCE", "REMOVE_INSTANCE",
1803 "ENABLE", "DISABLE", "ADMIN_RESTORE", "ADMIN_DEGRADED",
1804 "ADMIN_DEGRADE_IMMEDIATE", "ADMIN_REFRESH", "ADMIN_RESTART",
1805 "ADMIN_MAINT_OFF", "ADMIN_MAINT_ON", "ADMIN_MAINT_ON_IMMEDIATE",
1806 "STOP", "START", "DEPENDENCY_CYCLE", "INVALID_DEPENDENCY",
1807 "ADMIN_DISABLE", "STOP_RESET"
1808 };
1809
1810 /*
1811 * void *restarter_process_events()
1812 *
1813 * Called in a separate thread to process the events on an instance's
1814 * queue. Empties the queue completely, and tries to keep the thread
1815 * around for a little while after the queue is empty to save on
1816 * startup costs.
1817 */
1818 static void *
1819 restarter_process_events(void *arg)
1820 {
1821 scf_handle_t *h;
1822 restarter_instance_qentry_t *event;
1823 restarter_inst_t *rip;
1824 char *fmri = (char *)arg;
1825 struct timespec to;
1826
1827 assert(fmri != NULL);
1828
1829 h = libscf_handle_create_bound_loop();
1830
1831 /* grab the queue lock */
1832 rip = inst_lookup_queue(fmri);
1833 if (rip == NULL)
1834 goto out;
1835
1836 again:
1837
1838 while ((event = uu_list_first(rip->ri_queue)) != NULL) {
1839 restarter_inst_t *inst;
1840
1841 /* drop the queue lock */
1842 MUTEX_UNLOCK(&rip->ri_queue_lock);
1843
1844 /*
1845 * Grab the inst lock -- this waits until any outstanding
1846 * method finishes running.
1847 */
1848 inst = inst_lookup_by_name(fmri);
1849 if (inst == NULL) {
1850 /* Getting deleted in the middle isn't an error. */
1851 goto cont;
1852 }
1853
1854 assert(instance_in_transition(inst) == 0);
1855
1856 /* process the event */
1857 switch (event->riq_type) {
1858 case RESTARTER_EVENT_TYPE_ENABLE:
1859 case RESTARTER_EVENT_TYPE_DISABLE:
1860 (void) enable_inst(h, inst, event);
1861 break;
1862
1863 case RESTARTER_EVENT_TYPE_ADMIN_DISABLE:
1864 if (enable_inst(h, inst, event) == 0)
1865 reset_start_times(inst);
1866 break;
1867
1868 case RESTARTER_EVENT_TYPE_REMOVE_INSTANCE:
1869 restarter_delete_inst(inst);
1870 inst = NULL;
1871 goto cont;
1872
1873 case RESTARTER_EVENT_TYPE_STOP_RESET:
1874 reset_start_times(inst);
1875 /* FALLTHROUGH */
1876 case RESTARTER_EVENT_TYPE_STOP:
1877 (void) stop_instance(h, inst, RSTOP_DEPENDENCY);
1878 break;
1879
1880 case RESTARTER_EVENT_TYPE_START:
1881 start_instance(h, inst, event->riq_reason);
1882 break;
1883
1884 case RESTARTER_EVENT_TYPE_DEPENDENCY_CYCLE:
1885 maintain_instance(h, inst, 0,
1886 restarter_str_dependency_cycle);
1887 break;
1888
1889 case RESTARTER_EVENT_TYPE_INVALID_DEPENDENCY:
1890 maintain_instance(h, inst, 0,
1891 restarter_str_invalid_dependency);
1892 break;
1893
1894 case RESTARTER_EVENT_TYPE_ADMIN_MAINT_ON:
1895 if (event_from_tty(h, inst) == 0)
1896 maintain_instance(h, inst, 0,
1897 restarter_str_service_request);
1898 else
1899 maintain_instance(h, inst, 0,
1900 restarter_str_administrative_request);
1901 break;
1902
1903 case RESTARTER_EVENT_TYPE_ADMIN_MAINT_ON_IMMEDIATE:
1904 if (event_from_tty(h, inst) == 0)
1905 maintain_instance(h, inst, 1,
1906 restarter_str_service_request);
1907 else
1908 maintain_instance(h, inst, 1,
1909 restarter_str_administrative_request);
1910 break;
1911
1912 case RESTARTER_EVENT_TYPE_ADMIN_MAINT_OFF:
1913 unmaintain_instance(h, inst, RUNMAINT_CLEAR);
1914 reset_start_times(inst);
1915 break;
1916
1917 case RESTARTER_EVENT_TYPE_ADMIN_REFRESH:
1918 refresh_instance(h, inst);
1919 break;
1920
1921 case RESTARTER_EVENT_TYPE_ADMIN_DEGRADED:
1922 case RESTARTER_EVENT_TYPE_ADMIN_DEGRADE_IMMEDIATE:
1923 if (event_from_tty(h, inst) == 0)
1924 degrade_instance(h, inst,
1925 restarter_str_service_request);
1926 else
1927 degrade_instance(h, inst,
1928 restarter_str_administrative_request);
1929 break;
1930
1931 case RESTARTER_EVENT_TYPE_ADMIN_RESTORE:
1932 case RESTARTER_EVENT_TYPE_ADMIN_RESTART:
1933 if (!instance_started(inst)) {
1934 log_framework(LOG_DEBUG, "Restarter: "
1935 "Not restarting %s; not running.\n",
1936 inst->ri_i.i_fmri);
1937 } else {
1938 /*
1939 * Stop the instance. If it can be restarted,
1940 * the graph engine will send a new event.
1941 */
1942 if (restart_dump(h, inst)) {
1943 (void) contract_kill(
1944 inst->ri_i.i_primary_ctid, SIGABRT,
1945 inst->ri_i.i_fmri);
1946 } else if (stop_instance(h, inst,
1947 RSTOP_RESTART) == 0) {
1948 reset_start_times(inst);
1949 }
1950 }
1951 break;
1952
1953 case RESTARTER_EVENT_TYPE_ADD_INSTANCE:
1954 default:
1955 #ifndef NDEBUG
1956 uu_warn("%s:%d: Bad restarter event %d. "
1957 "Aborting.\n", __FILE__, __LINE__, event->riq_type);
1958 #endif
1959 abort();
1960 }
1961
1962 assert(inst != NULL);
1963 MUTEX_UNLOCK(&inst->ri_lock);
1964
1965 cont:
1966 /* grab the queue lock */
1967 rip = inst_lookup_queue(fmri);
1968 if (rip == NULL)
1969 goto out;
1970
1971 /* delete the event */
1972 uu_list_remove(rip->ri_queue, event);
1973 startd_free(event, sizeof (restarter_instance_qentry_t));
1974 }
1975
1976 assert(rip != NULL);
1977
1978 /*
1979 * Try to preserve the thread for a little while for future use.
1980 */
1981 to.tv_sec = 3;
1982 to.tv_nsec = 0;
1983 (void) pthread_cond_reltimedwait_np(&rip->ri_queue_cv,
1984 &rip->ri_queue_lock, &to);
1985
1986 if (uu_list_first(rip->ri_queue) != NULL)
1987 goto again;
1988
1989 rip->ri_queue_thread = 0;
1990 MUTEX_UNLOCK(&rip->ri_queue_lock);
1991
1992 out:
1993 (void) scf_handle_unbind(h);
1994 scf_handle_destroy(h);
1995 free(fmri);
1996 return (NULL);
1997 }
1998
1999 static int
2000 is_admin_event(restarter_event_type_t t) {
2001
2002 switch (t) {
2003 case RESTARTER_EVENT_TYPE_ADMIN_MAINT_ON:
2004 case RESTARTER_EVENT_TYPE_ADMIN_MAINT_ON_IMMEDIATE:
2005 case RESTARTER_EVENT_TYPE_ADMIN_MAINT_OFF:
2006 case RESTARTER_EVENT_TYPE_ADMIN_REFRESH:
2007 case RESTARTER_EVENT_TYPE_ADMIN_DEGRADED:
2008 case RESTARTER_EVENT_TYPE_ADMIN_DEGRADE_IMMEDIATE:
2009 case RESTARTER_EVENT_TYPE_ADMIN_RESTORE:
2010 case RESTARTER_EVENT_TYPE_ADMIN_RESTART:
2011 return (1);
2012 default:
2013 return (0);
2014 }
2015 }
2016
2017 static void
2018 restarter_queue_event(restarter_inst_t *ri, restarter_protocol_event_t *e)
2019 {
2020 restarter_instance_qentry_t *qe;
2021 int r;
2022
2023 assert(MUTEX_HELD(&ri->ri_queue_lock));
2024 assert(!MUTEX_HELD(&ri->ri_lock));
2025
2026 qe = startd_zalloc(sizeof (restarter_instance_qentry_t));
2027 qe->riq_type = e->rpe_type;
2028 qe->riq_reason = e->rpe_reason;
2029
2030 uu_list_node_init(qe, &qe->riq_link, restarter_queue_pool);
2031 r = uu_list_insert_before(ri->ri_queue, NULL, qe);
2032 assert(r == 0);
2033 }
2034
2035 /*
2036 * void *restarter_event_thread()
2037 *
2038 * Handle incoming graph events by placing them on a per-instance
2039 * queue. We can't lock the main part of the instance structure, so
2040 * just modify the seprarately locked event queue portion.
2041 */
2042 /*ARGSUSED*/
2043 static void *
2044 restarter_event_thread(void *unused)
2045 {
2046 scf_handle_t *h;
2047
2048 /*
2049 * This is a new thread, and thus, gets its own handle
2050 * to the repository.
2051 */
2052 h = libscf_handle_create_bound_loop();
2053
2054 MUTEX_LOCK(&ru->restarter_update_lock);
2055
2056 /*CONSTCOND*/
2057 while (1) {
2058 restarter_protocol_event_t *e;
2059
2060 while (ru->restarter_update_wakeup == 0)
2061 (void) pthread_cond_wait(&ru->restarter_update_cv,
2062 &ru->restarter_update_lock);
2063
2064 ru->restarter_update_wakeup = 0;
2065
2066 while ((e = restarter_event_dequeue()) != NULL) {
2067 restarter_inst_t *rip;
2068 char *fmri;
2069
2070 MUTEX_UNLOCK(&ru->restarter_update_lock);
2071
2072 /*
2073 * ADD_INSTANCE is special: there's likely no
2074 * instance structure yet, so we need to handle the
2075 * addition synchronously.
2076 */
2077 switch (e->rpe_type) {
2078 case RESTARTER_EVENT_TYPE_ADD_INSTANCE:
2079 if (restarter_insert_inst(h, e->rpe_inst) != 0)
2080 log_error(LOG_INFO, "Restarter: "
2081 "Could not add %s.\n", e->rpe_inst);
2082
2083 MUTEX_LOCK(&st->st_load_lock);
2084 if (--st->st_load_instances == 0)
2085 (void) pthread_cond_broadcast(
2086 &st->st_load_cv);
2087 MUTEX_UNLOCK(&st->st_load_lock);
2088
2089 goto nolookup;
2090 }
2091
2092 /*
2093 * Lookup the instance, locking only the event queue.
2094 * Can't grab ri_lock here because it might be held
2095 * by a long-running method.
2096 */
2097 rip = inst_lookup_queue(e->rpe_inst);
2098 if (rip == NULL) {
2099 log_error(LOG_INFO, "Restarter: "
2100 "Ignoring %s command for unknown service "
2101 "%s.\n", event_names[e->rpe_type],
2102 e->rpe_inst);
2103 goto nolookup;
2104 }
2105
2106 /* Keep ADMIN events from filling up the queue. */
2107 if (is_admin_event(e->rpe_type) &&
2108 uu_list_numnodes(rip->ri_queue) >
2109 RINST_QUEUE_THRESHOLD) {
2110 MUTEX_UNLOCK(&rip->ri_queue_lock);
2111 log_instance(rip, B_TRUE, "Instance event "
2112 "queue overflow. Dropping administrative "
2113 "request.");
2114 log_framework(LOG_DEBUG, "%s: Instance event "
2115 "queue overflow. Dropping administrative "
2116 "request.\n", rip->ri_i.i_fmri);
2117 goto nolookup;
2118 }
2119
2120 /* Now add the event to the instance queue. */
2121 restarter_queue_event(rip, e);
2122
2123 if (rip->ri_queue_thread == 0) {
2124 /*
2125 * Start a thread if one isn't already
2126 * running.
2127 */
2128 fmri = safe_strdup(e->rpe_inst);
2129 rip->ri_queue_thread = startd_thread_create(
2130 restarter_process_events, (void *)fmri);
2131 } else {
2132 /*
2133 * Signal the existing thread that there's
2134 * a new event.
2135 */
2136 (void) pthread_cond_broadcast(
2137 &rip->ri_queue_cv);
2138 }
2139
2140 MUTEX_UNLOCK(&rip->ri_queue_lock);
2141 nolookup:
2142 restarter_event_release(e);
2143
2144 MUTEX_LOCK(&ru->restarter_update_lock);
2145 }
2146 }
2147
2148 /*
2149 * Unreachable for now -- there's currently no graceful cleanup
2150 * called on exit().
2151 */
2152 (void) scf_handle_unbind(h);
2153 scf_handle_destroy(h);
2154 return (NULL);
2155 }
2156
2157 static restarter_inst_t *
2158 contract_to_inst(ctid_t ctid)
2159 {
2160 restarter_inst_t *inst;
2161 int id;
2162
2163 id = lookup_inst_by_contract(ctid);
2164 if (id == -1)
2165 return (NULL);
2166
2167 inst = inst_lookup_by_id(id);
2168 if (inst != NULL) {
2169 /*
2170 * Since ri_lock isn't held by the contract id lookup, this
2171 * instance may have been restarted and now be in a new
2172 * contract, making the old contract no longer valid for this
2173 * instance.
2174 */
2175 if (ctid != inst->ri_i.i_primary_ctid) {
2176 MUTEX_UNLOCK(&inst->ri_lock);
2177 inst = NULL;
2178 }
2179 }
2180 return (inst);
2181 }
2182
2183 /*
2184 * void contract_action()
2185 * Take action on contract events.
2186 */
2187 static void
2188 contract_action(scf_handle_t *h, restarter_inst_t *inst, ctid_t id,
2189 uint32_t type)
2190 {
2191 const char *fmri = inst->ri_i.i_fmri;
2192
2193 assert(MUTEX_HELD(&inst->ri_lock));
2194
2195 /*
2196 * If startd has stopped this contract, there is no need to
2197 * stop it again.
2198 */
2199 if (inst->ri_i.i_primary_ctid > 0 &&
2200 inst->ri_i.i_primary_ctid_stopped)
2201 return;
2202
2203 if ((type & (CT_PR_EV_EMPTY | CT_PR_EV_CORE | CT_PR_EV_SIGNAL
2204 | CT_PR_EV_HWERR)) == 0) {
2205 /*
2206 * There shouldn't be other events, since that's not how we set
2207 * the terms. Thus, just log an error and drive on.
2208 */
2209 log_framework(LOG_NOTICE,
2210 "%s: contract %ld received unexpected critical event "
2211 "(%d)\n", fmri, id, type);
2212 return;
2213 }
2214
2215 assert(instance_in_transition(inst) == 0);
2216
2217 if (instance_is_wait_style(inst)) {
2218 /*
2219 * We ignore all events; if they impact the
2220 * process we're monitoring, then the
2221 * wait_thread will stop the instance.
2222 */
2223 log_framework(LOG_DEBUG,
2224 "%s: ignoring contract event on wait-style service\n",
2225 fmri);
2226 } else {
2227 /*
2228 * A CT_PR_EV_EMPTY event is an RSTOP_EXIT request.
2229 */
2230 switch (type) {
2231 case CT_PR_EV_EMPTY:
2232 (void) stop_instance(h, inst, RSTOP_EXIT);
2233 break;
2234 case CT_PR_EV_CORE:
2235 (void) stop_instance(h, inst, RSTOP_CORE);
2236 break;
2237 case CT_PR_EV_SIGNAL:
2238 (void) stop_instance(h, inst, RSTOP_SIGNAL);
2239 break;
2240 case CT_PR_EV_HWERR:
2241 (void) stop_instance(h, inst, RSTOP_HWERR);
2242 break;
2243 }
2244 }
2245 }
2246
2247 /*
2248 * void *restarter_contract_event_thread(void *)
2249 * Listens to the process contract bundle for critical events, taking action
2250 * on events from contracts we know we are responsible for.
2251 */
2252 /*ARGSUSED*/
2253 static void *
2254 restarter_contracts_event_thread(void *unused)
2255 {
2256 int fd, err;
2257 scf_handle_t *local_handle;
2258
2259 /*
2260 * Await graph load completion. That is, stop here, until we've scanned
2261 * the repository for contract - instance associations.
2262 */
2263 MUTEX_LOCK(&st->st_load_lock);
2264 while (!(st->st_load_complete && st->st_load_instances == 0))
2265 (void) pthread_cond_wait(&st->st_load_cv, &st->st_load_lock);
2266 MUTEX_UNLOCK(&st->st_load_lock);
2267
2268 /*
2269 * This is a new thread, and thus, gets its own handle
2270 * to the repository.
2271 */
2272 if ((local_handle = libscf_handle_create_bound(SCF_VERSION)) == NULL)
2273 uu_die("Unable to bind a new repository handle: %s\n",
2274 scf_strerror(scf_error()));
2275
2276 fd = open64(CTFS_ROOT "/process/pbundle", O_RDONLY);
2277 if (fd == -1)
2278 uu_die("process bundle open failed");
2279
2280 /*
2281 * Make sure we get all events (including those generated by configd
2282 * before this thread was started).
2283 */
2284 err = ct_event_reset(fd);
2285 assert(err == 0);
2286
2287 for (;;) {
2288 int efd, sfd;
2289 ct_evthdl_t ev;
2290 uint32_t type;
2291 ctevid_t evid;
2292 ct_stathdl_t status;
2293 ctid_t ctid;
2294 restarter_inst_t *inst;
2295 uint64_t cookie;
2296
2297 if (err = ct_event_read_critical(fd, &ev)) {
2298 log_error(LOG_WARNING,
2299 "Error reading next contract event: %s",
2300 strerror(err));
2301 continue;
2302 }
2303
2304 evid = ct_event_get_evid(ev);
2305 ctid = ct_event_get_ctid(ev);
2306 type = ct_event_get_type(ev);
2307
2308 /* Fetch cookie. */
2309 if ((sfd = contract_open(ctid, "process", "status", O_RDONLY))
2310 < 0) {
2311 ct_event_free(ev);
2312 continue;
2313 }
2314
2315 if (err = ct_status_read(sfd, CTD_COMMON, &status)) {
2316 log_framework(LOG_WARNING, "Could not get status for "
2317 "contract %ld: %s\n", ctid, strerror(err));
2318
2319 startd_close(sfd);
2320 ct_event_free(ev);
2321 continue;
2322 }
2323
2324 cookie = ct_status_get_cookie(status);
2325
2326 log_framework(LOG_DEBUG, "Received event %d for ctid %ld "
2327 "cookie %lld\n", type, ctid, cookie);
2328
2329 ct_status_free(status);
2330
2331 startd_close(sfd);
2332
2333 /*
2334 * svc.configd(1M) restart handling performed by the
2335 * fork_configd_thread. We don't acknowledge, as that thread
2336 * will do so.
2337 */
2338 if (cookie == CONFIGD_COOKIE) {
2339 ct_event_free(ev);
2340 continue;
2341 }
2342
2343 inst = NULL;
2344 if (storing_contract != 0 &&
2345 (inst = contract_to_inst(ctid)) == NULL) {
2346 /*
2347 * This can happen for two reasons:
2348 * - method_run() has not yet stored the
2349 * the contract into the internal hash table.
2350 * - we receive an EMPTY event for an abandoned
2351 * contract.
2352 * If there is any contract in the process of
2353 * being stored into the hash table then re-read
2354 * the event later.
2355 */
2356 log_framework(LOG_DEBUG,
2357 "Reset event %d for unknown "
2358 "contract id %ld\n", type, ctid);
2359
2360 /* don't go too fast */
2361 (void) poll(NULL, 0, 100);
2362
2363 (void) ct_event_reset(fd);
2364 ct_event_free(ev);
2365 continue;
2366 }
2367
2368 /*
2369 * Do not call contract_to_inst() again if first
2370 * call succeeded.
2371 */
2372 if (inst == NULL)
2373 inst = contract_to_inst(ctid);
2374 if (inst == NULL) {
2375 /*
2376 * This can happen if we receive an EMPTY
2377 * event for an abandoned contract.
2378 */
2379 log_framework(LOG_DEBUG,
2380 "Received event %d for unknown contract id "
2381 "%ld\n", type, ctid);
2382 } else {
2383 log_framework(LOG_DEBUG,
2384 "Received event %d for contract id "
2385 "%ld (%s)\n", type, ctid,
2386 inst->ri_i.i_fmri);
2387
2388 contract_action(local_handle, inst, ctid, type);
2389
2390 MUTEX_UNLOCK(&inst->ri_lock);
2391 }
2392
2393 efd = contract_open(ct_event_get_ctid(ev), "process", "ctl",
2394 O_WRONLY);
2395 if (efd != -1) {
2396 (void) ct_ctl_ack(efd, evid);
2397 startd_close(efd);
2398 }
2399
2400 ct_event_free(ev);
2401
2402 }
2403
2404 /*NOTREACHED*/
2405 return (NULL);
2406 }
2407
2408 /*
2409 * Timeout queue, processed by restarter_timeouts_event_thread().
2410 */
2411 timeout_queue_t *timeouts;
2412 static uu_list_pool_t *timeout_pool;
2413
2414 typedef struct timeout_update {
2415 pthread_mutex_t tu_lock;
2416 pthread_cond_t tu_cv;
2417 int tu_wakeup;
2418 } timeout_update_t;
2419
2420 timeout_update_t *tu;
2421
2422 static const char *timeout_ovr_svcs[] = {
2423 "svc:/system/manifest-import:default",
2424 "svc:/network/initial:default",
2425 "svc:/network/service:default",
2426 "svc:/system/rmtmpfiles:default",
2427 "svc:/network/loopback:default",
2428 "svc:/network/physical:default",
2429 "svc:/system/device/local:default",
2430 "svc:/system/filesystem/usr:default",
2431 "svc:/system/filesystem/minimal:default",
2432 "svc:/system/filesystem/local:default",
2433 NULL
2434 };
2435
2436 int
2437 is_timeout_ovr(restarter_inst_t *inst)
2438 {
2439 int i;
2440
2441 for (i = 0; timeout_ovr_svcs[i] != NULL; ++i) {
2442 if (strcmp(inst->ri_i.i_fmri, timeout_ovr_svcs[i]) == 0) {
2443 log_instance(inst, B_TRUE, "Timeout override by "
2444 "svc.startd. Using infinite timeout.");
2445 return (1);
2446 }
2447 }
2448
2449 return (0);
2450 }
2451
2452 /*ARGSUSED*/
2453 static int
2454 timeout_compare(const void *lc_arg, const void *rc_arg, void *private)
2455 {
2456 hrtime_t t1 = ((const timeout_entry_t *)lc_arg)->te_timeout;
2457 hrtime_t t2 = ((const timeout_entry_t *)rc_arg)->te_timeout;
2458
2459 if (t1 > t2)
2460 return (1);
2461 else if (t1 < t2)
2462 return (-1);
2463 return (0);
2464 }
2465
2466 void
2467 timeout_init()
2468 {
2469 timeouts = startd_zalloc(sizeof (timeout_queue_t));
2470
2471 (void) pthread_mutex_init(&timeouts->tq_lock, &mutex_attrs);
2472
2473 timeout_pool = startd_list_pool_create("timeouts",
2474 sizeof (timeout_entry_t), offsetof(timeout_entry_t, te_link),
2475 timeout_compare, UU_LIST_POOL_DEBUG);
2476 assert(timeout_pool != NULL);
2477
2478 timeouts->tq_list = startd_list_create(timeout_pool,
2479 timeouts, UU_LIST_SORTED);
2480 assert(timeouts->tq_list != NULL);
2481
2482 tu = startd_zalloc(sizeof (timeout_update_t));
2483 (void) pthread_cond_init(&tu->tu_cv, NULL);
2484 (void) pthread_mutex_init(&tu->tu_lock, &mutex_attrs);
2485 }
2486
2487 void
2488 timeout_insert(restarter_inst_t *inst, ctid_t cid, uint64_t timeout_sec)
2489 {
2490 hrtime_t now, timeout;
2491 timeout_entry_t *entry;
2492 uu_list_index_t idx;
2493
2494 assert(MUTEX_HELD(&inst->ri_lock));
2495
2496 now = gethrtime();
2497
2498 /*
2499 * If we overflow LLONG_MAX, we're never timing out anyways, so
2500 * just return.
2501 */
2502 if (timeout_sec >= (LLONG_MAX - now) / 1000000000LL) {
2503 log_instance(inst, B_TRUE, "timeout_seconds too large, "
2504 "treating as infinite.");
2505 return;
2506 }
2507
2508 /* hrtime is in nanoseconds. Convert timeout_sec. */
2509 timeout = now + (timeout_sec * 1000000000LL);
2510
2511 entry = startd_alloc(sizeof (timeout_entry_t));
2512 entry->te_timeout = timeout;
2513 entry->te_ctid = cid;
2514 entry->te_fmri = safe_strdup(inst->ri_i.i_fmri);
2515 entry->te_logstem = safe_strdup(inst->ri_logstem);
2516 entry->te_fired = 0;
2517 /* Insert the calculated timeout time onto the queue. */
2518 MUTEX_LOCK(&timeouts->tq_lock);
2519 (void) uu_list_find(timeouts->tq_list, entry, NULL, &idx);
2520 uu_list_node_init(entry, &entry->te_link, timeout_pool);
2521 uu_list_insert(timeouts->tq_list, entry, idx);
2522 MUTEX_UNLOCK(&timeouts->tq_lock);
2523
2524 assert(inst->ri_timeout == NULL);
2525 inst->ri_timeout = entry;
2526
2527 MUTEX_LOCK(&tu->tu_lock);
2528 tu->tu_wakeup = 1;
2529 (void) pthread_cond_broadcast(&tu->tu_cv);
2530 MUTEX_UNLOCK(&tu->tu_lock);
2531 }
2532
2533
2534 void
2535 timeout_remove(restarter_inst_t *inst, ctid_t cid)
2536 {
2537 assert(MUTEX_HELD(&inst->ri_lock));
2538
2539 if (inst->ri_timeout == NULL)
2540 return;
2541
2542 assert(inst->ri_timeout->te_ctid == cid);
2543
2544 MUTEX_LOCK(&timeouts->tq_lock);
2545 uu_list_remove(timeouts->tq_list, inst->ri_timeout);
2546 MUTEX_UNLOCK(&timeouts->tq_lock);
2547
2548 free(inst->ri_timeout->te_fmri);
2549 free(inst->ri_timeout->te_logstem);
2550 startd_free(inst->ri_timeout, sizeof (timeout_entry_t));
2551 inst->ri_timeout = NULL;
2552 }
2553
2554 static int
2555 timeout_now()
2556 {
2557 timeout_entry_t *e;
2558 hrtime_t now;
2559 int ret;
2560
2561 now = gethrtime();
2562
2563 /*
2564 * Walk through the (sorted) timeouts list. While the timeout
2565 * at the head of the list is <= the current time, kill the
2566 * method.
2567 */
2568 MUTEX_LOCK(&timeouts->tq_lock);
2569
2570 for (e = uu_list_first(timeouts->tq_list);
2571 e != NULL && e->te_timeout <= now;
2572 e = uu_list_next(timeouts->tq_list, e)) {
2573 log_framework(LOG_WARNING, "%s: Method or service exit timed "
2574 "out. Killing contract %ld.\n", e->te_fmri, e->te_ctid);
2575 log_instance_fmri(e->te_fmri, e->te_logstem, B_TRUE,
2576 "Method or service exit timed out. Killing contract %ld.",
2577 e->te_ctid);
2578 e->te_fired = 1;
2579 (void) contract_kill(e->te_ctid, SIGKILL, e->te_fmri);
2580 }
2581
2582 if (uu_list_numnodes(timeouts->tq_list) > 0)
2583 ret = 0;
2584 else
2585 ret = -1;
2586
2587 MUTEX_UNLOCK(&timeouts->tq_lock);
2588
2589 return (ret);
2590 }
2591
2592 /*
2593 * void *restarter_timeouts_event_thread(void *)
2594 * Responsible for monitoring the method timeouts. This thread must
2595 * be started before any methods are called.
2596 */
2597 /*ARGSUSED*/
2598 static void *
2599 restarter_timeouts_event_thread(void *unused)
2600 {
2601 /*
2602 * Timeouts are entered on a priority queue, which is processed by
2603 * this thread. As timeouts are specified in seconds, we'll do
2604 * the necessary processing every second, as long as the queue
2605 * is not empty.
2606 */
2607
2608 /*CONSTCOND*/
2609 while (1) {
2610 /*
2611 * As long as the timeout list isn't empty, process it
2612 * every second.
2613 */
2614 if (timeout_now() == 0) {
2615 (void) sleep(1);
2616 continue;
2617 }
2618
2619 /* The list is empty, wait until we have more timeouts. */
2620 MUTEX_LOCK(&tu->tu_lock);
2621
2622 while (tu->tu_wakeup == 0)
2623 (void) pthread_cond_wait(&tu->tu_cv, &tu->tu_lock);
2624
2625 tu->tu_wakeup = 0;
2626 MUTEX_UNLOCK(&tu->tu_lock);
2627 }
2628
2629 return (NULL);
2630 }
2631
2632 void
2633 restarter_start()
2634 {
2635 (void) startd_thread_create(restarter_timeouts_event_thread, NULL);
2636 (void) startd_thread_create(restarter_event_thread, NULL);
2637 (void) startd_thread_create(restarter_contracts_event_thread, NULL);
2638 (void) startd_thread_create(wait_thread, NULL);
2639 }
2640
2641
2642 void
2643 restarter_init()
2644 {
2645 restarter_instance_pool = startd_list_pool_create("restarter_instances",
2646 sizeof (restarter_inst_t), offsetof(restarter_inst_t,
2647 ri_link), restarter_instance_compare, UU_LIST_POOL_DEBUG);
2648 (void) memset(&instance_list, 0, sizeof (instance_list));
2649
2650 (void) pthread_mutex_init(&instance_list.ril_lock, &mutex_attrs);
2651 instance_list.ril_instance_list = startd_list_create(
2652 restarter_instance_pool, &instance_list, UU_LIST_SORTED);
2653
2654 restarter_queue_pool = startd_list_pool_create(
2655 "restarter_instance_queue", sizeof (restarter_instance_qentry_t),
2656 offsetof(restarter_instance_qentry_t, riq_link), NULL,
2657 UU_LIST_POOL_DEBUG);
2658
2659 contract_list_pool = startd_list_pool_create(
2660 "contract_list", sizeof (contract_entry_t),
2661 offsetof(contract_entry_t, ce_link), NULL,
2662 UU_LIST_POOL_DEBUG);
2663 contract_hash_init();
2664
2665 log_framework(LOG_DEBUG, "Initialized restarter\n");
2666 }