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 */
25
26 /*
27 * graph.c - master restarter graph engine
28 *
29 * The graph engine keeps a dependency graph of all service instances on the
30 * system, as recorded in the repository. It decides when services should
31 * be brought up or down based on service states and dependencies and sends
32 * commands to restarters to effect any changes. It also executes
33 * administrator commands sent by svcadm via the repository.
34 *
35 * The graph is stored in uu_list_t *dgraph and its vertices are
36 * graph_vertex_t's, each of which has a name and an integer id unique to
37 * its name (see dict.c). A vertex's type attribute designates the type
38 * of object it represents: GVT_INST for service instances, GVT_SVC for
39 * service objects (since service instances may depend on another service,
40 * rather than service instance), GVT_FILE for files (which services may
41 * depend on), and GVT_GROUP for dependencies on multiple objects. GVT_GROUP
42 * vertices are necessary because dependency lists may have particular
43 * grouping types (require any, require all, optional, or exclude) and
44 * event-propagation characteristics.
45 *
46 * The initial graph is built by libscf_populate_graph() invoking
47 * dgraph_add_instance() for each instance in the repository. The function
48 * adds a GVT_SVC vertex for the service if one does not already exist, adds
49 * a GVT_INST vertex named by the FMRI of the instance, and sets up the edges.
50 * The resulting web of vertices & edges associated with an instance's vertex
51 * includes
52 *
53 * - an edge from the GVT_SVC vertex for the instance's service
54 *
55 * - an edge to the GVT_INST vertex of the instance's resarter, if its
56 * restarter is not svc.startd
57 *
58 * - edges from other GVT_INST vertices if the instance is a restarter
59 *
60 * - for each dependency property group in the instance's "running"
61 * snapshot, an edge to a GVT_GROUP vertex named by the FMRI of the
62 * instance and the name of the property group
63 *
64 * - for each value of the "entities" property in each dependency property
65 * group, an edge from the corresponding GVT_GROUP vertex to a
66 * GVT_INST, GVT_SVC, or GVT_FILE vertex
67 *
68 * - edges from GVT_GROUP vertices for each dependent instance
69 *
70 * After the edges are set up the vertex's GV_CONFIGURED flag is set. If
71 * there are problems, or if a service is mentioned in a dependency but does
72 * not exist in the repository, the GV_CONFIGURED flag will be clear.
73 *
74 * The graph and all of its vertices are protected by the dgraph_lock mutex.
75 * See restarter.c for more information.
76 *
77 * The properties of an instance fall into two classes: immediate and
78 * snapshotted. Immediate properties should have an immediate effect when
79 * changed. Snapshotted properties should be read from a snapshot, so they
80 * only change when the snapshot changes. The immediate properties used by
81 * the graph engine are general/enabled, general/restarter, and the properties
82 * in the restarter_actions property group. Since they are immediate, they
83 * are not read out of a snapshot. The snapshotted properties used by the
84 * graph engine are those in the property groups with type "dependency" and
85 * are read out of the "running" snapshot. The "running" snapshot is created
86 * by the the graph engine as soon as possible, and it is updated, along with
87 * in-core copies of the data (dependency information for the graph engine) on
88 * receipt of the refresh command from svcadm. In addition, the graph engine
89 * updates the "start" snapshot from the "running" snapshot whenever a service
90 * comes online.
91 *
92 * When a DISABLE event is requested by the administrator, svc.startd shutdown
93 * the dependents first before shutting down the requested service.
94 * In graph_enable_by_vertex, we create a subtree that contains the dependent
95 * vertices by marking those vertices with the GV_TOOFFLINE flag. And we mark
96 * the vertex to disable with the GV_TODISABLE flag. Once the tree is created,
97 * we send the _ADMIN_DISABLE event to the leaves. The leaves will then
98 * transition from STATE_ONLINE/STATE_DEGRADED to STATE_OFFLINE/STATE_MAINT.
99 * In gt_enter_offline and gt_enter_maint if the vertex was in a subtree then
100 * we clear the GV_TOOFFLINE flag and walk the dependencies to offline the new
101 * exposed leaves. We do the same until we reach the last leaf (the one with
102 * the GV_TODISABLE flag). If the vertex to disable is also part of a larger
103 * subtree (eg. multiple DISABLE events on vertices in the same subtree) then
104 * once the first vertex is disabled (GV_TODISABLE flag is removed), we
105 * continue to propagate the offline event to the vertex's dependencies.
106 *
107 *
108 * SMF state transition notifications
109 *
110 * When an instance of a service managed by SMF changes state, svc.startd may
111 * publish a GPEC sysevent. All transitions to or from maintenance, a
112 * transition cause by a hardware error will generate an event.
113 * Other transitions will generate an event if there exist notification
114 * parameter for that transition. Notification parameters are stored in the
115 * SMF repository for the service/instance they refer to. System-wide
116 * notification parameters are stored in the global instance.
117 * svc.startd can be told to send events for all SMF state transitions despite
118 * of notification parameters by setting options/info_events_all to true in
119 * restarter:default
120 *
121 * The set of transitions that generate events is cached in the
122 * dgraph_vertex_t gv_stn_tset for service/instance and in the global
123 * stn_global for the system-wide set. They are re-read when instances are
124 * refreshed.
125 *
126 * The GPEC events published by svc.startd are consumed by fmd(1M). After
127 * processing these events, fmd(1M) publishes the processed events to
128 * notification agents. The notification agents read the notification
129 * parameters from the SMF repository through libscf(3LIB) interfaces and send
130 * the notification, or not, based on those parameters.
131 *
132 * Subscription and publishing to the GPEC channels is done with the
133 * libfmevent(3LIB) wrappers fmev_[r]publish_*() and
134 * fmev_shdl_(un)subscribe().
135 *
136 */
137
138 #include <sys/uadmin.h>
139 #include <sys/wait.h>
140
141 #include <assert.h>
142 #include <errno.h>
143 #include <fcntl.h>
144 #include <fm/libfmevent.h>
145 #include <libscf.h>
146 #include <libscf_priv.h>
147 #include <librestart.h>
148 #include <libuutil.h>
149 #include <locale.h>
150 #include <poll.h>
151 #include <pthread.h>
152 #include <signal.h>
153 #include <stddef.h>
154 #include <stdio.h>
155 #include <stdlib.h>
156 #include <string.h>
157 #include <strings.h>
158 #include <sys/statvfs.h>
159 #include <sys/uadmin.h>
160 #include <zone.h>
161 #if defined(__i386)
162 #include <libgrubmgmt.h>
163 #endif /* __i386 */
164
165 #include "startd.h"
166 #include "protocol.h"
167
168
169 #define MILESTONE_NONE ((graph_vertex_t *)1)
170
171 #define CONSOLE_LOGIN_FMRI "svc:/system/console-login:default"
172 #define FS_MINIMAL_FMRI "svc:/system/filesystem/minimal:default"
173
174 #define VERTEX_REMOVED 0 /* vertex has been freed */
175 #define VERTEX_INUSE 1 /* vertex is still in use */
176
177 #define IS_ENABLED(v) ((v)->gv_flags & (GV_ENABLED | GV_ENBLD_NOOVR))
178
179 /*
180 * stn_global holds the tset for the system wide notification parameters.
181 * It is updated on refresh of svc:/system/svc/global:default
182 *
183 * There are two assumptions that relax the need for a mutex:
184 * 1. 32-bit value assignments are atomic
185 * 2. Its value is consumed only in one point at
186 * dgraph_state_transition_notify(). There are no test and set races.
187 *
188 * If either assumption is broken, we'll need a mutex to synchronize
189 * access to stn_global
190 */
191 int32_t stn_global;
192 /*
193 * info_events_all holds a flag to override notification parameters and send
194 * Information events for all state transitions.
195 * same about the need of a mutex here.
196 */
197 int info_events_all;
198
199 /*
200 * Services in these states are not considered 'down' by the
201 * milestone/shutdown code.
202 */
203 #define up_state(state) ((state) == RESTARTER_STATE_ONLINE || \
204 (state) == RESTARTER_STATE_DEGRADED || \
205 (state) == RESTARTER_STATE_OFFLINE)
206
207 #define is_depgrp_bypassed(v) ((v->gv_type == GVT_GROUP) && \
208 ((v->gv_depgroup == DEPGRP_EXCLUDE_ALL) || \
209 (v->gv_depgroup == DEPGRP_OPTIONAL_ALL) || \
210 (v->gv_restart < RERR_RESTART)))
211
212 static uu_list_pool_t *graph_edge_pool, *graph_vertex_pool;
213 static uu_list_t *dgraph;
214 static pthread_mutex_t dgraph_lock;
215
216 /*
217 * milestone indicates the current subgraph. When NULL, it is the entire
218 * graph. When MILESTONE_NONE, it is the empty graph. Otherwise, it is all
219 * services on which the target vertex depends.
220 */
221 static graph_vertex_t *milestone = NULL;
222 static boolean_t initial_milestone_set = B_FALSE;
223 static pthread_cond_t initial_milestone_cv = PTHREAD_COND_INITIALIZER;
224
225 /* protected by dgraph_lock */
226 static boolean_t sulogin_thread_running = B_FALSE;
227 static boolean_t sulogin_running = B_FALSE;
228 static boolean_t console_login_ready = B_FALSE;
229
230 /* Number of services to come down to complete milestone transition. */
231 static uint_t non_subgraph_svcs;
232
233 /*
234 * These variables indicate what should be done when we reach the milestone
235 * target milestone, i.e., when non_subgraph_svcs == 0. They are acted upon in
236 * dgraph_set_instance_state().
237 */
238 static int halting = -1;
239 static boolean_t go_single_user_mode = B_FALSE;
240 static boolean_t go_to_level1 = B_FALSE;
241
242 /*
243 * Tracks when we started halting.
244 */
245 static time_t halting_time = 0;
246
247 /*
248 * This tracks the legacy runlevel to ensure we signal init and manage
249 * utmpx entries correctly.
250 */
251 static char current_runlevel = '\0';
252
253 /* Number of single user threads currently running */
254 static pthread_mutex_t single_user_thread_lock;
255 static int single_user_thread_count = 0;
256
257 /* Statistics for dependency cycle-checking */
258 static u_longlong_t dep_inserts = 0;
259 static u_longlong_t dep_cycle_ns = 0;
260 static u_longlong_t dep_insert_ns = 0;
261
262
263 static const char * const emsg_invalid_restarter =
264 "Transitioning %s to maintenance, restarter FMRI %s is invalid "
265 "(see 'svcs -xv' for details).\n";
266 static const char * const console_login_fmri = CONSOLE_LOGIN_FMRI;
267 static const char * const single_user_fmri = SCF_MILESTONE_SINGLE_USER;
268 static const char * const multi_user_fmri = SCF_MILESTONE_MULTI_USER;
269 static const char * const multi_user_svr_fmri = SCF_MILESTONE_MULTI_USER_SERVER;
270
271
272 /*
273 * These services define the system being "up". If none of them can come
274 * online, then we will run sulogin on the console. Note that the install ones
275 * are for the miniroot and when installing CDs after the first. can_come_up()
276 * does the decision making, and an sulogin_thread() runs sulogin, which can be
277 * started by dgraph_set_instance_state() or single_user_thread().
278 *
279 * NOTE: can_come_up() relies on SCF_MILESTONE_SINGLE_USER being the first
280 * entry, which is only used when booting_to_single_user (boot -s) is set.
281 * This is because when doing a "boot -s", sulogin is started from specials.c
282 * after milestone/single-user comes online, for backwards compatibility.
283 * In this case, SCF_MILESTONE_SINGLE_USER needs to be part of up_svcs
284 * to ensure sulogin will be spawned if milestone/single-user cannot be reached.
285 */
286 static const char * const up_svcs[] = {
287 SCF_MILESTONE_SINGLE_USER,
288 CONSOLE_LOGIN_FMRI,
289 "svc:/system/install-setup:default",
290 "svc:/system/install:default",
291 NULL
292 };
293
294 /* This array must have an element for each non-NULL element of up_svcs[]. */
295 static graph_vertex_t *up_svcs_p[] = { NULL, NULL, NULL, NULL };
296
297 /* These are for seed repository magic. See can_come_up(). */
298 static const char * const manifest_import = SCF_INSTANCE_MI;
299 static graph_vertex_t *manifest_import_p = NULL;
300
301
302 static char target_milestone_as_runlevel(void);
303 static void graph_runlevel_changed(char rl, int online);
304 static int dgraph_set_milestone(const char *, scf_handle_t *, boolean_t);
305 static boolean_t should_be_in_subgraph(graph_vertex_t *v);
306 static int mark_subtree(graph_edge_t *, void *);
307 static boolean_t insubtree_dependents_down(graph_vertex_t *);
308
309 /*
310 * graph_vertex_compare()
311 * This function can compare either int *id or * graph_vertex_t *gv
312 * values, as the vertex id is always the first element of a
313 * graph_vertex structure.
314 */
315 /* ARGSUSED */
316 static int
graph_vertex_compare(const void * lc_arg,const void * rc_arg,void * private)317 graph_vertex_compare(const void *lc_arg, const void *rc_arg, void *private)
318 {
319 int lc_id = ((const graph_vertex_t *)lc_arg)->gv_id;
320 int rc_id = *(int *)rc_arg;
321
322 if (lc_id > rc_id)
323 return (1);
324 if (lc_id < rc_id)
325 return (-1);
326 return (0);
327 }
328
329 void
graph_init()330 graph_init()
331 {
332 graph_edge_pool = startd_list_pool_create("graph_edges",
333 sizeof (graph_edge_t), offsetof(graph_edge_t, ge_link), NULL,
334 UU_LIST_POOL_DEBUG);
335 assert(graph_edge_pool != NULL);
336
337 graph_vertex_pool = startd_list_pool_create("graph_vertices",
338 sizeof (graph_vertex_t), offsetof(graph_vertex_t, gv_link),
339 graph_vertex_compare, UU_LIST_POOL_DEBUG);
340 assert(graph_vertex_pool != NULL);
341
342 (void) pthread_mutex_init(&dgraph_lock, &mutex_attrs);
343 (void) pthread_mutex_init(&single_user_thread_lock, &mutex_attrs);
344 dgraph = startd_list_create(graph_vertex_pool, NULL, UU_LIST_SORTED);
345 assert(dgraph != NULL);
346
347 if (!st->st_initial)
348 current_runlevel = utmpx_get_runlevel();
349
350 log_framework(LOG_DEBUG, "Initialized graph\n");
351 }
352
353 static graph_vertex_t *
vertex_get_by_name(const char * name)354 vertex_get_by_name(const char *name)
355 {
356 int id;
357
358 assert(MUTEX_HELD(&dgraph_lock));
359
360 id = dict_lookup_byname(name);
361 if (id == -1)
362 return (NULL);
363
364 return (uu_list_find(dgraph, &id, NULL, NULL));
365 }
366
367 static graph_vertex_t *
vertex_get_by_id(int id)368 vertex_get_by_id(int id)
369 {
370 assert(MUTEX_HELD(&dgraph_lock));
371
372 if (id == -1)
373 return (NULL);
374
375 return (uu_list_find(dgraph, &id, NULL, NULL));
376 }
377
378 /*
379 * Creates a new vertex with the given name, adds it to the graph, and returns
380 * a pointer to it. The graph lock must be held by this thread on entry.
381 */
382 static graph_vertex_t *
graph_add_vertex(const char * name)383 graph_add_vertex(const char *name)
384 {
385 int id;
386 graph_vertex_t *v;
387 void *p;
388 uu_list_index_t idx;
389
390 assert(MUTEX_HELD(&dgraph_lock));
391
392 id = dict_insert(name);
393
394 v = startd_zalloc(sizeof (*v));
395
396 v->gv_id = id;
397
398 v->gv_name = startd_alloc(strlen(name) + 1);
399 (void) strcpy(v->gv_name, name);
400
401 v->gv_dependencies = startd_list_create(graph_edge_pool, v, 0);
402 v->gv_dependents = startd_list_create(graph_edge_pool, v, 0);
403
404 p = uu_list_find(dgraph, &id, NULL, &idx);
405 assert(p == NULL);
406
407 uu_list_node_init(v, &v->gv_link, graph_vertex_pool);
408 uu_list_insert(dgraph, v, idx);
409
410 return (v);
411 }
412
413 /*
414 * Removes v from the graph and frees it. The graph should be locked by this
415 * thread, and v should have no edges associated with it.
416 */
417 static void
graph_remove_vertex(graph_vertex_t * v)418 graph_remove_vertex(graph_vertex_t *v)
419 {
420 assert(MUTEX_HELD(&dgraph_lock));
421
422 assert(uu_list_numnodes(v->gv_dependencies) == 0);
423 assert(uu_list_numnodes(v->gv_dependents) == 0);
424 assert(v->gv_refs == 0);
425
426 startd_free(v->gv_name, strlen(v->gv_name) + 1);
427 uu_list_destroy(v->gv_dependencies);
428 uu_list_destroy(v->gv_dependents);
429 uu_list_remove(dgraph, v);
430
431 startd_free(v, sizeof (graph_vertex_t));
432 }
433
434 static void
graph_add_edge(graph_vertex_t * fv,graph_vertex_t * tv)435 graph_add_edge(graph_vertex_t *fv, graph_vertex_t *tv)
436 {
437 graph_edge_t *e, *re;
438 int r;
439
440 assert(MUTEX_HELD(&dgraph_lock));
441
442 e = startd_alloc(sizeof (graph_edge_t));
443 re = startd_alloc(sizeof (graph_edge_t));
444
445 e->ge_parent = fv;
446 e->ge_vertex = tv;
447
448 re->ge_parent = tv;
449 re->ge_vertex = fv;
450
451 uu_list_node_init(e, &e->ge_link, graph_edge_pool);
452 r = uu_list_insert_before(fv->gv_dependencies, NULL, e);
453 assert(r == 0);
454
455 uu_list_node_init(re, &re->ge_link, graph_edge_pool);
456 r = uu_list_insert_before(tv->gv_dependents, NULL, re);
457 assert(r == 0);
458 }
459
460 static void
graph_remove_edge(graph_vertex_t * v,graph_vertex_t * dv)461 graph_remove_edge(graph_vertex_t *v, graph_vertex_t *dv)
462 {
463 graph_edge_t *e;
464
465 for (e = uu_list_first(v->gv_dependencies);
466 e != NULL;
467 e = uu_list_next(v->gv_dependencies, e)) {
468 if (e->ge_vertex == dv) {
469 uu_list_remove(v->gv_dependencies, e);
470 startd_free(e, sizeof (graph_edge_t));
471 break;
472 }
473 }
474
475 for (e = uu_list_first(dv->gv_dependents);
476 e != NULL;
477 e = uu_list_next(dv->gv_dependents, e)) {
478 if (e->ge_vertex == v) {
479 uu_list_remove(dv->gv_dependents, e);
480 startd_free(e, sizeof (graph_edge_t));
481 break;
482 }
483 }
484 }
485
486 static void
remove_inst_vertex(graph_vertex_t * v)487 remove_inst_vertex(graph_vertex_t *v)
488 {
489 graph_edge_t *e;
490 graph_vertex_t *sv;
491 int i;
492
493 assert(MUTEX_HELD(&dgraph_lock));
494 assert(uu_list_numnodes(v->gv_dependents) == 1);
495 assert(uu_list_numnodes(v->gv_dependencies) == 0);
496 assert(v->gv_refs == 0);
497 assert((v->gv_flags & GV_CONFIGURED) == 0);
498
499 e = uu_list_first(v->gv_dependents);
500 sv = e->ge_vertex;
501 graph_remove_edge(sv, v);
502
503 for (i = 0; up_svcs[i] != NULL; ++i) {
504 if (up_svcs_p[i] == v)
505 up_svcs_p[i] = NULL;
506 }
507
508 if (manifest_import_p == v)
509 manifest_import_p = NULL;
510
511 graph_remove_vertex(v);
512
513 if (uu_list_numnodes(sv->gv_dependencies) == 0 &&
514 uu_list_numnodes(sv->gv_dependents) == 0 &&
515 sv->gv_refs == 0)
516 graph_remove_vertex(sv);
517 }
518
519 static void
graph_walk_dependents(graph_vertex_t * v,void (* func)(graph_vertex_t *,void *),void * arg)520 graph_walk_dependents(graph_vertex_t *v, void (*func)(graph_vertex_t *, void *),
521 void *arg)
522 {
523 graph_edge_t *e;
524
525 for (e = uu_list_first(v->gv_dependents);
526 e != NULL;
527 e = uu_list_next(v->gv_dependents, e))
528 func(e->ge_vertex, arg);
529 }
530
531 static void
graph_walk_dependencies(graph_vertex_t * v,void (* func)(graph_vertex_t *,void *),void * arg)532 graph_walk_dependencies(graph_vertex_t *v, void (*func)(graph_vertex_t *,
533 void *), void *arg)
534 {
535 graph_edge_t *e;
536
537 assert(MUTEX_HELD(&dgraph_lock));
538
539 for (e = uu_list_first(v->gv_dependencies);
540 e != NULL;
541 e = uu_list_next(v->gv_dependencies, e)) {
542
543 func(e->ge_vertex, arg);
544 }
545 }
546
547 /*
548 * Generic graph walking function.
549 *
550 * Given a vertex, this function will walk either dependencies
551 * (WALK_DEPENDENCIES) or dependents (WALK_DEPENDENTS) of a vertex recursively
552 * for the entire graph. It will avoid cycles and never visit the same vertex
553 * twice.
554 *
555 * We avoid traversing exclusion dependencies, because they are allowed to
556 * create cycles in the graph. When propagating satisfiability, there is no
557 * need to walk exclusion dependencies because exclude_all_satisfied() doesn't
558 * test for satisfiability.
559 *
560 * The walker takes two callbacks. The first is called before examining the
561 * dependents of each vertex. The second is called on each vertex after
562 * examining its dependents. This allows is_path_to() to construct a path only
563 * after the target vertex has been found.
564 */
565 typedef enum {
566 WALK_DEPENDENTS,
567 WALK_DEPENDENCIES
568 } graph_walk_dir_t;
569
570 typedef int (*graph_walk_cb_t)(graph_vertex_t *, void *);
571
572 typedef struct graph_walk_info {
573 graph_walk_dir_t gi_dir;
574 uchar_t *gi_visited; /* vertex bitmap */
575 int (*gi_pre)(graph_vertex_t *, void *);
576 void (*gi_post)(graph_vertex_t *, void *);
577 void *gi_arg; /* callback arg */
578 int gi_ret; /* return value */
579 } graph_walk_info_t;
580
581 static int
graph_walk_recurse(graph_edge_t * e,graph_walk_info_t * gip)582 graph_walk_recurse(graph_edge_t *e, graph_walk_info_t *gip)
583 {
584 uu_list_t *list;
585 int r;
586 graph_vertex_t *v = e->ge_vertex;
587 int i;
588 uint_t b;
589
590 i = v->gv_id / 8;
591 b = 1 << (v->gv_id % 8);
592
593 /*
594 * Check to see if we've visited this vertex already.
595 */
596 if (gip->gi_visited[i] & b)
597 return (UU_WALK_NEXT);
598
599 gip->gi_visited[i] |= b;
600
601 /*
602 * Don't follow exclusions.
603 */
604 if (v->gv_type == GVT_GROUP && v->gv_depgroup == DEPGRP_EXCLUDE_ALL)
605 return (UU_WALK_NEXT);
606
607 /*
608 * Call pre-visit callback. If this doesn't terminate the walk,
609 * continue search.
610 */
611 if ((gip->gi_ret = gip->gi_pre(v, gip->gi_arg)) == UU_WALK_NEXT) {
612 /*
613 * Recurse using appropriate list.
614 */
615 if (gip->gi_dir == WALK_DEPENDENTS)
616 list = v->gv_dependents;
617 else
618 list = v->gv_dependencies;
619
620 r = uu_list_walk(list, (uu_walk_fn_t *)graph_walk_recurse,
621 gip, 0);
622 assert(r == 0);
623 }
624
625 /*
626 * Callbacks must return either UU_WALK_NEXT or UU_WALK_DONE.
627 */
628 assert(gip->gi_ret == UU_WALK_NEXT || gip->gi_ret == UU_WALK_DONE);
629
630 /*
631 * If given a post-callback, call the function for every vertex.
632 */
633 if (gip->gi_post != NULL)
634 (void) gip->gi_post(v, gip->gi_arg);
635
636 /*
637 * Preserve the callback's return value. If the callback returns
638 * UU_WALK_DONE, then we propagate that to the caller in order to
639 * terminate the walk.
640 */
641 return (gip->gi_ret);
642 }
643
644 static void
graph_walk(graph_vertex_t * v,graph_walk_dir_t dir,int (* pre)(graph_vertex_t *,void *),void (* post)(graph_vertex_t *,void *),void * arg)645 graph_walk(graph_vertex_t *v, graph_walk_dir_t dir,
646 int (*pre)(graph_vertex_t *, void *),
647 void (*post)(graph_vertex_t *, void *), void *arg)
648 {
649 graph_walk_info_t gi;
650 graph_edge_t fake;
651 size_t sz = dictionary->dict_new_id / 8 + 1;
652
653 gi.gi_visited = startd_zalloc(sz);
654 gi.gi_pre = pre;
655 gi.gi_post = post;
656 gi.gi_arg = arg;
657 gi.gi_dir = dir;
658 gi.gi_ret = 0;
659
660 /*
661 * Fake up an edge for the first iteration
662 */
663 fake.ge_vertex = v;
664 (void) graph_walk_recurse(&fake, &gi);
665
666 startd_free(gi.gi_visited, sz);
667 }
668
669 typedef struct child_search {
670 int id; /* id of vertex to look for */
671 uint_t depth; /* recursion depth */
672 /*
673 * While the vertex is not found, path is NULL. After the search, if
674 * the vertex was found then path should point to a -1-terminated
675 * array of vertex id's which constitute the path to the vertex.
676 */
677 int *path;
678 } child_search_t;
679
680 static int
child_pre(graph_vertex_t * v,void * arg)681 child_pre(graph_vertex_t *v, void *arg)
682 {
683 child_search_t *cs = arg;
684
685 cs->depth++;
686
687 if (v->gv_id == cs->id) {
688 cs->path = startd_alloc((cs->depth + 1) * sizeof (int));
689 cs->path[cs->depth] = -1;
690 return (UU_WALK_DONE);
691 }
692
693 return (UU_WALK_NEXT);
694 }
695
696 static void
child_post(graph_vertex_t * v,void * arg)697 child_post(graph_vertex_t *v, void *arg)
698 {
699 child_search_t *cs = arg;
700
701 cs->depth--;
702
703 if (cs->path != NULL)
704 cs->path[cs->depth] = v->gv_id;
705 }
706
707 /*
708 * Look for a path from from to to. If one exists, returns a pointer to
709 * a NULL-terminated array of pointers to the vertices along the path. If
710 * there is no path, returns NULL.
711 */
712 static int *
is_path_to(graph_vertex_t * from,graph_vertex_t * to)713 is_path_to(graph_vertex_t *from, graph_vertex_t *to)
714 {
715 child_search_t cs;
716
717 cs.id = to->gv_id;
718 cs.depth = 0;
719 cs.path = NULL;
720
721 graph_walk(from, WALK_DEPENDENCIES, child_pre, child_post, &cs);
722
723 return (cs.path);
724 }
725
726 /*
727 * Given an array of int's as returned by is_path_to, allocates a string of
728 * their names joined by newlines. Returns the size of the allocated buffer
729 * in *sz and frees path.
730 */
731 static void
path_to_str(int * path,char ** cpp,size_t * sz)732 path_to_str(int *path, char **cpp, size_t *sz)
733 {
734 int i;
735 graph_vertex_t *v;
736 size_t allocd, new_allocd;
737 char *new, *name;
738
739 assert(MUTEX_HELD(&dgraph_lock));
740 assert(path[0] != -1);
741
742 allocd = 1;
743 *cpp = startd_alloc(1);
744 (*cpp)[0] = '\0';
745
746 for (i = 0; path[i] != -1; ++i) {
747 name = NULL;
748
749 v = vertex_get_by_id(path[i]);
750
751 if (v == NULL)
752 name = "<deleted>";
753 else if (v->gv_type == GVT_INST || v->gv_type == GVT_SVC)
754 name = v->gv_name;
755
756 if (name != NULL) {
757 new_allocd = allocd + strlen(name) + 1;
758 new = startd_alloc(new_allocd);
759 (void) strcpy(new, *cpp);
760 (void) strcat(new, name);
761 (void) strcat(new, "\n");
762
763 startd_free(*cpp, allocd);
764
765 *cpp = new;
766 allocd = new_allocd;
767 }
768 }
769
770 startd_free(path, sizeof (int) * (i + 1));
771
772 *sz = allocd;
773 }
774
775
776 /*
777 * This function along with run_sulogin() implements an exclusion relationship
778 * between system/console-login and sulogin. run_sulogin() will fail if
779 * system/console-login is online, and the graph engine should call
780 * graph_clogin_start() to bring system/console-login online, which defers the
781 * start if sulogin is running.
782 */
783 static void
graph_clogin_start(graph_vertex_t * v)784 graph_clogin_start(graph_vertex_t *v)
785 {
786 assert(MUTEX_HELD(&dgraph_lock));
787
788 if (sulogin_running)
789 console_login_ready = B_TRUE;
790 else
791 vertex_send_event(v, RESTARTER_EVENT_TYPE_START);
792 }
793
794 static void
graph_su_start(graph_vertex_t * v)795 graph_su_start(graph_vertex_t *v)
796 {
797 /*
798 * /etc/inittab used to have the initial /sbin/rcS as a 'sysinit'
799 * entry with a runlevel of 'S', before jumping to the final
800 * target runlevel (as set in initdefault). We mimic that legacy
801 * behavior here.
802 */
803 utmpx_set_runlevel('S', '0', B_FALSE);
804 vertex_send_event(v, RESTARTER_EVENT_TYPE_START);
805 }
806
807 static void
graph_post_su_online(void)808 graph_post_su_online(void)
809 {
810 graph_runlevel_changed('S', 1);
811 }
812
813 static void
graph_post_su_disable(void)814 graph_post_su_disable(void)
815 {
816 graph_runlevel_changed('S', 0);
817 }
818
819 static void
graph_post_mu_online(void)820 graph_post_mu_online(void)
821 {
822 graph_runlevel_changed('2', 1);
823 }
824
825 static void
graph_post_mu_disable(void)826 graph_post_mu_disable(void)
827 {
828 graph_runlevel_changed('2', 0);
829 }
830
831 static void
graph_post_mus_online(void)832 graph_post_mus_online(void)
833 {
834 graph_runlevel_changed('3', 1);
835 }
836
837 static void
graph_post_mus_disable(void)838 graph_post_mus_disable(void)
839 {
840 graph_runlevel_changed('3', 0);
841 }
842
843 static struct special_vertex_info {
844 const char *name;
845 void (*start_f)(graph_vertex_t *);
846 void (*post_online_f)(void);
847 void (*post_disable_f)(void);
848 } special_vertices[] = {
849 { CONSOLE_LOGIN_FMRI, graph_clogin_start, NULL, NULL },
850 { SCF_MILESTONE_SINGLE_USER, graph_su_start,
851 graph_post_su_online, graph_post_su_disable },
852 { SCF_MILESTONE_MULTI_USER, NULL,
853 graph_post_mu_online, graph_post_mu_disable },
854 { SCF_MILESTONE_MULTI_USER_SERVER, NULL,
855 graph_post_mus_online, graph_post_mus_disable },
856 { NULL },
857 };
858
859
860 void
vertex_send_event(graph_vertex_t * v,restarter_event_type_t e)861 vertex_send_event(graph_vertex_t *v, restarter_event_type_t e)
862 {
863 switch (e) {
864 case RESTARTER_EVENT_TYPE_ADD_INSTANCE:
865 assert(v->gv_state == RESTARTER_STATE_UNINIT);
866
867 MUTEX_LOCK(&st->st_load_lock);
868 st->st_load_instances++;
869 MUTEX_UNLOCK(&st->st_load_lock);
870 break;
871
872 case RESTARTER_EVENT_TYPE_ENABLE:
873 log_framework(LOG_DEBUG, "Enabling %s.\n", v->gv_name);
874 assert(v->gv_state == RESTARTER_STATE_UNINIT ||
875 v->gv_state == RESTARTER_STATE_DISABLED ||
876 v->gv_state == RESTARTER_STATE_MAINT);
877 break;
878
879 case RESTARTER_EVENT_TYPE_DISABLE:
880 case RESTARTER_EVENT_TYPE_ADMIN_DISABLE:
881 log_framework(LOG_DEBUG, "Disabling %s.\n", v->gv_name);
882 assert(v->gv_state != RESTARTER_STATE_DISABLED);
883 break;
884
885 case RESTARTER_EVENT_TYPE_STOP_RESET:
886 case RESTARTER_EVENT_TYPE_STOP:
887 log_framework(LOG_DEBUG, "Stopping %s.\n", v->gv_name);
888 assert(v->gv_state == RESTARTER_STATE_DEGRADED ||
889 v->gv_state == RESTARTER_STATE_ONLINE);
890 break;
891
892 case RESTARTER_EVENT_TYPE_START:
893 log_framework(LOG_DEBUG, "Starting %s.\n", v->gv_name);
894 assert(v->gv_state == RESTARTER_STATE_OFFLINE);
895 break;
896
897 case RESTARTER_EVENT_TYPE_REMOVE_INSTANCE:
898 case RESTARTER_EVENT_TYPE_ADMIN_DEGRADED:
899 case RESTARTER_EVENT_TYPE_ADMIN_REFRESH:
900 case RESTARTER_EVENT_TYPE_ADMIN_RESTART:
901 case RESTARTER_EVENT_TYPE_ADMIN_MAINT_OFF:
902 case RESTARTER_EVENT_TYPE_ADMIN_MAINT_ON:
903 case RESTARTER_EVENT_TYPE_ADMIN_MAINT_ON_IMMEDIATE:
904 case RESTARTER_EVENT_TYPE_DEPENDENCY_CYCLE:
905 case RESTARTER_EVENT_TYPE_INVALID_DEPENDENCY:
906 break;
907
908 default:
909 #ifndef NDEBUG
910 uu_warn("%s:%d: Bad event %d.\n", __FILE__, __LINE__, e);
911 #endif
912 abort();
913 }
914
915 restarter_protocol_send_event(v->gv_name, v->gv_restarter_channel, e,
916 v->gv_reason);
917 }
918
919 static void
graph_unset_restarter(graph_vertex_t * v)920 graph_unset_restarter(graph_vertex_t *v)
921 {
922 assert(MUTEX_HELD(&dgraph_lock));
923 assert(v->gv_flags & GV_CONFIGURED);
924
925 vertex_send_event(v, RESTARTER_EVENT_TYPE_REMOVE_INSTANCE);
926
927 if (v->gv_restarter_id != -1) {
928 graph_vertex_t *rv;
929
930 rv = vertex_get_by_id(v->gv_restarter_id);
931 graph_remove_edge(v, rv);
932 }
933
934 v->gv_restarter_id = -1;
935 v->gv_restarter_channel = NULL;
936 }
937
938 /*
939 * Return VERTEX_REMOVED when the vertex passed in argument is deleted from the
940 * dgraph otherwise return VERTEX_INUSE.
941 */
942 static int
free_if_unrefed(graph_vertex_t * v)943 free_if_unrefed(graph_vertex_t *v)
944 {
945 assert(MUTEX_HELD(&dgraph_lock));
946
947 if (v->gv_refs > 0)
948 return (VERTEX_INUSE);
949
950 if (v->gv_type == GVT_SVC &&
951 uu_list_numnodes(v->gv_dependents) == 0 &&
952 uu_list_numnodes(v->gv_dependencies) == 0) {
953 graph_remove_vertex(v);
954 return (VERTEX_REMOVED);
955 } else if (v->gv_type == GVT_INST &&
956 (v->gv_flags & GV_CONFIGURED) == 0 &&
957 uu_list_numnodes(v->gv_dependents) == 1 &&
958 uu_list_numnodes(v->gv_dependencies) == 0) {
959 remove_inst_vertex(v);
960 return (VERTEX_REMOVED);
961 }
962
963 return (VERTEX_INUSE);
964 }
965
966 static void
delete_depgroup(graph_vertex_t * v)967 delete_depgroup(graph_vertex_t *v)
968 {
969 graph_edge_t *e;
970 graph_vertex_t *dv;
971
972 assert(MUTEX_HELD(&dgraph_lock));
973 assert(v->gv_type == GVT_GROUP);
974 assert(uu_list_numnodes(v->gv_dependents) == 0);
975
976 while ((e = uu_list_first(v->gv_dependencies)) != NULL) {
977 dv = e->ge_vertex;
978
979 graph_remove_edge(v, dv);
980
981 switch (dv->gv_type) {
982 case GVT_INST: /* instance dependency */
983 case GVT_SVC: /* service dependency */
984 (void) free_if_unrefed(dv);
985 break;
986
987 case GVT_FILE: /* file dependency */
988 assert(uu_list_numnodes(dv->gv_dependencies) == 0);
989 if (uu_list_numnodes(dv->gv_dependents) == 0)
990 graph_remove_vertex(dv);
991 break;
992
993 default:
994 #ifndef NDEBUG
995 uu_warn("%s:%d: Unexpected node type %d", __FILE__,
996 __LINE__, dv->gv_type);
997 #endif
998 abort();
999 }
1000 }
1001
1002 graph_remove_vertex(v);
1003 }
1004
1005 static int
delete_instance_deps_cb(graph_edge_t * e,void ** ptrs)1006 delete_instance_deps_cb(graph_edge_t *e, void **ptrs)
1007 {
1008 graph_vertex_t *v = ptrs[0];
1009 boolean_t delete_restarter_dep = (boolean_t)ptrs[1];
1010 graph_vertex_t *dv;
1011
1012 dv = e->ge_vertex;
1013
1014 /*
1015 * We have four possibilities here:
1016 * - GVT_INST: restarter
1017 * - GVT_GROUP - GVT_INST: instance dependency
1018 * - GVT_GROUP - GVT_SVC - GV_INST: service dependency
1019 * - GVT_GROUP - GVT_FILE: file dependency
1020 */
1021 switch (dv->gv_type) {
1022 case GVT_INST: /* restarter */
1023 assert(dv->gv_id == v->gv_restarter_id);
1024 if (delete_restarter_dep)
1025 graph_remove_edge(v, dv);
1026 break;
1027
1028 case GVT_GROUP: /* pg dependency */
1029 graph_remove_edge(v, dv);
1030 delete_depgroup(dv);
1031 break;
1032
1033 case GVT_FILE:
1034 /* These are currently not direct dependencies */
1035
1036 default:
1037 #ifndef NDEBUG
1038 uu_warn("%s:%d: Bad vertex type %d.\n", __FILE__, __LINE__,
1039 dv->gv_type);
1040 #endif
1041 abort();
1042 }
1043
1044 return (UU_WALK_NEXT);
1045 }
1046
1047 static void
delete_instance_dependencies(graph_vertex_t * v,boolean_t delete_restarter_dep)1048 delete_instance_dependencies(graph_vertex_t *v, boolean_t delete_restarter_dep)
1049 {
1050 void *ptrs[2];
1051 int r;
1052
1053 assert(MUTEX_HELD(&dgraph_lock));
1054 assert(v->gv_type == GVT_INST);
1055
1056 ptrs[0] = v;
1057 ptrs[1] = (void *)delete_restarter_dep;
1058
1059 r = uu_list_walk(v->gv_dependencies,
1060 (uu_walk_fn_t *)delete_instance_deps_cb, &ptrs, UU_WALK_ROBUST);
1061 assert(r == 0);
1062 }
1063
1064 /*
1065 * int graph_insert_vertex_unconfigured()
1066 * Insert a vertex without sending any restarter events. If the vertex
1067 * already exists or creation is successful, return a pointer to it in *vp.
1068 *
1069 * If type is not GVT_GROUP, dt can remain unset.
1070 *
1071 * Returns 0, EEXIST, or EINVAL if the arguments are invalid (i.e., fmri
1072 * doesn't agree with type, or type doesn't agree with dt).
1073 */
1074 static int
graph_insert_vertex_unconfigured(const char * fmri,gv_type_t type,depgroup_type_t dt,restarter_error_t rt,graph_vertex_t ** vp)1075 graph_insert_vertex_unconfigured(const char *fmri, gv_type_t type,
1076 depgroup_type_t dt, restarter_error_t rt, graph_vertex_t **vp)
1077 {
1078 int r;
1079 int i;
1080
1081 assert(MUTEX_HELD(&dgraph_lock));
1082
1083 switch (type) {
1084 case GVT_SVC:
1085 case GVT_INST:
1086 if (strncmp(fmri, "svc:", sizeof ("svc:") - 1) != 0)
1087 return (EINVAL);
1088 break;
1089
1090 case GVT_FILE:
1091 if (strncmp(fmri, "file:", sizeof ("file:") - 1) != 0)
1092 return (EINVAL);
1093 break;
1094
1095 case GVT_GROUP:
1096 if (dt <= 0 || rt < 0)
1097 return (EINVAL);
1098 break;
1099
1100 default:
1101 #ifndef NDEBUG
1102 uu_warn("%s:%d: Unknown type %d.\n", __FILE__, __LINE__, type);
1103 #endif
1104 abort();
1105 }
1106
1107 *vp = vertex_get_by_name(fmri);
1108 if (*vp != NULL)
1109 return (EEXIST);
1110
1111 *vp = graph_add_vertex(fmri);
1112
1113 (*vp)->gv_type = type;
1114 (*vp)->gv_depgroup = dt;
1115 (*vp)->gv_restart = rt;
1116
1117 (*vp)->gv_flags = 0;
1118 (*vp)->gv_state = RESTARTER_STATE_NONE;
1119
1120 for (i = 0; special_vertices[i].name != NULL; ++i) {
1121 if (strcmp(fmri, special_vertices[i].name) == 0) {
1122 (*vp)->gv_start_f = special_vertices[i].start_f;
1123 (*vp)->gv_post_online_f =
1124 special_vertices[i].post_online_f;
1125 (*vp)->gv_post_disable_f =
1126 special_vertices[i].post_disable_f;
1127 break;
1128 }
1129 }
1130
1131 (*vp)->gv_restarter_id = -1;
1132 (*vp)->gv_restarter_channel = 0;
1133
1134 if (type == GVT_INST) {
1135 char *sfmri;
1136 graph_vertex_t *sv;
1137
1138 sfmri = inst_fmri_to_svc_fmri(fmri);
1139 sv = vertex_get_by_name(sfmri);
1140 if (sv == NULL) {
1141 r = graph_insert_vertex_unconfigured(sfmri, GVT_SVC, 0,
1142 0, &sv);
1143 assert(r == 0);
1144 }
1145 startd_free(sfmri, max_scf_fmri_size);
1146
1147 graph_add_edge(sv, *vp);
1148 }
1149
1150 /*
1151 * If this vertex is in the subgraph, mark it as so, for both
1152 * GVT_INST and GVT_SERVICE verteces.
1153 * A GVT_SERVICE vertex can only be in the subgraph if another instance
1154 * depends on it, in which case it's already been added to the graph
1155 * and marked as in the subgraph (by refresh_vertex()). If a
1156 * GVT_SERVICE vertex was freshly added (by the code above), it means
1157 * that it has no dependents, and cannot be in the subgraph.
1158 * Regardless of this, we still check that gv_flags includes
1159 * GV_INSUBGRAPH in the event that future behavior causes the above
1160 * code to add a GVT_SERVICE vertex which should be in the subgraph.
1161 */
1162
1163 (*vp)->gv_flags |= (should_be_in_subgraph(*vp)? GV_INSUBGRAPH : 0);
1164
1165 return (0);
1166 }
1167
1168 /*
1169 * Returns 0 on success or ELOOP if the dependency would create a cycle.
1170 */
1171 static int
graph_insert_dependency(graph_vertex_t * fv,graph_vertex_t * tv,int ** pathp)1172 graph_insert_dependency(graph_vertex_t *fv, graph_vertex_t *tv, int **pathp)
1173 {
1174 hrtime_t now;
1175
1176 assert(MUTEX_HELD(&dgraph_lock));
1177
1178 /* cycle detection */
1179 now = gethrtime();
1180
1181 /* Don't follow exclusions. */
1182 if (!(fv->gv_type == GVT_GROUP &&
1183 fv->gv_depgroup == DEPGRP_EXCLUDE_ALL)) {
1184 *pathp = is_path_to(tv, fv);
1185 if (*pathp)
1186 return (ELOOP);
1187 }
1188
1189 dep_cycle_ns += gethrtime() - now;
1190 ++dep_inserts;
1191 now = gethrtime();
1192
1193 graph_add_edge(fv, tv);
1194
1195 dep_insert_ns += gethrtime() - now;
1196
1197 /* Check if the dependency adds the "to" vertex to the subgraph */
1198 tv->gv_flags |= (should_be_in_subgraph(tv) ? GV_INSUBGRAPH : 0);
1199
1200 return (0);
1201 }
1202
1203 static int
inst_running(graph_vertex_t * v)1204 inst_running(graph_vertex_t *v)
1205 {
1206 assert(v->gv_type == GVT_INST);
1207
1208 if (v->gv_state == RESTARTER_STATE_ONLINE ||
1209 v->gv_state == RESTARTER_STATE_DEGRADED)
1210 return (1);
1211
1212 return (0);
1213 }
1214
1215 /*
1216 * The dependency evaluation functions return
1217 * 1 - dependency satisfied
1218 * 0 - dependency unsatisfied
1219 * -1 - dependency unsatisfiable (without administrator intervention)
1220 *
1221 * The functions also take a boolean satbility argument. When true, the
1222 * functions may recurse in order to determine satisfiability.
1223 */
1224 static int require_any_satisfied(graph_vertex_t *, boolean_t);
1225 static int dependency_satisfied(graph_vertex_t *, boolean_t);
1226
1227 /*
1228 * A require_all dependency is unsatisfied if any elements are unsatisfied. It
1229 * is unsatisfiable if any elements are unsatisfiable.
1230 */
1231 static int
require_all_satisfied(graph_vertex_t * groupv,boolean_t satbility)1232 require_all_satisfied(graph_vertex_t *groupv, boolean_t satbility)
1233 {
1234 graph_edge_t *edge;
1235 int i;
1236 boolean_t any_unsatisfied;
1237
1238 if (uu_list_numnodes(groupv->gv_dependencies) == 0)
1239 return (1);
1240
1241 any_unsatisfied = B_FALSE;
1242
1243 for (edge = uu_list_first(groupv->gv_dependencies);
1244 edge != NULL;
1245 edge = uu_list_next(groupv->gv_dependencies, edge)) {
1246 i = dependency_satisfied(edge->ge_vertex, satbility);
1247 if (i == 1)
1248 continue;
1249
1250 log_framework2(LOG_DEBUG, DEBUG_DEPENDENCIES,
1251 "require_all(%s): %s is unsatisfi%s.\n", groupv->gv_name,
1252 edge->ge_vertex->gv_name, i == 0 ? "ed" : "able");
1253
1254 if (!satbility)
1255 return (0);
1256
1257 if (i == -1)
1258 return (-1);
1259
1260 any_unsatisfied = B_TRUE;
1261 }
1262
1263 return (any_unsatisfied ? 0 : 1);
1264 }
1265
1266 /*
1267 * A require_any dependency is satisfied if any element is satisfied. It is
1268 * satisfiable if any element is satisfiable.
1269 */
1270 static int
require_any_satisfied(graph_vertex_t * groupv,boolean_t satbility)1271 require_any_satisfied(graph_vertex_t *groupv, boolean_t satbility)
1272 {
1273 graph_edge_t *edge;
1274 int s;
1275 boolean_t satisfiable;
1276
1277 if (uu_list_numnodes(groupv->gv_dependencies) == 0)
1278 return (1);
1279
1280 satisfiable = B_FALSE;
1281
1282 for (edge = uu_list_first(groupv->gv_dependencies);
1283 edge != NULL;
1284 edge = uu_list_next(groupv->gv_dependencies, edge)) {
1285 s = dependency_satisfied(edge->ge_vertex, satbility);
1286
1287 if (s == 1)
1288 return (1);
1289
1290 log_framework2(LOG_DEBUG, DEBUG_DEPENDENCIES,
1291 "require_any(%s): %s is unsatisfi%s.\n",
1292 groupv->gv_name, edge->ge_vertex->gv_name,
1293 s == 0 ? "ed" : "able");
1294
1295 if (satbility && s == 0)
1296 satisfiable = B_TRUE;
1297 }
1298
1299 return (!satbility || satisfiable ? 0 : -1);
1300 }
1301
1302 /*
1303 * An optional_all dependency only considers elements which are configured,
1304 * enabled, and not in maintenance. If any are unsatisfied, then the dependency
1305 * is unsatisfied.
1306 *
1307 * Offline dependencies which are waiting for a dependency to come online are
1308 * unsatisfied. Offline dependences which cannot possibly come online
1309 * (unsatisfiable) are always considered satisfied.
1310 */
1311 static int
optional_all_satisfied(graph_vertex_t * groupv,boolean_t satbility)1312 optional_all_satisfied(graph_vertex_t *groupv, boolean_t satbility)
1313 {
1314 graph_edge_t *edge;
1315 graph_vertex_t *v;
1316 boolean_t any_qualified;
1317 boolean_t any_unsatisfied;
1318 int i;
1319
1320 any_qualified = B_FALSE;
1321 any_unsatisfied = B_FALSE;
1322
1323 for (edge = uu_list_first(groupv->gv_dependencies);
1324 edge != NULL;
1325 edge = uu_list_next(groupv->gv_dependencies, edge)) {
1326 v = edge->ge_vertex;
1327
1328 switch (v->gv_type) {
1329 case GVT_INST:
1330 /* Skip missing or disabled instances */
1331 if ((v->gv_flags & (GV_CONFIGURED | GV_ENABLED)) !=
1332 (GV_CONFIGURED | GV_ENABLED))
1333 continue;
1334
1335 if (v->gv_state == RESTARTER_STATE_MAINT)
1336 continue;
1337
1338 if (v->gv_flags & GV_TOOFFLINE)
1339 continue;
1340
1341 any_qualified = B_TRUE;
1342 if (v->gv_state == RESTARTER_STATE_OFFLINE) {
1343 /*
1344 * For offline dependencies, treat unsatisfiable
1345 * as satisfied.
1346 */
1347 i = dependency_satisfied(v, B_TRUE);
1348 if (i == -1)
1349 i = 1;
1350 } else if (v->gv_state == RESTARTER_STATE_DISABLED) {
1351 /*
1352 * The service is enabled, but hasn't
1353 * transitioned out of disabled yet. Treat it
1354 * as unsatisfied (not unsatisfiable).
1355 */
1356 i = 0;
1357 } else {
1358 i = dependency_satisfied(v, satbility);
1359 }
1360 break;
1361
1362 case GVT_FILE:
1363 any_qualified = B_TRUE;
1364 i = dependency_satisfied(v, satbility);
1365
1366 break;
1367
1368 case GVT_SVC: {
1369 boolean_t svc_any_qualified;
1370 boolean_t svc_satisfied;
1371 boolean_t svc_satisfiable;
1372 graph_vertex_t *v2;
1373 graph_edge_t *e2;
1374
1375 svc_any_qualified = B_FALSE;
1376 svc_satisfied = B_FALSE;
1377 svc_satisfiable = B_FALSE;
1378
1379 for (e2 = uu_list_first(v->gv_dependencies);
1380 e2 != NULL;
1381 e2 = uu_list_next(v->gv_dependencies, e2)) {
1382 v2 = e2->ge_vertex;
1383 assert(v2->gv_type == GVT_INST);
1384
1385 if ((v2->gv_flags &
1386 (GV_CONFIGURED | GV_ENABLED)) !=
1387 (GV_CONFIGURED | GV_ENABLED))
1388 continue;
1389
1390 if (v2->gv_state == RESTARTER_STATE_MAINT)
1391 continue;
1392
1393 if (v2->gv_flags & GV_TOOFFLINE)
1394 continue;
1395
1396 svc_any_qualified = B_TRUE;
1397
1398 if (v2->gv_state == RESTARTER_STATE_OFFLINE) {
1399 /*
1400 * For offline dependencies, treat
1401 * unsatisfiable as satisfied.
1402 */
1403 i = dependency_satisfied(v2, B_TRUE);
1404 if (i == -1)
1405 i = 1;
1406 } else if (v2->gv_state ==
1407 RESTARTER_STATE_DISABLED) {
1408 i = 0;
1409 } else {
1410 i = dependency_satisfied(v2, satbility);
1411 }
1412
1413 if (i == 1) {
1414 svc_satisfied = B_TRUE;
1415 break;
1416 }
1417 if (i == 0)
1418 svc_satisfiable = B_TRUE;
1419 }
1420
1421 if (!svc_any_qualified)
1422 continue;
1423 any_qualified = B_TRUE;
1424 if (svc_satisfied) {
1425 i = 1;
1426 } else if (svc_satisfiable) {
1427 i = 0;
1428 } else {
1429 i = -1;
1430 }
1431 break;
1432 }
1433
1434 case GVT_GROUP:
1435 default:
1436 #ifndef NDEBUG
1437 uu_warn("%s:%d: Unexpected vertex type %d.\n", __FILE__,
1438 __LINE__, v->gv_type);
1439 #endif
1440 abort();
1441 }
1442
1443 if (i == 1)
1444 continue;
1445
1446 log_framework2(LOG_DEBUG, DEBUG_DEPENDENCIES,
1447 "optional_all(%s): %s is unsatisfi%s.\n", groupv->gv_name,
1448 v->gv_name, i == 0 ? "ed" : "able");
1449
1450 if (!satbility)
1451 return (0);
1452 if (i == -1)
1453 return (-1);
1454 any_unsatisfied = B_TRUE;
1455 }
1456
1457 if (!any_qualified)
1458 return (1);
1459
1460 return (any_unsatisfied ? 0 : 1);
1461 }
1462
1463 /*
1464 * An exclude_all dependency is unsatisfied if any non-service element is
1465 * satisfied or any service instance which is configured, enabled, and not in
1466 * maintenance is satisfied. Usually when unsatisfied, it is also
1467 * unsatisfiable.
1468 */
1469 #define LOG_EXCLUDE(u, v) \
1470 log_framework2(LOG_DEBUG, DEBUG_DEPENDENCIES, \
1471 "exclude_all(%s): %s is satisfied.\n", \
1472 (u)->gv_name, (v)->gv_name)
1473
1474 /* ARGSUSED */
1475 static int
exclude_all_satisfied(graph_vertex_t * groupv,boolean_t satbility)1476 exclude_all_satisfied(graph_vertex_t *groupv, boolean_t satbility)
1477 {
1478 graph_edge_t *edge, *e2;
1479 graph_vertex_t *v, *v2;
1480
1481 for (edge = uu_list_first(groupv->gv_dependencies);
1482 edge != NULL;
1483 edge = uu_list_next(groupv->gv_dependencies, edge)) {
1484 v = edge->ge_vertex;
1485
1486 switch (v->gv_type) {
1487 case GVT_INST:
1488 if ((v->gv_flags & GV_CONFIGURED) == 0)
1489 continue;
1490
1491 switch (v->gv_state) {
1492 case RESTARTER_STATE_ONLINE:
1493 case RESTARTER_STATE_DEGRADED:
1494 LOG_EXCLUDE(groupv, v);
1495 return (v->gv_flags & GV_ENABLED ? -1 : 0);
1496
1497 case RESTARTER_STATE_OFFLINE:
1498 case RESTARTER_STATE_UNINIT:
1499 LOG_EXCLUDE(groupv, v);
1500 return (0);
1501
1502 case RESTARTER_STATE_DISABLED:
1503 case RESTARTER_STATE_MAINT:
1504 continue;
1505
1506 default:
1507 #ifndef NDEBUG
1508 uu_warn("%s:%d: Unexpected vertex state %d.\n",
1509 __FILE__, __LINE__, v->gv_state);
1510 #endif
1511 abort();
1512 }
1513 /* NOTREACHED */
1514
1515 case GVT_SVC:
1516 break;
1517
1518 case GVT_FILE:
1519 if (!file_ready(v))
1520 continue;
1521 LOG_EXCLUDE(groupv, v);
1522 return (-1);
1523
1524 case GVT_GROUP:
1525 default:
1526 #ifndef NDEBUG
1527 uu_warn("%s:%d: Unexpected vertex type %d.\n", __FILE__,
1528 __LINE__, v->gv_type);
1529 #endif
1530 abort();
1531 }
1532
1533 /* v represents a service */
1534 if (uu_list_numnodes(v->gv_dependencies) == 0)
1535 continue;
1536
1537 for (e2 = uu_list_first(v->gv_dependencies);
1538 e2 != NULL;
1539 e2 = uu_list_next(v->gv_dependencies, e2)) {
1540 v2 = e2->ge_vertex;
1541 assert(v2->gv_type == GVT_INST);
1542
1543 if ((v2->gv_flags & GV_CONFIGURED) == 0)
1544 continue;
1545
1546 switch (v2->gv_state) {
1547 case RESTARTER_STATE_ONLINE:
1548 case RESTARTER_STATE_DEGRADED:
1549 LOG_EXCLUDE(groupv, v2);
1550 return (v2->gv_flags & GV_ENABLED ? -1 : 0);
1551
1552 case RESTARTER_STATE_OFFLINE:
1553 case RESTARTER_STATE_UNINIT:
1554 LOG_EXCLUDE(groupv, v2);
1555 return (0);
1556
1557 case RESTARTER_STATE_DISABLED:
1558 case RESTARTER_STATE_MAINT:
1559 continue;
1560
1561 default:
1562 #ifndef NDEBUG
1563 uu_warn("%s:%d: Unexpected vertex type %d.\n",
1564 __FILE__, __LINE__, v2->gv_type);
1565 #endif
1566 abort();
1567 }
1568 }
1569 }
1570
1571 return (1);
1572 }
1573
1574 /*
1575 * int instance_satisfied()
1576 * Determine if all the dependencies are satisfied for the supplied instance
1577 * vertex. Return 1 if they are, 0 if they aren't, and -1 if they won't be
1578 * without administrator intervention.
1579 */
1580 static int
instance_satisfied(graph_vertex_t * v,boolean_t satbility)1581 instance_satisfied(graph_vertex_t *v, boolean_t satbility)
1582 {
1583 assert(v->gv_type == GVT_INST);
1584 assert(!inst_running(v));
1585
1586 return (require_all_satisfied(v, satbility));
1587 }
1588
1589 /*
1590 * Decide whether v can satisfy a dependency. v can either be a child of
1591 * a group vertex, or of an instance vertex.
1592 */
1593 static int
dependency_satisfied(graph_vertex_t * v,boolean_t satbility)1594 dependency_satisfied(graph_vertex_t *v, boolean_t satbility)
1595 {
1596 switch (v->gv_type) {
1597 case GVT_INST:
1598 if ((v->gv_flags & GV_CONFIGURED) == 0) {
1599 if (v->gv_flags & GV_DEATHROW) {
1600 /*
1601 * A dependency on an instance with GV_DEATHROW
1602 * flag is always considered as satisfied.
1603 */
1604 return (1);
1605 }
1606 return (-1);
1607 }
1608
1609 /*
1610 * Any vertex with the GV_TOOFFLINE flag set is guaranteed
1611 * to have its dependencies unsatisfiable.
1612 */
1613 if (v->gv_flags & GV_TOOFFLINE)
1614 return (-1);
1615
1616 switch (v->gv_state) {
1617 case RESTARTER_STATE_ONLINE:
1618 case RESTARTER_STATE_DEGRADED:
1619 return (1);
1620
1621 case RESTARTER_STATE_OFFLINE:
1622 if (!satbility)
1623 return (0);
1624 return (instance_satisfied(v, satbility) != -1 ?
1625 0 : -1);
1626
1627 case RESTARTER_STATE_DISABLED:
1628 case RESTARTER_STATE_MAINT:
1629 return (-1);
1630
1631 case RESTARTER_STATE_UNINIT:
1632 return (0);
1633
1634 default:
1635 #ifndef NDEBUG
1636 uu_warn("%s:%d: Unexpected vertex state %d.\n",
1637 __FILE__, __LINE__, v->gv_state);
1638 #endif
1639 abort();
1640 /* NOTREACHED */
1641 }
1642
1643 case GVT_SVC:
1644 if (uu_list_numnodes(v->gv_dependencies) == 0)
1645 return (-1);
1646 return (require_any_satisfied(v, satbility));
1647
1648 case GVT_FILE:
1649 /* i.e., we assume files will not be automatically generated */
1650 return (file_ready(v) ? 1 : -1);
1651
1652 case GVT_GROUP:
1653 break;
1654
1655 default:
1656 #ifndef NDEBUG
1657 uu_warn("%s:%d: Unexpected node type %d.\n", __FILE__, __LINE__,
1658 v->gv_type);
1659 #endif
1660 abort();
1661 /* NOTREACHED */
1662 }
1663
1664 switch (v->gv_depgroup) {
1665 case DEPGRP_REQUIRE_ANY:
1666 return (require_any_satisfied(v, satbility));
1667
1668 case DEPGRP_REQUIRE_ALL:
1669 return (require_all_satisfied(v, satbility));
1670
1671 case DEPGRP_OPTIONAL_ALL:
1672 return (optional_all_satisfied(v, satbility));
1673
1674 case DEPGRP_EXCLUDE_ALL:
1675 return (exclude_all_satisfied(v, satbility));
1676
1677 default:
1678 #ifndef NDEBUG
1679 uu_warn("%s:%d: Unknown dependency grouping %d.\n", __FILE__,
1680 __LINE__, v->gv_depgroup);
1681 #endif
1682 abort();
1683 }
1684 }
1685
1686 void
graph_start_if_satisfied(graph_vertex_t * v)1687 graph_start_if_satisfied(graph_vertex_t *v)
1688 {
1689 if (v->gv_state == RESTARTER_STATE_OFFLINE &&
1690 instance_satisfied(v, B_FALSE) == 1) {
1691 if (v->gv_start_f == NULL)
1692 vertex_send_event(v, RESTARTER_EVENT_TYPE_START);
1693 else
1694 v->gv_start_f(v);
1695 }
1696 }
1697
1698 /*
1699 * propagate_satbility()
1700 *
1701 * This function is used when the given vertex changes state in such a way that
1702 * one of its dependents may become unsatisfiable. This happens when an
1703 * instance transitions between offline -> online, or from !running ->
1704 * maintenance, as well as when an instance is removed from the graph.
1705 *
1706 * We have to walk all the dependents, since optional_all dependencies several
1707 * levels up could become (un)satisfied, instead of unsatisfiable. For example,
1708 *
1709 * +-----+ optional_all +-----+ require_all +-----+
1710 * | A |--------------->| B |-------------->| C |
1711 * +-----+ +-----+ +-----+
1712 *
1713 * offline -> maintenance
1714 *
1715 * If C goes into maintenance, it's not enough simply to check B. Because A has
1716 * an optional dependency, what was previously an unsatisfiable situation is now
1717 * satisfied (B will never come online, even though its state hasn't changed).
1718 *
1719 * Note that it's not necessary to continue examining dependents after reaching
1720 * an optional_all dependency. It's not possible for an optional_all dependency
1721 * to change satisfiability without also coming online, in which case we get a
1722 * start event and propagation continues naturally. However, it does no harm to
1723 * continue propagating satisfiability (as it is a relatively rare event), and
1724 * keeps the walker code simple and generic.
1725 */
1726 /*ARGSUSED*/
1727 static int
satbility_cb(graph_vertex_t * v,void * arg)1728 satbility_cb(graph_vertex_t *v, void *arg)
1729 {
1730 if (v->gv_type == GVT_INST)
1731 graph_start_if_satisfied(v);
1732
1733 return (UU_WALK_NEXT);
1734 }
1735
1736 static void
propagate_satbility(graph_vertex_t * v)1737 propagate_satbility(graph_vertex_t *v)
1738 {
1739 graph_walk(v, WALK_DEPENDENTS, satbility_cb, NULL, NULL);
1740 }
1741
1742 static void propagate_stop(graph_vertex_t *, void *);
1743
1744 /* ARGSUSED */
1745 static void
propagate_start(graph_vertex_t * v,void * arg)1746 propagate_start(graph_vertex_t *v, void *arg)
1747 {
1748 switch (v->gv_type) {
1749 case GVT_INST:
1750 graph_start_if_satisfied(v);
1751 break;
1752
1753 case GVT_GROUP:
1754 if (v->gv_depgroup == DEPGRP_EXCLUDE_ALL) {
1755 graph_walk_dependents(v, propagate_stop,
1756 (void *)RERR_RESTART);
1757 break;
1758 }
1759 /* FALLTHROUGH */
1760
1761 case GVT_SVC:
1762 graph_walk_dependents(v, propagate_start, NULL);
1763 break;
1764
1765 case GVT_FILE:
1766 #ifndef NDEBUG
1767 uu_warn("%s:%d: propagate_start() encountered GVT_FILE.\n",
1768 __FILE__, __LINE__);
1769 #endif
1770 abort();
1771 /* NOTREACHED */
1772
1773 default:
1774 #ifndef NDEBUG
1775 uu_warn("%s:%d: Unknown vertex type %d.\n", __FILE__, __LINE__,
1776 v->gv_type);
1777 #endif
1778 abort();
1779 }
1780 }
1781
1782 static void
propagate_stop(graph_vertex_t * v,void * arg)1783 propagate_stop(graph_vertex_t *v, void *arg)
1784 {
1785 graph_edge_t *e;
1786 graph_vertex_t *svc;
1787 restarter_error_t err = (restarter_error_t)arg;
1788
1789 switch (v->gv_type) {
1790 case GVT_INST:
1791 /* Restarter */
1792 if (err > RERR_NONE && inst_running(v)) {
1793 if (err == RERR_RESTART || err == RERR_REFRESH) {
1794 vertex_send_event(v,
1795 RESTARTER_EVENT_TYPE_STOP_RESET);
1796 } else {
1797 vertex_send_event(v, RESTARTER_EVENT_TYPE_STOP);
1798 }
1799 }
1800 break;
1801
1802 case GVT_SVC:
1803 graph_walk_dependents(v, propagate_stop, arg);
1804 break;
1805
1806 case GVT_FILE:
1807 #ifndef NDEBUG
1808 uu_warn("%s:%d: propagate_stop() encountered GVT_FILE.\n",
1809 __FILE__, __LINE__);
1810 #endif
1811 abort();
1812 /* NOTREACHED */
1813
1814 case GVT_GROUP:
1815 if (v->gv_depgroup == DEPGRP_EXCLUDE_ALL) {
1816 graph_walk_dependents(v, propagate_start, NULL);
1817 break;
1818 }
1819
1820 if (err == RERR_NONE || err > v->gv_restart)
1821 break;
1822
1823 assert(uu_list_numnodes(v->gv_dependents) == 1);
1824 e = uu_list_first(v->gv_dependents);
1825 svc = e->ge_vertex;
1826
1827 if (inst_running(svc)) {
1828 if (err == RERR_RESTART || err == RERR_REFRESH) {
1829 vertex_send_event(svc,
1830 RESTARTER_EVENT_TYPE_STOP_RESET);
1831 } else {
1832 vertex_send_event(svc,
1833 RESTARTER_EVENT_TYPE_STOP);
1834 }
1835 }
1836 break;
1837
1838 default:
1839 #ifndef NDEBUG
1840 uu_warn("%s:%d: Unknown vertex type %d.\n", __FILE__, __LINE__,
1841 v->gv_type);
1842 #endif
1843 abort();
1844 }
1845 }
1846
1847 void
offline_vertex(graph_vertex_t * v)1848 offline_vertex(graph_vertex_t *v)
1849 {
1850 scf_handle_t *h = libscf_handle_create_bound_loop();
1851 scf_instance_t *scf_inst = safe_scf_instance_create(h);
1852 scf_propertygroup_t *pg = safe_scf_pg_create(h);
1853 restarter_instance_state_t state, next_state;
1854 int r;
1855
1856 assert(v->gv_type == GVT_INST);
1857
1858 if (scf_inst == NULL)
1859 bad_error("safe_scf_instance_create", scf_error());
1860 if (pg == NULL)
1861 bad_error("safe_scf_pg_create", scf_error());
1862
1863 /* if the vertex is already going offline, return */
1864 rep_retry:
1865 if (scf_handle_decode_fmri(h, v->gv_name, NULL, NULL, scf_inst, NULL,
1866 NULL, SCF_DECODE_FMRI_EXACT) != 0) {
1867 switch (scf_error()) {
1868 case SCF_ERROR_CONNECTION_BROKEN:
1869 libscf_handle_rebind(h);
1870 goto rep_retry;
1871
1872 case SCF_ERROR_NOT_FOUND:
1873 scf_pg_destroy(pg);
1874 scf_instance_destroy(scf_inst);
1875 (void) scf_handle_unbind(h);
1876 scf_handle_destroy(h);
1877 return;
1878 }
1879 uu_die("Can't decode FMRI %s: %s\n", v->gv_name,
1880 scf_strerror(scf_error()));
1881 }
1882
1883 r = scf_instance_get_pg(scf_inst, SCF_PG_RESTARTER, pg);
1884 if (r != 0) {
1885 switch (scf_error()) {
1886 case SCF_ERROR_CONNECTION_BROKEN:
1887 libscf_handle_rebind(h);
1888 goto rep_retry;
1889
1890 case SCF_ERROR_NOT_SET:
1891 case SCF_ERROR_NOT_FOUND:
1892 scf_pg_destroy(pg);
1893 scf_instance_destroy(scf_inst);
1894 (void) scf_handle_unbind(h);
1895 scf_handle_destroy(h);
1896 return;
1897
1898 default:
1899 bad_error("scf_instance_get_pg", scf_error());
1900 }
1901 } else {
1902 r = libscf_read_states(pg, &state, &next_state);
1903 if (r == 0 && (next_state == RESTARTER_STATE_OFFLINE ||
1904 next_state == RESTARTER_STATE_DISABLED)) {
1905 log_framework(LOG_DEBUG,
1906 "%s: instance is already going down.\n",
1907 v->gv_name);
1908 scf_pg_destroy(pg);
1909 scf_instance_destroy(scf_inst);
1910 (void) scf_handle_unbind(h);
1911 scf_handle_destroy(h);
1912 return;
1913 }
1914 }
1915
1916 scf_pg_destroy(pg);
1917 scf_instance_destroy(scf_inst);
1918 (void) scf_handle_unbind(h);
1919 scf_handle_destroy(h);
1920
1921 vertex_send_event(v, RESTARTER_EVENT_TYPE_STOP_RESET);
1922 }
1923
1924 /*
1925 * void graph_enable_by_vertex()
1926 * If admin is non-zero, this is an administrative request for change
1927 * of the enabled property. Thus, send the ADMIN_DISABLE rather than
1928 * a plain DISABLE restarter event.
1929 */
1930 void
graph_enable_by_vertex(graph_vertex_t * vertex,int enable,int admin)1931 graph_enable_by_vertex(graph_vertex_t *vertex, int enable, int admin)
1932 {
1933 graph_vertex_t *v;
1934 int r;
1935
1936 assert(MUTEX_HELD(&dgraph_lock));
1937 assert((vertex->gv_flags & GV_CONFIGURED));
1938
1939 vertex->gv_flags = (vertex->gv_flags & ~GV_ENABLED) |
1940 (enable ? GV_ENABLED : 0);
1941
1942 if (enable) {
1943 if (vertex->gv_state != RESTARTER_STATE_OFFLINE &&
1944 vertex->gv_state != RESTARTER_STATE_DEGRADED &&
1945 vertex->gv_state != RESTARTER_STATE_ONLINE) {
1946 /*
1947 * In case the vertex was notified to go down,
1948 * but now can return online, clear the _TOOFFLINE
1949 * and _TODISABLE flags.
1950 */
1951 vertex->gv_flags &= ~GV_TOOFFLINE;
1952 vertex->gv_flags &= ~GV_TODISABLE;
1953
1954 vertex_send_event(vertex, RESTARTER_EVENT_TYPE_ENABLE);
1955 }
1956
1957 /*
1958 * Wait for state update from restarter before sending _START or
1959 * _STOP.
1960 */
1961
1962 return;
1963 }
1964
1965 if (vertex->gv_state == RESTARTER_STATE_DISABLED)
1966 return;
1967
1968 if (!admin) {
1969 vertex_send_event(vertex, RESTARTER_EVENT_TYPE_DISABLE);
1970
1971 /*
1972 * Wait for state update from restarter before sending _START or
1973 * _STOP.
1974 */
1975
1976 return;
1977 }
1978
1979 /*
1980 * If it is a DISABLE event requested by the administrator then we are
1981 * offlining the dependents first.
1982 */
1983
1984 /*
1985 * Set GV_TOOFFLINE for the services we are offlining. We cannot
1986 * clear the GV_TOOFFLINE bits from all the services because
1987 * other DISABLE events might be handled at the same time.
1988 */
1989 vertex->gv_flags |= GV_TOOFFLINE;
1990
1991 /* remember which vertex to disable... */
1992 vertex->gv_flags |= GV_TODISABLE;
1993
1994 log_framework(LOG_DEBUG, "Marking in-subtree vertices before "
1995 "disabling %s.\n", vertex->gv_name);
1996
1997 /* set GV_TOOFFLINE for its dependents */
1998 r = uu_list_walk(vertex->gv_dependents, (uu_walk_fn_t *)mark_subtree,
1999 NULL, 0);
2000 assert(r == 0);
2001
2002 /* disable the instance now if there is nothing else to offline */
2003 if (insubtree_dependents_down(vertex) == B_TRUE) {
2004 vertex_send_event(vertex, RESTARTER_EVENT_TYPE_ADMIN_DISABLE);
2005 return;
2006 }
2007
2008 /*
2009 * This loop is similar to the one used for the graph reversal shutdown
2010 * and could be improved in term of performance for the subtree reversal
2011 * disable case.
2012 */
2013 for (v = uu_list_first(dgraph); v != NULL;
2014 v = uu_list_next(dgraph, v)) {
2015 /* skip the vertex we are disabling for now */
2016 if (v == vertex)
2017 continue;
2018
2019 if (v->gv_type != GVT_INST ||
2020 (v->gv_flags & GV_CONFIGURED) == 0 ||
2021 (v->gv_flags & GV_ENABLED) == 0 ||
2022 (v->gv_flags & GV_TOOFFLINE) == 0)
2023 continue;
2024
2025 if ((v->gv_state != RESTARTER_STATE_ONLINE) &&
2026 (v->gv_state != RESTARTER_STATE_DEGRADED)) {
2027 /* continue if there is nothing to offline */
2028 continue;
2029 }
2030
2031 /*
2032 * Instances which are up need to come down before we're
2033 * done, but we can only offline the leaves here. An
2034 * instance is a leaf when all its dependents are down.
2035 */
2036 if (insubtree_dependents_down(v) == B_TRUE) {
2037 log_framework(LOG_DEBUG, "Offlining in-subtree "
2038 "instance %s for %s.\n",
2039 v->gv_name, vertex->gv_name);
2040 offline_vertex(v);
2041 }
2042 }
2043 }
2044
2045 static int configure_vertex(graph_vertex_t *, scf_instance_t *);
2046
2047 /*
2048 * Set the restarter for v to fmri_arg. That is, make sure a vertex for
2049 * fmri_arg exists, make v depend on it, and send _ADD_INSTANCE for v. If
2050 * v is already configured and fmri_arg indicates the current restarter, do
2051 * nothing. If v is configured and fmri_arg is a new restarter, delete v's
2052 * dependency on the restarter, send _REMOVE_INSTANCE for v, and set the new
2053 * restarter. Returns 0 on success, EINVAL if the FMRI is invalid,
2054 * ECONNABORTED if the repository connection is broken, and ELOOP
2055 * if the dependency would create a cycle. In the last case, *pathp will
2056 * point to a -1-terminated array of ids which compose the path from v to
2057 * restarter_fmri.
2058 */
2059 int
graph_change_restarter(graph_vertex_t * v,const char * fmri_arg,scf_handle_t * h,int ** pathp)2060 graph_change_restarter(graph_vertex_t *v, const char *fmri_arg, scf_handle_t *h,
2061 int **pathp)
2062 {
2063 char *restarter_fmri = NULL;
2064 graph_vertex_t *rv;
2065 int err;
2066 int id;
2067
2068 assert(MUTEX_HELD(&dgraph_lock));
2069
2070 if (fmri_arg[0] != '\0') {
2071 err = fmri_canonify(fmri_arg, &restarter_fmri, B_TRUE);
2072 if (err != 0) {
2073 assert(err == EINVAL);
2074 return (err);
2075 }
2076 }
2077
2078 if (restarter_fmri == NULL ||
2079 strcmp(restarter_fmri, SCF_SERVICE_STARTD) == 0) {
2080 if (v->gv_flags & GV_CONFIGURED) {
2081 if (v->gv_restarter_id == -1) {
2082 if (restarter_fmri != NULL)
2083 startd_free(restarter_fmri,
2084 max_scf_fmri_size);
2085 return (0);
2086 }
2087
2088 graph_unset_restarter(v);
2089 }
2090
2091 /* Master restarter, nothing to do. */
2092 v->gv_restarter_id = -1;
2093 v->gv_restarter_channel = NULL;
2094 vertex_send_event(v, RESTARTER_EVENT_TYPE_ADD_INSTANCE);
2095 return (0);
2096 }
2097
2098 if (v->gv_flags & GV_CONFIGURED) {
2099 id = dict_lookup_byname(restarter_fmri);
2100 if (id != -1 && v->gv_restarter_id == id) {
2101 startd_free(restarter_fmri, max_scf_fmri_size);
2102 return (0);
2103 }
2104
2105 graph_unset_restarter(v);
2106 }
2107
2108 err = graph_insert_vertex_unconfigured(restarter_fmri, GVT_INST, 0,
2109 RERR_NONE, &rv);
2110 startd_free(restarter_fmri, max_scf_fmri_size);
2111 assert(err == 0 || err == EEXIST);
2112
2113 if (rv->gv_delegate_initialized == 0) {
2114 if ((rv->gv_delegate_channel = restarter_protocol_init_delegate(
2115 rv->gv_name)) == NULL)
2116 return (EINVAL);
2117 rv->gv_delegate_initialized = 1;
2118 }
2119 v->gv_restarter_id = rv->gv_id;
2120 v->gv_restarter_channel = rv->gv_delegate_channel;
2121
2122 err = graph_insert_dependency(v, rv, pathp);
2123 if (err != 0) {
2124 assert(err == ELOOP);
2125 return (ELOOP);
2126 }
2127
2128 vertex_send_event(v, RESTARTER_EVENT_TYPE_ADD_INSTANCE);
2129
2130 if (!(rv->gv_flags & GV_CONFIGURED)) {
2131 scf_instance_t *inst;
2132
2133 err = libscf_fmri_get_instance(h, rv->gv_name, &inst);
2134 switch (err) {
2135 case 0:
2136 err = configure_vertex(rv, inst);
2137 scf_instance_destroy(inst);
2138 switch (err) {
2139 case 0:
2140 case ECANCELED:
2141 break;
2142
2143 case ECONNABORTED:
2144 return (ECONNABORTED);
2145
2146 default:
2147 bad_error("configure_vertex", err);
2148 }
2149 break;
2150
2151 case ECONNABORTED:
2152 return (ECONNABORTED);
2153
2154 case ENOENT:
2155 break;
2156
2157 case ENOTSUP:
2158 /*
2159 * The fmri doesn't specify an instance - translate
2160 * to EINVAL.
2161 */
2162 return (EINVAL);
2163
2164 case EINVAL:
2165 default:
2166 bad_error("libscf_fmri_get_instance", err);
2167 }
2168 }
2169
2170 return (0);
2171 }
2172
2173
2174 /*
2175 * Add all of the instances of the service named by fmri to the graph.
2176 * Returns
2177 * 0 - success
2178 * ENOENT - service indicated by fmri does not exist
2179 *
2180 * In both cases *reboundp will be B_TRUE if the handle was rebound, or B_FALSE
2181 * otherwise.
2182 */
2183 static int
add_service(const char * fmri,scf_handle_t * h,boolean_t * reboundp)2184 add_service(const char *fmri, scf_handle_t *h, boolean_t *reboundp)
2185 {
2186 scf_service_t *svc;
2187 scf_instance_t *inst;
2188 scf_iter_t *iter;
2189 char *inst_fmri;
2190 int ret, r;
2191
2192 *reboundp = B_FALSE;
2193
2194 svc = safe_scf_service_create(h);
2195 inst = safe_scf_instance_create(h);
2196 iter = safe_scf_iter_create(h);
2197 inst_fmri = startd_alloc(max_scf_fmri_size);
2198
2199 rebound:
2200 if (scf_handle_decode_fmri(h, fmri, NULL, svc, NULL, NULL, NULL,
2201 SCF_DECODE_FMRI_EXACT) != 0) {
2202 switch (scf_error()) {
2203 case SCF_ERROR_CONNECTION_BROKEN:
2204 default:
2205 libscf_handle_rebind(h);
2206 *reboundp = B_TRUE;
2207 goto rebound;
2208
2209 case SCF_ERROR_NOT_FOUND:
2210 ret = ENOENT;
2211 goto out;
2212
2213 case SCF_ERROR_INVALID_ARGUMENT:
2214 case SCF_ERROR_CONSTRAINT_VIOLATED:
2215 case SCF_ERROR_NOT_BOUND:
2216 case SCF_ERROR_HANDLE_MISMATCH:
2217 bad_error("scf_handle_decode_fmri", scf_error());
2218 }
2219 }
2220
2221 if (scf_iter_service_instances(iter, svc) != 0) {
2222 switch (scf_error()) {
2223 case SCF_ERROR_CONNECTION_BROKEN:
2224 default:
2225 libscf_handle_rebind(h);
2226 *reboundp = B_TRUE;
2227 goto rebound;
2228
2229 case SCF_ERROR_DELETED:
2230 ret = ENOENT;
2231 goto out;
2232
2233 case SCF_ERROR_HANDLE_MISMATCH:
2234 case SCF_ERROR_NOT_BOUND:
2235 case SCF_ERROR_NOT_SET:
2236 bad_error("scf_iter_service_instances", scf_error());
2237 }
2238 }
2239
2240 for (;;) {
2241 r = scf_iter_next_instance(iter, inst);
2242 if (r == 0)
2243 break;
2244 if (r != 1) {
2245 switch (scf_error()) {
2246 case SCF_ERROR_CONNECTION_BROKEN:
2247 default:
2248 libscf_handle_rebind(h);
2249 *reboundp = B_TRUE;
2250 goto rebound;
2251
2252 case SCF_ERROR_DELETED:
2253 ret = ENOENT;
2254 goto out;
2255
2256 case SCF_ERROR_HANDLE_MISMATCH:
2257 case SCF_ERROR_NOT_BOUND:
2258 case SCF_ERROR_NOT_SET:
2259 case SCF_ERROR_INVALID_ARGUMENT:
2260 bad_error("scf_iter_next_instance",
2261 scf_error());
2262 }
2263 }
2264
2265 if (scf_instance_to_fmri(inst, inst_fmri, max_scf_fmri_size) <
2266 0) {
2267 switch (scf_error()) {
2268 case SCF_ERROR_CONNECTION_BROKEN:
2269 libscf_handle_rebind(h);
2270 *reboundp = B_TRUE;
2271 goto rebound;
2272
2273 case SCF_ERROR_DELETED:
2274 continue;
2275
2276 case SCF_ERROR_NOT_BOUND:
2277 case SCF_ERROR_NOT_SET:
2278 bad_error("scf_instance_to_fmri", scf_error());
2279 }
2280 }
2281
2282 r = dgraph_add_instance(inst_fmri, inst, B_FALSE);
2283 switch (r) {
2284 case 0:
2285 case ECANCELED:
2286 break;
2287
2288 case EEXIST:
2289 continue;
2290
2291 case ECONNABORTED:
2292 libscf_handle_rebind(h);
2293 *reboundp = B_TRUE;
2294 goto rebound;
2295
2296 case EINVAL:
2297 default:
2298 bad_error("dgraph_add_instance", r);
2299 }
2300 }
2301
2302 ret = 0;
2303
2304 out:
2305 startd_free(inst_fmri, max_scf_fmri_size);
2306 scf_iter_destroy(iter);
2307 scf_instance_destroy(inst);
2308 scf_service_destroy(svc);
2309 return (ret);
2310 }
2311
2312 struct depfmri_info {
2313 graph_vertex_t *v; /* GVT_GROUP vertex */
2314 gv_type_t type; /* type of dependency */
2315 const char *inst_fmri; /* FMRI of parental GVT_INST vert. */
2316 const char *pg_name; /* Name of dependency pg */
2317 scf_handle_t *h;
2318 int err; /* return error code */
2319 int **pathp; /* return circular dependency path */
2320 };
2321
2322 /*
2323 * Find or create a vertex for fmri and make info->v depend on it.
2324 * Returns
2325 * 0 - success
2326 * nonzero - failure
2327 *
2328 * On failure, sets info->err to
2329 * EINVAL - fmri is invalid
2330 * fmri does not match info->type
2331 * ELOOP - Adding the dependency creates a circular dependency. *info->pathp
2332 * will point to an array of the ids of the members of the cycle.
2333 * ECONNABORTED - repository connection was broken
2334 * ECONNRESET - succeeded, but repository connection was reset
2335 */
2336 static int
process_dependency_fmri(const char * fmri,struct depfmri_info * info)2337 process_dependency_fmri(const char *fmri, struct depfmri_info *info)
2338 {
2339 int err;
2340 graph_vertex_t *depgroup_v, *v;
2341 char *fmri_copy, *cfmri;
2342 size_t fmri_copy_sz;
2343 const char *scope, *service, *instance, *pg;
2344 scf_instance_t *inst;
2345 boolean_t rebound;
2346
2347 assert(MUTEX_HELD(&dgraph_lock));
2348
2349 /* Get or create vertex for FMRI */
2350 depgroup_v = info->v;
2351
2352 if (strncmp(fmri, "file:", sizeof ("file:") - 1) == 0) {
2353 if (info->type != GVT_FILE) {
2354 log_framework(LOG_NOTICE,
2355 "FMRI \"%s\" is not allowed for the \"%s\" "
2356 "dependency's type of instance %s.\n", fmri,
2357 info->pg_name, info->inst_fmri);
2358 return (info->err = EINVAL);
2359 }
2360
2361 err = graph_insert_vertex_unconfigured(fmri, info->type, 0,
2362 RERR_NONE, &v);
2363 switch (err) {
2364 case 0:
2365 break;
2366
2367 case EEXIST:
2368 assert(v->gv_type == GVT_FILE);
2369 break;
2370
2371 case EINVAL: /* prevented above */
2372 default:
2373 bad_error("graph_insert_vertex_unconfigured", err);
2374 }
2375 } else {
2376 if (info->type != GVT_INST) {
2377 log_framework(LOG_NOTICE,
2378 "FMRI \"%s\" is not allowed for the \"%s\" "
2379 "dependency's type of instance %s.\n", fmri,
2380 info->pg_name, info->inst_fmri);
2381 return (info->err = EINVAL);
2382 }
2383
2384 /*
2385 * We must canonify fmri & add a vertex for it.
2386 */
2387 fmri_copy_sz = strlen(fmri) + 1;
2388 fmri_copy = startd_alloc(fmri_copy_sz);
2389 (void) strcpy(fmri_copy, fmri);
2390
2391 /* Determine if the FMRI is a property group or instance */
2392 if (scf_parse_svc_fmri(fmri_copy, &scope, &service,
2393 &instance, &pg, NULL) != 0) {
2394 startd_free(fmri_copy, fmri_copy_sz);
2395 log_framework(LOG_NOTICE,
2396 "Dependency \"%s\" of %s has invalid FMRI "
2397 "\"%s\".\n", info->pg_name, info->inst_fmri,
2398 fmri);
2399 return (info->err = EINVAL);
2400 }
2401
2402 if (service == NULL || pg != NULL) {
2403 startd_free(fmri_copy, fmri_copy_sz);
2404 log_framework(LOG_NOTICE,
2405 "Dependency \"%s\" of %s does not designate a "
2406 "service or instance.\n", info->pg_name,
2407 info->inst_fmri);
2408 return (info->err = EINVAL);
2409 }
2410
2411 if (scope == NULL || strcmp(scope, SCF_SCOPE_LOCAL) == 0) {
2412 cfmri = uu_msprintf("svc:/%s%s%s",
2413 service, instance ? ":" : "", instance ? instance :
2414 "");
2415 } else {
2416 cfmri = uu_msprintf("svc://%s/%s%s%s",
2417 scope, service, instance ? ":" : "", instance ?
2418 instance : "");
2419 }
2420
2421 startd_free(fmri_copy, fmri_copy_sz);
2422
2423 err = graph_insert_vertex_unconfigured(cfmri, instance ?
2424 GVT_INST : GVT_SVC, instance ? 0 : DEPGRP_REQUIRE_ANY,
2425 RERR_NONE, &v);
2426 uu_free(cfmri);
2427 switch (err) {
2428 case 0:
2429 break;
2430
2431 case EEXIST:
2432 /* Verify v. */
2433 if (instance != NULL)
2434 assert(v->gv_type == GVT_INST);
2435 else
2436 assert(v->gv_type == GVT_SVC);
2437 break;
2438
2439 default:
2440 bad_error("graph_insert_vertex_unconfigured", err);
2441 }
2442 }
2443
2444 /* Add dependency from depgroup_v to new vertex */
2445 info->err = graph_insert_dependency(depgroup_v, v, info->pathp);
2446 switch (info->err) {
2447 case 0:
2448 break;
2449
2450 case ELOOP:
2451 return (ELOOP);
2452
2453 default:
2454 bad_error("graph_insert_dependency", info->err);
2455 }
2456
2457 /* This must be after we insert the dependency, to avoid looping. */
2458 switch (v->gv_type) {
2459 case GVT_INST:
2460 if ((v->gv_flags & GV_CONFIGURED) != 0)
2461 break;
2462
2463 inst = safe_scf_instance_create(info->h);
2464
2465 rebound = B_FALSE;
2466
2467 rebound:
2468 err = libscf_lookup_instance(v->gv_name, inst);
2469 switch (err) {
2470 case 0:
2471 err = configure_vertex(v, inst);
2472 switch (err) {
2473 case 0:
2474 case ECANCELED:
2475 break;
2476
2477 case ECONNABORTED:
2478 libscf_handle_rebind(info->h);
2479 rebound = B_TRUE;
2480 goto rebound;
2481
2482 default:
2483 bad_error("configure_vertex", err);
2484 }
2485 break;
2486
2487 case ENOENT:
2488 break;
2489
2490 case ECONNABORTED:
2491 libscf_handle_rebind(info->h);
2492 rebound = B_TRUE;
2493 goto rebound;
2494
2495 case EINVAL:
2496 case ENOTSUP:
2497 default:
2498 bad_error("libscf_fmri_get_instance", err);
2499 }
2500
2501 scf_instance_destroy(inst);
2502
2503 if (rebound)
2504 return (info->err = ECONNRESET);
2505 break;
2506
2507 case GVT_SVC:
2508 (void) add_service(v->gv_name, info->h, &rebound);
2509 if (rebound)
2510 return (info->err = ECONNRESET);
2511 }
2512
2513 return (0);
2514 }
2515
2516 struct deppg_info {
2517 graph_vertex_t *v; /* GVT_INST vertex */
2518 int err; /* return error */
2519 int **pathp; /* return circular dependency path */
2520 };
2521
2522 /*
2523 * Make info->v depend on a new GVT_GROUP node for this property group,
2524 * and then call process_dependency_fmri() for the values of the entity
2525 * property. Return 0 on success, or if something goes wrong return nonzero
2526 * and set info->err to ECONNABORTED, EINVAL, or the error code returned by
2527 * process_dependency_fmri().
2528 */
2529 static int
process_dependency_pg(scf_propertygroup_t * pg,struct deppg_info * info)2530 process_dependency_pg(scf_propertygroup_t *pg, struct deppg_info *info)
2531 {
2532 scf_handle_t *h;
2533 depgroup_type_t deptype;
2534 restarter_error_t rerr;
2535 struct depfmri_info linfo;
2536 char *fmri, *pg_name;
2537 size_t fmri_sz;
2538 graph_vertex_t *depgrp;
2539 scf_property_t *prop;
2540 int err;
2541 int empty;
2542 scf_error_t scferr;
2543 ssize_t len;
2544
2545 assert(MUTEX_HELD(&dgraph_lock));
2546
2547 h = scf_pg_handle(pg);
2548
2549 pg_name = startd_alloc(max_scf_name_size);
2550
2551 len = scf_pg_get_name(pg, pg_name, max_scf_name_size);
2552 if (len < 0) {
2553 startd_free(pg_name, max_scf_name_size);
2554 switch (scf_error()) {
2555 case SCF_ERROR_CONNECTION_BROKEN:
2556 default:
2557 return (info->err = ECONNABORTED);
2558
2559 case SCF_ERROR_DELETED:
2560 return (info->err = 0);
2561
2562 case SCF_ERROR_NOT_SET:
2563 bad_error("scf_pg_get_name", scf_error());
2564 }
2565 }
2566
2567 /*
2568 * Skip over empty dependency groups. Since dependency property
2569 * groups are updated atomically, they are either empty or
2570 * fully populated.
2571 */
2572 empty = depgroup_empty(h, pg);
2573 if (empty < 0) {
2574 log_error(LOG_INFO,
2575 "Error reading dependency group \"%s\" of %s: %s\n",
2576 pg_name, info->v->gv_name, scf_strerror(scf_error()));
2577 startd_free(pg_name, max_scf_name_size);
2578 return (info->err = EINVAL);
2579
2580 } else if (empty == 1) {
2581 log_framework(LOG_DEBUG,
2582 "Ignoring empty dependency group \"%s\" of %s\n",
2583 pg_name, info->v->gv_name);
2584 startd_free(pg_name, max_scf_name_size);
2585 return (info->err = 0);
2586 }
2587
2588 fmri_sz = strlen(info->v->gv_name) + 1 + len + 1;
2589 fmri = startd_alloc(fmri_sz);
2590
2591 (void) snprintf(fmri, max_scf_name_size, "%s>%s", info->v->gv_name,
2592 pg_name);
2593
2594 /* Validate the pg before modifying the graph */
2595 deptype = depgroup_read_grouping(h, pg);
2596 if (deptype == DEPGRP_UNSUPPORTED) {
2597 log_error(LOG_INFO,
2598 "Dependency \"%s\" of %s has an unknown grouping value.\n",
2599 pg_name, info->v->gv_name);
2600 startd_free(fmri, fmri_sz);
2601 startd_free(pg_name, max_scf_name_size);
2602 return (info->err = EINVAL);
2603 }
2604
2605 rerr = depgroup_read_restart(h, pg);
2606 if (rerr == RERR_UNSUPPORTED) {
2607 log_error(LOG_INFO,
2608 "Dependency \"%s\" of %s has an unknown restart_on value."
2609 "\n", pg_name, info->v->gv_name);
2610 startd_free(fmri, fmri_sz);
2611 startd_free(pg_name, max_scf_name_size);
2612 return (info->err = EINVAL);
2613 }
2614
2615 prop = safe_scf_property_create(h);
2616
2617 if (scf_pg_get_property(pg, SCF_PROPERTY_ENTITIES, prop) != 0) {
2618 scferr = scf_error();
2619 scf_property_destroy(prop);
2620 if (scferr == SCF_ERROR_DELETED) {
2621 startd_free(fmri, fmri_sz);
2622 startd_free(pg_name, max_scf_name_size);
2623 return (info->err = 0);
2624 } else if (scferr != SCF_ERROR_NOT_FOUND) {
2625 startd_free(fmri, fmri_sz);
2626 startd_free(pg_name, max_scf_name_size);
2627 return (info->err = ECONNABORTED);
2628 }
2629
2630 log_error(LOG_INFO,
2631 "Dependency \"%s\" of %s is missing a \"%s\" property.\n",
2632 pg_name, info->v->gv_name, SCF_PROPERTY_ENTITIES);
2633
2634 startd_free(fmri, fmri_sz);
2635 startd_free(pg_name, max_scf_name_size);
2636
2637 return (info->err = EINVAL);
2638 }
2639
2640 /* Create depgroup vertex for pg */
2641 err = graph_insert_vertex_unconfigured(fmri, GVT_GROUP, deptype,
2642 rerr, &depgrp);
2643 assert(err == 0);
2644 startd_free(fmri, fmri_sz);
2645
2646 /* Add dependency from inst vertex to new vertex */
2647 err = graph_insert_dependency(info->v, depgrp, info->pathp);
2648 /* ELOOP can't happen because this should be a new vertex */
2649 assert(err == 0);
2650
2651 linfo.v = depgrp;
2652 linfo.type = depgroup_read_scheme(h, pg);
2653 linfo.inst_fmri = info->v->gv_name;
2654 linfo.pg_name = pg_name;
2655 linfo.h = h;
2656 linfo.err = 0;
2657 linfo.pathp = info->pathp;
2658 err = walk_property_astrings(prop, (callback_t)process_dependency_fmri,
2659 &linfo);
2660
2661 scf_property_destroy(prop);
2662 startd_free(pg_name, max_scf_name_size);
2663
2664 switch (err) {
2665 case 0:
2666 case EINTR:
2667 return (info->err = linfo.err);
2668
2669 case ECONNABORTED:
2670 case EINVAL:
2671 return (info->err = err);
2672
2673 case ECANCELED:
2674 return (info->err = 0);
2675
2676 case ECONNRESET:
2677 return (info->err = ECONNABORTED);
2678
2679 default:
2680 bad_error("walk_property_astrings", err);
2681 /* NOTREACHED */
2682 }
2683 }
2684
2685 /*
2686 * Build the dependency info for v from the repository. Returns 0 on success,
2687 * ECONNABORTED on repository disconnection, EINVAL if the repository
2688 * configuration is invalid, and ELOOP if a dependency would cause a cycle.
2689 * In the last case, *pathp will point to a -1-terminated array of ids which
2690 * constitute the rest of the dependency cycle.
2691 */
2692 static int
set_dependencies(graph_vertex_t * v,scf_instance_t * inst,int ** pathp)2693 set_dependencies(graph_vertex_t *v, scf_instance_t *inst, int **pathp)
2694 {
2695 struct deppg_info info;
2696 int err;
2697 uint_t old_configured;
2698
2699 assert(MUTEX_HELD(&dgraph_lock));
2700
2701 /*
2702 * Mark the vertex as configured during dependency insertion to avoid
2703 * dependency cycles (which can appear in the graph if one of the
2704 * vertices is an exclusion-group).
2705 */
2706 old_configured = v->gv_flags & GV_CONFIGURED;
2707 v->gv_flags |= GV_CONFIGURED;
2708
2709 info.err = 0;
2710 info.v = v;
2711 info.pathp = pathp;
2712
2713 err = walk_dependency_pgs(inst, (callback_t)process_dependency_pg,
2714 &info);
2715
2716 if (!old_configured)
2717 v->gv_flags &= ~GV_CONFIGURED;
2718
2719 switch (err) {
2720 case 0:
2721 case EINTR:
2722 return (info.err);
2723
2724 case ECONNABORTED:
2725 return (ECONNABORTED);
2726
2727 case ECANCELED:
2728 /* Should get delete event, so return 0. */
2729 return (0);
2730
2731 default:
2732 bad_error("walk_dependency_pgs", err);
2733 /* NOTREACHED */
2734 }
2735 }
2736
2737
2738 static void
handle_cycle(const char * fmri,int * path)2739 handle_cycle(const char *fmri, int *path)
2740 {
2741 const char *cp;
2742 size_t sz;
2743
2744 assert(MUTEX_HELD(&dgraph_lock));
2745
2746 path_to_str(path, (char **)&cp, &sz);
2747
2748 log_error(LOG_ERR, "Transitioning %s to maintenance "
2749 "because it completes a dependency cycle (see svcs -xv for "
2750 "details):\n%s", fmri ? fmri : "?", cp);
2751
2752 startd_free((void *)cp, sz);
2753 }
2754
2755 /*
2756 * Increment the vertex's reference count to prevent the vertex removal
2757 * from the dgraph.
2758 */
2759 static void
vertex_ref(graph_vertex_t * v)2760 vertex_ref(graph_vertex_t *v)
2761 {
2762 assert(MUTEX_HELD(&dgraph_lock));
2763
2764 v->gv_refs++;
2765 }
2766
2767 /*
2768 * Decrement the vertex's reference count and remove the vertex from
2769 * the dgraph when possible.
2770 *
2771 * Return VERTEX_REMOVED when the vertex has been removed otherwise
2772 * return VERTEX_INUSE.
2773 */
2774 static int
vertex_unref(graph_vertex_t * v)2775 vertex_unref(graph_vertex_t *v)
2776 {
2777 assert(MUTEX_HELD(&dgraph_lock));
2778 assert(v->gv_refs > 0);
2779
2780 v->gv_refs--;
2781
2782 return (free_if_unrefed(v));
2783 }
2784
2785 /*
2786 * When run on the dependencies of a vertex, populates list with
2787 * graph_edge_t's which point to the service vertices or the instance
2788 * vertices (no GVT_GROUP nodes) on which the vertex depends.
2789 *
2790 * Increment the vertex's reference count once the vertex is inserted
2791 * in the list. The vertex won't be able to be deleted from the dgraph
2792 * while it is referenced.
2793 */
2794 static int
append_svcs_or_insts(graph_edge_t * e,uu_list_t * list)2795 append_svcs_or_insts(graph_edge_t *e, uu_list_t *list)
2796 {
2797 graph_vertex_t *v = e->ge_vertex;
2798 graph_edge_t *new;
2799 int r;
2800
2801 switch (v->gv_type) {
2802 case GVT_INST:
2803 case GVT_SVC:
2804 break;
2805
2806 case GVT_GROUP:
2807 r = uu_list_walk(v->gv_dependencies,
2808 (uu_walk_fn_t *)append_svcs_or_insts, list, 0);
2809 assert(r == 0);
2810 return (UU_WALK_NEXT);
2811
2812 case GVT_FILE:
2813 return (UU_WALK_NEXT);
2814
2815 default:
2816 #ifndef NDEBUG
2817 uu_warn("%s:%d: Unexpected vertex type %d.\n", __FILE__,
2818 __LINE__, v->gv_type);
2819 #endif
2820 abort();
2821 }
2822
2823 new = startd_alloc(sizeof (*new));
2824 new->ge_vertex = v;
2825 uu_list_node_init(new, &new->ge_link, graph_edge_pool);
2826 r = uu_list_insert_before(list, NULL, new);
2827 assert(r == 0);
2828
2829 /*
2830 * Because we are inserting the vertex in a list, we don't want
2831 * the vertex to be freed while the list is in use. In order to
2832 * achieve that, increment the vertex's reference count.
2833 */
2834 vertex_ref(v);
2835
2836 return (UU_WALK_NEXT);
2837 }
2838
2839 static boolean_t
should_be_in_subgraph(graph_vertex_t * v)2840 should_be_in_subgraph(graph_vertex_t *v)
2841 {
2842 graph_edge_t *e;
2843
2844 if (v == milestone)
2845 return (B_TRUE);
2846
2847 /*
2848 * v is in the subgraph if any of its dependents are in the subgraph.
2849 * Except for EXCLUDE_ALL dependents. And OPTIONAL dependents only
2850 * count if we're enabled.
2851 */
2852 for (e = uu_list_first(v->gv_dependents);
2853 e != NULL;
2854 e = uu_list_next(v->gv_dependents, e)) {
2855 graph_vertex_t *dv = e->ge_vertex;
2856
2857 if (!(dv->gv_flags & GV_INSUBGRAPH))
2858 continue;
2859
2860 /*
2861 * Don't include instances that are optional and disabled.
2862 */
2863 if (v->gv_type == GVT_INST && dv->gv_type == GVT_SVC) {
2864
2865 int in = 0;
2866 graph_edge_t *ee;
2867
2868 for (ee = uu_list_first(dv->gv_dependents);
2869 ee != NULL;
2870 ee = uu_list_next(dv->gv_dependents, ee)) {
2871
2872 graph_vertex_t *ddv = e->ge_vertex;
2873
2874 if (ddv->gv_type == GVT_GROUP &&
2875 ddv->gv_depgroup == DEPGRP_EXCLUDE_ALL)
2876 continue;
2877
2878 if (ddv->gv_type == GVT_GROUP &&
2879 ddv->gv_depgroup == DEPGRP_OPTIONAL_ALL &&
2880 !(v->gv_flags & GV_ENBLD_NOOVR))
2881 continue;
2882
2883 in = 1;
2884 }
2885 if (!in)
2886 continue;
2887 }
2888 if (v->gv_type == GVT_INST &&
2889 dv->gv_type == GVT_GROUP &&
2890 dv->gv_depgroup == DEPGRP_OPTIONAL_ALL &&
2891 !(v->gv_flags & GV_ENBLD_NOOVR))
2892 continue;
2893
2894 /* Don't include excluded services and instances */
2895 if (dv->gv_type == GVT_GROUP &&
2896 dv->gv_depgroup == DEPGRP_EXCLUDE_ALL)
2897 continue;
2898
2899 return (B_TRUE);
2900 }
2901
2902 return (B_FALSE);
2903 }
2904
2905 /*
2906 * Ensures that GV_INSUBGRAPH is set properly for v and its descendents. If
2907 * any bits change, manipulate the repository appropriately. Returns 0 or
2908 * ECONNABORTED.
2909 */
2910 static int
eval_subgraph(graph_vertex_t * v,scf_handle_t * h)2911 eval_subgraph(graph_vertex_t *v, scf_handle_t *h)
2912 {
2913 boolean_t old = (v->gv_flags & GV_INSUBGRAPH) != 0;
2914 boolean_t new;
2915 graph_edge_t *e;
2916 scf_instance_t *inst;
2917 int ret = 0, r;
2918
2919 assert(milestone != NULL && milestone != MILESTONE_NONE);
2920
2921 new = should_be_in_subgraph(v);
2922
2923 if (new == old)
2924 return (0);
2925
2926 log_framework(LOG_DEBUG, new ? "Adding %s to the subgraph.\n" :
2927 "Removing %s from the subgraph.\n", v->gv_name);
2928
2929 v->gv_flags = (v->gv_flags & ~GV_INSUBGRAPH) |
2930 (new ? GV_INSUBGRAPH : 0);
2931
2932 if (v->gv_type == GVT_INST && (v->gv_flags & GV_CONFIGURED)) {
2933 int err;
2934
2935 get_inst:
2936 err = libscf_fmri_get_instance(h, v->gv_name, &inst);
2937 if (err != 0) {
2938 switch (err) {
2939 case ECONNABORTED:
2940 libscf_handle_rebind(h);
2941 ret = ECONNABORTED;
2942 goto get_inst;
2943
2944 case ENOENT:
2945 break;
2946
2947 case EINVAL:
2948 case ENOTSUP:
2949 default:
2950 bad_error("libscf_fmri_get_instance", err);
2951 }
2952 } else {
2953 const char *f;
2954
2955 if (new) {
2956 err = libscf_delete_enable_ovr(inst);
2957 f = "libscf_delete_enable_ovr";
2958 } else {
2959 err = libscf_set_enable_ovr(inst, 0);
2960 f = "libscf_set_enable_ovr";
2961 }
2962 scf_instance_destroy(inst);
2963 switch (err) {
2964 case 0:
2965 case ECANCELED:
2966 break;
2967
2968 case ECONNABORTED:
2969 libscf_handle_rebind(h);
2970 /*
2971 * We must continue so the graph is updated,
2972 * but we must return ECONNABORTED so any
2973 * libscf state held by any callers is reset.
2974 */
2975 ret = ECONNABORTED;
2976 goto get_inst;
2977
2978 case EROFS:
2979 case EPERM:
2980 log_error(LOG_WARNING,
2981 "Could not set %s/%s for %s: %s.\n",
2982 SCF_PG_GENERAL_OVR, SCF_PROPERTY_ENABLED,
2983 v->gv_name, strerror(err));
2984 break;
2985
2986 default:
2987 bad_error(f, err);
2988 }
2989 }
2990 }
2991
2992 for (e = uu_list_first(v->gv_dependencies);
2993 e != NULL;
2994 e = uu_list_next(v->gv_dependencies, e)) {
2995 r = eval_subgraph(e->ge_vertex, h);
2996 if (r != 0) {
2997 assert(r == ECONNABORTED);
2998 ret = ECONNABORTED;
2999 }
3000 }
3001
3002 return (ret);
3003 }
3004
3005 /*
3006 * Delete the (property group) dependencies of v & create new ones based on
3007 * inst. If doing so would create a cycle, log a message and put the instance
3008 * into maintenance. Update GV_INSUBGRAPH flags as necessary. Returns 0 or
3009 * ECONNABORTED.
3010 */
3011 int
refresh_vertex(graph_vertex_t * v,scf_instance_t * inst)3012 refresh_vertex(graph_vertex_t *v, scf_instance_t *inst)
3013 {
3014 int err;
3015 int *path;
3016 char *fmri;
3017 int r;
3018 scf_handle_t *h = scf_instance_handle(inst);
3019 uu_list_t *old_deps;
3020 int ret = 0;
3021 graph_edge_t *e;
3022 graph_vertex_t *vv;
3023
3024 assert(MUTEX_HELD(&dgraph_lock));
3025 assert(v->gv_type == GVT_INST);
3026
3027 log_framework(LOG_DEBUG, "Graph engine: Refreshing %s.\n", v->gv_name);
3028
3029 if (milestone > MILESTONE_NONE) {
3030 /*
3031 * In case some of v's dependencies are being deleted we must
3032 * make a list of them now for GV_INSUBGRAPH-flag evaluation
3033 * after the new dependencies are in place.
3034 */
3035 old_deps = startd_list_create(graph_edge_pool, NULL, 0);
3036
3037 err = uu_list_walk(v->gv_dependencies,
3038 (uu_walk_fn_t *)append_svcs_or_insts, old_deps, 0);
3039 assert(err == 0);
3040 }
3041
3042 delete_instance_dependencies(v, B_FALSE);
3043
3044 err = set_dependencies(v, inst, &path);
3045 switch (err) {
3046 case 0:
3047 break;
3048
3049 case ECONNABORTED:
3050 ret = err;
3051 goto out;
3052
3053 case EINVAL:
3054 case ELOOP:
3055 r = libscf_instance_get_fmri(inst, &fmri);
3056 switch (r) {
3057 case 0:
3058 break;
3059
3060 case ECONNABORTED:
3061 ret = ECONNABORTED;
3062 goto out;
3063
3064 case ECANCELED:
3065 ret = 0;
3066 goto out;
3067
3068 default:
3069 bad_error("libscf_instance_get_fmri", r);
3070 }
3071
3072 if (err == EINVAL) {
3073 log_error(LOG_ERR, "Transitioning %s "
3074 "to maintenance due to misconfiguration.\n",
3075 fmri ? fmri : "?");
3076 vertex_send_event(v,
3077 RESTARTER_EVENT_TYPE_INVALID_DEPENDENCY);
3078 } else {
3079 handle_cycle(fmri, path);
3080 vertex_send_event(v,
3081 RESTARTER_EVENT_TYPE_DEPENDENCY_CYCLE);
3082 }
3083 startd_free(fmri, max_scf_fmri_size);
3084 ret = 0;
3085 goto out;
3086
3087 default:
3088 bad_error("set_dependencies", err);
3089 }
3090
3091 if (milestone > MILESTONE_NONE) {
3092 boolean_t aborted = B_FALSE;
3093
3094 for (e = uu_list_first(old_deps);
3095 e != NULL;
3096 e = uu_list_next(old_deps, e)) {
3097 vv = e->ge_vertex;
3098
3099 if (vertex_unref(vv) == VERTEX_INUSE &&
3100 eval_subgraph(vv, h) == ECONNABORTED)
3101 aborted = B_TRUE;
3102 }
3103
3104 for (e = uu_list_first(v->gv_dependencies);
3105 e != NULL;
3106 e = uu_list_next(v->gv_dependencies, e)) {
3107 if (eval_subgraph(e->ge_vertex, h) ==
3108 ECONNABORTED)
3109 aborted = B_TRUE;
3110 }
3111
3112 if (aborted) {
3113 ret = ECONNABORTED;
3114 goto out;
3115 }
3116 }
3117
3118 graph_start_if_satisfied(v);
3119
3120 ret = 0;
3121
3122 out:
3123 if (milestone > MILESTONE_NONE) {
3124 void *cookie = NULL;
3125
3126 while ((e = uu_list_teardown(old_deps, &cookie)) != NULL)
3127 startd_free(e, sizeof (*e));
3128
3129 uu_list_destroy(old_deps);
3130 }
3131
3132 return (ret);
3133 }
3134
3135 /*
3136 * Set up v according to inst. That is, make sure it depends on its
3137 * restarter and set up its dependencies. Send the ADD_INSTANCE command to
3138 * the restarter, and send ENABLE or DISABLE as appropriate.
3139 *
3140 * Returns 0 on success, ECONNABORTED on repository disconnection, or
3141 * ECANCELED if inst is deleted.
3142 */
3143 static int
configure_vertex(graph_vertex_t * v,scf_instance_t * inst)3144 configure_vertex(graph_vertex_t *v, scf_instance_t *inst)
3145 {
3146 scf_handle_t *h;
3147 scf_propertygroup_t *pg;
3148 scf_snapshot_t *snap;
3149 char *restarter_fmri = startd_alloc(max_scf_value_size);
3150 int enabled, enabled_ovr;
3151 int err;
3152 int *path;
3153 int deathrow;
3154 int32_t tset;
3155
3156 restarter_fmri[0] = '\0';
3157
3158 assert(MUTEX_HELD(&dgraph_lock));
3159 assert(v->gv_type == GVT_INST);
3160 assert((v->gv_flags & GV_CONFIGURED) == 0);
3161
3162 /* GV_INSUBGRAPH should already be set properly. */
3163 assert(should_be_in_subgraph(v) ==
3164 ((v->gv_flags & GV_INSUBGRAPH) != 0));
3165
3166 /*
3167 * If the instance fmri is in the deathrow list then set the
3168 * GV_DEATHROW flag on the vertex and create and set to true the
3169 * SCF_PROPERTY_DEATHROW boolean property in the non-persistent
3170 * repository for this instance fmri.
3171 */
3172 if ((v->gv_flags & GV_DEATHROW) ||
3173 (is_fmri_in_deathrow(v->gv_name) == B_TRUE)) {
3174 if ((v->gv_flags & GV_DEATHROW) == 0) {
3175 /*
3176 * Set flag GV_DEATHROW, create and set to true
3177 * the SCF_PROPERTY_DEATHROW property in the
3178 * non-persistent repository for this instance fmri.
3179 */
3180 v->gv_flags |= GV_DEATHROW;
3181
3182 switch (err = libscf_set_deathrow(inst, 1)) {
3183 case 0:
3184 break;
3185
3186 case ECONNABORTED:
3187 case ECANCELED:
3188 startd_free(restarter_fmri, max_scf_value_size);
3189 return (err);
3190
3191 case EROFS:
3192 log_error(LOG_WARNING, "Could not set %s/%s "
3193 "for deathrow %s: %s.\n",
3194 SCF_PG_DEATHROW, SCF_PROPERTY_DEATHROW,
3195 v->gv_name, strerror(err));
3196 break;
3197
3198 case EPERM:
3199 uu_die("Permission denied.\n");
3200 /* NOTREACHED */
3201
3202 default:
3203 bad_error("libscf_set_deathrow", err);
3204 }
3205 log_framework(LOG_DEBUG, "Deathrow, graph set %s.\n",
3206 v->gv_name);
3207 }
3208 startd_free(restarter_fmri, max_scf_value_size);
3209 return (0);
3210 }
3211
3212 h = scf_instance_handle(inst);
3213
3214 /*
3215 * Using a temporary deathrow boolean property, set through
3216 * libscf_set_deathrow(), only for fmris on deathrow, is necessary
3217 * because deathrow_fini() may already have been called, and in case
3218 * of a refresh, GV_DEATHROW may need to be set again.
3219 * libscf_get_deathrow() sets deathrow to 1 only if this instance
3220 * has a temporary boolean property named 'deathrow' valued true
3221 * in a property group 'deathrow', -1 or 0 in all other cases.
3222 */
3223 err = libscf_get_deathrow(h, inst, &deathrow);
3224 switch (err) {
3225 case 0:
3226 break;
3227
3228 case ECONNABORTED:
3229 case ECANCELED:
3230 startd_free(restarter_fmri, max_scf_value_size);
3231 return (err);
3232
3233 default:
3234 bad_error("libscf_get_deathrow", err);
3235 }
3236
3237 if (deathrow == 1) {
3238 v->gv_flags |= GV_DEATHROW;
3239 startd_free(restarter_fmri, max_scf_value_size);
3240 return (0);
3241 }
3242
3243 log_framework(LOG_DEBUG, "Graph adding %s.\n", v->gv_name);
3244
3245 /*
3246 * If the instance does not have a restarter property group,
3247 * initialize its state to uninitialized/none, in case the restarter
3248 * is not enabled.
3249 */
3250 pg = safe_scf_pg_create(h);
3251
3252 if (scf_instance_get_pg(inst, SCF_PG_RESTARTER, pg) != 0) {
3253 instance_data_t idata;
3254 uint_t count = 0, msecs = ALLOC_DELAY;
3255
3256 switch (scf_error()) {
3257 case SCF_ERROR_NOT_FOUND:
3258 break;
3259
3260 case SCF_ERROR_CONNECTION_BROKEN:
3261 default:
3262 scf_pg_destroy(pg);
3263 startd_free(restarter_fmri, max_scf_value_size);
3264 return (ECONNABORTED);
3265
3266 case SCF_ERROR_DELETED:
3267 scf_pg_destroy(pg);
3268 startd_free(restarter_fmri, max_scf_value_size);
3269 return (ECANCELED);
3270
3271 case SCF_ERROR_NOT_SET:
3272 bad_error("scf_instance_get_pg", scf_error());
3273 }
3274
3275 switch (err = libscf_instance_get_fmri(inst,
3276 (char **)&idata.i_fmri)) {
3277 case 0:
3278 break;
3279
3280 case ECONNABORTED:
3281 case ECANCELED:
3282 scf_pg_destroy(pg);
3283 startd_free(restarter_fmri, max_scf_value_size);
3284 return (err);
3285
3286 default:
3287 bad_error("libscf_instance_get_fmri", err);
3288 }
3289
3290 idata.i_state = RESTARTER_STATE_NONE;
3291 idata.i_next_state = RESTARTER_STATE_NONE;
3292
3293 init_state:
3294 switch (err = _restarter_commit_states(h, &idata,
3295 RESTARTER_STATE_UNINIT, RESTARTER_STATE_NONE,
3296 restarter_get_str_short(restarter_str_insert_in_graph))) {
3297 case 0:
3298 break;
3299
3300 case ENOMEM:
3301 ++count;
3302 if (count < ALLOC_RETRY) {
3303 (void) poll(NULL, 0, msecs);
3304 msecs *= ALLOC_DELAY_MULT;
3305 goto init_state;
3306 }
3307
3308 uu_die("Insufficient memory.\n");
3309 /* NOTREACHED */
3310
3311 case ECONNABORTED:
3312 startd_free((void *)idata.i_fmri, max_scf_fmri_size);
3313 scf_pg_destroy(pg);
3314 startd_free(restarter_fmri, max_scf_value_size);
3315 return (ECONNABORTED);
3316
3317 case ENOENT:
3318 startd_free((void *)idata.i_fmri, max_scf_fmri_size);
3319 scf_pg_destroy(pg);
3320 startd_free(restarter_fmri, max_scf_value_size);
3321 return (ECANCELED);
3322
3323 case EPERM:
3324 case EACCES:
3325 case EROFS:
3326 log_error(LOG_NOTICE, "Could not initialize state for "
3327 "%s: %s.\n", idata.i_fmri, strerror(err));
3328 break;
3329
3330 case EINVAL:
3331 default:
3332 bad_error("_restarter_commit_states", err);
3333 }
3334
3335 startd_free((void *)idata.i_fmri, max_scf_fmri_size);
3336 }
3337
3338 scf_pg_destroy(pg);
3339
3340 if (milestone != NULL) {
3341 /*
3342 * Make sure the enable-override is set properly before we
3343 * read whether we should be enabled.
3344 */
3345 if (milestone == MILESTONE_NONE ||
3346 !(v->gv_flags & GV_INSUBGRAPH)) {
3347 /*
3348 * This might seem unjustified after the milestone
3349 * transition has completed (non_subgraph_svcs == 0),
3350 * but it's important because when we boot to
3351 * a milestone, we set the milestone before populating
3352 * the graph, and all of the new non-subgraph services
3353 * need to be disabled here.
3354 */
3355 switch (err = libscf_set_enable_ovr(inst, 0)) {
3356 case 0:
3357 break;
3358
3359 case ECONNABORTED:
3360 case ECANCELED:
3361 startd_free(restarter_fmri, max_scf_value_size);
3362 return (err);
3363
3364 case EROFS:
3365 log_error(LOG_WARNING,
3366 "Could not set %s/%s for %s: %s.\n",
3367 SCF_PG_GENERAL_OVR, SCF_PROPERTY_ENABLED,
3368 v->gv_name, strerror(err));
3369 break;
3370
3371 case EPERM:
3372 uu_die("Permission denied.\n");
3373 /* NOTREACHED */
3374
3375 default:
3376 bad_error("libscf_set_enable_ovr", err);
3377 }
3378 } else {
3379 assert(v->gv_flags & GV_INSUBGRAPH);
3380 switch (err = libscf_delete_enable_ovr(inst)) {
3381 case 0:
3382 break;
3383
3384 case ECONNABORTED:
3385 case ECANCELED:
3386 startd_free(restarter_fmri, max_scf_value_size);
3387 return (err);
3388
3389 case EPERM:
3390 uu_die("Permission denied.\n");
3391 /* NOTREACHED */
3392
3393 default:
3394 bad_error("libscf_delete_enable_ovr", err);
3395 }
3396 }
3397 }
3398
3399 err = libscf_get_basic_instance_data(h, inst, v->gv_name, &enabled,
3400 &enabled_ovr, &restarter_fmri);
3401 switch (err) {
3402 case 0:
3403 break;
3404
3405 case ECONNABORTED:
3406 case ECANCELED:
3407 startd_free(restarter_fmri, max_scf_value_size);
3408 return (err);
3409
3410 case ENOENT:
3411 log_framework(LOG_DEBUG,
3412 "Ignoring %s because it has no general property group.\n",
3413 v->gv_name);
3414 startd_free(restarter_fmri, max_scf_value_size);
3415 return (0);
3416
3417 default:
3418 bad_error("libscf_get_basic_instance_data", err);
3419 }
3420
3421 if ((tset = libscf_get_stn_tset(inst)) == -1) {
3422 log_framework(LOG_WARNING,
3423 "Failed to get notification parameters for %s: %s\n",
3424 v->gv_name, scf_strerror(scf_error()));
3425 v->gv_stn_tset = 0;
3426 } else {
3427 v->gv_stn_tset = tset;
3428 }
3429 if (strcmp(v->gv_name, SCF_INSTANCE_GLOBAL) == 0)
3430 stn_global = v->gv_stn_tset;
3431
3432 if (enabled == -1) {
3433 startd_free(restarter_fmri, max_scf_value_size);
3434 return (0);
3435 }
3436
3437 v->gv_flags = (v->gv_flags & ~GV_ENBLD_NOOVR) |
3438 (enabled ? GV_ENBLD_NOOVR : 0);
3439
3440 if (enabled_ovr != -1)
3441 enabled = enabled_ovr;
3442
3443 v->gv_state = RESTARTER_STATE_UNINIT;
3444
3445 snap = libscf_get_or_make_running_snapshot(inst, v->gv_name, B_TRUE);
3446 scf_snapshot_destroy(snap);
3447
3448 /* Set up the restarter. (Sends _ADD_INSTANCE on success.) */
3449 err = graph_change_restarter(v, restarter_fmri, h, &path);
3450 if (err != 0) {
3451 instance_data_t idata;
3452 uint_t count = 0, msecs = ALLOC_DELAY;
3453 restarter_str_t reason;
3454
3455 if (err == ECONNABORTED) {
3456 startd_free(restarter_fmri, max_scf_value_size);
3457 return (err);
3458 }
3459
3460 assert(err == EINVAL || err == ELOOP);
3461
3462 if (err == EINVAL) {
3463 log_framework(LOG_ERR, emsg_invalid_restarter,
3464 v->gv_name, restarter_fmri);
3465 reason = restarter_str_invalid_restarter;
3466 } else {
3467 handle_cycle(v->gv_name, path);
3468 reason = restarter_str_dependency_cycle;
3469 }
3470
3471 startd_free(restarter_fmri, max_scf_value_size);
3472
3473 /*
3474 * We didn't register the instance with the restarter, so we
3475 * must set maintenance mode ourselves.
3476 */
3477 err = libscf_instance_get_fmri(inst, (char **)&idata.i_fmri);
3478 if (err != 0) {
3479 assert(err == ECONNABORTED || err == ECANCELED);
3480 return (err);
3481 }
3482
3483 idata.i_state = RESTARTER_STATE_NONE;
3484 idata.i_next_state = RESTARTER_STATE_NONE;
3485
3486 set_maint:
3487 switch (err = _restarter_commit_states(h, &idata,
3488 RESTARTER_STATE_MAINT, RESTARTER_STATE_NONE,
3489 restarter_get_str_short(reason))) {
3490 case 0:
3491 break;
3492
3493 case ENOMEM:
3494 ++count;
3495 if (count < ALLOC_RETRY) {
3496 (void) poll(NULL, 0, msecs);
3497 msecs *= ALLOC_DELAY_MULT;
3498 goto set_maint;
3499 }
3500
3501 uu_die("Insufficient memory.\n");
3502 /* NOTREACHED */
3503
3504 case ECONNABORTED:
3505 startd_free((void *)idata.i_fmri, max_scf_fmri_size);
3506 return (ECONNABORTED);
3507
3508 case ENOENT:
3509 startd_free((void *)idata.i_fmri, max_scf_fmri_size);
3510 return (ECANCELED);
3511
3512 case EPERM:
3513 case EACCES:
3514 case EROFS:
3515 log_error(LOG_NOTICE, "Could not initialize state for "
3516 "%s: %s.\n", idata.i_fmri, strerror(err));
3517 break;
3518
3519 case EINVAL:
3520 default:
3521 bad_error("_restarter_commit_states", err);
3522 }
3523
3524 startd_free((void *)idata.i_fmri, max_scf_fmri_size);
3525
3526 v->gv_state = RESTARTER_STATE_MAINT;
3527
3528 goto out;
3529 }
3530 startd_free(restarter_fmri, max_scf_value_size);
3531
3532 /* Add all the other dependencies. */
3533 err = refresh_vertex(v, inst);
3534 if (err != 0) {
3535 assert(err == ECONNABORTED);
3536 return (err);
3537 }
3538
3539 out:
3540 v->gv_flags |= GV_CONFIGURED;
3541
3542 graph_enable_by_vertex(v, enabled, 0);
3543
3544 return (0);
3545 }
3546
3547
3548 static void
kill_user_procs(void)3549 kill_user_procs(void)
3550 {
3551 (void) fputs("svc.startd: Killing user processes.\n", stdout);
3552
3553 /*
3554 * Despite its name, killall's role is to get select user processes--
3555 * basically those representing terminal-based logins-- to die. Victims
3556 * are located by killall in the utmp database. Since these are most
3557 * often shell based logins, and many shells mask SIGTERM (but are
3558 * responsive to SIGHUP) we first HUP and then shortly thereafter
3559 * kill -9.
3560 */
3561 (void) fork_with_timeout("/usr/sbin/killall HUP", 1, 5);
3562 (void) fork_with_timeout("/usr/sbin/killall KILL", 1, 5);
3563
3564 /*
3565 * Note the selection of user id's 0, 1 and 15, subsequently
3566 * inverted by -v. 15 is reserved for dladmd. Yes, this is a
3567 * kludge-- a better policy is needed.
3568 *
3569 * Note that fork_with_timeout will only wait out the 1 second
3570 * "grace time" if pkill actually returns 0. So if there are
3571 * no matches, this will run to completion much more quickly.
3572 */
3573 (void) fork_with_timeout("/usr/bin/pkill -TERM -v -u 0,1,15", 1, 5);
3574 (void) fork_with_timeout("/usr/bin/pkill -KILL -v -u 0,1,15", 1, 5);
3575 }
3576
3577 static void
do_uadmin(void)3578 do_uadmin(void)
3579 {
3580 const char * const resetting = "/etc/svc/volatile/resetting";
3581 int fd;
3582 struct statvfs vfs;
3583 time_t now;
3584 struct tm nowtm;
3585 char down_buf[256], time_buf[256];
3586 uintptr_t mdep;
3587 #if defined(__i386)
3588 grub_boot_args_t fbarg;
3589 #endif /* __i386 */
3590
3591 mdep = NULL;
3592 fd = creat(resetting, 0777);
3593 if (fd >= 0)
3594 startd_close(fd);
3595 else
3596 uu_warn("Could not create \"%s\"", resetting);
3597
3598 /* Kill dhcpagent if we're not using nfs for root */
3599 if ((statvfs("/", &vfs) == 0) &&
3600 (strncmp(vfs.f_basetype, "nfs", sizeof ("nfs") - 1) != 0))
3601 fork_with_timeout("/usr/bin/pkill -x -u 0 dhcpagent", 0, 5);
3602
3603 /*
3604 * Call sync(2) now, before we kill off user processes. This takes
3605 * advantage of the several seconds of pause we have before the
3606 * killalls are done. Time we can make good use of to get pages
3607 * moving out to disk.
3608 *
3609 * Inside non-global zones, we don't bother, and it's better not to
3610 * anyway, since sync(2) can have system-wide impact.
3611 */
3612 if (getzoneid() == 0)
3613 sync();
3614
3615 kill_user_procs();
3616
3617 /*
3618 * Note that this must come after the killing of user procs, since
3619 * killall relies on utmpx, and this command affects the contents of
3620 * said file.
3621 */
3622 if (access("/usr/lib/acct/closewtmp", X_OK) == 0)
3623 fork_with_timeout("/usr/lib/acct/closewtmp", 0, 5);
3624
3625 /*
3626 * For patches which may be installed as the system is shutting
3627 * down, we need to ensure, one more time, that the boot archive
3628 * really is up to date.
3629 */
3630 if (getzoneid() == 0 && access("/usr/sbin/bootadm", X_OK) == 0)
3631 fork_with_timeout("/usr/sbin/bootadm -ea update_all", 0, 3600);
3632
3633 /*
3634 * Right now, fast reboot is supported only on i386.
3635 * scf_is_fastboot_default() should take care of it.
3636 * If somehow we got there on unsupported platform -
3637 * print warning and fall back to regular reboot.
3638 */
3639 if (halting == AD_FASTREBOOT) {
3640 #if defined(__i386)
3641 int rc;
3642
3643 if ((rc = grub_get_boot_args(&fbarg, NULL,
3644 GRUB_ENTRY_DEFAULT)) == 0) {
3645 mdep = (uintptr_t)&fbarg.gba_bootargs;
3646 } else {
3647 /*
3648 * Failed to read GRUB menu, fall back to normal reboot
3649 */
3650 halting = AD_BOOT;
3651 uu_warn("Failed to process GRUB menu entry "
3652 "for fast reboot.\n\t%s\n"
3653 "Falling back to regular reboot.\n",
3654 grub_strerror(rc));
3655 }
3656 #else /* __i386 */
3657 halting = AD_BOOT;
3658 uu_warn("Fast reboot configured, but not supported by "
3659 "this ISA\n");
3660 #endif /* __i386 */
3661 }
3662
3663 fork_with_timeout("/sbin/umountall -l", 0, 5);
3664 fork_with_timeout("/sbin/umount /tmp /var/adm /var/run /var "
3665 ">/dev/null 2>&1", 0, 5);
3666
3667 /*
3668 * Try to get to consistency for whatever UFS filesystems are left.
3669 * This is pretty expensive, so we save it for the end in the hopes of
3670 * minimizing what it must do. The other option would be to start in
3671 * parallel with the killall's, but lockfs tends to throw out much more
3672 * than is needed, and so subsequent commands (like umountall) take a
3673 * long time to get going again.
3674 *
3675 * Inside of zones, we don't bother, since we're not about to terminate
3676 * the whole OS instance.
3677 *
3678 * On systems using only ZFS, this call to lockfs -fa is a no-op.
3679 */
3680 if (getzoneid() == 0) {
3681 if (access("/usr/sbin/lockfs", X_OK) == 0)
3682 fork_with_timeout("/usr/sbin/lockfs -fa", 0, 30);
3683
3684 sync(); /* once more, with feeling */
3685 }
3686
3687 fork_with_timeout("/sbin/umount /usr >/dev/null 2>&1", 0, 5);
3688
3689 /*
3690 * Construct and emit the last words from userland:
3691 * "<timestamp> The system is down. Shutdown took <N> seconds."
3692 *
3693 * Normally we'd use syslog, but with /var and other things
3694 * potentially gone, try to minimize the external dependencies.
3695 */
3696 now = time(NULL);
3697 (void) localtime_r(&now, &nowtm);
3698
3699 if (strftime(down_buf, sizeof (down_buf),
3700 "%b %e %T The system is down.", &nowtm) == 0) {
3701 (void) strlcpy(down_buf, "The system is down.",
3702 sizeof (down_buf));
3703 }
3704
3705 if (halting_time != 0 && halting_time <= now) {
3706 (void) snprintf(time_buf, sizeof (time_buf),
3707 " Shutdown took %lu seconds.", now - halting_time);
3708 } else {
3709 time_buf[0] = '\0';
3710 }
3711 (void) printf("%s%s\n", down_buf, time_buf);
3712
3713 (void) uadmin(A_SHUTDOWN, halting, mdep);
3714 uu_warn("uadmin() failed");
3715
3716 #if defined(__i386)
3717 /* uadmin fail, cleanup grub_boot_args */
3718 if (halting == AD_FASTREBOOT)
3719 grub_cleanup_boot_args(&fbarg);
3720 #endif /* __i386 */
3721
3722 if (remove(resetting) != 0 && errno != ENOENT)
3723 uu_warn("Could not remove \"%s\"", resetting);
3724 }
3725
3726 /*
3727 * If any of the up_svcs[] are online or satisfiable, return true. If they are
3728 * all missing, disabled, in maintenance, or unsatisfiable, return false.
3729 */
3730 boolean_t
can_come_up(void)3731 can_come_up(void)
3732 {
3733 int i;
3734
3735 assert(MUTEX_HELD(&dgraph_lock));
3736
3737 /*
3738 * If we are booting to single user (boot -s),
3739 * SCF_MILESTONE_SINGLE_USER is needed to come up because startd
3740 * spawns sulogin after single-user is online (see specials.c).
3741 */
3742 i = (booting_to_single_user ? 0 : 1);
3743
3744 for (; up_svcs[i] != NULL; ++i) {
3745 if (up_svcs_p[i] == NULL) {
3746 up_svcs_p[i] = vertex_get_by_name(up_svcs[i]);
3747
3748 if (up_svcs_p[i] == NULL)
3749 continue;
3750 }
3751
3752 /*
3753 * Ignore unconfigured services (the ones that have been
3754 * mentioned in a dependency from other services, but do
3755 * not exist in the repository). Services which exist
3756 * in the repository but don't have general/enabled
3757 * property will be also ignored.
3758 */
3759 if (!(up_svcs_p[i]->gv_flags & GV_CONFIGURED))
3760 continue;
3761
3762 switch (up_svcs_p[i]->gv_state) {
3763 case RESTARTER_STATE_ONLINE:
3764 case RESTARTER_STATE_DEGRADED:
3765 /*
3766 * Deactivate verbose boot once a login service has been
3767 * reached.
3768 */
3769 st->st_log_login_reached = 1;
3770 /*FALLTHROUGH*/
3771 case RESTARTER_STATE_UNINIT:
3772 return (B_TRUE);
3773
3774 case RESTARTER_STATE_OFFLINE:
3775 if (instance_satisfied(up_svcs_p[i], B_TRUE) != -1)
3776 return (B_TRUE);
3777 log_framework(LOG_DEBUG,
3778 "can_come_up(): %s is unsatisfiable.\n",
3779 up_svcs_p[i]->gv_name);
3780 continue;
3781
3782 case RESTARTER_STATE_DISABLED:
3783 case RESTARTER_STATE_MAINT:
3784 log_framework(LOG_DEBUG,
3785 "can_come_up(): %s is in state %s.\n",
3786 up_svcs_p[i]->gv_name,
3787 instance_state_str[up_svcs_p[i]->gv_state]);
3788 continue;
3789
3790 default:
3791 #ifndef NDEBUG
3792 uu_warn("%s:%d: Unexpected vertex state %d.\n",
3793 __FILE__, __LINE__, up_svcs_p[i]->gv_state);
3794 #endif
3795 abort();
3796 }
3797 }
3798
3799 /*
3800 * In the seed repository, console-login is unsatisfiable because
3801 * services are missing. To behave correctly in that case we don't want
3802 * to return false until manifest-import is online.
3803 */
3804
3805 if (manifest_import_p == NULL) {
3806 manifest_import_p = vertex_get_by_name(manifest_import);
3807
3808 if (manifest_import_p == NULL)
3809 return (B_FALSE);
3810 }
3811
3812 switch (manifest_import_p->gv_state) {
3813 case RESTARTER_STATE_ONLINE:
3814 case RESTARTER_STATE_DEGRADED:
3815 case RESTARTER_STATE_DISABLED:
3816 case RESTARTER_STATE_MAINT:
3817 break;
3818
3819 case RESTARTER_STATE_OFFLINE:
3820 if (instance_satisfied(manifest_import_p, B_TRUE) == -1)
3821 break;
3822 /* FALLTHROUGH */
3823
3824 case RESTARTER_STATE_UNINIT:
3825 return (B_TRUE);
3826 }
3827
3828 return (B_FALSE);
3829 }
3830
3831 /*
3832 * Runs sulogin. Returns
3833 * 0 - success
3834 * EALREADY - sulogin is already running
3835 * EBUSY - console-login is running
3836 */
3837 static int
run_sulogin(const char * msg)3838 run_sulogin(const char *msg)
3839 {
3840 graph_vertex_t *v;
3841
3842 assert(MUTEX_HELD(&dgraph_lock));
3843
3844 if (sulogin_running)
3845 return (EALREADY);
3846
3847 v = vertex_get_by_name(console_login_fmri);
3848 if (v != NULL && inst_running(v))
3849 return (EBUSY);
3850
3851 sulogin_running = B_TRUE;
3852
3853 MUTEX_UNLOCK(&dgraph_lock);
3854
3855 fork_sulogin(B_FALSE, msg);
3856
3857 MUTEX_LOCK(&dgraph_lock);
3858
3859 sulogin_running = B_FALSE;
3860
3861 if (console_login_ready) {
3862 v = vertex_get_by_name(console_login_fmri);
3863
3864 if (v != NULL && v->gv_state == RESTARTER_STATE_OFFLINE) {
3865 if (v->gv_start_f == NULL)
3866 vertex_send_event(v,
3867 RESTARTER_EVENT_TYPE_START);
3868 else
3869 v->gv_start_f(v);
3870 }
3871
3872 console_login_ready = B_FALSE;
3873 }
3874
3875 return (0);
3876 }
3877
3878 /*
3879 * The sulogin thread runs sulogin while can_come_up() is false. run_sulogin()
3880 * keeps sulogin from stepping on console-login's toes.
3881 */
3882 /* ARGSUSED */
3883 static void *
sulogin_thread(void * unused)3884 sulogin_thread(void *unused)
3885 {
3886 MUTEX_LOCK(&dgraph_lock);
3887
3888 assert(sulogin_thread_running);
3889
3890 do {
3891 (void) run_sulogin("Console login service(s) cannot run\n");
3892 } while (!can_come_up());
3893
3894 sulogin_thread_running = B_FALSE;
3895 MUTEX_UNLOCK(&dgraph_lock);
3896
3897 return (NULL);
3898 }
3899
3900 /* ARGSUSED */
3901 void *
single_user_thread(void * unused)3902 single_user_thread(void *unused)
3903 {
3904 uint_t left;
3905 scf_handle_t *h;
3906 scf_instance_t *inst;
3907 scf_property_t *prop;
3908 scf_value_t *val;
3909 const char *msg;
3910 char *buf;
3911 int r;
3912
3913 MUTEX_LOCK(&single_user_thread_lock);
3914 single_user_thread_count++;
3915
3916 if (!booting_to_single_user)
3917 kill_user_procs();
3918
3919 if (go_single_user_mode || booting_to_single_user) {
3920 msg = "SINGLE USER MODE\n";
3921 } else {
3922 assert(go_to_level1);
3923
3924 fork_rc_script('1', "start", B_TRUE);
3925
3926 uu_warn("The system is ready for administration.\n");
3927
3928 msg = "";
3929 }
3930
3931 MUTEX_UNLOCK(&single_user_thread_lock);
3932
3933 for (;;) {
3934 MUTEX_LOCK(&dgraph_lock);
3935 r = run_sulogin(msg);
3936 MUTEX_UNLOCK(&dgraph_lock);
3937 if (r == 0)
3938 break;
3939
3940 assert(r == EALREADY || r == EBUSY);
3941
3942 left = 3;
3943 while (left > 0)
3944 left = sleep(left);
3945 }
3946
3947 MUTEX_LOCK(&single_user_thread_lock);
3948
3949 /*
3950 * If another single user thread has started, let it finish changing
3951 * the run level.
3952 */
3953 if (single_user_thread_count > 1) {
3954 single_user_thread_count--;
3955 MUTEX_UNLOCK(&single_user_thread_lock);
3956 return (NULL);
3957 }
3958
3959 h = libscf_handle_create_bound_loop();
3960 inst = scf_instance_create(h);
3961 prop = safe_scf_property_create(h);
3962 val = safe_scf_value_create(h);
3963 buf = startd_alloc(max_scf_fmri_size);
3964
3965 lookup:
3966 if (scf_handle_decode_fmri(h, SCF_SERVICE_STARTD, NULL, NULL, inst,
3967 NULL, NULL, SCF_DECODE_FMRI_EXACT) != 0) {
3968 switch (scf_error()) {
3969 case SCF_ERROR_NOT_FOUND:
3970 r = libscf_create_self(h);
3971 if (r == 0)
3972 goto lookup;
3973 assert(r == ECONNABORTED);
3974 /* FALLTHROUGH */
3975
3976 case SCF_ERROR_CONNECTION_BROKEN:
3977 libscf_handle_rebind(h);
3978 goto lookup;
3979
3980 case SCF_ERROR_INVALID_ARGUMENT:
3981 case SCF_ERROR_CONSTRAINT_VIOLATED:
3982 case SCF_ERROR_NOT_BOUND:
3983 case SCF_ERROR_HANDLE_MISMATCH:
3984 default:
3985 bad_error("scf_handle_decode_fmri", scf_error());
3986 }
3987 }
3988
3989 MUTEX_LOCK(&dgraph_lock);
3990
3991 r = scf_instance_delete_prop(inst, SCF_PG_OPTIONS_OVR,
3992 SCF_PROPERTY_MILESTONE);
3993 switch (r) {
3994 case 0:
3995 case ECANCELED:
3996 break;
3997
3998 case ECONNABORTED:
3999 MUTEX_UNLOCK(&dgraph_lock);
4000 libscf_handle_rebind(h);
4001 goto lookup;
4002
4003 case EPERM:
4004 case EACCES:
4005 case EROFS:
4006 log_error(LOG_WARNING, "Could not clear temporary milestone: "
4007 "%s.\n", strerror(r));
4008 break;
4009
4010 default:
4011 bad_error("scf_instance_delete_prop", r);
4012 }
4013
4014 MUTEX_UNLOCK(&dgraph_lock);
4015
4016 r = libscf_get_milestone(inst, prop, val, buf, max_scf_fmri_size);
4017 switch (r) {
4018 case ECANCELED:
4019 case ENOENT:
4020 case EINVAL:
4021 (void) strcpy(buf, "all");
4022 /* FALLTHROUGH */
4023
4024 case 0:
4025 uu_warn("Returning to milestone %s.\n", buf);
4026 break;
4027
4028 case ECONNABORTED:
4029 libscf_handle_rebind(h);
4030 goto lookup;
4031
4032 default:
4033 bad_error("libscf_get_milestone", r);
4034 }
4035
4036 r = dgraph_set_milestone(buf, h, B_FALSE);
4037 switch (r) {
4038 case 0:
4039 case ECONNRESET:
4040 case EALREADY:
4041 case EINVAL:
4042 case ENOENT:
4043 break;
4044
4045 default:
4046 bad_error("dgraph_set_milestone", r);
4047 }
4048
4049 /*
4050 * See graph_runlevel_changed().
4051 */
4052 MUTEX_LOCK(&dgraph_lock);
4053 utmpx_set_runlevel(target_milestone_as_runlevel(), 'S', B_TRUE);
4054 MUTEX_UNLOCK(&dgraph_lock);
4055
4056 startd_free(buf, max_scf_fmri_size);
4057 scf_value_destroy(val);
4058 scf_property_destroy(prop);
4059 scf_instance_destroy(inst);
4060 scf_handle_destroy(h);
4061
4062 /*
4063 * We'll give ourselves 3 seconds to respond to all of the enablings
4064 * that setting the milestone should have created before checking
4065 * whether to run sulogin.
4066 */
4067 left = 3;
4068 while (left > 0)
4069 left = sleep(left);
4070
4071 MUTEX_LOCK(&dgraph_lock);
4072 /*
4073 * Clearing these variables will allow the sulogin thread to run. We
4074 * check here in case there aren't any more state updates anytime soon.
4075 */
4076 go_to_level1 = go_single_user_mode = booting_to_single_user = B_FALSE;
4077 if (!sulogin_thread_running && !can_come_up()) {
4078 (void) startd_thread_create(sulogin_thread, NULL);
4079 sulogin_thread_running = B_TRUE;
4080 }
4081 MUTEX_UNLOCK(&dgraph_lock);
4082 single_user_thread_count--;
4083 MUTEX_UNLOCK(&single_user_thread_lock);
4084 return (NULL);
4085 }
4086
4087
4088 /*
4089 * Dependency graph operations API. These are handle-independent thread-safe
4090 * graph manipulation functions which are the entry points for the event
4091 * threads below.
4092 */
4093
4094 /*
4095 * If a configured vertex exists for inst_fmri, return EEXIST. If no vertex
4096 * exists for inst_fmri, add one. Then fetch the restarter from inst, make
4097 * this vertex dependent on it, and send _ADD_INSTANCE to the restarter.
4098 * Fetch whether the instance should be enabled from inst and send _ENABLE or
4099 * _DISABLE as appropriate. Finally rummage through inst's dependency
4100 * property groups and add vertices and edges as appropriate. If anything
4101 * goes wrong after sending _ADD_INSTANCE, send _ADMIN_MAINT_ON to put the
4102 * instance in maintenance. Don't send _START or _STOP until we get a state
4103 * update in case we're being restarted and the service is already running.
4104 *
4105 * To support booting to a milestone, we must also make sure all dependencies
4106 * encountered are configured, if they exist in the repository.
4107 *
4108 * Returns 0 on success, ECONNABORTED on repository disconnection, EINVAL if
4109 * inst_fmri is an invalid (or not canonical) FMRI, ECANCELED if inst is
4110 * deleted, or EEXIST if a configured vertex for inst_fmri already exists.
4111 */
4112 int
dgraph_add_instance(const char * inst_fmri,scf_instance_t * inst,boolean_t lock_graph)4113 dgraph_add_instance(const char *inst_fmri, scf_instance_t *inst,
4114 boolean_t lock_graph)
4115 {
4116 graph_vertex_t *v;
4117 int err;
4118
4119 if (strcmp(inst_fmri, SCF_SERVICE_STARTD) == 0)
4120 return (0);
4121
4122 /* Check for a vertex for inst_fmri. */
4123 if (lock_graph) {
4124 MUTEX_LOCK(&dgraph_lock);
4125 } else {
4126 assert(MUTEX_HELD(&dgraph_lock));
4127 }
4128
4129 v = vertex_get_by_name(inst_fmri);
4130
4131 if (v != NULL) {
4132 assert(v->gv_type == GVT_INST);
4133
4134 if (v->gv_flags & GV_CONFIGURED) {
4135 if (lock_graph)
4136 MUTEX_UNLOCK(&dgraph_lock);
4137 return (EEXIST);
4138 }
4139 } else {
4140 /* Add the vertex. */
4141 err = graph_insert_vertex_unconfigured(inst_fmri, GVT_INST, 0,
4142 RERR_NONE, &v);
4143 if (err != 0) {
4144 assert(err == EINVAL);
4145 if (lock_graph)
4146 MUTEX_UNLOCK(&dgraph_lock);
4147 return (EINVAL);
4148 }
4149 }
4150
4151 err = configure_vertex(v, inst);
4152
4153 if (lock_graph)
4154 MUTEX_UNLOCK(&dgraph_lock);
4155
4156 return (err);
4157 }
4158
4159 /*
4160 * Locate the vertex for this property group's instance. If it doesn't exist
4161 * or is unconfigured, call dgraph_add_instance() & return. Otherwise fetch
4162 * the restarter for the instance, and if it has changed, send
4163 * _REMOVE_INSTANCE to the old restarter, remove the dependency, make sure the
4164 * new restarter has a vertex, add a new dependency, and send _ADD_INSTANCE to
4165 * the new restarter. Then fetch whether the instance should be enabled, and
4166 * if it is different from what we had, or if we changed the restarter, send
4167 * the appropriate _ENABLE or _DISABLE command.
4168 *
4169 * Returns 0 on success, ENOTSUP if the pg's parent is not an instance,
4170 * ECONNABORTED on repository disconnection, ECANCELED if the instance is
4171 * deleted, or -1 if the instance's general property group is deleted or if
4172 * its enabled property is misconfigured.
4173 */
4174 static int
dgraph_update_general(scf_propertygroup_t * pg)4175 dgraph_update_general(scf_propertygroup_t *pg)
4176 {
4177 scf_handle_t *h;
4178 scf_instance_t *inst;
4179 char *fmri;
4180 char *restarter_fmri;
4181 graph_vertex_t *v;
4182 int err;
4183 int enabled, enabled_ovr;
4184 int oldflags;
4185
4186 /* Find the vertex for this service */
4187 h = scf_pg_handle(pg);
4188
4189 inst = safe_scf_instance_create(h);
4190
4191 if (scf_pg_get_parent_instance(pg, inst) != 0) {
4192 switch (scf_error()) {
4193 case SCF_ERROR_CONSTRAINT_VIOLATED:
4194 return (ENOTSUP);
4195
4196 case SCF_ERROR_CONNECTION_BROKEN:
4197 default:
4198 return (ECONNABORTED);
4199
4200 case SCF_ERROR_DELETED:
4201 return (0);
4202
4203 case SCF_ERROR_NOT_SET:
4204 bad_error("scf_pg_get_parent_instance", scf_error());
4205 }
4206 }
4207
4208 err = libscf_instance_get_fmri(inst, &fmri);
4209 switch (err) {
4210 case 0:
4211 break;
4212
4213 case ECONNABORTED:
4214 scf_instance_destroy(inst);
4215 return (ECONNABORTED);
4216
4217 case ECANCELED:
4218 scf_instance_destroy(inst);
4219 return (0);
4220
4221 default:
4222 bad_error("libscf_instance_get_fmri", err);
4223 }
4224
4225 log_framework(LOG_DEBUG,
4226 "Graph engine: Reloading general properties for %s.\n", fmri);
4227
4228 MUTEX_LOCK(&dgraph_lock);
4229
4230 v = vertex_get_by_name(fmri);
4231 if (v == NULL || !(v->gv_flags & GV_CONFIGURED)) {
4232 /* Will get the up-to-date properties. */
4233 MUTEX_UNLOCK(&dgraph_lock);
4234 err = dgraph_add_instance(fmri, inst, B_TRUE);
4235 startd_free(fmri, max_scf_fmri_size);
4236 scf_instance_destroy(inst);
4237 return (err == ECANCELED ? 0 : err);
4238 }
4239
4240 /* Read enabled & restarter from repository. */
4241 restarter_fmri = startd_alloc(max_scf_value_size);
4242 err = libscf_get_basic_instance_data(h, inst, v->gv_name, &enabled,
4243 &enabled_ovr, &restarter_fmri);
4244 if (err != 0 || enabled == -1) {
4245 MUTEX_UNLOCK(&dgraph_lock);
4246 scf_instance_destroy(inst);
4247 startd_free(fmri, max_scf_fmri_size);
4248
4249 switch (err) {
4250 case ENOENT:
4251 case 0:
4252 startd_free(restarter_fmri, max_scf_value_size);
4253 return (-1);
4254
4255 case ECONNABORTED:
4256 case ECANCELED:
4257 startd_free(restarter_fmri, max_scf_value_size);
4258 return (err);
4259
4260 default:
4261 bad_error("libscf_get_basic_instance_data", err);
4262 }
4263 }
4264
4265 oldflags = v->gv_flags;
4266 v->gv_flags = (v->gv_flags & ~GV_ENBLD_NOOVR) |
4267 (enabled ? GV_ENBLD_NOOVR : 0);
4268
4269 if (enabled_ovr != -1)
4270 enabled = enabled_ovr;
4271
4272 /*
4273 * If GV_ENBLD_NOOVR has changed, then we need to re-evaluate the
4274 * subgraph.
4275 */
4276 if (milestone > MILESTONE_NONE && v->gv_flags != oldflags)
4277 (void) eval_subgraph(v, h);
4278
4279 scf_instance_destroy(inst);
4280
4281 /* Ignore restarter change for now. */
4282
4283 startd_free(restarter_fmri, max_scf_value_size);
4284 startd_free(fmri, max_scf_fmri_size);
4285
4286 /*
4287 * Always send _ENABLE or _DISABLE. We could avoid this if the
4288 * restarter didn't change and the enabled value didn't change, but
4289 * that's not easy to check and improbable anyway, so we'll just do
4290 * this.
4291 */
4292 graph_enable_by_vertex(v, enabled, 1);
4293
4294 MUTEX_UNLOCK(&dgraph_lock);
4295
4296 return (0);
4297 }
4298
4299 /*
4300 * Delete all of the property group dependencies of v, update inst's running
4301 * snapshot, and add the dependencies in the new snapshot. If any of the new
4302 * dependencies would create a cycle, send _ADMIN_MAINT_ON. Otherwise
4303 * reevaluate v's dependencies, send _START or _STOP as appropriate, and do
4304 * the same for v's dependents.
4305 *
4306 * Returns
4307 * 0 - success
4308 * ECONNABORTED - repository connection broken
4309 * ECANCELED - inst was deleted
4310 * EINVAL - inst is invalid (e.g., missing general/enabled)
4311 * -1 - libscf_snapshots_refresh() failed
4312 */
4313 static int
dgraph_refresh_instance(graph_vertex_t * v,scf_instance_t * inst)4314 dgraph_refresh_instance(graph_vertex_t *v, scf_instance_t *inst)
4315 {
4316 int r;
4317 int enabled;
4318 int32_t tset;
4319
4320 assert(MUTEX_HELD(&dgraph_lock));
4321 assert(v->gv_type == GVT_INST);
4322
4323 /* Only refresh services with valid general/enabled properties. */
4324 r = libscf_get_basic_instance_data(scf_instance_handle(inst), inst,
4325 v->gv_name, &enabled, NULL, NULL);
4326 switch (r) {
4327 case 0:
4328 break;
4329
4330 case ECONNABORTED:
4331 case ECANCELED:
4332 return (r);
4333
4334 case ENOENT:
4335 log_framework(LOG_DEBUG,
4336 "Ignoring %s because it has no general property group.\n",
4337 v->gv_name);
4338 return (EINVAL);
4339
4340 default:
4341 bad_error("libscf_get_basic_instance_data", r);
4342 }
4343
4344 if ((tset = libscf_get_stn_tset(inst)) == -1) {
4345 log_framework(LOG_WARNING,
4346 "Failed to get notification parameters for %s: %s\n",
4347 v->gv_name, scf_strerror(scf_error()));
4348 tset = 0;
4349 }
4350 v->gv_stn_tset = tset;
4351 if (strcmp(v->gv_name, SCF_INSTANCE_GLOBAL) == 0)
4352 stn_global = tset;
4353
4354 if (enabled == -1)
4355 return (EINVAL);
4356
4357 r = libscf_snapshots_refresh(inst, v->gv_name);
4358 if (r != 0) {
4359 if (r != -1)
4360 bad_error("libscf_snapshots_refresh", r);
4361
4362 /* error logged */
4363 return (r);
4364 }
4365
4366 r = refresh_vertex(v, inst);
4367 if (r != 0 && r != ECONNABORTED)
4368 bad_error("refresh_vertex", r);
4369 return (r);
4370 }
4371
4372 /*
4373 * Returns true only if none of this service's dependents are 'up' -- online
4374 * or degraded (offline is considered down in this situation). This function
4375 * is somehow similar to is_nonsubgraph_leaf() but works on subtrees.
4376 */
4377 static boolean_t
insubtree_dependents_down(graph_vertex_t * v)4378 insubtree_dependents_down(graph_vertex_t *v)
4379 {
4380 graph_vertex_t *vv;
4381 graph_edge_t *e;
4382
4383 assert(MUTEX_HELD(&dgraph_lock));
4384
4385 for (e = uu_list_first(v->gv_dependents); e != NULL;
4386 e = uu_list_next(v->gv_dependents, e)) {
4387 vv = e->ge_vertex;
4388 if (vv->gv_type == GVT_INST) {
4389 if ((vv->gv_flags & GV_CONFIGURED) == 0)
4390 continue;
4391
4392 if ((vv->gv_flags & GV_TOOFFLINE) == 0)
4393 continue;
4394
4395 if ((vv->gv_state == RESTARTER_STATE_ONLINE) ||
4396 (vv->gv_state == RESTARTER_STATE_DEGRADED))
4397 return (B_FALSE);
4398 } else {
4399 /*
4400 * Skip all excluded and optional_all dependencies
4401 * and decide whether to offline the service based
4402 * on restart_on attribute.
4403 */
4404 if (is_depgrp_bypassed(vv))
4405 continue;
4406
4407 /*
4408 * For dependency groups or service vertices, keep
4409 * traversing to see if instances are running.
4410 */
4411 if (insubtree_dependents_down(vv) == B_FALSE)
4412 return (B_FALSE);
4413 }
4414 }
4415
4416 return (B_TRUE);
4417 }
4418
4419 /*
4420 * Returns true only if none of this service's dependents are 'up' -- online,
4421 * degraded, or offline.
4422 */
4423 static int
is_nonsubgraph_leaf(graph_vertex_t * v)4424 is_nonsubgraph_leaf(graph_vertex_t *v)
4425 {
4426 graph_vertex_t *vv;
4427 graph_edge_t *e;
4428
4429 assert(MUTEX_HELD(&dgraph_lock));
4430
4431 for (e = uu_list_first(v->gv_dependents);
4432 e != NULL;
4433 e = uu_list_next(v->gv_dependents, e)) {
4434
4435 vv = e->ge_vertex;
4436 if (vv->gv_type == GVT_INST) {
4437 if ((vv->gv_flags & GV_CONFIGURED) == 0)
4438 continue;
4439
4440 if (vv->gv_flags & GV_INSUBGRAPH)
4441 continue;
4442
4443 if (up_state(vv->gv_state))
4444 return (0);
4445 } else {
4446 /*
4447 * For dependency group or service vertices, keep
4448 * traversing to see if instances are running.
4449 *
4450 * We should skip exclude_all dependencies otherwise
4451 * the vertex will never be considered as a leaf
4452 * if the dependent is offline. The main reason for
4453 * this is that disable_nonsubgraph_leaves() skips
4454 * exclusion dependencies.
4455 */
4456 if (vv->gv_type == GVT_GROUP &&
4457 vv->gv_depgroup == DEPGRP_EXCLUDE_ALL)
4458 continue;
4459
4460 if (!is_nonsubgraph_leaf(vv))
4461 return (0);
4462 }
4463 }
4464
4465 return (1);
4466 }
4467
4468 /*
4469 * Disable v temporarily. Attempt to do this by setting its enabled override
4470 * property in the repository. If that fails, send a _DISABLE command.
4471 * Returns 0 on success and ECONNABORTED if the repository connection is
4472 * broken.
4473 */
4474 static int
disable_service_temporarily(graph_vertex_t * v,scf_handle_t * h)4475 disable_service_temporarily(graph_vertex_t *v, scf_handle_t *h)
4476 {
4477 const char * const emsg = "Could not temporarily disable %s because "
4478 "%s. Will stop service anyways. Repository status for the "
4479 "service may be inaccurate.\n";
4480 const char * const emsg_cbroken =
4481 "the repository connection was broken";
4482
4483 scf_instance_t *inst;
4484 int r;
4485
4486 inst = scf_instance_create(h);
4487 if (inst == NULL) {
4488 char buf[100];
4489
4490 (void) snprintf(buf, sizeof (buf),
4491 "scf_instance_create() failed (%s)",
4492 scf_strerror(scf_error()));
4493 log_error(LOG_WARNING, emsg, v->gv_name, buf);
4494
4495 graph_enable_by_vertex(v, 0, 0);
4496 return (0);
4497 }
4498
4499 r = scf_handle_decode_fmri(h, v->gv_name, NULL, NULL, inst,
4500 NULL, NULL, SCF_DECODE_FMRI_EXACT);
4501 if (r != 0) {
4502 switch (scf_error()) {
4503 case SCF_ERROR_CONNECTION_BROKEN:
4504 log_error(LOG_WARNING, emsg, v->gv_name, emsg_cbroken);
4505 graph_enable_by_vertex(v, 0, 0);
4506 return (ECONNABORTED);
4507
4508 case SCF_ERROR_NOT_FOUND:
4509 return (0);
4510
4511 case SCF_ERROR_HANDLE_MISMATCH:
4512 case SCF_ERROR_INVALID_ARGUMENT:
4513 case SCF_ERROR_CONSTRAINT_VIOLATED:
4514 case SCF_ERROR_NOT_BOUND:
4515 default:
4516 bad_error("scf_handle_decode_fmri",
4517 scf_error());
4518 }
4519 }
4520
4521 r = libscf_set_enable_ovr(inst, 0);
4522 switch (r) {
4523 case 0:
4524 scf_instance_destroy(inst);
4525 return (0);
4526
4527 case ECANCELED:
4528 scf_instance_destroy(inst);
4529 return (0);
4530
4531 case ECONNABORTED:
4532 log_error(LOG_WARNING, emsg, v->gv_name, emsg_cbroken);
4533 graph_enable_by_vertex(v, 0, 0);
4534 return (ECONNABORTED);
4535
4536 case EPERM:
4537 log_error(LOG_WARNING, emsg, v->gv_name,
4538 "the repository denied permission");
4539 graph_enable_by_vertex(v, 0, 0);
4540 return (0);
4541
4542 case EROFS:
4543 log_error(LOG_WARNING, emsg, v->gv_name,
4544 "the repository is read-only");
4545 graph_enable_by_vertex(v, 0, 0);
4546 return (0);
4547
4548 default:
4549 bad_error("libscf_set_enable_ovr", r);
4550 /* NOTREACHED */
4551 }
4552 }
4553
4554 /*
4555 * Of the transitive instance dependencies of v, offline those which are
4556 * in the subtree and which are leaves (i.e., have no dependents which are
4557 * "up").
4558 */
4559 void
offline_subtree_leaves(graph_vertex_t * v,void * arg)4560 offline_subtree_leaves(graph_vertex_t *v, void *arg)
4561 {
4562 assert(MUTEX_HELD(&dgraph_lock));
4563
4564 /* If v isn't an instance, recurse on its dependencies. */
4565 if (v->gv_type != GVT_INST) {
4566 graph_walk_dependencies(v, offline_subtree_leaves, arg);
4567 return;
4568 }
4569
4570 /*
4571 * If v is not in the subtree, so should all of its dependencies,
4572 * so do nothing.
4573 */
4574 if ((v->gv_flags & GV_TOOFFLINE) == 0)
4575 return;
4576
4577 /* If v isn't a leaf because it's already down, recurse. */
4578 if (!up_state(v->gv_state)) {
4579 graph_walk_dependencies(v, offline_subtree_leaves, arg);
4580 return;
4581 }
4582
4583 /* if v is a leaf, offline it or disable it if it's the last one */
4584 if (insubtree_dependents_down(v) == B_TRUE) {
4585 if (v->gv_flags & GV_TODISABLE)
4586 vertex_send_event(v,
4587 RESTARTER_EVENT_TYPE_ADMIN_DISABLE);
4588 else
4589 offline_vertex(v);
4590 }
4591 }
4592
4593 void
graph_offline_subtree_leaves(graph_vertex_t * v,void * h)4594 graph_offline_subtree_leaves(graph_vertex_t *v, void *h)
4595 {
4596 graph_walk_dependencies(v, offline_subtree_leaves, (void *)h);
4597 }
4598
4599
4600 /*
4601 * Of the transitive instance dependencies of v, disable those which are not
4602 * in the subgraph and which are leaves (i.e., have no dependents which are
4603 * "up").
4604 */
4605 static void
disable_nonsubgraph_leaves(graph_vertex_t * v,void * arg)4606 disable_nonsubgraph_leaves(graph_vertex_t *v, void *arg)
4607 {
4608 assert(MUTEX_HELD(&dgraph_lock));
4609
4610 /*
4611 * We must skip exclusion dependencies because they are allowed to
4612 * complete dependency cycles. This is correct because A's exclusion
4613 * dependency on B doesn't bear on the order in which they should be
4614 * stopped. Indeed, the exclusion dependency should guarantee that
4615 * they are never online at the same time.
4616 */
4617 if (v->gv_type == GVT_GROUP && v->gv_depgroup == DEPGRP_EXCLUDE_ALL)
4618 return;
4619
4620 /* If v isn't an instance, recurse on its dependencies. */
4621 if (v->gv_type != GVT_INST)
4622 goto recurse;
4623
4624 if ((v->gv_flags & GV_CONFIGURED) == 0)
4625 /*
4626 * Unconfigured instances should have no dependencies, but in
4627 * case they ever get them,
4628 */
4629 goto recurse;
4630
4631 /*
4632 * If v is in the subgraph, so should all of its dependencies, so do
4633 * nothing.
4634 */
4635 if (v->gv_flags & GV_INSUBGRAPH)
4636 return;
4637
4638 /* If v isn't a leaf because it's already down, recurse. */
4639 if (!up_state(v->gv_state))
4640 goto recurse;
4641
4642 /* If v is disabled but not down yet, be patient. */
4643 if ((v->gv_flags & GV_ENABLED) == 0)
4644 return;
4645
4646 /* If v is a leaf, disable it. */
4647 if (is_nonsubgraph_leaf(v))
4648 (void) disable_service_temporarily(v, (scf_handle_t *)arg);
4649
4650 return;
4651
4652 recurse:
4653 graph_walk_dependencies(v, disable_nonsubgraph_leaves, arg);
4654 }
4655
4656 static int
stn_restarter_state(restarter_instance_state_t rstate)4657 stn_restarter_state(restarter_instance_state_t rstate)
4658 {
4659 static const struct statemap {
4660 restarter_instance_state_t restarter_state;
4661 int scf_state;
4662 } map[] = {
4663 { RESTARTER_STATE_UNINIT, SCF_STATE_UNINIT },
4664 { RESTARTER_STATE_MAINT, SCF_STATE_MAINT },
4665 { RESTARTER_STATE_OFFLINE, SCF_STATE_OFFLINE },
4666 { RESTARTER_STATE_DISABLED, SCF_STATE_DISABLED },
4667 { RESTARTER_STATE_ONLINE, SCF_STATE_ONLINE },
4668 { RESTARTER_STATE_DEGRADED, SCF_STATE_DEGRADED }
4669 };
4670
4671 int i;
4672
4673 for (i = 0; i < sizeof (map) / sizeof (map[0]); i++) {
4674 if (rstate == map[i].restarter_state)
4675 return (map[i].scf_state);
4676 }
4677
4678 return (-1);
4679 }
4680
4681 /*
4682 * State transition counters
4683 * Not incremented atomically - indicative only
4684 */
4685 static uint64_t stev_ct_maint;
4686 static uint64_t stev_ct_hwerr;
4687 static uint64_t stev_ct_service;
4688 static uint64_t stev_ct_global;
4689 static uint64_t stev_ct_noprefs;
4690 static uint64_t stev_ct_from_uninit;
4691 static uint64_t stev_ct_bad_state;
4692 static uint64_t stev_ct_ovr_prefs;
4693
4694 static void
dgraph_state_transition_notify(graph_vertex_t * v,restarter_instance_state_t old_state,restarter_str_t reason)4695 dgraph_state_transition_notify(graph_vertex_t *v,
4696 restarter_instance_state_t old_state, restarter_str_t reason)
4697 {
4698 restarter_instance_state_t new_state = v->gv_state;
4699 int stn_transition, maint;
4700 int from, to;
4701 nvlist_t *attr;
4702 fmev_pri_t pri = FMEV_LOPRI;
4703 int raise = 0;
4704
4705 if ((from = stn_restarter_state(old_state)) == -1 ||
4706 (to = stn_restarter_state(new_state)) == -1) {
4707 stev_ct_bad_state++;
4708 return;
4709 }
4710
4711 stn_transition = from << 16 | to;
4712
4713 maint = (to == SCF_STATE_MAINT || from == SCF_STATE_MAINT);
4714
4715 if (maint) {
4716 /*
4717 * All transitions to/from maintenance state must raise
4718 * an event.
4719 */
4720 raise++;
4721 pri = FMEV_HIPRI;
4722 stev_ct_maint++;
4723 } else if (reason == restarter_str_ct_ev_hwerr) {
4724 /*
4725 * All transitions caused by hardware fault must raise
4726 * an event
4727 */
4728 raise++;
4729 pri = FMEV_HIPRI;
4730 stev_ct_hwerr++;
4731 } else if (stn_transition & v->gv_stn_tset) {
4732 /*
4733 * Specifically enabled event.
4734 */
4735 raise++;
4736 stev_ct_service++;
4737 } else if (from == SCF_STATE_UNINIT) {
4738 /*
4739 * Only raise these if specifically selected above.
4740 */
4741 stev_ct_from_uninit++;
4742 } else if (stn_transition & stn_global &&
4743 (IS_ENABLED(v) == 1 || to == SCF_STATE_DISABLED)) {
4744 raise++;
4745 stev_ct_global++;
4746 } else {
4747 stev_ct_noprefs++;
4748 }
4749
4750 if (info_events_all) {
4751 stev_ct_ovr_prefs++;
4752 raise++;
4753 }
4754 if (!raise)
4755 return;
4756
4757 if (nvlist_alloc(&attr, NV_UNIQUE_NAME, 0) != 0 ||
4758 nvlist_add_string(attr, "fmri", v->gv_name) != 0 ||
4759 nvlist_add_uint32(attr, "reason-version",
4760 restarter_str_version()) || nvlist_add_string(attr, "reason-short",
4761 restarter_get_str_short(reason)) != 0 ||
4762 nvlist_add_string(attr, "reason-long",
4763 restarter_get_str_long(reason)) != 0 ||
4764 nvlist_add_int32(attr, "transition", stn_transition) != 0) {
4765 log_framework(LOG_WARNING,
4766 "FMEV: %s could not create nvlist for transition "
4767 "event: %s\n", v->gv_name, strerror(errno));
4768 nvlist_free(attr);
4769 return;
4770 }
4771
4772 if (fmev_rspublish_nvl(FMEV_RULESET_SMF, "state-transition",
4773 instance_state_str[new_state], pri, attr) != FMEV_SUCCESS) {
4774 log_framework(LOG_DEBUG,
4775 "FMEV: %s failed to publish transition event: %s\n",
4776 v->gv_name, fmev_strerror(fmev_errno));
4777 nvlist_free(attr);
4778 }
4779 }
4780
4781 /*
4782 * Find the vertex for inst_name. If it doesn't exist, return ENOENT.
4783 * Otherwise set its state to state. If the instance has entered a state
4784 * which requires automatic action, take it (Uninitialized: do
4785 * dgraph_refresh_instance() without the snapshot update. Disabled: if the
4786 * instance should be enabled, send _ENABLE. Offline: if the instance should
4787 * be disabled, send _DISABLE, and if its dependencies are satisfied, send
4788 * _START. Online, Degraded: if the instance wasn't running, update its start
4789 * snapshot. Maintenance: no action.)
4790 *
4791 * Also fails with ECONNABORTED, or EINVAL if state is invalid.
4792 */
4793 static int
dgraph_set_instance_state(scf_handle_t * h,const char * inst_name,protocol_states_t * states)4794 dgraph_set_instance_state(scf_handle_t *h, const char *inst_name,
4795 protocol_states_t *states)
4796 {
4797 graph_vertex_t *v;
4798 int err = 0;
4799 restarter_instance_state_t old_state;
4800 restarter_instance_state_t state = states->ps_state;
4801 restarter_error_t serr = states->ps_err;
4802
4803 MUTEX_LOCK(&dgraph_lock);
4804
4805 v = vertex_get_by_name(inst_name);
4806 if (v == NULL) {
4807 MUTEX_UNLOCK(&dgraph_lock);
4808 return (ENOENT);
4809 }
4810
4811 assert(v->gv_type == GVT_INST);
4812
4813 switch (state) {
4814 case RESTARTER_STATE_UNINIT:
4815 case RESTARTER_STATE_DISABLED:
4816 case RESTARTER_STATE_OFFLINE:
4817 case RESTARTER_STATE_ONLINE:
4818 case RESTARTER_STATE_DEGRADED:
4819 case RESTARTER_STATE_MAINT:
4820 break;
4821
4822 default:
4823 MUTEX_UNLOCK(&dgraph_lock);
4824 return (EINVAL);
4825 }
4826
4827 log_framework(LOG_DEBUG, "Graph noting %s %s -> %s.\n", v->gv_name,
4828 instance_state_str[v->gv_state], instance_state_str[state]);
4829
4830 old_state = v->gv_state;
4831 v->gv_state = state;
4832
4833 v->gv_reason = states->ps_reason;
4834 err = gt_transition(h, v, serr, old_state);
4835 if (err == 0 && v->gv_state != old_state) {
4836 dgraph_state_transition_notify(v, old_state, states->ps_reason);
4837 }
4838
4839 MUTEX_UNLOCK(&dgraph_lock);
4840 return (err);
4841 }
4842
4843 /*
4844 * Handle state changes during milestone shutdown. See
4845 * dgraph_set_milestone(). If the repository connection is broken,
4846 * ECONNABORTED will be returned, though a _DISABLE command will be sent for
4847 * the vertex anyway.
4848 */
4849 int
vertex_subgraph_dependencies_shutdown(scf_handle_t * h,graph_vertex_t * v,restarter_instance_state_t old_state)4850 vertex_subgraph_dependencies_shutdown(scf_handle_t *h, graph_vertex_t *v,
4851 restarter_instance_state_t old_state)
4852 {
4853 int was_up, now_up;
4854 int ret = 0;
4855
4856 assert(v->gv_type == GVT_INST);
4857
4858 /* Don't care if we're not going to a milestone. */
4859 if (milestone == NULL)
4860 return (0);
4861
4862 /* Don't care if we already finished coming down. */
4863 if (non_subgraph_svcs == 0)
4864 return (0);
4865
4866 /* Don't care if the service is in the subgraph. */
4867 if (v->gv_flags & GV_INSUBGRAPH)
4868 return (0);
4869
4870 /*
4871 * Update non_subgraph_svcs. It is the number of non-subgraph
4872 * services which are in online, degraded, or offline.
4873 */
4874
4875 was_up = up_state(old_state);
4876 now_up = up_state(v->gv_state);
4877
4878 if (!was_up && now_up) {
4879 ++non_subgraph_svcs;
4880 } else if (was_up && !now_up) {
4881 --non_subgraph_svcs;
4882
4883 if (non_subgraph_svcs == 0) {
4884 if (halting != -1) {
4885 do_uadmin();
4886 } else if (go_single_user_mode || go_to_level1) {
4887 (void) startd_thread_create(single_user_thread,
4888 NULL);
4889 }
4890 return (0);
4891 }
4892 }
4893
4894 /* If this service is a leaf, it should be disabled. */
4895 if ((v->gv_flags & GV_ENABLED) && is_nonsubgraph_leaf(v)) {
4896 int r;
4897
4898 r = disable_service_temporarily(v, h);
4899 switch (r) {
4900 case 0:
4901 break;
4902
4903 case ECONNABORTED:
4904 ret = ECONNABORTED;
4905 break;
4906
4907 default:
4908 bad_error("disable_service_temporarily", r);
4909 }
4910 }
4911
4912 /*
4913 * If the service just came down, propagate the disable to the newly
4914 * exposed leaves.
4915 */
4916 if (was_up && !now_up)
4917 graph_walk_dependencies(v, disable_nonsubgraph_leaves,
4918 (void *)h);
4919
4920 return (ret);
4921 }
4922
4923 /*
4924 * Decide whether to start up an sulogin thread after a service is
4925 * finished changing state. Only need to do the full can_come_up()
4926 * evaluation if an instance is changing state, we're not halfway through
4927 * loading the thread, and we aren't shutting down or going to the single
4928 * user milestone.
4929 */
4930 void
graph_transition_sulogin(restarter_instance_state_t state,restarter_instance_state_t old_state)4931 graph_transition_sulogin(restarter_instance_state_t state,
4932 restarter_instance_state_t old_state)
4933 {
4934 assert(MUTEX_HELD(&dgraph_lock));
4935
4936 if (state != old_state && st->st_load_complete &&
4937 !go_single_user_mode && !go_to_level1 &&
4938 halting == -1) {
4939 if (!sulogin_thread_running && !can_come_up()) {
4940 (void) startd_thread_create(sulogin_thread, NULL);
4941 sulogin_thread_running = B_TRUE;
4942 }
4943 }
4944 }
4945
4946 /*
4947 * Propagate a start, stop event, or a satisfiability event.
4948 *
4949 * PROPAGATE_START and PROPAGATE_STOP simply propagate the transition event
4950 * to direct dependents. PROPAGATE_SAT propagates a start then walks the
4951 * full dependent graph to check for newly satisfied nodes. This is
4952 * necessary for cases when non-direct dependents may be effected but direct
4953 * dependents may not (e.g. for optional_all evaluations, see the
4954 * propagate_satbility() comments).
4955 *
4956 * PROPAGATE_SAT should be used whenever a non-running service moves into
4957 * a state which can satisfy optional dependencies, like disabled or
4958 * maintenance.
4959 */
4960 void
graph_transition_propagate(graph_vertex_t * v,propagate_event_t type,restarter_error_t rerr)4961 graph_transition_propagate(graph_vertex_t *v, propagate_event_t type,
4962 restarter_error_t rerr)
4963 {
4964 if (type == PROPAGATE_STOP) {
4965 graph_walk_dependents(v, propagate_stop, (void *)rerr);
4966 } else if (type == PROPAGATE_START || type == PROPAGATE_SAT) {
4967 graph_walk_dependents(v, propagate_start, NULL);
4968
4969 if (type == PROPAGATE_SAT)
4970 propagate_satbility(v);
4971 } else {
4972 #ifndef NDEBUG
4973 uu_warn("%s:%d: Unexpected type value %d.\n", __FILE__,
4974 __LINE__, type);
4975 #endif
4976 abort();
4977 }
4978 }
4979
4980 /*
4981 * If a vertex for fmri exists and it is enabled, send _DISABLE to the
4982 * restarter. If it is running, send _STOP. Send _REMOVE_INSTANCE. Delete
4983 * all property group dependencies, and the dependency on the restarter,
4984 * disposing of vertices as appropriate. If other vertices depend on this
4985 * one, mark it unconfigured and return. Otherwise remove the vertex. Always
4986 * returns 0.
4987 */
4988 static int
dgraph_remove_instance(const char * fmri,scf_handle_t * h)4989 dgraph_remove_instance(const char *fmri, scf_handle_t *h)
4990 {
4991 graph_vertex_t *v;
4992 graph_edge_t *e;
4993 uu_list_t *old_deps;
4994 int err;
4995
4996 log_framework(LOG_DEBUG, "Graph engine: Removing %s.\n", fmri);
4997
4998 MUTEX_LOCK(&dgraph_lock);
4999
5000 v = vertex_get_by_name(fmri);
5001 if (v == NULL) {
5002 MUTEX_UNLOCK(&dgraph_lock);
5003 return (0);
5004 }
5005
5006 /* Send restarter delete event. */
5007 if (v->gv_flags & GV_CONFIGURED)
5008 graph_unset_restarter(v);
5009
5010 if (milestone > MILESTONE_NONE) {
5011 /*
5012 * Make a list of v's current dependencies so we can
5013 * reevaluate their GV_INSUBGRAPH flags after the dependencies
5014 * are removed.
5015 */
5016 old_deps = startd_list_create(graph_edge_pool, NULL, 0);
5017
5018 err = uu_list_walk(v->gv_dependencies,
5019 (uu_walk_fn_t *)append_svcs_or_insts, old_deps, 0);
5020 assert(err == 0);
5021 }
5022
5023 delete_instance_dependencies(v, B_TRUE);
5024
5025 /*
5026 * Deleting an instance can both satisfy and unsatisfy dependencies,
5027 * depending on their type. First propagate the stop as a RERR_RESTART
5028 * event -- deletion isn't a fault, just a normal stop. This gives
5029 * dependent services the chance to do a clean shutdown. Then, mark
5030 * the service as unconfigured and propagate the start event for the
5031 * optional_all dependencies that might have become satisfied.
5032 */
5033 graph_walk_dependents(v, propagate_stop, (void *)RERR_RESTART);
5034
5035 v->gv_flags &= ~GV_CONFIGURED;
5036 v->gv_flags &= ~GV_DEATHROW;
5037
5038 graph_walk_dependents(v, propagate_start, NULL);
5039 propagate_satbility(v);
5040
5041 /*
5042 * If there are no (non-service) dependents, the vertex can be
5043 * completely removed.
5044 */
5045 if (v != milestone && v->gv_refs == 0 &&
5046 uu_list_numnodes(v->gv_dependents) == 1)
5047 remove_inst_vertex(v);
5048
5049 if (milestone > MILESTONE_NONE) {
5050 void *cookie = NULL;
5051
5052 while ((e = uu_list_teardown(old_deps, &cookie)) != NULL) {
5053 v = e->ge_vertex;
5054
5055 if (vertex_unref(v) == VERTEX_INUSE)
5056 while (eval_subgraph(v, h) == ECONNABORTED)
5057 libscf_handle_rebind(h);
5058
5059 startd_free(e, sizeof (*e));
5060 }
5061
5062 uu_list_destroy(old_deps);
5063 }
5064
5065 MUTEX_UNLOCK(&dgraph_lock);
5066
5067 return (0);
5068 }
5069
5070 /*
5071 * Return the eventual (maybe current) milestone in the form of a
5072 * legacy runlevel.
5073 */
5074 static char
target_milestone_as_runlevel()5075 target_milestone_as_runlevel()
5076 {
5077 assert(MUTEX_HELD(&dgraph_lock));
5078
5079 if (milestone == NULL)
5080 return ('3');
5081 else if (milestone == MILESTONE_NONE)
5082 return ('0');
5083
5084 if (strcmp(milestone->gv_name, multi_user_fmri) == 0)
5085 return ('2');
5086 else if (strcmp(milestone->gv_name, single_user_fmri) == 0)
5087 return ('S');
5088 else if (strcmp(milestone->gv_name, multi_user_svr_fmri) == 0)
5089 return ('3');
5090
5091 #ifndef NDEBUG
5092 (void) fprintf(stderr, "%s:%d: Unknown milestone name \"%s\".\n",
5093 __FILE__, __LINE__, milestone->gv_name);
5094 #endif
5095 abort();
5096 /* NOTREACHED */
5097 }
5098
5099 static struct {
5100 char rl;
5101 int sig;
5102 } init_sigs[] = {
5103 { 'S', SIGBUS },
5104 { '0', SIGINT },
5105 { '1', SIGQUIT },
5106 { '2', SIGILL },
5107 { '3', SIGTRAP },
5108 { '4', SIGIOT },
5109 { '5', SIGEMT },
5110 { '6', SIGFPE },
5111 { 0, 0 }
5112 };
5113
5114 static void
signal_init(char rl)5115 signal_init(char rl)
5116 {
5117 pid_t init_pid;
5118 int i;
5119
5120 assert(MUTEX_HELD(&dgraph_lock));
5121
5122 if (zone_getattr(getzoneid(), ZONE_ATTR_INITPID, &init_pid,
5123 sizeof (init_pid)) != sizeof (init_pid)) {
5124 log_error(LOG_NOTICE, "Could not get pid to signal init.\n");
5125 return;
5126 }
5127
5128 for (i = 0; init_sigs[i].rl != 0; ++i)
5129 if (init_sigs[i].rl == rl)
5130 break;
5131
5132 if (init_sigs[i].rl != 0) {
5133 if (kill(init_pid, init_sigs[i].sig) != 0) {
5134 switch (errno) {
5135 case EPERM:
5136 case ESRCH:
5137 log_error(LOG_NOTICE, "Could not signal init: "
5138 "%s.\n", strerror(errno));
5139 break;
5140
5141 case EINVAL:
5142 default:
5143 bad_error("kill", errno);
5144 }
5145 }
5146 }
5147 }
5148
5149 /*
5150 * This is called when one of the major milestones changes state, or when
5151 * init is signalled and tells us it was told to change runlevel. We wait
5152 * to reach the milestone because this allows /etc/inittab entries to retain
5153 * some boot ordering: historically, entries could place themselves before/after
5154 * the running of /sbin/rcX scripts but we can no longer make the
5155 * distinction because the /sbin/rcX scripts no longer exist as punctuation
5156 * marks in /etc/inittab.
5157 *
5158 * Also, we only trigger an update when we reach the eventual target
5159 * milestone: without this, an /etc/inittab entry marked only for
5160 * runlevel 2 would be executed for runlevel 3, which is not how
5161 * /etc/inittab entries work.
5162 *
5163 * If we're single user coming online, then we set utmpx to the target
5164 * runlevel so that legacy scripts can work as expected.
5165 */
5166 static void
graph_runlevel_changed(char rl,int online)5167 graph_runlevel_changed(char rl, int online)
5168 {
5169 char trl;
5170
5171 assert(MUTEX_HELD(&dgraph_lock));
5172
5173 trl = target_milestone_as_runlevel();
5174
5175 if (online) {
5176 if (rl == trl) {
5177 current_runlevel = trl;
5178 signal_init(trl);
5179 } else if (rl == 'S') {
5180 /*
5181 * At boot, set the entry early for the benefit of the
5182 * legacy init scripts.
5183 */
5184 utmpx_set_runlevel(trl, 'S', B_FALSE);
5185 }
5186 } else {
5187 if (rl == '3' && trl == '2') {
5188 current_runlevel = trl;
5189 signal_init(trl);
5190 } else if (rl == '2' && trl == 'S') {
5191 current_runlevel = trl;
5192 signal_init(trl);
5193 }
5194 }
5195 }
5196
5197 /*
5198 * Move to a backwards-compatible runlevel by executing the appropriate
5199 * /etc/rc?.d/K* scripts and/or setting the milestone.
5200 *
5201 * Returns
5202 * 0 - success
5203 * ECONNRESET - success, but handle was reset
5204 * ECONNABORTED - repository connection broken
5205 * ECANCELED - pg was deleted
5206 */
5207 static int
dgraph_set_runlevel(scf_propertygroup_t * pg,scf_property_t * prop)5208 dgraph_set_runlevel(scf_propertygroup_t *pg, scf_property_t *prop)
5209 {
5210 char rl;
5211 scf_handle_t *h;
5212 int r;
5213 const char *ms = NULL; /* what to commit as options/milestone */
5214 boolean_t rebound = B_FALSE;
5215 int mark_rl = 0;
5216
5217 const char * const stop = "stop";
5218
5219 r = libscf_extract_runlevel(prop, &rl);
5220 switch (r) {
5221 case 0:
5222 break;
5223
5224 case ECONNABORTED:
5225 case ECANCELED:
5226 return (r);
5227
5228 case EINVAL:
5229 case ENOENT:
5230 log_error(LOG_WARNING, "runlevel property is misconfigured; "
5231 "ignoring.\n");
5232 /* delete the bad property */
5233 goto nolock_out;
5234
5235 default:
5236 bad_error("libscf_extract_runlevel", r);
5237 }
5238
5239 switch (rl) {
5240 case 's':
5241 rl = 'S';
5242 /* FALLTHROUGH */
5243
5244 case 'S':
5245 case '2':
5246 case '3':
5247 /*
5248 * These cases cause a milestone change, so
5249 * graph_runlevel_changed() will eventually deal with
5250 * signalling init.
5251 */
5252 break;
5253
5254 case '0':
5255 case '1':
5256 case '4':
5257 case '5':
5258 case '6':
5259 mark_rl = 1;
5260 break;
5261
5262 default:
5263 log_framework(LOG_NOTICE, "Unknown runlevel '%c'.\n", rl);
5264 ms = NULL;
5265 goto nolock_out;
5266 }
5267
5268 h = scf_pg_handle(pg);
5269
5270 MUTEX_LOCK(&dgraph_lock);
5271
5272 /*
5273 * Since this triggers no milestone changes, force it by hand.
5274 */
5275 if (current_runlevel == '4' && rl == '3')
5276 mark_rl = 1;
5277
5278 /*
5279 * 1. If we are here after an "init X":
5280 *
5281 * init X
5282 * init/lscf_set_runlevel()
5283 * process_pg_event()
5284 * dgraph_set_runlevel()
5285 *
5286 * then we haven't passed through graph_runlevel_changed() yet,
5287 * therefore 'current_runlevel' has not changed for sure but 'rl' has.
5288 * In consequence, if 'rl' is lower than 'current_runlevel', we change
5289 * the system runlevel and execute the appropriate /etc/rc?.d/K* scripts
5290 * past this test.
5291 *
5292 * 2. On the other hand, if we are here after a "svcadm milestone":
5293 *
5294 * svcadm milestone X
5295 * dgraph_set_milestone()
5296 * handle_graph_update_event()
5297 * dgraph_set_instance_state()
5298 * graph_post_X_[online|offline]()
5299 * graph_runlevel_changed()
5300 * signal_init()
5301 * init/lscf_set_runlevel()
5302 * process_pg_event()
5303 * dgraph_set_runlevel()
5304 *
5305 * then we already passed through graph_runlevel_changed() (by the way
5306 * of dgraph_set_milestone()) and 'current_runlevel' may have changed
5307 * and already be equal to 'rl' so we are going to return immediately
5308 * from dgraph_set_runlevel() without changing the system runlevel and
5309 * without executing the /etc/rc?.d/K* scripts.
5310 */
5311 if (rl == current_runlevel) {
5312 ms = NULL;
5313 goto out;
5314 }
5315
5316 log_framework(LOG_DEBUG, "Changing to runlevel '%c'.\n", rl);
5317
5318 /*
5319 * Make sure stop rc scripts see the new settings via who -r.
5320 */
5321 utmpx_set_runlevel(rl, current_runlevel, B_TRUE);
5322
5323 /*
5324 * Some run levels don't have a direct correspondence to any
5325 * milestones, so we have to signal init directly.
5326 */
5327 if (mark_rl) {
5328 current_runlevel = rl;
5329 signal_init(rl);
5330 }
5331
5332 switch (rl) {
5333 case 'S':
5334 uu_warn("The system is coming down for administration. "
5335 "Please wait.\n");
5336 fork_rc_script(rl, stop, B_FALSE);
5337 ms = single_user_fmri;
5338 go_single_user_mode = B_TRUE;
5339 break;
5340
5341 case '0':
5342 halting_time = time(NULL);
5343 fork_rc_script(rl, stop, B_TRUE);
5344 halting = AD_HALT;
5345 goto uadmin;
5346
5347 case '5':
5348 halting_time = time(NULL);
5349 fork_rc_script(rl, stop, B_TRUE);
5350 halting = AD_POWEROFF;
5351 goto uadmin;
5352
5353 case '6':
5354 halting_time = time(NULL);
5355 fork_rc_script(rl, stop, B_TRUE);
5356 if (scf_is_fastboot_default() && getzoneid() == GLOBAL_ZONEID)
5357 halting = AD_FASTREBOOT;
5358 else
5359 halting = AD_BOOT;
5360
5361 uadmin:
5362 uu_warn("The system is coming down. Please wait.\n");
5363 ms = "none";
5364
5365 /*
5366 * We can't wait until all services are offline since this
5367 * thread is responsible for taking them offline. Instead we
5368 * set halting to the second argument for uadmin() and call
5369 * do_uadmin() from dgraph_set_instance_state() when
5370 * appropriate.
5371 */
5372 break;
5373
5374 case '1':
5375 if (current_runlevel != 'S') {
5376 uu_warn("Changing to state 1.\n");
5377 fork_rc_script(rl, stop, B_FALSE);
5378 } else {
5379 uu_warn("The system is coming up for administration. "
5380 "Please wait.\n");
5381 }
5382 ms = single_user_fmri;
5383 go_to_level1 = B_TRUE;
5384 break;
5385
5386 case '2':
5387 if (current_runlevel == '3' || current_runlevel == '4')
5388 fork_rc_script(rl, stop, B_FALSE);
5389 ms = multi_user_fmri;
5390 break;
5391
5392 case '3':
5393 case '4':
5394 ms = "all";
5395 break;
5396
5397 default:
5398 #ifndef NDEBUG
5399 (void) fprintf(stderr, "%s:%d: Uncaught case %d ('%c').\n",
5400 __FILE__, __LINE__, rl, rl);
5401 #endif
5402 abort();
5403 }
5404
5405 out:
5406 MUTEX_UNLOCK(&dgraph_lock);
5407
5408 nolock_out:
5409 switch (r = libscf_clear_runlevel(pg, ms)) {
5410 case 0:
5411 break;
5412
5413 case ECONNABORTED:
5414 libscf_handle_rebind(h);
5415 rebound = B_TRUE;
5416 goto nolock_out;
5417
5418 case ECANCELED:
5419 break;
5420
5421 case EPERM:
5422 case EACCES:
5423 case EROFS:
5424 log_error(LOG_NOTICE, "Could not delete \"%s/%s\" property: "
5425 "%s.\n", SCF_PG_OPTIONS, "runlevel", strerror(r));
5426 break;
5427
5428 default:
5429 bad_error("libscf_clear_runlevel", r);
5430 }
5431
5432 return (rebound ? ECONNRESET : 0);
5433 }
5434
5435 /*
5436 * mark_subtree walks the dependents and add the GV_TOOFFLINE flag
5437 * to the instances that are supposed to go offline during an
5438 * administrative disable operation.
5439 */
5440 static int
mark_subtree(graph_edge_t * e,void * arg)5441 mark_subtree(graph_edge_t *e, void *arg)
5442 {
5443 graph_vertex_t *v;
5444 int r;
5445
5446 v = e->ge_vertex;
5447
5448 /* If it's already in the subgraph, skip. */
5449 if (v->gv_flags & GV_TOOFFLINE)
5450 return (UU_WALK_NEXT);
5451
5452 switch (v->gv_type) {
5453 case GVT_INST:
5454 /* If the instance is already disabled, skip it. */
5455 if (!(v->gv_flags & GV_ENABLED))
5456 return (UU_WALK_NEXT);
5457
5458 v->gv_flags |= GV_TOOFFLINE;
5459 log_framework(LOG_DEBUG, "%s added to subtree\n", v->gv_name);
5460 break;
5461 case GVT_GROUP:
5462 /*
5463 * Skip all excluded and optional_all dependencies and decide
5464 * whether to offline the service based on restart_on attribute.
5465 */
5466 if (is_depgrp_bypassed(v))
5467 return (UU_WALK_NEXT);
5468 break;
5469 }
5470
5471 r = uu_list_walk(v->gv_dependents, (uu_walk_fn_t *)mark_subtree, arg,
5472 0);
5473 assert(r == 0);
5474 return (UU_WALK_NEXT);
5475 }
5476
5477 static int
mark_subgraph(graph_edge_t * e,void * arg)5478 mark_subgraph(graph_edge_t *e, void *arg)
5479 {
5480 graph_vertex_t *v;
5481 int r;
5482 int optional = (int)arg;
5483
5484 v = e->ge_vertex;
5485
5486 /* If it's already in the subgraph, skip. */
5487 if (v->gv_flags & GV_INSUBGRAPH)
5488 return (UU_WALK_NEXT);
5489
5490 /*
5491 * Keep track if walk has entered an optional dependency group
5492 */
5493 if (v->gv_type == GVT_GROUP && v->gv_depgroup == DEPGRP_OPTIONAL_ALL) {
5494 optional = 1;
5495 }
5496 /*
5497 * Quit if we are in an optional dependency group and the instance
5498 * is disabled
5499 */
5500 if (optional && (v->gv_type == GVT_INST) &&
5501 (!(v->gv_flags & GV_ENBLD_NOOVR)))
5502 return (UU_WALK_NEXT);
5503
5504 v->gv_flags |= GV_INSUBGRAPH;
5505
5506 /* Skip all excluded dependencies. */
5507 if (v->gv_type == GVT_GROUP && v->gv_depgroup == DEPGRP_EXCLUDE_ALL)
5508 return (UU_WALK_NEXT);
5509
5510 r = uu_list_walk(v->gv_dependencies, (uu_walk_fn_t *)mark_subgraph,
5511 (void *)optional, 0);
5512 assert(r == 0);
5513 return (UU_WALK_NEXT);
5514 }
5515
5516 /*
5517 * Bring down all services which are not dependencies of fmri. The
5518 * dependencies of fmri (direct & indirect) will constitute the "subgraph",
5519 * and will have the GV_INSUBGRAPH flag set. The rest must be brought down,
5520 * which means the state is "disabled", "maintenance", or "uninitialized". We
5521 * could consider "offline" to be down, and refrain from sending start
5522 * commands for such services, but that's not strictly necessary, so we'll
5523 * decline to intrude on the state machine. It would probably confuse users
5524 * anyway.
5525 *
5526 * The services should be brought down in reverse-dependency order, so we
5527 * can't do it all at once here. We initiate by override-disabling the leaves
5528 * of the dependency tree -- those services which are up but have no
5529 * dependents which are up. When they come down,
5530 * vertex_subgraph_dependencies_shutdown() will override-disable the newly
5531 * exposed leaves. Perseverance will ensure completion.
5532 *
5533 * Sometimes we need to take action when the transition is complete, like
5534 * start sulogin or halt the system. To tell when we're done, we initialize
5535 * non_subgraph_svcs here to be the number of services which need to come
5536 * down. As each does, we decrement the counter. When it hits zero, we take
5537 * the appropriate action. See vertex_subgraph_dependencies_shutdown().
5538 *
5539 * In case we're coming up, we also remove any enable-overrides for the
5540 * services which are dependencies of fmri.
5541 *
5542 * If norepository is true, the function will not change the repository.
5543 *
5544 * The decision to change the system run level in accordance with the milestone
5545 * is taken in dgraph_set_runlevel().
5546 *
5547 * Returns
5548 * 0 - success
5549 * ECONNRESET - success, but handle was rebound
5550 * EINVAL - fmri is invalid (error is logged)
5551 * EALREADY - the milestone is already set to fmri
5552 * ENOENT - a configured vertex does not exist for fmri (an error is logged)
5553 */
5554 static int
dgraph_set_milestone(const char * fmri,scf_handle_t * h,boolean_t norepository)5555 dgraph_set_milestone(const char *fmri, scf_handle_t *h, boolean_t norepository)
5556 {
5557 const char *cfmri, *fs;
5558 graph_vertex_t *nm, *v;
5559 int ret = 0, r;
5560 scf_instance_t *inst;
5561 boolean_t isall, isnone, rebound = B_FALSE;
5562
5563 /* Validate fmri */
5564 isall = (strcmp(fmri, "all") == 0);
5565 isnone = (strcmp(fmri, "none") == 0);
5566
5567 if (!isall && !isnone) {
5568 if (fmri_canonify(fmri, (char **)&cfmri, B_FALSE) == EINVAL)
5569 goto reject;
5570
5571 if (strcmp(cfmri, single_user_fmri) != 0 &&
5572 strcmp(cfmri, multi_user_fmri) != 0 &&
5573 strcmp(cfmri, multi_user_svr_fmri) != 0) {
5574 startd_free((void *)cfmri, max_scf_fmri_size);
5575 reject:
5576 log_framework(LOG_WARNING,
5577 "Rejecting request for invalid milestone \"%s\".\n",
5578 fmri);
5579 return (EINVAL);
5580 }
5581 }
5582
5583 inst = safe_scf_instance_create(h);
5584
5585 MUTEX_LOCK(&dgraph_lock);
5586
5587 if (milestone == NULL) {
5588 if (isall) {
5589 log_framework(LOG_DEBUG,
5590 "Milestone already set to all.\n");
5591 ret = EALREADY;
5592 goto out;
5593 }
5594 } else if (milestone == MILESTONE_NONE) {
5595 if (isnone) {
5596 log_framework(LOG_DEBUG,
5597 "Milestone already set to none.\n");
5598 ret = EALREADY;
5599 goto out;
5600 }
5601 } else {
5602 if (!isall && !isnone &&
5603 strcmp(cfmri, milestone->gv_name) == 0) {
5604 log_framework(LOG_DEBUG,
5605 "Milestone already set to %s.\n", cfmri);
5606 ret = EALREADY;
5607 goto out;
5608 }
5609 }
5610
5611 if (!isall && !isnone) {
5612 nm = vertex_get_by_name(cfmri);
5613 if (nm == NULL || !(nm->gv_flags & GV_CONFIGURED)) {
5614 log_framework(LOG_WARNING, "Cannot set milestone to %s "
5615 "because no such service exists.\n", cfmri);
5616 ret = ENOENT;
5617 goto out;
5618 }
5619 }
5620
5621 log_framework(LOG_DEBUG, "Changing milestone to %s.\n", fmri);
5622
5623 /*
5624 * Set milestone, removing the old one if this was the last reference.
5625 */
5626 if (milestone > MILESTONE_NONE)
5627 (void) vertex_unref(milestone);
5628
5629 if (isall)
5630 milestone = NULL;
5631 else if (isnone)
5632 milestone = MILESTONE_NONE;
5633 else {
5634 milestone = nm;
5635 /* milestone should count as a reference */
5636 vertex_ref(milestone);
5637 }
5638
5639 /* Clear all GV_INSUBGRAPH bits. */
5640 for (v = uu_list_first(dgraph); v != NULL; v = uu_list_next(dgraph, v))
5641 v->gv_flags &= ~GV_INSUBGRAPH;
5642
5643 if (!isall && !isnone) {
5644 /* Set GV_INSUBGRAPH for milestone & descendents. */
5645 milestone->gv_flags |= GV_INSUBGRAPH;
5646
5647 r = uu_list_walk(milestone->gv_dependencies,
5648 (uu_walk_fn_t *)mark_subgraph, NULL, 0);
5649 assert(r == 0);
5650 }
5651
5652 /* Un-override services in the subgraph & override-disable the rest. */
5653 if (norepository)
5654 goto out;
5655
5656 non_subgraph_svcs = 0;
5657 for (v = uu_list_first(dgraph);
5658 v != NULL;
5659 v = uu_list_next(dgraph, v)) {
5660 if (v->gv_type != GVT_INST ||
5661 (v->gv_flags & GV_CONFIGURED) == 0)
5662 continue;
5663
5664 again:
5665 r = scf_handle_decode_fmri(h, v->gv_name, NULL, NULL, inst,
5666 NULL, NULL, SCF_DECODE_FMRI_EXACT);
5667 if (r != 0) {
5668 switch (scf_error()) {
5669 case SCF_ERROR_CONNECTION_BROKEN:
5670 default:
5671 libscf_handle_rebind(h);
5672 rebound = B_TRUE;
5673 goto again;
5674
5675 case SCF_ERROR_NOT_FOUND:
5676 continue;
5677
5678 case SCF_ERROR_HANDLE_MISMATCH:
5679 case SCF_ERROR_INVALID_ARGUMENT:
5680 case SCF_ERROR_CONSTRAINT_VIOLATED:
5681 case SCF_ERROR_NOT_BOUND:
5682 bad_error("scf_handle_decode_fmri",
5683 scf_error());
5684 }
5685 }
5686
5687 if (isall || (v->gv_flags & GV_INSUBGRAPH)) {
5688 r = libscf_delete_enable_ovr(inst);
5689 fs = "libscf_delete_enable_ovr";
5690 } else {
5691 assert(isnone || (v->gv_flags & GV_INSUBGRAPH) == 0);
5692
5693 /*
5694 * Services which are up need to come down before
5695 * we're done, but we can only disable the leaves
5696 * here.
5697 */
5698
5699 if (up_state(v->gv_state))
5700 ++non_subgraph_svcs;
5701
5702 /* If it's already disabled, don't bother. */
5703 if ((v->gv_flags & GV_ENABLED) == 0)
5704 continue;
5705
5706 if (!is_nonsubgraph_leaf(v))
5707 continue;
5708
5709 r = libscf_set_enable_ovr(inst, 0);
5710 fs = "libscf_set_enable_ovr";
5711 }
5712 switch (r) {
5713 case 0:
5714 case ECANCELED:
5715 break;
5716
5717 case ECONNABORTED:
5718 libscf_handle_rebind(h);
5719 rebound = B_TRUE;
5720 goto again;
5721
5722 case EPERM:
5723 case EROFS:
5724 log_error(LOG_WARNING,
5725 "Could not set %s/%s for %s: %s.\n",
5726 SCF_PG_GENERAL_OVR, SCF_PROPERTY_ENABLED,
5727 v->gv_name, strerror(r));
5728 break;
5729
5730 default:
5731 bad_error(fs, r);
5732 }
5733 }
5734
5735 if (halting != -1) {
5736 if (non_subgraph_svcs > 1)
5737 uu_warn("%d system services are now being stopped.\n",
5738 non_subgraph_svcs);
5739 else if (non_subgraph_svcs == 1)
5740 uu_warn("One system service is now being stopped.\n");
5741 else if (non_subgraph_svcs == 0)
5742 do_uadmin();
5743 }
5744
5745 ret = rebound ? ECONNRESET : 0;
5746
5747 out:
5748 MUTEX_UNLOCK(&dgraph_lock);
5749 if (!isall && !isnone)
5750 startd_free((void *)cfmri, max_scf_fmri_size);
5751 scf_instance_destroy(inst);
5752 return (ret);
5753 }
5754
5755
5756 /*
5757 * Returns 0, ECONNABORTED, or EINVAL.
5758 */
5759 static int
handle_graph_update_event(scf_handle_t * h,graph_protocol_event_t * e)5760 handle_graph_update_event(scf_handle_t *h, graph_protocol_event_t *e)
5761 {
5762 int r;
5763
5764 switch (e->gpe_type) {
5765 case GRAPH_UPDATE_RELOAD_GRAPH:
5766 log_error(LOG_WARNING,
5767 "graph_event: reload graph unimplemented\n");
5768 break;
5769
5770 case GRAPH_UPDATE_STATE_CHANGE: {
5771 protocol_states_t *states = e->gpe_data;
5772
5773 switch (r = dgraph_set_instance_state(h, e->gpe_inst, states)) {
5774 case 0:
5775 case ENOENT:
5776 break;
5777
5778 case ECONNABORTED:
5779 return (ECONNABORTED);
5780
5781 case EINVAL:
5782 default:
5783 #ifndef NDEBUG
5784 (void) fprintf(stderr, "dgraph_set_instance_state() "
5785 "failed with unexpected error %d at %s:%d.\n", r,
5786 __FILE__, __LINE__);
5787 #endif
5788 abort();
5789 }
5790
5791 startd_free(states, sizeof (protocol_states_t));
5792 break;
5793 }
5794
5795 default:
5796 log_error(LOG_WARNING,
5797 "graph_event_loop received an unknown event: %d\n",
5798 e->gpe_type);
5799 break;
5800 }
5801
5802 return (0);
5803 }
5804
5805 /*
5806 * graph_event_thread()
5807 * Wait for state changes from the restarters.
5808 */
5809 /*ARGSUSED*/
5810 void *
graph_event_thread(void * unused)5811 graph_event_thread(void *unused)
5812 {
5813 scf_handle_t *h;
5814 int err;
5815
5816 h = libscf_handle_create_bound_loop();
5817
5818 /*CONSTCOND*/
5819 while (1) {
5820 graph_protocol_event_t *e;
5821
5822 MUTEX_LOCK(&gu->gu_lock);
5823
5824 while (gu->gu_wakeup == 0)
5825 (void) pthread_cond_wait(&gu->gu_cv, &gu->gu_lock);
5826
5827 gu->gu_wakeup = 0;
5828
5829 while ((e = graph_event_dequeue()) != NULL) {
5830 MUTEX_LOCK(&e->gpe_lock);
5831 MUTEX_UNLOCK(&gu->gu_lock);
5832
5833 while ((err = handle_graph_update_event(h, e)) ==
5834 ECONNABORTED)
5835 libscf_handle_rebind(h);
5836
5837 if (err == 0)
5838 graph_event_release(e);
5839 else
5840 graph_event_requeue(e);
5841
5842 MUTEX_LOCK(&gu->gu_lock);
5843 }
5844
5845 MUTEX_UNLOCK(&gu->gu_lock);
5846 }
5847
5848 /*
5849 * Unreachable for now -- there's currently no graceful cleanup
5850 * called on exit().
5851 */
5852 MUTEX_UNLOCK(&gu->gu_lock);
5853 scf_handle_destroy(h);
5854 return (NULL);
5855 }
5856
5857 static void
set_initial_milestone(scf_handle_t * h)5858 set_initial_milestone(scf_handle_t *h)
5859 {
5860 scf_instance_t *inst;
5861 char *fmri, *cfmri;
5862 size_t sz;
5863 int r;
5864
5865 inst = safe_scf_instance_create(h);
5866 fmri = startd_alloc(max_scf_fmri_size);
5867
5868 /*
5869 * If -m milestone= was specified, we want to set options_ovr/milestone
5870 * to it. Otherwise we want to read what the milestone should be set
5871 * to. Either way we need our inst.
5872 */
5873 get_self:
5874 if (scf_handle_decode_fmri(h, SCF_SERVICE_STARTD, NULL, NULL, inst,
5875 NULL, NULL, SCF_DECODE_FMRI_EXACT) != 0) {
5876 switch (scf_error()) {
5877 case SCF_ERROR_CONNECTION_BROKEN:
5878 libscf_handle_rebind(h);
5879 goto get_self;
5880
5881 case SCF_ERROR_NOT_FOUND:
5882 if (st->st_subgraph != NULL &&
5883 st->st_subgraph[0] != '\0') {
5884 sz = strlcpy(fmri, st->st_subgraph,
5885 max_scf_fmri_size);
5886 assert(sz < max_scf_fmri_size);
5887 } else {
5888 fmri[0] = '\0';
5889 }
5890 break;
5891
5892 case SCF_ERROR_INVALID_ARGUMENT:
5893 case SCF_ERROR_CONSTRAINT_VIOLATED:
5894 case SCF_ERROR_HANDLE_MISMATCH:
5895 default:
5896 bad_error("scf_handle_decode_fmri", scf_error());
5897 }
5898 } else {
5899 if (st->st_subgraph != NULL && st->st_subgraph[0] != '\0') {
5900 scf_propertygroup_t *pg;
5901
5902 pg = safe_scf_pg_create(h);
5903
5904 sz = strlcpy(fmri, st->st_subgraph, max_scf_fmri_size);
5905 assert(sz < max_scf_fmri_size);
5906
5907 r = libscf_inst_get_or_add_pg(inst, SCF_PG_OPTIONS_OVR,
5908 SCF_PG_OPTIONS_OVR_TYPE, SCF_PG_OPTIONS_OVR_FLAGS,
5909 pg);
5910 switch (r) {
5911 case 0:
5912 break;
5913
5914 case ECONNABORTED:
5915 libscf_handle_rebind(h);
5916 goto get_self;
5917
5918 case EPERM:
5919 case EACCES:
5920 case EROFS:
5921 log_error(LOG_WARNING, "Could not set %s/%s: "
5922 "%s.\n", SCF_PG_OPTIONS_OVR,
5923 SCF_PROPERTY_MILESTONE, strerror(r));
5924 /* FALLTHROUGH */
5925
5926 case ECANCELED:
5927 sz = strlcpy(fmri, st->st_subgraph,
5928 max_scf_fmri_size);
5929 assert(sz < max_scf_fmri_size);
5930 break;
5931
5932 default:
5933 bad_error("libscf_inst_get_or_add_pg", r);
5934 }
5935
5936 r = libscf_clear_runlevel(pg, fmri);
5937 switch (r) {
5938 case 0:
5939 break;
5940
5941 case ECONNABORTED:
5942 libscf_handle_rebind(h);
5943 goto get_self;
5944
5945 case EPERM:
5946 case EACCES:
5947 case EROFS:
5948 log_error(LOG_WARNING, "Could not set %s/%s: "
5949 "%s.\n", SCF_PG_OPTIONS_OVR,
5950 SCF_PROPERTY_MILESTONE, strerror(r));
5951 /* FALLTHROUGH */
5952
5953 case ECANCELED:
5954 sz = strlcpy(fmri, st->st_subgraph,
5955 max_scf_fmri_size);
5956 assert(sz < max_scf_fmri_size);
5957 break;
5958
5959 default:
5960 bad_error("libscf_clear_runlevel", r);
5961 }
5962
5963 scf_pg_destroy(pg);
5964 } else {
5965 scf_property_t *prop;
5966 scf_value_t *val;
5967
5968 prop = safe_scf_property_create(h);
5969 val = safe_scf_value_create(h);
5970
5971 r = libscf_get_milestone(inst, prop, val, fmri,
5972 max_scf_fmri_size);
5973 switch (r) {
5974 case 0:
5975 break;
5976
5977 case ECONNABORTED:
5978 libscf_handle_rebind(h);
5979 goto get_self;
5980
5981 case EINVAL:
5982 log_error(LOG_WARNING, "Milestone property is "
5983 "misconfigured. Defaulting to \"all\".\n");
5984 /* FALLTHROUGH */
5985
5986 case ECANCELED:
5987 case ENOENT:
5988 fmri[0] = '\0';
5989 break;
5990
5991 default:
5992 bad_error("libscf_get_milestone", r);
5993 }
5994
5995 scf_value_destroy(val);
5996 scf_property_destroy(prop);
5997 }
5998 }
5999
6000 if (fmri[0] == '\0' || strcmp(fmri, "all") == 0)
6001 goto out;
6002
6003 if (strcmp(fmri, "none") != 0) {
6004 retry:
6005 if (scf_handle_decode_fmri(h, fmri, NULL, NULL, inst, NULL,
6006 NULL, SCF_DECODE_FMRI_EXACT) != 0) {
6007 switch (scf_error()) {
6008 case SCF_ERROR_INVALID_ARGUMENT:
6009 log_error(LOG_WARNING,
6010 "Requested milestone \"%s\" is invalid. "
6011 "Reverting to \"all\".\n", fmri);
6012 goto out;
6013
6014 case SCF_ERROR_CONSTRAINT_VIOLATED:
6015 log_error(LOG_WARNING, "Requested milestone "
6016 "\"%s\" does not specify an instance. "
6017 "Reverting to \"all\".\n", fmri);
6018 goto out;
6019
6020 case SCF_ERROR_CONNECTION_BROKEN:
6021 libscf_handle_rebind(h);
6022 goto retry;
6023
6024 case SCF_ERROR_NOT_FOUND:
6025 log_error(LOG_WARNING, "Requested milestone "
6026 "\"%s\" not in repository. Reverting to "
6027 "\"all\".\n", fmri);
6028 goto out;
6029
6030 case SCF_ERROR_HANDLE_MISMATCH:
6031 default:
6032 bad_error("scf_handle_decode_fmri",
6033 scf_error());
6034 }
6035 }
6036
6037 r = fmri_canonify(fmri, &cfmri, B_FALSE);
6038 assert(r == 0);
6039
6040 r = dgraph_add_instance(cfmri, inst, B_TRUE);
6041 startd_free(cfmri, max_scf_fmri_size);
6042 switch (r) {
6043 case 0:
6044 break;
6045
6046 case ECONNABORTED:
6047 goto retry;
6048
6049 case EINVAL:
6050 log_error(LOG_WARNING,
6051 "Requested milestone \"%s\" is invalid. "
6052 "Reverting to \"all\".\n", fmri);
6053 goto out;
6054
6055 case ECANCELED:
6056 log_error(LOG_WARNING,
6057 "Requested milestone \"%s\" not "
6058 "in repository. Reverting to \"all\".\n",
6059 fmri);
6060 goto out;
6061
6062 case EEXIST:
6063 default:
6064 bad_error("dgraph_add_instance", r);
6065 }
6066 }
6067
6068 log_console(LOG_INFO, "Booting to milestone \"%s\".\n", fmri);
6069
6070 r = dgraph_set_milestone(fmri, h, B_FALSE);
6071 switch (r) {
6072 case 0:
6073 case ECONNRESET:
6074 case EALREADY:
6075 break;
6076
6077 case EINVAL:
6078 case ENOENT:
6079 default:
6080 bad_error("dgraph_set_milestone", r);
6081 }
6082
6083 out:
6084 startd_free(fmri, max_scf_fmri_size);
6085 scf_instance_destroy(inst);
6086 }
6087
6088 void
set_restart_milestone(scf_handle_t * h)6089 set_restart_milestone(scf_handle_t *h)
6090 {
6091 scf_instance_t *inst;
6092 scf_property_t *prop;
6093 scf_value_t *val;
6094 char *fmri;
6095 int r;
6096
6097 inst = safe_scf_instance_create(h);
6098
6099 get_self:
6100 if (scf_handle_decode_fmri(h, SCF_SERVICE_STARTD, NULL, NULL,
6101 inst, NULL, NULL, SCF_DECODE_FMRI_EXACT) != 0) {
6102 switch (scf_error()) {
6103 case SCF_ERROR_CONNECTION_BROKEN:
6104 libscf_handle_rebind(h);
6105 goto get_self;
6106
6107 case SCF_ERROR_NOT_FOUND:
6108 break;
6109
6110 case SCF_ERROR_INVALID_ARGUMENT:
6111 case SCF_ERROR_CONSTRAINT_VIOLATED:
6112 case SCF_ERROR_HANDLE_MISMATCH:
6113 default:
6114 bad_error("scf_handle_decode_fmri", scf_error());
6115 }
6116
6117 scf_instance_destroy(inst);
6118 return;
6119 }
6120
6121 prop = safe_scf_property_create(h);
6122 val = safe_scf_value_create(h);
6123 fmri = startd_alloc(max_scf_fmri_size);
6124
6125 r = libscf_get_milestone(inst, prop, val, fmri, max_scf_fmri_size);
6126 switch (r) {
6127 case 0:
6128 break;
6129
6130 case ECONNABORTED:
6131 libscf_handle_rebind(h);
6132 goto get_self;
6133
6134 case ECANCELED:
6135 case ENOENT:
6136 case EINVAL:
6137 goto out;
6138
6139 default:
6140 bad_error("libscf_get_milestone", r);
6141 }
6142
6143 r = dgraph_set_milestone(fmri, h, B_TRUE);
6144 switch (r) {
6145 case 0:
6146 case ECONNRESET:
6147 case EALREADY:
6148 case EINVAL:
6149 case ENOENT:
6150 break;
6151
6152 default:
6153 bad_error("dgraph_set_milestone", r);
6154 }
6155
6156 out:
6157 startd_free(fmri, max_scf_fmri_size);
6158 scf_value_destroy(val);
6159 scf_property_destroy(prop);
6160 scf_instance_destroy(inst);
6161 }
6162
6163 /*
6164 * void *graph_thread(void *)
6165 *
6166 * Graph management thread.
6167 */
6168 /*ARGSUSED*/
6169 void *
graph_thread(void * arg)6170 graph_thread(void *arg)
6171 {
6172 scf_handle_t *h;
6173 int err;
6174
6175 h = libscf_handle_create_bound_loop();
6176
6177 if (st->st_initial)
6178 set_initial_milestone(h);
6179
6180 MUTEX_LOCK(&dgraph_lock);
6181 initial_milestone_set = B_TRUE;
6182 err = pthread_cond_broadcast(&initial_milestone_cv);
6183 assert(err == 0);
6184 MUTEX_UNLOCK(&dgraph_lock);
6185
6186 libscf_populate_graph(h);
6187
6188 if (!st->st_initial)
6189 set_restart_milestone(h);
6190
6191 MUTEX_LOCK(&st->st_load_lock);
6192 st->st_load_complete = 1;
6193 (void) pthread_cond_broadcast(&st->st_load_cv);
6194 MUTEX_UNLOCK(&st->st_load_lock);
6195
6196 MUTEX_LOCK(&dgraph_lock);
6197 /*
6198 * Now that we've set st_load_complete we need to check can_come_up()
6199 * since if we booted to a milestone, then there won't be any more
6200 * state updates.
6201 */
6202 if (!go_single_user_mode && !go_to_level1 &&
6203 halting == -1) {
6204 if (!sulogin_thread_running && !can_come_up()) {
6205 (void) startd_thread_create(sulogin_thread, NULL);
6206 sulogin_thread_running = B_TRUE;
6207 }
6208 }
6209 MUTEX_UNLOCK(&dgraph_lock);
6210
6211 (void) pthread_mutex_lock(&gu->gu_freeze_lock);
6212
6213 /*CONSTCOND*/
6214 while (1) {
6215 (void) pthread_cond_wait(&gu->gu_freeze_cv,
6216 &gu->gu_freeze_lock);
6217 }
6218
6219 /*
6220 * Unreachable for now -- there's currently no graceful cleanup
6221 * called on exit().
6222 */
6223 (void) pthread_mutex_unlock(&gu->gu_freeze_lock);
6224 scf_handle_destroy(h);
6225
6226 return (NULL);
6227 }
6228
6229
6230 /*
6231 * int next_action()
6232 * Given an array of timestamps 'a' with 'num' elements, find the
6233 * lowest non-zero timestamp and return its index. If there are no
6234 * non-zero elements, return -1.
6235 */
6236 static int
next_action(hrtime_t * a,int num)6237 next_action(hrtime_t *a, int num)
6238 {
6239 hrtime_t t = 0;
6240 int i = 0, smallest = -1;
6241
6242 for (i = 0; i < num; i++) {
6243 if (t == 0) {
6244 t = a[i];
6245 smallest = i;
6246 } else if (a[i] != 0 && a[i] < t) {
6247 t = a[i];
6248 smallest = i;
6249 }
6250 }
6251
6252 if (t == 0)
6253 return (-1);
6254 else
6255 return (smallest);
6256 }
6257
6258 /*
6259 * void process_actions()
6260 * Process actions requested by the administrator. Possibilities include:
6261 * refresh, restart, maintenance mode off, maintenance mode on,
6262 * maintenance mode immediate, and degraded.
6263 *
6264 * The set of pending actions is represented in the repository as a
6265 * per-instance property group, with each action being a single property
6266 * in that group. This property group is converted to an array, with each
6267 * action type having an array slot. The actions in the array at the
6268 * time process_actions() is called are acted on in the order of the
6269 * timestamp (which is the value stored in the slot). A value of zero
6270 * indicates that there is no pending action of the type associated with
6271 * a particular slot.
6272 *
6273 * Sending an action event multiple times before the restarter has a
6274 * chance to process that action will force it to be run at the last
6275 * timestamp where it appears in the ordering.
6276 *
6277 * Turning maintenance mode on trumps all other actions.
6278 *
6279 * Returns 0 or ECONNABORTED.
6280 */
6281 static int
process_actions(scf_handle_t * h,scf_propertygroup_t * pg,scf_instance_t * inst)6282 process_actions(scf_handle_t *h, scf_propertygroup_t *pg, scf_instance_t *inst)
6283 {
6284 scf_property_t *prop = NULL;
6285 scf_value_t *val = NULL;
6286 scf_type_t type;
6287 graph_vertex_t *vertex;
6288 admin_action_t a;
6289 int i, ret = 0, r;
6290 hrtime_t action_ts[NACTIONS];
6291 char *inst_name;
6292
6293 r = libscf_instance_get_fmri(inst, &inst_name);
6294 switch (r) {
6295 case 0:
6296 break;
6297
6298 case ECONNABORTED:
6299 return (ECONNABORTED);
6300
6301 case ECANCELED:
6302 return (0);
6303
6304 default:
6305 bad_error("libscf_instance_get_fmri", r);
6306 }
6307
6308 MUTEX_LOCK(&dgraph_lock);
6309
6310 vertex = vertex_get_by_name(inst_name);
6311 if (vertex == NULL) {
6312 MUTEX_UNLOCK(&dgraph_lock);
6313 log_framework(LOG_DEBUG, "%s: Can't find graph vertex. "
6314 "The instance must have been removed.\n", inst_name);
6315 startd_free(inst_name, max_scf_fmri_size);
6316 return (0);
6317 }
6318
6319 prop = safe_scf_property_create(h);
6320 val = safe_scf_value_create(h);
6321
6322 for (i = 0; i < NACTIONS; i++) {
6323 if (scf_pg_get_property(pg, admin_actions[i], prop) != 0) {
6324 switch (scf_error()) {
6325 case SCF_ERROR_CONNECTION_BROKEN:
6326 default:
6327 ret = ECONNABORTED;
6328 goto out;
6329
6330 case SCF_ERROR_DELETED:
6331 goto out;
6332
6333 case SCF_ERROR_NOT_FOUND:
6334 action_ts[i] = 0;
6335 continue;
6336
6337 case SCF_ERROR_HANDLE_MISMATCH:
6338 case SCF_ERROR_INVALID_ARGUMENT:
6339 case SCF_ERROR_NOT_SET:
6340 bad_error("scf_pg_get_property", scf_error());
6341 }
6342 }
6343
6344 if (scf_property_type(prop, &type) != 0) {
6345 switch (scf_error()) {
6346 case SCF_ERROR_CONNECTION_BROKEN:
6347 default:
6348 ret = ECONNABORTED;
6349 goto out;
6350
6351 case SCF_ERROR_DELETED:
6352 action_ts[i] = 0;
6353 continue;
6354
6355 case SCF_ERROR_NOT_SET:
6356 bad_error("scf_property_type", scf_error());
6357 }
6358 }
6359
6360 if (type != SCF_TYPE_INTEGER) {
6361 action_ts[i] = 0;
6362 continue;
6363 }
6364
6365 if (scf_property_get_value(prop, val) != 0) {
6366 switch (scf_error()) {
6367 case SCF_ERROR_CONNECTION_BROKEN:
6368 default:
6369 ret = ECONNABORTED;
6370 goto out;
6371
6372 case SCF_ERROR_DELETED:
6373 goto out;
6374
6375 case SCF_ERROR_NOT_FOUND:
6376 case SCF_ERROR_CONSTRAINT_VIOLATED:
6377 action_ts[i] = 0;
6378 continue;
6379
6380 case SCF_ERROR_NOT_SET:
6381 case SCF_ERROR_PERMISSION_DENIED:
6382 bad_error("scf_property_get_value",
6383 scf_error());
6384 }
6385 }
6386
6387 r = scf_value_get_integer(val, &action_ts[i]);
6388 assert(r == 0);
6389 }
6390
6391 a = ADMIN_EVENT_MAINT_ON_IMMEDIATE;
6392 if (action_ts[ADMIN_EVENT_MAINT_ON_IMMEDIATE] ||
6393 action_ts[ADMIN_EVENT_MAINT_ON]) {
6394 a = action_ts[ADMIN_EVENT_MAINT_ON_IMMEDIATE] ?
6395 ADMIN_EVENT_MAINT_ON_IMMEDIATE : ADMIN_EVENT_MAINT_ON;
6396
6397 vertex_send_event(vertex, admin_events[a]);
6398 r = libscf_unset_action(h, pg, a, action_ts[a]);
6399 switch (r) {
6400 case 0:
6401 case EACCES:
6402 break;
6403
6404 case ECONNABORTED:
6405 ret = ECONNABORTED;
6406 goto out;
6407
6408 case EPERM:
6409 uu_die("Insufficient privilege.\n");
6410 /* NOTREACHED */
6411
6412 default:
6413 bad_error("libscf_unset_action", r);
6414 }
6415 }
6416
6417 while ((a = next_action(action_ts, NACTIONS)) != -1) {
6418 log_framework(LOG_DEBUG,
6419 "Graph: processing %s action for %s.\n", admin_actions[a],
6420 inst_name);
6421
6422 if (a == ADMIN_EVENT_REFRESH) {
6423 r = dgraph_refresh_instance(vertex, inst);
6424 switch (r) {
6425 case 0:
6426 case ECANCELED:
6427 case EINVAL:
6428 case -1:
6429 break;
6430
6431 case ECONNABORTED:
6432 /* pg & inst are reset now, so just return. */
6433 ret = ECONNABORTED;
6434 goto out;
6435
6436 default:
6437 bad_error("dgraph_refresh_instance", r);
6438 }
6439 }
6440
6441 vertex_send_event(vertex, admin_events[a]);
6442
6443 r = libscf_unset_action(h, pg, a, action_ts[a]);
6444 switch (r) {
6445 case 0:
6446 case EACCES:
6447 break;
6448
6449 case ECONNABORTED:
6450 ret = ECONNABORTED;
6451 goto out;
6452
6453 case EPERM:
6454 uu_die("Insufficient privilege.\n");
6455 /* NOTREACHED */
6456
6457 default:
6458 bad_error("libscf_unset_action", r);
6459 }
6460
6461 action_ts[a] = 0;
6462 }
6463
6464 out:
6465 MUTEX_UNLOCK(&dgraph_lock);
6466
6467 scf_property_destroy(prop);
6468 scf_value_destroy(val);
6469 startd_free(inst_name, max_scf_fmri_size);
6470 return (ret);
6471 }
6472
6473 /*
6474 * inst and pg_name are scratch space, and are unset on entry.
6475 * Returns
6476 * 0 - success
6477 * ECONNRESET - success, but repository handle rebound
6478 * ECONNABORTED - repository connection broken
6479 */
6480 static int
process_pg_event(scf_handle_t * h,scf_propertygroup_t * pg,scf_instance_t * inst,char * pg_name)6481 process_pg_event(scf_handle_t *h, scf_propertygroup_t *pg, scf_instance_t *inst,
6482 char *pg_name)
6483 {
6484 int r;
6485 scf_property_t *prop;
6486 scf_value_t *val;
6487 char *fmri;
6488 boolean_t rebound = B_FALSE, rebind_inst = B_FALSE;
6489
6490 if (scf_pg_get_name(pg, pg_name, max_scf_value_size) < 0) {
6491 switch (scf_error()) {
6492 case SCF_ERROR_CONNECTION_BROKEN:
6493 default:
6494 return (ECONNABORTED);
6495
6496 case SCF_ERROR_DELETED:
6497 return (0);
6498
6499 case SCF_ERROR_NOT_SET:
6500 bad_error("scf_pg_get_name", scf_error());
6501 }
6502 }
6503
6504 if (strcmp(pg_name, SCF_PG_GENERAL) == 0 ||
6505 strcmp(pg_name, SCF_PG_GENERAL_OVR) == 0) {
6506 r = dgraph_update_general(pg);
6507 switch (r) {
6508 case 0:
6509 case ENOTSUP:
6510 case ECANCELED:
6511 return (0);
6512
6513 case ECONNABORTED:
6514 return (ECONNABORTED);
6515
6516 case -1:
6517 /* Error should have been logged. */
6518 return (0);
6519
6520 default:
6521 bad_error("dgraph_update_general", r);
6522 }
6523 } else if (strcmp(pg_name, SCF_PG_RESTARTER_ACTIONS) == 0) {
6524 if (scf_pg_get_parent_instance(pg, inst) != 0) {
6525 switch (scf_error()) {
6526 case SCF_ERROR_CONNECTION_BROKEN:
6527 return (ECONNABORTED);
6528
6529 case SCF_ERROR_DELETED:
6530 case SCF_ERROR_CONSTRAINT_VIOLATED:
6531 /* Ignore commands on services. */
6532 return (0);
6533
6534 case SCF_ERROR_NOT_BOUND:
6535 case SCF_ERROR_HANDLE_MISMATCH:
6536 case SCF_ERROR_NOT_SET:
6537 default:
6538 bad_error("scf_pg_get_parent_instance",
6539 scf_error());
6540 }
6541 }
6542
6543 return (process_actions(h, pg, inst));
6544 }
6545
6546 if (strcmp(pg_name, SCF_PG_OPTIONS) != 0 &&
6547 strcmp(pg_name, SCF_PG_OPTIONS_OVR) != 0)
6548 return (0);
6549
6550 /*
6551 * We only care about the options[_ovr] property groups of our own
6552 * instance, so get the fmri and compare. Plus, once we know it's
6553 * correct, if the repository connection is broken we know exactly what
6554 * property group we were operating on, and can look it up again.
6555 */
6556 if (scf_pg_get_parent_instance(pg, inst) != 0) {
6557 switch (scf_error()) {
6558 case SCF_ERROR_CONNECTION_BROKEN:
6559 return (ECONNABORTED);
6560
6561 case SCF_ERROR_DELETED:
6562 case SCF_ERROR_CONSTRAINT_VIOLATED:
6563 return (0);
6564
6565 case SCF_ERROR_HANDLE_MISMATCH:
6566 case SCF_ERROR_NOT_BOUND:
6567 case SCF_ERROR_NOT_SET:
6568 default:
6569 bad_error("scf_pg_get_parent_instance",
6570 scf_error());
6571 }
6572 }
6573
6574 switch (r = libscf_instance_get_fmri(inst, &fmri)) {
6575 case 0:
6576 break;
6577
6578 case ECONNABORTED:
6579 return (ECONNABORTED);
6580
6581 case ECANCELED:
6582 return (0);
6583
6584 default:
6585 bad_error("libscf_instance_get_fmri", r);
6586 }
6587
6588 if (strcmp(fmri, SCF_SERVICE_STARTD) != 0) {
6589 startd_free(fmri, max_scf_fmri_size);
6590 return (0);
6591 }
6592
6593 /*
6594 * update the information events flag
6595 */
6596 if (strcmp(pg_name, SCF_PG_OPTIONS) == 0)
6597 info_events_all = libscf_get_info_events_all(pg);
6598
6599 prop = safe_scf_property_create(h);
6600 val = safe_scf_value_create(h);
6601
6602 if (strcmp(pg_name, SCF_PG_OPTIONS_OVR) == 0) {
6603 /* See if we need to set the runlevel. */
6604 /* CONSTCOND */
6605 if (0) {
6606 rebind_pg:
6607 libscf_handle_rebind(h);
6608 rebound = B_TRUE;
6609
6610 r = libscf_lookup_instance(SCF_SERVICE_STARTD, inst);
6611 switch (r) {
6612 case 0:
6613 break;
6614
6615 case ECONNABORTED:
6616 goto rebind_pg;
6617
6618 case ENOENT:
6619 goto out;
6620
6621 case EINVAL:
6622 case ENOTSUP:
6623 bad_error("libscf_lookup_instance", r);
6624 }
6625
6626 if (scf_instance_get_pg(inst, pg_name, pg) != 0) {
6627 switch (scf_error()) {
6628 case SCF_ERROR_DELETED:
6629 case SCF_ERROR_NOT_FOUND:
6630 goto out;
6631
6632 case SCF_ERROR_CONNECTION_BROKEN:
6633 goto rebind_pg;
6634
6635 case SCF_ERROR_HANDLE_MISMATCH:
6636 case SCF_ERROR_NOT_BOUND:
6637 case SCF_ERROR_NOT_SET:
6638 case SCF_ERROR_INVALID_ARGUMENT:
6639 default:
6640 bad_error("scf_instance_get_pg",
6641 scf_error());
6642 }
6643 }
6644 }
6645
6646 if (scf_pg_get_property(pg, "runlevel", prop) == 0) {
6647 r = dgraph_set_runlevel(pg, prop);
6648 switch (r) {
6649 case ECONNRESET:
6650 rebound = B_TRUE;
6651 rebind_inst = B_TRUE;
6652 /* FALLTHROUGH */
6653
6654 case 0:
6655 break;
6656
6657 case ECONNABORTED:
6658 goto rebind_pg;
6659
6660 case ECANCELED:
6661 goto out;
6662
6663 default:
6664 bad_error("dgraph_set_runlevel", r);
6665 }
6666 } else {
6667 switch (scf_error()) {
6668 case SCF_ERROR_CONNECTION_BROKEN:
6669 default:
6670 goto rebind_pg;
6671
6672 case SCF_ERROR_DELETED:
6673 goto out;
6674
6675 case SCF_ERROR_NOT_FOUND:
6676 break;
6677
6678 case SCF_ERROR_INVALID_ARGUMENT:
6679 case SCF_ERROR_HANDLE_MISMATCH:
6680 case SCF_ERROR_NOT_BOUND:
6681 case SCF_ERROR_NOT_SET:
6682 bad_error("scf_pg_get_property", scf_error());
6683 }
6684 }
6685 }
6686
6687 if (rebind_inst) {
6688 lookup_inst:
6689 r = libscf_lookup_instance(SCF_SERVICE_STARTD, inst);
6690 switch (r) {
6691 case 0:
6692 break;
6693
6694 case ECONNABORTED:
6695 libscf_handle_rebind(h);
6696 rebound = B_TRUE;
6697 goto lookup_inst;
6698
6699 case ENOENT:
6700 goto out;
6701
6702 case EINVAL:
6703 case ENOTSUP:
6704 bad_error("libscf_lookup_instance", r);
6705 }
6706 }
6707
6708 r = libscf_get_milestone(inst, prop, val, fmri, max_scf_fmri_size);
6709 switch (r) {
6710 case 0:
6711 break;
6712
6713 case ECONNABORTED:
6714 libscf_handle_rebind(h);
6715 rebound = B_TRUE;
6716 goto lookup_inst;
6717
6718 case EINVAL:
6719 log_error(LOG_NOTICE,
6720 "%s/%s property of %s is misconfigured.\n", pg_name,
6721 SCF_PROPERTY_MILESTONE, SCF_SERVICE_STARTD);
6722 /* FALLTHROUGH */
6723
6724 case ECANCELED:
6725 case ENOENT:
6726 (void) strcpy(fmri, "all");
6727 break;
6728
6729 default:
6730 bad_error("libscf_get_milestone", r);
6731 }
6732
6733 r = dgraph_set_milestone(fmri, h, B_FALSE);
6734 switch (r) {
6735 case 0:
6736 case ECONNRESET:
6737 case EALREADY:
6738 break;
6739
6740 case EINVAL:
6741 log_error(LOG_WARNING, "Milestone %s is invalid.\n", fmri);
6742 break;
6743
6744 case ENOENT:
6745 log_error(LOG_WARNING, "Milestone %s does not exist.\n", fmri);
6746 break;
6747
6748 default:
6749 bad_error("dgraph_set_milestone", r);
6750 }
6751
6752 out:
6753 startd_free(fmri, max_scf_fmri_size);
6754 scf_value_destroy(val);
6755 scf_property_destroy(prop);
6756
6757 return (rebound ? ECONNRESET : 0);
6758 }
6759
6760 /*
6761 * process_delete() deletes an instance from the dgraph if 'fmri' is an
6762 * instance fmri or if 'fmri' matches the 'general' property group of an
6763 * instance (or the 'general/enabled' property).
6764 *
6765 * 'fmri' may be overwritten and cannot be trusted on return by the caller.
6766 */
6767 static void
process_delete(char * fmri,scf_handle_t * h)6768 process_delete(char *fmri, scf_handle_t *h)
6769 {
6770 char *lfmri, *end_inst_fmri;
6771 const char *inst_name = NULL;
6772 const char *pg_name = NULL;
6773 const char *prop_name = NULL;
6774
6775 lfmri = safe_strdup(fmri);
6776
6777 /* Determine if the FMRI is a property group or instance */
6778 if (scf_parse_svc_fmri(lfmri, NULL, NULL, &inst_name, &pg_name,
6779 &prop_name) != SCF_SUCCESS) {
6780 log_error(LOG_WARNING,
6781 "Received invalid FMRI \"%s\" from repository server.\n",
6782 fmri);
6783 } else if (inst_name != NULL && pg_name == NULL) {
6784 (void) dgraph_remove_instance(fmri, h);
6785 } else if (inst_name != NULL && pg_name != NULL) {
6786 /*
6787 * If we're deleting the 'general' property group or
6788 * 'general/enabled' property then the whole instance
6789 * must be removed from the dgraph.
6790 */
6791 if (strcmp(pg_name, SCF_PG_GENERAL) != 0) {
6792 free(lfmri);
6793 return;
6794 }
6795
6796 if (prop_name != NULL &&
6797 strcmp(prop_name, SCF_PROPERTY_ENABLED) != 0) {
6798 free(lfmri);
6799 return;
6800 }
6801
6802 /*
6803 * Because the instance has already been deleted from the
6804 * repository, we cannot use any scf_ functions to retrieve
6805 * the instance FMRI however we can easily reconstruct it
6806 * manually.
6807 */
6808 end_inst_fmri = strstr(fmri, SCF_FMRI_PROPERTYGRP_PREFIX);
6809 if (end_inst_fmri == NULL)
6810 bad_error("process_delete", 0);
6811
6812 end_inst_fmri[0] = '\0';
6813
6814 (void) dgraph_remove_instance(fmri, h);
6815 }
6816
6817 free(lfmri);
6818 }
6819
6820 /*ARGSUSED*/
6821 void *
repository_event_thread(void * unused)6822 repository_event_thread(void *unused)
6823 {
6824 scf_handle_t *h;
6825 scf_propertygroup_t *pg;
6826 scf_instance_t *inst;
6827 char *fmri = startd_alloc(max_scf_fmri_size);
6828 char *pg_name = startd_alloc(max_scf_value_size);
6829 int r;
6830
6831 h = libscf_handle_create_bound_loop();
6832
6833 pg = safe_scf_pg_create(h);
6834 inst = safe_scf_instance_create(h);
6835
6836 retry:
6837 if (_scf_notify_add_pgtype(h, SCF_GROUP_FRAMEWORK) != SCF_SUCCESS) {
6838 if (scf_error() == SCF_ERROR_CONNECTION_BROKEN) {
6839 libscf_handle_rebind(h);
6840 } else {
6841 log_error(LOG_WARNING,
6842 "Couldn't set up repository notification "
6843 "for property group type %s: %s\n",
6844 SCF_GROUP_FRAMEWORK, scf_strerror(scf_error()));
6845
6846 (void) sleep(1);
6847 }
6848
6849 goto retry;
6850 }
6851
6852 /*CONSTCOND*/
6853 while (1) {
6854 ssize_t res;
6855
6856 /* Note: fmri is only set on delete events. */
6857 res = _scf_notify_wait(pg, fmri, max_scf_fmri_size);
6858 if (res < 0) {
6859 libscf_handle_rebind(h);
6860 goto retry;
6861 } else if (res == 0) {
6862 /*
6863 * property group modified. inst and pg_name are
6864 * pre-allocated scratch space.
6865 */
6866 if (scf_pg_update(pg) < 0) {
6867 switch (scf_error()) {
6868 case SCF_ERROR_DELETED:
6869 continue;
6870
6871 case SCF_ERROR_CONNECTION_BROKEN:
6872 log_error(LOG_WARNING,
6873 "Lost repository event due to "
6874 "disconnection.\n");
6875 libscf_handle_rebind(h);
6876 goto retry;
6877
6878 case SCF_ERROR_NOT_BOUND:
6879 case SCF_ERROR_NOT_SET:
6880 default:
6881 bad_error("scf_pg_update", scf_error());
6882 }
6883 }
6884
6885 r = process_pg_event(h, pg, inst, pg_name);
6886 switch (r) {
6887 case 0:
6888 break;
6889
6890 case ECONNABORTED:
6891 log_error(LOG_WARNING, "Lost repository event "
6892 "due to disconnection.\n");
6893 libscf_handle_rebind(h);
6894 /* FALLTHROUGH */
6895
6896 case ECONNRESET:
6897 goto retry;
6898
6899 default:
6900 bad_error("process_pg_event", r);
6901 }
6902 } else {
6903 /*
6904 * Service, instance, or pg deleted.
6905 * Don't trust fmri on return.
6906 */
6907 process_delete(fmri, h);
6908 }
6909 }
6910
6911 /*NOTREACHED*/
6912 return (NULL);
6913 }
6914
6915 void
graph_engine_start()6916 graph_engine_start()
6917 {
6918 int err;
6919
6920 (void) startd_thread_create(graph_thread, NULL);
6921
6922 MUTEX_LOCK(&dgraph_lock);
6923 while (!initial_milestone_set) {
6924 err = pthread_cond_wait(&initial_milestone_cv, &dgraph_lock);
6925 assert(err == 0);
6926 }
6927 MUTEX_UNLOCK(&dgraph_lock);
6928
6929 (void) startd_thread_create(repository_event_thread, NULL);
6930 (void) startd_thread_create(graph_event_thread, NULL);
6931 }
6932