1 /* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21 22 /* 23 * Copyright (c) 2004, 2010, Oracle and/or its affiliates. All rights reserved. 24 * Copyright (c) 2015, Syneto S.R.L. All rights reserved. 25 * Copyright 2016 RackTop Systems. 26 */ 27 28 /* 29 * graph.c - master restarter graph engine 30 * 31 * The graph engine keeps a dependency graph of all service instances on the 32 * system, as recorded in the repository. It decides when services should 33 * be brought up or down based on service states and dependencies and sends 34 * commands to restarters to effect any changes. It also executes 35 * administrator commands sent by svcadm via the repository. 36 * 37 * The graph is stored in uu_list_t *dgraph and its vertices are 38 * graph_vertex_t's, each of which has a name and an integer id unique to 39 * its name (see dict.c). A vertex's type attribute designates the type 40 * of object it represents: GVT_INST for service instances, GVT_SVC for 41 * service objects (since service instances may depend on another service, 42 * rather than service instance), GVT_FILE for files (which services may 43 * depend on), and GVT_GROUP for dependencies on multiple objects. GVT_GROUP 44 * vertices are necessary because dependency lists may have particular 45 * grouping types (require any, require all, optional, or exclude) and 46 * event-propagation characteristics. 47 * 48 * The initial graph is built by libscf_populate_graph() invoking 49 * dgraph_add_instance() for each instance in the repository. The function 50 * adds a GVT_SVC vertex for the service if one does not already exist, adds 51 * a GVT_INST vertex named by the FMRI of the instance, and sets up the edges. 52 * The resulting web of vertices & edges associated with an instance's vertex 53 * includes 54 * 55 * - an edge from the GVT_SVC vertex for the instance's service 56 * 57 * - an edge to the GVT_INST vertex of the instance's resarter, if its 58 * restarter is not svc.startd 59 * 60 * - edges from other GVT_INST vertices if the instance is a restarter 61 * 62 * - for each dependency property group in the instance's "running" 63 * snapshot, an edge to a GVT_GROUP vertex named by the FMRI of the 64 * instance and the name of the property group 65 * 66 * - for each value of the "entities" property in each dependency property 67 * group, an edge from the corresponding GVT_GROUP vertex to a 68 * GVT_INST, GVT_SVC, or GVT_FILE vertex 69 * 70 * - edges from GVT_GROUP vertices for each dependent instance 71 * 72 * After the edges are set up the vertex's GV_CONFIGURED flag is set. If 73 * there are problems, or if a service is mentioned in a dependency but does 74 * not exist in the repository, the GV_CONFIGURED flag will be clear. 75 * 76 * The graph and all of its vertices are protected by the dgraph_lock mutex. 77 * See restarter.c for more information. 78 * 79 * The properties of an instance fall into two classes: immediate and 80 * snapshotted. Immediate properties should have an immediate effect when 81 * changed. Snapshotted properties should be read from a snapshot, so they 82 * only change when the snapshot changes. The immediate properties used by 83 * the graph engine are general/enabled, general/restarter, and the properties 84 * in the restarter_actions property group. Since they are immediate, they 85 * are not read out of a snapshot. The snapshotted properties used by the 86 * graph engine are those in the property groups with type "dependency" and 87 * are read out of the "running" snapshot. The "running" snapshot is created 88 * by the the graph engine as soon as possible, and it is updated, along with 89 * in-core copies of the data (dependency information for the graph engine) on 90 * receipt of the refresh command from svcadm. In addition, the graph engine 91 * updates the "start" snapshot from the "running" snapshot whenever a service 92 * comes online. 93 * 94 * When a DISABLE event is requested by the administrator, svc.startd shutdown 95 * the dependents first before shutting down the requested service. 96 * In graph_enable_by_vertex, we create a subtree that contains the dependent 97 * vertices by marking those vertices with the GV_TOOFFLINE flag. And we mark 98 * the vertex to disable with the GV_TODISABLE flag. Once the tree is created, 99 * we send the _ADMIN_DISABLE event to the leaves. The leaves will then 100 * transition from STATE_ONLINE/STATE_DEGRADED to STATE_OFFLINE/STATE_MAINT. 101 * In gt_enter_offline and gt_enter_maint if the vertex was in a subtree then 102 * we clear the GV_TOOFFLINE flag and walk the dependencies to offline the new 103 * exposed leaves. We do the same until we reach the last leaf (the one with 104 * the GV_TODISABLE flag). If the vertex to disable is also part of a larger 105 * subtree (eg. multiple DISABLE events on vertices in the same subtree) then 106 * once the first vertex is disabled (GV_TODISABLE flag is removed), we 107 * continue to propagate the offline event to the vertex's dependencies. 108 * 109 * 110 * SMF state transition notifications 111 * 112 * When an instance of a service managed by SMF changes state, svc.startd may 113 * publish a GPEC sysevent. All transitions to or from maintenance, a 114 * transition cause by a hardware error will generate an event. 115 * Other transitions will generate an event if there exist notification 116 * parameter for that transition. Notification parameters are stored in the 117 * SMF repository for the service/instance they refer to. System-wide 118 * notification parameters are stored in the global instance. 119 * svc.startd can be told to send events for all SMF state transitions despite 120 * of notification parameters by setting options/info_events_all to true in 121 * restarter:default 122 * 123 * The set of transitions that generate events is cached in the 124 * dgraph_vertex_t gv_stn_tset for service/instance and in the global 125 * stn_global for the system-wide set. They are re-read when instances are 126 * refreshed. 127 * 128 * The GPEC events published by svc.startd are consumed by fmd(1M). After 129 * processing these events, fmd(1M) publishes the processed events to 130 * notification agents. The notification agents read the notification 131 * parameters from the SMF repository through libscf(3LIB) interfaces and send 132 * the notification, or not, based on those parameters. 133 * 134 * Subscription and publishing to the GPEC channels is done with the 135 * libfmevent(3LIB) wrappers fmev_[r]publish_*() and 136 * fmev_shdl_(un)subscribe(). 137 * 138 */ 139 140 #include <sys/uadmin.h> 141 #include <sys/wait.h> 142 143 #include <assert.h> 144 #include <errno.h> 145 #include <fcntl.h> 146 #include <fm/libfmevent.h> 147 #include <libscf.h> 148 #include <libscf_priv.h> 149 #include <librestart.h> 150 #include <libuutil.h> 151 #include <locale.h> 152 #include <poll.h> 153 #include <pthread.h> 154 #include <signal.h> 155 #include <stddef.h> 156 #include <stdio.h> 157 #include <stdlib.h> 158 #include <string.h> 159 #include <strings.h> 160 #include <sys/statvfs.h> 161 #include <sys/uadmin.h> 162 #include <zone.h> 163 #if defined(__i386) 164 #include <libgrubmgmt.h> 165 #endif /* __i386 */ 166 167 #include "startd.h" 168 #include "protocol.h" 169 170 171 #define MILESTONE_NONE ((graph_vertex_t *)1) 172 173 #define CONSOLE_LOGIN_FMRI "svc:/system/console-login:default" 174 #define FS_MINIMAL_FMRI "svc:/system/filesystem/minimal:default" 175 176 #define VERTEX_REMOVED 0 /* vertex has been freed */ 177 #define VERTEX_INUSE 1 /* vertex is still in use */ 178 179 #define IS_ENABLED(v) ((v)->gv_flags & (GV_ENABLED | GV_ENBLD_NOOVR)) 180 181 /* 182 * stn_global holds the tset for the system wide notification parameters. 183 * It is updated on refresh of svc:/system/svc/global:default 184 * 185 * There are two assumptions that relax the need for a mutex: 186 * 1. 32-bit value assignments are atomic 187 * 2. Its value is consumed only in one point at 188 * dgraph_state_transition_notify(). There are no test and set races. 189 * 190 * If either assumption is broken, we'll need a mutex to synchronize 191 * access to stn_global 192 */ 193 int32_t stn_global; 194 /* 195 * info_events_all holds a flag to override notification parameters and send 196 * Information events for all state transitions. 197 * same about the need of a mutex here. 198 */ 199 int info_events_all; 200 201 /* 202 * Services in these states are not considered 'down' by the 203 * milestone/shutdown code. 204 */ 205 #define up_state(state) ((state) == RESTARTER_STATE_ONLINE || \ 206 (state) == RESTARTER_STATE_DEGRADED || \ 207 (state) == RESTARTER_STATE_OFFLINE) 208 209 #define is_depgrp_bypassed(v) ((v->gv_type == GVT_GROUP) && \ 210 ((v->gv_depgroup == DEPGRP_EXCLUDE_ALL) || \ 211 (v->gv_restart < RERR_RESTART))) 212 213 static uu_list_pool_t *graph_edge_pool, *graph_vertex_pool; 214 static uu_list_t *dgraph; 215 static pthread_mutex_t dgraph_lock; 216 217 /* 218 * milestone indicates the current subgraph. When NULL, it is the entire 219 * graph. When MILESTONE_NONE, it is the empty graph. Otherwise, it is all 220 * services on which the target vertex depends. 221 */ 222 static graph_vertex_t *milestone = NULL; 223 static boolean_t initial_milestone_set = B_FALSE; 224 static pthread_cond_t initial_milestone_cv = PTHREAD_COND_INITIALIZER; 225 226 /* protected by dgraph_lock */ 227 static boolean_t sulogin_thread_running = B_FALSE; 228 static boolean_t sulogin_running = B_FALSE; 229 static boolean_t console_login_ready = B_FALSE; 230 231 /* Number of services to come down to complete milestone transition. */ 232 static uint_t non_subgraph_svcs; 233 234 /* 235 * These variables indicate what should be done when we reach the milestone 236 * target milestone, i.e., when non_subgraph_svcs == 0. They are acted upon in 237 * dgraph_set_instance_state(). 238 */ 239 static int halting = -1; 240 static boolean_t go_single_user_mode = B_FALSE; 241 static boolean_t go_to_level1 = B_FALSE; 242 243 /* 244 * Tracks when we started halting. 245 */ 246 static time_t halting_time = 0; 247 248 /* 249 * This tracks the legacy runlevel to ensure we signal init and manage 250 * utmpx entries correctly. 251 */ 252 static char current_runlevel = '\0'; 253 254 /* Number of single user threads currently running */ 255 static pthread_mutex_t single_user_thread_lock; 256 static int single_user_thread_count = 0; 257 258 /* Statistics for dependency cycle-checking */ 259 static u_longlong_t dep_inserts = 0; 260 static u_longlong_t dep_cycle_ns = 0; 261 static u_longlong_t dep_insert_ns = 0; 262 263 264 static const char * const emsg_invalid_restarter = 265 "Transitioning %s to maintenance, restarter FMRI %s is invalid " 266 "(see 'svcs -xv' for details).\n"; 267 static const char * const console_login_fmri = CONSOLE_LOGIN_FMRI; 268 static const char * const single_user_fmri = SCF_MILESTONE_SINGLE_USER; 269 static const char * const multi_user_fmri = SCF_MILESTONE_MULTI_USER; 270 static const char * const multi_user_svr_fmri = SCF_MILESTONE_MULTI_USER_SERVER; 271 272 273 /* 274 * These services define the system being "up". If none of them can come 275 * online, then we will run sulogin on the console. Note that the install ones 276 * are for the miniroot and when installing CDs after the first. can_come_up() 277 * does the decision making, and an sulogin_thread() runs sulogin, which can be 278 * started by dgraph_set_instance_state() or single_user_thread(). 279 * 280 * NOTE: can_come_up() relies on SCF_MILESTONE_SINGLE_USER being the first 281 * entry, which is only used when booting_to_single_user (boot -s) is set. 282 * This is because when doing a "boot -s", sulogin is started from specials.c 283 * after milestone/single-user comes online, for backwards compatibility. 284 * In this case, SCF_MILESTONE_SINGLE_USER needs to be part of up_svcs 285 * to ensure sulogin will be spawned if milestone/single-user cannot be reached. 286 */ 287 static const char * const up_svcs[] = { 288 SCF_MILESTONE_SINGLE_USER, 289 CONSOLE_LOGIN_FMRI, 290 "svc:/system/install-setup:default", 291 "svc:/system/install:default", 292 NULL 293 }; 294 295 /* This array must have an element for each non-NULL element of up_svcs[]. */ 296 static graph_vertex_t *up_svcs_p[] = { NULL, NULL, NULL, NULL }; 297 298 /* These are for seed repository magic. See can_come_up(). */ 299 static const char * const manifest_import = SCF_INSTANCE_MI; 300 static graph_vertex_t *manifest_import_p = NULL; 301 302 303 static char target_milestone_as_runlevel(void); 304 static void graph_runlevel_changed(char rl, int online); 305 static int dgraph_set_milestone(const char *, scf_handle_t *, boolean_t); 306 static boolean_t should_be_in_subgraph(graph_vertex_t *v); 307 static int mark_subtree(graph_edge_t *, void *); 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 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 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 * 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 * 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 * 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 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 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 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 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 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 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 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 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 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 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 * 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 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 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 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 808 graph_post_su_online(void) 809 { 810 graph_runlevel_changed('S', 1); 811 } 812 813 static void 814 graph_post_su_disable(void) 815 { 816 graph_runlevel_changed('S', 0); 817 } 818 819 static void 820 graph_post_mu_online(void) 821 { 822 graph_runlevel_changed('2', 1); 823 } 824 825 static void 826 graph_post_mu_disable(void) 827 { 828 graph_runlevel_changed('2', 0); 829 } 830 831 static void 832 graph_post_mus_online(void) 833 { 834 graph_runlevel_changed('3', 1); 835 } 836 837 static 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 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 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 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 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 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 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 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 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 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 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 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 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 instances */ 1331 if ((v->gv_flags & GV_CONFIGURED) == 0) 1332 continue; 1333 1334 if (v->gv_state == RESTARTER_STATE_MAINT) 1335 continue; 1336 1337 any_qualified = B_TRUE; 1338 if (v->gv_state == RESTARTER_STATE_OFFLINE) { 1339 /* 1340 * For offline dependencies, treat unsatisfiable 1341 * as satisfied. 1342 */ 1343 i = dependency_satisfied(v, B_TRUE); 1344 if (i == -1) 1345 i = 1; 1346 } else if (v->gv_state == RESTARTER_STATE_DISABLED) { 1347 /* 1348 * The service is enabled, but hasn't 1349 * transitioned out of disabled yet. Treat it 1350 * as unsatisfied (not unsatisfiable). 1351 */ 1352 i = v->gv_flags & GV_ENABLED ? 0 : 1; 1353 } else { 1354 i = dependency_satisfied(v, satbility); 1355 } 1356 break; 1357 1358 case GVT_FILE: 1359 any_qualified = B_TRUE; 1360 i = dependency_satisfied(v, satbility); 1361 1362 break; 1363 1364 case GVT_SVC: { 1365 boolean_t svc_any_qualified; 1366 boolean_t svc_satisfied; 1367 boolean_t svc_satisfiable; 1368 graph_vertex_t *v2; 1369 graph_edge_t *e2; 1370 1371 svc_any_qualified = B_FALSE; 1372 svc_satisfied = B_FALSE; 1373 svc_satisfiable = B_FALSE; 1374 1375 for (e2 = uu_list_first(v->gv_dependencies); 1376 e2 != NULL; 1377 e2 = uu_list_next(v->gv_dependencies, e2)) { 1378 v2 = e2->ge_vertex; 1379 assert(v2->gv_type == GVT_INST); 1380 1381 if ((v2->gv_flags & GV_CONFIGURED) == 0) 1382 continue; 1383 1384 if (v2->gv_state == RESTARTER_STATE_MAINT) 1385 continue; 1386 1387 svc_any_qualified = B_TRUE; 1388 if (v2->gv_state == RESTARTER_STATE_OFFLINE) { 1389 /* 1390 * For offline dependencies, treat 1391 * unsatisfiable as satisfied. 1392 */ 1393 i = dependency_satisfied(v2, B_TRUE); 1394 if (i == -1) 1395 i = 1; 1396 } else if (v2->gv_state == 1397 RESTARTER_STATE_DISABLED) { 1398 i = v2->gv_flags & GV_ENABLED ? 0 : 1; 1399 } else { 1400 i = dependency_satisfied(v2, satbility); 1401 } 1402 1403 if (i == 1) { 1404 svc_satisfied = B_TRUE; 1405 break; 1406 } 1407 if (i == 0) 1408 svc_satisfiable = B_TRUE; 1409 } 1410 1411 if (!svc_any_qualified) 1412 continue; 1413 any_qualified = B_TRUE; 1414 if (svc_satisfied) { 1415 i = 1; 1416 } else if (svc_satisfiable) { 1417 i = 0; 1418 } else { 1419 i = -1; 1420 } 1421 break; 1422 } 1423 1424 case GVT_GROUP: 1425 default: 1426 #ifndef NDEBUG 1427 uu_warn("%s:%d: Unexpected vertex type %d.\n", __FILE__, 1428 __LINE__, v->gv_type); 1429 #endif 1430 abort(); 1431 } 1432 1433 if (i == 1) 1434 continue; 1435 1436 log_framework2(LOG_DEBUG, DEBUG_DEPENDENCIES, 1437 "optional_all(%s): %s is unsatisfi%s.\n", groupv->gv_name, 1438 v->gv_name, i == 0 ? "ed" : "able"); 1439 1440 if (!satbility) 1441 return (0); 1442 if (i == -1) 1443 return (-1); 1444 any_unsatisfied = B_TRUE; 1445 } 1446 1447 if (!any_qualified) 1448 return (1); 1449 1450 return (any_unsatisfied ? 0 : 1); 1451 } 1452 1453 /* 1454 * An exclude_all dependency is unsatisfied if any non-service element is 1455 * satisfied or any service instance which is configured, enabled, and not in 1456 * maintenance is satisfied. Usually when unsatisfied, it is also 1457 * unsatisfiable. 1458 */ 1459 #define LOG_EXCLUDE(u, v) \ 1460 log_framework2(LOG_DEBUG, DEBUG_DEPENDENCIES, \ 1461 "exclude_all(%s): %s is satisfied.\n", \ 1462 (u)->gv_name, (v)->gv_name) 1463 1464 /* ARGSUSED */ 1465 static int 1466 exclude_all_satisfied(graph_vertex_t *groupv, boolean_t satbility) 1467 { 1468 graph_edge_t *edge, *e2; 1469 graph_vertex_t *v, *v2; 1470 1471 for (edge = uu_list_first(groupv->gv_dependencies); 1472 edge != NULL; 1473 edge = uu_list_next(groupv->gv_dependencies, edge)) { 1474 v = edge->ge_vertex; 1475 1476 switch (v->gv_type) { 1477 case GVT_INST: 1478 if ((v->gv_flags & GV_CONFIGURED) == 0) 1479 continue; 1480 1481 switch (v->gv_state) { 1482 case RESTARTER_STATE_ONLINE: 1483 case RESTARTER_STATE_DEGRADED: 1484 LOG_EXCLUDE(groupv, v); 1485 return (v->gv_flags & GV_ENABLED ? -1 : 0); 1486 1487 case RESTARTER_STATE_OFFLINE: 1488 case RESTARTER_STATE_UNINIT: 1489 LOG_EXCLUDE(groupv, v); 1490 return (0); 1491 1492 case RESTARTER_STATE_DISABLED: 1493 case RESTARTER_STATE_MAINT: 1494 continue; 1495 1496 default: 1497 #ifndef NDEBUG 1498 uu_warn("%s:%d: Unexpected vertex state %d.\n", 1499 __FILE__, __LINE__, v->gv_state); 1500 #endif 1501 abort(); 1502 } 1503 /* NOTREACHED */ 1504 1505 case GVT_SVC: 1506 break; 1507 1508 case GVT_FILE: 1509 if (!file_ready(v)) 1510 continue; 1511 LOG_EXCLUDE(groupv, v); 1512 return (-1); 1513 1514 case GVT_GROUP: 1515 default: 1516 #ifndef NDEBUG 1517 uu_warn("%s:%d: Unexpected vertex type %d.\n", __FILE__, 1518 __LINE__, v->gv_type); 1519 #endif 1520 abort(); 1521 } 1522 1523 /* v represents a service */ 1524 if (uu_list_numnodes(v->gv_dependencies) == 0) 1525 continue; 1526 1527 for (e2 = uu_list_first(v->gv_dependencies); 1528 e2 != NULL; 1529 e2 = uu_list_next(v->gv_dependencies, e2)) { 1530 v2 = e2->ge_vertex; 1531 assert(v2->gv_type == GVT_INST); 1532 1533 if ((v2->gv_flags & GV_CONFIGURED) == 0) 1534 continue; 1535 1536 switch (v2->gv_state) { 1537 case RESTARTER_STATE_ONLINE: 1538 case RESTARTER_STATE_DEGRADED: 1539 LOG_EXCLUDE(groupv, v2); 1540 return (v2->gv_flags & GV_ENABLED ? -1 : 0); 1541 1542 case RESTARTER_STATE_OFFLINE: 1543 case RESTARTER_STATE_UNINIT: 1544 LOG_EXCLUDE(groupv, v2); 1545 return (0); 1546 1547 case RESTARTER_STATE_DISABLED: 1548 case RESTARTER_STATE_MAINT: 1549 continue; 1550 1551 default: 1552 #ifndef NDEBUG 1553 uu_warn("%s:%d: Unexpected vertex type %d.\n", 1554 __FILE__, __LINE__, v2->gv_type); 1555 #endif 1556 abort(); 1557 } 1558 } 1559 } 1560 1561 return (1); 1562 } 1563 1564 /* 1565 * int instance_satisfied() 1566 * Determine if all the dependencies are satisfied for the supplied instance 1567 * vertex. Return 1 if they are, 0 if they aren't, and -1 if they won't be 1568 * without administrator intervention. 1569 */ 1570 static int 1571 instance_satisfied(graph_vertex_t *v, boolean_t satbility) 1572 { 1573 assert(v->gv_type == GVT_INST); 1574 assert(!inst_running(v)); 1575 1576 return (require_all_satisfied(v, satbility)); 1577 } 1578 1579 /* 1580 * Decide whether v can satisfy a dependency. v can either be a child of 1581 * a group vertex, or of an instance vertex. 1582 */ 1583 static int 1584 dependency_satisfied(graph_vertex_t *v, boolean_t satbility) 1585 { 1586 switch (v->gv_type) { 1587 case GVT_INST: 1588 if ((v->gv_flags & GV_CONFIGURED) == 0) { 1589 if (v->gv_flags & GV_DEATHROW) { 1590 /* 1591 * A dependency on an instance with GV_DEATHROW 1592 * flag is always considered as satisfied. 1593 */ 1594 return (1); 1595 } 1596 return (-1); 1597 } 1598 1599 /* 1600 * Any vertex with the GV_TOOFFLINE flag set is guaranteed 1601 * to have its dependencies unsatisfiable. 1602 */ 1603 if (v->gv_flags & GV_TOOFFLINE) 1604 return (-1); 1605 1606 switch (v->gv_state) { 1607 case RESTARTER_STATE_ONLINE: 1608 case RESTARTER_STATE_DEGRADED: 1609 return (1); 1610 1611 case RESTARTER_STATE_OFFLINE: 1612 if (!satbility) 1613 return (0); 1614 return (instance_satisfied(v, satbility) != -1 ? 1615 0 : -1); 1616 1617 case RESTARTER_STATE_DISABLED: 1618 case RESTARTER_STATE_MAINT: 1619 return (-1); 1620 1621 case RESTARTER_STATE_UNINIT: 1622 return (0); 1623 1624 default: 1625 #ifndef NDEBUG 1626 uu_warn("%s:%d: Unexpected vertex state %d.\n", 1627 __FILE__, __LINE__, v->gv_state); 1628 #endif 1629 abort(); 1630 /* NOTREACHED */ 1631 } 1632 1633 case GVT_SVC: 1634 if (uu_list_numnodes(v->gv_dependencies) == 0) 1635 return (-1); 1636 return (require_any_satisfied(v, satbility)); 1637 1638 case GVT_FILE: 1639 /* i.e., we assume files will not be automatically generated */ 1640 return (file_ready(v) ? 1 : -1); 1641 1642 case GVT_GROUP: 1643 break; 1644 1645 default: 1646 #ifndef NDEBUG 1647 uu_warn("%s:%d: Unexpected node type %d.\n", __FILE__, __LINE__, 1648 v->gv_type); 1649 #endif 1650 abort(); 1651 /* NOTREACHED */ 1652 } 1653 1654 switch (v->gv_depgroup) { 1655 case DEPGRP_REQUIRE_ANY: 1656 return (require_any_satisfied(v, satbility)); 1657 1658 case DEPGRP_REQUIRE_ALL: 1659 return (require_all_satisfied(v, satbility)); 1660 1661 case DEPGRP_OPTIONAL_ALL: 1662 return (optional_all_satisfied(v, satbility)); 1663 1664 case DEPGRP_EXCLUDE_ALL: 1665 return (exclude_all_satisfied(v, satbility)); 1666 1667 default: 1668 #ifndef NDEBUG 1669 uu_warn("%s:%d: Unknown dependency grouping %d.\n", __FILE__, 1670 __LINE__, v->gv_depgroup); 1671 #endif 1672 abort(); 1673 } 1674 } 1675 1676 void 1677 graph_start_if_satisfied(graph_vertex_t *v) 1678 { 1679 if (v->gv_state == RESTARTER_STATE_OFFLINE && 1680 instance_satisfied(v, B_FALSE) == 1) { 1681 if (v->gv_start_f == NULL) 1682 vertex_send_event(v, RESTARTER_EVENT_TYPE_START); 1683 else 1684 v->gv_start_f(v); 1685 } 1686 } 1687 1688 /* 1689 * propagate_satbility() 1690 * 1691 * This function is used when the given vertex changes state in such a way that 1692 * one of its dependents may become unsatisfiable. This happens when an 1693 * instance transitions between offline -> online, or from !running -> 1694 * maintenance, as well as when an instance is removed from the graph. 1695 * 1696 * We have to walk all the dependents, since optional_all dependencies several 1697 * levels up could become (un)satisfied, instead of unsatisfiable. For example, 1698 * 1699 * +-----+ optional_all +-----+ require_all +-----+ 1700 * | A |--------------->| B |-------------->| C | 1701 * +-----+ +-----+ +-----+ 1702 * 1703 * offline -> maintenance 1704 * 1705 * If C goes into maintenance, it's not enough simply to check B. Because A has 1706 * an optional dependency, what was previously an unsatisfiable situation is now 1707 * satisfied (B will never come online, even though its state hasn't changed). 1708 * 1709 * Note that it's not necessary to continue examining dependents after reaching 1710 * an optional_all dependency. It's not possible for an optional_all dependency 1711 * to change satisfiability without also coming online, in which case we get a 1712 * start event and propagation continues naturally. However, it does no harm to 1713 * continue propagating satisfiability (as it is a relatively rare event), and 1714 * keeps the walker code simple and generic. 1715 */ 1716 /*ARGSUSED*/ 1717 static int 1718 satbility_cb(graph_vertex_t *v, void *arg) 1719 { 1720 if (v->gv_type == GVT_INST) 1721 graph_start_if_satisfied(v); 1722 1723 return (UU_WALK_NEXT); 1724 } 1725 1726 static void 1727 propagate_satbility(graph_vertex_t *v) 1728 { 1729 graph_walk(v, WALK_DEPENDENTS, satbility_cb, NULL, NULL); 1730 } 1731 1732 static void propagate_stop(graph_vertex_t *, void *); 1733 1734 /* ARGSUSED */ 1735 static void 1736 propagate_start(graph_vertex_t *v, void *arg) 1737 { 1738 switch (v->gv_type) { 1739 case GVT_INST: 1740 graph_start_if_satisfied(v); 1741 break; 1742 1743 case GVT_GROUP: 1744 if (v->gv_depgroup == DEPGRP_EXCLUDE_ALL) { 1745 graph_walk_dependents(v, propagate_stop, 1746 (void *)RERR_RESTART); 1747 break; 1748 } 1749 /* FALLTHROUGH */ 1750 1751 case GVT_SVC: 1752 graph_walk_dependents(v, propagate_start, NULL); 1753 break; 1754 1755 case GVT_FILE: 1756 #ifndef NDEBUG 1757 uu_warn("%s:%d: propagate_start() encountered GVT_FILE.\n", 1758 __FILE__, __LINE__); 1759 #endif 1760 abort(); 1761 /* NOTREACHED */ 1762 1763 default: 1764 #ifndef NDEBUG 1765 uu_warn("%s:%d: Unknown vertex type %d.\n", __FILE__, __LINE__, 1766 v->gv_type); 1767 #endif 1768 abort(); 1769 } 1770 } 1771 1772 static void 1773 propagate_stop(graph_vertex_t *v, void *arg) 1774 { 1775 graph_edge_t *e; 1776 graph_vertex_t *svc; 1777 restarter_error_t err = (restarter_error_t)arg; 1778 1779 switch (v->gv_type) { 1780 case GVT_INST: 1781 /* Restarter */ 1782 if (err > RERR_NONE && inst_running(v)) { 1783 if (err == RERR_RESTART || err == RERR_REFRESH) { 1784 vertex_send_event(v, 1785 RESTARTER_EVENT_TYPE_STOP_RESET); 1786 } else { 1787 vertex_send_event(v, RESTARTER_EVENT_TYPE_STOP); 1788 } 1789 } 1790 break; 1791 1792 case GVT_SVC: 1793 graph_walk_dependents(v, propagate_stop, arg); 1794 break; 1795 1796 case GVT_FILE: 1797 #ifndef NDEBUG 1798 uu_warn("%s:%d: propagate_stop() encountered GVT_FILE.\n", 1799 __FILE__, __LINE__); 1800 #endif 1801 abort(); 1802 /* NOTREACHED */ 1803 1804 case GVT_GROUP: 1805 if (v->gv_depgroup == DEPGRP_EXCLUDE_ALL) { 1806 graph_walk_dependents(v, propagate_start, NULL); 1807 break; 1808 } 1809 1810 if (err == RERR_NONE || err > v->gv_restart) 1811 break; 1812 1813 assert(uu_list_numnodes(v->gv_dependents) == 1); 1814 e = uu_list_first(v->gv_dependents); 1815 svc = e->ge_vertex; 1816 1817 if (inst_running(svc)) { 1818 if (err == RERR_RESTART || err == RERR_REFRESH) { 1819 vertex_send_event(svc, 1820 RESTARTER_EVENT_TYPE_STOP_RESET); 1821 } else { 1822 vertex_send_event(svc, 1823 RESTARTER_EVENT_TYPE_STOP); 1824 } 1825 } 1826 break; 1827 1828 default: 1829 #ifndef NDEBUG 1830 uu_warn("%s:%d: Unknown vertex type %d.\n", __FILE__, __LINE__, 1831 v->gv_type); 1832 #endif 1833 abort(); 1834 } 1835 } 1836 1837 void 1838 offline_vertex(graph_vertex_t *v) 1839 { 1840 scf_handle_t *h = libscf_handle_create_bound_loop(); 1841 scf_instance_t *scf_inst = safe_scf_instance_create(h); 1842 scf_propertygroup_t *pg = safe_scf_pg_create(h); 1843 restarter_instance_state_t state, next_state; 1844 int r; 1845 1846 assert(v->gv_type == GVT_INST); 1847 1848 if (scf_inst == NULL) 1849 bad_error("safe_scf_instance_create", scf_error()); 1850 if (pg == NULL) 1851 bad_error("safe_scf_pg_create", scf_error()); 1852 1853 /* if the vertex is already going offline, return */ 1854 rep_retry: 1855 if (scf_handle_decode_fmri(h, v->gv_name, NULL, NULL, scf_inst, NULL, 1856 NULL, SCF_DECODE_FMRI_EXACT) != 0) { 1857 switch (scf_error()) { 1858 case SCF_ERROR_CONNECTION_BROKEN: 1859 libscf_handle_rebind(h); 1860 goto rep_retry; 1861 1862 case SCF_ERROR_NOT_FOUND: 1863 scf_pg_destroy(pg); 1864 scf_instance_destroy(scf_inst); 1865 (void) scf_handle_unbind(h); 1866 scf_handle_destroy(h); 1867 return; 1868 } 1869 uu_die("Can't decode FMRI %s: %s\n", v->gv_name, 1870 scf_strerror(scf_error())); 1871 } 1872 1873 r = scf_instance_get_pg(scf_inst, SCF_PG_RESTARTER, pg); 1874 if (r != 0) { 1875 switch (scf_error()) { 1876 case SCF_ERROR_CONNECTION_BROKEN: 1877 libscf_handle_rebind(h); 1878 goto rep_retry; 1879 1880 case SCF_ERROR_NOT_SET: 1881 case SCF_ERROR_NOT_FOUND: 1882 scf_pg_destroy(pg); 1883 scf_instance_destroy(scf_inst); 1884 (void) scf_handle_unbind(h); 1885 scf_handle_destroy(h); 1886 return; 1887 1888 default: 1889 bad_error("scf_instance_get_pg", scf_error()); 1890 } 1891 } else { 1892 r = libscf_read_states(pg, &state, &next_state); 1893 if (r == 0 && (next_state == RESTARTER_STATE_OFFLINE || 1894 next_state == RESTARTER_STATE_DISABLED)) { 1895 log_framework(LOG_DEBUG, 1896 "%s: instance is already going down.\n", 1897 v->gv_name); 1898 scf_pg_destroy(pg); 1899 scf_instance_destroy(scf_inst); 1900 (void) scf_handle_unbind(h); 1901 scf_handle_destroy(h); 1902 return; 1903 } 1904 } 1905 1906 scf_pg_destroy(pg); 1907 scf_instance_destroy(scf_inst); 1908 (void) scf_handle_unbind(h); 1909 scf_handle_destroy(h); 1910 1911 vertex_send_event(v, RESTARTER_EVENT_TYPE_STOP_RESET); 1912 } 1913 1914 /* 1915 * void graph_enable_by_vertex() 1916 * If admin is non-zero, this is an administrative request for change 1917 * of the enabled property. Thus, send the ADMIN_DISABLE rather than 1918 * a plain DISABLE restarter event. 1919 */ 1920 void 1921 graph_enable_by_vertex(graph_vertex_t *vertex, int enable, int admin) 1922 { 1923 graph_vertex_t *v; 1924 int r; 1925 1926 assert(MUTEX_HELD(&dgraph_lock)); 1927 assert((vertex->gv_flags & GV_CONFIGURED)); 1928 1929 vertex->gv_flags = (vertex->gv_flags & ~GV_ENABLED) | 1930 (enable ? GV_ENABLED : 0); 1931 1932 if (enable) { 1933 if (vertex->gv_state != RESTARTER_STATE_OFFLINE && 1934 vertex->gv_state != RESTARTER_STATE_DEGRADED && 1935 vertex->gv_state != RESTARTER_STATE_ONLINE) { 1936 /* 1937 * In case the vertex was notified to go down, 1938 * but now can return online, clear the _TOOFFLINE 1939 * and _TODISABLE flags. 1940 */ 1941 vertex->gv_flags &= ~GV_TOOFFLINE; 1942 vertex->gv_flags &= ~GV_TODISABLE; 1943 1944 vertex_send_event(vertex, RESTARTER_EVENT_TYPE_ENABLE); 1945 } 1946 1947 /* 1948 * Wait for state update from restarter before sending _START or 1949 * _STOP. 1950 */ 1951 1952 return; 1953 } 1954 1955 if (vertex->gv_state == RESTARTER_STATE_DISABLED) 1956 return; 1957 1958 if (!admin) { 1959 vertex_send_event(vertex, RESTARTER_EVENT_TYPE_DISABLE); 1960 1961 /* 1962 * Wait for state update from restarter before sending _START or 1963 * _STOP. 1964 */ 1965 1966 return; 1967 } 1968 1969 /* 1970 * If it is a DISABLE event requested by the administrator then we are 1971 * offlining the dependents first. 1972 */ 1973 1974 /* 1975 * Set GV_TOOFFLINE for the services we are offlining. We cannot 1976 * clear the GV_TOOFFLINE bits from all the services because 1977 * other DISABLE events might be handled at the same time. 1978 */ 1979 vertex->gv_flags |= GV_TOOFFLINE; 1980 1981 /* remember which vertex to disable... */ 1982 vertex->gv_flags |= GV_TODISABLE; 1983 1984 log_framework(LOG_DEBUG, "Marking in-subtree vertices before " 1985 "disabling %s.\n", vertex->gv_name); 1986 1987 /* set GV_TOOFFLINE for its dependents */ 1988 r = uu_list_walk(vertex->gv_dependents, (uu_walk_fn_t *)mark_subtree, 1989 NULL, 0); 1990 assert(r == 0); 1991 1992 /* disable the instance now if there is nothing else to offline */ 1993 if (insubtree_dependents_down(vertex) == B_TRUE) { 1994 vertex_send_event(vertex, RESTARTER_EVENT_TYPE_ADMIN_DISABLE); 1995 return; 1996 } 1997 1998 /* 1999 * This loop is similar to the one used for the graph reversal shutdown 2000 * and could be improved in term of performance for the subtree reversal 2001 * disable case. 2002 */ 2003 for (v = uu_list_first(dgraph); v != NULL; 2004 v = uu_list_next(dgraph, v)) { 2005 /* skip the vertex we are disabling for now */ 2006 if (v == vertex) 2007 continue; 2008 2009 if (v->gv_type != GVT_INST || 2010 (v->gv_flags & GV_CONFIGURED) == 0 || 2011 (v->gv_flags & GV_ENABLED) == 0 || 2012 (v->gv_flags & GV_TOOFFLINE) == 0) 2013 continue; 2014 2015 if ((v->gv_state != RESTARTER_STATE_ONLINE) && 2016 (v->gv_state != RESTARTER_STATE_DEGRADED)) { 2017 /* continue if there is nothing to offline */ 2018 continue; 2019 } 2020 2021 /* 2022 * Instances which are up need to come down before we're 2023 * done, but we can only offline the leaves here. An 2024 * instance is a leaf when all its dependents are down. 2025 */ 2026 if (insubtree_dependents_down(v) == B_TRUE) { 2027 log_framework(LOG_DEBUG, "Offlining in-subtree " 2028 "instance %s for %s.\n", 2029 v->gv_name, vertex->gv_name); 2030 offline_vertex(v); 2031 } 2032 } 2033 } 2034 2035 static int configure_vertex(graph_vertex_t *, scf_instance_t *); 2036 2037 /* 2038 * Set the restarter for v to fmri_arg. That is, make sure a vertex for 2039 * fmri_arg exists, make v depend on it, and send _ADD_INSTANCE for v. If 2040 * v is already configured and fmri_arg indicates the current restarter, do 2041 * nothing. If v is configured and fmri_arg is a new restarter, delete v's 2042 * dependency on the restarter, send _REMOVE_INSTANCE for v, and set the new 2043 * restarter. Returns 0 on success, EINVAL if the FMRI is invalid, 2044 * ECONNABORTED if the repository connection is broken, and ELOOP 2045 * if the dependency would create a cycle. In the last case, *pathp will 2046 * point to a -1-terminated array of ids which compose the path from v to 2047 * restarter_fmri. 2048 */ 2049 int 2050 graph_change_restarter(graph_vertex_t *v, const char *fmri_arg, scf_handle_t *h, 2051 int **pathp) 2052 { 2053 char *restarter_fmri = NULL; 2054 graph_vertex_t *rv; 2055 int err; 2056 int id; 2057 2058 assert(MUTEX_HELD(&dgraph_lock)); 2059 2060 if (fmri_arg[0] != '\0') { 2061 err = fmri_canonify(fmri_arg, &restarter_fmri, B_TRUE); 2062 if (err != 0) { 2063 assert(err == EINVAL); 2064 return (err); 2065 } 2066 } 2067 2068 if (restarter_fmri == NULL || 2069 strcmp(restarter_fmri, SCF_SERVICE_STARTD) == 0) { 2070 if (v->gv_flags & GV_CONFIGURED) { 2071 if (v->gv_restarter_id == -1) { 2072 if (restarter_fmri != NULL) 2073 startd_free(restarter_fmri, 2074 max_scf_fmri_size); 2075 return (0); 2076 } 2077 2078 graph_unset_restarter(v); 2079 } 2080 2081 /* Master restarter, nothing to do. */ 2082 v->gv_restarter_id = -1; 2083 v->gv_restarter_channel = NULL; 2084 vertex_send_event(v, RESTARTER_EVENT_TYPE_ADD_INSTANCE); 2085 return (0); 2086 } 2087 2088 if (v->gv_flags & GV_CONFIGURED) { 2089 id = dict_lookup_byname(restarter_fmri); 2090 if (id != -1 && v->gv_restarter_id == id) { 2091 startd_free(restarter_fmri, max_scf_fmri_size); 2092 return (0); 2093 } 2094 2095 graph_unset_restarter(v); 2096 } 2097 2098 err = graph_insert_vertex_unconfigured(restarter_fmri, GVT_INST, 0, 2099 RERR_NONE, &rv); 2100 startd_free(restarter_fmri, max_scf_fmri_size); 2101 assert(err == 0 || err == EEXIST); 2102 2103 if (rv->gv_delegate_initialized == 0) { 2104 if ((rv->gv_delegate_channel = restarter_protocol_init_delegate( 2105 rv->gv_name)) == NULL) 2106 return (EINVAL); 2107 rv->gv_delegate_initialized = 1; 2108 } 2109 v->gv_restarter_id = rv->gv_id; 2110 v->gv_restarter_channel = rv->gv_delegate_channel; 2111 2112 err = graph_insert_dependency(v, rv, pathp); 2113 if (err != 0) { 2114 assert(err == ELOOP); 2115 return (ELOOP); 2116 } 2117 2118 vertex_send_event(v, RESTARTER_EVENT_TYPE_ADD_INSTANCE); 2119 2120 if (!(rv->gv_flags & GV_CONFIGURED)) { 2121 scf_instance_t *inst; 2122 2123 err = libscf_fmri_get_instance(h, rv->gv_name, &inst); 2124 switch (err) { 2125 case 0: 2126 err = configure_vertex(rv, inst); 2127 scf_instance_destroy(inst); 2128 switch (err) { 2129 case 0: 2130 case ECANCELED: 2131 break; 2132 2133 case ECONNABORTED: 2134 return (ECONNABORTED); 2135 2136 default: 2137 bad_error("configure_vertex", err); 2138 } 2139 break; 2140 2141 case ECONNABORTED: 2142 return (ECONNABORTED); 2143 2144 case ENOENT: 2145 break; 2146 2147 case ENOTSUP: 2148 /* 2149 * The fmri doesn't specify an instance - translate 2150 * to EINVAL. 2151 */ 2152 return (EINVAL); 2153 2154 case EINVAL: 2155 default: 2156 bad_error("libscf_fmri_get_instance", err); 2157 } 2158 } 2159 2160 return (0); 2161 } 2162 2163 2164 /* 2165 * Add all of the instances of the service named by fmri to the graph. 2166 * Returns 2167 * 0 - success 2168 * ENOENT - service indicated by fmri does not exist 2169 * 2170 * In both cases *reboundp will be B_TRUE if the handle was rebound, or B_FALSE 2171 * otherwise. 2172 */ 2173 static int 2174 add_service(const char *fmri, scf_handle_t *h, boolean_t *reboundp) 2175 { 2176 scf_service_t *svc; 2177 scf_instance_t *inst; 2178 scf_iter_t *iter; 2179 char *inst_fmri; 2180 int ret, r; 2181 2182 *reboundp = B_FALSE; 2183 2184 svc = safe_scf_service_create(h); 2185 inst = safe_scf_instance_create(h); 2186 iter = safe_scf_iter_create(h); 2187 inst_fmri = startd_alloc(max_scf_fmri_size); 2188 2189 rebound: 2190 if (scf_handle_decode_fmri(h, fmri, NULL, svc, NULL, NULL, NULL, 2191 SCF_DECODE_FMRI_EXACT) != 0) { 2192 switch (scf_error()) { 2193 case SCF_ERROR_CONNECTION_BROKEN: 2194 default: 2195 libscf_handle_rebind(h); 2196 *reboundp = B_TRUE; 2197 goto rebound; 2198 2199 case SCF_ERROR_NOT_FOUND: 2200 ret = ENOENT; 2201 goto out; 2202 2203 case SCF_ERROR_INVALID_ARGUMENT: 2204 case SCF_ERROR_CONSTRAINT_VIOLATED: 2205 case SCF_ERROR_NOT_BOUND: 2206 case SCF_ERROR_HANDLE_MISMATCH: 2207 bad_error("scf_handle_decode_fmri", scf_error()); 2208 } 2209 } 2210 2211 if (scf_iter_service_instances(iter, svc) != 0) { 2212 switch (scf_error()) { 2213 case SCF_ERROR_CONNECTION_BROKEN: 2214 default: 2215 libscf_handle_rebind(h); 2216 *reboundp = B_TRUE; 2217 goto rebound; 2218 2219 case SCF_ERROR_DELETED: 2220 ret = ENOENT; 2221 goto out; 2222 2223 case SCF_ERROR_HANDLE_MISMATCH: 2224 case SCF_ERROR_NOT_BOUND: 2225 case SCF_ERROR_NOT_SET: 2226 bad_error("scf_iter_service_instances", scf_error()); 2227 } 2228 } 2229 2230 for (;;) { 2231 r = scf_iter_next_instance(iter, inst); 2232 if (r == 0) 2233 break; 2234 if (r != 1) { 2235 switch (scf_error()) { 2236 case SCF_ERROR_CONNECTION_BROKEN: 2237 default: 2238 libscf_handle_rebind(h); 2239 *reboundp = B_TRUE; 2240 goto rebound; 2241 2242 case SCF_ERROR_DELETED: 2243 ret = ENOENT; 2244 goto out; 2245 2246 case SCF_ERROR_HANDLE_MISMATCH: 2247 case SCF_ERROR_NOT_BOUND: 2248 case SCF_ERROR_NOT_SET: 2249 case SCF_ERROR_INVALID_ARGUMENT: 2250 bad_error("scf_iter_next_instance", 2251 scf_error()); 2252 } 2253 } 2254 2255 if (scf_instance_to_fmri(inst, inst_fmri, max_scf_fmri_size) < 2256 0) { 2257 switch (scf_error()) { 2258 case SCF_ERROR_CONNECTION_BROKEN: 2259 libscf_handle_rebind(h); 2260 *reboundp = B_TRUE; 2261 goto rebound; 2262 2263 case SCF_ERROR_DELETED: 2264 continue; 2265 2266 case SCF_ERROR_NOT_BOUND: 2267 case SCF_ERROR_NOT_SET: 2268 bad_error("scf_instance_to_fmri", scf_error()); 2269 } 2270 } 2271 2272 r = dgraph_add_instance(inst_fmri, inst, B_FALSE); 2273 switch (r) { 2274 case 0: 2275 case ECANCELED: 2276 break; 2277 2278 case EEXIST: 2279 continue; 2280 2281 case ECONNABORTED: 2282 libscf_handle_rebind(h); 2283 *reboundp = B_TRUE; 2284 goto rebound; 2285 2286 case EINVAL: 2287 default: 2288 bad_error("dgraph_add_instance", r); 2289 } 2290 } 2291 2292 ret = 0; 2293 2294 out: 2295 startd_free(inst_fmri, max_scf_fmri_size); 2296 scf_iter_destroy(iter); 2297 scf_instance_destroy(inst); 2298 scf_service_destroy(svc); 2299 return (ret); 2300 } 2301 2302 struct depfmri_info { 2303 graph_vertex_t *v; /* GVT_GROUP vertex */ 2304 gv_type_t type; /* type of dependency */ 2305 const char *inst_fmri; /* FMRI of parental GVT_INST vert. */ 2306 const char *pg_name; /* Name of dependency pg */ 2307 scf_handle_t *h; 2308 int err; /* return error code */ 2309 int **pathp; /* return circular dependency path */ 2310 }; 2311 2312 /* 2313 * Find or create a vertex for fmri and make info->v depend on it. 2314 * Returns 2315 * 0 - success 2316 * nonzero - failure 2317 * 2318 * On failure, sets info->err to 2319 * EINVAL - fmri is invalid 2320 * fmri does not match info->type 2321 * ELOOP - Adding the dependency creates a circular dependency. *info->pathp 2322 * will point to an array of the ids of the members of the cycle. 2323 * ECONNABORTED - repository connection was broken 2324 * ECONNRESET - succeeded, but repository connection was reset 2325 */ 2326 static int 2327 process_dependency_fmri(const char *fmri, struct depfmri_info *info) 2328 { 2329 int err; 2330 graph_vertex_t *depgroup_v, *v; 2331 char *fmri_copy, *cfmri; 2332 size_t fmri_copy_sz; 2333 const char *scope, *service, *instance, *pg; 2334 scf_instance_t *inst; 2335 boolean_t rebound; 2336 2337 assert(MUTEX_HELD(&dgraph_lock)); 2338 2339 /* Get or create vertex for FMRI */ 2340 depgroup_v = info->v; 2341 2342 if (strncmp(fmri, "file:", sizeof ("file:") - 1) == 0) { 2343 if (info->type != GVT_FILE) { 2344 log_framework(LOG_NOTICE, 2345 "FMRI \"%s\" is not allowed for the \"%s\" " 2346 "dependency's type of instance %s.\n", fmri, 2347 info->pg_name, info->inst_fmri); 2348 return (info->err = EINVAL); 2349 } 2350 2351 err = graph_insert_vertex_unconfigured(fmri, info->type, 0, 2352 RERR_NONE, &v); 2353 switch (err) { 2354 case 0: 2355 break; 2356 2357 case EEXIST: 2358 assert(v->gv_type == GVT_FILE); 2359 break; 2360 2361 case EINVAL: /* prevented above */ 2362 default: 2363 bad_error("graph_insert_vertex_unconfigured", err); 2364 } 2365 } else { 2366 if (info->type != GVT_INST) { 2367 log_framework(LOG_NOTICE, 2368 "FMRI \"%s\" is not allowed for the \"%s\" " 2369 "dependency's type of instance %s.\n", fmri, 2370 info->pg_name, info->inst_fmri); 2371 return (info->err = EINVAL); 2372 } 2373 2374 /* 2375 * We must canonify fmri & add a vertex for it. 2376 */ 2377 fmri_copy_sz = strlen(fmri) + 1; 2378 fmri_copy = startd_alloc(fmri_copy_sz); 2379 (void) strcpy(fmri_copy, fmri); 2380 2381 /* Determine if the FMRI is a property group or instance */ 2382 if (scf_parse_svc_fmri(fmri_copy, &scope, &service, 2383 &instance, &pg, NULL) != 0) { 2384 startd_free(fmri_copy, fmri_copy_sz); 2385 log_framework(LOG_NOTICE, 2386 "Dependency \"%s\" of %s has invalid FMRI " 2387 "\"%s\".\n", info->pg_name, info->inst_fmri, 2388 fmri); 2389 return (info->err = EINVAL); 2390 } 2391 2392 if (service == NULL || pg != NULL) { 2393 startd_free(fmri_copy, fmri_copy_sz); 2394 log_framework(LOG_NOTICE, 2395 "Dependency \"%s\" of %s does not designate a " 2396 "service or instance.\n", info->pg_name, 2397 info->inst_fmri); 2398 return (info->err = EINVAL); 2399 } 2400 2401 if (scope == NULL || strcmp(scope, SCF_SCOPE_LOCAL) == 0) { 2402 cfmri = uu_msprintf("svc:/%s%s%s", 2403 service, instance ? ":" : "", instance ? instance : 2404 ""); 2405 } else { 2406 cfmri = uu_msprintf("svc://%s/%s%s%s", 2407 scope, service, instance ? ":" : "", instance ? 2408 instance : ""); 2409 } 2410 2411 startd_free(fmri_copy, fmri_copy_sz); 2412 2413 err = graph_insert_vertex_unconfigured(cfmri, instance ? 2414 GVT_INST : GVT_SVC, instance ? 0 : DEPGRP_REQUIRE_ANY, 2415 RERR_NONE, &v); 2416 uu_free(cfmri); 2417 switch (err) { 2418 case 0: 2419 break; 2420 2421 case EEXIST: 2422 /* Verify v. */ 2423 if (instance != NULL) 2424 assert(v->gv_type == GVT_INST); 2425 else 2426 assert(v->gv_type == GVT_SVC); 2427 break; 2428 2429 default: 2430 bad_error("graph_insert_vertex_unconfigured", err); 2431 } 2432 } 2433 2434 /* Add dependency from depgroup_v to new vertex */ 2435 info->err = graph_insert_dependency(depgroup_v, v, info->pathp); 2436 switch (info->err) { 2437 case 0: 2438 break; 2439 2440 case ELOOP: 2441 return (ELOOP); 2442 2443 default: 2444 bad_error("graph_insert_dependency", info->err); 2445 } 2446 2447 /* This must be after we insert the dependency, to avoid looping. */ 2448 switch (v->gv_type) { 2449 case GVT_INST: 2450 if ((v->gv_flags & GV_CONFIGURED) != 0) 2451 break; 2452 2453 inst = safe_scf_instance_create(info->h); 2454 2455 rebound = B_FALSE; 2456 2457 rebound: 2458 err = libscf_lookup_instance(v->gv_name, inst); 2459 switch (err) { 2460 case 0: 2461 err = configure_vertex(v, inst); 2462 switch (err) { 2463 case 0: 2464 case ECANCELED: 2465 break; 2466 2467 case ECONNABORTED: 2468 libscf_handle_rebind(info->h); 2469 rebound = B_TRUE; 2470 goto rebound; 2471 2472 default: 2473 bad_error("configure_vertex", err); 2474 } 2475 break; 2476 2477 case ENOENT: 2478 break; 2479 2480 case ECONNABORTED: 2481 libscf_handle_rebind(info->h); 2482 rebound = B_TRUE; 2483 goto rebound; 2484 2485 case EINVAL: 2486 case ENOTSUP: 2487 default: 2488 bad_error("libscf_fmri_get_instance", err); 2489 } 2490 2491 scf_instance_destroy(inst); 2492 2493 if (rebound) 2494 return (info->err = ECONNRESET); 2495 break; 2496 2497 case GVT_SVC: 2498 (void) add_service(v->gv_name, info->h, &rebound); 2499 if (rebound) 2500 return (info->err = ECONNRESET); 2501 } 2502 2503 return (0); 2504 } 2505 2506 struct deppg_info { 2507 graph_vertex_t *v; /* GVT_INST vertex */ 2508 int err; /* return error */ 2509 int **pathp; /* return circular dependency path */ 2510 }; 2511 2512 /* 2513 * Make info->v depend on a new GVT_GROUP node for this property group, 2514 * and then call process_dependency_fmri() for the values of the entity 2515 * property. Return 0 on success, or if something goes wrong return nonzero 2516 * and set info->err to ECONNABORTED, EINVAL, or the error code returned by 2517 * process_dependency_fmri(). 2518 */ 2519 static int 2520 process_dependency_pg(scf_propertygroup_t *pg, struct deppg_info *info) 2521 { 2522 scf_handle_t *h; 2523 depgroup_type_t deptype; 2524 restarter_error_t rerr; 2525 struct depfmri_info linfo; 2526 char *fmri, *pg_name; 2527 size_t fmri_sz; 2528 graph_vertex_t *depgrp; 2529 scf_property_t *prop; 2530 int err; 2531 int empty; 2532 scf_error_t scferr; 2533 ssize_t len; 2534 2535 assert(MUTEX_HELD(&dgraph_lock)); 2536 2537 h = scf_pg_handle(pg); 2538 2539 pg_name = startd_alloc(max_scf_name_size); 2540 2541 len = scf_pg_get_name(pg, pg_name, max_scf_name_size); 2542 if (len < 0) { 2543 startd_free(pg_name, max_scf_name_size); 2544 switch (scf_error()) { 2545 case SCF_ERROR_CONNECTION_BROKEN: 2546 default: 2547 return (info->err = ECONNABORTED); 2548 2549 case SCF_ERROR_DELETED: 2550 return (info->err = 0); 2551 2552 case SCF_ERROR_NOT_SET: 2553 bad_error("scf_pg_get_name", scf_error()); 2554 } 2555 } 2556 2557 /* 2558 * Skip over empty dependency groups. Since dependency property 2559 * groups are updated atomically, they are either empty or 2560 * fully populated. 2561 */ 2562 empty = depgroup_empty(h, pg); 2563 if (empty < 0) { 2564 log_error(LOG_INFO, 2565 "Error reading dependency group \"%s\" of %s: %s\n", 2566 pg_name, info->v->gv_name, scf_strerror(scf_error())); 2567 startd_free(pg_name, max_scf_name_size); 2568 return (info->err = EINVAL); 2569 2570 } else if (empty == 1) { 2571 log_framework(LOG_DEBUG, 2572 "Ignoring empty dependency group \"%s\" of %s\n", 2573 pg_name, info->v->gv_name); 2574 startd_free(pg_name, max_scf_name_size); 2575 return (info->err = 0); 2576 } 2577 2578 fmri_sz = strlen(info->v->gv_name) + 1 + len + 1; 2579 fmri = startd_alloc(fmri_sz); 2580 2581 (void) snprintf(fmri, fmri_sz, "%s>%s", info->v->gv_name, 2582 pg_name); 2583 2584 /* Validate the pg before modifying the graph */ 2585 deptype = depgroup_read_grouping(h, pg); 2586 if (deptype == DEPGRP_UNSUPPORTED) { 2587 log_error(LOG_INFO, 2588 "Dependency \"%s\" of %s has an unknown grouping value.\n", 2589 pg_name, info->v->gv_name); 2590 startd_free(fmri, fmri_sz); 2591 startd_free(pg_name, max_scf_name_size); 2592 return (info->err = EINVAL); 2593 } 2594 2595 rerr = depgroup_read_restart(h, pg); 2596 if (rerr == RERR_UNSUPPORTED) { 2597 log_error(LOG_INFO, 2598 "Dependency \"%s\" of %s has an unknown restart_on value." 2599 "\n", 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 prop = safe_scf_property_create(h); 2606 2607 if (scf_pg_get_property(pg, SCF_PROPERTY_ENTITIES, prop) != 0) { 2608 scferr = scf_error(); 2609 scf_property_destroy(prop); 2610 if (scferr == SCF_ERROR_DELETED) { 2611 startd_free(fmri, fmri_sz); 2612 startd_free(pg_name, max_scf_name_size); 2613 return (info->err = 0); 2614 } else if (scferr != SCF_ERROR_NOT_FOUND) { 2615 startd_free(fmri, fmri_sz); 2616 startd_free(pg_name, max_scf_name_size); 2617 return (info->err = ECONNABORTED); 2618 } 2619 2620 log_error(LOG_INFO, 2621 "Dependency \"%s\" of %s is missing a \"%s\" property.\n", 2622 pg_name, info->v->gv_name, SCF_PROPERTY_ENTITIES); 2623 2624 startd_free(fmri, fmri_sz); 2625 startd_free(pg_name, max_scf_name_size); 2626 2627 return (info->err = EINVAL); 2628 } 2629 2630 /* Create depgroup vertex for pg */ 2631 err = graph_insert_vertex_unconfigured(fmri, GVT_GROUP, deptype, 2632 rerr, &depgrp); 2633 assert(err == 0); 2634 startd_free(fmri, fmri_sz); 2635 2636 /* Add dependency from inst vertex to new vertex */ 2637 err = graph_insert_dependency(info->v, depgrp, info->pathp); 2638 /* ELOOP can't happen because this should be a new vertex */ 2639 assert(err == 0); 2640 2641 linfo.v = depgrp; 2642 linfo.type = depgroup_read_scheme(h, pg); 2643 linfo.inst_fmri = info->v->gv_name; 2644 linfo.pg_name = pg_name; 2645 linfo.h = h; 2646 linfo.err = 0; 2647 linfo.pathp = info->pathp; 2648 err = walk_property_astrings(prop, (callback_t)process_dependency_fmri, 2649 &linfo); 2650 2651 scf_property_destroy(prop); 2652 startd_free(pg_name, max_scf_name_size); 2653 2654 switch (err) { 2655 case 0: 2656 case EINTR: 2657 return (info->err = linfo.err); 2658 2659 case ECONNABORTED: 2660 case EINVAL: 2661 return (info->err = err); 2662 2663 case ECANCELED: 2664 return (info->err = 0); 2665 2666 case ECONNRESET: 2667 return (info->err = ECONNABORTED); 2668 2669 default: 2670 bad_error("walk_property_astrings", err); 2671 /* NOTREACHED */ 2672 } 2673 } 2674 2675 /* 2676 * Build the dependency info for v from the repository. Returns 0 on success, 2677 * ECONNABORTED on repository disconnection, EINVAL if the repository 2678 * configuration is invalid, and ELOOP if a dependency would cause a cycle. 2679 * In the last case, *pathp will point to a -1-terminated array of ids which 2680 * constitute the rest of the dependency cycle. 2681 */ 2682 static int 2683 set_dependencies(graph_vertex_t *v, scf_instance_t *inst, int **pathp) 2684 { 2685 struct deppg_info info; 2686 int err; 2687 uint_t old_configured; 2688 2689 assert(MUTEX_HELD(&dgraph_lock)); 2690 2691 /* 2692 * Mark the vertex as configured during dependency insertion to avoid 2693 * dependency cycles (which can appear in the graph if one of the 2694 * vertices is an exclusion-group). 2695 */ 2696 old_configured = v->gv_flags & GV_CONFIGURED; 2697 v->gv_flags |= GV_CONFIGURED; 2698 2699 info.err = 0; 2700 info.v = v; 2701 info.pathp = pathp; 2702 2703 err = walk_dependency_pgs(inst, (callback_t)process_dependency_pg, 2704 &info); 2705 2706 if (!old_configured) 2707 v->gv_flags &= ~GV_CONFIGURED; 2708 2709 switch (err) { 2710 case 0: 2711 case EINTR: 2712 return (info.err); 2713 2714 case ECONNABORTED: 2715 return (ECONNABORTED); 2716 2717 case ECANCELED: 2718 /* Should get delete event, so return 0. */ 2719 return (0); 2720 2721 default: 2722 bad_error("walk_dependency_pgs", err); 2723 /* NOTREACHED */ 2724 } 2725 } 2726 2727 2728 static void 2729 handle_cycle(const char *fmri, int *path) 2730 { 2731 const char *cp; 2732 size_t sz; 2733 2734 assert(MUTEX_HELD(&dgraph_lock)); 2735 2736 path_to_str(path, (char **)&cp, &sz); 2737 2738 log_error(LOG_ERR, "Transitioning %s to maintenance " 2739 "because it completes a dependency cycle (see svcs -xv for " 2740 "details):\n%s", fmri ? fmri : "?", cp); 2741 2742 startd_free((void *)cp, sz); 2743 } 2744 2745 /* 2746 * Increment the vertex's reference count to prevent the vertex removal 2747 * from the dgraph. 2748 */ 2749 static void 2750 vertex_ref(graph_vertex_t *v) 2751 { 2752 assert(MUTEX_HELD(&dgraph_lock)); 2753 2754 v->gv_refs++; 2755 } 2756 2757 /* 2758 * Decrement the vertex's reference count and remove the vertex from 2759 * the dgraph when possible. 2760 * 2761 * Return VERTEX_REMOVED when the vertex has been removed otherwise 2762 * return VERTEX_INUSE. 2763 */ 2764 static int 2765 vertex_unref(graph_vertex_t *v) 2766 { 2767 assert(MUTEX_HELD(&dgraph_lock)); 2768 assert(v->gv_refs > 0); 2769 2770 v->gv_refs--; 2771 2772 return (free_if_unrefed(v)); 2773 } 2774 2775 /* 2776 * When run on the dependencies of a vertex, populates list with 2777 * graph_edge_t's which point to the service vertices or the instance 2778 * vertices (no GVT_GROUP nodes) on which the vertex depends. 2779 * 2780 * Increment the vertex's reference count once the vertex is inserted 2781 * in the list. The vertex won't be able to be deleted from the dgraph 2782 * while it is referenced. 2783 */ 2784 static int 2785 append_svcs_or_insts(graph_edge_t *e, uu_list_t *list) 2786 { 2787 graph_vertex_t *v = e->ge_vertex; 2788 graph_edge_t *new; 2789 int r; 2790 2791 switch (v->gv_type) { 2792 case GVT_INST: 2793 case GVT_SVC: 2794 break; 2795 2796 case GVT_GROUP: 2797 r = uu_list_walk(v->gv_dependencies, 2798 (uu_walk_fn_t *)append_svcs_or_insts, list, 0); 2799 assert(r == 0); 2800 return (UU_WALK_NEXT); 2801 2802 case GVT_FILE: 2803 return (UU_WALK_NEXT); 2804 2805 default: 2806 #ifndef NDEBUG 2807 uu_warn("%s:%d: Unexpected vertex type %d.\n", __FILE__, 2808 __LINE__, v->gv_type); 2809 #endif 2810 abort(); 2811 } 2812 2813 new = startd_alloc(sizeof (*new)); 2814 new->ge_vertex = v; 2815 uu_list_node_init(new, &new->ge_link, graph_edge_pool); 2816 r = uu_list_insert_before(list, NULL, new); 2817 assert(r == 0); 2818 2819 /* 2820 * Because we are inserting the vertex in a list, we don't want 2821 * the vertex to be freed while the list is in use. In order to 2822 * achieve that, increment the vertex's reference count. 2823 */ 2824 vertex_ref(v); 2825 2826 return (UU_WALK_NEXT); 2827 } 2828 2829 static boolean_t 2830 should_be_in_subgraph(graph_vertex_t *v) 2831 { 2832 graph_edge_t *e; 2833 2834 if (v == milestone) 2835 return (B_TRUE); 2836 2837 /* 2838 * v is in the subgraph if any of its dependents are in the subgraph. 2839 * Except for EXCLUDE_ALL dependents. And OPTIONAL dependents only 2840 * count if we're enabled. 2841 */ 2842 for (e = uu_list_first(v->gv_dependents); 2843 e != NULL; 2844 e = uu_list_next(v->gv_dependents, e)) { 2845 graph_vertex_t *dv = e->ge_vertex; 2846 2847 if (!(dv->gv_flags & GV_INSUBGRAPH)) 2848 continue; 2849 2850 /* 2851 * Don't include instances that are optional and disabled. 2852 */ 2853 if (v->gv_type == GVT_INST && dv->gv_type == GVT_SVC) { 2854 2855 int in = 0; 2856 graph_edge_t *ee; 2857 2858 for (ee = uu_list_first(dv->gv_dependents); 2859 ee != NULL; 2860 ee = uu_list_next(dv->gv_dependents, ee)) { 2861 2862 graph_vertex_t *ddv = e->ge_vertex; 2863 2864 if (ddv->gv_type == GVT_GROUP && 2865 ddv->gv_depgroup == DEPGRP_EXCLUDE_ALL) 2866 continue; 2867 2868 if (ddv->gv_type == GVT_GROUP && 2869 ddv->gv_depgroup == DEPGRP_OPTIONAL_ALL && 2870 !(v->gv_flags & GV_ENBLD_NOOVR)) 2871 continue; 2872 2873 in = 1; 2874 } 2875 if (!in) 2876 continue; 2877 } 2878 if (v->gv_type == GVT_INST && 2879 dv->gv_type == GVT_GROUP && 2880 dv->gv_depgroup == DEPGRP_OPTIONAL_ALL && 2881 !(v->gv_flags & GV_ENBLD_NOOVR)) 2882 continue; 2883 2884 /* Don't include excluded services and instances */ 2885 if (dv->gv_type == GVT_GROUP && 2886 dv->gv_depgroup == DEPGRP_EXCLUDE_ALL) 2887 continue; 2888 2889 return (B_TRUE); 2890 } 2891 2892 return (B_FALSE); 2893 } 2894 2895 /* 2896 * Ensures that GV_INSUBGRAPH is set properly for v and its descendents. If 2897 * any bits change, manipulate the repository appropriately. Returns 0 or 2898 * ECONNABORTED. 2899 */ 2900 static int 2901 eval_subgraph(graph_vertex_t *v, scf_handle_t *h) 2902 { 2903 boolean_t old = (v->gv_flags & GV_INSUBGRAPH) != 0; 2904 boolean_t new; 2905 graph_edge_t *e; 2906 scf_instance_t *inst; 2907 int ret = 0, r; 2908 2909 assert(milestone != NULL && milestone != MILESTONE_NONE); 2910 2911 new = should_be_in_subgraph(v); 2912 2913 if (new == old) 2914 return (0); 2915 2916 log_framework(LOG_DEBUG, new ? "Adding %s to the subgraph.\n" : 2917 "Removing %s from the subgraph.\n", v->gv_name); 2918 2919 v->gv_flags = (v->gv_flags & ~GV_INSUBGRAPH) | 2920 (new ? GV_INSUBGRAPH : 0); 2921 2922 if (v->gv_type == GVT_INST && (v->gv_flags & GV_CONFIGURED)) { 2923 int err; 2924 2925 get_inst: 2926 err = libscf_fmri_get_instance(h, v->gv_name, &inst); 2927 if (err != 0) { 2928 switch (err) { 2929 case ECONNABORTED: 2930 libscf_handle_rebind(h); 2931 ret = ECONNABORTED; 2932 goto get_inst; 2933 2934 case ENOENT: 2935 break; 2936 2937 case EINVAL: 2938 case ENOTSUP: 2939 default: 2940 bad_error("libscf_fmri_get_instance", err); 2941 } 2942 } else { 2943 const char *f; 2944 2945 if (new) { 2946 err = libscf_delete_enable_ovr(inst); 2947 f = "libscf_delete_enable_ovr"; 2948 } else { 2949 err = libscf_set_enable_ovr(inst, 0); 2950 f = "libscf_set_enable_ovr"; 2951 } 2952 scf_instance_destroy(inst); 2953 switch (err) { 2954 case 0: 2955 case ECANCELED: 2956 break; 2957 2958 case ECONNABORTED: 2959 libscf_handle_rebind(h); 2960 /* 2961 * We must continue so the graph is updated, 2962 * but we must return ECONNABORTED so any 2963 * libscf state held by any callers is reset. 2964 */ 2965 ret = ECONNABORTED; 2966 goto get_inst; 2967 2968 case EROFS: 2969 case EPERM: 2970 log_error(LOG_WARNING, 2971 "Could not set %s/%s for %s: %s.\n", 2972 SCF_PG_GENERAL_OVR, SCF_PROPERTY_ENABLED, 2973 v->gv_name, strerror(err)); 2974 break; 2975 2976 default: 2977 bad_error(f, err); 2978 } 2979 } 2980 } 2981 2982 for (e = uu_list_first(v->gv_dependencies); 2983 e != NULL; 2984 e = uu_list_next(v->gv_dependencies, e)) { 2985 r = eval_subgraph(e->ge_vertex, h); 2986 if (r != 0) { 2987 assert(r == ECONNABORTED); 2988 ret = ECONNABORTED; 2989 } 2990 } 2991 2992 return (ret); 2993 } 2994 2995 /* 2996 * Delete the (property group) dependencies of v & create new ones based on 2997 * inst. If doing so would create a cycle, log a message and put the instance 2998 * into maintenance. Update GV_INSUBGRAPH flags as necessary. Returns 0 or 2999 * ECONNABORTED. 3000 */ 3001 int 3002 refresh_vertex(graph_vertex_t *v, scf_instance_t *inst) 3003 { 3004 int err; 3005 int *path; 3006 char *fmri; 3007 int r; 3008 scf_handle_t *h = scf_instance_handle(inst); 3009 uu_list_t *old_deps; 3010 int ret = 0; 3011 graph_edge_t *e; 3012 graph_vertex_t *vv; 3013 3014 assert(MUTEX_HELD(&dgraph_lock)); 3015 assert(v->gv_type == GVT_INST); 3016 3017 log_framework(LOG_DEBUG, "Graph engine: Refreshing %s.\n", v->gv_name); 3018 3019 if (milestone > MILESTONE_NONE) { 3020 /* 3021 * In case some of v's dependencies are being deleted we must 3022 * make a list of them now for GV_INSUBGRAPH-flag evaluation 3023 * after the new dependencies are in place. 3024 */ 3025 old_deps = startd_list_create(graph_edge_pool, NULL, 0); 3026 3027 err = uu_list_walk(v->gv_dependencies, 3028 (uu_walk_fn_t *)append_svcs_or_insts, old_deps, 0); 3029 assert(err == 0); 3030 } 3031 3032 delete_instance_dependencies(v, B_FALSE); 3033 3034 err = set_dependencies(v, inst, &path); 3035 switch (err) { 3036 case 0: 3037 break; 3038 3039 case ECONNABORTED: 3040 ret = err; 3041 goto out; 3042 3043 case EINVAL: 3044 case ELOOP: 3045 r = libscf_instance_get_fmri(inst, &fmri); 3046 switch (r) { 3047 case 0: 3048 break; 3049 3050 case ECONNABORTED: 3051 ret = ECONNABORTED; 3052 goto out; 3053 3054 case ECANCELED: 3055 ret = 0; 3056 goto out; 3057 3058 default: 3059 bad_error("libscf_instance_get_fmri", r); 3060 } 3061 3062 if (err == EINVAL) { 3063 log_error(LOG_ERR, "Transitioning %s " 3064 "to maintenance due to misconfiguration.\n", 3065 fmri ? fmri : "?"); 3066 vertex_send_event(v, 3067 RESTARTER_EVENT_TYPE_INVALID_DEPENDENCY); 3068 } else { 3069 handle_cycle(fmri, path); 3070 vertex_send_event(v, 3071 RESTARTER_EVENT_TYPE_DEPENDENCY_CYCLE); 3072 } 3073 startd_free(fmri, max_scf_fmri_size); 3074 ret = 0; 3075 goto out; 3076 3077 default: 3078 bad_error("set_dependencies", err); 3079 } 3080 3081 if (milestone > MILESTONE_NONE) { 3082 boolean_t aborted = B_FALSE; 3083 3084 for (e = uu_list_first(old_deps); 3085 e != NULL; 3086 e = uu_list_next(old_deps, e)) { 3087 vv = e->ge_vertex; 3088 3089 if (vertex_unref(vv) == VERTEX_INUSE && 3090 eval_subgraph(vv, h) == ECONNABORTED) 3091 aborted = B_TRUE; 3092 } 3093 3094 for (e = uu_list_first(v->gv_dependencies); 3095 e != NULL; 3096 e = uu_list_next(v->gv_dependencies, e)) { 3097 if (eval_subgraph(e->ge_vertex, h) == 3098 ECONNABORTED) 3099 aborted = B_TRUE; 3100 } 3101 3102 if (aborted) { 3103 ret = ECONNABORTED; 3104 goto out; 3105 } 3106 } 3107 3108 graph_start_if_satisfied(v); 3109 3110 ret = 0; 3111 3112 out: 3113 if (milestone > MILESTONE_NONE) { 3114 void *cookie = NULL; 3115 3116 while ((e = uu_list_teardown(old_deps, &cookie)) != NULL) 3117 startd_free(e, sizeof (*e)); 3118 3119 uu_list_destroy(old_deps); 3120 } 3121 3122 return (ret); 3123 } 3124 3125 /* 3126 * Set up v according to inst. That is, make sure it depends on its 3127 * restarter and set up its dependencies. Send the ADD_INSTANCE command to 3128 * the restarter, and send ENABLE or DISABLE as appropriate. 3129 * 3130 * Returns 0 on success, ECONNABORTED on repository disconnection, or 3131 * ECANCELED if inst is deleted. 3132 */ 3133 static int 3134 configure_vertex(graph_vertex_t *v, scf_instance_t *inst) 3135 { 3136 scf_handle_t *h; 3137 scf_propertygroup_t *pg; 3138 scf_snapshot_t *snap; 3139 char *restarter_fmri = startd_alloc(max_scf_value_size); 3140 int enabled, enabled_ovr; 3141 int err; 3142 int *path; 3143 int deathrow; 3144 int32_t tset; 3145 3146 restarter_fmri[0] = '\0'; 3147 3148 assert(MUTEX_HELD(&dgraph_lock)); 3149 assert(v->gv_type == GVT_INST); 3150 assert((v->gv_flags & GV_CONFIGURED) == 0); 3151 3152 /* GV_INSUBGRAPH should already be set properly. */ 3153 assert(should_be_in_subgraph(v) == 3154 ((v->gv_flags & GV_INSUBGRAPH) != 0)); 3155 3156 /* 3157 * If the instance fmri is in the deathrow list then set the 3158 * GV_DEATHROW flag on the vertex and create and set to true the 3159 * SCF_PROPERTY_DEATHROW boolean property in the non-persistent 3160 * repository for this instance fmri. 3161 */ 3162 if ((v->gv_flags & GV_DEATHROW) || 3163 (is_fmri_in_deathrow(v->gv_name) == B_TRUE)) { 3164 if ((v->gv_flags & GV_DEATHROW) == 0) { 3165 /* 3166 * Set flag GV_DEATHROW, create and set to true 3167 * the SCF_PROPERTY_DEATHROW property in the 3168 * non-persistent repository for this instance fmri. 3169 */ 3170 v->gv_flags |= GV_DEATHROW; 3171 3172 switch (err = libscf_set_deathrow(inst, 1)) { 3173 case 0: 3174 break; 3175 3176 case ECONNABORTED: 3177 case ECANCELED: 3178 startd_free(restarter_fmri, max_scf_value_size); 3179 return (err); 3180 3181 case EROFS: 3182 log_error(LOG_WARNING, "Could not set %s/%s " 3183 "for deathrow %s: %s.\n", 3184 SCF_PG_DEATHROW, SCF_PROPERTY_DEATHROW, 3185 v->gv_name, strerror(err)); 3186 break; 3187 3188 case EPERM: 3189 uu_die("Permission denied.\n"); 3190 /* NOTREACHED */ 3191 3192 default: 3193 bad_error("libscf_set_deathrow", err); 3194 } 3195 log_framework(LOG_DEBUG, "Deathrow, graph set %s.\n", 3196 v->gv_name); 3197 } 3198 startd_free(restarter_fmri, max_scf_value_size); 3199 return (0); 3200 } 3201 3202 h = scf_instance_handle(inst); 3203 3204 /* 3205 * Using a temporary deathrow boolean property, set through 3206 * libscf_set_deathrow(), only for fmris on deathrow, is necessary 3207 * because deathrow_fini() may already have been called, and in case 3208 * of a refresh, GV_DEATHROW may need to be set again. 3209 * libscf_get_deathrow() sets deathrow to 1 only if this instance 3210 * has a temporary boolean property named 'deathrow' valued true 3211 * in a property group 'deathrow', -1 or 0 in all other cases. 3212 */ 3213 err = libscf_get_deathrow(h, inst, &deathrow); 3214 switch (err) { 3215 case 0: 3216 break; 3217 3218 case ECONNABORTED: 3219 case ECANCELED: 3220 startd_free(restarter_fmri, max_scf_value_size); 3221 return (err); 3222 3223 default: 3224 bad_error("libscf_get_deathrow", err); 3225 } 3226 3227 if (deathrow == 1) { 3228 v->gv_flags |= GV_DEATHROW; 3229 startd_free(restarter_fmri, max_scf_value_size); 3230 return (0); 3231 } 3232 3233 log_framework(LOG_DEBUG, "Graph adding %s.\n", v->gv_name); 3234 3235 /* 3236 * If the instance does not have a restarter property group, 3237 * initialize its state to uninitialized/none, in case the restarter 3238 * is not enabled. 3239 */ 3240 pg = safe_scf_pg_create(h); 3241 3242 if (scf_instance_get_pg(inst, SCF_PG_RESTARTER, pg) != 0) { 3243 instance_data_t idata; 3244 uint_t count = 0, msecs = ALLOC_DELAY; 3245 3246 switch (scf_error()) { 3247 case SCF_ERROR_NOT_FOUND: 3248 break; 3249 3250 case SCF_ERROR_CONNECTION_BROKEN: 3251 default: 3252 scf_pg_destroy(pg); 3253 startd_free(restarter_fmri, max_scf_value_size); 3254 return (ECONNABORTED); 3255 3256 case SCF_ERROR_DELETED: 3257 scf_pg_destroy(pg); 3258 startd_free(restarter_fmri, max_scf_value_size); 3259 return (ECANCELED); 3260 3261 case SCF_ERROR_NOT_SET: 3262 bad_error("scf_instance_get_pg", scf_error()); 3263 } 3264 3265 switch (err = libscf_instance_get_fmri(inst, 3266 (char **)&idata.i_fmri)) { 3267 case 0: 3268 break; 3269 3270 case ECONNABORTED: 3271 case ECANCELED: 3272 scf_pg_destroy(pg); 3273 startd_free(restarter_fmri, max_scf_value_size); 3274 return (err); 3275 3276 default: 3277 bad_error("libscf_instance_get_fmri", err); 3278 } 3279 3280 idata.i_state = RESTARTER_STATE_NONE; 3281 idata.i_next_state = RESTARTER_STATE_NONE; 3282 3283 init_state: 3284 switch (err = _restarter_commit_states(h, &idata, 3285 RESTARTER_STATE_UNINIT, RESTARTER_STATE_NONE, 3286 restarter_get_str_short(restarter_str_insert_in_graph))) { 3287 case 0: 3288 break; 3289 3290 case ENOMEM: 3291 ++count; 3292 if (count < ALLOC_RETRY) { 3293 (void) poll(NULL, 0, msecs); 3294 msecs *= ALLOC_DELAY_MULT; 3295 goto init_state; 3296 } 3297 3298 uu_die("Insufficient memory.\n"); 3299 /* NOTREACHED */ 3300 3301 case ECONNABORTED: 3302 startd_free((void *)idata.i_fmri, max_scf_fmri_size); 3303 scf_pg_destroy(pg); 3304 startd_free(restarter_fmri, max_scf_value_size); 3305 return (ECONNABORTED); 3306 3307 case ENOENT: 3308 startd_free((void *)idata.i_fmri, max_scf_fmri_size); 3309 scf_pg_destroy(pg); 3310 startd_free(restarter_fmri, max_scf_value_size); 3311 return (ECANCELED); 3312 3313 case EPERM: 3314 case EACCES: 3315 case EROFS: 3316 log_error(LOG_NOTICE, "Could not initialize state for " 3317 "%s: %s.\n", idata.i_fmri, strerror(err)); 3318 break; 3319 3320 case EINVAL: 3321 default: 3322 bad_error("_restarter_commit_states", err); 3323 } 3324 3325 startd_free((void *)idata.i_fmri, max_scf_fmri_size); 3326 } 3327 3328 scf_pg_destroy(pg); 3329 3330 if (milestone != NULL) { 3331 /* 3332 * Make sure the enable-override is set properly before we 3333 * read whether we should be enabled. 3334 */ 3335 if (milestone == MILESTONE_NONE || 3336 !(v->gv_flags & GV_INSUBGRAPH)) { 3337 /* 3338 * This might seem unjustified after the milestone 3339 * transition has completed (non_subgraph_svcs == 0), 3340 * but it's important because when we boot to 3341 * a milestone, we set the milestone before populating 3342 * the graph, and all of the new non-subgraph services 3343 * need to be disabled here. 3344 */ 3345 switch (err = libscf_set_enable_ovr(inst, 0)) { 3346 case 0: 3347 break; 3348 3349 case ECONNABORTED: 3350 case ECANCELED: 3351 startd_free(restarter_fmri, max_scf_value_size); 3352 return (err); 3353 3354 case EROFS: 3355 log_error(LOG_WARNING, 3356 "Could not set %s/%s for %s: %s.\n", 3357 SCF_PG_GENERAL_OVR, SCF_PROPERTY_ENABLED, 3358 v->gv_name, strerror(err)); 3359 break; 3360 3361 case EPERM: 3362 uu_die("Permission denied.\n"); 3363 /* NOTREACHED */ 3364 3365 default: 3366 bad_error("libscf_set_enable_ovr", err); 3367 } 3368 } else { 3369 assert(v->gv_flags & GV_INSUBGRAPH); 3370 switch (err = libscf_delete_enable_ovr(inst)) { 3371 case 0: 3372 break; 3373 3374 case ECONNABORTED: 3375 case ECANCELED: 3376 startd_free(restarter_fmri, max_scf_value_size); 3377 return (err); 3378 3379 case EPERM: 3380 uu_die("Permission denied.\n"); 3381 /* NOTREACHED */ 3382 3383 default: 3384 bad_error("libscf_delete_enable_ovr", err); 3385 } 3386 } 3387 } 3388 3389 err = libscf_get_basic_instance_data(h, inst, v->gv_name, &enabled, 3390 &enabled_ovr, &restarter_fmri); 3391 switch (err) { 3392 case 0: 3393 break; 3394 3395 case ECONNABORTED: 3396 case ECANCELED: 3397 startd_free(restarter_fmri, max_scf_value_size); 3398 return (err); 3399 3400 case ENOENT: 3401 log_framework(LOG_DEBUG, 3402 "Ignoring %s because it has no general property group.\n", 3403 v->gv_name); 3404 startd_free(restarter_fmri, max_scf_value_size); 3405 return (0); 3406 3407 default: 3408 bad_error("libscf_get_basic_instance_data", err); 3409 } 3410 3411 if ((tset = libscf_get_stn_tset(inst)) == -1) { 3412 log_framework(LOG_WARNING, 3413 "Failed to get notification parameters for %s: %s\n", 3414 v->gv_name, scf_strerror(scf_error())); 3415 v->gv_stn_tset = 0; 3416 } else { 3417 v->gv_stn_tset = tset; 3418 } 3419 if (strcmp(v->gv_name, SCF_INSTANCE_GLOBAL) == 0) 3420 stn_global = v->gv_stn_tset; 3421 3422 if (enabled == -1) { 3423 startd_free(restarter_fmri, max_scf_value_size); 3424 return (0); 3425 } 3426 3427 v->gv_flags = (v->gv_flags & ~GV_ENBLD_NOOVR) | 3428 (enabled ? GV_ENBLD_NOOVR : 0); 3429 3430 if (enabled_ovr != -1) 3431 enabled = enabled_ovr; 3432 3433 v->gv_state = RESTARTER_STATE_UNINIT; 3434 3435 snap = libscf_get_or_make_running_snapshot(inst, v->gv_name, B_TRUE); 3436 scf_snapshot_destroy(snap); 3437 3438 /* Set up the restarter. (Sends _ADD_INSTANCE on success.) */ 3439 err = graph_change_restarter(v, restarter_fmri, h, &path); 3440 if (err != 0) { 3441 instance_data_t idata; 3442 uint_t count = 0, msecs = ALLOC_DELAY; 3443 restarter_str_t reason; 3444 3445 if (err == ECONNABORTED) { 3446 startd_free(restarter_fmri, max_scf_value_size); 3447 return (err); 3448 } 3449 3450 assert(err == EINVAL || err == ELOOP); 3451 3452 if (err == EINVAL) { 3453 log_framework(LOG_ERR, emsg_invalid_restarter, 3454 v->gv_name, restarter_fmri); 3455 reason = restarter_str_invalid_restarter; 3456 } else { 3457 handle_cycle(v->gv_name, path); 3458 reason = restarter_str_dependency_cycle; 3459 } 3460 3461 startd_free(restarter_fmri, max_scf_value_size); 3462 3463 /* 3464 * We didn't register the instance with the restarter, so we 3465 * must set maintenance mode ourselves. 3466 */ 3467 err = libscf_instance_get_fmri(inst, (char **)&idata.i_fmri); 3468 if (err != 0) { 3469 assert(err == ECONNABORTED || err == ECANCELED); 3470 return (err); 3471 } 3472 3473 idata.i_state = RESTARTER_STATE_NONE; 3474 idata.i_next_state = RESTARTER_STATE_NONE; 3475 3476 set_maint: 3477 switch (err = _restarter_commit_states(h, &idata, 3478 RESTARTER_STATE_MAINT, RESTARTER_STATE_NONE, 3479 restarter_get_str_short(reason))) { 3480 case 0: 3481 break; 3482 3483 case ENOMEM: 3484 ++count; 3485 if (count < ALLOC_RETRY) { 3486 (void) poll(NULL, 0, msecs); 3487 msecs *= ALLOC_DELAY_MULT; 3488 goto set_maint; 3489 } 3490 3491 uu_die("Insufficient memory.\n"); 3492 /* NOTREACHED */ 3493 3494 case ECONNABORTED: 3495 startd_free((void *)idata.i_fmri, max_scf_fmri_size); 3496 return (ECONNABORTED); 3497 3498 case ENOENT: 3499 startd_free((void *)idata.i_fmri, max_scf_fmri_size); 3500 return (ECANCELED); 3501 3502 case EPERM: 3503 case EACCES: 3504 case EROFS: 3505 log_error(LOG_NOTICE, "Could not initialize state for " 3506 "%s: %s.\n", idata.i_fmri, strerror(err)); 3507 break; 3508 3509 case EINVAL: 3510 default: 3511 bad_error("_restarter_commit_states", err); 3512 } 3513 3514 startd_free((void *)idata.i_fmri, max_scf_fmri_size); 3515 3516 v->gv_state = RESTARTER_STATE_MAINT; 3517 3518 goto out; 3519 } 3520 startd_free(restarter_fmri, max_scf_value_size); 3521 3522 /* Add all the other dependencies. */ 3523 err = refresh_vertex(v, inst); 3524 if (err != 0) { 3525 assert(err == ECONNABORTED); 3526 return (err); 3527 } 3528 3529 out: 3530 v->gv_flags |= GV_CONFIGURED; 3531 3532 graph_enable_by_vertex(v, enabled, 0); 3533 3534 return (0); 3535 } 3536 3537 3538 static void 3539 kill_user_procs(void) 3540 { 3541 (void) fputs("svc.startd: Killing user processes.\n", stdout); 3542 3543 /* 3544 * Despite its name, killall's role is to get select user processes-- 3545 * basically those representing terminal-based logins-- to die. Victims 3546 * are located by killall in the utmp database. Since these are most 3547 * often shell based logins, and many shells mask SIGTERM (but are 3548 * responsive to SIGHUP) we first HUP and then shortly thereafter 3549 * kill -9. 3550 */ 3551 (void) fork_with_timeout("/usr/sbin/killall HUP", 1, 5); 3552 (void) fork_with_timeout("/usr/sbin/killall KILL", 1, 5); 3553 3554 /* 3555 * Note the selection of user id's 0, 1 and 15, subsequently 3556 * inverted by -v. 15 is reserved for dladmd. Yes, this is a 3557 * kludge-- a better policy is needed. 3558 * 3559 * Note that fork_with_timeout will only wait out the 1 second 3560 * "grace time" if pkill actually returns 0. So if there are 3561 * no matches, this will run to completion much more quickly. 3562 */ 3563 (void) fork_with_timeout("/usr/bin/pkill -TERM -v -u 0,1,15", 1, 5); 3564 (void) fork_with_timeout("/usr/bin/pkill -KILL -v -u 0,1,15", 1, 5); 3565 } 3566 3567 static void 3568 do_uadmin(void) 3569 { 3570 const char * const resetting = "/etc/svc/volatile/resetting"; 3571 int fd; 3572 struct statvfs vfs; 3573 time_t now; 3574 struct tm nowtm; 3575 char down_buf[256], time_buf[256]; 3576 uintptr_t mdep; 3577 #if defined(__i386) 3578 grub_boot_args_t fbarg; 3579 #endif /* __i386 */ 3580 3581 mdep = NULL; 3582 fd = creat(resetting, 0777); 3583 if (fd >= 0) 3584 startd_close(fd); 3585 else 3586 uu_warn("Could not create \"%s\"", resetting); 3587 3588 /* Kill dhcpagent if we're not using nfs for root */ 3589 if ((statvfs("/", &vfs) == 0) && 3590 (strncmp(vfs.f_basetype, "nfs", sizeof ("nfs") - 1) != 0)) 3591 fork_with_timeout("/usr/bin/pkill -x -u 0 dhcpagent", 0, 5); 3592 3593 /* 3594 * Call sync(2) now, before we kill off user processes. This takes 3595 * advantage of the several seconds of pause we have before the 3596 * killalls are done. Time we can make good use of to get pages 3597 * moving out to disk. 3598 * 3599 * Inside non-global zones, we don't bother, and it's better not to 3600 * anyway, since sync(2) can have system-wide impact. 3601 */ 3602 if (getzoneid() == 0) 3603 sync(); 3604 3605 kill_user_procs(); 3606 3607 /* 3608 * Note that this must come after the killing of user procs, since 3609 * killall relies on utmpx, and this command affects the contents of 3610 * said file. 3611 */ 3612 if (access("/usr/lib/acct/closewtmp", X_OK) == 0) 3613 fork_with_timeout("/usr/lib/acct/closewtmp", 0, 5); 3614 3615 /* 3616 * For patches which may be installed as the system is shutting 3617 * down, we need to ensure, one more time, that the boot archive 3618 * really is up to date. 3619 */ 3620 if (getzoneid() == 0 && access("/usr/sbin/bootadm", X_OK) == 0) 3621 fork_with_timeout("/usr/sbin/bootadm -ea update_all", 0, 3600); 3622 3623 /* 3624 * Right now, fast reboot is supported only on i386. 3625 * scf_is_fastboot_default() should take care of it. 3626 * If somehow we got there on unsupported platform - 3627 * print warning and fall back to regular reboot. 3628 */ 3629 if (halting == AD_FASTREBOOT) { 3630 #if defined(__i386) 3631 int rc; 3632 3633 if ((rc = grub_get_boot_args(&fbarg, NULL, 3634 GRUB_ENTRY_DEFAULT)) == 0) { 3635 mdep = (uintptr_t)&fbarg.gba_bootargs; 3636 } else { 3637 /* 3638 * Failed to read GRUB menu, fall back to normal reboot 3639 */ 3640 halting = AD_BOOT; 3641 uu_warn("Failed to process GRUB menu entry " 3642 "for fast reboot.\n\t%s\n" 3643 "Falling back to regular reboot.\n", 3644 grub_strerror(rc)); 3645 } 3646 #else /* __i386 */ 3647 halting = AD_BOOT; 3648 uu_warn("Fast reboot configured, but not supported by " 3649 "this ISA\n"); 3650 #endif /* __i386 */ 3651 } 3652 3653 fork_with_timeout("/sbin/umountall -l", 0, 5); 3654 fork_with_timeout("/sbin/umount /tmp /var/adm /var/run /var " 3655 ">/dev/null 2>&1", 0, 5); 3656 3657 /* 3658 * Try to get to consistency for whatever UFS filesystems are left. 3659 * This is pretty expensive, so we save it for the end in the hopes of 3660 * minimizing what it must do. The other option would be to start in 3661 * parallel with the killall's, but lockfs tends to throw out much more 3662 * than is needed, and so subsequent commands (like umountall) take a 3663 * long time to get going again. 3664 * 3665 * Inside of zones, we don't bother, since we're not about to terminate 3666 * the whole OS instance. 3667 * 3668 * On systems using only ZFS, this call to lockfs -fa is a no-op. 3669 */ 3670 if (getzoneid() == 0) { 3671 if (access("/usr/sbin/lockfs", X_OK) == 0) 3672 fork_with_timeout("/usr/sbin/lockfs -fa", 0, 30); 3673 3674 sync(); /* once more, with feeling */ 3675 } 3676 3677 fork_with_timeout("/sbin/umount /usr >/dev/null 2>&1", 0, 5); 3678 3679 /* 3680 * Construct and emit the last words from userland: 3681 * "<timestamp> The system is down. Shutdown took <N> seconds." 3682 * 3683 * Normally we'd use syslog, but with /var and other things 3684 * potentially gone, try to minimize the external dependencies. 3685 */ 3686 now = time(NULL); 3687 (void) localtime_r(&now, &nowtm); 3688 3689 if (strftime(down_buf, sizeof (down_buf), 3690 "%b %e %T The system is down.", &nowtm) == 0) { 3691 (void) strlcpy(down_buf, "The system is down.", 3692 sizeof (down_buf)); 3693 } 3694 3695 if (halting_time != 0 && halting_time <= now) { 3696 (void) snprintf(time_buf, sizeof (time_buf), 3697 " Shutdown took %lu seconds.", now - halting_time); 3698 } else { 3699 time_buf[0] = '\0'; 3700 } 3701 (void) printf("%s%s\n", down_buf, time_buf); 3702 3703 (void) uadmin(A_SHUTDOWN, halting, mdep); 3704 uu_warn("uadmin() failed"); 3705 3706 #if defined(__i386) 3707 /* uadmin fail, cleanup grub_boot_args */ 3708 if (halting == AD_FASTREBOOT) 3709 grub_cleanup_boot_args(&fbarg); 3710 #endif /* __i386 */ 3711 3712 if (remove(resetting) != 0 && errno != ENOENT) 3713 uu_warn("Could not remove \"%s\"", resetting); 3714 } 3715 3716 /* 3717 * If any of the up_svcs[] are online or satisfiable, return true. If they are 3718 * all missing, disabled, in maintenance, or unsatisfiable, return false. 3719 */ 3720 boolean_t 3721 can_come_up(void) 3722 { 3723 int i; 3724 3725 assert(MUTEX_HELD(&dgraph_lock)); 3726 3727 /* 3728 * If we are booting to single user (boot -s), 3729 * SCF_MILESTONE_SINGLE_USER is needed to come up because startd 3730 * spawns sulogin after single-user is online (see specials.c). 3731 */ 3732 i = (booting_to_single_user ? 0 : 1); 3733 3734 for (; up_svcs[i] != NULL; ++i) { 3735 if (up_svcs_p[i] == NULL) { 3736 up_svcs_p[i] = vertex_get_by_name(up_svcs[i]); 3737 3738 if (up_svcs_p[i] == NULL) 3739 continue; 3740 } 3741 3742 /* 3743 * Ignore unconfigured services (the ones that have been 3744 * mentioned in a dependency from other services, but do 3745 * not exist in the repository). Services which exist 3746 * in the repository but don't have general/enabled 3747 * property will be also ignored. 3748 */ 3749 if (!(up_svcs_p[i]->gv_flags & GV_CONFIGURED)) 3750 continue; 3751 3752 switch (up_svcs_p[i]->gv_state) { 3753 case RESTARTER_STATE_ONLINE: 3754 case RESTARTER_STATE_DEGRADED: 3755 /* 3756 * Deactivate verbose boot once a login service has been 3757 * reached. 3758 */ 3759 st->st_log_login_reached = 1; 3760 /*FALLTHROUGH*/ 3761 case RESTARTER_STATE_UNINIT: 3762 return (B_TRUE); 3763 3764 case RESTARTER_STATE_OFFLINE: 3765 if (instance_satisfied(up_svcs_p[i], B_TRUE) != -1) 3766 return (B_TRUE); 3767 log_framework(LOG_DEBUG, 3768 "can_come_up(): %s is unsatisfiable.\n", 3769 up_svcs_p[i]->gv_name); 3770 continue; 3771 3772 case RESTARTER_STATE_DISABLED: 3773 case RESTARTER_STATE_MAINT: 3774 log_framework(LOG_DEBUG, 3775 "can_come_up(): %s is in state %s.\n", 3776 up_svcs_p[i]->gv_name, 3777 instance_state_str[up_svcs_p[i]->gv_state]); 3778 continue; 3779 3780 default: 3781 #ifndef NDEBUG 3782 uu_warn("%s:%d: Unexpected vertex state %d.\n", 3783 __FILE__, __LINE__, up_svcs_p[i]->gv_state); 3784 #endif 3785 abort(); 3786 } 3787 } 3788 3789 /* 3790 * In the seed repository, console-login is unsatisfiable because 3791 * services are missing. To behave correctly in that case we don't want 3792 * to return false until manifest-import is online. 3793 */ 3794 3795 if (manifest_import_p == NULL) { 3796 manifest_import_p = vertex_get_by_name(manifest_import); 3797 3798 if (manifest_import_p == NULL) 3799 return (B_FALSE); 3800 } 3801 3802 switch (manifest_import_p->gv_state) { 3803 case RESTARTER_STATE_ONLINE: 3804 case RESTARTER_STATE_DEGRADED: 3805 case RESTARTER_STATE_DISABLED: 3806 case RESTARTER_STATE_MAINT: 3807 break; 3808 3809 case RESTARTER_STATE_OFFLINE: 3810 if (instance_satisfied(manifest_import_p, B_TRUE) == -1) 3811 break; 3812 /* FALLTHROUGH */ 3813 3814 case RESTARTER_STATE_UNINIT: 3815 return (B_TRUE); 3816 } 3817 3818 return (B_FALSE); 3819 } 3820 3821 /* 3822 * Runs sulogin. Returns 3823 * 0 - success 3824 * EALREADY - sulogin is already running 3825 * EBUSY - console-login is running 3826 */ 3827 static int 3828 run_sulogin(const char *msg) 3829 { 3830 graph_vertex_t *v; 3831 3832 assert(MUTEX_HELD(&dgraph_lock)); 3833 3834 if (sulogin_running) 3835 return (EALREADY); 3836 3837 v = vertex_get_by_name(console_login_fmri); 3838 if (v != NULL && inst_running(v)) 3839 return (EBUSY); 3840 3841 sulogin_running = B_TRUE; 3842 3843 MUTEX_UNLOCK(&dgraph_lock); 3844 3845 fork_sulogin(B_FALSE, msg); 3846 3847 MUTEX_LOCK(&dgraph_lock); 3848 3849 sulogin_running = B_FALSE; 3850 3851 if (console_login_ready) { 3852 v = vertex_get_by_name(console_login_fmri); 3853 3854 if (v != NULL && v->gv_state == RESTARTER_STATE_OFFLINE) { 3855 if (v->gv_start_f == NULL) 3856 vertex_send_event(v, 3857 RESTARTER_EVENT_TYPE_START); 3858 else 3859 v->gv_start_f(v); 3860 } 3861 3862 console_login_ready = B_FALSE; 3863 } 3864 3865 return (0); 3866 } 3867 3868 /* 3869 * The sulogin thread runs sulogin while can_come_up() is false. run_sulogin() 3870 * keeps sulogin from stepping on console-login's toes. 3871 */ 3872 /* ARGSUSED */ 3873 static void * 3874 sulogin_thread(void *unused) 3875 { 3876 MUTEX_LOCK(&dgraph_lock); 3877 3878 assert(sulogin_thread_running); 3879 3880 do { 3881 (void) run_sulogin("Console login service(s) cannot run\n"); 3882 } while (!can_come_up()); 3883 3884 sulogin_thread_running = B_FALSE; 3885 MUTEX_UNLOCK(&dgraph_lock); 3886 3887 return (NULL); 3888 } 3889 3890 /* ARGSUSED */ 3891 void * 3892 single_user_thread(void *unused) 3893 { 3894 uint_t left; 3895 scf_handle_t *h; 3896 scf_instance_t *inst; 3897 scf_property_t *prop; 3898 scf_value_t *val; 3899 const char *msg; 3900 char *buf; 3901 int r; 3902 3903 MUTEX_LOCK(&single_user_thread_lock); 3904 single_user_thread_count++; 3905 3906 if (!booting_to_single_user) 3907 kill_user_procs(); 3908 3909 if (go_single_user_mode || booting_to_single_user) { 3910 msg = "SINGLE USER MODE\n"; 3911 } else { 3912 assert(go_to_level1); 3913 3914 fork_rc_script('1', "start", B_TRUE); 3915 3916 uu_warn("The system is ready for administration.\n"); 3917 3918 msg = ""; 3919 } 3920 3921 MUTEX_UNLOCK(&single_user_thread_lock); 3922 3923 for (;;) { 3924 MUTEX_LOCK(&dgraph_lock); 3925 r = run_sulogin(msg); 3926 MUTEX_UNLOCK(&dgraph_lock); 3927 if (r == 0) 3928 break; 3929 3930 assert(r == EALREADY || r == EBUSY); 3931 3932 left = 3; 3933 while (left > 0) 3934 left = sleep(left); 3935 } 3936 3937 MUTEX_LOCK(&single_user_thread_lock); 3938 3939 /* 3940 * If another single user thread has started, let it finish changing 3941 * the run level. 3942 */ 3943 if (single_user_thread_count > 1) { 3944 single_user_thread_count--; 3945 MUTEX_UNLOCK(&single_user_thread_lock); 3946 return (NULL); 3947 } 3948 3949 h = libscf_handle_create_bound_loop(); 3950 inst = scf_instance_create(h); 3951 prop = safe_scf_property_create(h); 3952 val = safe_scf_value_create(h); 3953 buf = startd_alloc(max_scf_fmri_size); 3954 3955 lookup: 3956 if (scf_handle_decode_fmri(h, SCF_SERVICE_STARTD, NULL, NULL, inst, 3957 NULL, NULL, SCF_DECODE_FMRI_EXACT) != 0) { 3958 switch (scf_error()) { 3959 case SCF_ERROR_NOT_FOUND: 3960 r = libscf_create_self(h); 3961 if (r == 0) 3962 goto lookup; 3963 assert(r == ECONNABORTED); 3964 /* FALLTHROUGH */ 3965 3966 case SCF_ERROR_CONNECTION_BROKEN: 3967 libscf_handle_rebind(h); 3968 goto lookup; 3969 3970 case SCF_ERROR_INVALID_ARGUMENT: 3971 case SCF_ERROR_CONSTRAINT_VIOLATED: 3972 case SCF_ERROR_NOT_BOUND: 3973 case SCF_ERROR_HANDLE_MISMATCH: 3974 default: 3975 bad_error("scf_handle_decode_fmri", scf_error()); 3976 } 3977 } 3978 3979 MUTEX_LOCK(&dgraph_lock); 3980 3981 r = scf_instance_delete_prop(inst, SCF_PG_OPTIONS_OVR, 3982 SCF_PROPERTY_MILESTONE); 3983 switch (r) { 3984 case 0: 3985 case ECANCELED: 3986 break; 3987 3988 case ECONNABORTED: 3989 MUTEX_UNLOCK(&dgraph_lock); 3990 libscf_handle_rebind(h); 3991 goto lookup; 3992 3993 case EPERM: 3994 case EACCES: 3995 case EROFS: 3996 log_error(LOG_WARNING, "Could not clear temporary milestone: " 3997 "%s.\n", strerror(r)); 3998 break; 3999 4000 default: 4001 bad_error("scf_instance_delete_prop", r); 4002 } 4003 4004 MUTEX_UNLOCK(&dgraph_lock); 4005 4006 r = libscf_get_milestone(inst, prop, val, buf, max_scf_fmri_size); 4007 switch (r) { 4008 case ECANCELED: 4009 case ENOENT: 4010 case EINVAL: 4011 (void) strcpy(buf, "all"); 4012 /* FALLTHROUGH */ 4013 4014 case 0: 4015 uu_warn("Returning to milestone %s.\n", buf); 4016 break; 4017 4018 case ECONNABORTED: 4019 libscf_handle_rebind(h); 4020 goto lookup; 4021 4022 default: 4023 bad_error("libscf_get_milestone", r); 4024 } 4025 4026 r = dgraph_set_milestone(buf, h, B_FALSE); 4027 switch (r) { 4028 case 0: 4029 case ECONNRESET: 4030 case EALREADY: 4031 case EINVAL: 4032 case ENOENT: 4033 break; 4034 4035 default: 4036 bad_error("dgraph_set_milestone", r); 4037 } 4038 4039 /* 4040 * See graph_runlevel_changed(). 4041 */ 4042 MUTEX_LOCK(&dgraph_lock); 4043 utmpx_set_runlevel(target_milestone_as_runlevel(), 'S', B_TRUE); 4044 MUTEX_UNLOCK(&dgraph_lock); 4045 4046 startd_free(buf, max_scf_fmri_size); 4047 scf_value_destroy(val); 4048 scf_property_destroy(prop); 4049 scf_instance_destroy(inst); 4050 scf_handle_destroy(h); 4051 4052 /* 4053 * We'll give ourselves 3 seconds to respond to all of the enablings 4054 * that setting the milestone should have created before checking 4055 * whether to run sulogin. 4056 */ 4057 left = 3; 4058 while (left > 0) 4059 left = sleep(left); 4060 4061 MUTEX_LOCK(&dgraph_lock); 4062 /* 4063 * Clearing these variables will allow the sulogin thread to run. We 4064 * check here in case there aren't any more state updates anytime soon. 4065 */ 4066 go_to_level1 = go_single_user_mode = booting_to_single_user = B_FALSE; 4067 if (!sulogin_thread_running && !can_come_up()) { 4068 (void) startd_thread_create(sulogin_thread, NULL); 4069 sulogin_thread_running = B_TRUE; 4070 } 4071 MUTEX_UNLOCK(&dgraph_lock); 4072 single_user_thread_count--; 4073 MUTEX_UNLOCK(&single_user_thread_lock); 4074 return (NULL); 4075 } 4076 4077 4078 /* 4079 * Dependency graph operations API. These are handle-independent thread-safe 4080 * graph manipulation functions which are the entry points for the event 4081 * threads below. 4082 */ 4083 4084 /* 4085 * If a configured vertex exists for inst_fmri, return EEXIST. If no vertex 4086 * exists for inst_fmri, add one. Then fetch the restarter from inst, make 4087 * this vertex dependent on it, and send _ADD_INSTANCE to the restarter. 4088 * Fetch whether the instance should be enabled from inst and send _ENABLE or 4089 * _DISABLE as appropriate. Finally rummage through inst's dependency 4090 * property groups and add vertices and edges as appropriate. If anything 4091 * goes wrong after sending _ADD_INSTANCE, send _ADMIN_MAINT_ON to put the 4092 * instance in maintenance. Don't send _START or _STOP until we get a state 4093 * update in case we're being restarted and the service is already running. 4094 * 4095 * To support booting to a milestone, we must also make sure all dependencies 4096 * encountered are configured, if they exist in the repository. 4097 * 4098 * Returns 0 on success, ECONNABORTED on repository disconnection, EINVAL if 4099 * inst_fmri is an invalid (or not canonical) FMRI, ECANCELED if inst is 4100 * deleted, or EEXIST if a configured vertex for inst_fmri already exists. 4101 */ 4102 int 4103 dgraph_add_instance(const char *inst_fmri, scf_instance_t *inst, 4104 boolean_t lock_graph) 4105 { 4106 graph_vertex_t *v; 4107 int err; 4108 4109 if (strcmp(inst_fmri, SCF_SERVICE_STARTD) == 0) 4110 return (0); 4111 4112 /* Check for a vertex for inst_fmri. */ 4113 if (lock_graph) { 4114 MUTEX_LOCK(&dgraph_lock); 4115 } else { 4116 assert(MUTEX_HELD(&dgraph_lock)); 4117 } 4118 4119 v = vertex_get_by_name(inst_fmri); 4120 4121 if (v != NULL) { 4122 assert(v->gv_type == GVT_INST); 4123 4124 if (v->gv_flags & GV_CONFIGURED) { 4125 if (lock_graph) 4126 MUTEX_UNLOCK(&dgraph_lock); 4127 return (EEXIST); 4128 } 4129 } else { 4130 /* Add the vertex. */ 4131 err = graph_insert_vertex_unconfigured(inst_fmri, GVT_INST, 0, 4132 RERR_NONE, &v); 4133 if (err != 0) { 4134 assert(err == EINVAL); 4135 if (lock_graph) 4136 MUTEX_UNLOCK(&dgraph_lock); 4137 return (EINVAL); 4138 } 4139 } 4140 4141 err = configure_vertex(v, inst); 4142 4143 if (lock_graph) 4144 MUTEX_UNLOCK(&dgraph_lock); 4145 4146 return (err); 4147 } 4148 4149 /* 4150 * Locate the vertex for this property group's instance. If it doesn't exist 4151 * or is unconfigured, call dgraph_add_instance() & return. Otherwise fetch 4152 * the restarter for the instance, and if it has changed, send 4153 * _REMOVE_INSTANCE to the old restarter, remove the dependency, make sure the 4154 * new restarter has a vertex, add a new dependency, and send _ADD_INSTANCE to 4155 * the new restarter. Then fetch whether the instance should be enabled, and 4156 * if it is different from what we had, or if we changed the restarter, send 4157 * the appropriate _ENABLE or _DISABLE command. 4158 * 4159 * Returns 0 on success, ENOTSUP if the pg's parent is not an instance, 4160 * ECONNABORTED on repository disconnection, ECANCELED if the instance is 4161 * deleted, or -1 if the instance's general property group is deleted or if 4162 * its enabled property is misconfigured. 4163 */ 4164 static int 4165 dgraph_update_general(scf_propertygroup_t *pg) 4166 { 4167 scf_handle_t *h; 4168 scf_instance_t *inst; 4169 char *fmri; 4170 char *restarter_fmri; 4171 graph_vertex_t *v; 4172 int err; 4173 int enabled, enabled_ovr; 4174 int oldflags; 4175 4176 /* Find the vertex for this service */ 4177 h = scf_pg_handle(pg); 4178 4179 inst = safe_scf_instance_create(h); 4180 4181 if (scf_pg_get_parent_instance(pg, inst) != 0) { 4182 switch (scf_error()) { 4183 case SCF_ERROR_CONSTRAINT_VIOLATED: 4184 return (ENOTSUP); 4185 4186 case SCF_ERROR_CONNECTION_BROKEN: 4187 default: 4188 return (ECONNABORTED); 4189 4190 case SCF_ERROR_DELETED: 4191 return (0); 4192 4193 case SCF_ERROR_NOT_SET: 4194 bad_error("scf_pg_get_parent_instance", scf_error()); 4195 } 4196 } 4197 4198 err = libscf_instance_get_fmri(inst, &fmri); 4199 switch (err) { 4200 case 0: 4201 break; 4202 4203 case ECONNABORTED: 4204 scf_instance_destroy(inst); 4205 return (ECONNABORTED); 4206 4207 case ECANCELED: 4208 scf_instance_destroy(inst); 4209 return (0); 4210 4211 default: 4212 bad_error("libscf_instance_get_fmri", err); 4213 } 4214 4215 log_framework(LOG_DEBUG, 4216 "Graph engine: Reloading general properties for %s.\n", fmri); 4217 4218 MUTEX_LOCK(&dgraph_lock); 4219 4220 v = vertex_get_by_name(fmri); 4221 if (v == NULL || !(v->gv_flags & GV_CONFIGURED)) { 4222 /* Will get the up-to-date properties. */ 4223 MUTEX_UNLOCK(&dgraph_lock); 4224 err = dgraph_add_instance(fmri, inst, B_TRUE); 4225 startd_free(fmri, max_scf_fmri_size); 4226 scf_instance_destroy(inst); 4227 return (err == ECANCELED ? 0 : err); 4228 } 4229 4230 /* Read enabled & restarter from repository. */ 4231 restarter_fmri = startd_alloc(max_scf_value_size); 4232 err = libscf_get_basic_instance_data(h, inst, v->gv_name, &enabled, 4233 &enabled_ovr, &restarter_fmri); 4234 if (err != 0 || enabled == -1) { 4235 MUTEX_UNLOCK(&dgraph_lock); 4236 scf_instance_destroy(inst); 4237 startd_free(fmri, max_scf_fmri_size); 4238 4239 switch (err) { 4240 case ENOENT: 4241 case 0: 4242 startd_free(restarter_fmri, max_scf_value_size); 4243 return (-1); 4244 4245 case ECONNABORTED: 4246 case ECANCELED: 4247 startd_free(restarter_fmri, max_scf_value_size); 4248 return (err); 4249 4250 default: 4251 bad_error("libscf_get_basic_instance_data", err); 4252 } 4253 } 4254 4255 oldflags = v->gv_flags; 4256 v->gv_flags = (v->gv_flags & ~GV_ENBLD_NOOVR) | 4257 (enabled ? GV_ENBLD_NOOVR : 0); 4258 4259 if (enabled_ovr != -1) 4260 enabled = enabled_ovr; 4261 4262 /* 4263 * If GV_ENBLD_NOOVR has changed, then we need to re-evaluate the 4264 * subgraph. 4265 */ 4266 if (milestone > MILESTONE_NONE && v->gv_flags != oldflags) 4267 (void) eval_subgraph(v, h); 4268 4269 scf_instance_destroy(inst); 4270 4271 /* Ignore restarter change for now. */ 4272 4273 startd_free(restarter_fmri, max_scf_value_size); 4274 startd_free(fmri, max_scf_fmri_size); 4275 4276 /* 4277 * Always send _ENABLE or _DISABLE. We could avoid this if the 4278 * restarter didn't change and the enabled value didn't change, but 4279 * that's not easy to check and improbable anyway, so we'll just do 4280 * this. 4281 */ 4282 graph_enable_by_vertex(v, enabled, 1); 4283 4284 MUTEX_UNLOCK(&dgraph_lock); 4285 4286 return (0); 4287 } 4288 4289 /* 4290 * Delete all of the property group dependencies of v, update inst's running 4291 * snapshot, and add the dependencies in the new snapshot. If any of the new 4292 * dependencies would create a cycle, send _ADMIN_MAINT_ON. Otherwise 4293 * reevaluate v's dependencies, send _START or _STOP as appropriate, and do 4294 * the same for v's dependents. 4295 * 4296 * Returns 4297 * 0 - success 4298 * ECONNABORTED - repository connection broken 4299 * ECANCELED - inst was deleted 4300 * EINVAL - inst is invalid (e.g., missing general/enabled) 4301 * -1 - libscf_snapshots_refresh() failed 4302 */ 4303 static int 4304 dgraph_refresh_instance(graph_vertex_t *v, scf_instance_t *inst) 4305 { 4306 int r; 4307 int enabled; 4308 int32_t tset; 4309 4310 assert(MUTEX_HELD(&dgraph_lock)); 4311 assert(v->gv_type == GVT_INST); 4312 4313 /* Only refresh services with valid general/enabled properties. */ 4314 r = libscf_get_basic_instance_data(scf_instance_handle(inst), inst, 4315 v->gv_name, &enabled, NULL, NULL); 4316 switch (r) { 4317 case 0: 4318 break; 4319 4320 case ECONNABORTED: 4321 case ECANCELED: 4322 return (r); 4323 4324 case ENOENT: 4325 log_framework(LOG_DEBUG, 4326 "Ignoring %s because it has no general property group.\n", 4327 v->gv_name); 4328 return (EINVAL); 4329 4330 default: 4331 bad_error("libscf_get_basic_instance_data", r); 4332 } 4333 4334 if ((tset = libscf_get_stn_tset(inst)) == -1) { 4335 log_framework(LOG_WARNING, 4336 "Failed to get notification parameters for %s: %s\n", 4337 v->gv_name, scf_strerror(scf_error())); 4338 tset = 0; 4339 } 4340 v->gv_stn_tset = tset; 4341 if (strcmp(v->gv_name, SCF_INSTANCE_GLOBAL) == 0) 4342 stn_global = tset; 4343 4344 if (enabled == -1) 4345 return (EINVAL); 4346 4347 r = libscf_snapshots_refresh(inst, v->gv_name); 4348 if (r != 0) { 4349 if (r != -1) 4350 bad_error("libscf_snapshots_refresh", r); 4351 4352 /* error logged */ 4353 return (r); 4354 } 4355 4356 r = refresh_vertex(v, inst); 4357 if (r != 0 && r != ECONNABORTED) 4358 bad_error("refresh_vertex", r); 4359 return (r); 4360 } 4361 4362 /* 4363 * Returns true only if none of this service's dependents are 'up' -- online 4364 * or degraded (offline is considered down in this situation). This function 4365 * is somehow similar to is_nonsubgraph_leaf() but works on subtrees. 4366 */ 4367 boolean_t 4368 insubtree_dependents_down(graph_vertex_t *v) 4369 { 4370 graph_vertex_t *vv; 4371 graph_edge_t *e; 4372 4373 assert(MUTEX_HELD(&dgraph_lock)); 4374 4375 for (e = uu_list_first(v->gv_dependents); e != NULL; 4376 e = uu_list_next(v->gv_dependents, e)) { 4377 vv = e->ge_vertex; 4378 if (vv->gv_type == GVT_INST) { 4379 if ((vv->gv_flags & GV_CONFIGURED) == 0) 4380 continue; 4381 4382 if ((vv->gv_flags & GV_TOOFFLINE) == 0) 4383 return (B_FALSE); 4384 4385 if ((vv->gv_state == RESTARTER_STATE_ONLINE) || 4386 (vv->gv_state == RESTARTER_STATE_DEGRADED)) 4387 return (B_FALSE); 4388 } else { 4389 /* 4390 * Skip all excluded dependencies and decide whether 4391 * decide whether to offline the service based on the 4392 * restart_on attribute. 4393 */ 4394 if (is_depgrp_bypassed(vv)) 4395 continue; 4396 4397 /* 4398 * For dependency groups or service vertices, keep 4399 * traversing to see if instances are running. 4400 */ 4401 if (insubtree_dependents_down(vv) == B_FALSE) 4402 return (B_FALSE); 4403 } 4404 } 4405 4406 return (B_TRUE); 4407 } 4408 4409 /* 4410 * Returns true only if none of this service's dependents are 'up' -- online, 4411 * degraded, or offline. 4412 */ 4413 static int 4414 is_nonsubgraph_leaf(graph_vertex_t *v) 4415 { 4416 graph_vertex_t *vv; 4417 graph_edge_t *e; 4418 4419 assert(MUTEX_HELD(&dgraph_lock)); 4420 4421 for (e = uu_list_first(v->gv_dependents); 4422 e != NULL; 4423 e = uu_list_next(v->gv_dependents, e)) { 4424 4425 vv = e->ge_vertex; 4426 if (vv->gv_type == GVT_INST) { 4427 if ((vv->gv_flags & GV_CONFIGURED) == 0) 4428 continue; 4429 4430 if (vv->gv_flags & GV_INSUBGRAPH) 4431 continue; 4432 4433 if (up_state(vv->gv_state)) 4434 return (0); 4435 } else { 4436 /* 4437 * For dependency group or service vertices, keep 4438 * traversing to see if instances are running. 4439 * 4440 * We should skip exclude_all dependencies otherwise 4441 * the vertex will never be considered as a leaf 4442 * if the dependent is offline. The main reason for 4443 * this is that disable_nonsubgraph_leaves() skips 4444 * exclusion dependencies. 4445 */ 4446 if (vv->gv_type == GVT_GROUP && 4447 vv->gv_depgroup == DEPGRP_EXCLUDE_ALL) 4448 continue; 4449 4450 if (!is_nonsubgraph_leaf(vv)) 4451 return (0); 4452 } 4453 } 4454 4455 return (1); 4456 } 4457 4458 /* 4459 * Disable v temporarily. Attempt to do this by setting its enabled override 4460 * property in the repository. If that fails, send a _DISABLE command. 4461 * Returns 0 on success and ECONNABORTED if the repository connection is 4462 * broken. 4463 */ 4464 static int 4465 disable_service_temporarily(graph_vertex_t *v, scf_handle_t *h) 4466 { 4467 const char * const emsg = "Could not temporarily disable %s because " 4468 "%s. Will stop service anyways. Repository status for the " 4469 "service may be inaccurate.\n"; 4470 const char * const emsg_cbroken = 4471 "the repository connection was broken"; 4472 4473 scf_instance_t *inst; 4474 int r; 4475 4476 inst = scf_instance_create(h); 4477 if (inst == NULL) { 4478 char buf[100]; 4479 4480 (void) snprintf(buf, sizeof (buf), 4481 "scf_instance_create() failed (%s)", 4482 scf_strerror(scf_error())); 4483 log_error(LOG_WARNING, emsg, v->gv_name, buf); 4484 4485 graph_enable_by_vertex(v, 0, 0); 4486 return (0); 4487 } 4488 4489 r = scf_handle_decode_fmri(h, v->gv_name, NULL, NULL, inst, 4490 NULL, NULL, SCF_DECODE_FMRI_EXACT); 4491 if (r != 0) { 4492 switch (scf_error()) { 4493 case SCF_ERROR_CONNECTION_BROKEN: 4494 log_error(LOG_WARNING, emsg, v->gv_name, emsg_cbroken); 4495 graph_enable_by_vertex(v, 0, 0); 4496 return (ECONNABORTED); 4497 4498 case SCF_ERROR_NOT_FOUND: 4499 return (0); 4500 4501 case SCF_ERROR_HANDLE_MISMATCH: 4502 case SCF_ERROR_INVALID_ARGUMENT: 4503 case SCF_ERROR_CONSTRAINT_VIOLATED: 4504 case SCF_ERROR_NOT_BOUND: 4505 default: 4506 bad_error("scf_handle_decode_fmri", 4507 scf_error()); 4508 } 4509 } 4510 4511 r = libscf_set_enable_ovr(inst, 0); 4512 switch (r) { 4513 case 0: 4514 scf_instance_destroy(inst); 4515 return (0); 4516 4517 case ECANCELED: 4518 scf_instance_destroy(inst); 4519 return (0); 4520 4521 case ECONNABORTED: 4522 log_error(LOG_WARNING, emsg, v->gv_name, emsg_cbroken); 4523 graph_enable_by_vertex(v, 0, 0); 4524 return (ECONNABORTED); 4525 4526 case EPERM: 4527 log_error(LOG_WARNING, emsg, v->gv_name, 4528 "the repository denied permission"); 4529 graph_enable_by_vertex(v, 0, 0); 4530 return (0); 4531 4532 case EROFS: 4533 log_error(LOG_WARNING, emsg, v->gv_name, 4534 "the repository is read-only"); 4535 graph_enable_by_vertex(v, 0, 0); 4536 return (0); 4537 4538 default: 4539 bad_error("libscf_set_enable_ovr", r); 4540 /* NOTREACHED */ 4541 } 4542 } 4543 4544 /* 4545 * Of the transitive instance dependencies of v, offline those which are 4546 * in the subtree and which are leaves (i.e., have no dependents which are 4547 * "up"). 4548 */ 4549 void 4550 offline_subtree_leaves(graph_vertex_t *v, void *arg) 4551 { 4552 assert(MUTEX_HELD(&dgraph_lock)); 4553 4554 /* If v has up dependents, try to bring them down first. */ 4555 if (insubtree_dependents_down(v) == B_FALSE) { 4556 graph_walk_dependents(v, offline_subtree_leaves, arg); 4557 4558 /* If we couldn't bring them down, return. */ 4559 if (insubtree_dependents_down(v) == B_FALSE) 4560 return; 4561 } 4562 4563 /* If v isn't an instance, recurse on its dependencies. */ 4564 if (v->gv_type != GVT_INST) { 4565 graph_walk_dependencies(v, offline_subtree_leaves, arg); 4566 return; 4567 } 4568 4569 /* 4570 * If v is not in the subtree, so should all of its dependencies, 4571 * so do nothing. 4572 */ 4573 if ((v->gv_flags & GV_TOOFFLINE) == 0) 4574 return; 4575 4576 /* If v isn't a leaf because it's already down, recurse. */ 4577 if (!up_state(v->gv_state)) { 4578 graph_walk_dependencies(v, offline_subtree_leaves, arg); 4579 return; 4580 } 4581 4582 /* if v is a leaf, offline it or disable it if it's the last one */ 4583 if (insubtree_dependents_down(v) == B_TRUE) { 4584 if (v->gv_flags & GV_TODISABLE) 4585 vertex_send_event(v, 4586 RESTARTER_EVENT_TYPE_ADMIN_DISABLE); 4587 else 4588 offline_vertex(v); 4589 } 4590 } 4591 4592 void 4593 graph_offline_subtree_leaves(graph_vertex_t *v, void *h) 4594 { 4595 graph_walk_dependencies(v, offline_subtree_leaves, (void *)h); 4596 } 4597 4598 4599 /* 4600 * Of the transitive instance dependencies of v, disable those which are not 4601 * in the subgraph and which are leaves (i.e., have no dependents which are 4602 * "up"). 4603 */ 4604 static void 4605 disable_nonsubgraph_leaves(graph_vertex_t *v, void *arg) 4606 { 4607 assert(MUTEX_HELD(&dgraph_lock)); 4608 4609 /* 4610 * We must skip exclusion dependencies because they are allowed to 4611 * complete dependency cycles. This is correct because A's exclusion 4612 * dependency on B doesn't bear on the order in which they should be 4613 * stopped. Indeed, the exclusion dependency should guarantee that 4614 * they are never online at the same time. 4615 */ 4616 if (v->gv_type == GVT_GROUP && v->gv_depgroup == DEPGRP_EXCLUDE_ALL) 4617 return; 4618 4619 /* If v isn't an instance, recurse on its dependencies. */ 4620 if (v->gv_type != GVT_INST) 4621 goto recurse; 4622 4623 if ((v->gv_flags & GV_CONFIGURED) == 0) 4624 /* 4625 * Unconfigured instances should have no dependencies, but in 4626 * case they ever get them, 4627 */ 4628 goto recurse; 4629 4630 /* 4631 * If v is in the subgraph, so should all of its dependencies, so do 4632 * nothing. 4633 */ 4634 if (v->gv_flags & GV_INSUBGRAPH) 4635 return; 4636 4637 /* If v isn't a leaf because it's already down, recurse. */ 4638 if (!up_state(v->gv_state)) 4639 goto recurse; 4640 4641 /* If v is disabled but not down yet, be patient. */ 4642 if ((v->gv_flags & GV_ENABLED) == 0) 4643 return; 4644 4645 /* If v is a leaf, disable it. */ 4646 if (is_nonsubgraph_leaf(v)) 4647 (void) disable_service_temporarily(v, (scf_handle_t *)arg); 4648 4649 return; 4650 4651 recurse: 4652 graph_walk_dependencies(v, disable_nonsubgraph_leaves, arg); 4653 } 4654 4655 static int 4656 stn_restarter_state(restarter_instance_state_t rstate) 4657 { 4658 static const struct statemap { 4659 restarter_instance_state_t restarter_state; 4660 int scf_state; 4661 } map[] = { 4662 { RESTARTER_STATE_UNINIT, SCF_STATE_UNINIT }, 4663 { RESTARTER_STATE_MAINT, SCF_STATE_MAINT }, 4664 { RESTARTER_STATE_OFFLINE, SCF_STATE_OFFLINE }, 4665 { RESTARTER_STATE_DISABLED, SCF_STATE_DISABLED }, 4666 { RESTARTER_STATE_ONLINE, SCF_STATE_ONLINE }, 4667 { RESTARTER_STATE_DEGRADED, SCF_STATE_DEGRADED } 4668 }; 4669 4670 int i; 4671 4672 for (i = 0; i < sizeof (map) / sizeof (map[0]); i++) { 4673 if (rstate == map[i].restarter_state) 4674 return (map[i].scf_state); 4675 } 4676 4677 return (-1); 4678 } 4679 4680 /* 4681 * State transition counters 4682 * Not incremented atomically - indicative only 4683 */ 4684 static uint64_t stev_ct_maint; 4685 static uint64_t stev_ct_hwerr; 4686 static uint64_t stev_ct_service; 4687 static uint64_t stev_ct_global; 4688 static uint64_t stev_ct_noprefs; 4689 static uint64_t stev_ct_from_uninit; 4690 static uint64_t stev_ct_bad_state; 4691 static uint64_t stev_ct_ovr_prefs; 4692 4693 static void 4694 dgraph_state_transition_notify(graph_vertex_t *v, 4695 restarter_instance_state_t old_state, restarter_str_t reason) 4696 { 4697 restarter_instance_state_t new_state = v->gv_state; 4698 int stn_transition, maint; 4699 int from, to; 4700 nvlist_t *attr; 4701 fmev_pri_t pri = FMEV_LOPRI; 4702 int raise = 0; 4703 4704 if ((from = stn_restarter_state(old_state)) == -1 || 4705 (to = stn_restarter_state(new_state)) == -1) { 4706 stev_ct_bad_state++; 4707 return; 4708 } 4709 4710 stn_transition = from << 16 | to; 4711 4712 maint = (to == SCF_STATE_MAINT || from == SCF_STATE_MAINT); 4713 4714 if (maint) { 4715 /* 4716 * All transitions to/from maintenance state must raise 4717 * an event. 4718 */ 4719 raise++; 4720 pri = FMEV_HIPRI; 4721 stev_ct_maint++; 4722 } else if (reason == restarter_str_ct_ev_hwerr) { 4723 /* 4724 * All transitions caused by hardware fault must raise 4725 * an event 4726 */ 4727 raise++; 4728 pri = FMEV_HIPRI; 4729 stev_ct_hwerr++; 4730 } else if (stn_transition & v->gv_stn_tset) { 4731 /* 4732 * Specifically enabled event. 4733 */ 4734 raise++; 4735 stev_ct_service++; 4736 } else if (from == SCF_STATE_UNINIT) { 4737 /* 4738 * Only raise these if specifically selected above. 4739 */ 4740 stev_ct_from_uninit++; 4741 } else if (stn_transition & stn_global && 4742 (IS_ENABLED(v) == 1 || to == SCF_STATE_DISABLED)) { 4743 raise++; 4744 stev_ct_global++; 4745 } else { 4746 stev_ct_noprefs++; 4747 } 4748 4749 if (info_events_all) { 4750 stev_ct_ovr_prefs++; 4751 raise++; 4752 } 4753 if (!raise) 4754 return; 4755 4756 if (nvlist_alloc(&attr, NV_UNIQUE_NAME, 0) != 0 || 4757 nvlist_add_string(attr, "fmri", v->gv_name) != 0 || 4758 nvlist_add_uint32(attr, "reason-version", 4759 restarter_str_version()) || nvlist_add_string(attr, "reason-short", 4760 restarter_get_str_short(reason)) != 0 || 4761 nvlist_add_string(attr, "reason-long", 4762 restarter_get_str_long(reason)) != 0 || 4763 nvlist_add_int32(attr, "transition", stn_transition) != 0) { 4764 log_framework(LOG_WARNING, 4765 "FMEV: %s could not create nvlist for transition " 4766 "event: %s\n", v->gv_name, strerror(errno)); 4767 nvlist_free(attr); 4768 return; 4769 } 4770 4771 if (fmev_rspublish_nvl(FMEV_RULESET_SMF, "state-transition", 4772 instance_state_str[new_state], pri, attr) != FMEV_SUCCESS) { 4773 log_framework(LOG_DEBUG, 4774 "FMEV: %s failed to publish transition event: %s\n", 4775 v->gv_name, fmev_strerror(fmev_errno)); 4776 nvlist_free(attr); 4777 } 4778 } 4779 4780 /* 4781 * Find the vertex for inst_name. If it doesn't exist, return ENOENT. 4782 * Otherwise set its state to state. If the instance has entered a state 4783 * which requires automatic action, take it (Uninitialized: do 4784 * dgraph_refresh_instance() without the snapshot update. Disabled: if the 4785 * instance should be enabled, send _ENABLE. Offline: if the instance should 4786 * be disabled, send _DISABLE, and if its dependencies are satisfied, send 4787 * _START. Online, Degraded: if the instance wasn't running, update its start 4788 * snapshot. Maintenance: no action.) 4789 * 4790 * Also fails with ECONNABORTED, or EINVAL if state is invalid. 4791 */ 4792 static int 4793 dgraph_set_instance_state(scf_handle_t *h, const char *inst_name, 4794 protocol_states_t *states) 4795 { 4796 graph_vertex_t *v; 4797 int err = 0; 4798 restarter_instance_state_t old_state; 4799 restarter_instance_state_t state = states->ps_state; 4800 restarter_error_t serr = states->ps_err; 4801 4802 MUTEX_LOCK(&dgraph_lock); 4803 4804 v = vertex_get_by_name(inst_name); 4805 if (v == NULL) { 4806 MUTEX_UNLOCK(&dgraph_lock); 4807 return (ENOENT); 4808 } 4809 4810 assert(v->gv_type == GVT_INST); 4811 4812 switch (state) { 4813 case RESTARTER_STATE_UNINIT: 4814 case RESTARTER_STATE_DISABLED: 4815 case RESTARTER_STATE_OFFLINE: 4816 case RESTARTER_STATE_ONLINE: 4817 case RESTARTER_STATE_DEGRADED: 4818 case RESTARTER_STATE_MAINT: 4819 break; 4820 4821 default: 4822 MUTEX_UNLOCK(&dgraph_lock); 4823 return (EINVAL); 4824 } 4825 4826 log_framework(LOG_DEBUG, "Graph noting %s %s -> %s.\n", v->gv_name, 4827 instance_state_str[v->gv_state], instance_state_str[state]); 4828 4829 old_state = v->gv_state; 4830 v->gv_state = state; 4831 4832 v->gv_reason = states->ps_reason; 4833 err = gt_transition(h, v, serr, old_state); 4834 if (err == 0 && v->gv_state != old_state) { 4835 dgraph_state_transition_notify(v, old_state, states->ps_reason); 4836 } 4837 4838 MUTEX_UNLOCK(&dgraph_lock); 4839 return (err); 4840 } 4841 4842 /* 4843 * Handle state changes during milestone shutdown. See 4844 * dgraph_set_milestone(). If the repository connection is broken, 4845 * ECONNABORTED will be returned, though a _DISABLE command will be sent for 4846 * the vertex anyway. 4847 */ 4848 int 4849 vertex_subgraph_dependencies_shutdown(scf_handle_t *h, graph_vertex_t *v, 4850 restarter_instance_state_t old_state) 4851 { 4852 int was_up, now_up; 4853 int ret = 0; 4854 4855 assert(v->gv_type == GVT_INST); 4856 4857 /* Don't care if we're not going to a milestone. */ 4858 if (milestone == NULL) 4859 return (0); 4860 4861 /* Don't care if we already finished coming down. */ 4862 if (non_subgraph_svcs == 0) 4863 return (0); 4864 4865 /* Don't care if the service is in the subgraph. */ 4866 if (v->gv_flags & GV_INSUBGRAPH) 4867 return (0); 4868 4869 /* 4870 * Update non_subgraph_svcs. It is the number of non-subgraph 4871 * services which are in online, degraded, or offline. 4872 */ 4873 4874 was_up = up_state(old_state); 4875 now_up = up_state(v->gv_state); 4876 4877 if (!was_up && now_up) { 4878 ++non_subgraph_svcs; 4879 } else if (was_up && !now_up) { 4880 --non_subgraph_svcs; 4881 4882 if (non_subgraph_svcs == 0) { 4883 if (halting != -1) { 4884 do_uadmin(); 4885 } else if (go_single_user_mode || go_to_level1) { 4886 (void) startd_thread_create(single_user_thread, 4887 NULL); 4888 } 4889 return (0); 4890 } 4891 } 4892 4893 /* If this service is a leaf, it should be disabled. */ 4894 if ((v->gv_flags & GV_ENABLED) && is_nonsubgraph_leaf(v)) { 4895 int r; 4896 4897 r = disable_service_temporarily(v, h); 4898 switch (r) { 4899 case 0: 4900 break; 4901 4902 case ECONNABORTED: 4903 ret = ECONNABORTED; 4904 break; 4905 4906 default: 4907 bad_error("disable_service_temporarily", r); 4908 } 4909 } 4910 4911 /* 4912 * If the service just came down, propagate the disable to the newly 4913 * exposed leaves. 4914 */ 4915 if (was_up && !now_up) 4916 graph_walk_dependencies(v, disable_nonsubgraph_leaves, 4917 (void *)h); 4918 4919 return (ret); 4920 } 4921 4922 /* 4923 * Decide whether to start up an sulogin thread after a service is 4924 * finished changing state. Only need to do the full can_come_up() 4925 * evaluation if an instance is changing state, we're not halfway through 4926 * loading the thread, and we aren't shutting down or going to the single 4927 * user milestone. 4928 */ 4929 void 4930 graph_transition_sulogin(restarter_instance_state_t state, 4931 restarter_instance_state_t old_state) 4932 { 4933 assert(MUTEX_HELD(&dgraph_lock)); 4934 4935 if (state != old_state && st->st_load_complete && 4936 !go_single_user_mode && !go_to_level1 && 4937 halting == -1) { 4938 if (!sulogin_thread_running && !can_come_up()) { 4939 (void) startd_thread_create(sulogin_thread, NULL); 4940 sulogin_thread_running = B_TRUE; 4941 } 4942 } 4943 } 4944 4945 /* 4946 * Propagate a start, stop event, or a satisfiability event. 4947 * 4948 * PROPAGATE_START and PROPAGATE_STOP simply propagate the transition event 4949 * to direct dependents. PROPAGATE_SAT propagates a start then walks the 4950 * full dependent graph to check for newly satisfied nodes. This is 4951 * necessary for cases when non-direct dependents may be effected but direct 4952 * dependents may not (e.g. for optional_all evaluations, see the 4953 * propagate_satbility() comments). 4954 * 4955 * PROPAGATE_SAT should be used whenever a non-running service moves into 4956 * a state which can satisfy optional dependencies, like disabled or 4957 * maintenance. 4958 */ 4959 void 4960 graph_transition_propagate(graph_vertex_t *v, propagate_event_t type, 4961 restarter_error_t rerr) 4962 { 4963 if (type == PROPAGATE_STOP) { 4964 graph_walk_dependents(v, propagate_stop, (void *)rerr); 4965 } else if (type == PROPAGATE_START || type == PROPAGATE_SAT) { 4966 graph_walk_dependents(v, propagate_start, NULL); 4967 4968 if (type == PROPAGATE_SAT) 4969 propagate_satbility(v); 4970 } else { 4971 #ifndef NDEBUG 4972 uu_warn("%s:%d: Unexpected type value %d.\n", __FILE__, 4973 __LINE__, type); 4974 #endif 4975 abort(); 4976 } 4977 } 4978 4979 /* 4980 * If a vertex for fmri exists and it is enabled, send _DISABLE to the 4981 * restarter. If it is running, send _STOP. Send _REMOVE_INSTANCE. Delete 4982 * all property group dependencies, and the dependency on the restarter, 4983 * disposing of vertices as appropriate. If other vertices depend on this 4984 * one, mark it unconfigured and return. Otherwise remove the vertex. Always 4985 * returns 0. 4986 */ 4987 static int 4988 dgraph_remove_instance(const char *fmri, scf_handle_t *h) 4989 { 4990 graph_vertex_t *v; 4991 graph_edge_t *e; 4992 uu_list_t *old_deps; 4993 int err; 4994 4995 log_framework(LOG_DEBUG, "Graph engine: Removing %s.\n", fmri); 4996 4997 MUTEX_LOCK(&dgraph_lock); 4998 4999 v = vertex_get_by_name(fmri); 5000 if (v == NULL) { 5001 MUTEX_UNLOCK(&dgraph_lock); 5002 return (0); 5003 } 5004 5005 /* Send restarter delete event. */ 5006 if (v->gv_flags & GV_CONFIGURED) 5007 graph_unset_restarter(v); 5008 5009 if (milestone > MILESTONE_NONE) { 5010 /* 5011 * Make a list of v's current dependencies so we can 5012 * reevaluate their GV_INSUBGRAPH flags after the dependencies 5013 * are removed. 5014 */ 5015 old_deps = startd_list_create(graph_edge_pool, NULL, 0); 5016 5017 err = uu_list_walk(v->gv_dependencies, 5018 (uu_walk_fn_t *)append_svcs_or_insts, old_deps, 0); 5019 assert(err == 0); 5020 } 5021 5022 delete_instance_dependencies(v, B_TRUE); 5023 5024 /* 5025 * Deleting an instance can both satisfy and unsatisfy dependencies, 5026 * depending on their type. First propagate the stop as a RERR_RESTART 5027 * event -- deletion isn't a fault, just a normal stop. This gives 5028 * dependent services the chance to do a clean shutdown. Then, mark 5029 * the service as unconfigured and propagate the start event for the 5030 * optional_all dependencies that might have become satisfied. 5031 */ 5032 graph_walk_dependents(v, propagate_stop, (void *)RERR_RESTART); 5033 5034 v->gv_flags &= ~GV_CONFIGURED; 5035 v->gv_flags &= ~GV_DEATHROW; 5036 5037 graph_walk_dependents(v, propagate_start, NULL); 5038 propagate_satbility(v); 5039 5040 /* 5041 * If there are no (non-service) dependents, the vertex can be 5042 * completely removed. 5043 */ 5044 if (v != milestone && v->gv_refs == 0 && 5045 uu_list_numnodes(v->gv_dependents) == 1) 5046 remove_inst_vertex(v); 5047 5048 if (milestone > MILESTONE_NONE) { 5049 void *cookie = NULL; 5050 5051 while ((e = uu_list_teardown(old_deps, &cookie)) != NULL) { 5052 v = e->ge_vertex; 5053 5054 if (vertex_unref(v) == VERTEX_INUSE) 5055 while (eval_subgraph(v, h) == ECONNABORTED) 5056 libscf_handle_rebind(h); 5057 5058 startd_free(e, sizeof (*e)); 5059 } 5060 5061 uu_list_destroy(old_deps); 5062 } 5063 5064 MUTEX_UNLOCK(&dgraph_lock); 5065 5066 return (0); 5067 } 5068 5069 /* 5070 * Return the eventual (maybe current) milestone in the form of a 5071 * legacy runlevel. 5072 */ 5073 static char 5074 target_milestone_as_runlevel() 5075 { 5076 assert(MUTEX_HELD(&dgraph_lock)); 5077 5078 if (milestone == NULL) 5079 return ('3'); 5080 else if (milestone == MILESTONE_NONE) 5081 return ('0'); 5082 5083 if (strcmp(milestone->gv_name, multi_user_fmri) == 0) 5084 return ('2'); 5085 else if (strcmp(milestone->gv_name, single_user_fmri) == 0) 5086 return ('S'); 5087 else if (strcmp(milestone->gv_name, multi_user_svr_fmri) == 0) 5088 return ('3'); 5089 5090 #ifndef NDEBUG 5091 (void) fprintf(stderr, "%s:%d: Unknown milestone name \"%s\".\n", 5092 __FILE__, __LINE__, milestone->gv_name); 5093 #endif 5094 abort(); 5095 /* NOTREACHED */ 5096 } 5097 5098 static struct { 5099 char rl; 5100 int sig; 5101 } init_sigs[] = { 5102 { 'S', SIGBUS }, 5103 { '0', SIGINT }, 5104 { '1', SIGQUIT }, 5105 { '2', SIGILL }, 5106 { '3', SIGTRAP }, 5107 { '4', SIGIOT }, 5108 { '5', SIGEMT }, 5109 { '6', SIGFPE }, 5110 { 0, 0 } 5111 }; 5112 5113 static void 5114 signal_init(char rl) 5115 { 5116 pid_t init_pid; 5117 int i; 5118 5119 assert(MUTEX_HELD(&dgraph_lock)); 5120 5121 if (zone_getattr(getzoneid(), ZONE_ATTR_INITPID, &init_pid, 5122 sizeof (init_pid)) != sizeof (init_pid)) { 5123 log_error(LOG_NOTICE, "Could not get pid to signal init.\n"); 5124 return; 5125 } 5126 5127 for (i = 0; init_sigs[i].rl != 0; ++i) 5128 if (init_sigs[i].rl == rl) 5129 break; 5130 5131 if (init_sigs[i].rl != 0) { 5132 if (kill(init_pid, init_sigs[i].sig) != 0) { 5133 switch (errno) { 5134 case EPERM: 5135 case ESRCH: 5136 log_error(LOG_NOTICE, "Could not signal init: " 5137 "%s.\n", strerror(errno)); 5138 break; 5139 5140 case EINVAL: 5141 default: 5142 bad_error("kill", errno); 5143 } 5144 } 5145 } 5146 } 5147 5148 /* 5149 * This is called when one of the major milestones changes state, or when 5150 * init is signalled and tells us it was told to change runlevel. We wait 5151 * to reach the milestone because this allows /etc/inittab entries to retain 5152 * some boot ordering: historically, entries could place themselves before/after 5153 * the running of /sbin/rcX scripts but we can no longer make the 5154 * distinction because the /sbin/rcX scripts no longer exist as punctuation 5155 * marks in /etc/inittab. 5156 * 5157 * Also, we only trigger an update when we reach the eventual target 5158 * milestone: without this, an /etc/inittab entry marked only for 5159 * runlevel 2 would be executed for runlevel 3, which is not how 5160 * /etc/inittab entries work. 5161 * 5162 * If we're single user coming online, then we set utmpx to the target 5163 * runlevel so that legacy scripts can work as expected. 5164 */ 5165 static void 5166 graph_runlevel_changed(char rl, int online) 5167 { 5168 char trl; 5169 5170 assert(MUTEX_HELD(&dgraph_lock)); 5171 5172 trl = target_milestone_as_runlevel(); 5173 5174 if (online) { 5175 if (rl == trl) { 5176 current_runlevel = trl; 5177 signal_init(trl); 5178 } else if (rl == 'S') { 5179 /* 5180 * At boot, set the entry early for the benefit of the 5181 * legacy init scripts. 5182 */ 5183 utmpx_set_runlevel(trl, 'S', B_FALSE); 5184 } 5185 } else { 5186 if (rl == '3' && trl == '2') { 5187 current_runlevel = trl; 5188 signal_init(trl); 5189 } else if (rl == '2' && trl == 'S') { 5190 current_runlevel = trl; 5191 signal_init(trl); 5192 } 5193 } 5194 } 5195 5196 /* 5197 * Move to a backwards-compatible runlevel by executing the appropriate 5198 * /etc/rc?.d/K* scripts and/or setting the milestone. 5199 * 5200 * Returns 5201 * 0 - success 5202 * ECONNRESET - success, but handle was reset 5203 * ECONNABORTED - repository connection broken 5204 * ECANCELED - pg was deleted 5205 */ 5206 static int 5207 dgraph_set_runlevel(scf_propertygroup_t *pg, scf_property_t *prop) 5208 { 5209 char rl; 5210 scf_handle_t *h; 5211 int r; 5212 const char *ms = NULL; /* what to commit as options/milestone */ 5213 boolean_t rebound = B_FALSE; 5214 int mark_rl = 0; 5215 5216 const char * const stop = "stop"; 5217 5218 r = libscf_extract_runlevel(prop, &rl); 5219 switch (r) { 5220 case 0: 5221 break; 5222 5223 case ECONNABORTED: 5224 case ECANCELED: 5225 return (r); 5226 5227 case EINVAL: 5228 case ENOENT: 5229 log_error(LOG_WARNING, "runlevel property is misconfigured; " 5230 "ignoring.\n"); 5231 /* delete the bad property */ 5232 goto nolock_out; 5233 5234 default: 5235 bad_error("libscf_extract_runlevel", r); 5236 } 5237 5238 switch (rl) { 5239 case 's': 5240 rl = 'S'; 5241 /* FALLTHROUGH */ 5242 5243 case 'S': 5244 case '2': 5245 case '3': 5246 /* 5247 * These cases cause a milestone change, so 5248 * graph_runlevel_changed() will eventually deal with 5249 * signalling init. 5250 */ 5251 break; 5252 5253 case '0': 5254 case '1': 5255 case '4': 5256 case '5': 5257 case '6': 5258 mark_rl = 1; 5259 break; 5260 5261 default: 5262 log_framework(LOG_NOTICE, "Unknown runlevel '%c'.\n", rl); 5263 ms = NULL; 5264 goto nolock_out; 5265 } 5266 5267 h = scf_pg_handle(pg); 5268 5269 MUTEX_LOCK(&dgraph_lock); 5270 5271 /* 5272 * Since this triggers no milestone changes, force it by hand. 5273 */ 5274 if (current_runlevel == '4' && rl == '3') 5275 mark_rl = 1; 5276 5277 /* 5278 * 1. If we are here after an "init X": 5279 * 5280 * init X 5281 * init/lscf_set_runlevel() 5282 * process_pg_event() 5283 * dgraph_set_runlevel() 5284 * 5285 * then we haven't passed through graph_runlevel_changed() yet, 5286 * therefore 'current_runlevel' has not changed for sure but 'rl' has. 5287 * In consequence, if 'rl' is lower than 'current_runlevel', we change 5288 * the system runlevel and execute the appropriate /etc/rc?.d/K* scripts 5289 * past this test. 5290 * 5291 * 2. On the other hand, if we are here after a "svcadm milestone": 5292 * 5293 * svcadm milestone X 5294 * dgraph_set_milestone() 5295 * handle_graph_update_event() 5296 * dgraph_set_instance_state() 5297 * graph_post_X_[online|offline]() 5298 * graph_runlevel_changed() 5299 * signal_init() 5300 * init/lscf_set_runlevel() 5301 * process_pg_event() 5302 * dgraph_set_runlevel() 5303 * 5304 * then we already passed through graph_runlevel_changed() (by the way 5305 * of dgraph_set_milestone()) and 'current_runlevel' may have changed 5306 * and already be equal to 'rl' so we are going to return immediately 5307 * from dgraph_set_runlevel() without changing the system runlevel and 5308 * without executing the /etc/rc?.d/K* scripts. 5309 */ 5310 if (rl == current_runlevel) { 5311 ms = NULL; 5312 goto out; 5313 } 5314 5315 log_framework(LOG_DEBUG, "Changing to runlevel '%c'.\n", rl); 5316 5317 /* 5318 * Make sure stop rc scripts see the new settings via who -r. 5319 */ 5320 utmpx_set_runlevel(rl, current_runlevel, B_TRUE); 5321 5322 /* 5323 * Some run levels don't have a direct correspondence to any 5324 * milestones, so we have to signal init directly. 5325 */ 5326 if (mark_rl) { 5327 current_runlevel = rl; 5328 signal_init(rl); 5329 } 5330 5331 switch (rl) { 5332 case 'S': 5333 uu_warn("The system is coming down for administration. " 5334 "Please wait.\n"); 5335 fork_rc_script(rl, stop, B_FALSE); 5336 ms = single_user_fmri; 5337 go_single_user_mode = B_TRUE; 5338 break; 5339 5340 case '0': 5341 halting_time = time(NULL); 5342 fork_rc_script(rl, stop, B_TRUE); 5343 halting = AD_HALT; 5344 goto uadmin; 5345 5346 case '5': 5347 halting_time = time(NULL); 5348 fork_rc_script(rl, stop, B_TRUE); 5349 halting = AD_POWEROFF; 5350 goto uadmin; 5351 5352 case '6': 5353 halting_time = time(NULL); 5354 fork_rc_script(rl, stop, B_TRUE); 5355 if (scf_is_fastboot_default() && getzoneid() == GLOBAL_ZONEID) 5356 halting = AD_FASTREBOOT; 5357 else 5358 halting = AD_BOOT; 5359 5360 uadmin: 5361 uu_warn("The system is coming down. Please wait.\n"); 5362 ms = "none"; 5363 5364 /* 5365 * We can't wait until all services are offline since this 5366 * thread is responsible for taking them offline. Instead we 5367 * set halting to the second argument for uadmin() and call 5368 * do_uadmin() from dgraph_set_instance_state() when 5369 * appropriate. 5370 */ 5371 break; 5372 5373 case '1': 5374 if (current_runlevel != 'S') { 5375 uu_warn("Changing to state 1.\n"); 5376 fork_rc_script(rl, stop, B_FALSE); 5377 } else { 5378 uu_warn("The system is coming up for administration. " 5379 "Please wait.\n"); 5380 } 5381 ms = single_user_fmri; 5382 go_to_level1 = B_TRUE; 5383 break; 5384 5385 case '2': 5386 if (current_runlevel == '3' || current_runlevel == '4') 5387 fork_rc_script(rl, stop, B_FALSE); 5388 ms = multi_user_fmri; 5389 break; 5390 5391 case '3': 5392 case '4': 5393 ms = "all"; 5394 break; 5395 5396 default: 5397 #ifndef NDEBUG 5398 (void) fprintf(stderr, "%s:%d: Uncaught case %d ('%c').\n", 5399 __FILE__, __LINE__, rl, rl); 5400 #endif 5401 abort(); 5402 } 5403 5404 out: 5405 MUTEX_UNLOCK(&dgraph_lock); 5406 5407 nolock_out: 5408 switch (r = libscf_clear_runlevel(pg, ms)) { 5409 case 0: 5410 break; 5411 5412 case ECONNABORTED: 5413 libscf_handle_rebind(h); 5414 rebound = B_TRUE; 5415 goto nolock_out; 5416 5417 case ECANCELED: 5418 break; 5419 5420 case EPERM: 5421 case EACCES: 5422 case EROFS: 5423 log_error(LOG_NOTICE, "Could not delete \"%s/%s\" property: " 5424 "%s.\n", SCF_PG_OPTIONS, "runlevel", strerror(r)); 5425 break; 5426 5427 default: 5428 bad_error("libscf_clear_runlevel", r); 5429 } 5430 5431 return (rebound ? ECONNRESET : 0); 5432 } 5433 5434 /* 5435 * mark_subtree walks the dependents and add the GV_TOOFFLINE flag 5436 * to the instances that are supposed to go offline during an 5437 * administrative disable operation. 5438 */ 5439 static int 5440 mark_subtree(graph_edge_t *e, void *arg) 5441 { 5442 graph_vertex_t *v; 5443 int r; 5444 5445 v = e->ge_vertex; 5446 5447 /* If it's already in the subgraph, skip. */ 5448 if (v->gv_flags & GV_TOOFFLINE) 5449 return (UU_WALK_NEXT); 5450 5451 switch (v->gv_type) { 5452 case GVT_INST: 5453 /* If the instance is already disabled, skip it. */ 5454 if (!(v->gv_flags & GV_ENABLED)) 5455 return (UU_WALK_NEXT); 5456 5457 v->gv_flags |= GV_TOOFFLINE; 5458 log_framework(LOG_DEBUG, "%s added to subtree\n", v->gv_name); 5459 break; 5460 case GVT_GROUP: 5461 /* 5462 * Skip all excluded dependencies and decide whether to offline 5463 * the service based on the restart_on attribute. 5464 */ 5465 if (is_depgrp_bypassed(v)) 5466 return (UU_WALK_NEXT); 5467 break; 5468 } 5469 5470 r = uu_list_walk(v->gv_dependents, (uu_walk_fn_t *)mark_subtree, arg, 5471 0); 5472 assert(r == 0); 5473 return (UU_WALK_NEXT); 5474 } 5475 5476 static int 5477 mark_subgraph(graph_edge_t *e, void *arg) 5478 { 5479 graph_vertex_t *v; 5480 int r; 5481 int optional = (int)arg; 5482 5483 v = e->ge_vertex; 5484 5485 /* If it's already in the subgraph, skip. */ 5486 if (v->gv_flags & GV_INSUBGRAPH) 5487 return (UU_WALK_NEXT); 5488 5489 /* 5490 * Keep track if walk has entered an optional dependency group 5491 */ 5492 if (v->gv_type == GVT_GROUP && v->gv_depgroup == DEPGRP_OPTIONAL_ALL) { 5493 optional = 1; 5494 } 5495 /* 5496 * Quit if we are in an optional dependency group and the instance 5497 * is disabled 5498 */ 5499 if (optional && (v->gv_type == GVT_INST) && 5500 (!(v->gv_flags & GV_ENBLD_NOOVR))) 5501 return (UU_WALK_NEXT); 5502 5503 v->gv_flags |= GV_INSUBGRAPH; 5504 5505 /* Skip all excluded dependencies. */ 5506 if (v->gv_type == GVT_GROUP && v->gv_depgroup == DEPGRP_EXCLUDE_ALL) 5507 return (UU_WALK_NEXT); 5508 5509 r = uu_list_walk(v->gv_dependencies, (uu_walk_fn_t *)mark_subgraph, 5510 (void *)optional, 0); 5511 assert(r == 0); 5512 return (UU_WALK_NEXT); 5513 } 5514 5515 /* 5516 * Bring down all services which are not dependencies of fmri. The 5517 * dependencies of fmri (direct & indirect) will constitute the "subgraph", 5518 * and will have the GV_INSUBGRAPH flag set. The rest must be brought down, 5519 * which means the state is "disabled", "maintenance", or "uninitialized". We 5520 * could consider "offline" to be down, and refrain from sending start 5521 * commands for such services, but that's not strictly necessary, so we'll 5522 * decline to intrude on the state machine. It would probably confuse users 5523 * anyway. 5524 * 5525 * The services should be brought down in reverse-dependency order, so we 5526 * can't do it all at once here. We initiate by override-disabling the leaves 5527 * of the dependency tree -- those services which are up but have no 5528 * dependents which are up. When they come down, 5529 * vertex_subgraph_dependencies_shutdown() will override-disable the newly 5530 * exposed leaves. Perseverance will ensure completion. 5531 * 5532 * Sometimes we need to take action when the transition is complete, like 5533 * start sulogin or halt the system. To tell when we're done, we initialize 5534 * non_subgraph_svcs here to be the number of services which need to come 5535 * down. As each does, we decrement the counter. When it hits zero, we take 5536 * the appropriate action. See vertex_subgraph_dependencies_shutdown(). 5537 * 5538 * In case we're coming up, we also remove any enable-overrides for the 5539 * services which are dependencies of fmri. 5540 * 5541 * If norepository is true, the function will not change the repository. 5542 * 5543 * The decision to change the system run level in accordance with the milestone 5544 * is taken in dgraph_set_runlevel(). 5545 * 5546 * Returns 5547 * 0 - success 5548 * ECONNRESET - success, but handle was rebound 5549 * EINVAL - fmri is invalid (error is logged) 5550 * EALREADY - the milestone is already set to fmri 5551 * ENOENT - a configured vertex does not exist for fmri (an error is logged) 5552 */ 5553 static int 5554 dgraph_set_milestone(const char *fmri, scf_handle_t *h, boolean_t norepository) 5555 { 5556 const char *cfmri, *fs; 5557 graph_vertex_t *nm, *v; 5558 int ret = 0, r; 5559 scf_instance_t *inst; 5560 boolean_t isall, isnone, rebound = B_FALSE; 5561 5562 /* Validate fmri */ 5563 isall = (strcmp(fmri, "all") == 0); 5564 isnone = (strcmp(fmri, "none") == 0); 5565 5566 if (!isall && !isnone) { 5567 if (fmri_canonify(fmri, (char **)&cfmri, B_FALSE) == EINVAL) 5568 goto reject; 5569 5570 if (strcmp(cfmri, single_user_fmri) != 0 && 5571 strcmp(cfmri, multi_user_fmri) != 0 && 5572 strcmp(cfmri, multi_user_svr_fmri) != 0) { 5573 startd_free((void *)cfmri, max_scf_fmri_size); 5574 reject: 5575 log_framework(LOG_WARNING, 5576 "Rejecting request for invalid milestone \"%s\".\n", 5577 fmri); 5578 return (EINVAL); 5579 } 5580 } 5581 5582 inst = safe_scf_instance_create(h); 5583 5584 MUTEX_LOCK(&dgraph_lock); 5585 5586 if (milestone == NULL) { 5587 if (isall) { 5588 log_framework(LOG_DEBUG, 5589 "Milestone already set to all.\n"); 5590 ret = EALREADY; 5591 goto out; 5592 } 5593 } else if (milestone == MILESTONE_NONE) { 5594 if (isnone) { 5595 log_framework(LOG_DEBUG, 5596 "Milestone already set to none.\n"); 5597 ret = EALREADY; 5598 goto out; 5599 } 5600 } else { 5601 if (!isall && !isnone && 5602 strcmp(cfmri, milestone->gv_name) == 0) { 5603 log_framework(LOG_DEBUG, 5604 "Milestone already set to %s.\n", cfmri); 5605 ret = EALREADY; 5606 goto out; 5607 } 5608 } 5609 5610 if (!isall && !isnone) { 5611 nm = vertex_get_by_name(cfmri); 5612 if (nm == NULL || !(nm->gv_flags & GV_CONFIGURED)) { 5613 log_framework(LOG_WARNING, "Cannot set milestone to %s " 5614 "because no such service exists.\n", cfmri); 5615 ret = ENOENT; 5616 goto out; 5617 } 5618 } 5619 5620 log_framework(LOG_DEBUG, "Changing milestone to %s.\n", fmri); 5621 5622 /* 5623 * Set milestone, removing the old one if this was the last reference. 5624 */ 5625 if (milestone > MILESTONE_NONE) 5626 (void) vertex_unref(milestone); 5627 5628 if (isall) 5629 milestone = NULL; 5630 else if (isnone) 5631 milestone = MILESTONE_NONE; 5632 else { 5633 milestone = nm; 5634 /* milestone should count as a reference */ 5635 vertex_ref(milestone); 5636 } 5637 5638 /* Clear all GV_INSUBGRAPH bits. */ 5639 for (v = uu_list_first(dgraph); v != NULL; v = uu_list_next(dgraph, v)) 5640 v->gv_flags &= ~GV_INSUBGRAPH; 5641 5642 if (!isall && !isnone) { 5643 /* Set GV_INSUBGRAPH for milestone & descendents. */ 5644 milestone->gv_flags |= GV_INSUBGRAPH; 5645 5646 r = uu_list_walk(milestone->gv_dependencies, 5647 (uu_walk_fn_t *)mark_subgraph, NULL, 0); 5648 assert(r == 0); 5649 } 5650 5651 /* Un-override services in the subgraph & override-disable the rest. */ 5652 if (norepository) 5653 goto out; 5654 5655 non_subgraph_svcs = 0; 5656 for (v = uu_list_first(dgraph); 5657 v != NULL; 5658 v = uu_list_next(dgraph, v)) { 5659 if (v->gv_type != GVT_INST || 5660 (v->gv_flags & GV_CONFIGURED) == 0) 5661 continue; 5662 5663 again: 5664 r = scf_handle_decode_fmri(h, v->gv_name, NULL, NULL, inst, 5665 NULL, NULL, SCF_DECODE_FMRI_EXACT); 5666 if (r != 0) { 5667 switch (scf_error()) { 5668 case SCF_ERROR_CONNECTION_BROKEN: 5669 default: 5670 libscf_handle_rebind(h); 5671 rebound = B_TRUE; 5672 goto again; 5673 5674 case SCF_ERROR_NOT_FOUND: 5675 continue; 5676 5677 case SCF_ERROR_HANDLE_MISMATCH: 5678 case SCF_ERROR_INVALID_ARGUMENT: 5679 case SCF_ERROR_CONSTRAINT_VIOLATED: 5680 case SCF_ERROR_NOT_BOUND: 5681 bad_error("scf_handle_decode_fmri", 5682 scf_error()); 5683 } 5684 } 5685 5686 if (isall || (v->gv_flags & GV_INSUBGRAPH)) { 5687 r = libscf_delete_enable_ovr(inst); 5688 fs = "libscf_delete_enable_ovr"; 5689 } else { 5690 assert(isnone || (v->gv_flags & GV_INSUBGRAPH) == 0); 5691 5692 /* 5693 * Services which are up need to come down before 5694 * we're done, but we can only disable the leaves 5695 * here. 5696 */ 5697 5698 if (up_state(v->gv_state)) 5699 ++non_subgraph_svcs; 5700 5701 /* If it's already disabled, don't bother. */ 5702 if ((v->gv_flags & GV_ENABLED) == 0) 5703 continue; 5704 5705 if (!is_nonsubgraph_leaf(v)) 5706 continue; 5707 5708 r = libscf_set_enable_ovr(inst, 0); 5709 fs = "libscf_set_enable_ovr"; 5710 } 5711 switch (r) { 5712 case 0: 5713 case ECANCELED: 5714 break; 5715 5716 case ECONNABORTED: 5717 libscf_handle_rebind(h); 5718 rebound = B_TRUE; 5719 goto again; 5720 5721 case EPERM: 5722 case EROFS: 5723 log_error(LOG_WARNING, 5724 "Could not set %s/%s for %s: %s.\n", 5725 SCF_PG_GENERAL_OVR, SCF_PROPERTY_ENABLED, 5726 v->gv_name, strerror(r)); 5727 break; 5728 5729 default: 5730 bad_error(fs, r); 5731 } 5732 } 5733 5734 if (halting != -1) { 5735 if (non_subgraph_svcs > 1) 5736 uu_warn("%d system services are now being stopped.\n", 5737 non_subgraph_svcs); 5738 else if (non_subgraph_svcs == 1) 5739 uu_warn("One system service is now being stopped.\n"); 5740 else if (non_subgraph_svcs == 0) 5741 do_uadmin(); 5742 } 5743 5744 ret = rebound ? ECONNRESET : 0; 5745 5746 out: 5747 MUTEX_UNLOCK(&dgraph_lock); 5748 if (!isall && !isnone) 5749 startd_free((void *)cfmri, max_scf_fmri_size); 5750 scf_instance_destroy(inst); 5751 return (ret); 5752 } 5753 5754 5755 /* 5756 * Returns 0, ECONNABORTED, or EINVAL. 5757 */ 5758 static int 5759 handle_graph_update_event(scf_handle_t *h, graph_protocol_event_t *e) 5760 { 5761 int r; 5762 5763 switch (e->gpe_type) { 5764 case GRAPH_UPDATE_RELOAD_GRAPH: 5765 log_error(LOG_WARNING, 5766 "graph_event: reload graph unimplemented\n"); 5767 break; 5768 5769 case GRAPH_UPDATE_STATE_CHANGE: { 5770 protocol_states_t *states = e->gpe_data; 5771 5772 switch (r = dgraph_set_instance_state(h, e->gpe_inst, states)) { 5773 case 0: 5774 case ENOENT: 5775 break; 5776 5777 case ECONNABORTED: 5778 return (ECONNABORTED); 5779 5780 case EINVAL: 5781 default: 5782 #ifndef NDEBUG 5783 (void) fprintf(stderr, "dgraph_set_instance_state() " 5784 "failed with unexpected error %d at %s:%d.\n", r, 5785 __FILE__, __LINE__); 5786 #endif 5787 abort(); 5788 } 5789 5790 startd_free(states, sizeof (protocol_states_t)); 5791 break; 5792 } 5793 5794 default: 5795 log_error(LOG_WARNING, 5796 "graph_event_loop received an unknown event: %d\n", 5797 e->gpe_type); 5798 break; 5799 } 5800 5801 return (0); 5802 } 5803 5804 /* 5805 * graph_event_thread() 5806 * Wait for state changes from the restarters. 5807 */ 5808 /*ARGSUSED*/ 5809 void * 5810 graph_event_thread(void *unused) 5811 { 5812 scf_handle_t *h; 5813 int err; 5814 5815 h = libscf_handle_create_bound_loop(); 5816 5817 /*CONSTCOND*/ 5818 while (1) { 5819 graph_protocol_event_t *e; 5820 5821 MUTEX_LOCK(&gu->gu_lock); 5822 5823 while (gu->gu_wakeup == 0) 5824 (void) pthread_cond_wait(&gu->gu_cv, &gu->gu_lock); 5825 5826 gu->gu_wakeup = 0; 5827 5828 while ((e = graph_event_dequeue()) != NULL) { 5829 MUTEX_LOCK(&e->gpe_lock); 5830 MUTEX_UNLOCK(&gu->gu_lock); 5831 5832 while ((err = handle_graph_update_event(h, e)) == 5833 ECONNABORTED) 5834 libscf_handle_rebind(h); 5835 5836 if (err == 0) 5837 graph_event_release(e); 5838 else 5839 graph_event_requeue(e); 5840 5841 MUTEX_LOCK(&gu->gu_lock); 5842 } 5843 5844 MUTEX_UNLOCK(&gu->gu_lock); 5845 } 5846 5847 /* 5848 * Unreachable for now -- there's currently no graceful cleanup 5849 * called on exit(). 5850 */ 5851 MUTEX_UNLOCK(&gu->gu_lock); 5852 scf_handle_destroy(h); 5853 return (NULL); 5854 } 5855 5856 static void 5857 set_initial_milestone(scf_handle_t *h) 5858 { 5859 scf_instance_t *inst; 5860 char *fmri, *cfmri; 5861 size_t sz; 5862 int r; 5863 5864 inst = safe_scf_instance_create(h); 5865 fmri = startd_alloc(max_scf_fmri_size); 5866 5867 /* 5868 * If -m milestone= was specified, we want to set options_ovr/milestone 5869 * to it. Otherwise we want to read what the milestone should be set 5870 * to. Either way we need our inst. 5871 */ 5872 get_self: 5873 if (scf_handle_decode_fmri(h, SCF_SERVICE_STARTD, NULL, NULL, inst, 5874 NULL, NULL, SCF_DECODE_FMRI_EXACT) != 0) { 5875 switch (scf_error()) { 5876 case SCF_ERROR_CONNECTION_BROKEN: 5877 libscf_handle_rebind(h); 5878 goto get_self; 5879 5880 case SCF_ERROR_NOT_FOUND: 5881 if (st->st_subgraph != NULL && 5882 st->st_subgraph[0] != '\0') { 5883 sz = strlcpy(fmri, st->st_subgraph, 5884 max_scf_fmri_size); 5885 assert(sz < max_scf_fmri_size); 5886 } else { 5887 fmri[0] = '\0'; 5888 } 5889 break; 5890 5891 case SCF_ERROR_INVALID_ARGUMENT: 5892 case SCF_ERROR_CONSTRAINT_VIOLATED: 5893 case SCF_ERROR_HANDLE_MISMATCH: 5894 default: 5895 bad_error("scf_handle_decode_fmri", scf_error()); 5896 } 5897 } else { 5898 if (st->st_subgraph != NULL && st->st_subgraph[0] != '\0') { 5899 scf_propertygroup_t *pg; 5900 5901 pg = safe_scf_pg_create(h); 5902 5903 sz = strlcpy(fmri, st->st_subgraph, max_scf_fmri_size); 5904 assert(sz < max_scf_fmri_size); 5905 5906 r = libscf_inst_get_or_add_pg(inst, SCF_PG_OPTIONS_OVR, 5907 SCF_PG_OPTIONS_OVR_TYPE, SCF_PG_OPTIONS_OVR_FLAGS, 5908 pg); 5909 switch (r) { 5910 case 0: 5911 break; 5912 5913 case ECONNABORTED: 5914 libscf_handle_rebind(h); 5915 goto get_self; 5916 5917 case EPERM: 5918 case EACCES: 5919 case EROFS: 5920 log_error(LOG_WARNING, "Could not set %s/%s: " 5921 "%s.\n", SCF_PG_OPTIONS_OVR, 5922 SCF_PROPERTY_MILESTONE, strerror(r)); 5923 /* FALLTHROUGH */ 5924 5925 case ECANCELED: 5926 sz = strlcpy(fmri, st->st_subgraph, 5927 max_scf_fmri_size); 5928 assert(sz < max_scf_fmri_size); 5929 break; 5930 5931 default: 5932 bad_error("libscf_inst_get_or_add_pg", r); 5933 } 5934 5935 r = libscf_clear_runlevel(pg, fmri); 5936 switch (r) { 5937 case 0: 5938 break; 5939 5940 case ECONNABORTED: 5941 libscf_handle_rebind(h); 5942 goto get_self; 5943 5944 case EPERM: 5945 case EACCES: 5946 case EROFS: 5947 log_error(LOG_WARNING, "Could not set %s/%s: " 5948 "%s.\n", SCF_PG_OPTIONS_OVR, 5949 SCF_PROPERTY_MILESTONE, strerror(r)); 5950 /* FALLTHROUGH */ 5951 5952 case ECANCELED: 5953 sz = strlcpy(fmri, st->st_subgraph, 5954 max_scf_fmri_size); 5955 assert(sz < max_scf_fmri_size); 5956 break; 5957 5958 default: 5959 bad_error("libscf_clear_runlevel", r); 5960 } 5961 5962 scf_pg_destroy(pg); 5963 } else { 5964 scf_property_t *prop; 5965 scf_value_t *val; 5966 5967 prop = safe_scf_property_create(h); 5968 val = safe_scf_value_create(h); 5969 5970 r = libscf_get_milestone(inst, prop, val, fmri, 5971 max_scf_fmri_size); 5972 switch (r) { 5973 case 0: 5974 break; 5975 5976 case ECONNABORTED: 5977 libscf_handle_rebind(h); 5978 goto get_self; 5979 5980 case EINVAL: 5981 log_error(LOG_WARNING, "Milestone property is " 5982 "misconfigured. Defaulting to \"all\".\n"); 5983 /* FALLTHROUGH */ 5984 5985 case ECANCELED: 5986 case ENOENT: 5987 fmri[0] = '\0'; 5988 break; 5989 5990 default: 5991 bad_error("libscf_get_milestone", r); 5992 } 5993 5994 scf_value_destroy(val); 5995 scf_property_destroy(prop); 5996 } 5997 } 5998 5999 if (fmri[0] == '\0' || strcmp(fmri, "all") == 0) 6000 goto out; 6001 6002 if (strcmp(fmri, "none") != 0) { 6003 retry: 6004 if (scf_handle_decode_fmri(h, fmri, NULL, NULL, inst, NULL, 6005 NULL, SCF_DECODE_FMRI_EXACT) != 0) { 6006 switch (scf_error()) { 6007 case SCF_ERROR_INVALID_ARGUMENT: 6008 log_error(LOG_WARNING, 6009 "Requested milestone \"%s\" is invalid. " 6010 "Reverting to \"all\".\n", fmri); 6011 goto out; 6012 6013 case SCF_ERROR_CONSTRAINT_VIOLATED: 6014 log_error(LOG_WARNING, "Requested milestone " 6015 "\"%s\" does not specify an instance. " 6016 "Reverting to \"all\".\n", fmri); 6017 goto out; 6018 6019 case SCF_ERROR_CONNECTION_BROKEN: 6020 libscf_handle_rebind(h); 6021 goto retry; 6022 6023 case SCF_ERROR_NOT_FOUND: 6024 log_error(LOG_WARNING, "Requested milestone " 6025 "\"%s\" not in repository. Reverting to " 6026 "\"all\".\n", fmri); 6027 goto out; 6028 6029 case SCF_ERROR_HANDLE_MISMATCH: 6030 default: 6031 bad_error("scf_handle_decode_fmri", 6032 scf_error()); 6033 } 6034 } 6035 6036 r = fmri_canonify(fmri, &cfmri, B_FALSE); 6037 assert(r == 0); 6038 6039 r = dgraph_add_instance(cfmri, inst, B_TRUE); 6040 startd_free(cfmri, max_scf_fmri_size); 6041 switch (r) { 6042 case 0: 6043 break; 6044 6045 case ECONNABORTED: 6046 goto retry; 6047 6048 case EINVAL: 6049 log_error(LOG_WARNING, 6050 "Requested milestone \"%s\" is invalid. " 6051 "Reverting to \"all\".\n", fmri); 6052 goto out; 6053 6054 case ECANCELED: 6055 log_error(LOG_WARNING, 6056 "Requested milestone \"%s\" not " 6057 "in repository. Reverting to \"all\".\n", 6058 fmri); 6059 goto out; 6060 6061 case EEXIST: 6062 default: 6063 bad_error("dgraph_add_instance", r); 6064 } 6065 } 6066 6067 log_console(LOG_INFO, "Booting to milestone \"%s\".\n", fmri); 6068 6069 r = dgraph_set_milestone(fmri, h, B_FALSE); 6070 switch (r) { 6071 case 0: 6072 case ECONNRESET: 6073 case EALREADY: 6074 break; 6075 6076 case EINVAL: 6077 case ENOENT: 6078 default: 6079 bad_error("dgraph_set_milestone", r); 6080 } 6081 6082 out: 6083 startd_free(fmri, max_scf_fmri_size); 6084 scf_instance_destroy(inst); 6085 } 6086 6087 void 6088 set_restart_milestone(scf_handle_t *h) 6089 { 6090 scf_instance_t *inst; 6091 scf_property_t *prop; 6092 scf_value_t *val; 6093 char *fmri; 6094 int r; 6095 6096 inst = safe_scf_instance_create(h); 6097 6098 get_self: 6099 if (scf_handle_decode_fmri(h, SCF_SERVICE_STARTD, NULL, NULL, 6100 inst, NULL, NULL, SCF_DECODE_FMRI_EXACT) != 0) { 6101 switch (scf_error()) { 6102 case SCF_ERROR_CONNECTION_BROKEN: 6103 libscf_handle_rebind(h); 6104 goto get_self; 6105 6106 case SCF_ERROR_NOT_FOUND: 6107 break; 6108 6109 case SCF_ERROR_INVALID_ARGUMENT: 6110 case SCF_ERROR_CONSTRAINT_VIOLATED: 6111 case SCF_ERROR_HANDLE_MISMATCH: 6112 default: 6113 bad_error("scf_handle_decode_fmri", scf_error()); 6114 } 6115 6116 scf_instance_destroy(inst); 6117 return; 6118 } 6119 6120 prop = safe_scf_property_create(h); 6121 val = safe_scf_value_create(h); 6122 fmri = startd_alloc(max_scf_fmri_size); 6123 6124 r = libscf_get_milestone(inst, prop, val, fmri, max_scf_fmri_size); 6125 switch (r) { 6126 case 0: 6127 break; 6128 6129 case ECONNABORTED: 6130 libscf_handle_rebind(h); 6131 goto get_self; 6132 6133 case ECANCELED: 6134 case ENOENT: 6135 case EINVAL: 6136 goto out; 6137 6138 default: 6139 bad_error("libscf_get_milestone", r); 6140 } 6141 6142 r = dgraph_set_milestone(fmri, h, B_TRUE); 6143 switch (r) { 6144 case 0: 6145 case ECONNRESET: 6146 case EALREADY: 6147 case EINVAL: 6148 case ENOENT: 6149 break; 6150 6151 default: 6152 bad_error("dgraph_set_milestone", r); 6153 } 6154 6155 out: 6156 startd_free(fmri, max_scf_fmri_size); 6157 scf_value_destroy(val); 6158 scf_property_destroy(prop); 6159 scf_instance_destroy(inst); 6160 } 6161 6162 /* 6163 * void *graph_thread(void *) 6164 * 6165 * Graph management thread. 6166 */ 6167 /*ARGSUSED*/ 6168 void * 6169 graph_thread(void *arg) 6170 { 6171 scf_handle_t *h; 6172 int err; 6173 6174 h = libscf_handle_create_bound_loop(); 6175 6176 if (st->st_initial) 6177 set_initial_milestone(h); 6178 6179 MUTEX_LOCK(&dgraph_lock); 6180 initial_milestone_set = B_TRUE; 6181 err = pthread_cond_broadcast(&initial_milestone_cv); 6182 assert(err == 0); 6183 MUTEX_UNLOCK(&dgraph_lock); 6184 6185 libscf_populate_graph(h); 6186 6187 if (!st->st_initial) 6188 set_restart_milestone(h); 6189 6190 MUTEX_LOCK(&st->st_load_lock); 6191 st->st_load_complete = 1; 6192 (void) pthread_cond_broadcast(&st->st_load_cv); 6193 MUTEX_UNLOCK(&st->st_load_lock); 6194 6195 MUTEX_LOCK(&dgraph_lock); 6196 /* 6197 * Now that we've set st_load_complete we need to check can_come_up() 6198 * since if we booted to a milestone, then there won't be any more 6199 * state updates. 6200 */ 6201 if (!go_single_user_mode && !go_to_level1 && 6202 halting == -1) { 6203 if (!sulogin_thread_running && !can_come_up()) { 6204 (void) startd_thread_create(sulogin_thread, NULL); 6205 sulogin_thread_running = B_TRUE; 6206 } 6207 } 6208 MUTEX_UNLOCK(&dgraph_lock); 6209 6210 (void) pthread_mutex_lock(&gu->gu_freeze_lock); 6211 6212 /*CONSTCOND*/ 6213 while (1) { 6214 (void) pthread_cond_wait(&gu->gu_freeze_cv, 6215 &gu->gu_freeze_lock); 6216 } 6217 6218 /* 6219 * Unreachable for now -- there's currently no graceful cleanup 6220 * called on exit(). 6221 */ 6222 (void) pthread_mutex_unlock(&gu->gu_freeze_lock); 6223 scf_handle_destroy(h); 6224 6225 return (NULL); 6226 } 6227 6228 6229 /* 6230 * int next_action() 6231 * Given an array of timestamps 'a' with 'num' elements, find the 6232 * lowest non-zero timestamp and return its index. If there are no 6233 * non-zero elements, return -1. 6234 */ 6235 static int 6236 next_action(hrtime_t *a, int num) 6237 { 6238 hrtime_t t = 0; 6239 int i = 0, smallest = -1; 6240 6241 for (i = 0; i < num; i++) { 6242 if (t == 0) { 6243 t = a[i]; 6244 smallest = i; 6245 } else if (a[i] != 0 && a[i] < t) { 6246 t = a[i]; 6247 smallest = i; 6248 } 6249 } 6250 6251 if (t == 0) 6252 return (-1); 6253 else 6254 return (smallest); 6255 } 6256 6257 /* 6258 * void process_actions() 6259 * Process actions requested by the administrator. Possibilities include: 6260 * refresh, restart, maintenance mode off, maintenance mode on, 6261 * maintenance mode immediate, and degraded. 6262 * 6263 * The set of pending actions is represented in the repository as a 6264 * per-instance property group, with each action being a single property 6265 * in that group. This property group is converted to an array, with each 6266 * action type having an array slot. The actions in the array at the 6267 * time process_actions() is called are acted on in the order of the 6268 * timestamp (which is the value stored in the slot). A value of zero 6269 * indicates that there is no pending action of the type associated with 6270 * a particular slot. 6271 * 6272 * Sending an action event multiple times before the restarter has a 6273 * chance to process that action will force it to be run at the last 6274 * timestamp where it appears in the ordering. 6275 * 6276 * Turning maintenance mode on trumps all other actions. 6277 * 6278 * Returns 0 or ECONNABORTED. 6279 */ 6280 static int 6281 process_actions(scf_handle_t *h, scf_propertygroup_t *pg, scf_instance_t *inst) 6282 { 6283 scf_property_t *prop = NULL; 6284 scf_value_t *val = NULL; 6285 scf_type_t type; 6286 graph_vertex_t *vertex; 6287 admin_action_t a; 6288 int i, ret = 0, r; 6289 hrtime_t action_ts[NACTIONS]; 6290 char *inst_name; 6291 6292 r = libscf_instance_get_fmri(inst, &inst_name); 6293 switch (r) { 6294 case 0: 6295 break; 6296 6297 case ECONNABORTED: 6298 return (ECONNABORTED); 6299 6300 case ECANCELED: 6301 return (0); 6302 6303 default: 6304 bad_error("libscf_instance_get_fmri", r); 6305 } 6306 6307 MUTEX_LOCK(&dgraph_lock); 6308 6309 vertex = vertex_get_by_name(inst_name); 6310 if (vertex == NULL) { 6311 MUTEX_UNLOCK(&dgraph_lock); 6312 log_framework(LOG_DEBUG, "%s: Can't find graph vertex. " 6313 "The instance must have been removed.\n", inst_name); 6314 startd_free(inst_name, max_scf_fmri_size); 6315 return (0); 6316 } 6317 6318 prop = safe_scf_property_create(h); 6319 val = safe_scf_value_create(h); 6320 6321 for (i = 0; i < NACTIONS; i++) { 6322 if (scf_pg_get_property(pg, admin_actions[i], prop) != 0) { 6323 switch (scf_error()) { 6324 case SCF_ERROR_CONNECTION_BROKEN: 6325 default: 6326 ret = ECONNABORTED; 6327 goto out; 6328 6329 case SCF_ERROR_DELETED: 6330 goto out; 6331 6332 case SCF_ERROR_NOT_FOUND: 6333 action_ts[i] = 0; 6334 continue; 6335 6336 case SCF_ERROR_HANDLE_MISMATCH: 6337 case SCF_ERROR_INVALID_ARGUMENT: 6338 case SCF_ERROR_NOT_SET: 6339 bad_error("scf_pg_get_property", scf_error()); 6340 } 6341 } 6342 6343 if (scf_property_type(prop, &type) != 0) { 6344 switch (scf_error()) { 6345 case SCF_ERROR_CONNECTION_BROKEN: 6346 default: 6347 ret = ECONNABORTED; 6348 goto out; 6349 6350 case SCF_ERROR_DELETED: 6351 action_ts[i] = 0; 6352 continue; 6353 6354 case SCF_ERROR_NOT_SET: 6355 bad_error("scf_property_type", scf_error()); 6356 } 6357 } 6358 6359 if (type != SCF_TYPE_INTEGER) { 6360 action_ts[i] = 0; 6361 continue; 6362 } 6363 6364 if (scf_property_get_value(prop, val) != 0) { 6365 switch (scf_error()) { 6366 case SCF_ERROR_CONNECTION_BROKEN: 6367 default: 6368 ret = ECONNABORTED; 6369 goto out; 6370 6371 case SCF_ERROR_DELETED: 6372 goto out; 6373 6374 case SCF_ERROR_NOT_FOUND: 6375 case SCF_ERROR_CONSTRAINT_VIOLATED: 6376 action_ts[i] = 0; 6377 continue; 6378 6379 case SCF_ERROR_NOT_SET: 6380 case SCF_ERROR_PERMISSION_DENIED: 6381 bad_error("scf_property_get_value", 6382 scf_error()); 6383 } 6384 } 6385 6386 r = scf_value_get_integer(val, &action_ts[i]); 6387 assert(r == 0); 6388 } 6389 6390 a = ADMIN_EVENT_MAINT_ON_IMMEDIATE; 6391 if (action_ts[ADMIN_EVENT_MAINT_ON_IMMEDIATE] || 6392 action_ts[ADMIN_EVENT_MAINT_ON]) { 6393 a = action_ts[ADMIN_EVENT_MAINT_ON_IMMEDIATE] ? 6394 ADMIN_EVENT_MAINT_ON_IMMEDIATE : ADMIN_EVENT_MAINT_ON; 6395 6396 vertex_send_event(vertex, admin_events[a]); 6397 r = libscf_unset_action(h, pg, a, action_ts[a]); 6398 switch (r) { 6399 case 0: 6400 case EACCES: 6401 break; 6402 6403 case ECONNABORTED: 6404 ret = ECONNABORTED; 6405 goto out; 6406 6407 case EPERM: 6408 uu_die("Insufficient privilege.\n"); 6409 /* NOTREACHED */ 6410 6411 default: 6412 bad_error("libscf_unset_action", r); 6413 } 6414 } 6415 6416 while ((a = next_action(action_ts, NACTIONS)) != -1) { 6417 log_framework(LOG_DEBUG, 6418 "Graph: processing %s action for %s.\n", admin_actions[a], 6419 inst_name); 6420 6421 if (a == ADMIN_EVENT_REFRESH) { 6422 r = dgraph_refresh_instance(vertex, inst); 6423 switch (r) { 6424 case 0: 6425 case ECANCELED: 6426 case EINVAL: 6427 case -1: 6428 break; 6429 6430 case ECONNABORTED: 6431 /* pg & inst are reset now, so just return. */ 6432 ret = ECONNABORTED; 6433 goto out; 6434 6435 default: 6436 bad_error("dgraph_refresh_instance", r); 6437 } 6438 } 6439 6440 vertex_send_event(vertex, admin_events[a]); 6441 6442 r = libscf_unset_action(h, pg, a, action_ts[a]); 6443 switch (r) { 6444 case 0: 6445 case EACCES: 6446 break; 6447 6448 case ECONNABORTED: 6449 ret = ECONNABORTED; 6450 goto out; 6451 6452 case EPERM: 6453 uu_die("Insufficient privilege.\n"); 6454 /* NOTREACHED */ 6455 6456 default: 6457 bad_error("libscf_unset_action", r); 6458 } 6459 6460 action_ts[a] = 0; 6461 } 6462 6463 out: 6464 MUTEX_UNLOCK(&dgraph_lock); 6465 6466 scf_property_destroy(prop); 6467 scf_value_destroy(val); 6468 startd_free(inst_name, max_scf_fmri_size); 6469 return (ret); 6470 } 6471 6472 /* 6473 * inst and pg_name are scratch space, and are unset on entry. 6474 * Returns 6475 * 0 - success 6476 * ECONNRESET - success, but repository handle rebound 6477 * ECONNABORTED - repository connection broken 6478 */ 6479 static int 6480 process_pg_event(scf_handle_t *h, scf_propertygroup_t *pg, scf_instance_t *inst, 6481 char *pg_name) 6482 { 6483 int r; 6484 scf_property_t *prop; 6485 scf_value_t *val; 6486 char *fmri; 6487 boolean_t rebound = B_FALSE, rebind_inst = B_FALSE; 6488 6489 if (scf_pg_get_name(pg, pg_name, max_scf_value_size) < 0) { 6490 switch (scf_error()) { 6491 case SCF_ERROR_CONNECTION_BROKEN: 6492 default: 6493 return (ECONNABORTED); 6494 6495 case SCF_ERROR_DELETED: 6496 return (0); 6497 6498 case SCF_ERROR_NOT_SET: 6499 bad_error("scf_pg_get_name", scf_error()); 6500 } 6501 } 6502 6503 if (strcmp(pg_name, SCF_PG_GENERAL) == 0 || 6504 strcmp(pg_name, SCF_PG_GENERAL_OVR) == 0) { 6505 r = dgraph_update_general(pg); 6506 switch (r) { 6507 case 0: 6508 case ENOTSUP: 6509 case ECANCELED: 6510 return (0); 6511 6512 case ECONNABORTED: 6513 return (ECONNABORTED); 6514 6515 case -1: 6516 /* Error should have been logged. */ 6517 return (0); 6518 6519 default: 6520 bad_error("dgraph_update_general", r); 6521 } 6522 } else if (strcmp(pg_name, SCF_PG_RESTARTER_ACTIONS) == 0) { 6523 if (scf_pg_get_parent_instance(pg, inst) != 0) { 6524 switch (scf_error()) { 6525 case SCF_ERROR_CONNECTION_BROKEN: 6526 return (ECONNABORTED); 6527 6528 case SCF_ERROR_DELETED: 6529 case SCF_ERROR_CONSTRAINT_VIOLATED: 6530 /* Ignore commands on services. */ 6531 return (0); 6532 6533 case SCF_ERROR_NOT_BOUND: 6534 case SCF_ERROR_HANDLE_MISMATCH: 6535 case SCF_ERROR_NOT_SET: 6536 default: 6537 bad_error("scf_pg_get_parent_instance", 6538 scf_error()); 6539 } 6540 } 6541 6542 return (process_actions(h, pg, inst)); 6543 } 6544 6545 if (strcmp(pg_name, SCF_PG_OPTIONS) != 0 && 6546 strcmp(pg_name, SCF_PG_OPTIONS_OVR) != 0) 6547 return (0); 6548 6549 /* 6550 * We only care about the options[_ovr] property groups of our own 6551 * instance, so get the fmri and compare. Plus, once we know it's 6552 * correct, if the repository connection is broken we know exactly what 6553 * property group we were operating on, and can look it up again. 6554 */ 6555 if (scf_pg_get_parent_instance(pg, inst) != 0) { 6556 switch (scf_error()) { 6557 case SCF_ERROR_CONNECTION_BROKEN: 6558 return (ECONNABORTED); 6559 6560 case SCF_ERROR_DELETED: 6561 case SCF_ERROR_CONSTRAINT_VIOLATED: 6562 return (0); 6563 6564 case SCF_ERROR_HANDLE_MISMATCH: 6565 case SCF_ERROR_NOT_BOUND: 6566 case SCF_ERROR_NOT_SET: 6567 default: 6568 bad_error("scf_pg_get_parent_instance", 6569 scf_error()); 6570 } 6571 } 6572 6573 switch (r = libscf_instance_get_fmri(inst, &fmri)) { 6574 case 0: 6575 break; 6576 6577 case ECONNABORTED: 6578 return (ECONNABORTED); 6579 6580 case ECANCELED: 6581 return (0); 6582 6583 default: 6584 bad_error("libscf_instance_get_fmri", r); 6585 } 6586 6587 if (strcmp(fmri, SCF_SERVICE_STARTD) != 0) { 6588 startd_free(fmri, max_scf_fmri_size); 6589 return (0); 6590 } 6591 6592 /* 6593 * update the information events flag 6594 */ 6595 if (strcmp(pg_name, SCF_PG_OPTIONS) == 0) 6596 info_events_all = libscf_get_info_events_all(pg); 6597 6598 prop = safe_scf_property_create(h); 6599 val = safe_scf_value_create(h); 6600 6601 if (strcmp(pg_name, SCF_PG_OPTIONS_OVR) == 0) { 6602 /* See if we need to set the runlevel. */ 6603 /* CONSTCOND */ 6604 if (0) { 6605 rebind_pg: 6606 libscf_handle_rebind(h); 6607 rebound = B_TRUE; 6608 6609 r = libscf_lookup_instance(SCF_SERVICE_STARTD, inst); 6610 switch (r) { 6611 case 0: 6612 break; 6613 6614 case ECONNABORTED: 6615 goto rebind_pg; 6616 6617 case ENOENT: 6618 goto out; 6619 6620 case EINVAL: 6621 case ENOTSUP: 6622 bad_error("libscf_lookup_instance", r); 6623 } 6624 6625 if (scf_instance_get_pg(inst, pg_name, pg) != 0) { 6626 switch (scf_error()) { 6627 case SCF_ERROR_DELETED: 6628 case SCF_ERROR_NOT_FOUND: 6629 goto out; 6630 6631 case SCF_ERROR_CONNECTION_BROKEN: 6632 goto rebind_pg; 6633 6634 case SCF_ERROR_HANDLE_MISMATCH: 6635 case SCF_ERROR_NOT_BOUND: 6636 case SCF_ERROR_NOT_SET: 6637 case SCF_ERROR_INVALID_ARGUMENT: 6638 default: 6639 bad_error("scf_instance_get_pg", 6640 scf_error()); 6641 } 6642 } 6643 } 6644 6645 if (scf_pg_get_property(pg, "runlevel", prop) == 0) { 6646 r = dgraph_set_runlevel(pg, prop); 6647 switch (r) { 6648 case ECONNRESET: 6649 rebound = B_TRUE; 6650 rebind_inst = B_TRUE; 6651 /* FALLTHROUGH */ 6652 6653 case 0: 6654 break; 6655 6656 case ECONNABORTED: 6657 goto rebind_pg; 6658 6659 case ECANCELED: 6660 goto out; 6661 6662 default: 6663 bad_error("dgraph_set_runlevel", r); 6664 } 6665 } else { 6666 switch (scf_error()) { 6667 case SCF_ERROR_CONNECTION_BROKEN: 6668 default: 6669 goto rebind_pg; 6670 6671 case SCF_ERROR_DELETED: 6672 goto out; 6673 6674 case SCF_ERROR_NOT_FOUND: 6675 break; 6676 6677 case SCF_ERROR_INVALID_ARGUMENT: 6678 case SCF_ERROR_HANDLE_MISMATCH: 6679 case SCF_ERROR_NOT_BOUND: 6680 case SCF_ERROR_NOT_SET: 6681 bad_error("scf_pg_get_property", scf_error()); 6682 } 6683 } 6684 } 6685 6686 if (rebind_inst) { 6687 lookup_inst: 6688 r = libscf_lookup_instance(SCF_SERVICE_STARTD, inst); 6689 switch (r) { 6690 case 0: 6691 break; 6692 6693 case ECONNABORTED: 6694 libscf_handle_rebind(h); 6695 rebound = B_TRUE; 6696 goto lookup_inst; 6697 6698 case ENOENT: 6699 goto out; 6700 6701 case EINVAL: 6702 case ENOTSUP: 6703 bad_error("libscf_lookup_instance", r); 6704 } 6705 } 6706 6707 r = libscf_get_milestone(inst, prop, val, fmri, max_scf_fmri_size); 6708 switch (r) { 6709 case 0: 6710 break; 6711 6712 case ECONNABORTED: 6713 libscf_handle_rebind(h); 6714 rebound = B_TRUE; 6715 goto lookup_inst; 6716 6717 case EINVAL: 6718 log_error(LOG_NOTICE, 6719 "%s/%s property of %s is misconfigured.\n", pg_name, 6720 SCF_PROPERTY_MILESTONE, SCF_SERVICE_STARTD); 6721 /* FALLTHROUGH */ 6722 6723 case ECANCELED: 6724 case ENOENT: 6725 (void) strcpy(fmri, "all"); 6726 break; 6727 6728 default: 6729 bad_error("libscf_get_milestone", r); 6730 } 6731 6732 r = dgraph_set_milestone(fmri, h, B_FALSE); 6733 switch (r) { 6734 case 0: 6735 case ECONNRESET: 6736 case EALREADY: 6737 break; 6738 6739 case EINVAL: 6740 log_error(LOG_WARNING, "Milestone %s is invalid.\n", fmri); 6741 break; 6742 6743 case ENOENT: 6744 log_error(LOG_WARNING, "Milestone %s does not exist.\n", fmri); 6745 break; 6746 6747 default: 6748 bad_error("dgraph_set_milestone", r); 6749 } 6750 6751 out: 6752 startd_free(fmri, max_scf_fmri_size); 6753 scf_value_destroy(val); 6754 scf_property_destroy(prop); 6755 6756 return (rebound ? ECONNRESET : 0); 6757 } 6758 6759 /* 6760 * process_delete() deletes an instance from the dgraph if 'fmri' is an 6761 * instance fmri or if 'fmri' matches the 'general' property group of an 6762 * instance (or the 'general/enabled' property). 6763 * 6764 * 'fmri' may be overwritten and cannot be trusted on return by the caller. 6765 */ 6766 static void 6767 process_delete(char *fmri, scf_handle_t *h) 6768 { 6769 char *lfmri, *end_inst_fmri; 6770 const char *inst_name = NULL; 6771 const char *pg_name = NULL; 6772 const char *prop_name = NULL; 6773 6774 lfmri = safe_strdup(fmri); 6775 6776 /* Determine if the FMRI is a property group or instance */ 6777 if (scf_parse_svc_fmri(lfmri, NULL, NULL, &inst_name, &pg_name, 6778 &prop_name) != SCF_SUCCESS) { 6779 log_error(LOG_WARNING, 6780 "Received invalid FMRI \"%s\" from repository server.\n", 6781 fmri); 6782 } else if (inst_name != NULL && pg_name == NULL) { 6783 (void) dgraph_remove_instance(fmri, h); 6784 } else if (inst_name != NULL && pg_name != NULL) { 6785 /* 6786 * If we're deleting the 'general' property group or 6787 * 'general/enabled' property then the whole instance 6788 * must be removed from the dgraph. 6789 */ 6790 if (strcmp(pg_name, SCF_PG_GENERAL) != 0) { 6791 free(lfmri); 6792 return; 6793 } 6794 6795 if (prop_name != NULL && 6796 strcmp(prop_name, SCF_PROPERTY_ENABLED) != 0) { 6797 free(lfmri); 6798 return; 6799 } 6800 6801 /* 6802 * Because the instance has already been deleted from the 6803 * repository, we cannot use any scf_ functions to retrieve 6804 * the instance FMRI however we can easily reconstruct it 6805 * manually. 6806 */ 6807 end_inst_fmri = strstr(fmri, SCF_FMRI_PROPERTYGRP_PREFIX); 6808 if (end_inst_fmri == NULL) 6809 bad_error("process_delete", 0); 6810 6811 end_inst_fmri[0] = '\0'; 6812 6813 (void) dgraph_remove_instance(fmri, h); 6814 } 6815 6816 free(lfmri); 6817 } 6818 6819 /*ARGSUSED*/ 6820 void * 6821 repository_event_thread(void *unused) 6822 { 6823 scf_handle_t *h; 6824 scf_propertygroup_t *pg; 6825 scf_instance_t *inst; 6826 char *fmri = startd_alloc(max_scf_fmri_size); 6827 char *pg_name = startd_alloc(max_scf_value_size); 6828 int r; 6829 6830 h = libscf_handle_create_bound_loop(); 6831 6832 pg = safe_scf_pg_create(h); 6833 inst = safe_scf_instance_create(h); 6834 6835 retry: 6836 if (_scf_notify_add_pgtype(h, SCF_GROUP_FRAMEWORK) != SCF_SUCCESS) { 6837 if (scf_error() == SCF_ERROR_CONNECTION_BROKEN) { 6838 libscf_handle_rebind(h); 6839 } else { 6840 log_error(LOG_WARNING, 6841 "Couldn't set up repository notification " 6842 "for property group type %s: %s\n", 6843 SCF_GROUP_FRAMEWORK, scf_strerror(scf_error())); 6844 6845 (void) sleep(1); 6846 } 6847 6848 goto retry; 6849 } 6850 6851 /*CONSTCOND*/ 6852 while (1) { 6853 ssize_t res; 6854 6855 /* Note: fmri is only set on delete events. */ 6856 res = _scf_notify_wait(pg, fmri, max_scf_fmri_size); 6857 if (res < 0) { 6858 libscf_handle_rebind(h); 6859 goto retry; 6860 } else if (res == 0) { 6861 /* 6862 * property group modified. inst and pg_name are 6863 * pre-allocated scratch space. 6864 */ 6865 if (scf_pg_update(pg) < 0) { 6866 switch (scf_error()) { 6867 case SCF_ERROR_DELETED: 6868 continue; 6869 6870 case SCF_ERROR_CONNECTION_BROKEN: 6871 log_error(LOG_WARNING, 6872 "Lost repository event due to " 6873 "disconnection.\n"); 6874 libscf_handle_rebind(h); 6875 goto retry; 6876 6877 case SCF_ERROR_NOT_BOUND: 6878 case SCF_ERROR_NOT_SET: 6879 default: 6880 bad_error("scf_pg_update", scf_error()); 6881 } 6882 } 6883 6884 r = process_pg_event(h, pg, inst, pg_name); 6885 switch (r) { 6886 case 0: 6887 break; 6888 6889 case ECONNABORTED: 6890 log_error(LOG_WARNING, "Lost repository event " 6891 "due to disconnection.\n"); 6892 libscf_handle_rebind(h); 6893 /* FALLTHROUGH */ 6894 6895 case ECONNRESET: 6896 goto retry; 6897 6898 default: 6899 bad_error("process_pg_event", r); 6900 } 6901 } else { 6902 /* 6903 * Service, instance, or pg deleted. 6904 * Don't trust fmri on return. 6905 */ 6906 process_delete(fmri, h); 6907 } 6908 } 6909 6910 /*NOTREACHED*/ 6911 return (NULL); 6912 } 6913 6914 void 6915 graph_engine_start() 6916 { 6917 int err; 6918 6919 (void) startd_thread_create(graph_thread, NULL); 6920 6921 MUTEX_LOCK(&dgraph_lock); 6922 while (!initial_milestone_set) { 6923 err = pthread_cond_wait(&initial_milestone_cv, &dgraph_lock); 6924 assert(err == 0); 6925 } 6926 MUTEX_UNLOCK(&dgraph_lock); 6927 6928 (void) startd_thread_create(repository_event_thread, NULL); 6929 (void) startd_thread_create(graph_event_thread, NULL); 6930 }