![[first]](../icons/first.png){kind=icon}
![[previous]](../icons/n_left.png){kind=icon}
![[next]](../icons/n_right.png){kind=icon}
![[last]](../icons/last.png){kind=icon}
![[top]](../icons/n_top.png){kind=icon}
![[bottom]](../icons/bottom.png){kind=icon}
![[index]](../icons/index.png){kind=icon}
![[help]](../icons/help.png){kind=icon}
*/
1 /*
2 * Copyright 2004-2024 the Pacemaker project contributors
3 *
4 * The version control history for this file may have further details.
5 *
6 * This source code is licensed under the GNU General Public License version 2
7 * or later (GPLv2+) WITHOUT ANY WARRANTY.
8 */
9
10 #include <crm_internal.h>
11
12 #include <glib.h>
13
14 #include <crm/crm.h>
15 #include <crm/pengine/status.h>
16 #include <pacemaker-internal.h>
17 #include "libpacemaker_private.h"
18
19 /*!
20 * \internal
21 * \brief Check whether a resource is known on a particular node
22 *
23 * \param[in] rsc Resource to check
24 * \param[in] node Node to check
25 *
26 * \return TRUE if resource (or parent if an anonymous clone) is known
27 */
28 static bool
29 rsc_is_known_on(const pcmk_resource_t *rsc, const pcmk_node_t *node)
/* ![[next]](../icons/right.png)
![[first]](../icons/n_first.png)
![[top]](../icons/top.png)
*/
30 {
31 if (g_hash_table_lookup(rsc->known_on, node->details->id) != NULL) {
32 return TRUE;
33
34 } else if (pcmk__is_primitive(rsc)
35 && pcmk__is_anonymous_clone(rsc->parent)
36 && (g_hash_table_lookup(rsc->parent->known_on,
37 node->details->id) != NULL)) {
38 /* We check only the parent, not the uber-parent, because we cannot
39 * assume that the resource is known if it is in an anonymously cloned
40 * group (which may be only partially known).
41 */
42 return TRUE;
43 }
44 return FALSE;
45 }
46
47 /*!
48 * \internal
49 * \brief Order a resource's start and promote actions relative to fencing
50 *
51 * \param[in,out] rsc Resource to be ordered
52 * \param[in,out] stonith_op Fence action
53 */
54 static void
55 order_start_vs_fencing(pcmk_resource_t *rsc, pcmk_action_t *stonith_op)
/* ![[previous]](../icons/left.png)
*/
56 {
57 pcmk_node_t *target;
58
59 CRM_CHECK(stonith_op && stonith_op->node, return);
60 target = stonith_op->node;
61
62 for (GList *iter = rsc->actions; iter != NULL; iter = iter->next) {
63 pcmk_action_t *action = iter->data;
64
65 switch (action->needs) {
66 case pcmk_requires_nothing:
67 // Anything other than start or promote requires nothing
68 break;
69
70 case pcmk_requires_fencing:
71 order_actions(stonith_op, action, pcmk__ar_ordered);
72 break;
73
74 case pcmk_requires_quorum:
75 if (pcmk__str_eq(action->task, PCMK_ACTION_START,
76 pcmk__str_none)
77 && (g_hash_table_lookup(rsc->allowed_nodes,
78 target->details->id) != NULL)
79 && !rsc_is_known_on(rsc, target)) {
80
81 /* If we don't know the status of the resource on the node
82 * we're about to shoot, we have to assume it may be active
83 * there. Order the resource start after the fencing. This
84 * is analogous to waiting for all the probes for a resource
85 * to complete before starting it.
86 *
87 * The most likely explanation is that the DC died and took
88 * its status with it.
89 */
90 pcmk__rsc_debug(rsc, "Ordering %s after %s recovery",
91 action->uuid, pcmk__node_name(target));
92 order_actions(stonith_op, action,
93 pcmk__ar_ordered
94 |pcmk__ar_unrunnable_first_blocks);
95 }
96 break;
97 }
98 }
99 }
100
101 /*!
102 * \internal
103 * \brief Order a resource's stop and demote actions relative to fencing
104 *
105 * \param[in,out] rsc Resource to be ordered
106 * \param[in,out] stonith_op Fence action
107 */
108 static void
109 order_stop_vs_fencing(pcmk_resource_t *rsc, pcmk_action_t *stonith_op)
/* */
110 {
111 GList *iter = NULL;
112 GList *action_list = NULL;
113 bool order_implicit = false;
114
115 pcmk_resource_t *top = uber_parent(rsc);
116 pcmk_action_t *parent_stop = NULL;
117 pcmk_node_t *target;
118
119 CRM_CHECK(stonith_op && stonith_op->node, return);
120 target = stonith_op->node;
121
122 /* Get a list of stop actions potentially implied by the fencing */
123 action_list = pe__resource_actions(rsc, target, PCMK_ACTION_STOP, FALSE);
124
125 /* If resource requires fencing, implicit actions must occur after fencing.
126 *
127 * Implied stops and demotes of resources running on guest nodes are always
128 * ordered after fencing, even if the resource does not require fencing,
129 * because guest node "fencing" is actually just a resource stop.
130 */
131 if (pcmk_is_set(rsc->flags, pcmk_rsc_needs_fencing)
132 || pcmk__is_guest_or_bundle_node(target)) {
133
134 order_implicit = true;
135 }
136
137 if (action_list && order_implicit) {
138 parent_stop = find_first_action(top->actions, NULL, PCMK_ACTION_STOP,
139 NULL);
140 }
141
142 for (iter = action_list; iter != NULL; iter = iter->next) {
143 pcmk_action_t *action = iter->data;
144
145 // The stop would never complete, so convert it into a pseudo-action.
146 pcmk__set_action_flags(action, pcmk_action_pseudo|pcmk_action_runnable);
147
148 if (order_implicit) {
149 /* Order the stonith before the parent stop (if any).
150 *
151 * Also order the stonith before the resource stop, unless the
152 * resource is inside a bundle -- that would cause a graph loop.
153 * We can rely on the parent stop's ordering instead.
154 *
155 * User constraints must not order a resource in a guest node
156 * relative to the guest node container resource. The
157 * pcmk__ar_guest_allowed flag marks constraints as generated by the
158 * cluster and thus immune to that check (and is irrelevant if
159 * target is not a guest).
160 */
161 if (!pcmk__is_bundled(rsc)) {
162 order_actions(stonith_op, action, pcmk__ar_guest_allowed);
163 }
164 order_actions(stonith_op, parent_stop, pcmk__ar_guest_allowed);
165 }
166
167 if (pcmk_is_set(rsc->flags, pcmk_rsc_failed)) {
168 crm_notice("Stop of failed resource %s is implicit %s %s is fenced",
169 rsc->id, (order_implicit? "after" : "because"),
170 pcmk__node_name(target));
171 } else {
172 crm_info("%s is implicit %s %s is fenced",
173 action->uuid, (order_implicit? "after" : "because"),
174 pcmk__node_name(target));
175 }
176
177 if (pcmk_is_set(rsc->flags, pcmk_rsc_notify)) {
178 pe__order_notifs_after_fencing(action, rsc, stonith_op);
179 }
180
181 #if 0
182 /* It might be a good idea to stop healthy resources on a node about to
183 * be fenced, when possible.
184 *
185 * However, fencing must be done before a failed resource's
186 * (pseudo-)stop action, so that could create a loop. For example, given
187 * a group of A and B running on node N with a failed stop of B:
188 *
189 * fence N -> stop B (pseudo-op) -> stop A -> fence N
190 *
191 * The block below creates the stop A -> fence N ordering and therefore
192 * must (at least for now) be disabled. Instead, run the block above and
193 * treat all resources on N as B would be (i.e., as a pseudo-op after
194 * the fencing).
195 *
196 * @TODO Maybe break the "A requires B" dependency in
197 * pcmk__update_action_for_orderings() and use this block for healthy
198 * resources instead of the above.
199 */
200 crm_info("Moving healthy resource %s off %s before fencing",
201 rsc->id, pcmk__node_name(node));
202 pcmk__new_ordering(rsc, stop_key(rsc), NULL, NULL,
203 strdup(PCMK_ACTION_STONITH), stonith_op,
204 pcmk__ar_ordered, rsc->cluster);
205 #endif
206 }
207
208 g_list_free(action_list);
209
210 /* Get a list of demote actions potentially implied by the fencing */
211 action_list = pe__resource_actions(rsc, target, PCMK_ACTION_DEMOTE, FALSE);
212
213 for (iter = action_list; iter != NULL; iter = iter->next) {
214 pcmk_action_t *action = iter->data;
215
216 if (!(action->node->details->online) || action->node->details->unclean
217 || pcmk_is_set(rsc->flags, pcmk_rsc_failed)) {
218
219 if (pcmk_is_set(rsc->flags, pcmk_rsc_failed)) {
220 pcmk__rsc_info(rsc,
221 "Demote of failed resource %s is implicit "
222 "after %s is fenced",
223 rsc->id, pcmk__node_name(target));
224 } else {
225 pcmk__rsc_info(rsc, "%s is implicit after %s is fenced",
226 action->uuid, pcmk__node_name(target));
227 }
228
229 /* The demote would never complete and is now implied by the
230 * fencing, so convert it into a pseudo-action.
231 */
232 pcmk__set_action_flags(action,
233 pcmk_action_pseudo|pcmk_action_runnable);
234
235 if (pcmk__is_bundled(rsc)) {
236 // Recovery will be ordered as usual after parent's implied stop
237
238 } else if (order_implicit) {
239 order_actions(stonith_op, action,
240 pcmk__ar_guest_allowed|pcmk__ar_ordered);
241 }
242 }
243 }
244
245 g_list_free(action_list);
246 }
247
248 /*!
249 * \internal
250 * \brief Order resource actions properly relative to fencing
251 *
252 * \param[in,out] rsc Resource whose actions should be ordered
253 * \param[in,out] stonith_op Fencing operation to be ordered against
254 */
255 static void
256 rsc_stonith_ordering(pcmk_resource_t *rsc, pcmk_action_t *stonith_op)
/* */
257 {
258 if (rsc->children) {
259 for (GList *iter = rsc->children; iter != NULL; iter = iter->next) {
260 pcmk_resource_t *child_rsc = iter->data;
261
262 rsc_stonith_ordering(child_rsc, stonith_op);
263 }
264
265 } else if (!pcmk_is_set(rsc->flags, pcmk_rsc_managed)) {
266 pcmk__rsc_trace(rsc,
267 "Skipping fencing constraints for unmanaged resource: "
268 "%s", rsc->id);
269
270 } else {
271 order_start_vs_fencing(rsc, stonith_op);
272 order_stop_vs_fencing(rsc, stonith_op);
273 }
274 }
275
276 /*!
277 * \internal
278 * \brief Order all actions appropriately relative to a fencing operation
279 *
280 * Ensure start operations of affected resources are ordered after fencing,
281 * imply stop and demote operations of affected resources by marking them as
282 * pseudo-actions, etc.
283 *
284 * \param[in,out] stonith_op Fencing operation
285 * \param[in,out] scheduler Scheduler data
286 */
287 void
288 pcmk__order_vs_fence(pcmk_action_t *stonith_op, pcmk_scheduler_t *scheduler)
/* */
289 {
290 CRM_CHECK(stonith_op && scheduler, return);
291 for (GList *r = scheduler->resources; r != NULL; r = r->next) {
292 rsc_stonith_ordering((pcmk_resource_t *) r->data, stonith_op);
293 }
294 }
295
296 /*!
297 * \internal
298 * \brief Order an action after unfencing
299 *
300 * \param[in] rsc Resource that action is for
301 * \param[in,out] node Node that action is on
302 * \param[in,out] action Action to be ordered after unfencing
303 * \param[in] order Ordering flags
304 */
305 void
306 pcmk__order_vs_unfence(const pcmk_resource_t *rsc, pcmk_node_t *node,
/* */
307 pcmk_action_t *action,
308 enum pcmk__action_relation_flags order)
309 {
310 /* When unfencing is in use, we order unfence actions before any probe or
311 * start of resources that require unfencing, and also of fence devices.
312 *
313 * This might seem to violate the principle that fence devices require
314 * only quorum. However, fence agents that unfence often don't have enough
315 * information to even probe or start unless the node is first unfenced.
316 */
317 if ((pcmk_is_set(rsc->flags, pcmk_rsc_fence_device)
318 && pcmk_is_set(rsc->cluster->flags, pcmk_sched_enable_unfencing))
319 || pcmk_is_set(rsc->flags, pcmk_rsc_needs_unfencing)) {
320
321 /* Start with an optional ordering. Requiring unfencing would result in
322 * the node being unfenced, and all its resources being stopped,
323 * whenever a new resource is added -- which would be highly suboptimal.
324 */
325 pcmk_action_t *unfence = pe_fence_op(node, PCMK_ACTION_ON, TRUE, NULL,
326 FALSE, node->details->data_set);
327
328 order_actions(unfence, action, order);
329
330 if (!pcmk__node_unfenced(node)) {
331 // But unfencing is required if it has never been done
332 char *reason = crm_strdup_printf("required by %s %s",
333 rsc->id, action->task);
334
335 trigger_unfencing(NULL, node, reason, NULL,
336 node->details->data_set);
337 free(reason);
338 }
339 }
340 }
341
342 /*!
343 * \internal
344 * \brief Create pseudo-op for guest node fence, and order relative to it
345 *
346 * \param[in,out] node Guest node to fence
347 */
348 void
349 pcmk__fence_guest(pcmk_node_t *node)
/* */
350 {
351 pcmk_resource_t *container = NULL;
352 pcmk_action_t *stop = NULL;
353 pcmk_action_t *stonith_op = NULL;
354
355 /* The fence action is just a label; we don't do anything differently for
356 * off vs. reboot. We specify it explicitly, rather than let it default to
357 * cluster's default action, because we are not _initiating_ fencing -- we
358 * are creating a pseudo-event to describe fencing that is already occurring
359 * by other means (container recovery).
360 */
361 const char *fence_action = PCMK_ACTION_OFF;
362
363 CRM_ASSERT(node != NULL);
364
365 /* Check whether guest's container resource has any explicit stop or
366 * start (the stop may be implied by fencing of the guest's host).
367 */
368 container = node->details->remote_rsc->container;
369 if (container) {
370 stop = find_first_action(container->actions, NULL, PCMK_ACTION_STOP,
371 NULL);
372
373 if (find_first_action(container->actions, NULL, PCMK_ACTION_START,
374 NULL)) {
375 fence_action = PCMK_ACTION_REBOOT;
376 }
377 }
378
379 /* Create a fence pseudo-event, so we have an event to order actions
380 * against, and the controller can always detect it.
381 */
382 stonith_op = pe_fence_op(node, fence_action, FALSE, "guest is unclean",
383 FALSE, node->details->data_set);
384 pcmk__set_action_flags(stonith_op, pcmk_action_pseudo|pcmk_action_runnable);
385
386 /* We want to imply stops/demotes after the guest is stopped, not wait until
387 * it is restarted, so we always order pseudo-fencing after stop, not start
388 * (even though start might be closer to what is done for a real reboot).
389 */
390 if ((stop != NULL) && pcmk_is_set(stop->flags, pcmk_action_pseudo)) {
391 pcmk_action_t *parent_stonith_op = pe_fence_op(stop->node, NULL, FALSE,
392 NULL, FALSE,
393 node->details->data_set);
394
395 crm_info("Implying guest %s is down (action %d) after %s fencing",
396 pcmk__node_name(node), stonith_op->id,
397 pcmk__node_name(stop->node));
398 order_actions(parent_stonith_op, stonith_op,
399 pcmk__ar_unrunnable_first_blocks
400 |pcmk__ar_first_implies_then);
401
402 } else if (stop) {
403 order_actions(stop, stonith_op,
404 pcmk__ar_unrunnable_first_blocks
405 |pcmk__ar_first_implies_then);
406 crm_info("Implying guest %s is down (action %d) "
407 "after container %s is stopped (action %d)",
408 pcmk__node_name(node), stonith_op->id,
409 container->id, stop->id);
410 } else {
411 /* If we're fencing the guest node but there's no stop for the guest
412 * resource, we must think the guest is already stopped. However, we may
413 * think so because its resource history was just cleaned. To avoid
414 * unnecessarily considering the guest node down if it's really up,
415 * order the pseudo-fencing after any stop of the connection resource,
416 * which will be ordered after any container (re-)probe.
417 */
418 stop = find_first_action(node->details->remote_rsc->actions, NULL,
419 PCMK_ACTION_STOP, NULL);
420
421 if (stop) {
422 order_actions(stop, stonith_op, pcmk__ar_ordered);
423 crm_info("Implying guest %s is down (action %d) "
424 "after connection is stopped (action %d)",
425 pcmk__node_name(node), stonith_op->id, stop->id);
426 } else {
427 /* Not sure why we're fencing, but everything must already be
428 * cleanly stopped.
429 */
430 crm_info("Implying guest %s is down (action %d) ",
431 pcmk__node_name(node), stonith_op->id);
432 }
433 }
434
435 // Order/imply other actions relative to pseudo-fence as with real fence
436 pcmk__order_vs_fence(stonith_op, node->details->data_set);
437 }
438
439 /*!
440 * \internal
441 * \brief Check whether node has already been unfenced
442 *
443 * \param[in] node Node to check
444 *
445 * \return true if node has a nonzero #node-unfenced attribute (or none),
446 * otherwise false
447 */
448 bool
449 pcmk__node_unfenced(const pcmk_node_t *node)
/* */
450 {
451 const char *unfenced = pcmk__node_attr(node, CRM_ATTR_UNFENCED, NULL,
452 pcmk__rsc_node_current);
453
454 return !pcmk__str_eq(unfenced, "0", pcmk__str_null_matches);
455 }
456
457 /*!
458 * \internal
459 * \brief Order a resource's start and stop relative to unfencing of a node
460 *
461 * \param[in,out] data Node that could be unfenced
462 * \param[in,out] user_data Resource to order
463 */
464 void
465 pcmk__order_restart_vs_unfence(gpointer data, gpointer user_data)
/* ![[last]](../icons/n_last.png)
*/
466 {
467 pcmk_node_t *node = (pcmk_node_t *) data;
468 pcmk_resource_t *rsc = (pcmk_resource_t *) user_data;
469
470 pcmk_action_t *unfence = pe_fence_op(node, PCMK_ACTION_ON, true, NULL,
471 false, rsc->cluster);
472
473 crm_debug("Ordering any stops of %s before %s, and any starts after",
474 rsc->id, unfence->uuid);
475
476 /*
477 * It would be more efficient to order clone resources once,
478 * rather than order each instance, but ordering the instance
479 * allows us to avoid unnecessary dependencies that might conflict
480 * with user constraints.
481 *
482 * @TODO: This constraint can still produce a transition loop if the
483 * resource has a stop scheduled on the node being unfenced, and
484 * there is a user ordering constraint to start some other resource
485 * (which will be ordered after the unfence) before stopping this
486 * resource. An example is "start some slow-starting cloned service
487 * before stopping an associated virtual IP that may be moving to
488 * it":
489 * stop this -> unfencing -> start that -> stop this
490 */
491 pcmk__new_ordering(rsc, stop_key(rsc), NULL,
492 NULL, strdup(unfence->uuid), unfence,
493 pcmk__ar_ordered|pcmk__ar_if_on_same_node,
494 rsc->cluster);
495
496 pcmk__new_ordering(NULL, strdup(unfence->uuid), unfence,
497 rsc, start_key(rsc), NULL,
498 pcmk__ar_first_implies_same_node_then
499 |pcmk__ar_if_on_same_node,
500 rsc->cluster);
501 }
![[bottom]](../icons/n_bottom.png){kind=icon}
*/