diff --git a/include/crm/common/action_relation_internal.h b/include/crm/common/action_relation_internal.h index 898ab417a7..e23c549b70 100644 --- a/include/crm/common/action_relation_internal.h +++ b/include/crm/common/action_relation_internal.h @@ -1,55 +1,58 @@ /* * Copyright 2023 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU Lesser General Public License * version 2.1 or later (LGPLv2.1+) WITHOUT ANY WARRANTY. */ #ifndef PCMK__CRM_COMMON_ACTION_RELATION_INTERNAL__H # define PCMK__CRM_COMMON_ACTION_RELATION_INTERNAL__H /*! * Flags to indicate the relationship between two actions * * @COMPAT The values and semantics of these flags should not be changed until * the deprecated enum pe_ordering is dropped from the public API. */ enum pcmk__action_relation_flags { //! No relation (compare with equality rather than bit set) pcmk__ar_none = 0U, //! Actions are ordered (optionally, if no other flags are set) pcmk__ar_ordered = (1U << 0), //! Relation applies only if 'first' cannot be part of a live migration pcmk__ar_if_first_unmigratable = (1U << 1), /*! * If 'then' is required, 'first' becomes required (and becomes unmigratable * if 'then' is); also, if 'first' is a stop of a blocked resource, 'then' * becomes unrunnable */ pcmk__ar_then_implies_first = (1U << 4), /*! * If 'first' is required, 'then' becomes required; if 'first' is a stop of * a blocked resource, 'then' becomes unrunnable */ pcmk__ar_first_implies_then = (1U << 5), /*! * If 'then' is required and for a promoted instance, 'first' becomes * required (and becomes unmigratable if 'then' is) */ pcmk__ar_promoted_then_implies_first = (1U << 6), /*! * 'first' is runnable only if 'then' is both runnable and migratable, * and 'first' becomes required if 'then' is */ pcmk__ar_unmigratable_then_blocks = (1U << 7), + + //! 'then' is runnable (and migratable) only if 'first' is runnable + pcmk__ar_unrunnable_first_blocks = (1U << 8), }; #endif // PCMK__CRM_COMMON_ACTION_RELATION_INTERNAL__H diff --git a/lib/pacemaker/pcmk_sched_actions.c b/lib/pacemaker/pcmk_sched_actions.c index 65782e2634..a5c9002852 100644 --- a/lib/pacemaker/pcmk_sched_actions.c +++ b/lib/pacemaker/pcmk_sched_actions.c @@ -1,1928 +1,1929 @@ /* * Copyright 2004-2023 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU General Public License version 2 * or later (GPLv2+) WITHOUT ANY WARRANTY. */ #include #include #include #include #include #include #include #include "libpacemaker_private.h" /*! * \internal * \brief Get the action flags relevant to ordering constraints * * \param[in,out] action Action to check * \param[in] node Node that *other* action in the ordering is on * (used only for clone resource actions) * * \return Action flags that should be used for orderings */ static uint32_t action_flags_for_ordering(pe_action_t *action, const pe_node_t *node) { bool runnable = false; uint32_t flags; // For non-resource actions, return the action flags if (action->rsc == NULL) { return action->flags; } /* For non-clone resources, or a clone action not assigned to a node, * return the flags as determined by the resource method without a node * specified. */ flags = action->rsc->cmds->action_flags(action, NULL); if ((node == NULL) || !pe_rsc_is_clone(action->rsc)) { return flags; } /* Otherwise (i.e., for clone resource actions on a specific node), first * remember whether the non-node-specific action is runnable. */ runnable = pcmk_is_set(flags, pcmk_action_runnable); // Then recheck the resource method with the node flags = action->rsc->cmds->action_flags(action, node); /* For clones in ordering constraints, the node-specific "runnable" doesn't * matter, just the non-node-specific setting (i.e., is the action runnable * anywhere). * * This applies only to runnable, and only for ordering constraints. This * function shouldn't be used for other types of constraints without * changes. Not very satisfying, but it's logical and appears to work well. */ if (runnable && !pcmk_is_set(flags, pcmk_action_runnable)) { pe__set_raw_action_flags(flags, action->rsc->id, pcmk_action_runnable); } return flags; } /*! * \internal * \brief Get action UUID that should be used with a resource ordering * * When an action is ordered relative to an action for a collective resource * (clone, group, or bundle), it actually needs to be ordered after all * instances of the collective have completed the relevant action (for example, * given "start CLONE then start RSC", RSC must wait until all instances of * CLONE have started). Given the UUID and resource of the first action in an * ordering, this returns the UUID of the action that should actually be used * for ordering (for example, "CLONE_started_0" instead of "CLONE_start_0"). * * \param[in] first_uuid UUID of first action in ordering * \param[in] first_rsc Resource of first action in ordering * * \return Newly allocated copy of UUID to use with ordering * \note It is the caller's responsibility to free the return value. */ static char * action_uuid_for_ordering(const char *first_uuid, const pe_resource_t *first_rsc) { guint interval_ms = 0; char *uuid = NULL; char *rid = NULL; char *first_task_str = NULL; enum action_tasks first_task = pcmk_action_unspecified; enum action_tasks remapped_task = pcmk_action_unspecified; // Only non-notify actions for collective resources need remapping if ((strstr(first_uuid, PCMK_ACTION_NOTIFY) != NULL) || (first_rsc->variant < pcmk_rsc_variant_group)) { goto done; } // Only non-recurring actions need remapping CRM_ASSERT(parse_op_key(first_uuid, &rid, &first_task_str, &interval_ms)); if (interval_ms > 0) { goto done; } first_task = text2task(first_task_str); switch (first_task) { case pcmk_action_stop: case pcmk_action_start: case pcmk_action_notify: case pcmk_action_promote: case pcmk_action_demote: remapped_task = first_task + 1; break; case pcmk_action_stopped: case pcmk_action_started: case pcmk_action_notified: case pcmk_action_promoted: case pcmk_action_demoted: remapped_task = first_task; break; case pcmk_action_monitor: case pcmk_action_shutdown: case pcmk_action_fence: break; default: crm_err("Unknown action '%s' in ordering", first_task_str); break; } if (remapped_task != pcmk_action_unspecified) { /* If a clone or bundle has notifications enabled, the ordering will be * relative to when notifications have been sent for the remapped task. */ if (pcmk_is_set(first_rsc->flags, pcmk_rsc_notify) && (pe_rsc_is_clone(first_rsc) || pe_rsc_is_bundled(first_rsc))) { uuid = pcmk__notify_key(rid, "confirmed-post", task2text(remapped_task)); } else { uuid = pcmk__op_key(rid, task2text(remapped_task), 0); } pe_rsc_trace(first_rsc, "Remapped action UUID %s to %s for ordering purposes", first_uuid, uuid); } done: if (uuid == NULL) { uuid = strdup(first_uuid); CRM_ASSERT(uuid != NULL); } free(first_task_str); free(rid); return uuid; } /*! * \internal * \brief Get actual action that should be used with an ordering * * When an action is ordered relative to an action for a collective resource * (clone, group, or bundle), it actually needs to be ordered after all * instances of the collective have completed the relevant action (for example, * given "start CLONE then start RSC", RSC must wait until all instances of * CLONE have started). Given the first action in an ordering, this returns the * the action that should actually be used for ordering (for example, the * started action instead of the start action). * * \param[in] action First action in an ordering * * \return Actual action that should be used for the ordering */ static pe_action_t * action_for_ordering(pe_action_t *action) { pe_action_t *result = action; pe_resource_t *rsc = action->rsc; if ((rsc != NULL) && (rsc->variant >= pcmk_rsc_variant_group) && (action->uuid != NULL)) { char *uuid = action_uuid_for_ordering(action->uuid, rsc); result = find_first_action(rsc->actions, uuid, NULL, NULL); if (result == NULL) { crm_warn("Not remapping %s to %s because %s does not have " "remapped action", action->uuid, uuid, rsc->id); result = action; } free(uuid); } return result; } /*! * \internal * \brief Wrapper for update_ordered_actions() method for readability * * \param[in,out] rsc Resource to call method for * \param[in,out] first 'First' action in an ordering * \param[in,out] then 'Then' action in an ordering * \param[in] node If not NULL, limit scope of ordering to this * node (only used when interleaving instances) * \param[in] flags Action flags for \p first for ordering purposes * \param[in] filter Action flags to limit scope of certain updates * (may include pcmk_action_optional to affect only * mandatory actions, and pe_action_runnable to * affect only runnable actions) * \param[in] type Group of enum pcmk__action_relation_flags to apply * \param[in,out] data_set Cluster working set * * \return Group of enum pcmk__updated flags indicating what was updated */ static inline uint32_t update(pe_resource_t *rsc, pe_action_t *first, pe_action_t *then, const pe_node_t *node, uint32_t flags, uint32_t filter, uint32_t type, pe_working_set_t *data_set) { return rsc->cmds->update_ordered_actions(first, then, node, flags, filter, type, data_set); } /*! * \internal * \brief Update flags for ordering's actions appropriately for ordering's flags * * \param[in,out] first First action in an ordering * \param[in,out] then Then action in an ordering * \param[in] first_flags Action flags for \p first for ordering purposes * \param[in] then_flags Action flags for \p then for ordering purposes * \param[in,out] order Action wrapper for \p first in ordering * \param[in,out] data_set Cluster working set * * \return Group of enum pcmk__updated flags */ static uint32_t update_action_for_ordering_flags(pe_action_t *first, pe_action_t *then, uint32_t first_flags, uint32_t then_flags, pe_action_wrapper_t *order, pe_working_set_t *data_set) { uint32_t changed = pcmk__updated_none; /* The node will only be used for clones. If interleaved, node will be NULL, * otherwise the ordering scope will be limited to the node. Normally, the * whole 'then' clone should restart if 'first' is restarted, so then->node * is needed. */ pe_node_t *node = then->node; if (pcmk_is_set(order->type, pe_order_implies_then_on_node)) { /* For unfencing, only instances of 'then' on the same node as 'first' * (the unfencing operation) should restart, so reset node to * first->node, at which point this case is handled like a normal * pcmk__ar_first_implies_then. */ pe__clear_order_flags(order->type, pe_order_implies_then_on_node); pe__set_order_flags(order->type, pcmk__ar_first_implies_then); node = first->node; pe_rsc_trace(then->rsc, "%s then %s: mapped pe_order_implies_then_on_node to " "pcmk__ar_first_implies_then on %s", first->uuid, then->uuid, pe__node_name(node)); } if (pcmk_is_set(order->type, pcmk__ar_first_implies_then)) { if (then->rsc != NULL) { changed |= update(then->rsc, first, then, node, first_flags & pcmk_action_optional, pcmk_action_optional, pcmk__ar_first_implies_then, data_set); } else if (!pcmk_is_set(first_flags, pcmk_action_optional) && pcmk_is_set(then->flags, pcmk_action_optional)) { pe__clear_action_flags(then, pcmk_action_optional); pcmk__set_updated_flags(changed, first, pcmk__updated_then); } pe_rsc_trace(then->rsc, "%s then %s: %s after pcmk__ar_first_implies_then", first->uuid, then->uuid, (changed? "changed" : "unchanged")); } if (pcmk_is_set(order->type, pe_order_restart) && (then->rsc != NULL)) { enum pe_action_flags restart = pcmk_action_optional |pcmk_action_runnable; changed |= update(then->rsc, first, then, node, first_flags, restart, pe_order_restart, data_set); pe_rsc_trace(then->rsc, "%s then %s: %s after pe_order_restart", first->uuid, then->uuid, (changed? "changed" : "unchanged")); } if (pcmk_is_set(order->type, pcmk__ar_then_implies_first)) { if (first->rsc != NULL) { changed |= update(first->rsc, first, then, node, first_flags, pcmk_action_optional, pcmk__ar_then_implies_first, data_set); } else if (!pcmk_is_set(first_flags, pcmk_action_optional) && pcmk_is_set(first->flags, pcmk_action_runnable)) { pe__clear_action_flags(first, pcmk_action_runnable); pcmk__set_updated_flags(changed, first, pcmk__updated_first); } pe_rsc_trace(then->rsc, "%s then %s: %s after pcmk__ar_then_implies_first", first->uuid, then->uuid, (changed? "changed" : "unchanged")); } if (pcmk_is_set(order->type, pcmk__ar_promoted_then_implies_first)) { if (then->rsc != NULL) { changed |= update(then->rsc, first, then, node, first_flags & pcmk_action_optional, pcmk_action_optional, pcmk__ar_promoted_then_implies_first, data_set); } pe_rsc_trace(then->rsc, "%s then %s: %s after pcmk__ar_promoted_then_implies_first", first->uuid, then->uuid, (changed? "changed" : "unchanged")); } if (pcmk_is_set(order->type, pe_order_one_or_more)) { if (then->rsc != NULL) { changed |= update(then->rsc, first, then, node, first_flags, pcmk_action_runnable, pe_order_one_or_more, data_set); } else if (pcmk_is_set(first_flags, pcmk_action_runnable)) { // We have another runnable instance of "first" then->runnable_before++; /* Mark "then" as runnable if it requires a certain number of * "before" instances to be runnable, and they now are. */ if ((then->runnable_before >= then->required_runnable_before) && !pcmk_is_set(then->flags, pcmk_action_runnable)) { pe__set_action_flags(then, pcmk_action_runnable); pcmk__set_updated_flags(changed, first, pcmk__updated_then); } } pe_rsc_trace(then->rsc, "%s then %s: %s after pe_order_one_or_more", first->uuid, then->uuid, (changed? "changed" : "unchanged")); } if (pcmk_is_set(order->type, pe_order_probe) && (then->rsc != NULL)) { if (!pcmk_is_set(first_flags, pcmk_action_runnable) && (first->rsc->running_on != NULL)) { pe_rsc_trace(then->rsc, "%s then %s: ignoring because first is stopping", first->uuid, then->uuid); order->type = pcmk__ar_none; } else { changed |= update(then->rsc, first, then, node, first_flags, - pcmk_action_runnable, pe_order_runnable_left, - data_set); + pcmk_action_runnable, + pcmk__ar_unrunnable_first_blocks, data_set); } pe_rsc_trace(then->rsc, "%s then %s: %s after pe_order_probe", first->uuid, then->uuid, (changed? "changed" : "unchanged")); } - if (pcmk_is_set(order->type, pe_order_runnable_left)) { + if (pcmk_is_set(order->type, pcmk__ar_unrunnable_first_blocks)) { if (then->rsc != NULL) { changed |= update(then->rsc, first, then, node, first_flags, - pcmk_action_runnable, pe_order_runnable_left, - data_set); + pcmk_action_runnable, + pcmk__ar_unrunnable_first_blocks, data_set); } else if (!pcmk_is_set(first_flags, pcmk_action_runnable) && pcmk_is_set(then->flags, pcmk_action_runnable)) { pe__clear_action_flags(then, pcmk_action_runnable); pcmk__set_updated_flags(changed, first, pcmk__updated_then); } - pe_rsc_trace(then->rsc, "%s then %s: %s after pe_order_runnable_left", + pe_rsc_trace(then->rsc, + "%s then %s: %s after pcmk__ar_unrunnable_first_blocks", first->uuid, then->uuid, (changed? "changed" : "unchanged")); } if (pcmk_is_set(order->type, pcmk__ar_unmigratable_then_blocks)) { if (then->rsc != NULL) { changed |= update(then->rsc, first, then, node, first_flags, pcmk_action_optional, pcmk__ar_unmigratable_then_blocks, data_set); } pe_rsc_trace(then->rsc, "%s then %s: %s after " "pcmk__ar_unmigratable_then_blocks", first->uuid, then->uuid, (changed? "changed" : "unchanged")); } if (pcmk_is_set(order->type, pe_order_pseudo_left)) { if (then->rsc != NULL) { changed |= update(then->rsc, first, then, node, first_flags, pcmk_action_optional, pe_order_pseudo_left, data_set); } pe_rsc_trace(then->rsc, "%s then %s: %s after pe_order_pseudo_left", first->uuid, then->uuid, (changed? "changed" : "unchanged")); } if (pcmk_is_set(order->type, pcmk__ar_ordered)) { if (then->rsc != NULL) { changed |= update(then->rsc, first, then, node, first_flags, pcmk_action_runnable, pcmk__ar_ordered, data_set); } pe_rsc_trace(then->rsc, "%s then %s: %s after pcmk__ar_ordered", first->uuid, then->uuid, (changed? "changed" : "unchanged")); } if (pcmk_is_set(order->type, pe_order_asymmetrical)) { if (then->rsc != NULL) { changed |= update(then->rsc, first, then, node, first_flags, pcmk_action_runnable, pe_order_asymmetrical, data_set); } pe_rsc_trace(then->rsc, "%s then %s: %s after pe_order_asymmetrical", first->uuid, then->uuid, (changed? "changed" : "unchanged")); } if (pcmk_is_set(first->flags, pcmk_action_runnable) && pcmk_is_set(order->type, pe_order_implies_then_printed) && !pcmk_is_set(first_flags, pcmk_action_optional)) { pe_rsc_trace(then->rsc, "%s will be in graph because %s is required", then->uuid, first->uuid); pe__set_action_flags(then, pcmk_action_always_in_graph); // Don't bother marking 'then' as changed just for this } if (pcmk_is_set(order->type, pe_order_implies_first_printed) && !pcmk_is_set(then_flags, pcmk_action_optional)) { pe_rsc_trace(then->rsc, "%s will be in graph because %s is required", first->uuid, then->uuid); pe__set_action_flags(first, pcmk_action_always_in_graph); // Don't bother marking 'first' as changed just for this } if (pcmk_any_flags_set(order->type, pcmk__ar_first_implies_then |pcmk__ar_then_implies_first |pe_order_restart) && (first->rsc != NULL) && !pcmk_is_set(first->rsc->flags, pcmk_rsc_managed) && pcmk_is_set(first->rsc->flags, pcmk_rsc_blocked) && !pcmk_is_set(first->flags, pcmk_action_runnable) && pcmk__str_eq(first->task, PCMK_ACTION_STOP, pcmk__str_none)) { if (pcmk_is_set(then->flags, pcmk_action_runnable)) { pe__clear_action_flags(then, pcmk_action_runnable); pcmk__set_updated_flags(changed, first, pcmk__updated_then); } pe_rsc_trace(then->rsc, "%s then %s: %s after checking whether first " "is blocked, unmanaged, unrunnable stop", first->uuid, then->uuid, (changed? "changed" : "unchanged")); } return changed; } // Convenience macros for logging action properties #define action_type_str(flags) \ (pcmk_is_set((flags), pcmk_action_pseudo)? "pseudo-action" : "action") #define action_optional_str(flags) \ (pcmk_is_set((flags), pcmk_action_optional)? "optional" : "required") #define action_runnable_str(flags) \ (pcmk_is_set((flags), pcmk_action_runnable)? "runnable" : "unrunnable") #define action_node_str(a) \ (((a)->node == NULL)? "no node" : (a)->node->details->uname) /*! * \internal * \brief Update an action's flags for all orderings where it is "then" * * \param[in,out] then Action to update * \param[in,out] data_set Cluster working set */ void pcmk__update_action_for_orderings(pe_action_t *then, pe_working_set_t *data_set) { GList *lpc = NULL; uint32_t changed = pcmk__updated_none; int last_flags = then->flags; pe_rsc_trace(then->rsc, "Updating %s %s (%s %s) on %s", action_type_str(then->flags), then->uuid, action_optional_str(then->flags), action_runnable_str(then->flags), action_node_str(then)); if (pcmk_is_set(then->flags, pcmk_action_min_runnable)) { /* Initialize current known "runnable before" actions. As * update_action_for_ordering_flags() is called for each of then's * before actions, this number will increment as runnable 'first' * actions are encountered. */ then->runnable_before = 0; if (then->required_runnable_before == 0) { /* @COMPAT This ordering constraint uses the deprecated * "require-all=false" attribute. Treat it like "clone-min=1". */ then->required_runnable_before = 1; } /* The pe_order_one_or_more clause of update_action_for_ordering_flags() * (called below) will reset runnable if appropriate. */ pe__clear_action_flags(then, pcmk_action_runnable); } for (lpc = then->actions_before; lpc != NULL; lpc = lpc->next) { pe_action_wrapper_t *other = (pe_action_wrapper_t *) lpc->data; pe_action_t *first = other->action; pe_node_t *then_node = then->node; pe_node_t *first_node = first->node; if ((first->rsc != NULL) && (first->rsc->variant == pcmk_rsc_variant_group) && pcmk__str_eq(first->task, PCMK_ACTION_START, pcmk__str_none)) { first_node = first->rsc->fns->location(first->rsc, NULL, FALSE); if (first_node != NULL) { pe_rsc_trace(first->rsc, "Found %s for 'first' %s", pe__node_name(first_node), first->uuid); } } if ((then->rsc != NULL) && (then->rsc->variant == pcmk_rsc_variant_group) && pcmk__str_eq(then->task, PCMK_ACTION_START, pcmk__str_none)) { then_node = then->rsc->fns->location(then->rsc, NULL, FALSE); if (then_node != NULL) { pe_rsc_trace(then->rsc, "Found %s for 'then' %s", pe__node_name(then_node), then->uuid); } } // Disable constraint if it only applies when on same node, but isn't if (pcmk_is_set(other->type, pe_order_same_node) && (first_node != NULL) && (then_node != NULL) && !pe__same_node(first_node, then_node)) { pe_rsc_trace(then->rsc, "Disabled ordering %s on %s then %s on %s: " "not same node", other->action->uuid, pe__node_name(first_node), then->uuid, pe__node_name(then_node)); other->type = pcmk__ar_none; continue; } pcmk__clear_updated_flags(changed, then, pcmk__updated_first); if ((first->rsc != NULL) && pcmk_is_set(other->type, pe_order_then_cancels_first) && !pcmk_is_set(then->flags, pcmk_action_optional)) { /* 'then' is required, so we must abandon 'first' * (e.g. a required stop cancels any agent reload). */ pe__set_action_flags(other->action, pcmk_action_optional); if (!strcmp(first->task, PCMK_ACTION_RELOAD_AGENT)) { pe__clear_resource_flags(first->rsc, pcmk_rsc_reload); } } if ((first->rsc != NULL) && (then->rsc != NULL) && (first->rsc != then->rsc) && !is_parent(then->rsc, first->rsc)) { first = action_for_ordering(first); } if (first != other->action) { pe_rsc_trace(then->rsc, "Ordering %s after %s instead of %s", then->uuid, first->uuid, other->action->uuid); } pe_rsc_trace(then->rsc, "%s (%#.6x) then %s (%#.6x): type=%#.6x node=%s", first->uuid, first->flags, then->uuid, then->flags, other->type, action_node_str(first)); if (first == other->action) { /* 'first' was not remapped (e.g. from 'start' to 'running'), which * could mean it is a non-resource action, a primitive resource * action, or already expanded. */ uint32_t first_flags, then_flags; first_flags = action_flags_for_ordering(first, then_node); then_flags = action_flags_for_ordering(then, first_node); changed |= update_action_for_ordering_flags(first, then, first_flags, then_flags, other, data_set); /* 'first' was for a complex resource (clone, group, etc), * create a new dependency if necessary */ } else if (order_actions(first, then, other->type)) { /* This was the first time 'first' and 'then' were associated, * start again to get the new actions_before list */ pcmk__set_updated_flags(changed, then, pcmk__updated_then); pe_rsc_trace(then->rsc, "Disabled ordering %s then %s in favor of %s then %s", other->action->uuid, then->uuid, first->uuid, then->uuid); other->type = pcmk__ar_none; } if (pcmk_is_set(changed, pcmk__updated_first)) { crm_trace("Re-processing %s and its 'after' actions " "because it changed", first->uuid); for (GList *lpc2 = first->actions_after; lpc2 != NULL; lpc2 = lpc2->next) { pe_action_wrapper_t *other = (pe_action_wrapper_t *) lpc2->data; pcmk__update_action_for_orderings(other->action, data_set); } pcmk__update_action_for_orderings(first, data_set); } } if (pcmk_is_set(then->flags, pcmk_action_min_runnable)) { if (last_flags == then->flags) { pcmk__clear_updated_flags(changed, then, pcmk__updated_then); } else { pcmk__set_updated_flags(changed, then, pcmk__updated_then); } } if (pcmk_is_set(changed, pcmk__updated_then)) { crm_trace("Re-processing %s and its 'after' actions because it changed", then->uuid); if (pcmk_is_set(last_flags, pcmk_action_runnable) && !pcmk_is_set(then->flags, pcmk_action_runnable)) { pcmk__block_colocation_dependents(then); } pcmk__update_action_for_orderings(then, data_set); for (lpc = then->actions_after; lpc != NULL; lpc = lpc->next) { pe_action_wrapper_t *other = (pe_action_wrapper_t *) lpc->data; pcmk__update_action_for_orderings(other->action, data_set); } } } static inline bool is_primitive_action(const pe_action_t *action) { return (action != NULL) && (action->rsc != NULL) && (action->rsc->variant == pcmk_rsc_variant_primitive); } /*! * \internal * \brief Clear a single action flag and set reason text * * \param[in,out] action Action whose flag should be cleared * \param[in] flag Action flag that should be cleared * \param[in] reason Action that is the reason why flag is being cleared */ #define clear_action_flag_because(action, flag, reason) do { \ if (pcmk_is_set((action)->flags, (flag))) { \ pe__clear_action_flags(action, flag); \ if ((action)->rsc != (reason)->rsc) { \ char *reason_text = pe__action2reason((reason), (flag)); \ pe_action_set_reason((action), reason_text, false); \ free(reason_text); \ } \ } \ } while (0) /*! * \internal * \brief Update actions in an asymmetric ordering * * If the "first" action in an asymmetric ordering is unrunnable, make the * "second" action unrunnable as well, if appropriate. * * \param[in] first 'First' action in an asymmetric ordering * \param[in,out] then 'Then' action in an asymmetric ordering */ static void handle_asymmetric_ordering(const pe_action_t *first, pe_action_t *then) { /* Only resource actions after an unrunnable 'first' action need updates for * asymmetric ordering. */ if ((then->rsc == NULL) || pcmk_is_set(first->flags, pcmk_action_runnable)) { return; } // Certain optional 'then' actions are unaffected by unrunnable 'first' if (pcmk_is_set(then->flags, pcmk_action_optional)) { enum rsc_role_e then_rsc_role = then->rsc->fns->state(then->rsc, TRUE); if ((then_rsc_role == pcmk_role_stopped) && pcmk__str_eq(then->task, PCMK_ACTION_STOP, pcmk__str_none)) { /* If 'then' should stop after 'first' but is already stopped, the * ordering is irrelevant. */ return; } else if ((then_rsc_role >= pcmk_role_started) && pcmk__str_eq(then->task, PCMK_ACTION_START, pcmk__str_none) && pe__rsc_running_on_only(then->rsc, then->node)) { /* Similarly if 'then' should start after 'first' but is already * started on a single node. */ return; } } // 'First' can't run, so 'then' can't either clear_action_flag_because(then, pcmk_action_optional, first); clear_action_flag_because(then, pcmk_action_runnable, first); } /*! * \internal * \brief Set action bits appropriately when pe_restart_order is used * * \param[in,out] first 'First' action in an ordering with pe_restart_order * \param[in,out] then 'Then' action in an ordering with pe_restart_order * \param[in] filter What action flags to care about * * \note pe_restart_order is set for "stop resource before starting it" and * "stop later group member before stopping earlier group member" */ static void handle_restart_ordering(pe_action_t *first, pe_action_t *then, uint32_t filter) { const char *reason = NULL; CRM_ASSERT(is_primitive_action(first)); CRM_ASSERT(is_primitive_action(then)); // We need to update the action in two cases: // ... if 'then' is required if (pcmk_is_set(filter, pcmk_action_optional) && !pcmk_is_set(then->flags, pcmk_action_optional)) { reason = "restart"; } /* ... if 'then' is unrunnable action on same resource (if a resource * should restart but can't start, we still want to stop) */ if (pcmk_is_set(filter, pcmk_action_runnable) && !pcmk_is_set(then->flags, pcmk_action_runnable) && pcmk_is_set(then->rsc->flags, pcmk_rsc_managed) && (first->rsc == then->rsc)) { reason = "stop"; } if (reason == NULL) { return; } pe_rsc_trace(first->rsc, "Handling %s -> %s for %s", first->uuid, then->uuid, reason); // Make 'first' required if it is runnable if (pcmk_is_set(first->flags, pcmk_action_runnable)) { clear_action_flag_because(first, pcmk_action_optional, then); } // Make 'first' required if 'then' is required if (!pcmk_is_set(then->flags, pcmk_action_optional)) { clear_action_flag_because(first, pcmk_action_optional, then); } // Make 'first' unmigratable if 'then' is unmigratable if (!pcmk_is_set(then->flags, pcmk_action_migratable)) { clear_action_flag_because(first, pcmk_action_migratable, then); } // Make 'then' unrunnable if 'first' is required but unrunnable if (!pcmk_is_set(first->flags, pcmk_action_optional) && !pcmk_is_set(first->flags, pcmk_action_runnable)) { clear_action_flag_because(then, pcmk_action_runnable, first); } } /*! * \internal * \brief Update two actions according to an ordering between them * * Given information about an ordering of two actions, update the actions' flags * (and runnable_before members if appropriate) as appropriate for the ordering. * Effects may cascade to other orderings involving the actions as well. * * \param[in,out] first 'First' action in an ordering * \param[in,out] then 'Then' action in an ordering * \param[in] node If not NULL, limit scope of ordering to this node * (ignored) * \param[in] flags Action flags for \p first for ordering purposes * \param[in] filter Action flags to limit scope of certain updates (may * include pcmk_action_optional to affect only * mandatory actions, and pcmk_action_runnable to * affect only runnable actions) * \param[in] type Group of enum pcmk__action_relation_flags to apply * \param[in,out] data_set Cluster working set * * \return Group of enum pcmk__updated flags indicating what was updated */ uint32_t pcmk__update_ordered_actions(pe_action_t *first, pe_action_t *then, const pe_node_t *node, uint32_t flags, uint32_t filter, uint32_t type, pe_working_set_t *data_set) { uint32_t changed = pcmk__updated_none; uint32_t then_flags = 0U; uint32_t first_flags = 0U; CRM_ASSERT((first != NULL) && (then != NULL) && (data_set != NULL)); then_flags = then->flags; first_flags = first->flags; if (pcmk_is_set(type, pe_order_asymmetrical)) { handle_asymmetric_ordering(first, then); } if (pcmk_is_set(type, pcmk__ar_then_implies_first) && !pcmk_is_set(then_flags, pcmk_action_optional)) { // Then is required, and implies first should be, too if (pcmk_is_set(filter, pcmk_action_optional) && !pcmk_is_set(flags, pcmk_action_optional) && pcmk_is_set(first_flags, pcmk_action_optional)) { clear_action_flag_because(first, pcmk_action_optional, then); } if (pcmk_is_set(flags, pcmk_action_migratable) && !pcmk_is_set(then->flags, pcmk_action_migratable)) { clear_action_flag_because(first, pcmk_action_migratable, then); } } if (pcmk_is_set(type, pcmk__ar_promoted_then_implies_first) && (then->rsc != NULL) && (then->rsc->role == pcmk_role_promoted) && pcmk_is_set(filter, pcmk_action_optional) && !pcmk_is_set(then->flags, pcmk_action_optional)) { clear_action_flag_because(first, pcmk_action_optional, then); if (pcmk_is_set(first->flags, pcmk_action_migratable) && !pcmk_is_set(then->flags, pcmk_action_migratable)) { clear_action_flag_because(first, pcmk_action_migratable, then); } } if (pcmk_is_set(type, pcmk__ar_unmigratable_then_blocks) && pcmk_is_set(filter, pcmk_action_optional)) { if (!pcmk_all_flags_set(then->flags, pcmk_action_migratable |pcmk_action_runnable)) { clear_action_flag_because(first, pcmk_action_runnable, then); } if (!pcmk_is_set(then->flags, pcmk_action_optional)) { clear_action_flag_because(first, pcmk_action_optional, then); } } if (pcmk_is_set(type, pe_order_pseudo_left) && pcmk_is_set(filter, pcmk_action_optional) && !pcmk_is_set(first->flags, pcmk_action_runnable)) { clear_action_flag_because(then, pcmk_action_migratable, first); pe__clear_action_flags(then, pcmk_action_pseudo); } - if (pcmk_is_set(type, pe_order_runnable_left) + if (pcmk_is_set(type, pcmk__ar_unrunnable_first_blocks) && pcmk_is_set(filter, pcmk_action_runnable) && pcmk_is_set(then->flags, pcmk_action_runnable) && !pcmk_is_set(flags, pcmk_action_runnable)) { clear_action_flag_because(then, pcmk_action_runnable, first); clear_action_flag_because(then, pcmk_action_migratable, first); } if (pcmk_is_set(type, pcmk__ar_first_implies_then) && pcmk_is_set(filter, pcmk_action_optional) && pcmk_is_set(then->flags, pcmk_action_optional) && !pcmk_is_set(flags, pcmk_action_optional) && !pcmk_is_set(first->flags, pcmk_action_migratable)) { clear_action_flag_because(then, pcmk_action_optional, first); } if (pcmk_is_set(type, pe_order_restart)) { handle_restart_ordering(first, then, filter); } if (then_flags != then->flags) { pcmk__set_updated_flags(changed, first, pcmk__updated_then); pe_rsc_trace(then->rsc, "%s on %s: flags are now %#.6x (was %#.6x) " "because of 'first' %s (%#.6x)", then->uuid, pe__node_name(then->node), then->flags, then_flags, first->uuid, first->flags); if ((then->rsc != NULL) && (then->rsc->parent != NULL)) { // Required to handle "X_stop then X_start" for cloned groups pcmk__update_action_for_orderings(then, data_set); } } if (first_flags != first->flags) { pcmk__set_updated_flags(changed, first, pcmk__updated_first); pe_rsc_trace(first->rsc, "%s on %s: flags are now %#.6x (was %#.6x) " "because of 'then' %s (%#.6x)", first->uuid, pe__node_name(first->node), first->flags, first_flags, then->uuid, then->flags); } return changed; } /*! * \internal * \brief Trace-log an action (optionally with its dependent actions) * * \param[in] pre_text If not NULL, prefix the log with this plus ": " * \param[in] action Action to log * \param[in] details If true, recursively log dependent actions */ void pcmk__log_action(const char *pre_text, const pe_action_t *action, bool details) { const char *node_uname = NULL; const char *node_uuid = NULL; const char *desc = NULL; CRM_CHECK(action != NULL, return); if (!pcmk_is_set(action->flags, pcmk_action_pseudo)) { if (action->node != NULL) { node_uname = action->node->details->uname; node_uuid = action->node->details->id; } else { node_uname = ""; } } switch (text2task(action->task)) { case pcmk_action_fence: case pcmk_action_shutdown: if (pcmk_is_set(action->flags, pcmk_action_pseudo)) { desc = "Pseudo "; } else if (pcmk_is_set(action->flags, pcmk_action_optional)) { desc = "Optional "; } else if (!pcmk_is_set(action->flags, pcmk_action_runnable)) { desc = "!!Non-Startable!! "; } else { desc = "(Provisional) "; } crm_trace("%s%s%sAction %d: %s%s%s%s%s%s", ((pre_text == NULL)? "" : pre_text), ((pre_text == NULL)? "" : ": "), desc, action->id, action->uuid, (node_uname? "\ton " : ""), (node_uname? node_uname : ""), (node_uuid? "\t\t(" : ""), (node_uuid? node_uuid : ""), (node_uuid? ")" : "")); break; default: if (pcmk_is_set(action->flags, pcmk_action_optional)) { desc = "Optional "; } else if (pcmk_is_set(action->flags, pcmk_action_pseudo)) { desc = "Pseudo "; } else if (!pcmk_is_set(action->flags, pcmk_action_runnable)) { desc = "!!Non-Startable!! "; } else { desc = "(Provisional) "; } crm_trace("%s%s%sAction %d: %s %s%s%s%s%s%s", ((pre_text == NULL)? "" : pre_text), ((pre_text == NULL)? "" : ": "), desc, action->id, action->uuid, (action->rsc? action->rsc->id : ""), (node_uname? "\ton " : ""), (node_uname? node_uname : ""), (node_uuid? "\t\t(" : ""), (node_uuid? node_uuid : ""), (node_uuid? ")" : "")); break; } if (details) { const GList *iter = NULL; const pe_action_wrapper_t *other = NULL; crm_trace("\t\t====== Preceding Actions"); for (iter = action->actions_before; iter != NULL; iter = iter->next) { other = (const pe_action_wrapper_t *) iter->data; pcmk__log_action("\t\t", other->action, false); } crm_trace("\t\t====== Subsequent Actions"); for (iter = action->actions_after; iter != NULL; iter = iter->next) { other = (const pe_action_wrapper_t *) iter->data; pcmk__log_action("\t\t", other->action, false); } crm_trace("\t\t====== End"); } else { crm_trace("\t\t(before=%d, after=%d)", g_list_length(action->actions_before), g_list_length(action->actions_after)); } } /*! * \internal * \brief Create a new shutdown action for a node * * \param[in,out] node Node being shut down * * \return Newly created shutdown action for \p node */ pe_action_t * pcmk__new_shutdown_action(pe_node_t *node) { char *shutdown_id = NULL; pe_action_t *shutdown_op = NULL; CRM_ASSERT(node != NULL); shutdown_id = crm_strdup_printf("%s-%s", PCMK_ACTION_DO_SHUTDOWN, node->details->uname); shutdown_op = custom_action(NULL, shutdown_id, PCMK_ACTION_DO_SHUTDOWN, node, FALSE, TRUE, node->details->data_set); pcmk__order_stops_before_shutdown(node, shutdown_op); add_hash_param(shutdown_op->meta, XML_ATTR_TE_NOWAIT, XML_BOOLEAN_TRUE); return shutdown_op; } /*! * \internal * \brief Calculate and add an operation digest to XML * * Calculate an operation digest, which enables us to later determine when a * restart is needed due to the resource's parameters being changed, and add it * to given XML. * * \param[in] op Operation result from executor * \param[in,out] update XML to add digest to */ static void add_op_digest_to_xml(const lrmd_event_data_t *op, xmlNode *update) { char *digest = NULL; xmlNode *args_xml = NULL; if (op->params == NULL) { return; } args_xml = create_xml_node(NULL, XML_TAG_PARAMS); g_hash_table_foreach(op->params, hash2field, args_xml); pcmk__filter_op_for_digest(args_xml); digest = calculate_operation_digest(args_xml, NULL); crm_xml_add(update, XML_LRM_ATTR_OP_DIGEST, digest); free_xml(args_xml); free(digest); } #define FAKE_TE_ID "xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxx" /*! * \internal * \brief Create XML for resource operation history update * * \param[in,out] parent Parent XML node to add to * \param[in,out] op Operation event data * \param[in] caller_version DC feature set * \param[in] target_rc Expected result of operation * \param[in] node Name of node on which operation was performed * \param[in] origin Arbitrary description of update source * * \return Newly created XML node for history update */ xmlNode * pcmk__create_history_xml(xmlNode *parent, lrmd_event_data_t *op, const char *caller_version, int target_rc, const char *node, const char *origin) { char *key = NULL; char *magic = NULL; char *op_id = NULL; char *op_id_additional = NULL; char *local_user_data = NULL; const char *exit_reason = NULL; xmlNode *xml_op = NULL; const char *task = NULL; CRM_CHECK(op != NULL, return NULL); crm_trace("Creating history XML for %s-interval %s action for %s on %s " "(DC version: %s, origin: %s)", pcmk__readable_interval(op->interval_ms), op->op_type, op->rsc_id, ((node == NULL)? "no node" : node), caller_version, origin); task = op->op_type; /* Record a successful agent reload as a start, and a failed one as a * monitor, to make life easier for the scheduler when determining the * current state. * * @COMPAT We should check "reload" here only if the operation was for a * pre-OCF-1.1 resource agent, but we don't know that here, and we should * only ever get results for actions scheduled by us, so we can reasonably * assume any "reload" is actually a pre-1.1 agent reload. */ if (pcmk__str_any_of(task, PCMK_ACTION_RELOAD, PCMK_ACTION_RELOAD_AGENT, NULL)) { if (op->op_status == PCMK_EXEC_DONE) { task = PCMK_ACTION_START; } else { task = PCMK_ACTION_MONITOR; } } key = pcmk__op_key(op->rsc_id, task, op->interval_ms); if (pcmk__str_eq(task, PCMK_ACTION_NOTIFY, pcmk__str_none)) { const char *n_type = crm_meta_value(op->params, "notify_type"); const char *n_task = crm_meta_value(op->params, "notify_operation"); CRM_LOG_ASSERT(n_type != NULL); CRM_LOG_ASSERT(n_task != NULL); op_id = pcmk__notify_key(op->rsc_id, n_type, n_task); if (op->op_status != PCMK_EXEC_PENDING) { /* Ignore notify errors. * * @TODO It might be better to keep the correct result here, and * ignore it in process_graph_event(). */ lrmd__set_result(op, PCMK_OCF_OK, PCMK_EXEC_DONE, NULL); } /* Migration history is preserved separately, which usually matters for * multiple nodes and is important for future cluster transitions. */ } else if (pcmk__str_any_of(op->op_type, PCMK_ACTION_MIGRATE_TO, PCMK_ACTION_MIGRATE_FROM, NULL)) { op_id = strdup(key); } else if (did_rsc_op_fail(op, target_rc)) { op_id = pcmk__op_key(op->rsc_id, "last_failure", 0); if (op->interval_ms == 0) { // Ensure 'last' gets updated, in case record-pending is true op_id_additional = pcmk__op_key(op->rsc_id, "last", 0); } exit_reason = op->exit_reason; } else if (op->interval_ms > 0) { op_id = strdup(key); } else { op_id = pcmk__op_key(op->rsc_id, "last", 0); } again: xml_op = pcmk__xe_match(parent, XML_LRM_TAG_RSC_OP, XML_ATTR_ID, op_id); if (xml_op == NULL) { xml_op = create_xml_node(parent, XML_LRM_TAG_RSC_OP); } if (op->user_data == NULL) { crm_debug("Generating fake transition key for: " PCMK__OP_FMT " %d from %s", op->rsc_id, op->op_type, op->interval_ms, op->call_id, origin); local_user_data = pcmk__transition_key(-1, op->call_id, target_rc, FAKE_TE_ID); op->user_data = local_user_data; } if (magic == NULL) { magic = crm_strdup_printf("%d:%d;%s", op->op_status, op->rc, (const char *) op->user_data); } crm_xml_add(xml_op, XML_ATTR_ID, op_id); crm_xml_add(xml_op, XML_LRM_ATTR_TASK_KEY, key); crm_xml_add(xml_op, XML_LRM_ATTR_TASK, task); crm_xml_add(xml_op, XML_ATTR_ORIGIN, origin); crm_xml_add(xml_op, XML_ATTR_CRM_VERSION, caller_version); crm_xml_add(xml_op, XML_ATTR_TRANSITION_KEY, op->user_data); crm_xml_add(xml_op, XML_ATTR_TRANSITION_MAGIC, magic); crm_xml_add(xml_op, XML_LRM_ATTR_EXIT_REASON, pcmk__s(exit_reason, "")); crm_xml_add(xml_op, XML_LRM_ATTR_TARGET, node); // For context during triage crm_xml_add_int(xml_op, XML_LRM_ATTR_CALLID, op->call_id); crm_xml_add_int(xml_op, XML_LRM_ATTR_RC, op->rc); crm_xml_add_int(xml_op, XML_LRM_ATTR_OPSTATUS, op->op_status); crm_xml_add_ms(xml_op, XML_LRM_ATTR_INTERVAL_MS, op->interval_ms); if (compare_version("2.1", caller_version) <= 0) { if (op->t_run || op->t_rcchange || op->exec_time || op->queue_time) { crm_trace("Timing data (" PCMK__OP_FMT "): last=%u change=%u exec=%u queue=%u", op->rsc_id, op->op_type, op->interval_ms, op->t_run, op->t_rcchange, op->exec_time, op->queue_time); if ((op->interval_ms != 0) && (op->t_rcchange != 0)) { // Recurring ops may have changed rc after initial run crm_xml_add_ll(xml_op, XML_RSC_OP_LAST_CHANGE, (long long) op->t_rcchange); } else { crm_xml_add_ll(xml_op, XML_RSC_OP_LAST_CHANGE, (long long) op->t_run); } crm_xml_add_int(xml_op, XML_RSC_OP_T_EXEC, op->exec_time); crm_xml_add_int(xml_op, XML_RSC_OP_T_QUEUE, op->queue_time); } } if (pcmk__str_any_of(op->op_type, PCMK_ACTION_MIGRATE_TO, PCMK_ACTION_MIGRATE_FROM, NULL)) { /* * Record migrate_source and migrate_target always for migrate ops. */ const char *name = XML_LRM_ATTR_MIGRATE_SOURCE; crm_xml_add(xml_op, name, crm_meta_value(op->params, name)); name = XML_LRM_ATTR_MIGRATE_TARGET; crm_xml_add(xml_op, name, crm_meta_value(op->params, name)); } add_op_digest_to_xml(op, xml_op); if (op_id_additional) { free(op_id); op_id = op_id_additional; op_id_additional = NULL; goto again; } if (local_user_data) { free(local_user_data); op->user_data = NULL; } free(magic); free(op_id); free(key); return xml_op; } /*! * \internal * \brief Check whether an action shutdown-locks a resource to a node * * If the shutdown-lock cluster property is set, resources will not be recovered * on a different node if cleanly stopped, and may start only on that same node. * This function checks whether that applies to a given action, so that the * transition graph can be marked appropriately. * * \param[in] action Action to check * * \return true if \p action locks its resource to the action's node, * otherwise false */ bool pcmk__action_locks_rsc_to_node(const pe_action_t *action) { // Only resource actions taking place on resource's lock node are locked if ((action == NULL) || (action->rsc == NULL) || !pe__same_node(action->node, action->rsc->lock_node)) { return false; } /* During shutdown, only stops are locked (otherwise, another action such as * a demote would cause the controller to clear the lock) */ if (action->node->details->shutdown && (action->task != NULL) && (strcmp(action->task, PCMK_ACTION_STOP) != 0)) { return false; } return true; } /* lowest to highest */ static gint sort_action_id(gconstpointer a, gconstpointer b) { const pe_action_wrapper_t *action_wrapper2 = (const pe_action_wrapper_t *)a; const pe_action_wrapper_t *action_wrapper1 = (const pe_action_wrapper_t *)b; if (a == NULL) { return 1; } if (b == NULL) { return -1; } if (action_wrapper1->action->id < action_wrapper2->action->id) { return 1; } if (action_wrapper1->action->id > action_wrapper2->action->id) { return -1; } return 0; } /*! * \internal * \brief Remove any duplicate action inputs, merging action flags * * \param[in,out] action Action whose inputs should be checked */ void pcmk__deduplicate_action_inputs(pe_action_t *action) { GList *item = NULL; GList *next = NULL; pe_action_wrapper_t *last_input = NULL; action->actions_before = g_list_sort(action->actions_before, sort_action_id); for (item = action->actions_before; item != NULL; item = next) { pe_action_wrapper_t *input = (pe_action_wrapper_t *) item->data; next = item->next; if ((last_input != NULL) && (input->action->id == last_input->action->id)) { crm_trace("Input %s (%d) duplicate skipped for action %s (%d)", input->action->uuid, input->action->id, action->uuid, action->id); /* For the purposes of scheduling, the ordering flags no longer * matter, but crm_simulate looks at certain ones when creating a * dot graph. Combining the flags is sufficient for that purpose. */ last_input->type |= input->type; if (input->state == pe_link_dumped) { last_input->state = pe_link_dumped; } free(item->data); action->actions_before = g_list_delete_link(action->actions_before, item); } else { last_input = input; input->state = pe_link_not_dumped; } } } /*! * \internal * \brief Output all scheduled actions * * \param[in,out] data_set Cluster working set */ void pcmk__output_actions(pe_working_set_t *data_set) { pcmk__output_t *out = data_set->priv; // Output node (non-resource) actions for (GList *iter = data_set->actions; iter != NULL; iter = iter->next) { char *node_name = NULL; char *task = NULL; pe_action_t *action = (pe_action_t *) iter->data; if (action->rsc != NULL) { continue; // Resource actions will be output later } else if (pcmk_is_set(action->flags, pcmk_action_optional)) { continue; // This action was not scheduled } if (pcmk__str_eq(action->task, PCMK_ACTION_DO_SHUTDOWN, pcmk__str_none)) { task = strdup("Shutdown"); } else if (pcmk__str_eq(action->task, PCMK_ACTION_STONITH, pcmk__str_none)) { const char *op = g_hash_table_lookup(action->meta, "stonith_action"); task = crm_strdup_printf("Fence (%s)", op); } else { continue; // Don't display other node action types } if (pe__is_guest_node(action->node)) { const pe_resource_t *remote = action->node->details->remote_rsc; node_name = crm_strdup_printf("%s (resource: %s)", pe__node_name(action->node), remote->container->id); } else if (action->node != NULL) { node_name = crm_strdup_printf("%s", pe__node_name(action->node)); } out->message(out, "node-action", task, node_name, action->reason); free(node_name); free(task); } // Output resource actions for (GList *iter = data_set->resources; iter != NULL; iter = iter->next) { pe_resource_t *rsc = (pe_resource_t *) iter->data; rsc->cmds->output_actions(rsc); } } /*! * \internal * \brief Check whether action from resource history is still in configuration * * \param[in] rsc Resource that action is for * \param[in] task Action's name * \param[in] interval_ms Action's interval (in milliseconds) * * \return true if action is still in resource configuration, otherwise false */ static bool action_in_config(const pe_resource_t *rsc, const char *task, guint interval_ms) { char *key = pcmk__op_key(rsc->id, task, interval_ms); bool config = (find_rsc_op_entry(rsc, key) != NULL); free(key); return config; } /*! * \internal * \brief Get action name needed to compare digest for configuration changes * * \param[in] task Action name from history * \param[in] interval_ms Action interval (in milliseconds) * * \return Action name whose digest should be compared */ static const char * task_for_digest(const char *task, guint interval_ms) { /* Certain actions need to be compared against the parameters used to start * the resource. */ if ((interval_ms == 0) && pcmk__str_any_of(task, PCMK_ACTION_MONITOR, PCMK_ACTION_MIGRATE_FROM, PCMK_ACTION_PROMOTE, NULL)) { task = PCMK_ACTION_START; } return task; } /*! * \internal * \brief Check whether only sanitized parameters to an action changed * * When collecting CIB files for troubleshooting, crm_report will mask * sensitive resource parameters. If simulations were run using that, affected * resources would appear to need a restart, which would complicate * troubleshooting. To avoid that, we save a "secure digest" of non-sensitive * parameters. This function used that digest to check whether only masked * parameters are different. * * \param[in] xml_op Resource history entry with secure digest * \param[in] digest_data Operation digest information being compared * \param[in] data_set Cluster working set * * \return true if only sanitized parameters changed, otherwise false */ static bool only_sanitized_changed(const xmlNode *xml_op, const op_digest_cache_t *digest_data, const pe_working_set_t *data_set) { const char *digest_secure = NULL; if (!pcmk_is_set(data_set->flags, pcmk_sched_sanitized)) { // The scheduler is not being run as a simulation return false; } digest_secure = crm_element_value(xml_op, XML_LRM_ATTR_SECURE_DIGEST); return (digest_data->rc != RSC_DIGEST_MATCH) && (digest_secure != NULL) && (digest_data->digest_secure_calc != NULL) && (strcmp(digest_data->digest_secure_calc, digest_secure) == 0); } /*! * \internal * \brief Force a restart due to a configuration change * * \param[in,out] rsc Resource that action is for * \param[in] task Name of action whose configuration changed * \param[in] interval_ms Action interval (in milliseconds) * \param[in,out] node Node where resource should be restarted */ static void force_restart(pe_resource_t *rsc, const char *task, guint interval_ms, pe_node_t *node) { char *key = pcmk__op_key(rsc->id, task, interval_ms); pe_action_t *required = custom_action(rsc, key, task, NULL, FALSE, TRUE, rsc->cluster); pe_action_set_reason(required, "resource definition change", true); trigger_unfencing(rsc, node, "Device parameters changed", NULL, rsc->cluster); } /*! * \internal * \brief Schedule a reload of a resource on a node * * \param[in,out] data Resource to reload * \param[in] user_data Where resource should be reloaded */ static void schedule_reload(gpointer data, gpointer user_data) { pe_resource_t *rsc = data; const pe_node_t *node = user_data; pe_action_t *reload = NULL; // For collective resources, just call recursively for children if (rsc->variant > pcmk_rsc_variant_primitive) { g_list_foreach(rsc->children, schedule_reload, user_data); return; } // Skip the reload in certain situations if ((node == NULL) || !pcmk_is_set(rsc->flags, pcmk_rsc_managed) || pcmk_is_set(rsc->flags, pcmk_rsc_failed)) { pe_rsc_trace(rsc, "Skip reload of %s:%s%s %s", rsc->id, pcmk_is_set(rsc->flags, pcmk_rsc_managed)? "" : " unmanaged", pcmk_is_set(rsc->flags, pcmk_rsc_failed)? " failed" : "", (node == NULL)? "inactive" : node->details->uname); return; } /* If a resource's configuration changed while a start was pending, * force a full restart instead of a reload. */ if (pcmk_is_set(rsc->flags, pcmk_rsc_start_pending)) { pe_rsc_trace(rsc, "%s: preventing agent reload because start pending", rsc->id); custom_action(rsc, stop_key(rsc), PCMK_ACTION_STOP, node, FALSE, TRUE, rsc->cluster); return; } // Schedule the reload pe__set_resource_flags(rsc, pcmk_rsc_reload); reload = custom_action(rsc, reload_key(rsc), PCMK_ACTION_RELOAD_AGENT, node, FALSE, TRUE, rsc->cluster); pe_action_set_reason(reload, "resource definition change", FALSE); // Set orderings so that a required stop or demote cancels the reload pcmk__new_ordering(NULL, NULL, reload, rsc, stop_key(rsc), NULL, pcmk__ar_ordered|pe_order_then_cancels_first, rsc->cluster); pcmk__new_ordering(NULL, NULL, reload, rsc, demote_key(rsc), NULL, pcmk__ar_ordered|pe_order_then_cancels_first, rsc->cluster); } /*! * \internal * \brief Handle any configuration change for an action * * Given an action from resource history, if the resource's configuration * changed since the action was done, schedule any actions needed (restart, * reload, unfencing, rescheduling recurring actions, etc.). * * \param[in,out] rsc Resource that action is for * \param[in,out] node Node that action was on * \param[in] xml_op Action XML from resource history * * \return true if action configuration changed, otherwise false */ bool pcmk__check_action_config(pe_resource_t *rsc, pe_node_t *node, const xmlNode *xml_op) { guint interval_ms = 0; const char *task = NULL; const op_digest_cache_t *digest_data = NULL; CRM_CHECK((rsc != NULL) && (node != NULL) && (xml_op != NULL), return false); task = crm_element_value(xml_op, XML_LRM_ATTR_TASK); CRM_CHECK(task != NULL, return false); crm_element_value_ms(xml_op, XML_LRM_ATTR_INTERVAL_MS, &interval_ms); // If this is a recurring action, check whether it has been orphaned if (interval_ms > 0) { if (action_in_config(rsc, task, interval_ms)) { pe_rsc_trace(rsc, "%s-interval %s for %s on %s is in configuration", pcmk__readable_interval(interval_ms), task, rsc->id, pe__node_name(node)); } else if (pcmk_is_set(rsc->cluster->flags, pcmk_sched_cancel_removed_actions)) { pcmk__schedule_cancel(rsc, crm_element_value(xml_op, XML_LRM_ATTR_CALLID), task, interval_ms, node, "orphan"); return true; } else { pe_rsc_debug(rsc, "%s-interval %s for %s on %s is orphaned", pcmk__readable_interval(interval_ms), task, rsc->id, pe__node_name(node)); return true; } } crm_trace("Checking %s-interval %s for %s on %s for configuration changes", pcmk__readable_interval(interval_ms), task, rsc->id, pe__node_name(node)); task = task_for_digest(task, interval_ms); digest_data = rsc_action_digest_cmp(rsc, xml_op, node, rsc->cluster); if (only_sanitized_changed(xml_op, digest_data, rsc->cluster)) { if (!pcmk__is_daemon && (rsc->cluster->priv != NULL)) { pcmk__output_t *out = rsc->cluster->priv; out->info(out, "Only 'private' parameters to %s-interval %s for %s " "on %s changed: %s", pcmk__readable_interval(interval_ms), task, rsc->id, pe__node_name(node), crm_element_value(xml_op, XML_ATTR_TRANSITION_MAGIC)); } return false; } switch (digest_data->rc) { case RSC_DIGEST_RESTART: crm_log_xml_debug(digest_data->params_restart, "params:restart"); force_restart(rsc, task, interval_ms, node); return true; case RSC_DIGEST_ALL: case RSC_DIGEST_UNKNOWN: // Changes that can potentially be handled by an agent reload if (interval_ms > 0) { /* Recurring actions aren't reloaded per se, they are just * re-scheduled so the next run uses the new parameters. * The old instance will be cancelled automatically. */ crm_log_xml_debug(digest_data->params_all, "params:reschedule"); pcmk__reschedule_recurring(rsc, task, interval_ms, node); } else if (crm_element_value(xml_op, XML_LRM_ATTR_RESTART_DIGEST) != NULL) { // Agent supports reload, so use it trigger_unfencing(rsc, node, "Device parameters changed (reload)", NULL, rsc->cluster); crm_log_xml_debug(digest_data->params_all, "params:reload"); schedule_reload((gpointer) rsc, (gpointer) node); } else { pe_rsc_trace(rsc, "Restarting %s " "because agent doesn't support reload", rsc->id); crm_log_xml_debug(digest_data->params_restart, "params:restart"); force_restart(rsc, task, interval_ms, node); } return true; default: break; } return false; } /*! * \internal * \brief Create a list of resource's action history entries, sorted by call ID * * \param[in] rsc_entry Resource's status XML * \param[out] start_index Where to store index of start-like action, if any * \param[out] stop_index Where to store index of stop action, if any */ static GList * rsc_history_as_list(const xmlNode *rsc_entry, int *start_index, int *stop_index) { GList *ops = NULL; for (xmlNode *rsc_op = first_named_child(rsc_entry, XML_LRM_TAG_RSC_OP); rsc_op != NULL; rsc_op = crm_next_same_xml(rsc_op)) { ops = g_list_prepend(ops, rsc_op); } ops = g_list_sort(ops, sort_op_by_callid); calculate_active_ops(ops, start_index, stop_index); return ops; } /*! * \internal * \brief Process a resource's action history from the CIB status * * Given a resource's action history, if the resource's configuration * changed since the actions were done, schedule any actions needed (restart, * reload, unfencing, rescheduling recurring actions, clean-up, etc.). * (This also cancels recurring actions for maintenance mode, which is not * entirely related but convenient to do here.) * * \param[in] rsc_entry Resource's status XML * \param[in,out] rsc Resource whose history is being processed * \param[in,out] node Node whose history is being processed */ static void process_rsc_history(const xmlNode *rsc_entry, pe_resource_t *rsc, pe_node_t *node) { int offset = -1; int stop_index = 0; int start_index = 0; GList *sorted_op_list = NULL; if (pcmk_is_set(rsc->flags, pcmk_rsc_removed)) { if (pe_rsc_is_anon_clone(pe__const_top_resource(rsc, false))) { pe_rsc_trace(rsc, "Skipping configuration check " "for orphaned clone instance %s", rsc->id); } else { pe_rsc_trace(rsc, "Skipping configuration check and scheduling clean-up " "for orphaned resource %s", rsc->id); pcmk__schedule_cleanup(rsc, node, false); } return; } if (pe_find_node_id(rsc->running_on, node->details->id) == NULL) { if (pcmk__rsc_agent_changed(rsc, node, rsc_entry, false)) { pcmk__schedule_cleanup(rsc, node, false); } pe_rsc_trace(rsc, "Skipping configuration check for %s " "because no longer active on %s", rsc->id, pe__node_name(node)); return; } pe_rsc_trace(rsc, "Checking for configuration changes for %s on %s", rsc->id, pe__node_name(node)); if (pcmk__rsc_agent_changed(rsc, node, rsc_entry, true)) { pcmk__schedule_cleanup(rsc, node, false); } sorted_op_list = rsc_history_as_list(rsc_entry, &start_index, &stop_index); if (start_index < stop_index) { return; // Resource is stopped } for (GList *iter = sorted_op_list; iter != NULL; iter = iter->next) { xmlNode *rsc_op = (xmlNode *) iter->data; const char *task = NULL; guint interval_ms = 0; if (++offset < start_index) { // Skip actions that happened before a start continue; } task = crm_element_value(rsc_op, XML_LRM_ATTR_TASK); crm_element_value_ms(rsc_op, XML_LRM_ATTR_INTERVAL_MS, &interval_ms); if ((interval_ms > 0) && (pcmk_is_set(rsc->flags, pcmk_rsc_maintenance) || node->details->maintenance)) { // Maintenance mode cancels recurring operations pcmk__schedule_cancel(rsc, crm_element_value(rsc_op, XML_LRM_ATTR_CALLID), task, interval_ms, node, "maintenance mode"); } else if ((interval_ms > 0) || pcmk__strcase_any_of(task, PCMK_ACTION_MONITOR, PCMK_ACTION_START, PCMK_ACTION_PROMOTE, PCMK_ACTION_MIGRATE_FROM, NULL)) { /* If a resource operation failed, and the operation's definition * has changed, clear any fail count so they can be retried fresh. */ if (pe__bundle_needs_remote_name(rsc)) { /* We haven't assigned resources to nodes yet, so if the * REMOTE_CONTAINER_HACK is used, we may calculate the digest * based on the literal "#uname" value rather than the properly * substituted value. That would mistakenly make the action * definition appear to have been changed. Defer the check until * later in this case. */ pe__add_param_check(rsc_op, rsc, node, pcmk__check_active, rsc->cluster); } else if (pcmk__check_action_config(rsc, node, rsc_op) && (pe_get_failcount(node, rsc, NULL, pe_fc_effective, NULL) != 0)) { pe__clear_failcount(rsc, node, "action definition changed", rsc->cluster); } } } g_list_free(sorted_op_list); } /*! * \internal * \brief Process a node's action history from the CIB status * * Given a node's resource history, if the resource's configuration changed * since the actions were done, schedule any actions needed (restart, * reload, unfencing, rescheduling recurring actions, clean-up, etc.). * (This also cancels recurring actions for maintenance mode, which is not * entirely related but convenient to do here.) * * \param[in,out] node Node whose history is being processed * \param[in] lrm_rscs Node's from CIB status XML */ static void process_node_history(pe_node_t *node, const xmlNode *lrm_rscs) { crm_trace("Processing node history for %s", pe__node_name(node)); for (const xmlNode *rsc_entry = first_named_child(lrm_rscs, XML_LRM_TAG_RESOURCE); rsc_entry != NULL; rsc_entry = crm_next_same_xml(rsc_entry)) { if (rsc_entry->children != NULL) { GList *result = pcmk__rscs_matching_id(ID(rsc_entry), node->details->data_set); for (GList *iter = result; iter != NULL; iter = iter->next) { pe_resource_t *rsc = (pe_resource_t *) iter->data; if (rsc->variant == pcmk_rsc_variant_primitive) { process_rsc_history(rsc_entry, rsc, node); } } g_list_free(result); } } } // XPath to find a node's resource history #define XPATH_NODE_HISTORY "/" XML_TAG_CIB "/" XML_CIB_TAG_STATUS \ "/" XML_CIB_TAG_STATE "[@" XML_ATTR_UNAME "='%s']" \ "/" XML_CIB_TAG_LRM "/" XML_LRM_TAG_RESOURCES /*! * \internal * \brief Process any resource configuration changes in the CIB status * * Go through all nodes' resource history, and if a resource's configuration * changed since its actions were done, schedule any actions needed (restart, * reload, unfencing, rescheduling recurring actions, clean-up, etc.). * (This also cancels recurring actions for maintenance mode, which is not * entirely related but convenient to do here.) * * \param[in,out] data_set Cluster working set */ void pcmk__handle_rsc_config_changes(pe_working_set_t *data_set) { crm_trace("Check resource and action configuration for changes"); /* Rather than iterate through the status section, iterate through the nodes * and search for the appropriate status subsection for each. This skips * orphaned nodes and lets us eliminate some cases before searching the XML. */ for (GList *iter = data_set->nodes; iter != NULL; iter = iter->next) { pe_node_t *node = (pe_node_t *) iter->data; /* Don't bother checking actions for a node that can't run actions ... * unless it's in maintenance mode, in which case we still need to * cancel any existing recurring monitors. */ if (node->details->maintenance || pcmk__node_available(node, false, false)) { char *xpath = NULL; xmlNode *history = NULL; xpath = crm_strdup_printf(XPATH_NODE_HISTORY, node->details->uname); history = get_xpath_object(xpath, data_set->input, LOG_NEVER); free(xpath); process_node_history(node, history); } } } diff --git a/lib/pacemaker/pcmk_sched_bundle.c b/lib/pacemaker/pcmk_sched_bundle.c index fdb7c19792..29868b087a 100644 --- a/lib/pacemaker/pcmk_sched_bundle.c +++ b/lib/pacemaker/pcmk_sched_bundle.c @@ -1,1053 +1,1054 @@ /* * Copyright 2004-2023 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU General Public License version 2 * or later (GPLv2+) WITHOUT ANY WARRANTY. */ #include #include #include #include #include "libpacemaker_private.h" struct assign_data { const pe_node_t *prefer; bool stop_if_fail; }; /*! * \internal * \brief Assign a single bundle replica's resources (other than container) * * \param[in,out] replica Replica to assign * \param[in] user_data Preferred node, if any * * \return true (to indicate that any further replicas should be processed) */ static bool assign_replica(pe__bundle_replica_t *replica, void *user_data) { pe_node_t *container_host = NULL; struct assign_data *assign_data = user_data; const pe_node_t *prefer = assign_data->prefer; bool stop_if_fail = assign_data->stop_if_fail; const pe_resource_t *bundle = pe__const_top_resource(replica->container, true); if (replica->ip != NULL) { pe_rsc_trace(bundle, "Assigning bundle %s IP %s", bundle->id, replica->ip->id); replica->ip->cmds->assign(replica->ip, prefer, stop_if_fail); } container_host = replica->container->allocated_to; if (replica->remote != NULL) { if (pe__is_guest_or_remote_node(container_host)) { /* REMOTE_CONTAINER_HACK: "Nested" connection resources must be on * the same host because Pacemaker Remote only supports a single * active connection. */ pcmk__new_colocation("#replica-remote-with-host-remote", NULL, INFINITY, replica->remote, container_host->details->remote_rsc, NULL, NULL, pcmk__coloc_influence); } pe_rsc_trace(bundle, "Assigning bundle %s connection %s", bundle->id, replica->remote->id); replica->remote->cmds->assign(replica->remote, prefer, stop_if_fail); } if (replica->child != NULL) { pe_node_t *node = NULL; GHashTableIter iter; g_hash_table_iter_init(&iter, replica->child->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (gpointer *) &node)) { if (!pe__same_node(node, replica->node)) { node->weight = -INFINITY; } else if (!pcmk__threshold_reached(replica->child, node, NULL)) { node->weight = INFINITY; } } pe__set_resource_flags(replica->child->parent, pcmk_rsc_assigning); pe_rsc_trace(bundle, "Assigning bundle %s replica child %s", bundle->id, replica->child->id); replica->child->cmds->assign(replica->child, replica->node, stop_if_fail); pe__clear_resource_flags(replica->child->parent, pcmk_rsc_assigning); } return true; } /*! * \internal * \brief Assign a bundle resource to a node * * \param[in,out] rsc Resource to assign to a node * \param[in] prefer Node to prefer, if all else is equal * \param[in] stop_if_fail If \c true and a primitive descendant of \p rsc * can't be assigned to a node, set the * descendant's next role to stopped and update * existing actions * * \return Node that \p rsc is assigned to, if assigned entirely to one node * * \note If \p stop_if_fail is \c false, then \c pcmk__unassign_resource() can * completely undo the assignment. A successful assignment can be either * undone or left alone as final. A failed assignment has the same effect * as calling pcmk__unassign_resource(); there are no side effects on * roles or actions. */ pe_node_t * pcmk__bundle_assign(pe_resource_t *rsc, const pe_node_t *prefer, bool stop_if_fail) { GList *containers = NULL; pe_resource_t *bundled_resource = NULL; struct assign_data assign_data = { prefer, stop_if_fail }; CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_bundle)); pe_rsc_trace(rsc, "Assigning bundle %s", rsc->id); pe__set_resource_flags(rsc, pcmk_rsc_assigning); pe__show_node_scores(!pcmk_is_set(rsc->cluster->flags, pcmk_sched_output_scores), rsc, __func__, rsc->allowed_nodes, rsc->cluster); // Assign all containers first, so we know what nodes the bundle will be on containers = g_list_sort(pe__bundle_containers(rsc), pcmk__cmp_instance); pcmk__assign_instances(rsc, containers, pe__bundle_max(rsc), rsc->fns->max_per_node(rsc)); g_list_free(containers); // Then assign remaining replica resources pe__foreach_bundle_replica(rsc, assign_replica, (void *) &assign_data); // Finally, assign the bundled resources to each bundle node bundled_resource = pe__bundled_resource(rsc); if (bundled_resource != NULL) { pe_node_t *node = NULL; GHashTableIter iter; g_hash_table_iter_init(&iter, bundled_resource->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (gpointer *) & node)) { if (pe__node_is_bundle_instance(rsc, node)) { node->weight = 0; } else { node->weight = -INFINITY; } } bundled_resource->cmds->assign(bundled_resource, prefer, stop_if_fail); } pe__clear_resource_flags(rsc, pcmk_rsc_assigning|pcmk_rsc_unassigned); return NULL; } /*! * \internal * \brief Create actions for a bundle replica's resources (other than child) * * \param[in,out] replica Replica to create actions for * \param[in] user_data Unused * * \return true (to indicate that any further replicas should be processed) */ static bool create_replica_actions(pe__bundle_replica_t *replica, void *user_data) { if (replica->ip != NULL) { replica->ip->cmds->create_actions(replica->ip); } if (replica->container != NULL) { replica->container->cmds->create_actions(replica->container); } if (replica->remote != NULL) { replica->remote->cmds->create_actions(replica->remote); } return true; } /*! * \internal * \brief Create all actions needed for a given bundle resource * * \param[in,out] rsc Bundle resource to create actions for */ void pcmk__bundle_create_actions(pe_resource_t *rsc) { pe_action_t *action = NULL; GList *containers = NULL; pe_resource_t *bundled_resource = NULL; CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_bundle)); pe__foreach_bundle_replica(rsc, create_replica_actions, NULL); containers = pe__bundle_containers(rsc); pcmk__create_instance_actions(rsc, containers); g_list_free(containers); bundled_resource = pe__bundled_resource(rsc); if (bundled_resource != NULL) { bundled_resource->cmds->create_actions(bundled_resource); if (pcmk_is_set(bundled_resource->flags, pcmk_rsc_promotable)) { pe__new_rsc_pseudo_action(rsc, PCMK_ACTION_PROMOTE, true, true); action = pe__new_rsc_pseudo_action(rsc, PCMK_ACTION_PROMOTED, true, true); action->priority = INFINITY; pe__new_rsc_pseudo_action(rsc, PCMK_ACTION_DEMOTE, true, true); action = pe__new_rsc_pseudo_action(rsc, PCMK_ACTION_DEMOTED, true, true); action->priority = INFINITY; } } } /*! * \internal * \brief Create internal constraints for a bundle replica's resources * * \param[in,out] replica Replica to create internal constraints for * \param[in,out] user_data Replica's parent bundle * * \return true (to indicate that any further replicas should be processed) */ static bool replica_internal_constraints(pe__bundle_replica_t *replica, void *user_data) { pe_resource_t *bundle = user_data; replica->container->cmds->internal_constraints(replica->container); // Start bundle -> start replica container pcmk__order_starts(bundle, replica->container, - pe_order_runnable_left|pe_order_implies_first_printed); + pcmk__ar_unrunnable_first_blocks + |pe_order_implies_first_printed); // Stop bundle -> stop replica child and container if (replica->child != NULL) { pcmk__order_stops(bundle, replica->child, pe_order_implies_first_printed); } pcmk__order_stops(bundle, replica->container, pe_order_implies_first_printed); // Start replica container -> bundle is started pcmk__order_resource_actions(replica->container, PCMK_ACTION_START, bundle, PCMK_ACTION_RUNNING, pe_order_implies_then_printed); // Stop replica container -> bundle is stopped pcmk__order_resource_actions(replica->container, PCMK_ACTION_STOP, bundle, PCMK_ACTION_STOPPED, pe_order_implies_then_printed); if (replica->ip != NULL) { replica->ip->cmds->internal_constraints(replica->ip); // Replica IP address -> replica container (symmetric) pcmk__order_starts(replica->ip, replica->container, - pe_order_runnable_left|pe_order_preserve); + pcmk__ar_unrunnable_first_blocks|pe_order_preserve); pcmk__order_stops(replica->container, replica->ip, pcmk__ar_then_implies_first|pe_order_preserve); pcmk__new_colocation("#ip-with-container", NULL, INFINITY, replica->ip, replica->container, NULL, NULL, pcmk__coloc_influence); } if (replica->remote != NULL) { /* This handles ordering and colocating remote relative to container * (via "#resource-with-container"). Since IP is also ordered and * colocated relative to the container, we don't need to do anything * explicit here with IP. */ replica->remote->cmds->internal_constraints(replica->remote); } if (replica->child != NULL) { CRM_ASSERT(replica->remote != NULL); // "Start remote then child" is implicit in scheduler's remote logic } return true; } /*! * \internal * \brief Create implicit constraints needed for a bundle resource * * \param[in,out] rsc Bundle resource to create implicit constraints for */ void pcmk__bundle_internal_constraints(pe_resource_t *rsc) { pe_resource_t *bundled_resource = NULL; CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_bundle)); pe__foreach_bundle_replica(rsc, replica_internal_constraints, rsc); bundled_resource = pe__bundled_resource(rsc); if (bundled_resource == NULL) { return; } // Start bundle -> start bundled clone pcmk__order_resource_actions(rsc, PCMK_ACTION_START, bundled_resource, PCMK_ACTION_START, pe_order_implies_first_printed); // Bundled clone is started -> bundle is started pcmk__order_resource_actions(bundled_resource, PCMK_ACTION_RUNNING, rsc, PCMK_ACTION_RUNNING, pe_order_implies_then_printed); // Stop bundle -> stop bundled clone pcmk__order_resource_actions(rsc, PCMK_ACTION_STOP, bundled_resource, PCMK_ACTION_STOP, pe_order_implies_first_printed); // Bundled clone is stopped -> bundle is stopped pcmk__order_resource_actions(bundled_resource, PCMK_ACTION_STOPPED, rsc, PCMK_ACTION_STOPPED, pe_order_implies_then_printed); bundled_resource->cmds->internal_constraints(bundled_resource); if (!pcmk_is_set(bundled_resource->flags, pcmk_rsc_promotable)) { return; } pcmk__promotable_restart_ordering(rsc); // Demote bundle -> demote bundled clone pcmk__order_resource_actions(rsc, PCMK_ACTION_DEMOTE, bundled_resource, PCMK_ACTION_DEMOTE, pe_order_implies_first_printed); // Bundled clone is demoted -> bundle is demoted pcmk__order_resource_actions(bundled_resource, PCMK_ACTION_DEMOTED, rsc, PCMK_ACTION_DEMOTED, pe_order_implies_then_printed); // Promote bundle -> promote bundled clone pcmk__order_resource_actions(rsc, PCMK_ACTION_PROMOTE, bundled_resource, PCMK_ACTION_PROMOTE, pe_order_implies_first_printed); // Bundled clone is promoted -> bundle is promoted pcmk__order_resource_actions(bundled_resource, PCMK_ACTION_PROMOTED, rsc, PCMK_ACTION_PROMOTED, pe_order_implies_then_printed); } struct match_data { const pe_node_t *node; // Node to compare against replica pe_resource_t *container; // Replica container corresponding to node }; /*! * \internal * \brief Check whether a replica container is assigned to a given node * * \param[in] replica Replica to check * \param[in,out] user_data struct match_data with node to compare against * * \return true if the replica does not match (to indicate further replicas * should be processed), otherwise false */ static bool match_replica_container(const pe__bundle_replica_t *replica, void *user_data) { struct match_data *match_data = user_data; if (pcmk__instance_matches(replica->container, match_data->node, pcmk_role_unknown, false)) { match_data->container = replica->container; return false; // Match found, don't bother searching further replicas } return true; // No match, keep searching } /*! * \internal * \brief Get the host to which a bundle node is assigned * * \param[in] node Possible bundle node to check * * \return Node to which the container for \p node is assigned if \p node is a * bundle node, otherwise \p node itself */ static const pe_node_t * get_bundle_node_host(const pe_node_t *node) { if (pe__is_bundle_node(node)) { const pe_resource_t *container = node->details->remote_rsc->container; return container->fns->location(container, NULL, 0); } return node; } /*! * \internal * \brief Find a bundle container compatible with a dependent resource * * \param[in] dependent Dependent resource in colocation with bundle * \param[in] bundle Bundle that \p dependent is colocated with * * \return A container from \p bundle assigned to the same node as \p dependent * if assigned, otherwise assigned to any of dependent's allowed nodes, * otherwise NULL. */ static pe_resource_t * compatible_container(const pe_resource_t *dependent, const pe_resource_t *bundle) { GList *scratch = NULL; struct match_data match_data = { NULL, NULL }; // If dependent is assigned, only check there match_data.node = dependent->fns->location(dependent, NULL, 0); match_data.node = get_bundle_node_host(match_data.node); if (match_data.node != NULL) { pe__foreach_const_bundle_replica(bundle, match_replica_container, &match_data); return match_data.container; } // Otherwise, check for any of the dependent's allowed nodes scratch = g_hash_table_get_values(dependent->allowed_nodes); scratch = pcmk__sort_nodes(scratch, NULL); for (const GList *iter = scratch; iter != NULL; iter = iter->next) { match_data.node = iter->data; match_data.node = get_bundle_node_host(match_data.node); if (match_data.node == NULL) { continue; } pe__foreach_const_bundle_replica(bundle, match_replica_container, &match_data); if (match_data.container != NULL) { break; } } g_list_free(scratch); return match_data.container; } struct coloc_data { const pcmk__colocation_t *colocation; pe_resource_t *dependent; GList *container_hosts; }; /*! * \internal * \brief Apply a colocation score to replica node scores or resource priority * * \param[in] replica Replica of primary bundle resource in colocation * \param[in,out] user_data struct coloc_data for colocation being applied * * \return true (to indicate that any further replicas should be processed) */ static bool replica_apply_coloc_score(const pe__bundle_replica_t *replica, void *user_data) { struct coloc_data *coloc_data = user_data; pe_node_t *chosen = NULL; if (coloc_data->colocation->score < INFINITY) { replica->container->cmds->apply_coloc_score(coloc_data->dependent, replica->container, coloc_data->colocation, false); return true; } chosen = replica->container->fns->location(replica->container, NULL, 0); if ((chosen == NULL) || is_set_recursive(replica->container, pcmk_rsc_blocked, true)) { return true; } if ((coloc_data->colocation->primary_role >= pcmk_role_promoted) && ((replica->child == NULL) || (replica->child->next_role < pcmk_role_promoted))) { return true; } pe_rsc_trace(pe__const_top_resource(replica->container, true), "Allowing mandatory colocation %s using %s @%d", coloc_data->colocation->id, pe__node_name(chosen), chosen->weight); coloc_data->container_hosts = g_list_prepend(coloc_data->container_hosts, chosen); return true; } /*! * \internal * \brief Apply a colocation's score to node scores or resource priority * * Given a colocation constraint, apply its score to the dependent's * allowed node scores (if we are still placing resources) or priority (if * we are choosing promotable clone instance roles). * * \param[in,out] dependent Dependent resource in colocation * \param[in] primary Primary resource in colocation * \param[in] colocation Colocation constraint to apply * \param[in] for_dependent true if called on behalf of dependent */ void pcmk__bundle_apply_coloc_score(pe_resource_t *dependent, const pe_resource_t *primary, const pcmk__colocation_t *colocation, bool for_dependent) { struct coloc_data coloc_data = { colocation, dependent, NULL }; /* This should never be called for the bundle itself as a dependent. * Instead, we add its colocation constraints to its containers and bundled * primitive and call the apply_coloc_score() method for them as dependents. */ CRM_ASSERT((primary != NULL) && (primary->variant == pcmk_rsc_variant_bundle) && (dependent != NULL) && (dependent->variant == pcmk_rsc_variant_primitive) && (colocation != NULL) && !for_dependent); if (pcmk_is_set(primary->flags, pcmk_rsc_unassigned)) { pe_rsc_trace(primary, "Skipping applying colocation %s " "because %s is still provisional", colocation->id, primary->id); return; } pe_rsc_trace(primary, "Applying colocation %s (%s with %s at %s)", colocation->id, dependent->id, primary->id, pcmk_readable_score(colocation->score)); /* If the constraint dependent is a clone or bundle, "dependent" here is one * of its instances. Look for a compatible instance of this bundle. */ if (colocation->dependent->variant > pcmk_rsc_variant_group) { const pe_resource_t *primary_container = compatible_container(dependent, primary); if (primary_container != NULL) { // Success, we found one pe_rsc_debug(primary, "Pairing %s with %s", dependent->id, primary_container->id); dependent->cmds->apply_coloc_score(dependent, primary_container, colocation, true); } else if (colocation->score >= INFINITY) { // Failure, and it's fatal crm_notice("%s cannot run because there is no compatible " "instance of %s to colocate with", dependent->id, primary->id); pcmk__assign_resource(dependent, NULL, true, true); } else { // Failure, but we can ignore it pe_rsc_debug(primary, "%s cannot be colocated with any instance of %s", dependent->id, primary->id); } return; } pe__foreach_const_bundle_replica(primary, replica_apply_coloc_score, &coloc_data); if (colocation->score >= INFINITY) { pcmk__colocation_intersect_nodes(dependent, primary, colocation, coloc_data.container_hosts, false); } g_list_free(coloc_data.container_hosts); } // Bundle implementation of resource_alloc_functions_t:with_this_colocations() void pcmk__with_bundle_colocations(const pe_resource_t *rsc, const pe_resource_t *orig_rsc, GList **list) { const pe_resource_t *bundled_rsc = NULL; CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_bundle) && (orig_rsc != NULL) && (list != NULL)); // The bundle itself and its containers always get its colocations if ((orig_rsc == rsc) || pcmk_is_set(orig_rsc->flags, pcmk_rsc_replica_container)) { pcmk__add_with_this_list(list, rsc->rsc_cons_lhs, orig_rsc); return; } /* The bundled resource gets the colocations if it's promotable and we've * begun choosing roles */ bundled_rsc = pe__bundled_resource(rsc); if ((bundled_rsc == NULL) || !pcmk_is_set(bundled_rsc->flags, pcmk_rsc_promotable) || (pe__const_top_resource(orig_rsc, false) != bundled_rsc)) { return; } if (orig_rsc == bundled_rsc) { if (pe__clone_flag_is_set(orig_rsc, pe__clone_promotion_constrained)) { /* orig_rsc is the clone and we're setting roles (or have already * done so) */ pcmk__add_with_this_list(list, rsc->rsc_cons_lhs, orig_rsc); } } else if (!pcmk_is_set(orig_rsc->flags, pcmk_rsc_unassigned)) { /* orig_rsc is an instance and is already assigned. If something * requests colocations for orig_rsc now, it's for setting roles. */ pcmk__add_with_this_list(list, rsc->rsc_cons_lhs, orig_rsc); } } // Bundle implementation of resource_alloc_functions_t:this_with_colocations() void pcmk__bundle_with_colocations(const pe_resource_t *rsc, const pe_resource_t *orig_rsc, GList **list) { const pe_resource_t *bundled_rsc = NULL; CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_bundle) && (orig_rsc != NULL) && (list != NULL)); // The bundle itself and its containers always get its colocations if ((orig_rsc == rsc) || pcmk_is_set(orig_rsc->flags, pcmk_rsc_replica_container)) { pcmk__add_this_with_list(list, rsc->rsc_cons, orig_rsc); return; } /* The bundled resource gets the colocations if it's promotable and we've * begun choosing roles */ bundled_rsc = pe__bundled_resource(rsc); if ((bundled_rsc == NULL) || !pcmk_is_set(bundled_rsc->flags, pcmk_rsc_promotable) || (pe__const_top_resource(orig_rsc, false) != bundled_rsc)) { return; } if (orig_rsc == bundled_rsc) { if (pe__clone_flag_is_set(orig_rsc, pe__clone_promotion_constrained)) { /* orig_rsc is the clone and we're setting roles (or have already * done so) */ pcmk__add_this_with_list(list, rsc->rsc_cons, orig_rsc); } } else if (!pcmk_is_set(orig_rsc->flags, pcmk_rsc_unassigned)) { /* orig_rsc is an instance and is already assigned. If something * requests colocations for orig_rsc now, it's for setting roles. */ pcmk__add_this_with_list(list, rsc->rsc_cons, orig_rsc); } } /*! * \internal * \brief Return action flags for a given bundle resource action * * \param[in,out] action Bundle resource action to get flags for * \param[in] node If not NULL, limit effects to this node * * \return Flags appropriate to \p action on \p node */ uint32_t pcmk__bundle_action_flags(pe_action_t *action, const pe_node_t *node) { GList *containers = NULL; uint32_t flags = 0; pe_resource_t *bundled_resource = NULL; CRM_ASSERT((action != NULL) && (action->rsc != NULL) && (action->rsc->variant == pcmk_rsc_variant_bundle)); bundled_resource = pe__bundled_resource(action->rsc); if (bundled_resource != NULL) { // Clone actions are done on the bundled clone resource, not container switch (get_complex_task(bundled_resource, action->task)) { case pcmk_action_unspecified: case pcmk_action_notify: case pcmk_action_notified: case pcmk_action_promote: case pcmk_action_promoted: case pcmk_action_demote: case pcmk_action_demoted: return pcmk__collective_action_flags(action, bundled_resource->children, node); default: break; } } containers = pe__bundle_containers(action->rsc); flags = pcmk__collective_action_flags(action, containers, node); g_list_free(containers); return flags; } /*! * \internal * \brief Apply a location constraint to a bundle replica * * \param[in,out] replica Replica to apply constraint to * \param[in,out] user_data Location constraint to apply * * \return true (to indicate that any further replicas should be processed) */ static bool apply_location_to_replica(pe__bundle_replica_t *replica, void *user_data) { pe__location_t *location = user_data; if (replica->container != NULL) { replica->container->cmds->apply_location(replica->container, location); } if (replica->ip != NULL) { replica->ip->cmds->apply_location(replica->ip, location); } return true; } /*! * \internal * \brief Apply a location constraint to a bundle resource's allowed node scores * * \param[in,out] rsc Bundle resource to apply constraint to * \param[in,out] location Location constraint to apply */ void pcmk__bundle_apply_location(pe_resource_t *rsc, pe__location_t *location) { pe_resource_t *bundled_resource = NULL; CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_bundle) && (location != NULL)); pcmk__apply_location(rsc, location); pe__foreach_bundle_replica(rsc, apply_location_to_replica, location); bundled_resource = pe__bundled_resource(rsc); if ((bundled_resource != NULL) && ((location->role_filter == pcmk_role_unpromoted) || (location->role_filter == pcmk_role_promoted))) { bundled_resource->cmds->apply_location(bundled_resource, location); bundled_resource->rsc_location = g_list_prepend( bundled_resource->rsc_location, location); } } #define XPATH_REMOTE "//nvpair[@name='" XML_RSC_ATTR_REMOTE_RA_ADDR "']" /*! * \internal * \brief Add a bundle replica's actions to transition graph * * \param[in,out] replica Replica to add to graph * \param[in] user_data Bundle that replica belongs to (for logging only) * * \return true (to indicate that any further replicas should be processed) */ static bool add_replica_actions_to_graph(pe__bundle_replica_t *replica, void *user_data) { if ((replica->remote != NULL) && (replica->container != NULL) && pe__bundle_needs_remote_name(replica->remote)) { /* REMOTE_CONTAINER_HACK: Allow remote nodes to run containers that * run pacemaker-remoted inside, without needing a separate IP for * the container. This is done by configuring the inner remote's * connection host as the magic string "#uname", then * replacing it with the underlying host when needed. */ xmlNode *nvpair = get_xpath_object(XPATH_REMOTE, replica->remote->xml, LOG_ERR); const char *calculated_addr = NULL; // Replace the value in replica->remote->xml (if appropriate) calculated_addr = pe__add_bundle_remote_name(replica->remote, replica->remote->cluster, nvpair, "value"); if (calculated_addr != NULL) { /* Since this is for the bundle as a resource, and not any * particular action, replace the value in the default * parameters (not evaluated for node). create_graph_action() * will grab it from there to replace it in node-evaluated * parameters. */ GHashTable *params = pe_rsc_params(replica->remote, NULL, replica->remote->cluster); g_hash_table_replace(params, strdup(XML_RSC_ATTR_REMOTE_RA_ADDR), strdup(calculated_addr)); } else { pe_resource_t *bundle = user_data; /* The only way to get here is if the remote connection is * neither currently running nor scheduled to run. That means we * won't be doing any operations that require addr (only start * requires it; we additionally use it to compare digests when * unpacking status, promote, and migrate_from history, but * that's already happened by this point). */ pe_rsc_info(bundle, "Unable to determine address for bundle %s " "remote connection", bundle->id); } } if (replica->ip != NULL) { replica->ip->cmds->add_actions_to_graph(replica->ip); } if (replica->container != NULL) { replica->container->cmds->add_actions_to_graph(replica->container); } if (replica->remote != NULL) { replica->remote->cmds->add_actions_to_graph(replica->remote); } return true; } /*! * \internal * \brief Add a bundle resource's actions to the transition graph * * \param[in,out] rsc Bundle resource whose actions should be added */ void pcmk__bundle_add_actions_to_graph(pe_resource_t *rsc) { pe_resource_t *bundled_resource = NULL; CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_bundle)); bundled_resource = pe__bundled_resource(rsc); if (bundled_resource != NULL) { bundled_resource->cmds->add_actions_to_graph(bundled_resource); } pe__foreach_bundle_replica(rsc, add_replica_actions_to_graph, rsc); } struct probe_data { pe_resource_t *bundle; // Bundle being probed pe_node_t *node; // Node to create probes on bool any_created; // Whether any probes have been created }; /*! * \internal * \brief Order a bundle replica's start after another replica's probe * * \param[in,out] replica Replica to order start for * \param[in,out] user_data Replica with probe to order after * * \return true (to indicate that any further replicas should be processed) */ static bool order_replica_start_after(pe__bundle_replica_t *replica, void *user_data) { pe__bundle_replica_t *probed_replica = user_data; if ((replica == probed_replica) || (replica->container == NULL)) { return true; } pcmk__new_ordering(probed_replica->container, pcmk__op_key(probed_replica->container->id, PCMK_ACTION_MONITOR, 0), NULL, replica->container, pcmk__op_key(replica->container->id, PCMK_ACTION_START, 0), NULL, pcmk__ar_ordered|pe_order_same_node, replica->container->cluster); return true; } /*! * \internal * \brief Create probes for a bundle replica's resources * * \param[in,out] replica Replica to create probes for * \param[in,out] user_data struct probe_data * * \return true (to indicate that any further replicas should be processed) */ static bool create_replica_probes(pe__bundle_replica_t *replica, void *user_data) { struct probe_data *probe_data = user_data; if ((replica->ip != NULL) && replica->ip->cmds->create_probe(replica->ip, probe_data->node)) { probe_data->any_created = true; } if ((replica->child != NULL) && pe__same_node(probe_data->node, replica->node) && replica->child->cmds->create_probe(replica->child, probe_data->node)) { probe_data->any_created = true; } if ((replica->container != NULL) && replica->container->cmds->create_probe(replica->container, probe_data->node)) { probe_data->any_created = true; /* If we're limited to one replica per host (due to * the lack of an IP range probably), then we don't * want any of our peer containers starting until * we've established that no other copies are already * running. * * Partly this is to ensure that the maximum replicas per host is * observed, but also to ensure that the containers * don't fail to start because the necessary port * mappings (which won't include an IP for uniqueness) * are already taken */ if (probe_data->bundle->fns->max_per_node(probe_data->bundle) == 1) { pe__foreach_bundle_replica(probe_data->bundle, order_replica_start_after, replica); } } if ((replica->container != NULL) && (replica->remote != NULL) && replica->remote->cmds->create_probe(replica->remote, probe_data->node)) { /* Do not probe the remote resource until we know where the container is * running. This is required for REMOTE_CONTAINER_HACK to correctly * probe remote resources. */ char *probe_uuid = pcmk__op_key(replica->remote->id, PCMK_ACTION_MONITOR, 0); pe_action_t *probe = find_first_action(replica->remote->actions, probe_uuid, NULL, probe_data->node); free(probe_uuid); if (probe != NULL) { probe_data->any_created = true; pe_rsc_trace(probe_data->bundle, "Ordering %s probe on %s", replica->remote->id, pe__node_name(probe_data->node)); pcmk__new_ordering(replica->container, pcmk__op_key(replica->container->id, PCMK_ACTION_START, 0), NULL, replica->remote, NULL, probe, pe_order_probe, probe_data->bundle->cluster); } } return true; } /*! * \internal * * \brief Schedule any probes needed for a bundle resource on a node * * \param[in,out] rsc Bundle resource to create probes for * \param[in,out] node Node to create probe on * * \return true if any probe was created, otherwise false */ bool pcmk__bundle_create_probe(pe_resource_t *rsc, pe_node_t *node) { struct probe_data probe_data = { rsc, node, false }; CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_bundle)); pe__foreach_bundle_replica(rsc, create_replica_probes, &probe_data); return probe_data.any_created; } /*! * \internal * \brief Output actions for one bundle replica * * \param[in,out] replica Replica to output actions for * \param[in] user_data Unused * * \return true (to indicate that any further replicas should be processed) */ static bool output_replica_actions(pe__bundle_replica_t *replica, void *user_data) { if (replica->ip != NULL) { replica->ip->cmds->output_actions(replica->ip); } if (replica->container != NULL) { replica->container->cmds->output_actions(replica->container); } if (replica->remote != NULL) { replica->remote->cmds->output_actions(replica->remote); } if (replica->child != NULL) { replica->child->cmds->output_actions(replica->child); } return true; } /*! * \internal * \brief Output a summary of scheduled actions for a bundle resource * * \param[in,out] rsc Bundle resource to output actions for */ void pcmk__output_bundle_actions(pe_resource_t *rsc) { CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_bundle)); pe__foreach_bundle_replica(rsc, output_replica_actions, NULL); } // Bundle implementation of resource_alloc_functions_t:add_utilization() void pcmk__bundle_add_utilization(const pe_resource_t *rsc, const pe_resource_t *orig_rsc, GList *all_rscs, GHashTable *utilization) { pe_resource_t *container = NULL; CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_bundle)); if (!pcmk_is_set(rsc->flags, pcmk_rsc_unassigned)) { return; } /* All bundle replicas are identical, so using the utilization of the first * is sufficient for any. Only the implicit container resource can have * utilization values. */ container = pe__first_container(rsc); if (container != NULL) { container->cmds->add_utilization(container, orig_rsc, all_rscs, utilization); } } // Bundle implementation of resource_alloc_functions_t:shutdown_lock() void pcmk__bundle_shutdown_lock(pe_resource_t *rsc) { CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_bundle)); // Bundles currently don't support shutdown locks } diff --git a/lib/pacemaker/pcmk_sched_clone.c b/lib/pacemaker/pcmk_sched_clone.c index d0c037340a..35c2a81e5d 100644 --- a/lib/pacemaker/pcmk_sched_clone.c +++ b/lib/pacemaker/pcmk_sched_clone.c @@ -1,709 +1,709 @@ /* * Copyright 2004-2023 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU General Public License version 2 * or later (GPLv2+) WITHOUT ANY WARRANTY. */ #include #include #include #include "libpacemaker_private.h" /*! * \internal * \brief Assign a clone resource's instances to nodes * * \param[in,out] rsc Clone resource to assign * \param[in] prefer Node to prefer, if all else is equal * \param[in] stop_if_fail If \c true and a primitive descendant of \p rsc * can't be assigned to a node, set the * descendant's next role to stopped and update * existing actions * * \return NULL (clones are not assigned to a single node) * * \note If \p stop_if_fail is \c false, then \c pcmk__unassign_resource() can * completely undo the assignment. A successful assignment can be either * undone or left alone as final. A failed assignment has the same effect * as calling pcmk__unassign_resource(); there are no side effects on * roles or actions. */ pe_node_t * pcmk__clone_assign(pe_resource_t *rsc, const pe_node_t *prefer, bool stop_if_fail) { GList *colocations = NULL; CRM_ASSERT(pe_rsc_is_clone(rsc)); if (!pcmk_is_set(rsc->flags, pcmk_rsc_unassigned)) { return NULL; // Assignment has already been done } // Detect assignment loops if (pcmk_is_set(rsc->flags, pcmk_rsc_assigning)) { pe_rsc_debug(rsc, "Breaking assignment loop involving %s", rsc->id); return NULL; } pe__set_resource_flags(rsc, pcmk_rsc_assigning); // If this clone is promotable, consider nodes' promotion scores if (pcmk_is_set(rsc->flags, pcmk_rsc_promotable)) { pcmk__add_promotion_scores(rsc); } // If this clone is colocated with any other resources, assign those first colocations = pcmk__this_with_colocations(rsc); for (GList *iter = colocations; iter != NULL; iter = iter->next) { pcmk__colocation_t *constraint = (pcmk__colocation_t *) iter->data; pe_rsc_trace(rsc, "%s: Assigning colocation %s primary %s first", rsc->id, constraint->id, constraint->primary->id); constraint->primary->cmds->assign(constraint->primary, prefer, stop_if_fail); } g_list_free(colocations); // If any resources are colocated with this one, consider their preferences colocations = pcmk__with_this_colocations(rsc); g_list_foreach(colocations, pcmk__add_dependent_scores, rsc); g_list_free(colocations); pe__show_node_scores(!pcmk_is_set(rsc->cluster->flags, pcmk_sched_output_scores), rsc, __func__, rsc->allowed_nodes, rsc->cluster); rsc->children = g_list_sort(rsc->children, pcmk__cmp_instance); pcmk__assign_instances(rsc, rsc->children, pe__clone_max(rsc), pe__clone_node_max(rsc)); if (pcmk_is_set(rsc->flags, pcmk_rsc_promotable)) { pcmk__set_instance_roles(rsc); } pe__clear_resource_flags(rsc, pcmk_rsc_unassigned|pcmk_rsc_assigning); pe_rsc_trace(rsc, "Assigned clone %s", rsc->id); return NULL; } /*! * \internal * \brief Create all actions needed for a given clone resource * * \param[in,out] rsc Clone resource to create actions for */ void pcmk__clone_create_actions(pe_resource_t *rsc) { CRM_ASSERT(pe_rsc_is_clone(rsc)); pe_rsc_trace(rsc, "Creating actions for clone %s", rsc->id); pcmk__create_instance_actions(rsc, rsc->children); if (pcmk_is_set(rsc->flags, pcmk_rsc_promotable)) { pcmk__create_promotable_actions(rsc); } } /*! * \internal * \brief Create implicit constraints needed for a clone resource * * \param[in,out] rsc Clone resource to create implicit constraints for */ void pcmk__clone_internal_constraints(pe_resource_t *rsc) { bool ordered = false; CRM_ASSERT(pe_rsc_is_clone(rsc)); pe_rsc_trace(rsc, "Creating internal constraints for clone %s", rsc->id); // Restart ordering: Stop -> stopped -> start -> started pcmk__order_resource_actions(rsc, PCMK_ACTION_STOPPED, rsc, PCMK_ACTION_START, pcmk__ar_ordered); pcmk__order_resource_actions(rsc, PCMK_ACTION_START, rsc, PCMK_ACTION_RUNNING, - pe_order_runnable_left); + pcmk__ar_unrunnable_first_blocks); pcmk__order_resource_actions(rsc, PCMK_ACTION_STOP, rsc, PCMK_ACTION_STOPPED, - pe_order_runnable_left); + pcmk__ar_unrunnable_first_blocks); // Demoted -> stop and started -> promote if (pcmk_is_set(rsc->flags, pcmk_rsc_promotable)) { pcmk__order_resource_actions(rsc, PCMK_ACTION_DEMOTED, rsc, PCMK_ACTION_STOP, pcmk__ar_ordered); pcmk__order_resource_actions(rsc, PCMK_ACTION_RUNNING, rsc, PCMK_ACTION_PROMOTE, - pe_order_runnable_left); + pcmk__ar_unrunnable_first_blocks); } ordered = pe__clone_is_ordered(rsc); if (ordered) { /* Ordered clone instances must start and stop by instance number. The * instances might have been previously shuffled for assignment or * promotion purposes, so re-sort them. */ rsc->children = g_list_sort(rsc->children, pcmk__cmp_instance_number); } for (GList *iter = rsc->children; iter != NULL; iter = iter->next) { pe_resource_t *instance = (pe_resource_t *) iter->data; instance->cmds->internal_constraints(instance); // Start clone -> start instance -> clone started - pcmk__order_starts(rsc, instance, pe_order_runnable_left + pcmk__order_starts(rsc, instance, pcmk__ar_unrunnable_first_blocks |pe_order_implies_first_printed); pcmk__order_resource_actions(instance, PCMK_ACTION_START, rsc, PCMK_ACTION_RUNNING, pe_order_implies_then_printed); // Stop clone -> stop instance -> clone stopped pcmk__order_stops(rsc, instance, pe_order_implies_first_printed); pcmk__order_resource_actions(instance, PCMK_ACTION_STOP, rsc, PCMK_ACTION_STOPPED, pe_order_implies_then_printed); /* Instances of ordered clones must be started and stopped by instance * number. Since only some instances may be starting or stopping, order * each instance relative to every later instance. */ if (ordered) { for (GList *later = iter->next; later != NULL; later = later->next) { pcmk__order_starts(instance, (pe_resource_t *) later->data, pcmk__ar_ordered); pcmk__order_stops((pe_resource_t *) later->data, instance, pcmk__ar_ordered); } } } if (pcmk_is_set(rsc->flags, pcmk_rsc_promotable)) { pcmk__order_promotable_instances(rsc); } } /*! * \internal * \brief Check whether colocated resources can be interleaved * * \param[in] colocation Colocation constraint with clone as primary * * \return true if colocated resources can be interleaved, otherwise false */ static bool can_interleave(const pcmk__colocation_t *colocation) { const pe_resource_t *dependent = colocation->dependent; // Only colocations between clone or bundle resources use interleaving if (dependent->variant <= pcmk_rsc_variant_group) { return false; } // Only the dependent needs to be marked for interleaving if (!crm_is_true(g_hash_table_lookup(dependent->meta, XML_RSC_ATTR_INTERLEAVE))) { return false; } /* @TODO Do we actually care about multiple primary instances sharing a * dependent instance? */ if (dependent->fns->max_per_node(dependent) != colocation->primary->fns->max_per_node(colocation->primary)) { pcmk__config_err("Cannot interleave %s and %s because they do not " "support the same number of instances per node", dependent->id, colocation->primary->id); return false; } return true; } /*! * \internal * \brief Apply a colocation's score to node scores or resource priority * * Given a colocation constraint, apply its score to the dependent's * allowed node scores (if we are still placing resources) or priority (if * we are choosing promotable clone instance roles). * * \param[in,out] dependent Dependent resource in colocation * \param[in] primary Primary resource in colocation * \param[in] colocation Colocation constraint to apply * \param[in] for_dependent true if called on behalf of dependent */ void pcmk__clone_apply_coloc_score(pe_resource_t *dependent, const pe_resource_t *primary, const pcmk__colocation_t *colocation, bool for_dependent) { const GList *iter = NULL; /* This should never be called for the clone itself as a dependent. Instead, * we add its colocation constraints to its instances and call the * apply_coloc_score() method for the instances as dependents. */ CRM_ASSERT(!for_dependent); CRM_ASSERT((colocation != NULL) && pe_rsc_is_clone(primary) && (dependent != NULL) && (dependent->variant == pcmk_rsc_variant_primitive)); if (pcmk_is_set(primary->flags, pcmk_rsc_unassigned)) { pe_rsc_trace(primary, "Delaying processing colocation %s " "because cloned primary %s is still provisional", colocation->id, primary->id); return; } pe_rsc_trace(primary, "Processing colocation %s (%s with clone %s @%s)", colocation->id, dependent->id, primary->id, pcmk_readable_score(colocation->score)); // Apply role-specific colocations if (pcmk_is_set(primary->flags, pcmk_rsc_promotable) && (colocation->primary_role != pcmk_role_unknown)) { if (pcmk_is_set(dependent->flags, pcmk_rsc_unassigned)) { // We're assigning the dependent to a node pcmk__update_dependent_with_promotable(primary, dependent, colocation); return; } if (colocation->dependent_role == pcmk_role_promoted) { // We're choosing a role for the dependent pcmk__update_promotable_dependent_priority(primary, dependent, colocation); return; } } // Apply interleaved colocations if (can_interleave(colocation)) { const pe_resource_t *primary_instance = NULL; primary_instance = pcmk__find_compatible_instance(dependent, primary, pcmk_role_unknown, false); if (primary_instance != NULL) { pe_rsc_debug(primary, "Interleaving %s with %s", dependent->id, primary_instance->id); dependent->cmds->apply_coloc_score(dependent, primary_instance, colocation, true); } else if (colocation->score >= INFINITY) { crm_notice("%s cannot run because it cannot interleave with " "any instance of %s", dependent->id, primary->id); pcmk__assign_resource(dependent, NULL, true, true); } else { pe_rsc_debug(primary, "%s will not colocate with %s " "because no instance can interleave with it", dependent->id, primary->id); } return; } // Apply mandatory colocations if (colocation->score >= INFINITY) { GList *primary_nodes = NULL; // Dependent can run only where primary will have unblocked instances for (iter = primary->children; iter != NULL; iter = iter->next) { const pe_resource_t *instance = iter->data; pe_node_t *chosen = instance->fns->location(instance, NULL, 0); if ((chosen != NULL) && !is_set_recursive(instance, pcmk_rsc_blocked, TRUE)) { pe_rsc_trace(primary, "Allowing %s: %s %d", colocation->id, pe__node_name(chosen), chosen->weight); primary_nodes = g_list_prepend(primary_nodes, chosen); } } pcmk__colocation_intersect_nodes(dependent, primary, colocation, primary_nodes, false); g_list_free(primary_nodes); return; } // Apply optional colocations for (iter = primary->children; iter != NULL; iter = iter->next) { const pe_resource_t *instance = iter->data; instance->cmds->apply_coloc_score(dependent, instance, colocation, false); } } // Clone implementation of resource_alloc_functions_t:with_this_colocations() void pcmk__with_clone_colocations(const pe_resource_t *rsc, const pe_resource_t *orig_rsc, GList **list) { CRM_CHECK((rsc != NULL) && (orig_rsc != NULL) && (list != NULL), return); pcmk__add_with_this_list(list, rsc->rsc_cons_lhs, orig_rsc); if (rsc->parent != NULL) { rsc->parent->cmds->with_this_colocations(rsc->parent, orig_rsc, list); } } // Clone implementation of resource_alloc_functions_t:this_with_colocations() void pcmk__clone_with_colocations(const pe_resource_t *rsc, const pe_resource_t *orig_rsc, GList **list) { CRM_CHECK((rsc != NULL) && (orig_rsc != NULL) && (list != NULL), return); pcmk__add_this_with_list(list, rsc->rsc_cons, orig_rsc); if (rsc->parent != NULL) { rsc->parent->cmds->this_with_colocations(rsc->parent, orig_rsc, list); } } /*! * \internal * \brief Return action flags for a given clone resource action * * \param[in,out] action Action to get flags for * \param[in] node If not NULL, limit effects to this node * * \return Flags appropriate to \p action on \p node */ uint32_t pcmk__clone_action_flags(pe_action_t *action, const pe_node_t *node) { CRM_ASSERT((action != NULL) && pe_rsc_is_clone(action->rsc)); return pcmk__collective_action_flags(action, action->rsc->children, node); } /*! * \internal * \brief Apply a location constraint to a clone resource's allowed node scores * * \param[in,out] rsc Clone resource to apply constraint to * \param[in,out] location Location constraint to apply */ void pcmk__clone_apply_location(pe_resource_t *rsc, pe__location_t *location) { CRM_CHECK((location != NULL) && pe_rsc_is_clone(rsc), return); pcmk__apply_location(rsc, location); for (GList *iter = rsc->children; iter != NULL; iter = iter->next) { pe_resource_t *instance = (pe_resource_t *) iter->data; instance->cmds->apply_location(instance, location); } } // GFunc wrapper for calling the action_flags() resource method static void call_action_flags(gpointer data, gpointer user_data) { pe_resource_t *rsc = user_data; rsc->cmds->action_flags((pe_action_t *) data, NULL); } /*! * \internal * \brief Add a clone resource's actions to the transition graph * * \param[in,out] rsc Resource whose actions should be added */ void pcmk__clone_add_actions_to_graph(pe_resource_t *rsc) { CRM_ASSERT(pe_rsc_is_clone(rsc)); g_list_foreach(rsc->actions, call_action_flags, rsc); pe__create_clone_notifications(rsc); for (GList *iter = rsc->children; iter != NULL; iter = iter->next) { pe_resource_t *child_rsc = (pe_resource_t *) iter->data; child_rsc->cmds->add_actions_to_graph(child_rsc); } pcmk__add_rsc_actions_to_graph(rsc); pe__free_clone_notification_data(rsc); } /*! * \internal * \brief Check whether a resource or any children have been probed on a node * * \param[in] rsc Resource to check * \param[in] node Node to check * * \return true if \p node is in the known_on table of \p rsc or any of its * children, otherwise false */ static bool rsc_probed_on(const pe_resource_t *rsc, const pe_node_t *node) { if (rsc->children != NULL) { for (GList *child_iter = rsc->children; child_iter != NULL; child_iter = child_iter->next) { pe_resource_t *child = (pe_resource_t *) child_iter->data; if (rsc_probed_on(child, node)) { return true; } } return false; } if (rsc->known_on != NULL) { GHashTableIter iter; pe_node_t *known_node = NULL; g_hash_table_iter_init(&iter, rsc->known_on); while (g_hash_table_iter_next(&iter, NULL, (gpointer *) &known_node)) { if (pe__same_node(node, known_node)) { return true; } } } return false; } /*! * \internal * \brief Find clone instance that has been probed on given node * * \param[in] clone Clone resource to check * \param[in] node Node to check * * \return Instance of \p clone that has been probed on \p node if any, * otherwise NULL */ static pe_resource_t * find_probed_instance_on(const pe_resource_t *clone, const pe_node_t *node) { for (GList *iter = clone->children; iter != NULL; iter = iter->next) { pe_resource_t *instance = (pe_resource_t *) iter->data; if (rsc_probed_on(instance, node)) { return instance; } } return NULL; } /*! * \internal * \brief Probe an anonymous clone on a node * * \param[in,out] clone Anonymous clone to probe * \param[in,out] node Node to probe \p clone on */ static bool probe_anonymous_clone(pe_resource_t *clone, pe_node_t *node) { // Check whether we already probed an instance on this node pe_resource_t *child = find_probed_instance_on(clone, node); // Otherwise, check if we plan to start an instance on this node for (GList *iter = clone->children; (iter != NULL) && (child == NULL); iter = iter->next) { pe_resource_t *instance = (pe_resource_t *) iter->data; const pe_node_t *instance_node = NULL; instance_node = instance->fns->location(instance, NULL, 0); if (pe__same_node(instance_node, node)) { child = instance; } } // Otherwise, use the first clone instance if (child == NULL) { child = clone->children->data; } // Anonymous clones only need to probe a single instance return child->cmds->create_probe(child, node); } /*! * \internal * \brief Schedule any probes needed for a resource on a node * * \param[in,out] rsc Resource to create probe for * \param[in,out] node Node to create probe on * * \return true if any probe was created, otherwise false */ bool pcmk__clone_create_probe(pe_resource_t *rsc, pe_node_t *node) { CRM_ASSERT((node != NULL) && pe_rsc_is_clone(rsc)); if (rsc->exclusive_discover) { /* The clone is configured to be probed only where a location constraint * exists with resource-discovery set to exclusive. * * This check is not strictly necessary here since the instance's * create_probe() method would also check, but doing it here is more * efficient (especially for unique clones with a large number of * instances), and affects the CRM_meta_notify_available_uname variable * passed with notify actions. */ pe_node_t *allowed = g_hash_table_lookup(rsc->allowed_nodes, node->details->id); if ((allowed == NULL) || (allowed->rsc_discover_mode != pcmk_probe_exclusive)) { /* This node is not marked for resource discovery. Remove it from * allowed_nodes so that notifications contain only nodes that the * clone can possibly run on. */ pe_rsc_trace(rsc, "Skipping probe for %s on %s because resource has " "exclusive discovery but is not allowed on node", rsc->id, pe__node_name(node)); g_hash_table_remove(rsc->allowed_nodes, node->details->id); return false; } } rsc->children = g_list_sort(rsc->children, pcmk__cmp_instance_number); if (pcmk_is_set(rsc->flags, pcmk_rsc_unique)) { return pcmk__probe_resource_list(rsc->children, node); } else { return probe_anonymous_clone(rsc, node); } } /*! * \internal * \brief Add meta-attributes relevant to transition graph actions to XML * * Add clone-specific meta-attributes needed for transition graph actions. * * \param[in] rsc Clone resource whose meta-attributes should be added * \param[in,out] xml Transition graph action attributes XML to add to */ void pcmk__clone_add_graph_meta(const pe_resource_t *rsc, xmlNode *xml) { char *name = NULL; CRM_ASSERT(pe_rsc_is_clone(rsc) && (xml != NULL)); name = crm_meta_name(XML_RSC_ATTR_UNIQUE); crm_xml_add(xml, name, pe__rsc_bool_str(rsc, pcmk_rsc_unique)); free(name); name = crm_meta_name(XML_RSC_ATTR_NOTIFY); crm_xml_add(xml, name, pe__rsc_bool_str(rsc, pcmk_rsc_notify)); free(name); name = crm_meta_name(PCMK_META_CLONE_MAX); crm_xml_add_int(xml, name, pe__clone_max(rsc)); free(name); name = crm_meta_name(PCMK_META_CLONE_NODE_MAX); crm_xml_add_int(xml, name, pe__clone_node_max(rsc)); free(name); if (pcmk_is_set(rsc->flags, pcmk_rsc_promotable)) { int promoted_max = pe__clone_promoted_max(rsc); int promoted_node_max = pe__clone_promoted_node_max(rsc); name = crm_meta_name(PCMK_META_PROMOTED_MAX); crm_xml_add_int(xml, name, promoted_max); free(name); name = crm_meta_name(PCMK_META_PROMOTED_NODE_MAX); crm_xml_add_int(xml, name, promoted_node_max); free(name); /* @COMPAT Maintain backward compatibility with resource agents that * expect the old names (deprecated since 2.0.0). */ name = crm_meta_name(PCMK_XA_PROMOTED_MAX_LEGACY); crm_xml_add_int(xml, name, promoted_max); free(name); name = crm_meta_name(PCMK_XA_PROMOTED_NODE_MAX_LEGACY); crm_xml_add_int(xml, name, promoted_node_max); free(name); } } // Clone implementation of resource_alloc_functions_t:add_utilization() void pcmk__clone_add_utilization(const pe_resource_t *rsc, const pe_resource_t *orig_rsc, GList *all_rscs, GHashTable *utilization) { bool existing = false; pe_resource_t *child = NULL; CRM_ASSERT(pe_rsc_is_clone(rsc) && (orig_rsc != NULL) && (utilization != NULL)); if (!pcmk_is_set(rsc->flags, pcmk_rsc_unassigned)) { return; } // Look for any child already existing in the list for (GList *iter = rsc->children; iter != NULL; iter = iter->next) { child = (pe_resource_t *) iter->data; if (g_list_find(all_rscs, child)) { existing = true; // Keep checking remaining children } else { // If this is a clone of a group, look for group's members for (GList *member_iter = child->children; member_iter != NULL; member_iter = member_iter->next) { pe_resource_t *member = (pe_resource_t *) member_iter->data; if (g_list_find(all_rscs, member) != NULL) { // Add *child's* utilization, not group member's child->cmds->add_utilization(child, orig_rsc, all_rscs, utilization); existing = true; break; } } } } if (!existing && (rsc->children != NULL)) { // If nothing was found, still add first child's utilization child = (pe_resource_t *) rsc->children->data; child->cmds->add_utilization(child, orig_rsc, all_rscs, utilization); } } // Clone implementation of resource_alloc_functions_t:shutdown_lock() void pcmk__clone_shutdown_lock(pe_resource_t *rsc) { CRM_ASSERT(pe_rsc_is_clone(rsc)); return; // Clones currently don't support shutdown locks } diff --git a/lib/pacemaker/pcmk_sched_fencing.c b/lib/pacemaker/pcmk_sched_fencing.c index 58e785c577..0ca81b97f3 100644 --- a/lib/pacemaker/pcmk_sched_fencing.c +++ b/lib/pacemaker/pcmk_sched_fencing.c @@ -1,496 +1,499 @@ /* * Copyright 2004-2023 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU General Public License version 2 * or later (GPLv2+) WITHOUT ANY WARRANTY. */ #include #include #include #include #include #include "libpacemaker_private.h" /*! * \internal * \brief Check whether a resource is known on a particular node * * \param[in] rsc Resource to check * \param[in] node Node to check * * \return TRUE if resource (or parent if an anonymous clone) is known */ static bool rsc_is_known_on(const pe_resource_t *rsc, const pe_node_t *node) { if (g_hash_table_lookup(rsc->known_on, node->details->id) != NULL) { return TRUE; } else if ((rsc->variant == pcmk_rsc_variant_primitive) && pe_rsc_is_anon_clone(rsc->parent) && (g_hash_table_lookup(rsc->parent->known_on, node->details->id) != NULL)) { /* We check only the parent, not the uber-parent, because we cannot * assume that the resource is known if it is in an anonymously cloned * group (which may be only partially known). */ return TRUE; } return FALSE; } /*! * \internal * \brief Order a resource's start and promote actions relative to fencing * * \param[in,out] rsc Resource to be ordered * \param[in,out] stonith_op Fence action */ static void order_start_vs_fencing(pe_resource_t *rsc, pe_action_t *stonith_op) { pe_node_t *target; CRM_CHECK(stonith_op && stonith_op->node, return); target = stonith_op->node; for (GList *iter = rsc->actions; iter != NULL; iter = iter->next) { pe_action_t *action = iter->data; switch (action->needs) { case pcmk_requires_nothing: // Anything other than start or promote requires nothing break; case pcmk_requires_fencing: order_actions(stonith_op, action, pcmk__ar_ordered); break; case pcmk_requires_quorum: if (pcmk__str_eq(action->task, PCMK_ACTION_START, pcmk__str_none) && (g_hash_table_lookup(rsc->allowed_nodes, target->details->id) != NULL) && !rsc_is_known_on(rsc, target)) { /* If we don't know the status of the resource on the node * we're about to shoot, we have to assume it may be active * there. Order the resource start after the fencing. This * is analogous to waiting for all the probes for a resource * to complete before starting it. * * The most likely explanation is that the DC died and took * its status with it. */ pe_rsc_debug(rsc, "Ordering %s after %s recovery", action->uuid, pe__node_name(target)); order_actions(stonith_op, action, - pcmk__ar_ordered|pe_order_runnable_left); + pcmk__ar_ordered + |pcmk__ar_unrunnable_first_blocks); } break; } } } /*! * \internal * \brief Order a resource's stop and demote actions relative to fencing * * \param[in,out] rsc Resource to be ordered * \param[in,out] stonith_op Fence action */ static void order_stop_vs_fencing(pe_resource_t *rsc, pe_action_t *stonith_op) { GList *iter = NULL; GList *action_list = NULL; bool order_implicit = false; pe_resource_t *top = uber_parent(rsc); pe_action_t *parent_stop = NULL; pe_node_t *target; CRM_CHECK(stonith_op && stonith_op->node, return); target = stonith_op->node; /* Get a list of stop actions potentially implied by the fencing */ action_list = pe__resource_actions(rsc, target, PCMK_ACTION_STOP, FALSE); /* If resource requires fencing, implicit actions must occur after fencing. * * Implied stops and demotes of resources running on guest nodes are always * ordered after fencing, even if the resource does not require fencing, * because guest node "fencing" is actually just a resource stop. */ if (pcmk_is_set(rsc->flags, pcmk_rsc_needs_fencing) || pe__is_guest_node(target)) { order_implicit = true; } if (action_list && order_implicit) { parent_stop = find_first_action(top->actions, NULL, PCMK_ACTION_STOP, NULL); } for (iter = action_list; iter != NULL; iter = iter->next) { pe_action_t *action = iter->data; // The stop would never complete, so convert it into a pseudo-action. pe__set_action_flags(action, pcmk_action_pseudo|pcmk_action_runnable); if (order_implicit) { /* Order the stonith before the parent stop (if any). * * Also order the stonith before the resource stop, unless the * resource is inside a bundle -- that would cause a graph loop. * We can rely on the parent stop's ordering instead. * * User constraints must not order a resource in a guest node * relative to the guest node container resource. The * pe_order_preserve flag marks constraints as generated by the * cluster and thus immune to that check (and is irrelevant if * target is not a guest). */ if (!pe_rsc_is_bundled(rsc)) { order_actions(stonith_op, action, pe_order_preserve); } order_actions(stonith_op, parent_stop, pe_order_preserve); } if (pcmk_is_set(rsc->flags, pcmk_rsc_failed)) { crm_notice("Stop of failed resource %s is implicit %s %s is fenced", rsc->id, (order_implicit? "after" : "because"), pe__node_name(target)); } else { crm_info("%s is implicit %s %s is fenced", action->uuid, (order_implicit? "after" : "because"), pe__node_name(target)); } if (pcmk_is_set(rsc->flags, pcmk_rsc_notify)) { pe__order_notifs_after_fencing(action, rsc, stonith_op); } #if 0 /* It might be a good idea to stop healthy resources on a node about to * be fenced, when possible. * * However, fencing must be done before a failed resource's * (pseudo-)stop action, so that could create a loop. For example, given * a group of A and B running on node N with a failed stop of B: * * fence N -> stop B (pseudo-op) -> stop A -> fence N * * The block below creates the stop A -> fence N ordering and therefore * must (at least for now) be disabled. Instead, run the block above and * treat all resources on N as B would be (i.e., as a pseudo-op after * the fencing). * * @TODO Maybe break the "A requires B" dependency in * pcmk__update_action_for_orderings() and use this block for healthy * resources instead of the above. */ crm_info("Moving healthy resource %s off %s before fencing", rsc->id, pe__node_name(node)); pcmk__new_ordering(rsc, stop_key(rsc), NULL, NULL, strdup(PCMK_ACTION_STONITH), stonith_op, pcmk__ar_ordered, rsc->cluster); #endif } g_list_free(action_list); /* Get a list of demote actions potentially implied by the fencing */ action_list = pe__resource_actions(rsc, target, PCMK_ACTION_DEMOTE, FALSE); for (iter = action_list; iter != NULL; iter = iter->next) { pe_action_t *action = iter->data; if (!(action->node->details->online) || action->node->details->unclean || pcmk_is_set(rsc->flags, pcmk_rsc_failed)) { if (pcmk_is_set(rsc->flags, pcmk_rsc_failed)) { pe_rsc_info(rsc, "Demote of failed resource %s is implicit " "after %s is fenced", rsc->id, pe__node_name(target)); } else { pe_rsc_info(rsc, "%s is implicit after %s is fenced", action->uuid, pe__node_name(target)); } /* The demote would never complete and is now implied by the * fencing, so convert it into a pseudo-action. */ pe__set_action_flags(action, pcmk_action_pseudo|pcmk_action_runnable); if (pe_rsc_is_bundled(rsc)) { // Recovery will be ordered as usual after parent's implied stop } else if (order_implicit) { order_actions(stonith_op, action, pe_order_preserve|pcmk__ar_ordered); } } } g_list_free(action_list); } /*! * \internal * \brief Order resource actions properly relative to fencing * * \param[in,out] rsc Resource whose actions should be ordered * \param[in,out] stonith_op Fencing operation to be ordered against */ static void rsc_stonith_ordering(pe_resource_t *rsc, pe_action_t *stonith_op) { if (rsc->children) { for (GList *iter = rsc->children; iter != NULL; iter = iter->next) { pe_resource_t *child_rsc = iter->data; rsc_stonith_ordering(child_rsc, stonith_op); } } else if (!pcmk_is_set(rsc->flags, pcmk_rsc_managed)) { pe_rsc_trace(rsc, "Skipping fencing constraints for unmanaged resource: %s", rsc->id); } else { order_start_vs_fencing(rsc, stonith_op); order_stop_vs_fencing(rsc, stonith_op); } } /*! * \internal * \brief Order all actions appropriately relative to a fencing operation * * Ensure start operations of affected resources are ordered after fencing, * imply stop and demote operations of affected resources by marking them as * pseudo-actions, etc. * * \param[in,out] stonith_op Fencing operation * \param[in,out] data_set Working set of cluster */ void pcmk__order_vs_fence(pe_action_t *stonith_op, pe_working_set_t *data_set) { CRM_CHECK(stonith_op && data_set, return); for (GList *r = data_set->resources; r != NULL; r = r->next) { rsc_stonith_ordering((pe_resource_t *) r->data, stonith_op); } } /*! * \internal * \brief Order an action after unfencing * * \param[in] rsc Resource that action is for * \param[in,out] node Node that action is on * \param[in,out] action Action to be ordered after unfencing * \param[in] order Ordering flags */ void pcmk__order_vs_unfence(const pe_resource_t *rsc, pe_node_t *node, pe_action_t *action, enum pcmk__action_relation_flags order) { /* When unfencing is in use, we order unfence actions before any probe or * start of resources that require unfencing, and also of fence devices. * * This might seem to violate the principle that fence devices require * only quorum. However, fence agents that unfence often don't have enough * information to even probe or start unless the node is first unfenced. */ if ((pcmk_is_set(rsc->flags, pcmk_rsc_fence_device) && pcmk_is_set(rsc->cluster->flags, pcmk_sched_enable_unfencing)) || pcmk_is_set(rsc->flags, pcmk_rsc_needs_unfencing)) { /* Start with an optional ordering. Requiring unfencing would result in * the node being unfenced, and all its resources being stopped, * whenever a new resource is added -- which would be highly suboptimal. */ pe_action_t *unfence = pe_fence_op(node, PCMK_ACTION_ON, TRUE, NULL, FALSE, node->details->data_set); order_actions(unfence, action, order); if (!pcmk__node_unfenced(node)) { // But unfencing is required if it has never been done char *reason = crm_strdup_printf("required by %s %s", rsc->id, action->task); trigger_unfencing(NULL, node, reason, NULL, node->details->data_set); free(reason); } } } /*! * \internal * \brief Create pseudo-op for guest node fence, and order relative to it * * \param[in,out] node Guest node to fence */ void pcmk__fence_guest(pe_node_t *node) { pe_resource_t *container = NULL; pe_action_t *stop = NULL; pe_action_t *stonith_op = NULL; /* The fence action is just a label; we don't do anything differently for * off vs. reboot. We specify it explicitly, rather than let it default to * cluster's default action, because we are not _initiating_ fencing -- we * are creating a pseudo-event to describe fencing that is already occurring * by other means (container recovery). */ const char *fence_action = PCMK_ACTION_OFF; CRM_ASSERT(node != NULL); /* Check whether guest's container resource has any explicit stop or * start (the stop may be implied by fencing of the guest's host). */ container = node->details->remote_rsc->container; if (container) { stop = find_first_action(container->actions, NULL, PCMK_ACTION_STOP, NULL); if (find_first_action(container->actions, NULL, PCMK_ACTION_START, NULL)) { fence_action = PCMK_ACTION_REBOOT; } } /* Create a fence pseudo-event, so we have an event to order actions * against, and the controller can always detect it. */ stonith_op = pe_fence_op(node, fence_action, FALSE, "guest is unclean", FALSE, node->details->data_set); pe__set_action_flags(stonith_op, pcmk_action_pseudo|pcmk_action_runnable); /* We want to imply stops/demotes after the guest is stopped, not wait until * it is restarted, so we always order pseudo-fencing after stop, not start * (even though start might be closer to what is done for a real reboot). */ if ((stop != NULL) && pcmk_is_set(stop->flags, pcmk_action_pseudo)) { pe_action_t *parent_stonith_op = pe_fence_op(stop->node, NULL, FALSE, NULL, FALSE, node->details->data_set); crm_info("Implying guest %s is down (action %d) after %s fencing", pe__node_name(node), stonith_op->id, pe__node_name(stop->node)); order_actions(parent_stonith_op, stonith_op, - pe_order_runnable_left|pcmk__ar_first_implies_then); + pcmk__ar_unrunnable_first_blocks + |pcmk__ar_first_implies_then); } else if (stop) { order_actions(stop, stonith_op, - pe_order_runnable_left|pcmk__ar_first_implies_then); + pcmk__ar_unrunnable_first_blocks + |pcmk__ar_first_implies_then); crm_info("Implying guest %s is down (action %d) " "after container %s is stopped (action %d)", pe__node_name(node), stonith_op->id, container->id, stop->id); } else { /* If we're fencing the guest node but there's no stop for the guest * resource, we must think the guest is already stopped. However, we may * think so because its resource history was just cleaned. To avoid * unnecessarily considering the guest node down if it's really up, * order the pseudo-fencing after any stop of the connection resource, * which will be ordered after any container (re-)probe. */ stop = find_first_action(node->details->remote_rsc->actions, NULL, PCMK_ACTION_STOP, NULL); if (stop) { order_actions(stop, stonith_op, pcmk__ar_ordered); crm_info("Implying guest %s is down (action %d) " "after connection is stopped (action %d)", pe__node_name(node), stonith_op->id, stop->id); } else { /* Not sure why we're fencing, but everything must already be * cleanly stopped. */ crm_info("Implying guest %s is down (action %d) ", pe__node_name(node), stonith_op->id); } } // Order/imply other actions relative to pseudo-fence as with real fence pcmk__order_vs_fence(stonith_op, node->details->data_set); } /*! * \internal * \brief Check whether node has already been unfenced * * \param[in] node Node to check * * \return true if node has a nonzero #node-unfenced attribute (or none), * otherwise false */ bool pcmk__node_unfenced(const pe_node_t *node) { const char *unfenced = pe_node_attribute_raw(node, CRM_ATTR_UNFENCED); return !pcmk__str_eq(unfenced, "0", pcmk__str_null_matches); } /*! * \internal * \brief Order a resource's start and stop relative to unfencing of a node * * \param[in,out] data Node that could be unfenced * \param[in,out] user_data Resource to order */ void pcmk__order_restart_vs_unfence(gpointer data, gpointer user_data) { pe_node_t *node = (pe_node_t *) data; pe_resource_t *rsc = (pe_resource_t *) user_data; pe_action_t *unfence = pe_fence_op(node, PCMK_ACTION_ON, true, NULL, false, rsc->cluster); crm_debug("Ordering any stops of %s before %s, and any starts after", rsc->id, unfence->uuid); /* * It would be more efficient to order clone resources once, * rather than order each instance, but ordering the instance * allows us to avoid unnecessary dependencies that might conflict * with user constraints. * * @TODO: This constraint can still produce a transition loop if the * resource has a stop scheduled on the node being unfenced, and * there is a user ordering constraint to start some other resource * (which will be ordered after the unfence) before stopping this * resource. An example is "start some slow-starting cloned service * before stopping an associated virtual IP that may be moving to * it": * stop this -> unfencing -> start that -> stop this */ pcmk__new_ordering(rsc, stop_key(rsc), NULL, NULL, strdup(unfence->uuid), unfence, pcmk__ar_ordered|pe_order_same_node, rsc->cluster); pcmk__new_ordering(NULL, strdup(unfence->uuid), unfence, rsc, start_key(rsc), NULL, pe_order_implies_then_on_node|pe_order_same_node, rsc->cluster); } diff --git a/lib/pacemaker/pcmk_sched_group.c b/lib/pacemaker/pcmk_sched_group.c index 94ea7b28b3..71d760c96a 100644 --- a/lib/pacemaker/pcmk_sched_group.c +++ b/lib/pacemaker/pcmk_sched_group.c @@ -1,950 +1,950 @@ /* * Copyright 2004-2023 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU General Public License version 2 * or later (GPLv2+) WITHOUT ANY WARRANTY. */ #include #include #include #include #include "libpacemaker_private.h" /*! * \internal * \brief Assign a group resource to a node * * \param[in,out] rsc Group resource to assign to a node * \param[in] prefer Node to prefer, if all else is equal * \param[in] stop_if_fail If \c true and a child of \p rsc can't be * assigned to a node, set the child's next role to * stopped and update existing actions * * \return Node that \p rsc is assigned to, if assigned entirely to one node * * \note If \p stop_if_fail is \c false, then \c pcmk__unassign_resource() can * completely undo the assignment. A successful assignment can be either * undone or left alone as final. A failed assignment has the same effect * as calling pcmk__unassign_resource(); there are no side effects on * roles or actions. */ pe_node_t * pcmk__group_assign(pe_resource_t *rsc, const pe_node_t *prefer, bool stop_if_fail) { pe_node_t *first_assigned_node = NULL; pe_resource_t *first_member = NULL; CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_group)); if (!pcmk_is_set(rsc->flags, pcmk_rsc_unassigned)) { return rsc->allocated_to; // Assignment already done } if (pcmk_is_set(rsc->flags, pcmk_rsc_assigning)) { pe_rsc_debug(rsc, "Assignment dependency loop detected involving %s", rsc->id); return NULL; } if (rsc->children == NULL) { // No members to assign pe__clear_resource_flags(rsc, pcmk_rsc_unassigned); return NULL; } pe__set_resource_flags(rsc, pcmk_rsc_assigning); first_member = (pe_resource_t *) rsc->children->data; rsc->role = first_member->role; pe__show_node_scores(!pcmk_is_set(rsc->cluster->flags, pcmk_sched_output_scores), rsc, __func__, rsc->allowed_nodes, rsc->cluster); for (GList *iter = rsc->children; iter != NULL; iter = iter->next) { pe_resource_t *member = (pe_resource_t *) iter->data; pe_node_t *node = NULL; pe_rsc_trace(rsc, "Assigning group %s member %s", rsc->id, member->id); node = member->cmds->assign(member, prefer, stop_if_fail); if (first_assigned_node == NULL) { first_assigned_node = node; } } pe__set_next_role(rsc, first_member->next_role, "first group member"); pe__clear_resource_flags(rsc, pcmk_rsc_assigning|pcmk_rsc_unassigned); if (!pe__group_flag_is_set(rsc, pe__group_colocated)) { return NULL; } return first_assigned_node; } /*! * \internal * \brief Create a pseudo-operation for a group as an ordering point * * \param[in,out] group Group resource to create action for * \param[in] action Action name * * \return Newly created pseudo-operation */ static pe_action_t * create_group_pseudo_op(pe_resource_t *group, const char *action) { pe_action_t *op = custom_action(group, pcmk__op_key(group->id, action, 0), action, NULL, TRUE, TRUE, group->cluster); pe__set_action_flags(op, pcmk_action_pseudo|pcmk_action_runnable); return op; } /*! * \internal * \brief Create all actions needed for a given group resource * * \param[in,out] rsc Group resource to create actions for */ void pcmk__group_create_actions(pe_resource_t *rsc) { CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_group)); pe_rsc_trace(rsc, "Creating actions for group %s", rsc->id); // Create actions for individual group members for (GList *iter = rsc->children; iter != NULL; iter = iter->next) { pe_resource_t *member = (pe_resource_t *) iter->data; member->cmds->create_actions(member); } // Create pseudo-actions for group itself to serve as ordering points create_group_pseudo_op(rsc, PCMK_ACTION_START); create_group_pseudo_op(rsc, PCMK_ACTION_RUNNING); create_group_pseudo_op(rsc, PCMK_ACTION_STOP); create_group_pseudo_op(rsc, PCMK_ACTION_STOPPED); if (crm_is_true(g_hash_table_lookup(rsc->meta, XML_RSC_ATTR_PROMOTABLE))) { create_group_pseudo_op(rsc, PCMK_ACTION_DEMOTE); create_group_pseudo_op(rsc, PCMK_ACTION_DEMOTED); create_group_pseudo_op(rsc, PCMK_ACTION_PROMOTE); create_group_pseudo_op(rsc, PCMK_ACTION_PROMOTED); } } // User data for member_internal_constraints() struct member_data { // These could be derived from member but this avoids some function calls bool ordered; bool colocated; bool promotable; pe_resource_t *last_active; pe_resource_t *previous_member; }; /*! * \internal * \brief Create implicit constraints needed for a group member * * \param[in,out] data Group member to create implicit constraints for * \param[in,out] user_data Member data (struct member_data *) */ static void member_internal_constraints(gpointer data, gpointer user_data) { pe_resource_t *member = (pe_resource_t *) data; struct member_data *member_data = (struct member_data *) user_data; // For ordering demote vs demote or stop vs stop uint32_t down_flags = pe_order_implies_first_printed; // For ordering demote vs demoted or stop vs stopped uint32_t post_down_flags = pe_order_implies_then_printed; // Create the individual member's implicit constraints member->cmds->internal_constraints(member); if (member_data->previous_member == NULL) { // This is first member if (member_data->ordered) { pe__set_order_flags(down_flags, pcmk__ar_ordered); post_down_flags = pcmk__ar_first_implies_then; } } else if (member_data->colocated) { uint32_t flags = pcmk__coloc_none; if (pcmk_is_set(member->flags, pcmk_rsc_critical)) { flags |= pcmk__coloc_influence; } // Colocate this member with the previous one pcmk__new_colocation("#group-members", NULL, INFINITY, member, member_data->previous_member, NULL, NULL, flags); } if (member_data->promotable) { // Demote group -> demote member -> group is demoted pcmk__order_resource_actions(member->parent, PCMK_ACTION_DEMOTE, member, PCMK_ACTION_DEMOTE, down_flags); pcmk__order_resource_actions(member, PCMK_ACTION_DEMOTE, member->parent, PCMK_ACTION_DEMOTED, post_down_flags); // Promote group -> promote member -> group is promoted pcmk__order_resource_actions(member, PCMK_ACTION_PROMOTE, member->parent, PCMK_ACTION_PROMOTED, - pe_order_runnable_left + pcmk__ar_unrunnable_first_blocks |pcmk__ar_first_implies_then |pe_order_implies_then_printed); pcmk__order_resource_actions(member->parent, PCMK_ACTION_PROMOTE, member, PCMK_ACTION_PROMOTE, pe_order_implies_first_printed); } // Stop group -> stop member -> group is stopped pcmk__order_stops(member->parent, member, down_flags); pcmk__order_resource_actions(member, PCMK_ACTION_STOP, member->parent, PCMK_ACTION_STOPPED, post_down_flags); // Start group -> start member -> group is started pcmk__order_starts(member->parent, member, pe_order_implies_first_printed); pcmk__order_resource_actions(member, PCMK_ACTION_START, member->parent, PCMK_ACTION_RUNNING, - pe_order_runnable_left + pcmk__ar_unrunnable_first_blocks |pcmk__ar_first_implies_then |pe_order_implies_then_printed); if (!member_data->ordered) { pcmk__order_starts(member->parent, member, pcmk__ar_first_implies_then - |pe_order_runnable_left + |pcmk__ar_unrunnable_first_blocks |pe_order_implies_first_printed); if (member_data->promotable) { pcmk__order_resource_actions(member->parent, PCMK_ACTION_PROMOTE, member, PCMK_ACTION_PROMOTE, pcmk__ar_first_implies_then - |pe_order_runnable_left + |pcmk__ar_unrunnable_first_blocks |pe_order_implies_first_printed); } } else if (member_data->previous_member == NULL) { pcmk__order_starts(member->parent, member, pcmk__ar_none); if (member_data->promotable) { pcmk__order_resource_actions(member->parent, PCMK_ACTION_PROMOTE, member, PCMK_ACTION_PROMOTE, pcmk__ar_none); } } else { // Order this member relative to the previous one pcmk__order_starts(member_data->previous_member, member, pcmk__ar_first_implies_then - |pe_order_runnable_left); + |pcmk__ar_unrunnable_first_blocks); pcmk__order_stops(member, member_data->previous_member, pcmk__ar_ordered|pe_order_restart); /* In unusual circumstances (such as adding a new member to the middle * of a group with unmanaged later members), this member may be active * while the previous (new) member is inactive. In this situation, the * usual restart orderings will be irrelevant, so we need to order this * member's stop before the previous member's start. */ if ((member->running_on != NULL) && (member_data->previous_member->running_on == NULL)) { pcmk__order_resource_actions(member, PCMK_ACTION_STOP, member_data->previous_member, PCMK_ACTION_START, pcmk__ar_then_implies_first - |pe_order_runnable_left); + |pcmk__ar_unrunnable_first_blocks); } if (member_data->promotable) { pcmk__order_resource_actions(member_data->previous_member, PCMK_ACTION_PROMOTE, member, PCMK_ACTION_PROMOTE, pcmk__ar_first_implies_then - |pe_order_runnable_left); + |pcmk__ar_unrunnable_first_blocks); pcmk__order_resource_actions(member, PCMK_ACTION_DEMOTE, member_data->previous_member, PCMK_ACTION_DEMOTE, pcmk__ar_ordered); } } // Make sure partially active groups shut down in sequence if (member->running_on != NULL) { if (member_data->ordered && (member_data->previous_member != NULL) && (member_data->previous_member->running_on == NULL) && (member_data->last_active != NULL) && (member_data->last_active->running_on != NULL)) { pcmk__order_stops(member, member_data->last_active, pcmk__ar_ordered); } member_data->last_active = member; } member_data->previous_member = member; } /*! * \internal * \brief Create implicit constraints needed for a group resource * * \param[in,out] rsc Group resource to create implicit constraints for */ void pcmk__group_internal_constraints(pe_resource_t *rsc) { struct member_data member_data = { false, }; const pe_resource_t *top = NULL; CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_group)); /* Order group pseudo-actions relative to each other for restarting: * stop group -> group is stopped -> start group -> group is started */ pcmk__order_resource_actions(rsc, PCMK_ACTION_STOP, rsc, PCMK_ACTION_STOPPED, - pe_order_runnable_left); + pcmk__ar_unrunnable_first_blocks); pcmk__order_resource_actions(rsc, PCMK_ACTION_STOPPED, rsc, PCMK_ACTION_START, pcmk__ar_ordered); pcmk__order_resource_actions(rsc, PCMK_ACTION_START, rsc, PCMK_ACTION_RUNNING, - pe_order_runnable_left); + pcmk__ar_unrunnable_first_blocks); top = pe__const_top_resource(rsc, false); member_data.ordered = pe__group_flag_is_set(rsc, pe__group_ordered); member_data.colocated = pe__group_flag_is_set(rsc, pe__group_colocated); member_data.promotable = pcmk_is_set(top->flags, pcmk_rsc_promotable); g_list_foreach(rsc->children, member_internal_constraints, &member_data); } /*! * \internal * \brief Apply a colocation's score to node scores or resource priority * * Given a colocation constraint for a group with some other resource, apply the * score to the dependent's allowed node scores (if we are still placing * resources) or priority (if we are choosing promotable clone instance roles). * * \param[in,out] dependent Dependent group resource in colocation * \param[in] primary Primary resource in colocation * \param[in] colocation Colocation constraint to apply */ static void colocate_group_with(pe_resource_t *dependent, const pe_resource_t *primary, const pcmk__colocation_t *colocation) { pe_resource_t *member = NULL; if (dependent->children == NULL) { return; } pe_rsc_trace(primary, "Processing %s (group %s with %s) for dependent", colocation->id, dependent->id, primary->id); if (pe__group_flag_is_set(dependent, pe__group_colocated)) { // Colocate first member (internal colocations will handle the rest) member = (pe_resource_t *) dependent->children->data; member->cmds->apply_coloc_score(member, primary, colocation, true); return; } if (colocation->score >= INFINITY) { pcmk__config_err("%s: Cannot perform mandatory colocation between " "non-colocated group and %s", dependent->id, primary->id); return; } // Colocate each member individually for (GList *iter = dependent->children; iter != NULL; iter = iter->next) { member = (pe_resource_t *) iter->data; member->cmds->apply_coloc_score(member, primary, colocation, true); } } /*! * \internal * \brief Apply a colocation's score to node scores or resource priority * * Given a colocation constraint for some other resource with a group, apply the * score to the dependent's allowed node scores (if we are still placing * resources) or priority (if we are choosing promotable clone instance roles). * * \param[in,out] dependent Dependent resource in colocation * \param[in] primary Primary group resource in colocation * \param[in] colocation Colocation constraint to apply */ static void colocate_with_group(pe_resource_t *dependent, const pe_resource_t *primary, const pcmk__colocation_t *colocation) { const pe_resource_t *member = NULL; pe_rsc_trace(primary, "Processing colocation %s (%s with group %s) for primary", colocation->id, dependent->id, primary->id); if (pcmk_is_set(primary->flags, pcmk_rsc_unassigned)) { return; } if (pe__group_flag_is_set(primary, pe__group_colocated)) { if (colocation->score >= INFINITY) { /* For mandatory colocations, the entire group must be assignable * (and in the specified role if any), so apply the colocation based * on the last member. */ member = pe__last_group_member(primary); } else if (primary->children != NULL) { /* For optional colocations, whether the group is partially or fully * up doesn't matter, so apply the colocation based on the first * member. */ member = (pe_resource_t *) primary->children->data; } if (member == NULL) { return; // Nothing to colocate with } member->cmds->apply_coloc_score(dependent, member, colocation, false); return; } if (colocation->score >= INFINITY) { pcmk__config_err("%s: Cannot perform mandatory colocation with" " non-colocated group %s", dependent->id, primary->id); return; } // Colocate dependent with each member individually for (const GList *iter = primary->children; iter != NULL; iter = iter->next) { member = iter->data; member->cmds->apply_coloc_score(dependent, member, colocation, false); } } /*! * \internal * \brief Apply a colocation's score to node scores or resource priority * * Given a colocation constraint, apply its score to the dependent's * allowed node scores (if we are still placing resources) or priority (if * we are choosing promotable clone instance roles). * * \param[in,out] dependent Dependent resource in colocation * \param[in] primary Primary resource in colocation * \param[in] colocation Colocation constraint to apply * \param[in] for_dependent true if called on behalf of dependent */ void pcmk__group_apply_coloc_score(pe_resource_t *dependent, const pe_resource_t *primary, const pcmk__colocation_t *colocation, bool for_dependent) { CRM_ASSERT((dependent != NULL) && (primary != NULL) && (colocation != NULL)); if (for_dependent) { colocate_group_with(dependent, primary, colocation); } else { // Method should only be called for primitive dependents CRM_ASSERT(dependent->variant == pcmk_rsc_variant_primitive); colocate_with_group(dependent, primary, colocation); } } /*! * \internal * \brief Return action flags for a given group resource action * * \param[in,out] action Group action to get flags for * \param[in] node If not NULL, limit effects to this node * * \return Flags appropriate to \p action on \p node */ uint32_t pcmk__group_action_flags(pe_action_t *action, const pe_node_t *node) { // Default flags for a group action uint32_t flags = pcmk_action_optional |pcmk_action_runnable |pcmk_action_pseudo; CRM_ASSERT(action != NULL); // Update flags considering each member's own flags for same action for (GList *iter = action->rsc->children; iter != NULL; iter = iter->next) { pe_resource_t *member = (pe_resource_t *) iter->data; // Check whether member has the same action enum action_tasks task = get_complex_task(member, action->task); const char *task_s = task2text(task); pe_action_t *member_action = find_first_action(member->actions, NULL, task_s, node); if (member_action != NULL) { uint32_t member_flags = member->cmds->action_flags(member_action, node); // Group action is mandatory if any member action is if (pcmk_is_set(flags, pcmk_action_optional) && !pcmk_is_set(member_flags, pcmk_action_optional)) { pe_rsc_trace(action->rsc, "%s is mandatory because %s is", action->uuid, member_action->uuid); pe__clear_raw_action_flags(flags, "group action", pcmk_action_optional); pe__clear_action_flags(action, pcmk_action_optional); } // Group action is unrunnable if any member action is if (!pcmk__str_eq(task_s, action->task, pcmk__str_none) && pcmk_is_set(flags, pcmk_action_runnable) && !pcmk_is_set(member_flags, pcmk_action_runnable)) { pe_rsc_trace(action->rsc, "%s is unrunnable because %s is", action->uuid, member_action->uuid); pe__clear_raw_action_flags(flags, "group action", pcmk_action_runnable); pe__clear_action_flags(action, pcmk_action_runnable); } /* Group (pseudo-)actions other than stop or demote are unrunnable * unless every member will do it. */ } else if ((task != pcmk_action_stop) && (task != pcmk_action_demote)) { pe_rsc_trace(action->rsc, "%s is not runnable because %s will not %s", action->uuid, member->id, task_s); pe__clear_raw_action_flags(flags, "group action", pcmk_action_runnable); } } return flags; } /*! * \internal * \brief Update two actions according to an ordering between them * * Given information about an ordering of two actions, update the actions' flags * (and runnable_before members if appropriate) as appropriate for the ordering. * Effects may cascade to other orderings involving the actions as well. * * \param[in,out] first 'First' action in an ordering * \param[in,out] then 'Then' action in an ordering * \param[in] node If not NULL, limit scope of ordering to this node * (only used when interleaving instances) * \param[in] flags Action flags for \p first for ordering purposes * \param[in] filter Action flags to limit scope of certain updates (may * include pcmk_action_optional to affect only * mandatory actions, and pcmk_action_runnable to * affect only runnable actions) * \param[in] type Group of enum pcmk__action_relation_flags to apply * \param[in,out] data_set Cluster working set * * \return Group of enum pcmk__updated flags indicating what was updated */ uint32_t pcmk__group_update_ordered_actions(pe_action_t *first, pe_action_t *then, const pe_node_t *node, uint32_t flags, uint32_t filter, uint32_t type, pe_working_set_t *data_set) { uint32_t changed = pcmk__updated_none; // Group method can be called only on behalf of "then" action CRM_ASSERT((first != NULL) && (then != NULL) && (then->rsc != NULL) && (data_set != NULL)); // Update the actions for the group itself changed |= pcmk__update_ordered_actions(first, then, node, flags, filter, type, data_set); // Update the actions for each group member for (GList *iter = then->rsc->children; iter != NULL; iter = iter->next) { pe_resource_t *member = (pe_resource_t *) iter->data; pe_action_t *member_action = find_first_action(member->actions, NULL, then->task, node); if (member_action != NULL) { changed |= member->cmds->update_ordered_actions(first, member_action, node, flags, filter, type, data_set); } } return changed; } /*! * \internal * \brief Apply a location constraint to a group's allowed node scores * * \param[in,out] rsc Group resource to apply constraint to * \param[in,out] location Location constraint to apply */ void pcmk__group_apply_location(pe_resource_t *rsc, pe__location_t *location) { GList *node_list_orig = NULL; GList *node_list_copy = NULL; bool reset_scores = true; CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_group) && (location != NULL)); node_list_orig = location->node_list_rh; node_list_copy = pcmk__copy_node_list(node_list_orig, true); reset_scores = pe__group_flag_is_set(rsc, pe__group_colocated); // Apply the constraint for the group itself (updates node scores) pcmk__apply_location(rsc, location); // Apply the constraint for each member for (GList *iter = rsc->children; iter != NULL; iter = iter->next) { pe_resource_t *member = (pe_resource_t *) iter->data; member->cmds->apply_location(member, location); if (reset_scores) { /* The first member of colocated groups needs to use the original * node scores, but subsequent members should work on a copy, since * the first member's scores already incorporate theirs. */ reset_scores = false; location->node_list_rh = node_list_copy; } } location->node_list_rh = node_list_orig; g_list_free_full(node_list_copy, free); } // Group implementation of resource_alloc_functions_t:colocated_resources() GList * pcmk__group_colocated_resources(const pe_resource_t *rsc, const pe_resource_t *orig_rsc, GList *colocated_rscs) { const pe_resource_t *member = NULL; CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_group)); if (orig_rsc == NULL) { orig_rsc = rsc; } if (pe__group_flag_is_set(rsc, pe__group_colocated) || pe_rsc_is_clone(rsc->parent)) { /* This group has colocated members and/or is cloned -- either way, * add every child's colocated resources to the list. The first and last * members will include the group's own colocations. */ colocated_rscs = g_list_prepend(colocated_rscs, (gpointer) rsc); for (const GList *iter = rsc->children; iter != NULL; iter = iter->next) { member = (const pe_resource_t *) iter->data; colocated_rscs = member->cmds->colocated_resources(member, orig_rsc, colocated_rscs); } } else if (rsc->children != NULL) { /* This group's members are not colocated, and the group is not cloned, * so just add the group's own colocations to the list. */ colocated_rscs = pcmk__colocated_resources(rsc, orig_rsc, colocated_rscs); } return colocated_rscs; } // Group implementation of resource_alloc_functions_t:with_this_colocations() void pcmk__with_group_colocations(const pe_resource_t *rsc, const pe_resource_t *orig_rsc, GList **list) { CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_group) && (orig_rsc != NULL) && (list != NULL)); // Ignore empty groups if (rsc->children == NULL) { return; } /* "With this" colocations are needed only for the group itself and for its * last member. (Previous members will chain via the group internal * colocations.) */ if ((orig_rsc != rsc) && (orig_rsc != pe__last_group_member(rsc))) { return; } pe_rsc_trace(rsc, "Adding 'with %s' colocations to list for %s", rsc->id, orig_rsc->id); // Add the group's own colocations pcmk__add_with_this_list(list, rsc->rsc_cons_lhs, orig_rsc); // If cloned, add any relevant colocations with the clone if (rsc->parent != NULL) { rsc->parent->cmds->with_this_colocations(rsc->parent, orig_rsc, list); } if (!pe__group_flag_is_set(rsc, pe__group_colocated)) { // @COMPAT Non-colocated groups are deprecated return; } // Add explicit colocations with the group's (other) children for (const GList *iter = rsc->children; iter != NULL; iter = iter->next) { const pe_resource_t *member = iter->data; if (member != orig_rsc) { member->cmds->with_this_colocations(member, orig_rsc, list); } } } // Group implementation of resource_alloc_functions_t:this_with_colocations() void pcmk__group_with_colocations(const pe_resource_t *rsc, const pe_resource_t *orig_rsc, GList **list) { const pe_resource_t *member = NULL; CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_group) && (orig_rsc != NULL) && (list != NULL)); // Ignore empty groups if (rsc->children == NULL) { return; } /* "This with" colocations are normally needed only for the group itself and * for its first member. */ if ((rsc == orig_rsc) || (orig_rsc == (const pe_resource_t *) rsc->children->data)) { pe_rsc_trace(rsc, "Adding '%s with' colocations to list for %s", rsc->id, orig_rsc->id); // Add the group's own colocations pcmk__add_this_with_list(list, rsc->rsc_cons, orig_rsc); // If cloned, add any relevant colocations involving the clone if (rsc->parent != NULL) { rsc->parent->cmds->this_with_colocations(rsc->parent, orig_rsc, list); } if (!pe__group_flag_is_set(rsc, pe__group_colocated)) { // @COMPAT Non-colocated groups are deprecated return; } // Add explicit colocations involving the group's (other) children for (const GList *iter = rsc->children; iter != NULL; iter = iter->next) { member = iter->data; if (member != orig_rsc) { member->cmds->this_with_colocations(member, orig_rsc, list); } } return; } /* Later group members honor the group's colocations indirectly, due to the * internal group colocations that chain everything from the first member. * However, if an earlier group member is unmanaged, this chaining will not * happen, so the group's mandatory colocations must be explicitly added. */ for (const GList *iter = rsc->children; iter != NULL; iter = iter->next) { member = iter->data; if (orig_rsc == member) { break; // We've seen all earlier members, and none are unmanaged } if (!pcmk_is_set(member->flags, pcmk_rsc_managed)) { crm_trace("Adding mandatory '%s with' colocations to list for " "member %s because earlier member %s is unmanaged", rsc->id, orig_rsc->id, member->id); for (const GList *cons_iter = rsc->rsc_cons; cons_iter != NULL; cons_iter = cons_iter->next) { const pcmk__colocation_t *colocation = NULL; colocation = (const pcmk__colocation_t *) cons_iter->data; if (colocation->score == INFINITY) { pcmk__add_this_with(list, colocation, orig_rsc); } } // @TODO Add mandatory (or all?) clone constraints if cloned break; } } } /*! * \internal * \brief Update nodes with scores of colocated resources' nodes * * Given a table of nodes and a resource, update the nodes' scores with the * scores of the best nodes matching the attribute used for each of the * resource's relevant colocations. * * \param[in,out] source_rsc Group resource whose node scores to add * \param[in] target_rsc Resource on whose behalf to update \p *nodes * \param[in] log_id Resource ID for logs (if \c NULL, use * \p source_rsc ID) * \param[in,out] nodes Nodes to update (set initial contents to \c NULL * to copy allowed nodes from \p source_rsc) * \param[in] colocation Original colocation constraint (used to get * configured primary resource's stickiness, and * to get colocation node attribute; if \c NULL, * source_rsc's own matching node scores will * not be added, and \p *nodes must be \c NULL as * well) * \param[in] factor Incorporate scores multiplied by this factor * \param[in] flags Bitmask of enum pcmk__coloc_select values * * \note \c NULL \p target_rsc, \c NULL \p *nodes, \c NULL \p colocation, and * the \c pcmk__coloc_select_this_with flag are used together (and only by * \c cmp_resources()). * \note The caller remains responsible for freeing \p *nodes. * \note This is the group implementation of * \c resource_alloc_functions_t:add_colocated_node_scores(). */ void pcmk__group_add_colocated_node_scores(pe_resource_t *source_rsc, const pe_resource_t *target_rsc, const char *log_id, GHashTable **nodes, const pcmk__colocation_t *colocation, float factor, uint32_t flags) { pe_resource_t *member = NULL; CRM_ASSERT((source_rsc != NULL) && (source_rsc->variant == pcmk_rsc_variant_group) && (nodes != NULL) && ((colocation != NULL) || ((target_rsc == NULL) && (*nodes == NULL)))); if (log_id == NULL) { log_id = source_rsc->id; } // Avoid infinite recursion if (pcmk_is_set(source_rsc->flags, pcmk_rsc_updating_nodes)) { pe_rsc_info(source_rsc, "%s: Breaking dependency loop at %s", log_id, source_rsc->id); return; } pe__set_resource_flags(source_rsc, pcmk_rsc_updating_nodes); // Ignore empty groups (only possible with schema validation disabled) if (source_rsc->children == NULL) { return; } /* Refer the operation to the first or last member as appropriate. * * cmp_resources() is the only caller that passes a NULL nodes table, * and is also the only caller using pcmk__coloc_select_this_with. * For "this with" colocations, the last member will recursively incorporate * all the other members' "this with" colocations via the internal group * colocations (and via the first member, the group's own colocations). * * For "with this" colocations, the first member works similarly. */ if (*nodes == NULL) { member = pe__last_group_member(source_rsc); } else { member = source_rsc->children->data; } pe_rsc_trace(source_rsc, "%s: Merging scores from group %s using member %s " "(at %.6f)", log_id, source_rsc->id, member->id, factor); member->cmds->add_colocated_node_scores(member, target_rsc, log_id, nodes, colocation, factor, flags); pe__clear_resource_flags(source_rsc, pcmk_rsc_updating_nodes); } // Group implementation of resource_alloc_functions_t:add_utilization() void pcmk__group_add_utilization(const pe_resource_t *rsc, const pe_resource_t *orig_rsc, GList *all_rscs, GHashTable *utilization) { pe_resource_t *member = NULL; CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_group) && (orig_rsc != NULL) && (utilization != NULL)); if (!pcmk_is_set(rsc->flags, pcmk_rsc_unassigned)) { return; } pe_rsc_trace(orig_rsc, "%s: Adding group %s as colocated utilization", orig_rsc->id, rsc->id); if (pe__group_flag_is_set(rsc, pe__group_colocated) || pe_rsc_is_clone(rsc->parent)) { // Every group member will be on same node, so sum all members for (GList *iter = rsc->children; iter != NULL; iter = iter->next) { member = (pe_resource_t *) iter->data; if (pcmk_is_set(member->flags, pcmk_rsc_unassigned) && (g_list_find(all_rscs, member) == NULL)) { member->cmds->add_utilization(member, orig_rsc, all_rscs, utilization); } } } else if (rsc->children != NULL) { // Just add first member's utilization member = (pe_resource_t *) rsc->children->data; if ((member != NULL) && pcmk_is_set(member->flags, pcmk_rsc_unassigned) && (g_list_find(all_rscs, member) == NULL)) { member->cmds->add_utilization(member, orig_rsc, all_rscs, utilization); } } } // Group implementation of resource_alloc_functions_t:shutdown_lock() void pcmk__group_shutdown_lock(pe_resource_t *rsc) { CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_group)); for (GList *iter = rsc->children; iter != NULL; iter = iter->next) { pe_resource_t *member = (pe_resource_t *) iter->data; member->cmds->shutdown_lock(member); } } diff --git a/lib/pacemaker/pcmk_sched_instances.c b/lib/pacemaker/pcmk_sched_instances.c index 9071ada3f2..6398000a33 100644 --- a/lib/pacemaker/pcmk_sched_instances.c +++ b/lib/pacemaker/pcmk_sched_instances.c @@ -1,1675 +1,1675 @@ /* * Copyright 2004-2023 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU General Public License version 2 * or later (GPLv2+) WITHOUT ANY WARRANTY. */ /* This file is intended for code usable with both clone instances and bundle * replica containers. */ #include #include #include #include "libpacemaker_private.h" /*! * \internal * \brief Check whether a node is allowed to run an instance * * \param[in] instance Clone instance or bundle container to check * \param[in] node Node to check * \param[in] max_per_node Maximum number of instances allowed to run on a node * * \return true if \p node is allowed to run \p instance, otherwise false */ static bool can_run_instance(const pe_resource_t *instance, const pe_node_t *node, int max_per_node) { pe_node_t *allowed_node = NULL; if (pcmk_is_set(instance->flags, pcmk_rsc_removed)) { pe_rsc_trace(instance, "%s cannot run on %s: orphaned", instance->id, pe__node_name(node)); return false; } if (!pcmk__node_available(node, false, false)) { pe_rsc_trace(instance, "%s cannot run on %s: node cannot run resources", instance->id, pe__node_name(node)); return false; } allowed_node = pcmk__top_allowed_node(instance, node); if (allowed_node == NULL) { crm_warn("%s cannot run on %s: node not allowed", instance->id, pe__node_name(node)); return false; } if (allowed_node->weight < 0) { pe_rsc_trace(instance, "%s cannot run on %s: parent score is %s there", instance->id, pe__node_name(node), pcmk_readable_score(allowed_node->weight)); return false; } if (allowed_node->count >= max_per_node) { pe_rsc_trace(instance, "%s cannot run on %s: node already has %d instance%s", instance->id, pe__node_name(node), max_per_node, pcmk__plural_s(max_per_node)); return false; } pe_rsc_trace(instance, "%s can run on %s (%d already running)", instance->id, pe__node_name(node), allowed_node->count); return true; } /*! * \internal * \brief Ban a clone instance or bundle replica from unavailable allowed nodes * * \param[in,out] instance Clone instance or bundle replica to ban * \param[in] max_per_node Maximum instances allowed to run on a node */ static void ban_unavailable_allowed_nodes(pe_resource_t *instance, int max_per_node) { if (instance->allowed_nodes != NULL) { GHashTableIter iter; pe_node_t *node = NULL; g_hash_table_iter_init(&iter, instance->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (void **) &node)) { if (!can_run_instance(instance, node, max_per_node)) { pe_rsc_trace(instance, "Banning %s from unavailable node %s", instance->id, pe__node_name(node)); node->weight = -INFINITY; for (GList *child_iter = instance->children; child_iter != NULL; child_iter = child_iter->next) { pe_resource_t *child = (pe_resource_t *) child_iter->data; pe_node_t *child_node = NULL; child_node = g_hash_table_lookup(child->allowed_nodes, node->details->id); if (child_node != NULL) { pe_rsc_trace(instance, "Banning %s child %s " "from unavailable node %s", instance->id, child->id, pe__node_name(node)); child_node->weight = -INFINITY; } } } } } } /*! * \internal * \brief Create a hash table with a single node in it * * \param[in] node Node to copy into new table * * \return Newly created hash table containing a copy of \p node * \note The caller is responsible for freeing the result with * g_hash_table_destroy(). */ static GHashTable * new_node_table(pe_node_t *node) { GHashTable *table = pcmk__strkey_table(NULL, free); node = pe__copy_node(node); g_hash_table_insert(table, (gpointer) node->details->id, node); return table; } /*! * \internal * \brief Apply a resource's parent's colocation scores to a node table * * \param[in] rsc Resource whose colocations should be applied * \param[in,out] nodes Node table to apply colocations to */ static void apply_parent_colocations(const pe_resource_t *rsc, GHashTable **nodes) { GList *colocations = pcmk__this_with_colocations(rsc); for (const GList *iter = colocations; iter != NULL; iter = iter->next) { const pcmk__colocation_t *colocation = iter->data; pe_resource_t *other = colocation->primary; float factor = colocation->score / (float) INFINITY; other->cmds->add_colocated_node_scores(other, rsc, rsc->id, nodes, colocation, factor, pcmk__coloc_select_default); } g_list_free(colocations); colocations = pcmk__with_this_colocations(rsc); for (const GList *iter = colocations; iter != NULL; iter = iter->next) { const pcmk__colocation_t *colocation = iter->data; pe_resource_t *other = colocation->dependent; float factor = colocation->score / (float) INFINITY; if (!pcmk__colocation_has_influence(colocation, rsc)) { continue; } other->cmds->add_colocated_node_scores(other, rsc, rsc->id, nodes, colocation, factor, pcmk__coloc_select_nonnegative); } g_list_free(colocations); } /*! * \internal * \brief Compare clone or bundle instances based on colocation scores * * Determine the relative order in which two clone or bundle instances should be * assigned to nodes, considering the scores of colocation constraints directly * or indirectly involving them. * * \param[in] instance1 First instance to compare * \param[in] instance2 Second instance to compare * * \return A negative number if \p instance1 should be assigned first, * a positive number if \p instance2 should be assigned first, * or 0 if assignment order doesn't matter */ static int cmp_instance_by_colocation(const pe_resource_t *instance1, const pe_resource_t *instance2) { int rc = 0; pe_node_t *node1 = NULL; pe_node_t *node2 = NULL; pe_node_t *current_node1 = pe__current_node(instance1); pe_node_t *current_node2 = pe__current_node(instance2); GHashTable *colocated_scores1 = NULL; GHashTable *colocated_scores2 = NULL; CRM_ASSERT((instance1 != NULL) && (instance1->parent != NULL) && (instance2 != NULL) && (instance2->parent != NULL) && (current_node1 != NULL) && (current_node2 != NULL)); // Create node tables initialized with each node colocated_scores1 = new_node_table(current_node1); colocated_scores2 = new_node_table(current_node2); // Apply parental colocations apply_parent_colocations(instance1, &colocated_scores1); apply_parent_colocations(instance2, &colocated_scores2); // Find original nodes again, with scores updated for colocations node1 = g_hash_table_lookup(colocated_scores1, current_node1->details->id); node2 = g_hash_table_lookup(colocated_scores2, current_node2->details->id); // Compare nodes by updated scores if (node1->weight < node2->weight) { crm_trace("Assign %s (%d on %s) after %s (%d on %s)", instance1->id, node1->weight, pe__node_name(node1), instance2->id, node2->weight, pe__node_name(node2)); rc = 1; } else if (node1->weight > node2->weight) { crm_trace("Assign %s (%d on %s) before %s (%d on %s)", instance1->id, node1->weight, pe__node_name(node1), instance2->id, node2->weight, pe__node_name(node2)); rc = -1; } g_hash_table_destroy(colocated_scores1); g_hash_table_destroy(colocated_scores2); return rc; } /*! * \internal * \brief Check whether a resource or any of its children are failed * * \param[in] rsc Resource to check * * \return true if \p rsc or any of its children are failed, otherwise false */ static bool did_fail(const pe_resource_t *rsc) { if (pcmk_is_set(rsc->flags, pcmk_rsc_failed)) { return true; } for (GList *iter = rsc->children; iter != NULL; iter = iter->next) { if (did_fail((const pe_resource_t *) iter->data)) { return true; } } return false; } /*! * \internal * \brief Check whether a node is allowed to run a resource * * \param[in] rsc Resource to check * \param[in,out] node Node to check (will be set NULL if not allowed) * * \return true if *node is either NULL or allowed for \p rsc, otherwise false */ static bool node_is_allowed(const pe_resource_t *rsc, pe_node_t **node) { if (*node != NULL) { pe_node_t *allowed = g_hash_table_lookup(rsc->allowed_nodes, (*node)->details->id); if ((allowed == NULL) || (allowed->weight < 0)) { pe_rsc_trace(rsc, "%s: current location (%s) is unavailable", rsc->id, pe__node_name(*node)); *node = NULL; return false; } } return true; } /*! * \internal * \brief Compare two clone or bundle instances' instance numbers * * \param[in] a First instance to compare * \param[in] b Second instance to compare * * \return A negative number if \p a's instance number is lower, * a positive number if \p b's instance number is lower, * or 0 if their instance numbers are the same */ gint pcmk__cmp_instance_number(gconstpointer a, gconstpointer b) { const pe_resource_t *instance1 = (const pe_resource_t *) a; const pe_resource_t *instance2 = (const pe_resource_t *) b; char *div1 = NULL; char *div2 = NULL; CRM_ASSERT((instance1 != NULL) && (instance2 != NULL)); // Clone numbers are after a colon, bundle numbers after a dash div1 = strrchr(instance1->id, ':'); if (div1 == NULL) { div1 = strrchr(instance1->id, '-'); } div2 = strrchr(instance2->id, ':'); if (div2 == NULL) { div2 = strrchr(instance2->id, '-'); } CRM_ASSERT((div1 != NULL) && (div2 != NULL)); return (gint) (strtol(div1 + 1, NULL, 10) - strtol(div2 + 1, NULL, 10)); } /*! * \internal * \brief Compare clone or bundle instances according to assignment order * * Compare two clone or bundle instances according to the order they should be * assigned to nodes, preferring (in order): * * - Active instance that is less multiply active * - Instance that is not active on a disallowed node * - Instance with higher configured priority * - Active instance whose current node can run resources * - Active instance whose parent is allowed on current node * - Active instance whose current node has fewer other instances * - Active instance * - Instance that isn't failed * - Instance whose colocations result in higher score on current node * - Instance with lower ID in lexicographic order * * \param[in] a First instance to compare * \param[in] b Second instance to compare * * \return A negative number if \p a should be assigned first, * a positive number if \p b should be assigned first, * or 0 if assignment order doesn't matter */ gint pcmk__cmp_instance(gconstpointer a, gconstpointer b) { int rc = 0; pe_node_t *node1 = NULL; pe_node_t *node2 = NULL; unsigned int nnodes1 = 0; unsigned int nnodes2 = 0; bool can1 = true; bool can2 = true; const pe_resource_t *instance1 = (const pe_resource_t *) a; const pe_resource_t *instance2 = (const pe_resource_t *) b; CRM_ASSERT((instance1 != NULL) && (instance2 != NULL)); node1 = instance1->fns->active_node(instance1, &nnodes1, NULL); node2 = instance2->fns->active_node(instance2, &nnodes2, NULL); /* If both instances are running and at least one is multiply * active, prefer instance that's running on fewer nodes. */ if ((nnodes1 > 0) && (nnodes2 > 0)) { if (nnodes1 < nnodes2) { crm_trace("Assign %s (active on %d) before %s (active on %d): " "less multiply active", instance1->id, nnodes1, instance2->id, nnodes2); return -1; } else if (nnodes1 > nnodes2) { crm_trace("Assign %s (active on %d) after %s (active on %d): " "more multiply active", instance1->id, nnodes1, instance2->id, nnodes2); return 1; } } /* An instance that is either inactive or active on an allowed node is * preferred over an instance that is active on a no-longer-allowed node. */ can1 = node_is_allowed(instance1, &node1); can2 = node_is_allowed(instance2, &node2); if (can1 && !can2) { crm_trace("Assign %s before %s: not active on a disallowed node", instance1->id, instance2->id); return -1; } else if (!can1 && can2) { crm_trace("Assign %s after %s: active on a disallowed node", instance1->id, instance2->id); return 1; } // Prefer instance with higher configured priority if (instance1->priority > instance2->priority) { crm_trace("Assign %s before %s: priority (%d > %d)", instance1->id, instance2->id, instance1->priority, instance2->priority); return -1; } else if (instance1->priority < instance2->priority) { crm_trace("Assign %s after %s: priority (%d < %d)", instance1->id, instance2->id, instance1->priority, instance2->priority); return 1; } // Prefer active instance if ((node1 == NULL) && (node2 == NULL)) { crm_trace("No assignment preference for %s vs. %s: inactive", instance1->id, instance2->id); return 0; } else if (node1 == NULL) { crm_trace("Assign %s after %s: active", instance1->id, instance2->id); return 1; } else if (node2 == NULL) { crm_trace("Assign %s before %s: active", instance1->id, instance2->id); return -1; } // Prefer instance whose current node can run resources can1 = pcmk__node_available(node1, false, false); can2 = pcmk__node_available(node2, false, false); if (can1 && !can2) { crm_trace("Assign %s before %s: current node can run resources", instance1->id, instance2->id); return -1; } else if (!can1 && can2) { crm_trace("Assign %s after %s: current node can't run resources", instance1->id, instance2->id); return 1; } // Prefer instance whose parent is allowed to run on instance's current node node1 = pcmk__top_allowed_node(instance1, node1); node2 = pcmk__top_allowed_node(instance2, node2); if ((node1 == NULL) && (node2 == NULL)) { crm_trace("No assignment preference for %s vs. %s: " "parent not allowed on either instance's current node", instance1->id, instance2->id); return 0; } else if (node1 == NULL) { crm_trace("Assign %s after %s: parent not allowed on current node", instance1->id, instance2->id); return 1; } else if (node2 == NULL) { crm_trace("Assign %s before %s: parent allowed on current node", instance1->id, instance2->id); return -1; } // Prefer instance whose current node is running fewer other instances if (node1->count < node2->count) { crm_trace("Assign %s before %s: fewer active instances on current node", instance1->id, instance2->id); return -1; } else if (node1->count > node2->count) { crm_trace("Assign %s after %s: more active instances on current node", instance1->id, instance2->id); return 1; } // Prefer instance that isn't failed can1 = did_fail(instance1); can2 = did_fail(instance2); if (!can1 && can2) { crm_trace("Assign %s before %s: not failed", instance1->id, instance2->id); return -1; } else if (can1 && !can2) { crm_trace("Assign %s after %s: failed", instance1->id, instance2->id); return 1; } // Prefer instance with higher cumulative colocation score on current node rc = cmp_instance_by_colocation(instance1, instance2); if (rc != 0) { return rc; } // Prefer instance with lower instance number rc = pcmk__cmp_instance_number(instance1, instance2); if (rc < 0) { crm_trace("Assign %s before %s: instance number", instance1->id, instance2->id); } else if (rc > 0) { crm_trace("Assign %s after %s: instance number", instance1->id, instance2->id); } else { crm_trace("No assignment preference for %s vs. %s", instance1->id, instance2->id); } return rc; } /*! * \internal * \brief Increment the parent's instance count after assigning an instance * * An instance's parent tracks how many instances have been assigned to each * node via its pe_node_t:count member. After assigning an instance to a node, * find the corresponding node in the parent's allowed table and increment it. * * \param[in,out] instance Instance whose parent to update * \param[in] assigned_to Node to which the instance was assigned */ static void increment_parent_count(pe_resource_t *instance, const pe_node_t *assigned_to) { pe_node_t *allowed = NULL; if (assigned_to == NULL) { return; } allowed = pcmk__top_allowed_node(instance, assigned_to); if (allowed == NULL) { /* The instance is allowed on the node, but its parent isn't. This * shouldn't be possible if the resource is managed, and we won't be * able to limit the number of instances assigned to the node. */ CRM_LOG_ASSERT(!pcmk_is_set(instance->flags, pcmk_rsc_managed)); } else { allowed->count++; } } /*! * \internal * \brief Assign an instance to a node * * \param[in,out] instance Clone instance or bundle replica container * \param[in] prefer If not NULL, attempt early assignment to this * node, if still the best choice; otherwise, * perform final assignment * \param[in] max_per_node Assign at most this many instances to one node * * \return Node to which \p instance is assigned */ static const pe_node_t * assign_instance(pe_resource_t *instance, const pe_node_t *prefer, int max_per_node) { pe_node_t *chosen = NULL; pe_rsc_trace(instance, "Assigning %s (preferring %s)", instance->id, ((prefer == NULL)? "no node" : prefer->details->uname)); if (pcmk_is_set(instance->flags, pcmk_rsc_assigning)) { pe_rsc_debug(instance, "Assignment loop detected involving %s colocations", instance->id); return NULL; } ban_unavailable_allowed_nodes(instance, max_per_node); // Failed early assignments are reversible (stop_if_fail=false) chosen = instance->cmds->assign(instance, prefer, (prefer == NULL)); increment_parent_count(instance, chosen); return chosen; } /*! * \internal * \brief Try to assign an instance to its current node early * * \param[in] rsc Clone or bundle being assigned (for logs only) * \param[in] instance Clone instance or bundle replica container * \param[in] current Instance's current node * \param[in] max_per_node Maximum number of instances per node * \param[in] available Number of instances still available for assignment * * \return \c true if \p instance was successfully assigned to its current node, * or \c false otherwise */ static bool assign_instance_early(const pe_resource_t *rsc, pe_resource_t *instance, const pe_node_t *current, int max_per_node, int available) { const pe_node_t *chosen = NULL; int reserved = 0; pe_resource_t *parent = instance->parent; GHashTable *allowed_orig = NULL; GHashTable *allowed_orig_parent = parent->allowed_nodes; const pe_node_t *allowed_node = g_hash_table_lookup(instance->allowed_nodes, current->details->id); pe_rsc_trace(instance, "Trying to assign %s to its current node %s", instance->id, pe__node_name(current)); if (!pcmk__node_available(allowed_node, true, false)) { pe_rsc_info(instance, "Not assigning %s to current node %s: unavailable", instance->id, pe__node_name(current)); return false; } /* On each iteration, if instance gets assigned to a node other than its * current one, we reserve one instance for the chosen node, unassign * instance, restore instance's original node tables, and try again. This * way, instances are proportionally assigned to nodes based on preferences, * but shuffling of specific instances is minimized. If a node will be * assigned instances at all, it preferentially receives instances that are * currently active there. * * parent->allowed_nodes tracks the number of instances assigned to each * node. If a node already has max_per_node instances assigned, * ban_unavailable_allowed_nodes() marks it as unavailable. * * In the end, we restore the original parent->allowed_nodes to undo the * changes to counts during tentative assignments. If we successfully * assigned instance to its current node, we increment that node's counter. */ // Back up the allowed node tables of instance and its children recursively pcmk__copy_node_tables(instance, &allowed_orig); // Update instances-per-node counts in a scratch table parent->allowed_nodes = pcmk__copy_node_table(parent->allowed_nodes); while (reserved < available) { chosen = assign_instance(instance, current, max_per_node); if (pe__same_node(chosen, current)) { // Successfully assigned to current node break; } // Assignment updates scores, so restore to original state pe_rsc_debug(instance, "Rolling back node scores for %s", instance->id); pcmk__restore_node_tables(instance, allowed_orig); if (chosen == NULL) { // Assignment failed, so give up pe_rsc_info(instance, "Not assigning %s to current node %s: unavailable", instance->id, pe__node_name(current)); pe__set_resource_flags(instance, pcmk_rsc_unassigned); break; } // We prefer more strongly to assign an instance to the chosen node pe_rsc_debug(instance, "Not assigning %s to current node %s: %s is better", instance->id, pe__node_name(current), pe__node_name(chosen)); // Reserve one instance for the chosen node and try again if (++reserved >= available) { pe_rsc_info(instance, "Not assigning %s to current node %s: " "other assignments are more important", instance->id, pe__node_name(current)); } else { pe_rsc_debug(instance, "Reserved an instance of %s for %s. Retrying " "assignment of %s to %s", rsc->id, pe__node_name(chosen), instance->id, pe__node_name(current)); } // Clear this assignment (frees chosen); leave instance counts in parent pcmk__unassign_resource(instance); chosen = NULL; } g_hash_table_destroy(allowed_orig); // Restore original instances-per-node counts g_hash_table_destroy(parent->allowed_nodes); parent->allowed_nodes = allowed_orig_parent; if (chosen == NULL) { // Couldn't assign instance to current node return false; } pe_rsc_trace(instance, "Assigned %s to current node %s", instance->id, pe__node_name(current)); increment_parent_count(instance, chosen); return true; } /*! * \internal * \brief Reset the node counts of a resource's allowed nodes to zero * * \param[in,out] rsc Resource to reset * * \return Number of nodes that are available to run resources */ static unsigned int reset_allowed_node_counts(pe_resource_t *rsc) { unsigned int available_nodes = 0; pe_node_t *node = NULL; GHashTableIter iter; g_hash_table_iter_init(&iter, rsc->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (gpointer *) &node)) { node->count = 0; if (pcmk__node_available(node, false, false)) { available_nodes++; } } return available_nodes; } /*! * \internal * \brief Check whether an instance has a preferred node * * \param[in] instance Clone instance or bundle replica container * \param[in] optimal_per_node Optimal number of instances per node * * \return Instance's current node if still available, otherwise NULL */ static const pe_node_t * preferred_node(const pe_resource_t *instance, int optimal_per_node) { const pe_node_t *node = NULL; const pe_node_t *parent_node = NULL; // Check whether instance is active, healthy, and not yet assigned if ((instance->running_on == NULL) || !pcmk_is_set(instance->flags, pcmk_rsc_unassigned) || pcmk_is_set(instance->flags, pcmk_rsc_failed)) { return NULL; } // Check whether instance's current node can run resources node = pe__current_node(instance); if (!pcmk__node_available(node, true, false)) { pe_rsc_trace(instance, "Not assigning %s to %s early (unavailable)", instance->id, pe__node_name(node)); return NULL; } // Check whether node already has optimal number of instances assigned parent_node = pcmk__top_allowed_node(instance, node); if ((parent_node != NULL) && (parent_node->count >= optimal_per_node)) { pe_rsc_trace(instance, "Not assigning %s to %s early " "(optimal instances already assigned)", instance->id, pe__node_name(node)); return NULL; } return node; } /*! * \internal * \brief Assign collective instances to nodes * * \param[in,out] collective Clone or bundle resource being assigned * \param[in,out] instances List of clone instances or bundle containers * \param[in] max_total Maximum instances to assign in total * \param[in] max_per_node Maximum instances to assign to any one node */ void pcmk__assign_instances(pe_resource_t *collective, GList *instances, int max_total, int max_per_node) { // Reuse node count to track number of assigned instances unsigned int available_nodes = reset_allowed_node_counts(collective); int optimal_per_node = 0; int assigned = 0; GList *iter = NULL; pe_resource_t *instance = NULL; const pe_node_t *current = NULL; if (available_nodes > 0) { optimal_per_node = max_total / available_nodes; } if (optimal_per_node < 1) { optimal_per_node = 1; } pe_rsc_debug(collective, "Assigning up to %d %s instance%s to up to %u node%s " "(at most %d per host, %d optimal)", max_total, collective->id, pcmk__plural_s(max_total), available_nodes, pcmk__plural_s(available_nodes), max_per_node, optimal_per_node); // Assign as many instances as possible to their current location for (iter = instances; (iter != NULL) && (assigned < max_total); iter = iter->next) { int available = max_total - assigned; instance = iter->data; if (!pcmk_is_set(instance->flags, pcmk_rsc_unassigned)) { continue; // Already assigned } current = preferred_node(instance, optimal_per_node); if ((current != NULL) && assign_instance_early(collective, instance, current, max_per_node, available)) { assigned++; } } pe_rsc_trace(collective, "Assigned %d of %d instance%s to current node", assigned, max_total, pcmk__plural_s(max_total)); for (iter = instances; iter != NULL; iter = iter->next) { instance = (pe_resource_t *) iter->data; if (!pcmk_is_set(instance->flags, pcmk_rsc_unassigned)) { continue; // Already assigned } if (instance->running_on != NULL) { current = pe__current_node(instance); if (pcmk__top_allowed_node(instance, current) == NULL) { const char *unmanaged = ""; if (!pcmk_is_set(instance->flags, pcmk_rsc_managed)) { unmanaged = "Unmanaged resource "; } crm_notice("%s%s is running on %s which is no longer allowed", unmanaged, instance->id, pe__node_name(current)); } } if (assigned >= max_total) { pe_rsc_debug(collective, "Not assigning %s because maximum %d instances " "already assigned", instance->id, max_total); resource_location(instance, NULL, -INFINITY, "collective_limit_reached", collective->cluster); } else if (assign_instance(instance, NULL, max_per_node) != NULL) { assigned++; } } pe_rsc_debug(collective, "Assigned %d of %d possible instance%s of %s", assigned, max_total, pcmk__plural_s(max_total), collective->id); } enum instance_state { instance_starting = (1 << 0), instance_stopping = (1 << 1), /* This indicates that some instance is restarting. It's not the same as * instance_starting|instance_stopping, which would indicate that some * instance is starting, and some instance (not necessarily the same one) is * stopping. */ instance_restarting = (1 << 2), instance_active = (1 << 3), instance_all = instance_starting|instance_stopping |instance_restarting|instance_active, }; /*! * \internal * \brief Check whether an instance is active, starting, and/or stopping * * \param[in] instance Clone instance or bundle replica container * \param[in,out] state Whether any instance is starting, stopping, etc. */ static void check_instance_state(const pe_resource_t *instance, uint32_t *state) { const GList *iter = NULL; uint32_t instance_state = 0; // State of just this instance // No need to check further if all conditions have already been detected if (pcmk_all_flags_set(*state, instance_all)) { return; } // If instance is a collective (a cloned group), check its children instead if (instance->variant > pcmk_rsc_variant_primitive) { for (iter = instance->children; (iter != NULL) && !pcmk_all_flags_set(*state, instance_all); iter = iter->next) { check_instance_state((const pe_resource_t *) iter->data, state); } return; } // If we get here, instance is a primitive if (instance->running_on != NULL) { instance_state |= instance_active; } // Check each of the instance's actions for runnable start or stop for (iter = instance->actions; (iter != NULL) && !pcmk_all_flags_set(instance_state, instance_starting |instance_stopping); iter = iter->next) { const pe_action_t *action = (const pe_action_t *) iter->data; const bool optional = pcmk_is_set(action->flags, pcmk_action_optional); if (pcmk__str_eq(PCMK_ACTION_START, action->task, pcmk__str_none)) { if (!optional && pcmk_is_set(action->flags, pcmk_action_runnable)) { pe_rsc_trace(instance, "Instance is starting due to %s", action->uuid); instance_state |= instance_starting; } else { pe_rsc_trace(instance, "%s doesn't affect %s state (%s)", action->uuid, instance->id, (optional? "optional" : "unrunnable")); } } else if (pcmk__str_eq(PCMK_ACTION_STOP, action->task, pcmk__str_none)) { /* Only stop actions can be pseudo-actions for primitives. That * indicates that the node they are on is being fenced, so the stop * is implied rather than actually executed. */ if (!optional && pcmk_any_flags_set(action->flags, pcmk_action_pseudo |pcmk_action_runnable)) { pe_rsc_trace(instance, "Instance is stopping due to %s", action->uuid); instance_state |= instance_stopping; } else { pe_rsc_trace(instance, "%s doesn't affect %s state (%s)", action->uuid, instance->id, (optional? "optional" : "unrunnable")); } } } if (pcmk_all_flags_set(instance_state, instance_starting|instance_stopping)) { instance_state |= instance_restarting; } *state |= instance_state; } /*! * \internal * \brief Create actions for collective resource instances * * \param[in,out] collective Clone or bundle resource to create actions for * \param[in,out] instances List of clone instances or bundle containers */ void pcmk__create_instance_actions(pe_resource_t *collective, GList *instances) { uint32_t state = 0; pe_action_t *stop = NULL; pe_action_t *stopped = NULL; pe_action_t *start = NULL; pe_action_t *started = NULL; pe_rsc_trace(collective, "Creating collective instance actions for %s", collective->id); // Create actions for each instance appropriate to its variant for (GList *iter = instances; iter != NULL; iter = iter->next) { pe_resource_t *instance = (pe_resource_t *) iter->data; instance->cmds->create_actions(instance); check_instance_state(instance, &state); } // Create pseudo-actions for rsc start and started start = pe__new_rsc_pseudo_action(collective, PCMK_ACTION_START, !pcmk_is_set(state, instance_starting), true); started = pe__new_rsc_pseudo_action(collective, PCMK_ACTION_RUNNING, !pcmk_is_set(state, instance_starting), false); started->priority = INFINITY; if (pcmk_any_flags_set(state, instance_active|instance_starting)) { pe__set_action_flags(started, pcmk_action_runnable); } // Create pseudo-actions for rsc stop and stopped stop = pe__new_rsc_pseudo_action(collective, PCMK_ACTION_STOP, !pcmk_is_set(state, instance_stopping), true); stopped = pe__new_rsc_pseudo_action(collective, PCMK_ACTION_STOPPED, !pcmk_is_set(state, instance_stopping), true); stopped->priority = INFINITY; if (!pcmk_is_set(state, instance_restarting)) { pe__set_action_flags(stop, pcmk_action_migratable); } if (collective->variant == pcmk_rsc_variant_clone) { pe__create_clone_notif_pseudo_ops(collective, start, started, stop, stopped); } } /*! * \internal * \brief Get a list of clone instances or bundle replica containers * * \param[in] rsc Clone or bundle resource * * \return Clone instances if \p rsc is a clone, or a newly created list of * \p rsc's replica containers if \p rsc is a bundle * \note The caller must call free_instance_list() on the result when the list * is no longer needed. */ static inline GList * get_instance_list(const pe_resource_t *rsc) { if (rsc->variant == pcmk_rsc_variant_bundle) { return pe__bundle_containers(rsc); } else { return rsc->children; } } /*! * \internal * \brief Free any memory created by get_instance_list() * * \param[in] rsc Clone or bundle resource passed to get_instance_list() * \param[in,out] list Return value of get_instance_list() for \p rsc */ static inline void free_instance_list(const pe_resource_t *rsc, GList *list) { if (list != rsc->children) { g_list_free(list); } } /*! * \internal * \brief Check whether an instance is compatible with a role and node * * \param[in] instance Clone instance or bundle replica container * \param[in] node Instance must match this node * \param[in] role If not pcmk_role_unknown, instance must match this role * \param[in] current If true, compare instance's original node and role, * otherwise compare assigned next node and role * * \return true if \p instance is compatible with \p node and \p role, * otherwise false */ bool pcmk__instance_matches(const pe_resource_t *instance, const pe_node_t *node, enum rsc_role_e role, bool current) { pe_node_t *instance_node = NULL; CRM_CHECK((instance != NULL) && (node != NULL), return false); if ((role != pcmk_role_unknown) && (role != instance->fns->state(instance, current))) { pe_rsc_trace(instance, "%s is not a compatible instance (role is not %s)", instance->id, role2text(role)); return false; } if (!is_set_recursive(instance, pcmk_rsc_blocked, true)) { // We only want instances that haven't failed instance_node = instance->fns->location(instance, NULL, current); } if (instance_node == NULL) { pe_rsc_trace(instance, "%s is not a compatible instance (not assigned to a node)", instance->id); return false; } if (!pe__same_node(instance_node, node)) { pe_rsc_trace(instance, "%s is not a compatible instance (assigned to %s not %s)", instance->id, pe__node_name(instance_node), pe__node_name(node)); return false; } return true; } /*! * \internal * \brief Find an instance that matches a given resource by node and role * * \param[in] match_rsc Resource that instance must match (for logging only) * \param[in] rsc Clone or bundle resource to check for matching instance * \param[in] node Instance must match this node * \param[in] role If not pcmk_role_unknown, instance must match this role * \param[in] current If true, compare instance's original node and role, * otherwise compare assigned next node and role * * \return \p rsc instance matching \p node and \p role if any, otherwise NULL */ static pe_resource_t * find_compatible_instance_on_node(const pe_resource_t *match_rsc, const pe_resource_t *rsc, const pe_node_t *node, enum rsc_role_e role, bool current) { GList *instances = NULL; instances = get_instance_list(rsc); for (GList *iter = instances; iter != NULL; iter = iter->next) { pe_resource_t *instance = (pe_resource_t *) iter->data; if (pcmk__instance_matches(instance, node, role, current)) { pe_rsc_trace(match_rsc, "Found %s %s instance %s compatible with %s on %s", role == pcmk_role_unknown? "matching" : role2text(role), rsc->id, instance->id, match_rsc->id, pe__node_name(node)); free_instance_list(rsc, instances); // Only frees list, not contents return instance; } } free_instance_list(rsc, instances); pe_rsc_trace(match_rsc, "No %s %s instance found compatible with %s on %s", ((role == pcmk_role_unknown)? "matching" : role2text(role)), rsc->id, match_rsc->id, pe__node_name(node)); return NULL; } /*! * \internal * \brief Find a clone instance or bundle container compatible with a resource * * \param[in] match_rsc Resource that instance must match * \param[in] rsc Clone or bundle resource to check for matching instance * \param[in] role If not pcmk_role_unknown, instance must match this role * \param[in] current If true, compare instance's original node and role, * otherwise compare assigned next node and role * * \return Compatible (by \p role and \p match_rsc location) instance of \p rsc * if any, otherwise NULL */ pe_resource_t * pcmk__find_compatible_instance(const pe_resource_t *match_rsc, const pe_resource_t *rsc, enum rsc_role_e role, bool current) { pe_resource_t *instance = NULL; GList *nodes = NULL; const pe_node_t *node = match_rsc->fns->location(match_rsc, NULL, current); // If match_rsc has a node, check only that node if (node != NULL) { return find_compatible_instance_on_node(match_rsc, rsc, node, role, current); } // Otherwise check for an instance matching any of match_rsc's allowed nodes nodes = pcmk__sort_nodes(g_hash_table_get_values(match_rsc->allowed_nodes), NULL); for (GList *iter = nodes; (iter != NULL) && (instance == NULL); iter = iter->next) { instance = find_compatible_instance_on_node(match_rsc, rsc, (pe_node_t *) iter->data, role, current); } if (instance == NULL) { pe_rsc_debug(rsc, "No %s instance found compatible with %s", rsc->id, match_rsc->id); } g_list_free(nodes); return instance; } /*! * \internal * \brief Unassign an instance if mandatory ordering has no interleave match * * \param[in] first 'First' action in an ordering * \param[in] then 'Then' action in an ordering * \param[in,out] then_instance 'Then' instance that has no interleave match * \param[in] type Group of enum pcmk__action_relation_flags * \param[in] current If true, "then" action is stopped or demoted * * \return true if \p then_instance was unassigned, otherwise false */ static bool unassign_if_mandatory(const pe_action_t *first, const pe_action_t *then, pe_resource_t *then_instance, uint32_t type, bool current) { // Allow "then" instance to go down even without an interleave match if (current) { pe_rsc_trace(then->rsc, "%s has no instance to order before stopping " "or demoting %s", first->rsc->id, then_instance->id); /* If the "first" action must be runnable, but there is no "first" * instance, the "then" instance must not be allowed to come up. */ - } else if (pcmk_any_flags_set(type, pe_order_runnable_left + } else if (pcmk_any_flags_set(type, pcmk__ar_unrunnable_first_blocks |pcmk__ar_first_implies_then)) { pe_rsc_info(then->rsc, "Inhibiting %s from being active " "because there is no %s instance to interleave", then_instance->id, first->rsc->id); return pcmk__assign_resource(then_instance, NULL, true, true); } return false; } /*! * \internal * \brief Find first matching action for a clone instance or bundle container * * \param[in] action Action in an interleaved ordering * \param[in] instance Clone instance or bundle container being interleaved * \param[in] action_name Action to look for * \param[in] node If not NULL, require action to be on this node * \param[in] for_first If true, \p instance is the 'first' resource in the * ordering, otherwise it is the 'then' resource * * \return First action for \p instance (or in some cases if \p instance is a * bundle container, its containerized resource) that matches * \p action_name and \p node if any, otherwise NULL */ static pe_action_t * find_instance_action(const pe_action_t *action, const pe_resource_t *instance, const char *action_name, const pe_node_t *node, bool for_first) { const pe_resource_t *rsc = NULL; pe_action_t *matching_action = NULL; /* If instance is a bundle container, sometimes we should interleave the * action for the container itself, and sometimes for the containerized * resource. * * For example, given "start bundle A then bundle B", B likely requires the * service inside A's container to be active, rather than just the * container, so we should interleave the action for A's containerized * resource. On the other hand, it's possible B's container itself requires * something from A, so we should interleave the action for B's container. * * Essentially, for 'first', we should use the containerized resource for * everything except stop, and for 'then', we should use the container for * everything except promote and demote (which can only be performed on the * containerized resource). */ if ((for_first && !pcmk__str_any_of(action->task, PCMK_ACTION_STOP, PCMK_ACTION_STOPPED, NULL)) || (!for_first && pcmk__str_any_of(action->task, PCMK_ACTION_PROMOTE, PCMK_ACTION_PROMOTED, PCMK_ACTION_DEMOTE, PCMK_ACTION_DEMOTED, NULL))) { rsc = pe__get_rsc_in_container(instance); } if (rsc == NULL) { rsc = instance; // No containerized resource, use instance itself } else { node = NULL; // Containerized actions are on bundle-created guest } matching_action = find_first_action(rsc->actions, NULL, action_name, node); if (matching_action != NULL) { return matching_action; } if (pcmk_is_set(instance->flags, pcmk_rsc_removed) || pcmk__str_any_of(action_name, PCMK_ACTION_STOP, PCMK_ACTION_DEMOTE, NULL)) { crm_trace("No %s action found for %s%s", action_name, pcmk_is_set(instance->flags, pcmk_rsc_removed)? "orphan " : "", instance->id); } else { crm_err("No %s action found for %s to interleave (bug?)", action_name, instance->id); } return NULL; } /*! * \internal * \brief Get the original action name of a bundle or clone action * * Given an action for a bundle or clone, get the original action name, * mapping notify to the action being notified, and if the instances are * primitives, mapping completion actions to the action that was completed * (for example, stopped to stop). * * \param[in] action Clone or bundle action to check * * \return Original action name for \p action */ static const char * orig_action_name(const pe_action_t *action) { const pe_resource_t *instance = action->rsc->children->data; // Any instance char *action_type = NULL; const char *action_name = action->task; enum action_tasks orig_task = pcmk_action_unspecified; if (pcmk__strcase_any_of(action->task, PCMK_ACTION_NOTIFY, PCMK_ACTION_NOTIFIED, NULL)) { // action->uuid is RSC_(confirmed-){pre,post}_notify_ACTION_INTERVAL CRM_CHECK(parse_op_key(action->uuid, NULL, &action_type, NULL), return task2text(pcmk_action_unspecified)); action_name = strstr(action_type, "_notify_"); CRM_CHECK(action_name != NULL, return task2text(pcmk_action_unspecified)); action_name += strlen("_notify_"); } orig_task = get_complex_task(instance, action_name); free(action_type); return task2text(orig_task); } /*! * \internal * \brief Update two interleaved actions according to an ordering between them * * Given information about an ordering of two interleaved actions, update the * actions' flags (and runnable_before members if appropriate) as appropriate * for the ordering. Effects may cascade to other orderings involving the * actions as well. * * \param[in,out] first 'First' action in an ordering * \param[in,out] then 'Then' action in an ordering * \param[in] node If not NULL, limit scope of ordering to this node * \param[in] filter Action flags to limit scope of certain updates (may * include pcmk_action_optional to affect only * mandatory actions, and pcmk_action_runnable to * affect only runnable actions) * \param[in] type Group of enum pcmk__action_relation_flags to apply * * \return Group of enum pcmk__updated flags indicating what was updated */ static uint32_t update_interleaved_actions(pe_action_t *first, pe_action_t *then, const pe_node_t *node, uint32_t filter, uint32_t type) { GList *instances = NULL; uint32_t changed = pcmk__updated_none; const char *orig_first_task = orig_action_name(first); // Stops and demotes must be interleaved with instance on current node bool current = pcmk__ends_with(first->uuid, "_" PCMK_ACTION_STOPPED "_0") || pcmk__ends_with(first->uuid, "_" PCMK_ACTION_DEMOTED "_0"); // Update the specified actions for each "then" instance individually instances = get_instance_list(then->rsc); for (GList *iter = instances; iter != NULL; iter = iter->next) { pe_resource_t *first_instance = NULL; pe_resource_t *then_instance = iter->data; pe_action_t *first_action = NULL; pe_action_t *then_action = NULL; // Find a "first" instance to interleave with this "then" instance first_instance = pcmk__find_compatible_instance(then_instance, first->rsc, pcmk_role_unknown, current); if (first_instance == NULL) { // No instance can be interleaved if (unassign_if_mandatory(first, then, then_instance, type, current)) { pcmk__set_updated_flags(changed, first, pcmk__updated_then); } continue; } first_action = find_instance_action(first, first_instance, orig_first_task, node, true); if (first_action == NULL) { continue; } then_action = find_instance_action(then, then_instance, then->task, node, false); if (then_action == NULL) { continue; } if (order_actions(first_action, then_action, type)) { pcmk__set_updated_flags(changed, first, pcmk__updated_first|pcmk__updated_then); } changed |= then_instance->cmds->update_ordered_actions( first_action, then_action, node, first_instance->cmds->action_flags(first_action, node), filter, type, then->rsc->cluster); } free_instance_list(then->rsc, instances); return changed; } /*! * \internal * \brief Check whether two actions in an ordering can be interleaved * * \param[in] first 'First' action in the ordering * \param[in] then 'Then' action in the ordering * * \return true if \p first and \p then can be interleaved, otherwise false */ static bool can_interleave_actions(const pe_action_t *first, const pe_action_t *then) { bool interleave = false; pe_resource_t *rsc = NULL; if ((first->rsc == NULL) || (then->rsc == NULL)) { crm_trace("Not interleaving %s with %s: not resource actions", first->uuid, then->uuid); return false; } if (first->rsc == then->rsc) { crm_trace("Not interleaving %s with %s: same resource", first->uuid, then->uuid); return false; } if ((first->rsc->variant < pcmk_rsc_variant_clone) || (then->rsc->variant < pcmk_rsc_variant_clone)) { crm_trace("Not interleaving %s with %s: not clones or bundles", first->uuid, then->uuid); return false; } if (pcmk__ends_with(then->uuid, "_stop_0") || pcmk__ends_with(then->uuid, "_demote_0")) { rsc = first->rsc; } else { rsc = then->rsc; } interleave = crm_is_true(g_hash_table_lookup(rsc->meta, XML_RSC_ATTR_INTERLEAVE)); pe_rsc_trace(rsc, "'%s then %s' will %sbe interleaved (based on %s)", first->uuid, then->uuid, (interleave? "" : "not "), rsc->id); return interleave; } /*! * \internal * \brief Update non-interleaved instance actions according to an ordering * * Given information about an ordering of two non-interleaved actions, update * the actions' flags (and runnable_before members if appropriate) as * appropriate for the ordering. Effects may cascade to other orderings * involving the actions as well. * * \param[in,out] instance Clone instance or bundle container * \param[in,out] first "First" action in ordering * \param[in] then "Then" action in ordering (for \p instance's parent) * \param[in] node If not NULL, limit scope of ordering to this node * \param[in] flags Action flags for \p first for ordering purposes * \param[in] filter Action flags to limit scope of certain updates (may * include pcmk_action_optional to affect only * mandatory actions, and pcmk_action_runnable to * affect only runnable actions) * \param[in] type Group of enum pcmk__action_relation_flags to apply * * \return Group of enum pcmk__updated flags indicating what was updated */ static uint32_t update_noninterleaved_actions(pe_resource_t *instance, pe_action_t *first, const pe_action_t *then, const pe_node_t *node, uint32_t flags, uint32_t filter, uint32_t type) { pe_action_t *instance_action = NULL; uint32_t instance_flags = 0; uint32_t changed = pcmk__updated_none; // Check whether instance has an equivalent of "then" action instance_action = find_first_action(instance->actions, NULL, then->task, node); if (instance_action == NULL) { return changed; } // Check whether action is runnable instance_flags = instance->cmds->action_flags(instance_action, node); if (!pcmk_is_set(instance_flags, pcmk_action_runnable)) { return changed; } // If so, update actions for the instance changed = instance->cmds->update_ordered_actions(first, instance_action, node, flags, filter, type, instance->cluster); // Propagate any changes to later actions if (pcmk_is_set(changed, pcmk__updated_then)) { for (GList *after_iter = instance_action->actions_after; after_iter != NULL; after_iter = after_iter->next) { pe_action_wrapper_t *after = after_iter->data; pcmk__update_action_for_orderings(after->action, instance->cluster); } } return changed; } /*! * \internal * \brief Update two actions according to an ordering between them * * Given information about an ordering of two clone or bundle actions, update * the actions' flags (and runnable_before members if appropriate) as * appropriate for the ordering. Effects may cascade to other orderings * involving the actions as well. * * \param[in,out] first 'First' action in an ordering * \param[in,out] then 'Then' action in an ordering * \param[in] node If not NULL, limit scope of ordering to this node * (only used when interleaving instances) * \param[in] flags Action flags for \p first for ordering purposes * \param[in] filter Action flags to limit scope of certain updates (may * include pcmk_action_optional to affect only * mandatory actions, and pcmk_action_runnable to * affect only runnable actions) * \param[in] type Group of enum pcmk__action_relation_flags to apply * \param[in,out] data_set Cluster working set * * \return Group of enum pcmk__updated flags indicating what was updated */ uint32_t pcmk__instance_update_ordered_actions(pe_action_t *first, pe_action_t *then, const pe_node_t *node, uint32_t flags, uint32_t filter, uint32_t type, pe_working_set_t *data_set) { CRM_ASSERT((first != NULL) && (then != NULL) && (data_set != NULL)); if (then->rsc == NULL) { return pcmk__updated_none; } else if (can_interleave_actions(first, then)) { return update_interleaved_actions(first, then, node, filter, type); } else { uint32_t changed = pcmk__updated_none; GList *instances = get_instance_list(then->rsc); // Update actions for the clone or bundle resource itself changed |= pcmk__update_ordered_actions(first, then, node, flags, filter, type, data_set); // Update the 'then' clone instances or bundle containers individually for (GList *iter = instances; iter != NULL; iter = iter->next) { pe_resource_t *instance = iter->data; changed |= update_noninterleaved_actions(instance, first, then, node, flags, filter, type); } free_instance_list(then->rsc, instances); return changed; } } #define pe__clear_action_summary_flags(flags, action, flag) do { \ flags = pcmk__clear_flags_as(__func__, __LINE__, LOG_TRACE, \ "Action summary", action->rsc->id, \ flags, flag, #flag); \ } while (0) /*! * \internal * \brief Return action flags for a given clone or bundle action * * \param[in,out] action Action for a clone or bundle * \param[in] instances Clone instances or bundle containers * \param[in] node If not NULL, limit effects to this node * * \return Flags appropriate to \p action on \p node */ uint32_t pcmk__collective_action_flags(pe_action_t *action, const GList *instances, const pe_node_t *node) { bool any_runnable = false; const char *action_name = orig_action_name(action); // Set original assumptions (optional and runnable may be cleared below) uint32_t flags = pcmk_action_optional |pcmk_action_runnable |pcmk_action_pseudo; for (const GList *iter = instances; iter != NULL; iter = iter->next) { const pe_resource_t *instance = iter->data; const pe_node_t *instance_node = NULL; pe_action_t *instance_action = NULL; uint32_t instance_flags; // Node is relevant only to primitive instances if (instance->variant == pcmk_rsc_variant_primitive) { instance_node = node; } instance_action = find_first_action(instance->actions, NULL, action_name, instance_node); if (instance_action == NULL) { pe_rsc_trace(action->rsc, "%s has no %s action on %s", instance->id, action_name, pe__node_name(node)); continue; } pe_rsc_trace(action->rsc, "%s has %s for %s on %s", instance->id, instance_action->uuid, action_name, pe__node_name(node)); instance_flags = instance->cmds->action_flags(instance_action, node); // If any instance action is mandatory, so is the collective action if (pcmk_is_set(flags, pcmk_action_optional) && !pcmk_is_set(instance_flags, pcmk_action_optional)) { pe_rsc_trace(instance, "%s is mandatory because %s is", action->uuid, instance_action->uuid); pe__clear_action_summary_flags(flags, action, pcmk_action_optional); pe__clear_action_flags(action, pcmk_action_optional); } // If any instance action is runnable, so is the collective action if (pcmk_is_set(instance_flags, pcmk_action_runnable)) { any_runnable = true; } } if (!any_runnable) { pe_rsc_trace(action->rsc, "%s is not runnable because no instance can run %s", action->uuid, action_name); pe__clear_action_summary_flags(flags, action, pcmk_action_runnable); if (node == NULL) { pe__clear_action_flags(action, pcmk_action_runnable); } } return flags; } diff --git a/lib/pacemaker/pcmk_sched_ordering.c b/lib/pacemaker/pcmk_sched_ordering.c index 6a9ad13ba2..7a51233514 100644 --- a/lib/pacemaker/pcmk_sched_ordering.c +++ b/lib/pacemaker/pcmk_sched_ordering.c @@ -1,1490 +1,1493 @@ /* * Copyright 2004-2023 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU General Public License version 2 * or later (GPLv2+) WITHOUT ANY WARRANTY. */ #include #include // PRIx32 #include #include #include #include #include "libpacemaker_private.h" enum pe_order_kind { pe_order_kind_optional, pe_order_kind_mandatory, pe_order_kind_serialize, }; enum ordering_symmetry { ordering_asymmetric, // the only relation in an asymmetric ordering ordering_symmetric, // the normal relation in a symmetric ordering ordering_symmetric_inverse, // the inverse relation in a symmetric ordering }; #define EXPAND_CONSTRAINT_IDREF(__set, __rsc, __name) do { \ __rsc = pcmk__find_constraint_resource(data_set->resources, \ __name); \ if (__rsc == NULL) { \ pcmk__config_err("%s: No resource found for %s", __set, __name);\ return pcmk_rc_unpack_error; \ } \ } while (0) static const char * invert_action(const char *action) { if (pcmk__str_eq(action, PCMK_ACTION_START, pcmk__str_none)) { return PCMK_ACTION_STOP; } else if (pcmk__str_eq(action, PCMK_ACTION_STOP, pcmk__str_none)) { return PCMK_ACTION_START; } else if (pcmk__str_eq(action, PCMK_ACTION_PROMOTE, pcmk__str_none)) { return PCMK_ACTION_DEMOTE; } else if (pcmk__str_eq(action, PCMK_ACTION_DEMOTE, pcmk__str_none)) { return PCMK_ACTION_PROMOTE; } else if (pcmk__str_eq(action, PCMK_ACTION_PROMOTED, pcmk__str_none)) { return PCMK_ACTION_DEMOTED; } else if (pcmk__str_eq(action, PCMK_ACTION_DEMOTED, pcmk__str_none)) { return PCMK_ACTION_PROMOTED; } else if (pcmk__str_eq(action, PCMK_ACTION_RUNNING, pcmk__str_none)) { return PCMK_ACTION_STOPPED; } else if (pcmk__str_eq(action, PCMK_ACTION_STOPPED, pcmk__str_none)) { return PCMK_ACTION_RUNNING; } crm_warn("Unknown action '%s' specified in order constraint", action); return NULL; } static enum pe_order_kind get_ordering_type(const xmlNode *xml_obj) { enum pe_order_kind kind_e = pe_order_kind_mandatory; const char *kind = crm_element_value(xml_obj, XML_ORDER_ATTR_KIND); if (kind == NULL) { const char *score = crm_element_value(xml_obj, XML_RULE_ATTR_SCORE); kind_e = pe_order_kind_mandatory; if (score) { // @COMPAT deprecated informally since 1.0.7, formally since 2.0.1 int score_i = char2score(score); if (score_i == 0) { kind_e = pe_order_kind_optional; } pe_warn_once(pe_wo_order_score, "Support for 'score' in rsc_order is deprecated " "and will be removed in a future release " "(use 'kind' instead)"); } } else if (pcmk__str_eq(kind, "Mandatory", pcmk__str_none)) { kind_e = pe_order_kind_mandatory; } else if (pcmk__str_eq(kind, "Optional", pcmk__str_none)) { kind_e = pe_order_kind_optional; } else if (pcmk__str_eq(kind, "Serialize", pcmk__str_none)) { kind_e = pe_order_kind_serialize; } else { pcmk__config_err("Resetting '" XML_ORDER_ATTR_KIND "' for constraint " "%s to 'Mandatory' because '%s' is not valid", pcmk__s(ID(xml_obj), "missing ID"), kind); } return kind_e; } /*! * \internal * \brief Get ordering symmetry from XML * * \param[in] xml_obj Ordering XML * \param[in] parent_kind Default ordering kind * \param[in] parent_symmetrical_s Parent element's symmetrical setting, if any * * \retval ordering_symmetric Ordering is symmetric * \retval ordering_asymmetric Ordering is asymmetric */ static enum ordering_symmetry get_ordering_symmetry(const xmlNode *xml_obj, enum pe_order_kind parent_kind, const char *parent_symmetrical_s) { int rc = pcmk_rc_ok; bool symmetric = false; enum pe_order_kind kind = parent_kind; // Default to parent's kind // Check ordering XML for explicit kind if ((crm_element_value(xml_obj, XML_ORDER_ATTR_KIND) != NULL) || (crm_element_value(xml_obj, XML_RULE_ATTR_SCORE) != NULL)) { kind = get_ordering_type(xml_obj); } // Check ordering XML (and parent) for explicit symmetrical setting rc = pcmk__xe_get_bool_attr(xml_obj, XML_CONS_ATTR_SYMMETRICAL, &symmetric); if (rc != pcmk_rc_ok && parent_symmetrical_s != NULL) { symmetric = crm_is_true(parent_symmetrical_s); rc = pcmk_rc_ok; } if (rc == pcmk_rc_ok) { if (symmetric) { if (kind == pe_order_kind_serialize) { pcmk__config_warn("Ignoring " XML_CONS_ATTR_SYMMETRICAL " for '%s' because not valid with " XML_ORDER_ATTR_KIND " of 'Serialize'", ID(xml_obj)); } else { return ordering_symmetric; } } return ordering_asymmetric; } // Use default symmetry if (kind == pe_order_kind_serialize) { return ordering_asymmetric; } return ordering_symmetric; } /*! * \internal * \brief Get ordering flags appropriate to ordering kind * * \param[in] kind Ordering kind * \param[in] first Action name for 'first' action * \param[in] symmetry This ordering's symmetry role * * \return Minimal ordering flags appropriate to \p kind */ static uint32_t ordering_flags_for_kind(enum pe_order_kind kind, const char *first, enum ordering_symmetry symmetry) { uint32_t flags = pcmk__ar_none; // so we trace-log all flags set pe__set_order_flags(flags, pcmk__ar_ordered); switch (kind) { case pe_order_kind_optional: break; case pe_order_kind_serialize: pe__set_order_flags(flags, pe_order_serialize_only); break; case pe_order_kind_mandatory: switch (symmetry) { case ordering_asymmetric: pe__set_order_flags(flags, pe_order_asymmetrical); break; case ordering_symmetric: pe__set_order_flags(flags, pcmk__ar_first_implies_then); if (pcmk__strcase_any_of(first, PCMK_ACTION_START, PCMK_ACTION_PROMOTE, NULL)) { - pe__set_order_flags(flags, pe_order_runnable_left); + pe__set_order_flags(flags, + pcmk__ar_unrunnable_first_blocks); } break; case ordering_symmetric_inverse: pe__set_order_flags(flags, pcmk__ar_then_implies_first); break; } break; } return flags; } /*! * \internal * \brief Find resource corresponding to ID specified in ordering * * \param[in] xml Ordering XML * \param[in] resource_attr XML attribute name for resource ID * \param[in] instance_attr XML attribute name for instance number. * This option is deprecated and will be removed in a * future release. * \param[in] data_set Cluster working set * * \return Resource corresponding to \p id, or NULL if none */ static pe_resource_t * get_ordering_resource(const xmlNode *xml, const char *resource_attr, const char *instance_attr, const pe_working_set_t *data_set) { // @COMPAT: instance_attr and instance_id variables deprecated since 2.1.5 pe_resource_t *rsc = NULL; const char *rsc_id = crm_element_value(xml, resource_attr); const char *instance_id = crm_element_value(xml, instance_attr); if (rsc_id == NULL) { pcmk__config_err("Ignoring constraint '%s' without %s", ID(xml), resource_attr); return NULL; } rsc = pcmk__find_constraint_resource(data_set->resources, rsc_id); if (rsc == NULL) { pcmk__config_err("Ignoring constraint '%s' because resource '%s' " "does not exist", ID(xml), rsc_id); return NULL; } if (instance_id != NULL) { pe_warn_once(pe_wo_order_inst, "Support for " XML_ORDER_ATTR_FIRST_INSTANCE " and " XML_ORDER_ATTR_THEN_INSTANCE " is deprecated and will be " "removed in a future release."); if (!pe_rsc_is_clone(rsc)) { pcmk__config_err("Ignoring constraint '%s' because resource '%s' " "is not a clone but instance '%s' was requested", ID(xml), rsc_id, instance_id); return NULL; } rsc = find_clone_instance(rsc, instance_id); if (rsc == NULL) { pcmk__config_err("Ignoring constraint '%s' because resource '%s' " "does not have an instance '%s'", "'%s'", ID(xml), rsc_id, instance_id); return NULL; } } return rsc; } /*! * \internal * \brief Determine minimum number of 'first' instances required in ordering * * \param[in] rsc 'First' resource in ordering * \param[in] xml Ordering XML * * \return Minimum 'first' instances required (or 0 if not applicable) */ static int get_minimum_first_instances(const pe_resource_t *rsc, const xmlNode *xml) { const char *clone_min = NULL; bool require_all = false; if (!pe_rsc_is_clone(rsc)) { return 0; } clone_min = g_hash_table_lookup(rsc->meta, PCMK_META_CLONE_MIN); if (clone_min != NULL) { int clone_min_int = 0; pcmk__scan_min_int(clone_min, &clone_min_int, 0); return clone_min_int; } /* @COMPAT 1.1.13: * require-all=false is deprecated equivalent of clone-min=1 */ if (pcmk__xe_get_bool_attr(xml, "require-all", &require_all) != ENODATA) { pe_warn_once(pe_wo_require_all, "Support for require-all in ordering constraints " "is deprecated and will be removed in a future release" " (use clone-min clone meta-attribute instead)"); if (!require_all) { return 1; } } return 0; } /*! * \internal * \brief Create orderings for a constraint with clone-min > 0 * * \param[in] id Ordering ID * \param[in,out] rsc_first 'First' resource in ordering (a clone) * \param[in] action_first 'First' action in ordering * \param[in] rsc_then 'Then' resource in ordering * \param[in] action_then 'Then' action in ordering * \param[in] flags Ordering flags * \param[in] clone_min Minimum required instances of 'first' */ static void clone_min_ordering(const char *id, pe_resource_t *rsc_first, const char *action_first, pe_resource_t *rsc_then, const char *action_then, uint32_t flags, int clone_min) { // Create a pseudo-action for when the minimum instances are active char *task = crm_strdup_printf(PCMK_ACTION_CLONE_ONE_OR_MORE ":%s", id); pe_action_t *clone_min_met = get_pseudo_op(task, rsc_first->cluster); free(task); /* Require the pseudo-action to have the required number of actions to be * considered runnable before allowing the pseudo-action to be runnable. */ clone_min_met->required_runnable_before = clone_min; pe__set_action_flags(clone_min_met, pcmk_action_min_runnable); // Order the actions for each clone instance before the pseudo-action for (GList *iter = rsc_first->children; iter != NULL; iter = iter->next) { pe_resource_t *child = iter->data; pcmk__new_ordering(child, pcmk__op_key(child->id, action_first, 0), NULL, NULL, NULL, clone_min_met, pe_order_one_or_more|pe_order_implies_then_printed, rsc_first->cluster); } // Order "then" action after the pseudo-action (if runnable) pcmk__new_ordering(NULL, NULL, clone_min_met, rsc_then, pcmk__op_key(rsc_then->id, action_then, 0), - NULL, flags|pe_order_runnable_left, rsc_first->cluster); + NULL, flags|pcmk__ar_unrunnable_first_blocks, + rsc_first->cluster); } /*! * \internal * \brief Update ordering flags for restart-type=restart * * \param[in] rsc 'Then' resource in ordering * \param[in] kind Ordering kind * \param[in] flag Ordering flag to set (when applicable) * \param[in,out] flags Ordering flag set to update * * \compat The restart-type resource meta-attribute is deprecated. Eventually, * it will be removed, and pe_restart_ignore will be the only behavior, * at which time this can just be removed entirely. */ #define handle_restart_type(rsc, kind, flag, flags) do { \ if (((kind) == pe_order_kind_optional) \ && ((rsc)->restart_type == pe_restart_restart)) { \ pe__set_order_flags((flags), (flag)); \ } \ } while (0) /*! * \internal * \brief Create new ordering for inverse of symmetric constraint * * \param[in] id Ordering ID (for logging only) * \param[in] kind Ordering kind * \param[in] rsc_first 'First' resource in ordering (a clone) * \param[in] action_first 'First' action in ordering * \param[in,out] rsc_then 'Then' resource in ordering * \param[in] action_then 'Then' action in ordering */ static void inverse_ordering(const char *id, enum pe_order_kind kind, pe_resource_t *rsc_first, const char *action_first, pe_resource_t *rsc_then, const char *action_then) { action_then = invert_action(action_then); action_first = invert_action(action_first); if ((action_then == NULL) || (action_first == NULL)) { pcmk__config_warn("Cannot invert constraint '%s' " "(please specify inverse manually)", id); } else { uint32_t flags = ordering_flags_for_kind(kind, action_first, ordering_symmetric_inverse); handle_restart_type(rsc_then, kind, pcmk__ar_then_implies_first, flags); pcmk__order_resource_actions(rsc_then, action_then, rsc_first, action_first, flags); } } static void unpack_simple_rsc_order(xmlNode *xml_obj, pe_working_set_t *data_set) { pe_resource_t *rsc_then = NULL; pe_resource_t *rsc_first = NULL; int min_required_before = 0; enum pe_order_kind kind = pe_order_kind_mandatory; uint32_t flags = pcmk__ar_none; enum ordering_symmetry symmetry; const char *action_then = NULL; const char *action_first = NULL; const char *id = NULL; CRM_CHECK(xml_obj != NULL, return); id = crm_element_value(xml_obj, XML_ATTR_ID); if (id == NULL) { pcmk__config_err("Ignoring <%s> constraint without " XML_ATTR_ID, xml_obj->name); return; } rsc_first = get_ordering_resource(xml_obj, XML_ORDER_ATTR_FIRST, XML_ORDER_ATTR_FIRST_INSTANCE, data_set); if (rsc_first == NULL) { return; } rsc_then = get_ordering_resource(xml_obj, XML_ORDER_ATTR_THEN, XML_ORDER_ATTR_THEN_INSTANCE, data_set); if (rsc_then == NULL) { return; } action_first = crm_element_value(xml_obj, XML_ORDER_ATTR_FIRST_ACTION); if (action_first == NULL) { action_first = PCMK_ACTION_START; } action_then = crm_element_value(xml_obj, XML_ORDER_ATTR_THEN_ACTION); if (action_then == NULL) { action_then = action_first; } kind = get_ordering_type(xml_obj); symmetry = get_ordering_symmetry(xml_obj, kind, NULL); flags = ordering_flags_for_kind(kind, action_first, symmetry); handle_restart_type(rsc_then, kind, pcmk__ar_first_implies_then, flags); /* If there is a minimum number of instances that must be runnable before * the 'then' action is runnable, we use a pseudo-action for convenience: * minimum number of clone instances have runnable actions -> * pseudo-action is runnable -> dependency is runnable. */ min_required_before = get_minimum_first_instances(rsc_first, xml_obj); if (min_required_before > 0) { clone_min_ordering(id, rsc_first, action_first, rsc_then, action_then, flags, min_required_before); } else { pcmk__order_resource_actions(rsc_first, action_first, rsc_then, action_then, flags); } if (symmetry == ordering_symmetric) { inverse_ordering(id, kind, rsc_first, action_first, rsc_then, action_then); } } /*! * \internal * \brief Create a new ordering between two actions * * \param[in,out] first_rsc Resource for 'first' action (if NULL and * \p first_action is a resource action, that * resource will be used) * \param[in,out] first_action_task Action key for 'first' action (if NULL and * \p first_action is not NULL, its UUID will * be used) * \param[in,out] first_action 'first' action (if NULL, \p first_rsc and * \p first_action_task must be set) * * \param[in] then_rsc Resource for 'then' action (if NULL and * \p then_action is a resource action, that * resource will be used) * \param[in,out] then_action_task Action key for 'then' action (if NULL and * \p then_action is not NULL, its UUID will * be used) * \param[in] then_action 'then' action (if NULL, \p then_rsc and * \p then_action_task must be set) * * \param[in] flags Group of enum pcmk__action_relation_flags * \param[in,out] sched Cluster working set to add ordering to * * \note This function takes ownership of first_action_task and * then_action_task, which do not need to be freed by the caller. */ void pcmk__new_ordering(pe_resource_t *first_rsc, char *first_action_task, pe_action_t *first_action, pe_resource_t *then_rsc, char *then_action_task, pe_action_t *then_action, uint32_t flags, pe_working_set_t *sched) { pe__ordering_t *order = NULL; // One of action or resource must be specified for each side CRM_CHECK(((first_action != NULL) || (first_rsc != NULL)) && ((then_action != NULL) || (then_rsc != NULL)), free(first_action_task); free(then_action_task); return); if ((first_rsc == NULL) && (first_action != NULL)) { first_rsc = first_action->rsc; } if ((then_rsc == NULL) && (then_action != NULL)) { then_rsc = then_action->rsc; } order = calloc(1, sizeof(pe__ordering_t)); CRM_ASSERT(order != NULL); order->id = sched->order_id++; order->flags = flags; order->lh_rsc = first_rsc; order->rh_rsc = then_rsc; order->lh_action = first_action; order->rh_action = then_action; order->lh_action_task = first_action_task; order->rh_action_task = then_action_task; if ((order->lh_action_task == NULL) && (first_action != NULL)) { order->lh_action_task = strdup(first_action->uuid); } if ((order->rh_action_task == NULL) && (then_action != NULL)) { order->rh_action_task = strdup(then_action->uuid); } if ((order->lh_rsc == NULL) && (first_action != NULL)) { order->lh_rsc = first_action->rsc; } if ((order->rh_rsc == NULL) && (then_action != NULL)) { order->rh_rsc = then_action->rsc; } pe_rsc_trace(first_rsc, "Created ordering %d for %s then %s", (sched->order_id - 1), pcmk__s(order->lh_action_task, "an underspecified action"), pcmk__s(order->rh_action_task, "an underspecified action")); sched->ordering_constraints = g_list_prepend(sched->ordering_constraints, order); pcmk__order_migration_equivalents(order); } /*! * \brief Unpack a set in an ordering constraint * * \param[in] set Set XML to unpack * \param[in] parent_kind rsc_order XML "kind" attribute * \param[in] parent_symmetrical_s rsc_order XML "symmetrical" attribute * \param[in,out] data_set Cluster working set * * \return Standard Pacemaker return code */ static int unpack_order_set(const xmlNode *set, enum pe_order_kind parent_kind, const char *parent_symmetrical_s, pe_working_set_t *data_set) { GList *set_iter = NULL; GList *resources = NULL; pe_resource_t *last = NULL; pe_resource_t *resource = NULL; int local_kind = parent_kind; bool sequential = false; uint32_t flags = pcmk__ar_ordered; enum ordering_symmetry symmetry; char *key = NULL; const char *id = ID(set); const char *action = crm_element_value(set, "action"); const char *sequential_s = crm_element_value(set, "sequential"); const char *kind_s = crm_element_value(set, XML_ORDER_ATTR_KIND); if (action == NULL) { action = PCMK_ACTION_START; } if (kind_s) { local_kind = get_ordering_type(set); } if (sequential_s == NULL) { sequential_s = "1"; } sequential = crm_is_true(sequential_s); symmetry = get_ordering_symmetry(set, parent_kind, parent_symmetrical_s); flags = ordering_flags_for_kind(local_kind, action, symmetry); for (const xmlNode *xml_rsc = first_named_child(set, XML_TAG_RESOURCE_REF); xml_rsc != NULL; xml_rsc = crm_next_same_xml(xml_rsc)) { EXPAND_CONSTRAINT_IDREF(id, resource, ID(xml_rsc)); resources = g_list_append(resources, resource); } if (pcmk__list_of_1(resources)) { crm_trace("Single set: %s", id); goto done; } set_iter = resources; while (set_iter != NULL) { resource = (pe_resource_t *) set_iter->data; set_iter = set_iter->next; key = pcmk__op_key(resource->id, action, 0); if (local_kind == pe_order_kind_serialize) { /* Serialize before everything that comes after */ for (GList *iter = set_iter; iter != NULL; iter = iter->next) { pe_resource_t *then_rsc = iter->data; char *then_key = pcmk__op_key(then_rsc->id, action, 0); pcmk__new_ordering(resource, strdup(key), NULL, then_rsc, then_key, NULL, flags, data_set); } } else if (sequential) { if (last != NULL) { pcmk__order_resource_actions(last, action, resource, action, flags); } last = resource; } free(key); } if (symmetry == ordering_asymmetric) { goto done; } last = NULL; action = invert_action(action); flags = ordering_flags_for_kind(local_kind, action, ordering_symmetric_inverse); set_iter = resources; while (set_iter != NULL) { resource = (pe_resource_t *) set_iter->data; set_iter = set_iter->next; if (sequential) { if (last != NULL) { pcmk__order_resource_actions(resource, action, last, action, flags); } last = resource; } } done: g_list_free(resources); return pcmk_rc_ok; } /*! * \brief Order two resource sets relative to each other * * \param[in] id Ordering ID (for logging) * \param[in] set1 First listed set * \param[in] set2 Second listed set * \param[in] kind Ordering kind * \param[in,out] data_set Cluster working set * \param[in] symmetry Which ordering symmetry applies to this relation * * \return Standard Pacemaker return code */ static int order_rsc_sets(const char *id, const xmlNode *set1, const xmlNode *set2, enum pe_order_kind kind, pe_working_set_t *data_set, enum ordering_symmetry symmetry) { const xmlNode *xml_rsc = NULL; const xmlNode *xml_rsc_2 = NULL; pe_resource_t *rsc_1 = NULL; pe_resource_t *rsc_2 = NULL; const char *action_1 = crm_element_value(set1, "action"); const char *action_2 = crm_element_value(set2, "action"); uint32_t flags = pcmk__ar_none; bool require_all = true; (void) pcmk__xe_get_bool_attr(set1, "require-all", &require_all); if (action_1 == NULL) { action_1 = PCMK_ACTION_START; } if (action_2 == NULL) { action_2 = PCMK_ACTION_START; } if (symmetry == ordering_symmetric_inverse) { action_1 = invert_action(action_1); action_2 = invert_action(action_2); } if (pcmk__str_eq(PCMK_ACTION_STOP, action_1, pcmk__str_none) || pcmk__str_eq(PCMK_ACTION_DEMOTE, action_1, pcmk__str_none)) { /* Assuming: A -> ( B || C) -> D * The one-or-more logic only applies during the start/promote phase. * During shutdown neither B nor can shutdown until D is down, so simply * turn require_all back on. */ require_all = true; } flags = ordering_flags_for_kind(kind, action_1, symmetry); /* If we have an unordered set1, whether it is sequential or not is * irrelevant in regards to set2. */ if (!require_all) { char *task = crm_strdup_printf(PCMK_ACTION_ONE_OR_MORE ":%s", ID(set1)); pe_action_t *unordered_action = get_pseudo_op(task, data_set); free(task); pe__set_action_flags(unordered_action, pcmk_action_min_runnable); for (xml_rsc = first_named_child(set1, XML_TAG_RESOURCE_REF); xml_rsc != NULL; xml_rsc = crm_next_same_xml(xml_rsc)) { EXPAND_CONSTRAINT_IDREF(id, rsc_1, ID(xml_rsc)); /* Add an ordering constraint between every element in set1 and the * pseudo action. If any action in set1 is runnable the pseudo * action will be runnable. */ pcmk__new_ordering(rsc_1, pcmk__op_key(rsc_1->id, action_1, 0), NULL, NULL, NULL, unordered_action, pe_order_one_or_more |pe_order_implies_then_printed, data_set); } for (xml_rsc_2 = first_named_child(set2, XML_TAG_RESOURCE_REF); xml_rsc_2 != NULL; xml_rsc_2 = crm_next_same_xml(xml_rsc_2)) { EXPAND_CONSTRAINT_IDREF(id, rsc_2, ID(xml_rsc_2)); /* Add an ordering constraint between the pseudo-action and every * element in set2. If the pseudo-action is runnable, every action * in set2 will be runnable. */ pcmk__new_ordering(NULL, NULL, unordered_action, rsc_2, pcmk__op_key(rsc_2->id, action_2, 0), - NULL, flags|pe_order_runnable_left, data_set); + NULL, flags|pcmk__ar_unrunnable_first_blocks, + data_set); } return pcmk_rc_ok; } if (pcmk__xe_attr_is_true(set1, "sequential")) { if (symmetry == ordering_symmetric_inverse) { // Get the first one xml_rsc = first_named_child(set1, XML_TAG_RESOURCE_REF); if (xml_rsc != NULL) { EXPAND_CONSTRAINT_IDREF(id, rsc_1, ID(xml_rsc)); } } else { // Get the last one const char *rid = NULL; for (xml_rsc = first_named_child(set1, XML_TAG_RESOURCE_REF); xml_rsc != NULL; xml_rsc = crm_next_same_xml(xml_rsc)) { rid = ID(xml_rsc); } EXPAND_CONSTRAINT_IDREF(id, rsc_1, rid); } } if (pcmk__xe_attr_is_true(set2, "sequential")) { if (symmetry == ordering_symmetric_inverse) { // Get the last one const char *rid = NULL; for (xml_rsc = first_named_child(set2, XML_TAG_RESOURCE_REF); xml_rsc != NULL; xml_rsc = crm_next_same_xml(xml_rsc)) { rid = ID(xml_rsc); } EXPAND_CONSTRAINT_IDREF(id, rsc_2, rid); } else { // Get the first one xml_rsc = first_named_child(set2, XML_TAG_RESOURCE_REF); if (xml_rsc != NULL) { EXPAND_CONSTRAINT_IDREF(id, rsc_2, ID(xml_rsc)); } } } if ((rsc_1 != NULL) && (rsc_2 != NULL)) { pcmk__order_resource_actions(rsc_1, action_1, rsc_2, action_2, flags); } else if (rsc_1 != NULL) { for (xml_rsc = first_named_child(set2, XML_TAG_RESOURCE_REF); xml_rsc != NULL; xml_rsc = crm_next_same_xml(xml_rsc)) { EXPAND_CONSTRAINT_IDREF(id, rsc_2, ID(xml_rsc)); pcmk__order_resource_actions(rsc_1, action_1, rsc_2, action_2, flags); } } else if (rsc_2 != NULL) { for (xml_rsc = first_named_child(set1, XML_TAG_RESOURCE_REF); xml_rsc != NULL; xml_rsc = crm_next_same_xml(xml_rsc)) { EXPAND_CONSTRAINT_IDREF(id, rsc_1, ID(xml_rsc)); pcmk__order_resource_actions(rsc_1, action_1, rsc_2, action_2, flags); } } else { for (xml_rsc = first_named_child(set1, XML_TAG_RESOURCE_REF); xml_rsc != NULL; xml_rsc = crm_next_same_xml(xml_rsc)) { EXPAND_CONSTRAINT_IDREF(id, rsc_1, ID(xml_rsc)); for (xmlNode *xml_rsc_2 = first_named_child(set2, XML_TAG_RESOURCE_REF); xml_rsc_2 != NULL; xml_rsc_2 = crm_next_same_xml(xml_rsc_2)) { EXPAND_CONSTRAINT_IDREF(id, rsc_2, ID(xml_rsc_2)); pcmk__order_resource_actions(rsc_1, action_1, rsc_2, action_2, flags); } } } return pcmk_rc_ok; } /*! * \internal * \brief If an ordering constraint uses resource tags, expand them * * \param[in,out] xml_obj Ordering constraint XML * \param[out] expanded_xml Equivalent XML with tags expanded * \param[in] data_set Cluster working set * * \return Standard Pacemaker return code (specifically, pcmk_rc_ok on success, * and pcmk_rc_unpack_error on invalid configuration) */ static int unpack_order_tags(xmlNode *xml_obj, xmlNode **expanded_xml, const pe_working_set_t *data_set) { const char *id_first = NULL; const char *id_then = NULL; const char *action_first = NULL; const char *action_then = NULL; pe_resource_t *rsc_first = NULL; pe_resource_t *rsc_then = NULL; pe_tag_t *tag_first = NULL; pe_tag_t *tag_then = NULL; xmlNode *rsc_set_first = NULL; xmlNode *rsc_set_then = NULL; bool any_sets = false; // Check whether there are any resource sets with template or tag references *expanded_xml = pcmk__expand_tags_in_sets(xml_obj, data_set); if (*expanded_xml != NULL) { crm_log_xml_trace(*expanded_xml, "Expanded rsc_order"); return pcmk_rc_ok; } id_first = crm_element_value(xml_obj, XML_ORDER_ATTR_FIRST); id_then = crm_element_value(xml_obj, XML_ORDER_ATTR_THEN); if ((id_first == NULL) || (id_then == NULL)) { return pcmk_rc_ok; } if (!pcmk__valid_resource_or_tag(data_set, id_first, &rsc_first, &tag_first)) { pcmk__config_err("Ignoring constraint '%s' because '%s' is not a " "valid resource or tag", ID(xml_obj), id_first); return pcmk_rc_unpack_error; } if (!pcmk__valid_resource_or_tag(data_set, id_then, &rsc_then, &tag_then)) { pcmk__config_err("Ignoring constraint '%s' because '%s' is not a " "valid resource or tag", ID(xml_obj), id_then); return pcmk_rc_unpack_error; } if ((rsc_first != NULL) && (rsc_then != NULL)) { // Neither side references a template or tag return pcmk_rc_ok; } action_first = crm_element_value(xml_obj, XML_ORDER_ATTR_FIRST_ACTION); action_then = crm_element_value(xml_obj, XML_ORDER_ATTR_THEN_ACTION); *expanded_xml = copy_xml(xml_obj); // Convert template/tag reference in "first" into constraint resource_set if (!pcmk__tag_to_set(*expanded_xml, &rsc_set_first, XML_ORDER_ATTR_FIRST, true, data_set)) { free_xml(*expanded_xml); *expanded_xml = NULL; return pcmk_rc_unpack_error; } if (rsc_set_first != NULL) { if (action_first != NULL) { // Move "first-action" into converted resource_set as "action" crm_xml_add(rsc_set_first, "action", action_first); xml_remove_prop(*expanded_xml, XML_ORDER_ATTR_FIRST_ACTION); } any_sets = true; } // Convert template/tag reference in "then" into constraint resource_set if (!pcmk__tag_to_set(*expanded_xml, &rsc_set_then, XML_ORDER_ATTR_THEN, true, data_set)) { free_xml(*expanded_xml); *expanded_xml = NULL; return pcmk_rc_unpack_error; } if (rsc_set_then != NULL) { if (action_then != NULL) { // Move "then-action" into converted resource_set as "action" crm_xml_add(rsc_set_then, "action", action_then); xml_remove_prop(*expanded_xml, XML_ORDER_ATTR_THEN_ACTION); } any_sets = true; } if (any_sets) { crm_log_xml_trace(*expanded_xml, "Expanded rsc_order"); } else { free_xml(*expanded_xml); *expanded_xml = NULL; } return pcmk_rc_ok; } /*! * \internal * \brief Unpack ordering constraint XML * * \param[in,out] xml_obj Ordering constraint XML to unpack * \param[in,out] data_set Cluster working set */ void pcmk__unpack_ordering(xmlNode *xml_obj, pe_working_set_t *data_set) { xmlNode *set = NULL; xmlNode *last = NULL; xmlNode *orig_xml = NULL; xmlNode *expanded_xml = NULL; const char *id = crm_element_value(xml_obj, XML_ATTR_ID); const char *invert = crm_element_value(xml_obj, XML_CONS_ATTR_SYMMETRICAL); enum pe_order_kind kind = get_ordering_type(xml_obj); enum ordering_symmetry symmetry = get_ordering_symmetry(xml_obj, kind, NULL); // Expand any resource tags in the constraint XML if (unpack_order_tags(xml_obj, &expanded_xml, data_set) != pcmk_rc_ok) { return; } if (expanded_xml != NULL) { orig_xml = xml_obj; xml_obj = expanded_xml; } // If the constraint has resource sets, unpack them for (set = first_named_child(xml_obj, XML_CONS_TAG_RSC_SET); set != NULL; set = crm_next_same_xml(set)) { set = expand_idref(set, data_set->input); if ((set == NULL) // Configuration error, message already logged || (unpack_order_set(set, kind, invert, data_set) != pcmk_rc_ok)) { if (expanded_xml != NULL) { free_xml(expanded_xml); } return; } if (last != NULL) { if (order_rsc_sets(id, last, set, kind, data_set, symmetry) != pcmk_rc_ok) { if (expanded_xml != NULL) { free_xml(expanded_xml); } return; } if ((symmetry == ordering_symmetric) && (order_rsc_sets(id, set, last, kind, data_set, ordering_symmetric_inverse) != pcmk_rc_ok)) { if (expanded_xml != NULL) { free_xml(expanded_xml); } return; } } last = set; } if (expanded_xml) { free_xml(expanded_xml); xml_obj = orig_xml; } // If the constraint has no resource sets, unpack it as a simple ordering if (last == NULL) { return unpack_simple_rsc_order(xml_obj, data_set); } } static bool ordering_is_invalid(pe_action_t *action, pe_action_wrapper_t *input) { /* Prevent user-defined ordering constraints between resources * running in a guest node and the resource that defines that node. */ if (!pcmk_is_set(input->type, pe_order_preserve) && (input->action->rsc != NULL) && pcmk__rsc_corresponds_to_guest(action->rsc, input->action->node)) { crm_warn("Invalid ordering constraint between %s and %s", input->action->rsc->id, action->rsc->id); return true; } /* If there's an order like * "rscB_stop node2"-> "load_stopped_node2" -> "rscA_migrate_to node1" * * then rscA is being migrated from node1 to node2, while rscB is being * migrated from node2 to node1. If there would be a graph loop, * break the order "load_stopped_node2" -> "rscA_migrate_to node1". */ if ((input->type == pe_order_load) && action->rsc && pcmk__str_eq(action->task, PCMK_ACTION_MIGRATE_TO, pcmk__str_none) && pcmk__graph_has_loop(action, action, input)) { return true; } return false; } void pcmk__disable_invalid_orderings(pe_working_set_t *data_set) { for (GList *iter = data_set->actions; iter != NULL; iter = iter->next) { pe_action_t *action = (pe_action_t *) iter->data; pe_action_wrapper_t *input = NULL; for (GList *input_iter = action->actions_before; input_iter != NULL; input_iter = input_iter->next) { input = (pe_action_wrapper_t *) input_iter->data; if (ordering_is_invalid(action, input)) { input->type = pcmk__ar_none; } } } } /*! * \internal * \brief Order stops on a node before the node's shutdown * * \param[in,out] node Node being shut down * \param[in] shutdown_op Shutdown action for node */ void pcmk__order_stops_before_shutdown(pe_node_t *node, pe_action_t *shutdown_op) { for (GList *iter = node->details->data_set->actions; iter != NULL; iter = iter->next) { pe_action_t *action = (pe_action_t *) iter->data; // Only stops on the node shutting down are relevant if (!pe__same_node(action->node, node) || !pcmk__str_eq(action->task, PCMK_ACTION_STOP, pcmk__str_none)) { continue; } // Resources and nodes in maintenance mode won't be touched if (pcmk_is_set(action->rsc->flags, pcmk_rsc_maintenance)) { pe_rsc_trace(action->rsc, "Not ordering %s before shutdown of %s because " "resource in maintenance mode", action->uuid, pe__node_name(node)); continue; } else if (node->details->maintenance) { pe_rsc_trace(action->rsc, "Not ordering %s before shutdown of %s because " "node in maintenance mode", action->uuid, pe__node_name(node)); continue; } /* Don't touch a resource that is unmanaged or blocked, to avoid * blocking the shutdown (though if another action depends on this one, * we may still end up blocking) */ if (!pcmk_any_flags_set(action->rsc->flags, pcmk_rsc_managed|pcmk_rsc_blocked)) { pe_rsc_trace(action->rsc, "Not ordering %s before shutdown of %s because " "resource is unmanaged or blocked", action->uuid, pe__node_name(node)); continue; } pe_rsc_trace(action->rsc, "Ordering %s before shutdown of %s", action->uuid, pe__node_name(node)); pe__clear_action_flags(action, pcmk_action_optional); pcmk__new_ordering(action->rsc, NULL, action, NULL, strdup(PCMK_ACTION_DO_SHUTDOWN), shutdown_op, - pcmk__ar_ordered|pe_order_runnable_left, + pcmk__ar_ordered|pcmk__ar_unrunnable_first_blocks, node->details->data_set); } } /*! * \brief Find resource actions matching directly or as child * * \param[in] rsc Resource to check * \param[in] original_key Action key to search for (possibly referencing * parent of \rsc) * * \return Newly allocated list of matching actions * \note It is the caller's responsibility to free the result with g_list_free() */ static GList * find_actions_by_task(const pe_resource_t *rsc, const char *original_key) { // Search under given task key directly GList *list = find_actions(rsc->actions, original_key, NULL); if (list == NULL) { // Search again using this resource's ID char *key = NULL; char *task = NULL; guint interval_ms = 0; if (parse_op_key(original_key, NULL, &task, &interval_ms)) { key = pcmk__op_key(rsc->id, task, interval_ms); list = find_actions(rsc->actions, key, NULL); free(key); free(task); } else { crm_err("Invalid operation key (bug?): %s", original_key); } } return list; } /*! * \internal * \brief Order relevant resource actions after a given action * * \param[in,out] first_action Action to order after (or NULL if none runnable) * \param[in] rsc Resource whose actions should be ordered * \param[in,out] order Ordering constraint being applied */ static void order_resource_actions_after(pe_action_t *first_action, const pe_resource_t *rsc, pe__ordering_t *order) { GList *then_actions = NULL; uint32_t flags = pcmk__ar_none; CRM_CHECK((rsc != NULL) && (order != NULL), return); flags = order->flags; pe_rsc_trace(rsc, "Applying ordering %d for 'then' resource %s", order->id, rsc->id); if (order->rh_action != NULL) { then_actions = g_list_prepend(NULL, order->rh_action); } else { then_actions = find_actions_by_task(rsc, order->rh_action_task); } if (then_actions == NULL) { pe_rsc_trace(rsc, "Ignoring ordering %d: no %s actions found for %s", order->id, order->rh_action_task, rsc->id); return; } if ((first_action != NULL) && (first_action->rsc == rsc) && pcmk_is_set(first_action->flags, pcmk_action_migration_abort)) { pe_rsc_trace(rsc, "Detected dangling migration ordering (%s then %s %s)", first_action->uuid, order->rh_action_task, rsc->id); pe__clear_order_flags(flags, pcmk__ar_first_implies_then); } if ((first_action == NULL) && !pcmk_is_set(flags, pcmk__ar_first_implies_then)) { pe_rsc_debug(rsc, "Ignoring ordering %d for %s: No first action found", order->id, rsc->id); g_list_free(then_actions); return; } for (GList *iter = then_actions; iter != NULL; iter = iter->next) { pe_action_t *then_action_iter = (pe_action_t *) iter->data; if (first_action != NULL) { order_actions(first_action, then_action_iter, flags); } else { pe__clear_action_flags(then_action_iter, pcmk_action_runnable); crm_warn("%s of %s is unrunnable because there is no %s of %s " "to order it after", then_action_iter->task, rsc->id, order->lh_action_task, order->lh_rsc->id); } } g_list_free(then_actions); } static void rsc_order_first(pe_resource_t *first_rsc, pe__ordering_t *order) { GList *first_actions = NULL; pe_action_t *first_action = order->lh_action; pe_resource_t *then_rsc = order->rh_rsc; CRM_ASSERT(first_rsc != NULL); pe_rsc_trace(first_rsc, "Applying ordering constraint %d (first: %s)", order->id, first_rsc->id); if (first_action != NULL) { first_actions = g_list_prepend(NULL, first_action); } else { first_actions = find_actions_by_task(first_rsc, order->lh_action_task); } if ((first_actions == NULL) && (first_rsc == then_rsc)) { pe_rsc_trace(first_rsc, "Ignoring constraint %d: first (%s for %s) not found", order->id, order->lh_action_task, first_rsc->id); } else if (first_actions == NULL) { char *key = NULL; char *op_type = NULL; guint interval_ms = 0; parse_op_key(order->lh_action_task, NULL, &op_type, &interval_ms); key = pcmk__op_key(first_rsc->id, op_type, interval_ms); if ((first_rsc->fns->state(first_rsc, TRUE) == pcmk_role_stopped) && pcmk__str_eq(op_type, PCMK_ACTION_STOP, pcmk__str_none)) { free(key); pe_rsc_trace(first_rsc, "Ignoring constraint %d: first (%s for %s) not found", order->id, order->lh_action_task, first_rsc->id); } else if ((first_rsc->fns->state(first_rsc, TRUE) == pcmk_role_unpromoted) && pcmk__str_eq(op_type, PCMK_ACTION_DEMOTE, pcmk__str_none)) { free(key); pe_rsc_trace(first_rsc, "Ignoring constraint %d: first (%s for %s) not found", order->id, order->lh_action_task, first_rsc->id); } else { pe_rsc_trace(first_rsc, "Creating first (%s for %s) for constraint %d ", order->lh_action_task, first_rsc->id, order->id); first_action = custom_action(first_rsc, key, op_type, NULL, TRUE, TRUE, first_rsc->cluster); first_actions = g_list_prepend(NULL, first_action); } free(op_type); } if (then_rsc == NULL) { if (order->rh_action == NULL) { pe_rsc_trace(first_rsc, "Ignoring constraint %d: then not found", order->id); return; } then_rsc = order->rh_action->rsc; } for (GList *iter = first_actions; iter != NULL; iter = iter->next) { first_action = iter->data; if (then_rsc == NULL) { order_actions(first_action, order->rh_action, order->flags); } else { order_resource_actions_after(first_action, then_rsc, order); } } g_list_free(first_actions); } // GFunc to call pcmk__block_colocation_dependents() static void block_colocation_dependents(gpointer data, gpointer user_data) { pcmk__block_colocation_dependents(data); } // GFunc to call pcmk__update_action_for_orderings() static void update_action_for_orderings(gpointer data, gpointer user_data) { pcmk__update_action_for_orderings((pe_action_t *) data, (pe_working_set_t *) user_data); } /*! * \internal * \brief Apply all ordering constraints * * \param[in,out] sched Cluster working set */ void pcmk__apply_orderings(pe_working_set_t *sched) { crm_trace("Applying ordering constraints"); /* Ordering constraints need to be processed in the order they were created. * rsc_order_first() and order_resource_actions_after() require the relevant * actions to already exist in some cases, but rsc_order_first() will create * the 'first' action in certain cases. Thus calling rsc_order_first() can * change the behavior of later-created orderings. * * Also, g_list_append() should be avoided for performance reasons, so we * prepend orderings when creating them and reverse the list here. * * @TODO This is brittle and should be carefully redesigned so that the * order of creation doesn't matter, and the reverse becomes unneeded. */ sched->ordering_constraints = g_list_reverse(sched->ordering_constraints); for (GList *iter = sched->ordering_constraints; iter != NULL; iter = iter->next) { pe__ordering_t *order = iter->data; pe_resource_t *rsc = order->lh_rsc; if (rsc != NULL) { rsc_order_first(rsc, order); continue; } rsc = order->rh_rsc; if (rsc != NULL) { order_resource_actions_after(order->lh_action, rsc, order); } else { crm_trace("Applying ordering constraint %d (non-resource actions)", order->id); order_actions(order->lh_action, order->rh_action, order->flags); } } g_list_foreach(sched->actions, block_colocation_dependents, NULL); crm_trace("Ordering probes"); pcmk__order_probes(sched); crm_trace("Updating %d actions", g_list_length(sched->actions)); g_list_foreach(sched->actions, update_action_for_orderings, sched); pcmk__disable_invalid_orderings(sched); } /*! * \internal * \brief Order a given action after each action in a given list * * \param[in,out] after "After" action * \param[in,out] list List of "before" actions */ void pcmk__order_after_each(pe_action_t *after, GList *list) { const char *after_desc = (after->task == NULL)? after->uuid : after->task; for (GList *iter = list; iter != NULL; iter = iter->next) { pe_action_t *before = (pe_action_t *) iter->data; const char *before_desc = before->task? before->task : before->uuid; crm_debug("Ordering %s on %s before %s on %s", before_desc, pe__node_name(before->node), after_desc, pe__node_name(after->node)); order_actions(before, after, pcmk__ar_ordered); } } /*! * \internal * \brief Order promotions and demotions for restarts of a clone or bundle * * \param[in,out] rsc Clone or bundle to order */ void pcmk__promotable_restart_ordering(pe_resource_t *rsc) { // Order start and promote after all instances are stopped pcmk__order_resource_actions(rsc, PCMK_ACTION_STOPPED, rsc, PCMK_ACTION_START, pcmk__ar_ordered); pcmk__order_resource_actions(rsc, PCMK_ACTION_STOPPED, rsc, PCMK_ACTION_PROMOTE, pcmk__ar_ordered); // Order stop, start, and promote after all instances are demoted pcmk__order_resource_actions(rsc, PCMK_ACTION_DEMOTED, rsc, PCMK_ACTION_STOP, pcmk__ar_ordered); pcmk__order_resource_actions(rsc, PCMK_ACTION_DEMOTED, rsc, PCMK_ACTION_START, pcmk__ar_ordered); pcmk__order_resource_actions(rsc, PCMK_ACTION_DEMOTED, rsc, PCMK_ACTION_PROMOTE, pcmk__ar_ordered); // Order promote after all instances are started pcmk__order_resource_actions(rsc, PCMK_ACTION_RUNNING, rsc, PCMK_ACTION_PROMOTE, pcmk__ar_ordered); // Order demote after all instances are demoted pcmk__order_resource_actions(rsc, PCMK_ACTION_DEMOTE, rsc, PCMK_ACTION_DEMOTED, pcmk__ar_ordered); } diff --git a/lib/pacemaker/pcmk_sched_primitive.c b/lib/pacemaker/pcmk_sched_primitive.c index 8172bef5f3..7dfbaae8db 100644 --- a/lib/pacemaker/pcmk_sched_primitive.c +++ b/lib/pacemaker/pcmk_sched_primitive.c @@ -1,1655 +1,1655 @@ /* * Copyright 2004-2023 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU General Public License version 2 * or later (GPLv2+) WITHOUT ANY WARRANTY. */ #include #include #include // uint8_t, uint32_t #include #include #include "libpacemaker_private.h" static void stop_resource(pe_resource_t *rsc, pe_node_t *node, bool optional); static void start_resource(pe_resource_t *rsc, pe_node_t *node, bool optional); static void demote_resource(pe_resource_t *rsc, pe_node_t *node, bool optional); static void promote_resource(pe_resource_t *rsc, pe_node_t *node, bool optional); static void assert_role_error(pe_resource_t *rsc, pe_node_t *node, bool optional); #define RSC_ROLE_MAX (pcmk_role_promoted + 1) static enum rsc_role_e rsc_state_matrix[RSC_ROLE_MAX][RSC_ROLE_MAX] = { /* This array lists the immediate next role when transitioning from one role * to a target role. For example, when going from Stopped to Promoted, the * next role is Unpromoted, because the resource must be started before it * can be promoted. The current state then becomes Started, which is fed * into this array again, giving a next role of Promoted. * * Current role Immediate next role Final target role * ------------ ------------------- ----------------- */ /* Unknown */ { pcmk_role_unknown, /* Unknown */ pcmk_role_stopped, /* Stopped */ pcmk_role_stopped, /* Started */ pcmk_role_stopped, /* Unpromoted */ pcmk_role_stopped, /* Promoted */ }, /* Stopped */ { pcmk_role_stopped, /* Unknown */ pcmk_role_stopped, /* Stopped */ pcmk_role_started, /* Started */ pcmk_role_unpromoted, /* Unpromoted */ pcmk_role_unpromoted, /* Promoted */ }, /* Started */ { pcmk_role_stopped, /* Unknown */ pcmk_role_stopped, /* Stopped */ pcmk_role_started, /* Started */ pcmk_role_unpromoted, /* Unpromoted */ pcmk_role_promoted, /* Promoted */ }, /* Unpromoted */ { pcmk_role_stopped, /* Unknown */ pcmk_role_stopped, /* Stopped */ pcmk_role_stopped, /* Started */ pcmk_role_unpromoted, /* Unpromoted */ pcmk_role_promoted, /* Promoted */ }, /* Promoted */ { pcmk_role_stopped, /* Unknown */ pcmk_role_unpromoted, /* Stopped */ pcmk_role_unpromoted, /* Started */ pcmk_role_unpromoted, /* Unpromoted */ pcmk_role_promoted, /* Promoted */ }, }; /*! * \internal * \brief Function to schedule actions needed for a role change * * \param[in,out] rsc Resource whose role is changing * \param[in,out] node Node where resource will be in its next role * \param[in] optional Whether scheduled actions should be optional */ typedef void (*rsc_transition_fn)(pe_resource_t *rsc, pe_node_t *node, bool optional); static rsc_transition_fn rsc_action_matrix[RSC_ROLE_MAX][RSC_ROLE_MAX] = { /* This array lists the function needed to transition directly from one role * to another. NULL indicates that nothing is needed. * * Current role Transition function Next role * ------------ ------------------- ---------- */ /* Unknown */ { assert_role_error, /* Unknown */ stop_resource, /* Stopped */ assert_role_error, /* Started */ assert_role_error, /* Unpromoted */ assert_role_error, /* Promoted */ }, /* Stopped */ { assert_role_error, /* Unknown */ NULL, /* Stopped */ start_resource, /* Started */ start_resource, /* Unpromoted */ assert_role_error, /* Promoted */ }, /* Started */ { assert_role_error, /* Unknown */ stop_resource, /* Stopped */ NULL, /* Started */ NULL, /* Unpromoted */ promote_resource, /* Promoted */ }, /* Unpromoted */ { assert_role_error, /* Unknown */ stop_resource, /* Stopped */ stop_resource, /* Started */ NULL, /* Unpromoted */ promote_resource, /* Promoted */ }, /* Promoted */ { assert_role_error, /* Unknown */ demote_resource, /* Stopped */ demote_resource, /* Started */ demote_resource, /* Unpromoted */ NULL, /* Promoted */ }, }; /*! * \internal * \brief Get a list of a resource's allowed nodes sorted by node score * * \param[in] rsc Resource to check * * \return List of allowed nodes sorted by node score */ static GList * sorted_allowed_nodes(const pe_resource_t *rsc) { if (rsc->allowed_nodes != NULL) { GList *nodes = g_hash_table_get_values(rsc->allowed_nodes); if (nodes != NULL) { return pcmk__sort_nodes(nodes, pe__current_node(rsc)); } } return NULL; } /*! * \internal * \brief Assign a resource to its best allowed node, if possible * * \param[in,out] rsc Resource to choose a node for * \param[in] prefer If not \c NULL, prefer this node when all else * equal * \param[in] stop_if_fail If \c true and \p rsc can't be assigned to a * node, set next role to stopped and update * existing actions * * \return true if \p rsc could be assigned to a node, otherwise false * * \note If \p stop_if_fail is \c false, then \c pcmk__unassign_resource() can * completely undo the assignment. A successful assignment can be either * undone or left alone as final. A failed assignment has the same effect * as calling pcmk__unassign_resource(); there are no side effects on * roles or actions. */ static bool assign_best_node(pe_resource_t *rsc, const pe_node_t *prefer, bool stop_if_fail) { GList *nodes = NULL; pe_node_t *chosen = NULL; pe_node_t *best = NULL; const pe_node_t *most_free_node = pcmk__ban_insufficient_capacity(rsc); if (prefer == NULL) { prefer = most_free_node; } if (!pcmk_is_set(rsc->flags, pcmk_rsc_unassigned)) { // We've already finished assignment of resources to nodes return rsc->allocated_to != NULL; } // Sort allowed nodes by score nodes = sorted_allowed_nodes(rsc); if (nodes != NULL) { best = (pe_node_t *) nodes->data; // First node has best score } if ((prefer != NULL) && (nodes != NULL)) { // Get the allowed node version of prefer chosen = g_hash_table_lookup(rsc->allowed_nodes, prefer->details->id); if (chosen == NULL) { pe_rsc_trace(rsc, "Preferred node %s for %s was unknown", pe__node_name(prefer), rsc->id); /* Favor the preferred node as long as its score is at least as good as * the best allowed node's. * * An alternative would be to favor the preferred node even if the best * node is better, when the best node's score is less than INFINITY. */ } else if (chosen->weight < best->weight) { pe_rsc_trace(rsc, "Preferred node %s for %s was unsuitable", pe__node_name(chosen), rsc->id); chosen = NULL; } else if (!pcmk__node_available(chosen, true, false)) { pe_rsc_trace(rsc, "Preferred node %s for %s was unavailable", pe__node_name(chosen), rsc->id); chosen = NULL; } else { pe_rsc_trace(rsc, "Chose preferred node %s for %s " "(ignoring %d candidates)", pe__node_name(chosen), rsc->id, g_list_length(nodes)); } } if ((chosen == NULL) && (best != NULL)) { /* Either there is no preferred node, or the preferred node is not * suitable, but another node is allowed to run the resource. */ chosen = best; if (!pe_rsc_is_unique_clone(rsc->parent) && (chosen->weight > 0) // Zero not acceptable && pcmk__node_available(chosen, false, false)) { /* If the resource is already running on a node, prefer that node if * it is just as good as the chosen node. * * We don't do this for unique clone instances, because * pcmk__assign_instances() has already assigned instances to their * running nodes when appropriate, and if we get here, we don't want * remaining unassigned instances to prefer a node that's already * running another instance. */ pe_node_t *running = pe__current_node(rsc); if (running == NULL) { // Nothing to do } else if (!pcmk__node_available(running, true, false)) { pe_rsc_trace(rsc, "Current node for %s (%s) can't run resources", rsc->id, pe__node_name(running)); } else { int nodes_with_best_score = 1; for (GList *iter = nodes->next; iter; iter = iter->next) { pe_node_t *allowed = (pe_node_t *) iter->data; if (allowed->weight != chosen->weight) { // The nodes are sorted by score, so no more are equal break; } if (pe__same_node(allowed, running)) { // Scores are equal, so prefer the current node chosen = allowed; } nodes_with_best_score++; } if (nodes_with_best_score > 1) { uint8_t log_level = LOG_INFO; if (chosen->weight >= INFINITY) { log_level = LOG_WARNING; } do_crm_log(log_level, "Chose %s for %s from %d nodes with score %s", pe__node_name(chosen), rsc->id, nodes_with_best_score, pcmk_readable_score(chosen->weight)); } } } pe_rsc_trace(rsc, "Chose %s for %s from %d candidates", pe__node_name(chosen), rsc->id, g_list_length(nodes)); } pcmk__assign_resource(rsc, chosen, false, stop_if_fail); g_list_free(nodes); return rsc->allocated_to != NULL; } /*! * \internal * \brief Apply a "this with" colocation to a node's allowed node scores * * \param[in,out] colocation Colocation to apply * \param[in,out] rsc Resource being assigned */ static void apply_this_with(pcmk__colocation_t *colocation, pe_resource_t *rsc) { GHashTable *archive = NULL; pe_resource_t *other = colocation->primary; // In certain cases, we will need to revert the node scores if ((colocation->dependent_role >= pcmk_role_promoted) || ((colocation->score < 0) && (colocation->score > -INFINITY))) { archive = pcmk__copy_node_table(rsc->allowed_nodes); } if (pcmk_is_set(other->flags, pcmk_rsc_unassigned)) { pe_rsc_trace(rsc, "%s: Assigning colocation %s primary %s first" "(score=%d role=%s)", rsc->id, colocation->id, other->id, colocation->score, role2text(colocation->dependent_role)); other->cmds->assign(other, NULL, true); } // Apply the colocation score to this resource's allowed node scores rsc->cmds->apply_coloc_score(rsc, other, colocation, true); if ((archive != NULL) && !pcmk__any_node_available(rsc->allowed_nodes)) { pe_rsc_info(rsc, "%s: Reverting scores from colocation with %s " "because no nodes allowed", rsc->id, other->id); g_hash_table_destroy(rsc->allowed_nodes); rsc->allowed_nodes = archive; archive = NULL; } if (archive != NULL) { g_hash_table_destroy(archive); } } /*! * \internal * \brief Update a Pacemaker Remote node once its connection has been assigned * * \param[in] connection Connection resource that has been assigned */ static void remote_connection_assigned(const pe_resource_t *connection) { pe_node_t *remote_node = pe_find_node(connection->cluster->nodes, connection->id); CRM_CHECK(remote_node != NULL, return); if ((connection->allocated_to != NULL) && (connection->next_role != pcmk_role_stopped)) { crm_trace("Pacemaker Remote node %s will be online", remote_node->details->id); remote_node->details->online = TRUE; if (remote_node->details->unseen) { // Avoid unnecessary fence, since we will attempt connection remote_node->details->unclean = FALSE; } } else { crm_trace("Pacemaker Remote node %s will be shut down " "(%sassigned connection's next role is %s)", remote_node->details->id, ((connection->allocated_to == NULL)? "un" : ""), role2text(connection->next_role)); remote_node->details->shutdown = TRUE; } } /*! * \internal * \brief Assign a primitive resource to a node * * \param[in,out] rsc Resource to assign to a node * \param[in] prefer Node to prefer, if all else is equal * \param[in] stop_if_fail If \c true and \p rsc can't be assigned to a * node, set next role to stopped and update * existing actions * * \return Node that \p rsc is assigned to, if assigned entirely to one node * * \note If \p stop_if_fail is \c false, then \c pcmk__unassign_resource() can * completely undo the assignment. A successful assignment can be either * undone or left alone as final. A failed assignment has the same effect * as calling pcmk__unassign_resource(); there are no side effects on * roles or actions. */ pe_node_t * pcmk__primitive_assign(pe_resource_t *rsc, const pe_node_t *prefer, bool stop_if_fail) { GList *this_with_colocations = NULL; GList *with_this_colocations = NULL; GList *iter = NULL; pcmk__colocation_t *colocation = NULL; CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_primitive)); // Never assign a child without parent being assigned first if ((rsc->parent != NULL) && !pcmk_is_set(rsc->parent->flags, pcmk_rsc_assigning)) { pe_rsc_debug(rsc, "%s: Assigning parent %s first", rsc->id, rsc->parent->id); rsc->parent->cmds->assign(rsc->parent, prefer, stop_if_fail); } if (!pcmk_is_set(rsc->flags, pcmk_rsc_unassigned)) { // Assignment has already been done const char *node_name = "no node"; if (rsc->allocated_to != NULL) { node_name = pe__node_name(rsc->allocated_to); } pe_rsc_debug(rsc, "%s: pre-assigned to %s", rsc->id, node_name); return rsc->allocated_to; } // Ensure we detect assignment loops if (pcmk_is_set(rsc->flags, pcmk_rsc_assigning)) { pe_rsc_debug(rsc, "Breaking assignment loop involving %s", rsc->id); return NULL; } pe__set_resource_flags(rsc, pcmk_rsc_assigning); pe__show_node_scores(true, rsc, "Pre-assignment", rsc->allowed_nodes, rsc->cluster); this_with_colocations = pcmk__this_with_colocations(rsc); with_this_colocations = pcmk__with_this_colocations(rsc); // Apply mandatory colocations first, to satisfy as many as possible for (iter = this_with_colocations; iter != NULL; iter = iter->next) { colocation = iter->data; if ((colocation->score <= -CRM_SCORE_INFINITY) || (colocation->score >= CRM_SCORE_INFINITY)) { apply_this_with(colocation, rsc); } } for (iter = with_this_colocations; iter != NULL; iter = iter->next) { colocation = iter->data; if ((colocation->score <= -CRM_SCORE_INFINITY) || (colocation->score >= CRM_SCORE_INFINITY)) { pcmk__add_dependent_scores(colocation, rsc); } } pe__show_node_scores(true, rsc, "Mandatory-colocations", rsc->allowed_nodes, rsc->cluster); // Then apply optional colocations for (iter = this_with_colocations; iter != NULL; iter = iter->next) { colocation = iter->data; if ((colocation->score > -CRM_SCORE_INFINITY) && (colocation->score < CRM_SCORE_INFINITY)) { apply_this_with(colocation, rsc); } } for (iter = with_this_colocations; iter != NULL; iter = iter->next) { colocation = iter->data; if ((colocation->score > -CRM_SCORE_INFINITY) && (colocation->score < CRM_SCORE_INFINITY)) { pcmk__add_dependent_scores(colocation, rsc); } } g_list_free(this_with_colocations); g_list_free(with_this_colocations); if (rsc->next_role == pcmk_role_stopped) { pe_rsc_trace(rsc, "Banning %s from all nodes because it will be stopped", rsc->id); resource_location(rsc, NULL, -INFINITY, XML_RSC_ATTR_TARGET_ROLE, rsc->cluster); } else if ((rsc->next_role > rsc->role) && !pcmk_is_set(rsc->cluster->flags, pcmk_sched_quorate) && (rsc->cluster->no_quorum_policy == pcmk_no_quorum_freeze)) { crm_notice("Resource %s cannot be elevated from %s to %s due to " "no-quorum-policy=freeze", rsc->id, role2text(rsc->role), role2text(rsc->next_role)); pe__set_next_role(rsc, rsc->role, "no-quorum-policy=freeze"); } pe__show_node_scores(!pcmk_is_set(rsc->cluster->flags, pcmk_sched_output_scores), rsc, __func__, rsc->allowed_nodes, rsc->cluster); // Unmanage resource if fencing is enabled but no device is configured if (pcmk_is_set(rsc->cluster->flags, pcmk_sched_fencing_enabled) && !pcmk_is_set(rsc->cluster->flags, pcmk_sched_have_fencing)) { pe__clear_resource_flags(rsc, pcmk_rsc_managed); } if (!pcmk_is_set(rsc->flags, pcmk_rsc_managed)) { // Unmanaged resources stay on their current node const char *reason = NULL; pe_node_t *assign_to = NULL; pe__set_next_role(rsc, rsc->role, "unmanaged"); assign_to = pe__current_node(rsc); if (assign_to == NULL) { reason = "inactive"; } else if (rsc->role == pcmk_role_promoted) { reason = "promoted"; } else if (pcmk_is_set(rsc->flags, pcmk_rsc_failed)) { reason = "failed"; } else { reason = "active"; } pe_rsc_info(rsc, "Unmanaged resource %s assigned to %s: %s", rsc->id, (assign_to? assign_to->details->uname : "no node"), reason); pcmk__assign_resource(rsc, assign_to, true, stop_if_fail); } else if (pcmk_is_set(rsc->cluster->flags, pcmk_sched_stop_all)) { // Must stop at some point, but be consistent with stop_if_fail if (stop_if_fail) { pe_rsc_debug(rsc, "Forcing %s to stop: stop-all-resources", rsc->id); } pcmk__assign_resource(rsc, NULL, true, stop_if_fail); } else if (!assign_best_node(rsc, prefer, stop_if_fail)) { // Assignment failed if (!pcmk_is_set(rsc->flags, pcmk_rsc_removed)) { pe_rsc_info(rsc, "Resource %s cannot run anywhere", rsc->id); } else if ((rsc->running_on != NULL) && stop_if_fail) { pe_rsc_info(rsc, "Stopping orphan resource %s", rsc->id); } } pe__clear_resource_flags(rsc, pcmk_rsc_assigning); if (rsc->is_remote_node) { remote_connection_assigned(rsc); } return rsc->allocated_to; } /*! * \internal * \brief Schedule actions to bring resource down and back to current role * * \param[in,out] rsc Resource to restart * \param[in,out] current Node that resource should be brought down on * \param[in] need_stop Whether the resource must be stopped * \param[in] need_promote Whether the resource must be promoted * * \return Role that resource would have after scheduled actions are taken */ static void schedule_restart_actions(pe_resource_t *rsc, pe_node_t *current, bool need_stop, bool need_promote) { enum rsc_role_e role = rsc->role; enum rsc_role_e next_role; rsc_transition_fn fn = NULL; pe__set_resource_flags(rsc, pcmk_rsc_restarting); // Bring resource down to a stop on its current node while (role != pcmk_role_stopped) { next_role = rsc_state_matrix[role][pcmk_role_stopped]; pe_rsc_trace(rsc, "Creating %s action to take %s down from %s to %s", (need_stop? "required" : "optional"), rsc->id, role2text(role), role2text(next_role)); fn = rsc_action_matrix[role][next_role]; if (fn == NULL) { break; } fn(rsc, current, !need_stop); role = next_role; } // Bring resource up to its next role on its next node while ((rsc->role <= rsc->next_role) && (role != rsc->role) && !pcmk_is_set(rsc->flags, pcmk_rsc_blocked)) { bool required = need_stop; next_role = rsc_state_matrix[role][rsc->role]; if ((next_role == pcmk_role_promoted) && need_promote) { required = true; } pe_rsc_trace(rsc, "Creating %s action to take %s up from %s to %s", (required? "required" : "optional"), rsc->id, role2text(role), role2text(next_role)); fn = rsc_action_matrix[role][next_role]; if (fn == NULL) { break; } fn(rsc, rsc->allocated_to, !required); role = next_role; } pe__clear_resource_flags(rsc, pcmk_rsc_restarting); } /*! * \internal * \brief If a resource's next role is not explicitly specified, set a default * * \param[in,out] rsc Resource to set next role for * * \return "explicit" if next role was explicitly set, otherwise "implicit" */ static const char * set_default_next_role(pe_resource_t *rsc) { if (rsc->next_role != pcmk_role_unknown) { return "explicit"; } if (rsc->allocated_to == NULL) { pe__set_next_role(rsc, pcmk_role_stopped, "assignment"); } else { pe__set_next_role(rsc, pcmk_role_started, "assignment"); } return "implicit"; } /*! * \internal * \brief Create an action to represent an already pending start * * \param[in,out] rsc Resource to create start action for */ static void create_pending_start(pe_resource_t *rsc) { pe_action_t *start = NULL; pe_rsc_trace(rsc, "Creating action for %s to represent already pending start", rsc->id); start = start_action(rsc, rsc->allocated_to, TRUE); pe__set_action_flags(start, pcmk_action_always_in_graph); } /*! * \internal * \brief Schedule actions needed to take a resource to its next role * * \param[in,out] rsc Resource to schedule actions for */ static void schedule_role_transition_actions(pe_resource_t *rsc) { enum rsc_role_e role = rsc->role; while (role != rsc->next_role) { enum rsc_role_e next_role = rsc_state_matrix[role][rsc->next_role]; rsc_transition_fn fn = NULL; pe_rsc_trace(rsc, "Creating action to take %s from %s to %s (ending at %s)", rsc->id, role2text(role), role2text(next_role), role2text(rsc->next_role)); fn = rsc_action_matrix[role][next_role]; if (fn == NULL) { break; } fn(rsc, rsc->allocated_to, false); role = next_role; } } /*! * \internal * \brief Create all actions needed for a given primitive resource * * \param[in,out] rsc Primitive resource to create actions for */ void pcmk__primitive_create_actions(pe_resource_t *rsc) { bool need_stop = false; bool need_promote = false; bool is_moving = false; bool allow_migrate = false; bool multiply_active = false; pe_node_t *current = NULL; unsigned int num_all_active = 0; unsigned int num_clean_active = 0; const char *next_role_source = NULL; CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_primitive)); next_role_source = set_default_next_role(rsc); pe_rsc_trace(rsc, "Creating all actions for %s transition from %s to %s " "(%s) on %s", rsc->id, role2text(rsc->role), role2text(rsc->next_role), next_role_source, pe__node_name(rsc->allocated_to)); current = rsc->fns->active_node(rsc, &num_all_active, &num_clean_active); g_list_foreach(rsc->dangling_migrations, pcmk__abort_dangling_migration, rsc); if ((current != NULL) && (rsc->allocated_to != NULL) && !pe__same_node(current, rsc->allocated_to) && (rsc->next_role >= pcmk_role_started)) { pe_rsc_trace(rsc, "Moving %s from %s to %s", rsc->id, pe__node_name(current), pe__node_name(rsc->allocated_to)); is_moving = true; allow_migrate = pcmk__rsc_can_migrate(rsc, current); // This is needed even if migrating (though I'm not sure why ...) need_stop = true; } // Check whether resource is partially migrated and/or multiply active if ((rsc->partial_migration_source != NULL) && (rsc->partial_migration_target != NULL) && allow_migrate && (num_all_active == 2) && pe__same_node(current, rsc->partial_migration_source) && pe__same_node(rsc->allocated_to, rsc->partial_migration_target)) { /* A partial migration is in progress, and the migration target remains * the same as when the migration began. */ pe_rsc_trace(rsc, "Partial migration of %s from %s to %s will continue", rsc->id, pe__node_name(rsc->partial_migration_source), pe__node_name(rsc->partial_migration_target)); } else if ((rsc->partial_migration_source != NULL) || (rsc->partial_migration_target != NULL)) { // A partial migration is in progress but can't be continued if (num_all_active > 2) { // The resource is migrating *and* multiply active! crm_notice("Forcing recovery of %s because it is migrating " "from %s to %s and possibly active elsewhere", rsc->id, pe__node_name(rsc->partial_migration_source), pe__node_name(rsc->partial_migration_target)); } else { // The migration source or target isn't available crm_notice("Forcing recovery of %s because it can no longer " "migrate from %s to %s", rsc->id, pe__node_name(rsc->partial_migration_source), pe__node_name(rsc->partial_migration_target)); } need_stop = true; rsc->partial_migration_source = rsc->partial_migration_target = NULL; allow_migrate = false; } else if (pcmk_is_set(rsc->flags, pcmk_rsc_needs_fencing)) { multiply_active = (num_all_active > 1); } else { /* If a resource has "requires" set to nothing or quorum, don't consider * it active on unclean nodes (similar to how all resources behave when * stonith-enabled is false). We can start such resources elsewhere * before fencing completes, and if we considered the resource active on * the failed node, we would attempt recovery for being active on * multiple nodes. */ multiply_active = (num_clean_active > 1); } if (multiply_active) { const char *class = crm_element_value(rsc->xml, XML_AGENT_ATTR_CLASS); // Resource was (possibly) incorrectly multiply active pe_proc_err("%s resource %s might be active on %u nodes (%s)", pcmk__s(class, "Untyped"), rsc->id, num_all_active, recovery2text(rsc->recovery_type)); crm_notice("See https://wiki.clusterlabs.org/wiki/FAQ" "#Resource_is_Too_Active for more information"); switch (rsc->recovery_type) { case pcmk_multiply_active_restart: need_stop = true; break; case pcmk_multiply_active_unexpected: need_stop = true; // stop_resource() will skip expected node pe__set_resource_flags(rsc, pcmk_rsc_stop_unexpected); break; default: break; } } else { pe__clear_resource_flags(rsc, pcmk_rsc_stop_unexpected); } if (pcmk_is_set(rsc->flags, pcmk_rsc_start_pending)) { create_pending_start(rsc); } if (is_moving) { // Remaining tests are only for resources staying where they are } else if (pcmk_is_set(rsc->flags, pcmk_rsc_failed)) { if (pcmk_is_set(rsc->flags, pcmk_rsc_stop_if_failed)) { need_stop = true; pe_rsc_trace(rsc, "Recovering %s", rsc->id); } else { pe_rsc_trace(rsc, "Recovering %s by demotion", rsc->id); if (rsc->next_role == pcmk_role_promoted) { need_promote = true; } } } else if (pcmk_is_set(rsc->flags, pcmk_rsc_blocked)) { pe_rsc_trace(rsc, "Blocking further actions on %s", rsc->id); need_stop = true; } else if ((rsc->role > pcmk_role_started) && (current != NULL) && (rsc->allocated_to != NULL)) { pe_action_t *start = NULL; pe_rsc_trace(rsc, "Creating start action for promoted resource %s", rsc->id); start = start_action(rsc, rsc->allocated_to, TRUE); if (!pcmk_is_set(start->flags, pcmk_action_optional)) { // Recovery of a promoted resource pe_rsc_trace(rsc, "%s restart is required for recovery", rsc->id); need_stop = true; } } // Create any actions needed to bring resource down and back up to same role schedule_restart_actions(rsc, current, need_stop, need_promote); // Create any actions needed to take resource from this role to the next schedule_role_transition_actions(rsc); pcmk__create_recurring_actions(rsc); if (allow_migrate) { pcmk__create_migration_actions(rsc, current); } } /*! * \internal * \brief Ban a resource from any allowed nodes that are Pacemaker Remote nodes * * \param[in] rsc Resource to check */ static void rsc_avoids_remote_nodes(const pe_resource_t *rsc) { GHashTableIter iter; pe_node_t *node = NULL; g_hash_table_iter_init(&iter, rsc->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (void **) &node)) { if (node->details->remote_rsc != NULL) { node->weight = -INFINITY; } } } /*! * \internal * \brief Return allowed nodes as (possibly sorted) list * * Convert a resource's hash table of allowed nodes to a list. If printing to * stdout, sort the list, to keep action ID numbers consistent for regression * test output (while avoiding the performance hit on a live cluster). * * \param[in] rsc Resource to check for allowed nodes * * \return List of resource's allowed nodes * \note Callers should take care not to rely on the list being sorted. */ static GList * allowed_nodes_as_list(const pe_resource_t *rsc) { GList *allowed_nodes = NULL; if (rsc->allowed_nodes) { allowed_nodes = g_hash_table_get_values(rsc->allowed_nodes); } if (!pcmk__is_daemon) { allowed_nodes = g_list_sort(allowed_nodes, pe__cmp_node_name); } return allowed_nodes; } /*! * \internal * \brief Create implicit constraints needed for a primitive resource * * \param[in,out] rsc Primitive resource to create implicit constraints for */ void pcmk__primitive_internal_constraints(pe_resource_t *rsc) { GList *allowed_nodes = NULL; bool check_unfencing = false; bool check_utilization = false; CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_primitive)); if (!pcmk_is_set(rsc->flags, pcmk_rsc_managed)) { pe_rsc_trace(rsc, "Skipping implicit constraints for unmanaged resource %s", rsc->id); return; } // Whether resource requires unfencing check_unfencing = !pcmk_is_set(rsc->flags, pcmk_rsc_fence_device) && pcmk_is_set(rsc->cluster->flags, pcmk_sched_enable_unfencing) && pcmk_is_set(rsc->flags, pcmk_rsc_needs_unfencing); // Whether a non-default placement strategy is used check_utilization = (g_hash_table_size(rsc->utilization) > 0) && !pcmk__str_eq(rsc->cluster->placement_strategy, "default", pcmk__str_casei); // Order stops before starts (i.e. restart) pcmk__new_ordering(rsc, pcmk__op_key(rsc->id, PCMK_ACTION_STOP, 0), NULL, rsc, pcmk__op_key(rsc->id, PCMK_ACTION_START, 0), NULL, pcmk__ar_ordered |pcmk__ar_first_implies_then |pe_order_restart, rsc->cluster); // Promotable ordering: demote before stop, start before promote if (pcmk_is_set(pe__const_top_resource(rsc, false)->flags, pcmk_rsc_promotable) || (rsc->role > pcmk_role_unpromoted)) { pcmk__new_ordering(rsc, pcmk__op_key(rsc->id, PCMK_ACTION_DEMOTE, 0), NULL, rsc, pcmk__op_key(rsc->id, PCMK_ACTION_STOP, 0), NULL, pcmk__ar_promoted_then_implies_first, rsc->cluster); pcmk__new_ordering(rsc, pcmk__op_key(rsc->id, PCMK_ACTION_START, 0), NULL, rsc, pcmk__op_key(rsc->id, PCMK_ACTION_PROMOTE, 0), NULL, - pe_order_runnable_left, rsc->cluster); + pcmk__ar_unrunnable_first_blocks, rsc->cluster); } // Don't clear resource history if probing on same node pcmk__new_ordering(rsc, pcmk__op_key(rsc->id, PCMK_ACTION_LRM_DELETE, 0), NULL, rsc, pcmk__op_key(rsc->id, PCMK_ACTION_MONITOR, 0), NULL, pe_order_same_node|pe_order_then_cancels_first, rsc->cluster); // Certain checks need allowed nodes if (check_unfencing || check_utilization || (rsc->container != NULL)) { allowed_nodes = allowed_nodes_as_list(rsc); } if (check_unfencing) { g_list_foreach(allowed_nodes, pcmk__order_restart_vs_unfence, rsc); } if (check_utilization) { pcmk__create_utilization_constraints(rsc, allowed_nodes); } if (rsc->container != NULL) { pe_resource_t *remote_rsc = NULL; if (rsc->is_remote_node) { // rsc is the implicit remote connection for a guest or bundle node /* Guest resources are not allowed to run on Pacemaker Remote nodes, * to avoid nesting remotes. However, bundles are allowed. */ if (!pcmk_is_set(rsc->flags, pcmk_rsc_remote_nesting_allowed)) { rsc_avoids_remote_nodes(rsc->container); } /* If someone cleans up a guest or bundle node's container, we will * likely schedule a (re-)probe of the container and recovery of the * connection. Order the connection stop after the container probe, * so that if we detect the container running, we will trigger a new * transition and avoid the unnecessary recovery. */ pcmk__order_resource_actions(rsc->container, PCMK_ACTION_MONITOR, rsc, PCMK_ACTION_STOP, pcmk__ar_ordered); /* A user can specify that a resource must start on a Pacemaker Remote * node by explicitly configuring it with the container=NODENAME * meta-attribute. This is of questionable merit, since location * constraints can accomplish the same thing. But we support it, so here * we check whether a resource (that is not itself a remote connection) * has container set to a remote node or guest node resource. */ } else if (rsc->container->is_remote_node) { remote_rsc = rsc->container; } else { remote_rsc = pe__resource_contains_guest_node(rsc->cluster, rsc->container); } if (remote_rsc != NULL) { /* Force the resource on the Pacemaker Remote node instead of * colocating the resource with the container resource. */ for (GList *item = allowed_nodes; item; item = item->next) { pe_node_t *node = item->data; if (node->details->remote_rsc != remote_rsc) { node->weight = -INFINITY; } } } else { /* This resource is either a filler for a container that does NOT * represent a Pacemaker Remote node, or a Pacemaker Remote * connection resource for a guest node or bundle. */ int score; crm_trace("Order and colocate %s relative to its container %s", rsc->id, rsc->container->id); pcmk__new_ordering(rsc->container, pcmk__op_key(rsc->container->id, PCMK_ACTION_START, 0), NULL, rsc, pcmk__op_key(rsc->id, PCMK_ACTION_START, 0), NULL, pcmk__ar_first_implies_then - |pe_order_runnable_left, + |pcmk__ar_unrunnable_first_blocks, rsc->cluster); pcmk__new_ordering(rsc, pcmk__op_key(rsc->id, PCMK_ACTION_STOP, 0), NULL, rsc->container, pcmk__op_key(rsc->container->id, PCMK_ACTION_STOP, 0), NULL, pcmk__ar_then_implies_first, rsc->cluster); if (pcmk_is_set(rsc->flags, pcmk_rsc_remote_nesting_allowed)) { score = 10000; /* Highly preferred but not essential */ } else { score = INFINITY; /* Force them to run on the same host */ } pcmk__new_colocation("#resource-with-container", NULL, score, rsc, rsc->container, NULL, NULL, pcmk__coloc_influence); } } if (rsc->is_remote_node || pcmk_is_set(rsc->flags, pcmk_rsc_fence_device)) { /* Remote connections and fencing devices are not allowed to run on * Pacemaker Remote nodes */ rsc_avoids_remote_nodes(rsc); } g_list_free(allowed_nodes); } /*! * \internal * \brief Apply a colocation's score to node scores or resource priority * * Given a colocation constraint, apply its score to the dependent's * allowed node scores (if we are still placing resources) or priority (if * we are choosing promotable clone instance roles). * * \param[in,out] dependent Dependent resource in colocation * \param[in] primary Primary resource in colocation * \param[in] colocation Colocation constraint to apply * \param[in] for_dependent true if called on behalf of dependent */ void pcmk__primitive_apply_coloc_score(pe_resource_t *dependent, const pe_resource_t *primary, const pcmk__colocation_t *colocation, bool for_dependent) { enum pcmk__coloc_affects filter_results; CRM_ASSERT((dependent != NULL) && (primary != NULL) && (colocation != NULL)); if (for_dependent) { // Always process on behalf of primary resource primary->cmds->apply_coloc_score(dependent, primary, colocation, false); return; } filter_results = pcmk__colocation_affects(dependent, primary, colocation, false); pe_rsc_trace(dependent, "%s %s with %s (%s, score=%d, filter=%d)", ((colocation->score > 0)? "Colocating" : "Anti-colocating"), dependent->id, primary->id, colocation->id, colocation->score, filter_results); switch (filter_results) { case pcmk__coloc_affects_role: pcmk__apply_coloc_to_priority(dependent, primary, colocation); break; case pcmk__coloc_affects_location: pcmk__apply_coloc_to_scores(dependent, primary, colocation); break; default: // pcmk__coloc_affects_nothing return; } } /* Primitive implementation of * resource_alloc_functions_t:with_this_colocations() */ void pcmk__with_primitive_colocations(const pe_resource_t *rsc, const pe_resource_t *orig_rsc, GList **list) { CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_primitive) && (list != NULL)); if (rsc == orig_rsc) { /* For the resource itself, add all of its own colocations and relevant * colocations from its parent (if any). */ pcmk__add_with_this_list(list, rsc->rsc_cons_lhs, orig_rsc); if (rsc->parent != NULL) { rsc->parent->cmds->with_this_colocations(rsc->parent, orig_rsc, list); } } else { // For an ancestor, add only explicitly configured constraints for (GList *iter = rsc->rsc_cons_lhs; iter != NULL; iter = iter->next) { pcmk__colocation_t *colocation = iter->data; if (pcmk_is_set(colocation->flags, pcmk__coloc_explicit)) { pcmk__add_with_this(list, colocation, orig_rsc); } } } } /* Primitive implementation of * resource_alloc_functions_t:this_with_colocations() */ void pcmk__primitive_with_colocations(const pe_resource_t *rsc, const pe_resource_t *orig_rsc, GList **list) { CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_primitive) && (list != NULL)); if (rsc == orig_rsc) { /* For the resource itself, add all of its own colocations and relevant * colocations from its parent (if any). */ pcmk__add_this_with_list(list, rsc->rsc_cons, orig_rsc); if (rsc->parent != NULL) { rsc->parent->cmds->this_with_colocations(rsc->parent, orig_rsc, list); } } else { // For an ancestor, add only explicitly configured constraints for (GList *iter = rsc->rsc_cons; iter != NULL; iter = iter->next) { pcmk__colocation_t *colocation = iter->data; if (pcmk_is_set(colocation->flags, pcmk__coloc_explicit)) { pcmk__add_this_with(list, colocation, orig_rsc); } } } } /*! * \internal * \brief Return action flags for a given primitive resource action * * \param[in,out] action Action to get flags for * \param[in] node If not NULL, limit effects to this node (ignored) * * \return Flags appropriate to \p action on \p node */ uint32_t pcmk__primitive_action_flags(pe_action_t *action, const pe_node_t *node) { CRM_ASSERT(action != NULL); return (uint32_t) action->flags; } /*! * \internal * \brief Check whether a node is a multiply active resource's expected node * * \param[in] rsc Resource to check * \param[in] node Node to check * * \return true if \p rsc is multiply active with multiple-active set to * stop_unexpected, and \p node is the node where it will remain active * \note This assumes that the resource's next role cannot be changed to stopped * after this is called, which should be reasonable if status has already * been unpacked and resources have been assigned to nodes. */ static bool is_expected_node(const pe_resource_t *rsc, const pe_node_t *node) { return pcmk_all_flags_set(rsc->flags, pcmk_rsc_stop_unexpected|pcmk_rsc_restarting) && (rsc->next_role > pcmk_role_stopped) && pe__same_node(rsc->allocated_to, node); } /*! * \internal * \brief Schedule actions needed to stop a resource wherever it is active * * \param[in,out] rsc Resource being stopped * \param[in] node Node where resource is being stopped (ignored) * \param[in] optional Whether actions should be optional */ static void stop_resource(pe_resource_t *rsc, pe_node_t *node, bool optional) { for (GList *iter = rsc->running_on; iter != NULL; iter = iter->next) { pe_node_t *current = (pe_node_t *) iter->data; pe_action_t *stop = NULL; if (is_expected_node(rsc, current)) { /* We are scheduling restart actions for a multiply active resource * with multiple-active=stop_unexpected, and this is where it should * not be stopped. */ pe_rsc_trace(rsc, "Skipping stop of multiply active resource %s " "on expected node %s", rsc->id, pe__node_name(current)); continue; } if (rsc->partial_migration_target != NULL) { // Continue migration if node originally was and remains target if (pe__same_node(current, rsc->partial_migration_target) && pe__same_node(current, rsc->allocated_to)) { pe_rsc_trace(rsc, "Skipping stop of %s on %s " "because partial migration there will continue", rsc->id, pe__node_name(current)); continue; } else { pe_rsc_trace(rsc, "Forcing stop of %s on %s " "because migration target changed", rsc->id, pe__node_name(current)); optional = false; } } pe_rsc_trace(rsc, "Scheduling stop of %s on %s", rsc->id, pe__node_name(current)); stop = stop_action(rsc, current, optional); if (rsc->allocated_to == NULL) { pe_action_set_reason(stop, "node availability", true); } else if (pcmk_all_flags_set(rsc->flags, pcmk_rsc_restarting |pcmk_rsc_stop_unexpected)) { /* We are stopping a multiply active resource on a node that is * not its expected node, and we are still scheduling restart * actions, so the stop is for being multiply active. */ pe_action_set_reason(stop, "being multiply active", true); } if (!pcmk_is_set(rsc->flags, pcmk_rsc_managed)) { pe__clear_action_flags(stop, pcmk_action_runnable); } if (pcmk_is_set(rsc->cluster->flags, pcmk_sched_remove_after_stop)) { pcmk__schedule_cleanup(rsc, current, optional); } if (pcmk_is_set(rsc->flags, pcmk_rsc_needs_unfencing)) { pe_action_t *unfence = pe_fence_op(current, PCMK_ACTION_ON, true, NULL, false, rsc->cluster); order_actions(stop, unfence, pcmk__ar_then_implies_first); if (!pcmk__node_unfenced(current)) { pe_proc_err("Stopping %s until %s can be unfenced", rsc->id, pe__node_name(current)); } } } } /*! * \internal * \brief Schedule actions needed to start a resource on a node * * \param[in,out] rsc Resource being started * \param[in,out] node Node where resource should be started * \param[in] optional Whether actions should be optional */ static void start_resource(pe_resource_t *rsc, pe_node_t *node, bool optional) { pe_action_t *start = NULL; CRM_ASSERT(node != NULL); pe_rsc_trace(rsc, "Scheduling %s start of %s on %s (score %d)", (optional? "optional" : "required"), rsc->id, pe__node_name(node), node->weight); start = start_action(rsc, node, TRUE); pcmk__order_vs_unfence(rsc, node, start, pcmk__ar_first_implies_then); if (pcmk_is_set(start->flags, pcmk_action_runnable) && !optional) { pe__clear_action_flags(start, pcmk_action_optional); } if (is_expected_node(rsc, node)) { /* This could be a problem if the start becomes necessary for other * reasons later. */ pe_rsc_trace(rsc, "Start of multiply active resouce %s " "on expected node %s will be a pseudo-action", rsc->id, pe__node_name(node)); pe__set_action_flags(start, pcmk_action_pseudo); } } /*! * \internal * \brief Schedule actions needed to promote a resource on a node * * \param[in,out] rsc Resource being promoted * \param[in] node Node where resource should be promoted * \param[in] optional Whether actions should be optional */ static void promote_resource(pe_resource_t *rsc, pe_node_t *node, bool optional) { GList *iter = NULL; GList *action_list = NULL; bool runnable = true; CRM_ASSERT(node != NULL); // Any start must be runnable for promotion to be runnable action_list = pe__resource_actions(rsc, node, PCMK_ACTION_START, true); for (iter = action_list; iter != NULL; iter = iter->next) { pe_action_t *start = (pe_action_t *) iter->data; if (!pcmk_is_set(start->flags, pcmk_action_runnable)) { runnable = false; } } g_list_free(action_list); if (runnable) { pe_action_t *promote = promote_action(rsc, node, optional); pe_rsc_trace(rsc, "Scheduling %s promotion of %s on %s", (optional? "optional" : "required"), rsc->id, pe__node_name(node)); if (is_expected_node(rsc, node)) { /* This could be a problem if the promote becomes necessary for * other reasons later. */ pe_rsc_trace(rsc, "Promotion of multiply active resouce %s " "on expected node %s will be a pseudo-action", rsc->id, pe__node_name(node)); pe__set_action_flags(promote, pcmk_action_pseudo); } } else { pe_rsc_trace(rsc, "Not promoting %s on %s: start unrunnable", rsc->id, pe__node_name(node)); action_list = pe__resource_actions(rsc, node, PCMK_ACTION_PROMOTE, true); for (iter = action_list; iter != NULL; iter = iter->next) { pe_action_t *promote = (pe_action_t *) iter->data; pe__clear_action_flags(promote, pcmk_action_runnable); } g_list_free(action_list); } } /*! * \internal * \brief Schedule actions needed to demote a resource wherever it is active * * \param[in,out] rsc Resource being demoted * \param[in] node Node where resource should be demoted (ignored) * \param[in] optional Whether actions should be optional */ static void demote_resource(pe_resource_t *rsc, pe_node_t *node, bool optional) { /* Since this will only be called for a primitive (possibly as an instance * of a collective resource), the resource is multiply active if it is * running on more than one node, so we want to demote on all of them as * part of recovery, regardless of which one is the desired node. */ for (GList *iter = rsc->running_on; iter != NULL; iter = iter->next) { pe_node_t *current = (pe_node_t *) iter->data; if (is_expected_node(rsc, current)) { pe_rsc_trace(rsc, "Skipping demote of multiply active resource %s " "on expected node %s", rsc->id, pe__node_name(current)); } else { pe_rsc_trace(rsc, "Scheduling %s demotion of %s on %s", (optional? "optional" : "required"), rsc->id, pe__node_name(current)); demote_action(rsc, current, optional); } } } static void assert_role_error(pe_resource_t *rsc, pe_node_t *node, bool optional) { CRM_ASSERT(false); } /*! * \internal * \brief Schedule cleanup of a resource * * \param[in,out] rsc Resource to clean up * \param[in] node Node to clean up on * \param[in] optional Whether clean-up should be optional */ void pcmk__schedule_cleanup(pe_resource_t *rsc, const pe_node_t *node, bool optional) { /* If the cleanup is required, its orderings are optional, because they're * relevant only if both actions are required. Conversely, if the cleanup is * optional, the orderings make the then action required if the first action * becomes required. */ uint32_t flag = optional? pcmk__ar_first_implies_then : pcmk__ar_ordered; CRM_CHECK((rsc != NULL) && (node != NULL), return); if (pcmk_is_set(rsc->flags, pcmk_rsc_failed)) { pe_rsc_trace(rsc, "Skipping clean-up of %s on %s: resource failed", rsc->id, pe__node_name(node)); return; } if (node->details->unclean || !node->details->online) { pe_rsc_trace(rsc, "Skipping clean-up of %s on %s: node unavailable", rsc->id, pe__node_name(node)); return; } crm_notice("Scheduling clean-up of %s on %s", rsc->id, pe__node_name(node)); delete_action(rsc, node, optional); // stop -> clean-up -> start pcmk__order_resource_actions(rsc, PCMK_ACTION_STOP, rsc, PCMK_ACTION_DELETE, flag); pcmk__order_resource_actions(rsc, PCMK_ACTION_DELETE, rsc, PCMK_ACTION_START, flag); } /*! * \internal * \brief Add primitive meta-attributes relevant to graph actions to XML * * \param[in] rsc Primitive resource whose meta-attributes should be added * \param[in,out] xml Transition graph action attributes XML to add to */ void pcmk__primitive_add_graph_meta(const pe_resource_t *rsc, xmlNode *xml) { char *name = NULL; char *value = NULL; const pe_resource_t *parent = NULL; CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_primitive) && (xml != NULL)); /* Clone instance numbers get set internally as meta-attributes, and are * needed in the transition graph (for example, to tell unique clone * instances apart). */ value = g_hash_table_lookup(rsc->meta, XML_RSC_ATTR_INCARNATION); if (value != NULL) { name = crm_meta_name(XML_RSC_ATTR_INCARNATION); crm_xml_add(xml, name, value); free(name); } // Not sure if this one is really needed ... value = g_hash_table_lookup(rsc->meta, XML_RSC_ATTR_REMOTE_NODE); if (value != NULL) { name = crm_meta_name(XML_RSC_ATTR_REMOTE_NODE); crm_xml_add(xml, name, value); free(name); } /* The container meta-attribute can be set on the primitive itself or one of * its parents (for example, a group inside a container resource), so check * them all, and keep the highest one found. */ for (parent = rsc; parent != NULL; parent = parent->parent) { if (parent->container != NULL) { crm_xml_add(xml, CRM_META "_" XML_RSC_ATTR_CONTAINER, parent->container->id); } } /* Bundle replica children will get their external-ip set internally as a * meta-attribute. The graph action needs it, but under a different naming * convention than other meta-attributes. */ value = g_hash_table_lookup(rsc->meta, "external-ip"); if (value != NULL) { crm_xml_add(xml, "pcmk_external_ip", value); } } // Primitive implementation of resource_alloc_functions_t:add_utilization() void pcmk__primitive_add_utilization(const pe_resource_t *rsc, const pe_resource_t *orig_rsc, GList *all_rscs, GHashTable *utilization) { CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_primitive) && (orig_rsc != NULL) && (utilization != NULL)); if (!pcmk_is_set(rsc->flags, pcmk_rsc_unassigned)) { return; } pe_rsc_trace(orig_rsc, "%s: Adding primitive %s as colocated utilization", orig_rsc->id, rsc->id); pcmk__release_node_capacity(utilization, rsc); } /*! * \internal * \brief Get epoch time of node's shutdown attribute (or now if none) * * \param[in,out] node Node to check * * \return Epoch time corresponding to shutdown attribute if set or now if not */ static time_t shutdown_time(pe_node_t *node) { const char *shutdown = pe_node_attribute_raw(node, XML_CIB_ATTR_SHUTDOWN); time_t result = 0; if (shutdown != NULL) { long long result_ll; if (pcmk__scan_ll(shutdown, &result_ll, 0LL) == pcmk_rc_ok) { result = (time_t) result_ll; } } return (result == 0)? get_effective_time(node->details->data_set) : result; } /*! * \internal * \brief Ban a resource from a node if it's not locked to the node * * \param[in] data Node to check * \param[in,out] user_data Resource to check */ static void ban_if_not_locked(gpointer data, gpointer user_data) { const pe_node_t *node = (const pe_node_t *) data; pe_resource_t *rsc = (pe_resource_t *) user_data; if (strcmp(node->details->uname, rsc->lock_node->details->uname) != 0) { resource_location(rsc, node, -CRM_SCORE_INFINITY, XML_CONFIG_ATTR_SHUTDOWN_LOCK, rsc->cluster); } } // Primitive implementation of resource_alloc_functions_t:shutdown_lock() void pcmk__primitive_shutdown_lock(pe_resource_t *rsc) { const char *class = NULL; CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_primitive)); class = crm_element_value(rsc->xml, XML_AGENT_ATTR_CLASS); // Fence devices and remote connections can't be locked if (pcmk__str_eq(class, PCMK_RESOURCE_CLASS_STONITH, pcmk__str_null_matches) || pe__resource_is_remote_conn(rsc)) { return; } if (rsc->lock_node != NULL) { // The lock was obtained from resource history if (rsc->running_on != NULL) { /* The resource was started elsewhere even though it is now * considered locked. This shouldn't be possible, but as a * failsafe, we don't want to disturb the resource now. */ pe_rsc_info(rsc, "Cancelling shutdown lock because %s is already active", rsc->id); pe__clear_resource_history(rsc, rsc->lock_node); rsc->lock_node = NULL; rsc->lock_time = 0; } // Only a resource active on exactly one node can be locked } else if (pcmk__list_of_1(rsc->running_on)) { pe_node_t *node = rsc->running_on->data; if (node->details->shutdown) { if (node->details->unclean) { pe_rsc_debug(rsc, "Not locking %s to unclean %s for shutdown", rsc->id, pe__node_name(node)); } else { rsc->lock_node = node; rsc->lock_time = shutdown_time(node); } } } if (rsc->lock_node == NULL) { // No lock needed return; } if (rsc->cluster->shutdown_lock > 0) { time_t lock_expiration = rsc->lock_time + rsc->cluster->shutdown_lock; pe_rsc_info(rsc, "Locking %s to %s due to shutdown (expires @%lld)", rsc->id, pe__node_name(rsc->lock_node), (long long) lock_expiration); pe__update_recheck_time(++lock_expiration, rsc->cluster); } else { pe_rsc_info(rsc, "Locking %s to %s due to shutdown", rsc->id, pe__node_name(rsc->lock_node)); } // If resource is locked to one node, ban it from all other nodes g_list_foreach(rsc->cluster->nodes, ban_if_not_locked, rsc); } diff --git a/lib/pacemaker/pcmk_sched_probes.c b/lib/pacemaker/pcmk_sched_probes.c index 1438bff079..1d60eb78ee 100644 --- a/lib/pacemaker/pcmk_sched_probes.c +++ b/lib/pacemaker/pcmk_sched_probes.c @@ -1,902 +1,902 @@ /* * Copyright 2004-2023 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU General Public License version 2 * or later (GPLv2+) WITHOUT ANY WARRANTY. */ #include #include #include #include #include #include "libpacemaker_private.h" /*! * \internal * \brief Add the expected result to a newly created probe * * \param[in,out] probe Probe action to add expected result to * \param[in] rsc Resource that probe is for * \param[in] node Node that probe will run on */ static void add_expected_result(pe_action_t *probe, const pe_resource_t *rsc, const pe_node_t *node) { // Check whether resource is currently active on node pe_node_t *running = pe_find_node_id(rsc->running_on, node->details->id); // The expected result is what we think the resource's current state is if (running == NULL) { pe__add_action_expected_result(probe, CRM_EX_NOT_RUNNING); } else if (rsc->role == pcmk_role_promoted) { pe__add_action_expected_result(probe, CRM_EX_PROMOTED); } } /*! * \internal * \brief Create any needed robes on a node for a list of resources * * \param[in,out] rscs List of resources to create probes for * \param[in,out] node Node to create probes on * * \return true if any probe was created, otherwise false */ bool pcmk__probe_resource_list(GList *rscs, pe_node_t *node) { bool any_created = false; for (GList *iter = rscs; iter != NULL; iter = iter->next) { pe_resource_t *rsc = (pe_resource_t *) iter->data; if (rsc->cmds->create_probe(rsc, node)) { any_created = true; } } return any_created; } /*! * \internal * \brief Order one resource's start after another's start-up probe * * \param[in,out] rsc1 Resource that might get start-up probe * \param[in] rsc2 Resource that might be started */ static void probe_then_start(pe_resource_t *rsc1, pe_resource_t *rsc2) { if ((rsc1->allocated_to != NULL) && (g_hash_table_lookup(rsc1->known_on, rsc1->allocated_to->details->id) == NULL)) { pcmk__new_ordering(rsc1, pcmk__op_key(rsc1->id, PCMK_ACTION_MONITOR, 0), NULL, rsc2, pcmk__op_key(rsc2->id, PCMK_ACTION_START, 0), NULL, pcmk__ar_ordered, rsc1->cluster); } } /*! * \internal * \brief Check whether a guest resource will stop * * \param[in] node Guest node to check * * \return true if guest resource will likely stop, otherwise false */ static bool guest_resource_will_stop(const pe_node_t *node) { const pe_resource_t *guest_rsc = node->details->remote_rsc->container; /* Ideally, we'd check whether the guest has a required stop, but that * information doesn't exist yet, so approximate it ... */ return node->details->remote_requires_reset || node->details->unclean || pcmk_is_set(guest_rsc->flags, pcmk_rsc_failed) || (guest_rsc->next_role == pcmk_role_stopped) // Guest is moving || ((guest_rsc->role > pcmk_role_stopped) && (guest_rsc->allocated_to != NULL) && (pe_find_node(guest_rsc->running_on, guest_rsc->allocated_to->details->uname) == NULL)); } /*! * \internal * \brief Create a probe action for a resource on a node * * \param[in,out] rsc Resource to create probe for * \param[in,out] node Node to create probe on * * \return Newly created probe action */ static pe_action_t * probe_action(pe_resource_t *rsc, pe_node_t *node) { pe_action_t *probe = NULL; char *key = pcmk__op_key(rsc->id, PCMK_ACTION_MONITOR, 0); crm_debug("Scheduling probe of %s %s on %s", role2text(rsc->role), rsc->id, pe__node_name(node)); probe = custom_action(rsc, key, PCMK_ACTION_MONITOR, node, FALSE, TRUE, rsc->cluster); pe__clear_action_flags(probe, pcmk_action_optional); pcmk__order_vs_unfence(rsc, node, probe, pcmk__ar_ordered); add_expected_result(probe, rsc, node); return probe; } /*! * \internal * \brief Create probes for a resource on a node, if needed * * \brief Schedule any probes needed for a resource on a node * * \param[in,out] rsc Resource to create probe for * \param[in,out] node Node to create probe on * * \return true if any probe was created, otherwise false */ bool pcmk__probe_rsc_on_node(pe_resource_t *rsc, pe_node_t *node) { uint32_t flags = pcmk__ar_ordered; pe_action_t *probe = NULL; pe_node_t *allowed = NULL; pe_resource_t *top = uber_parent(rsc); const char *reason = NULL; CRM_ASSERT((rsc != NULL) && (node != NULL)); if (!pcmk_is_set(rsc->cluster->flags, pcmk_sched_probe_resources)) { reason = "start-up probes are disabled"; goto no_probe; } if (pe__is_guest_or_remote_node(node)) { const char *class = crm_element_value(rsc->xml, XML_AGENT_ATTR_CLASS); if (pcmk__str_eq(class, PCMK_RESOURCE_CLASS_STONITH, pcmk__str_none)) { reason = "Pacemaker Remote nodes cannot run stonith agents"; goto no_probe; } else if (pe__is_guest_node(node) && pe__resource_contains_guest_node(rsc->cluster, rsc)) { reason = "guest nodes cannot run resources containing guest nodes"; goto no_probe; } else if (rsc->is_remote_node) { reason = "Pacemaker Remote nodes cannot host remote connections"; goto no_probe; } } // If this is a collective resource, probes are created for its children if (rsc->children != NULL) { return pcmk__probe_resource_list(rsc->children, node); } if ((rsc->container != NULL) && !rsc->is_remote_node) { reason = "resource is inside a container"; goto no_probe; } else if (pcmk_is_set(rsc->flags, pcmk_rsc_removed)) { reason = "resource is orphaned"; goto no_probe; } else if (g_hash_table_lookup(rsc->known_on, node->details->id) != NULL) { reason = "resource state is already known"; goto no_probe; } allowed = g_hash_table_lookup(rsc->allowed_nodes, node->details->id); if (rsc->exclusive_discover || top->exclusive_discover) { // Exclusive discovery is enabled ... if (allowed == NULL) { // ... but this node is not allowed to run the resource reason = "resource has exclusive discovery but is not allowed " "on node"; goto no_probe; } else if (allowed->rsc_discover_mode != pcmk_probe_exclusive) { // ... but no constraint marks this node for discovery of resource reason = "resource has exclusive discovery but is not enabled " "on node"; goto no_probe; } } if (allowed == NULL) { allowed = node; } if (allowed->rsc_discover_mode == pcmk_probe_never) { reason = "node has discovery disabled"; goto no_probe; } if (pe__is_guest_node(node)) { pe_resource_t *guest = node->details->remote_rsc->container; if (guest->role == pcmk_role_stopped) { // The guest is stopped, so we know no resource is active there reason = "node's guest is stopped"; probe_then_start(guest, top); goto no_probe; } else if (guest_resource_will_stop(node)) { reason = "node's guest will stop"; // Order resource start after guest stop (in case it's restarting) pcmk__new_ordering(guest, pcmk__op_key(guest->id, PCMK_ACTION_STOP, 0), NULL, top, pcmk__op_key(top->id, PCMK_ACTION_START, 0), NULL, pcmk__ar_ordered, rsc->cluster); goto no_probe; } } // We've eliminated all cases where a probe is not needed, so now it is probe = probe_action(rsc, node); /* Below, we will order the probe relative to start or reload. If this is a * clone instance, the start or reload is for the entire clone rather than * just the instance. Otherwise, the start or reload is for the resource * itself. */ if (!pe_rsc_is_clone(top)) { top = rsc; } /* Prevent a start if the resource can't be probed, but don't cause the * resource or entire clone to stop if already active. */ if (!pcmk_is_set(probe->flags, pcmk_action_runnable) && (top->running_on == NULL)) { - pe__set_order_flags(flags, pe_order_runnable_left); + pe__set_order_flags(flags, pcmk__ar_unrunnable_first_blocks); } // Start or reload after probing the resource pcmk__new_ordering(rsc, NULL, probe, top, pcmk__op_key(top->id, PCMK_ACTION_START, 0), NULL, flags, rsc->cluster); pcmk__new_ordering(rsc, NULL, probe, top, reload_key(rsc), NULL, pcmk__ar_ordered, rsc->cluster); return true; no_probe: pe_rsc_trace(rsc, "Skipping probe for %s on %s because %s", rsc->id, node->details->id, reason); return false; } /*! * \internal * \brief Check whether a probe should be ordered before another action * * \param[in] probe Probe action to check * \param[in] then Other action to check * * \return true if \p probe should be ordered before \p then, otherwise false */ static bool probe_needed_before_action(const pe_action_t *probe, const pe_action_t *then) { // Probes on a node are performed after unfencing it, not before if (pcmk__str_eq(then->task, PCMK_ACTION_STONITH, pcmk__str_none) && pe__same_node(probe->node, then->node)) { const char *op = g_hash_table_lookup(then->meta, "stonith_action"); if (pcmk__str_eq(op, PCMK_ACTION_ON, pcmk__str_casei)) { return false; } } // Probes should be done on a node before shutting it down if (pcmk__str_eq(then->task, PCMK_ACTION_DO_SHUTDOWN, pcmk__str_none) && (probe->node != NULL) && (then->node != NULL) && !pe__same_node(probe->node, then->node)) { return false; } // Otherwise probes should always be done before any other action return true; } /*! * \internal * \brief Add implicit "probe then X" orderings for "stop then X" orderings * * If the state of a resource is not known yet, a probe will be scheduled, * expecting a "not running" result. If the probe fails, a stop will not be * scheduled until the next transition. Thus, if there are ordering constraints * like "stop this resource then do something else that's not for the same * resource", add implicit "probe this resource then do something" equivalents * so the relation is upheld until we know whether a stop is needed. * * \param[in,out] data_set Cluster working set */ static void add_probe_orderings_for_stops(pe_working_set_t *data_set) { for (GList *iter = data_set->ordering_constraints; iter != NULL; iter = iter->next) { pe__ordering_t *order = iter->data; uint32_t order_flags = pcmk__ar_ordered; GList *probes = NULL; GList *then_actions = NULL; pe_action_t *first = NULL; pe_action_t *then = NULL; // Skip disabled orderings if (order->flags == pcmk__ar_none) { continue; } // Skip non-resource orderings, and orderings for the same resource if ((order->lh_rsc == NULL) || (order->lh_rsc == order->rh_rsc)) { continue; } // Skip invalid orderings (shouldn't be possible) first = order->lh_action; then = order->rh_action; if (((first == NULL) && (order->lh_action_task == NULL)) || ((then == NULL) && (order->rh_action_task == NULL))) { continue; } // Skip orderings for first actions other than stop if ((first != NULL) && !pcmk__str_eq(first->task, PCMK_ACTION_STOP, pcmk__str_none)) { continue; } else if ((first == NULL) && !pcmk__ends_with(order->lh_action_task, "_" PCMK_ACTION_STOP "_0")) { continue; } /* Do not imply a probe ordering for a resource inside of a stopping * container. Otherwise, it might introduce a transition loop, since a * probe could be scheduled after the container starts again. */ if ((order->rh_rsc != NULL) && (order->lh_rsc->container == order->rh_rsc)) { if ((then != NULL) && pcmk__str_eq(then->task, PCMK_ACTION_STOP, pcmk__str_none)) { continue; } else if ((then == NULL) && pcmk__ends_with(order->rh_action_task, "_" PCMK_ACTION_STOP "_0")) { continue; } } // Preserve certain order options for future filtering if (pcmk_is_set(order->flags, pcmk__ar_if_first_unmigratable)) { pe__set_order_flags(order_flags, pcmk__ar_if_first_unmigratable); } if (pcmk_is_set(order->flags, pe_order_same_node)) { pe__set_order_flags(order_flags, pe_order_same_node); } // Preserve certain order types for future filtering if ((order->flags == pe_order_anti_colocation) || (order->flags == pe_order_load)) { order_flags = order->flags; } // List all scheduled probes for the first resource probes = pe__resource_actions(order->lh_rsc, NULL, PCMK_ACTION_MONITOR, FALSE); if (probes == NULL) { // There aren't any continue; } // List all relevant "then" actions if (then != NULL) { then_actions = g_list_prepend(NULL, then); } else if (order->rh_rsc != NULL) { then_actions = find_actions(order->rh_rsc->actions, order->rh_action_task, NULL); if (then_actions == NULL) { // There aren't any g_list_free(probes); continue; } } crm_trace("Implying 'probe then' orderings for '%s then %s' " "(id=%d, type=%.6x)", ((first == NULL)? order->lh_action_task : first->uuid), ((then == NULL)? order->rh_action_task : then->uuid), order->id, order->flags); for (GList *probe_iter = probes; probe_iter != NULL; probe_iter = probe_iter->next) { pe_action_t *probe = (pe_action_t *) probe_iter->data; for (GList *then_iter = then_actions; then_iter != NULL; then_iter = then_iter->next) { pe_action_t *then = (pe_action_t *) then_iter->data; if (probe_needed_before_action(probe, then)) { order_actions(probe, then, order_flags); } } } g_list_free(then_actions); g_list_free(probes); } } /*! * \internal * \brief Add necessary orderings between probe and starts of clone instances * * , in additon to the ordering with the parent resource added upon creating * the probe. * * \param[in,out] probe Probe as 'first' action in an ordering * \param[in,out] after 'then' action wrapper in the ordering */ static void add_start_orderings_for_probe(pe_action_t *probe, pe_action_wrapper_t *after) { - uint32_t flags = pcmk__ar_ordered|pe_order_runnable_left; + uint32_t flags = pcmk__ar_ordered|pcmk__ar_unrunnable_first_blocks; /* Although the ordering between the probe of the clone instance and the * start of its parent has been added in pcmk__probe_rsc_on_node(), we - * avoided enforcing `pe_order_runnable_left` order type for that as long as - * any of the clone instances are running to prevent them from being - * unexpectedly stopped. + * avoided enforcing `pcmk__ar_unrunnable_first_blocks` order type for that + * as long as any of the clone instances are running to prevent them from + * being unexpectedly stopped. * * On the other hand, we still need to prevent any inactive instances from * starting unless the probe is runnable so that we don't risk starting too * many instances before we know the state on all nodes. */ if ((after->action->rsc->variant <= pcmk_rsc_variant_group) || pcmk_is_set(probe->flags, pcmk_action_runnable) // The order type is already enforced for its parent. - || pcmk_is_set(after->type, pe_order_runnable_left) + || pcmk_is_set(after->type, pcmk__ar_unrunnable_first_blocks) || (pe__const_top_resource(probe->rsc, false) != after->action->rsc) || !pcmk__str_eq(after->action->task, PCMK_ACTION_START, pcmk__str_none)) { return; } crm_trace("Adding probe start orderings for 'unrunnable %s@%s " "then instances of %s@%s'", probe->uuid, pe__node_name(probe->node), after->action->uuid, pe__node_name(after->action->node)); for (GList *then_iter = after->action->actions_after; then_iter != NULL; then_iter = then_iter->next) { pe_action_wrapper_t *then = (pe_action_wrapper_t *) then_iter->data; if (then->action->rsc->running_on || (pe__const_top_resource(then->action->rsc, false) != after->action->rsc) || !pcmk__str_eq(then->action->task, PCMK_ACTION_START, pcmk__str_none)) { continue; } crm_trace("Adding probe start ordering for 'unrunnable %s@%s " "then %s@%s' (type=%#.6x)", probe->uuid, pe__node_name(probe->node), then->action->uuid, pe__node_name(then->action->node), flags); /* Prevent the instance from starting if the instance can't, but don't * cause any other intances to stop if already active. */ order_actions(probe, then->action, flags); } return; } /*! * \internal * \brief Order probes before restarts and re-promotes * * If a given ordering is a "probe then start" or "probe then promote" ordering, * add an implicit "probe then stop/demote" ordering in case the action is part * of a restart/re-promote, and do the same recursively for all actions ordered * after the "then" action. * * \param[in,out] probe Probe as 'first' action in an ordering * \param[in,out] after 'then' action in the ordering */ static void add_restart_orderings_for_probe(pe_action_t *probe, pe_action_t *after) { GList *iter = NULL; bool interleave = false; pe_resource_t *compatible_rsc = NULL; // Validate that this is a resource probe followed by some action if ((after == NULL) || (probe == NULL) || (probe->rsc == NULL) || (probe->rsc->variant != pcmk_rsc_variant_primitive) || !pcmk__str_eq(probe->task, PCMK_ACTION_MONITOR, pcmk__str_none)) { return; } // Avoid running into any possible loop if (pcmk_is_set(after->flags, pcmk_action_detect_loop)) { return; } pe__set_action_flags(after, pcmk_action_detect_loop); crm_trace("Adding probe restart orderings for '%s@%s then %s@%s'", probe->uuid, pe__node_name(probe->node), after->uuid, pe__node_name(after->node)); /* Add restart orderings if "then" is for a different primitive. * Orderings for collective resources will be added later. */ if ((after->rsc != NULL) && (after->rsc->variant == pcmk_rsc_variant_primitive) && (probe->rsc != after->rsc)) { GList *then_actions = NULL; if (pcmk__str_eq(after->task, PCMK_ACTION_START, pcmk__str_none)) { then_actions = pe__resource_actions(after->rsc, NULL, PCMK_ACTION_STOP, FALSE); } else if (pcmk__str_eq(after->task, PCMK_ACTION_PROMOTE, pcmk__str_none)) { then_actions = pe__resource_actions(after->rsc, NULL, PCMK_ACTION_DEMOTE, FALSE); } for (iter = then_actions; iter != NULL; iter = iter->next) { pe_action_t *then = (pe_action_t *) iter->data; // Skip pseudo-actions (for example, those implied by fencing) if (!pcmk_is_set(then->flags, pcmk_action_pseudo)) { order_actions(probe, then, pcmk__ar_ordered); } } g_list_free(then_actions); } /* Detect whether "then" is an interleaved clone action. For these, we want * to add orderings only for the relevant instance. */ if ((after->rsc != NULL) && (after->rsc->variant > pcmk_rsc_variant_group)) { const char *interleave_s = g_hash_table_lookup(after->rsc->meta, XML_RSC_ATTR_INTERLEAVE); interleave = crm_is_true(interleave_s); if (interleave) { compatible_rsc = pcmk__find_compatible_instance(probe->rsc, after->rsc, pcmk_role_unknown, false); } } /* Now recursively do the same for all actions ordered after "then". This * also handles collective resources since the collective action will be * ordered before its individual instances' actions. */ for (iter = after->actions_after; iter != NULL; iter = iter->next) { pe_action_wrapper_t *after_wrapper = (pe_action_wrapper_t *) iter->data; /* pcmk__ar_first_implies_then is the reason why a required A.start * implies/enforces B.start to be required too, which is the cause of * B.restart/re-promote. * * Not sure about pe_order_implies_then_on_node though. It's now only * used for unfencing case, which tends to introduce transition * loops... */ if (!pcmk_is_set(after_wrapper->type, pcmk__ar_first_implies_then)) { /* The order type between a group/clone and its child such as * B.start-> B_child.start is: - * pe_order_implies_first_printed | pe_order_runnable_left + * pe_order_implies_first_printed|pcmk__ar_unrunnable_first_blocks * * Proceed through the ordering chain and build dependencies with * its children. */ if ((after->rsc == NULL) || (after->rsc->variant < pcmk_rsc_variant_group) || (probe->rsc->parent == after->rsc) || (after_wrapper->action->rsc == NULL) || (after_wrapper->action->rsc->variant > pcmk_rsc_variant_group) || (after->rsc != after_wrapper->action->rsc->parent)) { continue; } /* Proceed to the children of a group or a non-interleaved clone. * For an interleaved clone, proceed only to the relevant child. */ if ((after->rsc->variant > pcmk_rsc_variant_group) && interleave && ((compatible_rsc == NULL) || (compatible_rsc != after_wrapper->action->rsc))) { continue; } } crm_trace("Recursively adding probe restart orderings for " "'%s@%s then %s@%s' (type=%#.6x)", after->uuid, pe__node_name(after->node), after_wrapper->action->uuid, pe__node_name(after_wrapper->action->node), after_wrapper->type); add_restart_orderings_for_probe(probe, after_wrapper->action); } } /*! * \internal * \brief Clear the tracking flag on all scheduled actions * * \param[in,out] data_set Cluster working set */ static void clear_actions_tracking_flag(pe_working_set_t *data_set) { for (GList *iter = data_set->actions; iter != NULL; iter = iter->next) { pe_action_t *action = iter->data; pe__clear_action_flags(action, pcmk_action_detect_loop); } } /*! * \internal * \brief Add start and restart orderings for probes scheduled for a resource * * \param[in,out] data Resource whose probes should be ordered * \param[in] user_data Unused */ static void add_start_restart_orderings_for_rsc(gpointer data, gpointer user_data) { pe_resource_t *rsc = data; GList *probes = NULL; // For collective resources, order each instance recursively if (rsc->variant != pcmk_rsc_variant_primitive) { g_list_foreach(rsc->children, add_start_restart_orderings_for_rsc, NULL); return; } // Find all probes for given resource probes = pe__resource_actions(rsc, NULL, PCMK_ACTION_MONITOR, FALSE); // Add probe restart orderings for each probe found for (GList *iter = probes; iter != NULL; iter = iter->next) { pe_action_t *probe = (pe_action_t *) iter->data; for (GList *then_iter = probe->actions_after; then_iter != NULL; then_iter = then_iter->next) { pe_action_wrapper_t *then = (pe_action_wrapper_t *) then_iter->data; add_start_orderings_for_probe(probe, then); add_restart_orderings_for_probe(probe, then->action); clear_actions_tracking_flag(rsc->cluster); } } g_list_free(probes); } /*! * \internal * \brief Add "A then probe B" orderings for "A then B" orderings * * \param[in,out] data_set Cluster working set * * \note This function is currently disabled (see next comment). */ static void order_then_probes(pe_working_set_t *data_set) { #if 0 /* Given an ordering "A then B", we would prefer to wait for A to be started * before probing B. * * For example, if A is a filesystem which B can't even run without, it * would be helpful if the author of B's agent could assume that A is * running before B.monitor will be called. * * However, we can't _only_ probe after A is running, otherwise we wouldn't * detect the state of B if A could not be started. We can't even do an * opportunistic version of this, because B may be moving: * * A.stop -> A.start -> B.probe -> B.stop -> B.start * * and if we add B.stop -> A.stop here, we get a loop: * * A.stop -> A.start -> B.probe -> B.stop -> A.stop * * We could kill the "B.probe -> B.stop" dependency, but that could mean * stopping B "too" soon, because B.start must wait for the probe, and * we don't want to stop B if we can't start it. * * We could add the ordering only if A is an anonymous clone with * clone-max == node-max (since we'll never be moving it). However, we could * still be stopping one instance at the same time as starting another. * * The complexity of checking for allowed conditions combined with the ever * narrowing use case suggests that this code should remain disabled until * someone gets smarter. */ for (GList *iter = data_set->resources; iter != NULL; iter = iter->next) { pe_resource_t *rsc = (pe_resource_t *) iter->data; pe_action_t *start = NULL; GList *actions = NULL; GList *probes = NULL; actions = pe__resource_actions(rsc, NULL, PCMK_ACTION_START, FALSE); if (actions) { start = actions->data; g_list_free(actions); } if (start == NULL) { crm_err("No start action for %s", rsc->id); continue; } probes = pe__resource_actions(rsc, NULL, PCMK_ACTION_MONITOR, FALSE); for (actions = start->actions_before; actions != NULL; actions = actions->next) { pe_action_wrapper_t *before = (pe_action_wrapper_t *) actions->data; pe_action_t *first = before->action; pe_resource_t *first_rsc = first->rsc; if (first->required_runnable_before) { for (GList *clone_actions = first->actions_before; clone_actions != NULL; clone_actions = clone_actions->next) { before = (pe_action_wrapper_t *) clone_actions->data; crm_trace("Testing '%s then %s' for %s", first->uuid, before->action->uuid, start->uuid); CRM_ASSERT(before->action->rsc != NULL); first_rsc = before->action->rsc; break; } } else if (!pcmk__str_eq(first->task, PCMK_ACTION_START, pcmk__str_none)) { crm_trace("Not a start op %s for %s", first->uuid, start->uuid); } if (first_rsc == NULL) { continue; } else if (pe__const_top_resource(first_rsc, false) == pe__const_top_resource(start->rsc, false)) { crm_trace("Same parent %s for %s", first_rsc->id, start->uuid); continue; } else if (!pe_rsc_is_clone(pe__const_top_resource(first_rsc, false))) { crm_trace("Not a clone %s for %s", first_rsc->id, start->uuid); continue; } crm_err("Applying %s before %s %d", first->uuid, start->uuid, pe__const_top_resource(first_rsc, false)->variant); for (GList *probe_iter = probes; probe_iter != NULL; probe_iter = probe_iter->next) { pe_action_t *probe = (pe_action_t *) probe_iter->data; crm_err("Ordering %s before %s", first->uuid, probe->uuid); order_actions(first, probe, pcmk__ar_ordered); } } } #endif } void pcmk__order_probes(pe_working_set_t *data_set) { // Add orderings for "probe then X" g_list_foreach(data_set->resources, add_start_restart_orderings_for_rsc, NULL); add_probe_orderings_for_stops(data_set); order_then_probes(data_set); } /*! * \internal * \brief Schedule any probes needed * * \param[in,out] data_set Cluster working set * * \note This may also schedule fencing of failed remote nodes. */ void pcmk__schedule_probes(pe_working_set_t *data_set) { // Schedule probes on each node in the cluster as needed for (GList *iter = data_set->nodes; iter != NULL; iter = iter->next) { pe_node_t *node = (pe_node_t *) iter->data; const char *probed = NULL; if (!node->details->online) { // Don't probe offline nodes if (pcmk__is_failed_remote_node(node)) { pe_fence_node(data_set, node, "the connection is unrecoverable", FALSE); } continue; } else if (node->details->unclean) { // ... or nodes that need fencing continue; } else if (!node->details->rsc_discovery_enabled) { // The user requested that probes not be done on this node continue; } /* This is no longer needed for live clusters, since the probe_complete * node attribute will never be in the CIB. However this is still useful * for processing old saved CIBs (< 1.1.14), including the * reprobe-target_rc regression test. */ probed = pe_node_attribute_raw(node, CRM_OP_PROBED); if (probed != NULL && crm_is_true(probed) == FALSE) { pe_action_t *probe_op = NULL; probe_op = custom_action(NULL, crm_strdup_printf("%s-%s", CRM_OP_REPROBE, node->details->uname), CRM_OP_REPROBE, node, FALSE, TRUE, data_set); add_hash_param(probe_op->meta, XML_ATTR_TE_NOWAIT, XML_BOOLEAN_TRUE); continue; } // Probe each resource in the cluster on this node, as needed pcmk__probe_resource_list(data_set->resources, node); } } diff --git a/lib/pacemaker/pcmk_sched_recurring.c b/lib/pacemaker/pcmk_sched_recurring.c index f8cf010065..9eb827af0b 100644 --- a/lib/pacemaker/pcmk_sched_recurring.c +++ b/lib/pacemaker/pcmk_sched_recurring.c @@ -1,719 +1,724 @@ /* * Copyright 2004-2023 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU General Public License version 2 * or later (GPLv2+) WITHOUT ANY WARRANTY. */ #include #include #include #include #include #include "libpacemaker_private.h" // Information parsed from an operation history entry in the CIB struct op_history { // XML attributes const char *id; // ID of history entry const char *name; // Action name // Parsed information char *key; // Operation key for action enum rsc_role_e role; // Action role (or pcmk_role_unknown for default) guint interval_ms; // Action interval }; /*! * \internal * \brief Parse an interval from XML * * \param[in] xml XML containing an interval attribute * * \return Interval parsed from XML (or 0 as default) */ static guint xe_interval(const xmlNode *xml) { return crm_parse_interval_spec(crm_element_value(xml, XML_LRM_ATTR_INTERVAL)); } /*! * \internal * \brief Check whether an operation exists multiple times in resource history * * \param[in] rsc Resource with history to search * \param[in] name Name of action to search for * \param[in] interval_ms Interval (in milliseconds) of action to search for * * \return true if an operation with \p name and \p interval_ms exists more than * once in the operation history of \p rsc, otherwise false */ static bool is_op_dup(const pe_resource_t *rsc, const char *name, guint interval_ms) { const char *id = NULL; for (xmlNode *op = first_named_child(rsc->ops_xml, "op"); op != NULL; op = crm_next_same_xml(op)) { // Check whether action name and interval match if (!pcmk__str_eq(crm_element_value(op, "name"), name, pcmk__str_none) || (xe_interval(op) != interval_ms)) { continue; } if (ID(op) == NULL) { continue; // Shouldn't be possible } if (id == NULL) { id = ID(op); // First matching op } else { pcmk__config_err("Operation %s is duplicate of %s (do not use " "same name and interval combination more " "than once per resource)", ID(op), id); return true; } } return false; } /*! * \internal * \brief Check whether an action name is one that can be recurring * * \param[in] name Action name to check * * \return true if \p name is an action known to be unsuitable as a recurring * operation, otherwise false * * \note Pacemaker's current philosophy is to allow users to configure recurring * operations except for a short list of actions known not to be suitable * for that (as opposed to allowing only actions known to be suitable, * which includes only monitor). Among other things, this approach allows * users to define their own custom operations and make them recurring, * though that use case is not well tested. */ static bool op_cannot_recur(const char *name) { return pcmk__str_any_of(name, PCMK_ACTION_STOP, PCMK_ACTION_START, PCMK_ACTION_DEMOTE, PCMK_ACTION_PROMOTE, PCMK_ACTION_RELOAD_AGENT, PCMK_ACTION_MIGRATE_TO, PCMK_ACTION_MIGRATE_FROM, NULL); } /*! * \internal * \brief Check whether a resource history entry is for a recurring action * * \param[in] rsc Resource that history entry is for * \param[in] xml XML of resource history entry to check * \param[out] op Where to store parsed info if recurring * * \return true if \p xml is for a recurring action, otherwise false */ static bool is_recurring_history(const pe_resource_t *rsc, const xmlNode *xml, struct op_history *op) { const char *role = NULL; op->interval_ms = xe_interval(xml); if (op->interval_ms == 0) { return false; // Not recurring } op->id = ID(xml); if (pcmk__str_empty(op->id)) { pcmk__config_err("Ignoring resource history entry without ID"); return false; // Shouldn't be possible (unless CIB was manually edited) } op->name = crm_element_value(xml, "name"); if (op_cannot_recur(op->name)) { pcmk__config_err("Ignoring %s because %s action cannot be recurring", op->id, pcmk__s(op->name, "unnamed")); return false; } // There should only be one recurring operation per action/interval if (is_op_dup(rsc, op->name, op->interval_ms)) { return false; } // Ensure role is valid if specified role = crm_element_value(xml, "role"); if (role == NULL) { op->role = pcmk_role_unknown; } else { op->role = text2role(role); if (op->role == pcmk_role_unknown) { pcmk__config_err("Ignoring %s because %s is not a valid role", op->id, role); } } // Disabled resources don't get monitored op->key = pcmk__op_key(rsc->id, op->name, op->interval_ms); if (find_rsc_op_entry(rsc, op->key) == NULL) { crm_trace("Not creating recurring action %s for disabled resource %s", op->id, rsc->id); free(op->key); return false; } return true; } /*! * \internal * \brief Check whether a recurring action for an active role should be optional * * \param[in] rsc Resource that recurring action is for * \param[in] node Node that \p rsc will be active on (if any) * \param[in] key Operation key for recurring action to check * \param[in,out] start Start action for \p rsc * * \return true if recurring action should be optional, otherwise false */ static bool active_recurring_should_be_optional(const pe_resource_t *rsc, const pe_node_t *node, const char *key, pe_action_t *start) { GList *possible_matches = NULL; if (node == NULL) { // Should only be possible if unmanaged and stopped pe_rsc_trace(rsc, "%s will be mandatory because resource is unmanaged", key); return false; } if (!pcmk_is_set(rsc->cmds->action_flags(start, NULL), pcmk_action_optional)) { pe_rsc_trace(rsc, "%s will be mandatory because %s is", key, start->uuid); return false; } possible_matches = find_actions_exact(rsc->actions, key, node); if (possible_matches == NULL) { pe_rsc_trace(rsc, "%s will be mandatory because it is not active on %s", key, pe__node_name(node)); return false; } for (const GList *iter = possible_matches; iter != NULL; iter = iter->next) { const pe_action_t *op = (const pe_action_t *) iter->data; if (pcmk_is_set(op->flags, pcmk_action_reschedule)) { pe_rsc_trace(rsc, "%s will be mandatory because " "it needs to be rescheduled", key); g_list_free(possible_matches); return false; } } g_list_free(possible_matches); return true; } /*! * \internal * \brief Create recurring action from resource history entry for an active role * * \param[in,out] rsc Resource that resource history is for * \param[in,out] start Start action for \p rsc on \p node * \param[in] node Node that resource will be active on (if any) * \param[in] op Resource history entry */ static void recurring_op_for_active(pe_resource_t *rsc, pe_action_t *start, const pe_node_t *node, const struct op_history *op) { pe_action_t *mon = NULL; bool is_optional = true; const bool is_default_role = (op->role == pcmk_role_unknown); // We're only interested in recurring actions for active roles if (op->role == pcmk_role_stopped) { return; } is_optional = active_recurring_should_be_optional(rsc, node, op->key, start); if ((!is_default_role && (rsc->next_role != op->role)) || (is_default_role && (rsc->next_role == pcmk_role_promoted))) { // Configured monitor role doesn't match role resource will have if (is_optional) { // It's running, so cancel it char *after_key = NULL; pe_action_t *cancel_op = pcmk__new_cancel_action(rsc, op->name, op->interval_ms, node); switch (rsc->role) { case pcmk_role_unpromoted: case pcmk_role_started: if (rsc->next_role == pcmk_role_promoted) { after_key = promote_key(rsc); } else if (rsc->next_role == pcmk_role_stopped) { after_key = stop_key(rsc); } break; case pcmk_role_promoted: after_key = demote_key(rsc); break; default: break; } if (after_key) { pcmk__new_ordering(rsc, NULL, cancel_op, rsc, after_key, NULL, - pe_order_runnable_left, rsc->cluster); + pcmk__ar_unrunnable_first_blocks, + rsc->cluster); } } do_crm_log((is_optional? LOG_INFO : LOG_TRACE), "%s recurring action %s because %s configured for %s role " "(not %s)", (is_optional? "Cancelling" : "Ignoring"), op->key, op->id, role2text(is_default_role? pcmk_role_unpromoted : op->role), role2text(rsc->next_role)); return; } pe_rsc_trace(rsc, "Creating %s recurring action %s for %s (%s %s on %s)", (is_optional? "optional" : "mandatory"), op->key, op->id, rsc->id, role2text(rsc->next_role), pe__node_name(node)); mon = custom_action(rsc, strdup(op->key), op->name, node, is_optional, TRUE, rsc->cluster); if (!pcmk_is_set(start->flags, pcmk_action_runnable)) { pe_rsc_trace(rsc, "%s is unrunnable because start is", mon->uuid); pe__clear_action_flags(mon, pcmk_action_runnable); } else if ((node == NULL) || !node->details->online || node->details->unclean) { pe_rsc_trace(rsc, "%s is unrunnable because no node is available", mon->uuid); pe__clear_action_flags(mon, pcmk_action_runnable); } else if (!pcmk_is_set(mon->flags, pcmk_action_optional)) { pe_rsc_info(rsc, "Start %s-interval %s for %s on %s", pcmk__readable_interval(op->interval_ms), mon->task, rsc->id, pe__node_name(node)); } if (rsc->next_role == pcmk_role_promoted) { pe__add_action_expected_result(mon, CRM_EX_PROMOTED); } // Order monitor relative to other actions if ((node == NULL) || pcmk_is_set(rsc->flags, pcmk_rsc_managed)) { pcmk__new_ordering(rsc, start_key(rsc), NULL, NULL, strdup(mon->uuid), mon, - pcmk__ar_first_implies_then|pe_order_runnable_left, + pcmk__ar_first_implies_then + |pcmk__ar_unrunnable_first_blocks, rsc->cluster); pcmk__new_ordering(rsc, reload_key(rsc), NULL, NULL, strdup(mon->uuid), mon, - pcmk__ar_first_implies_then|pe_order_runnable_left, + pcmk__ar_first_implies_then + |pcmk__ar_unrunnable_first_blocks, rsc->cluster); if (rsc->next_role == pcmk_role_promoted) { pcmk__new_ordering(rsc, promote_key(rsc), NULL, rsc, NULL, mon, - pcmk__ar_ordered|pe_order_runnable_left, + pcmk__ar_ordered + |pcmk__ar_unrunnable_first_blocks, rsc->cluster); } else if (rsc->role == pcmk_role_promoted) { pcmk__new_ordering(rsc, demote_key(rsc), NULL, rsc, NULL, mon, - pcmk__ar_ordered|pe_order_runnable_left, + pcmk__ar_ordered + |pcmk__ar_unrunnable_first_blocks, rsc->cluster); } } } /*! * \internal * \brief Cancel a recurring action if running on a node * * \param[in,out] rsc Resource that action is for * \param[in] node Node to cancel action on * \param[in] key Operation key for action * \param[in] name Action name * \param[in] interval_ms Action interval (in milliseconds) */ static void cancel_if_running(pe_resource_t *rsc, const pe_node_t *node, const char *key, const char *name, guint interval_ms) { GList *possible_matches = find_actions_exact(rsc->actions, key, node); pe_action_t *cancel_op = NULL; if (possible_matches == NULL) { return; // Recurring action isn't running on this node } g_list_free(possible_matches); cancel_op = pcmk__new_cancel_action(rsc, name, interval_ms, node); switch (rsc->next_role) { case pcmk_role_started: case pcmk_role_unpromoted: /* Order starts after cancel. If the current role is * stopped, this cancels the monitor before the resource * starts; if the current role is started, then this cancels * the monitor on a migration target before starting there. */ pcmk__new_ordering(rsc, NULL, cancel_op, rsc, start_key(rsc), NULL, - pe_order_runnable_left, rsc->cluster); + pcmk__ar_unrunnable_first_blocks, rsc->cluster); break; default: break; } pe_rsc_info(rsc, "Cancelling %s-interval %s action for %s on %s because " "configured for " PCMK__ROLE_STOPPED " role (not %s)", pcmk__readable_interval(interval_ms), name, rsc->id, pe__node_name(node), role2text(rsc->next_role)); } /*! * \internal * \brief Order an action after all probes of a resource on a node * * \param[in,out] rsc Resource to check for probes * \param[in] node Node to check for probes of \p rsc * \param[in,out] action Action to order after probes of \p rsc on \p node */ static void order_after_probes(pe_resource_t *rsc, const pe_node_t *node, pe_action_t *action) { GList *probes = pe__resource_actions(rsc, node, PCMK_ACTION_MONITOR, FALSE); for (GList *iter = probes; iter != NULL; iter = iter->next) { order_actions((pe_action_t *) iter->data, action, - pe_order_runnable_left); + pcmk__ar_unrunnable_first_blocks); } g_list_free(probes); } /*! * \internal * \brief Order an action after all stops of a resource on a node * * \param[in,out] rsc Resource to check for stops * \param[in] node Node to check for stops of \p rsc * \param[in,out] action Action to order after stops of \p rsc on \p node */ static void order_after_stops(pe_resource_t *rsc, const pe_node_t *node, pe_action_t *action) { GList *stop_ops = pe__resource_actions(rsc, node, PCMK_ACTION_STOP, TRUE); for (GList *iter = stop_ops; iter != NULL; iter = iter->next) { pe_action_t *stop = (pe_action_t *) iter->data; if (!pcmk_is_set(stop->flags, pcmk_action_optional) && !pcmk_is_set(action->flags, pcmk_action_optional) && !pcmk_is_set(rsc->flags, pcmk_rsc_managed)) { pe_rsc_trace(rsc, "%s optional on %s: unmanaged", action->uuid, pe__node_name(node)); pe__set_action_flags(action, pcmk_action_optional); } if (!pcmk_is_set(stop->flags, pcmk_action_runnable)) { crm_debug("%s unrunnable on %s: stop is unrunnable", action->uuid, pe__node_name(node)); pe__clear_action_flags(action, pcmk_action_runnable); } if (pcmk_is_set(rsc->flags, pcmk_rsc_managed)) { pcmk__new_ordering(rsc, stop_key(rsc), stop, NULL, NULL, action, pcmk__ar_first_implies_then - |pe_order_runnable_left, + |pcmk__ar_unrunnable_first_blocks, rsc->cluster); } } g_list_free(stop_ops); } /*! * \internal * \brief Create recurring action from resource history entry for inactive role * * \param[in,out] rsc Resource that resource history is for * \param[in] node Node that resource will be active on (if any) * \param[in] op Resource history entry */ static void recurring_op_for_inactive(pe_resource_t *rsc, const pe_node_t *node, const struct op_history *op) { GList *possible_matches = NULL; // We're only interested in recurring actions for the inactive role if (op->role != pcmk_role_stopped) { return; } if (!pcmk_is_set(rsc->flags, pcmk_rsc_unique)) { crm_notice("Ignoring %s (recurring monitors for " PCMK__ROLE_STOPPED " role are not supported for anonymous clones)", op->id); return; // @TODO add support } pe_rsc_trace(rsc, "Creating recurring action %s for %s on nodes " "where it should not be running", op->id, rsc->id); for (GList *iter = rsc->cluster->nodes; iter != NULL; iter = iter->next) { pe_node_t *stop_node = (pe_node_t *) iter->data; bool is_optional = true; pe_action_t *stopped_mon = NULL; // Cancel action on node where resource will be active if ((node != NULL) && pcmk__str_eq(stop_node->details->uname, node->details->uname, pcmk__str_casei)) { cancel_if_running(rsc, node, op->key, op->name, op->interval_ms); continue; } // Recurring action on this node is optional if it's already active here possible_matches = find_actions_exact(rsc->actions, op->key, stop_node); is_optional = (possible_matches != NULL); g_list_free(possible_matches); pe_rsc_trace(rsc, "Creating %s recurring action %s for %s (%s " PCMK__ROLE_STOPPED " on %s)", (is_optional? "optional" : "mandatory"), op->key, op->id, rsc->id, pe__node_name(stop_node)); stopped_mon = custom_action(rsc, strdup(op->key), op->name, stop_node, is_optional, TRUE, rsc->cluster); pe__add_action_expected_result(stopped_mon, CRM_EX_NOT_RUNNING); if (pcmk_is_set(rsc->flags, pcmk_rsc_managed)) { order_after_probes(rsc, stop_node, stopped_mon); } /* The recurring action is for the inactive role, so it shouldn't be * performed until the resource is inactive. */ order_after_stops(rsc, stop_node, stopped_mon); if (!stop_node->details->online || stop_node->details->unclean) { pe_rsc_debug(rsc, "%s unrunnable on %s: node unavailable)", stopped_mon->uuid, pe__node_name(stop_node)); pe__clear_action_flags(stopped_mon, pcmk_action_runnable); } if (pcmk_is_set(stopped_mon->flags, pcmk_action_runnable) && !pcmk_is_set(stopped_mon->flags, pcmk_action_optional)) { crm_notice("Start recurring %s-interval %s for " PCMK__ROLE_STOPPED " %s on %s", pcmk__readable_interval(op->interval_ms), stopped_mon->task, rsc->id, pe__node_name(stop_node)); } } } /*! * \internal * \brief Create recurring actions for a resource * * \param[in,out] rsc Resource to create recurring actions for */ void pcmk__create_recurring_actions(pe_resource_t *rsc) { pe_action_t *start = NULL; if (pcmk_is_set(rsc->flags, pcmk_rsc_blocked)) { pe_rsc_trace(rsc, "Skipping recurring actions for blocked resource %s", rsc->id); return; } if (pcmk_is_set(rsc->flags, pcmk_rsc_maintenance)) { pe_rsc_trace(rsc, "Skipping recurring actions for %s " "in maintenance mode", rsc->id); return; } if (rsc->allocated_to == NULL) { // Recurring actions for active roles not needed } else if (rsc->allocated_to->details->maintenance) { pe_rsc_trace(rsc, "Skipping recurring actions for %s on %s " "in maintenance mode", rsc->id, pe__node_name(rsc->allocated_to)); } else if ((rsc->next_role != pcmk_role_stopped) || !pcmk_is_set(rsc->flags, pcmk_rsc_managed)) { // Recurring actions for active roles needed start = start_action(rsc, rsc->allocated_to, TRUE); } pe_rsc_trace(rsc, "Creating any recurring actions needed for %s", rsc->id); for (xmlNode *op = first_named_child(rsc->ops_xml, "op"); op != NULL; op = crm_next_same_xml(op)) { struct op_history op_history = { NULL, }; if (!is_recurring_history(rsc, op, &op_history)) { continue; } if (start != NULL) { recurring_op_for_active(rsc, start, rsc->allocated_to, &op_history); } recurring_op_for_inactive(rsc, rsc->allocated_to, &op_history); free(op_history.key); } } /*! * \internal * \brief Create an executor cancel action * * \param[in,out] rsc Resource of action to cancel * \param[in] task Name of action to cancel * \param[in] interval_ms Interval of action to cancel * \param[in] node Node of action to cancel * * \return Created op */ pe_action_t * pcmk__new_cancel_action(pe_resource_t *rsc, const char *task, guint interval_ms, const pe_node_t *node) { pe_action_t *cancel_op = NULL; char *key = NULL; char *interval_ms_s = NULL; CRM_ASSERT((rsc != NULL) && (task != NULL) && (node != NULL)); // @TODO dangerous if possible to schedule another action with this key key = pcmk__op_key(rsc->id, task, interval_ms); cancel_op = custom_action(rsc, key, PCMK_ACTION_CANCEL, node, FALSE, TRUE, rsc->cluster); pcmk__str_update(&cancel_op->task, PCMK_ACTION_CANCEL); pcmk__str_update(&cancel_op->cancel_task, task); interval_ms_s = crm_strdup_printf("%u", interval_ms); add_hash_param(cancel_op->meta, XML_LRM_ATTR_TASK, task); add_hash_param(cancel_op->meta, XML_LRM_ATTR_INTERVAL_MS, interval_ms_s); free(interval_ms_s); return cancel_op; } /*! * \internal * \brief Schedule cancellation of a recurring action * * \param[in,out] rsc Resource that action is for * \param[in] call_id Action's call ID from history * \param[in] task Action name * \param[in] interval_ms Action interval * \param[in] node Node that history entry is for * \param[in] reason Short description of why action is cancelled */ void pcmk__schedule_cancel(pe_resource_t *rsc, const char *call_id, const char *task, guint interval_ms, const pe_node_t *node, const char *reason) { pe_action_t *cancel = NULL; CRM_CHECK((rsc != NULL) && (task != NULL) && (node != NULL) && (reason != NULL), return); crm_info("Recurring %s-interval %s for %s will be stopped on %s: %s", pcmk__readable_interval(interval_ms), task, rsc->id, pe__node_name(node), reason); cancel = pcmk__new_cancel_action(rsc, task, interval_ms, node); add_hash_param(cancel->meta, XML_LRM_ATTR_CALLID, call_id); // Cancellations happen after stops pcmk__new_ordering(rsc, stop_key(rsc), NULL, rsc, NULL, cancel, pcmk__ar_ordered, rsc->cluster); } /*! * \internal * \brief Create a recurring action marked as needing rescheduling if active * * \param[in,out] rsc Resource that action is for * \param[in] task Name of action being rescheduled * \param[in] interval_ms Action interval (in milliseconds) * \param[in,out] node Node where action should be rescheduled */ void pcmk__reschedule_recurring(pe_resource_t *rsc, const char *task, guint interval_ms, pe_node_t *node) { pe_action_t *op = NULL; trigger_unfencing(rsc, node, "Device parameters changed (reschedule)", NULL, rsc->cluster); op = custom_action(rsc, pcmk__op_key(rsc->id, task, interval_ms), task, node, TRUE, TRUE, rsc->cluster); pe__set_action_flags(op, pcmk_action_reschedule); } /*! * \internal * \brief Check whether an action is recurring * * \param[in] action Action to check * * \return true if \p action has a nonzero interval, otherwise false */ bool pcmk__action_is_recurring(const pe_action_t *action) { guint interval_ms = 0; if (pcmk__guint_from_hash(action->meta, XML_LRM_ATTR_INTERVAL_MS, 0, &interval_ms) != pcmk_rc_ok) { return false; } return (interval_ms > 0); } diff --git a/lib/pacemaker/pcmk_sched_remote.c b/lib/pacemaker/pcmk_sched_remote.c index af6aba8197..57f8655a6f 100644 --- a/lib/pacemaker/pcmk_sched_remote.c +++ b/lib/pacemaker/pcmk_sched_remote.c @@ -1,726 +1,728 @@ /* * Copyright 2004-2023 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU General Public License version 2 * or later (GPLv2+) WITHOUT ANY WARRANTY. */ #include #include #include #include #include #include #include #include #include #include #include "libpacemaker_private.h" enum remote_connection_state { remote_state_unknown = 0, remote_state_alive = 1, remote_state_resting = 2, remote_state_failed = 3, remote_state_stopped = 4 }; static const char * state2text(enum remote_connection_state state) { switch (state) { case remote_state_unknown: return "unknown"; case remote_state_alive: return "alive"; case remote_state_resting: return "resting"; case remote_state_failed: return "failed"; case remote_state_stopped: return "stopped"; } return "impossible"; } /* We always use pe_order_preserve with these convenience functions to exempt * internally generated constraints from the prohibition of user constraints * involving remote connection resources. * - * The start ordering additionally uses pe_order_runnable_left so that the - * specified action is not runnable if the start is not runnable. + * The start ordering additionally uses pcmk__ar_unrunnable_first_blocks so that + * the specified action is not runnable if the start is not runnable. */ static inline void order_start_then_action(pe_resource_t *first_rsc, pe_action_t *then_action, uint32_t extra) { if ((first_rsc != NULL) && (then_action != NULL)) { pcmk__new_ordering(first_rsc, start_key(first_rsc), NULL, then_action->rsc, NULL, then_action, - pe_order_preserve|pe_order_runnable_left|extra, + pe_order_preserve + |pcmk__ar_unrunnable_first_blocks + |extra, first_rsc->cluster); } } static inline void order_action_then_stop(pe_action_t *first_action, pe_resource_t *then_rsc, uint32_t extra) { if ((first_action != NULL) && (then_rsc != NULL)) { pcmk__new_ordering(first_action->rsc, NULL, first_action, then_rsc, stop_key(then_rsc), NULL, pe_order_preserve|extra, then_rsc->cluster); } } static enum remote_connection_state get_remote_node_state(const pe_node_t *node) { const pe_resource_t *remote_rsc = NULL; const pe_node_t *cluster_node = NULL; CRM_ASSERT(node != NULL); remote_rsc = node->details->remote_rsc; CRM_ASSERT(remote_rsc != NULL); cluster_node = pe__current_node(remote_rsc); /* If the cluster node the remote connection resource resides on * is unclean or went offline, we can't process any operations * on that remote node until after it starts elsewhere. */ if ((remote_rsc->next_role == pcmk_role_stopped) || (remote_rsc->allocated_to == NULL)) { // The connection resource is not going to run anywhere if ((cluster_node != NULL) && cluster_node->details->unclean) { /* The remote connection is failed because its resource is on a * failed node and can't be recovered elsewhere, so we must fence. */ return remote_state_failed; } if (!pcmk_is_set(remote_rsc->flags, pcmk_rsc_failed)) { /* Connection resource is cleanly stopped */ return remote_state_stopped; } /* Connection resource is failed */ if ((remote_rsc->next_role == pcmk_role_stopped) && remote_rsc->remote_reconnect_ms && node->details->remote_was_fenced && !pe__shutdown_requested(node)) { /* We won't know whether the connection is recoverable until the * reconnect interval expires and we reattempt connection. */ return remote_state_unknown; } /* The remote connection is in a failed state. If there are any * resources known to be active on it (stop) or in an unknown state * (probe), we must assume the worst and fence it. */ return remote_state_failed; } else if (cluster_node == NULL) { /* Connection is recoverable but not currently running anywhere, so see * if we can recover it first */ return remote_state_unknown; } else if (cluster_node->details->unclean || !(cluster_node->details->online)) { // Connection is running on a dead node, see if we can recover it first return remote_state_resting; } else if (pcmk__list_of_multiple(remote_rsc->running_on) && (remote_rsc->partial_migration_source != NULL) && (remote_rsc->partial_migration_target != NULL)) { /* We're in the middle of migrating a connection resource, so wait until * after the migration completes before performing any actions. */ return remote_state_resting; } return remote_state_alive; } /*! * \internal * \brief Order actions on remote node relative to actions for the connection * * \param[in,out] action An action scheduled on a Pacemaker Remote node */ static void apply_remote_ordering(pe_action_t *action) { pe_resource_t *remote_rsc = NULL; enum action_tasks task = text2task(action->task); enum remote_connection_state state = get_remote_node_state(action->node); uint32_t order_opts = pcmk__ar_none; if (action->rsc == NULL) { return; } CRM_ASSERT(pe__is_guest_or_remote_node(action->node)); remote_rsc = action->node->details->remote_rsc; CRM_ASSERT(remote_rsc != NULL); crm_trace("Order %s action %s relative to %s%s (state: %s)", action->task, action->uuid, pcmk_is_set(remote_rsc->flags, pcmk_rsc_failed)? "failed " : "", remote_rsc->id, state2text(state)); if (pcmk__strcase_any_of(action->task, PCMK_ACTION_MIGRATE_TO, PCMK_ACTION_MIGRATE_FROM, NULL)) { /* Migration ops map to pcmk_action_unspecified, but we need to apply * the same ordering as for stop or demote (see get_router_node()). */ task = pcmk_action_stop; } switch (task) { case pcmk_action_start: case pcmk_action_promote: order_opts = pcmk__ar_none; if (state == remote_state_failed) { /* Force recovery, by making this action required */ pe__set_order_flags(order_opts, pcmk__ar_first_implies_then); } /* Ensure connection is up before running this action */ order_start_then_action(remote_rsc, action, order_opts); break; case pcmk_action_stop: if (state == remote_state_alive) { order_action_then_stop(action, remote_rsc, pcmk__ar_then_implies_first); } else if (state == remote_state_failed) { /* The resource is active on the node, but since we don't have a * valid connection, the only way to stop the resource is by * fencing the node. There is no need to order the stop relative * to the remote connection, since the stop will become implied * by the fencing. */ pe_fence_node(remote_rsc->cluster, action->node, "resources are active but " "connection is unrecoverable", FALSE); } else if (remote_rsc->next_role == pcmk_role_stopped) { /* State must be remote_state_unknown or remote_state_stopped. * Since the connection is not coming back up in this * transition, stop this resource first. */ order_action_then_stop(action, remote_rsc, pcmk__ar_then_implies_first); } else { /* The connection is going to be started somewhere else, so * stop this resource after that completes. */ order_start_then_action(remote_rsc, action, pcmk__ar_none); } break; case pcmk_action_demote: /* Only order this demote relative to the connection start if the * connection isn't being torn down. Otherwise, the demote would be * blocked because the connection start would not be allowed. */ if ((state == remote_state_resting) || (state == remote_state_unknown)) { order_start_then_action(remote_rsc, action, pcmk__ar_none); } /* Otherwise we can rely on the stop ordering */ break; default: /* Wait for the connection resource to be up */ if (pcmk__action_is_recurring(action)) { /* In case we ever get the recovery logic wrong, force * recurring monitors to be restarted, even if just * the connection was re-established */ order_start_then_action(remote_rsc, action, pcmk__ar_first_implies_then); } else { pe_node_t *cluster_node = pe__current_node(remote_rsc); if ((task == pcmk_action_monitor) && (state == remote_state_failed)) { /* We would only be here if we do not know the state of the * resource on the remote node. Since we have no way to find * out, it is necessary to fence the node. */ pe_fence_node(remote_rsc->cluster, action->node, "resources are in unknown state " "and connection is unrecoverable", FALSE); } if ((cluster_node != NULL) && (state == remote_state_stopped)) { /* The connection is currently up, but is going down * permanently. Make sure we check services are actually * stopped _before_ we let the connection get closed. */ order_action_then_stop(action, remote_rsc, - pe_order_runnable_left); + pcmk__ar_unrunnable_first_blocks); } else { order_start_then_action(remote_rsc, action, pcmk__ar_none); } } break; } } static void apply_container_ordering(pe_action_t *action) { /* VMs are also classified as containers for these purposes... in * that they both involve a 'thing' running on a real or remote * cluster node. * * This allows us to be smarter about the type and extent of * recovery actions required in various scenarios */ pe_resource_t *remote_rsc = NULL; pe_resource_t *container = NULL; enum action_tasks task = text2task(action->task); CRM_ASSERT(action->rsc != NULL); CRM_ASSERT(action->node != NULL); CRM_ASSERT(pe__is_guest_or_remote_node(action->node)); remote_rsc = action->node->details->remote_rsc; CRM_ASSERT(remote_rsc != NULL); container = remote_rsc->container; CRM_ASSERT(container != NULL); if (pcmk_is_set(container->flags, pcmk_rsc_failed)) { pe_fence_node(action->rsc->cluster, action->node, "container failed", FALSE); } crm_trace("Order %s action %s relative to %s%s for %s%s", action->task, action->uuid, pcmk_is_set(remote_rsc->flags, pcmk_rsc_failed)? "failed " : "", remote_rsc->id, pcmk_is_set(container->flags, pcmk_rsc_failed)? "failed " : "", container->id); if (pcmk__strcase_any_of(action->task, PCMK_ACTION_MIGRATE_TO, PCMK_ACTION_MIGRATE_FROM, NULL)) { /* Migration ops map to pcmk_action_unspecified, but we need to apply * the same ordering as for stop or demote (see get_router_node()). */ task = pcmk_action_stop; } switch (task) { case pcmk_action_start: case pcmk_action_promote: // Force resource recovery if the container is recovered order_start_then_action(container, action, pcmk__ar_first_implies_then); // Wait for the connection resource to be up, too order_start_then_action(remote_rsc, action, pcmk__ar_none); break; case pcmk_action_stop: case pcmk_action_demote: if (pcmk_is_set(container->flags, pcmk_rsc_failed)) { /* When the container representing a guest node fails, any stop * or demote actions for resources running on the guest node * are implied by the container stopping. This is similar to * how fencing operations work for cluster nodes and remote * nodes. */ } else { /* Ensure the operation happens before the connection is brought * down. * * If we really wanted to, we could order these after the * connection start, IFF the container's current role was * stopped (otherwise we re-introduce an ordering loop when the * connection is restarting). */ order_action_then_stop(action, remote_rsc, pcmk__ar_none); } break; default: /* Wait for the connection resource to be up */ if (pcmk__action_is_recurring(action)) { /* In case we ever get the recovery logic wrong, force * recurring monitors to be restarted, even if just * the connection was re-established */ if (task != pcmk_action_unspecified) { order_start_then_action(remote_rsc, action, pcmk__ar_first_implies_then); } } else { order_start_then_action(remote_rsc, action, pcmk__ar_none); } break; } } /*! * \internal * \brief Order all relevant actions relative to remote connection actions * * \param[in,out] data_set Cluster working set */ void pcmk__order_remote_connection_actions(pe_working_set_t *data_set) { if (!pcmk_is_set(data_set->flags, pcmk_sched_have_remote_nodes)) { return; } crm_trace("Creating remote connection orderings"); for (GList *iter = data_set->actions; iter != NULL; iter = iter->next) { pe_action_t *action = iter->data; pe_resource_t *remote = NULL; // We are only interested in resource actions if (action->rsc == NULL) { continue; } /* Special case: If we are clearing the failcount of an actual * remote connection resource, then make sure this happens before * any start of the resource in this transition. */ if (action->rsc->is_remote_node && pcmk__str_eq(action->task, PCMK_ACTION_CLEAR_FAILCOUNT, pcmk__str_none)) { pcmk__new_ordering(action->rsc, NULL, action, action->rsc, pcmk__op_key(action->rsc->id, PCMK_ACTION_START, 0), NULL, pcmk__ar_ordered, data_set); continue; } // We are only interested in actions assigned to a node if (action->node == NULL) { continue; } if (!pe__is_guest_or_remote_node(action->node)) { continue; } /* We are only interested in real actions. * * @TODO This is probably wrong; pseudo-actions might be converted to * real actions and vice versa later in update_actions() at the end of * pcmk__apply_orderings(). */ if (pcmk_is_set(action->flags, pcmk_action_pseudo)) { continue; } remote = action->node->details->remote_rsc; if (remote == NULL) { // Orphaned continue; } /* Another special case: if a resource is moving to a Pacemaker Remote * node, order the stop on the original node after any start of the * remote connection. This ensures that if the connection fails to * start, we leave the resource running on the original node. */ if (pcmk__str_eq(action->task, PCMK_ACTION_START, pcmk__str_none)) { for (GList *item = action->rsc->actions; item != NULL; item = item->next) { pe_action_t *rsc_action = item->data; if (!pe__same_node(rsc_action->node, action->node) && pcmk__str_eq(rsc_action->task, PCMK_ACTION_STOP, pcmk__str_none)) { pcmk__new_ordering(remote, start_key(remote), NULL, action->rsc, NULL, rsc_action, pcmk__ar_ordered, data_set); } } } /* The action occurs across a remote connection, so create * ordering constraints that guarantee the action occurs while the node * is active (after start, before stop ... things like that). * * This is somewhat brittle in that we need to make sure the results of * this ordering are compatible with the result of get_router_node(). * It would probably be better to add XML_LRM_ATTR_ROUTER_NODE as part * of this logic rather than create_graph_action(). */ if (remote->container) { crm_trace("Container ordering for %s", action->uuid); apply_container_ordering(action); } else { crm_trace("Remote ordering for %s", action->uuid); apply_remote_ordering(action); } } } /*! * \internal * \brief Check whether a node is a failed remote node * * \param[in] node Node to check * * \return true if \p node is a failed remote node, false otherwise */ bool pcmk__is_failed_remote_node(const pe_node_t *node) { return pe__is_remote_node(node) && (node->details->remote_rsc != NULL) && (get_remote_node_state(node) == remote_state_failed); } /*! * \internal * \brief Check whether a given resource corresponds to a given node as guest * * \param[in] rsc Resource to check * \param[in] node Node to check * * \return true if \p node is a guest node and \p rsc is its containing * resource, otherwise false */ bool pcmk__rsc_corresponds_to_guest(const pe_resource_t *rsc, const pe_node_t *node) { return (rsc != NULL) && (rsc->fillers != NULL) && (node != NULL) && (node->details->remote_rsc != NULL) && (node->details->remote_rsc->container == rsc); } /*! * \internal * \brief Get proper connection host that a remote action must be routed through * * A remote connection resource might be starting, stopping, or migrating in the * same transition that an action needs to be executed on its Pacemaker Remote * node. Determine the proper node that the remote action should be routed * through. * * \param[in] action (Potentially remote) action to route * * \return Connection host that action should be routed through if remote, * otherwise NULL */ pe_node_t * pcmk__connection_host_for_action(const pe_action_t *action) { pe_node_t *began_on = NULL; pe_node_t *ended_on = NULL; bool partial_migration = false; const char *task = action->task; if (pcmk__str_eq(task, PCMK_ACTION_STONITH, pcmk__str_none) || !pe__is_guest_or_remote_node(action->node)) { return NULL; } CRM_ASSERT(action->node->details->remote_rsc != NULL); began_on = pe__current_node(action->node->details->remote_rsc); ended_on = action->node->details->remote_rsc->allocated_to; if (action->node->details->remote_rsc && (action->node->details->remote_rsc->container == NULL) && action->node->details->remote_rsc->partial_migration_target) { partial_migration = true; } if (began_on == NULL) { crm_trace("Routing %s for %s through remote connection's " "next node %s (starting)%s", action->task, (action->rsc? action->rsc->id : "no resource"), (ended_on? ended_on->details->uname : "none"), partial_migration? " (partial migration)" : ""); return ended_on; } if (ended_on == NULL) { crm_trace("Routing %s for %s through remote connection's " "current node %s (stopping)%s", action->task, (action->rsc? action->rsc->id : "no resource"), (began_on? began_on->details->uname : "none"), partial_migration? " (partial migration)" : ""); return began_on; } if (pe__same_node(began_on, ended_on)) { crm_trace("Routing %s for %s through remote connection's " "current node %s (not moving)%s", action->task, (action->rsc? action->rsc->id : "no resource"), (began_on? began_on->details->uname : "none"), partial_migration? " (partial migration)" : ""); return began_on; } /* If we get here, the remote connection is moving during this transition. * This means some actions for resources behind the connection will get * routed through the cluster node the connection resource is currently on, * and others are routed through the cluster node the connection will end up * on. */ if (pcmk__str_eq(task, PCMK_ACTION_NOTIFY, pcmk__str_none)) { task = g_hash_table_lookup(action->meta, "notify_operation"); } /* * Stop, demote, and migration actions must occur before the connection can * move (these actions are required before the remote resource can stop). In * this case, we know these actions have to be routed through the initial * cluster node the connection resource lived on before the move takes * place. * * The exception is a partial migration of a (non-guest) remote connection * resource; in that case, all actions (even these) will be ordered after * the connection's pseudo-start on the migration target, so the target is * the router node. */ if (pcmk__strcase_any_of(task, PCMK_ACTION_CANCEL, PCMK_ACTION_STOP, PCMK_ACTION_DEMOTE, PCMK_ACTION_MIGRATE_FROM, PCMK_ACTION_MIGRATE_TO, NULL) && !partial_migration) { crm_trace("Routing %s for %s through remote connection's " "current node %s (moving)%s", action->task, (action->rsc? action->rsc->id : "no resource"), (began_on? began_on->details->uname : "none"), partial_migration? " (partial migration)" : ""); return began_on; } /* Everything else (start, promote, monitor, probe, refresh, * clear failcount, delete, ...) must occur after the connection starts on * the node it is moving to. */ crm_trace("Routing %s for %s through remote connection's " "next node %s (moving)%s", action->task, (action->rsc? action->rsc->id : "no resource"), (ended_on? ended_on->details->uname : "none"), partial_migration? " (partial migration)" : ""); return ended_on; } /*! * \internal * \brief Replace remote connection's addr="#uname" with actual address * * REMOTE_CONTAINER_HACK: If a given resource is a remote connection resource * with its "addr" parameter set to "#uname", pull the actual value from the * parameters evaluated without a node (which was put there earlier in * pcmk__create_graph() when the bundle's expand() method was called). * * \param[in,out] rsc Resource to check * \param[in,out] params Resource parameters evaluated per node */ void pcmk__substitute_remote_addr(pe_resource_t *rsc, GHashTable *params) { const char *remote_addr = g_hash_table_lookup(params, XML_RSC_ATTR_REMOTE_RA_ADDR); if (pcmk__str_eq(remote_addr, "#uname", pcmk__str_none)) { GHashTable *base = pe_rsc_params(rsc, NULL, rsc->cluster); remote_addr = g_hash_table_lookup(base, XML_RSC_ATTR_REMOTE_RA_ADDR); if (remote_addr != NULL) { g_hash_table_insert(params, strdup(XML_RSC_ATTR_REMOTE_RA_ADDR), strdup(remote_addr)); } } } /*! * \brief Add special bundle meta-attributes to XML * * If a given action will be executed on a guest node (including a bundle), * add the special bundle meta-attribute "container-attribute-target" and * environment variable "physical_host" as XML attributes (using meta-attribute * naming). * * \param[in,out] args_xml XML to add attributes to * \param[in] action Action to check */ void pcmk__add_bundle_meta_to_xml(xmlNode *args_xml, const pe_action_t *action) { const pe_node_t *guest = action->node; const pe_node_t *host = NULL; enum action_tasks task; if (!pe__is_guest_node(guest)) { return; } task = text2task(action->task); if ((task == pcmk_action_notify) || (task == pcmk_action_notified)) { task = text2task(g_hash_table_lookup(action->meta, "notify_operation")); } switch (task) { case pcmk_action_stop: case pcmk_action_stopped: case pcmk_action_demote: case pcmk_action_demoted: // "Down" actions take place on guest's current host host = pe__current_node(guest->details->remote_rsc->container); break; case pcmk_action_start: case pcmk_action_started: case pcmk_action_monitor: case pcmk_action_promote: case pcmk_action_promoted: // "Up" actions take place on guest's next host host = guest->details->remote_rsc->container->allocated_to; break; default: break; } if (host != NULL) { hash2metafield((gpointer) XML_RSC_ATTR_TARGET, (gpointer) g_hash_table_lookup(action->rsc->meta, XML_RSC_ATTR_TARGET), (gpointer) args_xml); hash2metafield((gpointer) PCMK__ENV_PHYSICAL_HOST, (gpointer) host->details->uname, (gpointer) args_xml); } }