diff --git a/lib/pacemaker/libpacemaker_private.h b/lib/pacemaker/libpacemaker_private.h index 58435a6217..0e9cb25b62 100644 --- a/lib/pacemaker/libpacemaker_private.h +++ b/lib/pacemaker/libpacemaker_private.h @@ -1,1142 +1,1156 @@ /* - * Copyright 2021-2024 the Pacemaker project contributors + * Copyright 2021-2025 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__PACEMAKER_LIBPACEMAKER_PRIVATE__H #define PCMK__PACEMAKER_LIBPACEMAKER_PRIVATE__H /* This header is for the sole use of libpacemaker, so that functions can be * declared with G_GNUC_INTERNAL for efficiency. */ #include // NULL #include // uint32_t #include // bool, false #include // guint, gpointer, GList, GHashTable #include // xmlNode #include // pcmk_action_t, pcmk_node_t, etc. #include // pcmk__location_t, etc. #include // cib_t #include // lrmd_event_data_t #include // pe__const_top_resource(), etc. #include // pcmk_injections_t #include // pcmk__colocation_t #ifdef __cplusplus extern "C" { #endif // Colocation flags enum pcmk__coloc_flags { pcmk__coloc_none = 0U, // Primary is affected even if already active pcmk__coloc_influence = (1U << 0), // Colocation was explicitly configured in CIB pcmk__coloc_explicit = (1U << 1), }; // Flags to modify the behavior of add_colocated_node_scores() enum pcmk__coloc_select { // With no other flags, apply all "with this" colocations pcmk__coloc_select_default = 0, // Apply "this with" colocations instead of "with this" colocations pcmk__coloc_select_this_with = (1 << 0), // Apply only colocations with non-negative scores pcmk__coloc_select_nonnegative = (1 << 1), // Apply only colocations with at least one matching node pcmk__coloc_select_active = (1 << 2), }; // Flags the update_ordered_actions() method can return enum pcmk__updated { pcmk__updated_none = 0, // Nothing changed pcmk__updated_first = (1 << 0), // First action was updated pcmk__updated_then = (1 << 1), // Then action was updated }; #define pcmk__set_updated_flags(au_flags, action, flags_to_set) do { \ au_flags = pcmk__set_flags_as(__func__, __LINE__, \ LOG_TRACE, "Action update", \ (action)->uuid, au_flags, \ (flags_to_set), #flags_to_set); \ } while (0) #define pcmk__clear_updated_flags(au_flags, action, flags_to_clear) do { \ au_flags = pcmk__clear_flags_as(__func__, __LINE__, \ LOG_TRACE, "Action update", \ (action)->uuid, au_flags, \ (flags_to_clear), #flags_to_clear); \ } while (0) // Resource assignment methods struct pcmk__assignment_methods { /*! * \internal * \brief Assign a 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 (if \p rsc is not a * primitive, this applies to its primitive * descendants instead) * * \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. */ pcmk_node_t *(*assign)(pcmk_resource_t *rsc, const pcmk_node_t *prefer, bool stop_if_fail); /*! * \internal * \brief Create all actions needed for a given resource * * \param[in,out] rsc Resource to create actions for */ void (*create_actions)(pcmk_resource_t *rsc); /*! * \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 (*create_probe)(pcmk_resource_t *rsc, pcmk_node_t *node); /*! * \internal * \brief Create implicit constraints needed for a resource * * \param[in,out] rsc Resource to create implicit constraints for */ void (*internal_constraints)(pcmk_resource_t *rsc); /*! * \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 * * \return The score added to the dependent's priority */ int (*apply_coloc_score)(pcmk_resource_t *dependent, const pcmk_resource_t *primary, const pcmk__colocation_t *colocation, bool for_dependent); /*! * \internal * \brief Create list of all resources in colocations with a given resource * * Given a resource, create a list of all resources involved in mandatory * colocations with it, whether directly or via chained colocations. * * \param[in] rsc Resource to add to colocated list * \param[in] orig_rsc Resource originally requested * \param[in,out] colocated_rscs Existing list * * \return List of given resource and all resources involved in colocations * * \note This function is recursive; top-level callers should pass NULL as * \p colocated_rscs and \p orig_rsc, and the desired resource as * \p rsc. The recursive calls will use other values. */ GList *(*colocated_resources)(const pcmk_resource_t *rsc, const pcmk_resource_t *orig_rsc, GList *colocated_rscs); /*! * \internal * \brief Add colocations affecting a resource as primary to a list * * Given a resource being assigned (\p orig_rsc) and a resource somewhere in * its chain of ancestors (\p rsc, which may be \p orig_rsc), get * colocations that affect the ancestor as primary and should affect the * resource, and add them to a given list. * * \param[in] rsc Resource whose colocations should be added * \param[in] orig_rsc Affected resource (\p rsc or a descendant) * \param[in,out] list List of colocations to add to * * \note All arguments should be non-NULL. * \note The pcmk__with_this_colocations() wrapper should usually be used * instead of using this method directly. */ void (*with_this_colocations)(const pcmk_resource_t *rsc, const pcmk_resource_t *orig_rsc, GList **list); /*! * \internal * \brief Add colocations affecting a resource as dependent to a list * * Given a resource being assigned (\p orig_rsc) and a resource somewhere in * its chain of ancestors (\p rsc, which may be \p orig_rsc), get * colocations that affect the ancestor as dependent and should affect the * resource, and add them to a given list. * * * \param[in] rsc Resource whose colocations should be added * \param[in] orig_rsc Affected resource (\p rsc or a descendant) * \param[in,out] list List of colocations to add to * * \note All arguments should be non-NULL. * \note The pcmk__this_with_colocations() wrapper should usually be used * instead of using this method directly. */ void (*this_with_colocations)(const pcmk_resource_t *rsc, const pcmk_resource_t *orig_rsc, GList **list); /*! * \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 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. */ void (*add_colocated_node_scores)(pcmk_resource_t *source_rsc, const pcmk_resource_t *target_rsc, const char *log_id, GHashTable **nodes, const pcmk__colocation_t *colocation, float factor, uint32_t flags); /*! * \internal * \brief Apply a location constraint to a resource's allowed node scores * * \param[in,out] rsc Resource to apply constraint to * \param[in,out] location Location constraint to apply */ void (*apply_location)(pcmk_resource_t *rsc, pcmk__location_t *location); /*! * \internal * \brief Return action flags for a given 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 * \note For primitives, this will be the same as action->flags regardless * of node. For collective resources, the flags can differ due to * multiple instances possibly being involved. */ uint32_t (*action_flags)(pcmk_action_t *action, const pcmk_node_t *node); /*! * \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 * \param[in,out] scheduler Scheduler data * * \return Group of enum pcmk__updated flags indicating what was updated */ uint32_t (*update_ordered_actions)(pcmk_action_t *first, pcmk_action_t *then, const pcmk_node_t *node, uint32_t flags, uint32_t filter, uint32_t type, pcmk_scheduler_t *scheduler); /*! * \internal * \brief Output a summary of scheduled actions for a resource * * \param[in,out] rsc Resource to output actions for */ void (*output_actions)(pcmk_resource_t *rsc); /*! * \internal * \brief Add a resource's actions to the transition graph * * \param[in,out] rsc Resource whose actions should be added */ void (*add_actions_to_graph)(pcmk_resource_t *rsc); /*! * \internal * \brief Add meta-attributes relevant to transition graph actions to XML * * If a given resource supports variant-specific meta-attributes that are * needed for transition graph actions, add them to a given XML element. * * \param[in] rsc Resource whose meta-attributes should be added * \param[in,out] xml Transition graph action attributes XML to add to */ void (*add_graph_meta)(const pcmk_resource_t *rsc, xmlNode *xml); /*! * \internal * \brief Add a resource's utilization to a table of utilization values * * This function is used when summing the utilization of a resource and all * resources colocated with it, to determine whether a node has sufficient * capacity. Given a resource and a table of utilization values, it will add * the resource's utilization to the existing values, if the resource has * not yet been assigned to a node. * * \param[in] rsc Resource with utilization to add * \param[in] orig_rsc Resource being assigned (for logging only) * \param[in] all_rscs List of all resources that will be summed * \param[in,out] utilization Table of utilization values to add to */ void (*add_utilization)(const pcmk_resource_t *rsc, const pcmk_resource_t *orig_rsc, GList *all_rscs, GHashTable *utilization); /*! * \internal * \brief Apply a shutdown lock for a resource, if appropriate * * \param[in,out] rsc Resource to check for shutdown lock */ void (*shutdown_lock)(pcmk_resource_t *rsc); }; // Actions (pcmk_sched_actions.c) G_GNUC_INTERNAL void pcmk__update_action_for_orderings(pcmk_action_t *action, pcmk_scheduler_t *scheduler); G_GNUC_INTERNAL uint32_t pcmk__update_ordered_actions(pcmk_action_t *first, pcmk_action_t *then, const pcmk_node_t *node, uint32_t flags, uint32_t filter, uint32_t type, pcmk_scheduler_t *scheduler); G_GNUC_INTERNAL void pcmk__log_action(const char *pre_text, const pcmk_action_t *action, bool details); G_GNUC_INTERNAL pcmk_action_t *pcmk__new_cancel_action(pcmk_resource_t *rsc, const char *name, guint interval_ms, const pcmk_node_t *node); G_GNUC_INTERNAL pcmk_action_t *pcmk__new_shutdown_action(pcmk_node_t *node); G_GNUC_INTERNAL bool pcmk__action_locks_rsc_to_node(const pcmk_action_t *action); G_GNUC_INTERNAL void pcmk__deduplicate_action_inputs(pcmk_action_t *action); G_GNUC_INTERNAL void pcmk__output_actions(pcmk_scheduler_t *scheduler); G_GNUC_INTERNAL bool pcmk__check_action_config(pcmk_resource_t *rsc, pcmk_node_t *node, const xmlNode *xml_op); G_GNUC_INTERNAL void pcmk__handle_rsc_config_changes(pcmk_scheduler_t *scheduler); // Recurring actions (pcmk_sched_recurring.c) G_GNUC_INTERNAL void pcmk__create_recurring_actions(pcmk_resource_t *rsc); G_GNUC_INTERNAL void pcmk__schedule_cancel(pcmk_resource_t *rsc, const char *call_id, const char *task, guint interval_ms, const pcmk_node_t *node, const char *reason); G_GNUC_INTERNAL void pcmk__reschedule_recurring(pcmk_resource_t *rsc, const char *task, guint interval_ms, pcmk_node_t *node); G_GNUC_INTERNAL bool pcmk__action_is_recurring(const pcmk_action_t *action); // Producing transition graphs (pcmk_graph_producer.c) G_GNUC_INTERNAL bool pcmk__graph_has_loop(const pcmk_action_t *init_action, const pcmk_action_t *action, pcmk__related_action_t *input); G_GNUC_INTERNAL void pcmk__add_rsc_actions_to_graph(pcmk_resource_t *rsc); G_GNUC_INTERNAL void pcmk__create_graph(pcmk_scheduler_t *scheduler); // Fencing (pcmk_sched_fencing.c) G_GNUC_INTERNAL void pcmk__order_vs_fence(pcmk_action_t *stonith_op, pcmk_scheduler_t *scheduler); G_GNUC_INTERNAL void pcmk__order_vs_unfence(const pcmk_resource_t *rsc, pcmk_node_t *node, pcmk_action_t *action, enum pcmk__action_relation_flags order); G_GNUC_INTERNAL void pcmk__fence_guest(pcmk_node_t *node); G_GNUC_INTERNAL bool pcmk__node_unfenced(const pcmk_node_t *node); G_GNUC_INTERNAL void pcmk__order_restart_vs_unfence(gpointer data, gpointer user_data); // Injected scheduler inputs (pcmk_sched_injections.c) G_GNUC_INTERNAL void pcmk__inject_scheduler_input(pcmk_scheduler_t *scheduler, cib_t *cib, const pcmk_injections_t *injections); // Constraints of any type (pcmk_sched_constraints.c) G_GNUC_INTERNAL pcmk_resource_t *pcmk__find_constraint_resource(GList *rsc_list, const char *id); G_GNUC_INTERNAL int pcmk__parse_constraint_role(const char *id, const char *role_spec, enum rsc_role_e *role); G_GNUC_INTERNAL xmlNode *pcmk__expand_tags_in_sets(xmlNode *xml_obj, const pcmk_scheduler_t *scheduler); G_GNUC_INTERNAL bool pcmk__valid_resource_or_tag(const pcmk_scheduler_t *scheduler, const char *id, pcmk_resource_t **rsc, pcmk__idref_t **tag); G_GNUC_INTERNAL bool pcmk__tag_to_set(xmlNode *xml_obj, xmlNode **rsc_set, const char *attr, bool convert_rsc, const pcmk_scheduler_t *scheduler); G_GNUC_INTERNAL void pcmk__create_internal_constraints(pcmk_scheduler_t *scheduler); // Location constraints G_GNUC_INTERNAL void pcmk__unpack_location(xmlNode *xml_obj, pcmk_scheduler_t *scheduler); G_GNUC_INTERNAL pcmk__location_t *pcmk__new_location(const char *id, pcmk_resource_t *rsc, int node_score, const char *discover_mode, pcmk_node_t *foo_node); G_GNUC_INTERNAL void pcmk__apply_locations(pcmk_scheduler_t *scheduler); G_GNUC_INTERNAL void pcmk__apply_location(pcmk_resource_t *rsc, pcmk__location_t *constraint); // Colocation constraints (pcmk_sched_colocation.c) enum pcmk__coloc_affects { pcmk__coloc_affects_nothing = 0, pcmk__coloc_affects_location, pcmk__coloc_affects_role, }; G_GNUC_INTERNAL const char *pcmk__colocation_node_attr(const pcmk_node_t *node, const char *attr, const pcmk_resource_t *rsc); G_GNUC_INTERNAL enum pcmk__coloc_affects pcmk__colocation_affects(const pcmk_resource_t *dependent, const pcmk_resource_t *primary, const pcmk__colocation_t *colocation, bool preview); G_GNUC_INTERNAL void pcmk__apply_coloc_to_scores(pcmk_resource_t *dependent, const pcmk_resource_t *primary, const pcmk__colocation_t *colocation); G_GNUC_INTERNAL int pcmk__apply_coloc_to_priority(pcmk_resource_t *dependent, const pcmk_resource_t *primary, const pcmk__colocation_t *colocation); G_GNUC_INTERNAL void pcmk__add_colocated_node_scores(pcmk_resource_t *source_rsc, const pcmk_resource_t *target_rsc, const char *log_id, GHashTable **nodes, const pcmk__colocation_t *colocation, float factor, uint32_t flags); G_GNUC_INTERNAL void pcmk__add_dependent_scores(gpointer data, gpointer user_data); G_GNUC_INTERNAL void pcmk__colocation_intersect_nodes(pcmk_resource_t *dependent, const pcmk_resource_t *primary, const pcmk__colocation_t *colocation, const GList *primary_nodes, bool merge_scores); G_GNUC_INTERNAL void pcmk__unpack_colocation(xmlNode *xml_obj, pcmk_scheduler_t *scheduler); G_GNUC_INTERNAL void pcmk__add_this_with(GList **list, const pcmk__colocation_t *colocation, const pcmk_resource_t *rsc); G_GNUC_INTERNAL void pcmk__add_this_with_list(GList **list, GList *addition, const pcmk_resource_t *rsc); G_GNUC_INTERNAL void pcmk__add_with_this(GList **list, const pcmk__colocation_t *colocation, const pcmk_resource_t *rsc); G_GNUC_INTERNAL void pcmk__add_with_this_list(GList **list, GList *addition, const pcmk_resource_t *rsc); G_GNUC_INTERNAL GList *pcmk__with_this_colocations(const pcmk_resource_t *rsc); G_GNUC_INTERNAL GList *pcmk__this_with_colocations(const pcmk_resource_t *rsc); G_GNUC_INTERNAL void pcmk__new_colocation(const char *id, const char *node_attr, int score, pcmk_resource_t *dependent, pcmk_resource_t *primary, const char *dependent_role_spec, const char *primary_role_spec, uint32_t flags); G_GNUC_INTERNAL void pcmk__block_colocation_dependents(pcmk_action_t *action); G_GNUC_INTERNAL bool pcmk__colocation_has_influence(const pcmk__colocation_t *colocation, const pcmk_resource_t *rsc); // Ordering constraints (pcmk_sched_ordering.c) G_GNUC_INTERNAL void pcmk__new_ordering(pcmk_resource_t *first_rsc, char *first_task, pcmk_action_t *first_action, pcmk_resource_t *then_rsc, char *then_task, pcmk_action_t *then_action, uint32_t flags, pcmk_scheduler_t *sched); G_GNUC_INTERNAL void pcmk__unpack_ordering(xmlNode *xml_obj, pcmk_scheduler_t *scheduler); G_GNUC_INTERNAL void pcmk__disable_invalid_orderings(pcmk_scheduler_t *scheduler); G_GNUC_INTERNAL void pcmk__order_stops_before_shutdown(pcmk_node_t *node, pcmk_action_t *shutdown_op); G_GNUC_INTERNAL void pcmk__apply_orderings(pcmk_scheduler_t *sched); G_GNUC_INTERNAL void pcmk__order_after_each(pcmk_action_t *after, GList *list); /*! * \internal * \brief Create a new ordering between two resource actions * * \param[in,out] first_rsc Resource for 'first' action * \param[in,out] first_task Action key for 'first' action * \param[in] then_rsc Resource for 'then' action * \param[in,out] then_task Action key for 'then' action * \param[in] flags Group of enum pcmk__action_relation_flags */ #define pcmk__order_resource_actions(first_rsc, first_task, \ then_rsc, then_task, flags) \ pcmk__new_ordering((first_rsc), \ pcmk__op_key((first_rsc)->id, (first_task), 0), \ NULL, \ (then_rsc), \ pcmk__op_key((then_rsc)->id, (then_task), 0), \ NULL, (flags), (first_rsc)->priv->scheduler) #define pcmk__order_starts(rsc1, rsc2, flags) \ pcmk__order_resource_actions((rsc1), PCMK_ACTION_START, \ (rsc2), PCMK_ACTION_START, (flags)) #define pcmk__order_stops(rsc1, rsc2, flags) \ pcmk__order_resource_actions((rsc1), PCMK_ACTION_STOP, \ (rsc2), PCMK_ACTION_STOP, (flags)) // Ticket constraints (pcmk_sched_tickets.c) G_GNUC_INTERNAL void pcmk__unpack_rsc_ticket(xmlNode *xml_obj, pcmk_scheduler_t *scheduler); // Promotable clone resources (pcmk_sched_promotable.c) G_GNUC_INTERNAL void pcmk__add_promotion_scores(pcmk_resource_t *rsc); G_GNUC_INTERNAL void pcmk__require_promotion_tickets(pcmk_resource_t *rsc); G_GNUC_INTERNAL void pcmk__set_instance_roles(pcmk_resource_t *rsc); G_GNUC_INTERNAL void pcmk__create_promotable_actions(pcmk_resource_t *clone); G_GNUC_INTERNAL void pcmk__promotable_restart_ordering(pcmk_resource_t *rsc); G_GNUC_INTERNAL void pcmk__order_promotable_instances(pcmk_resource_t *clone); G_GNUC_INTERNAL void pcmk__update_dependent_with_promotable(const pcmk_resource_t *primary, pcmk_resource_t *dependent, const pcmk__colocation_t *colocation); G_GNUC_INTERNAL int pcmk__update_promotable_dependent_priority(const pcmk_resource_t *primary, pcmk_resource_t *dependent, const pcmk__colocation_t *colocation); // Pacemaker Remote nodes (pcmk_sched_remote.c) G_GNUC_INTERNAL bool pcmk__is_failed_remote_node(const pcmk_node_t *node); G_GNUC_INTERNAL void pcmk__order_remote_connection_actions(pcmk_scheduler_t *scheduler); G_GNUC_INTERNAL bool pcmk__rsc_corresponds_to_guest(const pcmk_resource_t *rsc, const pcmk_node_t *node); G_GNUC_INTERNAL pcmk_node_t *pcmk__connection_host_for_action(const pcmk_action_t *action); G_GNUC_INTERNAL void pcmk__substitute_remote_addr(pcmk_resource_t *rsc, GHashTable *params); G_GNUC_INTERNAL void pcmk__add_guest_meta_to_xml(xmlNode *args_xml, const pcmk_action_t *action); // Primitives (pcmk_sched_primitive.c) G_GNUC_INTERNAL pcmk_node_t *pcmk__primitive_assign(pcmk_resource_t *rsc, const pcmk_node_t *prefer, bool stop_if_fail); G_GNUC_INTERNAL void pcmk__primitive_create_actions(pcmk_resource_t *rsc); G_GNUC_INTERNAL void pcmk__primitive_internal_constraints(pcmk_resource_t *rsc); G_GNUC_INTERNAL uint32_t pcmk__primitive_action_flags(pcmk_action_t *action, const pcmk_node_t *node); G_GNUC_INTERNAL int pcmk__primitive_apply_coloc_score(pcmk_resource_t *dependent, const pcmk_resource_t *primary, const pcmk__colocation_t *colocation, bool for_dependent); G_GNUC_INTERNAL void pcmk__with_primitive_colocations(const pcmk_resource_t *rsc, const pcmk_resource_t *orig_rsc, GList **list); G_GNUC_INTERNAL void pcmk__primitive_with_colocations(const pcmk_resource_t *rsc, const pcmk_resource_t *orig_rsc, GList **list); G_GNUC_INTERNAL void pcmk__schedule_cleanup(pcmk_resource_t *rsc, const pcmk_node_t *node, bool optional); G_GNUC_INTERNAL void pcmk__primitive_add_graph_meta(const pcmk_resource_t *rsc, xmlNode *xml); G_GNUC_INTERNAL void pcmk__primitive_add_utilization(const pcmk_resource_t *rsc, const pcmk_resource_t *orig_rsc, GList *all_rscs, GHashTable *utilization); G_GNUC_INTERNAL void pcmk__primitive_shutdown_lock(pcmk_resource_t *rsc); // Groups (pcmk_sched_group.c) G_GNUC_INTERNAL pcmk_node_t *pcmk__group_assign(pcmk_resource_t *rsc, const pcmk_node_t *prefer, bool stop_if_fail); G_GNUC_INTERNAL void pcmk__group_create_actions(pcmk_resource_t *rsc); G_GNUC_INTERNAL void pcmk__group_internal_constraints(pcmk_resource_t *rsc); G_GNUC_INTERNAL int pcmk__group_apply_coloc_score(pcmk_resource_t *dependent, const pcmk_resource_t *primary, const pcmk__colocation_t *colocation, bool for_dependent); G_GNUC_INTERNAL void pcmk__with_group_colocations(const pcmk_resource_t *rsc, const pcmk_resource_t *orig_rsc, GList **list); G_GNUC_INTERNAL void pcmk__group_with_colocations(const pcmk_resource_t *rsc, const pcmk_resource_t *orig_rsc, GList **list); G_GNUC_INTERNAL void pcmk__group_add_colocated_node_scores(pcmk_resource_t *source_rsc, const pcmk_resource_t *target_rsc, const char *log_id, GHashTable **nodes, const pcmk__colocation_t *colocation, float factor, uint32_t flags); G_GNUC_INTERNAL void pcmk__group_apply_location(pcmk_resource_t *rsc, pcmk__location_t *location); G_GNUC_INTERNAL uint32_t pcmk__group_action_flags(pcmk_action_t *action, const pcmk_node_t *node); G_GNUC_INTERNAL uint32_t pcmk__group_update_ordered_actions(pcmk_action_t *first, pcmk_action_t *then, const pcmk_node_t *node, uint32_t flags, uint32_t filter, uint32_t type, pcmk_scheduler_t *scheduler); G_GNUC_INTERNAL GList *pcmk__group_colocated_resources(const pcmk_resource_t *rsc, const pcmk_resource_t *orig_rsc, GList *colocated_rscs); G_GNUC_INTERNAL void pcmk__group_add_utilization(const pcmk_resource_t *rsc, const pcmk_resource_t *orig_rsc, GList *all_rscs, GHashTable *utilization); G_GNUC_INTERNAL void pcmk__group_shutdown_lock(pcmk_resource_t *rsc); // Clones (pcmk_sched_clone.c) G_GNUC_INTERNAL pcmk_node_t *pcmk__clone_assign(pcmk_resource_t *rsc, const pcmk_node_t *prefer, bool stop_if_fail); G_GNUC_INTERNAL void pcmk__clone_create_actions(pcmk_resource_t *rsc); G_GNUC_INTERNAL bool pcmk__clone_create_probe(pcmk_resource_t *rsc, pcmk_node_t *node); G_GNUC_INTERNAL void pcmk__clone_internal_constraints(pcmk_resource_t *rsc); G_GNUC_INTERNAL int pcmk__clone_apply_coloc_score(pcmk_resource_t *dependent, const pcmk_resource_t *primary, const pcmk__colocation_t *colocation, bool for_dependent); G_GNUC_INTERNAL void pcmk__with_clone_colocations(const pcmk_resource_t *rsc, const pcmk_resource_t *orig_rsc, GList **list); G_GNUC_INTERNAL void pcmk__clone_with_colocations(const pcmk_resource_t *rsc, const pcmk_resource_t *orig_rsc, GList **list); G_GNUC_INTERNAL void pcmk__clone_apply_location(pcmk_resource_t *rsc, pcmk__location_t *constraint); G_GNUC_INTERNAL uint32_t pcmk__clone_action_flags(pcmk_action_t *action, const pcmk_node_t *node); G_GNUC_INTERNAL void pcmk__clone_add_actions_to_graph(pcmk_resource_t *rsc); G_GNUC_INTERNAL void pcmk__clone_add_graph_meta(const pcmk_resource_t *rsc, xmlNode *xml); G_GNUC_INTERNAL void pcmk__clone_add_utilization(const pcmk_resource_t *rsc, const pcmk_resource_t *orig_rsc, GList *all_rscs, GHashTable *utilization); G_GNUC_INTERNAL void pcmk__clone_shutdown_lock(pcmk_resource_t *rsc); // Bundles (pcmk_sched_bundle.c) G_GNUC_INTERNAL pcmk_node_t *pcmk__bundle_assign(pcmk_resource_t *rsc, const pcmk_node_t *prefer, bool stop_if_fail); G_GNUC_INTERNAL void pcmk__bundle_create_actions(pcmk_resource_t *rsc); G_GNUC_INTERNAL bool pcmk__bundle_create_probe(pcmk_resource_t *rsc, pcmk_node_t *node); G_GNUC_INTERNAL void pcmk__bundle_internal_constraints(pcmk_resource_t *rsc); G_GNUC_INTERNAL int pcmk__bundle_apply_coloc_score(pcmk_resource_t *dependent, const pcmk_resource_t *primary, const pcmk__colocation_t *colocation, bool for_dependent); G_GNUC_INTERNAL void pcmk__with_bundle_colocations(const pcmk_resource_t *rsc, const pcmk_resource_t *orig_rsc, GList **list); G_GNUC_INTERNAL void pcmk__bundle_with_colocations(const pcmk_resource_t *rsc, const pcmk_resource_t *orig_rsc, GList **list); G_GNUC_INTERNAL void pcmk__bundle_apply_location(pcmk_resource_t *rsc, pcmk__location_t *constraint); G_GNUC_INTERNAL uint32_t pcmk__bundle_action_flags(pcmk_action_t *action, const pcmk_node_t *node); G_GNUC_INTERNAL void pcmk__output_bundle_actions(pcmk_resource_t *rsc); G_GNUC_INTERNAL void pcmk__bundle_add_actions_to_graph(pcmk_resource_t *rsc); G_GNUC_INTERNAL void pcmk__bundle_add_utilization(const pcmk_resource_t *rsc, const pcmk_resource_t *orig_rsc, GList *all_rscs, GHashTable *utilization); G_GNUC_INTERNAL void pcmk__bundle_shutdown_lock(pcmk_resource_t *rsc); // Clone instances or bundle replica containers (pcmk_sched_instances.c) G_GNUC_INTERNAL void pcmk__assign_instances(pcmk_resource_t *collective, GList *instances, int max_total, int max_per_node); G_GNUC_INTERNAL void pcmk__create_instance_actions(pcmk_resource_t *rsc, GList *instances); G_GNUC_INTERNAL bool pcmk__instance_matches(const pcmk_resource_t *instance, const pcmk_node_t *node, enum rsc_role_e role, bool current); G_GNUC_INTERNAL pcmk_resource_t *pcmk__find_compatible_instance(const pcmk_resource_t *match_rsc, const pcmk_resource_t *rsc, enum rsc_role_e role, bool current); G_GNUC_INTERNAL uint32_t pcmk__instance_update_ordered_actions(pcmk_action_t *first, pcmk_action_t *then, const pcmk_node_t *node, uint32_t flags, uint32_t filter, uint32_t type, pcmk_scheduler_t *scheduler); G_GNUC_INTERNAL uint32_t pcmk__collective_action_flags(pcmk_action_t *action, const GList *instances, const pcmk_node_t *node); // Injections (pcmk_injections.c) G_GNUC_INTERNAL xmlNode *pcmk__inject_node(cib_t *cib_conn, const char *node, const char *uuid); G_GNUC_INTERNAL xmlNode *pcmk__inject_node_state_change(cib_t *cib_conn, const char *node, bool up); G_GNUC_INTERNAL xmlNode *pcmk__inject_resource_history(pcmk__output_t *out, xmlNode *cib_node, const char *resource, const char *lrm_name, const char *rclass, const char *rtype, const char *rprovider); G_GNUC_INTERNAL void pcmk__inject_failcount(pcmk__output_t *out, cib_t *cib_conn, xmlNode *cib_node, const char *resource, const char *task, guint interval_ms, int rc, bool infinity); G_GNUC_INTERNAL xmlNode *pcmk__inject_action_result(xmlNode *cib_resource, lrmd_event_data_t *op, const char *node, int target_rc); // Nodes (pcmk_sched_nodes.c) +//! Options for checking node availability +enum pcmk__node_availability { + //! Disallow offline or unclean nodes (always implied) + pcmk__node_alive = 0, + + //! Disallow shutting down, standby, and maintenance nodes + pcmk__node_usable = (1 << 0), + + //! Disallow nodes with negative scores + pcmk__node_no_negative = (1 << 2), + + //! Disallow guest nodes whose guest resource is unrunnable + pcmk__node_no_unrunnable_guest = (1 << 4), +}; + G_GNUC_INTERNAL -bool pcmk__node_available(const pcmk_node_t *node, bool consider_score, - bool consider_guest); +bool pcmk__node_available(const pcmk_node_t *node, uint32_t flags); G_GNUC_INTERNAL bool pcmk__any_node_available(GHashTable *nodes); G_GNUC_INTERNAL GHashTable *pcmk__copy_node_table(GHashTable *nodes); G_GNUC_INTERNAL void pcmk__copy_node_tables(const pcmk_resource_t *rsc, GHashTable **copy); G_GNUC_INTERNAL void pcmk__restore_node_tables(pcmk_resource_t *rsc, GHashTable *backup); G_GNUC_INTERNAL GList *pcmk__sort_nodes(GList *nodes, pcmk_node_t *active_node); G_GNUC_INTERNAL void pcmk__apply_node_health(pcmk_scheduler_t *scheduler); G_GNUC_INTERNAL pcmk_node_t *pcmk__top_allowed_node(const pcmk_resource_t *rsc, const pcmk_node_t *node); // Functions applying to more than one variant (pcmk_sched_resource.c) G_GNUC_INTERNAL void pcmk__set_assignment_methods(pcmk_scheduler_t *scheduler); G_GNUC_INTERNAL bool pcmk__rsc_agent_changed(pcmk_resource_t *rsc, pcmk_node_t *node, const xmlNode *rsc_entry, bool active_on_node); G_GNUC_INTERNAL GList *pcmk__rscs_matching_id(const char *id, const pcmk_scheduler_t *scheduler); G_GNUC_INTERNAL GList *pcmk__colocated_resources(const pcmk_resource_t *rsc, const pcmk_resource_t *orig_rsc, GList *colocated_rscs); G_GNUC_INTERNAL void pcmk__noop_add_graph_meta(const pcmk_resource_t *rsc, xmlNode *xml); G_GNUC_INTERNAL void pcmk__output_resource_actions(pcmk_resource_t *rsc); G_GNUC_INTERNAL bool pcmk__assign_resource(pcmk_resource_t *rsc, pcmk_node_t *node, bool force, bool stop_if_fail); G_GNUC_INTERNAL void pcmk__unassign_resource(pcmk_resource_t *rsc); G_GNUC_INTERNAL bool pcmk__threshold_reached(pcmk_resource_t *rsc, const pcmk_node_t *node, pcmk_resource_t **failed); G_GNUC_INTERNAL void pcmk__sort_resources(pcmk_scheduler_t *scheduler); G_GNUC_INTERNAL gint pcmk__cmp_instance(gconstpointer a, gconstpointer b); G_GNUC_INTERNAL gint pcmk__cmp_instance_number(gconstpointer a, gconstpointer b); // Functions related to probes (pcmk_sched_probes.c) G_GNUC_INTERNAL bool pcmk__probe_rsc_on_node(pcmk_resource_t *rsc, pcmk_node_t *node); G_GNUC_INTERNAL void pcmk__order_probes(pcmk_scheduler_t *scheduler); G_GNUC_INTERNAL bool pcmk__probe_resource_list(GList *rscs, pcmk_node_t *node); G_GNUC_INTERNAL void pcmk__schedule_probes(pcmk_scheduler_t *scheduler); // Functions related to live migration (pcmk_sched_migration.c) void pcmk__create_migration_actions(pcmk_resource_t *rsc, const pcmk_node_t *current); void pcmk__abort_dangling_migration(void *data, void *user_data); bool pcmk__rsc_can_migrate(const pcmk_resource_t *rsc, const pcmk_node_t *current); void pcmk__order_migration_equivalents(pcmk__action_relation_t *order); // Functions related to node utilization (pcmk_sched_utilization.c) G_GNUC_INTERNAL int pcmk__compare_node_capacities(const pcmk_node_t *node1, const pcmk_node_t *node2); G_GNUC_INTERNAL void pcmk__consume_node_capacity(GHashTable *current_utilization, const pcmk_resource_t *rsc); G_GNUC_INTERNAL void pcmk__release_node_capacity(GHashTable *current_utilization, const pcmk_resource_t *rsc); G_GNUC_INTERNAL const pcmk_node_t *pcmk__ban_insufficient_capacity(pcmk_resource_t *rsc); G_GNUC_INTERNAL void pcmk__create_utilization_constraints(pcmk_resource_t *rsc, const GList *allowed_nodes); G_GNUC_INTERNAL void pcmk__show_node_capacities(const char *desc, pcmk_scheduler_t *scheduler); // Functions related to the scheduler (pcmk_scheduler.c) G_GNUC_INTERNAL int pcmk__init_scheduler(pcmk__output_t *out, xmlNodePtr input, const crm_time_t *date, pcmk_scheduler_t **scheduler); // General setup functions (pcmk_setup.c) G_GNUC_INTERNAL int pcmk__setup_output_cib_sched(pcmk__output_t **out, cib_t **cib, pcmk_scheduler_t **scheduler, xmlNode **xml); G_GNUC_INTERNAL int pcmk__setup_output_fencing(pcmk__output_t **out, stonith_t **st, xmlNode **xml); #ifdef __cplusplus } #endif #endif // PCMK__PACEMAKER_LIBPACEMAKER_PRIVATE__H diff --git a/lib/pacemaker/pcmk_sched_actions.c b/lib/pacemaker/pcmk_sched_actions.c index f8699db46b..36ecbeef42 100644 --- a/lib/pacemaker/pcmk_sched_actions.c +++ b/lib/pacemaker/pcmk_sched_actions.c @@ -1,1955 +1,1955 @@ /* * Copyright 2004-2025 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(pcmk_action_t *action, const pcmk_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->priv->cmds->action_flags(action, NULL); if ((node == NULL) || !pcmk__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->priv->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)) { pcmk__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 pcmk_resource_t *first_rsc) { guint interval_ms = 0; char *uuid = NULL; char *rid = NULL; char *first_task_str = NULL; enum pcmk__action_type first_task = pcmk__action_unspecified; enum pcmk__action_type remapped_task = pcmk__action_unspecified; // Only non-notify actions for collective resources need remapping if ((strstr(first_uuid, PCMK_ACTION_NOTIFY) != NULL) || (first_rsc->priv->variant < pcmk__rsc_variant_group)) { goto done; } // Only non-recurring actions need remapping pcmk__assert(parse_op_key(first_uuid, &rid, &first_task_str, &interval_ms)); if (interval_ms > 0) { goto done; } first_task = pcmk__parse_action(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) && (pcmk__is_clone(first_rsc) || pcmk__is_bundled(first_rsc))) { uuid = pcmk__notify_key(rid, "confirmed-post", pcmk__action_text(remapped_task)); } else { uuid = pcmk__op_key(rid, pcmk__action_text(remapped_task), 0); } pcmk__rsc_trace(first_rsc, "Remapped action UUID %s to %s for ordering purposes", first_uuid, uuid); } done: free(first_task_str); free(rid); return (uuid != NULL)? uuid : pcmk__str_copy(first_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 pcmk_action_t * action_for_ordering(pcmk_action_t *action) { pcmk_action_t *result = action; pcmk_resource_t *rsc = action->rsc; if (rsc == NULL) { return result; } if ((rsc->priv->variant >= pcmk__rsc_variant_group) && (action->uuid != NULL)) { char *uuid = action_uuid_for_ordering(action->uuid, rsc); result = find_first_action(rsc->priv->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] scheduler Scheduler data * * \return Group of enum pcmk__updated flags indicating what was updated */ static inline uint32_t update(pcmk_resource_t *rsc, pcmk_action_t *first, pcmk_action_t *then, const pcmk_node_t *node, uint32_t flags, uint32_t filter, uint32_t type, pcmk_scheduler_t *scheduler) { return rsc->priv->cmds->update_ordered_actions(first, then, node, flags, filter, type, scheduler); } /*! * \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] scheduler Scheduler data * * \return Group of enum pcmk__updated flags */ static uint32_t update_action_for_ordering_flags(pcmk_action_t *first, pcmk_action_t *then, uint32_t first_flags, uint32_t then_flags, pcmk__related_action_t *order, pcmk_scheduler_t *scheduler) { 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. */ pcmk_node_t *node = then->node; if (pcmk_is_set(order->flags, pcmk__ar_first_implies_same_node_then)) { /* 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. */ pcmk__clear_relation_flags(order->flags, pcmk__ar_first_implies_same_node_then); pcmk__set_relation_flags(order->flags, pcmk__ar_first_implies_then); node = first->node; pcmk__rsc_trace(then->rsc, "%s then %s: mapped " "pcmk__ar_first_implies_same_node_then to " "pcmk__ar_first_implies_then on %s", first->uuid, then->uuid, pcmk__node_name(node)); } if (pcmk_is_set(order->flags, 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, scheduler); } else if (!pcmk_is_set(first_flags, pcmk__action_optional) && pcmk_is_set(then->flags, pcmk__action_optional)) { pcmk__clear_action_flags(then, pcmk__action_optional); pcmk__set_updated_flags(changed, first, pcmk__updated_then); } pcmk__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->flags, pcmk__ar_intermediate_stop) && (then->rsc != NULL)) { enum pcmk__action_flags restart = pcmk__action_optional |pcmk__action_runnable; changed |= update(then->rsc, first, then, node, first_flags, restart, pcmk__ar_intermediate_stop, scheduler); pcmk__rsc_trace(then->rsc, "%s then %s: %s after pcmk__ar_intermediate_stop", first->uuid, then->uuid, (changed? "changed" : "unchanged")); } if (pcmk_is_set(order->flags, 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, scheduler); } else if (!pcmk_is_set(first_flags, pcmk__action_optional) && pcmk_is_set(first->flags, pcmk__action_runnable)) { pcmk__clear_action_flags(first, pcmk__action_runnable); pcmk__set_updated_flags(changed, first, pcmk__updated_first); } pcmk__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->flags, 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, scheduler); } pcmk__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->flags, pcmk__ar_min_runnable)) { if (then->rsc != NULL) { changed |= update(then->rsc, first, then, node, first_flags, pcmk__action_runnable, pcmk__ar_min_runnable, scheduler); } 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)) { pcmk__set_action_flags(then, pcmk__action_runnable); pcmk__set_updated_flags(changed, first, pcmk__updated_then); } } pcmk__rsc_trace(then->rsc, "%s then %s: %s after pcmk__ar_min_runnable", first->uuid, then->uuid, (changed? "changed" : "unchanged")); } if (pcmk_is_set(order->flags, pcmk__ar_nested_remote_probe) && (then->rsc != NULL)) { if (!pcmk_is_set(first_flags, pcmk__action_runnable) && (first->rsc != NULL) && (first->rsc->priv->active_nodes != NULL)) { pcmk__rsc_trace(then->rsc, "%s then %s: ignoring because first is stopping", first->uuid, then->uuid); order->flags = pcmk__ar_none; } else { changed |= update(then->rsc, first, then, node, first_flags, pcmk__action_runnable, pcmk__ar_unrunnable_first_blocks, scheduler); } pcmk__rsc_trace(then->rsc, "%s then %s: %s after pcmk__ar_nested_remote_probe", first->uuid, then->uuid, (changed? "changed" : "unchanged")); } if (pcmk_is_set(order->flags, pcmk__ar_unrunnable_first_blocks)) { if (then->rsc != NULL) { changed |= update(then->rsc, first, then, node, first_flags, pcmk__action_runnable, pcmk__ar_unrunnable_first_blocks, scheduler); } else if (!pcmk_is_set(first_flags, pcmk__action_runnable) && pcmk_is_set(then->flags, pcmk__action_runnable)) { pcmk__clear_action_flags(then, pcmk__action_runnable); pcmk__set_updated_flags(changed, first, pcmk__updated_then); } pcmk__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->flags, 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, scheduler); } pcmk__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->flags, pcmk__ar_first_else_then)) { if (then->rsc != NULL) { changed |= update(then->rsc, first, then, node, first_flags, pcmk__action_optional, pcmk__ar_first_else_then, scheduler); } pcmk__rsc_trace(then->rsc, "%s then %s: %s after pcmk__ar_first_else_then", first->uuid, then->uuid, (changed? "changed" : "unchanged")); } if (pcmk_is_set(order->flags, pcmk__ar_ordered)) { if (then->rsc != NULL) { changed |= update(then->rsc, first, then, node, first_flags, pcmk__action_runnable, pcmk__ar_ordered, scheduler); } pcmk__rsc_trace(then->rsc, "%s then %s: %s after pcmk__ar_ordered", first->uuid, then->uuid, (changed? "changed" : "unchanged")); } if (pcmk_is_set(order->flags, pcmk__ar_asymmetric)) { if (then->rsc != NULL) { changed |= update(then->rsc, first, then, node, first_flags, pcmk__action_runnable, pcmk__ar_asymmetric, scheduler); } pcmk__rsc_trace(then->rsc, "%s then %s: %s after pcmk__ar_asymmetric", first->uuid, then->uuid, (changed? "changed" : "unchanged")); } if (pcmk_is_set(first->flags, pcmk__action_runnable) && pcmk_is_set(order->flags, pcmk__ar_first_implies_then_graphed) && !pcmk_is_set(first_flags, pcmk__action_optional)) { pcmk__rsc_trace(then->rsc, "%s will be in graph because %s is required", then->uuid, first->uuid); pcmk__set_action_flags(then, pcmk__action_always_in_graph); // Don't bother marking 'then' as changed just for this } if (pcmk_is_set(order->flags, pcmk__ar_then_implies_first_graphed) && !pcmk_is_set(then_flags, pcmk__action_optional)) { pcmk__rsc_trace(then->rsc, "%s will be in graph because %s is required", first->uuid, then->uuid); pcmk__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->flags, pcmk__ar_first_implies_then |pcmk__ar_then_implies_first |pcmk__ar_intermediate_stop) && (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)) { /* @TODO This seems odd; why wouldn't an unrunnable "first" already * block "then" before this? Note that the unmanaged-stop-{1,2} * scheduler regression tests and the test CIB for T209 have tests for * "stop then stop" relations that would be good for checking any * changes. */ if (pcmk_is_set(then->flags, pcmk__action_runnable)) { pcmk__clear_action_flags(then, pcmk__action_runnable); pcmk__set_updated_flags(changed, first, pcmk__updated_then); } pcmk__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->priv->name) /*! * \internal * \brief Update an action's flags for all orderings where it is "then" * * \param[in,out] then Action to update * \param[in,out] scheduler Scheduler data */ void pcmk__update_action_for_orderings(pcmk_action_t *then, pcmk_scheduler_t *scheduler) { GList *lpc = NULL; uint32_t changed = pcmk__updated_none; int last_flags = then->flags; pcmk__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 (then->required_runnable_before > 0) { /* 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; /* The pcmk__ar_min_runnable clause of * update_action_for_ordering_flags() (called below) * will reset runnable if appropriate. */ pcmk__clear_action_flags(then, pcmk__action_runnable); } for (lpc = then->actions_before; lpc != NULL; lpc = lpc->next) { pcmk__related_action_t *other = lpc->data; pcmk_action_t *first = other->action; pcmk_node_t *then_node = then->node; pcmk_node_t *first_node = first->node; const uint32_t target = pcmk__rsc_node_assigned; if ((first->rsc != NULL) && pcmk__is_group(first->rsc) && pcmk__str_eq(first->task, PCMK_ACTION_START, pcmk__str_none)) { first_node = first->rsc->priv->fns->location(first->rsc, NULL, target); if (first_node != NULL) { pcmk__rsc_trace(first->rsc, "Found %s for 'first' %s", pcmk__node_name(first_node), first->uuid); } } if (pcmk__is_group(then->rsc) && pcmk__str_eq(then->task, PCMK_ACTION_START, pcmk__str_none)) { then_node = then->rsc->priv->fns->location(then->rsc, NULL, target); if (then_node != NULL) { pcmk__rsc_trace(then->rsc, "Found %s for 'then' %s", pcmk__node_name(then_node), then->uuid); } } // Disable constraint if it only applies when on same node, but isn't if (pcmk_is_set(other->flags, pcmk__ar_if_on_same_node) && (first_node != NULL) && (then_node != NULL) && !pcmk__same_node(first_node, then_node)) { pcmk__rsc_trace(then->rsc, "Disabled ordering %s on %s then %s on %s: " "not same node", other->action->uuid, pcmk__node_name(first_node), then->uuid, pcmk__node_name(then_node)); other->flags = pcmk__ar_none; continue; } pcmk__clear_updated_flags(changed, then, pcmk__updated_first); if ((first->rsc != NULL) && pcmk_is_set(other->flags, pcmk__ar_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). */ pcmk__set_action_flags(other->action, pcmk__action_optional); if (!strcmp(first->task, PCMK_ACTION_RELOAD_AGENT)) { pcmk__clear_rsc_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) { pcmk__rsc_trace(then->rsc, "Ordering %s after %s instead of %s", then->uuid, first->uuid, other->action->uuid); } pcmk__rsc_trace(then->rsc, "%s (%#.6x) then %s (%#.6x): type=%#.6x node=%s", first->uuid, first->flags, then->uuid, then->flags, other->flags, 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, scheduler); /* 'first' was for a complex resource (clone, group, etc), * create a new dependency if necessary */ } else if (order_actions(first, then, other->flags)) { /* 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); pcmk__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->flags = 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) { pcmk__related_action_t *other = lpc2->data; pcmk__update_action_for_orderings(other->action, scheduler); } pcmk__update_action_for_orderings(first, scheduler); } } if (then->required_runnable_before > 0) { 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, scheduler); for (lpc = then->actions_after; lpc != NULL; lpc = lpc->next) { pcmk__related_action_t *other = lpc->data; pcmk__update_action_for_orderings(other->action, scheduler); } } } static inline bool is_primitive_action(const pcmk_action_t *action) { return (action != NULL) && pcmk__is_primitive(action->rsc); } /*! * \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))) { \ pcmk__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 pcmk_action_t *first, pcmk_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_role = then->rsc->priv->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 pcmk__ar_intermediate_stop is used * * \param[in,out] first 'First' action in ordering * \param[in,out] then 'Then' action in ordering * \param[in] filter What action flags to care about * * \note pcmk__ar_intermediate_stop is set for "stop resource before starting * it" and "stop later group member before stopping earlier group member" */ static void handle_restart_ordering(pcmk_action_t *first, pcmk_action_t *then, uint32_t filter) { const char *reason = NULL; pcmk__assert(is_primitive_action(first) && 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; } pcmk__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] scheduler Scheduler data * * \return Group of enum pcmk__updated flags indicating what was updated */ uint32_t pcmk__update_ordered_actions(pcmk_action_t *first, pcmk_action_t *then, const pcmk_node_t *node, uint32_t flags, uint32_t filter, uint32_t type, pcmk_scheduler_t *scheduler) { uint32_t changed = pcmk__updated_none; uint32_t then_flags = 0U; uint32_t first_flags = 0U; pcmk__assert((first != NULL) && (then != NULL) && (scheduler != NULL)); then_flags = then->flags; first_flags = first->flags; if (pcmk_is_set(type, pcmk__ar_asymmetric)) { 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->priv->orig_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, pcmk__ar_first_else_then) && pcmk_is_set(filter, pcmk__action_optional) && !pcmk_is_set(first->flags, pcmk__action_runnable)) { clear_action_flag_because(then, pcmk__action_migratable, first); pcmk__clear_action_flags(then, pcmk__action_pseudo); } 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, pcmk__ar_intermediate_stop)) { handle_restart_ordering(first, then, filter); } if (then_flags != then->flags) { pcmk__set_updated_flags(changed, first, pcmk__updated_then); pcmk__rsc_trace(then->rsc, "%s on %s: flags are now %#.6x (was %#.6x) " "because of 'first' %s (%#.6x)", then->uuid, pcmk__node_name(then->node), then->flags, then_flags, first->uuid, first->flags); if ((then->rsc != NULL) && (then->rsc->priv->parent != NULL)) { // Required to handle "X_stop then X_start" for cloned groups pcmk__update_action_for_orderings(then, scheduler); } } if (first_flags != first->flags) { pcmk__set_updated_flags(changed, first, pcmk__updated_first); pcmk__rsc_trace(first->rsc, "%s on %s: flags are now %#.6x (was %#.6x) " "because of 'then' %s (%#.6x)", first->uuid, pcmk__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 pcmk_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->priv->name; node_uuid = action->node->priv->id; } else { node_uname = ""; } } switch (pcmk__parse_action(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 pcmk__related_action_t *other = NULL; crm_trace("\t\t====== Preceding Actions"); for (iter = action->actions_before; iter != NULL; iter = iter->next) { other = (const pcmk__related_action_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 pcmk__related_action_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 */ pcmk_action_t * pcmk__new_shutdown_action(pcmk_node_t *node) { char *shutdown_id = NULL; pcmk_action_t *shutdown_op = NULL; pcmk__assert(node != NULL); shutdown_id = crm_strdup_printf("%s-%s", PCMK_ACTION_DO_SHUTDOWN, node->priv->name); shutdown_op = custom_action(NULL, shutdown_id, PCMK_ACTION_DO_SHUTDOWN, node, FALSE, node->priv->scheduler); pcmk__order_stops_before_shutdown(node, shutdown_op); pcmk__insert_meta(shutdown_op, PCMK__META_OP_NO_WAIT, PCMK_VALUE_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 = pcmk__xe_create(NULL, PCMK_XE_PARAMETERS); g_hash_table_foreach(op->params, hash2field, args_xml); pcmk__filter_op_for_digest(args_xml); digest = pcmk__digest_operation(args_xml); crm_xml_add(update, PCMK__XA_OP_DIGEST, digest); pcmk__xml_free(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. * * @TODO This remapping can make log messages with task confusing for users * (for example, an "Initiating reload ..." followed by "... start ... * confirmed"). Either do this remapping in the scheduler if possible, or * store the original task in a new XML attribute for later logging. */ 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 PCMK_META_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_first_child(parent, PCMK__XE_LRM_RSC_OP, PCMK_XA_ID, op_id); if (xml_op == NULL) { xml_op = pcmk__xe_create(parent, PCMK__XE_LRM_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, PCMK_XA_ID, op_id); crm_xml_add(xml_op, PCMK__XA_OPERATION_KEY, key); crm_xml_add(xml_op, PCMK_XA_OPERATION, task); crm_xml_add(xml_op, PCMK_XA_CRM_DEBUG_ORIGIN, origin); crm_xml_add(xml_op, PCMK_XA_CRM_FEATURE_SET, caller_version); crm_xml_add(xml_op, PCMK__XA_TRANSITION_KEY, op->user_data); crm_xml_add(xml_op, PCMK__XA_TRANSITION_MAGIC, magic); crm_xml_add(xml_op, PCMK_XA_EXIT_REASON, pcmk__s(exit_reason, "")); crm_xml_add(xml_op, PCMK__META_ON_NODE, node); // For context during triage crm_xml_add_int(xml_op, PCMK__XA_CALL_ID, op->call_id); crm_xml_add_int(xml_op, PCMK__XA_RC_CODE, op->rc); crm_xml_add_int(xml_op, PCMK__XA_OP_STATUS, op->op_status); crm_xml_add_ms(xml_op, PCMK_META_INTERVAL, op->interval_ms); if ((op->t_run > 0) || (op->t_rcchange > 0) || (op->exec_time > 0) || (op->queue_time > 0)) { crm_trace("Timing data (" PCMK__OP_FMT "): " "last=%lld change=%lld exec=%u queue=%u", op->rsc_id, op->op_type, op->interval_ms, (long long) op->t_run, (long long) 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, PCMK_XA_LAST_RC_CHANGE, (long long) op->t_rcchange); } else { crm_xml_add_ll(xml_op, PCMK_XA_LAST_RC_CHANGE, (long long) op->t_run); } crm_xml_add_int(xml_op, PCMK_XA_EXEC_TIME, op->exec_time); crm_xml_add_int(xml_op, PCMK_XA_QUEUE_TIME, op->queue_time); } if (pcmk__str_any_of(op->op_type, PCMK_ACTION_MIGRATE_TO, PCMK_ACTION_MIGRATE_FROM, NULL)) { /* Record PCMK__META_MIGRATE_SOURCE and PCMK__META_MIGRATE_TARGET always * for migrate ops. */ const char *name = PCMK__META_MIGRATE_SOURCE; crm_xml_add(xml_op, name, crm_meta_value(op->params, name)); name = PCMK__META_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 PCMK_OPT_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 pcmk_action_t *action) { // Only resource actions taking place on resource's lock node are locked if ((action == NULL) || (action->rsc == NULL) || !pcmk__same_node(action->node, action->rsc->priv->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 pcmk__related_action_t *action_wrapper2 = a; const pcmk__related_action_t *action_wrapper1 = 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(pcmk_action_t *action) { GList *item = NULL; GList *next = NULL; pcmk__related_action_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) { pcmk__related_action_t *input = 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. */ pcmk__set_relation_flags(last_input->flags, input->flags); if (input->graphed) { last_input->graphed = true; } free(item->data); action->actions_before = g_list_delete_link(action->actions_before, item); } else { last_input = input; input->graphed = false; } } } /*! * \internal * \brief Output all scheduled actions * * \param[in,out] scheduler Scheduler data */ void pcmk__output_actions(pcmk_scheduler_t *scheduler) { pcmk__output_t *out = scheduler->priv->out; // Output node (non-resource) actions for (GList *iter = scheduler->priv->actions; iter != NULL; iter = iter->next) { char *node_name = NULL; char *task = NULL; pcmk_action_t *action = (pcmk_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, PCMK__META_STONITH_ACTION); task = crm_strdup_printf("Fence (%s)", op); } else { continue; // Don't display other node action types } if (pcmk__is_guest_or_bundle_node(action->node)) { const pcmk_resource_t *remote = action->node->priv->remote; node_name = crm_strdup_printf("%s (resource: %s)", pcmk__node_name(action->node), remote->priv->launcher->id); } else if (action->node != NULL) { node_name = crm_strdup_printf("%s", pcmk__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 = scheduler->priv->resources; iter != NULL; iter = iter->next) { pcmk_resource_t *rsc = (pcmk_resource_t *) iter->data; rsc->priv->cmds->output_actions(rsc); } } /*! * \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] scheduler Scheduler data * * \return true if only sanitized parameters changed, otherwise false */ static bool only_sanitized_changed(const xmlNode *xml_op, const pcmk__op_digest_t *digest_data, const pcmk_scheduler_t *scheduler) { const char *digest_secure = NULL; if (!pcmk_is_set(scheduler->flags, pcmk__sched_sanitized)) { // The scheduler is not being run as a simulation return false; } digest_secure = crm_element_value(xml_op, PCMK__XA_OP_SECURE_DIGEST); return (digest_data->rc != pcmk__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(pcmk_resource_t *rsc, const char *task, guint interval_ms, pcmk_node_t *node) { char *key = pcmk__op_key(rsc->id, task, interval_ms); pcmk_action_t *required = custom_action(rsc, key, task, NULL, FALSE, rsc->priv->scheduler); pe_action_set_reason(required, "resource definition change", true); trigger_unfencing(rsc, node, "Device parameters changed", NULL, rsc->priv->scheduler); } /*! * \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) { pcmk_resource_t *rsc = data; const pcmk_node_t *node = user_data; pcmk_action_t *reload = NULL; // For collective resources, just call recursively for children if (rsc->priv->variant > pcmk__rsc_variant_primitive) { g_list_foreach(rsc->priv->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)) { pcmk__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->priv->name); 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)) { pcmk__rsc_trace(rsc, "%s: preventing agent reload because start pending", rsc->id); custom_action(rsc, stop_key(rsc), PCMK_ACTION_STOP, node, FALSE, rsc->priv->scheduler); return; } // Schedule the reload pcmk__set_rsc_flags(rsc, pcmk__rsc_reload); reload = custom_action(rsc, reload_key(rsc), PCMK_ACTION_RELOAD_AGENT, node, FALSE, rsc->priv->scheduler); 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|pcmk__ar_then_cancels_first, rsc->priv->scheduler); pcmk__new_ordering(NULL, NULL, reload, rsc, demote_key(rsc), NULL, pcmk__ar_ordered|pcmk__ar_then_cancels_first, rsc->priv->scheduler); } /*! * \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(pcmk_resource_t *rsc, pcmk_node_t *node, const xmlNode *xml_op) { guint interval_ms = 0; const char *task = NULL; const pcmk__op_digest_t *digest_data = NULL; CRM_CHECK((rsc != NULL) && (node != NULL) && (xml_op != NULL), return false); task = crm_element_value(xml_op, PCMK_XA_OPERATION); CRM_CHECK(task != NULL, return false); crm_element_value_ms(xml_op, PCMK_META_INTERVAL, &interval_ms); // If this is a recurring action, check whether it has been orphaned if (interval_ms > 0) { if (pcmk__find_action_config(rsc, task, interval_ms, false) != NULL) { pcmk__rsc_trace(rsc, "%s-interval %s for %s on %s is in configuration", pcmk__readable_interval(interval_ms), task, rsc->id, pcmk__node_name(node)); } else if (pcmk_is_set(rsc->priv->scheduler->flags, pcmk__sched_cancel_removed_actions)) { pcmk__schedule_cancel(rsc, crm_element_value(xml_op, PCMK__XA_CALL_ID), task, interval_ms, node, "orphan"); return true; } else { pcmk__rsc_debug(rsc, "%s-interval %s for %s on %s is orphaned", pcmk__readable_interval(interval_ms), task, rsc->id, pcmk__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, pcmk__node_name(node)); task = task_for_digest(task, interval_ms); digest_data = rsc_action_digest_cmp(rsc, xml_op, node, rsc->priv->scheduler); if (only_sanitized_changed(xml_op, digest_data, rsc->priv->scheduler)) { if (!pcmk__is_daemon && (rsc->priv->scheduler->priv->out != NULL)) { pcmk__output_t *out = rsc->priv->scheduler->priv->out; out->info(out, "Only 'private' parameters to %s-interval %s for %s " "on %s changed: %s", pcmk__readable_interval(interval_ms), task, rsc->id, pcmk__node_name(node), crm_element_value(xml_op, PCMK__XA_TRANSITION_MAGIC)); } return false; } switch (digest_data->rc) { case pcmk__digest_restart: crm_log_xml_debug(digest_data->params_restart, "params:restart"); force_restart(rsc, task, interval_ms, node); return true; case pcmk__digest_unknown: case pcmk__digest_mismatch: // 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, PCMK__XA_OP_RESTART_DIGEST) != NULL) { // Agent supports reload, so use it trigger_unfencing(rsc, node, "Device parameters changed (reload)", NULL, rsc->priv->scheduler); crm_log_xml_debug(digest_data->params_all, "params:reload"); schedule_reload((gpointer) rsc, (gpointer) node); } else { pcmk__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 \c PCMK__XE_LRM_RSC_OP 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 = pcmk__xe_first_child(rsc_entry, PCMK__XE_LRM_RSC_OP, NULL, NULL); rsc_op != NULL; rsc_op = pcmk__xe_next(rsc_op, PCMK__XE_LRM_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 \c PCMK__XE_LRM_RSC_OP 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, pcmk_resource_t *rsc, pcmk_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 (pcmk__is_anonymous_clone(pe__const_top_resource(rsc, false))) { /* @TODO Should this be done for bundled primitives as well? Added * by 2ac43ae31 */ pcmk__rsc_trace(rsc, "Skipping configuration check " "for orphaned clone instance %s", rsc->id); } else { pcmk__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->priv->active_nodes, node->priv->id) == NULL) { if (pcmk__rsc_agent_changed(rsc, node, rsc_entry, false)) { pcmk__schedule_cleanup(rsc, node, false); } pcmk__rsc_trace(rsc, "Skipping configuration check for %s " "because no longer active on %s", rsc->id, pcmk__node_name(node)); return; } pcmk__rsc_trace(rsc, "Checking for configuration changes for %s on %s", rsc->id, pcmk__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, PCMK_XA_OPERATION); crm_element_value_ms(rsc_op, PCMK_META_INTERVAL, &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, PCMK__XA_CALL_ID), 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. */ pcmk__add_param_check(rsc_op, rsc, node, pcmk__check_active); } else if (pcmk__check_action_config(rsc, node, rsc_op) && (pe_get_failcount(node, rsc, NULL, pcmk__fc_effective, NULL) != 0)) { pe__clear_failcount(rsc, node, "action definition changed", rsc->priv->scheduler); } } } 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 \c PCMK__XE_LRM_RESOURCES from CIB status XML */ static void process_node_history(pcmk_node_t *node, const xmlNode *lrm_rscs) { crm_trace("Processing node history for %s", pcmk__node_name(node)); for (const xmlNode *rsc_entry = pcmk__xe_first_child(lrm_rscs, PCMK__XE_LRM_RESOURCE, NULL, NULL); rsc_entry != NULL; rsc_entry = pcmk__xe_next(rsc_entry, PCMK__XE_LRM_RESOURCE)) { if (rsc_entry->children != NULL) { GList *result = pcmk__rscs_matching_id(pcmk__xe_id(rsc_entry), node->priv->scheduler); for (GList *iter = result; iter != NULL; iter = iter->next) { pcmk_resource_t *rsc = (pcmk_resource_t *) iter->data; if (pcmk__is_primitive(rsc)) { process_rsc_history(rsc_entry, rsc, node); } } g_list_free(result); } } } // XPath to find a node's resource history #define XPATH_NODE_HISTORY "/" PCMK_XE_CIB "/" PCMK_XE_STATUS \ "/" PCMK__XE_NODE_STATE \ "[@" PCMK_XA_UNAME "='%s']" \ "/" PCMK__XE_LRM "/" PCMK__XE_LRM_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] scheduler Scheduler data */ void pcmk__handle_rsc_config_changes(pcmk_scheduler_t *scheduler) { 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 = scheduler->nodes; iter != NULL; iter = iter->next) { pcmk_node_t *node = (pcmk_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)) { + || pcmk__node_available(node, pcmk__node_alive|pcmk__node_usable)) { char *xpath = NULL; xmlNode *history = NULL; xpath = crm_strdup_printf(XPATH_NODE_HISTORY, node->priv->name); history = get_xpath_object(xpath, scheduler->input, LOG_NEVER); free(xpath); process_node_history(node, history); } } } diff --git a/lib/pacemaker/pcmk_sched_colocation.c b/lib/pacemaker/pcmk_sched_colocation.c index 3c4c4c6b23..3472d28adc 100644 --- a/lib/pacemaker/pcmk_sched_colocation.c +++ b/lib/pacemaker/pcmk_sched_colocation.c @@ -1,2044 +1,2044 @@ /* - * Copyright 2004-2024 the Pacemaker project contributors + * Copyright 2004-2025 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 "crm/common/util.h" #include "crm/common/xml_internal.h" #include "crm/common/xml.h" #include "libpacemaker_private.h" // Used to temporarily mark a node as unusable #define INFINITY_HACK (PCMK_SCORE_INFINITY * -100) /*! * \internal * \brief Get the value of a colocation's node attribute * * \param[in] node Node on which to look up the attribute * \param[in] attr Name of attribute to look up * \param[in] rsc Resource on whose behalf to look up the attribute * * \return Value of \p attr on \p node or on the host of \p node, as appropriate */ const char * pcmk__colocation_node_attr(const pcmk_node_t *node, const char *attr, const pcmk_resource_t *rsc) { const char *target = NULL; /* A resource colocated with a bundle or its primitive can't run on the * bundle node itself (where only the primitive, if any, can run). Instead, * we treat it as a colocation with the bundle's containers, so always look * up colocation node attributes on the container host. */ if (pcmk__is_bundle_node(node) && pcmk__is_bundled(rsc) && (pe__const_top_resource(rsc, false) == pe__bundled_resource(rsc))) { target = PCMK_VALUE_HOST; } else if (rsc != NULL) { target = g_hash_table_lookup(rsc->priv->meta, PCMK_META_CONTAINER_ATTRIBUTE_TARGET); } return pcmk__node_attr(node, attr, target, pcmk__rsc_node_assigned); } /*! * \internal * \brief Compare two colocations according to priority * * Compare two colocations according to the order in which they should be * considered, based on either their dependent resources or their primary * resources -- preferring (in order): * * Colocation that is not \c NULL * * Colocation whose resource has higher priority * * Colocation whose resource is of a higher-level variant * (bundle > clone > group > primitive) * * Colocation whose resource is promotable, if both are clones * * Colocation whose resource has lower ID in lexicographic order * * \param[in] colocation1 First colocation to compare * \param[in] colocation2 Second colocation to compare * \param[in] dependent If \c true, compare colocations by dependent * priority; otherwise compare them by primary priority * * \return A negative number if \p colocation1 should be considered first, * a positive number if \p colocation2 should be considered first, * or 0 if order doesn't matter */ static gint cmp_colocation_priority(const pcmk__colocation_t *colocation1, const pcmk__colocation_t *colocation2, bool dependent) { const pcmk_resource_t *rsc1 = NULL; const pcmk_resource_t *rsc2 = NULL; if (colocation1 == NULL) { return 1; } if (colocation2 == NULL) { return -1; } if (dependent) { rsc1 = colocation1->dependent; rsc2 = colocation2->dependent; pcmk__assert(colocation1->primary != NULL); } else { rsc1 = colocation1->primary; rsc2 = colocation2->primary; pcmk__assert(colocation1->dependent != NULL); } pcmk__assert((rsc1 != NULL) && (rsc2 != NULL)); if (rsc1->priv->priority > rsc2->priv->priority) { return -1; } if (rsc1->priv->priority < rsc2->priv->priority) { return 1; } // Process clones before primitives and groups if (rsc1->priv->variant > rsc2->priv->variant) { return -1; } if (rsc1->priv->variant < rsc2->priv->variant) { return 1; } /* @COMPAT scheduler <2.0.0: Process promotable clones before nonpromotable * clones (probably unnecessary, but avoids having to update regression * tests) */ if (pcmk__is_clone(rsc1)) { if (pcmk_is_set(rsc1->flags, pcmk__rsc_promotable) && !pcmk_is_set(rsc2->flags, pcmk__rsc_promotable)) { return -1; } if (!pcmk_is_set(rsc1->flags, pcmk__rsc_promotable) && pcmk_is_set(rsc2->flags, pcmk__rsc_promotable)) { return 1; } } return strcmp(rsc1->id, rsc2->id); } /*! * \internal * \brief Compare two colocations according to priority based on dependents * * Compare two colocations according to the order in which they should be * considered, based on their dependent resources -- preferring (in order): * * Colocation that is not \c NULL * * Colocation whose resource has higher priority * * Colocation whose resource is of a higher-level variant * (bundle > clone > group > primitive) * * Colocation whose resource is promotable, if both are clones * * Colocation whose resource has lower ID in lexicographic order * * \param[in] a First colocation to compare * \param[in] b Second colocation to compare * * \return A negative number if \p a should be considered first, * a positive number if \p b should be considered first, * or 0 if order doesn't matter */ static gint cmp_dependent_priority(gconstpointer a, gconstpointer b) { return cmp_colocation_priority(a, b, true); } /*! * \internal * \brief Compare two colocations according to priority based on primaries * * Compare two colocations according to the order in which they should be * considered, based on their primary resources -- preferring (in order): * * Colocation that is not \c NULL * * Colocation whose primary has higher priority * * Colocation whose primary is of a higher-level variant * (bundle > clone > group > primitive) * * Colocation whose primary is promotable, if both are clones * * Colocation whose primary has lower ID in lexicographic order * * \param[in] a First colocation to compare * \param[in] b Second colocation to compare * * \return A negative number if \p a should be considered first, * a positive number if \p b should be considered first, * or 0 if order doesn't matter */ static gint cmp_primary_priority(gconstpointer a, gconstpointer b) { return cmp_colocation_priority(a, b, false); } /*! * \internal * \brief Add a "this with" colocation constraint to a sorted list * * \param[in,out] list List of constraints to add \p colocation to * \param[in] colocation Colocation constraint to add to \p list * \param[in] rsc Resource whose colocations we're getting (for * logging only) * * \note The list will be sorted using cmp_primary_priority(). */ void pcmk__add_this_with(GList **list, const pcmk__colocation_t *colocation, const pcmk_resource_t *rsc) { pcmk__assert((list != NULL) && (colocation != NULL) && (rsc != NULL)); pcmk__rsc_trace(rsc, "Adding colocation %s (%s with %s using %s @%s) to " "'this with' list for %s", colocation->id, colocation->dependent->id, colocation->primary->id, colocation->node_attribute, pcmk_readable_score(colocation->score), rsc->id); *list = g_list_insert_sorted(*list, (gpointer) colocation, cmp_primary_priority); } /*! * \internal * \brief Add a list of "this with" colocation constraints to a list * * \param[in,out] list List of constraints to add \p addition to * \param[in] addition List of colocation constraints to add to \p list * \param[in] rsc Resource whose colocations we're getting (for * logging only) * * \note The lists must be pre-sorted by cmp_primary_priority(). */ void pcmk__add_this_with_list(GList **list, GList *addition, const pcmk_resource_t *rsc) { pcmk__assert((list != NULL) && (rsc != NULL)); pcmk__if_tracing( {}, // Always add each colocation individually if tracing { if (*list == NULL) { // Trivial case for efficiency if not tracing *list = g_list_copy(addition); return; } } ); for (const GList *iter = addition; iter != NULL; iter = iter->next) { pcmk__add_this_with(list, addition->data, rsc); } } /*! * \internal * \brief Add a "with this" colocation constraint to a sorted list * * \param[in,out] list List of constraints to add \p colocation to * \param[in] colocation Colocation constraint to add to \p list * \param[in] rsc Resource whose colocations we're getting (for * logging only) * * \note The list will be sorted using cmp_dependent_priority(). */ void pcmk__add_with_this(GList **list, const pcmk__colocation_t *colocation, const pcmk_resource_t *rsc) { pcmk__assert((list != NULL) && (colocation != NULL) && (rsc != NULL)); pcmk__rsc_trace(rsc, "Adding colocation %s (%s with %s using %s @%s) to " "'with this' list for %s", colocation->id, colocation->dependent->id, colocation->primary->id, colocation->node_attribute, pcmk_readable_score(colocation->score), rsc->id); *list = g_list_insert_sorted(*list, (gpointer) colocation, cmp_dependent_priority); } /*! * \internal * \brief Add a list of "with this" colocation constraints to a list * * \param[in,out] list List of constraints to add \p addition to * \param[in] addition List of colocation constraints to add to \p list * \param[in] rsc Resource whose colocations we're getting (for * logging only) * * \note The lists must be pre-sorted by cmp_dependent_priority(). */ void pcmk__add_with_this_list(GList **list, GList *addition, const pcmk_resource_t *rsc) { pcmk__assert((list != NULL) && (rsc != NULL)); pcmk__if_tracing( {}, // Always add each colocation individually if tracing { if (*list == NULL) { // Trivial case for efficiency if not tracing *list = g_list_copy(addition); return; } } ); for (const GList *iter = addition; iter != NULL; iter = iter->next) { pcmk__add_with_this(list, addition->data, rsc); } } /*! * \internal * \brief Add orderings necessary for an anti-colocation constraint * * \param[in,out] first_rsc One resource in an anti-colocation * \param[in] first_role Anti-colocation role of \p first_rsc * \param[in] then_rsc Other resource in the anti-colocation * \param[in] then_role Anti-colocation role of \p then_rsc */ static void anti_colocation_order(pcmk_resource_t *first_rsc, int first_role, pcmk_resource_t *then_rsc, int then_role) { const char *first_tasks[] = { NULL, NULL }; const char *then_tasks[] = { NULL, NULL }; /* Actions to make first_rsc lose first_role */ if (first_role == pcmk_role_promoted) { first_tasks[0] = PCMK_ACTION_DEMOTE; } else { first_tasks[0] = PCMK_ACTION_STOP; if (first_role == pcmk_role_unpromoted) { first_tasks[1] = PCMK_ACTION_PROMOTE; } } /* Actions to make then_rsc gain then_role */ if (then_role == pcmk_role_promoted) { then_tasks[0] = PCMK_ACTION_PROMOTE; } else { then_tasks[0] = PCMK_ACTION_START; if (then_role == pcmk_role_unpromoted) { then_tasks[1] = PCMK_ACTION_DEMOTE; } } for (int first_lpc = 0; (first_lpc <= 1) && (first_tasks[first_lpc] != NULL); first_lpc++) { for (int then_lpc = 0; (then_lpc <= 1) && (then_tasks[then_lpc] != NULL); then_lpc++) { pcmk__order_resource_actions(first_rsc, first_tasks[first_lpc], then_rsc, then_tasks[then_lpc], pcmk__ar_if_required_on_same_node); } } } /*! * \internal * \brief Add a new colocation constraint to scheduler data * * \param[in] id XML ID for this constraint * \param[in] node_attr Colocate by this attribute (NULL for #uname) * \param[in] score Constraint score * \param[in,out] dependent Resource to be colocated * \param[in,out] primary Resource to colocate \p dependent with * \param[in] dependent_role_spec If not NULL, only \p dependent instances * with this role should be colocated * \param[in] primary_role_spec If not NULL, only \p primary instances * with this role should be colocated * \param[in] flags Group of enum pcmk__coloc_flags */ void pcmk__new_colocation(const char *id, const char *node_attr, int score, pcmk_resource_t *dependent, pcmk_resource_t *primary, const char *dependent_role_spec, const char *primary_role_spec, uint32_t flags) { pcmk__colocation_t *new_con = NULL; enum rsc_role_e dependent_role = pcmk_role_unknown; enum rsc_role_e primary_role = pcmk_role_unknown; CRM_CHECK(id != NULL, return); if ((dependent == NULL) || (primary == NULL)) { pcmk__config_err("Ignoring colocation '%s' because resource " "does not exist", id); return; } if ((pcmk__parse_constraint_role(id, dependent_role_spec, &dependent_role) != pcmk_rc_ok) || (pcmk__parse_constraint_role(id, primary_role_spec, &primary_role) != pcmk_rc_ok)) { // Not possible with schema validation enabled (error already logged) return; } if (score == 0) { pcmk__rsc_trace(dependent, "Ignoring colocation '%s' (%s with %s) because score is 0", id, dependent->id, primary->id); return; } new_con = pcmk__assert_alloc(1, sizeof(pcmk__colocation_t)); new_con->id = id; new_con->dependent = dependent; new_con->primary = primary; new_con->score = score; new_con->dependent_role = dependent_role; new_con->primary_role = primary_role; new_con->node_attribute = pcmk__s(node_attr, CRM_ATTR_UNAME); new_con->flags = flags; pcmk__add_this_with(&(dependent->priv->this_with_colocations), new_con, dependent); pcmk__add_with_this(&(primary->priv->with_this_colocations), new_con, primary); dependent->priv->scheduler->priv->colocation_constraints = g_list_prepend(dependent->priv->scheduler->priv->colocation_constraints, new_con); if (score <= -PCMK_SCORE_INFINITY) { anti_colocation_order(dependent, new_con->dependent_role, primary, new_con->primary_role); anti_colocation_order(primary, new_con->primary_role, dependent, new_con->dependent_role); } } /*! * \internal * \brief Return the boolean influence corresponding to configuration * * \param[in] coloc_id Colocation XML ID (for error logging) * \param[in] rsc Resource involved in constraint (for default) * \param[in] influence_s String value of \c PCMK_XA_INFLUENCE option * * \return \c pcmk__coloc_influence if string evaluates true, or string is * \c NULL or invalid and resource's \c PCMK_META_CRITICAL option * evaluates true, otherwise \c pcmk__coloc_none */ static uint32_t unpack_influence(const char *coloc_id, const pcmk_resource_t *rsc, const char *influence_s) { if (influence_s != NULL) { int influence_i = 0; if (crm_str_to_boolean(influence_s, &influence_i) < 0) { pcmk__config_err("Constraint '%s' has invalid value for " PCMK_XA_INFLUENCE " (using default)", coloc_id); } else { return (influence_i == 0)? pcmk__coloc_none : pcmk__coloc_influence; } } if (pcmk_is_set(rsc->flags, pcmk__rsc_critical)) { return pcmk__coloc_influence; } return pcmk__coloc_none; } static void unpack_colocation_set(xmlNode *set, int score, const char *coloc_id, const char *influence_s, pcmk_scheduler_t *scheduler) { xmlNode *xml_rsc = NULL; pcmk_resource_t *other = NULL; pcmk_resource_t *resource = NULL; const char *set_id = pcmk__xe_id(set); const char *role = crm_element_value(set, PCMK_XA_ROLE); bool with_previous = false; int local_score = score; bool sequential = false; uint32_t flags = pcmk__coloc_none; const char *xml_rsc_id = NULL; const char *score_s = crm_element_value(set, PCMK_XA_SCORE); if (score_s != NULL) { int rc = pcmk_parse_score(score_s, &local_score, 0); if (rc != pcmk_rc_ok) { // Not possible with schema validation enabled pcmk__config_err("Ignoring colocation '%s' for set '%s' " "because '%s' is not a valid score", coloc_id, set_id, score_s); return; } } if (local_score == 0) { crm_trace("Ignoring colocation '%s' for set '%s' because score is 0", coloc_id, set_id); return; } /* @COMPAT The deprecated PCMK__XA_ORDERING attribute specifies whether * resources in a positive-score set are colocated with the previous or next * resource. */ if (pcmk__str_eq(crm_element_value(set, PCMK__XA_ORDERING), PCMK__VALUE_GROUP, pcmk__str_null_matches|pcmk__str_casei)) { with_previous = true; } else { pcmk__warn_once(pcmk__wo_set_ordering, "Support for '" PCMK__XA_ORDERING "' other than" " '" PCMK__VALUE_GROUP "' in " PCMK_XE_RESOURCE_SET " (such as %s) is deprecated and will be removed in a" " future release", set_id); } if ((pcmk__xe_get_bool_attr(set, PCMK_XA_SEQUENTIAL, &sequential) == pcmk_rc_ok) && !sequential) { return; } if (local_score > 0) { for (xml_rsc = pcmk__xe_first_child(set, PCMK_XE_RESOURCE_REF, NULL, NULL); xml_rsc != NULL; xml_rsc = pcmk__xe_next(xml_rsc, PCMK_XE_RESOURCE_REF)) { xml_rsc_id = pcmk__xe_id(xml_rsc); resource = pcmk__find_constraint_resource(scheduler->priv->resources, xml_rsc_id); if (resource == NULL) { // Should be possible only with validation disabled pcmk__config_err("Ignoring %s and later resources in set %s: " "No such resource", xml_rsc_id, set_id); return; } if (other != NULL) { flags = pcmk__coloc_explicit | unpack_influence(coloc_id, resource, influence_s); if (with_previous) { pcmk__rsc_trace(resource, "Colocating %s with %s in set %s", resource->id, other->id, set_id); pcmk__new_colocation(set_id, NULL, local_score, resource, other, role, role, flags); } else { pcmk__rsc_trace(resource, "Colocating %s with %s in set %s", other->id, resource->id, set_id); pcmk__new_colocation(set_id, NULL, local_score, other, resource, role, role, flags); } } other = resource; } } else { /* Anti-colocating with every prior resource is * the only way to ensure the intuitive result * (i.e. that no one in the set can run with anyone else in the set) */ for (xml_rsc = pcmk__xe_first_child(set, PCMK_XE_RESOURCE_REF, NULL, NULL); xml_rsc != NULL; xml_rsc = pcmk__xe_next(xml_rsc, PCMK_XE_RESOURCE_REF)) { xmlNode *xml_rsc_with = NULL; xml_rsc_id = pcmk__xe_id(xml_rsc); resource = pcmk__find_constraint_resource(scheduler->priv->resources, xml_rsc_id); if (resource == NULL) { // Should be possible only with validation disabled pcmk__config_err("Ignoring %s and later resources in set %s: " "No such resource", xml_rsc_id, set_id); return; } flags = pcmk__coloc_explicit | unpack_influence(coloc_id, resource, influence_s); for (xml_rsc_with = pcmk__xe_first_child(set, PCMK_XE_RESOURCE_REF, NULL, NULL); xml_rsc_with != NULL; xml_rsc_with = pcmk__xe_next(xml_rsc_with, PCMK_XE_RESOURCE_REF)) { xml_rsc_id = pcmk__xe_id(xml_rsc_with); if (pcmk__str_eq(resource->id, xml_rsc_id, pcmk__str_none)) { break; } other = pcmk__find_constraint_resource(scheduler->priv->resources, xml_rsc_id); pcmk__assert(other != NULL); // We already processed it pcmk__new_colocation(set_id, NULL, local_score, resource, other, role, role, flags); } } } } /*! * \internal * \brief Colocate two resource sets relative to each other * * \param[in] id Colocation XML ID * \param[in] set1 Dependent set * \param[in] set2 Primary set * \param[in] score Colocation score * \param[in] influence_s Value of colocation's \c PCMK_XA_INFLUENCE * attribute * \param[in,out] scheduler Scheduler data */ static void colocate_rsc_sets(const char *id, const xmlNode *set1, const xmlNode *set2, int score, const char *influence_s, pcmk_scheduler_t *scheduler) { xmlNode *xml_rsc = NULL; pcmk_resource_t *rsc_1 = NULL; pcmk_resource_t *rsc_2 = NULL; const char *xml_rsc_id = NULL; const char *role_1 = crm_element_value(set1, PCMK_XA_ROLE); const char *role_2 = crm_element_value(set2, PCMK_XA_ROLE); int rc = pcmk_rc_ok; bool sequential = false; uint32_t flags = pcmk__coloc_none; if (score == 0) { crm_trace("Ignoring colocation '%s' between sets %s and %s " "because score is 0", id, pcmk__xe_id(set1), pcmk__xe_id(set2)); return; } rc = pcmk__xe_get_bool_attr(set1, PCMK_XA_SEQUENTIAL, &sequential); if ((rc != pcmk_rc_ok) || sequential) { // Get the first one xml_rsc = pcmk__xe_first_child(set1, PCMK_XE_RESOURCE_REF, NULL, NULL); if (xml_rsc != NULL) { xml_rsc_id = pcmk__xe_id(xml_rsc); rsc_1 = pcmk__find_constraint_resource(scheduler->priv->resources, xml_rsc_id); if (rsc_1 == NULL) { // Should be possible only with validation disabled pcmk__config_err("Ignoring colocation of set %s with set %s " "because first resource %s not found", pcmk__xe_id(set1), pcmk__xe_id(set2), xml_rsc_id); return; } } } rc = pcmk__xe_get_bool_attr(set2, PCMK_XA_SEQUENTIAL, &sequential); if ((rc != pcmk_rc_ok) || sequential) { // Get the last one for (xml_rsc = pcmk__xe_first_child(set2, PCMK_XE_RESOURCE_REF, NULL, NULL); xml_rsc != NULL; xml_rsc = pcmk__xe_next(xml_rsc, PCMK_XE_RESOURCE_REF)) { xml_rsc_id = pcmk__xe_id(xml_rsc); } rsc_2 = pcmk__find_constraint_resource(scheduler->priv->resources, xml_rsc_id); if (rsc_2 == NULL) { // Should be possible only with validation disabled pcmk__config_err("Ignoring colocation of set %s with set %s " "because last resource %s not found", pcmk__xe_id(set1), pcmk__xe_id(set2), xml_rsc_id); return; } } if ((rsc_1 != NULL) && (rsc_2 != NULL)) { // Both sets are sequential flags = pcmk__coloc_explicit | unpack_influence(id, rsc_1, influence_s); pcmk__new_colocation(id, NULL, score, rsc_1, rsc_2, role_1, role_2, flags); } else if (rsc_1 != NULL) { // Only set1 is sequential flags = pcmk__coloc_explicit | unpack_influence(id, rsc_1, influence_s); for (xml_rsc = pcmk__xe_first_child(set2, PCMK_XE_RESOURCE_REF, NULL, NULL); xml_rsc != NULL; xml_rsc = pcmk__xe_next(xml_rsc, PCMK_XE_RESOURCE_REF)) { xml_rsc_id = pcmk__xe_id(xml_rsc); rsc_2 = pcmk__find_constraint_resource(scheduler->priv->resources, xml_rsc_id); if (rsc_2 == NULL) { // Should be possible only with validation disabled pcmk__config_err("Ignoring set %s colocation with resource %s " "in set %s: No such resource", pcmk__xe_id(set1), xml_rsc_id, pcmk__xe_id(set2)); continue; } pcmk__new_colocation(id, NULL, score, rsc_1, rsc_2, role_1, role_2, flags); } } else if (rsc_2 != NULL) { // Only set2 is sequential for (xml_rsc = pcmk__xe_first_child(set1, PCMK_XE_RESOURCE_REF, NULL, NULL); xml_rsc != NULL; xml_rsc = pcmk__xe_next(xml_rsc, PCMK_XE_RESOURCE_REF)) { xml_rsc_id = pcmk__xe_id(xml_rsc); rsc_1 = pcmk__find_constraint_resource(scheduler->priv->resources, xml_rsc_id); if (rsc_1 == NULL) { // Should be possible only with validation disabled pcmk__config_err("Ignoring colocation of set %s resource %s " "with set %s: No such resource", pcmk__xe_id(set1), xml_rsc_id, pcmk__xe_id(set2)); continue; } flags = pcmk__coloc_explicit | unpack_influence(id, rsc_1, influence_s); pcmk__new_colocation(id, NULL, score, rsc_1, rsc_2, role_1, role_2, flags); } } else { // Neither set is sequential for (xml_rsc = pcmk__xe_first_child(set1, PCMK_XE_RESOURCE_REF, NULL, NULL); xml_rsc != NULL; xml_rsc = pcmk__xe_next(xml_rsc, PCMK_XE_RESOURCE_REF)) { xmlNode *xml_rsc_2 = NULL; xml_rsc_id = pcmk__xe_id(xml_rsc); rsc_1 = pcmk__find_constraint_resource(scheduler->priv->resources, xml_rsc_id); if (rsc_1 == NULL) { // Should be possible only with validation disabled pcmk__config_err("Ignoring colocation of set %s resource %s " "with set %s: No such resource", pcmk__xe_id(set1), xml_rsc_id, pcmk__xe_id(set2)); continue; } flags = pcmk__coloc_explicit | unpack_influence(id, rsc_1, influence_s); for (xml_rsc_2 = pcmk__xe_first_child(set2, PCMK_XE_RESOURCE_REF, NULL, NULL); xml_rsc_2 != NULL; xml_rsc_2 = pcmk__xe_next(xml_rsc_2, PCMK_XE_RESOURCE_REF)) { xml_rsc_id = pcmk__xe_id(xml_rsc_2); rsc_2 = pcmk__find_constraint_resource(scheduler->priv->resources, xml_rsc_id); if (rsc_2 == NULL) { // Should be possible only with validation disabled pcmk__config_err("Ignoring colocation of set %s resource " "%s with set %s resource %s: No such " "resource", pcmk__xe_id(set1), pcmk__xe_id(xml_rsc), pcmk__xe_id(set2), xml_rsc_id); continue; } pcmk__new_colocation(id, NULL, score, rsc_1, rsc_2, role_1, role_2, flags); } } } } /*! * \internal * \brief Unpack a colocation constraint that contains no resource sets * * \param[in] xml_obj Colocation constraint XML * \param[in] id Colocation constraint XML ID (non-NULL) * \param[in] score Integer score parsed from score attribute * \param[in] influence_s Colocation constraint's influence attribute value * \param[in,out] scheduler Scheduler data */ static void unpack_simple_colocation(const xmlNode *xml_obj, const char *id, int score, const char *influence_s, pcmk_scheduler_t *scheduler) { uint32_t flags = pcmk__coloc_none; const char *dependent_id = crm_element_value(xml_obj, PCMK_XA_RSC); const char *primary_id = crm_element_value(xml_obj, PCMK_XA_WITH_RSC); const char *dependent_role = crm_element_value(xml_obj, PCMK_XA_RSC_ROLE); const char *primary_role = crm_element_value(xml_obj, PCMK_XA_WITH_RSC_ROLE); const char *attr = crm_element_value(xml_obj, PCMK_XA_NODE_ATTRIBUTE); pcmk_resource_t *primary = NULL; pcmk_resource_t *dependent = NULL; primary = pcmk__find_constraint_resource(scheduler->priv->resources, primary_id); dependent = pcmk__find_constraint_resource(scheduler->priv->resources, dependent_id); if (dependent == NULL) { pcmk__config_err("Ignoring constraint '%s' because resource '%s' " "does not exist", id, dependent_id); return; } else if (primary == NULL) { pcmk__config_err("Ignoring constraint '%s' because resource '%s' " "does not exist", id, primary_id); return; } if (pcmk__xe_attr_is_true(xml_obj, PCMK_XA_SYMMETRICAL)) { pcmk__config_warn("The colocation constraint " "'" PCMK_XA_SYMMETRICAL "' attribute has been " "removed"); } flags = pcmk__coloc_explicit | unpack_influence(id, dependent, influence_s); pcmk__new_colocation(id, attr, score, dependent, primary, dependent_role, primary_role, flags); } // \return Standard Pacemaker return code static int unpack_colocation_tags(xmlNode *xml_obj, xmlNode **expanded_xml, pcmk_scheduler_t *scheduler) { const char *id = NULL; const char *dependent_id = NULL; const char *primary_id = NULL; const char *dependent_role = NULL; const char *primary_role = NULL; pcmk_resource_t *dependent = NULL; pcmk_resource_t *primary = NULL; pcmk__idref_t *dependent_tag = NULL; pcmk__idref_t *primary_tag = NULL; xmlNode *dependent_set = NULL; xmlNode *primary_set = NULL; bool any_sets = false; *expanded_xml = NULL; CRM_CHECK(xml_obj != NULL, return EINVAL); id = pcmk__xe_id(xml_obj); if (id == NULL) { pcmk__config_err("Ignoring <%s> constraint without " PCMK_XA_ID, xml_obj->name); return pcmk_rc_unpack_error; } // Check whether there are any resource sets with template or tag references *expanded_xml = pcmk__expand_tags_in_sets(xml_obj, scheduler); if (*expanded_xml != NULL) { crm_log_xml_trace(*expanded_xml, "Expanded " PCMK_XE_RSC_COLOCATION); return pcmk_rc_ok; } dependent_id = crm_element_value(xml_obj, PCMK_XA_RSC); primary_id = crm_element_value(xml_obj, PCMK_XA_WITH_RSC); if ((dependent_id == NULL) || (primary_id == NULL)) { return pcmk_rc_ok; } if (!pcmk__valid_resource_or_tag(scheduler, dependent_id, &dependent, &dependent_tag)) { pcmk__config_err("Ignoring constraint '%s' because '%s' is not a " "valid resource or tag", id, dependent_id); return pcmk_rc_unpack_error; } if (!pcmk__valid_resource_or_tag(scheduler, primary_id, &primary, &primary_tag)) { pcmk__config_err("Ignoring constraint '%s' because '%s' is not a " "valid resource or tag", id, primary_id); return pcmk_rc_unpack_error; } if ((dependent != NULL) && (primary != NULL)) { /* Neither side references any template/tag. */ return pcmk_rc_ok; } if ((dependent_tag != NULL) && (primary_tag != NULL)) { // A colocation constraint between two templates/tags makes no sense pcmk__config_err("Ignoring constraint '%s' because two templates or " "tags cannot be colocated", id); return pcmk_rc_unpack_error; } dependent_role = crm_element_value(xml_obj, PCMK_XA_RSC_ROLE); primary_role = crm_element_value(xml_obj, PCMK_XA_WITH_RSC_ROLE); *expanded_xml = pcmk__xml_copy(NULL, xml_obj); /* Convert dependent's template/tag reference into constraint * PCMK_XE_RESOURCE_SET */ if (!pcmk__tag_to_set(*expanded_xml, &dependent_set, PCMK_XA_RSC, true, scheduler)) { pcmk__xml_free(*expanded_xml); *expanded_xml = NULL; return pcmk_rc_unpack_error; } if (dependent_set != NULL) { if (dependent_role != NULL) { /* Move PCMK_XA_RSC_ROLE into converted PCMK_XE_RESOURCE_SET as * PCMK_XA_ROLE */ crm_xml_add(dependent_set, PCMK_XA_ROLE, dependent_role); pcmk__xe_remove_attr(*expanded_xml, PCMK_XA_RSC_ROLE); } any_sets = true; } /* Convert primary's template/tag reference into constraint * PCMK_XE_RESOURCE_SET */ if (!pcmk__tag_to_set(*expanded_xml, &primary_set, PCMK_XA_WITH_RSC, true, scheduler)) { pcmk__xml_free(*expanded_xml); *expanded_xml = NULL; return pcmk_rc_unpack_error; } if (primary_set != NULL) { if (primary_role != NULL) { /* Move PCMK_XA_WITH_RSC_ROLE into converted PCMK_XE_RESOURCE_SET as * PCMK_XA_ROLE */ crm_xml_add(primary_set, PCMK_XA_ROLE, primary_role); pcmk__xe_remove_attr(*expanded_xml, PCMK_XA_WITH_RSC_ROLE); } any_sets = true; } if (any_sets) { crm_log_xml_trace(*expanded_xml, "Expanded " PCMK_XE_RSC_COLOCATION); } else { pcmk__xml_free(*expanded_xml); *expanded_xml = NULL; } return pcmk_rc_ok; } /*! * \internal * \brief Parse a colocation constraint from XML into scheduler data * * \param[in,out] xml_obj Colocation constraint XML to unpack * \param[in,out] scheduler Scheduler data to add constraint to */ void pcmk__unpack_colocation(xmlNode *xml_obj, pcmk_scheduler_t *scheduler) { int score_i = 0; xmlNode *set = NULL; xmlNode *last = NULL; xmlNode *orig_xml = NULL; xmlNode *expanded_xml = NULL; const char *id = crm_element_value(xml_obj, PCMK_XA_ID); const char *score = NULL; const char *influence_s = NULL; if (pcmk__str_empty(id)) { pcmk__config_err("Ignoring " PCMK_XE_RSC_COLOCATION " without " CRM_ATTR_ID); return; } if (unpack_colocation_tags(xml_obj, &expanded_xml, scheduler) != pcmk_rc_ok) { return; } if (expanded_xml != NULL) { orig_xml = xml_obj; xml_obj = expanded_xml; } score = crm_element_value(xml_obj, PCMK_XA_SCORE); if (score != NULL) { int rc = pcmk_parse_score(score, &score_i, 0); if (rc != pcmk_rc_ok) { // Not possible with schema validation enabled pcmk__config_err("Ignoring colocation %s because '%s' " "is not a valid score", id, score); return; } } influence_s = crm_element_value(xml_obj, PCMK_XA_INFLUENCE); for (set = pcmk__xe_first_child(xml_obj, PCMK_XE_RESOURCE_SET, NULL, NULL); set != NULL; set = pcmk__xe_next(set, PCMK_XE_RESOURCE_SET)) { set = pcmk__xe_resolve_idref(set, scheduler->input); if (set == NULL) { // Configuration error, message already logged if (expanded_xml != NULL) { pcmk__xml_free(expanded_xml); } return; } if (pcmk__str_empty(pcmk__xe_id(set))) { pcmk__config_err("Ignoring " PCMK_XE_RESOURCE_SET " without " CRM_ATTR_ID); continue; } unpack_colocation_set(set, score_i, id, influence_s, scheduler); if (last != NULL) { colocate_rsc_sets(id, last, set, score_i, influence_s, scheduler); } last = set; } if (expanded_xml) { pcmk__xml_free(expanded_xml); xml_obj = orig_xml; } if (last == NULL) { unpack_simple_colocation(xml_obj, id, score_i, influence_s, scheduler); } } /*! * \internal * \brief Check whether colocation's dependent preferences should be considered * * \param[in] colocation Colocation constraint * \param[in] rsc Primary instance (normally this will be * colocation->primary, which NULL will be treated as, * but for clones or bundles with multiple instances * this can be a particular instance) * * \return true if colocation influence should be effective, otherwise false */ bool pcmk__colocation_has_influence(const pcmk__colocation_t *colocation, const pcmk_resource_t *rsc) { if (rsc == NULL) { rsc = colocation->primary; } /* A bundle replica colocates its remote connection with its container, * using a finite score so that the container can run on Pacemaker Remote * nodes. * * Moving a connection is lightweight and does not interrupt the service, * while moving a container is heavyweight and does interrupt the service, * so don't move a clean, active container based solely on the preferences * of its connection. * * This also avoids problematic scenarios where two containers want to * perpetually swap places. */ if (pcmk_is_set(colocation->dependent->flags, pcmk__rsc_remote_nesting_allowed) && !pcmk_is_set(rsc->flags, pcmk__rsc_failed) && pcmk__list_of_1(rsc->priv->active_nodes)) { return false; } /* The dependent in a colocation influences the primary's location * if the PCMK_XA_INFLUENCE option is true or the primary is not yet active. */ return pcmk_is_set(colocation->flags, pcmk__coloc_influence) || (rsc->priv->active_nodes == NULL); } /*! * \internal * \brief Make actions of a given type unrunnable for a given resource * * \param[in,out] rsc Resource whose actions should be blocked * \param[in] task Name of action to block * \param[in] reason Unrunnable start action causing the block */ static void mark_action_blocked(pcmk_resource_t *rsc, const char *task, const pcmk_resource_t *reason) { GList *iter = NULL; char *reason_text = crm_strdup_printf("colocation with %s", reason->id); for (iter = rsc->priv->actions; iter != NULL; iter = iter->next) { pcmk_action_t *action = iter->data; if (pcmk_is_set(action->flags, pcmk__action_runnable) && pcmk__str_eq(action->task, task, pcmk__str_none)) { pcmk__clear_action_flags(action, pcmk__action_runnable); pe_action_set_reason(action, reason_text, false); pcmk__block_colocation_dependents(action); pcmk__update_action_for_orderings(action, rsc->priv->scheduler); } } // If parent resource can't perform an action, neither can any children for (iter = rsc->priv->children; iter != NULL; iter = iter->next) { mark_action_blocked((pcmk_resource_t *) (iter->data), task, reason); } free(reason_text); } /*! * \internal * \brief If an action is unrunnable, block any relevant dependent actions * * If a given action is an unrunnable start or promote, block the start or * promote actions of resources colocated with it, as appropriate to the * colocations' configured roles. * * \param[in,out] action Action to check */ void pcmk__block_colocation_dependents(pcmk_action_t *action) { GList *iter = NULL; GList *colocations = NULL; pcmk_resource_t *rsc = NULL; bool is_start = false; if (pcmk_is_set(action->flags, pcmk__action_runnable)) { return; // Only unrunnable actions block dependents } is_start = pcmk__str_eq(action->task, PCMK_ACTION_START, pcmk__str_none); if (!is_start && !pcmk__str_eq(action->task, PCMK_ACTION_PROMOTE, pcmk__str_none)) { return; // Only unrunnable starts and promotes block dependents } pcmk__assert(action->rsc != NULL); // Start and promote are resource actions /* If this resource is part of a collective resource, dependents are blocked * only if all instances of the collective are unrunnable, so check the * collective resource. */ rsc = uber_parent(action->rsc); if (rsc->priv->parent != NULL) { rsc = rsc->priv->parent; // Bundle } // Colocation fails only if entire primary can't reach desired role for (iter = rsc->priv->children; iter != NULL; iter = iter->next) { pcmk_resource_t *child = iter->data; pcmk_action_t *child_action = NULL; child_action = find_first_action(child->priv->actions, NULL, action->task, NULL); if ((child_action == NULL) || pcmk_is_set(child_action->flags, pcmk__action_runnable)) { crm_trace("Not blocking %s colocation dependents because " "at least %s has runnable %s", rsc->id, child->id, action->task); return; // At least one child can reach desired role } } crm_trace("Blocking %s colocation dependents due to unrunnable %s %s", rsc->id, action->rsc->id, action->task); // Check each colocation where this resource is primary colocations = pcmk__with_this_colocations(rsc); for (iter = colocations; iter != NULL; iter = iter->next) { pcmk__colocation_t *colocation = iter->data; if (colocation->score < PCMK_SCORE_INFINITY) { continue; // Only mandatory colocations block dependent } /* If the primary can't start, the dependent can't reach its colocated * role, regardless of what the primary or dependent colocation role is. * * If the primary can't be promoted, the dependent can't reach its * colocated role if the primary's colocation role is promoted. */ if (!is_start && (colocation->primary_role != pcmk_role_promoted)) { continue; } // Block the dependent from reaching its colocated role if (colocation->dependent_role == pcmk_role_promoted) { mark_action_blocked(colocation->dependent, PCMK_ACTION_PROMOTE, action->rsc); } else { mark_action_blocked(colocation->dependent, PCMK_ACTION_START, action->rsc); } } g_list_free(colocations); } /*! * \internal * \brief Get the resource to use for role comparisons * * A bundle replica includes a container and possibly an instance of the bundled * resource. The dependent in a "with bundle" colocation is colocated with a * particular bundle container. However, if the colocation includes a role, then * the role must be checked on the bundled resource instance inside the * container. The container itself will never be promoted; the bundled resource * may be. * * If the given resource is a bundle replica container, return the resource * inside it, if any. Otherwise, return the resource itself. * * \param[in] rsc Resource to check * * \return Resource to use for role comparisons */ static const pcmk_resource_t * get_resource_for_role(const pcmk_resource_t *rsc) { if (pcmk_is_set(rsc->flags, pcmk__rsc_replica_container)) { const pcmk_resource_t *child = pe__get_rsc_in_container(rsc); if (child != NULL) { return child; } } return rsc; } /*! * \internal * \brief Determine how a colocation constraint should affect a resource * * Colocation constraints have different effects at different points in the * scheduler sequence. Initially, they affect a resource's location; once that * is determined, then for promotable clones they can affect a resource * instance's role; after both are determined, the constraints no longer matter. * Given a specific colocation constraint, check what has been done so far to * determine what should be affected at the current point in the scheduler. * * \param[in] dependent Dependent resource in colocation * \param[in] primary Primary resource in colocation * \param[in] colocation Colocation constraint * \param[in] preview If true, pretend resources have already been assigned * * \return How colocation constraint should be applied at this point */ enum pcmk__coloc_affects pcmk__colocation_affects(const pcmk_resource_t *dependent, const pcmk_resource_t *primary, const pcmk__colocation_t *colocation, bool preview) { const pcmk_resource_t *dependent_role_rsc = NULL; const pcmk_resource_t *primary_role_rsc = NULL; pcmk__assert((dependent != NULL) && (primary != NULL) && (colocation != NULL)); if (!preview && pcmk_is_set(primary->flags, pcmk__rsc_unassigned)) { // Primary resource has not been assigned yet, so we can't do anything return pcmk__coloc_affects_nothing; } dependent_role_rsc = get_resource_for_role(dependent); primary_role_rsc = get_resource_for_role(primary); if ((colocation->dependent_role >= pcmk_role_unpromoted) && (dependent_role_rsc->priv->parent != NULL) && pcmk_is_set(dependent_role_rsc->priv->parent->flags, pcmk__rsc_promotable) && !pcmk_is_set(dependent_role_rsc->flags, pcmk__rsc_unassigned)) { /* This is a colocation by role, and the dependent is a promotable clone * that has already been assigned, so the colocation should now affect * the role. */ return pcmk__coloc_affects_role; } if (!preview && !pcmk_is_set(dependent->flags, pcmk__rsc_unassigned)) { /* The dependent resource has already been through assignment, so the * constraint no longer matters. */ return pcmk__coloc_affects_nothing; } if ((colocation->dependent_role != pcmk_role_unknown) && (colocation->dependent_role != dependent_role_rsc->priv->next_role)) { crm_trace("Skipping %scolocation '%s': dependent limited to %s role " "but %s next role is %s", ((colocation->score < 0)? "anti-" : ""), colocation->id, pcmk_role_text(colocation->dependent_role), dependent_role_rsc->id, pcmk_role_text(dependent_role_rsc->priv->next_role)); return pcmk__coloc_affects_nothing; } if ((colocation->primary_role != pcmk_role_unknown) && (colocation->primary_role != primary_role_rsc->priv->next_role)) { crm_trace("Skipping %scolocation '%s': primary limited to %s role " "but %s next role is %s", ((colocation->score < 0)? "anti-" : ""), colocation->id, pcmk_role_text(colocation->primary_role), primary_role_rsc->id, pcmk_role_text(primary_role_rsc->priv->next_role)); return pcmk__coloc_affects_nothing; } return pcmk__coloc_affects_location; } /*! * \internal * \brief Apply colocation to dependent for assignment purposes * * Update the allowed node scores of the dependent resource in a colocation, * for the purposes of assigning it to a node. * * \param[in,out] dependent Dependent resource in colocation * \param[in] primary Primary resource in colocation * \param[in] colocation Colocation constraint */ void pcmk__apply_coloc_to_scores(pcmk_resource_t *dependent, const pcmk_resource_t *primary, const pcmk__colocation_t *colocation) { const char *attr = colocation->node_attribute; const char *value = NULL; GHashTable *work = NULL; GHashTableIter iter; pcmk_node_t *node = NULL; if (primary->priv->assigned_node != NULL) { value = pcmk__colocation_node_attr(primary->priv->assigned_node, attr, primary); } else if (colocation->score < 0) { // Nothing to do (anti-colocation with something that is not running) return; } work = pcmk__copy_node_table(dependent->priv->allowed_nodes); g_hash_table_iter_init(&iter, work); while (g_hash_table_iter_next(&iter, NULL, (void **)&node)) { if (primary->priv->assigned_node == NULL) { node->assign->score = pcmk__add_scores(-colocation->score, node->assign->score); pcmk__rsc_trace(dependent, "Applied %s to %s score on %s (now %s after " "subtracting %s because primary %s inactive)", colocation->id, dependent->id, pcmk__node_name(node), pcmk_readable_score(node->assign->score), pcmk_readable_score(colocation->score), primary->id); continue; } if (pcmk__str_eq(pcmk__colocation_node_attr(node, attr, dependent), value, pcmk__str_casei)) { /* Add colocation score only if optional (or minus infinity). A * mandatory colocation is a requirement rather than a preference, * so we don't need to consider it for relative assignment purposes. * The resource will simply be forbidden from running on the node if * the primary isn't active there (via the condition above). */ if (colocation->score < PCMK_SCORE_INFINITY) { node->assign->score = pcmk__add_scores(colocation->score, node->assign->score); pcmk__rsc_trace(dependent, "Applied %s to %s score on %s (now %s after " "adding %s)", colocation->id, dependent->id, pcmk__node_name(node), pcmk_readable_score(node->assign->score), pcmk_readable_score(colocation->score)); } continue; } if (colocation->score >= PCMK_SCORE_INFINITY) { /* Only mandatory colocations are relevant when the colocation * attribute doesn't match, because an attribute not matching is not * a negative preference -- the colocation is simply relevant only * where it matches. */ node->assign->score = -PCMK_SCORE_INFINITY; pcmk__rsc_trace(dependent, "Banned %s from %s because colocation %s attribute %s " "does not match", dependent->id, pcmk__node_name(node), colocation->id, attr); } } if ((colocation->score <= -PCMK_SCORE_INFINITY) || (colocation->score >= PCMK_SCORE_INFINITY) || pcmk__any_node_available(work)) { g_hash_table_destroy(dependent->priv->allowed_nodes); dependent->priv->allowed_nodes = work; work = NULL; } else { pcmk__rsc_info(dependent, "%s: Rolling back scores from %s (no available nodes)", dependent->id, primary->id); } if (work != NULL) { g_hash_table_destroy(work); } } /*! * \internal * \brief Apply colocation to dependent for role purposes * * Update the priority of the dependent resource in a colocation, for the * purposes of selecting its role * * \param[in,out] dependent Dependent resource in colocation * \param[in] primary Primary resource in colocation * \param[in] colocation Colocation constraint * * \return The score added to the dependent's priority */ int pcmk__apply_coloc_to_priority(pcmk_resource_t *dependent, const pcmk_resource_t *primary, const pcmk__colocation_t *colocation) { const char *dependent_value = NULL; const char *primary_value = NULL; const char *attr = colocation->node_attribute; int score_multiplier = 1; int priority_delta = 0; const pcmk_node_t *primary_node = NULL; const pcmk_node_t *dependent_node = NULL; pcmk__assert((dependent != NULL) && (primary != NULL) && (colocation != NULL)); primary_node = primary->priv->assigned_node; dependent_node = dependent->priv->assigned_node; if (dependent_node == NULL) { return 0; } if ((primary_node != NULL) && (colocation->primary_role != pcmk_role_unknown)) { /* Colocation applies only if the primary's next role matches. * * If primary_node == NULL, we want to proceed past this block, so that * dependent_node is marked ineligible for promotion. * * @TODO Why ignore a mandatory colocation in this case when we apply * its negation in the mismatched value case? */ const pcmk_resource_t *role_rsc = get_resource_for_role(primary); if (colocation->primary_role != role_rsc->priv->next_role) { return 0; } } dependent_value = pcmk__colocation_node_attr(dependent_node, attr, dependent); primary_value = pcmk__colocation_node_attr(primary_node, attr, primary); if (!pcmk__str_eq(dependent_value, primary_value, pcmk__str_casei)) { if ((colocation->score == PCMK_SCORE_INFINITY) && (colocation->dependent_role == pcmk_role_promoted)) { /* For a mandatory promoted-role colocation, mark the dependent node * ineligible to promote the dependent if its attribute value * doesn't match the primary node's */ score_multiplier = -1; } else { // Otherwise, ignore the colocation if attribute values don't match return 0; } } else if (colocation->dependent_role == pcmk_role_unpromoted) { /* Node attribute values matched, so we want to avoid promoting the * dependent on this node */ score_multiplier = -1; } priority_delta = score_multiplier * colocation->score; dependent->priv->priority = pcmk__add_scores(priority_delta, dependent->priv->priority); pcmk__rsc_trace(dependent, "Applied %s to %s promotion priority (now %s after %s %d)", colocation->id, dependent->id, pcmk_readable_score(dependent->priv->priority), ((score_multiplier == 1)? "adding" : "subtracting"), colocation->score); return priority_delta; } /*! * \internal * \brief Find score of highest-scored node that matches colocation attribute * * \param[in] colocation Colocation constraint being applied * \param[in,out] rsc Resource whose allowed nodes should be searched * \param[in] attr Colocation attribute name (must not be NULL) * \param[in] value Colocation attribute value to require */ static int best_node_score_matching_attr(const pcmk__colocation_t *colocation, pcmk_resource_t *rsc, const char *attr, const char *value) { GHashTable *allowed_nodes_orig = NULL; GHashTableIter iter; pcmk_node_t *node = NULL; int best_score = -PCMK_SCORE_INFINITY; const char *best_node = NULL; if ((colocation != NULL) && (rsc == colocation->dependent) && pcmk_is_set(colocation->flags, pcmk__coloc_explicit) && pcmk__is_group(rsc->priv->parent) && (rsc != rsc->priv->parent->priv->children->data)) { /* The resource is a user-configured colocation's explicit dependent, * and a group member other than the first, which means the group's * location constraint scores were not applied to it (see * pcmk__group_apply_location()). Explicitly consider those scores now. * * @TODO This does leave one suboptimal case: if the group itself or * another member other than the first is explicitly colocated with * the same primary, the primary will count the group's location scores * multiple times. This is much less likely than a single member being * explicitly colocated, so it's an acceptable tradeoff for now. */ allowed_nodes_orig = rsc->priv->allowed_nodes; rsc->priv->allowed_nodes = pcmk__copy_node_table(allowed_nodes_orig); for (GList *loc_iter = rsc->priv->scheduler->priv->location_constraints; loc_iter != NULL; loc_iter = loc_iter->next) { pcmk__location_t *location = loc_iter->data; if (location->rsc == rsc->priv->parent) { rsc->priv->cmds->apply_location(rsc, location); } } } // Find best allowed node with matching attribute g_hash_table_iter_init(&iter, rsc->priv->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (void **) &node)) { if ((node->assign->score > best_score) - && pcmk__node_available(node, false, false) + && pcmk__node_available(node, pcmk__node_alive|pcmk__node_usable) && pcmk__str_eq(value, pcmk__colocation_node_attr(node, attr, rsc), pcmk__str_casei)) { best_score = node->assign->score; best_node = node->priv->name; } } if (!pcmk__str_eq(attr, CRM_ATTR_UNAME, pcmk__str_none)) { if (best_node == NULL) { crm_info("No allowed node for %s matches node attribute %s=%s", rsc->id, attr, value); } else { crm_info("Allowed node %s for %s had best score (%d) " "of those matching node attribute %s=%s", best_node, rsc->id, best_score, attr, value); } } if (allowed_nodes_orig != NULL) { g_hash_table_destroy(rsc->priv->allowed_nodes); rsc->priv->allowed_nodes = allowed_nodes_orig; } return best_score; } /*! * \internal * \brief Check whether a resource is allowed only on a single node * * \param[in] rsc Resource to check * * \return \c true if \p rsc is allowed only on one node, otherwise \c false */ static bool allowed_on_one(const pcmk_resource_t *rsc) { GHashTableIter iter; pcmk_node_t *allowed_node = NULL; int allowed_nodes = 0; g_hash_table_iter_init(&iter, rsc->priv->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (gpointer *) &allowed_node)) { if ((allowed_node->assign->score >= 0) && (++allowed_nodes > 1)) { pcmk__rsc_trace(rsc, "%s is allowed on multiple nodes", rsc->id); return false; } } pcmk__rsc_trace(rsc, "%s is allowed %s", rsc->id, ((allowed_nodes == 1)? "on a single node" : "nowhere")); return (allowed_nodes == 1); } /*! * \internal * \brief Add resource's colocation matches to current node assignment scores * * For each node in a given table, if any of a given resource's allowed nodes * have a matching value for the colocation attribute, add the highest of those * nodes' scores to the node's score. * * \param[in,out] nodes Table of nodes with assignment scores so far * \param[in,out] source_rsc Resource whose node scores to add * \param[in] target_rsc Resource on whose behalf to update \p nodes * \param[in] colocation Original colocation constraint (used to get * configured primary resource's stickiness, and * to get colocation node attribute; pass NULL to * ignore stickiness and use default attribute) * \param[in] factor Factor by which to multiply scores being added * \param[in] only_positive Whether to add only positive scores */ static void add_node_scores_matching_attr(GHashTable *nodes, pcmk_resource_t *source_rsc, const pcmk_resource_t *target_rsc, const pcmk__colocation_t *colocation, float factor, bool only_positive) { GHashTableIter iter; pcmk_node_t *node = NULL; const char *attr = colocation->node_attribute; // Iterate through each node g_hash_table_iter_init(&iter, nodes); while (g_hash_table_iter_next(&iter, NULL, (void **)&node)) { float delta_f = 0; int delta = 0; int score = 0; int new_score = 0; const char *value = pcmk__colocation_node_attr(node, attr, target_rsc); score = best_node_score_matching_attr(colocation, source_rsc, attr, value); if ((factor < 0) && (score < 0)) { /* If the dependent is anti-colocated, we generally don't want the * primary to prefer nodes that the dependent avoids. That could * lead to unnecessary shuffling of the primary when the dependent * hits its migration threshold somewhere, for example. * * However, there are cases when it is desirable. If the dependent * can't run anywhere but where the primary is, it would be * worthwhile to move the primary for the sake of keeping the * dependent active. * * We can't know that exactly at this point since we don't know * where the primary will be assigned, but we can limit considering * the preference to when the dependent is allowed only on one node. * This is less than ideal for multiple reasons: * * - the dependent could be allowed on more than one node but have * anti-colocation primaries on each; * - the dependent could be a clone or bundle with multiple * instances, and the dependent as a whole is allowed on multiple * nodes but some instance still can't run * - the dependent has considered node-specific criteria such as * location constraints and stickiness by this point, but might * have other factors that end up disallowing a node * * but the alternative is making the primary move when it doesn't * need to. * * We also consider the primary's stickiness and influence, so the * user has some say in the matter. (This is the configured primary, * not a particular instance of the primary, but that doesn't matter * unless stickiness uses a rule to vary by node, and that seems * acceptable to ignore.) */ if ((colocation->primary->priv->stickiness >= -score) || !pcmk__colocation_has_influence(colocation, NULL) || !allowed_on_one(colocation->dependent)) { crm_trace("%s: Filtering %d + %f * %d " "(double negative disallowed)", pcmk__node_name(node), node->assign->score, factor, score); continue; } } if (node->assign->score == INFINITY_HACK) { crm_trace("%s: Filtering %d + %f * %d (node was marked unusable)", pcmk__node_name(node), node->assign->score, factor, score); continue; } delta_f = factor * score; // Round the number; see http://c-faq.com/fp/round.html delta = (int) ((delta_f < 0)? (delta_f - 0.5) : (delta_f + 0.5)); /* Small factors can obliterate the small scores that are often actually * used in configurations. If the score and factor are nonzero, ensure * that the result is nonzero as well. */ if ((delta == 0) && (score != 0)) { if (factor > 0.0) { delta = 1; } else if (factor < 0.0) { delta = -1; } } new_score = pcmk__add_scores(delta, node->assign->score); if (only_positive && (new_score < 0) && (node->assign->score > 0)) { crm_trace("%s: Filtering %d + %f * %d = %d " "(negative disallowed, marking node unusable)", pcmk__node_name(node), node->assign->score, factor, score, new_score); node->assign->score = INFINITY_HACK; continue; } if (only_positive && (new_score < 0) && (node->assign->score == 0)) { crm_trace("%s: Filtering %d + %f * %d = %d (negative disallowed)", pcmk__node_name(node), node->assign->score, factor, score, new_score); continue; } crm_trace("%s: %d + %f * %d = %d", pcmk__node_name(node), node->assign->score, factor, score, new_score); node->assign->score = new_score; } } /*! * \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 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 shared implementation of * \c pcmk__assignment_methods_t:add_colocated_node_scores(). */ void pcmk__add_colocated_node_scores(pcmk_resource_t *source_rsc, const pcmk_resource_t *target_rsc, const char *log_id, GHashTable **nodes, const pcmk__colocation_t *colocation, float factor, uint32_t flags) { GHashTable *work = NULL; pcmk__assert((source_rsc != NULL) && (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)) { pcmk__rsc_info(source_rsc, "%s: Breaking dependency loop at %s", log_id, source_rsc->id); return; } pcmk__set_rsc_flags(source_rsc, pcmk__rsc_updating_nodes); if (*nodes == NULL) { work = pcmk__copy_node_table(source_rsc->priv->allowed_nodes); target_rsc = source_rsc; } else { const bool pos = pcmk_is_set(flags, pcmk__coloc_select_nonnegative); pcmk__rsc_trace(source_rsc, "%s: Merging %s scores from %s (at %.6f)", log_id, (pos? "positive" : "all"), source_rsc->id, factor); work = pcmk__copy_node_table(*nodes); add_node_scores_matching_attr(work, source_rsc, target_rsc, colocation, factor, pos); } if (work == NULL) { pcmk__clear_rsc_flags(source_rsc, pcmk__rsc_updating_nodes); return; } if (pcmk__any_node_available(work)) { GList *colocations = NULL; if (pcmk_is_set(flags, pcmk__coloc_select_this_with)) { colocations = pcmk__this_with_colocations(source_rsc); pcmk__rsc_trace(source_rsc, "Checking additional %d optional '%s with' " "constraints", g_list_length(colocations), source_rsc->id); } else { colocations = pcmk__with_this_colocations(source_rsc); pcmk__rsc_trace(source_rsc, "Checking additional %d optional 'with %s' " "constraints", g_list_length(colocations), source_rsc->id); } flags |= pcmk__coloc_select_active; for (GList *iter = colocations; iter != NULL; iter = iter->next) { pcmk__colocation_t *constraint = iter->data; pcmk_resource_t *other = NULL; float other_factor = factor * constraint->score / (float) PCMK_SCORE_INFINITY; if (pcmk_is_set(flags, pcmk__coloc_select_this_with)) { other = constraint->primary; } else if (!pcmk__colocation_has_influence(constraint, NULL)) { continue; } else { other = constraint->dependent; } pcmk__rsc_trace(source_rsc, "Optionally merging score of '%s' constraint " "(%s with %s)", constraint->id, constraint->dependent->id, constraint->primary->id); other->priv->cmds->add_colocated_node_scores(other, target_rsc, log_id, &work, constraint, other_factor, flags); pe__show_node_scores(true, NULL, log_id, work, source_rsc->priv->scheduler); } g_list_free(colocations); } else if (pcmk_is_set(flags, pcmk__coloc_select_active)) { pcmk__rsc_info(source_rsc, "%s: Rolling back optional scores from %s", log_id, source_rsc->id); g_hash_table_destroy(work); pcmk__clear_rsc_flags(source_rsc, pcmk__rsc_updating_nodes); return; } if (pcmk_is_set(flags, pcmk__coloc_select_nonnegative)) { pcmk_node_t *node = NULL; GHashTableIter iter; g_hash_table_iter_init(&iter, work); while (g_hash_table_iter_next(&iter, NULL, (void **)&node)) { if (node->assign->score == INFINITY_HACK) { node->assign->score = 1; } } } if (*nodes != NULL) { g_hash_table_destroy(*nodes); } *nodes = work; pcmk__clear_rsc_flags(source_rsc, pcmk__rsc_updating_nodes); } /*! * \internal * \brief Apply a "with this" colocation to a resource's allowed node scores * * \param[in,out] data Colocation to apply * \param[in,out] user_data Resource being assigned */ void pcmk__add_dependent_scores(gpointer data, gpointer user_data) { pcmk__colocation_t *colocation = data; pcmk_resource_t *primary = user_data; pcmk_resource_t *dependent = colocation->dependent; const float factor = colocation->score / (float) PCMK_SCORE_INFINITY; uint32_t flags = pcmk__coloc_select_active; if (!pcmk__colocation_has_influence(colocation, NULL)) { return; } if (pcmk__is_clone(primary)) { flags |= pcmk__coloc_select_nonnegative; } pcmk__rsc_trace(primary, "%s: Incorporating attenuated %s assignment scores due " "to colocation %s", primary->id, dependent->id, colocation->id); dependent->priv->cmds->add_colocated_node_scores(dependent, primary, dependent->id, &(primary->priv->allowed_nodes), colocation, factor, flags); } /*! * \internal * \brief Exclude nodes from a dependent's node table if not in a given list * * Given a dependent resource in a colocation and a list of nodes where the * primary resource will run, set a node's score to \c -INFINITY in the * dependent's node table if not found in the primary nodes list. * * \param[in,out] dependent Dependent resource * \param[in] primary Primary resource (for logging only) * \param[in] colocation Colocation constraint (for logging only) * \param[in] primary_nodes List of nodes where the primary will have * unblocked instances in a suitable role * \param[in] merge_scores If \c true and a node is found in both \p table * and \p list, add the node's score in \p list to * the node's score in \p table */ void pcmk__colocation_intersect_nodes(pcmk_resource_t *dependent, const pcmk_resource_t *primary, const pcmk__colocation_t *colocation, const GList *primary_nodes, bool merge_scores) { GHashTableIter iter; pcmk_node_t *dependent_node = NULL; pcmk__assert((dependent != NULL) && (primary != NULL) && (colocation != NULL)); g_hash_table_iter_init(&iter, dependent->priv->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (gpointer *) &dependent_node)) { const pcmk_node_t *primary_node = NULL; primary_node = pe_find_node_id(primary_nodes, dependent_node->priv->id); if (primary_node == NULL) { dependent_node->assign->score = -PCMK_SCORE_INFINITY; pcmk__rsc_trace(dependent, "Banning %s from %s (no primary instance) for %s", dependent->id, pcmk__node_name(dependent_node), colocation->id); } else if (merge_scores) { dependent_node->assign->score = pcmk__add_scores(dependent_node->assign->score, primary_node->assign->score); pcmk__rsc_trace(dependent, "Added %s's score %s to %s's score for %s (now %d) " "for colocation %s", primary->id, pcmk_readable_score(primary_node->assign->score), dependent->id, pcmk__node_name(dependent_node), dependent_node->assign->score, colocation->id); } } } /*! * \internal * \brief Get all colocations affecting a resource as the primary * * \param[in] rsc Resource to get colocations for * * \return Newly allocated list of colocations affecting \p rsc as primary * * \note This is a convenience wrapper for the with_this_colocations() method. */ GList * pcmk__with_this_colocations(const pcmk_resource_t *rsc) { GList *list = NULL; rsc->priv->cmds->with_this_colocations(rsc, rsc, &list); return list; } /*! * \internal * \brief Get all colocations affecting a resource as the dependent * * \param[in] rsc Resource to get colocations for * * \return Newly allocated list of colocations affecting \p rsc as dependent * * \note This is a convenience wrapper for the this_with_colocations() method. */ GList * pcmk__this_with_colocations(const pcmk_resource_t *rsc) { GList *list = NULL; rsc->priv->cmds->this_with_colocations(rsc, rsc, &list); return list; } diff --git a/lib/pacemaker/pcmk_sched_instances.c b/lib/pacemaker/pcmk_sched_instances.c index f2bc1a432c..af53fc0726 100644 --- a/lib/pacemaker/pcmk_sched_instances.c +++ b/lib/pacemaker/pcmk_sched_instances.c @@ -1,1713 +1,1717 @@ /* - * Copyright 2004-2024 the Pacemaker project contributors + * Copyright 2004-2025 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 pcmk_resource_t *instance, const pcmk_node_t *node, int max_per_node) { pcmk_node_t *allowed_node = NULL; if (pcmk_is_set(instance->flags, pcmk__rsc_removed)) { pcmk__rsc_trace(instance, "%s cannot run on %s: orphaned", instance->id, pcmk__node_name(node)); return false; } - if (!pcmk__node_available(node, false, false)) { + if (!pcmk__node_available(node, pcmk__node_alive|pcmk__node_usable)) { pcmk__rsc_trace(instance, "%s cannot run on %s: node cannot run resources", instance->id, pcmk__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, pcmk__node_name(node)); return false; } if (allowed_node->assign->score < 0) { pcmk__rsc_trace(instance, "%s cannot run on %s: parent score is %s there", instance->id, pcmk__node_name(node), pcmk_readable_score(allowed_node->assign->score)); return false; } if (allowed_node->assign->count >= max_per_node) { pcmk__rsc_trace(instance, "%s cannot run on %s: node already has %d instance%s", instance->id, pcmk__node_name(node), max_per_node, pcmk__plural_s(max_per_node)); return false; } pcmk__rsc_trace(instance, "%s can run on %s (%d already running)", instance->id, pcmk__node_name(node), allowed_node->assign->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(pcmk_resource_t *instance, int max_per_node) { if (instance->priv->allowed_nodes != NULL) { GHashTableIter iter; pcmk_node_t *node = NULL; g_hash_table_iter_init(&iter, instance->priv->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (void **) &node)) { if (!can_run_instance(instance, node, max_per_node)) { pcmk__rsc_trace(instance, "Banning %s from unavailable node %s", instance->id, pcmk__node_name(node)); node->assign->score = -PCMK_SCORE_INFINITY; for (GList *child_iter = instance->priv->children; child_iter != NULL; child_iter = child_iter->next) { pcmk_resource_t *child = child_iter->data; pcmk_node_t *child_node = NULL; child_node = g_hash_table_lookup(child->priv->allowed_nodes, node->priv->id); if (child_node != NULL) { pcmk__rsc_trace(instance, "Banning %s child %s " "from unavailable node %s", instance->id, child->id, pcmk__node_name(node)); child_node->assign->score = -PCMK_SCORE_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(pcmk_node_t *node) { GHashTable *table = pcmk__strkey_table(NULL, pcmk__free_node_copy); node = pe__copy_node(node); g_hash_table_insert(table, (gpointer) node->priv->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 pcmk_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; pcmk_resource_t *other = colocation->primary; float factor = colocation->score / (float) PCMK_SCORE_INFINITY; other->priv->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; pcmk_resource_t *other = colocation->dependent; float factor = colocation->score / (float) PCMK_SCORE_INFINITY; if (!pcmk__colocation_has_influence(colocation, rsc)) { continue; } other->priv->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 pcmk_resource_t *instance1, const pcmk_resource_t *instance2) { int rc = 0; pcmk_node_t *node1 = NULL; pcmk_node_t *node2 = NULL; pcmk_node_t *current_node1 = pcmk__current_node(instance1); pcmk_node_t *current_node2 = pcmk__current_node(instance2); GHashTable *colocated_scores1 = NULL; GHashTable *colocated_scores2 = NULL; pcmk__assert((instance1 != NULL) && (instance1->priv->parent != NULL) && (instance2 != NULL) && (instance2->priv->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->priv->id); node2 = g_hash_table_lookup(colocated_scores2, current_node2->priv->id); // Compare nodes by updated scores if (node1->assign->score < node2->assign->score) { crm_trace("Assign %s (%d on %s) after %s (%d on %s)", instance1->id, node1->assign->score, pcmk__node_name(node1), instance2->id, node2->assign->score, pcmk__node_name(node2)); rc = 1; } else if (node1->assign->score > node2->assign->score) { crm_trace("Assign %s (%d on %s) before %s (%d on %s)", instance1->id, node1->assign->score, pcmk__node_name(node1), instance2->id, node2->assign->score, pcmk__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 pcmk_resource_t *rsc) { if (pcmk_is_set(rsc->flags, pcmk__rsc_failed)) { return true; } for (GList *iter = rsc->priv->children; iter != NULL; iter = iter->next) { if (did_fail((const pcmk_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 pcmk_resource_t *rsc, pcmk_node_t **node) { if (*node != NULL) { pcmk_node_t *allowed = g_hash_table_lookup(rsc->priv->allowed_nodes, (*node)->priv->id); if ((allowed == NULL) || (allowed->assign->score < 0)) { pcmk__rsc_trace(rsc, "%s: current location (%s) is unavailable", rsc->id, pcmk__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 pcmk_resource_t *instance1 = (const pcmk_resource_t *) a; const pcmk_resource_t *instance2 = (const pcmk_resource_t *) b; char *div1 = NULL; char *div2 = NULL; pcmk__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, '-'); } pcmk__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; pcmk_node_t *node1 = NULL; pcmk_node_t *node2 = NULL; unsigned int nnodes1 = 0; unsigned int nnodes2 = 0; bool can1 = true; bool can2 = true; const pcmk_resource_t *instance1 = (const pcmk_resource_t *) a; const pcmk_resource_t *instance2 = (const pcmk_resource_t *) b; pcmk__assert((instance1 != NULL) && (instance2 != NULL)); node1 = instance1->priv->fns->active_node(instance1, &nnodes1, NULL); node2 = instance2->priv->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->priv->priority > instance2->priv->priority) { crm_trace("Assign %s before %s: priority (%d > %d)", instance1->id, instance2->id, instance1->priv->priority, instance2->priv->priority); return -1; } else if (instance1->priv->priority < instance2->priv->priority) { crm_trace("Assign %s after %s: priority (%d < %d)", instance1->id, instance2->id, instance1->priv->priority, instance2->priv->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); + can1 = pcmk__node_available(node1, pcmk__node_alive|pcmk__node_usable); + can2 = pcmk__node_available(node2, pcmk__node_alive|pcmk__node_usable); 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->assign->count < node2->assign->count) { crm_trace("Assign %s before %s: fewer active instances on current node", instance1->id, instance2->id); return -1; } else if (node1->assign->count > node2->assign->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 pcmk_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(pcmk_resource_t *instance, const pcmk_node_t *assigned_to) { pcmk_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->assign->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 pcmk_node_t * assign_instance(pcmk_resource_t *instance, const pcmk_node_t *prefer, int max_per_node) { pcmk_node_t *chosen = NULL; pcmk__rsc_trace(instance, "Assigning %s (preferring %s)", instance->id, ((prefer == NULL)? "no node" : prefer->priv->name)); if (pcmk_is_set(instance->flags, pcmk__rsc_assigning)) { pcmk__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->priv->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 pcmk_resource_t *rsc, pcmk_resource_t *instance, const pcmk_node_t *current, int max_per_node, int available) { const pcmk_node_t *chosen = NULL; int reserved = 0; pcmk_resource_t *parent = instance->priv->parent; GHashTable *allowed_orig = NULL; GHashTable *allowed_orig_parent = parent->priv->allowed_nodes; const pcmk_node_t *allowed_node = NULL; pcmk__rsc_trace(instance, "Trying to assign %s to its current node %s", instance->id, pcmk__node_name(current)); allowed_node = g_hash_table_lookup(instance->priv->allowed_nodes, current->priv->id); - if (!pcmk__node_available(allowed_node, true, false)) { + if (!pcmk__node_available(allowed_node, pcmk__node_alive + |pcmk__node_usable + |pcmk__node_no_negative)) { pcmk__rsc_info(instance, "Not assigning %s to current node %s: unavailable", instance->id, pcmk__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->private->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->private->allowed_nodes to * undo the changes to counts during tentative assignments. If we * successfully assigned an 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->priv->allowed_nodes = pcmk__copy_node_table(allowed_orig_parent); while (reserved < available) { chosen = assign_instance(instance, current, max_per_node); if (pcmk__same_node(chosen, current)) { // Successfully assigned to current node break; } // Assignment updates scores, so restore to original state pcmk__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 pcmk__rsc_info(instance, "Not assigning %s to current node %s: unavailable", instance->id, pcmk__node_name(current)); pcmk__set_rsc_flags(instance, pcmk__rsc_unassigned); break; } // We prefer more strongly to assign an instance to the chosen node pcmk__rsc_debug(instance, "Not assigning %s to current node %s: %s is better", instance->id, pcmk__node_name(current), pcmk__node_name(chosen)); // Reserve one instance for the chosen node and try again if (++reserved >= available) { pcmk__rsc_info(instance, "Not assigning %s to current node %s: " "other assignments are more important", instance->id, pcmk__node_name(current)); } else { pcmk__rsc_debug(instance, "Reserved an instance of %s for %s. Retrying " "assignment of %s to %s", rsc->id, pcmk__node_name(chosen), instance->id, pcmk__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->priv->allowed_nodes); parent->priv->allowed_nodes = allowed_orig_parent; if (chosen == NULL) { // Couldn't assign instance to current node return false; } pcmk__rsc_trace(instance, "Assigned %s to current node %s", instance->id, pcmk__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(pcmk_resource_t *rsc) { unsigned int available_nodes = 0; pcmk_node_t *node = NULL; GHashTableIter iter; g_hash_table_iter_init(&iter, rsc->priv->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (gpointer *) &node)) { node->assign->count = 0; - if (pcmk__node_available(node, false, false)) { + if (pcmk__node_available(node, pcmk__node_alive|pcmk__node_usable)) { 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 pcmk_node_t * preferred_node(const pcmk_resource_t *instance, int optimal_per_node) { const pcmk_node_t *node = NULL; const pcmk_node_t *parent_node = NULL; // Check whether instance is active, healthy, and not yet assigned if ((instance->priv->active_nodes == 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 = pcmk__current_node(instance); - if (!pcmk__node_available(node, true, false)) { + if (!pcmk__node_available(node, pcmk__node_alive + |pcmk__node_usable + |pcmk__node_no_negative)) { pcmk__rsc_trace(instance, "Not assigning %s to %s early (unavailable)", instance->id, pcmk__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->assign->count >= optimal_per_node)) { pcmk__rsc_trace(instance, "Not assigning %s to %s early " "(optimal instances already assigned)", instance->id, pcmk__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(pcmk_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; pcmk_resource_t *instance = NULL; const pcmk_node_t *current = NULL; if (available_nodes > 0) { optimal_per_node = max_total / available_nodes; } if (optimal_per_node < 1) { optimal_per_node = 1; } pcmk__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++; } } pcmk__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 = (pcmk_resource_t *) iter->data; if (!pcmk_is_set(instance->flags, pcmk__rsc_unassigned)) { continue; // Already assigned } if (instance->priv->active_nodes != NULL) { current = pcmk__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, pcmk__node_name(current)); } } if (assigned >= max_total) { pcmk__rsc_debug(collective, "Not assigning %s because maximum %d instances " "already assigned", instance->id, max_total); resource_location(instance, NULL, -PCMK_SCORE_INFINITY, "collective_limit_reached", collective->priv->scheduler); } else if (assign_instance(instance, NULL, max_per_node) != NULL) { assigned++; } } pcmk__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 pcmk_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->priv->variant > pcmk__rsc_variant_primitive) { for (iter = instance->priv->children; (iter != NULL) && !pcmk_all_flags_set(*state, instance_all); iter = iter->next) { check_instance_state((const pcmk_resource_t *) iter->data, state); } return; } // If we get here, instance is a primitive if (instance->priv->active_nodes != NULL) { instance_state |= instance_active; } // Check each of the instance's actions for runnable start or stop for (iter = instance->priv->actions; (iter != NULL) && !pcmk_all_flags_set(instance_state, instance_starting |instance_stopping); iter = iter->next) { const pcmk_action_t *action = (const pcmk_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)) { pcmk__rsc_trace(instance, "Instance is starting due to %s", action->uuid); instance_state |= instance_starting; } else { pcmk__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)) { pcmk__rsc_trace(instance, "Instance is stopping due to %s", action->uuid); instance_state |= instance_stopping; } else { pcmk__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(pcmk_resource_t *collective, GList *instances) { uint32_t state = 0; pcmk_action_t *stop = NULL; pcmk_action_t *stopped = NULL; pcmk_action_t *start = NULL; pcmk_action_t *started = NULL; pcmk__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) { pcmk_resource_t *instance = (pcmk_resource_t *) iter->data; instance->priv->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 = PCMK_SCORE_INFINITY; if (pcmk_any_flags_set(state, instance_active|instance_starting)) { pcmk__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 = PCMK_SCORE_INFINITY; if (!pcmk_is_set(state, instance_restarting)) { pcmk__set_action_flags(stop, pcmk__action_migratable); } if (pcmk__is_clone(collective)) { 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 pcmk_resource_t *rsc) { if (pcmk__is_bundle(rsc)) { return pe__bundle_containers(rsc); } else { return rsc->priv->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 pcmk_resource_t *rsc, GList *list) { if (list != rsc->priv->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 pcmk_resource_t *instance, const pcmk_node_t *node, enum rsc_role_e role, bool current) { pcmk_node_t *instance_node = NULL; CRM_CHECK((instance != NULL) && (node != NULL), return false); if ((role != pcmk_role_unknown) && (role != instance->priv->fns->state(instance, current))) { pcmk__rsc_trace(instance, "%s is not a compatible instance (role is not %s)", instance->id, pcmk_role_text(role)); return false; } if (!is_set_recursive(instance, pcmk__rsc_blocked, true)) { uint32_t target = pcmk__rsc_node_assigned; if (current) { target = pcmk__rsc_node_current; } // We only want instances that haven't failed instance_node = instance->priv->fns->location(instance, NULL, target); } if (instance_node == NULL) { pcmk__rsc_trace(instance, "%s is not a compatible instance " "(not assigned to a node)", instance->id); return false; } if (!pcmk__same_node(instance_node, node)) { pcmk__rsc_trace(instance, "%s is not a compatible instance " "(assigned to %s not %s)", instance->id, pcmk__node_name(instance_node), pcmk__node_name(node)); return false; } return true; } #define display_role(r) \ (((r) == pcmk_role_unknown)? "matching" : pcmk_role_text(r)) /*! * \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 pcmk_resource_t * find_compatible_instance_on_node(const pcmk_resource_t *match_rsc, const pcmk_resource_t *rsc, const pcmk_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) { pcmk_resource_t *instance = (pcmk_resource_t *) iter->data; if (pcmk__instance_matches(instance, node, role, current)) { pcmk__rsc_trace(match_rsc, "Found %s %s instance %s compatible with %s on %s", display_role(role), rsc->id, instance->id, match_rsc->id, pcmk__node_name(node)); free_instance_list(rsc, instances); // Only frees list, not contents return instance; } } free_instance_list(rsc, instances); pcmk__rsc_trace(match_rsc, "No %s %s instance found compatible with %s on %s", display_role(role), rsc->id, match_rsc->id, pcmk__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 */ pcmk_resource_t * pcmk__find_compatible_instance(const pcmk_resource_t *match_rsc, const pcmk_resource_t *rsc, enum rsc_role_e role, bool current) { pcmk_resource_t *instance = NULL; GList *nodes = NULL; const pcmk_node_t *node = NULL; GHashTable *allowed_nodes = match_rsc->priv->allowed_nodes; uint32_t target = pcmk__rsc_node_assigned; if (current) { target = pcmk__rsc_node_current; } // If match_rsc has a node, check only that node node = match_rsc->priv->fns->location(match_rsc, NULL, target); 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(allowed_nodes), NULL); for (GList *iter = nodes; (iter != NULL) && (instance == NULL); iter = iter->next) { instance = find_compatible_instance_on_node(match_rsc, rsc, (pcmk_node_t *) iter->data, role, current); } if (instance == NULL) { pcmk__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 pcmk_action_t *first, const pcmk_action_t *then, pcmk_resource_t *then_instance, uint32_t type, bool current) { // Allow "then" instance to go down even without an interleave match if (current) { pcmk__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, pcmk__ar_unrunnable_first_blocks |pcmk__ar_first_implies_then)) { pcmk__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 pcmk_action_t * find_instance_action(const pcmk_action_t *action, const pcmk_resource_t *instance, const char *action_name, const pcmk_node_t *node, bool for_first) { const pcmk_resource_t *rsc = NULL; pcmk_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->priv->actions, NULL, action_name, node); if (matching_action != NULL) { return matching_action; } if (pcmk_is_set(instance->flags, pcmk__rsc_removed) || pcmk__is_down_action(action_name)) { 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 pcmk_action_t *action) { // Any instance will do const pcmk_resource_t *instance = action->rsc->priv->children->data; char *action_type = NULL; const char *action_name = action->task; enum pcmk__action_type 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 pcmk__action_text(pcmk__action_unspecified)); action_name = strstr(action_type, "_notify_"); CRM_CHECK(action_name != NULL, return pcmk__action_text(pcmk__action_unspecified)); action_name += strlen("_notify_"); } orig_task = get_complex_task(instance, action_name); free(action_type); return pcmk__action_text(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(pcmk_action_t *first, pcmk_action_t *then, const pcmk_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) { pcmk_resource_t *first_instance = NULL; pcmk_resource_t *then_instance = iter->data; pcmk_action_t *first_action = NULL; pcmk_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->priv->cmds->update_ordered_actions( first_action, then_action, node, first_instance->priv->cmds->action_flags(first_action, node), filter, type, then->rsc->priv->scheduler); } 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 pcmk_action_t *first, const pcmk_action_t *then) { bool interleave = false; pcmk_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->priv->variant < pcmk__rsc_variant_clone) || (then->rsc->priv->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->priv->meta, PCMK_META_INTERLEAVE)); pcmk__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(pcmk_resource_t *instance, pcmk_action_t *first, const pcmk_action_t *then, const pcmk_node_t *node, uint32_t flags, uint32_t filter, uint32_t type) { pcmk_action_t *instance_action = NULL; pcmk_scheduler_t *scheduler = instance->priv->scheduler; 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->priv->actions, NULL, then->task, node); if (instance_action == NULL) { return changed; } // Check whether action is runnable instance_flags = instance->priv->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->priv->cmds->update_ordered_actions(first, instance_action, node, flags, filter, type, scheduler); // 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) { pcmk__related_action_t *after = after_iter->data; pcmk__update_action_for_orderings(after->action, scheduler); } } 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] scheduler Scheduler data * * \return Group of enum pcmk__updated flags indicating what was updated */ uint32_t pcmk__instance_update_ordered_actions(pcmk_action_t *first, pcmk_action_t *then, const pcmk_node_t *node, uint32_t flags, uint32_t filter, uint32_t type, pcmk_scheduler_t *scheduler) { pcmk__assert((first != NULL) && (then != NULL) && (scheduler != 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, scheduler); // Update the 'then' clone instances or bundle containers individually for (GList *iter = instances; iter != NULL; iter = iter->next) { pcmk_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(pcmk_action_t *action, const GList *instances, const pcmk_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 pcmk_resource_t *instance = iter->data; const pcmk_node_t *instance_node = NULL; pcmk_action_t *instance_action = NULL; uint32_t instance_flags; // Node is relevant only to primitive instances if (pcmk__is_primitive(instance)) { instance_node = node; } instance_action = find_first_action(instance->priv->actions, NULL, action_name, instance_node); if (instance_action == NULL) { pcmk__rsc_trace(action->rsc, "%s has no %s action on %s", instance->id, action_name, pcmk__node_name(node)); continue; } pcmk__rsc_trace(action->rsc, "%s has %s for %s on %s", instance->id, instance_action->uuid, action_name, pcmk__node_name(node)); instance_flags = instance->priv->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)) { pcmk__rsc_trace(instance, "%s is mandatory because %s is", action->uuid, instance_action->uuid); pe__clear_action_summary_flags(flags, action, pcmk__action_optional); pcmk__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) { pcmk__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) { pcmk__clear_action_flags(action, pcmk__action_runnable); } } return flags; } diff --git a/lib/pacemaker/pcmk_sched_nodes.c b/lib/pacemaker/pcmk_sched_nodes.c index 544a8ced6b..f866cb3254 100644 --- a/lib/pacemaker/pcmk_sched_nodes.c +++ b/lib/pacemaker/pcmk_sched_nodes.c @@ -1,446 +1,451 @@ /* - * Copyright 2004-2024 the Pacemaker project contributors + * Copyright 2004-2025 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 node is available to run resources * - * \param[in] node Node to check - * \param[in] consider_score If true, consider a negative score unavailable - * \param[in] consider_guest If true, consider a guest node unavailable whose - * resource will not be active + * \param[in] node Node to check + * \param[in] flags Group of enum pcmk__node_availability flags * - * \return true if node is online and not shutting down, unclean, or in standby - * or maintenance mode, otherwise false + * \return true if node is available per flags, otherwise false */ bool -pcmk__node_available(const pcmk_node_t *node, bool consider_score, - bool consider_guest) +pcmk__node_available(const pcmk_node_t *node, uint32_t flags) { - if ((node == NULL) || (node->details == NULL) || !node->details->online - || node->details->shutdown || node->details->unclean + // pcmk__node_alive is implicit + if ((node == NULL) || (node->details == NULL) + || !node->details->online || node->details->unclean) { + return false; + } + + if (pcmk_is_set(flags, pcmk__node_usable) + && (node->details->shutdown || pcmk_is_set(node->priv->flags, pcmk__node_standby) - || node->details->maintenance) { + || node->details->maintenance)) { return false; } - if (consider_score && (node->assign->score < 0)) { + if (pcmk_is_set(flags, pcmk__node_no_negative) + && (node->assign->score < 0)) { return false; } - // @TODO Go through all callers to see which should set consider_guest - if (consider_guest && pcmk__is_guest_or_bundle_node(node)) { + if (pcmk_is_set(flags, pcmk__node_no_unrunnable_guest) + && pcmk__is_guest_or_bundle_node(node)) { pcmk_resource_t *guest = node->priv->remote->priv->launcher; if (guest->priv->fns->location(guest, NULL, pcmk__rsc_node_assigned) == NULL) { return false; } } return true; } /*! * \internal * \brief Create a hash table with copies of another table's nodes * * \param[in] nodes Hash table to copy * * \return New table with copies of nodes in \p nodes, or \c NULL if \p nodes is * \c NULL */ GHashTable * pcmk__copy_node_table(GHashTable *nodes) { GHashTable *new_table = NULL; GHashTableIter iter; pcmk_node_t *node = NULL; if (nodes == NULL) { return NULL; } new_table = pcmk__strkey_table(NULL, pcmk__free_node_copy); g_hash_table_iter_init(&iter, nodes); while (g_hash_table_iter_next(&iter, NULL, (gpointer *) &node)) { pcmk_node_t *new_node = pe__copy_node(node); g_hash_table_insert(new_table, (gpointer) new_node->priv->id, new_node); } return new_table; } /*! * \internal * \brief Free a table of node tables * * \param[in,out] data Table to free * * \note This is a \c GDestroyNotify wrapper for \c g_hash_table_destroy(). */ static void destroy_node_tables(gpointer data) { g_hash_table_destroy((GHashTable *) data); } /*! * \internal * \brief Recursively copy the node tables of a resource * * Build a hash table containing copies of the allowed nodes tables of \p rsc * and its entire tree of descendants. The key is the resource ID, and the value * is a copy of the resource's node table. * * \param[in] rsc Resource whose node table to copy * \param[in,out] copy Where to store the copied node tables * * \note \p *copy should be \c NULL for the top-level call. * \note The caller is responsible for freeing \p copy using * \c g_hash_table_destroy(). */ void pcmk__copy_node_tables(const pcmk_resource_t *rsc, GHashTable **copy) { pcmk__assert((rsc != NULL) && (copy != NULL)); if (*copy == NULL) { *copy = pcmk__strkey_table(NULL, destroy_node_tables); } g_hash_table_insert(*copy, rsc->id, pcmk__copy_node_table(rsc->priv->allowed_nodes)); for (const GList *iter = rsc->priv->children; iter != NULL; iter = iter->next) { pcmk__copy_node_tables((const pcmk_resource_t *) iter->data, copy); } } /*! * \internal * \brief Recursively restore the node tables of a resource from backup * * Given a hash table containing backup copies of the allowed nodes tables of * \p rsc and its entire tree of descendants, replace the resources' current * node tables with the backed-up copies. * * \param[in,out] rsc Resource whose node tables to restore * \param[in] backup Table of backup node tables (created by * \c pcmk__copy_node_tables()) * * \note This function frees the resources' current node tables. */ void pcmk__restore_node_tables(pcmk_resource_t *rsc, GHashTable *backup) { pcmk__assert((rsc != NULL) && (backup != NULL)); g_hash_table_destroy(rsc->priv->allowed_nodes); // Copy to avoid danger with multiple restores rsc->priv->allowed_nodes = pcmk__copy_node_table(g_hash_table_lookup(backup, rsc->id)); for (GList *iter = rsc->priv->children; iter != NULL; iter = iter->next) { pcmk__restore_node_tables((pcmk_resource_t *) iter->data, backup); } } /*! * \internal * \brief Copy a list of node objects * * \param[in] list List to copy * \param[in] reset Set copies' scores to 0 * * \return New list of shallow copies of nodes in original list */ GList * pcmk__copy_node_list(const GList *list, bool reset) { GList *result = NULL; for (const GList *iter = list; iter != NULL; iter = iter->next) { pcmk_node_t *new_node = NULL; pcmk_node_t *this_node = iter->data; new_node = pe__copy_node(this_node); if (reset) { new_node->assign->score = 0; } result = g_list_prepend(result, new_node); } return result; } /*! * \internal * \brief Compare two nodes for assignment preference * * Given two nodes, check which one is more preferred by assignment criteria * such as node score and utilization. * * \param[in] a First node to compare * \param[in] b Second node to compare * \param[in] data Node to prefer if all else equal * * \return -1 if \p a is preferred, +1 if \p b is preferred, or 0 if they are * equally preferred */ static gint compare_nodes(gconstpointer a, gconstpointer b, gpointer data) { const pcmk_node_t *node1 = (const pcmk_node_t *) a; const pcmk_node_t *node2 = (const pcmk_node_t *) b; const pcmk_node_t *preferred = (const pcmk_node_t *) data; int node1_score = -PCMK_SCORE_INFINITY; int node2_score = -PCMK_SCORE_INFINITY; int result = 0; if (a == NULL) { return 1; } if (b == NULL) { return -1; } // Compare node scores - if (pcmk__node_available(node1, false, false)) { + if (pcmk__node_available(node1, pcmk__node_alive|pcmk__node_usable)) { node1_score = node1->assign->score; } - if (pcmk__node_available(node2, false, false)) { + if (pcmk__node_available(node2, pcmk__node_alive|pcmk__node_usable)) { node2_score = node2->assign->score; } if (node1_score > node2_score) { crm_trace("%s before %s (score %d > %d)", pcmk__node_name(node1), pcmk__node_name(node2), node1_score, node2_score); return -1; } if (node1_score < node2_score) { crm_trace("%s after %s (score %d < %d)", pcmk__node_name(node1), pcmk__node_name(node2), node1_score, node2_score); return 1; } // If appropriate, compare node utilization if (pcmk__str_eq(node1->priv->scheduler->priv->placement_strategy, PCMK_VALUE_MINIMAL, pcmk__str_casei)) { goto equal; } if (pcmk__str_eq(node1->priv->scheduler->priv->placement_strategy, PCMK_VALUE_BALANCED, pcmk__str_casei)) { result = pcmk__compare_node_capacities(node1, node2); if (result < 0) { crm_trace("%s before %s (greater capacity by %d attributes)", pcmk__node_name(node1), pcmk__node_name(node2), result * -1); return -1; } else if (result > 0) { crm_trace("%s after %s (lower capacity by %d attributes)", pcmk__node_name(node1), pcmk__node_name(node2), result); return 1; } } // Compare number of resources already assigned to node if (node1->priv->num_resources < node2->priv->num_resources) { crm_trace("%s before %s (%d resources < %d)", pcmk__node_name(node1), pcmk__node_name(node2), node1->priv->num_resources, node2->priv->num_resources); return -1; } else if (node1->priv->num_resources > node2->priv->num_resources) { crm_trace("%s after %s (%d resources > %d)", pcmk__node_name(node1), pcmk__node_name(node2), node1->priv->num_resources, node2->priv->num_resources); return 1; } // Check whether one node is already running desired resource if (preferred != NULL) { if (pcmk__same_node(preferred, node1)) { crm_trace("%s before %s (preferred node)", pcmk__node_name(node1), pcmk__node_name(node2)); return -1; } else if (pcmk__same_node(preferred, node2)) { crm_trace("%s after %s (not preferred node)", pcmk__node_name(node1), pcmk__node_name(node2)); return 1; } } // If all else is equal, prefer node with lowest-sorting name equal: result = strcmp(node1->priv->name, node2->priv->name); if (result < 0) { crm_trace("%s before %s (name)", pcmk__node_name(node1), pcmk__node_name(node2)); return -1; } else if (result > 0) { crm_trace("%s after %s (name)", pcmk__node_name(node1), pcmk__node_name(node2)); return 1; } crm_trace("%s == %s", pcmk__node_name(node1), pcmk__node_name(node2)); return 0; } /*! * \internal * \brief Sort a list of nodes by assigment preference * * \param[in,out] nodes Node list to sort * \param[in] active_node Node where resource being assigned is active * * \return New head of sorted list */ GList * pcmk__sort_nodes(GList *nodes, pcmk_node_t *active_node) { return g_list_sort_with_data(nodes, compare_nodes, active_node); } /*! * \internal * \brief Check whether any node is available to run resources * * \param[in] nodes Nodes to check * * \return true if any node in \p nodes is available to run resources, * otherwise false */ bool pcmk__any_node_available(GHashTable *nodes) { GHashTableIter iter; const pcmk_node_t *node = NULL; if (nodes == NULL) { return false; } g_hash_table_iter_init(&iter, nodes); while (g_hash_table_iter_next(&iter, NULL, (void **) &node)) { - if (pcmk__node_available(node, true, false)) { + if (pcmk__node_available(node, pcmk__node_alive + |pcmk__node_usable + |pcmk__node_no_negative)) { return true; } } return false; } /*! * \internal * \brief Apply node health values for all nodes in cluster * * \param[in,out] scheduler Scheduler data */ void pcmk__apply_node_health(pcmk_scheduler_t *scheduler) { int base_health = 0; enum pcmk__health_strategy strategy; const char *strategy_str = pcmk__cluster_option(scheduler->priv->options, PCMK_OPT_NODE_HEALTH_STRATEGY); strategy = pcmk__parse_health_strategy(strategy_str); if (strategy == pcmk__health_strategy_none) { return; } crm_info("Applying node health strategy '%s'", strategy_str); // The progressive strategy can use a base health score if (strategy == pcmk__health_strategy_progressive) { base_health = pcmk__health_score(PCMK_OPT_NODE_HEALTH_BASE, scheduler); } for (GList *iter = scheduler->nodes; iter != NULL; iter = iter->next) { pcmk_node_t *node = (pcmk_node_t *) iter->data; int health = pe__sum_node_health_scores(node, base_health); // An overall health score of 0 has no effect if (health == 0) { continue; } crm_info("Overall system health of %s is %d", pcmk__node_name(node), health); // Use node health as a location score for each resource on the node for (GList *r = scheduler->priv->resources; r != NULL; r = r->next) { pcmk_resource_t *rsc = (pcmk_resource_t *) r->data; bool constrain = true; if (health < 0) { /* Negative health scores do not apply to resources with * PCMK_META_ALLOW_UNHEALTHY_NODES=true. */ constrain = !crm_is_true(g_hash_table_lookup(rsc->priv->meta, PCMK_META_ALLOW_UNHEALTHY_NODES)); } if (constrain) { pcmk__new_location(strategy_str, rsc, health, NULL, node); } else { pcmk__rsc_trace(rsc, "%s is immune from health ban on %s", rsc->id, pcmk__node_name(node)); } } } } /*! * \internal * \brief Check for a node in a resource's parent's allowed nodes * * \param[in] rsc Resource whose parent should be checked * \param[in] node Node to check for * * \return Equivalent of \p node from \p rsc's parent's allowed nodes if any, * otherwise NULL */ pcmk_node_t * pcmk__top_allowed_node(const pcmk_resource_t *rsc, const pcmk_node_t *node) { GHashTable *allowed_nodes = NULL; if ((rsc == NULL) || (node == NULL)) { return NULL; } if (rsc->priv->parent == NULL) { allowed_nodes = rsc->priv->allowed_nodes; } else { allowed_nodes = rsc->priv->parent->priv->allowed_nodes; } return g_hash_table_lookup(allowed_nodes, node->priv->id); } diff --git a/lib/pacemaker/pcmk_sched_primitive.c b/lib/pacemaker/pcmk_sched_primitive.c index 390072a214..8bff84b8a3 100644 --- a/lib/pacemaker/pcmk_sched_primitive.c +++ b/lib/pacemaker/pcmk_sched_primitive.c @@ -1,1720 +1,1725 @@ /* - * Copyright 2004-2024 the Pacemaker project contributors + * Copyright 2004-2025 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(pcmk_resource_t *rsc, pcmk_node_t *node, bool optional); static void start_resource(pcmk_resource_t *rsc, pcmk_node_t *node, bool optional); static void demote_resource(pcmk_resource_t *rsc, pcmk_node_t *node, bool optional); static void promote_resource(pcmk_resource_t *rsc, pcmk_node_t *node, bool optional); static void assert_role_error(pcmk_resource_t *rsc, pcmk_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)(pcmk_resource_t *rsc, pcmk_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 pcmk_resource_t *rsc) { if (rsc->priv->allowed_nodes != NULL) { GList *nodes = g_hash_table_get_values(rsc->priv->allowed_nodes); if (nodes != NULL) { return pcmk__sort_nodes(nodes, pcmk__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(pcmk_resource_t *rsc, const pcmk_node_t *prefer, bool stop_if_fail) { GList *nodes = NULL; pcmk_node_t *chosen = NULL; pcmk_node_t *best = NULL; const pcmk_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->priv->assigned_node != NULL; } // Sort allowed nodes by score nodes = sorted_allowed_nodes(rsc); if (nodes != NULL) { best = (pcmk_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->priv->allowed_nodes, prefer->priv->id); if (chosen == NULL) { pcmk__rsc_trace(rsc, "Preferred node %s for %s was unknown", pcmk__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->assign->score < best->assign->score) { pcmk__rsc_trace(rsc, "Preferred node %s for %s was unsuitable", pcmk__node_name(chosen), rsc->id); chosen = NULL; - } else if (!pcmk__node_available(chosen, true, false)) { + } else if (!pcmk__node_available(chosen, pcmk__node_alive + |pcmk__node_usable + |pcmk__node_no_negative)) { pcmk__rsc_trace(rsc, "Preferred node %s for %s was unavailable", pcmk__node_name(chosen), rsc->id); chosen = NULL; } else { pcmk__rsc_trace(rsc, "Chose preferred node %s for %s " "(ignoring %d candidates)", pcmk__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 (!pcmk__is_unique_clone(rsc->priv->parent) && (chosen->assign->score > 0) // Zero not acceptable - && pcmk__node_available(chosen, false, false)) { + && pcmk__node_available(chosen, + pcmk__node_alive|pcmk__node_usable)) { /* 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. */ pcmk_node_t *running = pcmk__current_node(rsc); if (running == NULL) { // Nothing to do - } else if (!pcmk__node_available(running, true, false)) { + } else if (!pcmk__node_available(running, pcmk__node_alive + |pcmk__node_usable + |pcmk__node_no_negative)) { pcmk__rsc_trace(rsc, "Current node for %s (%s) can't run resources", rsc->id, pcmk__node_name(running)); } else { int nodes_with_best_score = 1; for (GList *iter = nodes->next; iter; iter = iter->next) { pcmk_node_t *allowed = (pcmk_node_t *) iter->data; if (allowed->assign->score != chosen->assign->score) { // The nodes are sorted by score, so no more are equal break; } if (pcmk__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->assign->score >= PCMK_SCORE_INFINITY) { log_level = LOG_WARNING; } do_crm_log(log_level, "Chose %s for %s from %d nodes with score %s", pcmk__node_name(chosen), rsc->id, nodes_with_best_score, pcmk_readable_score(chosen->assign->score)); } } } pcmk__rsc_trace(rsc, "Chose %s for %s from %d candidates", pcmk__node_name(chosen), rsc->id, g_list_length(nodes)); } pcmk__assign_resource(rsc, chosen, false, stop_if_fail); g_list_free(nodes); return rsc->priv->assigned_node != 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, pcmk_resource_t *rsc) { GHashTable *archive = NULL; pcmk_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 > -PCMK_SCORE_INFINITY))) { archive = pcmk__copy_node_table(rsc->priv->allowed_nodes); } if (pcmk_is_set(other->flags, pcmk__rsc_unassigned)) { pcmk__rsc_trace(rsc, "%s: Assigning colocation %s primary %s first" "(score=%d role=%s)", rsc->id, colocation->id, other->id, colocation->score, pcmk_role_text(colocation->dependent_role)); other->priv->cmds->assign(other, NULL, true); } // Apply the colocation score to this resource's allowed node scores rsc->priv->cmds->apply_coloc_score(rsc, other, colocation, true); if ((archive != NULL) && !pcmk__any_node_available(rsc->priv->allowed_nodes)) { pcmk__rsc_info(rsc, "%s: Reverting scores from colocation with %s " "because no nodes allowed", rsc->id, other->id); g_hash_table_destroy(rsc->priv->allowed_nodes); rsc->priv->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 pcmk_resource_t *connection) { pcmk_node_t *remote_node = pcmk_find_node(connection->priv->scheduler, connection->id); CRM_CHECK(remote_node != NULL, return); if ((connection->priv->assigned_node != NULL) && (connection->priv->next_role != pcmk_role_stopped)) { crm_trace("Pacemaker Remote node %s will be online", remote_node->priv->id); remote_node->details->online = TRUE; if (!pcmk_is_set(remote_node->priv->flags, pcmk__node_seen)) { // 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->priv->id, ((connection->priv->assigned_node == NULL)? "un" : ""), pcmk_role_text(connection->priv->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. */ pcmk_node_t * pcmk__primitive_assign(pcmk_resource_t *rsc, const pcmk_node_t *prefer, bool stop_if_fail) { GList *this_with_colocations = NULL; GList *with_this_colocations = NULL; GList *iter = NULL; pcmk_resource_t *parent = NULL; pcmk__colocation_t *colocation = NULL; pcmk_scheduler_t *scheduler = NULL; pcmk__assert(pcmk__is_primitive(rsc)); scheduler = rsc->priv->scheduler; parent = rsc->priv->parent; // Never assign a child without parent being assigned first if ((parent != NULL) && !pcmk_is_set(parent->flags, pcmk__rsc_assigning)) { pcmk__rsc_debug(rsc, "%s: Assigning parent %s first", rsc->id, parent->id); parent->priv->cmds->assign(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->priv->assigned_node != NULL) { node_name = pcmk__node_name(rsc->priv->assigned_node); } pcmk__rsc_debug(rsc, "%s: pre-assigned to %s", rsc->id, node_name); return rsc->priv->assigned_node; } // Ensure we detect assignment loops if (pcmk_is_set(rsc->flags, pcmk__rsc_assigning)) { pcmk__rsc_debug(rsc, "Breaking assignment loop involving %s", rsc->id); return NULL; } pcmk__set_rsc_flags(rsc, pcmk__rsc_assigning); pe__show_node_scores(true, rsc, "Pre-assignment", rsc->priv->allowed_nodes, scheduler); 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 <= -PCMK_SCORE_INFINITY) || (colocation->score >= PCMK_SCORE_INFINITY)) { apply_this_with(colocation, rsc); } } for (iter = with_this_colocations; iter != NULL; iter = iter->next) { colocation = iter->data; if ((colocation->score <= -PCMK_SCORE_INFINITY) || (colocation->score >= PCMK_SCORE_INFINITY)) { pcmk__add_dependent_scores(colocation, rsc); } } pe__show_node_scores(true, rsc, "Mandatory-colocations", rsc->priv->allowed_nodes, scheduler); // Then apply optional colocations for (iter = this_with_colocations; iter != NULL; iter = iter->next) { colocation = iter->data; if ((colocation->score > -PCMK_SCORE_INFINITY) && (colocation->score < PCMK_SCORE_INFINITY)) { apply_this_with(colocation, rsc); } } for (iter = with_this_colocations; iter != NULL; iter = iter->next) { colocation = iter->data; if ((colocation->score > -PCMK_SCORE_INFINITY) && (colocation->score < PCMK_SCORE_INFINITY)) { pcmk__add_dependent_scores(colocation, rsc); } } g_list_free(this_with_colocations); g_list_free(with_this_colocations); if (rsc->priv->next_role == pcmk_role_stopped) { pcmk__rsc_trace(rsc, "Banning %s from all nodes because it will be stopped", rsc->id); resource_location(rsc, NULL, -PCMK_SCORE_INFINITY, PCMK_META_TARGET_ROLE, scheduler); } else if ((rsc->priv->next_role > rsc->priv->orig_role) && !pcmk_is_set(scheduler->flags, pcmk__sched_quorate) && (scheduler->no_quorum_policy == pcmk_no_quorum_freeze)) { crm_notice("Resource %s cannot be elevated from %s to %s due to " PCMK_OPT_NO_QUORUM_POLICY "=" PCMK_VALUE_FREEZE, rsc->id, pcmk_role_text(rsc->priv->orig_role), pcmk_role_text(rsc->priv->next_role)); pe__set_next_role(rsc, rsc->priv->orig_role, PCMK_OPT_NO_QUORUM_POLICY "=" PCMK_VALUE_FREEZE); } pe__show_node_scores(!pcmk_is_set(scheduler->flags, pcmk__sched_output_scores), rsc, __func__, rsc->priv->allowed_nodes, scheduler); // Unmanage resource if fencing is enabled but no device is configured if (pcmk_is_set(scheduler->flags, pcmk__sched_fencing_enabled) && !pcmk_is_set(scheduler->flags, pcmk__sched_have_fencing)) { pcmk__clear_rsc_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; pcmk_node_t *assign_to = NULL; pe__set_next_role(rsc, rsc->priv->orig_role, "unmanaged"); assign_to = pcmk__current_node(rsc); if (assign_to == NULL) { reason = "inactive"; } else if (rsc->priv->orig_role == pcmk_role_promoted) { reason = "promoted"; } else if (pcmk_is_set(rsc->flags, pcmk__rsc_failed)) { reason = "failed"; } else { reason = "active"; } pcmk__rsc_info(rsc, "Unmanaged resource %s assigned to %s: %s", rsc->id, (assign_to? assign_to->priv->name : "no node"), reason); pcmk__assign_resource(rsc, assign_to, true, stop_if_fail); } else if (pcmk_is_set(scheduler->flags, pcmk__sched_stop_all)) { // Must stop at some point, but be consistent with stop_if_fail if (stop_if_fail) { pcmk__rsc_debug(rsc, "Forcing %s to stop: " PCMK_OPT_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)) { pcmk__rsc_info(rsc, "Resource %s cannot run anywhere", rsc->id); } else if ((rsc->priv->active_nodes != NULL) && stop_if_fail) { pcmk__rsc_info(rsc, "Stopping removed resource %s", rsc->id); } } pcmk__clear_rsc_flags(rsc, pcmk__rsc_assigning); if (pcmk_is_set(rsc->flags, pcmk__rsc_is_remote_connection)) { remote_connection_assigned(rsc); } return rsc->priv->assigned_node; } /*! * \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(pcmk_resource_t *rsc, pcmk_node_t *current, bool need_stop, bool need_promote) { enum rsc_role_e role = rsc->priv->orig_role; enum rsc_role_e next_role; rsc_transition_fn fn = NULL; pcmk__set_rsc_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]; pcmk__rsc_trace(rsc, "Creating %s action to take %s down from %s to %s", (need_stop? "required" : "optional"), rsc->id, pcmk_role_text(role), pcmk_role_text(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->priv->orig_role <= rsc->priv->next_role) && (role != rsc->priv->orig_role) && !pcmk_is_set(rsc->flags, pcmk__rsc_blocked)) { bool required = need_stop; next_role = rsc_state_matrix[role][rsc->priv->orig_role]; if ((next_role == pcmk_role_promoted) && need_promote) { required = true; } pcmk__rsc_trace(rsc, "Creating %s action to take %s up from %s to %s", (required? "required" : "optional"), rsc->id, pcmk_role_text(role), pcmk_role_text(next_role)); fn = rsc_action_matrix[role][next_role]; if (fn == NULL) { break; } fn(rsc, rsc->priv->assigned_node, !required); role = next_role; } pcmk__clear_rsc_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(pcmk_resource_t *rsc) { if (rsc->priv->next_role != pcmk_role_unknown) { return "explicit"; } if (rsc->priv->assigned_node == 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(pcmk_resource_t *rsc) { pcmk_action_t *start = NULL; pcmk__rsc_trace(rsc, "Creating action for %s to represent already pending start", rsc->id); start = start_action(rsc, rsc->priv->assigned_node, TRUE); pcmk__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(pcmk_resource_t *rsc) { enum rsc_role_e role = rsc->priv->orig_role; while (role != rsc->priv->next_role) { enum rsc_role_e next_role = rsc_state_matrix[role][rsc->priv->next_role]; rsc_transition_fn fn = NULL; pcmk__rsc_trace(rsc, "Creating action to take %s from %s to %s " "(ending at %s)", rsc->id, pcmk_role_text(role), pcmk_role_text(next_role), pcmk_role_text(rsc->priv->next_role)); fn = rsc_action_matrix[role][next_role]; if (fn == NULL) { break; } fn(rsc, rsc->priv->assigned_node, 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(pcmk_resource_t *rsc) { bool need_stop = false; bool need_promote = false; bool is_moving = false; bool allow_migrate = false; bool multiply_active = false; pcmk_node_t *current = NULL; pcmk_node_t *migration_target = NULL; unsigned int num_all_active = 0; unsigned int num_clean_active = 0; const char *next_role_source = NULL; pcmk__assert(pcmk__is_primitive(rsc)); next_role_source = set_default_next_role(rsc); pcmk__rsc_trace(rsc, "Creating all actions for %s transition from %s to %s " "(%s) on %s", rsc->id, pcmk_role_text(rsc->priv->orig_role), pcmk_role_text(rsc->priv->next_role), next_role_source, pcmk__node_name(rsc->priv->assigned_node)); current = rsc->priv->fns->active_node(rsc, &num_all_active, &num_clean_active); g_list_foreach(rsc->priv->dangling_migration_sources, pcmk__abort_dangling_migration, rsc); if ((current != NULL) && (rsc->priv->assigned_node != NULL) && !pcmk__same_node(current, rsc->priv->assigned_node) && (rsc->priv->next_role >= pcmk_role_started)) { pcmk__rsc_trace(rsc, "Moving %s from %s to %s", rsc->id, pcmk__node_name(current), pcmk__node_name(rsc->priv->assigned_node)); 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 migration_target = rsc->priv->partial_migration_target; if ((rsc->priv->partial_migration_source != NULL) && (migration_target != NULL) && allow_migrate && (num_all_active == 2) && pcmk__same_node(current, rsc->priv->partial_migration_source) && pcmk__same_node(rsc->priv->assigned_node, migration_target)) { /* A partial migration is in progress, and the migration target remains * the same as when the migration began. */ pcmk__rsc_trace(rsc, "Partial migration of %s from %s to %s will continue", rsc->id, pcmk__node_name(rsc->priv->partial_migration_source), pcmk__node_name(migration_target)); } else if ((rsc->priv->partial_migration_source != NULL) || (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, pcmk__node_name(rsc->priv->partial_migration_source), pcmk__node_name(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, pcmk__node_name(rsc->priv->partial_migration_source), pcmk__node_name(migration_target)); } need_stop = true; rsc->priv->partial_migration_source = NULL; rsc->priv->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 PCMK_META_REQUIRES set to PCMK_VALUE_NOTHING or * PCMK_VALUE_QUORUM, don't consider it active on unclean nodes (similar * to how all resources behave when PCMK_OPT_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->priv->xml, PCMK_XA_CLASS); // Resource was (possibly) incorrectly multiply active pcmk__sched_err(rsc->priv->scheduler, "%s resource %s might be active on %u nodes (%s)", pcmk__s(class, "Untyped"), rsc->id, num_all_active, pcmk__multiply_active_text(rsc)); crm_notice("For more information, see \"What are multiply active " "resources?\" at " "https://projects.clusterlabs.org/w/clusterlabs/faq/"); switch (rsc->priv->multiply_active_policy) { case pcmk__multiply_active_restart: need_stop = true; break; case pcmk__multiply_active_unexpected: need_stop = true; // stop_resource() will skip expected node pcmk__set_rsc_flags(rsc, pcmk__rsc_stop_unexpected); break; default: break; } } else { pcmk__clear_rsc_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; pcmk__rsc_trace(rsc, "Recovering %s", rsc->id); } else { pcmk__rsc_trace(rsc, "Recovering %s by demotion", rsc->id); if (rsc->priv->next_role == pcmk_role_promoted) { need_promote = true; } } } else if (pcmk_is_set(rsc->flags, pcmk__rsc_blocked)) { pcmk__rsc_trace(rsc, "Blocking further actions on %s", rsc->id); need_stop = true; } else if ((rsc->priv->orig_role > pcmk_role_started) && (current != NULL) && (rsc->priv->assigned_node != NULL)) { pcmk_action_t *start = NULL; pcmk__rsc_trace(rsc, "Creating start action for promoted resource %s", rsc->id); start = start_action(rsc, rsc->priv->assigned_node, TRUE); if (!pcmk_is_set(start->flags, pcmk__action_optional)) { // Recovery of a promoted resource pcmk__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 pcmk_resource_t *rsc) { GHashTableIter iter; pcmk_node_t *node = NULL; g_hash_table_iter_init(&iter, rsc->priv->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (void **) &node)) { if (node->priv->remote != NULL) { node->assign->score = -PCMK_SCORE_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 pcmk_resource_t *rsc) { GList *allowed_nodes = NULL; if (rsc->priv->allowed_nodes != NULL) { allowed_nodes = g_hash_table_get_values(rsc->priv->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(pcmk_resource_t *rsc) { GList *allowed_nodes = NULL; bool check_unfencing = false; bool check_utilization = false; pcmk_scheduler_t *scheduler = NULL; pcmk__assert(pcmk__is_primitive(rsc)); scheduler = rsc->priv->scheduler; if (!pcmk_is_set(rsc->flags, pcmk__rsc_managed)) { pcmk__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(scheduler->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->priv->utilization) > 0) && !pcmk__str_eq(scheduler->priv->placement_strategy, PCMK_VALUE_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 |pcmk__ar_intermediate_stop, scheduler); // Promotable ordering: demote before stop, start before promote if (pcmk_is_set(pe__const_top_resource(rsc, false)->flags, pcmk__rsc_promotable) || (rsc->priv->orig_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, scheduler); 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, pcmk__ar_unrunnable_first_blocks, scheduler); } // 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, pcmk__ar_if_on_same_node|pcmk__ar_then_cancels_first, scheduler); // Certain checks need allowed nodes if (check_unfencing || check_utilization || (rsc->priv->launcher != 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->priv->launcher != NULL) { pcmk_resource_t *remote_rsc = NULL; if (pcmk_is_set(rsc->flags, pcmk__rsc_is_remote_connection)) { // 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->priv->launcher); } /* If someone cleans up a guest or bundle node's launcher, we will * likely schedule a (re-)probe of the launcher and recovery of the * connection. Order the connection stop after the launcher probe, * so that if we detect the launcher running, we will trigger a new * transition and avoid the unnecessary recovery. */ pcmk__order_resource_actions(rsc->priv->launcher, 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 PCMK__META_CONTAINER * 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 PCMK__META_CONTAINER set to a remote node or guest node resource. */ } else if (pcmk_is_set(rsc->priv->launcher->flags, pcmk__rsc_is_remote_connection)) { remote_rsc = rsc->priv->launcher; } else { remote_rsc = pe__resource_contains_guest_node(scheduler, rsc->priv->launcher); } if (remote_rsc != NULL) { /* Force the resource on the Pacemaker Remote node instead of * colocating the resource with the launcher. */ for (GList *item = allowed_nodes; item; item = item->next) { pcmk_node_t *node = item->data; if (node->priv->remote != remote_rsc) { node->assign->score = -PCMK_SCORE_INFINITY; } } } else { /* This resource is either launched by a resource 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 launcher %s", rsc->id, rsc->priv->launcher->id); pcmk__new_ordering(rsc->priv->launcher, pcmk__op_key(rsc->priv->launcher->id, PCMK_ACTION_START, 0), NULL, rsc, pcmk__op_key(rsc->id, PCMK_ACTION_START, 0), NULL, pcmk__ar_first_implies_then |pcmk__ar_unrunnable_first_blocks, scheduler); pcmk__new_ordering(rsc, pcmk__op_key(rsc->id, PCMK_ACTION_STOP, 0), NULL, rsc->priv->launcher, pcmk__op_key(rsc->priv->launcher->id, PCMK_ACTION_STOP, 0), NULL, pcmk__ar_then_implies_first, scheduler); if (pcmk_is_set(rsc->flags, pcmk__rsc_remote_nesting_allowed) /* @TODO: && non-bundle Pacemaker Remote nodes exist */) { score = 10000; /* Highly preferred but not essential */ } else { score = PCMK_SCORE_INFINITY; // Force to run on same host } pcmk__new_colocation("#resource-with-container", NULL, score, rsc, rsc->priv->launcher, NULL, NULL, pcmk__coloc_influence); } } if (pcmk_is_set(rsc->flags, pcmk__rsc_is_remote_connection) || 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 * * \return The score added to the dependent's priority */ int pcmk__primitive_apply_coloc_score(pcmk_resource_t *dependent, const pcmk_resource_t *primary, const pcmk__colocation_t *colocation, bool for_dependent) { enum pcmk__coloc_affects filter_results; pcmk__assert((dependent != NULL) && (primary != NULL) && (colocation != NULL)); if (for_dependent) { // Always process on behalf of primary resource return primary->priv->cmds->apply_coloc_score(dependent, primary, colocation, false); } filter_results = pcmk__colocation_affects(dependent, primary, colocation, false); pcmk__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: return pcmk__apply_coloc_to_priority(dependent, primary, colocation); case pcmk__coloc_affects_location: pcmk__apply_coloc_to_scores(dependent, primary, colocation); return 0; default: // pcmk__coloc_affects_nothing return 0; } } /* Primitive implementation of * pcmk__assignment_methods_t:with_this_colocations() */ void pcmk__with_primitive_colocations(const pcmk_resource_t *rsc, const pcmk_resource_t *orig_rsc, GList **list) { const pcmk_resource_t *parent = NULL; pcmk__assert(pcmk__is_primitive(rsc) && (list != NULL)); parent = rsc->priv->parent; 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->priv->with_this_colocations, orig_rsc); if (parent != NULL) { parent->priv->cmds->with_this_colocations(parent, orig_rsc, list); } } else { // For an ancestor, add only explicitly configured constraints for (GList *iter = rsc->priv->with_this_colocations; 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 * pcmk__assignment_methods_t:this_with_colocations() */ void pcmk__primitive_with_colocations(const pcmk_resource_t *rsc, const pcmk_resource_t *orig_rsc, GList **list) { const pcmk_resource_t *parent = NULL; pcmk__assert(pcmk__is_primitive(rsc) && (list != NULL)); parent = rsc->priv->parent; 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->priv->this_with_colocations, orig_rsc); if (parent != NULL) { parent->priv->cmds->this_with_colocations(parent, orig_rsc, list); } } else { // For an ancestor, add only explicitly configured constraints for (GList *iter = rsc->priv->this_with_colocations; 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(pcmk_action_t *action, const pcmk_node_t *node) { pcmk__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 \c true if \p rsc is multiply active with * \c PCMK_META_MULTIPLE_ACTIVE set to \c PCMK_VALUE_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 pcmk_resource_t *rsc, const pcmk_node_t *node) { return pcmk_all_flags_set(rsc->flags, pcmk__rsc_stop_unexpected|pcmk__rsc_restarting) && (rsc->priv->next_role > pcmk_role_stopped) && pcmk__same_node(rsc->priv->assigned_node, 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(pcmk_resource_t *rsc, pcmk_node_t *node, bool optional) { for (GList *iter = rsc->priv->active_nodes; iter != NULL; iter = iter->next) { pcmk_node_t *current = (pcmk_node_t *) iter->data; pcmk_action_t *stop = NULL; if (is_expected_node(rsc, current)) { /* We are scheduling restart actions for a multiply active resource * with PCMK_META_MULTIPLE_ACTIVE=PCMK_VALUE_STOP_UNEXPECTED, and * this is where it should not be stopped. */ pcmk__rsc_trace(rsc, "Skipping stop of multiply active resource %s " "on expected node %s", rsc->id, pcmk__node_name(current)); continue; } if (rsc->priv->partial_migration_target != NULL) { // Continue migration if node originally was and remains target if (pcmk__same_node(current, rsc->priv->partial_migration_target) && pcmk__same_node(current, rsc->priv->assigned_node)) { pcmk__rsc_trace(rsc, "Skipping stop of %s on %s " "because partial migration there will continue", rsc->id, pcmk__node_name(current)); continue; } else { pcmk__rsc_trace(rsc, "Forcing stop of %s on %s " "because migration target changed", rsc->id, pcmk__node_name(current)); optional = false; } } pcmk__rsc_trace(rsc, "Scheduling stop of %s on %s", rsc->id, pcmk__node_name(current)); stop = stop_action(rsc, current, optional); if (rsc->priv->assigned_node == 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)) { pcmk__clear_action_flags(stop, pcmk__action_runnable); } if (pcmk_is_set(rsc->flags, pcmk__rsc_needs_unfencing)) { pcmk_action_t *unfence = pe_fence_op(current, PCMK_ACTION_ON, true, NULL, false, rsc->priv->scheduler); order_actions(stop, unfence, pcmk__ar_then_implies_first); if (!pcmk__node_unfenced(current)) { pcmk__sched_err(rsc->priv->scheduler, "Stopping %s until %s can be unfenced", rsc->id, pcmk__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(pcmk_resource_t *rsc, pcmk_node_t *node, bool optional) { pcmk_action_t *start = NULL; pcmk__assert(node != NULL); pcmk__rsc_trace(rsc, "Scheduling %s start of %s on %s (score %d)", (optional? "optional" : "required"), rsc->id, pcmk__node_name(node), node->assign->score); 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) { pcmk__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. */ pcmk__rsc_trace(rsc, "Start of multiply active resouce %s " "on expected node %s will be a pseudo-action", rsc->id, pcmk__node_name(node)); pcmk__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(pcmk_resource_t *rsc, pcmk_node_t *node, bool optional) { GList *iter = NULL; GList *action_list = NULL; bool runnable = true; pcmk__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) { pcmk_action_t *start = (pcmk_action_t *) iter->data; if (!pcmk_is_set(start->flags, pcmk__action_runnable)) { runnable = false; } } g_list_free(action_list); if (runnable) { pcmk_action_t *promote = promote_action(rsc, node, optional); pcmk__rsc_trace(rsc, "Scheduling %s promotion of %s on %s", (optional? "optional" : "required"), rsc->id, pcmk__node_name(node)); if (is_expected_node(rsc, node)) { /* This could be a problem if the promote becomes necessary for * other reasons later. */ pcmk__rsc_trace(rsc, "Promotion of multiply active resouce %s " "on expected node %s will be a pseudo-action", rsc->id, pcmk__node_name(node)); pcmk__set_action_flags(promote, pcmk__action_pseudo); } } else { pcmk__rsc_trace(rsc, "Not promoting %s on %s: start unrunnable", rsc->id, pcmk__node_name(node)); action_list = pe__resource_actions(rsc, node, PCMK_ACTION_PROMOTE, true); for (iter = action_list; iter != NULL; iter = iter->next) { pcmk_action_t *promote = (pcmk_action_t *) iter->data; pcmk__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(pcmk_resource_t *rsc, pcmk_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->priv->active_nodes; iter != NULL; iter = iter->next) { pcmk_node_t *current = (pcmk_node_t *) iter->data; if (is_expected_node(rsc, current)) { pcmk__rsc_trace(rsc, "Skipping demote of multiply active resource %s " "on expected node %s", rsc->id, pcmk__node_name(current)); } else { pcmk__rsc_trace(rsc, "Scheduling %s demotion of %s on %s", (optional? "optional" : "required"), rsc->id, pcmk__node_name(current)); demote_action(rsc, current, optional); } } } static void assert_role_error(pcmk_resource_t *rsc, pcmk_node_t *node, bool optional) { pcmk__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(pcmk_resource_t *rsc, const pcmk_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)) { pcmk__rsc_trace(rsc, "Skipping clean-up of %s on %s: resource failed", rsc->id, pcmk__node_name(node)); return; } if (node->details->unclean || !node->details->online) { pcmk__rsc_trace(rsc, "Skipping clean-up of %s on %s: node unavailable", rsc->id, pcmk__node_name(node)); return; } crm_notice("Scheduling clean-up of %s on %s", rsc->id, pcmk__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 pcmk_resource_t *rsc, xmlNode *xml) { char *name = NULL; char *value = NULL; const pcmk_resource_t *parent = NULL; pcmk__assert(pcmk__is_primitive(rsc) && (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->priv->meta, PCMK__META_CLONE); if (value != NULL) { name = crm_meta_name(PCMK__META_CLONE); crm_xml_add(xml, name, value); free(name); } // Not sure if this one is really needed ... value = g_hash_table_lookup(rsc->priv->meta, PCMK_META_REMOTE_NODE); if (value != NULL) { name = crm_meta_name(PCMK_META_REMOTE_NODE); crm_xml_add(xml, name, value); free(name); } /* The PCMK__META_CONTAINER meta-attribute can be set on the primitive * itself or one of its ancestors, so check them all and keep the highest. */ for (parent = rsc; parent != NULL; parent = parent->priv->parent) { if (parent->priv->launcher != NULL) { crm_xml_add(xml, CRM_META "_" PCMK__META_CONTAINER, parent->priv->launcher->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->priv->meta, "external-ip"); if (value != NULL) { crm_xml_add(xml, "pcmk_external_ip", value); } } // Primitive implementation of pcmk__assignment_methods_t:add_utilization() void pcmk__primitive_add_utilization(const pcmk_resource_t *rsc, const pcmk_resource_t *orig_rsc, GList *all_rscs, GHashTable *utilization) { pcmk__assert(pcmk__is_primitive(rsc) && (orig_rsc != NULL) && (utilization != NULL)); if (!pcmk_is_set(rsc->flags, pcmk__rsc_unassigned)) { return; } pcmk__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(pcmk_node_t *node) { const char *shutdown = pcmk__node_attr(node, PCMK__NODE_ATTR_SHUTDOWN, NULL, pcmk__rsc_node_current); time_t result = 0; if (shutdown != NULL) { long long result_ll; int rc = pcmk__scan_ll(shutdown, &result_ll, 0LL); if (rc == pcmk_rc_ok) { result = (time_t) result_ll; } else { crm_warn("Ignoring invalid value '%s' for %s " PCMK__NODE_ATTR_SHUTDOWN " attribute: %s", shutdown, pcmk__node_name(node), pcmk_rc_str(rc)); } } if (result == 0) { result = pcmk__scheduler_epoch_time(node->priv->scheduler); } return 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 pcmk_node_t *node = (const pcmk_node_t *) data; pcmk_resource_t *rsc = (pcmk_resource_t *) user_data; if (!pcmk__same_node(node, rsc->priv->lock_node)) { resource_location(rsc, node, -PCMK_SCORE_INFINITY, PCMK_OPT_SHUTDOWN_LOCK, rsc->priv->scheduler); } } // Primitive implementation of pcmk__assignment_methods_t:shutdown_lock() void pcmk__primitive_shutdown_lock(pcmk_resource_t *rsc) { pcmk_scheduler_t *scheduler = NULL; pcmk__assert(pcmk__is_primitive(rsc)); scheduler = rsc->priv->scheduler; // Fence devices and remote connections can't be locked if (pcmk_any_flags_set(rsc->flags, pcmk__rsc_fence_device |pcmk__rsc_is_remote_connection)) { return; } if (rsc->priv->lock_node != NULL) { // The lock was obtained from resource history if (rsc->priv->active_nodes != 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. */ pcmk__rsc_info(rsc, "Cancelling shutdown lock " "because %s is already active", rsc->id); pe__clear_resource_history(rsc, rsc->priv->lock_node); rsc->priv->lock_node = NULL; rsc->priv->lock_time = 0; } // Only a resource active on exactly one node can be locked } else if (pcmk__list_of_1(rsc->priv->active_nodes)) { pcmk_node_t *node = rsc->priv->active_nodes->data; if (node->details->shutdown) { if (node->details->unclean) { pcmk__rsc_debug(rsc, "Not locking %s to unclean %s for shutdown", rsc->id, pcmk__node_name(node)); } else { rsc->priv->lock_node = node; rsc->priv->lock_time = shutdown_time(node); } } } if (rsc->priv->lock_node == NULL) { // No lock needed return; } if (scheduler->priv->shutdown_lock_ms > 0U) { time_t lock_expiration = rsc->priv->lock_time + pcmk__timeout_ms2s(scheduler->priv->shutdown_lock_ms); pcmk__rsc_info(rsc, "Locking %s to %s due to shutdown (expires @%lld)", rsc->id, pcmk__node_name(rsc->priv->lock_node), (long long) lock_expiration); pcmk__update_recheck_time(++lock_expiration, scheduler, "shutdown lock expiration"); } else { pcmk__rsc_info(rsc, "Locking %s to %s due to shutdown", rsc->id, pcmk__node_name(rsc->priv->lock_node)); } // If resource is locked to one node, ban it from all other nodes g_list_foreach(scheduler->nodes, ban_if_not_locked, rsc); } diff --git a/lib/pacemaker/pcmk_sched_promotable.c b/lib/pacemaker/pcmk_sched_promotable.c index 6a1bf9cd9d..e345554b47 100644 --- a/lib/pacemaker/pcmk_sched_promotable.c +++ b/lib/pacemaker/pcmk_sched_promotable.c @@ -1,1363 +1,1366 @@ /* - * Copyright 2004-2024 the Pacemaker project contributors + * Copyright 2004-2025 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 Add implicit promotion ordering for a promotable instance * * \param[in,out] clone Clone resource * \param[in,out] child Instance of \p clone being ordered * \param[in,out] last Previous instance ordered (NULL if \p child is first) */ static void order_instance_promotion(pcmk_resource_t *clone, pcmk_resource_t *child, pcmk_resource_t *last) { // "Promote clone" -> promote instance -> "clone promoted" pcmk__order_resource_actions(clone, PCMK_ACTION_PROMOTE, child, PCMK_ACTION_PROMOTE, pcmk__ar_ordered); pcmk__order_resource_actions(child, PCMK_ACTION_PROMOTE, clone, PCMK_ACTION_PROMOTED, pcmk__ar_ordered); // If clone is ordered, order this instance relative to last if ((last != NULL) && pe__clone_is_ordered(clone)) { pcmk__order_resource_actions(last, PCMK_ACTION_PROMOTE, child, PCMK_ACTION_PROMOTE, pcmk__ar_ordered); } } /*! * \internal * \brief Add implicit demotion ordering for a promotable instance * * \param[in,out] clone Clone resource * \param[in,out] child Instance of \p clone being ordered * \param[in] last Previous instance ordered (NULL if \p child is first) */ static void order_instance_demotion(pcmk_resource_t *clone, pcmk_resource_t *child, pcmk_resource_t *last) { // "Demote clone" -> demote instance -> "clone demoted" pcmk__order_resource_actions(clone, PCMK_ACTION_DEMOTE, child, PCMK_ACTION_DEMOTE, pcmk__ar_then_implies_first_graphed); pcmk__order_resource_actions(child, PCMK_ACTION_DEMOTE, clone, PCMK_ACTION_DEMOTED, pcmk__ar_first_implies_then_graphed); // If clone is ordered, order this instance relative to last if ((last != NULL) && pe__clone_is_ordered(clone)) { pcmk__order_resource_actions(child, PCMK_ACTION_DEMOTE, last, PCMK_ACTION_DEMOTE, pcmk__ar_ordered); } } /*! * \internal * \brief Check whether an instance will be promoted or demoted * * \param[in] rsc Instance to check * \param[out] demoting If \p rsc will be demoted, this will be set to true * \param[out] promoting If \p rsc will be promoted, this will be set to true */ static void check_for_role_change(const pcmk_resource_t *rsc, bool *demoting, bool *promoting) { const GList *iter = NULL; // If this is a cloned group, check group members recursively if (rsc->priv->children != NULL) { for (iter = rsc->priv->children; iter != NULL; iter = iter->next) { check_for_role_change((const pcmk_resource_t *) iter->data, demoting, promoting); } return; } for (iter = rsc->priv->actions; iter != NULL; iter = iter->next) { const pcmk_action_t *action = (const pcmk_action_t *) iter->data; if (*promoting && *demoting) { return; } else if (pcmk_is_set(action->flags, pcmk__action_optional)) { continue; } else if (pcmk__str_eq(PCMK_ACTION_DEMOTE, action->task, pcmk__str_none)) { *demoting = true; } else if (pcmk__str_eq(PCMK_ACTION_PROMOTE, action->task, pcmk__str_none)) { *promoting = true; } } } /*! * \internal * \brief Add promoted-role location constraint scores to an instance's priority * * Adjust a promotable clone instance's promotion priority by the scores of any * location constraints in a list that are both limited to the promoted role and * for the node where the instance will be placed. * * \param[in,out] child Promotable clone instance * \param[in] location_constraints List of location constraints to apply * \param[in] chosen Node where \p child will be placed */ static void apply_promoted_locations(pcmk_resource_t *child, const GList *location_constraints, const pcmk_node_t *chosen) { for (const GList *iter = location_constraints; iter; iter = iter->next) { const pcmk__location_t *location = iter->data; const pcmk_node_t *constraint_node = NULL; if (location->role_filter == pcmk_role_promoted) { constraint_node = pe_find_node_id(location->nodes, chosen->priv->id); } if (constraint_node != NULL) { int new_priority = pcmk__add_scores(child->priv->priority, constraint_node->assign->score); pcmk__rsc_trace(child, "Applying location %s to %s promotion priority on " "%s: %s + %s = %s", location->id, child->id, pcmk__node_name(constraint_node), pcmk_readable_score(child->priv->priority), pcmk_readable_score(constraint_node->assign->score), pcmk_readable_score(new_priority)); child->priv->priority = new_priority; } } } /*! * \internal * \brief Get the node that an instance will be promoted on * * \param[in] rsc Promotable clone instance to check * * \return Node that \p rsc will be promoted on, or NULL if none */ static pcmk_node_t * node_to_be_promoted_on(const pcmk_resource_t *rsc) { pcmk_node_t *node = NULL; pcmk_node_t *local_node = NULL; const pcmk_resource_t *parent = NULL; // If this is a cloned group, bail if any group member can't be promoted for (GList *iter = rsc->priv->children; iter != NULL; iter = iter->next) { pcmk_resource_t *child = (pcmk_resource_t *) iter->data; if (node_to_be_promoted_on(child) == NULL) { pcmk__rsc_trace(rsc, "%s can't be promoted because member %s can't", rsc->id, child->id); return NULL; } } node = rsc->priv->fns->location(rsc, NULL, pcmk__rsc_node_assigned); if (node == NULL) { pcmk__rsc_trace(rsc, "%s can't be promoted because it won't be active", rsc->id); return NULL; } else if (!pcmk_is_set(rsc->flags, pcmk__rsc_managed)) { if (rsc->priv->fns->state(rsc, TRUE) == pcmk_role_promoted) { crm_notice("Unmanaged instance %s will be left promoted on %s", rsc->id, pcmk__node_name(node)); } else { pcmk__rsc_trace(rsc, "%s can't be promoted because it is unmanaged", rsc->id); return NULL; } } else if (rsc->priv->priority < 0) { pcmk__rsc_trace(rsc, "%s can't be promoted because its promotion priority " "%d is negative", rsc->id, rsc->priv->priority); return NULL; - } else if (!pcmk__node_available(node, false, true)) { + } else if (!pcmk__node_available(node, pcmk__node_alive + |pcmk__node_usable + |pcmk__node_no_unrunnable_guest)) { pcmk__rsc_trace(rsc, "%s can't be promoted because %s can't run resources", rsc->id, pcmk__node_name(node)); return NULL; } parent = pe__const_top_resource(rsc, false); local_node = g_hash_table_lookup(parent->priv->allowed_nodes, node->priv->id); if (local_node == NULL) { /* It should not be possible for the scheduler to have assigned the * instance to a node where its parent is not allowed, but it's good to * have a fail-safe. */ if (pcmk_is_set(rsc->flags, pcmk__rsc_managed)) { pcmk__sched_err(node->priv->scheduler, "%s can't be promoted because %s is not allowed " "on %s (scheduler bug?)", rsc->id, parent->id, pcmk__node_name(node)); } // else the instance is unmanaged and already promoted return NULL; } else if ((local_node->assign->count >= pe__clone_promoted_node_max(parent)) && pcmk_is_set(rsc->flags, pcmk__rsc_managed)) { pcmk__rsc_trace(rsc, "%s can't be promoted because %s has " "maximum promoted instances already", rsc->id, pcmk__node_name(node)); return NULL; } return local_node; } /*! * \internal * \brief Compare two promotable clone instances by promotion priority * * \param[in] a First instance to compare * \param[in] b Second instance to compare * * \return A negative number if \p a has higher promotion priority, * a positive number if \p b has higher promotion priority, * or 0 if promotion priorities are equal */ static gint cmp_promotable_instance(gconstpointer a, gconstpointer b) { const pcmk_resource_t *rsc1 = (const pcmk_resource_t *) a; const pcmk_resource_t *rsc2 = (const pcmk_resource_t *) b; enum rsc_role_e role1 = pcmk_role_unknown; enum rsc_role_e role2 = pcmk_role_unknown; pcmk__assert((rsc1 != NULL) && (rsc2 != NULL)); // Check promotion priority set by pcmk__set_instance_roles() if (rsc1->priv->promotion_priority > rsc2->priv->promotion_priority) { pcmk__rsc_trace(rsc1, "%s has higher promotion priority (%s) than %s (%d)", rsc1->id, pcmk_readable_score(rsc1->priv->promotion_priority), rsc2->id, rsc2->priv->promotion_priority); return -1; } if (rsc1->priv->promotion_priority < rsc2->priv->promotion_priority) { pcmk__rsc_trace(rsc1, "%s has lower promotion priority (%s) than %s (%d)", rsc1->id, pcmk_readable_score(rsc1->priv->promotion_priority), rsc2->id, rsc2->priv->promotion_priority); return 1; } // If those are the same, prefer instance whose current role is higher role1 = rsc1->priv->fns->state(rsc1, TRUE); role2 = rsc2->priv->fns->state(rsc2, TRUE); if (role1 > role2) { pcmk__rsc_trace(rsc1, "%s has higher promotion priority than %s " "(higher current role)", rsc1->id, rsc2->id); return -1; } else if (role1 < role2) { pcmk__rsc_trace(rsc1, "%s has lower promotion priority than %s " "(lower current role)", rsc1->id, rsc2->id); return 1; } // Finally, do normal clone instance sorting return pcmk__cmp_instance(a, b); } /*! * \internal * \brief Add promotable clone instance's promotion priority to its node's score * * Add a promotable clone instance's promotion priority (which sums its * promotion preferences and scores of relevant location constraints for the * promoted role) to the node score of the instance's assigned node. * * \param[in] data Promotable clone instance * \param[in,out] user_data Clone parent of \p data */ static void add_promotion_priority_to_node_score(gpointer data, gpointer user_data) { const pcmk_resource_t *child = (const pcmk_resource_t *) data; pcmk_resource_t *clone = (pcmk_resource_t *) user_data; pcmk_node_t *node = NULL; const pcmk_node_t *chosen = NULL; const int promotion_priority = child->priv->promotion_priority; if (promotion_priority < 0) { pcmk__rsc_trace(clone, "Not adding promotion priority of %s: negative (%s)", child->id, pcmk_readable_score(promotion_priority)); return; } chosen = child->priv->fns->location(child, NULL, pcmk__rsc_node_assigned); if (chosen == NULL) { pcmk__rsc_trace(clone, "Not adding promotion priority of %s: inactive", child->id); return; } node = g_hash_table_lookup(clone->priv->allowed_nodes, chosen->priv->id); pcmk__assert(node != NULL); node->assign->score = pcmk__add_scores(promotion_priority, node->assign->score); pcmk__rsc_trace(clone, "Added cumulative priority of %s (%s) to score on %s " "(now %d)", child->id, pcmk_readable_score(promotion_priority), pcmk__node_name(node), node->assign->score); } /*! * \internal * \brief Apply colocation to primary's node scores if for promoted role * * \param[in,out] data Colocation constraint to apply * \param[in,out] user_data Promotable clone that is constraint's primary */ static void apply_coloc_to_primary(gpointer data, gpointer user_data) { pcmk__colocation_t *colocation = data; pcmk_resource_t *clone = user_data; pcmk_resource_t *dependent = colocation->dependent; const float factor = colocation->score / (float) PCMK_SCORE_INFINITY; const uint32_t flags = pcmk__coloc_select_active |pcmk__coloc_select_nonnegative; if ((colocation->primary_role != pcmk_role_promoted) || !pcmk__colocation_has_influence(colocation, NULL)) { return; } pcmk__rsc_trace(clone, "Applying colocation %s (%s with promoted %s) @%s", colocation->id, colocation->dependent->id, colocation->primary->id, pcmk_readable_score(colocation->score)); dependent->priv->cmds->add_colocated_node_scores(dependent, clone, clone->id, &(clone->priv->allowed_nodes), colocation, factor, flags); } /*! * \internal * \brief Set clone instance's promotion priority to its node's score * * \param[in,out] data Promotable clone instance * \param[in] user_data Parent clone of \p data */ static void set_promotion_priority_to_node_score(gpointer data, gpointer user_data) { pcmk_resource_t *child = (pcmk_resource_t *) data; const pcmk_resource_t *clone = (const pcmk_resource_t *) user_data; pcmk_node_t *chosen = child->priv->fns->location(child, NULL, pcmk__rsc_node_assigned); if (!pcmk_is_set(child->flags, pcmk__rsc_managed) && (child->priv->next_role == pcmk_role_promoted)) { child->priv->promotion_priority = PCMK_SCORE_INFINITY; pcmk__rsc_trace(clone, "Final promotion priority for %s is %s " "(unmanaged promoted)", child->id, pcmk_readable_score(PCMK_SCORE_INFINITY)); } else if (chosen == NULL) { child->priv->promotion_priority = -PCMK_SCORE_INFINITY; pcmk__rsc_trace(clone, "Final promotion priority for %s is %s " "(will not be active)", child->id, pcmk_readable_score(-PCMK_SCORE_INFINITY)); } else if (child->priv->promotion_priority < 0) { pcmk__rsc_trace(clone, "Final promotion priority for %s is %s " "(ignoring node score)", child->id, pcmk_readable_score(child->priv->promotion_priority)); } else { const pcmk_node_t *node = NULL; node = g_hash_table_lookup(clone->priv->allowed_nodes, chosen->priv->id); pcmk__assert(node != NULL); child->priv->promotion_priority = node->assign->score; pcmk__rsc_trace(clone, "Adding scores for %s: " "final promotion priority for %s is %s", clone->id, child->id, pcmk_readable_score(child->priv->promotion_priority)); } } /*! * \internal * \brief Sort a promotable clone's instances by descending promotion priority * * \param[in,out] clone Promotable clone to sort */ static void sort_promotable_instances(pcmk_resource_t *clone) { GList *colocations = NULL; if (pe__set_clone_flag(clone, pcmk__clone_promotion_constrained) == pcmk_rc_already) { return; } pcmk__set_rsc_flags(clone, pcmk__rsc_updating_nodes); for (GList *iter = clone->priv->children; iter != NULL; iter = iter->next) { pcmk_resource_t *child = (pcmk_resource_t *) iter->data; pcmk__rsc_trace(clone, "Adding scores for %s: " "initial promotion priority for %s is %s", clone->id, child->id, pcmk_readable_score(child->priv->promotion_priority)); } pe__show_node_scores(true, clone, "Before", clone->priv->allowed_nodes, clone->priv->scheduler); g_list_foreach(clone->priv->children, add_promotion_priority_to_node_score, clone); // "this with" colocations were already applied via set_instance_priority() colocations = pcmk__with_this_colocations(clone); g_list_foreach(colocations, apply_coloc_to_primary, clone); g_list_free(colocations); // Ban resource from all nodes if it needs a ticket but doesn't have it pcmk__require_promotion_tickets(clone); pe__show_node_scores(true, clone, "After", clone->priv->allowed_nodes, clone->priv->scheduler); // Reset promotion priorities to final node scores g_list_foreach(clone->priv->children, set_promotion_priority_to_node_score, clone); // Finally, sort instances in descending order of promotion priority clone->priv->children = g_list_sort(clone->priv->children, cmp_promotable_instance); pcmk__clear_rsc_flags(clone, pcmk__rsc_updating_nodes); } /*! * \internal * \brief Find the active instance (if any) of an anonymous clone on a node * * \param[in] clone Anonymous clone to check * \param[in] id Instance ID (without instance number) to check * \param[in] node Node to check * * \return */ static pcmk_resource_t * find_active_anon_instance(const pcmk_resource_t *clone, const char *id, const pcmk_node_t *node) { for (GList *iter = clone->priv->children; iter; iter = iter->next) { pcmk_resource_t *child = iter->data; pcmk_resource_t *active = NULL; // Use ->find_rsc() in case this is a cloned group active = clone->priv->fns->find_rsc(child, id, node, pcmk_rsc_match_clone_only |pcmk_rsc_match_current_node); if (active != NULL) { return active; } } return NULL; } /* * \brief Check whether an anonymous clone instance is known on a node * * \param[in] clone Anonymous clone to check * \param[in] id Instance ID (without instance number) to check * \param[in] node Node to check * * \return true if \p id instance of \p clone is known on \p node, * otherwise false */ static bool anonymous_known_on(const pcmk_resource_t *clone, const char *id, const pcmk_node_t *node) { for (GList *iter = clone->priv->children; iter; iter = iter->next) { pcmk_resource_t *child = iter->data; /* Use ->find_rsc() because this might be a cloned group, and knowing * that other members of the group are known here implies nothing. */ child = clone->priv->fns->find_rsc(child, id, NULL, pcmk_rsc_match_clone_only); CRM_LOG_ASSERT(child != NULL); if (child != NULL) { if (g_hash_table_lookup(child->priv->probed_nodes, node->priv->id)) { return true; } } } return false; } /*! * \internal * \brief Check whether a node is allowed to run a resource * * \param[in] rsc Resource to check * \param[in] node Node to check * * \return true if \p node is allowed to run \p rsc, otherwise false */ static bool is_allowed(const pcmk_resource_t *rsc, const pcmk_node_t *node) { pcmk_node_t *allowed = g_hash_table_lookup(rsc->priv->allowed_nodes, node->priv->id); return (allowed != NULL) && (allowed->assign->score >= 0); } /*! * \brief Check whether a clone instance's promotion score should be considered * * \param[in] rsc Promotable clone instance to check * \param[in] node Node where score would be applied * * \return true if \p rsc's promotion score should be considered on \p node, * otherwise false */ static bool promotion_score_applies(const pcmk_resource_t *rsc, const pcmk_node_t *node) { char *id = clone_strip(rsc->id); const pcmk_resource_t *parent = pe__const_top_resource(rsc, false); pcmk_resource_t *active = NULL; const char *reason = "allowed"; // Some checks apply only to anonymous clone instances if (!pcmk_is_set(rsc->flags, pcmk__rsc_unique)) { // If instance is active on the node, its score definitely applies active = find_active_anon_instance(parent, id, node); if (active == rsc) { reason = "active"; goto check_allowed; } /* If *no* instance is active on this node, this instance's score will * count if it has been probed on this node. */ if ((active == NULL) && anonymous_known_on(parent, id, node)) { reason = "probed"; goto check_allowed; } } /* If this clone's status is unknown on *all* nodes (e.g. cluster startup), * take all instances' scores into account, to make sure we use any * permanent promotion scores. */ if ((rsc->priv->active_nodes == NULL) && (g_hash_table_size(rsc->priv->probed_nodes) == 0)) { reason = "none probed"; goto check_allowed; } /* Otherwise, we've probed and/or started the resource *somewhere*, so * consider promotion scores on nodes where we know the status. */ if ((g_hash_table_lookup(rsc->priv->probed_nodes, node->priv->id) != NULL) || (pe_find_node_id(rsc->priv->active_nodes, node->priv->id) != NULL)) { reason = "known"; } else { pcmk__rsc_trace(rsc, "Ignoring %s promotion score (for %s) on %s: " "not probed", rsc->id, id, pcmk__node_name(node)); free(id); return false; } check_allowed: if (is_allowed(rsc, node)) { pcmk__rsc_trace(rsc, "Counting %s promotion score (for %s) on %s: %s", rsc->id, id, pcmk__node_name(node), reason); free(id); return true; } pcmk__rsc_trace(rsc, "Ignoring %s promotion score (for %s) on %s: not allowed", rsc->id, id, pcmk__node_name(node)); free(id); return false; } /*! * \internal * \brief Get the value of a promotion score node attribute * * \param[in] rsc Promotable clone instance to get promotion score for * \param[in] node Node to get promotion score for * \param[in] name Resource name to use in promotion score attribute name * * \return Value of promotion score node attribute for \p rsc on \p node */ static const char * promotion_attr_value(const pcmk_resource_t *rsc, const pcmk_node_t *node, const char *name) { char *attr_name = NULL; const char *attr_value = NULL; const char *target = NULL; enum pcmk__rsc_node node_type = pcmk__rsc_node_assigned; if (pcmk_is_set(rsc->flags, pcmk__rsc_unassigned)) { // Not assigned yet node_type = pcmk__rsc_node_current; } target = g_hash_table_lookup(rsc->priv->meta, PCMK_META_CONTAINER_ATTRIBUTE_TARGET); attr_name = pcmk_promotion_score_name(name); attr_value = pcmk__node_attr(node, attr_name, target, node_type); free(attr_name); return attr_value; } /*! * \internal * \brief Get the promotion score for a clone instance on a node * * \param[in] rsc Promotable clone instance to get score for * \param[in] node Node to get score for * \param[out] is_default If non-NULL, will be set true if no score available * * \return Promotion score for \p rsc on \p node (or 0 if none) */ static int promotion_score(const pcmk_resource_t *rsc, const pcmk_node_t *node, bool *is_default) { int score = 0; int rc = pcmk_rc_ok; const char *name = NULL; const char *attr_value = NULL; if (is_default != NULL) { *is_default = true; } CRM_CHECK((rsc != NULL) && (node != NULL), return 0); /* If this is an instance of a cloned group, the promotion score is the sum * of all members' promotion scores. */ if (rsc->priv->children != NULL) { int score = 0; for (const GList *iter = rsc->priv->children; iter != NULL; iter = iter->next) { const pcmk_resource_t *child = (const pcmk_resource_t *) iter->data; bool child_default = false; int child_score = promotion_score(child, node, &child_default); if (!child_default && (is_default != NULL)) { *is_default = false; } score += child_score; } return score; } if (!promotion_score_applies(rsc, node)) { return 0; } /* For the promotion score attribute name, use the name the resource is * known as in resource history, since that's what crm_attribute --promotion * would have used. */ name = pcmk__s(rsc->priv->history_id, rsc->id); attr_value = promotion_attr_value(rsc, node, name); if (attr_value != NULL) { pcmk__rsc_trace(rsc, "Promotion score for %s on %s = %s", name, pcmk__node_name(node), pcmk__s(attr_value, "(unset)")); } else if (!pcmk_is_set(rsc->flags, pcmk__rsc_unique)) { /* If we don't have any resource history yet, we won't have history_id. * In that case, for anonymous clones, try the resource name without * any instance number. */ char *rsc_name = clone_strip(rsc->id); if (strcmp(rsc->id, rsc_name) != 0) { attr_value = promotion_attr_value(rsc, node, rsc_name); pcmk__rsc_trace(rsc, "Promotion score for %s on %s (for %s) = %s", rsc_name, pcmk__node_name(node), rsc->id, pcmk__s(attr_value, "(unset)")); } free(rsc_name); } if (attr_value == NULL) { return 0; } if (is_default != NULL) { *is_default = false; } rc = pcmk_parse_score(attr_value, &score, 0); if (rc != pcmk_rc_ok) { crm_warn("Using 0 as promotion score for %s on %s " "because '%s' is not a valid score", rsc->id, pcmk__node_name(node), attr_value); } return score; } /*! * \internal * \brief Include promotion scores in instances' node scores and priorities * * \param[in,out] rsc Promotable clone resource to update */ void pcmk__add_promotion_scores(pcmk_resource_t *rsc) { if (pe__set_clone_flag(rsc, pcmk__clone_promotion_added) == pcmk_rc_already) { return; } for (GList *iter = rsc->priv->children; iter != NULL; iter = iter->next) { pcmk_resource_t *child_rsc = (pcmk_resource_t *) iter->data; GHashTableIter iter; pcmk_node_t *node = NULL; int score, new_score; g_hash_table_iter_init(&iter, child_rsc->priv->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (void **) &node)) { - if (!pcmk__node_available(node, false, false)) { + if (!pcmk__node_available(node, + pcmk__node_alive|pcmk__node_usable)) { /* This node will never be promoted, so don't apply the * promotion score, as that may lead to clone shuffling. */ continue; } score = promotion_score(child_rsc, node, NULL); if (score > 0) { new_score = pcmk__add_scores(node->assign->score, score); if (new_score != node->assign->score) { // Could remain INFINITY node->assign->score = new_score; pcmk__rsc_trace(rsc, "Added %s promotion priority (%s) to score " "on %s (now %s)", child_rsc->id, pcmk_readable_score(score), pcmk__node_name(node), pcmk_readable_score(new_score)); } } if (score > child_rsc->priv->priority) { pcmk__rsc_trace(rsc, "Updating %s priority to promotion score " "(%d->%d)", child_rsc->id, child_rsc->priv->priority, score); child_rsc->priv->priority = score; } } } } /*! * \internal * \brief If a resource's current role is started, change it to unpromoted * * \param[in,out] data Resource to update * \param[in] user_data Ignored */ static void set_current_role_unpromoted(void *data, void *user_data) { pcmk_resource_t *rsc = (pcmk_resource_t *) data; if (rsc->priv->orig_role == pcmk_role_started) { // Promotable clones should use unpromoted role instead of started rsc->priv->orig_role = pcmk_role_unpromoted; } g_list_foreach(rsc->priv->children, set_current_role_unpromoted, NULL); } /*! * \internal * \brief Set a resource's next role to unpromoted (or stopped if unassigned) * * \param[in,out] data Resource to update * \param[in] user_data Ignored */ static void set_next_role_unpromoted(void *data, void *user_data) { pcmk_resource_t *rsc = (pcmk_resource_t *) data; GList *assigned = NULL; rsc->priv->fns->location(rsc, &assigned, pcmk__rsc_node_assigned); if (assigned == NULL) { pe__set_next_role(rsc, pcmk_role_stopped, "stopped instance"); } else { pe__set_next_role(rsc, pcmk_role_unpromoted, "unpromoted instance"); g_list_free(assigned); } g_list_foreach(rsc->priv->children, set_next_role_unpromoted, NULL); } /*! * \internal * \brief Set a resource's next role to promoted if not already set * * \param[in,out] data Resource to update * \param[in] user_data Ignored */ static void set_next_role_promoted(void *data, gpointer user_data) { pcmk_resource_t *rsc = (pcmk_resource_t *) data; if (rsc->priv->next_role == pcmk_role_unknown) { pe__set_next_role(rsc, pcmk_role_promoted, "promoted instance"); } g_list_foreach(rsc->priv->children, set_next_role_promoted, NULL); } /*! * \internal * \brief Show instance's promotion score on node where it will be active * * \param[in,out] instance Promotable clone instance to show */ static void show_promotion_score(pcmk_resource_t *instance) { pcmk_node_t *chosen = NULL; const char *score_s = NULL; chosen = instance->priv->fns->location(instance, NULL, pcmk__rsc_node_assigned); score_s = pcmk_readable_score(instance->priv->promotion_priority); if (pcmk_is_set(instance->priv->scheduler->flags, pcmk__sched_output_scores) && !pcmk__is_daemon && (instance->priv->scheduler->priv->out != NULL)) { pcmk__output_t *out = instance->priv->scheduler->priv->out; out->message(out, "promotion-score", instance, chosen, score_s); } else if (chosen == NULL) { pcmk__rsc_debug(pe__const_top_resource(instance, false), "%s promotion score (inactive): %s (priority=%d)", instance->id, score_s, instance->priv->priority); } else { pcmk__rsc_debug(pe__const_top_resource(instance, false), "%s promotion score on %s: %s (priority=%d)", instance->id, pcmk__node_name(chosen), score_s, instance->priv->priority); } } /*! * \internal * \brief Set a clone instance's promotion priority * * \param[in,out] data Promotable clone instance to update * \param[in] user_data Instance's parent clone */ static void set_instance_priority(gpointer data, gpointer user_data) { pcmk_resource_t *instance = (pcmk_resource_t *) data; const pcmk_resource_t *clone = (const pcmk_resource_t *) user_data; const pcmk_node_t *chosen = NULL; enum rsc_role_e next_role = pcmk_role_unknown; GList *list = NULL; pcmk__rsc_trace(clone, "Assigning priority for %s: %s", instance->id, pcmk_role_text(instance->priv->next_role)); if (instance->priv->fns->state(instance, TRUE) == pcmk_role_started) { set_current_role_unpromoted(instance, NULL); } // Only an instance that will be active can be promoted chosen = instance->priv->fns->location(instance, &list, pcmk__rsc_node_assigned); if (pcmk__list_of_multiple(list)) { pcmk__config_err("Cannot promote non-colocated child %s", instance->id); } g_list_free(list); if (chosen == NULL) { return; } next_role = instance->priv->fns->state(instance, FALSE); switch (next_role) { case pcmk_role_started: case pcmk_role_unknown: // Set instance priority to its promotion score (or -1 if none) { bool is_default = false; instance->priv->priority = promotion_score(instance, chosen, &is_default); if (is_default) { /* Default to -1 if no value is set. This allows instances * eligible for promotion to be specified based solely on * PCMK_XE_RSC_LOCATION constraints, but prevents any * instance from being promoted if neither a constraint nor * a promotion score is present. */ instance->priv->priority = -1; } } break; case pcmk_role_unpromoted: case pcmk_role_stopped: // Instance can't be promoted instance->priv->priority = -PCMK_SCORE_INFINITY; break; case pcmk_role_promoted: // Nothing needed (re-creating actions after scheduling fencing) break; default: CRM_CHECK(FALSE, crm_err("Unknown resource role %d for %s", next_role, instance->id)); } // Add relevant location constraint scores for promoted role apply_promoted_locations(instance, instance->priv->location_constraints, chosen); apply_promoted_locations(instance, clone->priv->location_constraints, chosen); // Consider instance's role-based colocations with other resources list = pcmk__this_with_colocations(instance); for (GList *iter = list; iter != NULL; iter = iter->next) { pcmk__colocation_t *cons = (pcmk__colocation_t *) iter->data; instance->priv->cmds->apply_coloc_score(instance, cons->primary, cons, true); } g_list_free(list); instance->priv->promotion_priority = instance->priv->priority; if (next_role == pcmk_role_promoted) { instance->priv->promotion_priority = PCMK_SCORE_INFINITY; } pcmk__rsc_trace(clone, "Assigning %s priority = %d", instance->id, instance->priv->priority); } /*! * \internal * \brief Set a promotable clone instance's role * * \param[in,out] data Promotable clone instance to update * \param[in,out] user_data Pointer to count of instances chosen for promotion */ static void set_instance_role(gpointer data, gpointer user_data) { pcmk_resource_t *instance = (pcmk_resource_t *) data; int *count = (int *) user_data; const pcmk_resource_t *clone = pe__const_top_resource(instance, false); const pcmk_scheduler_t *scheduler = instance->priv->scheduler; pcmk_node_t *chosen = NULL; show_promotion_score(instance); if (instance->priv->promotion_priority < 0) { pcmk__rsc_trace(clone, "Not supposed to promote instance %s", instance->id); } else if ((*count < pe__clone_promoted_max(instance)) || !pcmk_is_set(clone->flags, pcmk__rsc_managed)) { chosen = node_to_be_promoted_on(instance); } if (chosen == NULL) { set_next_role_unpromoted(instance, NULL); return; } if ((instance->priv->orig_role < pcmk_role_promoted) && !pcmk_is_set(scheduler->flags, pcmk__sched_quorate) && (scheduler->no_quorum_policy == pcmk_no_quorum_freeze)) { crm_notice("Clone instance %s cannot be promoted without quorum", instance->id); set_next_role_unpromoted(instance, NULL); return; } chosen->assign->count++; pcmk__rsc_info(clone, "Choosing %s (%s) on %s for promotion", instance->id, pcmk_role_text(instance->priv->orig_role), pcmk__node_name(chosen)); set_next_role_promoted(instance, NULL); (*count)++; } /*! * \internal * \brief Set roles for all instances of a promotable clone * * \param[in,out] rsc Promotable clone resource to update */ void pcmk__set_instance_roles(pcmk_resource_t *rsc) { int promoted = 0; GHashTableIter iter; pcmk_node_t *node = NULL; // Repurpose count to track the number of promoted instances assigned g_hash_table_iter_init(&iter, rsc->priv->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (void **)&node)) { node->assign->count = 0; } // Set instances' promotion priorities and sort by highest priority first g_list_foreach(rsc->priv->children, set_instance_priority, rsc); sort_promotable_instances(rsc); // Choose the first N eligible instances to be promoted g_list_foreach(rsc->priv->children, set_instance_role, &promoted); pcmk__rsc_info(rsc, "%s: Promoted %d instances of a possible %d", rsc->id, promoted, pe__clone_promoted_max(rsc)); } /*! * * \internal * \brief Create actions for promotable clone instances * * \param[in,out] clone Promotable clone to create actions for * \param[out] any_promoting Will be set true if any instance is promoting * \param[out] any_demoting Will be set true if any instance is demoting */ static void create_promotable_instance_actions(pcmk_resource_t *clone, bool *any_promoting, bool *any_demoting) { for (GList *iter = clone->priv->children; iter != NULL; iter = iter->next) { pcmk_resource_t *instance = (pcmk_resource_t *) iter->data; instance->priv->cmds->create_actions(instance); check_for_role_change(instance, any_demoting, any_promoting); } } /*! * \internal * \brief Reset each promotable instance's resource priority * * Reset the priority of each instance of a promotable clone to the clone's * priority (after promotion actions are scheduled, when instance priorities * were repurposed as promotion scores). * * \param[in,out] clone Promotable clone to reset */ static void reset_instance_priorities(pcmk_resource_t *clone) { for (GList *iter = clone->priv->children; iter != NULL; iter = iter->next) { pcmk_resource_t *instance = (pcmk_resource_t *) iter->data; instance->priv->priority = clone->priv->priority; } } /*! * \internal * \brief Create actions specific to promotable clones * * \param[in,out] clone Promotable clone to create actions for */ void pcmk__create_promotable_actions(pcmk_resource_t *clone) { bool any_promoting = false; bool any_demoting = false; // Create actions for each clone instance individually create_promotable_instance_actions(clone, &any_promoting, &any_demoting); // Create pseudo-actions for clone as a whole pe__create_promotable_pseudo_ops(clone, any_promoting, any_demoting); // Undo our temporary repurposing of resource priority for instances reset_instance_priorities(clone); } /*! * \internal * \brief Create internal orderings for a promotable clone's instances * * \param[in,out] clone Promotable clone instance to order */ void pcmk__order_promotable_instances(pcmk_resource_t *clone) { pcmk_resource_t *previous = NULL; // Needed for ordered clones pcmk__promotable_restart_ordering(clone); for (GList *iter = clone->priv->children; iter != NULL; iter = iter->next) { pcmk_resource_t *instance = (pcmk_resource_t *) iter->data; // Demote before promote pcmk__order_resource_actions(instance, PCMK_ACTION_DEMOTE, instance, PCMK_ACTION_PROMOTE, pcmk__ar_ordered); order_instance_promotion(clone, instance, previous); order_instance_demotion(clone, instance, previous); previous = instance; } } /*! * \internal * \brief Update dependent's allowed nodes for colocation with promotable * * \param[in,out] dependent Dependent resource to update * \param[in] primary Primary resource * \param[in] primary_node Node where an instance of the primary will be * \param[in] colocation Colocation constraint to apply */ static void update_dependent_allowed_nodes(pcmk_resource_t *dependent, const pcmk_resource_t *primary, const pcmk_node_t *primary_node, const pcmk__colocation_t *colocation) { GHashTableIter iter; pcmk_node_t *node = NULL; const char *primary_value = NULL; const char *attr = colocation->node_attribute; if (colocation->score >= PCMK_SCORE_INFINITY) { return; // Colocation is mandatory, so allowed node scores don't matter } primary_value = pcmk__colocation_node_attr(primary_node, attr, primary); pcmk__rsc_trace(colocation->primary, "Applying %s (%s with %s on %s by %s @%d) to %s", colocation->id, colocation->dependent->id, colocation->primary->id, pcmk__node_name(primary_node), attr, colocation->score, dependent->id); g_hash_table_iter_init(&iter, dependent->priv->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (void **) &node)) { const char *dependent_value = pcmk__colocation_node_attr(node, attr, dependent); if (pcmk__str_eq(primary_value, dependent_value, pcmk__str_casei)) { node->assign->score = pcmk__add_scores(node->assign->score, colocation->score); pcmk__rsc_trace(colocation->primary, "Added %s score (%s) to %s (now %s)", colocation->id, pcmk_readable_score(colocation->score), pcmk__node_name(node), pcmk_readable_score(node->assign->score)); } } } /*! * \brief Update dependent for a colocation with a promotable clone * * \param[in] primary Primary resource in the colocation * \param[in,out] dependent Dependent resource in the colocation * \param[in] colocation Colocation constraint to apply */ void pcmk__update_dependent_with_promotable(const pcmk_resource_t *primary, pcmk_resource_t *dependent, const pcmk__colocation_t *colocation) { GList *affected_nodes = NULL; /* Build a list of all nodes where an instance of the primary will be, and * (for optional colocations) update the dependent's allowed node scores for * each one. */ for (GList *iter = primary->priv->children; iter != NULL; iter = iter->next) { pcmk_resource_t *instance = (pcmk_resource_t *) iter->data; pcmk_node_t *node = NULL; node = instance->priv->fns->location(instance, NULL, pcmk__rsc_node_assigned); if (node == NULL) { continue; } if (instance->priv->fns->state(instance, FALSE) == colocation->primary_role) { update_dependent_allowed_nodes(dependent, primary, node, colocation); affected_nodes = g_list_prepend(affected_nodes, node); } } /* For mandatory colocations, add the primary's node score to the * dependent's node score for each affected node, and ban the dependent * from all other nodes. * * However, skip this for promoted-with-promoted colocations, otherwise * inactive dependent instances can't start (in the unpromoted role). */ if ((colocation->score >= PCMK_SCORE_INFINITY) && ((colocation->dependent_role != pcmk_role_promoted) || (colocation->primary_role != pcmk_role_promoted))) { pcmk__rsc_trace(colocation->primary, "Applying %s (mandatory %s with %s) to %s", colocation->id, colocation->dependent->id, colocation->primary->id, dependent->id); pcmk__colocation_intersect_nodes(dependent, primary, colocation, affected_nodes, true); } g_list_free(affected_nodes); } /*! * \internal * \brief Update dependent priority for colocation with promotable * * \param[in] primary Primary resource in the colocation * \param[in,out] dependent Dependent resource in the colocation * \param[in] colocation Colocation constraint to apply * * \return The score added to the dependent's priority */ int pcmk__update_promotable_dependent_priority(const pcmk_resource_t *primary, pcmk_resource_t *dependent, const pcmk__colocation_t *colocation) { pcmk_resource_t *primary_instance = NULL; // Look for a primary instance where dependent will be primary_instance = pcmk__find_compatible_instance(dependent, primary, colocation->primary_role, false); if (primary_instance != NULL) { // Add primary instance's priority to dependent's int new_priority = pcmk__add_scores(dependent->priv->priority, colocation->score); pcmk__rsc_trace(colocation->primary, "Applying %s (%s with %s) to %s priority " "(%s + %s = %s)", colocation->id, colocation->dependent->id, colocation->primary->id, dependent->id, pcmk_readable_score(dependent->priv->priority), pcmk_readable_score(colocation->score), pcmk_readable_score(new_priority)); dependent->priv->priority = new_priority; return colocation->score; } if (colocation->score >= PCMK_SCORE_INFINITY) { // Mandatory colocation, but primary won't be here pcmk__rsc_trace(colocation->primary, "Applying %s (%s with %s) to %s: can't be promoted", colocation->id, colocation->dependent->id, colocation->primary->id, dependent->id); dependent->priv->priority = -PCMK_SCORE_INFINITY; return -PCMK_SCORE_INFINITY; } return 0; } diff --git a/lib/pacemaker/pcmk_sched_resource.c b/lib/pacemaker/pcmk_sched_resource.c index 2bc843f05f..fa4330bf3c 100644 --- a/lib/pacemaker/pcmk_sched_resource.c +++ b/lib/pacemaker/pcmk_sched_resource.c @@ -1,800 +1,805 @@ /* - * Copyright 2014-2024 the Pacemaker project contributors + * Copyright 2014-2025 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" // Resource assignment methods by resource variant static pcmk__assignment_methods_t assignment_methods[] = { { pcmk__primitive_assign, pcmk__primitive_create_actions, pcmk__probe_rsc_on_node, pcmk__primitive_internal_constraints, pcmk__primitive_apply_coloc_score, pcmk__colocated_resources, pcmk__with_primitive_colocations, pcmk__primitive_with_colocations, pcmk__add_colocated_node_scores, pcmk__apply_location, pcmk__primitive_action_flags, pcmk__update_ordered_actions, pcmk__output_resource_actions, pcmk__add_rsc_actions_to_graph, pcmk__primitive_add_graph_meta, pcmk__primitive_add_utilization, pcmk__primitive_shutdown_lock, }, { pcmk__group_assign, pcmk__group_create_actions, pcmk__probe_rsc_on_node, pcmk__group_internal_constraints, pcmk__group_apply_coloc_score, pcmk__group_colocated_resources, pcmk__with_group_colocations, pcmk__group_with_colocations, pcmk__group_add_colocated_node_scores, pcmk__group_apply_location, pcmk__group_action_flags, pcmk__group_update_ordered_actions, pcmk__output_resource_actions, pcmk__add_rsc_actions_to_graph, pcmk__noop_add_graph_meta, pcmk__group_add_utilization, pcmk__group_shutdown_lock, }, { pcmk__clone_assign, pcmk__clone_create_actions, pcmk__clone_create_probe, pcmk__clone_internal_constraints, pcmk__clone_apply_coloc_score, pcmk__colocated_resources, pcmk__with_clone_colocations, pcmk__clone_with_colocations, pcmk__add_colocated_node_scores, pcmk__clone_apply_location, pcmk__clone_action_flags, pcmk__instance_update_ordered_actions, pcmk__output_resource_actions, pcmk__clone_add_actions_to_graph, pcmk__clone_add_graph_meta, pcmk__clone_add_utilization, pcmk__clone_shutdown_lock, }, { pcmk__bundle_assign, pcmk__bundle_create_actions, pcmk__bundle_create_probe, pcmk__bundle_internal_constraints, pcmk__bundle_apply_coloc_score, pcmk__colocated_resources, pcmk__with_bundle_colocations, pcmk__bundle_with_colocations, pcmk__add_colocated_node_scores, pcmk__bundle_apply_location, pcmk__bundle_action_flags, pcmk__instance_update_ordered_actions, pcmk__output_bundle_actions, pcmk__bundle_add_actions_to_graph, pcmk__noop_add_graph_meta, pcmk__bundle_add_utilization, pcmk__bundle_shutdown_lock, } }; /*! * \internal * \brief Check whether a resource's agent standard, provider, or type changed * * \param[in,out] rsc Resource to check * \param[in,out] node Node needing unfencing if agent changed * \param[in] rsc_entry XML with previously known agent information * \param[in] active_on_node Whether \p rsc is active on \p node * * \return true if agent for \p rsc changed, otherwise false */ bool pcmk__rsc_agent_changed(pcmk_resource_t *rsc, pcmk_node_t *node, const xmlNode *rsc_entry, bool active_on_node) { bool changed = false; const char *attr_list[] = { PCMK_XA_TYPE, PCMK_XA_CLASS, PCMK_XA_PROVIDER, }; for (int i = 0; i < PCMK__NELEM(attr_list); i++) { const char *value = crm_element_value(rsc->priv->xml, attr_list[i]); const char *old_value = crm_element_value(rsc_entry, attr_list[i]); if (!pcmk__str_eq(value, old_value, pcmk__str_none)) { changed = true; trigger_unfencing(rsc, node, "Device definition changed", NULL, rsc->priv->scheduler); if (active_on_node) { crm_notice("Forcing restart of %s on %s " "because %s changed from '%s' to '%s'", rsc->id, pcmk__node_name(node), attr_list[i], pcmk__s(old_value, ""), pcmk__s(value, "")); } } } if (changed && active_on_node) { // Make sure the resource is restarted custom_action(rsc, stop_key(rsc), PCMK_ACTION_STOP, node, FALSE, rsc->priv->scheduler); pcmk__set_rsc_flags(rsc, pcmk__rsc_start_pending); } return changed; } /*! * \internal * \brief Add resource (and any matching children) to list if it matches ID * * \param[in] result List to add resource to * \param[in] rsc Resource to check * \param[in] id ID to match * * \return (Possibly new) head of list */ static GList * add_rsc_if_matching(GList *result, pcmk_resource_t *rsc, const char *id) { if (pcmk__str_eq(id, rsc->id, pcmk__str_none) || pcmk__str_eq(id, rsc->priv->history_id, pcmk__str_none)) { result = g_list_prepend(result, rsc); } for (GList *iter = rsc->priv->children; iter != NULL; iter = iter->next) { pcmk_resource_t *child = (pcmk_resource_t *) iter->data; result = add_rsc_if_matching(result, child, id); } return result; } /*! * \internal * \brief Find all resources matching a given ID by either ID or clone name * * \param[in] id Resource ID to check * \param[in] scheduler Scheduler data * * \return List of all resources that match \p id * \note The caller is responsible for freeing the return value with * g_list_free(). */ GList * pcmk__rscs_matching_id(const char *id, const pcmk_scheduler_t *scheduler) { GList *result = NULL; CRM_CHECK((id != NULL) && (scheduler != NULL), return NULL); for (GList *iter = scheduler->priv->resources; iter != NULL; iter = iter->next) { result = add_rsc_if_matching(result, (pcmk_resource_t *) iter->data, id); } return result; } /*! * \internal * \brief Set the variant-appropriate assignment methods for a resource * * \param[in,out] data Resource to set assignment methods for * \param[in] user_data Ignored */ static void set_assignment_methods_for_rsc(gpointer data, gpointer user_data) { pcmk_resource_t *rsc = data; rsc->priv->cmds = &assignment_methods[rsc->priv->variant]; g_list_foreach(rsc->priv->children, set_assignment_methods_for_rsc, NULL); } /*! * \internal * \brief Set the variant-appropriate assignment methods for all resources * * \param[in,out] scheduler Scheduler data */ void pcmk__set_assignment_methods(pcmk_scheduler_t *scheduler) { g_list_foreach(scheduler->priv->resources, set_assignment_methods_for_rsc, NULL); } /*! * \internal * \brief Wrapper for colocated_resources() method for readability * * \param[in] rsc Resource to add to colocated list * \param[in] orig_rsc Resource originally requested * \param[in,out] list Pointer to list to add to * * \return (Possibly new) head of list */ static inline void add_colocated_resources(const pcmk_resource_t *rsc, const pcmk_resource_t *orig_rsc, GList **list) { *list = rsc->priv->cmds->colocated_resources(rsc, orig_rsc, *list); } // Shared implementation of pcmk__assignment_methods_t:colocated_resources() GList * pcmk__colocated_resources(const pcmk_resource_t *rsc, const pcmk_resource_t *orig_rsc, GList *colocated_rscs) { const GList *iter = NULL; GList *colocations = NULL; if (orig_rsc == NULL) { orig_rsc = rsc; } if ((rsc == NULL) || (g_list_find(colocated_rscs, rsc) != NULL)) { return colocated_rscs; } pcmk__rsc_trace(orig_rsc, "%s is in colocation chain with %s", rsc->id, orig_rsc->id); colocated_rscs = g_list_prepend(colocated_rscs, (gpointer) rsc); // Follow colocations where this resource is the dependent resource colocations = pcmk__this_with_colocations(rsc); for (iter = colocations; iter != NULL; iter = iter->next) { const pcmk__colocation_t *constraint = iter->data; const pcmk_resource_t *primary = constraint->primary; if (primary == orig_rsc) { continue; // Break colocation loop } if ((constraint->score == PCMK_SCORE_INFINITY) && (pcmk__colocation_affects(rsc, primary, constraint, true) == pcmk__coloc_affects_location)) { add_colocated_resources(primary, orig_rsc, &colocated_rscs); } } g_list_free(colocations); // Follow colocations where this resource is the primary resource colocations = pcmk__with_this_colocations(rsc); for (iter = colocations; iter != NULL; iter = iter->next) { const pcmk__colocation_t *constraint = iter->data; const pcmk_resource_t *dependent = constraint->dependent; if (dependent == orig_rsc) { continue; // Break colocation loop } if (pcmk__is_clone(rsc) && !pcmk__is_clone(dependent)) { continue; // We can't be sure whether dependent will be colocated } if ((constraint->score == PCMK_SCORE_INFINITY) && (pcmk__colocation_affects(dependent, rsc, constraint, true) == pcmk__coloc_affects_location)) { add_colocated_resources(dependent, orig_rsc, &colocated_rscs); } } g_list_free(colocations); return colocated_rscs; } // No-op function for variants that don't need to implement add_graph_meta() void pcmk__noop_add_graph_meta(const pcmk_resource_t *rsc, xmlNode *xml) { } /*! * \internal * \brief Output a summary of scheduled actions for a resource * * \param[in,out] rsc Resource to output actions for */ void pcmk__output_resource_actions(pcmk_resource_t *rsc) { pcmk_node_t *next = NULL; pcmk_node_t *current = NULL; pcmk__output_t *out = NULL; pcmk__assert(rsc != NULL); out = rsc->priv->scheduler->priv->out; if (rsc->priv->children != NULL) { for (GList *iter = rsc->priv->children; iter != NULL; iter = iter->next) { pcmk_resource_t *child = (pcmk_resource_t *) iter->data; child->priv->cmds->output_actions(child); } return; } next = rsc->priv->assigned_node; if (rsc->priv->active_nodes != NULL) { current = pcmk__current_node(rsc); if (rsc->priv->orig_role == pcmk_role_stopped) { /* This can occur when resources are being recovered because * the current role can change in pcmk__primitive_create_actions() */ rsc->priv->orig_role = pcmk_role_started; } } if ((current == NULL) && pcmk_is_set(rsc->flags, pcmk__rsc_removed)) { /* Don't log stopped orphans */ return; } out->message(out, "rsc-action", rsc, current, next); } /*! * \internal * \brief Add a resource to a node's list of assigned resources * * \param[in,out] node Node to add resource to * \param[in] rsc Resource to add */ static inline void add_assigned_resource(pcmk_node_t *node, pcmk_resource_t *rsc) { node->priv->assigned_resources = g_list_prepend(node->priv->assigned_resources, rsc); } /*! * \internal * \brief Assign a specified resource (of any variant) to a node * * Assign a specified resource and its children (if any) to a specified node, if * the node can run the resource (or unconditionally, if \p force is true). Mark * the resources as no longer provisional. * * If a resource can't be assigned (or \p node is \c NULL), unassign any * previous assignment. If \p stop_if_fail is \c true, set next role to stopped * and update any existing actions scheduled for the resource. * * \param[in,out] rsc Resource to assign * \param[in,out] node Node to assign \p rsc to * \param[in] force If true, assign to \p node even if unavailable * \param[in] stop_if_fail If \c true and either \p rsc can't be assigned * or \p chosen is \c NULL, set next role to * stopped and update existing actions (if \p rsc * is not a primitive, this applies to its * primitive descendants instead) * * \return \c true if the assignment of \p rsc changed, or \c false otherwise * * \note Assigning a resource to the NULL node using this function is different * from calling pcmk__unassign_resource(), in that it may also update any * actions created for the resource. * \note The \c pcmk__assignment_methods_t:assign() method is preferred, unless * a resource should be assigned to the \c NULL node or every resource in * a tree should be assigned to the same 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. */ bool pcmk__assign_resource(pcmk_resource_t *rsc, pcmk_node_t *node, bool force, bool stop_if_fail) { bool changed = false; pcmk_scheduler_t *scheduler = NULL; pcmk__assert(rsc != NULL); scheduler = rsc->priv->scheduler; if (rsc->priv->children != NULL) { for (GList *iter = rsc->priv->children; iter != NULL; iter = iter->next) { pcmk_resource_t *child_rsc = iter->data; changed |= pcmk__assign_resource(child_rsc, node, force, stop_if_fail); } return changed; } // Assigning a primitive - if (!force && (node != NULL) - && ((node->assign->score < 0) + if (!force && (node != NULL)) { + bool available = pcmk__node_available(node, pcmk__node_alive + |pcmk__node_usable + |pcmk__node_no_negative); + + if ((node->assign->score < 0) // Allow graph to assume that guest node connections will come up - || (!pcmk__node_available(node, true, false) - && !pcmk__is_guest_or_bundle_node(node)))) { - - pcmk__rsc_debug(rsc, - "All nodes for resource %s are unavailable, unclean or " - "shutting down (%s can%s run resources, with score %s)", - rsc->id, pcmk__node_name(node), - (pcmk__node_available(node, true, false)? "" : "not"), - pcmk_readable_score(node->assign->score)); - - if (stop_if_fail) { - pe__set_next_role(rsc, pcmk_role_stopped, "node availability"); + || (!available && !pcmk__is_guest_or_bundle_node(node))) { + + pcmk__rsc_debug(rsc, + "All nodes for resource %s are unavailable, " + "unclean or shutting down (%s can%s run " + "resources, with score %s)", + rsc->id, pcmk__node_name(node), + (available? "" : "not"), + pcmk_readable_score(node->assign->score)); + + if (stop_if_fail) { + pe__set_next_role(rsc, pcmk_role_stopped, "node availability"); + } + node = NULL; } - node = NULL; } if (rsc->priv->assigned_node != NULL) { changed = !pcmk__same_node(rsc->priv->assigned_node, node); } else { changed = (node != NULL); } pcmk__unassign_resource(rsc); pcmk__clear_rsc_flags(rsc, pcmk__rsc_unassigned); if (node == NULL) { char *rc_stopped = NULL; pcmk__rsc_debug(rsc, "Could not assign %s to a node", rsc->id); if (!stop_if_fail) { return changed; } pe__set_next_role(rsc, pcmk_role_stopped, "unable to assign"); for (GList *iter = rsc->priv->actions; iter != NULL; iter = iter->next) { pcmk_action_t *op = (pcmk_action_t *) iter->data; pcmk__rsc_debug(rsc, "Updating %s for %s assignment failure", op->uuid, rsc->id); if (pcmk__str_eq(op->task, PCMK_ACTION_STOP, pcmk__str_none)) { pcmk__clear_action_flags(op, pcmk__action_optional); } else if (pcmk__str_eq(op->task, PCMK_ACTION_START, pcmk__str_none)) { pcmk__clear_action_flags(op, pcmk__action_runnable); } else { // Cancel recurring actions, unless for stopped state const char *interval_ms_s = NULL; const char *target_rc_s = NULL; interval_ms_s = g_hash_table_lookup(op->meta, PCMK_META_INTERVAL); target_rc_s = g_hash_table_lookup(op->meta, PCMK__META_OP_TARGET_RC); if (rc_stopped == NULL) { rc_stopped = pcmk__itoa(PCMK_OCF_NOT_RUNNING); } if (!pcmk__str_eq(interval_ms_s, "0", pcmk__str_null_matches) && !pcmk__str_eq(rc_stopped, target_rc_s, pcmk__str_none)) { pcmk__clear_action_flags(op, pcmk__action_runnable); } } } free(rc_stopped); return changed; } pcmk__rsc_debug(rsc, "Assigning %s to %s", rsc->id, pcmk__node_name(node)); rsc->priv->assigned_node = pe__copy_node(node); add_assigned_resource(node, rsc); node->priv->num_resources++; node->assign->count++; pcmk__consume_node_capacity(node->priv->utilization, rsc); if (pcmk_is_set(scheduler->flags, pcmk__sched_show_utilization)) { pcmk__output_t *out = scheduler->priv->out; out->message(out, "resource-util", rsc, node, __func__); } return changed; } /*! * \internal * \brief Remove any node assignment from a specified resource and its children * * If a specified resource has been assigned to a node, remove that assignment * and mark the resource as provisional again. * * \param[in,out] rsc Resource to unassign * * \note This function is called recursively on \p rsc and its children. */ void pcmk__unassign_resource(pcmk_resource_t *rsc) { pcmk_node_t *old = rsc->priv->assigned_node; if (old == NULL) { crm_info("Unassigning %s", rsc->id); } else { crm_info("Unassigning %s from %s", rsc->id, pcmk__node_name(old)); } pcmk__set_rsc_flags(rsc, pcmk__rsc_unassigned); if (rsc->priv->children == NULL) { if (old == NULL) { return; } rsc->priv->assigned_node = NULL; /* We're going to free the pcmk_node_t copy, but its priv member is * shared and will remain, so update that appropriately first. */ old->priv->assigned_resources = g_list_remove(old->priv->assigned_resources, rsc); old->priv->num_resources--; pcmk__release_node_capacity(old->priv->utilization, rsc); pcmk__free_node_copy(old); return; } for (GList *iter = rsc->priv->children; iter != NULL; iter = iter->next) { pcmk__unassign_resource((pcmk_resource_t *) iter->data); } } /*! * \internal * \brief Check whether a resource has reached its migration threshold on a node * * \param[in,out] rsc Resource to check * \param[in] node Node to check * \param[out] failed If threshold has been reached, this will be set to * resource that failed (possibly a parent of \p rsc) * * \return true if the migration threshold has been reached, false otherwise */ bool pcmk__threshold_reached(pcmk_resource_t *rsc, const pcmk_node_t *node, pcmk_resource_t **failed) { int fail_count, remaining_tries; pcmk_resource_t *rsc_to_ban = rsc; // Migration threshold of 0 means never force away if (rsc->priv->ban_after_failures == 0) { return false; } // If we're ignoring failures, also ignore the migration threshold if (pcmk_is_set(rsc->flags, pcmk__rsc_ignore_failure)) { return false; } // If there are no failures, there's no need to force away fail_count = pe_get_failcount(node, rsc, NULL, pcmk__fc_effective|pcmk__fc_launched, NULL); if (fail_count <= 0) { return false; } // If failed resource is anonymous clone instance, we'll force clone away if (!pcmk_is_set(rsc->flags, pcmk__rsc_unique)) { rsc_to_ban = uber_parent(rsc); } // How many more times recovery will be tried on this node remaining_tries = rsc->priv->ban_after_failures - fail_count; if (remaining_tries <= 0) { pcmk__sched_warn(rsc->priv->scheduler, "%s cannot run on %s due to reaching migration " "threshold (clean up resource to allow again) " QB_XS " failures=%d " PCMK_META_MIGRATION_THRESHOLD "=%d", rsc_to_ban->id, pcmk__node_name(node), fail_count, rsc->priv->ban_after_failures); if (failed != NULL) { *failed = rsc_to_ban; } return true; } crm_info("%s can fail %d more time%s on " "%s before reaching migration threshold (%d)", rsc_to_ban->id, remaining_tries, pcmk__plural_s(remaining_tries), pcmk__node_name(node), rsc->priv->ban_after_failures); return false; } /*! * \internal * \brief Get a node's score * * \param[in] node Node with ID to check * \param[in] nodes List of nodes to look for \p node score in * * \return Node's score, or -INFINITY if not found */ static int get_node_score(const pcmk_node_t *node, GHashTable *nodes) { pcmk_node_t *found_node = NULL; if ((node != NULL) && (nodes != NULL)) { found_node = g_hash_table_lookup(nodes, node->priv->id); } if (found_node == NULL) { return -PCMK_SCORE_INFINITY; } return found_node->assign->score; } /*! * \internal * \brief Compare two resources according to which should be assigned first * * \param[in] a First resource to compare * \param[in] b Second resource to compare * \param[in] data Sorted list of all nodes in cluster * * \return -1 if \p a should be assigned before \b, 0 if they are equal, * or +1 if \p a should be assigned after \b */ static gint cmp_resources(gconstpointer a, gconstpointer b, gpointer data) { /* GLib insists that this function require gconstpointer arguments, but we * make a small, temporary change to each argument (setting the * pe_rsc_merging flag) during comparison */ pcmk_resource_t *resource1 = (pcmk_resource_t *) a; pcmk_resource_t *resource2 = (pcmk_resource_t *) b; const GList *nodes = data; int rc = 0; int r1_score = -PCMK_SCORE_INFINITY; int r2_score = -PCMK_SCORE_INFINITY; pcmk_node_t *r1_node = NULL; pcmk_node_t *r2_node = NULL; GHashTable *r1_nodes = NULL; GHashTable *r2_nodes = NULL; const char *reason = NULL; // Resources with highest priority should be assigned first reason = "priority"; r1_score = resource1->priv->priority; r2_score = resource2->priv->priority; if (r1_score > r2_score) { rc = -1; goto done; } if (r1_score < r2_score) { rc = 1; goto done; } // We need nodes to make any other useful comparisons reason = "no node list"; if (nodes == NULL) { goto done; } // Calculate and log node scores resource1->priv->cmds->add_colocated_node_scores(resource1, NULL, resource1->id, &r1_nodes, NULL, 1, pcmk__coloc_select_this_with); resource2->priv->cmds->add_colocated_node_scores(resource2, NULL, resource2->id, &r2_nodes, NULL, 1, pcmk__coloc_select_this_with); pe__show_node_scores(true, NULL, resource1->id, r1_nodes, resource1->priv->scheduler); pe__show_node_scores(true, NULL, resource2->id, r2_nodes, resource2->priv->scheduler); // The resource with highest score on its current node goes first reason = "current location"; if (resource1->priv->active_nodes != NULL) { r1_node = pcmk__current_node(resource1); } if (resource2->priv->active_nodes != NULL) { r2_node = pcmk__current_node(resource2); } r1_score = get_node_score(r1_node, r1_nodes); r2_score = get_node_score(r2_node, r2_nodes); if (r1_score > r2_score) { rc = -1; goto done; } if (r1_score < r2_score) { rc = 1; goto done; } // Otherwise a higher score on any node will do reason = "score"; for (const GList *iter = nodes; iter != NULL; iter = iter->next) { const pcmk_node_t *node = (const pcmk_node_t *) iter->data; r1_score = get_node_score(node, r1_nodes); r2_score = get_node_score(node, r2_nodes); if (r1_score > r2_score) { rc = -1; goto done; } if (r1_score < r2_score) { rc = 1; goto done; } } done: crm_trace("%s (%d)%s%s %c %s (%d)%s%s: %s", resource1->id, r1_score, ((r1_node == NULL)? "" : " on "), ((r1_node == NULL)? "" : r1_node->priv->id), ((rc < 0)? '>' : ((rc > 0)? '<' : '=')), resource2->id, r2_score, ((r2_node == NULL)? "" : " on "), ((r2_node == NULL)? "" : r2_node->priv->id), reason); if (r1_nodes != NULL) { g_hash_table_destroy(r1_nodes); } if (r2_nodes != NULL) { g_hash_table_destroy(r2_nodes); } return rc; } /*! * \internal * \brief Sort resources in the order they should be assigned to nodes * * \param[in,out] scheduler Scheduler data */ void pcmk__sort_resources(pcmk_scheduler_t *scheduler) { GList *nodes = g_list_copy(scheduler->nodes); nodes = pcmk__sort_nodes(nodes, NULL); scheduler->priv->resources = g_list_sort_with_data(scheduler->priv->resources, cmp_resources, nodes); g_list_free(nodes); } diff --git a/lib/pacemaker/pcmk_sched_utilization.c b/lib/pacemaker/pcmk_sched_utilization.c index 4e4132b4da..a3627c7efb 100644 --- a/lib/pacemaker/pcmk_sched_utilization.c +++ b/lib/pacemaker/pcmk_sched_utilization.c @@ -1,485 +1,491 @@ /* - * Copyright 2014-2024 the Pacemaker project contributors + * Copyright 2014-2025 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 // INT_MIN, INT_MAX #include #include #include "libpacemaker_private.h" /*! * \internal * \brief Get integer utilization from a string * * \param[in] s String representation of a node utilization value * * \return Integer equivalent of \p s * \todo It would make sense to restrict utilization values to nonnegative * integers, but the documentation just says "integers" and we didn't * restrict them initially, so for backward compatibility, allow any * integer. */ static int utilization_value(const char *s) { int value = 0; if ((s != NULL) && (pcmk__scan_min_int(s, &value, INT_MIN) == EINVAL)) { pcmk__config_warn("Using 0 for utilization instead of " "invalid value '%s'", s); value = 0; } return value; } /* * Functions for comparing node capacities */ struct compare_data { const pcmk_node_t *node1; const pcmk_node_t *node2; bool node2_only; int result; }; /*! * \internal * \brief Compare a single utilization attribute for two nodes * * Compare one utilization attribute for two nodes, decrementing the result if * the first node has greater capacity, and incrementing it if the second node * has greater capacity. * * \param[in] key Utilization attribute name to compare * \param[in] value Utilization attribute value to compare * \param[in,out] user_data Comparison data (as struct compare_data*) */ static void compare_utilization_value(gpointer key, gpointer value, gpointer user_data) { int node1_capacity = 0; int node2_capacity = 0; struct compare_data *data = user_data; const char *node2_value = NULL; if (data->node2_only) { if (g_hash_table_lookup(data->node1->priv->utilization, key)) { return; // We've already compared this attribute } } else { node1_capacity = utilization_value((const char *) value); } node2_value = g_hash_table_lookup(data->node2->priv->utilization, key); node2_capacity = utilization_value(node2_value); if (node1_capacity > node2_capacity) { data->result--; } else if (node1_capacity < node2_capacity) { data->result++; } } /*! * \internal * \brief Compare utilization capacities of two nodes * * \param[in] node1 First node to compare * \param[in] node2 Second node to compare * * \return Negative integer if node1 has more free capacity, * 0 if the capacities are equal, or a positive integer * if node2 has more free capacity */ int pcmk__compare_node_capacities(const pcmk_node_t *node1, const pcmk_node_t *node2) { struct compare_data data = { .node1 = node1, .node2 = node2, .node2_only = false, .result = 0, }; // Compare utilization values that node1 and maybe node2 have g_hash_table_foreach(node1->priv->utilization, compare_utilization_value, &data); // Compare utilization values that only node2 has data.node2_only = true; g_hash_table_foreach(node2->priv->utilization, compare_utilization_value, &data); return data.result; } /* * Functions for updating node capacities */ struct calculate_data { GHashTable *current_utilization; bool plus; }; /*! * \internal * \brief Update a single utilization attribute with a new value * * \param[in] key Name of utilization attribute to update * \param[in] value Value to add or substract * \param[in,out] user_data Calculation data (as struct calculate_data *) */ static void update_utilization_value(gpointer key, gpointer value, gpointer user_data) { struct calculate_data *data = user_data; const char *current = g_hash_table_lookup(data->current_utilization, key); long long result = utilization_value(current) + (data->plus? 1LL : -1LL) * utilization_value(value); if (result < INT_MIN) { result = INT_MIN; } else if (result > INT_MAX) { result = INT_MAX; } g_hash_table_replace(data->current_utilization, strdup(key), pcmk__itoa((int) result)); } /*! * \internal * \brief Subtract a resource's utilization from node capacity * * \param[in,out] current_utilization Current node utilization attributes * \param[in] rsc Resource with utilization to subtract */ void pcmk__consume_node_capacity(GHashTable *current_utilization, const pcmk_resource_t *rsc) { struct calculate_data data = { .current_utilization = current_utilization, .plus = false, }; g_hash_table_foreach(rsc->priv->utilization, update_utilization_value, &data); } /*! * \internal * \brief Add a resource's utilization to node capacity * * \param[in,out] current_utilization Current node utilization attributes * \param[in] rsc Resource with utilization to add */ void pcmk__release_node_capacity(GHashTable *current_utilization, const pcmk_resource_t *rsc) { struct calculate_data data = { .current_utilization = current_utilization, .plus = true, }; g_hash_table_foreach(rsc->priv->utilization, update_utilization_value, &data); } /* * Functions for checking for sufficient node capacity */ struct capacity_data { const pcmk_node_t *node; const char *rsc_id; bool is_enough; }; /*! * \internal * \brief Check whether a single utilization attribute has sufficient capacity * * \param[in] key Name of utilization attribute to check * \param[in] value Amount of utilization required * \param[in,out] user_data Capacity data (as struct capacity_data *) */ static void check_capacity(gpointer key, gpointer value, gpointer user_data) { int required = 0; int remaining = 0; const char *node_value_s = NULL; struct capacity_data *data = user_data; node_value_s = g_hash_table_lookup(data->node->priv->utilization, key); required = utilization_value(value); remaining = utilization_value(node_value_s); if (required > remaining) { crm_debug("Remaining capacity for %s on %s (%d) is insufficient " "for resource %s usage (%d)", (const char *) key, pcmk__node_name(data->node), remaining, data->rsc_id, required); data->is_enough = false; } } /*! * \internal * \brief Check whether a node has sufficient capacity for a resource * * \param[in] node Node to check * \param[in] rsc_id ID of resource to check (for debug logs only) * \param[in] utilization Required utilization amounts * * \return true if node has sufficient capacity for resource, otherwise false */ static bool have_enough_capacity(const pcmk_node_t *node, const char *rsc_id, GHashTable *utilization) { struct capacity_data data = { .node = node, .rsc_id = rsc_id, .is_enough = true, }; g_hash_table_foreach(utilization, check_capacity, &data); return data.is_enough; } /*! * \internal * \brief Sum the utilization requirements of a list of resources * * \param[in] orig_rsc Resource being assigned (for logging purposes) * \param[in] rscs Resources whose utilization should be summed * * \return Newly allocated hash table with sum of all utilization values * \note It is the caller's responsibility to free the return value using * g_hash_table_destroy(). */ static GHashTable * sum_resource_utilization(const pcmk_resource_t *orig_rsc, GList *rscs) { GHashTable *utilization = pcmk__strkey_table(free, free); for (GList *iter = rscs; iter != NULL; iter = iter->next) { pcmk_resource_t *rsc = (pcmk_resource_t *) iter->data; rsc->priv->cmds->add_utilization(rsc, orig_rsc, rscs, utilization); } return utilization; } /*! * \internal * \brief Ban resource from nodes with insufficient utilization capacity * * \param[in,out] rsc Resource to check * * \return Allowed node for \p rsc with most spare capacity, if there are no * nodes with enough capacity for \p rsc and all its colocated resources */ const pcmk_node_t * pcmk__ban_insufficient_capacity(pcmk_resource_t *rsc) { bool any_capable = false; char *rscs_id = NULL; pcmk_node_t *node = NULL; const pcmk_node_t *most_capable_node = NULL; GList *colocated_rscs = NULL; GHashTable *unassigned_utilization = NULL; GHashTableIter iter; CRM_CHECK(rsc != NULL, return NULL); // The default placement strategy ignores utilization if (pcmk__str_eq(rsc->priv->scheduler->priv->placement_strategy, PCMK_VALUE_DEFAULT, pcmk__str_casei)) { return NULL; } // Check whether any resources are colocated with this one colocated_rscs = rsc->priv->cmds->colocated_resources(rsc, NULL, NULL); if (colocated_rscs == NULL) { return NULL; } rscs_id = crm_strdup_printf("%s and its colocated resources", rsc->id); // If rsc isn't in the list, add it so we include its utilization if (g_list_find(colocated_rscs, rsc) == NULL) { colocated_rscs = g_list_append(colocated_rscs, rsc); } // Sum utilization of colocated resources that haven't been assigned yet unassigned_utilization = sum_resource_utilization(rsc, colocated_rscs); // Check whether any node has enough capacity for all the resources g_hash_table_iter_init(&iter, rsc->priv->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (void **) &node)) { - if (!pcmk__node_available(node, true, false)) { + if (!pcmk__node_available(node, pcmk__node_alive + |pcmk__node_usable + |pcmk__node_no_negative)) { continue; } if (have_enough_capacity(node, rscs_id, unassigned_utilization)) { any_capable = true; } // Keep track of node with most free capacity if ((most_capable_node == NULL) || (pcmk__compare_node_capacities(node, most_capable_node) < 0)) { most_capable_node = node; } } if (any_capable) { // If so, ban resource from any node with insufficient capacity g_hash_table_iter_init(&iter, rsc->priv->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (void **) &node)) { - if (pcmk__node_available(node, true, false) + if (pcmk__node_available(node, pcmk__node_alive + |pcmk__node_usable + |pcmk__node_no_negative) && !have_enough_capacity(node, rscs_id, unassigned_utilization)) { pcmk__rsc_debug(rsc, "%s does not have enough capacity for %s", pcmk__node_name(node), rscs_id); resource_location(rsc, node, -PCMK_SCORE_INFINITY, "__limit_utilization__", rsc->priv->scheduler); } } most_capable_node = NULL; } else { // Otherwise, ban from nodes with insufficient capacity for rsc alone g_hash_table_iter_init(&iter, rsc->priv->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (void **) &node)) { - if (pcmk__node_available(node, true, false) + if (pcmk__node_available(node, pcmk__node_alive + |pcmk__node_usable + |pcmk__node_no_negative) && !have_enough_capacity(node, rsc->id, rsc->priv->utilization)) { pcmk__rsc_debug(rsc, "%s does not have enough capacity for %s", pcmk__node_name(node), rsc->id); resource_location(rsc, node, -PCMK_SCORE_INFINITY, "__limit_utilization__", rsc->priv->scheduler); } } } g_hash_table_destroy(unassigned_utilization); g_list_free(colocated_rscs); free(rscs_id); pe__show_node_scores(true, rsc, "Post-utilization", rsc->priv->allowed_nodes, rsc->priv->scheduler); return most_capable_node; } /*! * \internal * \brief Create a new load_stopped pseudo-op for a node * * \param[in,out] node Node to create op for * * \return Newly created load_stopped op */ static pcmk_action_t * new_load_stopped_op(pcmk_node_t *node) { char *load_stopped_task = crm_strdup_printf(PCMK_ACTION_LOAD_STOPPED "_%s", node->priv->name); pcmk_action_t *load_stopped = get_pseudo_op(load_stopped_task, node->priv->scheduler); if (load_stopped->node == NULL) { load_stopped->node = pe__copy_node(node); pcmk__clear_action_flags(load_stopped, pcmk__action_optional); } free(load_stopped_task); return load_stopped; } /*! * \internal * \brief Create utilization-related internal constraints for a resource * * \param[in,out] rsc Resource to create constraints for * \param[in] allowed_nodes List of allowed next nodes for \p rsc */ void pcmk__create_utilization_constraints(pcmk_resource_t *rsc, const GList *allowed_nodes) { const GList *iter = NULL; pcmk_action_t *load_stopped = NULL; pcmk__rsc_trace(rsc, "Creating utilization constraints for %s - strategy: %s", rsc->id, rsc->priv->scheduler->priv->placement_strategy); // "stop rsc then load_stopped" constraints for current nodes for (iter = rsc->priv->active_nodes; iter != NULL; iter = iter->next) { load_stopped = new_load_stopped_op(iter->data); pcmk__new_ordering(rsc, stop_key(rsc), NULL, NULL, NULL, load_stopped, pcmk__ar_if_on_same_node_or_target, rsc->priv->scheduler); } // "load_stopped then start/migrate_to rsc" constraints for allowed nodes for (iter = allowed_nodes; iter; iter = iter->next) { load_stopped = new_load_stopped_op(iter->data); pcmk__new_ordering(NULL, NULL, load_stopped, rsc, start_key(rsc), NULL, pcmk__ar_if_on_same_node_or_target, rsc->priv->scheduler); pcmk__new_ordering(NULL, NULL, load_stopped, rsc, pcmk__op_key(rsc->id, PCMK_ACTION_MIGRATE_TO, 0), NULL, pcmk__ar_if_on_same_node_or_target, rsc->priv->scheduler); } } /*! * \internal * \brief Output node capacities if enabled * * \param[in] desc Prefix for output * \param[in,out] scheduler Scheduler data */ void pcmk__show_node_capacities(const char *desc, pcmk_scheduler_t *scheduler) { if (!pcmk_is_set(scheduler->flags, pcmk__sched_show_utilization)) { return; } for (const GList *iter = scheduler->nodes; iter != NULL; iter = iter->next) { const pcmk_node_t *node = (const pcmk_node_t *) iter->data; pcmk__output_t *out = scheduler->priv->out; // Utilization doesn't apply to bundle nodes if (pcmk__is_bundle_node(node)) { continue; } out->message(out, "node-capacity", node, desc); } }