diff --git a/lib/pacemaker/libpacemaker_private.h b/lib/pacemaker/libpacemaker_private.h index b677a7fb85..8358519b91 100644 --- a/lib/pacemaker/libpacemaker_private.h +++ b/lib/pacemaker/libpacemaker_private.h @@ -1,956 +1,977 @@ /* * Copyright 2021-2023 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU Lesser General Public License * version 2.1 or later (LGPLv2.1+) WITHOUT ANY WARRANTY. */ #ifndef PCMK__LIBPACEMAKER_PRIVATE__H # define PCMK__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 // pe_action_t, pe_node_t, pe_working_set_t -// Flags to modify the behavior of pcmk__add_colocated_node_scores() +// 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 allocation methods struct resource_alloc_functions_s { /*! * \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 * * \return Node that \p rsc is assigned to, if assigned entirely to one node */ pe_node_t *(*assign)(pe_resource_t *rsc, const pe_node_t *prefer); /*! * \internal * \brief Create all actions needed for a given resource * * \param[in,out] rsc Resource to create actions for */ void (*create_actions)(pe_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)(pe_resource_t *rsc, pe_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)(pe_resource_t *rsc); /*! * \internal * \brief Apply a colocation's score to node weights or resource priority * * Given a colocation constraint, apply its score to the dependent's * allowed node weights (if we are still placing resources) or priority (if * we are choosing promotable clone instance roles). * * \param[in,out] dependent Dependent resource in colocation * \param[in] primary Primary resource in colocation * \param[in] colocation Colocation constraint to apply * \param[in] for_dependent true if called on behalf of dependent */ void (*apply_coloc_score) (pe_resource_t *dependent, const pe_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 indirectly 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 pe_resource_t *rsc, const pe_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 pe_resource_t *rsc, const pe_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 pe_resource_t *rsc, const pe_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] rsc Resource to check colocations for + * \param[in] log_id Resource ID to use in logs (if NULL, use \p rsc ID) + * \param[in,out] nodes Nodes to update + * \param[in] attr Colocation attribute (NULL to use default) + * \param[in] factor Incorporate scores multiplied by this factor + * \param[in] flags Bitmask of enum pcmk__coloc_select values + * + * \note The caller remains responsible for freeing \p *nodes. + */ + void (*add_colocated_node_scores)(pe_resource_t *rsc, const char *log_id, + GHashTable **nodes, const char *attr, + 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)(pe_resource_t *rsc, pe__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. */ enum pe_action_flags (*action_flags)(pe_action_t *action, const pe_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 pe_action_optional to affect only * mandatory actions, and pe_action_runnable to * affect only runnable actions) * \param[in] type Group of enum pe_ordering flags to apply * \param[in,out] data_set Cluster working set * * \return Group of enum pcmk__updated flags indicating what was updated */ uint32_t (*update_ordered_actions)(pe_action_t *first, pe_action_t *then, const pe_node_t *node, uint32_t flags, uint32_t filter, uint32_t type, pe_working_set_t *data_set); void (*output_actions)(pe_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)(pe_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 pe_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 allocated to a node. * * \param[in] rsc Resource with utilization to add * \param[in] orig_rsc Resource being allocated (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 pe_resource_t *rsc, const pe_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)(pe_resource_t *rsc); }; // Actions (pcmk_sched_actions.c) G_GNUC_INTERNAL void pcmk__update_action_for_orderings(pe_action_t *action, pe_working_set_t *data_set); G_GNUC_INTERNAL uint32_t pcmk__update_ordered_actions(pe_action_t *first, pe_action_t *then, const pe_node_t *node, uint32_t flags, uint32_t filter, uint32_t type, pe_working_set_t *data_set); G_GNUC_INTERNAL void pcmk__log_action(const char *pre_text, const pe_action_t *action, bool details); G_GNUC_INTERNAL pe_action_t *pcmk__new_cancel_action(pe_resource_t *rsc, const char *name, guint interval_ms, const pe_node_t *node); G_GNUC_INTERNAL pe_action_t *pcmk__new_shutdown_action(pe_node_t *node); G_GNUC_INTERNAL bool pcmk__action_locks_rsc_to_node(const pe_action_t *action); G_GNUC_INTERNAL void pcmk__deduplicate_action_inputs(pe_action_t *action); G_GNUC_INTERNAL void pcmk__output_actions(pe_working_set_t *data_set); G_GNUC_INTERNAL bool pcmk__check_action_config(pe_resource_t *rsc, pe_node_t *node, const xmlNode *xml_op); G_GNUC_INTERNAL void pcmk__handle_rsc_config_changes(pe_working_set_t *data_set); // Recurring actions (pcmk_sched_recurring.c) G_GNUC_INTERNAL void pcmk__create_recurring_actions(pe_resource_t *rsc); G_GNUC_INTERNAL void pcmk__schedule_cancel(pe_resource_t *rsc, const char *call_id, const char *task, guint interval_ms, const pe_node_t *node, const char *reason); G_GNUC_INTERNAL void pcmk__reschedule_recurring(pe_resource_t *rsc, const char *task, guint interval_ms, pe_node_t *node); G_GNUC_INTERNAL bool pcmk__action_is_recurring(const pe_action_t *action); // Producing transition graphs (pcmk_graph_producer.c) G_GNUC_INTERNAL bool pcmk__graph_has_loop(const pe_action_t *init_action, const pe_action_t *action, pe_action_wrapper_t *input); G_GNUC_INTERNAL void pcmk__add_rsc_actions_to_graph(pe_resource_t *rsc); G_GNUC_INTERNAL void pcmk__create_graph(pe_working_set_t *data_set); // Fencing (pcmk_sched_fencing.c) G_GNUC_INTERNAL void pcmk__order_vs_fence(pe_action_t *stonith_op, pe_working_set_t *data_set); G_GNUC_INTERNAL void pcmk__order_vs_unfence(const pe_resource_t *rsc, pe_node_t *node, pe_action_t *action, enum pe_ordering order); G_GNUC_INTERNAL void pcmk__fence_guest(pe_node_t *node); G_GNUC_INTERNAL bool pcmk__node_unfenced(const pe_node_t *node); G_GNUC_INTERNAL void pcmk__order_restart_vs_unfence(gpointer data, gpointer user_data); // Injected scheduler inputs (pcmk_sched_injections.c) void pcmk__inject_scheduler_input(pe_working_set_t *data_set, cib_t *cib, const pcmk_injections_t *injections); // Constraints of any type (pcmk_sched_constraints.c) G_GNUC_INTERNAL pe_resource_t *pcmk__find_constraint_resource(GList *rsc_list, const char *id); G_GNUC_INTERNAL xmlNode *pcmk__expand_tags_in_sets(xmlNode *xml_obj, const pe_working_set_t *data_set); G_GNUC_INTERNAL bool pcmk__valid_resource_or_tag(const pe_working_set_t *data_set, const char *id, pe_resource_t **rsc, pe_tag_t **tag); G_GNUC_INTERNAL bool pcmk__tag_to_set(xmlNode *xml_obj, xmlNode **rsc_set, const char *attr, bool convert_rsc, const pe_working_set_t *data_set); G_GNUC_INTERNAL void pcmk__create_internal_constraints(pe_working_set_t *data_set); // Location constraints G_GNUC_INTERNAL void pcmk__unpack_location(xmlNode *xml_obj, pe_working_set_t *data_set); G_GNUC_INTERNAL pe__location_t *pcmk__new_location(const char *id, pe_resource_t *rsc, int node_weight, const char *discover_mode, pe_node_t *foo_node, pe_working_set_t *data_set); G_GNUC_INTERNAL void pcmk__apply_locations(pe_working_set_t *data_set); G_GNUC_INTERNAL void pcmk__apply_location(pe_resource_t *rsc, pe__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 enum pcmk__coloc_affects pcmk__colocation_affects(const pe_resource_t *dependent, const pe_resource_t *primary, const pcmk__colocation_t *colocation, bool preview); G_GNUC_INTERNAL void pcmk__apply_coloc_to_weights(pe_resource_t *dependent, const pe_resource_t *primary, const pcmk__colocation_t *colocation); G_GNUC_INTERNAL void pcmk__apply_coloc_to_priority(pe_resource_t *dependent, const pe_resource_t *primary, const pcmk__colocation_t *colocation); G_GNUC_INTERNAL void pcmk__add_colocated_node_scores(pe_resource_t *rsc, const char *log_id, GHashTable **nodes, const char *attr, float factor, uint32_t flags); G_GNUC_INTERNAL void pcmk__unpack_colocation(xmlNode *xml_obj, pe_working_set_t *data_set); G_GNUC_INTERNAL void pcmk__add_this_with(GList **list, const pcmk__colocation_t *colocation); G_GNUC_INTERNAL void pcmk__add_this_with_list(GList **list, GList *addition); G_GNUC_INTERNAL void pcmk__add_with_this(GList **list, const pcmk__colocation_t *colocation); G_GNUC_INTERNAL void pcmk__add_with_this_list(GList **list, GList *addition); G_GNUC_INTERNAL void pcmk__new_colocation(const char *id, const char *node_attr, int score, pe_resource_t *dependent, pe_resource_t *primary, const char *dependent_role, const char *primary_role, bool influence, pe_working_set_t *data_set); G_GNUC_INTERNAL void pcmk__block_colocation_dependents(pe_action_t *action, pe_working_set_t *data_set); /*! * \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 */ static inline bool pcmk__colocation_has_influence(const pcmk__colocation_t *colocation, const pe_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, pe_rsc_allow_remote_remotes) && !pcmk_is_set(rsc->flags, pe_rsc_failed) && pcmk__list_of_1(rsc->running_on)) { return false; } /* The dependent in a colocation influences the primary's location * if the influence option is true or the primary is not yet active. */ return colocation->influence || (rsc->running_on == NULL); } // Ordering constraints (pcmk_sched_ordering.c) G_GNUC_INTERNAL void pcmk__new_ordering(pe_resource_t *first_rsc, char *first_task, pe_action_t *first_action, pe_resource_t *then_rsc, char *then_task, pe_action_t *then_action, uint32_t flags, pe_working_set_t *data_set); G_GNUC_INTERNAL void pcmk__unpack_ordering(xmlNode *xml_obj, pe_working_set_t *data_set); G_GNUC_INTERNAL void pcmk__disable_invalid_orderings(pe_working_set_t *data_set); G_GNUC_INTERNAL void pcmk__order_stops_before_shutdown(pe_node_t *node, pe_action_t *shutdown_op); G_GNUC_INTERNAL void pcmk__apply_orderings(pe_working_set_t *data_set); G_GNUC_INTERNAL void pcmk__order_after_each(pe_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 Bitmask of enum pe_ordering 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)->cluster) #define pcmk__order_starts(rsc1, rsc2, flags) \ pcmk__order_resource_actions((rsc1), CRMD_ACTION_START, \ (rsc2), CRMD_ACTION_START, (flags)) #define pcmk__order_stops(rsc1, rsc2, flags) \ pcmk__order_resource_actions((rsc1), CRMD_ACTION_STOP, \ (rsc2), CRMD_ACTION_STOP, (flags)) // Ticket constraints (pcmk_sched_tickets.c) G_GNUC_INTERNAL void pcmk__unpack_rsc_ticket(xmlNode *xml_obj, pe_working_set_t *data_set); // Promotable clone resources (pcmk_sched_promotable.c) G_GNUC_INTERNAL void pcmk__add_promotion_scores(pe_resource_t *rsc); G_GNUC_INTERNAL void pcmk__require_promotion_tickets(pe_resource_t *rsc); G_GNUC_INTERNAL void pcmk__set_instance_roles(pe_resource_t *rsc); G_GNUC_INTERNAL void pcmk__create_promotable_actions(pe_resource_t *clone); G_GNUC_INTERNAL void pcmk__promotable_restart_ordering(pe_resource_t *rsc); G_GNUC_INTERNAL void pcmk__order_promotable_instances(pe_resource_t *clone); G_GNUC_INTERNAL void pcmk__update_dependent_with_promotable(const pe_resource_t *primary, pe_resource_t *dependent, const pcmk__colocation_t *colocation); G_GNUC_INTERNAL void pcmk__update_promotable_dependent_priority(const pe_resource_t *primary, pe_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 pe_node_t *node); G_GNUC_INTERNAL void pcmk__order_remote_connection_actions(pe_working_set_t *data_set); G_GNUC_INTERNAL bool pcmk__rsc_corresponds_to_guest(const pe_resource_t *rsc, const pe_node_t *node); G_GNUC_INTERNAL pe_node_t *pcmk__connection_host_for_action(const pe_action_t *action); G_GNUC_INTERNAL void pcmk__substitute_remote_addr(pe_resource_t *rsc, GHashTable *params); G_GNUC_INTERNAL void pcmk__add_bundle_meta_to_xml(xmlNode *args_xml, const pe_action_t *action); // Primitives (pcmk_sched_primitive.c) G_GNUC_INTERNAL pe_node_t *pcmk__primitive_assign(pe_resource_t *rsc, const pe_node_t *prefer); G_GNUC_INTERNAL void pcmk__primitive_create_actions(pe_resource_t *rsc); G_GNUC_INTERNAL void pcmk__primitive_internal_constraints(pe_resource_t *rsc); G_GNUC_INTERNAL enum pe_action_flags pcmk__primitive_action_flags(pe_action_t *action, const pe_node_t *node); G_GNUC_INTERNAL void pcmk__primitive_apply_coloc_score(pe_resource_t *dependent, const pe_resource_t *primary, const pcmk__colocation_t *colocation, bool for_dependent); G_GNUC_INTERNAL void pcmk__with_primitive_colocations(const pe_resource_t *rsc, const pe_resource_t *orig_rsc, GList **list); G_GNUC_INTERNAL void pcmk__primitive_with_colocations(const pe_resource_t *rsc, const pe_resource_t *orig_rsc, GList **list); G_GNUC_INTERNAL void pcmk__schedule_cleanup(pe_resource_t *rsc, const pe_node_t *node, bool optional); G_GNUC_INTERNAL void pcmk__primitive_add_graph_meta(const pe_resource_t *rsc, xmlNode *xml); G_GNUC_INTERNAL void pcmk__primitive_add_utilization(const pe_resource_t *rsc, const pe_resource_t *orig_rsc, GList *all_rscs, GHashTable *utilization); G_GNUC_INTERNAL void pcmk__primitive_shutdown_lock(pe_resource_t *rsc); // Groups (pcmk_sched_group.c) G_GNUC_INTERNAL pe_node_t *pcmk__group_assign(pe_resource_t *rsc, const pe_node_t *prefer); G_GNUC_INTERNAL void pcmk__group_create_actions(pe_resource_t *rsc); G_GNUC_INTERNAL void pcmk__group_internal_constraints(pe_resource_t *rsc); G_GNUC_INTERNAL void pcmk__group_apply_coloc_score(pe_resource_t *dependent, const pe_resource_t *primary, const pcmk__colocation_t *colocation, bool for_dependent); G_GNUC_INTERNAL void pcmk__with_group_colocations(const pe_resource_t *rsc, const pe_resource_t *orig_rsc, GList **list); G_GNUC_INTERNAL void pcmk__group_with_colocations(const pe_resource_t *rsc, const pe_resource_t *orig_rsc, GList **list); G_GNUC_INTERNAL void pcmk__group_apply_location(pe_resource_t *rsc, pe__location_t *location); G_GNUC_INTERNAL enum pe_action_flags pcmk__group_action_flags(pe_action_t *action, const pe_node_t *node); G_GNUC_INTERNAL uint32_t pcmk__group_update_ordered_actions(pe_action_t *first, pe_action_t *then, const pe_node_t *node, uint32_t flags, uint32_t filter, uint32_t type, pe_working_set_t *data_set); G_GNUC_INTERNAL GList *pcmk__group_colocated_resources(const pe_resource_t *rsc, const pe_resource_t *orig_rsc, GList *colocated_rscs); G_GNUC_INTERNAL void pcmk__group_add_utilization(const pe_resource_t *rsc, const pe_resource_t *orig_rsc, GList *all_rscs, GHashTable *utilization); G_GNUC_INTERNAL void pcmk__group_shutdown_lock(pe_resource_t *rsc); // Clones (pcmk_sched_clone.c) G_GNUC_INTERNAL pe_node_t *pcmk__clone_assign(pe_resource_t *rsc, const pe_node_t *prefer); G_GNUC_INTERNAL void pcmk__clone_apply_coloc_score(pe_resource_t *dependent, const pe_resource_t *primary, const pcmk__colocation_t *colocation, bool for_dependent); G_GNUC_INTERNAL void pcmk__with_clone_colocations(const pe_resource_t *rsc, const pe_resource_t *orig_rsc, GList **list); G_GNUC_INTERNAL void pcmk__clone_with_colocations(const pe_resource_t *rsc, const pe_resource_t *orig_rsc, GList **list); // Bundles (pcmk_sched_bundle.c) G_GNUC_INTERNAL const pe_resource_t *pcmk__get_rsc_in_container(const pe_resource_t *instance); G_GNUC_INTERNAL void pcmk__bundle_apply_coloc_score(pe_resource_t *dependent, const pe_resource_t *primary, const pcmk__colocation_t *colocation, bool for_dependent); G_GNUC_INTERNAL void pcmk__with_bundle_colocations(const pe_resource_t *rsc, const pe_resource_t *orig_rsc, GList **list); G_GNUC_INTERNAL void pcmk__bundle_with_colocations(const pe_resource_t *rsc, const pe_resource_t *orig_rsc, GList **list); G_GNUC_INTERNAL void pcmk__output_bundle_actions(pe_resource_t *rsc); // Clone instances or bundle replica containers (pcmk_sched_instances.c) G_GNUC_INTERNAL void pcmk__assign_instances(pe_resource_t *collective, GList *instances, int max_total, int max_per_node); G_GNUC_INTERNAL void pcmk__create_instance_actions(pe_resource_t *rsc, GList *instances); G_GNUC_INTERNAL bool pcmk__instance_matches(const pe_resource_t *instance, const pe_node_t *node, enum rsc_role_e role, bool current); G_GNUC_INTERNAL pe_resource_t *pcmk__find_compatible_instance(const pe_resource_t *match_rsc, const pe_resource_t *rsc, enum rsc_role_e role, bool current); G_GNUC_INTERNAL uint32_t pcmk__instance_update_ordered_actions(pe_action_t *first, pe_action_t *then, const pe_node_t *node, uint32_t flags, uint32_t filter, uint32_t type, pe_working_set_t *data_set); G_GNUC_INTERNAL enum pe_action_flags pcmk__collective_action_flags(pe_action_t *action, const GList *instances, const pe_node_t *node); G_GNUC_INTERNAL void pcmk__add_collective_constraints(GList **list, const pe_resource_t *instance, const pe_resource_t *collective, bool with_this); // 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, xmlNode *cib_node, const char *resource, const char *task, guint interval_ms, int rc); G_GNUC_INTERNAL xmlNode *pcmk__inject_action_result(xmlNode *cib_resource, lrmd_event_data_t *op, int target_rc); // Nodes (pcmk_sched_nodes.c) G_GNUC_INTERNAL bool pcmk__node_available(const pe_node_t *node, bool consider_score, bool consider_guest); G_GNUC_INTERNAL bool pcmk__any_node_available(GHashTable *nodes); G_GNUC_INTERNAL GHashTable *pcmk__copy_node_table(GHashTable *nodes); G_GNUC_INTERNAL GList *pcmk__sort_nodes(GList *nodes, pe_node_t *active_node); G_GNUC_INTERNAL void pcmk__apply_node_health(pe_working_set_t *data_set); G_GNUC_INTERNAL pe_node_t *pcmk__top_allowed_node(const pe_resource_t *rsc, const pe_node_t *node); // Functions applying to more than one variant (pcmk_sched_resource.c) G_GNUC_INTERNAL void pcmk__set_allocation_methods(pe_working_set_t *data_set); G_GNUC_INTERNAL bool pcmk__rsc_agent_changed(pe_resource_t *rsc, pe_node_t *node, const xmlNode *rsc_entry, bool active_on_node); G_GNUC_INTERNAL GList *pcmk__rscs_matching_id(const char *id, const pe_working_set_t *data_set); G_GNUC_INTERNAL GList *pcmk__colocated_resources(const pe_resource_t *rsc, const pe_resource_t *orig_rsc, GList *colocated_rscs); G_GNUC_INTERNAL void pcmk__noop_add_graph_meta(const pe_resource_t *rsc, xmlNode *xml); G_GNUC_INTERNAL void pcmk__output_resource_actions(pe_resource_t *rsc); G_GNUC_INTERNAL bool pcmk__finalize_assignment(pe_resource_t *rsc, pe_node_t *chosen, bool force); G_GNUC_INTERNAL bool pcmk__assign_resource(pe_resource_t *rsc, pe_node_t *node, bool force); G_GNUC_INTERNAL void pcmk__unassign_resource(pe_resource_t *rsc); G_GNUC_INTERNAL bool pcmk__threshold_reached(pe_resource_t *rsc, const pe_node_t *node, pe_resource_t **failed); G_GNUC_INTERNAL void pcmk__sort_resources(pe_working_set_t *data_set); 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(pe_resource_t *rsc, pe_node_t *node); G_GNUC_INTERNAL void pcmk__order_probes(pe_working_set_t *data_set); G_GNUC_INTERNAL bool pcmk__probe_resource_list(GList *rscs, pe_node_t *node); G_GNUC_INTERNAL void pcmk__schedule_probes(pe_working_set_t *data_set); // Functions related to live migration (pcmk_sched_migration.c) void pcmk__create_migration_actions(pe_resource_t *rsc, const pe_node_t *current); void pcmk__abort_dangling_migration(void *data, void *user_data); bool pcmk__rsc_can_migrate(const pe_resource_t *rsc, const pe_node_t *current); void pcmk__order_migration_equivalents(pe__ordering_t *order); // Functions related to node utilization (pcmk_sched_utilization.c) G_GNUC_INTERNAL int pcmk__compare_node_capacities(const pe_node_t *node1, const pe_node_t *node2); G_GNUC_INTERNAL void pcmk__consume_node_capacity(GHashTable *current_utilization, const pe_resource_t *rsc); G_GNUC_INTERNAL void pcmk__release_node_capacity(GHashTable *current_utilization, const pe_resource_t *rsc); G_GNUC_INTERNAL const pe_node_t *pcmk__ban_insufficient_capacity(pe_resource_t *rsc); G_GNUC_INTERNAL void pcmk__create_utilization_constraints(pe_resource_t *rsc, const GList *allowed_nodes); G_GNUC_INTERNAL void pcmk__show_node_capacities(const char *desc, pe_working_set_t *data_set); #endif // PCMK__LIBPACEMAKER_PRIVATE__H diff --git a/lib/pacemaker/pcmk_sched_clone.c b/lib/pacemaker/pcmk_sched_clone.c index b529bb43d4..cea66142bf 100644 --- a/lib/pacemaker/pcmk_sched_clone.c +++ b/lib/pacemaker/pcmk_sched_clone.c @@ -1,666 +1,669 @@ /* * Copyright 2004-2023 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU General Public License version 2 * or later (GPLv2+) WITHOUT ANY WARRANTY. */ #include #include #include #include "libpacemaker_private.h" /*! * \internal * \brief Consider a colocation dependent's node scores * * \param[in,out] data Colocation constraint to consider * \param[in,out] user_data Clone resource that is colocation primary */ static void add_dependent_scores(gpointer data, gpointer user_data) { pcmk__colocation_t *colocation = (pcmk__colocation_t *) data; pe_resource_t *clone = (pe_resource_t *) user_data; if (pcmk__colocation_has_influence(colocation, NULL)) { - pcmk__add_colocated_node_scores(colocation->dependent, clone->id, - &clone->allowed_nodes, - colocation->node_attribute, - colocation->score / (float) INFINITY, - pcmk__coloc_select_active - |pcmk__coloc_select_nonnegative); + pe_resource_t *rsc = colocation->dependent; + const float factor = colocation->score / (float) INFINITY; + + rsc->cmds->add_colocated_node_scores(rsc, clone->id, + &clone->allowed_nodes, + colocation->node_attribute, + factor, + pcmk__coloc_select_active + |pcmk__coloc_select_nonnegative); } } /*! * \internal * \brief Assign a clone resource's instances to nodes * * \param[in,out] rsc Clone resource to assign * \param[in] prefer Node to prefer, if all else is equal * * \return NULL (clones are not assigned to a single node) */ pe_node_t * pcmk__clone_assign(pe_resource_t *rsc, const pe_node_t *prefer) { CRM_ASSERT(pe_rsc_is_clone(rsc)); if (!pcmk_is_set(rsc->flags, pe_rsc_provisional)) { return NULL; // Assignment has already been done } // Detect assignment loops if (pcmk_is_set(rsc->flags, pe_rsc_allocating)) { pe_rsc_debug(rsc, "Breaking assignment loop involving %s", rsc->id); return NULL; } pe__set_resource_flags(rsc, pe_rsc_allocating); // If this clone is promotable, consider nodes' promotion scores if (pcmk_is_set(rsc->flags, pe_rsc_promotable)) { pcmk__add_promotion_scores(rsc); } /* If this clone is colocated with any other resources, assign those first. * Since the this_with_colocations() method boils down to a copy of rsc_cons * for clones, we can use that here directly for efficiency. */ for (GList *iter = rsc->rsc_cons; iter != NULL; iter = iter->next) { pcmk__colocation_t *constraint = (pcmk__colocation_t *) iter->data; pe_rsc_trace(rsc, "%s: Assigning colocation %s primary %s first", rsc->id, constraint->id, constraint->primary->id); constraint->primary->cmds->assign(constraint->primary, prefer); } /* If any resources are colocated with this one, consider their preferences. * Because the with_this_colocations() method boils down to a copy of * rsc_cons_lhs for clones, we can use that here directly for efficiency. */ g_list_foreach(rsc->rsc_cons_lhs, add_dependent_scores, rsc); pe__show_node_weights(!pcmk_is_set(rsc->cluster->flags, pe_flag_show_scores), rsc, __func__, rsc->allowed_nodes, rsc->cluster); rsc->children = g_list_sort(rsc->children, pcmk__cmp_instance); pcmk__assign_instances(rsc, rsc->children, pe__clone_max(rsc), pe__clone_node_max(rsc)); if (pcmk_is_set(rsc->flags, pe_rsc_promotable)) { pcmk__set_instance_roles(rsc); } pe__clear_resource_flags(rsc, pe_rsc_provisional|pe_rsc_allocating); pe_rsc_trace(rsc, "Assigned clone %s", rsc->id); return NULL; } static pe_action_t * find_rsc_action(pe_resource_t *rsc, const char *task) { pe_action_t *match = NULL; GList *actions = pe__resource_actions(rsc, NULL, task, FALSE); for (GList *item = actions; item != NULL; item = item->next) { pe_action_t *op = (pe_action_t *) item->data; if (!pcmk_is_set(op->flags, pe_action_optional)) { if (match != NULL) { // More than one match, don't return any match = NULL; break; } match = op; } } g_list_free(actions); return match; } /*! * \internal * \brief Order starts and stops of an ordered clone's instances * * \param[in,out] rsc Clone resource */ static void order_instance_starts_stops(pe_resource_t *rsc) { pe_action_t *last_stop = NULL; pe_action_t *last_start = NULL; // Instances must be ordered by ascending instance number, so sort them rsc->children = g_list_sort(rsc->children, pcmk__cmp_instance_number); for (GList *iter = rsc->children; iter != NULL; iter = iter->next) { pe_resource_t *child = (pe_resource_t *) iter->data; pe_action_t *action = NULL; // Order this instance's stop after previous instance's stop // @TODO: Should instances be stopped in reverse order instead? action = find_rsc_action(child, RSC_STOP); if (action != NULL) { if (last_stop != NULL) { order_actions(action, last_stop, pe_order_optional); } last_stop = action; } // Order this instance's start after previous instance's start action = find_rsc_action(child, RSC_START); if (action != NULL) { if (last_start != NULL) { order_actions(last_start, action, pe_order_optional); } last_start = action; } } } void clone_create_actions(pe_resource_t *rsc) { pe_rsc_debug(rsc, "Creating actions for clone %s", rsc->id); pcmk__create_instance_actions(rsc, rsc->children); if (pe__clone_is_ordered(rsc)) { order_instance_starts_stops(rsc); } if (pcmk_is_set(rsc->flags, pe_rsc_promotable)) { pcmk__create_promotable_actions(rsc); } } void clone_internal_constraints(pe_resource_t *rsc) { pe_resource_t *last_rsc = NULL; GList *gIter; bool ordered = pe__clone_is_ordered(rsc); pe_rsc_trace(rsc, "Internal constraints for %s", rsc->id); pcmk__order_resource_actions(rsc, RSC_STOPPED, rsc, RSC_START, pe_order_optional); pcmk__order_resource_actions(rsc, RSC_START, rsc, RSC_STARTED, pe_order_runnable_left); pcmk__order_resource_actions(rsc, RSC_STOP, rsc, RSC_STOPPED, pe_order_runnable_left); if (pcmk_is_set(rsc->flags, pe_rsc_promotable)) { pcmk__order_resource_actions(rsc, RSC_DEMOTED, rsc, RSC_STOP, pe_order_optional); pcmk__order_resource_actions(rsc, RSC_STARTED, rsc, RSC_PROMOTE, pe_order_runnable_left); } if (ordered) { /* we have to maintain a consistent sorted child list when building order constraints */ rsc->children = g_list_sort(rsc->children, pcmk__cmp_instance_number); } for (gIter = rsc->children; gIter != NULL; gIter = gIter->next) { pe_resource_t *child_rsc = (pe_resource_t *) gIter->data; child_rsc->cmds->internal_constraints(child_rsc); pcmk__order_starts(rsc, child_rsc, pe_order_runnable_left|pe_order_implies_first_printed); pcmk__order_resource_actions(child_rsc, RSC_START, rsc, RSC_STARTED, pe_order_implies_then_printed); if (ordered && (last_rsc != NULL)) { pcmk__order_starts(last_rsc, child_rsc, pe_order_optional); } pcmk__order_stops(rsc, child_rsc, pe_order_implies_first_printed); pcmk__order_resource_actions(child_rsc, RSC_STOP, rsc, RSC_STOPPED, pe_order_implies_then_printed); if (ordered && (last_rsc != NULL)) { pcmk__order_stops(child_rsc, last_rsc, pe_order_optional); } last_rsc = child_rsc; } if (pcmk_is_set(rsc->flags, pe_rsc_promotable)) { pcmk__order_promotable_instances(rsc); } } /*! * \internal * \brief Apply a colocation's score to node weights or resource priority * * Given a colocation constraint, apply its score to the dependent's * allowed node weights (if we are still placing resources) or priority (if * we are choosing promotable clone instance roles). * * \param[in,out] dependent Dependent resource in colocation * \param[in] primary Primary resource in colocation * \param[in] colocation Colocation constraint to apply * \param[in] for_dependent true if called on behalf of dependent */ void pcmk__clone_apply_coloc_score(pe_resource_t *dependent, const pe_resource_t *primary, const pcmk__colocation_t *colocation, bool for_dependent) { GList *gIter = NULL; gboolean do_interleave = FALSE; const char *interleave_s = NULL; /* This should never be called for the clone itself as a dependent. Instead, * we add its colocation constraints to its instances and call the * apply_coloc_score() for the instances as dependents. */ CRM_ASSERT(!for_dependent); CRM_CHECK((colocation != NULL) && (dependent != NULL) && (primary != NULL), return); CRM_CHECK(dependent->variant == pe_native, return); pe_rsc_trace(primary, "Processing constraint %s: %s -> %s %d", colocation->id, dependent->id, primary->id, colocation->score); if (pcmk_is_set(primary->flags, pe_rsc_promotable)) { if (pcmk_is_set(primary->flags, pe_rsc_provisional)) { // We haven't placed the primary yet, so we can't apply colocation pe_rsc_trace(primary, "%s is still provisional", primary->id); return; } else if (colocation->primary_role == RSC_ROLE_UNKNOWN) { // This isn't a role-specfic colocation, so handle normally pe_rsc_trace(primary, "Handling %s as a clone colocation", colocation->id); } else if (pcmk_is_set(dependent->flags, pe_rsc_provisional)) { // We're placing the dependent pcmk__update_dependent_with_promotable(primary, dependent, colocation); return; } else if (colocation->dependent_role == RSC_ROLE_PROMOTED) { // We're choosing roles for the dependent pcmk__update_promotable_dependent_priority(primary, dependent, colocation); return; } } // Only the dependent needs to be marked for interleave interleave_s = g_hash_table_lookup(colocation->dependent->meta, XML_RSC_ATTR_INTERLEAVE); if (crm_is_true(interleave_s) && (colocation->dependent->variant > pe_group)) { /* @TODO Do we actually care about multiple primary copies sharing a * dependent copy anymore? */ if (copies_per_node(colocation->dependent) != copies_per_node(colocation->primary)) { pcmk__config_err("Cannot interleave %s and %s because they do not " "support the same number of instances per node", colocation->dependent->id, colocation->primary->id); } else { do_interleave = TRUE; } } if (pcmk_is_set(primary->flags, pe_rsc_provisional)) { pe_rsc_trace(primary, "%s is still provisional", primary->id); return; } else if (do_interleave) { pe_resource_t *primary_instance = NULL; primary_instance = pcmk__find_compatible_instance(dependent, primary, RSC_ROLE_UNKNOWN, false); if (primary_instance != NULL) { pe_rsc_debug(primary, "Pairing %s with %s", dependent->id, primary_instance->id); dependent->cmds->apply_coloc_score(dependent, primary_instance, colocation, true); } else if (colocation->score >= INFINITY) { crm_notice("Cannot pair %s with instance of %s", dependent->id, primary->id); pcmk__assign_resource(dependent, NULL, true); } else { pe_rsc_debug(primary, "Cannot pair %s with instance of %s", dependent->id, primary->id); } return; } else if (colocation->score >= INFINITY) { GList *affected_nodes = NULL; gIter = primary->children; for (; gIter != NULL; gIter = gIter->next) { pe_resource_t *child_rsc = (pe_resource_t *) gIter->data; pe_node_t *chosen = child_rsc->fns->location(child_rsc, NULL, FALSE); if (chosen != NULL && is_set_recursive(child_rsc, pe_rsc_block, TRUE) == FALSE) { pe_rsc_trace(primary, "Allowing %s: %s %d", colocation->id, pe__node_name(chosen), chosen->weight); affected_nodes = g_list_prepend(affected_nodes, chosen); } } node_list_exclude(dependent->allowed_nodes, affected_nodes, FALSE); g_list_free(affected_nodes); return; } gIter = primary->children; for (; gIter != NULL; gIter = gIter->next) { pe_resource_t *child_rsc = (pe_resource_t *) gIter->data; child_rsc->cmds->apply_coloc_score(dependent, child_rsc, colocation, false); } } // Clone implementation of resource_alloc_functions_t:with_this_colocations() void pcmk__with_clone_colocations(const pe_resource_t *rsc, const pe_resource_t *orig_rsc, GList **list) { CRM_CHECK((rsc != NULL) && (orig_rsc != NULL) && (list != NULL), return); if (rsc == orig_rsc) { // Colocations are wanted for clone itself pcmk__add_with_this_list(list, rsc->rsc_cons_lhs); } else { pcmk__add_collective_constraints(list, orig_rsc, rsc, true); } } // Clone implementation of resource_alloc_functions_t:this_with_colocations() void pcmk__clone_with_colocations(const pe_resource_t *rsc, const pe_resource_t *orig_rsc, GList **list) { CRM_CHECK((rsc != NULL) && (orig_rsc != NULL) && (list != NULL), return); if (rsc == orig_rsc) { // Colocations are wanted for clone itself pcmk__add_this_with_list(list, rsc->rsc_cons); } else { pcmk__add_collective_constraints(list, orig_rsc, rsc, false); } } enum pe_action_flags clone_action_flags(pe_action_t *action, const pe_node_t *node) { return pcmk__collective_action_flags(action, action->rsc->children, node); } void clone_rsc_location(pe_resource_t *rsc, pe__location_t *constraint) { GList *gIter = rsc->children; pe_rsc_trace(rsc, "Processing location constraint %s for %s", constraint->id, rsc->id); pcmk__apply_location(rsc, constraint); for (; gIter != NULL; gIter = gIter->next) { pe_resource_t *child_rsc = (pe_resource_t *) gIter->data; child_rsc->cmds->apply_location(child_rsc, constraint); } } /*! * \internal * \brief Add a resource's actions to the transition graph * * \param[in,out] rsc Resource whose actions should be added */ void clone_expand(pe_resource_t *rsc) { GList *gIter = NULL; g_list_foreach(rsc->actions, (GFunc) rsc->cmds->action_flags, NULL); pe__create_clone_notifications(rsc); /* Now that the notifcations have been created we can expand the children */ gIter = rsc->children; for (; gIter != NULL; gIter = gIter->next) { pe_resource_t *child_rsc = (pe_resource_t *) gIter->data; child_rsc->cmds->add_actions_to_graph(child_rsc); } pcmk__add_rsc_actions_to_graph(rsc); /* The notifications are in the graph now, we can destroy the notify_data */ pe__free_clone_notification_data(rsc); } // Check whether a resource or any of its children is known on node static bool rsc_known_on(const pe_resource_t *rsc, const pe_node_t *node) { if (rsc->children) { for (GList *child_iter = rsc->children; child_iter != NULL; child_iter = child_iter->next) { pe_resource_t *child = (pe_resource_t *) child_iter->data; if (rsc_known_on(child, node)) { return TRUE; } } } else if (rsc->known_on) { GHashTableIter iter; pe_node_t *known_node = NULL; g_hash_table_iter_init(&iter, rsc->known_on); while (g_hash_table_iter_next(&iter, NULL, (gpointer *) &known_node)) { if (node->details == known_node->details) { return TRUE; } } } return FALSE; } // Look for an instance of clone that is known on node static pe_resource_t * find_instance_on(const pe_resource_t *clone, const pe_node_t *node) { for (GList *gIter = clone->children; gIter != NULL; gIter = gIter->next) { pe_resource_t *child = (pe_resource_t *) gIter->data; if (rsc_known_on(child, node)) { return child; } } return NULL; } // For anonymous clones, only a single instance needs to be probed static bool probe_anonymous_clone(pe_resource_t *rsc, pe_node_t *node, pe_working_set_t *data_set) { // First, check if we probed an instance on this node last time pe_resource_t *child = find_instance_on(rsc, node); // Otherwise, check if we plan to start an instance on this node if (child == NULL) { for (GList *child_iter = rsc->children; child_iter && !child; child_iter = child_iter->next) { pe_node_t *local_node = NULL; pe_resource_t *child_rsc = (pe_resource_t *) child_iter->data; if (child_rsc) { /* make clang analyzer happy */ local_node = child_rsc->fns->location(child_rsc, NULL, FALSE); if (local_node && (local_node->details == node->details)) { child = child_rsc; } } } } // Otherwise, use the first clone instance if (child == NULL) { child = rsc->children->data; } CRM_ASSERT(child); return child->cmds->create_probe(child, node); } /*! * \internal * * \brief Schedule any probes needed for a resource on a node * * \param[in,out] rsc Resource to create probe for * \param[in,out] node Node to create probe on * * \return true if any probe was created, otherwise false */ bool clone_create_probe(pe_resource_t *rsc, pe_node_t *node) { CRM_ASSERT(rsc); rsc->children = g_list_sort(rsc->children, pcmk__cmp_instance_number); if (rsc->children == NULL) { pe_warn("Clone %s has no children", rsc->id); return false; } if (rsc->exclusive_discover) { pe_node_t *allowed = g_hash_table_lookup(rsc->allowed_nodes, node->details->id); if (allowed && allowed->rsc_discover_mode != pe_discover_exclusive) { /* exclusive discover is enabled and this node is not marked * as a node this resource should be discovered on * * remove the node from allowed_nodes so that the * notification contains only nodes that we might ever run * on */ g_hash_table_remove(rsc->allowed_nodes, node->details->id); /* Bit of a shortcut - might as well take it */ return false; } } if (pcmk_is_set(rsc->flags, pe_rsc_unique)) { return pcmk__probe_resource_list(rsc->children, node); } else { return probe_anonymous_clone(rsc, node, rsc->cluster); } } void clone_append_meta(const pe_resource_t *rsc, xmlNode *xml) { char *name = NULL; name = crm_meta_name(XML_RSC_ATTR_UNIQUE); crm_xml_add(xml, name, pe__rsc_bool_str(rsc, pe_rsc_unique)); free(name); name = crm_meta_name(XML_RSC_ATTR_NOTIFY); crm_xml_add(xml, name, pe__rsc_bool_str(rsc, pe_rsc_notify)); free(name); name = crm_meta_name(XML_RSC_ATTR_INCARNATION_MAX); crm_xml_add_int(xml, name, pe__clone_max(rsc)); free(name); name = crm_meta_name(XML_RSC_ATTR_INCARNATION_NODEMAX); crm_xml_add_int(xml, name, pe__clone_node_max(rsc)); free(name); if (pcmk_is_set(rsc->flags, pe_rsc_promotable)) { int promoted_max = pe__clone_promoted_max(rsc); int promoted_node_max = pe__clone_promoted_node_max(rsc); name = crm_meta_name(XML_RSC_ATTR_PROMOTED_MAX); crm_xml_add_int(xml, name, promoted_max); free(name); name = crm_meta_name(XML_RSC_ATTR_PROMOTED_NODEMAX); crm_xml_add_int(xml, name, promoted_node_max); free(name); /* @COMPAT Maintain backward compatibility with resource agents that * expect the old names (deprecated since 2.0.0). */ name = crm_meta_name(PCMK_XA_PROMOTED_MAX_LEGACY); crm_xml_add_int(xml, name, promoted_max); free(name); name = crm_meta_name(PCMK_XA_PROMOTED_NODE_MAX_LEGACY); crm_xml_add_int(xml, name, promoted_node_max); free(name); } } // Clone implementation of resource_alloc_functions_t:add_utilization() void pcmk__clone_add_utilization(const pe_resource_t *rsc, const pe_resource_t *orig_rsc, GList *all_rscs, GHashTable *utilization) { bool existing = false; pe_resource_t *child = NULL; if (!pcmk_is_set(rsc->flags, pe_rsc_provisional)) { return; } // Look for any child already existing in the list for (GList *iter = rsc->children; iter != NULL; iter = iter->next) { child = (pe_resource_t *) iter->data; if (g_list_find(all_rscs, child)) { existing = true; // Keep checking remaining children } else { // If this is a clone of a group, look for group's members for (GList *member_iter = child->children; member_iter != NULL; member_iter = member_iter->next) { pe_resource_t *member = (pe_resource_t *) member_iter->data; if (g_list_find(all_rscs, member) != NULL) { // Add *child's* utilization, not group member's child->cmds->add_utilization(child, orig_rsc, all_rscs, utilization); existing = true; break; } } } } if (!existing && (rsc->children != NULL)) { // If nothing was found, still add first child's utilization child = (pe_resource_t *) rsc->children->data; child->cmds->add_utilization(child, orig_rsc, all_rscs, utilization); } } // Clone implementation of resource_alloc_functions_t:shutdown_lock() void pcmk__clone_shutdown_lock(pe_resource_t *rsc) { return; // Clones currently don't support shutdown locks } diff --git a/lib/pacemaker/pcmk_sched_colocation.c b/lib/pacemaker/pcmk_sched_colocation.c index f1e62d4d17..d3284905a9 100644 --- a/lib/pacemaker/pcmk_sched_colocation.c +++ b/lib/pacemaker/pcmk_sched_colocation.c @@ -1,1650 +1,1651 @@ /* * Copyright 2004-2023 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU General Public License version 2 * or later (GPLv2+) WITHOUT ANY WARRANTY. */ #include #include #include #include #include #include #include "crm/common/util.h" #include "crm/common/xml_internal.h" #include "crm/msg_xml.h" #include "libpacemaker_private.h" #define EXPAND_CONSTRAINT_IDREF(__set, __rsc, __name) do { \ __rsc = pcmk__find_constraint_resource(data_set->resources, __name); \ if (__rsc == NULL) { \ pcmk__config_err("%s: No resource found for %s", __set, __name); \ return; \ } \ } while(0) // Used to temporarily mark a node as unusable #define INFINITY_HACK (INFINITY * -100) static gint cmp_dependent_priority(gconstpointer a, gconstpointer b) { const pcmk__colocation_t *rsc_constraint1 = (const pcmk__colocation_t *) a; const pcmk__colocation_t *rsc_constraint2 = (const pcmk__colocation_t *) b; if (a == NULL) { return 1; } if (b == NULL) { return -1; } CRM_ASSERT(rsc_constraint1->dependent != NULL); CRM_ASSERT(rsc_constraint1->primary != NULL); if (rsc_constraint1->dependent->priority > rsc_constraint2->dependent->priority) { return -1; } if (rsc_constraint1->dependent->priority < rsc_constraint2->dependent->priority) { return 1; } /* Process clones before primitives and groups */ if (rsc_constraint1->dependent->variant > rsc_constraint2->dependent->variant) { return -1; } if (rsc_constraint1->dependent->variant < rsc_constraint2->dependent->variant) { return 1; } /* @COMPAT scheduler <2.0.0: Process promotable clones before nonpromotable * clones (probably unnecessary, but avoids having to update regression * tests) */ if (rsc_constraint1->dependent->variant == pe_clone) { if (pcmk_is_set(rsc_constraint1->dependent->flags, pe_rsc_promotable) && !pcmk_is_set(rsc_constraint2->dependent->flags, pe_rsc_promotable)) { return -1; } else if (!pcmk_is_set(rsc_constraint1->dependent->flags, pe_rsc_promotable) && pcmk_is_set(rsc_constraint2->dependent->flags, pe_rsc_promotable)) { return 1; } } return strcmp(rsc_constraint1->dependent->id, rsc_constraint2->dependent->id); } static gint cmp_primary_priority(gconstpointer a, gconstpointer b) { const pcmk__colocation_t *rsc_constraint1 = (const pcmk__colocation_t *) a; const pcmk__colocation_t *rsc_constraint2 = (const pcmk__colocation_t *) b; if (a == NULL) { return 1; } if (b == NULL) { return -1; } CRM_ASSERT(rsc_constraint1->dependent != NULL); CRM_ASSERT(rsc_constraint1->primary != NULL); if (rsc_constraint1->primary->priority > rsc_constraint2->primary->priority) { return -1; } if (rsc_constraint1->primary->priority < rsc_constraint2->primary->priority) { return 1; } /* Process clones before primitives and groups */ if (rsc_constraint1->primary->variant > rsc_constraint2->primary->variant) { return -1; } else if (rsc_constraint1->primary->variant < rsc_constraint2->primary->variant) { return 1; } /* @COMPAT scheduler <2.0.0: Process promotable clones before nonpromotable * clones (probably unnecessary, but avoids having to update regression * tests) */ if (rsc_constraint1->primary->variant == pe_clone) { if (pcmk_is_set(rsc_constraint1->primary->flags, pe_rsc_promotable) && !pcmk_is_set(rsc_constraint2->primary->flags, pe_rsc_promotable)) { return -1; } else if (!pcmk_is_set(rsc_constraint1->primary->flags, pe_rsc_promotable) && pcmk_is_set(rsc_constraint2->primary->flags, pe_rsc_promotable)) { return 1; } } return strcmp(rsc_constraint1->primary->id, rsc_constraint2->primary->id); } /*! * \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 * * \note The list will be sorted using cmp_primary_priority(). */ void pcmk__add_this_with(GList **list, const pcmk__colocation_t *colocation) { CRM_ASSERT((list != NULL) && (colocation != NULL)); crm_trace("Adding colocation %s (%s with %s%s%s @%d) " "to 'this with' list", colocation->id, colocation->dependent->id, colocation->primary->id, (colocation->node_attribute == NULL)? "" : " using ", pcmk__s(colocation->node_attribute, ""), colocation->score); *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 * * \note The lists must be pre-sorted by cmp_primary_priority(). */ void pcmk__add_this_with_list(GList **list, GList *addition) { CRM_CHECK((list != NULL), return); if (*list == NULL) { // Trivial case for efficiency crm_trace("Copying %u 'this with' colocations to new list", g_list_length(addition)); *list = g_list_copy(addition); } else { while (addition != NULL) { pcmk__add_this_with(list, addition->data); addition = addition->next; } } } /*! * \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 * * \note The list will be sorted using cmp_dependent_priority(). */ void pcmk__add_with_this(GList **list, const pcmk__colocation_t *colocation) { CRM_ASSERT((list != NULL) && (colocation != NULL)); crm_trace("Adding colocation %s (%s with %s%s%s @%d) " "to 'with this' list", colocation->id, colocation->dependent->id, colocation->primary->id, (colocation->node_attribute == NULL)? "" : " using ", pcmk__s(colocation->node_attribute, ""), colocation->score); *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 * * \note The lists must be pre-sorted by cmp_dependent_priority(). */ void pcmk__add_with_this_list(GList **list, GList *addition) { CRM_CHECK((list != NULL), return); if (*list == NULL) { // Trivial case for efficiency crm_trace("Copying %u 'with this' colocations to new list", g_list_length(addition)); *list = g_list_copy(addition); } else { while (addition != NULL) { pcmk__add_with_this(list, addition->data); addition = addition->next; } } } /*! * \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(pe_resource_t *first_rsc, int first_role, pe_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 == RSC_ROLE_PROMOTED) { first_tasks[0] = CRMD_ACTION_DEMOTE; } else { first_tasks[0] = CRMD_ACTION_STOP; if (first_role == RSC_ROLE_UNPROMOTED) { first_tasks[1] = CRMD_ACTION_PROMOTE; } } /* Actions to make then_rsc gain then_role */ if (then_role == RSC_ROLE_PROMOTED) { then_tasks[0] = CRMD_ACTION_PROMOTE; } else { then_tasks[0] = CRMD_ACTION_START; if (then_role == RSC_ROLE_UNPROMOTED) { then_tasks[1] = CRMD_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], pe_order_anti_colocation); } } } /*! * \internal * \brief Add a new colocation constraint to a cluster working set * * \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 Current role of \p dependent * \param[in] primary_role Current role of \p primary * \param[in] influence Whether colocation constraint has influence * \param[in,out] data_set Cluster working set to add constraint to */ void pcmk__new_colocation(const char *id, const char *node_attr, int score, pe_resource_t *dependent, pe_resource_t *primary, const char *dependent_role, const char *primary_role, bool influence, pe_working_set_t *data_set) { pcmk__colocation_t *new_con = NULL; if (score == 0) { crm_trace("Ignoring colocation '%s' because score is 0", id); return; } if ((dependent == NULL) || (primary == NULL)) { pcmk__config_err("Ignoring colocation '%s' because resource " "does not exist", id); return; } new_con = calloc(1, sizeof(pcmk__colocation_t)); if (new_con == NULL) { return; } if (pcmk__str_eq(dependent_role, RSC_ROLE_STARTED_S, pcmk__str_null_matches|pcmk__str_casei)) { dependent_role = RSC_ROLE_UNKNOWN_S; } if (pcmk__str_eq(primary_role, RSC_ROLE_STARTED_S, pcmk__str_null_matches|pcmk__str_casei)) { primary_role = RSC_ROLE_UNKNOWN_S; } new_con->id = id; new_con->dependent = dependent; new_con->primary = primary; new_con->score = score; new_con->dependent_role = text2role(dependent_role); new_con->primary_role = text2role(primary_role); new_con->node_attribute = node_attr; new_con->influence = influence; if (node_attr == NULL) { node_attr = CRM_ATTR_UNAME; } pe_rsc_trace(dependent, "%s ==> %s (%s %d)", dependent->id, primary->id, node_attr, score); pcmk__add_this_with(&(dependent->rsc_cons), new_con); pcmk__add_with_this(&(primary->rsc_cons_lhs), new_con); data_set->colocation_constraints = g_list_append(data_set->colocation_constraints, new_con); if (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 influence option * * \return true if string evaluates true, false if string evaluates false, * or value of resource's critical option if string is NULL or invalid */ static bool unpack_influence(const char *coloc_id, const pe_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 " XML_COLOC_ATTR_INFLUENCE " (using default)", coloc_id); } else { return (influence_i != 0); } } return pcmk_is_set(rsc->flags, pe_rsc_critical); } static void unpack_colocation_set(xmlNode *set, int score, const char *coloc_id, const char *influence_s, pe_working_set_t *data_set) { xmlNode *xml_rsc = NULL; pe_resource_t *with = NULL; pe_resource_t *resource = NULL; const char *set_id = ID(set); const char *role = crm_element_value(set, "role"); const char *ordering = crm_element_value(set, "ordering"); int local_score = score; bool sequential = false; const char *score_s = crm_element_value(set, XML_RULE_ATTR_SCORE); if (score_s) { local_score = char2score(score_s); } if (local_score == 0) { crm_trace("Ignoring colocation '%s' for set '%s' because score is 0", coloc_id, set_id); return; } if (ordering == NULL) { ordering = "group"; } if (pcmk__xe_get_bool_attr(set, "sequential", &sequential) == pcmk_rc_ok && !sequential) { return; } else if ((local_score > 0) && pcmk__str_eq(ordering, "group", pcmk__str_casei)) { for (xml_rsc = first_named_child(set, XML_TAG_RESOURCE_REF); xml_rsc != NULL; xml_rsc = crm_next_same_xml(xml_rsc)) { EXPAND_CONSTRAINT_IDREF(set_id, resource, ID(xml_rsc)); if (with != NULL) { pe_rsc_trace(resource, "Colocating %s with %s", resource->id, with->id); pcmk__new_colocation(set_id, NULL, local_score, resource, with, role, role, unpack_influence(coloc_id, resource, influence_s), data_set); } with = resource; } } else if (local_score > 0) { pe_resource_t *last = NULL; for (xml_rsc = first_named_child(set, XML_TAG_RESOURCE_REF); xml_rsc != NULL; xml_rsc = crm_next_same_xml(xml_rsc)) { EXPAND_CONSTRAINT_IDREF(set_id, resource, ID(xml_rsc)); if (last != NULL) { pe_rsc_trace(resource, "Colocating %s with %s", last->id, resource->id); pcmk__new_colocation(set_id, NULL, local_score, last, resource, role, role, unpack_influence(coloc_id, last, influence_s), data_set); } last = 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 = first_named_child(set, XML_TAG_RESOURCE_REF); xml_rsc != NULL; xml_rsc = crm_next_same_xml(xml_rsc)) { xmlNode *xml_rsc_with = NULL; bool influence = true; EXPAND_CONSTRAINT_IDREF(set_id, resource, ID(xml_rsc)); influence = unpack_influence(coloc_id, resource, influence_s); for (xml_rsc_with = first_named_child(set, XML_TAG_RESOURCE_REF); xml_rsc_with != NULL; xml_rsc_with = crm_next_same_xml(xml_rsc_with)) { if (pcmk__str_eq(resource->id, ID(xml_rsc_with), pcmk__str_casei)) { break; } EXPAND_CONSTRAINT_IDREF(set_id, with, ID(xml_rsc_with)); pe_rsc_trace(resource, "Anti-Colocating %s with %s", resource->id, with->id); pcmk__new_colocation(set_id, NULL, local_score, resource, with, role, role, influence, data_set); } } } } static void colocate_rsc_sets(const char *id, xmlNode *set1, xmlNode *set2, int score, const char *influence_s, pe_working_set_t *data_set) { xmlNode *xml_rsc = NULL; pe_resource_t *rsc_1 = NULL; pe_resource_t *rsc_2 = NULL; const char *role_1 = crm_element_value(set1, "role"); const char *role_2 = crm_element_value(set2, "role"); int rc = pcmk_rc_ok; bool sequential = false; if (score == 0) { crm_trace("Ignoring colocation '%s' between sets because score is 0", id); return; } rc = pcmk__xe_get_bool_attr(set1, "sequential", &sequential); if (rc != pcmk_rc_ok || sequential) { // Get the first one xml_rsc = first_named_child(set1, XML_TAG_RESOURCE_REF); if (xml_rsc != NULL) { EXPAND_CONSTRAINT_IDREF(id, rsc_1, ID(xml_rsc)); } } rc = pcmk__xe_get_bool_attr(set2, "sequential", &sequential); if (rc != pcmk_rc_ok || sequential) { // Get the last one const char *rid = NULL; for (xml_rsc = first_named_child(set2, XML_TAG_RESOURCE_REF); xml_rsc != NULL; xml_rsc = crm_next_same_xml(xml_rsc)) { rid = ID(xml_rsc); } EXPAND_CONSTRAINT_IDREF(id, rsc_2, rid); } if ((rsc_1 != NULL) && (rsc_2 != NULL)) { pcmk__new_colocation(id, NULL, score, rsc_1, rsc_2, role_1, role_2, unpack_influence(id, rsc_1, influence_s), data_set); } else if (rsc_1 != NULL) { bool influence = unpack_influence(id, rsc_1, influence_s); for (xml_rsc = first_named_child(set2, XML_TAG_RESOURCE_REF); xml_rsc != NULL; xml_rsc = crm_next_same_xml(xml_rsc)) { EXPAND_CONSTRAINT_IDREF(id, rsc_2, ID(xml_rsc)); pcmk__new_colocation(id, NULL, score, rsc_1, rsc_2, role_1, role_2, influence, data_set); } } else if (rsc_2 != NULL) { for (xml_rsc = first_named_child(set1, XML_TAG_RESOURCE_REF); xml_rsc != NULL; xml_rsc = crm_next_same_xml(xml_rsc)) { EXPAND_CONSTRAINT_IDREF(id, rsc_1, ID(xml_rsc)); pcmk__new_colocation(id, NULL, score, rsc_1, rsc_2, role_1, role_2, unpack_influence(id, rsc_1, influence_s), data_set); } } else { for (xml_rsc = first_named_child(set1, XML_TAG_RESOURCE_REF); xml_rsc != NULL; xml_rsc = crm_next_same_xml(xml_rsc)) { xmlNode *xml_rsc_2 = NULL; bool influence = true; EXPAND_CONSTRAINT_IDREF(id, rsc_1, ID(xml_rsc)); influence = unpack_influence(id, rsc_1, influence_s); for (xml_rsc_2 = first_named_child(set2, XML_TAG_RESOURCE_REF); xml_rsc_2 != NULL; xml_rsc_2 = crm_next_same_xml(xml_rsc_2)) { EXPAND_CONSTRAINT_IDREF(id, rsc_2, ID(xml_rsc_2)); pcmk__new_colocation(id, NULL, score, rsc_1, rsc_2, role_1, role_2, influence, data_set); } } } } static void unpack_simple_colocation(xmlNode *xml_obj, const char *id, const char *influence_s, pe_working_set_t *data_set) { int score_i = 0; const char *score = crm_element_value(xml_obj, XML_RULE_ATTR_SCORE); const char *dependent_id = crm_element_value(xml_obj, XML_COLOC_ATTR_SOURCE); const char *primary_id = crm_element_value(xml_obj, XML_COLOC_ATTR_TARGET); const char *dependent_role = crm_element_value(xml_obj, XML_COLOC_ATTR_SOURCE_ROLE); const char *primary_role = crm_element_value(xml_obj, XML_COLOC_ATTR_TARGET_ROLE); const char *attr = crm_element_value(xml_obj, XML_COLOC_ATTR_NODE_ATTR); // @COMPAT: Deprecated since 2.1.5 const char *dependent_instance = crm_element_value(xml_obj, XML_COLOC_ATTR_SOURCE_INSTANCE); // @COMPAT: Deprecated since 2.1.5 const char *primary_instance = crm_element_value(xml_obj, XML_COLOC_ATTR_TARGET_INSTANCE); pe_resource_t *dependent = pcmk__find_constraint_resource(data_set->resources, dependent_id); pe_resource_t *primary = pcmk__find_constraint_resource(data_set->resources, primary_id); if (dependent_instance != NULL) { pe_warn_once(pe_wo_coloc_inst, "Support for " XML_COLOC_ATTR_SOURCE_INSTANCE " is " "deprecated and will be removed in a future release."); } if (primary_instance != NULL) { pe_warn_once(pe_wo_coloc_inst, "Support for " XML_COLOC_ATTR_TARGET_INSTANCE " is " "deprecated and will be removed in a future release."); } 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; } else if ((dependent_instance != NULL) && !pe_rsc_is_clone(dependent)) { pcmk__config_err("Ignoring constraint '%s' because resource '%s' " "is not a clone but instance '%s' was requested", id, dependent_id, dependent_instance); return; } else if ((primary_instance != NULL) && !pe_rsc_is_clone(primary)) { pcmk__config_err("Ignoring constraint '%s' because resource '%s' " "is not a clone but instance '%s' was requested", id, primary_id, primary_instance); return; } if (dependent_instance != NULL) { dependent = find_clone_instance(dependent, dependent_instance); if (dependent == NULL) { pcmk__config_warn("Ignoring constraint '%s' because resource '%s' " "does not have an instance '%s'", id, dependent_id, dependent_instance); return; } } if (primary_instance != NULL) { primary = find_clone_instance(primary, primary_instance); if (primary == NULL) { pcmk__config_warn("Ignoring constraint '%s' because resource '%s' " "does not have an instance '%s'", "'%s'", id, primary_id, primary_instance); return; } } if (pcmk__xe_attr_is_true(xml_obj, XML_CONS_ATTR_SYMMETRICAL)) { pcmk__config_warn("The colocation constraint '" XML_CONS_ATTR_SYMMETRICAL "' attribute has been removed"); } if (score) { score_i = char2score(score); } pcmk__new_colocation(id, attr, score_i, dependent, primary, dependent_role, primary_role, unpack_influence(id, dependent, influence_s), data_set); } // \return Standard Pacemaker return code static int unpack_colocation_tags(xmlNode *xml_obj, xmlNode **expanded_xml, pe_working_set_t *data_set) { const char *id = NULL; const char *dependent_id = NULL; const char *primary_id = NULL; const char *dependent_role = NULL; const char *primary_role = NULL; pe_resource_t *dependent = NULL; pe_resource_t *primary = NULL; pe_tag_t *dependent_tag = NULL; pe_tag_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 = ID(xml_obj); if (id == NULL) { pcmk__config_err("Ignoring <%s> constraint without " XML_ATTR_ID, crm_element_name(xml_obj)); 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, data_set); if (*expanded_xml != NULL) { crm_log_xml_trace(*expanded_xml, "Expanded rsc_colocation"); return pcmk_rc_ok; } dependent_id = crm_element_value(xml_obj, XML_COLOC_ATTR_SOURCE); primary_id = crm_element_value(xml_obj, XML_COLOC_ATTR_TARGET); if ((dependent_id == NULL) || (primary_id == NULL)) { return pcmk_rc_ok; } if (!pcmk__valid_resource_or_tag(data_set, 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(data_set, 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, XML_COLOC_ATTR_SOURCE_ROLE); primary_role = crm_element_value(xml_obj, XML_COLOC_ATTR_TARGET_ROLE); *expanded_xml = copy_xml(xml_obj); // Convert template/tag reference in "rsc" into resource_set under constraint if (!pcmk__tag_to_set(*expanded_xml, &dependent_set, XML_COLOC_ATTR_SOURCE, true, data_set)) { free_xml(*expanded_xml); *expanded_xml = NULL; return pcmk_rc_unpack_error; } if (dependent_set != NULL) { if (dependent_role != NULL) { // Move "rsc-role" into converted resource_set as "role" crm_xml_add(dependent_set, "role", dependent_role); xml_remove_prop(*expanded_xml, XML_COLOC_ATTR_SOURCE_ROLE); } any_sets = true; } // Convert template/tag reference in "with-rsc" into resource_set under constraint if (!pcmk__tag_to_set(*expanded_xml, &primary_set, XML_COLOC_ATTR_TARGET, true, data_set)) { free_xml(*expanded_xml); *expanded_xml = NULL; return pcmk_rc_unpack_error; } if (primary_set != NULL) { if (primary_role != NULL) { // Move "with-rsc-role" into converted resource_set as "role" crm_xml_add(primary_set, "role", primary_role); xml_remove_prop(*expanded_xml, XML_COLOC_ATTR_TARGET_ROLE); } any_sets = true; } if (any_sets) { crm_log_xml_trace(*expanded_xml, "Expanded rsc_colocation"); } else { free_xml(*expanded_xml); *expanded_xml = NULL; } return pcmk_rc_ok; } /*! * \internal * \brief Parse a colocation constraint from XML into a cluster working set * * \param[in,out] xml_obj Colocation constraint XML to unpack * \param[in,out] data_set Cluster working set to add constraint to */ void pcmk__unpack_colocation(xmlNode *xml_obj, pe_working_set_t *data_set) { 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, XML_ATTR_ID); const char *score = crm_element_value(xml_obj, XML_RULE_ATTR_SCORE); const char *influence_s = crm_element_value(xml_obj, XML_COLOC_ATTR_INFLUENCE); if (score) { score_i = char2score(score); } if (unpack_colocation_tags(xml_obj, &expanded_xml, data_set) != pcmk_rc_ok) { return; } if (expanded_xml) { orig_xml = xml_obj; xml_obj = expanded_xml; } for (set = first_named_child(xml_obj, XML_CONS_TAG_RSC_SET); set != NULL; set = crm_next_same_xml(set)) { set = expand_idref(set, data_set->input); if (set == NULL) { // Configuration error, message already logged if (expanded_xml != NULL) { free_xml(expanded_xml); } return; } unpack_colocation_set(set, score_i, id, influence_s, data_set); if (last != NULL) { colocate_rsc_sets(id, last, set, score_i, influence_s, data_set); } last = set; } if (expanded_xml) { free_xml(expanded_xml); xml_obj = orig_xml; } if (last == NULL) { unpack_simple_colocation(xml_obj, id, influence_s, data_set); } } /*! * \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(pe_resource_t *rsc, const char *task, const pe_resource_t *reason) { char *reason_text = crm_strdup_printf("colocation with %s", reason->id); for (GList *gIter = rsc->actions; gIter != NULL; gIter = gIter->next) { pe_action_t *action = (pe_action_t *) gIter->data; if (pcmk_is_set(action->flags, pe_action_runnable) && pcmk__str_eq(action->task, task, pcmk__str_casei)) { pe__clear_action_flags(action, pe_action_runnable); pe_action_set_reason(action, reason_text, false); pcmk__block_colocation_dependents(action, rsc->cluster); pcmk__update_action_for_orderings(action, rsc->cluster); } } // If parent resource can't perform an action, neither can any children for (GList *iter = rsc->children; iter != NULL; iter = iter->next) { mark_action_blocked((pe_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 * \param[in] data_set Cluster working set (ignored) */ void pcmk__block_colocation_dependents(pe_action_t *action, pe_working_set_t *data_set) { GList *gIter = NULL; GList *colocations = NULL; pe_resource_t *rsc = NULL; bool is_start = false; if (pcmk_is_set(action->flags, pe_action_runnable)) { return; // Only unrunnable actions block dependents } is_start = pcmk__str_eq(action->task, RSC_START, pcmk__str_none); if (!is_start && !pcmk__str_eq(action->task, RSC_PROMOTE, pcmk__str_none)) { return; // Only unrunnable starts and promotes block dependents } CRM_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->parent != NULL) { rsc = rsc->parent; // Bundle } // Colocation fails only if entire primary can't reach desired role for (gIter = rsc->children; gIter != NULL; gIter = gIter->next) { pe_resource_t *child = (pe_resource_t *) gIter->data; pe_action_t *child_action = find_first_action(child->actions, NULL, action->task, NULL); if ((child_action == NULL) || pcmk_is_set(child_action->flags, pe_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 (gIter = colocations; gIter != NULL; gIter = gIter->next) { pcmk__colocation_t *colocation = (pcmk__colocation_t *) gIter->data; if (colocation->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 != RSC_ROLE_PROMOTED)) { continue; } // Block the dependent from reaching its colocated role if (colocation->dependent_role == RSC_ROLE_PROMOTED) { mark_action_blocked(colocation->dependent, RSC_PROMOTE, action->rsc); } else { mark_action_blocked(colocation->dependent, RSC_START, action->rsc); } } g_list_free(colocations); } /*! * \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 allocated * * \return How colocation constraint should be applied at this point */ enum pcmk__coloc_affects pcmk__colocation_affects(const pe_resource_t *dependent, const pe_resource_t *primary, const pcmk__colocation_t *colocation, bool preview) { if (!preview && pcmk_is_set(primary->flags, pe_rsc_provisional)) { // Primary resource has not been allocated yet, so we can't do anything return pcmk__coloc_affects_nothing; } if ((colocation->dependent_role >= RSC_ROLE_UNPROMOTED) && (dependent->parent != NULL) && pcmk_is_set(dependent->parent->flags, pe_rsc_promotable) && !pcmk_is_set(dependent->flags, pe_rsc_provisional)) { /* This is a colocation by role, and the dependent is a promotable clone * that has already been allocated, so the colocation should now affect * the role. */ return pcmk__coloc_affects_role; } if (!preview && !pcmk_is_set(dependent->flags, pe_rsc_provisional)) { /* The dependent resource has already been through allocation, so the * constraint no longer has any effect. Log an error if a mandatory * colocation constraint has been violated. */ const pe_node_t *primary_node = primary->allocated_to; if (dependent->allocated_to == NULL) { crm_trace("Skipping colocation '%s': %s will not run anywhere", colocation->id, dependent->id); } else if (colocation->score >= INFINITY) { // Dependent resource must colocate with primary resource if ((primary_node == NULL) || (primary_node->details != dependent->allocated_to->details)) { crm_err("%s must be colocated with %s but is not (%s vs. %s)", dependent->id, primary->id, pe__node_name(dependent->allocated_to), pe__node_name(primary_node)); } } else if (colocation->score <= -CRM_SCORE_INFINITY) { // Dependent resource must anti-colocate with primary resource if ((primary_node != NULL) && (dependent->allocated_to->details == primary_node->details)) { crm_err("%s and %s must be anti-colocated but are allocated " "to the same node (%s)", dependent->id, primary->id, pe__node_name(primary_node)); } } return pcmk__coloc_affects_nothing; } if ((colocation->score > 0) && (colocation->dependent_role != RSC_ROLE_UNKNOWN) && (colocation->dependent_role != dependent->next_role)) { crm_trace("Skipping colocation '%s': dependent limited to %s role " "but %s next role is %s", colocation->id, role2text(colocation->dependent_role), dependent->id, role2text(dependent->next_role)); return pcmk__coloc_affects_nothing; } if ((colocation->score > 0) && (colocation->primary_role != RSC_ROLE_UNKNOWN) && (colocation->primary_role != primary->next_role)) { crm_trace("Skipping colocation '%s': primary limited to %s role " "but %s next role is %s", colocation->id, role2text(colocation->primary_role), primary->id, role2text(primary->next_role)); return pcmk__coloc_affects_nothing; } if ((colocation->score < 0) && (colocation->dependent_role != RSC_ROLE_UNKNOWN) && (colocation->dependent_role == dependent->next_role)) { crm_trace("Skipping anti-colocation '%s': dependent role %s matches", colocation->id, role2text(colocation->dependent_role)); return pcmk__coloc_affects_nothing; } if ((colocation->score < 0) && (colocation->primary_role != RSC_ROLE_UNKNOWN) && (colocation->primary_role == primary->next_role)) { crm_trace("Skipping anti-colocation '%s': primary role %s matches", colocation->id, role2text(colocation->primary_role)); return pcmk__coloc_affects_nothing; } return pcmk__coloc_affects_location; } /*! * \internal * \brief Apply colocation to dependent for allocation purposes * * Update the allowed node weights of the dependent resource in a colocation, * for the purposes of allocating 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_weights(pe_resource_t *dependent, const pe_resource_t *primary, const pcmk__colocation_t *colocation) { const char *attribute = CRM_ATTR_ID; const char *value = NULL; GHashTable *work = NULL; GHashTableIter iter; pe_node_t *node = NULL; if (colocation->node_attribute != NULL) { attribute = colocation->node_attribute; } if (primary->allocated_to != NULL) { value = pe_node_attribute_raw(primary->allocated_to, attribute); } else if (colocation->score < 0) { // Nothing to do (anti-colocation with something that is not running) return; } work = pcmk__copy_node_table(dependent->allowed_nodes); g_hash_table_iter_init(&iter, work); while (g_hash_table_iter_next(&iter, NULL, (void **)&node)) { if (primary->allocated_to == NULL) { node->weight = pcmk__add_scores(-colocation->score, node->weight); pe_rsc_trace(dependent, "Applied %s to %s score on %s (now %s after " "subtracting %s because primary %s inactive)", colocation->id, dependent->id, pe__node_name(node), pcmk_readable_score(node->weight), pcmk_readable_score(colocation->score), primary->id); } else if (pcmk__str_eq(pe_node_attribute_raw(node, attribute), 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 < CRM_SCORE_INFINITY) { node->weight = pcmk__add_scores(colocation->score, node->weight); pe_rsc_trace(dependent, "Applied %s to %s score on %s (now %s after " "adding %s)", colocation->id, dependent->id, pe__node_name(node), pcmk_readable_score(node->weight), pcmk_readable_score(colocation->score)); } } else if (colocation->score >= CRM_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->weight = -CRM_SCORE_INFINITY; pe_rsc_trace(dependent, "Banned %s from %s because colocation %s attribute %s " "does not match", dependent->id, pe__node_name(node), colocation->id, attribute); } } if ((colocation->score <= -INFINITY) || (colocation->score >= INFINITY) || pcmk__any_node_available(work)) { g_hash_table_destroy(dependent->allowed_nodes); dependent->allowed_nodes = work; work = NULL; } else { pe_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 */ void pcmk__apply_coloc_to_priority(pe_resource_t *dependent, const pe_resource_t *primary, const pcmk__colocation_t *colocation) { const char *dependent_value = NULL; const char *primary_value = NULL; const char *attribute = CRM_ATTR_ID; int score_multiplier = 1; if ((primary->allocated_to == NULL) || (dependent->allocated_to == NULL)) { return; } if (colocation->node_attribute != NULL) { attribute = colocation->node_attribute; } dependent_value = pe_node_attribute_raw(dependent->allocated_to, attribute); primary_value = pe_node_attribute_raw(primary->allocated_to, attribute); if (!pcmk__str_eq(dependent_value, primary_value, pcmk__str_casei)) { if ((colocation->score == INFINITY) && (colocation->dependent_role == RSC_ROLE_PROMOTED)) { dependent->priority = -INFINITY; } return; } if ((colocation->primary_role != RSC_ROLE_UNKNOWN) && (colocation->primary_role != primary->next_role)) { return; } if (colocation->dependent_role == RSC_ROLE_UNPROMOTED) { score_multiplier = -1; } dependent->priority = pcmk__add_scores(score_multiplier * colocation->score, dependent->priority); pe_rsc_trace(dependent, "Applied %s to %s promotion priority (now %s after %s %s)", colocation->id, dependent->id, pcmk_readable_score(dependent->priority), ((score_multiplier == 1)? "adding" : "subtracting"), pcmk_readable_score(colocation->score)); } /*! * \internal * \brief Find score of highest-scored node that matches colocation attribute * * \param[in] 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 pe_resource_t *rsc, const char *attr, const char *value) { GHashTableIter iter; pe_node_t *node = NULL; int best_score = -INFINITY; const char *best_node = NULL; // Find best allowed node with matching attribute g_hash_table_iter_init(&iter, rsc->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (void **) &node)) { if ((node->weight > best_score) && pcmk__node_available(node, false, false) && pcmk__str_eq(value, pe_node_attribute_raw(node, attr), pcmk__str_casei)) { best_score = node->weight; best_node = node->details->uname; } } if (!pcmk__str_eq(attr, CRM_ATTR_UNAME, pcmk__str_casei)) { 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); } } return best_score; } /*! * \internal * \brief Add resource's colocation matches to current node allocation 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 Hash table of nodes with allocation scores so far * \param[in] rsc Resource whose allowed nodes should be compared * \param[in] attr Colocation attribute that must match (NULL for default) * \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, const pe_resource_t *rsc, const char *attr, float factor, bool only_positive) { GHashTableIter iter; pe_node_t *node = NULL; if (attr == NULL) { attr = CRM_ATTR_UNAME; } // Iterate through each node g_hash_table_iter_init(&iter, nodes); while (g_hash_table_iter_next(&iter, NULL, (void **)&node)) { float weight_f = 0; int weight = 0; int score = 0; int new_score = 0; score = best_node_score_matching_attr(rsc, attr, pe_node_attribute_raw(node, attr)); if ((factor < 0) && (score < 0)) { /* Negative preference for a node with a negative score * should not become a positive preference. * * @TODO Consider filtering only if weight is -INFINITY */ crm_trace("%s: Filtering %d + %f * %d (double negative disallowed)", pe__node_name(node), node->weight, factor, score); continue; } if (node->weight == INFINITY_HACK) { crm_trace("%s: Filtering %d + %f * %d (node was marked unusable)", pe__node_name(node), node->weight, factor, score); continue; } weight_f = factor * score; // Round the number; see http://c-faq.com/fp/round.html weight = (int) ((weight_f < 0)? (weight_f - 0.5) : (weight_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 ((weight == 0) && (score != 0)) { if (factor > 0.0) { weight = 1; } else if (factor < 0.0) { weight = -1; } } new_score = pcmk__add_scores(weight, node->weight); if (only_positive && (new_score < 0) && (node->weight > 0)) { crm_trace("%s: Filtering %d + %f * %d = %d " "(negative disallowed, marking node unusable)", pe__node_name(node), node->weight, factor, score, new_score); node->weight = INFINITY_HACK; continue; } if (only_positive && (new_score < 0) && (node->weight == 0)) { crm_trace("%s: Filtering %d + %f * %d = %d (negative disallowed)", pe__node_name(node), node->weight, factor, score, new_score); continue; } crm_trace("%s: %d + %f * %d = %d", pe__node_name(node), node->weight, factor, score, new_score); node->weight = new_score; } } /*! * \internal * \brief Initialize colocated node table for a group resource * * \param[in] rsc Group resource being colocated with another resource * \param[in] log_id Resource ID to use in log messages * \param[in] nodes Nodes to update * \param[in] attr Colocation attribute (NULL to use default) * \param[in] factor Incorporate scores multiplied by this factor * \param[in] flags Bitmask of enum pcmk__coloc_select values * * \return Table of node scores initialized for colocation, or NULL if resource * should be ignored for colocation purposes * \note The caller is responsible for freeing a non-NULL return value using * g_hash_table_destroy(). */ static GHashTable * init_group_colocated_nodes(const pe_resource_t *rsc, const char *log_id, GHashTable **nodes, const char *attr, float factor, uint32_t flags) { pe_resource_t *member = NULL; // Ignore empty groups (only possible with schema validation disabled) if (rsc->children == NULL) { return NULL; } if (*nodes == NULL) { // Only cmp_resources() passes a NULL nodes table member = pe__last_group_member(rsc); pe_rsc_trace(rsc, "%s: Merging scores from group %s using member %s " "(at %.6f)", log_id, rsc->id, member->id, factor); - pcmk__add_colocated_node_scores(member, log_id, nodes, attr, factor, - flags); + member->cmds->add_colocated_node_scores(member, log_id, nodes, attr, + factor, flags); } else { /* The first member of the group will recursively incorporate any * constraints involving other members (including the group internal * colocation). * * @TODO The indirect colocations from the dependent group's other * members will be incorporated at full strength rather than by * factor, so the group's combined stickiness will be treated as * (factor + (#members - 1)) * stickiness. It is questionable what * the right approach should be. */ member = rsc->children->data; - pcmk__add_colocated_node_scores(member, log_id, nodes, attr, factor, - flags); + member->cmds->add_colocated_node_scores(member, log_id, nodes, attr, + factor, flags); // Above handles everything, so work is left as NULL } return NULL; } /*! * \internal * \brief Initialize colocated node table for a non-group resource * * \param[in] rsc Non-group resource being colocated with another resource * \param[in] log_id Resource ID to use in log messages * \param[in] nodes Nodes to update * \param[in] attr Colocation attribute (NULL to use default) * \param[in] factor Incorporate scores multiplied by this factor * \param[in] flags Bitmask of enum pcmk__coloc_select values * * \return Table of node scores initialized for colocation, or NULL if resource * should be ignored for colocation purposes * \note The caller is responsible for freeing a non-NULL return value using * g_hash_table_destroy(). */ static GHashTable * init_nongroup_colocated_nodes(const pe_resource_t *rsc, const char *log_id, GHashTable **nodes, const char *attr, float factor, uint32_t flags) { GHashTable *work = NULL; if (*nodes == NULL) { /* Only cmp_resources() passes a NULL nodes table, which indicates we * should initialize it with the resource's allowed node scores. */ work = pcmk__copy_node_table(rsc->allowed_nodes); } else { pe_rsc_trace(rsc, "%s: Merging scores from %s (at %.6f)", log_id, rsc->id, factor); work = pcmk__copy_node_table(*nodes); add_node_scores_matching_attr(work, rsc, attr, factor, pcmk_is_set(flags, pcmk__coloc_select_nonnegative)); } return work; } /*! * \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] rsc Resource to check colocations for * \param[in] log_id Resource ID to use in logs (if NULL, use \p rsc ID) * \param[in,out] nodes Nodes to update * \param[in] attr Colocation attribute (NULL to use default) * \param[in] factor Incorporate scores multiplied by this factor * \param[in] flags Bitmask of enum pcmk__coloc_select values * * \note The caller remains responsible for freeing \p *nodes. */ void pcmk__add_colocated_node_scores(pe_resource_t *rsc, const char *log_id, GHashTable **nodes, const char *attr, float factor, uint32_t flags) { GHashTable *work = NULL; CRM_CHECK((rsc != NULL) && (nodes != NULL), return); if (log_id == NULL) { log_id = rsc->id; } // Avoid infinite recursion if (pcmk_is_set(rsc->flags, pe_rsc_merging)) { pe_rsc_info(rsc, "%s: Breaking dependency loop at %s", log_id, rsc->id); return; } pe__set_resource_flags(rsc, pe_rsc_merging); if (rsc->variant == pe_group) { work = init_group_colocated_nodes(rsc, log_id, nodes, attr, factor, flags); } else { work = init_nongroup_colocated_nodes(rsc, log_id, nodes, attr, factor, flags); } if (work == NULL) { pe__clear_resource_flags(rsc, pe_rsc_merging); return; } if (pcmk__any_node_available(work)) { GList *colocations = NULL; float multiplier = (factor < 0.0)? -1.0 : 1.0; if (pcmk_is_set(flags, pcmk__coloc_select_this_with)) { colocations = pcmk__this_with_colocations(rsc); pe_rsc_trace(rsc, "Checking additional %d optional '%s with' constraints", g_list_length(colocations), rsc->id); } else { colocations = pcmk__with_this_colocations(rsc); pe_rsc_trace(rsc, "Checking additional %d optional 'with %s' constraints", g_list_length(colocations), rsc->id); } + flags |= pcmk__coloc_select_active; for (GList *iter = colocations; iter != NULL; iter = iter->next) { pe_resource_t *other = NULL; pcmk__colocation_t *constraint = (pcmk__colocation_t *) iter->data; 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; } pe_rsc_trace(rsc, "Optionally merging score of '%s' constraint (%s with %s)", constraint->id, constraint->dependent->id, constraint->primary->id); factor = multiplier * constraint->score / (float) INFINITY; - pcmk__add_colocated_node_scores(other, log_id, &work, - constraint->node_attribute, factor, - flags|pcmk__coloc_select_active); + other->cmds->add_colocated_node_scores(other, log_id, &work, + constraint->node_attribute, + factor, flags); pe__show_node_weights(true, NULL, log_id, work, rsc->cluster); } g_list_free(colocations); } else if (pcmk_is_set(flags, pcmk__coloc_select_active)) { pe_rsc_info(rsc, "%s: Rolling back optional scores from %s", log_id, rsc->id); g_hash_table_destroy(work); pe__clear_resource_flags(rsc, pe_rsc_merging); return; } if (pcmk_is_set(flags, pcmk__coloc_select_nonnegative)) { pe_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->weight == INFINITY_HACK) { node->weight = 1; } } } if (*nodes != NULL) { g_hash_table_destroy(*nodes); } *nodes = work; pe__clear_resource_flags(rsc, pe_rsc_merging); } /*! * \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 pe_resource_t *rsc) { GList *list = NULL; rsc->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 pe_resource_t *rsc) { GList *list = NULL; rsc->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 8e32e1fba1..c880196f70 100644 --- a/lib/pacemaker/pcmk_sched_instances.c +++ b/lib/pacemaker/pcmk_sched_instances.c @@ -1,1653 +1,1659 @@ /* * Copyright 2004-2023 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU General Public License version 2 * or later (GPLv2+) WITHOUT ANY WARRANTY. */ /* This file is intended for code usable with both clone instances and bundle * replica containers. */ #include #include #include #include "libpacemaker_private.h" /*! * \internal * \brief Check whether a clone or bundle has instances for all available nodes * * \param[in] collective Clone or bundle to check * * \return true if \p collective has enough instances for all of its available * allowed nodes, otherwise false */ static bool can_run_everywhere(const pe_resource_t *collective) { GHashTableIter iter; pe_node_t *node = NULL; int available_nodes = 0; int max_instances = 0; switch (collective->variant) { case pe_clone: max_instances = pe__clone_max(collective); break; case pe_container: max_instances = pe__bundle_max(collective); break; default: return false; // Not actually possible } g_hash_table_iter_init(&iter, collective->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (gpointer *) &node)) { if (pcmk__node_available(node, false, false) && (max_instances < ++available_nodes)) { return false; } } return true; } /*! * \internal * \brief Check whether a node is allowed to run an instance * * \param[in] instance Clone instance or bundle container to check * \param[in] node Node to check * \param[in] max_per_node Maximum number of instances allowed to run on a node * * \return true if \p node is allowed to run \p instance, otherwise false */ static bool can_run_instance(const pe_resource_t *instance, const pe_node_t *node, int max_per_node) { pe_node_t *allowed_node = NULL; if (pcmk_is_set(instance->flags, pe_rsc_orphan)) { pe_rsc_trace(instance, "%s cannot run on %s: orphaned", instance->id, pe__node_name(node)); return false; } if (!pcmk__node_available(node, false, false)) { pe_rsc_trace(instance, "%s cannot run on %s: node cannot run resources", instance->id, pe__node_name(node)); return false; } allowed_node = pcmk__top_allowed_node(instance, node); if (allowed_node == NULL) { crm_warn("%s cannot run on %s: node not allowed", instance->id, pe__node_name(node)); return false; } if (allowed_node->weight < 0) { pe_rsc_trace(instance, "%s cannot run on %s: parent score is %s there", instance->id, pe__node_name(node), pcmk_readable_score(allowed_node->weight)); return false; } if (allowed_node->count >= max_per_node) { pe_rsc_trace(instance, "%s cannot run on %s: node already has %d instance%s", instance->id, pe__node_name(node), max_per_node, pcmk__plural_s(max_per_node)); return false; } pe_rsc_trace(instance, "%s can run on %s (%d already running)", instance->id, pe__node_name(node), allowed_node->count); return true; } /*! * \internal * \brief Ban a clone instance or bundle replica from unavailable allowed nodes * * \param[in,out] instance Clone instance or bundle replica to ban * \param[in] max_per_node Maximum instances allowed to run on a node */ static void ban_unavailable_allowed_nodes(pe_resource_t *instance, int max_per_node) { if (instance->allowed_nodes != NULL) { GHashTableIter iter; pe_node_t *node = NULL; g_hash_table_iter_init(&iter, instance->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (void **) &node)) { if (!can_run_instance(instance, node, max_per_node)) { pe_rsc_trace(instance, "Banning %s from unavailable node %s", instance->id, pe__node_name(node)); node->weight = -INFINITY; for (GList *child_iter = instance->children; child_iter != NULL; child_iter = child_iter->next) { pe_resource_t *child = (pe_resource_t *) child_iter->data; pe_node_t *child_node = NULL; child_node = pe_hash_table_lookup(child->allowed_nodes, node->details->id); if (child_node != NULL) { pe_rsc_trace(instance, "Banning %s child %s " "from unavailable node %s", instance->id, child->id, pe__node_name(node)); child_node->weight = -INFINITY; } } } } } } /*! * \internal * \brief Create a hash table with a single node in it * * \param[in] node Node to copy into new table * * \return Newly created hash table containing a copy of \p node * \note The caller is responsible for freeing the result with * g_hash_table_destroy(). */ static GHashTable * new_node_table(pe_node_t *node) { GHashTable *table = pcmk__strkey_table(NULL, free); node = pe__copy_node(node); g_hash_table_insert(table, (gpointer) node->details->id, node); return table; } /*! * \internal * \brief Apply a resource's parent's colocation scores to a node table * * \param[in] rsc Resource whose colocations should be applied * \param[in,out] nodes Node table to apply colocations to */ static void apply_parent_colocations(const pe_resource_t *rsc, GHashTable **nodes) { GList *iter = NULL; pcmk__colocation_t *colocation = NULL; + pe_resource_t *other = NULL; + float factor = 0.0; /* Because the this_with_colocations() and with_this_colocations() methods * boil down to copies of rsc_cons and rsc_cons_lhs for clones and bundles, * we can use those here directly for efficiency. */ for (iter = rsc->parent->rsc_cons; iter != NULL; iter = iter->next) { colocation = (pcmk__colocation_t *) iter->data; - pcmk__add_colocated_node_scores(colocation->primary, rsc->id, nodes, - colocation->node_attribute, - colocation->score / (float) INFINITY, - pcmk__coloc_select_default); + other = colocation->primary; + factor = colocation->score / (float) INFINITY, + other->cmds->add_colocated_node_scores(other, rsc->id, nodes, + colocation->node_attribute, + factor, + pcmk__coloc_select_default); } for (iter = rsc->parent->rsc_cons_lhs; iter != NULL; iter = iter->next) { colocation = (pcmk__colocation_t *) iter->data; if (!pcmk__colocation_has_influence(colocation, rsc)) { continue; } - pcmk__add_colocated_node_scores(colocation->dependent, rsc->id, nodes, - colocation->node_attribute, - colocation->score / (float) INFINITY, - pcmk__coloc_select_nonnegative); + other = colocation->dependent; + factor = colocation->score / (float) INFINITY, + other->cmds->add_colocated_node_scores(other, rsc->id, nodes, + colocation->node_attribute, + factor, + pcmk__coloc_select_nonnegative); } } /*! * \internal * \brief Compare clone or bundle instances based on colocation scores * * Determine the relative order in which two clone or bundle instances should be * assigned to nodes, considering the scores of colocation constraints directly * or indirectly involving them. * * \param[in] instance1 First instance to compare * \param[in] instance2 Second instance to compare * * \return A negative number if \p instance1 should be assigned first, * a positive number if \p instance2 should be assigned first, * or 0 if assignment order doesn't matter */ static int cmp_instance_by_colocation(const pe_resource_t *instance1, const pe_resource_t *instance2) { int rc = 0; pe_node_t *node1 = NULL; pe_node_t *node2 = NULL; pe_node_t *current_node1 = pe__current_node(instance1); pe_node_t *current_node2 = pe__current_node(instance2); GHashTable *colocated_scores1 = NULL; GHashTable *colocated_scores2 = NULL; CRM_ASSERT((instance1 != NULL) && (instance1->parent != NULL) && (instance2 != NULL) && (instance2->parent != NULL) && (current_node1 != NULL) && (current_node2 != NULL)); // Create node tables initialized with each node colocated_scores1 = new_node_table(current_node1); colocated_scores2 = new_node_table(current_node2); // Apply parental colocations apply_parent_colocations(instance1, &colocated_scores1); apply_parent_colocations(instance2, &colocated_scores2); // Find original nodes again, with scores updated for colocations node1 = g_hash_table_lookup(colocated_scores1, current_node1->details->id); node2 = g_hash_table_lookup(colocated_scores2, current_node2->details->id); // Compare nodes by updated scores if (node1->weight < node2->weight) { crm_trace("Assign %s (%d on %s) after %s (%d on %s)", instance1->id, node1->weight, pe__node_name(node1), instance2->id, node2->weight, pe__node_name(node2)); rc = 1; } else if (node1->weight > node2->weight) { crm_trace("Assign %s (%d on %s) before %s (%d on %s)", instance1->id, node1->weight, pe__node_name(node1), instance2->id, node2->weight, pe__node_name(node2)); rc = -1; } g_hash_table_destroy(colocated_scores1); g_hash_table_destroy(colocated_scores2); return rc; } /*! * \internal * \brief Check whether a resource or any of its children are failed * * \param[in] rsc Resource to check * * \return true if \p rsc or any of its children are failed, otherwise false */ static bool did_fail(const pe_resource_t *rsc) { if (pcmk_is_set(rsc->flags, pe_rsc_failed)) { return true; } for (GList *iter = rsc->children; iter != NULL; iter = iter->next) { if (did_fail((const pe_resource_t *) iter->data)) { return true; } } return false; } /*! * \internal * \brief Check whether a node is allowed to run a resource * * \param[in] rsc Resource to check * \param[in,out] node Node to check (will be set NULL if not allowed) * * \return true if *node is either NULL or allowed for \p rsc, otherwise false */ static bool node_is_allowed(const pe_resource_t *rsc, pe_node_t **node) { if (*node != NULL) { pe_node_t *allowed = pe_hash_table_lookup(rsc->allowed_nodes, (*node)->details->id); if ((allowed == NULL) || (allowed->weight < 0)) { pe_rsc_trace(rsc, "%s: current location (%s) is unavailable", rsc->id, pe__node_name(*node)); *node = NULL; return false; } } return true; } /*! * \internal * \brief Compare two clone or bundle instances' instance numbers * * \param[in] a First instance to compare * \param[in] b Second instance to compare * * \return A negative number if \p a's instance number is lower, * a positive number if \p b's instance number is lower, * or 0 if their instance numbers are the same */ gint pcmk__cmp_instance_number(gconstpointer a, gconstpointer b) { const pe_resource_t *instance1 = (const pe_resource_t *) a; const pe_resource_t *instance2 = (const pe_resource_t *) b; char *div1 = NULL; char *div2 = NULL; CRM_ASSERT((instance1 != NULL) && (instance2 != NULL)); // Clone numbers are after a colon, bundle numbers after a dash div1 = strrchr(instance1->id, ':'); if (div1 == NULL) { div1 = strrchr(instance1->id, '-'); } div2 = strrchr(instance2->id, ':'); if (div2 == NULL) { div2 = strrchr(instance2->id, '-'); } CRM_ASSERT((div1 != NULL) && (div2 != NULL)); return (gint) (strtol(div1 + 1, NULL, 10) - strtol(div2 + 1, NULL, 10)); } /*! * \internal * \brief Compare clone or bundle instances according to assignment order * * Compare two clone or bundle instances according to the order they should be * assigned to nodes, preferring (in order): * * - Active instance that is less multiply active * - Instance that is not active on a disallowed node * - Instance with higher configured priority * - Active instance whose current node can run resources * - Active instance whose parent is allowed on current node * - Active instance whose current node has fewer other instances * - Active instance * - Instance that isn't failed * - Instance whose colocations result in higher score on current node * - Instance with lower ID in lexicographic order * * \param[in] a First instance to compare * \param[in] b Second instance to compare * * \return A negative number if \p a should be assigned first, * a positive number if \p b should be assigned first, * or 0 if assignment order doesn't matter */ gint pcmk__cmp_instance(gconstpointer a, gconstpointer b) { int rc = 0; pe_node_t *node1 = NULL; pe_node_t *node2 = NULL; unsigned int nnodes1 = 0; unsigned int nnodes2 = 0; bool can1 = true; bool can2 = true; const pe_resource_t *instance1 = (const pe_resource_t *) a; const pe_resource_t *instance2 = (const pe_resource_t *) b; CRM_ASSERT((instance1 != NULL) && (instance2 != NULL)); node1 = instance1->fns->active_node(instance1, &nnodes1, NULL); node2 = instance2->fns->active_node(instance2, &nnodes2, NULL); /* If both instances are running and at least one is multiply * active, prefer instance that's running on fewer nodes. */ if ((nnodes1 > 0) && (nnodes2 > 0)) { if (nnodes1 < nnodes2) { crm_trace("Assign %s (active on %d) before %s (active on %d): " "less multiply active", instance1->id, nnodes1, instance2->id, nnodes2); return -1; } else if (nnodes1 > nnodes2) { crm_trace("Assign %s (active on %d) after %s (active on %d): " "more multiply active", instance1->id, nnodes1, instance2->id, nnodes2); return 1; } } /* An instance that is either inactive or active on an allowed node is * preferred over an instance that is active on a no-longer-allowed node. */ can1 = node_is_allowed(instance1, &node1); can2 = node_is_allowed(instance2, &node2); if (can1 && !can2) { crm_trace("Assign %s before %s: not active on a disallowed node", instance1->id, instance2->id); return -1; } else if (!can1 && can2) { crm_trace("Assign %s after %s: active on a disallowed node", instance1->id, instance2->id); return 1; } // Prefer instance with higher configured priority if (instance1->priority > instance2->priority) { crm_trace("Assign %s before %s: priority (%d > %d)", instance1->id, instance2->id, instance1->priority, instance2->priority); return -1; } else if (instance1->priority < instance2->priority) { crm_trace("Assign %s after %s: priority (%d < %d)", instance1->id, instance2->id, instance1->priority, instance2->priority); return 1; } // Prefer active instance if ((node1 == NULL) && (node2 == NULL)) { crm_trace("No assignment preference for %s vs. %s: inactive", instance1->id, instance2->id); return 0; } else if (node1 == NULL) { crm_trace("Assign %s after %s: active", instance1->id, instance2->id); return 1; } else if (node2 == NULL) { crm_trace("Assign %s before %s: active", instance1->id, instance2->id); return -1; } // Prefer instance whose current node can run resources can1 = pcmk__node_available(node1, false, false); can2 = pcmk__node_available(node2, false, false); if (can1 && !can2) { crm_trace("Assign %s before %s: current node can run resources", instance1->id, instance2->id); return -1; } else if (!can1 && can2) { crm_trace("Assign %s after %s: current node can't run resources", instance1->id, instance2->id); return 1; } // Prefer instance whose parent is allowed to run on instance's current node node1 = pcmk__top_allowed_node(instance1, node1); node2 = pcmk__top_allowed_node(instance2, node2); if ((node1 == NULL) && (node2 == NULL)) { crm_trace("No assignment preference for %s vs. %s: " "parent not allowed on either instance's current node", instance1->id, instance2->id); return 0; } else if (node1 == NULL) { crm_trace("Assign %s after %s: parent not allowed on current node", instance1->id, instance2->id); return 1; } else if (node2 == NULL) { crm_trace("Assign %s before %s: parent allowed on current node", instance1->id, instance2->id); return -1; } // Prefer instance whose current node is running fewer other instances if (node1->count < node2->count) { crm_trace("Assign %s before %s: fewer active instances on current node", instance1->id, instance2->id); return -1; } else if (node1->count > node2->count) { crm_trace("Assign %s after %s: more active instances on current node", instance1->id, instance2->id); return 1; } // Prefer instance that isn't failed can1 = did_fail(instance1); can2 = did_fail(instance2); if (!can1 && can2) { crm_trace("Assign %s before %s: not failed", instance1->id, instance2->id); return -1; } else if (can1 && !can2) { crm_trace("Assign %s after %s: failed", instance1->id, instance2->id); return 1; } // Prefer instance with higher cumulative colocation score on current node rc = cmp_instance_by_colocation(instance1, instance2); if (rc != 0) { return rc; } // Prefer instance with lower instance number rc = pcmk__cmp_instance_number(instance1, instance2); if (rc < 0) { crm_trace("Assign %s before %s: instance number", instance1->id, instance2->id); } else if (rc > 0) { crm_trace("Assign %s after %s: instance number", instance1->id, instance2->id); } else { crm_trace("No assignment preference for %s vs. %s", instance1->id, instance2->id); } return rc; } /*! * \internal * \brief Choose a node for an instance * * \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 true if \p instance could be assigned to a node, otherwise false */ static bool assign_instance(pe_resource_t *instance, const pe_node_t *prefer, int max_per_node) { pe_node_t *chosen = NULL; pe_node_t *allowed = NULL; CRM_ASSERT(instance != NULL); pe_rsc_trace(instance, "Assigning %s (preferring %s)", instance->id, ((prefer == NULL)? "no node" : prefer->details->uname)); if (!pcmk_is_set(instance->flags, pe_rsc_provisional)) { // Instance is already assigned return instance->fns->location(instance, NULL, FALSE) != NULL; } if (pcmk_is_set(instance->flags, pe_rsc_allocating)) { pe_rsc_debug(instance, "Assignment loop detected involving %s colocations", instance->id); return false; } if (prefer != NULL) { // Possible early assignment to preferred node // Get preferred node with instance's scores allowed = g_hash_table_lookup(instance->allowed_nodes, prefer->details->id); if ((allowed == NULL) || (allowed->weight < 0)) { pe_rsc_trace(instance, "Not assigning %s to preferred node %s: unavailable", instance->id, pe__node_name(prefer)); return false; } } ban_unavailable_allowed_nodes(instance, max_per_node); if (prefer == NULL) { // Final assignment chosen = instance->cmds->assign(instance, NULL); } else { // Possible early assignment to preferred node GHashTable *backup = pcmk__copy_node_table(instance->allowed_nodes); chosen = instance->cmds->assign(instance, prefer); // Revert nodes if preferred node won't be assigned if ((chosen != NULL) && (chosen->details != prefer->details)) { crm_info("Not assigning %s to preferred node %s: %s is better", instance->id, pe__node_name(prefer), pe__node_name(chosen)); g_hash_table_destroy(instance->allowed_nodes); instance->allowed_nodes = backup; pcmk__unassign_resource(instance); chosen = NULL; } else if (backup != NULL) { g_hash_table_destroy(backup); } } // The parent tracks how many instances have been assigned to each node if (chosen != NULL) { allowed = pcmk__top_allowed_node(instance, chosen); 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, pe_rsc_managed)); } else { allowed->count++; } } return chosen != NULL; } /*! * \internal * \brief Reset the node counts of a resource's allowed nodes to zero * * \param[in,out] rsc Resource to reset * * \return Number of nodes that are available to run resources */ static unsigned int reset_allowed_node_counts(pe_resource_t *rsc) { unsigned int available_nodes = 0; pe_node_t *node = NULL; GHashTableIter iter; g_hash_table_iter_init(&iter, rsc->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (gpointer *) &node)) { node->count = 0; if (pcmk__node_available(node, false, false)) { available_nodes++; } } return available_nodes; } /*! * \internal * \brief Check whether an instance has a preferred node * * \param[in] rsc Clone or bundle being assigned (for logs only) * \param[in] instance Clone instance or bundle replica container * \param[in] optimal_per_node Optimal number of instances per node * * \return Instance's current node if still available, otherwise NULL */ static const pe_node_t * preferred_node(const pe_resource_t *rsc, const pe_resource_t *instance, int optimal_per_node) { const pe_node_t *node = NULL; const pe_node_t *parent_node = NULL; // Check whether instance is active, healthy, and not yet assigned if ((instance->running_on == NULL) || !pcmk_is_set(instance->flags, pe_rsc_provisional) || pcmk_is_set(instance->flags, pe_rsc_failed)) { return NULL; } // Check whether instance's current node can run resources node = pe__current_node(instance); if (!pcmk__node_available(node, true, false)) { pe_rsc_trace(rsc, "Not assigning %s to %s early (unavailable)", instance->id, pe__node_name(node)); return NULL; } // Check whether node already has optimal number of instances assigned parent_node = pcmk__top_allowed_node(instance, node); if ((parent_node != NULL) && (parent_node->count >= optimal_per_node)) { pe_rsc_trace(rsc, "Not assigning %s to %s early " "(optimal instances already assigned)", instance->id, pe__node_name(node)); return NULL; } return node; } /*! * \internal * \brief Assign collective instances to nodes * * \param[in,out] collective Clone or bundle resource being assigned * \param[in,out] instances List of clone instances or bundle containers * \param[in] max_total Maximum instances to assign in total * \param[in] max_per_node Maximum instances to assign to any one node */ void pcmk__assign_instances(pe_resource_t *collective, GList *instances, int max_total, int max_per_node) { // Reuse node count to track number of assigned instances unsigned int available_nodes = reset_allowed_node_counts(collective); int optimal_per_node = 0; int assigned = 0; GList *iter = NULL; pe_resource_t *instance = NULL; const pe_node_t *current = NULL; if (available_nodes > 0) { optimal_per_node = max_total / available_nodes; } if (optimal_per_node < 1) { optimal_per_node = 1; } pe_rsc_debug(collective, "Assigning up to %d %s instance%s to up to %u node%s " "(at most %d per host, %d optimal)", max_total, collective->id, pcmk__plural_s(max_total), available_nodes, pcmk__plural_s(available_nodes), max_per_node, optimal_per_node); // Assign as many instances as possible to their current location for (iter = instances; (iter != NULL) && (assigned < max_total); iter = iter->next) { instance = (pe_resource_t *) iter->data; current = preferred_node(collective, instance, optimal_per_node); if ((current != NULL) && assign_instance(instance, current, max_per_node)) { pe_rsc_trace(collective, "Assigned %s to current node %s", instance->id, pe__node_name(current)); assigned++; } } pe_rsc_trace(collective, "Assigned %d of %d instance%s to current node", assigned, max_total, pcmk__plural_s(max_total)); for (iter = instances; iter != NULL; iter = iter->next) { instance = (pe_resource_t *) iter->data; if (!pcmk_is_set(instance->flags, pe_rsc_provisional)) { continue; // Already assigned } if (instance->running_on != NULL) { current = pe__current_node(instance); if (pcmk__top_allowed_node(instance, current) == NULL) { const char *unmanaged = ""; if (!pcmk_is_set(instance->flags, pe_rsc_managed)) { unmanaged = "Unmanaged resource "; } crm_notice("%s%s is running on %s which is no longer allowed", unmanaged, instance->id, pe__node_name(current)); } } if (assigned >= max_total) { pe_rsc_debug(collective, "Not assigning %s because maximum %d instances " "already assigned", instance->id, max_total); resource_location(instance, NULL, -INFINITY, "collective_limit_reached", collective->cluster); } else if (assign_instance(instance, NULL, max_per_node)) { assigned++; } } pe_rsc_debug(collective, "Assigned %d of %d possible instance%s of %s", assigned, max_total, pcmk__plural_s(max_total), collective->id); } enum instance_state { instance_starting = (1 << 0), instance_stopping = (1 << 1), /* This indicates that some instance is restarting. It's not the same as * instance_starting|instance_stopping, which would indicate that some * instance is starting, and some instance (not necessarily the same one) is * stopping. */ instance_restarting = (1 << 2), instance_active = (1 << 3), instance_all = instance_starting|instance_stopping |instance_restarting|instance_active, }; /*! * \internal * \brief Check whether an instance is active, starting, and/or stopping * * \param[in] instance Clone instance or bundle replica container * \param[in,out] state Whether any instance is starting, stopping, etc. */ static void check_instance_state(const pe_resource_t *instance, uint32_t *state) { const GList *iter = NULL; uint32_t instance_state = 0; // State of just this instance // No need to check further if all conditions have already been detected if (pcmk_all_flags_set(*state, instance_all)) { return; } // If instance is a collective (a cloned group), check its children instead if (instance->variant > pe_native) { for (iter = instance->children; (iter != NULL) && !pcmk_all_flags_set(*state, instance_all); iter = iter->next) { check_instance_state((const pe_resource_t *) iter->data, state); } return; } // If we get here, instance is a primitive if (instance->running_on != NULL) { instance_state |= instance_active; } // Check each of the instance's actions for runnable start or stop for (iter = instance->actions; (iter != NULL) && !pcmk_all_flags_set(instance_state, instance_starting |instance_stopping); iter = iter->next) { const pe_action_t *action = (const pe_action_t *) iter->data; const bool optional = pcmk_is_set(action->flags, pe_action_optional); if (pcmk__str_eq(RSC_START, action->task, pcmk__str_none)) { if (!optional && pcmk_is_set(action->flags, pe_action_runnable)) { pe_rsc_trace(instance, "Instance is starting due to %s", action->uuid); instance_state |= instance_starting; } else { pe_rsc_trace(instance, "%s doesn't affect %s state (%s)", action->uuid, instance->id, (optional? "optional" : "unrunnable")); } } else if (pcmk__str_eq(RSC_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, pe_action_pseudo|pe_action_runnable)) { pe_rsc_trace(instance, "Instance is stopping due to %s", action->uuid); instance_state |= instance_stopping; } else { pe_rsc_trace(instance, "%s doesn't affect %s state (%s)", action->uuid, instance->id, (optional? "optional" : "unrunnable")); } } } if (pcmk_all_flags_set(instance_state, instance_starting|instance_stopping)) { instance_state |= instance_restarting; } *state |= instance_state; } /*! * \internal * \brief Create actions for collective resource instances * * \param[in,out] collective Clone or bundle resource to create actions for * \param[in,out] instances List of clone instances or bundle containers */ void pcmk__create_instance_actions(pe_resource_t *collective, GList *instances) { uint32_t state = 0; pe_action_t *stop = NULL; pe_action_t *stopped = NULL; pe_action_t *start = NULL; pe_action_t *started = NULL; pe_rsc_trace(collective, "Creating collective instance actions for %s", collective->id); // Create actions for each instance appropriate to its variant for (GList *iter = instances; iter != NULL; iter = iter->next) { pe_resource_t *instance = (pe_resource_t *) iter->data; instance->cmds->create_actions(instance); check_instance_state(instance, &state); } // Create pseudo-actions for rsc start and started start = pe__new_rsc_pseudo_action(collective, RSC_START, !pcmk_is_set(state, instance_starting), true); started = pe__new_rsc_pseudo_action(collective, RSC_STARTED, !pcmk_is_set(state, instance_starting), false); started->priority = INFINITY; if (pcmk_any_flags_set(state, instance_active|instance_starting)) { pe__set_action_flags(started, pe_action_runnable); } // Create pseudo-actions for rsc stop and stopped stop = pe__new_rsc_pseudo_action(collective, RSC_STOP, !pcmk_is_set(state, instance_stopping), true); stopped = pe__new_rsc_pseudo_action(collective, RSC_STOPPED, !pcmk_is_set(state, instance_stopping), true); stopped->priority = INFINITY; if (!pcmk_is_set(state, instance_restarting)) { pe__set_action_flags(stop, pe_action_migrate_runnable); } if (collective->variant == pe_clone) { pe__create_clone_notif_pseudo_ops(collective, start, started, stop, stopped); } } /*! * \internal * \brief Get a list of clone instances or bundle replica containers * * \param[in] rsc Clone or bundle resource * * \return Clone instances if \p rsc is a clone, or a newly created list of * \p rsc's replica containers if \p rsc is a bundle * \note The caller must call free_instance_list() on the result when the list * is no longer needed. */ static inline GList * get_instance_list(const pe_resource_t *rsc) { if (rsc->variant == pe_container) { return pe__bundle_containers(rsc); } else { return rsc->children; } } /*! * \internal * \brief Free any memory created by get_instance_list() * * \param[in] rsc Clone or bundle resource passed to get_instance_list() * \param[in,out] list Return value of get_instance_list() for \p rsc */ static inline void free_instance_list(const pe_resource_t *rsc, GList *list) { if (list != rsc->children) { g_list_free(list); } } /*! * \internal * \brief Check whether an instance is compatible with a role and node * * \param[in] instance Clone instance or bundle replica container * \param[in] node Instance must match this node * \param[in] role If not RSC_ROLE_UNKNOWN, instance must match this role * \param[in] current If true, compare instance's original node and role, * otherwise compare assigned next node and role * * \return true if \p instance is compatible with \p node and \p role, * otherwise false */ bool pcmk__instance_matches(const pe_resource_t *instance, const pe_node_t *node, enum rsc_role_e role, bool current) { pe_node_t *instance_node = NULL; CRM_CHECK((instance != NULL) && (node != NULL), return false); if ((role != RSC_ROLE_UNKNOWN) && (role != instance->fns->state(instance, current))) { pe_rsc_trace(instance, "%s is not a compatible instance (role is not %s)", instance->id, role2text(role)); return false; } if (!is_set_recursive(instance, pe_rsc_block, true)) { // We only want instances that haven't failed instance_node = instance->fns->location(instance, NULL, current); } if (instance_node == NULL) { pe_rsc_trace(instance, "%s is not a compatible instance (not assigned to a node)", instance->id); return false; } if (instance_node->details != node->details) { pe_rsc_trace(instance, "%s is not a compatible instance (assigned to %s not %s)", instance->id, pe__node_name(instance_node), pe__node_name(node)); return false; } return true; } /*! * \internal * \brief Find an instance that matches a given resource by node and role * * \param[in] match_rsc Resource that instance must match (for logging only) * \param[in] rsc Clone or bundle resource to check for matching instance * \param[in] node Instance must match this node * \param[in] role If not RSC_ROLE_UNKNOWN, instance must match this role * \param[in] current If true, compare instance's original node and role, * otherwise compare assigned next node and role * * \return \p rsc instance matching \p node and \p role if any, otherwise NULL */ static pe_resource_t * find_compatible_instance_on_node(const pe_resource_t *match_rsc, const pe_resource_t *rsc, const pe_node_t *node, enum rsc_role_e role, bool current) { GList *instances = NULL; instances = get_instance_list(rsc); for (GList *iter = instances; iter != NULL; iter = iter->next) { pe_resource_t *instance = (pe_resource_t *) iter->data; if (pcmk__instance_matches(instance, node, role, current)) { pe_rsc_trace(match_rsc, "Found %s %s instance %s compatible with %s on %s", role == RSC_ROLE_UNKNOWN? "matching" : role2text(role), rsc->id, instance->id, match_rsc->id, pe__node_name(node)); free_instance_list(rsc, instances); // Only frees list, not contents return instance; } } free_instance_list(rsc, instances); pe_rsc_trace(match_rsc, "No %s %s instance found compatible with %s on %s", ((role == RSC_ROLE_UNKNOWN)? "matching" : role2text(role)), rsc->id, match_rsc->id, pe__node_name(node)); return NULL; } /*! * \internal * \brief Find a clone instance or bundle container compatible with a resource * * \param[in] match_rsc Resource that instance must match * \param[in] rsc Clone or bundle resource to check for matching instance * \param[in] role If not RSC_ROLE_UNKNOWN, instance must match this role * \param[in] current If true, compare instance's original node and role, * otherwise compare assigned next node and role * * \return Compatible (by \p role and \p match_rsc location) instance of \p rsc * if any, otherwise NULL */ pe_resource_t * pcmk__find_compatible_instance(const pe_resource_t *match_rsc, const pe_resource_t *rsc, enum rsc_role_e role, bool current) { pe_resource_t *instance = NULL; GList *nodes = NULL; const pe_node_t *node = match_rsc->fns->location(match_rsc, NULL, current); // If match_rsc has a node, check only that node if (node != NULL) { return find_compatible_instance_on_node(match_rsc, rsc, node, role, current); } // Otherwise check for an instance matching any of match_rsc's allowed nodes nodes = pcmk__sort_nodes(g_hash_table_get_values(match_rsc->allowed_nodes), NULL); for (GList *iter = nodes; (iter != NULL) && (instance == NULL); iter = iter->next) { instance = find_compatible_instance_on_node(match_rsc, rsc, (pe_node_t *) iter->data, role, current); } if (instance == NULL) { pe_rsc_debug(rsc, "No %s instance found compatible with %s", rsc->id, match_rsc->id); } g_list_free(nodes); return instance; } /*! * \internal * \brief Unassign an instance if mandatory ordering has no interleave match * * \param[in] first 'First' action in an ordering * \param[in] then 'Then' action in an ordering * \param[in,out] then_instance 'Then' instance that has no interleave match * \param[in] type Group of enum pe_ordering flags to apply * \param[in] current If true, "then" action is stopped or demoted * * \return true if \p then_instance was unassigned, otherwise false */ static bool unassign_if_mandatory(const pe_action_t *first, const pe_action_t *then, pe_resource_t *then_instance, uint32_t type, bool current) { // Allow "then" instance to go down even without an interleave match if (current) { pe_rsc_trace(then->rsc, "%s has no instance to order before stopping " "or demoting %s", first->rsc->id, then_instance->id); /* If the "first" action must be runnable, but there is no "first" * instance, the "then" instance must not be allowed to come up. */ } else if (pcmk_any_flags_set(type, pe_order_runnable_left |pe_order_implies_then)) { pe_rsc_info(then->rsc, "Inhibiting %s from being active " "because there is no %s instance to interleave", then_instance->id, first->rsc->id); return pcmk__assign_resource(then_instance, NULL, true); } return false; } /*! * \internal * \brief Find first matching action for a clone instance or bundle container * * \param[in] action Action in an interleaved ordering * \param[in] instance Clone instance or bundle container being interleaved * \param[in] action_name Action to look for * \param[in] node If not NULL, require action to be on this node * \param[in] for_first If true, \p instance is the 'first' resource in the * ordering, otherwise it is the 'then' resource * * \return First action for \p instance (or in some cases if \p instance is a * bundle container, its containerized resource) that matches * \p action_name and \p node if any, otherwise NULL */ static pe_action_t * find_instance_action(const pe_action_t *action, const pe_resource_t *instance, const char *action_name, const pe_node_t *node, bool for_first) { const pe_resource_t *rsc = NULL; pe_action_t *matching_action = NULL; /* If instance is a bundle container, sometimes we should interleave the * action for the container itself, and sometimes for the containerized * resource. * * For example, given "start bundle A then bundle B", B likely requires the * service inside A's container to be active, rather than just the * container, so we should interleave the action for A's containerized * resource. On the other hand, it's possible B's container itself requires * something from A, so we should interleave the action for B's container. * * Essentially, for 'first', we should use the containerized resource for * everything except stop, and for 'then', we should use the container for * everything except promote and demote (which can only be performed on the * containerized resource). */ if ((for_first && !pcmk__str_any_of(action->task, CRMD_ACTION_STOP, CRMD_ACTION_STOPPED, NULL)) || (!for_first && pcmk__str_any_of(action->task, CRMD_ACTION_PROMOTE, CRMD_ACTION_PROMOTED, CRMD_ACTION_DEMOTE, CRMD_ACTION_DEMOTED, NULL))) { rsc = pcmk__get_rsc_in_container(instance); } if (rsc == NULL) { rsc = instance; // No containerized resource, use instance itself } else { node = NULL; // Containerized actions are on bundle-created guest } matching_action = find_first_action(rsc->actions, NULL, action_name, node); if (matching_action != NULL) { return matching_action; } if (pcmk_is_set(instance->flags, pe_rsc_orphan) || pcmk__str_any_of(action_name, RSC_STOP, RSC_DEMOTE, NULL)) { crm_trace("No %s action found for %s%s", action_name, pcmk_is_set(instance->flags, pe_rsc_orphan)? "orphan " : "", instance->id); } else { crm_err("No %s action found for %s to interleave (bug?)", action_name, instance->id); } return NULL; } /*! * \internal * \brief Get the original action name of a bundle or clone action * * Given an action for a bundle or clone, get the original action name, * mapping notify to the action being notified, and if the instances are * primitives, mapping completion actions to the action that was completed * (for example, stopped to stop). * * \param[in] action Clone or bundle action to check * * \return Original action name for \p action */ static const char * orig_action_name(const pe_action_t *action) { const pe_resource_t *instance = action->rsc->children->data; // Any instance char *action_type = NULL; const char *action_name = action->task; enum action_tasks orig_task = no_action; if (pcmk__strcase_any_of(action->task, CRMD_ACTION_NOTIFY, CRMD_ACTION_NOTIFIED, NULL)) { // action->uuid is RSC_(confirmed-){pre,post}_notify_ACTION_INTERVAL CRM_CHECK(parse_op_key(action->uuid, NULL, &action_type, NULL), return task2text(no_action)); action_name = strstr(action_type, "_notify_"); CRM_CHECK(action_name != NULL, return task2text(no_action)); action_name += strlen("_notify_"); } orig_task = get_complex_task(instance, action_name); free(action_type); return task2text(orig_task); } /*! * \internal * \brief Update two interleaved actions according to an ordering between them * * Given information about an ordering of two interleaved actions, update the * actions' flags (and runnable_before members if appropriate) as appropriate * for the ordering. Effects may cascade to other orderings involving the * actions as well. * * \param[in,out] first 'First' action in an ordering * \param[in,out] then 'Then' action in an ordering * \param[in] node If not NULL, limit scope of ordering to this node * \param[in] filter Action flags to limit scope of certain updates (may * include pe_action_optional to affect only mandatory * actions, and pe_action_runnable to affect only * runnable actions) * \param[in] type Group of enum pe_ordering flags to apply * * \return Group of enum pcmk__updated flags indicating what was updated */ static uint32_t update_interleaved_actions(pe_action_t *first, pe_action_t *then, const pe_node_t *node, uint32_t filter, uint32_t type) { GList *instances = NULL; uint32_t changed = pcmk__updated_none; const char *orig_first_task = orig_action_name(first); // Stops and demotes must be interleaved with instance on current node bool current = pcmk__ends_with(first->uuid, "_" CRMD_ACTION_STOPPED "_0") || pcmk__ends_with(first->uuid, "_" CRMD_ACTION_DEMOTED "_0"); // Update the specified actions for each "then" instance individually instances = get_instance_list(then->rsc); for (GList *iter = instances; iter != NULL; iter = iter->next) { pe_resource_t *first_instance = NULL; pe_resource_t *then_instance = iter->data; pe_action_t *first_action = NULL; pe_action_t *then_action = NULL; // Find a "first" instance to interleave with this "then" instance first_instance = pcmk__find_compatible_instance(then_instance, first->rsc, RSC_ROLE_UNKNOWN, current); if (first_instance == NULL) { // No instance can be interleaved if (unassign_if_mandatory(first, then, then_instance, type, current)) { pcmk__set_updated_flags(changed, first, pcmk__updated_then); } continue; } first_action = find_instance_action(first, first_instance, orig_first_task, node, true); if (first_action == NULL) { continue; } then_action = find_instance_action(then, then_instance, then->task, node, false); if (then_action == NULL) { continue; } if (order_actions(first_action, then_action, type)) { pcmk__set_updated_flags(changed, first, pcmk__updated_first|pcmk__updated_then); } changed |= then_instance->cmds->update_ordered_actions( first_action, then_action, node, first_instance->cmds->action_flags(first_action, node), filter, type, then->rsc->cluster); } free_instance_list(then->rsc, instances); return changed; } /*! * \internal * \brief Check whether two actions in an ordering can be interleaved * * \param[in] first 'First' action in the ordering * \param[in] then 'Then' action in the ordering * * \return true if \p first and \p then can be interleaved, otherwise false */ static bool can_interleave_actions(const pe_action_t *first, const pe_action_t *then) { bool interleave = false; pe_resource_t *rsc = NULL; if ((first->rsc == NULL) || (then->rsc == NULL)) { crm_trace("Not interleaving %s with %s: not resource actions", first->uuid, then->uuid); return false; } if (first->rsc == then->rsc) { crm_trace("Not interleaving %s with %s: same resource", first->uuid, then->uuid); return false; } if ((first->rsc->variant < pe_clone) || (then->rsc->variant < pe_clone)) { crm_trace("Not interleaving %s with %s: not clones or bundles", first->uuid, then->uuid); return false; } if (pcmk__ends_with(then->uuid, "_stop_0") || pcmk__ends_with(then->uuid, "_demote_0")) { rsc = first->rsc; } else { rsc = then->rsc; } interleave = crm_is_true(g_hash_table_lookup(rsc->meta, XML_RSC_ATTR_INTERLEAVE)); pe_rsc_trace(rsc, "'%s then %s' will %sbe interleaved (based on %s)", first->uuid, then->uuid, (interleave? "" : "not "), rsc->id); return interleave; } /*! * \internal * \brief Update non-interleaved instance actions according to an ordering * * Given information about an ordering of two non-interleaved actions, update * the actions' flags (and runnable_before members if appropriate) as * appropriate for the ordering. Effects may cascade to other orderings * involving the actions as well. * * \param[in,out] instance Clone instance or bundle container * \param[in,out] first "First" action in ordering * \param[in] then "Then" action in ordering (for \p instance's parent) * \param[in] node If not NULL, limit scope of ordering to this node * \param[in] flags Action flags for \p first for ordering purposes * \param[in] filter Action flags to limit scope of certain updates (may * include pe_action_optional to affect only mandatory * actions, and pe_action_runnable to affect only * runnable actions) * \param[in] type Group of enum pe_ordering flags to apply * * \return Group of enum pcmk__updated flags indicating what was updated */ static uint32_t update_noninterleaved_actions(pe_resource_t *instance, pe_action_t *first, const pe_action_t *then, const pe_node_t *node, uint32_t flags, uint32_t filter, uint32_t type) { pe_action_t *instance_action = NULL; uint32_t instance_flags = 0; uint32_t changed = pcmk__updated_none; // Check whether instance has an equivalent of "then" action instance_action = find_first_action(instance->actions, NULL, then->task, node); if (instance_action == NULL) { return changed; } // Check whether action is runnable instance_flags = instance->cmds->action_flags(instance_action, node); if (!pcmk_is_set(instance_flags, pe_action_runnable)) { return changed; } // If so, update actions for the instance changed = instance->cmds->update_ordered_actions(first, instance_action, node, flags, filter, type, instance->cluster); // Propagate any changes to later actions if (pcmk_is_set(changed, pcmk__updated_then)) { for (GList *after_iter = instance_action->actions_after; after_iter != NULL; after_iter = after_iter->next) { pe_action_wrapper_t *after = after_iter->data; pcmk__update_action_for_orderings(after->action, instance->cluster); } } return changed; } /*! * \internal * \brief Update two actions according to an ordering between them * * Given information about an ordering of two clone or bundle actions, update * the actions' flags (and runnable_before members if appropriate) as * appropriate for the ordering. Effects may cascade to other orderings * involving the actions as well. * * \param[in,out] first 'First' action in an ordering * \param[in,out] then 'Then' action in an ordering * \param[in] node If not NULL, limit scope of ordering to this node * (only used when interleaving instances) * \param[in] flags Action flags for \p first for ordering purposes * \param[in] filter Action flags to limit scope of certain updates (may * include pe_action_optional to affect only mandatory * actions, and pe_action_runnable to affect only * runnable actions) * \param[in] type Group of enum pe_ordering flags to apply * \param[in,out] data_set Cluster working set * * \return Group of enum pcmk__updated flags indicating what was updated */ uint32_t pcmk__instance_update_ordered_actions(pe_action_t *first, pe_action_t *then, const pe_node_t *node, uint32_t flags, uint32_t filter, uint32_t type, pe_working_set_t *data_set) { if (then->rsc == NULL) { return pcmk__updated_none; } else if (can_interleave_actions(first, then)) { return update_interleaved_actions(first, then, node, filter, type); } else { uint32_t changed = pcmk__updated_none; GList *instances = get_instance_list(then->rsc); // Update actions for the clone or bundle resource itself changed |= pcmk__update_ordered_actions(first, then, node, flags, filter, type, data_set); // Update the 'then' clone instances or bundle containers individually for (GList *iter = instances; iter != NULL; iter = iter->next) { pe_resource_t *instance = iter->data; changed |= update_noninterleaved_actions(instance, first, then, node, flags, filter, type); } free_instance_list(then->rsc, instances); return changed; } } #define pe__clear_action_summary_flags(flags, action, flag) do { \ flags = pcmk__clear_flags_as(__func__, __LINE__, LOG_TRACE, \ "Action summary", action->rsc->id, \ flags, flag, #flag); \ } while (0) /*! * \internal * \brief Return action flags for a given clone or bundle action * * \param[in,out] action Action for a clone or bundle * \param[in] instances Clone instances or bundle containers * \param[in] node If not NULL, limit effects to this node * * \return Flags appropriate to \p action on \p node */ enum pe_action_flags pcmk__collective_action_flags(pe_action_t *action, const GList *instances, const pe_node_t *node) { bool any_runnable = false; enum pe_action_flags flags; const char *action_name = orig_action_name(action); // Set original assumptions (optional and runnable may be cleared below) flags = pe_action_optional|pe_action_runnable|pe_action_pseudo; for (const GList *iter = instances; iter != NULL; iter = iter->next) { const pe_resource_t *instance = iter->data; const pe_node_t *instance_node = NULL; pe_action_t *instance_action = NULL; enum pe_action_flags instance_flags; // Node is relevant only to primitive instances if (instance->variant == pe_native) { instance_node = node; } instance_action = find_first_action(instance->actions, NULL, action_name, instance_node); if (instance_action == NULL) { pe_rsc_trace(action->rsc, "%s has no %s action on %s", instance->id, action_name, pe__node_name(node)); continue; } pe_rsc_trace(action->rsc, "%s has %s for %s on %s", instance->id, instance_action->uuid, action_name, pe__node_name(node)); instance_flags = instance->cmds->action_flags(instance_action, node); // If any instance action is mandatory, so is the collective action if (pcmk_is_set(flags, pe_action_optional) && !pcmk_is_set(instance_flags, pe_action_optional)) { pe_rsc_trace(instance, "%s is mandatory because %s is", action->uuid, instance_action->uuid); pe__clear_action_summary_flags(flags, action, pe_action_optional); pe__clear_action_flags(action, pe_action_optional); } // If any instance action is runnable, so is the collective action if (pcmk_is_set(instance_flags, pe_action_runnable)) { any_runnable = true; } } if (!any_runnable) { pe_rsc_trace(action->rsc, "%s is not runnable because no instance can run %s", action->uuid, action_name); pe__clear_action_summary_flags(flags, action, pe_action_runnable); if (node == NULL) { pe__clear_action_flags(action, pe_action_runnable); } } return flags; } /*! * \internal * \brief Add a collective resource's colocations to a list for an instance * * \param[in,out] list Colocation list to add to * \param[in] instance Clone or bundle instance or instance group member * \param[in] collective Clone or bundle resource with colocations to add * \param[in] with_this If true, add collective's "with this" colocations, * otherwise add its "this with" colocations */ void pcmk__add_collective_constraints(GList **list, const pe_resource_t *instance, const pe_resource_t *collective, bool with_this) { const GList *colocations = NULL; bool everywhere = false; CRM_CHECK((list != NULL) && (instance != NULL), return); if (collective == NULL) { return; } switch (collective->variant) { case pe_clone: case pe_container: break; default: return; } everywhere = can_run_everywhere(collective); if (with_this) { colocations = collective->rsc_cons_lhs; } else { colocations = collective->rsc_cons; } for (const GList *iter = colocations; iter != NULL; iter = iter->next) { const pcmk__colocation_t *colocation = iter->data; if (with_this && !pcmk__colocation_has_influence(colocation, instance)) { continue; } if (!everywhere || (colocation->score < 0) || (!with_this && (colocation->score == INFINITY))) { if (with_this) { pcmk__add_with_this(list, colocation); } else { pcmk__add_this_with(list, colocation); } } } } diff --git a/lib/pacemaker/pcmk_sched_primitive.c b/lib/pacemaker/pcmk_sched_primitive.c index e6d5b92d25..d3b0f92bf0 100644 --- a/lib/pacemaker/pcmk_sched_primitive.c +++ b/lib/pacemaker/pcmk_sched_primitive.c @@ -1,1564 +1,1564 @@ /* * Copyright 2004-2023 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU General Public License version 2 * or later (GPLv2+) WITHOUT ANY WARRANTY. */ #include #include #include #include #include "libpacemaker_private.h" static void stop_resource(pe_resource_t *rsc, pe_node_t *node, bool optional); static void start_resource(pe_resource_t *rsc, pe_node_t *node, bool optional); static void demote_resource(pe_resource_t *rsc, pe_node_t *node, bool optional); static void promote_resource(pe_resource_t *rsc, pe_node_t *node, bool optional); static void assert_role_error(pe_resource_t *rsc, pe_node_t *node, bool optional); 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 */ { RSC_ROLE_UNKNOWN, /* Unknown */ RSC_ROLE_STOPPED, /* Stopped */ RSC_ROLE_STOPPED, /* Started */ RSC_ROLE_STOPPED, /* Unpromoted */ RSC_ROLE_STOPPED, /* Promoted */ }, /* Stopped */ { RSC_ROLE_STOPPED, /* Unknown */ RSC_ROLE_STOPPED, /* Stopped */ RSC_ROLE_STARTED, /* Started */ RSC_ROLE_UNPROMOTED, /* Unpromoted */ RSC_ROLE_UNPROMOTED, /* Promoted */ }, /* Started */ { RSC_ROLE_STOPPED, /* Unknown */ RSC_ROLE_STOPPED, /* Stopped */ RSC_ROLE_STARTED, /* Started */ RSC_ROLE_UNPROMOTED, /* Unpromoted */ RSC_ROLE_PROMOTED, /* Promoted */ }, /* Unpromoted */ { RSC_ROLE_STOPPED, /* Unknown */ RSC_ROLE_STOPPED, /* Stopped */ RSC_ROLE_STOPPED, /* Started */ RSC_ROLE_UNPROMOTED, /* Unpromoted */ RSC_ROLE_PROMOTED, /* Promoted */ }, /* Promoted */ { RSC_ROLE_STOPPED, /* Unknown */ RSC_ROLE_UNPROMOTED, /* Stopped */ RSC_ROLE_UNPROMOTED, /* Started */ RSC_ROLE_UNPROMOTED, /* Unpromoted */ RSC_ROLE_PROMOTED, /* Promoted */ }, }; /*! * \internal * \brief Function to schedule actions needed for a role change * * \param[in,out] rsc Resource whose role is changing * \param[in,out] node Node where resource will be in its next role * \param[in] optional Whether scheduled actions should be optional */ typedef void (*rsc_transition_fn)(pe_resource_t *rsc, pe_node_t *node, bool optional); static rsc_transition_fn rsc_action_matrix[RSC_ROLE_MAX][RSC_ROLE_MAX] = { /* This array lists the function needed to transition directly from one role * to another. NULL indicates that nothing is needed. * * Current role Transition function Next role * ------------ ------------------- ---------- */ /* Unknown */ { assert_role_error, /* Unknown */ stop_resource, /* Stopped */ assert_role_error, /* Started */ assert_role_error, /* Unpromoted */ assert_role_error, /* Promoted */ }, /* Stopped */ { assert_role_error, /* Unknown */ NULL, /* Stopped */ start_resource, /* Started */ start_resource, /* Unpromoted */ assert_role_error, /* Promoted */ }, /* Started */ { assert_role_error, /* Unknown */ stop_resource, /* Stopped */ NULL, /* Started */ NULL, /* Unpromoted */ promote_resource, /* Promoted */ }, /* Unpromoted */ { assert_role_error, /* Unknown */ stop_resource, /* Stopped */ stop_resource, /* Started */ NULL, /* Unpromoted */ promote_resource, /* Promoted */ }, /* Promoted */ { assert_role_error, /* Unknown */ demote_resource, /* Stopped */ demote_resource, /* Started */ demote_resource, /* Unpromoted */ NULL, /* Promoted */ }, }; /*! * \internal * \brief Get a list of a resource's allowed nodes sorted by node weight * * \param[in] rsc Resource to check * * \return List of allowed nodes sorted by node weight */ static GList * sorted_allowed_nodes(const pe_resource_t *rsc) { if (rsc->allowed_nodes != NULL) { GList *nodes = g_hash_table_get_values(rsc->allowed_nodes); if (nodes != NULL) { return pcmk__sort_nodes(nodes, pe__current_node(rsc)); } } return NULL; } /*! * \internal * \brief Assign a resource to its best allowed node, if possible * * \param[in,out] rsc Resource to choose a node for * \param[in] prefer If not NULL, prefer this node when all else equal * * \return true if \p rsc could be assigned to a node, otherwise false */ static bool assign_best_node(pe_resource_t *rsc, const pe_node_t *prefer) { GList *nodes = NULL; pe_node_t *chosen = NULL; pe_node_t *best = NULL; bool result = false; const pe_node_t *most_free_node = pcmk__ban_insufficient_capacity(rsc); if (prefer == NULL) { prefer = most_free_node; } if (!pcmk_is_set(rsc->flags, pe_rsc_provisional)) { // We've already finished assignment of resources to nodes return rsc->allocated_to != NULL; } // Sort allowed nodes by weight nodes = sorted_allowed_nodes(rsc); if (nodes != NULL) { best = (pe_node_t *) nodes->data; // First node has best score } if ((prefer != NULL) && (nodes != NULL)) { // Get the allowed node version of prefer chosen = g_hash_table_lookup(rsc->allowed_nodes, prefer->details->id); if (chosen == NULL) { pe_rsc_trace(rsc, "Preferred node %s for %s was unknown", pe__node_name(prefer), rsc->id); /* Favor the preferred node as long as its weight 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 weight is less than INFINITY. */ } else if (chosen->weight < best->weight) { pe_rsc_trace(rsc, "Preferred node %s for %s was unsuitable", pe__node_name(chosen), rsc->id); chosen = NULL; } else if (!pcmk__node_available(chosen, true, false)) { pe_rsc_trace(rsc, "Preferred node %s for %s was unavailable", pe__node_name(chosen), rsc->id); chosen = NULL; } else { pe_rsc_trace(rsc, "Chose preferred node %s for %s (ignoring %d candidates)", pe__node_name(chosen), rsc->id, g_list_length(nodes)); } } if ((chosen == NULL) && (best != NULL)) { /* Either there is no preferred node, or the preferred node is not * suitable, but another node is allowed to run the resource. */ chosen = best; if (!pe_rsc_is_unique_clone(rsc->parent) && (chosen->weight > 0) // Zero not acceptable && pcmk__node_available(chosen, false, false)) { /* If the resource is already running on a node, prefer that node if * it is just as good as the chosen node. * * We don't do this for unique clone instances, because * pcmk__assign_instances() has already assigned instances to their * running nodes when appropriate, and if we get here, we don't want * remaining unassigned instances to prefer a node that's already * running another instance. */ pe_node_t *running = pe__current_node(rsc); if (running == NULL) { // Nothing to do } else if (!pcmk__node_available(running, true, false)) { pe_rsc_trace(rsc, "Current node for %s (%s) can't run resources", rsc->id, pe__node_name(running)); } else { int nodes_with_best_score = 1; for (GList *iter = nodes->next; iter; iter = iter->next) { pe_node_t *allowed = (pe_node_t *) iter->data; if (allowed->weight != chosen->weight) { // The nodes are sorted by weight, so no more are equal break; } if (pe__same_node(allowed, running)) { // Scores are equal, so prefer the current node chosen = allowed; } nodes_with_best_score++; } if (nodes_with_best_score > 1) { do_crm_log(((chosen->weight >= INFINITY)? LOG_WARNING : LOG_INFO), "Chose %s for %s from %d nodes with score %s", pe__node_name(chosen), rsc->id, nodes_with_best_score, pcmk_readable_score(chosen->weight)); } } } pe_rsc_trace(rsc, "Chose %s for %s from %d candidates", pe__node_name(chosen), rsc->id, g_list_length(nodes)); } result = pcmk__finalize_assignment(rsc, chosen, false); g_list_free(nodes); return result; } /*! * \internal * \brief Apply a "this with" colocation to a node's allowed node scores * * \param[in,out] data Colocation to apply * \param[in,out] user_data Resource being assigned */ static void apply_this_with(gpointer data, gpointer user_data) { pcmk__colocation_t *colocation = (pcmk__colocation_t *) data; pe_resource_t *rsc = (pe_resource_t *) user_data; GHashTable *archive = NULL; pe_resource_t *other = colocation->primary; // In certain cases, we will need to revert the node scores if ((colocation->dependent_role >= RSC_ROLE_PROMOTED) || ((colocation->score < 0) && (colocation->score > -INFINITY))) { archive = pcmk__copy_node_table(rsc->allowed_nodes); } if (pcmk_is_set(other->flags, pe_rsc_provisional)) { pe_rsc_trace(rsc, "%s: Assigning colocation %s primary %s first" "(score=%d role=%s)", rsc->id, colocation->id, other->id, colocation->score, role2text(colocation->dependent_role)); other->cmds->assign(other, NULL); } // Apply the colocation score to this resource's allowed node scores rsc->cmds->apply_coloc_score(rsc, other, colocation, true); if ((archive != NULL) && !pcmk__any_node_available(rsc->allowed_nodes)) { pe_rsc_info(rsc, "%s: Reverting scores from colocation with %s " "because no nodes allowed", rsc->id, other->id); g_hash_table_destroy(rsc->allowed_nodes); rsc->allowed_nodes = archive; archive = NULL; } if (archive != NULL) { g_hash_table_destroy(archive); } } /*! * \internal * \brief Apply a "with this" colocation to a node's allowed node scores * * \param[in,out] data Colocation to apply * \param[in,out] user_data Resource being assigned */ static void apply_with_this(void *data, void *user_data) { pcmk__colocation_t *colocation = (pcmk__colocation_t *) data; pe_resource_t *rsc = (pe_resource_t *) user_data; pe_resource_t *other = colocation->dependent; const float factor = colocation->score / (float) INFINITY; if (!pcmk__colocation_has_influence(colocation, NULL)) { return; } pe_rsc_trace(rsc, "%s: Incorporating attenuated %s assignment scores due " "to colocation %s", rsc->id, other->id, colocation->id); - pcmk__add_colocated_node_scores(other, rsc->id, &rsc->allowed_nodes, - colocation->node_attribute, factor, - pcmk__coloc_select_active); + other->cmds->add_colocated_node_scores(other, rsc->id, &rsc->allowed_nodes, + colocation->node_attribute, factor, + pcmk__coloc_select_active); } /*! * \internal * \brief Update a Pacemaker Remote node once its connection has been assigned * * \param[in] connection Connection resource that has been assigned */ static void remote_connection_assigned(const pe_resource_t *connection) { pe_node_t *remote_node = pe_find_node(connection->cluster->nodes, connection->id); CRM_CHECK(remote_node != NULL, return); if ((connection->allocated_to != NULL) && (connection->next_role != RSC_ROLE_STOPPED)) { crm_trace("Pacemaker Remote node %s will be online", remote_node->details->id); remote_node->details->online = TRUE; if (remote_node->details->unseen) { // Avoid unnecessary fence, since we will attempt connection remote_node->details->unclean = FALSE; } } else { crm_trace("Pacemaker Remote node %s will be shut down " "(%sassigned connection's next role is %s)", remote_node->details->id, ((connection->allocated_to == NULL)? "un" : ""), role2text(connection->next_role)); remote_node->details->shutdown = TRUE; } } /*! * \internal * \brief Assign a primitive resource to a node * * \param[in,out] rsc Resource to assign to a node * \param[in] prefer Node to prefer, if all else is equal * * \return Node that \p rsc is assigned to, if assigned entirely to one node */ pe_node_t * pcmk__primitive_assign(pe_resource_t *rsc, const pe_node_t *prefer) { GList *colocations = NULL; CRM_ASSERT(rsc != NULL); // Never assign a child without parent being assigned first if ((rsc->parent != NULL) && !pcmk_is_set(rsc->parent->flags, pe_rsc_allocating)) { pe_rsc_debug(rsc, "%s: Assigning parent %s first", rsc->id, rsc->parent->id); rsc->parent->cmds->assign(rsc->parent, prefer); } if (!pcmk_is_set(rsc->flags, pe_rsc_provisional)) { return rsc->allocated_to; // Assignment has already been done } // Ensure we detect assignment loops if (pcmk_is_set(rsc->flags, pe_rsc_allocating)) { pe_rsc_debug(rsc, "Breaking assignment loop involving %s", rsc->id); return NULL; } pe__set_resource_flags(rsc, pe_rsc_allocating); pe__show_node_weights(true, rsc, "Pre-assignment", rsc->allowed_nodes, rsc->cluster); colocations = pcmk__this_with_colocations(rsc); g_list_foreach(colocations, apply_this_with, rsc); g_list_free(colocations); pe__show_node_weights(true, rsc, "Post-this-with", rsc->allowed_nodes, rsc->cluster); colocations = pcmk__with_this_colocations(rsc); g_list_foreach(colocations, apply_with_this, rsc); g_list_free(colocations); if (rsc->next_role == RSC_ROLE_STOPPED) { pe_rsc_trace(rsc, "Banning %s from all nodes because it will be stopped", rsc->id); resource_location(rsc, NULL, -INFINITY, XML_RSC_ATTR_TARGET_ROLE, rsc->cluster); } else if ((rsc->next_role > rsc->role) && !pcmk_is_set(rsc->cluster->flags, pe_flag_have_quorum) && (rsc->cluster->no_quorum_policy == no_quorum_freeze)) { crm_notice("Resource %s cannot be elevated from %s to %s due to " "no-quorum-policy=freeze", rsc->id, role2text(rsc->role), role2text(rsc->next_role)); pe__set_next_role(rsc, rsc->role, "no-quorum-policy=freeze"); } pe__show_node_weights(!pcmk_is_set(rsc->cluster->flags, pe_flag_show_scores), rsc, __func__, rsc->allowed_nodes, rsc->cluster); // Unmanage resource if fencing is enabled but no device is configured if (pcmk_is_set(rsc->cluster->flags, pe_flag_stonith_enabled) && !pcmk_is_set(rsc->cluster->flags, pe_flag_have_stonith_resource)) { pe__clear_resource_flags(rsc, pe_rsc_managed); } if (!pcmk_is_set(rsc->flags, pe_rsc_managed)) { // Unmanaged resources stay on their current node const char *reason = NULL; pe_node_t *assign_to = NULL; pe__set_next_role(rsc, rsc->role, "unmanaged"); assign_to = pe__current_node(rsc); if (assign_to == NULL) { reason = "inactive"; } else if (rsc->role == RSC_ROLE_PROMOTED) { reason = "promoted"; } else if (pcmk_is_set(rsc->flags, pe_rsc_failed)) { reason = "failed"; } else { reason = "active"; } pe_rsc_info(rsc, "Unmanaged resource %s assigned to %s: %s", rsc->id, (assign_to? assign_to->details->uname : "no node"), reason); pcmk__finalize_assignment(rsc, assign_to, true); } else if (pcmk_is_set(rsc->cluster->flags, pe_flag_stop_everything)) { pe_rsc_debug(rsc, "Forcing %s to stop: stop-all-resources", rsc->id); pcmk__finalize_assignment(rsc, NULL, true); } else if (pcmk_is_set(rsc->flags, pe_rsc_provisional) && assign_best_node(rsc, prefer)) { // Assignment successful } else if (rsc->allocated_to == NULL) { if (!pcmk_is_set(rsc->flags, pe_rsc_orphan)) { pe_rsc_info(rsc, "Resource %s cannot run anywhere", rsc->id); } else if (rsc->running_on != NULL) { pe_rsc_info(rsc, "Stopping orphan resource %s", rsc->id); } } else { pe_rsc_debug(rsc, "%s: pre-assigned to %s", rsc->id, pe__node_name(rsc->allocated_to)); } pe__clear_resource_flags(rsc, pe_rsc_allocating); if (rsc->is_remote_node) { remote_connection_assigned(rsc); } return rsc->allocated_to; } /*! * \internal * \brief Schedule actions to bring resource down and back to current role * * \param[in,out] rsc Resource to restart * \param[in,out] current Node that resource should be brought down on * \param[in] need_stop Whether the resource must be stopped * \param[in] need_promote Whether the resource must be promoted * * \return Role that resource would have after scheduled actions are taken */ static void schedule_restart_actions(pe_resource_t *rsc, pe_node_t *current, bool need_stop, bool need_promote) { enum rsc_role_e role = rsc->role; enum rsc_role_e next_role; rsc_transition_fn fn = NULL; pe__set_resource_flags(rsc, pe_rsc_restarting); // Bring resource down to a stop on its current node while (role != RSC_ROLE_STOPPED) { next_role = rsc_state_matrix[role][RSC_ROLE_STOPPED]; pe_rsc_trace(rsc, "Creating %s action to take %s down from %s to %s", (need_stop? "required" : "optional"), rsc->id, role2text(role), role2text(next_role)); fn = rsc_action_matrix[role][next_role]; if (fn == NULL) { break; } fn(rsc, current, !need_stop); role = next_role; } // Bring resource up to its next role on its next node while ((rsc->role <= rsc->next_role) && (role != rsc->role) && !pcmk_is_set(rsc->flags, pe_rsc_block)) { bool required = need_stop; next_role = rsc_state_matrix[role][rsc->role]; if ((next_role == RSC_ROLE_PROMOTED) && need_promote) { required = true; } pe_rsc_trace(rsc, "Creating %s action to take %s up from %s to %s", (required? "required" : "optional"), rsc->id, role2text(role), role2text(next_role)); fn = rsc_action_matrix[role][next_role]; if (fn == NULL) { break; } fn(rsc, rsc->allocated_to, !required); role = next_role; } pe__clear_resource_flags(rsc, pe_rsc_restarting); } /*! * \internal * \brief If a resource's next role is not explicitly specified, set a default * * \param[in,out] rsc Resource to set next role for * * \return "explicit" if next role was explicitly set, otherwise "implicit" */ static const char * set_default_next_role(pe_resource_t *rsc) { if (rsc->next_role != RSC_ROLE_UNKNOWN) { return "explicit"; } if (rsc->allocated_to == NULL) { pe__set_next_role(rsc, RSC_ROLE_STOPPED, "assignment"); } else { pe__set_next_role(rsc, RSC_ROLE_STARTED, "assignment"); } return "implicit"; } /*! * \internal * \brief Create an action to represent an already pending start * * \param[in,out] rsc Resource to create start action for */ static void create_pending_start(pe_resource_t *rsc) { pe_action_t *start = NULL; pe_rsc_trace(rsc, "Creating action for %s to represent already pending start", rsc->id); start = start_action(rsc, rsc->allocated_to, TRUE); pe__set_action_flags(start, pe_action_print_always); } /*! * \internal * \brief Schedule actions needed to take a resource to its next role * * \param[in,out] rsc Resource to schedule actions for */ static void schedule_role_transition_actions(pe_resource_t *rsc) { enum rsc_role_e role = rsc->role; while (role != rsc->next_role) { enum rsc_role_e next_role = rsc_state_matrix[role][rsc->next_role]; rsc_transition_fn fn = NULL; pe_rsc_trace(rsc, "Creating action to take %s from %s to %s (ending at %s)", rsc->id, role2text(role), role2text(next_role), role2text(rsc->next_role)); fn = rsc_action_matrix[role][next_role]; if (fn == NULL) { break; } fn(rsc, rsc->allocated_to, false); role = next_role; } } /*! * \internal * \brief Create all actions needed for a given primitive resource * * \param[in,out] rsc Primitive resource to create actions for */ void pcmk__primitive_create_actions(pe_resource_t *rsc) { bool need_stop = false; bool need_promote = false; bool is_moving = false; bool allow_migrate = false; bool multiply_active = false; pe_node_t *current = NULL; unsigned int num_all_active = 0; unsigned int num_clean_active = 0; const char *next_role_source = NULL; CRM_ASSERT(rsc != NULL); next_role_source = set_default_next_role(rsc); pe_rsc_trace(rsc, "Creating all actions for %s transition from %s to %s " "(%s) on %s", rsc->id, role2text(rsc->role), role2text(rsc->next_role), next_role_source, pe__node_name(rsc->allocated_to)); current = rsc->fns->active_node(rsc, &num_all_active, &num_clean_active); g_list_foreach(rsc->dangling_migrations, pcmk__abort_dangling_migration, rsc); if ((current != NULL) && (rsc->allocated_to != NULL) && (current->details != rsc->allocated_to->details) && (rsc->next_role >= RSC_ROLE_STARTED)) { pe_rsc_trace(rsc, "Moving %s from %s to %s", rsc->id, pe__node_name(current), pe__node_name(rsc->allocated_to)); is_moving = true; allow_migrate = pcmk__rsc_can_migrate(rsc, current); // This is needed even if migrating (though I'm not sure why ...) need_stop = true; } // Check whether resource is partially migrated and/or multiply active if ((rsc->partial_migration_source != NULL) && (rsc->partial_migration_target != NULL) && allow_migrate && (num_all_active == 2) && pe__same_node(current, rsc->partial_migration_source) && pe__same_node(rsc->allocated_to, rsc->partial_migration_target)) { /* A partial migration is in progress, and the migration target remains * the same as when the migration began. */ pe_rsc_trace(rsc, "Partial migration of %s from %s to %s will continue", rsc->id, pe__node_name(rsc->partial_migration_source), pe__node_name(rsc->partial_migration_target)); } else if ((rsc->partial_migration_source != NULL) || (rsc->partial_migration_target != NULL)) { // A partial migration is in progress but can't be continued if (num_all_active > 2) { // The resource is migrating *and* multiply active! crm_notice("Forcing recovery of %s because it is migrating " "from %s to %s and possibly active elsewhere", rsc->id, pe__node_name(rsc->partial_migration_source), pe__node_name(rsc->partial_migration_target)); } else { // The migration source or target isn't available crm_notice("Forcing recovery of %s because it can no longer " "migrate from %s to %s", rsc->id, pe__node_name(rsc->partial_migration_source), pe__node_name(rsc->partial_migration_target)); } need_stop = true; rsc->partial_migration_source = rsc->partial_migration_target = NULL; allow_migrate = false; } else if (pcmk_is_set(rsc->flags, pe_rsc_needs_fencing)) { multiply_active = (num_all_active > 1); } else { /* If a resource has "requires" set to nothing or quorum, don't consider * it active on unclean nodes (similar to how all resources behave when * stonith-enabled is false). We can start such resources elsewhere * before fencing completes, and if we considered the resource active on * the failed node, we would attempt recovery for being active on * multiple nodes. */ multiply_active = (num_clean_active > 1); } if (multiply_active) { const char *class = crm_element_value(rsc->xml, XML_AGENT_ATTR_CLASS); // Resource was (possibly) incorrectly multiply active pe_proc_err("%s resource %s might be active on %u nodes (%s)", pcmk__s(class, "Untyped"), rsc->id, num_all_active, recovery2text(rsc->recovery_type)); crm_notice("See https://wiki.clusterlabs.org/wiki/FAQ" "#Resource_is_Too_Active for more information"); switch (rsc->recovery_type) { case recovery_stop_start: need_stop = true; break; case recovery_stop_unexpected: need_stop = true; // stop_resource() will skip expected node pe__set_resource_flags(rsc, pe_rsc_stop_unexpected); break; default: break; } } else { pe__clear_resource_flags(rsc, pe_rsc_stop_unexpected); } if (pcmk_is_set(rsc->flags, pe_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, pe_rsc_failed)) { if (pcmk_is_set(rsc->flags, pe_rsc_stop)) { need_stop = true; pe_rsc_trace(rsc, "Recovering %s", rsc->id); } else { pe_rsc_trace(rsc, "Recovering %s by demotion", rsc->id); if (rsc->next_role == RSC_ROLE_PROMOTED) { need_promote = true; } } } else if (pcmk_is_set(rsc->flags, pe_rsc_block)) { pe_rsc_trace(rsc, "Blocking further actions on %s", rsc->id); need_stop = true; } else if ((rsc->role > RSC_ROLE_STARTED) && (current != NULL) && (rsc->allocated_to != NULL)) { pe_action_t *start = NULL; pe_rsc_trace(rsc, "Creating start action for promoted resource %s", rsc->id); start = start_action(rsc, rsc->allocated_to, TRUE); if (!pcmk_is_set(start->flags, pe_action_optional)) { // Recovery of a promoted resource pe_rsc_trace(rsc, "%s restart is required for recovery", rsc->id); need_stop = true; } } // Create any actions needed to bring resource down and back up to same role schedule_restart_actions(rsc, current, need_stop, need_promote); // Create any actions needed to take resource from this role to the next schedule_role_transition_actions(rsc); pcmk__create_recurring_actions(rsc); if (allow_migrate) { pcmk__create_migration_actions(rsc, current); } } /*! * \internal * \brief Ban a resource from any allowed nodes that are Pacemaker Remote nodes * * \param[in] rsc Resource to check */ static void rsc_avoids_remote_nodes(const pe_resource_t *rsc) { GHashTableIter iter; pe_node_t *node = NULL; g_hash_table_iter_init(&iter, rsc->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (void **) &node)) { if (node->details->remote_rsc != NULL) { node->weight = -INFINITY; } } } /*! * \internal * \brief Return allowed nodes as (possibly sorted) list * * Convert a resource's hash table of allowed nodes to a list. If printing to * stdout, sort the list, to keep action ID numbers consistent for regression * test output (while avoiding the performance hit on a live cluster). * * \param[in] rsc Resource to check for allowed nodes * * \return List of resource's allowed nodes * \note Callers should take care not to rely on the list being sorted. */ static GList * allowed_nodes_as_list(const pe_resource_t *rsc) { GList *allowed_nodes = NULL; if (rsc->allowed_nodes) { allowed_nodes = g_hash_table_get_values(rsc->allowed_nodes); } if (!pcmk__is_daemon) { allowed_nodes = g_list_sort(allowed_nodes, pe__cmp_node_name); } return allowed_nodes; } /*! * \internal * \brief Create implicit constraints needed for a primitive resource * * \param[in,out] rsc Primitive resource to create implicit constraints for */ void pcmk__primitive_internal_constraints(pe_resource_t *rsc) { GList *allowed_nodes = NULL; bool check_unfencing = false; bool check_utilization = false; CRM_ASSERT(rsc != NULL); if (!pcmk_is_set(rsc->flags, pe_rsc_managed)) { pe_rsc_trace(rsc, "Skipping implicit constraints for unmanaged resource %s", rsc->id); return; } // Whether resource requires unfencing check_unfencing = !pcmk_is_set(rsc->flags, pe_rsc_fence_device) && pcmk_is_set(rsc->cluster->flags, pe_flag_enable_unfencing) && pcmk_is_set(rsc->flags, pe_rsc_needs_unfencing); // Whether a non-default placement strategy is used check_utilization = (g_hash_table_size(rsc->utilization) > 0) && !pcmk__str_eq(rsc->cluster->placement_strategy, "default", pcmk__str_casei); // Order stops before starts (i.e. restart) pcmk__new_ordering(rsc, pcmk__op_key(rsc->id, RSC_STOP, 0), NULL, rsc, pcmk__op_key(rsc->id, RSC_START, 0), NULL, pe_order_optional|pe_order_implies_then|pe_order_restart, rsc->cluster); // Promotable ordering: demote before stop, start before promote if (pcmk_is_set(pe__const_top_resource(rsc, false)->flags, pe_rsc_promotable) || (rsc->role > RSC_ROLE_UNPROMOTED)) { pcmk__new_ordering(rsc, pcmk__op_key(rsc->id, RSC_DEMOTE, 0), NULL, rsc, pcmk__op_key(rsc->id, RSC_STOP, 0), NULL, pe_order_promoted_implies_first, rsc->cluster); pcmk__new_ordering(rsc, pcmk__op_key(rsc->id, RSC_START, 0), NULL, rsc, pcmk__op_key(rsc->id, RSC_PROMOTE, 0), NULL, pe_order_runnable_left, rsc->cluster); } // Don't clear resource history if probing on same node pcmk__new_ordering(rsc, pcmk__op_key(rsc->id, CRM_OP_LRM_DELETE, 0), NULL, rsc, pcmk__op_key(rsc->id, RSC_STATUS, 0), NULL, pe_order_same_node|pe_order_then_cancels_first, rsc->cluster); // Certain checks need allowed nodes if (check_unfencing || check_utilization || (rsc->container != NULL)) { allowed_nodes = allowed_nodes_as_list(rsc); } if (check_unfencing) { g_list_foreach(allowed_nodes, pcmk__order_restart_vs_unfence, rsc); } if (check_utilization) { pcmk__create_utilization_constraints(rsc, allowed_nodes); } if (rsc->container != NULL) { pe_resource_t *remote_rsc = NULL; if (rsc->is_remote_node) { // rsc is the implicit remote connection for a guest or bundle node /* Guest resources are not allowed to run on Pacemaker Remote nodes, * to avoid nesting remotes. However, bundles are allowed. */ if (!pcmk_is_set(rsc->flags, pe_rsc_allow_remote_remotes)) { rsc_avoids_remote_nodes(rsc->container); } /* If someone cleans up a guest or bundle node's container, we will * likely schedule a (re-)probe of the container and recovery of the * connection. Order the connection stop after the container probe, * so that if we detect the container running, we will trigger a new * transition and avoid the unnecessary recovery. */ pcmk__order_resource_actions(rsc->container, RSC_STATUS, rsc, RSC_STOP, pe_order_optional); /* A user can specify that a resource must start on a Pacemaker Remote * node by explicitly configuring it with the container=NODENAME * meta-attribute. This is of questionable merit, since location * constraints can accomplish the same thing. But we support it, so here * we check whether a resource (that is not itself a remote connection) * has container set to a remote node or guest node resource. */ } else if (rsc->container->is_remote_node) { remote_rsc = rsc->container; } else { remote_rsc = pe__resource_contains_guest_node(rsc->cluster, rsc->container); } if (remote_rsc != NULL) { /* Force the resource on the Pacemaker Remote node instead of * colocating the resource with the container resource. */ for (GList *item = allowed_nodes; item; item = item->next) { pe_node_t *node = item->data; if (node->details->remote_rsc != remote_rsc) { node->weight = -INFINITY; } } } else { /* This resource is either a filler for a container that does NOT * represent a Pacemaker Remote node, or a Pacemaker Remote * connection resource for a guest node or bundle. */ int score; crm_trace("Order and colocate %s relative to its container %s", rsc->id, rsc->container->id); pcmk__new_ordering(rsc->container, pcmk__op_key(rsc->container->id, RSC_START, 0), NULL, rsc, pcmk__op_key(rsc->id, RSC_START, 0), NULL, pe_order_implies_then|pe_order_runnable_left, rsc->cluster); pcmk__new_ordering(rsc, pcmk__op_key(rsc->id, RSC_STOP, 0), NULL, rsc->container, pcmk__op_key(rsc->container->id, RSC_STOP, 0), NULL, pe_order_implies_first, rsc->cluster); if (pcmk_is_set(rsc->flags, pe_rsc_allow_remote_remotes)) { score = 10000; /* Highly preferred but not essential */ } else { score = INFINITY; /* Force them to run on the same host */ } pcmk__new_colocation("resource-with-container", NULL, score, rsc, rsc->container, NULL, NULL, true, rsc->cluster); } } if (rsc->is_remote_node || pcmk_is_set(rsc->flags, pe_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 weights or resource priority * * Given a colocation constraint, apply its score to the dependent's * allowed node weights (if we are still placing resources) or priority (if * we are choosing promotable clone instance roles). * * \param[in,out] dependent Dependent resource in colocation * \param[in] primary Primary resource in colocation * \param[in] colocation Colocation constraint to apply * \param[in] for_dependent true if called on behalf of dependent */ void pcmk__primitive_apply_coloc_score(pe_resource_t *dependent, const pe_resource_t *primary, const pcmk__colocation_t *colocation, bool for_dependent) { enum pcmk__coloc_affects filter_results; CRM_CHECK((colocation != NULL) && (dependent != NULL) && (primary != NULL), return); if (for_dependent) { // Always process on behalf of primary resource primary->cmds->apply_coloc_score(dependent, primary, colocation, false); return; } filter_results = pcmk__colocation_affects(dependent, primary, colocation, false); pe_rsc_trace(dependent, "%s %s with %s (%s, score=%d, filter=%d)", ((colocation->score > 0)? "Colocating" : "Anti-colocating"), dependent->id, primary->id, colocation->id, colocation->score, filter_results); switch (filter_results) { case pcmk__coloc_affects_role: pcmk__apply_coloc_to_priority(dependent, primary, colocation); break; case pcmk__coloc_affects_location: pcmk__apply_coloc_to_weights(dependent, primary, colocation); break; default: // pcmk__coloc_affects_nothing return; } } /* Primitive implementation of * resource_alloc_functions_t:with_this_colocations() */ void pcmk__with_primitive_colocations(const pe_resource_t *rsc, const pe_resource_t *orig_rsc, GList **list) { // Primitives don't have children, so rsc should also be orig_rsc CRM_CHECK((rsc != NULL) && (rsc->variant == pe_native) && (rsc == orig_rsc) && (list != NULL), return); // Add primitive's own colocations plus any relevant ones from parent pcmk__add_with_this_list(list, rsc->rsc_cons_lhs); if (rsc->parent != NULL) { rsc->parent->cmds->with_this_colocations(rsc->parent, rsc, list); } } /* Primitive implementation of * resource_alloc_functions_t:this_with_colocations() */ void pcmk__primitive_with_colocations(const pe_resource_t *rsc, const pe_resource_t *orig_rsc, GList **list) { // Primitives don't have children, so rsc should also be orig_rsc CRM_CHECK((rsc != NULL) && (rsc->variant == pe_native) && (rsc == orig_rsc) && (list != NULL), return); // Add primitive's own colocations plus any relevant ones from parent pcmk__add_this_with_list(list, rsc->rsc_cons); if (rsc->parent != NULL) { rsc->parent->cmds->this_with_colocations(rsc->parent, rsc, list); } } /*! * \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 */ enum pe_action_flags pcmk__primitive_action_flags(pe_action_t *action, const pe_node_t *node) { CRM_ASSERT(action != NULL); return action->flags; } /*! * \internal * \brief Check whether a node is a multiply active resource's expected node * * \param[in] rsc Resource to check * \param[in] node Node to check * * \return true if \p rsc is multiply active with multiple-active set to * stop_unexpected, and \p node is the node where it will remain active * \note This assumes that the resource's next role cannot be changed to stopped * after this is called, which should be reasonable if status has already * been unpacked and resources have been assigned to nodes. */ static bool is_expected_node(const pe_resource_t *rsc, const pe_node_t *node) { return pcmk_all_flags_set(rsc->flags, pe_rsc_stop_unexpected|pe_rsc_restarting) && (rsc->next_role > RSC_ROLE_STOPPED) && pe__same_node(rsc->allocated_to, node); } /*! * \internal * \brief Schedule actions needed to stop a resource wherever it is active * * \param[in,out] rsc Resource being stopped * \param[in] node Node where resource is being stopped (ignored) * \param[in] optional Whether actions should be optional */ static void stop_resource(pe_resource_t *rsc, pe_node_t *node, bool optional) { for (GList *iter = rsc->running_on; iter != NULL; iter = iter->next) { pe_node_t *current = (pe_node_t *) iter->data; pe_action_t *stop = NULL; if (is_expected_node(rsc, current)) { /* We are scheduling restart actions for a multiply active resource * with multiple-active=stop_unexpected, and this is where it should * not be stopped. */ pe_rsc_trace(rsc, "Skipping stop of multiply active resource %s " "on expected node %s", rsc->id, pe__node_name(current)); continue; } if (rsc->partial_migration_target != NULL) { // Continue migration if node originally was and remains target if (pe__same_node(current, rsc->partial_migration_target) && pe__same_node(current, rsc->allocated_to)) { pe_rsc_trace(rsc, "Skipping stop of %s on %s " "because partial migration there will continue", rsc->id, pe__node_name(current)); continue; } else { pe_rsc_trace(rsc, "Forcing stop of %s on %s " "because migration target changed", rsc->id, pe__node_name(current)); optional = false; } } pe_rsc_trace(rsc, "Scheduling stop of %s on %s", rsc->id, pe__node_name(current)); stop = stop_action(rsc, current, optional); if (rsc->allocated_to == NULL) { pe_action_set_reason(stop, "node availability", true); } else if (pcmk_all_flags_set(rsc->flags, pe_rsc_restarting |pe_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, pe_rsc_managed)) { pe__clear_action_flags(stop, pe_action_runnable); } if (pcmk_is_set(rsc->cluster->flags, pe_flag_remove_after_stop)) { pcmk__schedule_cleanup(rsc, current, optional); } if (pcmk_is_set(rsc->flags, pe_rsc_needs_unfencing)) { pe_action_t *unfence = pe_fence_op(current, "on", true, NULL, false, rsc->cluster); order_actions(stop, unfence, pe_order_implies_first); if (!pcmk__node_unfenced(current)) { pe_proc_err("Stopping %s until %s can be unfenced", rsc->id, pe__node_name(current)); } } } } /*! * \internal * \brief Schedule actions needed to start a resource on a node * * \param[in,out] rsc Resource being started * \param[in,out] node Node where resource should be started * \param[in] optional Whether actions should be optional */ static void start_resource(pe_resource_t *rsc, pe_node_t *node, bool optional) { pe_action_t *start = NULL; CRM_ASSERT(node != NULL); pe_rsc_trace(rsc, "Scheduling %s start of %s on %s (score %d)", (optional? "optional" : "required"), rsc->id, pe__node_name(node), node->weight); start = start_action(rsc, node, TRUE); pcmk__order_vs_unfence(rsc, node, start, pe_order_implies_then); if (pcmk_is_set(start->flags, pe_action_runnable) && !optional) { pe__clear_action_flags(start, pe_action_optional); } if (is_expected_node(rsc, node)) { /* This could be a problem if the start becomes necessary for other * reasons later. */ pe_rsc_trace(rsc, "Start of multiply active resouce %s " "on expected node %s will be a pseudo-action", rsc->id, pe__node_name(node)); pe__set_action_flags(start, pe_action_pseudo); } } /*! * \internal * \brief Schedule actions needed to promote a resource on a node * * \param[in,out] rsc Resource being promoted * \param[in] node Node where resource should be promoted * \param[in] optional Whether actions should be optional */ static void promote_resource(pe_resource_t *rsc, pe_node_t *node, bool optional) { GList *iter = NULL; GList *action_list = NULL; bool runnable = true; CRM_ASSERT(node != NULL); // Any start must be runnable for promotion to be runnable action_list = pe__resource_actions(rsc, node, RSC_START, true); for (iter = action_list; iter != NULL; iter = iter->next) { pe_action_t *start = (pe_action_t *) iter->data; if (!pcmk_is_set(start->flags, pe_action_runnable)) { runnable = false; } } g_list_free(action_list); if (runnable) { pe_action_t *promote = promote_action(rsc, node, optional); pe_rsc_trace(rsc, "Scheduling %s promotion of %s on %s", (optional? "optional" : "required"), rsc->id, pe__node_name(node)); if (is_expected_node(rsc, node)) { /* This could be a problem if the promote becomes necessary for * other reasons later. */ pe_rsc_trace(rsc, "Promotion of multiply active resouce %s " "on expected node %s will be a pseudo-action", rsc->id, pe__node_name(node)); pe__set_action_flags(promote, pe_action_pseudo); } } else { pe_rsc_trace(rsc, "Not promoting %s on %s: start unrunnable", rsc->id, pe__node_name(node)); action_list = pe__resource_actions(rsc, node, RSC_PROMOTE, true); for (iter = action_list; iter != NULL; iter = iter->next) { pe_action_t *promote = (pe_action_t *) iter->data; pe__clear_action_flags(promote, pe_action_runnable); } g_list_free(action_list); } } /*! * \internal * \brief Schedule actions needed to demote a resource wherever it is active * * \param[in,out] rsc Resource being demoted * \param[in] node Node where resource should be demoted (ignored) * \param[in] optional Whether actions should be optional */ static void demote_resource(pe_resource_t *rsc, pe_node_t *node, bool optional) { /* Since this will only be called for a primitive (possibly as an instance * of a collective resource), the resource is multiply active if it is * running on more than one node, so we want to demote on all of them as * part of recovery, regardless of which one is the desired node. */ for (GList *iter = rsc->running_on; iter != NULL; iter = iter->next) { pe_node_t *current = (pe_node_t *) iter->data; if (is_expected_node(rsc, current)) { pe_rsc_trace(rsc, "Skipping demote of multiply active resource %s " "on expected node %s", rsc->id, pe__node_name(current)); } else { pe_rsc_trace(rsc, "Scheduling %s demotion of %s on %s", (optional? "optional" : "required"), rsc->id, pe__node_name(current)); demote_action(rsc, current, optional); } } } static void assert_role_error(pe_resource_t *rsc, pe_node_t *node, bool optional) { CRM_ASSERT(false); } /*! * \internal * \brief Schedule cleanup of a resource * * \param[in,out] rsc Resource to clean up * \param[in] node Node to clean up on * \param[in] optional Whether clean-up should be optional */ void pcmk__schedule_cleanup(pe_resource_t *rsc, const pe_node_t *node, bool optional) { /* If the cleanup is required, its orderings are optional, because they're * relevant only if both actions are required. Conversely, if the cleanup is * optional, the orderings make the then action required if the first action * becomes required. */ uint32_t flag = optional? pe_order_implies_then : pe_order_optional; CRM_CHECK((rsc != NULL) && (node != NULL), return); if (pcmk_is_set(rsc->flags, pe_rsc_failed)) { pe_rsc_trace(rsc, "Skipping clean-up of %s on %s: resource failed", rsc->id, pe__node_name(node)); return; } if (node->details->unclean || !node->details->online) { pe_rsc_trace(rsc, "Skipping clean-up of %s on %s: node unavailable", rsc->id, pe__node_name(node)); return; } crm_notice("Scheduling clean-up of %s on %s", rsc->id, pe__node_name(node)); delete_action(rsc, node, optional); // stop -> clean-up -> start pcmk__order_resource_actions(rsc, RSC_STOP, rsc, RSC_DELETE, flag); pcmk__order_resource_actions(rsc, RSC_DELETE, rsc, RSC_START, flag); } /*! * \internal * \brief Add primitive meta-attributes relevant to graph actions to XML * * \param[in] rsc Primitive resource whose meta-attributes should be added * \param[in,out] xml Transition graph action attributes XML to add to */ void pcmk__primitive_add_graph_meta(const pe_resource_t *rsc, xmlNode *xml) { char *name = NULL; char *value = NULL; const pe_resource_t *parent = NULL; CRM_ASSERT((rsc != NULL) && (xml != NULL)); /* Clone instance numbers get set internally as meta-attributes, and are * needed in the transition graph (for example, to tell unique clone * instances apart). */ value = g_hash_table_lookup(rsc->meta, XML_RSC_ATTR_INCARNATION); if (value != NULL) { name = crm_meta_name(XML_RSC_ATTR_INCARNATION); crm_xml_add(xml, name, value); free(name); } // Not sure if this one is really needed ... value = g_hash_table_lookup(rsc->meta, XML_RSC_ATTR_REMOTE_NODE); if (value != NULL) { name = crm_meta_name(XML_RSC_ATTR_REMOTE_NODE); crm_xml_add(xml, name, value); free(name); } /* The container meta-attribute can be set on the primitive itself or one of * its parents (for example, a group inside a container resource), so check * them all, and keep the highest one found. */ for (parent = rsc; parent != NULL; parent = parent->parent) { if (parent->container != NULL) { crm_xml_add(xml, CRM_META "_" XML_RSC_ATTR_CONTAINER, parent->container->id); } } /* Bundle replica children will get their external-ip set internally as a * meta-attribute. The graph action needs it, but under a different naming * convention than other meta-attributes. */ value = g_hash_table_lookup(rsc->meta, "external-ip"); if (value != NULL) { crm_xml_add(xml, "pcmk_external_ip", value); } } // Primitive implementation of resource_alloc_functions_t:add_utilization() void pcmk__primitive_add_utilization(const pe_resource_t *rsc, const pe_resource_t *orig_rsc, GList *all_rscs, GHashTable *utilization) { if (!pcmk_is_set(rsc->flags, pe_rsc_provisional)) { return; } pe_rsc_trace(orig_rsc, "%s: Adding primitive %s as colocated utilization", orig_rsc->id, rsc->id); pcmk__release_node_capacity(utilization, rsc); } /*! * \internal * \brief Get epoch time of node's shutdown attribute (or now if none) * * \param[in,out] node Node to check * * \return Epoch time corresponding to shutdown attribute if set or now if not */ static time_t shutdown_time(pe_node_t *node) { const char *shutdown = pe_node_attribute_raw(node, XML_CIB_ATTR_SHUTDOWN); time_t result = 0; if (shutdown != NULL) { long long result_ll; if (pcmk__scan_ll(shutdown, &result_ll, 0LL) == pcmk_rc_ok) { result = (time_t) result_ll; } } return (result == 0)? get_effective_time(node->details->data_set) : result; } /*! * \internal * \brief Ban a resource from a node if it's not locked to the node * * \param[in] data Node to check * \param[in,out] user_data Resource to check */ static void ban_if_not_locked(gpointer data, gpointer user_data) { const pe_node_t *node = (const pe_node_t *) data; pe_resource_t *rsc = (pe_resource_t *) user_data; if (strcmp(node->details->uname, rsc->lock_node->details->uname) != 0) { resource_location(rsc, node, -CRM_SCORE_INFINITY, XML_CONFIG_ATTR_SHUTDOWN_LOCK, rsc->cluster); } } // Primitive implementation of resource_alloc_functions_t:shutdown_lock() void pcmk__primitive_shutdown_lock(pe_resource_t *rsc) { const char *class = crm_element_value(rsc->xml, XML_AGENT_ATTR_CLASS); // Fence devices and remote connections can't be locked if (pcmk__str_eq(class, PCMK_RESOURCE_CLASS_STONITH, pcmk__str_null_matches) || pe__resource_is_remote_conn(rsc, rsc->cluster)) { return; } if (rsc->lock_node != NULL) { // The lock was obtained from resource history if (rsc->running_on != NULL) { /* The resource was started elsewhere even though it is now * considered locked. This shouldn't be possible, but as a * failsafe, we don't want to disturb the resource now. */ pe_rsc_info(rsc, "Cancelling shutdown lock because %s is already active", rsc->id); pe__clear_resource_history(rsc, rsc->lock_node, rsc->cluster); rsc->lock_node = NULL; rsc->lock_time = 0; } // Only a resource active on exactly one node can be locked } else if (pcmk__list_of_1(rsc->running_on)) { pe_node_t *node = rsc->running_on->data; if (node->details->shutdown) { if (node->details->unclean) { pe_rsc_debug(rsc, "Not locking %s to unclean %s for shutdown", rsc->id, pe__node_name(node)); } else { rsc->lock_node = node; rsc->lock_time = shutdown_time(node); } } } if (rsc->lock_node == NULL) { // No lock needed return; } if (rsc->cluster->shutdown_lock > 0) { time_t lock_expiration = rsc->lock_time + rsc->cluster->shutdown_lock; pe_rsc_info(rsc, "Locking %s to %s due to shutdown (expires @%lld)", rsc->id, pe__node_name(rsc->lock_node), (long long) lock_expiration); pe__update_recheck_time(++lock_expiration, rsc->cluster); } else { pe_rsc_info(rsc, "Locking %s to %s due to shutdown", rsc->id, pe__node_name(rsc->lock_node)); } // If resource is locked to one node, ban it from all other nodes g_list_foreach(rsc->cluster->nodes, ban_if_not_locked, rsc); } diff --git a/lib/pacemaker/pcmk_sched_promotable.c b/lib/pacemaker/pcmk_sched_promotable.c index ed461c0cc2..d12d017bab 100644 --- a/lib/pacemaker/pcmk_sched_promotable.c +++ b/lib/pacemaker/pcmk_sched_promotable.c @@ -1,1282 +1,1286 @@ /* * Copyright 2004-2023 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU General Public License version 2 * or later (GPLv2+) WITHOUT ANY WARRANTY. */ #include #include #include #include "libpacemaker_private.h" /*! * \internal * \brief 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(pe_resource_t *clone, pe_resource_t *child, pe_resource_t *last) { // "Promote clone" -> promote instance -> "clone promoted" pcmk__order_resource_actions(clone, RSC_PROMOTE, child, RSC_PROMOTE, pe_order_optional); pcmk__order_resource_actions(child, RSC_PROMOTE, clone, RSC_PROMOTED, pe_order_optional); // If clone is ordered, order this instance relative to last if ((last != NULL) && pe__clone_is_ordered(clone)) { pcmk__order_resource_actions(last, RSC_PROMOTE, child, RSC_PROMOTE, pe_order_optional); } } /*! * \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(pe_resource_t *clone, pe_resource_t *child, pe_resource_t *last) { // "Demote clone" -> demote instance -> "clone demoted" pcmk__order_resource_actions(clone, RSC_DEMOTE, child, RSC_DEMOTE, pe_order_implies_first_printed); pcmk__order_resource_actions(child, RSC_DEMOTE, clone, RSC_DEMOTED, pe_order_implies_then_printed); // If clone is ordered, order this instance relative to last if ((last != NULL) && pe__clone_is_ordered(clone)) { pcmk__order_resource_actions(child, RSC_DEMOTE, last, RSC_DEMOTE, pe_order_optional); } } /*! * \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 pe_resource_t *rsc, bool *demoting, bool *promoting) { const GList *iter = NULL; // If this is a cloned group, check group members recursively if (rsc->children != NULL) { for (iter = rsc->children; iter != NULL; iter = iter->next) { check_for_role_change((const pe_resource_t *) iter->data, demoting, promoting); } return; } for (iter = rsc->actions; iter != NULL; iter = iter->next) { const pe_action_t *action = (const pe_action_t *) iter->data; if (*promoting && *demoting) { return; } else if (pcmk_is_set(action->flags, pe_action_optional)) { continue; } else if (pcmk__str_eq(RSC_DEMOTE, action->task, pcmk__str_none)) { *demoting = true; } else if (pcmk__str_eq(RSC_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(pe_resource_t *child, const GList *location_constraints, const pe_node_t *chosen) { for (const GList *iter = location_constraints; iter; iter = iter->next) { const pe__location_t *location = iter->data; pe_node_t *weighted_node = NULL; if (location->role_filter == RSC_ROLE_PROMOTED) { weighted_node = pe_find_node_id(location->node_list_rh, chosen->details->id); } if (weighted_node != NULL) { int new_priority = pcmk__add_scores(child->priority, weighted_node->weight); pe_rsc_trace(child, "Applying location %s to %s promotion priority on %s: " "%s + %s = %s", location->id, child->id, pe__node_name(weighted_node), pcmk_readable_score(child->priority), pcmk_readable_score(weighted_node->weight), pcmk_readable_score(new_priority)); child->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 pe_node_t * node_to_be_promoted_on(const pe_resource_t *rsc) { pe_node_t *node = NULL; pe_node_t *local_node = NULL; const pe_resource_t *parent = NULL; // If this is a cloned group, bail if any group member can't be promoted for (GList *iter = rsc->children; iter != NULL; iter = iter->next) { pe_resource_t *child = (pe_resource_t *) iter->data; if (node_to_be_promoted_on(child) == NULL) { pe_rsc_trace(rsc, "%s can't be promoted because member %s can't", rsc->id, child->id); return NULL; } } node = rsc->fns->location(rsc, NULL, FALSE); if (node == NULL) { pe_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, pe_rsc_managed)) { if (rsc->fns->state(rsc, TRUE) == RSC_ROLE_PROMOTED) { crm_notice("Unmanaged instance %s will be left promoted on %s", rsc->id, pe__node_name(node)); } else { pe_rsc_trace(rsc, "%s can't be promoted because it is unmanaged", rsc->id); return NULL; } } else if (rsc->priority < 0) { pe_rsc_trace(rsc, "%s can't be promoted because its promotion priority %d " "is negative", rsc->id, rsc->priority); return NULL; } else if (!pcmk__node_available(node, false, true)) { pe_rsc_trace(rsc, "%s can't be promoted because %s can't run resources", rsc->id, pe__node_name(node)); return NULL; } parent = pe__const_top_resource(rsc, false); local_node = pe_hash_table_lookup(parent->allowed_nodes, node->details->id); if (local_node == NULL) { /* It should not be possible for the scheduler to have allocated 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, pe_rsc_managed)) { crm_warn("%s can't be promoted because %s is not allowed on %s " "(scheduler bug?)", rsc->id, parent->id, pe__node_name(node)); } // else the instance is unmanaged and already promoted return NULL; } else if ((local_node->count >= pe__clone_promoted_node_max(parent)) && pcmk_is_set(rsc->flags, pe_rsc_managed)) { pe_rsc_trace(rsc, "%s can't be promoted because %s has " "maximum promoted instances already", rsc->id, pe__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 pe_resource_t *rsc1 = (const pe_resource_t *) a; const pe_resource_t *rsc2 = (const pe_resource_t *) b; enum rsc_role_e role1 = RSC_ROLE_UNKNOWN; enum rsc_role_e role2 = RSC_ROLE_UNKNOWN; CRM_ASSERT((rsc1 != NULL) && (rsc2 != NULL)); // Check sort index set by pcmk__set_instance_roles() if (rsc1->sort_index > rsc2->sort_index) { pe_rsc_trace(rsc1, "%s has higher promotion priority than %s " "(sort index %d > %d)", rsc1->id, rsc2->id, rsc1->sort_index, rsc2->sort_index); return -1; } else if (rsc1->sort_index < rsc2->sort_index) { pe_rsc_trace(rsc1, "%s has lower promotion priority than %s " "(sort index %d < %d)", rsc1->id, rsc2->id, rsc1->sort_index, rsc2->sort_index); return 1; } // If those are the same, prefer instance whose current role is higher role1 = rsc1->fns->state(rsc1, TRUE); role2 = rsc2->fns->state(rsc2, TRUE); if (role1 > role2) { pe_rsc_trace(rsc1, "%s has higher promotion priority than %s " "(higher current role)", rsc1->id, rsc2->id); return -1; } else if (role1 < role2) { pe_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 a promotable clone instance's sort index to its node's weight * * Add a promotable clone instance's sort index (which sums its promotion * preferences and scores of relevant location constraints for the promoted * role) to the node weight of the instance's allocated node. * * \param[in] data Promotable clone instance * \param[in,out] user_data Clone parent of \p data */ static void add_sort_index_to_node_weight(gpointer data, gpointer user_data) { const pe_resource_t *child = (const pe_resource_t *) data; pe_resource_t *clone = (pe_resource_t *) user_data; pe_node_t *node = NULL; const pe_node_t *chosen = NULL; if (child->sort_index < 0) { pe_rsc_trace(clone, "Not adding sort index of %s: negative", child->id); return; } chosen = child->fns->location(child, NULL, FALSE); if (chosen == NULL) { pe_rsc_trace(clone, "Not adding sort index of %s: inactive", child->id); return; } node = (pe_node_t *) pe_hash_table_lookup(clone->allowed_nodes, chosen->details->id); CRM_ASSERT(node != NULL); node->weight = pcmk__add_scores(child->sort_index, node->weight); pe_rsc_trace(clone, "Added cumulative priority of %s (%s) to score on %s (now %s)", child->id, pcmk_readable_score(child->sort_index), pe__node_name(node), pcmk_readable_score(node->weight)); } /*! * \internal * \brief Apply colocation to dependent's node weights if for promoted role * * \param[in,out] data Colocation constraint to apply * \param[in,out] user_data Promotable clone that is constraint's dependent */ static void apply_coloc_to_dependent(gpointer data, gpointer user_data) { pcmk__colocation_t *constraint = (pcmk__colocation_t *) data; pe_resource_t *clone = (pe_resource_t *) user_data; pe_resource_t *primary = constraint->primary; uint32_t flags = pcmk__coloc_select_default; float factor = constraint->score / (float) INFINITY; if (constraint->dependent_role != RSC_ROLE_PROMOTED) { return; } if (constraint->score < INFINITY) { flags = pcmk__coloc_select_active; } pe_rsc_trace(clone, "Applying colocation %s (promoted %s with %s) @%s", constraint->id, constraint->dependent->id, constraint->primary->id, pcmk_readable_score(constraint->score)); - pcmk__add_colocated_node_scores(primary, clone->id, &clone->allowed_nodes, - constraint->node_attribute, factor, flags); + primary->cmds->add_colocated_node_scores(primary, clone->id, + &clone->allowed_nodes, + constraint->node_attribute, factor, + flags); } /*! * \internal * \brief Apply colocation to primary's node weights 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 *constraint = (pcmk__colocation_t *) data; pe_resource_t *clone = (pe_resource_t *) user_data; pe_resource_t *dependent = constraint->dependent; const float factor = constraint->score / (float) INFINITY; const uint32_t flags = pcmk__coloc_select_active |pcmk__coloc_select_nonnegative; if ((constraint->primary_role != RSC_ROLE_PROMOTED) || !pcmk__colocation_has_influence(constraint, NULL)) { return; } pe_rsc_trace(clone, "Applying colocation %s (%s with promoted %s) @%s", constraint->id, constraint->dependent->id, constraint->primary->id, pcmk_readable_score(constraint->score)); - pcmk__add_colocated_node_scores(dependent, clone->id, &clone->allowed_nodes, - constraint->node_attribute, factor, flags); + dependent->cmds->add_colocated_node_scores(dependent, clone->id, + &clone->allowed_nodes, + constraint->node_attribute, + factor, flags); } /*! * \internal * \brief Set clone instance's sort index to its node's weight * * \param[in,out] data Promotable clone instance * \param[in] user_data Parent clone of \p data */ static void set_sort_index_to_node_weight(gpointer data, gpointer user_data) { pe_resource_t *child = (pe_resource_t *) data; const pe_resource_t *clone = (const pe_resource_t *) user_data; pe_node_t *chosen = child->fns->location(child, NULL, FALSE); if (!pcmk_is_set(child->flags, pe_rsc_managed) && (child->next_role == RSC_ROLE_PROMOTED)) { child->sort_index = INFINITY; pe_rsc_trace(clone, "Final sort index for %s is INFINITY (unmanaged promoted)", child->id); } else if ((chosen == NULL) || (child->sort_index < 0)) { pe_rsc_trace(clone, "Final sort index for %s is %d (ignoring node weight)", child->id, child->sort_index); } else { const pe_node_t *node = NULL; node = pe_hash_table_lookup(clone->allowed_nodes, chosen->details->id); CRM_ASSERT(node != NULL); child->sort_index = node->weight; pe_rsc_trace(clone, "Merging weights for %s: final sort index for %s is %d", clone->id, child->id, child->sort_index); } } /*! * \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(pe_resource_t *clone) { if (pe__set_clone_flag(clone, pe__clone_promotion_constrained) == pcmk_rc_already) { return; } pe__set_resource_flags(clone, pe_rsc_merging); for (GList *iter = clone->children; iter != NULL; iter = iter->next) { pe_resource_t *child = (pe_resource_t *) iter->data; pe_rsc_trace(clone, "Merging weights for %s: initial sort index for %s is %d", clone->id, child->id, child->sort_index); } pe__show_node_weights(true, clone, "Before", clone->allowed_nodes, clone->cluster); /* Because the this_with_colocations() and with_this_colocations() methods * boil down to copies of rsc_cons and rsc_cons_lhs for clones, we can use * those here directly for efficiency. */ g_list_foreach(clone->children, add_sort_index_to_node_weight, clone); g_list_foreach(clone->rsc_cons, apply_coloc_to_dependent, clone); g_list_foreach(clone->rsc_cons_lhs, apply_coloc_to_primary, clone); // Ban resource from all nodes if it needs a ticket but doesn't have it pcmk__require_promotion_tickets(clone); pe__show_node_weights(true, clone, "After", clone->allowed_nodes, clone->cluster); // Reset sort indexes to final node weights g_list_foreach(clone->children, set_sort_index_to_node_weight, clone); // Finally, sort instances in descending order of promotion priority clone->children = g_list_sort(clone->children, cmp_promotable_instance); pe__clear_resource_flags(clone, pe_rsc_merging); } /*! * \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 pe_resource_t * find_active_anon_instance(const pe_resource_t *clone, const char *id, const pe_node_t *node) { for (GList *iter = clone->children; iter; iter = iter->next) { pe_resource_t *child = iter->data; pe_resource_t *active = NULL; // Use ->find_rsc() in case this is a cloned group active = clone->fns->find_rsc(child, id, node, pe_find_clone|pe_find_current); 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 pe_resource_t *clone, const char *id, const pe_node_t *node) { for (GList *iter = clone->children; iter; iter = iter->next) { pe_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->fns->find_rsc(child, id, NULL, pe_find_clone); CRM_LOG_ASSERT(child != NULL); if (child != NULL) { if (g_hash_table_lookup(child->known_on, node->details->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 pe_resource_t *rsc, const pe_node_t *node) { pe_node_t *allowed = pe_hash_table_lookup(rsc->allowed_nodes, node->details->id); return (allowed != NULL) && (allowed->weight >= 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 pe_resource_t *rsc, const pe_node_t *node) { char *id = clone_strip(rsc->id); const pe_resource_t *parent = pe__const_top_resource(rsc, false); pe_resource_t *active = NULL; const char *reason = "allowed"; // Some checks apply only to anonymous clone instances if (!pcmk_is_set(rsc->flags, pe_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->running_on == NULL) && (g_hash_table_size(rsc->known_on) == 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 ((pe_hash_table_lookup(rsc->known_on, node->details->id) != NULL) || (pe_find_node_id(rsc->running_on, node->details->id) != NULL)) { reason = "known"; } else { pe_rsc_trace(rsc, "Ignoring %s promotion score (for %s) on %s: not probed", rsc->id, id, pe__node_name(node)); free(id); return false; } check_allowed: if (is_allowed(rsc, node)) { pe_rsc_trace(rsc, "Counting %s promotion score (for %s) on %s: %s", rsc->id, id, pe__node_name(node), reason); free(id); return true; } pe_rsc_trace(rsc, "Ignoring %s promotion score (for %s) on %s: not allowed", rsc->id, id, pe__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 pe_resource_t *rsc, const pe_node_t *node, const char *name) { char *attr_name = NULL; const char *attr_value = NULL; CRM_CHECK((rsc != NULL) && (node != NULL) && (name != NULL), return NULL); attr_name = pcmk_promotion_score_name(name); attr_value = pe_node_attribute_calculated(node, attr_name, rsc); 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 pe_resource_t *rsc, const pe_node_t *node, bool *is_default) { 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->children != NULL) { int score = 0; for (const GList *iter = rsc->children; iter != NULL; iter = iter->next) { const pe_resource_t *child = (const pe_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 = (rsc->clone_name == NULL)? rsc->id : rsc->clone_name; attr_value = promotion_attr_value(rsc, node, name); if (attr_value != NULL) { pe_rsc_trace(rsc, "Promotion score for %s on %s = %s", name, pe__node_name(node), pcmk__s(attr_value, "(unset)")); } else if (!pcmk_is_set(rsc->flags, pe_rsc_unique)) { /* If we don't have any resource history yet, we won't have clone_name. * In that case, for anonymous clones, try the resource name without * any instance number. */ name = clone_strip(rsc->id); if (strcmp(rsc->id, name) != 0) { attr_value = promotion_attr_value(rsc, node, name); pe_rsc_trace(rsc, "Promotion score for %s on %s (for %s) = %s", name, pe__node_name(node), rsc->id, pcmk__s(attr_value, "(unset)")); } free(name); } if (attr_value == NULL) { return 0; } if (is_default != NULL) { *is_default = false; } return char2score(attr_value); } /*! * \internal * \brief Include promotion scores in instances' node weights and priorities * * \param[in,out] rsc Promotable clone resource to update */ void pcmk__add_promotion_scores(pe_resource_t *rsc) { if (pe__set_clone_flag(rsc, pe__clone_promotion_added) == pcmk_rc_already) { return; } for (GList *iter = rsc->children; iter != NULL; iter = iter->next) { pe_resource_t *child_rsc = (pe_resource_t *) iter->data; GHashTableIter iter; pe_node_t *node = NULL; int score, new_score; g_hash_table_iter_init(&iter, child_rsc->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (void **) &node)) { if (!pcmk__node_available(node, false, false)) { /* 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->weight, score); if (new_score != node->weight) { // Could remain INFINITY node->weight = new_score; pe_rsc_trace(rsc, "Added %s promotion priority (%s) to score " "on %s (now %s)", child_rsc->id, pcmk_readable_score(score), pe__node_name(node), pcmk_readable_score(new_score)); } } if (score > child_rsc->priority) { pe_rsc_trace(rsc, "Updating %s priority to promotion score (%d->%d)", child_rsc->id, child_rsc->priority, score); child_rsc->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) { pe_resource_t *rsc = (pe_resource_t *) data; if (rsc->role == RSC_ROLE_STARTED) { // Promotable clones should use unpromoted role instead of started rsc->role = RSC_ROLE_UNPROMOTED; } g_list_foreach(rsc->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) { pe_resource_t *rsc = (pe_resource_t *) data; GList *assigned = NULL; rsc->fns->location(rsc, &assigned, FALSE); if (assigned == NULL) { pe__set_next_role(rsc, RSC_ROLE_STOPPED, "stopped instance"); } else { pe__set_next_role(rsc, RSC_ROLE_UNPROMOTED, "unpromoted instance"); g_list_free(assigned); } g_list_foreach(rsc->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) { pe_resource_t *rsc = (pe_resource_t *) data; if (rsc->next_role == RSC_ROLE_UNKNOWN) { pe__set_next_role(rsc, RSC_ROLE_PROMOTED, "promoted instance"); } g_list_foreach(rsc->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(pe_resource_t *instance) { pe_node_t *chosen = instance->fns->location(instance, NULL, FALSE); if (pcmk_is_set(instance->cluster->flags, pe_flag_show_scores) && !pcmk__is_daemon && (instance->cluster->priv != NULL)) { pcmk__output_t *out = instance->cluster->priv; out->message(out, "promotion-score", instance, chosen, pcmk_readable_score(instance->sort_index)); } else { pe_rsc_debug(pe__const_top_resource(instance, false), "%s promotion score on %s: sort=%s priority=%s", instance->id, ((chosen == NULL)? "none" : pe__node_name(chosen)), pcmk_readable_score(instance->sort_index), pcmk_readable_score(instance->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) { pe_resource_t *instance = (pe_resource_t *) data; const pe_resource_t *clone = (const pe_resource_t *) user_data; const pe_node_t *chosen = NULL; enum rsc_role_e next_role = RSC_ROLE_UNKNOWN; GList *list = NULL; pe_rsc_trace(clone, "Assigning priority for %s: %s", instance->id, role2text(instance->next_role)); if (instance->fns->state(instance, TRUE) == RSC_ROLE_STARTED) { set_current_role_unpromoted(instance, NULL); } // Only an instance that will be active can be promoted chosen = instance->fns->location(instance, &list, FALSE); 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->fns->state(instance, FALSE); switch (next_role) { case RSC_ROLE_STARTED: case RSC_ROLE_UNKNOWN: // Set instance priority to its promotion score (or -1 if none) { bool is_default = false; instance->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 rsc_location constraints, but * prevents any instance from being promoted if neither * a constraint nor a promotion score is present */ instance->priority = -1; } } break; case RSC_ROLE_UNPROMOTED: case RSC_ROLE_STOPPED: // Instance can't be promoted instance->priority = -INFINITY; break; case RSC_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->rsc_location, chosen); apply_promoted_locations(instance, clone->rsc_location, 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->cmds->apply_coloc_score(instance, cons->primary, cons, true); } g_list_free(list); instance->sort_index = instance->priority; if (next_role == RSC_ROLE_PROMOTED) { instance->sort_index = INFINITY; } pe_rsc_trace(clone, "Assigning %s priority = %d", instance->id, instance->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) { pe_resource_t *instance = (pe_resource_t *) data; int *count = (int *) user_data; const pe_resource_t *clone = pe__const_top_resource(instance, false); pe_node_t *chosen = NULL; show_promotion_score(instance); if (instance->sort_index < 0) { pe_rsc_trace(clone, "Not supposed to promote instance %s", instance->id); } else if ((*count < pe__clone_promoted_max(instance)) || !pcmk_is_set(clone->flags, pe_rsc_managed)) { chosen = node_to_be_promoted_on(instance); } if (chosen == NULL) { set_next_role_unpromoted(instance, NULL); return; } if ((instance->role < RSC_ROLE_PROMOTED) && !pcmk_is_set(instance->cluster->flags, pe_flag_have_quorum) && (instance->cluster->no_quorum_policy == no_quorum_freeze)) { crm_notice("Clone instance %s cannot be promoted without quorum", instance->id); set_next_role_unpromoted(instance, NULL); return; } chosen->count++; pe_rsc_info(clone, "Choosing %s (%s) on %s for promotion", instance->id, role2text(instance->role), pe__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(pe_resource_t *rsc) { int promoted = 0; GHashTableIter iter; pe_node_t *node = NULL; // Repurpose count to track the number of promoted instances allocated g_hash_table_iter_init(&iter, rsc->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (void **)&node)) { node->count = 0; } // Set instances' promotion priorities and sort by highest priority first g_list_foreach(rsc->children, set_instance_priority, rsc); sort_promotable_instances(rsc); // Choose the first N eligible instances to be promoted g_list_foreach(rsc->children, set_instance_role, &promoted); pe_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(pe_resource_t *clone, bool *any_promoting, bool *any_demoting) { for (GList *iter = clone->children; iter != NULL; iter = iter->next) { pe_resource_t *instance = (pe_resource_t *) iter->data; instance->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(pe_resource_t *clone) { for (GList *iter = clone->children; iter != NULL; iter = iter->next) { pe_resource_t *instance = (pe_resource_t *) iter->data; instance->priority = clone->priority; } } /*! * \internal * \brief Create actions specific to promotable clones * * \param[in,out] clone Promotable clone to create actions for */ void pcmk__create_promotable_actions(pe_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(pe_resource_t *clone) { pe_resource_t *previous = NULL; // Needed for ordered clones pcmk__promotable_restart_ordering(clone); for (GList *iter = clone->children; iter != NULL; iter = iter->next) { pe_resource_t *instance = (pe_resource_t *) iter->data; // Demote before promote pcmk__order_resource_actions(instance, RSC_DEMOTE, instance, RSC_PROMOTE, pe_order_optional); 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_node Node where an instance of the primary will be * \param[in] colocation Colocation constraint to apply */ static void update_dependent_allowed_nodes(pe_resource_t *dependent, const pe_node_t *primary_node, const pcmk__colocation_t *colocation) { GHashTableIter iter; pe_node_t *node = NULL; const char *primary_value = NULL; const char *attr = NULL; if (colocation->score >= INFINITY) { return; // Colocation is mandatory, so allowed node scores don't matter } // Get value of primary's colocation node attribute attr = colocation->node_attribute; if (attr == NULL) { attr = CRM_ATTR_UNAME; } primary_value = pe_node_attribute_raw(primary_node, attr); pe_rsc_trace(colocation->primary, "Applying %s (%s with %s on %s by %s @%d) to %s", colocation->id, colocation->dependent->id, colocation->primary->id, pe__node_name(primary_node), attr, colocation->score, dependent->id); g_hash_table_iter_init(&iter, dependent->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (void **) &node)) { const char *dependent_value = pe_node_attribute_raw(node, attr); if (pcmk__str_eq(primary_value, dependent_value, pcmk__str_casei)) { node->weight = pcmk__add_scores(node->weight, colocation->score); pe_rsc_trace(colocation->primary, "Added %s score (%s) to %s (now %s)", colocation->id, pcmk_readable_score(colocation->score), pe__node_name(node), pcmk_readable_score(node->weight)); } } } /*! * \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 pe_resource_t *primary, pe_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->children; iter != NULL; iter = iter->next) { pe_resource_t *instance = (pe_resource_t *) iter->data; pe_node_t *node = instance->fns->location(instance, NULL, FALSE); if (node == NULL) { continue; } if (instance->fns->state(instance, FALSE) == colocation->primary_role) { update_dependent_allowed_nodes(dependent, node, colocation); affected_nodes = g_list_prepend(affected_nodes, node); } } /* For mandatory colocations, add the primary's node weight to the * dependent's node weight 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 >= INFINITY) && ((colocation->dependent_role != RSC_ROLE_PROMOTED) || (colocation->primary_role != RSC_ROLE_PROMOTED))) { pe_rsc_trace(colocation->primary, "Applying %s (mandatory %s with %s) to %s", colocation->id, colocation->dependent->id, colocation->primary->id, dependent->id); node_list_exclude(dependent->allowed_nodes, 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 */ void pcmk__update_promotable_dependent_priority(const pe_resource_t *primary, pe_resource_t *dependent, const pcmk__colocation_t *colocation) { pe_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->priority, colocation->score); pe_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->priority), pcmk_readable_score(colocation->score), pcmk_readable_score(new_priority)); dependent->priority = new_priority; } else if (colocation->score >= INFINITY) { // Mandatory colocation, but primary won't be here pe_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->priority = -INFINITY; } } diff --git a/lib/pacemaker/pcmk_sched_resource.c b/lib/pacemaker/pcmk_sched_resource.c index da73b1d6a4..754f252b01 100644 --- a/lib/pacemaker/pcmk_sched_resource.c +++ b/lib/pacemaker/pcmk_sched_resource.c @@ -1,718 +1,722 @@ /* * Copyright 2014-2023 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU General Public License version 2 * or later (GPLv2+) WITHOUT ANY WARRANTY. */ #include #include #include #include #include #include "libpacemaker_private.h" // Resource allocation methods that vary by resource variant static resource_alloc_functions_t allocation_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__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, clone_create_actions, clone_create_probe, clone_internal_constraints, pcmk__clone_apply_coloc_score, pcmk__colocated_resources, pcmk__with_clone_colocations, pcmk__clone_with_colocations, + pcmk__add_colocated_node_scores, clone_rsc_location, clone_action_flags, pcmk__instance_update_ordered_actions, pcmk__output_resource_actions, clone_expand, clone_append_meta, pcmk__clone_add_utilization, pcmk__clone_shutdown_lock, }, { pcmk__bundle_allocate, 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_rsc_location, pcmk__bundle_action_flags, pcmk__instance_update_ordered_actions, pcmk__output_bundle_actions, pcmk__bundle_expand, 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(pe_resource_t *rsc, pe_node_t *node, const xmlNode *rsc_entry, bool active_on_node) { bool changed = false; const char *attr_list[] = { XML_ATTR_TYPE, XML_AGENT_ATTR_CLASS, XML_AGENT_ATTR_PROVIDER }; for (int i = 0; i < PCMK__NELEM(attr_list); i++) { const char *value = crm_element_value(rsc->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->cluster); if (active_on_node) { crm_notice("Forcing restart of %s on %s " "because %s changed from '%s' to '%s'", rsc->id, pe__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), CRMD_ACTION_STOP, node, FALSE, TRUE, rsc->cluster); pe__set_resource_flags(rsc, pe_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, pe_resource_t *rsc, const char *id) { if ((strcmp(rsc->id, id) == 0) || ((rsc->clone_name != NULL) && (strcmp(rsc->clone_name, id) == 0))) { result = g_list_prepend(result, rsc); } for (GList *iter = rsc->children; iter != NULL; iter = iter->next) { pe_resource_t *child = (pe_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] data_set Cluster working set * * \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 pe_working_set_t *data_set) { GList *result = NULL; CRM_CHECK((id != NULL) && (data_set != NULL), return NULL); for (GList *iter = data_set->resources; iter != NULL; iter = iter->next) { result = add_rsc_if_matching(result, (pe_resource_t *) iter->data, id); } return result; } /*! * \internal * \brief Set the variant-appropriate allocation methods for a resource * * \param[in,out] rsc Resource to set allocation methods for * \param[in] ignored Here so function can be used with g_list_foreach() */ static void set_allocation_methods_for_rsc(pe_resource_t *rsc, void *ignored) { rsc->cmds = &allocation_methods[rsc->variant]; g_list_foreach(rsc->children, (GFunc) set_allocation_methods_for_rsc, NULL); } /*! * \internal * \brief Set the variant-appropriate allocation methods for all resources * * \param[in,out] data_set Cluster working set */ void pcmk__set_allocation_methods(pe_working_set_t *data_set) { g_list_foreach(data_set->resources, (GFunc) set_allocation_methods_for_rsc, NULL); } // Shared implementation of resource_alloc_functions_t:colocated_resources() GList * pcmk__colocated_resources(const pe_resource_t *rsc, const pe_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; } pe_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 pe_resource_t *primary = constraint->primary; if (primary == orig_rsc) { continue; // Break colocation loop } if ((constraint->score == INFINITY) && (pcmk__colocation_affects(rsc, primary, constraint, true) == pcmk__coloc_affects_location)) { colocated_rscs = primary->cmds->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 pe_resource_t *dependent = constraint->dependent; if (dependent == orig_rsc) { continue; // Break colocation loop } if (pe_rsc_is_clone(rsc) && !pe_rsc_is_clone(dependent)) { continue; // We can't be sure whether dependent will be colocated } if ((constraint->score == INFINITY) && (pcmk__colocation_affects(dependent, rsc, constraint, true) == pcmk__coloc_affects_location)) { colocated_rscs = dependent->cmds->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 pe_resource_t *rsc, xmlNode *xml) { } void pcmk__output_resource_actions(pe_resource_t *rsc) { pcmk__output_t *out = rsc->cluster->priv; pe_node_t *next = NULL; pe_node_t *current = NULL; if (rsc->children != NULL) { for (GList *iter = rsc->children; iter != NULL; iter = iter->next) { pe_resource_t *child = (pe_resource_t *) iter->data; child->cmds->output_actions(child); } return; } next = rsc->allocated_to; if (rsc->running_on) { current = pe__current_node(rsc); if (rsc->role == RSC_ROLE_STOPPED) { /* This can occur when resources are being recovered because * the current role can change in pcmk__primitive_create_actions() */ rsc->role = RSC_ROLE_STARTED; } } if ((current == NULL) && pcmk_is_set(rsc->flags, pe_rsc_orphan)) { /* Don't log stopped orphans */ return; } out->message(out, "rsc-action", rsc, current, next); } /*! * \internal * \brief Assign a specified primitive resource to a node * * Assign a specified primitive resource to a specified node, if the node can * run the resource (or unconditionally, if \p force is true). Mark the resource * as no longer provisional. If the primitive can't be assigned (or \p chosen is * NULL), unassign any previous assignment for it, set its next role to stopped, * and update any existing actions scheduled for it. This is not done * recursively for children, so it should be called only for primitives. * * \param[in,out] rsc Resource to assign * \param[in,out] chosen Node to assign \p rsc to * \param[in] force If true, assign to \p chosen even if unavailable * * \return true if \p rsc could be assigned, otherwise false * * \note Assigning a resource to the NULL node using this function is different * from calling pcmk__unassign_resource(), in that it will also update any * actions created for the resource. */ bool pcmk__finalize_assignment(pe_resource_t *rsc, pe_node_t *chosen, bool force) { pcmk__output_t *out = rsc->cluster->priv; CRM_ASSERT(rsc->variant == pe_native); if (!force && (chosen != NULL)) { if ((chosen->weight < 0) // Allow the graph to assume that guest node connections will come up || (!pcmk__node_available(chosen, true, false) && !pe__is_guest_node(chosen))) { crm_debug("All nodes for resource %s are unavailable, unclean or " "shutting down (%s can%s run resources, with weight %d)", rsc->id, pe__node_name(chosen), (pcmk__node_available(chosen, true, false)? "" : "not"), chosen->weight); pe__set_next_role(rsc, RSC_ROLE_STOPPED, "node availability"); chosen = NULL; } } pcmk__unassign_resource(rsc); pe__clear_resource_flags(rsc, pe_rsc_provisional); if (chosen == NULL) { crm_debug("Could not allocate a node for %s", rsc->id); pe__set_next_role(rsc, RSC_ROLE_STOPPED, "unable to allocate"); for (GList *iter = rsc->actions; iter != NULL; iter = iter->next) { pe_action_t *op = (pe_action_t *) iter->data; crm_debug("Updating %s for allocation failure", op->uuid); if (pcmk__str_eq(op->task, RSC_STOP, pcmk__str_casei)) { pe__clear_action_flags(op, pe_action_optional); } else if (pcmk__str_eq(op->task, RSC_START, pcmk__str_casei)) { pe__clear_action_flags(op, pe_action_runnable); //pe__set_resource_flags(rsc, pe_rsc_block); } else { // Cancel recurring actions, unless for stopped state const char *interval_ms_s = NULL; const char *target_rc_s = NULL; char *rc_stopped = pcmk__itoa(PCMK_OCF_NOT_RUNNING); interval_ms_s = g_hash_table_lookup(op->meta, XML_LRM_ATTR_INTERVAL_MS); target_rc_s = g_hash_table_lookup(op->meta, XML_ATTR_TE_TARGET_RC); if ((interval_ms_s != NULL) && !pcmk__str_eq(interval_ms_s, "0", pcmk__str_none) && !pcmk__str_eq(rc_stopped, target_rc_s, pcmk__str_none)) { pe__clear_action_flags(op, pe_action_runnable); } free(rc_stopped); } } return false; } crm_debug("Assigning %s to %s", rsc->id, pe__node_name(chosen)); rsc->allocated_to = pe__copy_node(chosen); chosen->details->allocated_rsc = g_list_prepend(chosen->details->allocated_rsc, rsc); chosen->details->num_resources++; chosen->count++; pcmk__consume_node_capacity(chosen->details->utilization, rsc); if (pcmk_is_set(rsc->cluster->flags, pe_flag_show_utilization)) { out->message(out, "resource-util", rsc, chosen, __func__); } return true; } /*! * \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 the resources can't be assigned * (or \p chosen is NULL), unassign any previous assignments, set next role to * stopped, and update any existing actions scheduled for them. * * \param[in,out] rsc Resource to assign * \param[in,out] chosen Node to assign \p rsc to * \param[in] force If true, assign to \p chosen even if unavailable * * \return true if \p rsc could be assigned, otherwise false * * \note Assigning a resource to the NULL node using this function is different * from calling pcmk__unassign_resource(), in that it will also update any * actions created for the resource. */ bool pcmk__assign_resource(pe_resource_t *rsc, pe_node_t *node, bool force) { bool changed = false; if (rsc->children == NULL) { if (rsc->allocated_to != NULL) { changed = true; } pcmk__finalize_assignment(rsc, node, force); } else { for (GList *iter = rsc->children; iter != NULL; iter = iter->next) { pe_resource_t *child_rsc = (pe_resource_t *) iter->data; changed |= pcmk__assign_resource(child_rsc, node, force); } } return changed; } /*! * \internal * \brief Remove any assignment of a specified resource to a node * * If a specified resource has been assigned to a node, remove that assignment * and mark the resource as provisional again. This is not done recursively for * children, so it should be called only for primitives. * * \param[in,out] rsc Resource to unassign */ void pcmk__unassign_resource(pe_resource_t *rsc) { pe_node_t *old = rsc->allocated_to; if (old == NULL) { return; } crm_info("Unassigning %s from %s", rsc->id, pe__node_name(old)); pe__set_resource_flags(rsc, pe_rsc_provisional); rsc->allocated_to = NULL; /* We're going to free the pe_node_t, but its details member is shared and * will remain, so update that appropriately first. */ old->details->allocated_rsc = g_list_remove(old->details->allocated_rsc, rsc); old->details->num_resources--; pcmk__release_node_capacity(old->details->utilization, rsc); free(old); } /*! * \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(pe_resource_t *rsc, const pe_node_t *node, pe_resource_t **failed) { int fail_count, remaining_tries; pe_resource_t *rsc_to_ban = rsc; // Migration threshold of 0 means never force away if (rsc->migration_threshold == 0) { return false; } // If we're ignoring failures, also ignore the migration threshold if (pcmk_is_set(rsc->flags, pe_rsc_failure_ignored)) { return false; } // If there are no failures, there's no need to force away fail_count = pe_get_failcount(node, rsc, NULL, pe_fc_effective|pe_fc_fillers, 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, pe_rsc_unique)) { rsc_to_ban = uber_parent(rsc); } // How many more times recovery will be tried on this node remaining_tries = rsc->migration_threshold - fail_count; if (remaining_tries <= 0) { crm_warn("%s cannot run on %s due to reaching migration threshold " "(clean up resource to allow again)" CRM_XS " failures=%d migration-threshold=%d", rsc_to_ban->id, pe__node_name(node), fail_count, rsc->migration_threshold); 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), pe__node_name(node), rsc->migration_threshold); return false; } static void * convert_const_pointer(const void *ptr) { /* Worst function ever */ return (void *)ptr; } /*! * \internal * \brief Get a node's weight * * \param[in] node Unweighted node to check (for node ID) * \param[in] nodes List of weighted nodes to look for \p node in * * \return Node's weight, or -INFINITY if not found */ static int get_node_weight(const pe_node_t *node, GHashTable *nodes) { pe_node_t *weighted_node = NULL; if ((node != NULL) && (nodes != NULL)) { weighted_node = g_hash_table_lookup(nodes, node->details->id); } return (weighted_node == NULL)? -INFINITY : weighted_node->weight; } /*! * \internal * \brief Compare two resources according to which should be allocated 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 allocated before \b, 0 if they are equal, * or +1 if \p a should be allocated after \b */ static gint cmp_resources(gconstpointer a, gconstpointer b, gpointer data) { const pe_resource_t *resource1 = a; const pe_resource_t *resource2 = b; const GList *nodes = (const GList *) data; int rc = 0; int r1_weight = -INFINITY; int r2_weight = -INFINITY; pe_node_t *r1_node = NULL; pe_node_t *r2_node = NULL; GHashTable *r1_nodes = NULL; GHashTable *r2_nodes = NULL; const char *reason = NULL; // Resources with highest priority should be allocated first reason = "priority"; r1_weight = resource1->priority; r2_weight = resource2->priority; if (r1_weight > r2_weight) { rc = -1; goto done; } if (r1_weight < r2_weight) { 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 weights - pcmk__add_colocated_node_scores(convert_const_pointer(resource1), - resource1->id, &r1_nodes, NULL, 1, - pcmk__coloc_select_this_with); - pcmk__add_colocated_node_scores(convert_const_pointer(resource2), - resource2->id, &r2_nodes, NULL, 1, - pcmk__coloc_select_this_with); + resource1->cmds->add_colocated_node_scores(convert_const_pointer(resource1), + resource1->id, &r1_nodes, NULL, + 1, pcmk__coloc_select_this_with); + resource2->cmds->add_colocated_node_scores(convert_const_pointer(resource2), + resource2->id, &r2_nodes, NULL, + 1, pcmk__coloc_select_this_with); pe__show_node_weights(true, NULL, resource1->id, r1_nodes, resource1->cluster); pe__show_node_weights(true, NULL, resource2->id, r2_nodes, resource2->cluster); // The resource with highest score on its current node goes first reason = "current location"; if (resource1->running_on != NULL) { r1_node = pe__current_node(resource1); } if (resource2->running_on != NULL) { r2_node = pe__current_node(resource2); } r1_weight = get_node_weight(r1_node, r1_nodes); r2_weight = get_node_weight(r2_node, r2_nodes); if (r1_weight > r2_weight) { rc = -1; goto done; } if (r1_weight < r2_weight) { rc = 1; goto done; } // Otherwise a higher weight on any node will do reason = "score"; for (const GList *iter = nodes; iter != NULL; iter = iter->next) { const pe_node_t *node = (const pe_node_t *) iter->data; r1_weight = get_node_weight(node, r1_nodes); r2_weight = get_node_weight(node, r2_nodes); if (r1_weight > r2_weight) { rc = -1; goto done; } if (r1_weight < r2_weight) { rc = 1; goto done; } } done: crm_trace("%s (%d)%s%s %c %s (%d)%s%s: %s", resource1->id, r1_weight, ((r1_node == NULL)? "" : " on "), ((r1_node == NULL)? "" : r1_node->details->id), ((rc < 0)? '>' : ((rc > 0)? '<' : '=')), resource2->id, r2_weight, ((r2_node == NULL)? "" : " on "), ((r2_node == NULL)? "" : r2_node->details->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 allocated to nodes * * \param[in,out] data_set Cluster working set */ void pcmk__sort_resources(pe_working_set_t *data_set) { GList *nodes = g_list_copy(data_set->nodes); nodes = pcmk__sort_nodes(nodes, NULL); data_set->resources = g_list_sort_with_data(data_set->resources, cmp_resources, nodes); g_list_free(nodes); }