diff --git a/include/pcmki/pcmki_sched_allocate.h b/include/pcmki/pcmki_sched_allocate.h index 64fa0b8e55..41c4b7bb07 100644 --- a/include/pcmki/pcmki_sched_allocate.h +++ b/include/pcmki/pcmki_sched_allocate.h @@ -1,85 +1,84 @@ /* * Copyright 2004-2022 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__PCMKI_PCMKI_SCHED_ALLOCATE__H # define PCMK__PCMKI_PCMKI_SCHED_ALLOCATE__H # include # include # include # include # include # include # include # include pe_node_t *pcmk__native_allocate(pe_resource_t *rsc, pe_node_t *prefer); void native_create_actions(pe_resource_t *rsc); void native_internal_constraints(pe_resource_t *rsc); extern enum pe_action_flags native_action_flags(pe_action_t * action, pe_node_t * node); void native_rsc_location(pe_resource_t *rsc, pe__location_t *constraint); -bool native_create_probe(pe_resource_t *rsc, pe_node_t *node); extern void native_append_meta(pe_resource_t * rsc, xmlNode * xml); void pcmk__primitive_add_utilization(pe_resource_t *rsc, pe_resource_t *orig_rsc, GList *all_rscs, GHashTable *utilization); void pcmk__primitive_shutdown_lock(pe_resource_t *rsc); pe_node_t *pcmk__group_allocate(pe_resource_t *rsc, pe_node_t *prefer); void group_create_actions(pe_resource_t *rsc); void group_internal_constraints(pe_resource_t *rsc); extern enum pe_action_flags group_action_flags(pe_action_t * action, pe_node_t * node); void group_rsc_location(pe_resource_t *rsc, pe__location_t *constraint); extern void group_append_meta(pe_resource_t * rsc, xmlNode * xml); void pcmk__group_add_utilization(pe_resource_t *rsc, pe_resource_t *orig_rsc, GList *all_rscs, GHashTable *utilization); void pcmk__group_shutdown_lock(pe_resource_t *rsc); pe_node_t *pcmk__bundle_allocate(pe_resource_t *rsc, pe_node_t *prefer); void pcmk__bundle_create_actions(pe_resource_t *rsc); bool pcmk__bundle_create_probe(pe_resource_t *rsc, pe_node_t *node); void pcmk__bundle_internal_constraints(pe_resource_t *rsc); void pcmk__bundle_rsc_location(pe_resource_t *rsc, pe__location_t *constraint); enum pe_action_flags pcmk__bundle_action_flags(pe_action_t *action, pe_node_t *node); void pcmk__bundle_expand(pe_resource_t *rsc); void pcmk__bundle_append_meta(pe_resource_t *rsc, xmlNode *xml); void pcmk__bundle_add_utilization(pe_resource_t *rsc, pe_resource_t *orig_rsc, GList *all_rscs, GHashTable *utilization); void pcmk__bundle_shutdown_lock(pe_resource_t *rsc); pe_node_t *pcmk__clone_allocate(pe_resource_t *rsc, pe_node_t *prefer); void clone_create_actions(pe_resource_t *rsc); void clone_internal_constraints(pe_resource_t *rsc); void clone_rsc_location(pe_resource_t *rsc, pe__location_t *constraint); extern enum pe_action_flags clone_action_flags(pe_action_t * action, pe_node_t * node); void clone_expand(pe_resource_t *rsc); bool clone_create_probe(pe_resource_t *rsc, pe_node_t *node); extern void clone_append_meta(pe_resource_t * rsc, xmlNode * xml); void pcmk__clone_add_utilization(pe_resource_t *rsc, pe_resource_t *orig_rsc, GList *all_rscs, GHashTable *utilization); void pcmk__clone_shutdown_lock(pe_resource_t *rsc); void pcmk__add_promotion_scores(pe_resource_t *rsc); uint32_t group_update_actions(pe_action_t *first, pe_action_t *then, pe_node_t *node, uint32_t flags, uint32_t filter, uint32_t type, pe_working_set_t *data_set); uint32_t pcmk__multi_update_actions(pe_action_t *first, pe_action_t *then, pe_node_t *node, uint32_t flags, uint32_t filter, uint32_t type, pe_working_set_t *data_set); void pcmk__log_transition_summary(const char *filename); void clone_create_pseudo_actions(pe_resource_t *rsc, GList *children, notify_data_t **start_notify, notify_data_t **stop_notify); #endif diff --git a/lib/pacemaker/libpacemaker_private.h b/lib/pacemaker/libpacemaker_private.h index 968195fcbb..7cb49fb08c 100644 --- a/lib/pacemaker/libpacemaker_private.h +++ b/lib/pacemaker/libpacemaker_private.h @@ -1,688 +1,691 @@ /* * Copyright 2021-2022 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 the add_colocated_node_scores() method 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 { pe_node_t *(*allocate)(pe_resource_t *rsc, pe_node_t *prefer); void (*create_actions)(pe_resource_t *rsc); /*! * \internal * \brief Schedule any probes needed for a resource on a node * * \param[in] rsc Resource to create probe for * \param[in] 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); 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] 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, pe_resource_t *primary, pcmk__colocation_t *colocation, bool for_dependent); /*! * \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 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, enum pcmk__coloc_select flags); /*! * \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] 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)(pe_resource_t *rsc, pe_resource_t *orig_rsc, GList *colocated_rscs); void (*rsc_location) (pe_resource_t *, pe__location_t *); enum pe_action_flags (*action_flags) (pe_action_t *, pe_node_t *); /*! * \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. In some cases, the ordering could be disabled as well. * * \param[in] first 'First' action in an ordering * \param[in] 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] 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, 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] rsc Resource whose actions should be added */ void (*add_actions_to_graph)(pe_resource_t *rsc); void (*append_meta) (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] utilization Table of utilization values to add to */ void (*add_utilization)(pe_resource_t *rsc, pe_resource_t *orig_rsc, GList *all_rscs, GHashTable *utilization); /*! * \internal * \brief Apply a shutdown lock for a resource, if appropriate * * \param[in] 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, 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, 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, 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, xmlNode *xml_op); G_GNUC_INTERNAL void pcmk__handle_rsc_config_changes(pe_working_set_t *data_set); // Producing transition graphs (pcmk_graph_producer.c) G_GNUC_INTERNAL bool pcmk__graph_has_loop(pe_action_t *init_action, pe_action_t *action, pe_action_wrapper_t *input); G_GNUC_INTERNAL void pcmk__add_action_to_graph(pe_action_t *action, pe_working_set_t *data_set); 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(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(pe_node_t *node); // Injected scheduler inputs (pcmk_sched_injections.c) void pcmk__inject_scheduler_input(pe_working_set_t *data_set, cib_t *cib, 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, pe_working_set_t *data_set); G_GNUC_INTERNAL bool pcmk__valid_resource_or_tag(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, 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__location_t *constraint, pe_resource_t *rsc); // 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(pe_resource_t *dependent, pe_resource_t *primary, pcmk__colocation_t *constraint, bool preview); G_GNUC_INTERNAL void pcmk__apply_coloc_to_weights(pe_resource_t *dependent, pe_resource_t *primary, pcmk__colocation_t *constraint); G_GNUC_INTERNAL void pcmk__apply_coloc_to_priority(pe_resource_t *dependent, pe_resource_t *primary, pcmk__colocation_t *constraint); 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__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_colocated_starts(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, enum pe_ordering type, 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] first_rsc Resource for 'first' action * \param[in] then_rsc Resource for 'then' action * \param[in] first_task Action key for 'first' action * \param[in] then_task Action key for 'then' action * \param[in] flags Bitmask of enum pe_ordering flags * \param[in] data_set Cluster working set to add ordering to */ #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, type) \ pcmk__order_resource_actions((rsc1), CRMD_ACTION_START, \ (rsc2), CRMD_ACTION_START, (type)) #define pcmk__order_stops(rsc1, rsc2, type) \ pcmk__order_resource_actions((rsc1), CRMD_ACTION_STOP, \ (rsc2), CRMD_ACTION_STOP, (type)) // 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__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(pe_resource_t *primary, pe_resource_t *dependent, pcmk__colocation_t *colocation); G_GNUC_INTERNAL void pcmk__update_promotable_dependent_priority(pe_resource_t *primary, pe_resource_t *dependent, pcmk__colocation_t *colocation); // Pacemaker Remote nodes (pcmk_sched_remote.c) G_GNUC_INTERNAL bool pcmk__is_failed_remote_node(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(pe_resource_t *rsc, pe_node_t *node); G_GNUC_INTERNAL pe_node_t *pcmk__connection_host_for_action(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, pe_action_t *action); // Primitives (pcmk_sched_native.c) G_GNUC_INTERNAL void pcmk__primitive_apply_coloc_score(pe_resource_t *dependent, pe_resource_t *primary, pcmk__colocation_t *colocation, bool for_dependent); // Groups (pcmk_sched_group.c) G_GNUC_INTERNAL void pcmk__group_apply_coloc_score(pe_resource_t *dependent, pe_resource_t *primary, pcmk__colocation_t *colocation, bool for_dependent); G_GNUC_INTERNAL void pcmk__group_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 GList *pcmk__group_colocated_resources(pe_resource_t *rsc, pe_resource_t *orig_rsc, GList *colocated_rscs); // Clones (pcmk_sched_clone.c) G_GNUC_INTERNAL void pcmk__clone_apply_coloc_score(pe_resource_t *dependent, pe_resource_t *primary, pcmk__colocation_t *colocation, bool for_dependent); // Bundles (pcmk_sched_bundle.c) G_GNUC_INTERNAL void pcmk__bundle_apply_coloc_score(pe_resource_t *dependent, pe_resource_t *primary, pcmk__colocation_t *colocation, bool for_dependent); G_GNUC_INTERNAL void pcmk__output_bundle_actions(pe_resource_t *rsc); // 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, pe_working_set_t *data_set); G_GNUC_INTERNAL GList *pcmk__colocated_resources(pe_resource_t *rsc, pe_resource_t *orig_rsc, GList *colocated_rscs); G_GNUC_INTERNAL void pcmk__output_resource_actions(pe_resource_t *rsc); G_GNUC_INTERNAL bool pcmk__assign_primitive(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, 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 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, pe_resource_t *rsc); G_GNUC_INTERNAL void pcmk__release_node_capacity(GHashTable *current_utilization, pe_resource_t *rsc); G_GNUC_INTERNAL void pcmk__ban_insufficient_capacity(pe_resource_t *rsc, pe_node_t **prefer); G_GNUC_INTERNAL void pcmk__create_utilization_constraints(pe_resource_t *rsc, 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_probes.c b/lib/pacemaker/pcmk_sched_probes.c index eac5fb0206..672178abc8 100644 --- a/lib/pacemaker/pcmk_sched_probes.c +++ b/lib/pacemaker/pcmk_sched_probes.c @@ -1,859 +1,836 @@ /* * Copyright 2004-2022 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU General Public License version 2 * or later (GPLv2+) WITHOUT ANY WARRANTY. */ #include #include #include #include #include #include "libpacemaker_private.h" /*! * \internal * \brief Add the expected result to a newly created probe * * \param[in] probe Probe action to add expected result to * \param[in] rsc Resource that probe is for * \param[in] node Node that probe will run on */ static void add_expected_result(pe_action_t *probe, pe_resource_t *rsc, pe_node_t *node) { // Check whether resource is currently active on node pe_node_t *running = pe_find_node_id(rsc->running_on, node->details->id); // The expected result is what we think the resource's current state is if (running == NULL) { static const char *rc_inactive = NULL; if (rc_inactive == NULL) { rc_inactive = pcmk__itoa(PCMK_OCF_NOT_RUNNING); } add_hash_param(probe->meta, XML_ATTR_TE_TARGET_RC, rc_inactive); } else if (rsc->role == RSC_ROLE_PROMOTED) { static const char *rc_promoted = NULL; if (rc_promoted == NULL) { rc_promoted = pcmk__itoa(PCMK_OCF_RUNNING_PROMOTED); } add_hash_param(probe->meta, XML_ATTR_TE_TARGET_RC, rc_promoted); } } /*! * \internal * \brief Create any needed robes on a node for a list of resources * * \param[in] rscs List of resources to create probes for * \param[in] node Node to create probes on * * \return true if any probe was created, otherwise false */ bool pcmk__probe_resource_list(GList *rscs, pe_node_t *node) { bool any_created = false; for (GList *iter = rscs; iter != NULL; iter = iter->next) { pe_resource_t *rsc = (pe_resource_t *) iter->data; if (rsc->cmds->create_probe(rsc, node)) { any_created = true; } } return any_created; } /*! * \internal * \brief Order one resource's start after another's start-up probe * * \param[in] rsc1 Resource that might get start-up probe * \param[in] rsc2 Resource that might be started */ static void probe_then_start(pe_resource_t *rsc1, pe_resource_t *rsc2) { if ((rsc1->allocated_to != NULL) && (g_hash_table_lookup(rsc1->known_on, rsc1->allocated_to->details->id) == NULL)) { pcmk__new_ordering(rsc1, pcmk__op_key(rsc1->id, RSC_STATUS, 0), NULL, rsc2, pcmk__op_key(rsc2->id, RSC_START, 0), NULL, pe_order_optional, rsc1->cluster); } } /*! * \internal * \brief Check whether a guest resource will stop * * \param[in] node Guest node to check * * \return true if guest resource will likely stop, otherwise false */ static bool guest_resource_will_stop(pe_node_t *node) { pe_resource_t *guest_rsc = node->details->remote_rsc->container; /* Ideally, we'd check whether the guest has a required stop, but that * information doesn't exist yet, so approximate it ... */ return node->details->remote_requires_reset || node->details->unclean || pcmk_is_set(guest_rsc->flags, pe_rsc_failed) || (guest_rsc->next_role == RSC_ROLE_STOPPED) // Guest is moving || ((guest_rsc->role > RSC_ROLE_STOPPED) && (guest_rsc->allocated_to != NULL) && (pe_find_node(guest_rsc->running_on, guest_rsc->allocated_to->details->uname) == NULL)); } /*! * \internal * \brief Create a probe action for a resource on a node * * \param[in] rsc Resource to create probe for * \param[in[ node Node to create probe on * * \return Newly created probe action */ static pe_action_t * probe_action(pe_resource_t *rsc, pe_node_t *node) { pe_action_t *probe = NULL; char *key = pcmk__op_key(rsc->id, RSC_STATUS, 0); crm_debug("Scheduling probe of %s %s on %s", role2text(rsc->role), rsc->id, node->details->uname); probe = custom_action(rsc, key, RSC_STATUS, node, FALSE, TRUE, rsc->cluster); pe__clear_action_flags(probe, pe_action_optional); pcmk__order_vs_unfence(rsc, node, probe, pe_order_optional); add_expected_result(probe, rsc, node); return probe; } /*! * \internal * \brief Create probes for a resource on a node, if needed * * \brief Schedule any probes needed for a resource on a node * * \param[in] rsc Resource to create probe for * \param[in] node Node to create probe on * * \return true if any probe was created, otherwise false */ bool -native_create_probe(pe_resource_t *rsc, pe_node_t *node) +pcmk__probe_rsc_on_node(pe_resource_t *rsc, pe_node_t *node) { enum pe_ordering flags = pe_order_optional; pe_action_t *probe = NULL; pe_node_t *allowed = NULL; pe_resource_t *top = uber_parent(rsc); + const char *reason = NULL; + + CRM_CHECK((rsc != NULL) && (node != NULL), return false); - CRM_CHECK(node != NULL, return false); if (!pcmk_is_set(rsc->cluster->flags, pe_flag_startup_probes)) { - pe_rsc_trace(rsc, "Skipping active resource detection for %s", rsc->id); - return false; + reason = "start-up probes are disabled"; + goto no_probe; } if (pe__is_guest_or_remote_node(node)) { const char *class = crm_element_value(rsc->xml, XML_AGENT_ATTR_CLASS); - if (pcmk__str_eq(class, PCMK_RESOURCE_CLASS_STONITH, pcmk__str_casei)) { - pe_rsc_trace(rsc, - "Skipping probe for %s on %s because Pacemaker Remote nodes cannot run stonith agents", - rsc->id, node->details->id); - return false; + if (pcmk__str_eq(class, PCMK_RESOURCE_CLASS_STONITH, pcmk__str_none)) { + reason = "Pacemaker Remote nodes cannot run stonith agents"; + goto no_probe; + } else if (pe__is_guest_node(node) && pe__resource_contains_guest_node(rsc->cluster, rsc)) { - pe_rsc_trace(rsc, - "Skipping probe for %s on %s because guest nodes cannot run resources containing guest nodes", - rsc->id, node->details->id); - return false; + reason = "guest nodes cannot run resources containing guest nodes"; + goto no_probe; + } else if (rsc->is_remote_node) { - pe_rsc_trace(rsc, - "Skipping probe for %s on %s because Pacemaker Remote nodes cannot host remote connections", - rsc->id, node->details->id); - return false; + reason = "Pacemaker Remote nodes cannot host remote connections"; + goto no_probe; } } // If this is a collective resource, probes are created for its children if (rsc->children != NULL) { return pcmk__probe_resource_list(rsc->children, node); - - } else if ((rsc->container) && (!rsc->is_remote_node)) { - pe_rsc_trace(rsc, "Skipping %s: it is within container %s", rsc->id, rsc->container->id); - return false; } - if (pcmk_is_set(rsc->flags, pe_rsc_orphan)) { - pe_rsc_trace(rsc, "Skipping orphan: %s", rsc->id); - return false; - } + if ((rsc->container != NULL) && !rsc->is_remote_node) { + reason = "resource is inside a container"; + goto no_probe; - // Check whether resource is already known on node - if (g_hash_table_lookup(rsc->known_on, node->details->id)) { - pe_rsc_trace(rsc, "Skipping known: %s on %s", rsc->id, node->details->uname); - return false; + } else if (pcmk_is_set(rsc->flags, pe_rsc_orphan)) { + reason = "resource is orphaned"; + goto no_probe; + + } else if (g_hash_table_lookup(rsc->known_on, node->details->id) != NULL) { + reason = "resource state is already known"; + goto no_probe; } allowed = g_hash_table_lookup(rsc->allowed_nodes, node->details->id); if (rsc->exclusive_discover || top->exclusive_discover) { + // Exclusive discovery is enabled ... + if (allowed == NULL) { - /* exclusive discover is enabled and this node is not in the allowed list. */ - pe_rsc_trace(rsc, "Skipping probe for %s on node %s, A", rsc->id, node->details->id); - return false; + // ... but this node is not allowed to run the resource + reason = "resource has exclusive discovery but is not allowed " + "on node"; + goto no_probe; + } else if (allowed->rsc_discover_mode != pe_discover_exclusive) { - /* exclusive discover is enabled and this node is not marked - * as a node this resource should be discovered on */ - pe_rsc_trace(rsc, "Skipping probe for %s on node %s, B", rsc->id, node->details->id); - return false; + // ... but no constraint marks this node for discovery of resource + reason = "resource has exclusive discovery but is not enabled " + "on node"; + goto no_probe; } } - if(allowed == NULL && node->rsc_discover_mode == pe_discover_never) { - /* If this node was allowed to host this resource it would - * have been explicitly added to the 'allowed_nodes' list. - * However it wasn't and the node has discovery disabled, so - * no need to probe for this resource. - */ - pe_rsc_trace(rsc, "Skipping probe for %s on node %s, C", rsc->id, node->details->id); - return false; + if (allowed == NULL) { + allowed = node; } - - if (allowed && allowed->rsc_discover_mode == pe_discover_never) { - /* this resource is marked as not needing to be discovered on this node */ - pe_rsc_trace(rsc, "Skipping probe for %s on node %s, discovery mode", rsc->id, node->details->id); - return false; + if (allowed->rsc_discover_mode == pe_discover_never) { + reason = "node has discovery disabled"; + goto no_probe; } if (pe__is_guest_node(node)) { - pe_resource_t *remote = node->details->remote_rsc->container; + pe_resource_t *guest = node->details->remote_rsc->container; - if(remote->role == RSC_ROLE_STOPPED) { - /* If the container is stopped, then we know anything that - * might have been inside it is also stopped and there is - * no need to probe. - * - * If we don't know the container's state on the target - * either: - * - * - the container is running, the transition will abort - * and we'll end up in a different case next time, or - * - * - the container is stopped - * - * Either way there is no need to probe. - * - */ - pe_rsc_trace(rsc, "Skipping probe for %s on node %s, %s is stopped", - rsc->id, node->details->id, remote->id); - probe_then_start(remote, top); - return false; + if (guest->role == RSC_ROLE_STOPPED) { + // The guest is stopped, so we know no resource is active there + reason = "node's guest is stopped"; + probe_then_start(guest, top); + goto no_probe; } else if (guest_resource_will_stop(node)) { - /* The container is stopping or restarting, don't start - * 'rsc' until 'remote' stops as this also implies that - * 'rsc' is stopped - avoiding the need to probe - */ - pcmk__new_ordering(remote, pcmk__op_key(remote->id, RSC_STOP, 0), + reason = "node's guest will stop"; + + // Order resource start after guest stop (in case it's restarting) + pcmk__new_ordering(guest, pcmk__op_key(guest->id, RSC_STOP, 0), NULL, top, pcmk__op_key(top->id, RSC_START, 0), NULL, pe_order_optional, rsc->cluster); - pe_rsc_trace(rsc, "Skipping probe for %s on node %s, %s is stopping, restarting or moving", - rsc->id, node->details->id, remote->id); - return false; -/* } else { - * The container is running so there is no problem probing it - */ + goto no_probe; } } // We've eliminated all cases where a probe is not needed, so now it is probe = probe_action(rsc, node); + /* Order the probe relative to the parent -- or the resource itself if + * cloned or a fence device when unfencing is used. + */ if ((pcmk_is_set(rsc->flags, pe_rsc_fence_device) && pcmk_is_set(rsc->cluster->flags, pe_flag_enable_unfencing)) || !pe_rsc_is_clone(top)) { top = rsc; - } else { - crm_trace("Probing %s on %s (%s) as %s", rsc->id, node->details->uname, role2text(rsc->role), top->id); } if (!pcmk_is_set(probe->flags, pe_action_runnable) && (rsc->running_on == NULL)) { - /* Prevent the start from occurring if rsc isn't active, but - * don't cause it to stop if it was active already + /* Prevent the parent from starting if the resource can't, but don't + * cause the parent to stop if already active. */ pe__set_order_flags(flags, pe_order_runnable_left); } - pcmk__new_ordering(rsc, NULL, probe, top, - pcmk__op_key(top->id, RSC_START, 0), NULL, flags, - rsc->cluster); - - // Order the probe before any agent reload + // Start or reload the parent after probing the resource + pcmk__new_ordering(rsc, NULL, probe, + top, pcmk__op_key(top->id, RSC_START, 0), NULL, + flags, rsc->cluster); pcmk__new_ordering(rsc, NULL, probe, top, reload_key(rsc), NULL, pe_order_optional, rsc->cluster); return true; + +no_probe: + pe_rsc_trace(rsc, + "Skipping probe for %s on %s because %s", + rsc->id, node->details->id, reason); + return false; } /*! * \internal * \brief Check whether a probe should be ordered before another action * * \param[in] probe Probe action to check * \param[in] then Other action to check * * \return true if \p probe should be ordered before \p then, otherwise false */ static bool probe_needed_before_action(pe_action_t *probe, pe_action_t *then) { // Probes on a node are performed after unfencing it, not before if (pcmk__str_eq(then->task, CRM_OP_FENCE, pcmk__str_casei) && (probe->node != NULL) && (then->node != NULL) && (probe->node->details == then->node->details)) { const char *op = g_hash_table_lookup(then->meta, "stonith_action"); if (pcmk__str_eq(op, "on", pcmk__str_casei)) { return false; } } // Probes should be done on a node before shutting it down if (pcmk__str_eq(then->task, CRM_OP_SHUTDOWN, pcmk__str_none) && (probe->node != NULL) && (then->node != NULL) && (probe->node->details != then->node->details)) { return false; } // Otherwise probes should always be done before any other action return true; } /*! * \internal * \brief Add implicit "probe then X" orderings for "stop then X" orderings * * If the state of a resource is not known yet, a probe will be scheduled, * expecting a "not running" result. If the probe fails, a stop will not be * scheduled until the next transition. Thus, if there are ordering constraints * like "stop this resource then do something else that's not for the same * resource", add implicit "probe this resource then do something" equivalents * so the relation is upheld until we know whether a stop is needed. * * \param[in] data_set Cluster working set */ static void add_probe_orderings_for_stops(pe_working_set_t *data_set) { for (GList *iter = data_set->ordering_constraints; iter != NULL; iter = iter->next) { pe__ordering_t *order = iter->data; enum pe_ordering order_type = pe_order_optional; GList *probes = NULL; GList *then_actions = NULL; // Skip disabled orderings if (order->type == pe_order_none) { continue; } // Skip non-resource orderings, and orderings for the same resource if ((order->lh_rsc == NULL) || (order->lh_rsc == order->rh_rsc)) { continue; } // Skip invalid orderings (shouldn't be possible) if (((order->lh_action == NULL) && (order->lh_action_task == NULL)) || ((order->rh_action == NULL) && (order->rh_action_task == NULL))) { continue; } // Skip orderings for first actions other than stop if ((order->lh_action != NULL) && !pcmk__str_eq(order->lh_action->task, RSC_STOP, pcmk__str_none)) { continue; } else if ((order->lh_action == NULL) && !pcmk__ends_with(order->lh_action_task, "_" RSC_STOP "_0")) { continue; } /* Do not imply a probe ordering for a resource inside of a stopping * container. Otherwise, it might introduce a transition loop, since a * probe could be scheduled after the container starts again. */ if ((order->rh_rsc != NULL) && (order->lh_rsc->container == order->rh_rsc)) { if ((order->rh_action != NULL) && pcmk__str_eq(order->rh_action->task, RSC_STOP, pcmk__str_none)) { continue; } else if ((order->rh_action == NULL) && pcmk__ends_with(order->rh_action_task, "_" RSC_STOP "_0")) { continue; } } // Preserve certain order options for future filtering if (pcmk_is_set(order->type, pe_order_apply_first_non_migratable)) { pe__set_order_flags(order_type, pe_order_apply_first_non_migratable); } if (pcmk_is_set(order->type, pe_order_same_node)) { pe__set_order_flags(order_type, pe_order_same_node); } // Preserve certain order types for future filtering if ((order->type == pe_order_anti_colocation) || (order->type == pe_order_load)) { order_type = order->type; } // List all scheduled probes for the first resource probes = pe__resource_actions(order->lh_rsc, NULL, RSC_STATUS, FALSE); if (probes == NULL) { // There aren't any continue; } // List all relevant "then" actions if (order->rh_action != NULL) { then_actions = g_list_prepend(NULL, order->rh_action); } else if (order->rh_rsc != NULL) { then_actions = find_actions(order->rh_rsc->actions, order->rh_action_task, NULL); if (then_actions == NULL) { // There aren't any g_list_free(probes); continue; } } crm_trace("Implying 'probe then' orderings for '%s then %s' " "(id=%d, type=%.6x)", order->lh_action? order->lh_action->uuid : order->lh_action_task, order->rh_action? order->rh_action->uuid : order->rh_action_task, order->id, order->type); for (GList *probe_iter = probes; probe_iter != NULL; probe_iter = probe_iter->next) { pe_action_t *probe = (pe_action_t *) probe_iter->data; for (GList *then_iter = then_actions; then_iter != NULL; then_iter = then_iter->next) { pe_action_t *then = (pe_action_t *) then_iter->data; if (probe_needed_before_action(probe, then)) { order_actions(probe, then, order_type); } } } g_list_free(then_actions); g_list_free(probes); } } /*! * \internal * \brief Order probes before restarts and re-promotes * * If a given ordering is a "probe then start" or "probe then promote" ordering, * add an implicit "probe then stop/demote" ordering in case the action is part * of a restart/re-promote, and do the same recursively for all actions ordered * after the "then" action. * * \param[in] probe Probe as 'first' action in an ordering * \param[in] after 'then' action in the ordering * \param[in] data_set Cluster working set */ static void add_restart_orderings_for_probe(pe_action_t *probe, pe_action_t *after, pe_working_set_t *data_set) { GList *iter = NULL; bool interleave = false; pe_resource_t *compatible_rsc = NULL; // Validate that this is a resource probe followed by some action if ((after == NULL) || (probe == NULL) || (probe->rsc == NULL) || (probe->rsc->variant != pe_native) || !pcmk__str_eq(probe->task, RSC_STATUS, pcmk__str_casei)) { return; } // Avoid running into any possible loop if (pcmk_is_set(after->flags, pe_action_tracking)) { return; } pe__set_action_flags(after, pe_action_tracking); crm_trace("Adding probe restart orderings for '%s@%s then %s@%s'", probe->uuid, ((probe->node == NULL)? "" : probe->node->details->uname), after->uuid, ((after->node == NULL)? "" : after->node->details->uname)); /* Add restart orderings if "then" is for a different primitive. * Orderings for collective resources will be added later. */ if ((after->rsc != NULL) && (after->rsc->variant == pe_native) && (probe->rsc != after->rsc)) { GList *then_actions = NULL; if (pcmk__str_eq(after->task, RSC_START, pcmk__str_casei)) { then_actions = pe__resource_actions(after->rsc, NULL, RSC_STOP, FALSE); } else if (pcmk__str_eq(after->task, RSC_PROMOTE, pcmk__str_casei)) { then_actions = pe__resource_actions(after->rsc, NULL, RSC_DEMOTE, FALSE); } for (iter = then_actions; iter != NULL; iter = iter->next) { pe_action_t *then = (pe_action_t *) iter->data; // Skip pseudo-actions (for example, those implied by fencing) if (!pcmk_is_set(then->flags, pe_action_pseudo)) { order_actions(probe, then, pe_order_optional); } } g_list_free(then_actions); } /* Detect whether "then" is an interleaved clone action. For these, we want * to add orderings only for the relevant instance. */ if ((after->rsc != NULL) && (after->rsc->variant > pe_group)) { const char *interleave_s = g_hash_table_lookup(after->rsc->meta, XML_RSC_ATTR_INTERLEAVE); interleave = crm_is_true(interleave_s); if (interleave) { compatible_rsc = find_compatible_child(probe->rsc, after->rsc, RSC_ROLE_UNKNOWN, FALSE); } } /* Now recursively do the same for all actions ordered after "then". This * also handles collective resources since the collective action will be * ordered before its individual instances' actions. */ for (iter = after->actions_after; iter != NULL; iter = iter->next) { pe_action_wrapper_t *after_wrapper = (pe_action_wrapper_t *) iter->data; /* pe_order_implies_then is the reason why a required A.start * implies/enforces B.start to be required too, which is the cause of * B.restart/re-promote. * * Not sure about pe_order_implies_then_on_node though. It's now only * used for unfencing case, which tends to introduce transition * loops... */ if (!pcmk_is_set(after_wrapper->type, pe_order_implies_then)) { /* The order type between a group/clone and its child such as * B.start-> B_child.start is: * pe_order_implies_first_printed | pe_order_runnable_left * * Proceed through the ordering chain and build dependencies with * its children. */ if ((after->rsc == NULL) || (after->rsc->variant < pe_group) || (probe->rsc->parent == after->rsc) || (after_wrapper->action->rsc == NULL) || (after_wrapper->action->rsc->variant > pe_group) || (after->rsc != after_wrapper->action->rsc->parent)) { continue; } /* Proceed to the children of a group or a non-interleaved clone. * For an interleaved clone, proceed only to the relevant child. */ if ((after->rsc->variant > pe_group) && interleave && ((compatible_rsc == NULL) || (compatible_rsc != after_wrapper->action->rsc))) { continue; } } crm_trace("Recursively adding probe restart orderings for " "'%s@%s then %s@%s' (type=%#.6x)", after->uuid, ((after->node == NULL)? "" : after->node->details->uname), after_wrapper->action->uuid, ((after_wrapper->action->node == NULL)? "" : after_wrapper->action->node->details->uname), after_wrapper->type); add_restart_orderings_for_probe(probe, after_wrapper->action, data_set); } } /*! * \internal * \brief Clear the tracking flag on all scheduled actions * * \param[in] data_set Cluster working set */ static void clear_actions_tracking_flag(pe_working_set_t *data_set) { GList *gIter = NULL; for (gIter = data_set->actions; gIter != NULL; gIter = gIter->next) { pe_action_t *action = (pe_action_t *) gIter->data; pe__clear_action_flags(action, pe_action_tracking); } } /*! * \internal * \brief Add restart orderings for any scheduled probes for a given resource * * \param[in] rsc Resource whose probes should be ordered * \param[in] data_set Cluster working set */ static void add_restart_orderings_for_rsc(pe_resource_t *rsc, pe_working_set_t *data_set) { GList *probes = NULL; // For collective resources, order each instance recursively if (rsc->variant != pe_native) { g_list_foreach(rsc->children, (GFunc) add_restart_orderings_for_rsc, data_set); return; } // Find all probes for given resource probes = pe__resource_actions(rsc, NULL, RSC_STATUS, FALSE); // Add probe restart orderings for each probe found for (GList *iter = probes; iter != NULL; iter = iter->next) { pe_action_t *probe = (pe_action_t *) iter->data; for (GList *then_iter = probe->actions_after; then_iter != NULL; then_iter = then_iter->next) { pe_action_wrapper_t *then = (pe_action_wrapper_t *) then_iter->data; add_restart_orderings_for_probe(probe, then->action, data_set); clear_actions_tracking_flag(data_set); } } g_list_free(probes); } /*! * \internal * \brief Add "A then probe B" orderings for "A then B" orderings * * \param[in] data_set Cluster working set * * \note This function is currently disabled (see next comment). */ static void order_then_probes(pe_working_set_t *data_set) { #if 0 /* Given an ordering "A then B", we would prefer to wait for A to be started * before probing B. * * For example, if A is a filesystem which B can't even run without, it * would be helpful if the author of B's agent could assume that A is * running before B.monitor will be called. * * However, we can't _only_ probe after A is running, otherwise we wouldn't * detect the state of B if A could not be started. We can't even do an * opportunistic version of this, because B may be moving: * * A.stop -> A.start -> B.probe -> B.stop -> B.start * * and if we add B.stop -> A.stop here, we get a loop: * * A.stop -> A.start -> B.probe -> B.stop -> A.stop * * We could kill the "B.probe -> B.stop" dependency, but that could mean * stopping B "too" soon, because B.start must wait for the probe, and * we don't want to stop B if we can't start it. * * We could add the ordering only if A is an anonymous clone with * clone-max == node-max (since we'll never be moving it). However, we could * still be stopping one instance at the same time as starting another. * * The complexity of checking for allowed conditions combined with the ever * narrowing use case suggests that this code should remain disabled until * someone gets smarter. */ for (GList *iter = data_set->resources; iter != NULL; iter = iter->next) { pe_resource_t *rsc = (pe_resource_t *) iter->data; pe_action_t *start = NULL; GList *actions = NULL; GList *probes = NULL; actions = pe__resource_actions(rsc, NULL, RSC_START, FALSE); if (actions) { start = actions->data; g_list_free(actions); } if (start == NULL) { crm_err("No start action for %s", rsc->id); continue; } probes = pe__resource_actions(rsc, NULL, RSC_STATUS, FALSE); for (actions = start->actions_before; actions != NULL; actions = actions->next) { pe_action_wrapper_t *before = (pe_action_wrapper_t *) actions->data; pe_action_t *first = before->action; pe_resource_t *first_rsc = first->rsc; if (first->required_runnable_before) { for (GList *clone_actions = first->actions_before; clone_actions != NULL; clone_actions = clone_actions->next) { before = (pe_action_wrapper_t *) clone_actions->data; crm_trace("Testing '%s then %s' for %s", first->uuid, before->action->uuid, start->uuid); CRM_ASSERT(before->action->rsc != NULL); first_rsc = before->action->rsc; break; } } else if (!pcmk__str_eq(first->task, RSC_START, pcmk__str_none)) { crm_trace("Not a start op %s for %s", first->uuid, start->uuid); } if (first_rsc == NULL) { continue; } else if (uber_parent(first_rsc) == uber_parent(start->rsc)) { crm_trace("Same parent %s for %s", first_rsc->id, start->uuid); continue; } else if (!pe_rsc_is_clone(uber_parent(first_rsc))) { crm_trace("Not a clone %s for %s", first_rsc->id, start->uuid); continue; } crm_err("Applying %s before %s %d", first->uuid, start->uuid, uber_parent(first_rsc)->variant); for (GList *probe_iter = probes; probe_iter != NULL; probe_iter = probe_iter->next) { pe_action_t *probe = (pe_action_t *) probe_iter->data; crm_err("Ordering %s before %s", first->uuid, probe->uuid); order_actions(first, probe, pe_order_optional); } } } #endif } void pcmk__order_probes(pe_working_set_t *data_set) { // Add orderings for "probe then X" g_list_foreach(data_set->resources, (GFunc) add_restart_orderings_for_rsc, data_set); add_probe_orderings_for_stops(data_set); order_then_probes(data_set); } /*! * \internal * \brief Schedule any probes needed * * \param[in] data_set Cluster working set * * \note This may also schedule fencing of failed remote nodes. */ void pcmk__schedule_probes(pe_working_set_t *data_set) { // Schedule probes on each node in the cluster as needed for (GList *iter = data_set->nodes; iter != NULL; iter = iter->next) { pe_node_t *node = (pe_node_t *) iter->data; const char *probed = NULL; if (!node->details->online) { // Don't probe offline nodes if (pcmk__is_failed_remote_node(node)) { pe_fence_node(data_set, node, "the connection is unrecoverable", FALSE); } continue; } else if (node->details->unclean) { // ... or nodes that need fencing continue; } else if (!node->details->rsc_discovery_enabled) { // The user requested that probes not be done on this node continue; } /* This is no longer needed for live clusters, since the probe_complete * node attribute will never be in the CIB. However this is still useful * for processing old saved CIBs (< 1.1.14), including the * reprobe-target_rc regression test. */ probed = pe_node_attribute_raw(node, CRM_OP_PROBED); if (probed != NULL && crm_is_true(probed) == FALSE) { pe_action_t *probe_op = NULL; probe_op = custom_action(NULL, crm_strdup_printf("%s-%s", CRM_OP_REPROBE, node->details->uname), CRM_OP_REPROBE, node, FALSE, TRUE, data_set); add_hash_param(probe_op->meta, XML_ATTR_TE_NOWAIT, XML_BOOLEAN_TRUE); continue; } // Probe each resource in the cluster on this node, as needed pcmk__probe_resource_list(data_set->resources, node); } } diff --git a/lib/pacemaker/pcmk_sched_resource.c b/lib/pacemaker/pcmk_sched_resource.c index 0915cc2632..9999f33823 100644 --- a/lib/pacemaker/pcmk_sched_resource.c +++ b/lib/pacemaker/pcmk_sched_resource.c @@ -1,1088 +1,1088 @@ /* * Copyright 2014-2022 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__native_allocate, native_create_actions, - native_create_probe, + pcmk__probe_rsc_on_node, native_internal_constraints, pcmk__primitive_apply_coloc_score, pcmk__add_colocated_node_scores, pcmk__colocated_resources, native_rsc_location, native_action_flags, pcmk__update_ordered_actions, pcmk__output_resource_actions, pcmk__add_rsc_actions_to_graph, native_append_meta, pcmk__primitive_add_utilization, pcmk__primitive_shutdown_lock, }, { pcmk__group_allocate, group_create_actions, - native_create_probe, + pcmk__probe_rsc_on_node, group_internal_constraints, pcmk__group_apply_coloc_score, pcmk__group_add_colocated_node_scores, pcmk__group_colocated_resources, group_rsc_location, group_action_flags, group_update_actions, pcmk__output_resource_actions, pcmk__add_rsc_actions_to_graph, group_append_meta, pcmk__group_add_utilization, pcmk__group_shutdown_lock, }, { pcmk__clone_allocate, clone_create_actions, clone_create_probe, clone_internal_constraints, pcmk__clone_apply_coloc_score, pcmk__add_colocated_node_scores, pcmk__colocated_resources, clone_rsc_location, clone_action_flags, pcmk__multi_update_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__add_colocated_node_scores, pcmk__colocated_resources, pcmk__bundle_rsc_location, pcmk__bundle_action_flags, pcmk__multi_update_actions, pcmk__output_bundle_actions, pcmk__bundle_expand, pcmk__bundle_append_meta, pcmk__bundle_add_utilization, pcmk__bundle_shutdown_lock, } }; /*! * \internal * \brief Check whether a resource's agent standard, provider, or type changed * * \param[in] rsc Resource to check * \param[in] node Node needing unfencing/restart 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, node->details->uname, 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, 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] rsc Resource to set allocation methods for * \param[in] ignored Only 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] 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(pe_resource_t *rsc, pe_resource_t *orig_rsc, GList *colocated_rscs) { GList *gIter = 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_append(colocated_rscs, rsc); // Follow colocations where this resource is the dependent resource for (gIter = rsc->rsc_cons; gIter != NULL; gIter = gIter->next) { pcmk__colocation_t *constraint = (pcmk__colocation_t *) gIter->data; 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); } } // Follow colocations where this resource is the primary resource for (gIter = rsc->rsc_cons_lhs; gIter != NULL; gIter = gIter->next) { pcmk__colocation_t *constraint = (pcmk__colocation_t *) gIter->data; 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); } } return colocated_rscs; } 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 * We fiddle with the current role in native_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] rsc Resource to assign * \param[in] 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_primitive(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, chosen->details->uname, (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, chosen->details->uname); 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] rsc Resource to assign * \param[in] 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__assign_primitive(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] 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, old->details->uname); 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] rsc Resource to check * \param[in] node Node to check * \param[out] failed If the threshold has been reached, this will be set to * the 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, 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, rsc->cluster); 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, node->details->uname, 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), node->details->uname, 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(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; GList *nodes = (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); 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 (GList *iter = nodes; iter != NULL; iter = iter->next) { pe_node_t *node = (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] 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); } /*! * \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; 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); } 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); } } /*! * \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, node1->details->uname, instance2->id, node2->weight, node2->details->uname); rc = 1; } else if (node1->weight > node2->weight) { crm_trace("Assign %s (%d on %s) before %s (%d on %s)", instance1->id, node1->weight, node1->details->uname, instance2->id, node2->weight, node2->details->uname); 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((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, (*node)->details->uname); *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 * - Failed instance * - 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 = pe__find_active_on(instance1, &nnodes1, NULL); node2 = pe__find_active_on(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 failed instance can1 = did_fail(instance1); can2 = did_fail(instance2); if (!can1 && can2) { crm_trace("Assign %s before %s: failed", instance1->id, instance2->id); return -1; } else if (can1 && !can2) { crm_trace("Assign %s after %s: not 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; }