diff --git a/lib/pacemaker/libpacemaker_private.h b/lib/pacemaker/libpacemaker_private.h index f222f6eacc..a46aa34291 100644 --- a/lib/pacemaker/libpacemaker_private.h +++ b/lib/pacemaker/libpacemaker_private.h @@ -1,386 +1,390 @@ /* * 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 // 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 void pcmk__log_action(const char *pre_text, pe_action_t *action, bool details); G_GNUC_INTERNAL pe_action_t *pcmk__new_rsc_pseudo_action(pe_resource_t *rsc, const char *task, bool optional, bool runnable); 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, pe_working_set_t *data_set); 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__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, pe_working_set_t *data_set); G_GNUC_INTERNAL void pcmk__fence_guest(pe_node_t *node, pe_working_set_t *data_set); G_GNUC_INTERNAL bool pcmk__node_unfenced(pe_node_t *node); G_GNUC_INTERNAL bool pcmk__is_unfence_device(const pe_resource_t *rsc, const pe_working_set_t *data_set); // 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 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__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); G_GNUC_INTERNAL void pcmk__new_ordering(pe_resource_t *lh_rsc, char *lh_task, pe_action_t *lh_action, pe_resource_t *rh_rsc, char *rh_task, pe_action_t *rh_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, pe_working_set_t *data_set); G_GNUC_INTERNAL void pcmk__apply_orderings(pe_working_set_t *data_set); /*! * \internal * \brief Create a new ordering between two resource actions * * \param[in] lh_rsc Resource for 'first' action * \param[in] rh_rsc Resource for 'then' action * \param[in] lh_task Action key for 'first' action * \param[in] rh_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(lh_rsc, lh_task, rh_rsc, rh_task, \ flags, data_set) \ pcmk__new_ordering((lh_rsc), pcmk__op_key((lh_rsc)->id, (lh_task), 0), \ NULL, \ (rh_rsc), pcmk__op_key((rh_rsc)->id, (rh_task), 0), \ NULL, (flags), (data_set)) #define pcmk__order_starts(rsc1, rsc2, type, data_set) \ pcmk__order_resource_actions((rsc1), CRMD_ACTION_START, \ (rsc2), CRMD_ACTION_START, (type), (data_set)) #define pcmk__order_stops(rsc1, rsc2, type, data_set) \ pcmk__order_resource_actions((rsc1), CRMD_ACTION_STOP, \ (rsc2), CRMD_ACTION_STOP, (type), (data_set)) G_GNUC_INTERNAL void pcmk__unpack_rsc_ticket(xmlNode *xml_obj, pe_working_set_t *data_set); 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, pe_working_set_t *data_set); G_GNUC_INTERNAL void pcmk__add_bundle_meta_to_xml(xmlNode *args_xml, pe_action_t *action); // Groups (pcmk_sched_group.c) G_GNUC_INTERNAL GList *pcmk__group_colocated_resources(pe_resource_t *rsc, pe_resource_t *orig_rsc, GList *colocated_rscs); // Bundles (pcmk_sched_bundle.c) 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); 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, pe_working_set_t *data_set); G_GNUC_INTERNAL void pcmk__apply_node_health(pe_working_set_t *data_set); // Clone notifictions (pcmk_sched_notif.c) G_GNUC_INTERNAL void pcmk__create_notifications(pe_resource_t *rsc, notify_data_t *n_data); G_GNUC_INTERNAL notify_data_t *pcmk__clone_notif_pseudo_ops(pe_resource_t *rsc, const char *task, pe_action_t *action, pe_action_t *complete); G_GNUC_INTERNAL void pcmk__free_notification_data(notify_data_t *n_data); G_GNUC_INTERNAL void pcmk__order_notifs_after_fencing(pe_action_t *action, pe_resource_t *rsc, pe_action_t *stonith_op); // 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); // Functions related to probes (pcmk_sched_probes.c) G_GNUC_INTERNAL void pcmk__order_probes(pe_working_set_t *data_set); 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, pe_working_set_t *data_set); -G_GNUC_INTERNAL void pcmk__create_utilization_constraints(pe_resource_t *rsc, - GList *allowed_nodes); +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_allocate.c b/lib/pacemaker/pcmk_sched_allocate.c index aa28748c36..4a426b0da4 100644 --- a/lib/pacemaker/pcmk_sched_allocate.c +++ b/lib/pacemaker/pcmk_sched_allocate.c @@ -1,770 +1,755 @@ /* * 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 #include #include #include #include "libpacemaker_private.h" CRM_TRACE_INIT_DATA(pacemaker); /*! * \internal * \brief Do deferred action checks after allocation * * When unpacking the resource history, the scheduler checks for resource * configurations that have changed since an action was run. However, at that * time, bundles using the REMOTE_CONTAINER_HACK don't have their final * parameter information, so instead they add a deferred check to a list. This * function processes one entry in that list. * * \param[in] rsc Resource that action history is for * \param[in] node Node that action history is for * \param[in] rsc_op Action history entry * \param[in] check Type of deferred check to do * \param[in] data_set Working set for cluster */ static void check_params(pe_resource_t *rsc, pe_node_t *node, xmlNode *rsc_op, enum pe_check_parameters check, pe_working_set_t *data_set) { const char *reason = NULL; op_digest_cache_t *digest_data = NULL; switch (check) { case pe_check_active: if (pcmk__check_action_config(rsc, node, rsc_op) && pe_get_failcount(node, rsc, NULL, pe_fc_effective, NULL, data_set)) { reason = "action definition changed"; } break; case pe_check_last_failure: digest_data = rsc_action_digest_cmp(rsc, rsc_op, node, data_set); switch (digest_data->rc) { case RSC_DIGEST_UNKNOWN: crm_trace("Resource %s history entry %s on %s has " "no digest to compare", rsc->id, ID(rsc_op), node->details->id); break; case RSC_DIGEST_MATCH: break; default: reason = "resource parameters have changed"; break; } break; } if (reason != NULL) { pe__clear_failcount(rsc, node, reason, data_set); } } /*! * \internal * \brief Check whether a resource has failcount clearing scheduled on a node * * \param[in] node Node to check * \param[in] rsc Resource to check * * \return true if \p rsc has failcount clearing scheduled on \p node, * otherwise false */ static bool failcount_clear_action_exists(pe_node_t *node, pe_resource_t *rsc) { GList *list = pe__resource_actions(rsc, node, CRM_OP_CLEAR_FAILCOUNT, TRUE); if (list != NULL) { g_list_free(list); return true; } return false; } /*! * \internal * \brief Ban a resource from a node if it reached its failure threshold there * * \param[in] rsc Resource to check failure threshold for * \param[in] node Node to check \p rsc on */ static void check_failure_threshold(pe_resource_t *rsc, pe_node_t *node) { // If this is a collective resource, apply recursively to children instead if (rsc->children != NULL) { g_list_foreach(rsc->children, (GFunc) check_failure_threshold, node); return; } else if (failcount_clear_action_exists(node, rsc)) { /* Don't force the resource away from this node due to a failcount * that's going to be cleared. * * @TODO Failcount clearing can be scheduled in * pcmk__handle_rsc_config_changes() via process_rsc_history(), or in * stage5() via check_params(). This runs well before then, so it cannot * detect those, meaning we might check the migration threshold when we * shouldn't. Worst case, we stop or move the resource, then move it * back in the next transition. */ return; } else { pe_resource_t *failed = NULL; if (pcmk__threshold_reached(rsc, node, &failed)) { resource_location(failed, node, -INFINITY, "__fail_limit__", rsc->cluster); } } } /*! * \internal * \brief If resource has exclusive discovery, ban node if not allowed * * Location constraints have a resource-discovery option that allows users to * specify where probes are done for the affected resource. If this is set to * exclusive, probes will only be done on nodes listed in exclusive constraints. * This function bans the resource from the node if the node is not listed. * * \param[in] rsc Resource to check * \param[in] node Node to check \p rsc on */ static void apply_exclusive_discovery(pe_resource_t *rsc, pe_node_t *node) { if (rsc->exclusive_discover || uber_parent(rsc)->exclusive_discover) { pe_node_t *match = NULL; // If this is a collective resource, apply recursively to children g_list_foreach(rsc->children, (GFunc) apply_exclusive_discovery, node); match = g_hash_table_lookup(rsc->allowed_nodes, node->details->id); if ((match != NULL) && (match->rsc_discover_mode != pe_discover_exclusive)) { match->weight = -INFINITY; } } } /*! * \internal * \brief Apply stickiness to a resource if appropriate * * \param[in] rsc Resource to check for stickiness * \param[in] data_set Cluster working set */ static void apply_stickiness(pe_resource_t *rsc, pe_working_set_t *data_set) { pe_node_t *node = NULL; // If this is a collective resource, apply recursively to children instead if (rsc->children != NULL) { g_list_foreach(rsc->children, (GFunc) apply_stickiness, data_set); return; } /* A resource is sticky if it is managed, has stickiness configured, and is * active on a single node. */ if (!pcmk_is_set(rsc->flags, pe_rsc_managed) || (rsc->stickiness < 1) || !pcmk__list_of_1(rsc->running_on)) { return; } node = rsc->running_on->data; /* In a symmetric cluster, stickiness can always be used. In an * asymmetric cluster, we have to check whether the resource is still * allowed on the node, so we don't keep the resource somewhere it is no * longer explicitly enabled. */ if (!pcmk_is_set(rsc->cluster->flags, pe_flag_symmetric_cluster) && (pe_hash_table_lookup(rsc->allowed_nodes, node->details->id) == NULL)) { pe_rsc_debug(rsc, "Ignoring %s stickiness because the cluster is " "asymmetric and node %s is not explicitly allowed", rsc->id, node->details->uname); return; } pe_rsc_debug(rsc, "Resource %s has %d stickiness on node %s", rsc->id, rsc->stickiness, node->details->uname); resource_location(rsc, node, rsc->stickiness, "stickiness", rsc->cluster); } /*! * \internal * \brief Apply shutdown locks for all resources as appropriate * * \param[in] data_set Cluster working set */ static void apply_shutdown_locks(pe_working_set_t *data_set) { if (pcmk_is_set(data_set->flags, pe_flag_shutdown_lock)) { for (GList *iter = data_set->resources; iter != NULL; iter = iter->next) { pe_resource_t *rsc = (pe_resource_t *) iter->data; rsc->cmds->shutdown_lock(rsc); } } } /*! * \internal * \brief Calculate the number of available nodes in the cluster * * \param[in] data_set Cluster working set */ static void count_available_nodes(pe_working_set_t *data_set) { if (pcmk_is_set(data_set->flags, pe_flag_no_compat)) { return; } // @COMPAT for API backward compatibility only (cluster does not use value) for (GList *iter = data_set->nodes; iter != NULL; iter = iter->next) { pe_node_t *node = (pe_node_t *) iter->data; if ((node != NULL) && (node->weight >= 0) && node->details->online && (node->details->type != node_ping)) { data_set->max_valid_nodes++; } } crm_trace("Online node count: %d", data_set->max_valid_nodes); } /* * \internal * \brief Apply node-specific scheduling criteria * * After the CIB has been unpacked, process node-specific scheduling criteria * including shutdown locks, location constraints, resource stickiness, * migration thresholds, and exclusive resource discovery. */ static void apply_node_criteria(pe_working_set_t *data_set) { crm_trace("Applying node-specific scheduling criteria"); apply_shutdown_locks(data_set); count_available_nodes(data_set); pcmk__apply_locations(data_set); g_list_foreach(data_set->resources, (GFunc) apply_stickiness, data_set); for (GList *node_iter = data_set->nodes; node_iter != NULL; node_iter = node_iter->next) { for (GList *rsc_iter = data_set->resources; rsc_iter != NULL; rsc_iter = rsc_iter->next) { pe_node_t *node = (pe_node_t *) node_iter->data; pe_resource_t *rsc = (pe_resource_t *) rsc_iter->data; check_failure_threshold(rsc, node); apply_exclusive_discovery(rsc, node); } } } /*! * \internal * \brief Allocate resources to nodes * * \param[in] data_set Cluster working set */ static void allocate_resources(pe_working_set_t *data_set) { GList *iter = NULL; crm_trace("Allocating resources to nodes"); if (pcmk_is_set(data_set->flags, pe_flag_have_remote_nodes)) { /* Allocate remote connection resources first (which will also allocate * any colocation dependencies). If the connection is migrating, always * prefer the partial migration target. */ for (iter = data_set->resources; iter != NULL; iter = iter->next) { pe_resource_t *rsc = (pe_resource_t *) iter->data; if (rsc->is_remote_node) { pe_rsc_trace(rsc, "Allocating remote connection resource '%s'", rsc->id); rsc->cmds->allocate(rsc, rsc->partial_migration_target, data_set); } } } /* now do the rest of the resources */ for (iter = data_set->resources; iter != NULL; iter = iter->next) { pe_resource_t *rsc = (pe_resource_t *) iter->data; if (!rsc->is_remote_node) { pe_rsc_trace(rsc, "Allocating %s resource '%s'", crm_element_name(rsc->xml), rsc->id); rsc->cmds->allocate(rsc, NULL, data_set); } } } /*! * \internal * \brief Schedule fail count clearing on online nodes if resource is orphaned * * \param[in] rsc Resource to check * \param[in] data_set Cluster working set */ static void clear_failcounts_if_orphaned(pe_resource_t *rsc, pe_working_set_t *data_set) { if (!pcmk_is_set(rsc->flags, pe_rsc_orphan)) { return; } crm_trace("Clear fail counts for orphaned resource %s", rsc->id); /* There's no need to recurse into rsc->children because those * should just be unallocated clone instances. */ for (GList *iter = data_set->nodes; iter != NULL; iter = iter->next) { pe_node_t *node = (pe_node_t *) iter->data; pe_action_t *clear_op = NULL; if (!node->details->online) { continue; } if (pe_get_failcount(node, rsc, NULL, pe_fc_effective, NULL, data_set) == 0) { continue; } clear_op = pe__clear_failcount(rsc, node, "it is orphaned", data_set); /* We can't use order_action_then_stop() here because its * pe_order_preserve breaks things */ pcmk__new_ordering(clear_op->rsc, NULL, clear_op, rsc, stop_key(rsc), NULL, pe_order_optional, data_set); } } gboolean stage5(pe_working_set_t * data_set) { - pcmk__output_t *out = data_set->priv; GList *gIter = NULL; if (!pcmk__str_eq(data_set->placement_strategy, "default", pcmk__str_casei)) { pcmk__sort_resources(data_set); } - gIter = data_set->nodes; - for (; gIter != NULL; gIter = gIter->next) { - pe_node_t *node = (pe_node_t *) gIter->data; - - if (pcmk_is_set(data_set->flags, pe_flag_show_utilization)) { - out->message(out, "node-capacity", node, "Original"); - } - } + pcmk__show_node_capacities("Original", data_set); /* Take (next) highest resource, assign it and create its actions */ allocate_resources(data_set); - gIter = data_set->nodes; - for (; gIter != NULL; gIter = gIter->next) { - pe_node_t *node = (pe_node_t *) gIter->data; - - if (pcmk_is_set(data_set->flags, pe_flag_show_utilization)) { - out->message(out, "node-capacity", node, "Remaining"); - } - } + pcmk__show_node_capacities("Remaining", data_set); // Process deferred action checks pe__foreach_param_check(data_set, check_params); pe__free_param_checks(data_set); if (pcmk_is_set(data_set->flags, pe_flag_startup_probes)) { crm_trace("Calculating needed probes"); pcmk__schedule_probes(data_set); } if (pcmk_is_set(data_set->flags, pe_flag_stop_rsc_orphans)) { g_list_foreach(data_set->resources, (GFunc) clear_failcounts_if_orphaned, data_set); } crm_trace("Creating actions"); for (gIter = data_set->resources; gIter != NULL; gIter = gIter->next) { pe_resource_t *rsc = (pe_resource_t *) gIter->data; rsc->cmds->create_actions(rsc, data_set); } crm_trace("Creating done"); return TRUE; } static gboolean is_managed(const pe_resource_t * rsc) { GList *gIter = rsc->children; if (pcmk_is_set(rsc->flags, pe_rsc_managed)) { return TRUE; } for (; gIter != NULL; gIter = gIter->next) { pe_resource_t *child_rsc = (pe_resource_t *) gIter->data; if (is_managed(child_rsc)) { return TRUE; } } return FALSE; } static gboolean any_managed_resources(pe_working_set_t * data_set) { GList *gIter = data_set->resources; for (; gIter != NULL; gIter = gIter->next) { pe_resource_t *rsc = (pe_resource_t *) gIter->data; if (is_managed(rsc)) { return TRUE; } } return FALSE; } /* * Create dependencies for stonith and shutdown operations */ gboolean stage6(pe_working_set_t * data_set) { pe_action_t *dc_down = NULL; pe_action_t *stonith_op = NULL; gboolean integrity_lost = FALSE; gboolean need_stonith = TRUE; GList *gIter; GList *stonith_ops = NULL; GList *shutdown_ops = NULL; /* Remote ordering constraints need to happen prior to calculating fencing * because it is one more place we can mark nodes as needing fencing. */ pcmk__order_remote_connection_actions(data_set); crm_trace("Processing fencing and shutdown cases"); if (any_managed_resources(data_set) == FALSE) { crm_notice("Delaying fencing operations until there are resources to manage"); need_stonith = FALSE; } /* Check each node for stonith/shutdown */ for (gIter = data_set->nodes; gIter != NULL; gIter = gIter->next) { pe_node_t *node = (pe_node_t *) gIter->data; /* Guest nodes are "fenced" by recovering their container resource, * so handle them separately. */ if (pe__is_guest_node(node)) { if (node->details->remote_requires_reset && need_stonith && pe_can_fence(data_set, node)) { pcmk__fence_guest(node, data_set); } continue; } stonith_op = NULL; if (node->details->unclean && need_stonith && pe_can_fence(data_set, node)) { stonith_op = pe_fence_op(node, NULL, FALSE, "node is unclean", FALSE, data_set); pe_warn("Scheduling Node %s for STONITH", node->details->uname); pcmk__order_vs_fence(stonith_op, data_set); if (node->details->is_dc) { // Remember if the DC is being fenced dc_down = stonith_op; } else { if (!pcmk_is_set(data_set->flags, pe_flag_concurrent_fencing) && (stonith_ops != NULL)) { /* Concurrent fencing is disabled, so order each non-DC * fencing in a chain. If there is any DC fencing or * shutdown, it will be ordered after the last action in the * chain later. */ order_actions((pe_action_t *) stonith_ops->data, stonith_op, pe_order_optional); } // Remember all non-DC fencing actions in a separate list stonith_ops = g_list_prepend(stonith_ops, stonith_op); } } else if (node->details->online && node->details->shutdown && /* TODO define what a shutdown op means for a remote node. * For now we do not send shutdown operations for remote nodes, but * if we can come up with a good use for this in the future, we will. */ pe__is_guest_or_remote_node(node) == FALSE) { pe_action_t *down_op = pcmk__new_shutdown_action(node, data_set); if (node->details->is_dc) { // Remember if the DC is being shut down dc_down = down_op; } else { // Remember non-DC shutdowns for later ordering shutdown_ops = g_list_prepend(shutdown_ops, down_op); } } if (node->details->unclean && stonith_op == NULL) { integrity_lost = TRUE; pe_warn("Node %s is unclean!", node->details->uname); } } if (integrity_lost) { if (!pcmk_is_set(data_set->flags, pe_flag_stonith_enabled)) { pe_warn("YOUR RESOURCES ARE NOW LIKELY COMPROMISED"); pe_err("ENABLE STONITH TO KEEP YOUR RESOURCES SAFE"); } else if (!pcmk_is_set(data_set->flags, pe_flag_have_quorum)) { crm_notice("Cannot fence unclean nodes until quorum is" " attained (or no-quorum-policy is set to ignore)"); } } if (dc_down != NULL) { /* Order any non-DC shutdowns before any DC shutdown, to avoid repeated * DC elections. However, we don't want to order non-DC shutdowns before * a DC *fencing*, because even though we don't want a node that's * shutting down to become DC, the DC fencing could be ordered before a * clone stop that's also ordered before the shutdowns, thus leading to * a graph loop. */ if (pcmk__str_eq(dc_down->task, CRM_OP_SHUTDOWN, pcmk__str_casei)) { for (gIter = shutdown_ops; gIter != NULL; gIter = gIter->next) { pe_action_t *node_stop = (pe_action_t *) gIter->data; crm_debug("Ordering shutdown on %s before %s on DC %s", node_stop->node->details->uname, dc_down->task, dc_down->node->details->uname); order_actions(node_stop, dc_down, pe_order_optional); } } // Order any non-DC fencing before any DC fencing or shutdown if (pcmk_is_set(data_set->flags, pe_flag_concurrent_fencing)) { /* With concurrent fencing, order each non-DC fencing action * separately before any DC fencing or shutdown. */ for (gIter = stonith_ops; gIter != NULL; gIter = gIter->next) { order_actions((pe_action_t *) gIter->data, dc_down, pe_order_optional); } } else if (stonith_ops) { /* Without concurrent fencing, the non-DC fencing actions are * already ordered relative to each other, so we just need to order * the DC fencing after the last action in the chain (which is the * first item in the list). */ order_actions((pe_action_t *) stonith_ops->data, dc_down, pe_order_optional); } } g_list_free(stonith_ops); g_list_free(shutdown_ops); return TRUE; } static void log_resource_details(pe_working_set_t *data_set) { pcmk__output_t *out = data_set->priv; GList *all = NULL; /* We need a list of nodes that we are allowed to output information for. * This is necessary because out->message for all the resource-related * messages expects such a list, due to the `crm_mon --node=` feature. Here, * we just make it a list of all the nodes. */ all = g_list_prepend(all, (gpointer) "*"); for (GList *item = data_set->resources; item != NULL; item = item->next) { pe_resource_t *rsc = (pe_resource_t *) item->data; // Log all resources except inactive orphans if (!pcmk_is_set(rsc->flags, pe_rsc_orphan) || (rsc->role != RSC_ROLE_STOPPED)) { out->message(out, crm_map_element_name(rsc->xml), 0, rsc, all, all); } } g_list_free(all); } static void log_all_actions(pe_working_set_t *data_set) { /* This only ever outputs to the log, so ignore whatever output object was * previously set and just log instead. */ pcmk__output_t *prev_out = data_set->priv; pcmk__output_t *out = pcmk__new_logger(); if (out == NULL) { return; } pcmk__output_set_log_level(out, LOG_NOTICE); data_set->priv = out; out->begin_list(out, NULL, NULL, "Actions"); pcmk__output_actions(data_set); out->end_list(out); out->finish(out, CRM_EX_OK, true, NULL); pcmk__output_free(out); data_set->priv = prev_out; } /*! * \internal * \brief Log all required but unrunnable actions at trace level * * \param[in] data_set Cluster working set */ static void log_unrunnable_actions(pe_working_set_t *data_set) { const uint64_t flags = pe_action_optional|pe_action_runnable|pe_action_pseudo; crm_trace("Required but unrunnable actions:"); for (GList *iter = data_set->actions; iter != NULL; iter = iter->next) { pe_action_t *action = (pe_action_t *) iter->data; if (!pcmk_any_flags_set(action->flags, flags)) { pcmk__log_action("\t", action, true); } } } /*! * \internal * \brief Unpack the CIB for scheduling * * \param[in] cib CIB XML to unpack (may be NULL if previously unpacked) * \param[in] flags Working set flags to set in addition to defaults * \param[in] data_set Cluster working set */ static void unpack_cib(xmlNode *cib, unsigned long long flags, pe_working_set_t *data_set) { if (pcmk_is_set(data_set->flags, pe_flag_have_status)) { crm_trace("Reusing previously calculated cluster status"); pe__set_working_set_flags(data_set, flags); return; } CRM_ASSERT(cib != NULL); crm_trace("Calculating cluster status"); /* This will zero the entire struct without freeing anything first, so * callers should never call pcmk__schedule_actions() with a populated data * set unless pe_flag_have_status is set (i.e. cluster_status() was * previously called, whether directly or via pcmk__schedule_actions()). */ set_working_set_defaults(data_set); pe__set_working_set_flags(data_set, flags); data_set->input = cib; cluster_status(data_set); // Sets pe_flag_have_status } /*! * \internal * \brief Run the scheduler for a given CIB * * \param[in] cib CIB XML to use as scheduler input * \param[in] flags Working set flags to set in addition to defaults * \param[in,out] data_set Cluster working set */ void pcmk__schedule_actions(xmlNode *cib, unsigned long long flags, pe_working_set_t *data_set) { unpack_cib(cib, flags, data_set); pcmk__set_allocation_methods(data_set); pcmk__apply_node_health(data_set); pcmk__unpack_constraints(data_set); if (pcmk_is_set(data_set->flags, pe_flag_check_config)) { return; } if (!pcmk_is_set(data_set->flags, pe_flag_quick_location) && pcmk__is_daemon) { log_resource_details(data_set); } apply_node_criteria(data_set); if (pcmk_is_set(data_set->flags, pe_flag_quick_location)) { return; } pcmk__create_internal_constraints(data_set); pcmk__handle_rsc_config_changes(data_set); crm_trace("Allocate resources"); stage5(data_set); crm_trace("Processing fencing and shutdown cases"); stage6(data_set); pcmk__apply_orderings(data_set); log_all_actions(data_set); crm_trace("Create transition graph"); pcmk__create_graph(data_set); if (get_crm_log_level() == LOG_TRACE) { log_unrunnable_actions(data_set); } } diff --git a/lib/pacemaker/pcmk_sched_utilization.c b/lib/pacemaker/pcmk_sched_utilization.c index ecbcc33c4b..c55d374473 100644 --- a/lib/pacemaker/pcmk_sched_utilization.c +++ b/lib/pacemaker/pcmk_sched_utilization.c @@ -1,443 +1,463 @@ /* * 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 "libpacemaker_private.h" /*! * \internal * \brief Get integer utilization from a string * * \param[in] s String representation of a node utilization value * * \return Integer equivalent of \p s * \todo It would make sense to restrict utilization values to nonnegative * integers, but the documentation just says "integers" and we didn't * restrict them initially, so for backward compatibility, allow any * integer. */ static int utilization_value(const char *s) { int value = 0; if ((s != NULL) && (pcmk__scan_min_int(s, &value, INT_MIN) == EINVAL)) { pe_warn("Using 0 for utilization instead of invalid value '%s'", value); value = 0; } return value; } /* * Functions for comparing node capacities */ struct compare_data { const pe_node_t *node1; const pe_node_t *node2; bool node2_only; int result; }; /*! * \internal * \brief Compare a single utilization attribute for two nodes * * Compare one utilization attribute for two nodes, incrementing the result if * the first node has greater capacity, and decrementing it if the second node * has greater capacity. * * \param[in] key Utilization attribute name to compare * \param[in] value Utilization attribute value to compare * \param[in] user_data Comparison data (as struct compare_data*) */ static void compare_utilization_value(gpointer key, gpointer value, gpointer user_data) { int node1_capacity = 0; int node2_capacity = 0; struct compare_data *data = user_data; const char *node2_value = NULL; if (data->node2_only) { if (g_hash_table_lookup(data->node1->details->utilization, key)) { return; // We've already compared this attribute } } else { node1_capacity = utilization_value((const char *) value); } node2_value = g_hash_table_lookup(data->node2->details->utilization, key); node2_capacity = utilization_value(node2_value); if (node1_capacity > node2_capacity) { data->result--; } else if (node1_capacity < node2_capacity) { data->result++; } } /*! * \internal * \brief Compare utilization capacities of two nodes * * \param[in] node1 First node to compare * \param[in] node2 Second node to compare * * \return Negative integer if node1 has more free capacity, * 0 if the capacities are equal, or a positive integer * if node2 has more free capacity */ int pcmk__compare_node_capacities(const pe_node_t *node1, const pe_node_t *node2) { struct compare_data data = { .node1 = node1, .node2 = node2, .node2_only = false, .result = 0, }; // Compare utilization values that node1 and maybe node2 have g_hash_table_foreach(node1->details->utilization, compare_utilization_value, &data); // Compare utilization values that only node2 has data.node2_only = true; g_hash_table_foreach(node2->details->utilization, compare_utilization_value, &data); return data.result; } /* * Functions for updating node capacities */ struct calculate_data { GHashTable *current_utilization; bool plus; }; /*! * \internal * \brief Update a single utilization attribute with a new value * * \param[in] key Name of utilization attribute to update * \param[in] value Value to add or substract * \param[in] user_data Calculation data (as struct calculate_data *) */ static void update_utilization_value(gpointer key, gpointer value, gpointer user_data) { int result = 0; const char *current = NULL; struct calculate_data *data = user_data; current = g_hash_table_lookup(data->current_utilization, key); if (data->plus) { result = utilization_value(current) + utilization_value(value); } else if (current) { result = utilization_value(current) - utilization_value(value); } g_hash_table_replace(data->current_utilization, strdup(key), pcmk__itoa(result)); } /*! * \internal * \brief Subtract a resource's utilization from node capacity * * \param[in] current_utilization Current node utilization attributes * \param[in] rsc Resource with utilization to subtract */ void pcmk__consume_node_capacity(GHashTable *current_utilization, pe_resource_t *rsc) { struct calculate_data data = { .current_utilization = current_utilization, .plus = false, }; g_hash_table_foreach(rsc->utilization, update_utilization_value, &data); } /*! * \internal * \brief Add a resource's utilization to node capacity * * \param[in] current_utilization Current node utilization attributes * \param[in] rsc Resource with utilization to add */ void pcmk__release_node_capacity(GHashTable *current_utilization, pe_resource_t *rsc) { struct calculate_data data = { .current_utilization = current_utilization, .plus = true, }; g_hash_table_foreach(rsc->utilization, update_utilization_value, &data); } /* * Functions for checking for sufficient node capacity */ struct capacity_data { pe_node_t *node; const char *rsc_id; bool is_enough; }; /*! * \internal * \brief Check whether a single utilization attribute has sufficient capacity * * \param[in] key Name of utilization attribute to check * \param[in] value Amount of utilization required * \param[in] user_data Capacity data (as struct capacity_data *) */ static void check_capacity(gpointer key, gpointer value, gpointer user_data) { int required = 0; int remaining = 0; const char *node_value_s = NULL; struct capacity_data *data = user_data; node_value_s = g_hash_table_lookup(data->node->details->utilization, key); required = utilization_value(value); remaining = utilization_value(node_value_s); if (required > remaining) { crm_debug("Remaining capacity for %s on %s (%d) is insufficient " "for resource %s usage (%d)", (const char *) key, data->node->details->uname, remaining, data->rsc_id, required); data->is_enough = false; } } /*! * \internal * \brief Check whether a node has sufficient capacity for a resource * * \param[in] node Node to check * \param[in] rsc_id ID of resource to check (for debug logs only) * \param[in] utilization Required utilization amounts * * \return true if node has sufficient capacity for resource, otherwise false */ static bool have_enough_capacity(pe_node_t *node, const char *rsc_id, GHashTable *utilization) { struct capacity_data data = { .node = node, .rsc_id = rsc_id, .is_enough = true, }; g_hash_table_foreach(utilization, check_capacity, &data); return data.is_enough; } /*! * \internal * \brief Sum the utilization requirements of a list of resources * * \param[in] orig_rsc Resource being allocated (for logging purposes) * \param[in] rscs Resources whose utilization should be summed * * \return Newly allocated hash table with sum of all utilization values * \note It is the caller's responsibility to free the return value using * g_hash_table_destroy(). */ static GHashTable * sum_resource_utilization(pe_resource_t *orig_rsc, GList *rscs) { GHashTable *utilization = pcmk__strkey_table(free, free); for (GList *iter = rscs; iter != NULL; iter = iter->next) { pe_resource_t *rsc = (pe_resource_t *) iter->data; rsc->cmds->add_utilization(rsc, orig_rsc, rscs, utilization); } return utilization; } /*! * \internal * \brief Ban resource from nodes with insufficient utilization capacity * * \param[in] rsc Resource to check * \param[in,out] prefer Resource's preferred node (might be updated) * \param[in] data_set Cluster working set */ void pcmk__ban_insufficient_capacity(pe_resource_t *rsc, pe_node_t **prefer, pe_working_set_t *data_set) { bool any_capable = false; char *rscs_id = NULL; pe_node_t *node = NULL; pe_node_t *most_capable_node = NULL; GList *colocated_rscs = NULL; GHashTable *unallocated_utilization = NULL; GHashTableIter iter; CRM_CHECK((rsc != NULL) && (prefer != NULL) && (data_set != NULL), return); // The default placement strategy ignores utilization if (pcmk__str_eq(data_set->placement_strategy, "default", pcmk__str_casei)) { return; } // Check whether any resources are colocated with this one colocated_rscs = rsc->cmds->colocated_resources(rsc, NULL, NULL); if (colocated_rscs == NULL) { return; } rscs_id = crm_strdup_printf("%s and its colocated resources", rsc->id); // If rsc isn't in the list, add it so we include its utilization if (g_list_find(colocated_rscs, rsc) == NULL) { colocated_rscs = g_list_append(colocated_rscs, rsc); } // Sum utilization of colocated resources that haven't been allocated yet unallocated_utilization = sum_resource_utilization(rsc, colocated_rscs); // Check whether any node has enough capacity for all the resources g_hash_table_iter_init(&iter, rsc->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (void **) &node)) { if (!pcmk__node_available(node) || (node->weight < 0)) { continue; } if (have_enough_capacity(node, rscs_id, unallocated_utilization)) { any_capable = true; } // Keep track of node with most free capacity if ((most_capable_node == NULL) || (pcmk__compare_node_capacities(node, most_capable_node) < 0)) { most_capable_node = node; } } if (any_capable) { // If so, ban resource from any node with insufficient capacity g_hash_table_iter_init(&iter, rsc->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (void **) &node)) { if ((node->weight >= 0) && pcmk__node_available(node) && !have_enough_capacity(node, rscs_id, unallocated_utilization)) { pe_rsc_debug(rsc, "%s does not have enough capacity for %s", node->details->uname, rscs_id); resource_location(rsc, node, -INFINITY, "__limit_utilization__", data_set); } } } else { // Otherwise, ban from nodes with insufficient capacity for rsc alone if (*prefer == NULL) { *prefer = most_capable_node; } g_hash_table_iter_init(&iter, rsc->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (void **) &node)) { if ((node->weight >= 0) && pcmk__node_available(node) && !have_enough_capacity(node, rsc->id, rsc->utilization)) { pe_rsc_debug(rsc, "%s does not have enough capacity for %s", node->details->uname, rsc->id); resource_location(rsc, node, -INFINITY, "__limit_utilization__", data_set); } } } g_hash_table_destroy(unallocated_utilization); g_list_free(colocated_rscs); free(rscs_id); pe__show_node_weights(true, rsc, "Post-utilization", rsc->allowed_nodes, data_set); } /*! * \internal * \brief Create a new load_stopped pseudo-op for a node * * \param[in] node Node to create op for * \param[in] data_set Cluster working set * * \return Newly created load_stopped op */ static pe_action_t * new_load_stopped_op(const pe_node_t *node, pe_working_set_t *data_set) { char *load_stopped_task = crm_strdup_printf(LOAD_STOPPED "_%s", node->details->uname); pe_action_t *load_stopped = get_pseudo_op(load_stopped_task, data_set); if (load_stopped->node == NULL) { load_stopped->node = pe__copy_node(node); pe__clear_action_flags(load_stopped, pe_action_optional); } free(load_stopped_task); return load_stopped; } /*! * \internal * \brief Create utilization-related internal constraints for a resource * * \param[in] rsc Resource to create constraints for * \param[in] allowed_nodes List of allowed next nodes for \p rsc */ void pcmk__create_utilization_constraints(pe_resource_t *rsc, GList *allowed_nodes) { GList *iter = NULL; pe_node_t *node = NULL; pe_action_t *load_stopped = NULL; pe_rsc_trace(rsc, "Creating utilization constraints for %s - strategy: %s", rsc->id, rsc->cluster->placement_strategy); // "stop rsc then load_stopped" constraints for current nodes for (iter = rsc->running_on; iter != NULL; iter = iter->next) { node = (pe_node_t *) iter->data; load_stopped = new_load_stopped_op(node, rsc->cluster); pcmk__new_ordering(rsc, stop_key(rsc), NULL, NULL, NULL, load_stopped, pe_order_load, rsc->cluster); } // "load_stopped then start/migrate_to rsc" constraints for allowed nodes for (GList *iter = allowed_nodes; iter; iter = iter->next) { node = (pe_node_t *) iter->data; load_stopped = new_load_stopped_op(node, rsc->cluster); pcmk__new_ordering(NULL, NULL, load_stopped, rsc, start_key(rsc), NULL, pe_order_load, rsc->cluster); pcmk__new_ordering(NULL, NULL, load_stopped, rsc, pcmk__op_key(rsc->id, RSC_MIGRATE, 0), NULL, pe_order_load, rsc->cluster); } } + +/*! + * \internal + * \brief Output node capacities if enabled + * + * \param[in] desc Prefix for output + * \param[in] data_set Cluster working set + */ +void +pcmk__show_node_capacities(const char *desc, pe_working_set_t *data_set) +{ + if (pcmk_is_set(data_set->flags, pe_flag_show_utilization)) { + for (GList *iter = data_set->nodes; iter != NULL; iter = iter->next) { + pe_node_t *node = (pe_node_t *) iter->data; + pcmk__output_t *out = data_set->priv; + + out->message(out, "node-capacity", node, desc); + } + } +}