diff --git a/include/pcmki/pcmki_sched_utils.h b/include/pcmki/pcmki_sched_utils.h index 9076030921..8bd01601dd 100644 --- a/include/pcmki/pcmki_sched_utils.h +++ b/include/pcmki/pcmki_sched_utils.h @@ -1,59 +1,56 @@ /* * Copyright 2004-2021 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 PENGINE_AUTILS__H # define PENGINE_AUTILS__H #include // bool #include // GList, GHashTable, gboolean, guint #include // lrmd_event_data_t #include // cib_t #include #include #include #include #include /* Constraint helper functions */ pcmk__colocation_t *invert_constraint(pcmk__colocation_t *constraint); pe__location_t *copy_constraint(pe__location_t *constraint); GHashTable *pcmk__copy_node_table(GHashTable *nodes); GList *pcmk__copy_node_list(const GList *list, bool reset); GList *pcmk__sort_nodes(GList *nodes, pe_node_t *active_node, pe_working_set_t *data_set); bool pcmk__node_available(const pe_node_t *node); bool pcmk__any_node_available(GHashTable *nodes); pe_resource_t *find_compatible_child(pe_resource_t *local_child, pe_resource_t *rsc, enum rsc_role_e filter, gboolean current, pe_working_set_t *data_set); pe_resource_t *find_compatible_child_by_node(pe_resource_t * local_child, pe_node_t * local_node, pe_resource_t * rsc, enum rsc_role_e filter, gboolean current); gboolean is_child_compatible(pe_resource_t *child_rsc, pe_node_t * local_node, enum rsc_role_e filter, gboolean current); enum pe_action_flags summary_action_flags(pe_action_t * action, GList *children, pe_node_t * node); enum action_tasks clone_child_action(pe_action_t * action); int copies_per_node(pe_resource_t * rsc); -extern void calculate_utilization(GHashTable * current_utilization, - GHashTable * utilization, gboolean plus); - extern void process_utilization(pe_resource_t * rsc, pe_node_t ** prefer, pe_working_set_t * data_set); xmlNode *pcmk__create_history_xml(xmlNode *parent, lrmd_event_data_t *event, const char *caller_version, int target_rc, const char *node, const char *origin, int level); # define LOAD_STOPPED "load_stopped" #endif diff --git a/lib/pacemaker/libpacemaker_private.h b/lib/pacemaker/libpacemaker_private.h index 8a4bd056cb..4699b5e931 100644 --- a/lib/pacemaker/libpacemaker_private.h +++ b/lib/pacemaker/libpacemaker_private.h @@ -1,316 +1,323 @@ /* * Copyright 2021 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 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); // 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); // Functions applying to more than one variant (pcmk_sched_resource.c) 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_working_set_t *data_set, pe_resource_t **failed); // 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); #endif // PCMK__LIBPACEMAKER_PRIVATE__H diff --git a/lib/pacemaker/pcmk_sched_resource.c b/lib/pacemaker/pcmk_sched_resource.c index 3794c3634c..de19d67f9b 100644 --- a/lib/pacemaker/pcmk_sched_resource.c +++ b/lib/pacemaker/pcmk_sched_resource.c @@ -1,355 +1,354 @@ /* * Copyright 2014-2021 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" // 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; gboolean moving = FALSE; 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, moving); } /*! * \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) && !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)? "" : "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++; - calculate_utilization(chosen->details->utilization, rsc->utilization, - FALSE); + 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--; - calculate_utilization(old->details->utilization, rsc->utilization, TRUE); + 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[in] data_set Cluster working set * \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_working_set_t *data_set, 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, data_set); 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; } diff --git a/lib/pacemaker/pcmk_sched_utilization.c b/lib/pacemaker/pcmk_sched_utilization.c index 6230649873..ec239786a2 100644 --- a/lib/pacemaker/pcmk_sched_utilization.c +++ b/lib/pacemaker/pcmk_sched_utilization.c @@ -1,419 +1,449 @@ /* * Copyright 2014-2021 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" static void group_add_unallocated_utilization(GHashTable * all_utilization, pe_resource_t * rsc, GList *all_rscs); /*! * \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; - gboolean plus; + 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 -do_calculate_utilization(gpointer key, gpointer value, gpointer user_data) +update_utilization_value(gpointer key, gpointer value, gpointer user_data) { + int result = 0; const char *current = NULL; - char *result = NULL; struct calculate_data *data = user_data; current = g_hash_table_lookup(data->current_utilization, key); if (data->plus) { - result = pcmk__itoa(utilization_value(current) + utilization_value(value)); - g_hash_table_replace(data->current_utilization, strdup(key), result); - + result = utilization_value(current) + utilization_value(value); } else if (current) { - result = pcmk__itoa(utilization_value(current) - utilization_value(value)); - g_hash_table_replace(data->current_utilization, strdup(key), result); + result = utilization_value(current) - utilization_value(value); } + g_hash_table_replace(data->current_utilization, + strdup(key), pcmk__itoa(result)); } -/* Specify 'plus' to FALSE when allocating - * Otherwise to TRUE when deallocating +/*! + * \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 -calculate_utilization(GHashTable * current_utilization, - GHashTable * utilization, gboolean plus) +pcmk__consume_node_capacity(GHashTable *current_utilization, pe_resource_t *rsc) { - struct calculate_data data; + struct calculate_data data = { + .current_utilization = current_utilization, + .plus = false, + }; - data.current_utilization = current_utilization; - data.plus = plus; + 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(utilization, do_calculate_utilization, &data); + g_hash_table_foreach(rsc->utilization, update_utilization_value, &data); } struct capacity_data { pe_node_t *node; const char *rsc_id; gboolean is_enough; }; static void check_capacity(gpointer key, gpointer value, gpointer user_data) { int required = 0; int remaining = 0; struct capacity_data *data = user_data; required = utilization_value(value); remaining = utilization_value(g_hash_table_lookup(data->node->details->utilization, key)); if (required > remaining) { CRM_ASSERT(data->rsc_id); CRM_ASSERT(data->node); crm_debug("Node %s does not have enough %s for %s: required=%d remaining=%d", data->node->details->uname, (char *)key, data->rsc_id, required, remaining); data->is_enough = FALSE; } } static gboolean have_enough_capacity(pe_node_t * node, const char * rsc_id, GHashTable * utilization) { struct capacity_data data; data.node = node; data.rsc_id = rsc_id; data.is_enough = TRUE; g_hash_table_foreach(utilization, check_capacity, &data); return data.is_enough; } static void native_add_unallocated_utilization(GHashTable * all_utilization, pe_resource_t * rsc) { - if (!pcmk_is_set(rsc->flags, pe_rsc_provisional)) { - return; + if (pcmk_is_set(rsc->flags, pe_rsc_provisional)) { + pcmk__release_node_capacity(all_utilization, rsc); } - - calculate_utilization(all_utilization, rsc->utilization, TRUE); } static void add_unallocated_utilization(GHashTable * all_utilization, pe_resource_t * rsc, GList *all_rscs, pe_resource_t * orig_rsc) { if (!pcmk_is_set(rsc->flags, pe_rsc_provisional)) { return; } if (rsc->variant == pe_native) { pe_rsc_trace(orig_rsc, "%s: Adding %s as colocated utilization", orig_rsc->id, rsc->id); native_add_unallocated_utilization(all_utilization, rsc); } else if (rsc->variant == pe_group) { pe_rsc_trace(orig_rsc, "%s: Adding %s as colocated utilization", orig_rsc->id, rsc->id); group_add_unallocated_utilization(all_utilization, rsc, all_rscs); } else if (pe_rsc_is_clone(rsc)) { GList *gIter1 = NULL; gboolean existing = FALSE; /* Check if there's any child already existing in the list */ gIter1 = rsc->children; for (; gIter1 != NULL; gIter1 = gIter1->next) { pe_resource_t *child = (pe_resource_t *) gIter1->data; GList *gIter2 = NULL; if (g_list_find(all_rscs, child)) { existing = TRUE; } else { /* Check if there's any child of another cloned group already existing in the list */ gIter2 = child->children; for (; gIter2 != NULL; gIter2 = gIter2->next) { pe_resource_t *grandchild = (pe_resource_t *) gIter2->data; if (g_list_find(all_rscs, grandchild)) { pe_rsc_trace(orig_rsc, "%s: Adding %s as colocated utilization", orig_rsc->id, child->id); add_unallocated_utilization(all_utilization, child, all_rscs, orig_rsc); existing = TRUE; break; } } } } // rsc->children is always non-NULL but this makes static analysis happy if (!existing && (rsc->children != NULL)) { pe_resource_t *first_child = (pe_resource_t *) rsc->children->data; pe_rsc_trace(orig_rsc, "%s: Adding %s as colocated utilization", orig_rsc->id, ID(first_child->xml)); add_unallocated_utilization(all_utilization, first_child, all_rscs, orig_rsc); } } } static GHashTable * sum_unallocated_utilization(pe_resource_t * rsc, GList *colocated_rscs) { GList *gIter = NULL; GList *all_rscs = NULL; GHashTable *all_utilization = pcmk__strkey_table(free, free); all_rscs = g_list_copy(colocated_rscs); if (g_list_find(all_rscs, rsc) == FALSE) { all_rscs = g_list_append(all_rscs, rsc); } for (gIter = all_rscs; gIter != NULL; gIter = gIter->next) { pe_resource_t *listed_rsc = (pe_resource_t *) gIter->data; if (!pcmk_is_set(listed_rsc->flags, pe_rsc_provisional)) { continue; } pe_rsc_trace(rsc, "%s: Processing unallocated colocated %s", rsc->id, listed_rsc->id); add_unallocated_utilization(all_utilization, listed_rsc, all_rscs, rsc); } g_list_free(all_rscs); return all_utilization; } void process_utilization(pe_resource_t * rsc, pe_node_t ** prefer, pe_working_set_t * data_set) { CRM_CHECK(rsc && prefer && data_set, return); if (!pcmk__str_eq(data_set->placement_strategy, "default", pcmk__str_casei)) { GHashTableIter iter; GList *colocated_rscs = NULL; gboolean any_capable = FALSE; pe_node_t *node = NULL; colocated_rscs = rsc->cmds->colocated_resources(rsc, NULL, NULL); if (colocated_rscs) { GHashTable *unallocated_utilization = NULL; char *rscs_id = crm_strdup_printf("%s and its colocated resources", rsc->id); pe_node_t *most_capable_node = NULL; unallocated_utilization = sum_unallocated_utilization(rsc, colocated_rscs); 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; } if (most_capable_node == NULL || pcmk__compare_node_capacities(node, most_capable_node) < 0) { /* < 0 means 'node' is more capable */ most_capable_node = node; } } if (any_capable) { 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) == FALSE) { pe_rsc_debug(rsc, "Resource %s and its colocated resources" " cannot be allocated to node %s: not enough capacity", rsc->id, node->details->uname); resource_location(rsc, node, -INFINITY, "__limit_utilization__", data_set); } } } else if (*prefer == NULL) { *prefer = most_capable_node; } if (unallocated_utilization) { g_hash_table_destroy(unallocated_utilization); } g_list_free(colocated_rscs); free(rscs_id); } if (any_capable == FALSE) { 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, rsc->id, rsc->utilization) == FALSE) { pe_rsc_debug(rsc, "Resource %s cannot be allocated to node %s:" " not enough capacity", rsc->id, node->details->uname); resource_location(rsc, node, -INFINITY, "__limit_utilization__", data_set); } } } pe__show_node_weights(true, rsc, "Post-utilization", rsc->allowed_nodes, data_set); } } #define VARIANT_GROUP 1 #include static void group_add_unallocated_utilization(GHashTable * all_utilization, pe_resource_t * rsc, GList *all_rscs) { group_variant_data_t *group_data = NULL; get_group_variant_data(group_data, rsc); if (group_data->colocated || pe_rsc_is_clone(rsc->parent)) { GList *gIter = rsc->children; for (; gIter != NULL; gIter = gIter->next) { pe_resource_t *child_rsc = (pe_resource_t *) gIter->data; if (pcmk_is_set(child_rsc->flags, pe_rsc_provisional) && g_list_find(all_rscs, child_rsc) == FALSE) { native_add_unallocated_utilization(all_utilization, child_rsc); } } } else { if (group_data->first_child && pcmk_is_set(group_data->first_child->flags, pe_rsc_provisional) && g_list_find(all_rscs, group_data->first_child) == FALSE) { native_add_unallocated_utilization(all_utilization, group_data->first_child); } } }