diff --git a/lib/pacemaker/pcmk_sched_bundle.c b/lib/pacemaker/pcmk_sched_bundle.c index fc6ae8c487..b0724bcece 100644 --- a/lib/pacemaker/pcmk_sched_bundle.c +++ b/lib/pacemaker/pcmk_sched_bundle.c @@ -1,961 +1,1016 @@ /* * Copyright 2004-2023 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU General Public License version 2 * or later (GPLv2+) WITHOUT ANY WARRANTY. */ #include #include #include #include #include "libpacemaker_private.h" struct assign_data { const pe_node_t *prefer; bool stop_if_fail; }; /*! * \internal * \brief Assign a single bundle replica's resources (other than container) * * \param[in,out] replica Replica to assign * \param[in] user_data Preferred node, if any * * \return true (to indicate that any further replicas should be processed) */ static bool assign_replica(pe__bundle_replica_t *replica, void *user_data) { pe_node_t *container_host = NULL; struct assign_data *assign_data = user_data; const pe_node_t *prefer = assign_data->prefer; bool stop_if_fail = assign_data->stop_if_fail; const pe_resource_t *bundle = pe__const_top_resource(replica->container, true); if (replica->ip != NULL) { pe_rsc_trace(bundle, "Assigning bundle %s IP %s", bundle->id, replica->ip->id); replica->ip->cmds->assign(replica->ip, prefer, stop_if_fail); } container_host = replica->container->allocated_to; if (replica->remote != NULL) { if (pe__is_guest_or_remote_node(container_host)) { /* REMOTE_CONTAINER_HACK: "Nested" connection resources must be on * the same host because Pacemaker Remote only supports a single * active connection. */ pcmk__new_colocation("#replica-remote-with-host-remote", NULL, INFINITY, replica->remote, container_host->details->remote_rsc, NULL, NULL, pcmk__coloc_influence); } pe_rsc_trace(bundle, "Assigning bundle %s connection %s", bundle->id, replica->remote->id); replica->remote->cmds->assign(replica->remote, prefer, stop_if_fail); } if (replica->child != NULL) { pe_node_t *node = NULL; GHashTableIter iter; g_hash_table_iter_init(&iter, replica->child->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (gpointer *) &node)) { if (!pe__same_node(node, replica->node)) { node->weight = -INFINITY; } else if (!pcmk__threshold_reached(replica->child, node, NULL)) { node->weight = INFINITY; } } pe__set_resource_flags(replica->child->parent, pe_rsc_allocating); pe_rsc_trace(bundle, "Assigning bundle %s replica child %s", bundle->id, replica->child->id); replica->child->cmds->assign(replica->child, replica->node, stop_if_fail); pe__clear_resource_flags(replica->child->parent, pe_rsc_allocating); } return true; } /*! * \internal * \brief Assign a bundle resource to a node * * \param[in,out] rsc Resource to assign to a node * \param[in] prefer Node to prefer, if all else is equal * \param[in] stop_if_fail If \c true and a primitive descendant of \p rsc * can't be assigned to a node, set the * descendant's next role to stopped and update * existing actions * * \return Node that \p rsc is assigned to, if assigned entirely to one node * * \note If \p stop_if_fail is \c false, then \c pcmk__unassign_resource() can * completely undo the assignment. A successful assignment can be either * undone or left alone as final. A failed assignment has the same effect * as calling pcmk__unassign_resource(); there are no side effects on * roles or actions. */ pe_node_t * pcmk__bundle_assign(pe_resource_t *rsc, const pe_node_t *prefer, bool stop_if_fail) { GList *containers = NULL; pe_resource_t *bundled_resource = NULL; struct assign_data assign_data = { prefer, stop_if_fail }; CRM_ASSERT((rsc != NULL) && (rsc->variant == pe_container)); pe_rsc_trace(rsc, "Assigning bundle %s", rsc->id); pe__set_resource_flags(rsc, pe_rsc_allocating); pe__show_node_scores(!pcmk_is_set(rsc->cluster->flags, pe_flag_show_scores), rsc, __func__, rsc->allowed_nodes, rsc->cluster); // Assign all containers first, so we know what nodes the bundle will be on containers = g_list_sort(pe__bundle_containers(rsc), pcmk__cmp_instance); pcmk__assign_instances(rsc, containers, pe__bundle_max(rsc), rsc->fns->max_per_node(rsc)); g_list_free(containers); // Then assign remaining replica resources pe__foreach_bundle_replica(rsc, assign_replica, (void *) &assign_data); // Finally, assign the bundled resources to each bundle node bundled_resource = pe__bundled_resource(rsc); if (bundled_resource != NULL) { pe_node_t *node = NULL; GHashTableIter iter; g_hash_table_iter_init(&iter, bundled_resource->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (gpointer *) & node)) { if (pe__node_is_bundle_instance(rsc, node)) { node->weight = 0; } else { node->weight = -INFINITY; } } bundled_resource->cmds->assign(bundled_resource, prefer, stop_if_fail); } pe__clear_resource_flags(rsc, pe_rsc_allocating|pe_rsc_provisional); return NULL; } /*! * \internal * \brief Create actions for a bundle replica's resources (other than child) * * \param[in,out] replica Replica to create actions for * \param[in] user_data Unused * * \return true (to indicate that any further replicas should be processed) */ static bool create_replica_actions(pe__bundle_replica_t *replica, void *user_data) { if (replica->ip != NULL) { replica->ip->cmds->create_actions(replica->ip); } if (replica->container != NULL) { replica->container->cmds->create_actions(replica->container); } if (replica->remote != NULL) { replica->remote->cmds->create_actions(replica->remote); } return true; } /*! * \internal * \brief Create all actions needed for a given bundle resource * * \param[in,out] rsc Bundle resource to create actions for */ void pcmk__bundle_create_actions(pe_resource_t *rsc) { pe_action_t *action = NULL; GList *containers = NULL; pe_resource_t *bundled_resource = NULL; CRM_ASSERT((rsc != NULL) && (rsc->variant == pe_container)); pe__foreach_bundle_replica(rsc, create_replica_actions, NULL); containers = pe__bundle_containers(rsc); pcmk__create_instance_actions(rsc, containers); g_list_free(containers); bundled_resource = pe__bundled_resource(rsc); if (bundled_resource != NULL) { bundled_resource->cmds->create_actions(bundled_resource); if (pcmk_is_set(bundled_resource->flags, pe_rsc_promotable)) { pe__new_rsc_pseudo_action(rsc, RSC_PROMOTE, true, true); action = pe__new_rsc_pseudo_action(rsc, RSC_PROMOTED, true, true); action->priority = INFINITY; pe__new_rsc_pseudo_action(rsc, RSC_DEMOTE, true, true); action = pe__new_rsc_pseudo_action(rsc, RSC_DEMOTED, true, true); action->priority = INFINITY; } } } /*! * \internal * \brief Create internal constraints for a bundle replica's resources * * \param[in,out] replica Replica to create internal constraints for * \param[in,out] user_data Replica's parent bundle * * \return true (to indicate that any further replicas should be processed) */ static bool replica_internal_constraints(pe__bundle_replica_t *replica, void *user_data) { pe_resource_t *bundle = user_data; replica->container->cmds->internal_constraints(replica->container); // Start bundle -> start replica container pcmk__order_starts(bundle, replica->container, pe_order_runnable_left|pe_order_implies_first_printed); // Stop bundle -> stop replica child and container if (replica->child != NULL) { pcmk__order_stops(bundle, replica->child, pe_order_implies_first_printed); } pcmk__order_stops(bundle, replica->container, pe_order_implies_first_printed); // Start replica container -> bundle is started pcmk__order_resource_actions(replica->container, RSC_START, bundle, RSC_STARTED, pe_order_implies_then_printed); // Stop replica container -> bundle is stopped pcmk__order_resource_actions(replica->container, RSC_STOP, bundle, RSC_STOPPED, pe_order_implies_then_printed); if (replica->ip != NULL) { replica->ip->cmds->internal_constraints(replica->ip); // Replica IP address -> replica container (symmetric) pcmk__order_starts(replica->ip, replica->container, pe_order_runnable_left|pe_order_preserve); pcmk__order_stops(replica->container, replica->ip, pe_order_implies_first|pe_order_preserve); pcmk__new_colocation("#ip-with-container", NULL, INFINITY, replica->ip, replica->container, NULL, NULL, pcmk__coloc_influence); } if (replica->remote != NULL) { /* This handles ordering and colocating remote relative to container * (via "#resource-with-container"). Since IP is also ordered and * colocated relative to the container, we don't need to do anything * explicit here with IP. */ replica->remote->cmds->internal_constraints(replica->remote); } if (replica->child != NULL) { CRM_ASSERT(replica->remote != NULL); // "Start remote then child" is implicit in scheduler's remote logic } return true; } /*! * \internal * \brief Create implicit constraints needed for a bundle resource * * \param[in,out] rsc Bundle resource to create implicit constraints for */ void pcmk__bundle_internal_constraints(pe_resource_t *rsc) { pe_resource_t *bundled_resource = NULL; CRM_ASSERT((rsc != NULL) && (rsc->variant == pe_container)); pe__foreach_bundle_replica(rsc, replica_internal_constraints, rsc); bundled_resource = pe__bundled_resource(rsc); if (bundled_resource == NULL) { return; } // Start bundle -> start bundled clone pcmk__order_resource_actions(rsc, RSC_START, bundled_resource, RSC_START, pe_order_implies_first_printed); // Bundled clone is started -> bundle is started pcmk__order_resource_actions(bundled_resource, RSC_STARTED, rsc, RSC_STARTED, pe_order_implies_then_printed); // Stop bundle -> stop bundled clone pcmk__order_resource_actions(rsc, RSC_STOP, bundled_resource, RSC_STOP, pe_order_implies_first_printed); // Bundled clone is stopped -> bundle is stopped pcmk__order_resource_actions(bundled_resource, RSC_STOPPED, rsc, RSC_STOPPED, pe_order_implies_then_printed); bundled_resource->cmds->internal_constraints(bundled_resource); if (!pcmk_is_set(bundled_resource->flags, pe_rsc_promotable)) { return; } pcmk__promotable_restart_ordering(rsc); // Demote bundle -> demote bundled clone pcmk__order_resource_actions(rsc, RSC_DEMOTE, bundled_resource, RSC_DEMOTE, pe_order_implies_first_printed); // Bundled clone is demoted -> bundle is demoted pcmk__order_resource_actions(bundled_resource, RSC_DEMOTED, rsc, RSC_DEMOTED, pe_order_implies_then_printed); // Promote bundle -> promote bundled clone pcmk__order_resource_actions(rsc, RSC_PROMOTE, bundled_resource, RSC_PROMOTE, pe_order_implies_first_printed); // Bundled clone is promoted -> bundle is promoted pcmk__order_resource_actions(bundled_resource, RSC_PROMOTED, rsc, RSC_PROMOTED, pe_order_implies_then_printed); } struct match_data { const pe_node_t *node; // Node to compare against replica pe_resource_t *container; // Replica container corresponding to node }; /*! * \internal * \brief Check whether a replica container is assigned to a given node * * \param[in] replica Replica to check * \param[in,out] user_data struct match_data with node to compare against * * \return true if the replica does not match (to indicate further replicas * should be processed), otherwise false */ static bool match_replica_container(const pe__bundle_replica_t *replica, void *user_data) { struct match_data *match_data = user_data; if (pcmk__instance_matches(replica->container, match_data->node, RSC_ROLE_UNKNOWN, false)) { match_data->container = replica->container; return false; // Match found, don't bother searching further replicas } return true; // No match, keep searching } /*! * \internal * \brief Find a bundle container compatible with a dependent resource * * \param[in] dependent Dependent resource in colocation with bundle * \param[in] bundle Bundle that \p dependent is colocated with * * \return A container from \p bundle assigned to the same node as \p dependent * if assigned, otherwise assigned to any of dependent's allowed nodes, * otherwise NULL. */ static pe_resource_t * compatible_container(const pe_resource_t *dependent, const pe_resource_t *bundle) { GList *scratch = NULL; struct match_data match_data = { NULL, NULL }; // If dependent is assigned, only check there match_data.node = dependent->fns->location(dependent, NULL, 0); if (match_data.node != NULL) { pe__foreach_const_bundle_replica(bundle, match_replica_container, &match_data); return match_data.container; } // Otherwise, check for any of the dependent's allowed nodes scratch = g_hash_table_get_values(dependent->allowed_nodes); scratch = pcmk__sort_nodes(scratch, NULL); for (const GList *iter = scratch; iter != NULL; iter = iter->next) { match_data.node = iter->data; pe__foreach_const_bundle_replica(bundle, match_replica_container, &match_data); if (match_data.container != NULL) { break; } } g_list_free(scratch); return match_data.container; } struct coloc_data { const pcmk__colocation_t *colocation; pe_resource_t *dependent; GList *container_hosts; }; /*! * \internal * \brief Apply a colocation score to replica node scores or resource priority * * \param[in] replica Replica of primary bundle resource in colocation * \param[in,out] user_data struct coloc_data for colocation being applied * * \return true (to indicate that any further replicas should be processed) */ static bool replica_apply_coloc_score(const pe__bundle_replica_t *replica, void *user_data) { struct coloc_data *coloc_data = user_data; pe_node_t *chosen = NULL; if (coloc_data->colocation->score < INFINITY) { replica->container->cmds->apply_coloc_score(coloc_data->dependent, replica->container, coloc_data->colocation, false); return true; } chosen = replica->container->fns->location(replica->container, NULL, 0); if ((chosen == NULL) || is_set_recursive(replica->container, pe_rsc_block, true)) { return true; } if ((coloc_data->colocation->primary_role >= RSC_ROLE_PROMOTED) && ((replica->child == NULL) || (replica->child->next_role < RSC_ROLE_PROMOTED))) { return true; } pe_rsc_trace(pe__const_top_resource(replica->container, true), "Allowing mandatory colocation %s using %s @%d", coloc_data->colocation->id, pe__node_name(chosen), chosen->weight); coloc_data->container_hosts = g_list_prepend(coloc_data->container_hosts, chosen); return true; } /*! * \internal * \brief Apply a colocation's score to node scores or resource priority * * Given a colocation constraint, apply its score to the dependent's * allowed node scores (if we are still placing resources) or priority (if * we are choosing promotable clone instance roles). * * \param[in,out] dependent Dependent resource in colocation * \param[in] primary Primary resource in colocation * \param[in] colocation Colocation constraint to apply * \param[in] for_dependent true if called on behalf of dependent */ void pcmk__bundle_apply_coloc_score(pe_resource_t *dependent, const pe_resource_t *primary, const pcmk__colocation_t *colocation, bool for_dependent) { struct coloc_data coloc_data = { colocation, dependent, NULL }; /* This should never be called for the bundle itself as a dependent. - * Instead, we add its colocation constraints to its containers and call the - * apply_coloc_score() method for the containers as dependents. + * Instead, we add its colocation constraints to its containers and bundled + * primitive and call the apply_coloc_score() method for them as dependents. */ CRM_ASSERT((primary != NULL) && (primary->variant == pe_container) && (dependent != NULL) && (dependent->variant == pe_native) && (colocation != NULL) && !for_dependent); if (pcmk_is_set(primary->flags, pe_rsc_provisional)) { pe_rsc_trace(primary, "Skipping applying colocation %s " "because %s is still provisional", colocation->id, primary->id); return; } pe_rsc_trace(primary, "Applying colocation %s (%s with %s at %s)", colocation->id, dependent->id, primary->id, pcmk_readable_score(colocation->score)); /* If the constraint dependent is a clone or bundle, "dependent" here is one * of its instances. Look for a compatible instance of this bundle. */ if (colocation->dependent->variant > pe_group) { const pe_resource_t *primary_container = compatible_container(dependent, primary); if (primary_container != NULL) { // Success, we found one pe_rsc_debug(primary, "Pairing %s with %s", dependent->id, primary_container->id); dependent->cmds->apply_coloc_score(dependent, primary_container, colocation, true); } else if (colocation->score >= INFINITY) { // Failure, and it's fatal crm_notice("%s cannot run because there is no compatible " "instance of %s to colocate with", dependent->id, primary->id); pcmk__assign_resource(dependent, NULL, true, true); } else { // Failure, but we can ignore it pe_rsc_debug(primary, "%s cannot be colocated with any instance of %s", dependent->id, primary->id); } return; } pe__foreach_const_bundle_replica(primary, replica_apply_coloc_score, &coloc_data); if (colocation->score >= INFINITY) { node_list_exclude(dependent->allowed_nodes, coloc_data.container_hosts, FALSE); } g_list_free(coloc_data.container_hosts); } // Bundle implementation of resource_alloc_functions_t:with_this_colocations() void pcmk__with_bundle_colocations(const pe_resource_t *rsc, const pe_resource_t *orig_rsc, GList **list) { + const pe_resource_t *bundled_rsc = NULL; + CRM_ASSERT((rsc != NULL) && (rsc->variant == pe_container) && (orig_rsc != NULL) && (list != NULL)); - // Only the bundle itself and its containers get the bundle's constraints + // The bundle itself and its containers always get its colocations if ((orig_rsc == rsc) || pcmk_is_set(orig_rsc->flags, pe_rsc_replica_container)) { pcmk__add_with_this_list(list, rsc->rsc_cons_lhs, orig_rsc); + return; + } + + /* The bundled resource gets the colocations if it's promotable and we've + * begun choosing roles + */ + bundled_rsc = pe__bundled_resource(rsc); + if ((bundled_rsc == NULL) + || !pcmk_is_set(bundled_rsc->flags, pe_rsc_promotable) + || (pe__const_top_resource(orig_rsc, false) != bundled_rsc)) { + return; + } + + if (orig_rsc == bundled_rsc) { + if (pe__clone_flag_is_set(orig_rsc, pe__clone_promotion_constrained)) { + /* orig_rsc is the clone and we're setting roles (or have already + * done so) + */ + pcmk__add_with_this_list(list, rsc->rsc_cons_lhs, orig_rsc); + } + + } else if (!pcmk_is_set(orig_rsc->flags, pe_rsc_provisional)) { + /* orig_rsc is an instance and is already assigned. If something + * requests colocations for orig_rsc now, it's for setting roles. + */ + pcmk__add_with_this_list(list, rsc->rsc_cons_lhs, orig_rsc); } } // Bundle implementation of resource_alloc_functions_t:this_with_colocations() void pcmk__bundle_with_colocations(const pe_resource_t *rsc, const pe_resource_t *orig_rsc, GList **list) { + const pe_resource_t *bundled_rsc = NULL; + CRM_ASSERT((rsc != NULL) && (rsc->variant == pe_container) && (orig_rsc != NULL) && (list != NULL)); - // Only the bundle itself and its containers get the bundle's constraints + // The bundle itself and its containers always get its colocations if ((orig_rsc == rsc) || pcmk_is_set(orig_rsc->flags, pe_rsc_replica_container)) { pcmk__add_this_with_list(list, rsc->rsc_cons, orig_rsc); } + + /* The bundled resource gets the colocations if it's promotable and we've + * begun choosing roles + */ + bundled_rsc = pe__bundled_resource(rsc); + if ((bundled_rsc == NULL) + || !pcmk_is_set(bundled_rsc->flags, pe_rsc_promotable) + || (pe__const_top_resource(orig_rsc, false) != bundled_rsc)) { + return; + } + + if (orig_rsc == bundled_rsc) { + if (pe__clone_flag_is_set(orig_rsc, pe__clone_promotion_constrained)) { + /* orig_rsc is the clone and we're setting roles (or have already + * done so) + */ + pcmk__add_this_with_list(list, rsc->rsc_cons, orig_rsc); + } + + } else if (!pcmk_is_set(orig_rsc->flags, pe_rsc_provisional)) { + /* orig_rsc is an instance and is already assigned. If something + * requests colocations for orig_rsc now, it's for setting roles. + */ + pcmk__add_this_with_list(list, rsc->rsc_cons, orig_rsc); + } } /*! * \internal * \brief Return action flags for a given bundle resource action * * \param[in,out] action Bundle resource action to get flags for * \param[in] node If not NULL, limit effects to this node * * \return Flags appropriate to \p action on \p node */ uint32_t pcmk__bundle_action_flags(pe_action_t *action, const pe_node_t *node) { GList *containers = NULL; uint32_t flags = 0; pe_resource_t *bundled_resource = NULL; CRM_ASSERT((action != NULL) && (action->rsc != NULL) && (action->rsc->variant == pe_container)); bundled_resource = pe__bundled_resource(action->rsc); if (bundled_resource != NULL) { // Clone actions are done on the bundled clone resource, not container switch (get_complex_task(bundled_resource, action->task)) { case no_action: case action_notify: case action_notified: case action_promote: case action_promoted: case action_demote: case action_demoted: return pcmk__collective_action_flags(action, bundled_resource->children, node); default: break; } } containers = pe__bundle_containers(action->rsc); flags = pcmk__collective_action_flags(action, containers, node); g_list_free(containers); return flags; } /*! * \internal * \brief Apply a location constraint to a bundle replica * * \param[in,out] replica Replica to apply constraint to * \param[in,out] user_data Location constraint to apply * * \return true (to indicate that any further replicas should be processed) */ static bool apply_location_to_replica(pe__bundle_replica_t *replica, void *user_data) { pe__location_t *location = user_data; if (replica->container != NULL) { replica->container->cmds->apply_location(replica->container, location); } if (replica->ip != NULL) { replica->ip->cmds->apply_location(replica->ip, location); } return true; } /*! * \internal * \brief Apply a location constraint to a bundle resource's allowed node scores * * \param[in,out] rsc Bundle resource to apply constraint to * \param[in,out] location Location constraint to apply */ void pcmk__bundle_apply_location(pe_resource_t *rsc, pe__location_t *location) { pe_resource_t *bundled_resource = NULL; CRM_ASSERT((rsc != NULL) && (rsc->variant == pe_container) && (location != NULL)); pcmk__apply_location(rsc, location); pe__foreach_bundle_replica(rsc, apply_location_to_replica, location); bundled_resource = pe__bundled_resource(rsc); if ((bundled_resource != NULL) && ((location->role_filter == RSC_ROLE_UNPROMOTED) || (location->role_filter == RSC_ROLE_PROMOTED))) { bundled_resource->cmds->apply_location(bundled_resource, location); bundled_resource->rsc_location = g_list_prepend( bundled_resource->rsc_location, location); } } #define XPATH_REMOTE "//nvpair[@name='" XML_RSC_ATTR_REMOTE_RA_ADDR "']" /*! * \internal * \brief Add a bundle replica's actions to transition graph * * \param[in,out] replica Replica to add to graph * \param[in] user_data Bundle that replica belongs to (for logging only) * * \return true (to indicate that any further replicas should be processed) */ static bool add_replica_actions_to_graph(pe__bundle_replica_t *replica, void *user_data) { if ((replica->remote != NULL) && (replica->container != NULL) && pe__bundle_needs_remote_name(replica->remote)) { /* REMOTE_CONTAINER_HACK: Allow remote nodes to run containers that * run pacemaker-remoted inside, without needing a separate IP for * the container. This is done by configuring the inner remote's * connection host as the magic string "#uname", then * replacing it with the underlying host when needed. */ xmlNode *nvpair = get_xpath_object(XPATH_REMOTE, replica->remote->xml, LOG_ERR); const char *calculated_addr = NULL; // Replace the value in replica->remote->xml (if appropriate) calculated_addr = pe__add_bundle_remote_name(replica->remote, replica->remote->cluster, nvpair, "value"); if (calculated_addr != NULL) { /* Since this is for the bundle as a resource, and not any * particular action, replace the value in the default * parameters (not evaluated for node). create_graph_action() * will grab it from there to replace it in node-evaluated * parameters. */ GHashTable *params = pe_rsc_params(replica->remote, NULL, replica->remote->cluster); g_hash_table_replace(params, strdup(XML_RSC_ATTR_REMOTE_RA_ADDR), strdup(calculated_addr)); } else { pe_resource_t *bundle = user_data; /* The only way to get here is if the remote connection is * neither currently running nor scheduled to run. That means we * won't be doing any operations that require addr (only start * requires it; we additionally use it to compare digests when * unpacking status, promote, and migrate_from history, but * that's already happened by this point). */ pe_rsc_info(bundle, "Unable to determine address for bundle %s " "remote connection", bundle->id); } } if (replica->ip != NULL) { replica->ip->cmds->add_actions_to_graph(replica->ip); } if (replica->container != NULL) { replica->container->cmds->add_actions_to_graph(replica->container); } if (replica->remote != NULL) { replica->remote->cmds->add_actions_to_graph(replica->remote); } return true; } /*! * \internal * \brief Add a bundle resource's actions to the transition graph * * \param[in,out] rsc Bundle resource whose actions should be added */ void pcmk__bundle_add_actions_to_graph(pe_resource_t *rsc) { pe_resource_t *bundled_resource = NULL; CRM_ASSERT((rsc != NULL) && (rsc->variant == pe_container)); bundled_resource = pe__bundled_resource(rsc); if (bundled_resource != NULL) { bundled_resource->cmds->add_actions_to_graph(bundled_resource); } pe__foreach_bundle_replica(rsc, add_replica_actions_to_graph, rsc); } struct probe_data { pe_resource_t *bundle; // Bundle being probed pe_node_t *node; // Node to create probes on bool any_created; // Whether any probes have been created }; /*! * \internal * \brief Order a bundle replica's start after another replica's probe * * \param[in,out] replica Replica to order start for * \param[in,out] user_data Replica with probe to order after * * \return true (to indicate that any further replicas should be processed) */ static bool order_replica_start_after(pe__bundle_replica_t *replica, void *user_data) { pe__bundle_replica_t *probed_replica = user_data; if ((replica == probed_replica) || (replica->container == NULL)) { return true; } pcmk__new_ordering(probed_replica->container, pcmk__op_key(probed_replica->container->id, RSC_STATUS, 0), NULL, replica->container, pcmk__op_key(replica->container->id, RSC_START, 0), NULL, pe_order_optional|pe_order_same_node, replica->container->cluster); return true; } /*! * \internal * \brief Create probes for a bundle replica's resources * * \param[in,out] replica Replica to create probes for * \param[in,out] user_data struct probe_data * * \return true (to indicate that any further replicas should be processed) */ static bool create_replica_probes(pe__bundle_replica_t *replica, void *user_data) { struct probe_data *probe_data = user_data; if ((replica->ip != NULL) && replica->ip->cmds->create_probe(replica->ip, probe_data->node)) { probe_data->any_created = true; } if ((replica->child != NULL) && pe__same_node(probe_data->node, replica->node) && replica->child->cmds->create_probe(replica->child, probe_data->node)) { probe_data->any_created = true; } if ((replica->container != NULL) && replica->container->cmds->create_probe(replica->container, probe_data->node)) { probe_data->any_created = true; /* If we're limited to one replica per host (due to * the lack of an IP range probably), then we don't * want any of our peer containers starting until * we've established that no other copies are already * running. * * Partly this is to ensure that the maximum replicas per host is * observed, but also to ensure that the containers * don't fail to start because the necessary port * mappings (which won't include an IP for uniqueness) * are already taken */ if (probe_data->bundle->fns->max_per_node(probe_data->bundle) == 1) { pe__foreach_bundle_replica(probe_data->bundle, order_replica_start_after, replica); } } if ((replica->container != NULL) && (replica->remote != NULL) && replica->remote->cmds->create_probe(replica->remote, probe_data->node)) { /* Do not probe the remote resource until we know where the container is * running. This is required for REMOTE_CONTAINER_HACK to correctly * probe remote resources. */ char *probe_uuid = pcmk__op_key(replica->remote->id, RSC_STATUS, 0); pe_action_t *probe = find_first_action(replica->remote->actions, probe_uuid, NULL, probe_data->node); free(probe_uuid); if (probe != NULL) { probe_data->any_created = true; pe_rsc_trace(probe_data->bundle, "Ordering %s probe on %s", replica->remote->id, pe__node_name(probe_data->node)); pcmk__new_ordering(replica->container, pcmk__op_key(replica->container->id, RSC_START, 0), NULL, replica->remote, NULL, probe, pe_order_probe, probe_data->bundle->cluster); } } return true; } /*! * \internal * * \brief Schedule any probes needed for a bundle resource on a node * * \param[in,out] rsc Bundle resource to create probes for * \param[in,out] node Node to create probe on * * \return true if any probe was created, otherwise false */ bool pcmk__bundle_create_probe(pe_resource_t *rsc, pe_node_t *node) { struct probe_data probe_data = { rsc, node, false }; CRM_ASSERT((rsc != NULL) && (rsc->variant == pe_container)); pe__foreach_bundle_replica(rsc, create_replica_probes, &probe_data); return probe_data.any_created; } /*! * \internal * \brief Output actions for one bundle replica * * \param[in,out] replica Replica to output actions for * \param[in] user_data Unused * * \return true (to indicate that any further replicas should be processed) */ static bool output_replica_actions(pe__bundle_replica_t *replica, void *user_data) { if (replica->ip != NULL) { replica->ip->cmds->output_actions(replica->ip); } if (replica->container != NULL) { replica->container->cmds->output_actions(replica->container); } if (replica->remote != NULL) { replica->remote->cmds->output_actions(replica->remote); } if (replica->child != NULL) { replica->child->cmds->output_actions(replica->child); } return true; } /*! * \internal * \brief Output a summary of scheduled actions for a bundle resource * * \param[in,out] rsc Bundle resource to output actions for */ void pcmk__output_bundle_actions(pe_resource_t *rsc) { CRM_ASSERT((rsc != NULL) && (rsc->variant == pe_container)); pe__foreach_bundle_replica(rsc, output_replica_actions, NULL); } // Bundle implementation of resource_alloc_functions_t:add_utilization() void pcmk__bundle_add_utilization(const pe_resource_t *rsc, const pe_resource_t *orig_rsc, GList *all_rscs, GHashTable *utilization) { pe_resource_t *container = NULL; CRM_ASSERT((rsc != NULL) && (rsc->variant == pe_container)); if (!pcmk_is_set(rsc->flags, pe_rsc_provisional)) { return; } /* All bundle replicas are identical, so using the utilization of the first * is sufficient for any. Only the implicit container resource can have * utilization values. */ container = pe__first_container(rsc); if (container != NULL) { container->cmds->add_utilization(container, orig_rsc, all_rscs, utilization); } } // Bundle implementation of resource_alloc_functions_t:shutdown_lock() void pcmk__bundle_shutdown_lock(pe_resource_t *rsc) { CRM_ASSERT((rsc != NULL) && (rsc->variant == pe_container)); // Bundles currently don't support shutdown locks } diff --git a/lib/pacemaker/pcmk_sched_clone.c b/lib/pacemaker/pcmk_sched_clone.c index 5633f42b27..35f2e3d4d6 100644 --- a/lib/pacemaker/pcmk_sched_clone.c +++ b/lib/pacemaker/pcmk_sched_clone.c @@ -1,691 +1,699 @@ /* * Copyright 2004-2023 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU General Public License version 2 * or later (GPLv2+) WITHOUT ANY WARRANTY. */ #include #include #include #include "libpacemaker_private.h" /*! * \internal * \brief Assign a clone resource's instances to nodes * * \param[in,out] rsc Clone resource to assign * \param[in] prefer Node to prefer, if all else is equal * \param[in] stop_if_fail If \c true and a primitive descendant of \p rsc * can't be assigned to a node, set the * descendant's next role to stopped and update * existing actions * * \return NULL (clones are not assigned to a single node) * * \note If \p stop_if_fail is \c false, then \c pcmk__unassign_resource() can * completely undo the assignment. A successful assignment can be either * undone or left alone as final. A failed assignment has the same effect * as calling pcmk__unassign_resource(); there are no side effects on * roles or actions. */ pe_node_t * pcmk__clone_assign(pe_resource_t *rsc, const pe_node_t *prefer, bool stop_if_fail) { + GList *colocations = NULL; + CRM_ASSERT(pe_rsc_is_clone(rsc)); if (!pcmk_is_set(rsc->flags, pe_rsc_provisional)) { return NULL; // Assignment has already been done } // Detect assignment loops if (pcmk_is_set(rsc->flags, pe_rsc_allocating)) { pe_rsc_debug(rsc, "Breaking assignment loop involving %s", rsc->id); return NULL; } pe__set_resource_flags(rsc, pe_rsc_allocating); // If this clone is promotable, consider nodes' promotion scores if (pcmk_is_set(rsc->flags, pe_rsc_promotable)) { pcmk__add_promotion_scores(rsc); } - /* If this clone is colocated with any other resources, assign those first. - * Since the this_with_colocations() method boils down to a copy of rsc_cons - * for clones, we can use that here directly for efficiency. - */ - for (GList *iter = rsc->rsc_cons; iter != NULL; iter = iter->next) { + // If this clone is colocated with any other resources, assign those first + colocations = pcmk__this_with_colocations(rsc); + for (GList *iter = colocations; iter != NULL; iter = iter->next) { pcmk__colocation_t *constraint = (pcmk__colocation_t *) iter->data; pe_rsc_trace(rsc, "%s: Assigning colocation %s primary %s first", rsc->id, constraint->id, constraint->primary->id); constraint->primary->cmds->assign(constraint->primary, prefer, stop_if_fail); } + g_list_free(colocations); - /* If any resources are colocated with this one, consider their preferences. - * Because the with_this_colocations() method boils down to a copy of - * rsc_cons_lhs for clones, we can use that here directly for efficiency. - */ - g_list_foreach(rsc->rsc_cons_lhs, pcmk__add_dependent_scores, rsc); + // If any resources are colocated with this one, consider their preferences + colocations = pcmk__with_this_colocations(rsc); + g_list_foreach(colocations, pcmk__add_dependent_scores, rsc); + g_list_free(colocations); pe__show_node_scores(!pcmk_is_set(rsc->cluster->flags, pe_flag_show_scores), rsc, __func__, rsc->allowed_nodes, rsc->cluster); rsc->children = g_list_sort(rsc->children, pcmk__cmp_instance); pcmk__assign_instances(rsc, rsc->children, pe__clone_max(rsc), pe__clone_node_max(rsc)); if (pcmk_is_set(rsc->flags, pe_rsc_promotable)) { pcmk__set_instance_roles(rsc); } pe__clear_resource_flags(rsc, pe_rsc_provisional|pe_rsc_allocating); pe_rsc_trace(rsc, "Assigned clone %s", rsc->id); return NULL; } /*! * \internal * \brief Create all actions needed for a given clone resource * * \param[in,out] rsc Clone resource to create actions for */ void pcmk__clone_create_actions(pe_resource_t *rsc) { CRM_ASSERT(pe_rsc_is_clone(rsc)); pe_rsc_trace(rsc, "Creating actions for clone %s", rsc->id); pcmk__create_instance_actions(rsc, rsc->children); if (pcmk_is_set(rsc->flags, pe_rsc_promotable)) { pcmk__create_promotable_actions(rsc); } } /*! * \internal * \brief Create implicit constraints needed for a clone resource * * \param[in,out] rsc Clone resource to create implicit constraints for */ void pcmk__clone_internal_constraints(pe_resource_t *rsc) { bool ordered = false; CRM_ASSERT(pe_rsc_is_clone(rsc)); pe_rsc_trace(rsc, "Creating internal constraints for clone %s", rsc->id); // Restart ordering: Stop -> stopped -> start -> started pcmk__order_resource_actions(rsc, RSC_STOPPED, rsc, RSC_START, pe_order_optional); pcmk__order_resource_actions(rsc, RSC_START, rsc, RSC_STARTED, pe_order_runnable_left); pcmk__order_resource_actions(rsc, RSC_STOP, rsc, RSC_STOPPED, pe_order_runnable_left); // Demoted -> stop and started -> promote if (pcmk_is_set(rsc->flags, pe_rsc_promotable)) { pcmk__order_resource_actions(rsc, RSC_DEMOTED, rsc, RSC_STOP, pe_order_optional); pcmk__order_resource_actions(rsc, RSC_STARTED, rsc, RSC_PROMOTE, pe_order_runnable_left); } ordered = pe__clone_is_ordered(rsc); if (ordered) { /* Ordered clone instances must start and stop by instance number. The * instances might have been previously shuffled for assignment or * promotion purposes, so re-sort them. */ rsc->children = g_list_sort(rsc->children, pcmk__cmp_instance_number); } for (GList *iter = rsc->children; iter != NULL; iter = iter->next) { pe_resource_t *instance = (pe_resource_t *) iter->data; instance->cmds->internal_constraints(instance); // Start clone -> start instance -> clone started pcmk__order_starts(rsc, instance, pe_order_runnable_left |pe_order_implies_first_printed); pcmk__order_resource_actions(instance, RSC_START, rsc, RSC_STARTED, pe_order_implies_then_printed); // Stop clone -> stop instance -> clone stopped pcmk__order_stops(rsc, instance, pe_order_implies_first_printed); pcmk__order_resource_actions(instance, RSC_STOP, rsc, RSC_STOPPED, pe_order_implies_then_printed); /* Instances of ordered clones must be started and stopped by instance * number. Since only some instances may be starting or stopping, order * each instance relative to every later instance. */ if (ordered) { for (GList *later = iter->next; later != NULL; later = later->next) { pcmk__order_starts(instance, (pe_resource_t *) later->data, pe_order_optional); pcmk__order_stops((pe_resource_t *) later->data, instance, pe_order_optional); } } } if (pcmk_is_set(rsc->flags, pe_rsc_promotable)) { pcmk__order_promotable_instances(rsc); } } /*! * \internal * \brief Check whether colocated resources can be interleaved * * \param[in] colocation Colocation constraint with clone as primary * * \return true if colocated resources can be interleaved, otherwise false */ static bool can_interleave(const pcmk__colocation_t *colocation) { const pe_resource_t *dependent = colocation->dependent; // Only colocations between clone or bundle resources use interleaving if (dependent->variant <= pe_group) { return false; } // Only the dependent needs to be marked for interleaving if (!crm_is_true(g_hash_table_lookup(dependent->meta, XML_RSC_ATTR_INTERLEAVE))) { return false; } /* @TODO Do we actually care about multiple primary instances sharing a * dependent instance? */ if (dependent->fns->max_per_node(dependent) != colocation->primary->fns->max_per_node(colocation->primary)) { pcmk__config_err("Cannot interleave %s and %s because they do not " "support the same number of instances per node", dependent->id, colocation->primary->id); return false; } return true; } /*! * \internal * \brief Apply a colocation's score to node scores or resource priority * * Given a colocation constraint, apply its score to the dependent's * allowed node scores (if we are still placing resources) or priority (if * we are choosing promotable clone instance roles). * * \param[in,out] dependent Dependent resource in colocation * \param[in] primary Primary resource in colocation * \param[in] colocation Colocation constraint to apply * \param[in] for_dependent true if called on behalf of dependent */ void pcmk__clone_apply_coloc_score(pe_resource_t *dependent, const pe_resource_t *primary, const pcmk__colocation_t *colocation, bool for_dependent) { const GList *iter = NULL; /* This should never be called for the clone itself as a dependent. Instead, * we add its colocation constraints to its instances and call the * apply_coloc_score() method for the instances as dependents. */ CRM_ASSERT(!for_dependent); CRM_ASSERT((colocation != NULL) && pe_rsc_is_clone(primary) && (dependent != NULL) && (dependent->variant == pe_native)); if (pcmk_is_set(primary->flags, pe_rsc_provisional)) { pe_rsc_trace(primary, "Delaying processing colocation %s " "because cloned primary %s is still provisional", colocation->id, primary->id); return; } pe_rsc_trace(primary, "Processing colocation %s (%s with clone %s @%s)", colocation->id, dependent->id, primary->id, pcmk_readable_score(colocation->score)); // Apply role-specific colocations if (pcmk_is_set(primary->flags, pe_rsc_promotable) && (colocation->primary_role != RSC_ROLE_UNKNOWN)) { if (pcmk_is_set(dependent->flags, pe_rsc_provisional)) { // We're assigning the dependent to a node pcmk__update_dependent_with_promotable(primary, dependent, colocation); return; } if (colocation->dependent_role == RSC_ROLE_PROMOTED) { // We're choosing a role for the dependent pcmk__update_promotable_dependent_priority(primary, dependent, colocation); return; } } // Apply interleaved colocations if (can_interleave(colocation)) { const pe_resource_t *primary_instance = NULL; primary_instance = pcmk__find_compatible_instance(dependent, primary, RSC_ROLE_UNKNOWN, false); if (primary_instance != NULL) { pe_rsc_debug(primary, "Interleaving %s with %s", dependent->id, primary_instance->id); dependent->cmds->apply_coloc_score(dependent, primary_instance, colocation, true); } else if (colocation->score >= INFINITY) { crm_notice("%s cannot run because it cannot interleave with " "any instance of %s", dependent->id, primary->id); pcmk__assign_resource(dependent, NULL, true, true); } else { pe_rsc_debug(primary, "%s will not colocate with %s " "because no instance can interleave with it", dependent->id, primary->id); } return; } // Apply mandatory colocations if (colocation->score >= INFINITY) { GList *primary_nodes = NULL; // Dependent can run only where primary will have unblocked instances for (iter = primary->children; iter != NULL; iter = iter->next) { const pe_resource_t *instance = iter->data; pe_node_t *chosen = instance->fns->location(instance, NULL, 0); if ((chosen != NULL) && !is_set_recursive(instance, pe_rsc_block, TRUE)) { pe_rsc_trace(primary, "Allowing %s: %s %d", colocation->id, pe__node_name(chosen), chosen->weight); primary_nodes = g_list_prepend(primary_nodes, chosen); } } node_list_exclude(dependent->allowed_nodes, primary_nodes, FALSE); g_list_free(primary_nodes); return; } // Apply optional colocations for (iter = primary->children; iter != NULL; iter = iter->next) { const pe_resource_t *instance = iter->data; instance->cmds->apply_coloc_score(dependent, instance, colocation, false); } } // Clone implementation of resource_alloc_functions_t:with_this_colocations() void pcmk__with_clone_colocations(const pe_resource_t *rsc, const pe_resource_t *orig_rsc, GList **list) { CRM_CHECK((rsc != NULL) && (orig_rsc != NULL) && (list != NULL), return); pcmk__add_with_this_list(list, rsc->rsc_cons_lhs, orig_rsc); + + if (rsc->parent != NULL) { + rsc->parent->cmds->with_this_colocations(rsc->parent, orig_rsc, list); + } } // Clone implementation of resource_alloc_functions_t:this_with_colocations() void pcmk__clone_with_colocations(const pe_resource_t *rsc, const pe_resource_t *orig_rsc, GList **list) { CRM_CHECK((rsc != NULL) && (orig_rsc != NULL) && (list != NULL), return); pcmk__add_this_with_list(list, rsc->rsc_cons, orig_rsc); + + if (rsc->parent != NULL) { + rsc->parent->cmds->this_with_colocations(rsc->parent, orig_rsc, list); + } } /*! * \internal * \brief Return action flags for a given clone resource action * * \param[in,out] action Action to get flags for * \param[in] node If not NULL, limit effects to this node * * \return Flags appropriate to \p action on \p node */ uint32_t pcmk__clone_action_flags(pe_action_t *action, const pe_node_t *node) { CRM_ASSERT((action != NULL) && pe_rsc_is_clone(action->rsc)); return pcmk__collective_action_flags(action, action->rsc->children, node); } /*! * \internal * \brief Apply a location constraint to a clone resource's allowed node scores * * \param[in,out] rsc Clone resource to apply constraint to * \param[in,out] location Location constraint to apply */ void pcmk__clone_apply_location(pe_resource_t *rsc, pe__location_t *location) { CRM_CHECK((location != NULL) && pe_rsc_is_clone(rsc), return); pcmk__apply_location(rsc, location); for (GList *iter = rsc->children; iter != NULL; iter = iter->next) { pe_resource_t *instance = (pe_resource_t *) iter->data; instance->cmds->apply_location(instance, location); } } // GFunc wrapper for calling the action_flags() resource method static void call_action_flags(gpointer data, gpointer user_data) { pe_resource_t *rsc = user_data; rsc->cmds->action_flags((pe_action_t *) data, NULL); } /*! * \internal * \brief Add a clone resource's actions to the transition graph * * \param[in,out] rsc Resource whose actions should be added */ void pcmk__clone_add_actions_to_graph(pe_resource_t *rsc) { CRM_ASSERT(pe_rsc_is_clone(rsc)); g_list_foreach(rsc->actions, call_action_flags, rsc); pe__create_clone_notifications(rsc); for (GList *iter = rsc->children; iter != NULL; iter = iter->next) { pe_resource_t *child_rsc = (pe_resource_t *) iter->data; child_rsc->cmds->add_actions_to_graph(child_rsc); } pcmk__add_rsc_actions_to_graph(rsc); pe__free_clone_notification_data(rsc); } /*! * \internal * \brief Check whether a resource or any children have been probed on a node * * \param[in] rsc Resource to check * \param[in] node Node to check * * \return true if \p node is in the known_on table of \p rsc or any of its * children, otherwise false */ static bool rsc_probed_on(const pe_resource_t *rsc, const pe_node_t *node) { if (rsc->children != NULL) { for (GList *child_iter = rsc->children; child_iter != NULL; child_iter = child_iter->next) { pe_resource_t *child = (pe_resource_t *) child_iter->data; if (rsc_probed_on(child, node)) { return true; } } return false; } if (rsc->known_on != NULL) { GHashTableIter iter; pe_node_t *known_node = NULL; g_hash_table_iter_init(&iter, rsc->known_on); while (g_hash_table_iter_next(&iter, NULL, (gpointer *) &known_node)) { if (pe__same_node(node, known_node)) { return true; } } } return false; } /*! * \internal * \brief Find clone instance that has been probed on given node * * \param[in] clone Clone resource to check * \param[in] node Node to check * * \return Instance of \p clone that has been probed on \p node if any, * otherwise NULL */ static pe_resource_t * find_probed_instance_on(const pe_resource_t *clone, const pe_node_t *node) { for (GList *iter = clone->children; iter != NULL; iter = iter->next) { pe_resource_t *instance = (pe_resource_t *) iter->data; if (rsc_probed_on(instance, node)) { return instance; } } return NULL; } /*! * \internal * \brief Probe an anonymous clone on a node * * \param[in,out] clone Anonymous clone to probe * \param[in,out] node Node to probe \p clone on */ static bool probe_anonymous_clone(pe_resource_t *clone, pe_node_t *node) { // Check whether we already probed an instance on this node pe_resource_t *child = find_probed_instance_on(clone, node); // Otherwise, check if we plan to start an instance on this node for (GList *iter = clone->children; (iter != NULL) && (child == NULL); iter = iter->next) { pe_resource_t *instance = (pe_resource_t *) iter->data; const pe_node_t *instance_node = NULL; instance_node = instance->fns->location(instance, NULL, 0); if (pe__same_node(instance_node, node)) { child = instance; } } // Otherwise, use the first clone instance if (child == NULL) { child = clone->children->data; } // Anonymous clones only need to probe a single instance return child->cmds->create_probe(child, node); } /*! * \internal * \brief Schedule any probes needed for a resource on a node * * \param[in,out] rsc Resource to create probe for * \param[in,out] node Node to create probe on * * \return true if any probe was created, otherwise false */ bool pcmk__clone_create_probe(pe_resource_t *rsc, pe_node_t *node) { CRM_ASSERT((node != NULL) && pe_rsc_is_clone(rsc)); if (rsc->exclusive_discover) { /* The clone is configured to be probed only where a location constraint * exists with resource-discovery set to exclusive. * * This check is not strictly necessary here since the instance's * create_probe() method would also check, but doing it here is more * efficient (especially for unique clones with a large number of * instances), and affects the CRM_meta_notify_available_uname variable * passed with notify actions. */ pe_node_t *allowed = g_hash_table_lookup(rsc->allowed_nodes, node->details->id); if ((allowed == NULL) || (allowed->rsc_discover_mode != pe_discover_exclusive)) { /* This node is not marked for resource discovery. Remove it from * allowed_nodes so that notifications contain only nodes that the * clone can possibly run on. */ pe_rsc_trace(rsc, "Skipping probe for %s on %s because resource has " "exclusive discovery but is not allowed on node", rsc->id, pe__node_name(node)); g_hash_table_remove(rsc->allowed_nodes, node->details->id); return false; } } rsc->children = g_list_sort(rsc->children, pcmk__cmp_instance_number); if (pcmk_is_set(rsc->flags, pe_rsc_unique)) { return pcmk__probe_resource_list(rsc->children, node); } else { return probe_anonymous_clone(rsc, node); } } /*! * \internal * \brief Add meta-attributes relevant to transition graph actions to XML * * Add clone-specific meta-attributes needed for transition graph actions. * * \param[in] rsc Clone resource whose meta-attributes should be added * \param[in,out] xml Transition graph action attributes XML to add to */ void pcmk__clone_add_graph_meta(const pe_resource_t *rsc, xmlNode *xml) { char *name = NULL; CRM_ASSERT(pe_rsc_is_clone(rsc) && (xml != NULL)); name = crm_meta_name(XML_RSC_ATTR_UNIQUE); crm_xml_add(xml, name, pe__rsc_bool_str(rsc, pe_rsc_unique)); free(name); name = crm_meta_name(XML_RSC_ATTR_NOTIFY); crm_xml_add(xml, name, pe__rsc_bool_str(rsc, pe_rsc_notify)); free(name); name = crm_meta_name(XML_RSC_ATTR_INCARNATION_MAX); crm_xml_add_int(xml, name, pe__clone_max(rsc)); free(name); name = crm_meta_name(XML_RSC_ATTR_INCARNATION_NODEMAX); crm_xml_add_int(xml, name, pe__clone_node_max(rsc)); free(name); if (pcmk_is_set(rsc->flags, pe_rsc_promotable)) { int promoted_max = pe__clone_promoted_max(rsc); int promoted_node_max = pe__clone_promoted_node_max(rsc); name = crm_meta_name(XML_RSC_ATTR_PROMOTED_MAX); crm_xml_add_int(xml, name, promoted_max); free(name); name = crm_meta_name(XML_RSC_ATTR_PROMOTED_NODEMAX); crm_xml_add_int(xml, name, promoted_node_max); free(name); /* @COMPAT Maintain backward compatibility with resource agents that * expect the old names (deprecated since 2.0.0). */ name = crm_meta_name(PCMK_XA_PROMOTED_MAX_LEGACY); crm_xml_add_int(xml, name, promoted_max); free(name); name = crm_meta_name(PCMK_XA_PROMOTED_NODE_MAX_LEGACY); crm_xml_add_int(xml, name, promoted_node_max); free(name); } } // Clone implementation of resource_alloc_functions_t:add_utilization() void pcmk__clone_add_utilization(const pe_resource_t *rsc, const pe_resource_t *orig_rsc, GList *all_rscs, GHashTable *utilization) { bool existing = false; pe_resource_t *child = NULL; CRM_ASSERT(pe_rsc_is_clone(rsc) && (orig_rsc != NULL) && (utilization != NULL)); if (!pcmk_is_set(rsc->flags, pe_rsc_provisional)) { return; } // Look for any child already existing in the list for (GList *iter = rsc->children; iter != NULL; iter = iter->next) { child = (pe_resource_t *) iter->data; if (g_list_find(all_rscs, child)) { existing = true; // Keep checking remaining children } else { // If this is a clone of a group, look for group's members for (GList *member_iter = child->children; member_iter != NULL; member_iter = member_iter->next) { pe_resource_t *member = (pe_resource_t *) member_iter->data; if (g_list_find(all_rscs, member) != NULL) { // Add *child's* utilization, not group member's child->cmds->add_utilization(child, orig_rsc, all_rscs, utilization); existing = true; break; } } } } if (!existing && (rsc->children != NULL)) { // If nothing was found, still add first child's utilization child = (pe_resource_t *) rsc->children->data; child->cmds->add_utilization(child, orig_rsc, all_rscs, utilization); } } // Clone implementation of resource_alloc_functions_t:shutdown_lock() void pcmk__clone_shutdown_lock(pe_resource_t *rsc) { CRM_ASSERT(pe_rsc_is_clone(rsc)); return; // Clones currently don't support shutdown locks } diff --git a/lib/pacemaker/pcmk_sched_instances.c b/lib/pacemaker/pcmk_sched_instances.c index 72b0b3e3a1..357e4f97c6 100644 --- a/lib/pacemaker/pcmk_sched_instances.c +++ b/lib/pacemaker/pcmk_sched_instances.c @@ -1,1668 +1,1667 @@ /* * Copyright 2004-2023 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU General Public License version 2 * or later (GPLv2+) WITHOUT ANY WARRANTY. */ /* This file is intended for code usable with both clone instances and bundle * replica containers. */ #include #include #include #include "libpacemaker_private.h" /*! * \internal * \brief Check whether a node is allowed to run an instance * * \param[in] instance Clone instance or bundle container to check * \param[in] node Node to check * \param[in] max_per_node Maximum number of instances allowed to run on a node * * \return true if \p node is allowed to run \p instance, otherwise false */ static bool can_run_instance(const pe_resource_t *instance, const pe_node_t *node, int max_per_node) { pe_node_t *allowed_node = NULL; if (pcmk_is_set(instance->flags, pe_rsc_orphan)) { pe_rsc_trace(instance, "%s cannot run on %s: orphaned", instance->id, pe__node_name(node)); return false; } if (!pcmk__node_available(node, false, false)) { pe_rsc_trace(instance, "%s cannot run on %s: node cannot run resources", instance->id, pe__node_name(node)); return false; } allowed_node = pcmk__top_allowed_node(instance, node); if (allowed_node == NULL) { crm_warn("%s cannot run on %s: node not allowed", instance->id, pe__node_name(node)); return false; } if (allowed_node->weight < 0) { pe_rsc_trace(instance, "%s cannot run on %s: parent score is %s there", instance->id, pe__node_name(node), pcmk_readable_score(allowed_node->weight)); return false; } if (allowed_node->count >= max_per_node) { pe_rsc_trace(instance, "%s cannot run on %s: node already has %d instance%s", instance->id, pe__node_name(node), max_per_node, pcmk__plural_s(max_per_node)); return false; } pe_rsc_trace(instance, "%s can run on %s (%d already running)", instance->id, pe__node_name(node), allowed_node->count); return true; } /*! * \internal * \brief Ban a clone instance or bundle replica from unavailable allowed nodes * * \param[in,out] instance Clone instance or bundle replica to ban * \param[in] max_per_node Maximum instances allowed to run on a node */ static void ban_unavailable_allowed_nodes(pe_resource_t *instance, int max_per_node) { if (instance->allowed_nodes != NULL) { GHashTableIter iter; pe_node_t *node = NULL; g_hash_table_iter_init(&iter, instance->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (void **) &node)) { if (!can_run_instance(instance, node, max_per_node)) { pe_rsc_trace(instance, "Banning %s from unavailable node %s", instance->id, pe__node_name(node)); node->weight = -INFINITY; for (GList *child_iter = instance->children; child_iter != NULL; child_iter = child_iter->next) { pe_resource_t *child = (pe_resource_t *) child_iter->data; pe_node_t *child_node = NULL; child_node = pe_hash_table_lookup(child->allowed_nodes, node->details->id); if (child_node != NULL) { pe_rsc_trace(instance, "Banning %s child %s " "from unavailable node %s", instance->id, child->id, pe__node_name(node)); child_node->weight = -INFINITY; } } } } } } /*! * \internal * \brief Create a hash table with a single node in it * * \param[in] node Node to copy into new table * * \return Newly created hash table containing a copy of \p node * \note The caller is responsible for freeing the result with * g_hash_table_destroy(). */ static GHashTable * new_node_table(pe_node_t *node) { GHashTable *table = pcmk__strkey_table(NULL, free); node = pe__copy_node(node); g_hash_table_insert(table, (gpointer) node->details->id, node); return table; } /*! * \internal * \brief Apply a resource's parent's colocation scores to a node table * * \param[in] rsc Resource whose colocations should be applied * \param[in,out] nodes Node table to apply colocations to */ static void apply_parent_colocations(const pe_resource_t *rsc, GHashTable **nodes) { - GList *iter = NULL; - pcmk__colocation_t *colocation = NULL; - pe_resource_t *other = NULL; - float factor = 0.0; + GList *colocations = pcmk__this_with_colocations(rsc); + + for (const GList *iter = colocations; iter != NULL; iter = iter->next) { + const pcmk__colocation_t *colocation = iter->data; + pe_resource_t *other = colocation->primary; + float factor = colocation->score / (float) INFINITY; - /* Because the this_with_colocations() and with_this_colocations() methods - * boil down to copies of rsc_cons and rsc_cons_lhs for clones and bundles, - * we can use those here directly for efficiency. - */ - for (iter = rsc->parent->rsc_cons; iter != NULL; iter = iter->next) { - colocation = (pcmk__colocation_t *) iter->data; - other = colocation->primary; - factor = colocation->score / (float) INFINITY, other->cmds->add_colocated_node_scores(other, rsc, rsc->id, nodes, colocation, factor, pcmk__coloc_select_default); } - for (iter = rsc->parent->rsc_cons_lhs; iter != NULL; iter = iter->next) { - colocation = (pcmk__colocation_t *) iter->data; + g_list_free(colocations); + colocations = pcmk__with_this_colocations(rsc); + + for (const GList *iter = colocations; iter != NULL; iter = iter->next) { + const pcmk__colocation_t *colocation = iter->data; + pe_resource_t *other = colocation->dependent; + float factor = colocation->score / (float) INFINITY; + if (!pcmk__colocation_has_influence(colocation, rsc)) { continue; } - other = colocation->dependent; - factor = colocation->score / (float) INFINITY, other->cmds->add_colocated_node_scores(other, rsc, rsc->id, nodes, colocation, factor, pcmk__coloc_select_nonnegative); } + g_list_free(colocations); } /*! * \internal * \brief Compare clone or bundle instances based on colocation scores * * Determine the relative order in which two clone or bundle instances should be * assigned to nodes, considering the scores of colocation constraints directly * or indirectly involving them. * * \param[in] instance1 First instance to compare * \param[in] instance2 Second instance to compare * * \return A negative number if \p instance1 should be assigned first, * a positive number if \p instance2 should be assigned first, * or 0 if assignment order doesn't matter */ static int cmp_instance_by_colocation(const pe_resource_t *instance1, const pe_resource_t *instance2) { int rc = 0; pe_node_t *node1 = NULL; pe_node_t *node2 = NULL; pe_node_t *current_node1 = pe__current_node(instance1); pe_node_t *current_node2 = pe__current_node(instance2); GHashTable *colocated_scores1 = NULL; GHashTable *colocated_scores2 = NULL; CRM_ASSERT((instance1 != NULL) && (instance1->parent != NULL) && (instance2 != NULL) && (instance2->parent != NULL) && (current_node1 != NULL) && (current_node2 != NULL)); // Create node tables initialized with each node colocated_scores1 = new_node_table(current_node1); colocated_scores2 = new_node_table(current_node2); // Apply parental colocations apply_parent_colocations(instance1, &colocated_scores1); apply_parent_colocations(instance2, &colocated_scores2); // Find original nodes again, with scores updated for colocations node1 = g_hash_table_lookup(colocated_scores1, current_node1->details->id); node2 = g_hash_table_lookup(colocated_scores2, current_node2->details->id); // Compare nodes by updated scores if (node1->weight < node2->weight) { crm_trace("Assign %s (%d on %s) after %s (%d on %s)", instance1->id, node1->weight, pe__node_name(node1), instance2->id, node2->weight, pe__node_name(node2)); rc = 1; } else if (node1->weight > node2->weight) { crm_trace("Assign %s (%d on %s) before %s (%d on %s)", instance1->id, node1->weight, pe__node_name(node1), instance2->id, node2->weight, pe__node_name(node2)); rc = -1; } g_hash_table_destroy(colocated_scores1); g_hash_table_destroy(colocated_scores2); return rc; } /*! * \internal * \brief Check whether a resource or any of its children are failed * * \param[in] rsc Resource to check * * \return true if \p rsc or any of its children are failed, otherwise false */ static bool did_fail(const pe_resource_t *rsc) { if (pcmk_is_set(rsc->flags, pe_rsc_failed)) { return true; } for (GList *iter = rsc->children; iter != NULL; iter = iter->next) { if (did_fail((const pe_resource_t *) iter->data)) { return true; } } return false; } /*! * \internal * \brief Check whether a node is allowed to run a resource * * \param[in] rsc Resource to check * \param[in,out] node Node to check (will be set NULL if not allowed) * * \return true if *node is either NULL or allowed for \p rsc, otherwise false */ static bool node_is_allowed(const pe_resource_t *rsc, pe_node_t **node) { if (*node != NULL) { pe_node_t *allowed = pe_hash_table_lookup(rsc->allowed_nodes, (*node)->details->id); if ((allowed == NULL) || (allowed->weight < 0)) { pe_rsc_trace(rsc, "%s: current location (%s) is unavailable", rsc->id, pe__node_name(*node)); *node = NULL; return false; } } return true; } /*! * \internal * \brief Compare two clone or bundle instances' instance numbers * * \param[in] a First instance to compare * \param[in] b Second instance to compare * * \return A negative number if \p a's instance number is lower, * a positive number if \p b's instance number is lower, * or 0 if their instance numbers are the same */ gint pcmk__cmp_instance_number(gconstpointer a, gconstpointer b) { const pe_resource_t *instance1 = (const pe_resource_t *) a; const pe_resource_t *instance2 = (const pe_resource_t *) b; char *div1 = NULL; char *div2 = NULL; CRM_ASSERT((instance1 != NULL) && (instance2 != NULL)); // Clone numbers are after a colon, bundle numbers after a dash div1 = strrchr(instance1->id, ':'); if (div1 == NULL) { div1 = strrchr(instance1->id, '-'); } div2 = strrchr(instance2->id, ':'); if (div2 == NULL) { div2 = strrchr(instance2->id, '-'); } CRM_ASSERT((div1 != NULL) && (div2 != NULL)); return (gint) (strtol(div1 + 1, NULL, 10) - strtol(div2 + 1, NULL, 10)); } /*! * \internal * \brief Compare clone or bundle instances according to assignment order * * Compare two clone or bundle instances according to the order they should be * assigned to nodes, preferring (in order): * * - Active instance that is less multiply active * - Instance that is not active on a disallowed node * - Instance with higher configured priority * - Active instance whose current node can run resources * - Active instance whose parent is allowed on current node * - Active instance whose current node has fewer other instances * - Active instance * - Instance that isn't failed * - Instance whose colocations result in higher score on current node * - Instance with lower ID in lexicographic order * * \param[in] a First instance to compare * \param[in] b Second instance to compare * * \return A negative number if \p a should be assigned first, * a positive number if \p b should be assigned first, * or 0 if assignment order doesn't matter */ gint pcmk__cmp_instance(gconstpointer a, gconstpointer b) { int rc = 0; pe_node_t *node1 = NULL; pe_node_t *node2 = NULL; unsigned int nnodes1 = 0; unsigned int nnodes2 = 0; bool can1 = true; bool can2 = true; const pe_resource_t *instance1 = (const pe_resource_t *) a; const pe_resource_t *instance2 = (const pe_resource_t *) b; CRM_ASSERT((instance1 != NULL) && (instance2 != NULL)); node1 = instance1->fns->active_node(instance1, &nnodes1, NULL); node2 = instance2->fns->active_node(instance2, &nnodes2, NULL); /* If both instances are running and at least one is multiply * active, prefer instance that's running on fewer nodes. */ if ((nnodes1 > 0) && (nnodes2 > 0)) { if (nnodes1 < nnodes2) { crm_trace("Assign %s (active on %d) before %s (active on %d): " "less multiply active", instance1->id, nnodes1, instance2->id, nnodes2); return -1; } else if (nnodes1 > nnodes2) { crm_trace("Assign %s (active on %d) after %s (active on %d): " "more multiply active", instance1->id, nnodes1, instance2->id, nnodes2); return 1; } } /* An instance that is either inactive or active on an allowed node is * preferred over an instance that is active on a no-longer-allowed node. */ can1 = node_is_allowed(instance1, &node1); can2 = node_is_allowed(instance2, &node2); if (can1 && !can2) { crm_trace("Assign %s before %s: not active on a disallowed node", instance1->id, instance2->id); return -1; } else if (!can1 && can2) { crm_trace("Assign %s after %s: active on a disallowed node", instance1->id, instance2->id); return 1; } // Prefer instance with higher configured priority if (instance1->priority > instance2->priority) { crm_trace("Assign %s before %s: priority (%d > %d)", instance1->id, instance2->id, instance1->priority, instance2->priority); return -1; } else if (instance1->priority < instance2->priority) { crm_trace("Assign %s after %s: priority (%d < %d)", instance1->id, instance2->id, instance1->priority, instance2->priority); return 1; } // Prefer active instance if ((node1 == NULL) && (node2 == NULL)) { crm_trace("No assignment preference for %s vs. %s: inactive", instance1->id, instance2->id); return 0; } else if (node1 == NULL) { crm_trace("Assign %s after %s: active", instance1->id, instance2->id); return 1; } else if (node2 == NULL) { crm_trace("Assign %s before %s: active", instance1->id, instance2->id); return -1; } // Prefer instance whose current node can run resources can1 = pcmk__node_available(node1, false, false); can2 = pcmk__node_available(node2, false, false); if (can1 && !can2) { crm_trace("Assign %s before %s: current node can run resources", instance1->id, instance2->id); return -1; } else if (!can1 && can2) { crm_trace("Assign %s after %s: current node can't run resources", instance1->id, instance2->id); return 1; } // Prefer instance whose parent is allowed to run on instance's current node node1 = pcmk__top_allowed_node(instance1, node1); node2 = pcmk__top_allowed_node(instance2, node2); if ((node1 == NULL) && (node2 == NULL)) { crm_trace("No assignment preference for %s vs. %s: " "parent not allowed on either instance's current node", instance1->id, instance2->id); return 0; } else if (node1 == NULL) { crm_trace("Assign %s after %s: parent not allowed on current node", instance1->id, instance2->id); return 1; } else if (node2 == NULL) { crm_trace("Assign %s before %s: parent allowed on current node", instance1->id, instance2->id); return -1; } // Prefer instance whose current node is running fewer other instances if (node1->count < node2->count) { crm_trace("Assign %s before %s: fewer active instances on current node", instance1->id, instance2->id); return -1; } else if (node1->count > node2->count) { crm_trace("Assign %s after %s: more active instances on current node", instance1->id, instance2->id); return 1; } // Prefer instance that isn't failed can1 = did_fail(instance1); can2 = did_fail(instance2); if (!can1 && can2) { crm_trace("Assign %s before %s: not failed", instance1->id, instance2->id); return -1; } else if (can1 && !can2) { crm_trace("Assign %s after %s: failed", instance1->id, instance2->id); return 1; } // Prefer instance with higher cumulative colocation score on current node rc = cmp_instance_by_colocation(instance1, instance2); if (rc != 0) { return rc; } // Prefer instance with lower instance number rc = pcmk__cmp_instance_number(instance1, instance2); if (rc < 0) { crm_trace("Assign %s before %s: instance number", instance1->id, instance2->id); } else if (rc > 0) { crm_trace("Assign %s after %s: instance number", instance1->id, instance2->id); } else { crm_trace("No assignment preference for %s vs. %s", instance1->id, instance2->id); } return rc; } /*! * \internal * \brief Increment the parent's instance count after assigning an instance * * An instance's parent tracks how many instances have been assigned to each * node via its pe_node_t:count member. After assigning an instance to a node, * find the corresponding node in the parent's allowed table and increment it. * * \param[in,out] instance Instance whose parent to update * \param[in] assigned_to Node to which the instance was assigned */ static void increment_parent_count(pe_resource_t *instance, const pe_node_t *assigned_to) { pe_node_t *allowed = NULL; if (assigned_to == NULL) { return; } allowed = pcmk__top_allowed_node(instance, assigned_to); if (allowed == NULL) { /* The instance is allowed on the node, but its parent isn't. This * shouldn't be possible if the resource is managed, and we won't be * able to limit the number of instances assigned to the node. */ CRM_LOG_ASSERT(!pcmk_is_set(instance->flags, pe_rsc_managed)); } else { allowed->count++; } } /*! * \internal * \brief Assign an instance to a node * * \param[in,out] instance Clone instance or bundle replica container * \param[in] prefer If not NULL, attempt early assignment to this * node, if still the best choice; otherwise, * perform final assignment * \param[in] max_per_node Assign at most this many instances to one node * * \return Node to which \p instance is assigned */ static const pe_node_t * assign_instance(pe_resource_t *instance, const pe_node_t *prefer, int max_per_node) { pe_node_t *chosen = NULL; pe_rsc_trace(instance, "Assigning %s (preferring %s)", instance->id, ((prefer == NULL)? "no node" : prefer->details->uname)); if (pcmk_is_set(instance->flags, pe_rsc_allocating)) { pe_rsc_debug(instance, "Assignment loop detected involving %s colocations", instance->id); return NULL; } ban_unavailable_allowed_nodes(instance, max_per_node); // Failed early assignments are reversible (stop_if_fail=false) chosen = instance->cmds->assign(instance, prefer, (prefer == NULL)); increment_parent_count(instance, chosen); return chosen; } /*! * \internal * \brief Try to assign an instance to its current node early * * \param[in] rsc Clone or bundle being assigned (for logs only) * \param[in] instance Clone instance or bundle replica container * \param[in] current Instance's current node * \param[in] max_per_node Maximum number of instances per node * \param[in] available Number of instances still available for assignment * * \return \c true if \p instance was successfully assigned to its current node, * or \c false otherwise */ static bool assign_instance_early(const pe_resource_t *rsc, pe_resource_t *instance, const pe_node_t *current, int max_per_node, int available) { const pe_node_t *chosen = NULL; int reserved = 0; pe_resource_t *parent = instance->parent; GHashTable *allowed_orig = NULL; GHashTable *allowed_orig_parent = parent->allowed_nodes; const pe_node_t *allowed_node = g_hash_table_lookup(instance->allowed_nodes, current->details->id); pe_rsc_trace(instance, "Trying to assign %s to its current node %s", instance->id, pe__node_name(current)); if (!pcmk__node_available(allowed_node, true, false)) { pe_rsc_info(instance, "Not assigning %s to current node %s: unavailable", instance->id, pe__node_name(current)); return false; } /* On each iteration, if instance gets assigned to a node other than its * current one, we reserve one instance for the chosen node, unassign * instance, restore instance's original node tables, and try again. This * way, instances are proportionally assigned to nodes based on preferences, * but shuffling of specific instances is minimized. If a node will be * assigned instances at all, it preferentially receives instances that are * currently active there. * * parent->allowed_nodes tracks the number of instances assigned to each * node. If a node already has max_per_node instances assigned, * ban_unavailable_allowed_nodes() marks it as unavailable. * * In the end, we restore the original parent->allowed_nodes to undo the * changes to counts during tentative assignments. If we successfully * assigned instance to its current node, we increment that node's counter. */ // Back up the allowed node tables of instance and its children recursively pcmk__copy_node_tables(instance, &allowed_orig); // Update instances-per-node counts in a scratch table parent->allowed_nodes = pcmk__copy_node_table(parent->allowed_nodes); while (reserved < available) { chosen = assign_instance(instance, current, max_per_node); if (pe__same_node(chosen, current)) { // Successfully assigned to current node break; } // Assignment updates scores, so restore to original state pe_rsc_debug(instance, "Rolling back node scores for %s", instance->id); pcmk__restore_node_tables(instance, allowed_orig); if (chosen == NULL) { // Assignment failed, so give up pe_rsc_info(instance, "Not assigning %s to current node %s: unavailable", instance->id, pe__node_name(current)); pe__set_resource_flags(instance, pe_rsc_provisional); break; } // We prefer more strongly to assign an instance to the chosen node pe_rsc_debug(instance, "Not assigning %s to current node %s: %s is better", instance->id, pe__node_name(current), pe__node_name(chosen)); // Reserve one instance for the chosen node and try again if (++reserved >= available) { pe_rsc_info(instance, "Not assigning %s to current node %s: " "other assignments are more important", instance->id, pe__node_name(current)); } else { pe_rsc_debug(instance, "Reserved an instance of %s for %s. Retrying " "assignment of %s to %s", rsc->id, pe__node_name(chosen), instance->id, pe__node_name(current)); } // Clear this assignment (frees chosen); leave instance counts in parent pcmk__unassign_resource(instance); chosen = NULL; } g_hash_table_destroy(allowed_orig); // Restore original instances-per-node counts g_hash_table_destroy(parent->allowed_nodes); parent->allowed_nodes = allowed_orig_parent; if (chosen == NULL) { // Couldn't assign instance to current node return false; } pe_rsc_trace(instance, "Assigned %s to current node %s", instance->id, pe__node_name(current)); increment_parent_count(instance, chosen); return true; } /*! * \internal * \brief Reset the node counts of a resource's allowed nodes to zero * * \param[in,out] rsc Resource to reset * * \return Number of nodes that are available to run resources */ static unsigned int reset_allowed_node_counts(pe_resource_t *rsc) { unsigned int available_nodes = 0; pe_node_t *node = NULL; GHashTableIter iter; g_hash_table_iter_init(&iter, rsc->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (gpointer *) &node)) { node->count = 0; if (pcmk__node_available(node, false, false)) { available_nodes++; } } return available_nodes; } /*! * \internal * \brief Check whether an instance has a preferred node * * \param[in] rsc Clone or bundle being assigned (for logs only) * \param[in] instance Clone instance or bundle replica container * \param[in] optimal_per_node Optimal number of instances per node * * \return Instance's current node if still available, otherwise NULL */ static const pe_node_t * preferred_node(const pe_resource_t *rsc, const pe_resource_t *instance, int optimal_per_node) { const pe_node_t *node = NULL; const pe_node_t *parent_node = NULL; // Check whether instance is active, healthy, and not yet assigned if ((instance->running_on == NULL) || !pcmk_is_set(instance->flags, pe_rsc_provisional) || pcmk_is_set(instance->flags, pe_rsc_failed)) { return NULL; } // Check whether instance's current node can run resources node = pe__current_node(instance); if (!pcmk__node_available(node, true, false)) { pe_rsc_trace(rsc, "Not assigning %s to %s early (unavailable)", instance->id, pe__node_name(node)); return NULL; } // Check whether node already has optimal number of instances assigned parent_node = pcmk__top_allowed_node(instance, node); if ((parent_node != NULL) && (parent_node->count >= optimal_per_node)) { pe_rsc_trace(rsc, "Not assigning %s to %s early " "(optimal instances already assigned)", instance->id, pe__node_name(node)); return NULL; } return node; } /*! * \internal * \brief Assign collective instances to nodes * * \param[in,out] collective Clone or bundle resource being assigned * \param[in,out] instances List of clone instances or bundle containers * \param[in] max_total Maximum instances to assign in total * \param[in] max_per_node Maximum instances to assign to any one node */ void pcmk__assign_instances(pe_resource_t *collective, GList *instances, int max_total, int max_per_node) { // Reuse node count to track number of assigned instances unsigned int available_nodes = reset_allowed_node_counts(collective); int optimal_per_node = 0; int assigned = 0; GList *iter = NULL; pe_resource_t *instance = NULL; const pe_node_t *current = NULL; if (available_nodes > 0) { optimal_per_node = max_total / available_nodes; } if (optimal_per_node < 1) { optimal_per_node = 1; } pe_rsc_debug(collective, "Assigning up to %d %s instance%s to up to %u node%s " "(at most %d per host, %d optimal)", max_total, collective->id, pcmk__plural_s(max_total), available_nodes, pcmk__plural_s(available_nodes), max_per_node, optimal_per_node); // Assign as many instances as possible to their current location for (iter = instances; (iter != NULL) && (assigned < max_total); iter = iter->next) { int available = max_total - assigned; instance = iter->data; if (!pcmk_is_set(instance->flags, pe_rsc_provisional)) { continue; // Already assigned } current = preferred_node(collective, instance, optimal_per_node); if ((current != NULL) && assign_instance_early(collective, instance, current, max_per_node, available)) { assigned++; } } pe_rsc_trace(collective, "Assigned %d of %d instance%s to current node", assigned, max_total, pcmk__plural_s(max_total)); for (iter = instances; iter != NULL; iter = iter->next) { instance = (pe_resource_t *) iter->data; if (!pcmk_is_set(instance->flags, pe_rsc_provisional)) { continue; // Already assigned } if (instance->running_on != NULL) { current = pe__current_node(instance); if (pcmk__top_allowed_node(instance, current) == NULL) { const char *unmanaged = ""; if (!pcmk_is_set(instance->flags, pe_rsc_managed)) { unmanaged = "Unmanaged resource "; } crm_notice("%s%s is running on %s which is no longer allowed", unmanaged, instance->id, pe__node_name(current)); } } if (assigned >= max_total) { pe_rsc_debug(collective, "Not assigning %s because maximum %d instances " "already assigned", instance->id, max_total); resource_location(instance, NULL, -INFINITY, "collective_limit_reached", collective->cluster); } else if (assign_instance(instance, NULL, max_per_node) != NULL) { assigned++; } } pe_rsc_debug(collective, "Assigned %d of %d possible instance%s of %s", assigned, max_total, pcmk__plural_s(max_total), collective->id); } enum instance_state { instance_starting = (1 << 0), instance_stopping = (1 << 1), /* This indicates that some instance is restarting. It's not the same as * instance_starting|instance_stopping, which would indicate that some * instance is starting, and some instance (not necessarily the same one) is * stopping. */ instance_restarting = (1 << 2), instance_active = (1 << 3), instance_all = instance_starting|instance_stopping |instance_restarting|instance_active, }; /*! * \internal * \brief Check whether an instance is active, starting, and/or stopping * * \param[in] instance Clone instance or bundle replica container * \param[in,out] state Whether any instance is starting, stopping, etc. */ static void check_instance_state(const pe_resource_t *instance, uint32_t *state) { const GList *iter = NULL; uint32_t instance_state = 0; // State of just this instance // No need to check further if all conditions have already been detected if (pcmk_all_flags_set(*state, instance_all)) { return; } // If instance is a collective (a cloned group), check its children instead if (instance->variant > pe_native) { for (iter = instance->children; (iter != NULL) && !pcmk_all_flags_set(*state, instance_all); iter = iter->next) { check_instance_state((const pe_resource_t *) iter->data, state); } return; } // If we get here, instance is a primitive if (instance->running_on != NULL) { instance_state |= instance_active; } // Check each of the instance's actions for runnable start or stop for (iter = instance->actions; (iter != NULL) && !pcmk_all_flags_set(instance_state, instance_starting |instance_stopping); iter = iter->next) { const pe_action_t *action = (const pe_action_t *) iter->data; const bool optional = pcmk_is_set(action->flags, pe_action_optional); if (pcmk__str_eq(RSC_START, action->task, pcmk__str_none)) { if (!optional && pcmk_is_set(action->flags, pe_action_runnable)) { pe_rsc_trace(instance, "Instance is starting due to %s", action->uuid); instance_state |= instance_starting; } else { pe_rsc_trace(instance, "%s doesn't affect %s state (%s)", action->uuid, instance->id, (optional? "optional" : "unrunnable")); } } else if (pcmk__str_eq(RSC_STOP, action->task, pcmk__str_none)) { /* Only stop actions can be pseudo-actions for primitives. That * indicates that the node they are on is being fenced, so the stop * is implied rather than actually executed. */ if (!optional && pcmk_any_flags_set(action->flags, pe_action_pseudo|pe_action_runnable)) { pe_rsc_trace(instance, "Instance is stopping due to %s", action->uuid); instance_state |= instance_stopping; } else { pe_rsc_trace(instance, "%s doesn't affect %s state (%s)", action->uuid, instance->id, (optional? "optional" : "unrunnable")); } } } if (pcmk_all_flags_set(instance_state, instance_starting|instance_stopping)) { instance_state |= instance_restarting; } *state |= instance_state; } /*! * \internal * \brief Create actions for collective resource instances * * \param[in,out] collective Clone or bundle resource to create actions for * \param[in,out] instances List of clone instances or bundle containers */ void pcmk__create_instance_actions(pe_resource_t *collective, GList *instances) { uint32_t state = 0; pe_action_t *stop = NULL; pe_action_t *stopped = NULL; pe_action_t *start = NULL; pe_action_t *started = NULL; pe_rsc_trace(collective, "Creating collective instance actions for %s", collective->id); // Create actions for each instance appropriate to its variant for (GList *iter = instances; iter != NULL; iter = iter->next) { pe_resource_t *instance = (pe_resource_t *) iter->data; instance->cmds->create_actions(instance); check_instance_state(instance, &state); } // Create pseudo-actions for rsc start and started start = pe__new_rsc_pseudo_action(collective, RSC_START, !pcmk_is_set(state, instance_starting), true); started = pe__new_rsc_pseudo_action(collective, RSC_STARTED, !pcmk_is_set(state, instance_starting), false); started->priority = INFINITY; if (pcmk_any_flags_set(state, instance_active|instance_starting)) { pe__set_action_flags(started, pe_action_runnable); } // Create pseudo-actions for rsc stop and stopped stop = pe__new_rsc_pseudo_action(collective, RSC_STOP, !pcmk_is_set(state, instance_stopping), true); stopped = pe__new_rsc_pseudo_action(collective, RSC_STOPPED, !pcmk_is_set(state, instance_stopping), true); stopped->priority = INFINITY; if (!pcmk_is_set(state, instance_restarting)) { pe__set_action_flags(stop, pe_action_migrate_runnable); } if (collective->variant == pe_clone) { pe__create_clone_notif_pseudo_ops(collective, start, started, stop, stopped); } } /*! * \internal * \brief Get a list of clone instances or bundle replica containers * * \param[in] rsc Clone or bundle resource * * \return Clone instances if \p rsc is a clone, or a newly created list of * \p rsc's replica containers if \p rsc is a bundle * \note The caller must call free_instance_list() on the result when the list * is no longer needed. */ static inline GList * get_instance_list(const pe_resource_t *rsc) { if (rsc->variant == pe_container) { return pe__bundle_containers(rsc); } else { return rsc->children; } } /*! * \internal * \brief Free any memory created by get_instance_list() * * \param[in] rsc Clone or bundle resource passed to get_instance_list() * \param[in,out] list Return value of get_instance_list() for \p rsc */ static inline void free_instance_list(const pe_resource_t *rsc, GList *list) { if (list != rsc->children) { g_list_free(list); } } /*! * \internal * \brief Check whether an instance is compatible with a role and node * * \param[in] instance Clone instance or bundle replica container * \param[in] node Instance must match this node * \param[in] role If not RSC_ROLE_UNKNOWN, instance must match this role * \param[in] current If true, compare instance's original node and role, * otherwise compare assigned next node and role * * \return true if \p instance is compatible with \p node and \p role, * otherwise false */ bool pcmk__instance_matches(const pe_resource_t *instance, const pe_node_t *node, enum rsc_role_e role, bool current) { pe_node_t *instance_node = NULL; CRM_CHECK((instance != NULL) && (node != NULL), return false); if ((role != RSC_ROLE_UNKNOWN) && (role != instance->fns->state(instance, current))) { pe_rsc_trace(instance, "%s is not a compatible instance (role is not %s)", instance->id, role2text(role)); return false; } if (!is_set_recursive(instance, pe_rsc_block, true)) { // We only want instances that haven't failed instance_node = instance->fns->location(instance, NULL, current); } if (instance_node == NULL) { pe_rsc_trace(instance, "%s is not a compatible instance (not assigned to a node)", instance->id); return false; } if (!pe__same_node(instance_node, node)) { pe_rsc_trace(instance, "%s is not a compatible instance (assigned to %s not %s)", instance->id, pe__node_name(instance_node), pe__node_name(node)); return false; } return true; } /*! * \internal * \brief Find an instance that matches a given resource by node and role * * \param[in] match_rsc Resource that instance must match (for logging only) * \param[in] rsc Clone or bundle resource to check for matching instance * \param[in] node Instance must match this node * \param[in] role If not RSC_ROLE_UNKNOWN, instance must match this role * \param[in] current If true, compare instance's original node and role, * otherwise compare assigned next node and role * * \return \p rsc instance matching \p node and \p role if any, otherwise NULL */ static pe_resource_t * find_compatible_instance_on_node(const pe_resource_t *match_rsc, const pe_resource_t *rsc, const pe_node_t *node, enum rsc_role_e role, bool current) { GList *instances = NULL; instances = get_instance_list(rsc); for (GList *iter = instances; iter != NULL; iter = iter->next) { pe_resource_t *instance = (pe_resource_t *) iter->data; if (pcmk__instance_matches(instance, node, role, current)) { pe_rsc_trace(match_rsc, "Found %s %s instance %s compatible with %s on %s", role == RSC_ROLE_UNKNOWN? "matching" : role2text(role), rsc->id, instance->id, match_rsc->id, pe__node_name(node)); free_instance_list(rsc, instances); // Only frees list, not contents return instance; } } free_instance_list(rsc, instances); pe_rsc_trace(match_rsc, "No %s %s instance found compatible with %s on %s", ((role == RSC_ROLE_UNKNOWN)? "matching" : role2text(role)), rsc->id, match_rsc->id, pe__node_name(node)); return NULL; } /*! * \internal * \brief Find a clone instance or bundle container compatible with a resource * * \param[in] match_rsc Resource that instance must match * \param[in] rsc Clone or bundle resource to check for matching instance * \param[in] role If not RSC_ROLE_UNKNOWN, instance must match this role * \param[in] current If true, compare instance's original node and role, * otherwise compare assigned next node and role * * \return Compatible (by \p role and \p match_rsc location) instance of \p rsc * if any, otherwise NULL */ pe_resource_t * pcmk__find_compatible_instance(const pe_resource_t *match_rsc, const pe_resource_t *rsc, enum rsc_role_e role, bool current) { pe_resource_t *instance = NULL; GList *nodes = NULL; const pe_node_t *node = match_rsc->fns->location(match_rsc, NULL, current); // If match_rsc has a node, check only that node if (node != NULL) { return find_compatible_instance_on_node(match_rsc, rsc, node, role, current); } // Otherwise check for an instance matching any of match_rsc's allowed nodes nodes = pcmk__sort_nodes(g_hash_table_get_values(match_rsc->allowed_nodes), NULL); for (GList *iter = nodes; (iter != NULL) && (instance == NULL); iter = iter->next) { instance = find_compatible_instance_on_node(match_rsc, rsc, (pe_node_t *) iter->data, role, current); } if (instance == NULL) { pe_rsc_debug(rsc, "No %s instance found compatible with %s", rsc->id, match_rsc->id); } g_list_free(nodes); return instance; } /*! * \internal * \brief Unassign an instance if mandatory ordering has no interleave match * * \param[in] first 'First' action in an ordering * \param[in] then 'Then' action in an ordering * \param[in,out] then_instance 'Then' instance that has no interleave match * \param[in] type Group of enum pe_ordering flags to apply * \param[in] current If true, "then" action is stopped or demoted * * \return true if \p then_instance was unassigned, otherwise false */ static bool unassign_if_mandatory(const pe_action_t *first, const pe_action_t *then, pe_resource_t *then_instance, uint32_t type, bool current) { // Allow "then" instance to go down even without an interleave match if (current) { pe_rsc_trace(then->rsc, "%s has no instance to order before stopping " "or demoting %s", first->rsc->id, then_instance->id); /* If the "first" action must be runnable, but there is no "first" * instance, the "then" instance must not be allowed to come up. */ } else if (pcmk_any_flags_set(type, pe_order_runnable_left |pe_order_implies_then)) { pe_rsc_info(then->rsc, "Inhibiting %s from being active " "because there is no %s instance to interleave", then_instance->id, first->rsc->id); return pcmk__assign_resource(then_instance, NULL, true, true); } return false; } /*! * \internal * \brief Find first matching action for a clone instance or bundle container * * \param[in] action Action in an interleaved ordering * \param[in] instance Clone instance or bundle container being interleaved * \param[in] action_name Action to look for * \param[in] node If not NULL, require action to be on this node * \param[in] for_first If true, \p instance is the 'first' resource in the * ordering, otherwise it is the 'then' resource * * \return First action for \p instance (or in some cases if \p instance is a * bundle container, its containerized resource) that matches * \p action_name and \p node if any, otherwise NULL */ static pe_action_t * find_instance_action(const pe_action_t *action, const pe_resource_t *instance, const char *action_name, const pe_node_t *node, bool for_first) { const pe_resource_t *rsc = NULL; pe_action_t *matching_action = NULL; /* If instance is a bundle container, sometimes we should interleave the * action for the container itself, and sometimes for the containerized * resource. * * For example, given "start bundle A then bundle B", B likely requires the * service inside A's container to be active, rather than just the * container, so we should interleave the action for A's containerized * resource. On the other hand, it's possible B's container itself requires * something from A, so we should interleave the action for B's container. * * Essentially, for 'first', we should use the containerized resource for * everything except stop, and for 'then', we should use the container for * everything except promote and demote (which can only be performed on the * containerized resource). */ if ((for_first && !pcmk__str_any_of(action->task, CRMD_ACTION_STOP, CRMD_ACTION_STOPPED, NULL)) || (!for_first && pcmk__str_any_of(action->task, CRMD_ACTION_PROMOTE, CRMD_ACTION_PROMOTED, CRMD_ACTION_DEMOTE, CRMD_ACTION_DEMOTED, NULL))) { rsc = pe__get_rsc_in_container(instance); } if (rsc == NULL) { rsc = instance; // No containerized resource, use instance itself } else { node = NULL; // Containerized actions are on bundle-created guest } matching_action = find_first_action(rsc->actions, NULL, action_name, node); if (matching_action != NULL) { return matching_action; } if (pcmk_is_set(instance->flags, pe_rsc_orphan) || pcmk__str_any_of(action_name, RSC_STOP, RSC_DEMOTE, NULL)) { crm_trace("No %s action found for %s%s", action_name, pcmk_is_set(instance->flags, pe_rsc_orphan)? "orphan " : "", instance->id); } else { crm_err("No %s action found for %s to interleave (bug?)", action_name, instance->id); } return NULL; } /*! * \internal * \brief Get the original action name of a bundle or clone action * * Given an action for a bundle or clone, get the original action name, * mapping notify to the action being notified, and if the instances are * primitives, mapping completion actions to the action that was completed * (for example, stopped to stop). * * \param[in] action Clone or bundle action to check * * \return Original action name for \p action */ static const char * orig_action_name(const pe_action_t *action) { const pe_resource_t *instance = action->rsc->children->data; // Any instance char *action_type = NULL; const char *action_name = action->task; enum action_tasks orig_task = no_action; if (pcmk__strcase_any_of(action->task, CRMD_ACTION_NOTIFY, CRMD_ACTION_NOTIFIED, NULL)) { // action->uuid is RSC_(confirmed-){pre,post}_notify_ACTION_INTERVAL CRM_CHECK(parse_op_key(action->uuid, NULL, &action_type, NULL), return task2text(no_action)); action_name = strstr(action_type, "_notify_"); CRM_CHECK(action_name != NULL, return task2text(no_action)); action_name += strlen("_notify_"); } orig_task = get_complex_task(instance, action_name); free(action_type); return task2text(orig_task); } /*! * \internal * \brief Update two interleaved actions according to an ordering between them * * Given information about an ordering of two interleaved actions, update the * actions' flags (and runnable_before members if appropriate) as appropriate * for the ordering. Effects may cascade to other orderings involving the * actions as well. * * \param[in,out] first 'First' action in an ordering * \param[in,out] then 'Then' action in an ordering * \param[in] node If not NULL, limit scope of ordering to this node * \param[in] filter Action flags to limit scope of certain updates (may * include pe_action_optional to affect only mandatory * actions, and pe_action_runnable to affect only * runnable actions) * \param[in] type Group of enum pe_ordering flags to apply * * \return Group of enum pcmk__updated flags indicating what was updated */ static uint32_t update_interleaved_actions(pe_action_t *first, pe_action_t *then, const pe_node_t *node, uint32_t filter, uint32_t type) { GList *instances = NULL; uint32_t changed = pcmk__updated_none; const char *orig_first_task = orig_action_name(first); // Stops and demotes must be interleaved with instance on current node bool current = pcmk__ends_with(first->uuid, "_" CRMD_ACTION_STOPPED "_0") || pcmk__ends_with(first->uuid, "_" CRMD_ACTION_DEMOTED "_0"); // Update the specified actions for each "then" instance individually instances = get_instance_list(then->rsc); for (GList *iter = instances; iter != NULL; iter = iter->next) { pe_resource_t *first_instance = NULL; pe_resource_t *then_instance = iter->data; pe_action_t *first_action = NULL; pe_action_t *then_action = NULL; // Find a "first" instance to interleave with this "then" instance first_instance = pcmk__find_compatible_instance(then_instance, first->rsc, RSC_ROLE_UNKNOWN, current); if (first_instance == NULL) { // No instance can be interleaved if (unassign_if_mandatory(first, then, then_instance, type, current)) { pcmk__set_updated_flags(changed, first, pcmk__updated_then); } continue; } first_action = find_instance_action(first, first_instance, orig_first_task, node, true); if (first_action == NULL) { continue; } then_action = find_instance_action(then, then_instance, then->task, node, false); if (then_action == NULL) { continue; } if (order_actions(first_action, then_action, type)) { pcmk__set_updated_flags(changed, first, pcmk__updated_first|pcmk__updated_then); } changed |= then_instance->cmds->update_ordered_actions( first_action, then_action, node, first_instance->cmds->action_flags(first_action, node), filter, type, then->rsc->cluster); } free_instance_list(then->rsc, instances); return changed; } /*! * \internal * \brief Check whether two actions in an ordering can be interleaved * * \param[in] first 'First' action in the ordering * \param[in] then 'Then' action in the ordering * * \return true if \p first and \p then can be interleaved, otherwise false */ static bool can_interleave_actions(const pe_action_t *first, const pe_action_t *then) { bool interleave = false; pe_resource_t *rsc = NULL; if ((first->rsc == NULL) || (then->rsc == NULL)) { crm_trace("Not interleaving %s with %s: not resource actions", first->uuid, then->uuid); return false; } if (first->rsc == then->rsc) { crm_trace("Not interleaving %s with %s: same resource", first->uuid, then->uuid); return false; } if ((first->rsc->variant < pe_clone) || (then->rsc->variant < pe_clone)) { crm_trace("Not interleaving %s with %s: not clones or bundles", first->uuid, then->uuid); return false; } if (pcmk__ends_with(then->uuid, "_stop_0") || pcmk__ends_with(then->uuid, "_demote_0")) { rsc = first->rsc; } else { rsc = then->rsc; } interleave = crm_is_true(g_hash_table_lookup(rsc->meta, XML_RSC_ATTR_INTERLEAVE)); pe_rsc_trace(rsc, "'%s then %s' will %sbe interleaved (based on %s)", first->uuid, then->uuid, (interleave? "" : "not "), rsc->id); return interleave; } /*! * \internal * \brief Update non-interleaved instance actions according to an ordering * * Given information about an ordering of two non-interleaved actions, update * the actions' flags (and runnable_before members if appropriate) as * appropriate for the ordering. Effects may cascade to other orderings * involving the actions as well. * * \param[in,out] instance Clone instance or bundle container * \param[in,out] first "First" action in ordering * \param[in] then "Then" action in ordering (for \p instance's parent) * \param[in] node If not NULL, limit scope of ordering to this node * \param[in] flags Action flags for \p first for ordering purposes * \param[in] filter Action flags to limit scope of certain updates (may * include pe_action_optional to affect only mandatory * actions, and pe_action_runnable to affect only * runnable actions) * \param[in] type Group of enum pe_ordering flags to apply * * \return Group of enum pcmk__updated flags indicating what was updated */ static uint32_t update_noninterleaved_actions(pe_resource_t *instance, pe_action_t *first, const pe_action_t *then, const pe_node_t *node, uint32_t flags, uint32_t filter, uint32_t type) { pe_action_t *instance_action = NULL; uint32_t instance_flags = 0; uint32_t changed = pcmk__updated_none; // Check whether instance has an equivalent of "then" action instance_action = find_first_action(instance->actions, NULL, then->task, node); if (instance_action == NULL) { return changed; } // Check whether action is runnable instance_flags = instance->cmds->action_flags(instance_action, node); if (!pcmk_is_set(instance_flags, pe_action_runnable)) { return changed; } // If so, update actions for the instance changed = instance->cmds->update_ordered_actions(first, instance_action, node, flags, filter, type, instance->cluster); // Propagate any changes to later actions if (pcmk_is_set(changed, pcmk__updated_then)) { for (GList *after_iter = instance_action->actions_after; after_iter != NULL; after_iter = after_iter->next) { pe_action_wrapper_t *after = after_iter->data; pcmk__update_action_for_orderings(after->action, instance->cluster); } } return changed; } /*! * \internal * \brief Update two actions according to an ordering between them * * Given information about an ordering of two clone or bundle actions, update * the actions' flags (and runnable_before members if appropriate) as * appropriate for the ordering. Effects may cascade to other orderings * involving the actions as well. * * \param[in,out] first 'First' action in an ordering * \param[in,out] then 'Then' action in an ordering * \param[in] node If not NULL, limit scope of ordering to this node * (only used when interleaving instances) * \param[in] flags Action flags for \p first for ordering purposes * \param[in] filter Action flags to limit scope of certain updates (may * include pe_action_optional to affect only mandatory * actions, and pe_action_runnable to affect only * runnable actions) * \param[in] type Group of enum pe_ordering flags to apply * \param[in,out] data_set Cluster working set * * \return Group of enum pcmk__updated flags indicating what was updated */ uint32_t pcmk__instance_update_ordered_actions(pe_action_t *first, pe_action_t *then, const pe_node_t *node, uint32_t flags, uint32_t filter, uint32_t type, pe_working_set_t *data_set) { CRM_ASSERT((first != NULL) && (then != NULL) && (data_set != NULL)); if (then->rsc == NULL) { return pcmk__updated_none; } else if (can_interleave_actions(first, then)) { return update_interleaved_actions(first, then, node, filter, type); } else { uint32_t changed = pcmk__updated_none; GList *instances = get_instance_list(then->rsc); // Update actions for the clone or bundle resource itself changed |= pcmk__update_ordered_actions(first, then, node, flags, filter, type, data_set); // Update the 'then' clone instances or bundle containers individually for (GList *iter = instances; iter != NULL; iter = iter->next) { pe_resource_t *instance = iter->data; changed |= update_noninterleaved_actions(instance, first, then, node, flags, filter, type); } free_instance_list(then->rsc, instances); return changed; } } #define pe__clear_action_summary_flags(flags, action, flag) do { \ flags = pcmk__clear_flags_as(__func__, __LINE__, LOG_TRACE, \ "Action summary", action->rsc->id, \ flags, flag, #flag); \ } while (0) /*! * \internal * \brief Return action flags for a given clone or bundle action * * \param[in,out] action Action for a clone or bundle * \param[in] instances Clone instances or bundle containers * \param[in] node If not NULL, limit effects to this node * * \return Flags appropriate to \p action on \p node */ uint32_t pcmk__collective_action_flags(pe_action_t *action, const GList *instances, const pe_node_t *node) { bool any_runnable = false; const char *action_name = orig_action_name(action); // Set original assumptions (optional and runnable may be cleared below) uint32_t flags = pe_action_optional|pe_action_runnable|pe_action_pseudo; for (const GList *iter = instances; iter != NULL; iter = iter->next) { const pe_resource_t *instance = iter->data; const pe_node_t *instance_node = NULL; pe_action_t *instance_action = NULL; uint32_t instance_flags; // Node is relevant only to primitive instances if (instance->variant == pe_native) { instance_node = node; } instance_action = find_first_action(instance->actions, NULL, action_name, instance_node); if (instance_action == NULL) { pe_rsc_trace(action->rsc, "%s has no %s action on %s", instance->id, action_name, pe__node_name(node)); continue; } pe_rsc_trace(action->rsc, "%s has %s for %s on %s", instance->id, instance_action->uuid, action_name, pe__node_name(node)); instance_flags = instance->cmds->action_flags(instance_action, node); // If any instance action is mandatory, so is the collective action if (pcmk_is_set(flags, pe_action_optional) && !pcmk_is_set(instance_flags, pe_action_optional)) { pe_rsc_trace(instance, "%s is mandatory because %s is", action->uuid, instance_action->uuid); pe__clear_action_summary_flags(flags, action, pe_action_optional); pe__clear_action_flags(action, pe_action_optional); } // If any instance action is runnable, so is the collective action if (pcmk_is_set(instance_flags, pe_action_runnable)) { any_runnable = true; } } if (!any_runnable) { pe_rsc_trace(action->rsc, "%s is not runnable because no instance can run %s", action->uuid, action_name); pe__clear_action_summary_flags(flags, action, pe_action_runnable); if (node == NULL) { pe__clear_action_flags(action, pe_action_runnable); } } return flags; } diff --git a/lib/pacemaker/pcmk_sched_primitive.c b/lib/pacemaker/pcmk_sched_primitive.c index b130e15c62..10ed831cbc 100644 --- a/lib/pacemaker/pcmk_sched_primitive.c +++ b/lib/pacemaker/pcmk_sched_primitive.c @@ -1,1631 +1,1631 @@ /* * Copyright 2004-2023 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU General Public License version 2 * or later (GPLv2+) WITHOUT ANY WARRANTY. */ #include #include #include // uint8_t, uint32_t #include #include #include "libpacemaker_private.h" static void stop_resource(pe_resource_t *rsc, pe_node_t *node, bool optional); static void start_resource(pe_resource_t *rsc, pe_node_t *node, bool optional); static void demote_resource(pe_resource_t *rsc, pe_node_t *node, bool optional); static void promote_resource(pe_resource_t *rsc, pe_node_t *node, bool optional); static void assert_role_error(pe_resource_t *rsc, pe_node_t *node, bool optional); static enum rsc_role_e rsc_state_matrix[RSC_ROLE_MAX][RSC_ROLE_MAX] = { /* This array lists the immediate next role when transitioning from one role * to a target role. For example, when going from Stopped to Promoted, the * next role is Unpromoted, because the resource must be started before it * can be promoted. The current state then becomes Started, which is fed * into this array again, giving a next role of Promoted. * * Current role Immediate next role Final target role * ------------ ------------------- ----------------- */ /* Unknown */ { RSC_ROLE_UNKNOWN, /* Unknown */ RSC_ROLE_STOPPED, /* Stopped */ RSC_ROLE_STOPPED, /* Started */ RSC_ROLE_STOPPED, /* Unpromoted */ RSC_ROLE_STOPPED, /* Promoted */ }, /* Stopped */ { RSC_ROLE_STOPPED, /* Unknown */ RSC_ROLE_STOPPED, /* Stopped */ RSC_ROLE_STARTED, /* Started */ RSC_ROLE_UNPROMOTED, /* Unpromoted */ RSC_ROLE_UNPROMOTED, /* Promoted */ }, /* Started */ { RSC_ROLE_STOPPED, /* Unknown */ RSC_ROLE_STOPPED, /* Stopped */ RSC_ROLE_STARTED, /* Started */ RSC_ROLE_UNPROMOTED, /* Unpromoted */ RSC_ROLE_PROMOTED, /* Promoted */ }, /* Unpromoted */ { RSC_ROLE_STOPPED, /* Unknown */ RSC_ROLE_STOPPED, /* Stopped */ RSC_ROLE_STOPPED, /* Started */ RSC_ROLE_UNPROMOTED, /* Unpromoted */ RSC_ROLE_PROMOTED, /* Promoted */ }, /* Promoted */ { RSC_ROLE_STOPPED, /* Unknown */ RSC_ROLE_UNPROMOTED, /* Stopped */ RSC_ROLE_UNPROMOTED, /* Started */ RSC_ROLE_UNPROMOTED, /* Unpromoted */ RSC_ROLE_PROMOTED, /* Promoted */ }, }; /*! * \internal * \brief Function to schedule actions needed for a role change * * \param[in,out] rsc Resource whose role is changing * \param[in,out] node Node where resource will be in its next role * \param[in] optional Whether scheduled actions should be optional */ typedef void (*rsc_transition_fn)(pe_resource_t *rsc, pe_node_t *node, bool optional); static rsc_transition_fn rsc_action_matrix[RSC_ROLE_MAX][RSC_ROLE_MAX] = { /* This array lists the function needed to transition directly from one role * to another. NULL indicates that nothing is needed. * * Current role Transition function Next role * ------------ ------------------- ---------- */ /* Unknown */ { assert_role_error, /* Unknown */ stop_resource, /* Stopped */ assert_role_error, /* Started */ assert_role_error, /* Unpromoted */ assert_role_error, /* Promoted */ }, /* Stopped */ { assert_role_error, /* Unknown */ NULL, /* Stopped */ start_resource, /* Started */ start_resource, /* Unpromoted */ assert_role_error, /* Promoted */ }, /* Started */ { assert_role_error, /* Unknown */ stop_resource, /* Stopped */ NULL, /* Started */ NULL, /* Unpromoted */ promote_resource, /* Promoted */ }, /* Unpromoted */ { assert_role_error, /* Unknown */ stop_resource, /* Stopped */ stop_resource, /* Started */ NULL, /* Unpromoted */ promote_resource, /* Promoted */ }, /* Promoted */ { assert_role_error, /* Unknown */ demote_resource, /* Stopped */ demote_resource, /* Started */ demote_resource, /* Unpromoted */ NULL, /* Promoted */ }, }; /*! * \internal * \brief Get a list of a resource's allowed nodes sorted by node score * * \param[in] rsc Resource to check * * \return List of allowed nodes sorted by node score */ static GList * sorted_allowed_nodes(const pe_resource_t *rsc) { if (rsc->allowed_nodes != NULL) { GList *nodes = g_hash_table_get_values(rsc->allowed_nodes); if (nodes != NULL) { return pcmk__sort_nodes(nodes, pe__current_node(rsc)); } } return NULL; } /*! * \internal * \brief Assign a resource to its best allowed node, if possible * * \param[in,out] rsc Resource to choose a node for * \param[in] prefer If not \c NULL, prefer this node when all else * equal * \param[in] stop_if_fail If \c true and \p rsc can't be assigned to a * node, set next role to stopped and update * existing actions * * \return true if \p rsc could be assigned to a node, otherwise false * * \note If \p stop_if_fail is \c false, then \c pcmk__unassign_resource() can * completely undo the assignment. A successful assignment can be either * undone or left alone as final. A failed assignment has the same effect * as calling pcmk__unassign_resource(); there are no side effects on * roles or actions. */ static bool assign_best_node(pe_resource_t *rsc, const pe_node_t *prefer, bool stop_if_fail) { GList *nodes = NULL; pe_node_t *chosen = NULL; pe_node_t *best = NULL; const pe_node_t *most_free_node = pcmk__ban_insufficient_capacity(rsc); if (prefer == NULL) { prefer = most_free_node; } if (!pcmk_is_set(rsc->flags, pe_rsc_provisional)) { // We've already finished assignment of resources to nodes return rsc->allocated_to != NULL; } // Sort allowed nodes by score nodes = sorted_allowed_nodes(rsc); if (nodes != NULL) { best = (pe_node_t *) nodes->data; // First node has best score } if ((prefer != NULL) && (nodes != NULL)) { // Get the allowed node version of prefer chosen = g_hash_table_lookup(rsc->allowed_nodes, prefer->details->id); if (chosen == NULL) { pe_rsc_trace(rsc, "Preferred node %s for %s was unknown", pe__node_name(prefer), rsc->id); /* Favor the preferred node as long as its score is at least as good as * the best allowed node's. * * An alternative would be to favor the preferred node even if the best * node is better, when the best node's score is less than INFINITY. */ } else if (chosen->weight < best->weight) { pe_rsc_trace(rsc, "Preferred node %s for %s was unsuitable", pe__node_name(chosen), rsc->id); chosen = NULL; } else if (!pcmk__node_available(chosen, true, false)) { pe_rsc_trace(rsc, "Preferred node %s for %s was unavailable", pe__node_name(chosen), rsc->id); chosen = NULL; } else { pe_rsc_trace(rsc, "Chose preferred node %s for %s " "(ignoring %d candidates)", pe__node_name(chosen), rsc->id, g_list_length(nodes)); } } if ((chosen == NULL) && (best != NULL)) { /* Either there is no preferred node, or the preferred node is not * suitable, but another node is allowed to run the resource. */ chosen = best; if (!pe_rsc_is_unique_clone(rsc->parent) && (chosen->weight > 0) // Zero not acceptable && pcmk__node_available(chosen, false, false)) { /* If the resource is already running on a node, prefer that node if * it is just as good as the chosen node. * * We don't do this for unique clone instances, because * pcmk__assign_instances() has already assigned instances to their * running nodes when appropriate, and if we get here, we don't want * remaining unassigned instances to prefer a node that's already * running another instance. */ pe_node_t *running = pe__current_node(rsc); if (running == NULL) { // Nothing to do } else if (!pcmk__node_available(running, true, false)) { pe_rsc_trace(rsc, "Current node for %s (%s) can't run resources", rsc->id, pe__node_name(running)); } else { int nodes_with_best_score = 1; for (GList *iter = nodes->next; iter; iter = iter->next) { pe_node_t *allowed = (pe_node_t *) iter->data; if (allowed->weight != chosen->weight) { // The nodes are sorted by score, so no more are equal break; } if (pe__same_node(allowed, running)) { // Scores are equal, so prefer the current node chosen = allowed; } nodes_with_best_score++; } if (nodes_with_best_score > 1) { uint8_t log_level = LOG_INFO; if (chosen->weight >= INFINITY) { log_level = LOG_WARNING; } do_crm_log(log_level, "Chose %s for %s from %d nodes with score %s", pe__node_name(chosen), rsc->id, nodes_with_best_score, pcmk_readable_score(chosen->weight)); } } } pe_rsc_trace(rsc, "Chose %s for %s from %d candidates", pe__node_name(chosen), rsc->id, g_list_length(nodes)); } pcmk__assign_resource(rsc, chosen, false, stop_if_fail); g_list_free(nodes); return rsc->allocated_to != NULL; } /*! * \internal * \brief Apply a "this with" colocation to a node's allowed node scores * * \param[in,out] colocation Colocation to apply * \param[in,out] rsc Resource being assigned */ static void apply_this_with(pcmk__colocation_t *colocation, pe_resource_t *rsc) { GHashTable *archive = NULL; pe_resource_t *other = colocation->primary; // In certain cases, we will need to revert the node scores if ((colocation->dependent_role >= RSC_ROLE_PROMOTED) || ((colocation->score < 0) && (colocation->score > -INFINITY))) { archive = pcmk__copy_node_table(rsc->allowed_nodes); } if (pcmk_is_set(other->flags, pe_rsc_provisional)) { pe_rsc_trace(rsc, "%s: Assigning colocation %s primary %s first" "(score=%d role=%s)", rsc->id, colocation->id, other->id, colocation->score, role2text(colocation->dependent_role)); other->cmds->assign(other, NULL, true); } // Apply the colocation score to this resource's allowed node scores rsc->cmds->apply_coloc_score(rsc, other, colocation, true); if ((archive != NULL) && !pcmk__any_node_available(rsc->allowed_nodes)) { pe_rsc_info(rsc, "%s: Reverting scores from colocation with %s " "because no nodes allowed", rsc->id, other->id); g_hash_table_destroy(rsc->allowed_nodes); rsc->allowed_nodes = archive; archive = NULL; } if (archive != NULL) { g_hash_table_destroy(archive); } } /*! * \internal * \brief Update a Pacemaker Remote node once its connection has been assigned * * \param[in] connection Connection resource that has been assigned */ static void remote_connection_assigned(const pe_resource_t *connection) { pe_node_t *remote_node = pe_find_node(connection->cluster->nodes, connection->id); CRM_CHECK(remote_node != NULL, return); if ((connection->allocated_to != NULL) && (connection->next_role != RSC_ROLE_STOPPED)) { crm_trace("Pacemaker Remote node %s will be online", remote_node->details->id); remote_node->details->online = TRUE; if (remote_node->details->unseen) { // Avoid unnecessary fence, since we will attempt connection remote_node->details->unclean = FALSE; } } else { crm_trace("Pacemaker Remote node %s will be shut down " "(%sassigned connection's next role is %s)", remote_node->details->id, ((connection->allocated_to == NULL)? "un" : ""), role2text(connection->next_role)); remote_node->details->shutdown = TRUE; } } /*! * \internal * \brief Assign a primitive resource to a node * * \param[in,out] rsc Resource to assign to a node * \param[in] prefer Node to prefer, if all else is equal * \param[in] stop_if_fail If \c true and \p rsc can't be assigned to a * node, set next role to stopped and update * existing actions * * \return Node that \p rsc is assigned to, if assigned entirely to one node * * \note If \p stop_if_fail is \c false, then \c pcmk__unassign_resource() can * completely undo the assignment. A successful assignment can be either * undone or left alone as final. A failed assignment has the same effect * as calling pcmk__unassign_resource(); there are no side effects on * roles or actions. */ pe_node_t * pcmk__primitive_assign(pe_resource_t *rsc, const pe_node_t *prefer, bool stop_if_fail) { GList *this_with_colocations = NULL; GList *with_this_colocations = NULL; GList *iter = NULL; pcmk__colocation_t *colocation = NULL; CRM_ASSERT((rsc != NULL) && (rsc->variant == pe_native)); // Never assign a child without parent being assigned first if ((rsc->parent != NULL) && !pcmk_is_set(rsc->parent->flags, pe_rsc_allocating)) { pe_rsc_debug(rsc, "%s: Assigning parent %s first", rsc->id, rsc->parent->id); rsc->parent->cmds->assign(rsc->parent, prefer, stop_if_fail); } if (!pcmk_is_set(rsc->flags, pe_rsc_provisional)) { // Assignment has already been done const char *node_name = "no node"; if (rsc->allocated_to != NULL) { node_name = pe__node_name(rsc->allocated_to); } pe_rsc_debug(rsc, "%s: pre-assigned to %s", rsc->id, node_name); return rsc->allocated_to; } // Ensure we detect assignment loops if (pcmk_is_set(rsc->flags, pe_rsc_allocating)) { pe_rsc_debug(rsc, "Breaking assignment loop involving %s", rsc->id); return NULL; } pe__set_resource_flags(rsc, pe_rsc_allocating); pe__show_node_scores(true, rsc, "Pre-assignment", rsc->allowed_nodes, rsc->cluster); this_with_colocations = pcmk__this_with_colocations(rsc); with_this_colocations = pcmk__with_this_colocations(rsc); // Apply mandatory colocations first, to satisfy as many as possible for (iter = this_with_colocations; iter != NULL; iter = iter->next) { colocation = iter->data; if ((colocation->score <= -CRM_SCORE_INFINITY) || (colocation->score >= CRM_SCORE_INFINITY)) { apply_this_with(colocation, rsc); } } for (iter = with_this_colocations; iter != NULL; iter = iter->next) { colocation = iter->data; if ((colocation->score <= -CRM_SCORE_INFINITY) || (colocation->score >= CRM_SCORE_INFINITY)) { pcmk__add_dependent_scores(colocation, rsc); } } pe__show_node_scores(true, rsc, "Mandatory-colocations", rsc->allowed_nodes, rsc->cluster); // Then apply optional colocations for (iter = this_with_colocations; iter != NULL; iter = iter->next) { colocation = iter->data; if ((colocation->score > -CRM_SCORE_INFINITY) && (colocation->score < CRM_SCORE_INFINITY)) { apply_this_with(colocation, rsc); } } for (iter = with_this_colocations; iter != NULL; iter = iter->next) { colocation = iter->data; if ((colocation->score > -CRM_SCORE_INFINITY) && (colocation->score < CRM_SCORE_INFINITY)) { pcmk__add_dependent_scores(colocation, rsc); } } g_list_free(this_with_colocations); g_list_free(with_this_colocations); if (rsc->next_role == RSC_ROLE_STOPPED) { pe_rsc_trace(rsc, "Banning %s from all nodes because it will be stopped", rsc->id); resource_location(rsc, NULL, -INFINITY, XML_RSC_ATTR_TARGET_ROLE, rsc->cluster); } else if ((rsc->next_role > rsc->role) && !pcmk_is_set(rsc->cluster->flags, pe_flag_have_quorum) && (rsc->cluster->no_quorum_policy == no_quorum_freeze)) { crm_notice("Resource %s cannot be elevated from %s to %s due to " "no-quorum-policy=freeze", rsc->id, role2text(rsc->role), role2text(rsc->next_role)); pe__set_next_role(rsc, rsc->role, "no-quorum-policy=freeze"); } pe__show_node_scores(!pcmk_is_set(rsc->cluster->flags, pe_flag_show_scores), rsc, __func__, rsc->allowed_nodes, rsc->cluster); // Unmanage resource if fencing is enabled but no device is configured if (pcmk_is_set(rsc->cluster->flags, pe_flag_stonith_enabled) && !pcmk_is_set(rsc->cluster->flags, pe_flag_have_stonith_resource)) { pe__clear_resource_flags(rsc, pe_rsc_managed); } if (!pcmk_is_set(rsc->flags, pe_rsc_managed)) { // Unmanaged resources stay on their current node const char *reason = NULL; pe_node_t *assign_to = NULL; pe__set_next_role(rsc, rsc->role, "unmanaged"); assign_to = pe__current_node(rsc); if (assign_to == NULL) { reason = "inactive"; } else if (rsc->role == RSC_ROLE_PROMOTED) { reason = "promoted"; } else if (pcmk_is_set(rsc->flags, pe_rsc_failed)) { reason = "failed"; } else { reason = "active"; } pe_rsc_info(rsc, "Unmanaged resource %s assigned to %s: %s", rsc->id, (assign_to? assign_to->details->uname : "no node"), reason); pcmk__assign_resource(rsc, assign_to, true, stop_if_fail); } else if (pcmk_is_set(rsc->cluster->flags, pe_flag_stop_everything)) { // Must stop at some point, but be consistent with stop_if_fail if (stop_if_fail) { pe_rsc_debug(rsc, "Forcing %s to stop: stop-all-resources", rsc->id); } pcmk__assign_resource(rsc, NULL, true, stop_if_fail); } else if (!assign_best_node(rsc, prefer, stop_if_fail)) { // Assignment failed if (!pcmk_is_set(rsc->flags, pe_rsc_orphan)) { pe_rsc_info(rsc, "Resource %s cannot run anywhere", rsc->id); } else if ((rsc->running_on != NULL) && stop_if_fail) { pe_rsc_info(rsc, "Stopping orphan resource %s", rsc->id); } } pe__clear_resource_flags(rsc, pe_rsc_allocating); if (rsc->is_remote_node) { remote_connection_assigned(rsc); } return rsc->allocated_to; } /*! * \internal * \brief Schedule actions to bring resource down and back to current role * * \param[in,out] rsc Resource to restart * \param[in,out] current Node that resource should be brought down on * \param[in] need_stop Whether the resource must be stopped * \param[in] need_promote Whether the resource must be promoted * * \return Role that resource would have after scheduled actions are taken */ static void schedule_restart_actions(pe_resource_t *rsc, pe_node_t *current, bool need_stop, bool need_promote) { enum rsc_role_e role = rsc->role; enum rsc_role_e next_role; rsc_transition_fn fn = NULL; pe__set_resource_flags(rsc, pe_rsc_restarting); // Bring resource down to a stop on its current node while (role != RSC_ROLE_STOPPED) { next_role = rsc_state_matrix[role][RSC_ROLE_STOPPED]; pe_rsc_trace(rsc, "Creating %s action to take %s down from %s to %s", (need_stop? "required" : "optional"), rsc->id, role2text(role), role2text(next_role)); fn = rsc_action_matrix[role][next_role]; if (fn == NULL) { break; } fn(rsc, current, !need_stop); role = next_role; } // Bring resource up to its next role on its next node while ((rsc->role <= rsc->next_role) && (role != rsc->role) && !pcmk_is_set(rsc->flags, pe_rsc_block)) { bool required = need_stop; next_role = rsc_state_matrix[role][rsc->role]; if ((next_role == RSC_ROLE_PROMOTED) && need_promote) { required = true; } pe_rsc_trace(rsc, "Creating %s action to take %s up from %s to %s", (required? "required" : "optional"), rsc->id, role2text(role), role2text(next_role)); fn = rsc_action_matrix[role][next_role]; if (fn == NULL) { break; } fn(rsc, rsc->allocated_to, !required); role = next_role; } pe__clear_resource_flags(rsc, pe_rsc_restarting); } /*! * \internal * \brief If a resource's next role is not explicitly specified, set a default * * \param[in,out] rsc Resource to set next role for * * \return "explicit" if next role was explicitly set, otherwise "implicit" */ static const char * set_default_next_role(pe_resource_t *rsc) { if (rsc->next_role != RSC_ROLE_UNKNOWN) { return "explicit"; } if (rsc->allocated_to == NULL) { pe__set_next_role(rsc, RSC_ROLE_STOPPED, "assignment"); } else { pe__set_next_role(rsc, RSC_ROLE_STARTED, "assignment"); } return "implicit"; } /*! * \internal * \brief Create an action to represent an already pending start * * \param[in,out] rsc Resource to create start action for */ static void create_pending_start(pe_resource_t *rsc) { pe_action_t *start = NULL; pe_rsc_trace(rsc, "Creating action for %s to represent already pending start", rsc->id); start = start_action(rsc, rsc->allocated_to, TRUE); pe__set_action_flags(start, pe_action_print_always); } /*! * \internal * \brief Schedule actions needed to take a resource to its next role * * \param[in,out] rsc Resource to schedule actions for */ static void schedule_role_transition_actions(pe_resource_t *rsc) { enum rsc_role_e role = rsc->role; while (role != rsc->next_role) { enum rsc_role_e next_role = rsc_state_matrix[role][rsc->next_role]; rsc_transition_fn fn = NULL; pe_rsc_trace(rsc, "Creating action to take %s from %s to %s (ending at %s)", rsc->id, role2text(role), role2text(next_role), role2text(rsc->next_role)); fn = rsc_action_matrix[role][next_role]; if (fn == NULL) { break; } fn(rsc, rsc->allocated_to, false); role = next_role; } } /*! * \internal * \brief Create all actions needed for a given primitive resource * * \param[in,out] rsc Primitive resource to create actions for */ void pcmk__primitive_create_actions(pe_resource_t *rsc) { bool need_stop = false; bool need_promote = false; bool is_moving = false; bool allow_migrate = false; bool multiply_active = false; pe_node_t *current = NULL; unsigned int num_all_active = 0; unsigned int num_clean_active = 0; const char *next_role_source = NULL; CRM_ASSERT((rsc != NULL) && (rsc->variant == pe_native)); next_role_source = set_default_next_role(rsc); pe_rsc_trace(rsc, "Creating all actions for %s transition from %s to %s " "(%s) on %s", rsc->id, role2text(rsc->role), role2text(rsc->next_role), next_role_source, pe__node_name(rsc->allocated_to)); current = rsc->fns->active_node(rsc, &num_all_active, &num_clean_active); g_list_foreach(rsc->dangling_migrations, pcmk__abort_dangling_migration, rsc); if ((current != NULL) && (rsc->allocated_to != NULL) && !pe__same_node(current, rsc->allocated_to) && (rsc->next_role >= RSC_ROLE_STARTED)) { pe_rsc_trace(rsc, "Moving %s from %s to %s", rsc->id, pe__node_name(current), pe__node_name(rsc->allocated_to)); is_moving = true; allow_migrate = pcmk__rsc_can_migrate(rsc, current); // This is needed even if migrating (though I'm not sure why ...) need_stop = true; } // Check whether resource is partially migrated and/or multiply active if ((rsc->partial_migration_source != NULL) && (rsc->partial_migration_target != NULL) && allow_migrate && (num_all_active == 2) && pe__same_node(current, rsc->partial_migration_source) && pe__same_node(rsc->allocated_to, rsc->partial_migration_target)) { /* A partial migration is in progress, and the migration target remains * the same as when the migration began. */ pe_rsc_trace(rsc, "Partial migration of %s from %s to %s will continue", rsc->id, pe__node_name(rsc->partial_migration_source), pe__node_name(rsc->partial_migration_target)); } else if ((rsc->partial_migration_source != NULL) || (rsc->partial_migration_target != NULL)) { // A partial migration is in progress but can't be continued if (num_all_active > 2) { // The resource is migrating *and* multiply active! crm_notice("Forcing recovery of %s because it is migrating " "from %s to %s and possibly active elsewhere", rsc->id, pe__node_name(rsc->partial_migration_source), pe__node_name(rsc->partial_migration_target)); } else { // The migration source or target isn't available crm_notice("Forcing recovery of %s because it can no longer " "migrate from %s to %s", rsc->id, pe__node_name(rsc->partial_migration_source), pe__node_name(rsc->partial_migration_target)); } need_stop = true; rsc->partial_migration_source = rsc->partial_migration_target = NULL; allow_migrate = false; } else if (pcmk_is_set(rsc->flags, pe_rsc_needs_fencing)) { multiply_active = (num_all_active > 1); } else { /* If a resource has "requires" set to nothing or quorum, don't consider * it active on unclean nodes (similar to how all resources behave when * stonith-enabled is false). We can start such resources elsewhere * before fencing completes, and if we considered the resource active on * the failed node, we would attempt recovery for being active on * multiple nodes. */ multiply_active = (num_clean_active > 1); } if (multiply_active) { const char *class = crm_element_value(rsc->xml, XML_AGENT_ATTR_CLASS); // Resource was (possibly) incorrectly multiply active pe_proc_err("%s resource %s might be active on %u nodes (%s)", pcmk__s(class, "Untyped"), rsc->id, num_all_active, recovery2text(rsc->recovery_type)); crm_notice("See https://wiki.clusterlabs.org/wiki/FAQ" "#Resource_is_Too_Active for more information"); switch (rsc->recovery_type) { case recovery_stop_start: need_stop = true; break; case recovery_stop_unexpected: need_stop = true; // stop_resource() will skip expected node pe__set_resource_flags(rsc, pe_rsc_stop_unexpected); break; default: break; } } else { pe__clear_resource_flags(rsc, pe_rsc_stop_unexpected); } if (pcmk_is_set(rsc->flags, pe_rsc_start_pending)) { create_pending_start(rsc); } if (is_moving) { // Remaining tests are only for resources staying where they are } else if (pcmk_is_set(rsc->flags, pe_rsc_failed)) { if (pcmk_is_set(rsc->flags, pe_rsc_stop)) { need_stop = true; pe_rsc_trace(rsc, "Recovering %s", rsc->id); } else { pe_rsc_trace(rsc, "Recovering %s by demotion", rsc->id); if (rsc->next_role == RSC_ROLE_PROMOTED) { need_promote = true; } } } else if (pcmk_is_set(rsc->flags, pe_rsc_block)) { pe_rsc_trace(rsc, "Blocking further actions on %s", rsc->id); need_stop = true; } else if ((rsc->role > RSC_ROLE_STARTED) && (current != NULL) && (rsc->allocated_to != NULL)) { pe_action_t *start = NULL; pe_rsc_trace(rsc, "Creating start action for promoted resource %s", rsc->id); start = start_action(rsc, rsc->allocated_to, TRUE); if (!pcmk_is_set(start->flags, pe_action_optional)) { // Recovery of a promoted resource pe_rsc_trace(rsc, "%s restart is required for recovery", rsc->id); need_stop = true; } } // Create any actions needed to bring resource down and back up to same role schedule_restart_actions(rsc, current, need_stop, need_promote); // Create any actions needed to take resource from this role to the next schedule_role_transition_actions(rsc); pcmk__create_recurring_actions(rsc); if (allow_migrate) { pcmk__create_migration_actions(rsc, current); } } /*! * \internal * \brief Ban a resource from any allowed nodes that are Pacemaker Remote nodes * * \param[in] rsc Resource to check */ static void rsc_avoids_remote_nodes(const pe_resource_t *rsc) { GHashTableIter iter; pe_node_t *node = NULL; g_hash_table_iter_init(&iter, rsc->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (void **) &node)) { if (node->details->remote_rsc != NULL) { node->weight = -INFINITY; } } } /*! * \internal * \brief Return allowed nodes as (possibly sorted) list * * Convert a resource's hash table of allowed nodes to a list. If printing to * stdout, sort the list, to keep action ID numbers consistent for regression * test output (while avoiding the performance hit on a live cluster). * * \param[in] rsc Resource to check for allowed nodes * * \return List of resource's allowed nodes * \note Callers should take care not to rely on the list being sorted. */ static GList * allowed_nodes_as_list(const pe_resource_t *rsc) { GList *allowed_nodes = NULL; if (rsc->allowed_nodes) { allowed_nodes = g_hash_table_get_values(rsc->allowed_nodes); } if (!pcmk__is_daemon) { allowed_nodes = g_list_sort(allowed_nodes, pe__cmp_node_name); } return allowed_nodes; } /*! * \internal * \brief Create implicit constraints needed for a primitive resource * * \param[in,out] rsc Primitive resource to create implicit constraints for */ void pcmk__primitive_internal_constraints(pe_resource_t *rsc) { GList *allowed_nodes = NULL; bool check_unfencing = false; bool check_utilization = false; CRM_ASSERT((rsc != NULL) && (rsc->variant == pe_native)); if (!pcmk_is_set(rsc->flags, pe_rsc_managed)) { pe_rsc_trace(rsc, "Skipping implicit constraints for unmanaged resource %s", rsc->id); return; } // Whether resource requires unfencing check_unfencing = !pcmk_is_set(rsc->flags, pe_rsc_fence_device) && pcmk_is_set(rsc->cluster->flags, pe_flag_enable_unfencing) && pcmk_is_set(rsc->flags, pe_rsc_needs_unfencing); // Whether a non-default placement strategy is used check_utilization = (g_hash_table_size(rsc->utilization) > 0) && !pcmk__str_eq(rsc->cluster->placement_strategy, "default", pcmk__str_casei); // Order stops before starts (i.e. restart) pcmk__new_ordering(rsc, pcmk__op_key(rsc->id, RSC_STOP, 0), NULL, rsc, pcmk__op_key(rsc->id, RSC_START, 0), NULL, pe_order_optional|pe_order_implies_then|pe_order_restart, rsc->cluster); // Promotable ordering: demote before stop, start before promote if (pcmk_is_set(pe__const_top_resource(rsc, false)->flags, pe_rsc_promotable) || (rsc->role > RSC_ROLE_UNPROMOTED)) { pcmk__new_ordering(rsc, pcmk__op_key(rsc->id, RSC_DEMOTE, 0), NULL, rsc, pcmk__op_key(rsc->id, RSC_STOP, 0), NULL, pe_order_promoted_implies_first, rsc->cluster); pcmk__new_ordering(rsc, pcmk__op_key(rsc->id, RSC_START, 0), NULL, rsc, pcmk__op_key(rsc->id, RSC_PROMOTE, 0), NULL, pe_order_runnable_left, rsc->cluster); } // Don't clear resource history if probing on same node pcmk__new_ordering(rsc, pcmk__op_key(rsc->id, CRM_OP_LRM_DELETE, 0), NULL, rsc, pcmk__op_key(rsc->id, RSC_STATUS, 0), NULL, pe_order_same_node|pe_order_then_cancels_first, rsc->cluster); // Certain checks need allowed nodes if (check_unfencing || check_utilization || (rsc->container != NULL)) { allowed_nodes = allowed_nodes_as_list(rsc); } if (check_unfencing) { g_list_foreach(allowed_nodes, pcmk__order_restart_vs_unfence, rsc); } if (check_utilization) { pcmk__create_utilization_constraints(rsc, allowed_nodes); } if (rsc->container != NULL) { pe_resource_t *remote_rsc = NULL; if (rsc->is_remote_node) { // rsc is the implicit remote connection for a guest or bundle node /* Guest resources are not allowed to run on Pacemaker Remote nodes, * to avoid nesting remotes. However, bundles are allowed. */ if (!pcmk_is_set(rsc->flags, pe_rsc_allow_remote_remotes)) { rsc_avoids_remote_nodes(rsc->container); } /* If someone cleans up a guest or bundle node's container, we will * likely schedule a (re-)probe of the container and recovery of the * connection. Order the connection stop after the container probe, * so that if we detect the container running, we will trigger a new * transition and avoid the unnecessary recovery. */ pcmk__order_resource_actions(rsc->container, RSC_STATUS, rsc, RSC_STOP, pe_order_optional); /* A user can specify that a resource must start on a Pacemaker Remote * node by explicitly configuring it with the container=NODENAME * meta-attribute. This is of questionable merit, since location * constraints can accomplish the same thing. But we support it, so here * we check whether a resource (that is not itself a remote connection) * has container set to a remote node or guest node resource. */ } else if (rsc->container->is_remote_node) { remote_rsc = rsc->container; } else { remote_rsc = pe__resource_contains_guest_node(rsc->cluster, rsc->container); } if (remote_rsc != NULL) { /* Force the resource on the Pacemaker Remote node instead of * colocating the resource with the container resource. */ for (GList *item = allowed_nodes; item; item = item->next) { pe_node_t *node = item->data; if (node->details->remote_rsc != remote_rsc) { node->weight = -INFINITY; } } } else { /* This resource is either a filler for a container that does NOT * represent a Pacemaker Remote node, or a Pacemaker Remote * connection resource for a guest node or bundle. */ int score; crm_trace("Order and colocate %s relative to its container %s", rsc->id, rsc->container->id); pcmk__new_ordering(rsc->container, pcmk__op_key(rsc->container->id, RSC_START, 0), NULL, rsc, pcmk__op_key(rsc->id, RSC_START, 0), NULL, pe_order_implies_then|pe_order_runnable_left, rsc->cluster); pcmk__new_ordering(rsc, pcmk__op_key(rsc->id, RSC_STOP, 0), NULL, rsc->container, pcmk__op_key(rsc->container->id, RSC_STOP, 0), NULL, pe_order_implies_first, rsc->cluster); if (pcmk_is_set(rsc->flags, pe_rsc_allow_remote_remotes)) { score = 10000; /* Highly preferred but not essential */ } else { score = INFINITY; /* Force them to run on the same host */ } pcmk__new_colocation("#resource-with-container", NULL, score, rsc, rsc->container, NULL, NULL, pcmk__coloc_influence); } } if (rsc->is_remote_node || pcmk_is_set(rsc->flags, pe_rsc_fence_device)) { /* Remote connections and fencing devices are not allowed to run on * Pacemaker Remote nodes */ rsc_avoids_remote_nodes(rsc); } g_list_free(allowed_nodes); } /*! * \internal * \brief Apply a colocation's score to node scores or resource priority * * Given a colocation constraint, apply its score to the dependent's * allowed node scores (if we are still placing resources) or priority (if * we are choosing promotable clone instance roles). * * \param[in,out] dependent Dependent resource in colocation * \param[in] primary Primary resource in colocation * \param[in] colocation Colocation constraint to apply * \param[in] for_dependent true if called on behalf of dependent */ void pcmk__primitive_apply_coloc_score(pe_resource_t *dependent, const pe_resource_t *primary, const pcmk__colocation_t *colocation, bool for_dependent) { enum pcmk__coloc_affects filter_results; CRM_ASSERT((dependent != NULL) && (primary != NULL) && (colocation != NULL)); if (for_dependent) { // Always process on behalf of primary resource primary->cmds->apply_coloc_score(dependent, primary, colocation, false); return; } filter_results = pcmk__colocation_affects(dependent, primary, colocation, false); pe_rsc_trace(dependent, "%s %s with %s (%s, score=%d, filter=%d)", ((colocation->score > 0)? "Colocating" : "Anti-colocating"), dependent->id, primary->id, colocation->id, colocation->score, filter_results); switch (filter_results) { case pcmk__coloc_affects_role: pcmk__apply_coloc_to_priority(dependent, primary, colocation); break; case pcmk__coloc_affects_location: pcmk__apply_coloc_to_scores(dependent, primary, colocation); break; default: // pcmk__coloc_affects_nothing return; } } /* Primitive implementation of * resource_alloc_functions_t:with_this_colocations() */ void pcmk__with_primitive_colocations(const pe_resource_t *rsc, const pe_resource_t *orig_rsc, GList **list) { CRM_ASSERT((rsc != NULL) && (rsc->variant == pe_native) && (list != NULL)); if (rsc == orig_rsc) { /* For the resource itself, add all of its own colocations and relevant * colocations from its parent (if any). */ pcmk__add_with_this_list(list, rsc->rsc_cons_lhs, orig_rsc); if (rsc->parent != NULL) { - rsc->parent->cmds->with_this_colocations(rsc->parent, rsc, list); + rsc->parent->cmds->with_this_colocations(rsc->parent, orig_rsc, list); } } else { // For an ancestor, add only explicitly configured constraints for (GList *iter = rsc->rsc_cons_lhs; iter != NULL; iter = iter->next) { pcmk__colocation_t *colocation = iter->data; if (pcmk_is_set(colocation->flags, pcmk__coloc_explicit)) { pcmk__add_with_this(list, colocation, orig_rsc); } } } } /* Primitive implementation of * resource_alloc_functions_t:this_with_colocations() */ void pcmk__primitive_with_colocations(const pe_resource_t *rsc, const pe_resource_t *orig_rsc, GList **list) { CRM_ASSERT((rsc != NULL) && (rsc->variant == pe_native) && (list != NULL)); if (rsc == orig_rsc) { /* For the resource itself, add all of its own colocations and relevant * colocations from its parent (if any). */ pcmk__add_this_with_list(list, rsc->rsc_cons, orig_rsc); if (rsc->parent != NULL) { - rsc->parent->cmds->this_with_colocations(rsc->parent, rsc, list); + rsc->parent->cmds->this_with_colocations(rsc->parent, orig_rsc, list); } } else { // For an ancestor, add only explicitly configured constraints for (GList *iter = rsc->rsc_cons; iter != NULL; iter = iter->next) { pcmk__colocation_t *colocation = iter->data; if (pcmk_is_set(colocation->flags, pcmk__coloc_explicit)) { pcmk__add_this_with(list, colocation, orig_rsc); } } } } /*! * \internal * \brief Return action flags for a given primitive resource action * * \param[in,out] action Action to get flags for * \param[in] node If not NULL, limit effects to this node (ignored) * * \return Flags appropriate to \p action on \p node */ uint32_t pcmk__primitive_action_flags(pe_action_t *action, const pe_node_t *node) { CRM_ASSERT(action != NULL); return (uint32_t) action->flags; } /*! * \internal * \brief Check whether a node is a multiply active resource's expected node * * \param[in] rsc Resource to check * \param[in] node Node to check * * \return true if \p rsc is multiply active with multiple-active set to * stop_unexpected, and \p node is the node where it will remain active * \note This assumes that the resource's next role cannot be changed to stopped * after this is called, which should be reasonable if status has already * been unpacked and resources have been assigned to nodes. */ static bool is_expected_node(const pe_resource_t *rsc, const pe_node_t *node) { return pcmk_all_flags_set(rsc->flags, pe_rsc_stop_unexpected|pe_rsc_restarting) && (rsc->next_role > RSC_ROLE_STOPPED) && pe__same_node(rsc->allocated_to, node); } /*! * \internal * \brief Schedule actions needed to stop a resource wherever it is active * * \param[in,out] rsc Resource being stopped * \param[in] node Node where resource is being stopped (ignored) * \param[in] optional Whether actions should be optional */ static void stop_resource(pe_resource_t *rsc, pe_node_t *node, bool optional) { for (GList *iter = rsc->running_on; iter != NULL; iter = iter->next) { pe_node_t *current = (pe_node_t *) iter->data; pe_action_t *stop = NULL; if (is_expected_node(rsc, current)) { /* We are scheduling restart actions for a multiply active resource * with multiple-active=stop_unexpected, and this is where it should * not be stopped. */ pe_rsc_trace(rsc, "Skipping stop of multiply active resource %s " "on expected node %s", rsc->id, pe__node_name(current)); continue; } if (rsc->partial_migration_target != NULL) { // Continue migration if node originally was and remains target if (pe__same_node(current, rsc->partial_migration_target) && pe__same_node(current, rsc->allocated_to)) { pe_rsc_trace(rsc, "Skipping stop of %s on %s " "because partial migration there will continue", rsc->id, pe__node_name(current)); continue; } else { pe_rsc_trace(rsc, "Forcing stop of %s on %s " "because migration target changed", rsc->id, pe__node_name(current)); optional = false; } } pe_rsc_trace(rsc, "Scheduling stop of %s on %s", rsc->id, pe__node_name(current)); stop = stop_action(rsc, current, optional); if (rsc->allocated_to == NULL) { pe_action_set_reason(stop, "node availability", true); } else if (pcmk_all_flags_set(rsc->flags, pe_rsc_restarting |pe_rsc_stop_unexpected)) { /* We are stopping a multiply active resource on a node that is * not its expected node, and we are still scheduling restart * actions, so the stop is for being multiply active. */ pe_action_set_reason(stop, "being multiply active", true); } if (!pcmk_is_set(rsc->flags, pe_rsc_managed)) { pe__clear_action_flags(stop, pe_action_runnable); } if (pcmk_is_set(rsc->cluster->flags, pe_flag_remove_after_stop)) { pcmk__schedule_cleanup(rsc, current, optional); } if (pcmk_is_set(rsc->flags, pe_rsc_needs_unfencing)) { pe_action_t *unfence = pe_fence_op(current, "on", true, NULL, false, rsc->cluster); order_actions(stop, unfence, pe_order_implies_first); if (!pcmk__node_unfenced(current)) { pe_proc_err("Stopping %s until %s can be unfenced", rsc->id, pe__node_name(current)); } } } } /*! * \internal * \brief Schedule actions needed to start a resource on a node * * \param[in,out] rsc Resource being started * \param[in,out] node Node where resource should be started * \param[in] optional Whether actions should be optional */ static void start_resource(pe_resource_t *rsc, pe_node_t *node, bool optional) { pe_action_t *start = NULL; CRM_ASSERT(node != NULL); pe_rsc_trace(rsc, "Scheduling %s start of %s on %s (score %d)", (optional? "optional" : "required"), rsc->id, pe__node_name(node), node->weight); start = start_action(rsc, node, TRUE); pcmk__order_vs_unfence(rsc, node, start, pe_order_implies_then); if (pcmk_is_set(start->flags, pe_action_runnable) && !optional) { pe__clear_action_flags(start, pe_action_optional); } if (is_expected_node(rsc, node)) { /* This could be a problem if the start becomes necessary for other * reasons later. */ pe_rsc_trace(rsc, "Start of multiply active resouce %s " "on expected node %s will be a pseudo-action", rsc->id, pe__node_name(node)); pe__set_action_flags(start, pe_action_pseudo); } } /*! * \internal * \brief Schedule actions needed to promote a resource on a node * * \param[in,out] rsc Resource being promoted * \param[in] node Node where resource should be promoted * \param[in] optional Whether actions should be optional */ static void promote_resource(pe_resource_t *rsc, pe_node_t *node, bool optional) { GList *iter = NULL; GList *action_list = NULL; bool runnable = true; CRM_ASSERT(node != NULL); // Any start must be runnable for promotion to be runnable action_list = pe__resource_actions(rsc, node, RSC_START, true); for (iter = action_list; iter != NULL; iter = iter->next) { pe_action_t *start = (pe_action_t *) iter->data; if (!pcmk_is_set(start->flags, pe_action_runnable)) { runnable = false; } } g_list_free(action_list); if (runnable) { pe_action_t *promote = promote_action(rsc, node, optional); pe_rsc_trace(rsc, "Scheduling %s promotion of %s on %s", (optional? "optional" : "required"), rsc->id, pe__node_name(node)); if (is_expected_node(rsc, node)) { /* This could be a problem if the promote becomes necessary for * other reasons later. */ pe_rsc_trace(rsc, "Promotion of multiply active resouce %s " "on expected node %s will be a pseudo-action", rsc->id, pe__node_name(node)); pe__set_action_flags(promote, pe_action_pseudo); } } else { pe_rsc_trace(rsc, "Not promoting %s on %s: start unrunnable", rsc->id, pe__node_name(node)); action_list = pe__resource_actions(rsc, node, RSC_PROMOTE, true); for (iter = action_list; iter != NULL; iter = iter->next) { pe_action_t *promote = (pe_action_t *) iter->data; pe__clear_action_flags(promote, pe_action_runnable); } g_list_free(action_list); } } /*! * \internal * \brief Schedule actions needed to demote a resource wherever it is active * * \param[in,out] rsc Resource being demoted * \param[in] node Node where resource should be demoted (ignored) * \param[in] optional Whether actions should be optional */ static void demote_resource(pe_resource_t *rsc, pe_node_t *node, bool optional) { /* Since this will only be called for a primitive (possibly as an instance * of a collective resource), the resource is multiply active if it is * running on more than one node, so we want to demote on all of them as * part of recovery, regardless of which one is the desired node. */ for (GList *iter = rsc->running_on; iter != NULL; iter = iter->next) { pe_node_t *current = (pe_node_t *) iter->data; if (is_expected_node(rsc, current)) { pe_rsc_trace(rsc, "Skipping demote of multiply active resource %s " "on expected node %s", rsc->id, pe__node_name(current)); } else { pe_rsc_trace(rsc, "Scheduling %s demotion of %s on %s", (optional? "optional" : "required"), rsc->id, pe__node_name(current)); demote_action(rsc, current, optional); } } } static void assert_role_error(pe_resource_t *rsc, pe_node_t *node, bool optional) { CRM_ASSERT(false); } /*! * \internal * \brief Schedule cleanup of a resource * * \param[in,out] rsc Resource to clean up * \param[in] node Node to clean up on * \param[in] optional Whether clean-up should be optional */ void pcmk__schedule_cleanup(pe_resource_t *rsc, const pe_node_t *node, bool optional) { /* If the cleanup is required, its orderings are optional, because they're * relevant only if both actions are required. Conversely, if the cleanup is * optional, the orderings make the then action required if the first action * becomes required. */ uint32_t flag = optional? pe_order_implies_then : pe_order_optional; CRM_CHECK((rsc != NULL) && (node != NULL), return); if (pcmk_is_set(rsc->flags, pe_rsc_failed)) { pe_rsc_trace(rsc, "Skipping clean-up of %s on %s: resource failed", rsc->id, pe__node_name(node)); return; } if (node->details->unclean || !node->details->online) { pe_rsc_trace(rsc, "Skipping clean-up of %s on %s: node unavailable", rsc->id, pe__node_name(node)); return; } crm_notice("Scheduling clean-up of %s on %s", rsc->id, pe__node_name(node)); delete_action(rsc, node, optional); // stop -> clean-up -> start pcmk__order_resource_actions(rsc, RSC_STOP, rsc, RSC_DELETE, flag); pcmk__order_resource_actions(rsc, RSC_DELETE, rsc, RSC_START, flag); } /*! * \internal * \brief Add primitive meta-attributes relevant to graph actions to XML * * \param[in] rsc Primitive resource whose meta-attributes should be added * \param[in,out] xml Transition graph action attributes XML to add to */ void pcmk__primitive_add_graph_meta(const pe_resource_t *rsc, xmlNode *xml) { char *name = NULL; char *value = NULL; const pe_resource_t *parent = NULL; CRM_ASSERT((rsc != NULL) && (rsc->variant == pe_native) && (xml != NULL)); /* Clone instance numbers get set internally as meta-attributes, and are * needed in the transition graph (for example, to tell unique clone * instances apart). */ value = g_hash_table_lookup(rsc->meta, XML_RSC_ATTR_INCARNATION); if (value != NULL) { name = crm_meta_name(XML_RSC_ATTR_INCARNATION); crm_xml_add(xml, name, value); free(name); } // Not sure if this one is really needed ... value = g_hash_table_lookup(rsc->meta, XML_RSC_ATTR_REMOTE_NODE); if (value != NULL) { name = crm_meta_name(XML_RSC_ATTR_REMOTE_NODE); crm_xml_add(xml, name, value); free(name); } /* The container meta-attribute can be set on the primitive itself or one of * its parents (for example, a group inside a container resource), so check * them all, and keep the highest one found. */ for (parent = rsc; parent != NULL; parent = parent->parent) { if (parent->container != NULL) { crm_xml_add(xml, CRM_META "_" XML_RSC_ATTR_CONTAINER, parent->container->id); } } /* Bundle replica children will get their external-ip set internally as a * meta-attribute. The graph action needs it, but under a different naming * convention than other meta-attributes. */ value = g_hash_table_lookup(rsc->meta, "external-ip"); if (value != NULL) { crm_xml_add(xml, "pcmk_external_ip", value); } } // Primitive implementation of resource_alloc_functions_t:add_utilization() void pcmk__primitive_add_utilization(const pe_resource_t *rsc, const pe_resource_t *orig_rsc, GList *all_rscs, GHashTable *utilization) { CRM_ASSERT((rsc != NULL) && (rsc->variant == pe_native) && (orig_rsc != NULL) && (utilization != NULL)); if (!pcmk_is_set(rsc->flags, pe_rsc_provisional)) { return; } pe_rsc_trace(orig_rsc, "%s: Adding primitive %s as colocated utilization", orig_rsc->id, rsc->id); pcmk__release_node_capacity(utilization, rsc); } /*! * \internal * \brief Get epoch time of node's shutdown attribute (or now if none) * * \param[in,out] node Node to check * * \return Epoch time corresponding to shutdown attribute if set or now if not */ static time_t shutdown_time(pe_node_t *node) { const char *shutdown = pe_node_attribute_raw(node, XML_CIB_ATTR_SHUTDOWN); time_t result = 0; if (shutdown != NULL) { long long result_ll; if (pcmk__scan_ll(shutdown, &result_ll, 0LL) == pcmk_rc_ok) { result = (time_t) result_ll; } } return (result == 0)? get_effective_time(node->details->data_set) : result; } /*! * \internal * \brief Ban a resource from a node if it's not locked to the node * * \param[in] data Node to check * \param[in,out] user_data Resource to check */ static void ban_if_not_locked(gpointer data, gpointer user_data) { const pe_node_t *node = (const pe_node_t *) data; pe_resource_t *rsc = (pe_resource_t *) user_data; if (strcmp(node->details->uname, rsc->lock_node->details->uname) != 0) { resource_location(rsc, node, -CRM_SCORE_INFINITY, XML_CONFIG_ATTR_SHUTDOWN_LOCK, rsc->cluster); } } // Primitive implementation of resource_alloc_functions_t:shutdown_lock() void pcmk__primitive_shutdown_lock(pe_resource_t *rsc) { const char *class = NULL; CRM_ASSERT((rsc != NULL) && (rsc->variant == pe_native)); class = crm_element_value(rsc->xml, XML_AGENT_ATTR_CLASS); // Fence devices and remote connections can't be locked if (pcmk__str_eq(class, PCMK_RESOURCE_CLASS_STONITH, pcmk__str_null_matches) || pe__resource_is_remote_conn(rsc, rsc->cluster)) { return; } if (rsc->lock_node != NULL) { // The lock was obtained from resource history if (rsc->running_on != NULL) { /* The resource was started elsewhere even though it is now * considered locked. This shouldn't be possible, but as a * failsafe, we don't want to disturb the resource now. */ pe_rsc_info(rsc, "Cancelling shutdown lock because %s is already active", rsc->id); pe__clear_resource_history(rsc, rsc->lock_node, rsc->cluster); rsc->lock_node = NULL; rsc->lock_time = 0; } // Only a resource active on exactly one node can be locked } else if (pcmk__list_of_1(rsc->running_on)) { pe_node_t *node = rsc->running_on->data; if (node->details->shutdown) { if (node->details->unclean) { pe_rsc_debug(rsc, "Not locking %s to unclean %s for shutdown", rsc->id, pe__node_name(node)); } else { rsc->lock_node = node; rsc->lock_time = shutdown_time(node); } } } if (rsc->lock_node == NULL) { // No lock needed return; } if (rsc->cluster->shutdown_lock > 0) { time_t lock_expiration = rsc->lock_time + rsc->cluster->shutdown_lock; pe_rsc_info(rsc, "Locking %s to %s due to shutdown (expires @%lld)", rsc->id, pe__node_name(rsc->lock_node), (long long) lock_expiration); pe__update_recheck_time(++lock_expiration, rsc->cluster); } else { pe_rsc_info(rsc, "Locking %s to %s due to shutdown", rsc->id, pe__node_name(rsc->lock_node)); } // If resource is locked to one node, ban it from all other nodes g_list_foreach(rsc->cluster->nodes, ban_if_not_locked, rsc); } diff --git a/lib/pacemaker/pcmk_sched_promotable.c b/lib/pacemaker/pcmk_sched_promotable.c index a4a27b4edf..afdcdb5880 100644 --- a/lib/pacemaker/pcmk_sched_promotable.c +++ b/lib/pacemaker/pcmk_sched_promotable.c @@ -1,1288 +1,1292 @@ /* * Copyright 2004-2023 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU General Public License version 2 * or later (GPLv2+) WITHOUT ANY WARRANTY. */ #include #include #include #include "libpacemaker_private.h" /*! * \internal * \brief Add implicit promotion ordering for a promotable instance * * \param[in,out] clone Clone resource * \param[in,out] child Instance of \p clone being ordered * \param[in,out] last Previous instance ordered (NULL if \p child is first) */ static void order_instance_promotion(pe_resource_t *clone, pe_resource_t *child, pe_resource_t *last) { // "Promote clone" -> promote instance -> "clone promoted" pcmk__order_resource_actions(clone, RSC_PROMOTE, child, RSC_PROMOTE, pe_order_optional); pcmk__order_resource_actions(child, RSC_PROMOTE, clone, RSC_PROMOTED, pe_order_optional); // If clone is ordered, order this instance relative to last if ((last != NULL) && pe__clone_is_ordered(clone)) { pcmk__order_resource_actions(last, RSC_PROMOTE, child, RSC_PROMOTE, pe_order_optional); } } /*! * \internal * \brief Add implicit demotion ordering for a promotable instance * * \param[in,out] clone Clone resource * \param[in,out] child Instance of \p clone being ordered * \param[in] last Previous instance ordered (NULL if \p child is first) */ static void order_instance_demotion(pe_resource_t *clone, pe_resource_t *child, pe_resource_t *last) { // "Demote clone" -> demote instance -> "clone demoted" pcmk__order_resource_actions(clone, RSC_DEMOTE, child, RSC_DEMOTE, pe_order_implies_first_printed); pcmk__order_resource_actions(child, RSC_DEMOTE, clone, RSC_DEMOTED, pe_order_implies_then_printed); // If clone is ordered, order this instance relative to last if ((last != NULL) && pe__clone_is_ordered(clone)) { pcmk__order_resource_actions(child, RSC_DEMOTE, last, RSC_DEMOTE, pe_order_optional); } } /*! * \internal * \brief Check whether an instance will be promoted or demoted * * \param[in] rsc Instance to check * \param[out] demoting If \p rsc will be demoted, this will be set to true * \param[out] promoting If \p rsc will be promoted, this will be set to true */ static void check_for_role_change(const pe_resource_t *rsc, bool *demoting, bool *promoting) { const GList *iter = NULL; // If this is a cloned group, check group members recursively if (rsc->children != NULL) { for (iter = rsc->children; iter != NULL; iter = iter->next) { check_for_role_change((const pe_resource_t *) iter->data, demoting, promoting); } return; } for (iter = rsc->actions; iter != NULL; iter = iter->next) { const pe_action_t *action = (const pe_action_t *) iter->data; if (*promoting && *demoting) { return; } else if (pcmk_is_set(action->flags, pe_action_optional)) { continue; } else if (pcmk__str_eq(RSC_DEMOTE, action->task, pcmk__str_none)) { *demoting = true; } else if (pcmk__str_eq(RSC_PROMOTE, action->task, pcmk__str_none)) { *promoting = true; } } } /*! * \internal * \brief Add promoted-role location constraint scores to an instance's priority * * Adjust a promotable clone instance's promotion priority by the scores of any * location constraints in a list that are both limited to the promoted role and * for the node where the instance will be placed. * * \param[in,out] child Promotable clone instance * \param[in] location_constraints List of location constraints to apply * \param[in] chosen Node where \p child will be placed */ static void apply_promoted_locations(pe_resource_t *child, const GList *location_constraints, const pe_node_t *chosen) { for (const GList *iter = location_constraints; iter; iter = iter->next) { const pe__location_t *location = iter->data; const pe_node_t *constraint_node = NULL; if (location->role_filter == RSC_ROLE_PROMOTED) { constraint_node = pe_find_node_id(location->node_list_rh, chosen->details->id); } if (constraint_node != NULL) { int new_priority = pcmk__add_scores(child->priority, constraint_node->weight); pe_rsc_trace(child, "Applying location %s to %s promotion priority on %s: " "%s + %s = %s", location->id, child->id, pe__node_name(constraint_node), pcmk_readable_score(child->priority), pcmk_readable_score(constraint_node->weight), pcmk_readable_score(new_priority)); child->priority = new_priority; } } } /*! * \internal * \brief Get the node that an instance will be promoted on * * \param[in] rsc Promotable clone instance to check * * \return Node that \p rsc will be promoted on, or NULL if none */ static pe_node_t * node_to_be_promoted_on(const pe_resource_t *rsc) { pe_node_t *node = NULL; pe_node_t *local_node = NULL; const pe_resource_t *parent = NULL; // If this is a cloned group, bail if any group member can't be promoted for (GList *iter = rsc->children; iter != NULL; iter = iter->next) { pe_resource_t *child = (pe_resource_t *) iter->data; if (node_to_be_promoted_on(child) == NULL) { pe_rsc_trace(rsc, "%s can't be promoted because member %s can't", rsc->id, child->id); return NULL; } } node = rsc->fns->location(rsc, NULL, FALSE); if (node == NULL) { pe_rsc_trace(rsc, "%s can't be promoted because it won't be active", rsc->id); return NULL; } else if (!pcmk_is_set(rsc->flags, pe_rsc_managed)) { if (rsc->fns->state(rsc, TRUE) == RSC_ROLE_PROMOTED) { crm_notice("Unmanaged instance %s will be left promoted on %s", rsc->id, pe__node_name(node)); } else { pe_rsc_trace(rsc, "%s can't be promoted because it is unmanaged", rsc->id); return NULL; } } else if (rsc->priority < 0) { pe_rsc_trace(rsc, "%s can't be promoted because its promotion priority %d " "is negative", rsc->id, rsc->priority); return NULL; } else if (!pcmk__node_available(node, false, true)) { pe_rsc_trace(rsc, "%s can't be promoted because %s can't run resources", rsc->id, pe__node_name(node)); return NULL; } parent = pe__const_top_resource(rsc, false); local_node = pe_hash_table_lookup(parent->allowed_nodes, node->details->id); if (local_node == NULL) { /* It should not be possible for the scheduler to have assigned the * instance to a node where its parent is not allowed, but it's good to * have a fail-safe. */ if (pcmk_is_set(rsc->flags, pe_rsc_managed)) { crm_warn("%s can't be promoted because %s is not allowed on %s " "(scheduler bug?)", rsc->id, parent->id, pe__node_name(node)); } // else the instance is unmanaged and already promoted return NULL; } else if ((local_node->count >= pe__clone_promoted_node_max(parent)) && pcmk_is_set(rsc->flags, pe_rsc_managed)) { pe_rsc_trace(rsc, "%s can't be promoted because %s has " "maximum promoted instances already", rsc->id, pe__node_name(node)); return NULL; } return local_node; } /*! * \internal * \brief Compare two promotable clone instances by promotion priority * * \param[in] a First instance to compare * \param[in] b Second instance to compare * * \return A negative number if \p a has higher promotion priority, * a positive number if \p b has higher promotion priority, * or 0 if promotion priorities are equal */ static gint cmp_promotable_instance(gconstpointer a, gconstpointer b) { const pe_resource_t *rsc1 = (const pe_resource_t *) a; const pe_resource_t *rsc2 = (const pe_resource_t *) b; enum rsc_role_e role1 = RSC_ROLE_UNKNOWN; enum rsc_role_e role2 = RSC_ROLE_UNKNOWN; CRM_ASSERT((rsc1 != NULL) && (rsc2 != NULL)); // Check sort index set by pcmk__set_instance_roles() if (rsc1->sort_index > rsc2->sort_index) { pe_rsc_trace(rsc1, "%s has higher promotion priority than %s " "(sort index %d > %d)", rsc1->id, rsc2->id, rsc1->sort_index, rsc2->sort_index); return -1; } else if (rsc1->sort_index < rsc2->sort_index) { pe_rsc_trace(rsc1, "%s has lower promotion priority than %s " "(sort index %d < %d)", rsc1->id, rsc2->id, rsc1->sort_index, rsc2->sort_index); return 1; } // If those are the same, prefer instance whose current role is higher role1 = rsc1->fns->state(rsc1, TRUE); role2 = rsc2->fns->state(rsc2, TRUE); if (role1 > role2) { pe_rsc_trace(rsc1, "%s has higher promotion priority than %s " "(higher current role)", rsc1->id, rsc2->id); return -1; } else if (role1 < role2) { pe_rsc_trace(rsc1, "%s has lower promotion priority than %s " "(lower current role)", rsc1->id, rsc2->id); return 1; } // Finally, do normal clone instance sorting return pcmk__cmp_instance(a, b); } /*! * \internal * \brief Add a promotable clone instance's sort index to its node's score * * Add a promotable clone instance's sort index (which sums its promotion * preferences and scores of relevant location constraints for the promoted * role) to the node score of the instance's assigned node. * * \param[in] data Promotable clone instance * \param[in,out] user_data Clone parent of \p data */ static void add_sort_index_to_node_score(gpointer data, gpointer user_data) { const pe_resource_t *child = (const pe_resource_t *) data; pe_resource_t *clone = (pe_resource_t *) user_data; pe_node_t *node = NULL; const pe_node_t *chosen = NULL; if (child->sort_index < 0) { pe_rsc_trace(clone, "Not adding sort index of %s: negative", child->id); return; } chosen = child->fns->location(child, NULL, FALSE); if (chosen == NULL) { pe_rsc_trace(clone, "Not adding sort index of %s: inactive", child->id); return; } node = (pe_node_t *) pe_hash_table_lookup(clone->allowed_nodes, chosen->details->id); CRM_ASSERT(node != NULL); node->weight = pcmk__add_scores(child->sort_index, node->weight); pe_rsc_trace(clone, "Added cumulative priority of %s (%s) to score on %s (now %s)", child->id, pcmk_readable_score(child->sort_index), pe__node_name(node), pcmk_readable_score(node->weight)); } /*! * \internal * \brief Apply colocation to dependent's node scores if for promoted role * * \param[in,out] data Colocation constraint to apply * \param[in,out] user_data Promotable clone that is constraint's dependent */ static void apply_coloc_to_dependent(gpointer data, gpointer user_data) { pcmk__colocation_t *colocation = data; pe_resource_t *clone = user_data; pe_resource_t *primary = colocation->primary; uint32_t flags = pcmk__coloc_select_default; float factor = colocation->score / (float) INFINITY; if (colocation->dependent_role != RSC_ROLE_PROMOTED) { return; } if (colocation->score < INFINITY) { flags = pcmk__coloc_select_active; } pe_rsc_trace(clone, "Applying colocation %s (promoted %s with %s) @%s", colocation->id, colocation->dependent->id, colocation->primary->id, pcmk_readable_score(colocation->score)); primary->cmds->add_colocated_node_scores(primary, clone, clone->id, &clone->allowed_nodes, colocation, factor, flags); } /*! * \internal * \brief Apply colocation to primary's node scores if for promoted role * * \param[in,out] data Colocation constraint to apply * \param[in,out] user_data Promotable clone that is constraint's primary */ static void apply_coloc_to_primary(gpointer data, gpointer user_data) { pcmk__colocation_t *colocation = data; pe_resource_t *clone = user_data; pe_resource_t *dependent = colocation->dependent; const float factor = colocation->score / (float) INFINITY; const uint32_t flags = pcmk__coloc_select_active |pcmk__coloc_select_nonnegative; if ((colocation->primary_role != RSC_ROLE_PROMOTED) || !pcmk__colocation_has_influence(colocation, NULL)) { return; } pe_rsc_trace(clone, "Applying colocation %s (%s with promoted %s) @%s", colocation->id, colocation->dependent->id, colocation->primary->id, pcmk_readable_score(colocation->score)); dependent->cmds->add_colocated_node_scores(dependent, clone, clone->id, &clone->allowed_nodes, colocation, factor, flags); } /*! * \internal * \brief Set clone instance's sort index to its node's score * * \param[in,out] data Promotable clone instance * \param[in] user_data Parent clone of \p data */ static void set_sort_index_to_node_score(gpointer data, gpointer user_data) { pe_resource_t *child = (pe_resource_t *) data; const pe_resource_t *clone = (const pe_resource_t *) user_data; pe_node_t *chosen = child->fns->location(child, NULL, FALSE); if (!pcmk_is_set(child->flags, pe_rsc_managed) && (child->next_role == RSC_ROLE_PROMOTED)) { child->sort_index = INFINITY; pe_rsc_trace(clone, "Final sort index for %s is INFINITY (unmanaged promoted)", child->id); } else if ((chosen == NULL) || (child->sort_index < 0)) { pe_rsc_trace(clone, "Final sort index for %s is %d (ignoring node score)", child->id, child->sort_index); } else { const pe_node_t *node = NULL; node = pe_hash_table_lookup(clone->allowed_nodes, chosen->details->id); CRM_ASSERT(node != NULL); child->sort_index = node->weight; pe_rsc_trace(clone, "Adding scores for %s: final sort index for %s is %d", clone->id, child->id, child->sort_index); } } /*! * \internal * \brief Sort a promotable clone's instances by descending promotion priority * * \param[in,out] clone Promotable clone to sort */ static void sort_promotable_instances(pe_resource_t *clone) { + GList *colocations = NULL; + if (pe__set_clone_flag(clone, pe__clone_promotion_constrained) == pcmk_rc_already) { return; } pe__set_resource_flags(clone, pe_rsc_merging); for (GList *iter = clone->children; iter != NULL; iter = iter->next) { pe_resource_t *child = (pe_resource_t *) iter->data; pe_rsc_trace(clone, "Adding scores for %s: initial sort index for %s is %d", clone->id, child->id, child->sort_index); } pe__show_node_scores(true, clone, "Before", clone->allowed_nodes, clone->cluster); - /* Because the this_with_colocations() and with_this_colocations() methods - * boil down to copies of rsc_cons and rsc_cons_lhs for clones, we can use - * those here directly for efficiency. - */ g_list_foreach(clone->children, add_sort_index_to_node_score, clone); - g_list_foreach(clone->rsc_cons, apply_coloc_to_dependent, clone); - g_list_foreach(clone->rsc_cons_lhs, apply_coloc_to_primary, clone); + + colocations = pcmk__this_with_colocations(clone); + g_list_foreach(colocations, apply_coloc_to_dependent, clone); + g_list_free(colocations); + + colocations = pcmk__with_this_colocations(clone); + g_list_foreach(colocations, apply_coloc_to_primary, clone); + g_list_free(colocations); // Ban resource from all nodes if it needs a ticket but doesn't have it pcmk__require_promotion_tickets(clone); pe__show_node_scores(true, clone, "After", clone->allowed_nodes, clone->cluster); // Reset sort indexes to final node scores g_list_foreach(clone->children, set_sort_index_to_node_score, clone); // Finally, sort instances in descending order of promotion priority clone->children = g_list_sort(clone->children, cmp_promotable_instance); pe__clear_resource_flags(clone, pe_rsc_merging); } /*! * \internal * \brief Find the active instance (if any) of an anonymous clone on a node * * \param[in] clone Anonymous clone to check * \param[in] id Instance ID (without instance number) to check * \param[in] node Node to check * * \return */ static pe_resource_t * find_active_anon_instance(const pe_resource_t *clone, const char *id, const pe_node_t *node) { for (GList *iter = clone->children; iter; iter = iter->next) { pe_resource_t *child = iter->data; pe_resource_t *active = NULL; // Use ->find_rsc() in case this is a cloned group active = clone->fns->find_rsc(child, id, node, pe_find_clone|pe_find_current); if (active != NULL) { return active; } } return NULL; } /* * \brief Check whether an anonymous clone instance is known on a node * * \param[in] clone Anonymous clone to check * \param[in] id Instance ID (without instance number) to check * \param[in] node Node to check * * \return true if \p id instance of \p clone is known on \p node, * otherwise false */ static bool anonymous_known_on(const pe_resource_t *clone, const char *id, const pe_node_t *node) { for (GList *iter = clone->children; iter; iter = iter->next) { pe_resource_t *child = iter->data; /* Use ->find_rsc() because this might be a cloned group, and knowing * that other members of the group are known here implies nothing. */ child = clone->fns->find_rsc(child, id, NULL, pe_find_clone); CRM_LOG_ASSERT(child != NULL); if (child != NULL) { if (g_hash_table_lookup(child->known_on, node->details->id)) { return true; } } } return false; } /*! * \internal * \brief Check whether a node is allowed to run a resource * * \param[in] rsc Resource to check * \param[in] node Node to check * * \return true if \p node is allowed to run \p rsc, otherwise false */ static bool is_allowed(const pe_resource_t *rsc, const pe_node_t *node) { pe_node_t *allowed = pe_hash_table_lookup(rsc->allowed_nodes, node->details->id); return (allowed != NULL) && (allowed->weight >= 0); } /*! * \brief Check whether a clone instance's promotion score should be considered * * \param[in] rsc Promotable clone instance to check * \param[in] node Node where score would be applied * * \return true if \p rsc's promotion score should be considered on \p node, * otherwise false */ static bool promotion_score_applies(const pe_resource_t *rsc, const pe_node_t *node) { char *id = clone_strip(rsc->id); const pe_resource_t *parent = pe__const_top_resource(rsc, false); pe_resource_t *active = NULL; const char *reason = "allowed"; // Some checks apply only to anonymous clone instances if (!pcmk_is_set(rsc->flags, pe_rsc_unique)) { // If instance is active on the node, its score definitely applies active = find_active_anon_instance(parent, id, node); if (active == rsc) { reason = "active"; goto check_allowed; } /* If *no* instance is active on this node, this instance's score will * count if it has been probed on this node. */ if ((active == NULL) && anonymous_known_on(parent, id, node)) { reason = "probed"; goto check_allowed; } } /* If this clone's status is unknown on *all* nodes (e.g. cluster startup), * take all instances' scores into account, to make sure we use any * permanent promotion scores. */ if ((rsc->running_on == NULL) && (g_hash_table_size(rsc->known_on) == 0)) { reason = "none probed"; goto check_allowed; } /* Otherwise, we've probed and/or started the resource *somewhere*, so * consider promotion scores on nodes where we know the status. */ if ((pe_hash_table_lookup(rsc->known_on, node->details->id) != NULL) || (pe_find_node_id(rsc->running_on, node->details->id) != NULL)) { reason = "known"; } else { pe_rsc_trace(rsc, "Ignoring %s promotion score (for %s) on %s: not probed", rsc->id, id, pe__node_name(node)); free(id); return false; } check_allowed: if (is_allowed(rsc, node)) { pe_rsc_trace(rsc, "Counting %s promotion score (for %s) on %s: %s", rsc->id, id, pe__node_name(node), reason); free(id); return true; } pe_rsc_trace(rsc, "Ignoring %s promotion score (for %s) on %s: not allowed", rsc->id, id, pe__node_name(node)); free(id); return false; } /*! * \internal * \brief Get the value of a promotion score node attribute * * \param[in] rsc Promotable clone instance to get promotion score for * \param[in] node Node to get promotion score for * \param[in] name Resource name to use in promotion score attribute name * * \return Value of promotion score node attribute for \p rsc on \p node */ static const char * promotion_attr_value(const pe_resource_t *rsc, const pe_node_t *node, const char *name) { char *attr_name = NULL; const char *attr_value = NULL; enum pe__rsc_node node_type = pe__rsc_node_assigned; if (pcmk_is_set(rsc->flags, pe_rsc_provisional)) { // Not assigned yet node_type = pe__rsc_node_current; } attr_name = pcmk_promotion_score_name(name); attr_value = pe__node_attribute_calculated(node, attr_name, rsc, node_type, false); free(attr_name); return attr_value; } /*! * \internal * \brief Get the promotion score for a clone instance on a node * * \param[in] rsc Promotable clone instance to get score for * \param[in] node Node to get score for * \param[out] is_default If non-NULL, will be set true if no score available * * \return Promotion score for \p rsc on \p node (or 0 if none) */ static int promotion_score(const pe_resource_t *rsc, const pe_node_t *node, bool *is_default) { char *name = NULL; const char *attr_value = NULL; if (is_default != NULL) { *is_default = true; } CRM_CHECK((rsc != NULL) && (node != NULL), return 0); /* If this is an instance of a cloned group, the promotion score is the sum * of all members' promotion scores. */ if (rsc->children != NULL) { int score = 0; for (const GList *iter = rsc->children; iter != NULL; iter = iter->next) { const pe_resource_t *child = (const pe_resource_t *) iter->data; bool child_default = false; int child_score = promotion_score(child, node, &child_default); if (!child_default && (is_default != NULL)) { *is_default = false; } score += child_score; } return score; } if (!promotion_score_applies(rsc, node)) { return 0; } /* For the promotion score attribute name, use the name the resource is * known as in resource history, since that's what crm_attribute --promotion * would have used. */ name = (rsc->clone_name == NULL)? rsc->id : rsc->clone_name; attr_value = promotion_attr_value(rsc, node, name); if (attr_value != NULL) { pe_rsc_trace(rsc, "Promotion score for %s on %s = %s", name, pe__node_name(node), pcmk__s(attr_value, "(unset)")); } else if (!pcmk_is_set(rsc->flags, pe_rsc_unique)) { /* If we don't have any resource history yet, we won't have clone_name. * In that case, for anonymous clones, try the resource name without * any instance number. */ name = clone_strip(rsc->id); if (strcmp(rsc->id, name) != 0) { attr_value = promotion_attr_value(rsc, node, name); pe_rsc_trace(rsc, "Promotion score for %s on %s (for %s) = %s", name, pe__node_name(node), rsc->id, pcmk__s(attr_value, "(unset)")); } free(name); } if (attr_value == NULL) { return 0; } if (is_default != NULL) { *is_default = false; } return char2score(attr_value); } /*! * \internal * \brief Include promotion scores in instances' node scores and priorities * * \param[in,out] rsc Promotable clone resource to update */ void pcmk__add_promotion_scores(pe_resource_t *rsc) { if (pe__set_clone_flag(rsc, pe__clone_promotion_added) == pcmk_rc_already) { return; } for (GList *iter = rsc->children; iter != NULL; iter = iter->next) { pe_resource_t *child_rsc = (pe_resource_t *) iter->data; GHashTableIter iter; pe_node_t *node = NULL; int score, new_score; g_hash_table_iter_init(&iter, child_rsc->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (void **) &node)) { if (!pcmk__node_available(node, false, false)) { /* This node will never be promoted, so don't apply the * promotion score, as that may lead to clone shuffling. */ continue; } score = promotion_score(child_rsc, node, NULL); if (score > 0) { new_score = pcmk__add_scores(node->weight, score); if (new_score != node->weight) { // Could remain INFINITY node->weight = new_score; pe_rsc_trace(rsc, "Added %s promotion priority (%s) to score " "on %s (now %s)", child_rsc->id, pcmk_readable_score(score), pe__node_name(node), pcmk_readable_score(new_score)); } } if (score > child_rsc->priority) { pe_rsc_trace(rsc, "Updating %s priority to promotion score (%d->%d)", child_rsc->id, child_rsc->priority, score); child_rsc->priority = score; } } } } /*! * \internal * \brief If a resource's current role is started, change it to unpromoted * * \param[in,out] data Resource to update * \param[in] user_data Ignored */ static void set_current_role_unpromoted(void *data, void *user_data) { pe_resource_t *rsc = (pe_resource_t *) data; if (rsc->role == RSC_ROLE_STARTED) { // Promotable clones should use unpromoted role instead of started rsc->role = RSC_ROLE_UNPROMOTED; } g_list_foreach(rsc->children, set_current_role_unpromoted, NULL); } /*! * \internal * \brief Set a resource's next role to unpromoted (or stopped if unassigned) * * \param[in,out] data Resource to update * \param[in] user_data Ignored */ static void set_next_role_unpromoted(void *data, void *user_data) { pe_resource_t *rsc = (pe_resource_t *) data; GList *assigned = NULL; rsc->fns->location(rsc, &assigned, FALSE); if (assigned == NULL) { pe__set_next_role(rsc, RSC_ROLE_STOPPED, "stopped instance"); } else { pe__set_next_role(rsc, RSC_ROLE_UNPROMOTED, "unpromoted instance"); g_list_free(assigned); } g_list_foreach(rsc->children, set_next_role_unpromoted, NULL); } /*! * \internal * \brief Set a resource's next role to promoted if not already set * * \param[in,out] data Resource to update * \param[in] user_data Ignored */ static void set_next_role_promoted(void *data, gpointer user_data) { pe_resource_t *rsc = (pe_resource_t *) data; if (rsc->next_role == RSC_ROLE_UNKNOWN) { pe__set_next_role(rsc, RSC_ROLE_PROMOTED, "promoted instance"); } g_list_foreach(rsc->children, set_next_role_promoted, NULL); } /*! * \internal * \brief Show instance's promotion score on node where it will be active * * \param[in,out] instance Promotable clone instance to show */ static void show_promotion_score(pe_resource_t *instance) { pe_node_t *chosen = instance->fns->location(instance, NULL, FALSE); if (pcmk_is_set(instance->cluster->flags, pe_flag_show_scores) && !pcmk__is_daemon && (instance->cluster->priv != NULL)) { pcmk__output_t *out = instance->cluster->priv; out->message(out, "promotion-score", instance, chosen, pcmk_readable_score(instance->sort_index)); } else { pe_rsc_debug(pe__const_top_resource(instance, false), "%s promotion score on %s: sort=%s priority=%s", instance->id, ((chosen == NULL)? "none" : pe__node_name(chosen)), pcmk_readable_score(instance->sort_index), pcmk_readable_score(instance->priority)); } } /*! * \internal * \brief Set a clone instance's promotion priority * * \param[in,out] data Promotable clone instance to update * \param[in] user_data Instance's parent clone */ static void set_instance_priority(gpointer data, gpointer user_data) { pe_resource_t *instance = (pe_resource_t *) data; const pe_resource_t *clone = (const pe_resource_t *) user_data; const pe_node_t *chosen = NULL; enum rsc_role_e next_role = RSC_ROLE_UNKNOWN; GList *list = NULL; pe_rsc_trace(clone, "Assigning priority for %s: %s", instance->id, role2text(instance->next_role)); if (instance->fns->state(instance, TRUE) == RSC_ROLE_STARTED) { set_current_role_unpromoted(instance, NULL); } // Only an instance that will be active can be promoted chosen = instance->fns->location(instance, &list, FALSE); if (pcmk__list_of_multiple(list)) { pcmk__config_err("Cannot promote non-colocated child %s", instance->id); } g_list_free(list); if (chosen == NULL) { return; } next_role = instance->fns->state(instance, FALSE); switch (next_role) { case RSC_ROLE_STARTED: case RSC_ROLE_UNKNOWN: // Set instance priority to its promotion score (or -1 if none) { bool is_default = false; instance->priority = promotion_score(instance, chosen, &is_default); if (is_default) { /* * Default to -1 if no value is set. This allows * instances eligible for promotion to be specified * based solely on rsc_location constraints, but * prevents any instance from being promoted if neither * a constraint nor a promotion score is present */ instance->priority = -1; } } break; case RSC_ROLE_UNPROMOTED: case RSC_ROLE_STOPPED: // Instance can't be promoted instance->priority = -INFINITY; break; case RSC_ROLE_PROMOTED: // Nothing needed (re-creating actions after scheduling fencing) break; default: CRM_CHECK(FALSE, crm_err("Unknown resource role %d for %s", next_role, instance->id)); } // Add relevant location constraint scores for promoted role apply_promoted_locations(instance, instance->rsc_location, chosen); apply_promoted_locations(instance, clone->rsc_location, chosen); // Consider instance's role-based colocations with other resources list = pcmk__this_with_colocations(instance); for (GList *iter = list; iter != NULL; iter = iter->next) { pcmk__colocation_t *cons = (pcmk__colocation_t *) iter->data; instance->cmds->apply_coloc_score(instance, cons->primary, cons, true); } g_list_free(list); instance->sort_index = instance->priority; if (next_role == RSC_ROLE_PROMOTED) { instance->sort_index = INFINITY; } pe_rsc_trace(clone, "Assigning %s priority = %d", instance->id, instance->priority); } /*! * \internal * \brief Set a promotable clone instance's role * * \param[in,out] data Promotable clone instance to update * \param[in,out] user_data Pointer to count of instances chosen for promotion */ static void set_instance_role(gpointer data, gpointer user_data) { pe_resource_t *instance = (pe_resource_t *) data; int *count = (int *) user_data; const pe_resource_t *clone = pe__const_top_resource(instance, false); pe_node_t *chosen = NULL; show_promotion_score(instance); if (instance->sort_index < 0) { pe_rsc_trace(clone, "Not supposed to promote instance %s", instance->id); } else if ((*count < pe__clone_promoted_max(instance)) || !pcmk_is_set(clone->flags, pe_rsc_managed)) { chosen = node_to_be_promoted_on(instance); } if (chosen == NULL) { set_next_role_unpromoted(instance, NULL); return; } if ((instance->role < RSC_ROLE_PROMOTED) && !pcmk_is_set(instance->cluster->flags, pe_flag_have_quorum) && (instance->cluster->no_quorum_policy == no_quorum_freeze)) { crm_notice("Clone instance %s cannot be promoted without quorum", instance->id); set_next_role_unpromoted(instance, NULL); return; } chosen->count++; pe_rsc_info(clone, "Choosing %s (%s) on %s for promotion", instance->id, role2text(instance->role), pe__node_name(chosen)); set_next_role_promoted(instance, NULL); (*count)++; } /*! * \internal * \brief Set roles for all instances of a promotable clone * * \param[in,out] rsc Promotable clone resource to update */ void pcmk__set_instance_roles(pe_resource_t *rsc) { int promoted = 0; GHashTableIter iter; pe_node_t *node = NULL; // Repurpose count to track the number of promoted instances assigned g_hash_table_iter_init(&iter, rsc->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (void **)&node)) { node->count = 0; } // Set instances' promotion priorities and sort by highest priority first g_list_foreach(rsc->children, set_instance_priority, rsc); sort_promotable_instances(rsc); // Choose the first N eligible instances to be promoted g_list_foreach(rsc->children, set_instance_role, &promoted); pe_rsc_info(rsc, "%s: Promoted %d instances of a possible %d", rsc->id, promoted, pe__clone_promoted_max(rsc)); } /*! * * \internal * \brief Create actions for promotable clone instances * * \param[in,out] clone Promotable clone to create actions for * \param[out] any_promoting Will be set true if any instance is promoting * \param[out] any_demoting Will be set true if any instance is demoting */ static void create_promotable_instance_actions(pe_resource_t *clone, bool *any_promoting, bool *any_demoting) { for (GList *iter = clone->children; iter != NULL; iter = iter->next) { pe_resource_t *instance = (pe_resource_t *) iter->data; instance->cmds->create_actions(instance); check_for_role_change(instance, any_demoting, any_promoting); } } /*! * \internal * \brief Reset each promotable instance's resource priority * * Reset the priority of each instance of a promotable clone to the clone's * priority (after promotion actions are scheduled, when instance priorities * were repurposed as promotion scores). * * \param[in,out] clone Promotable clone to reset */ static void reset_instance_priorities(pe_resource_t *clone) { for (GList *iter = clone->children; iter != NULL; iter = iter->next) { pe_resource_t *instance = (pe_resource_t *) iter->data; instance->priority = clone->priority; } } /*! * \internal * \brief Create actions specific to promotable clones * * \param[in,out] clone Promotable clone to create actions for */ void pcmk__create_promotable_actions(pe_resource_t *clone) { bool any_promoting = false; bool any_demoting = false; // Create actions for each clone instance individually create_promotable_instance_actions(clone, &any_promoting, &any_demoting); // Create pseudo-actions for clone as a whole pe__create_promotable_pseudo_ops(clone, any_promoting, any_demoting); // Undo our temporary repurposing of resource priority for instances reset_instance_priorities(clone); } /*! * \internal * \brief Create internal orderings for a promotable clone's instances * * \param[in,out] clone Promotable clone instance to order */ void pcmk__order_promotable_instances(pe_resource_t *clone) { pe_resource_t *previous = NULL; // Needed for ordered clones pcmk__promotable_restart_ordering(clone); for (GList *iter = clone->children; iter != NULL; iter = iter->next) { pe_resource_t *instance = (pe_resource_t *) iter->data; // Demote before promote pcmk__order_resource_actions(instance, RSC_DEMOTE, instance, RSC_PROMOTE, pe_order_optional); order_instance_promotion(clone, instance, previous); order_instance_demotion(clone, instance, previous); previous = instance; } } /*! * \internal * \brief Update dependent's allowed nodes for colocation with promotable * * \param[in,out] dependent Dependent resource to update * \param[in] primary Primary resource * \param[in] primary_node Node where an instance of the primary will be * \param[in] colocation Colocation constraint to apply */ static void update_dependent_allowed_nodes(pe_resource_t *dependent, const pe_resource_t *primary, const pe_node_t *primary_node, const pcmk__colocation_t *colocation) { GHashTableIter iter; pe_node_t *node = NULL; const char *primary_value = NULL; const char *attr = colocation->node_attribute; if (colocation->score >= INFINITY) { return; // Colocation is mandatory, so allowed node scores don't matter } primary_value = pcmk__colocation_node_attr(primary_node, attr, primary); pe_rsc_trace(colocation->primary, "Applying %s (%s with %s on %s by %s @%d) to %s", colocation->id, colocation->dependent->id, colocation->primary->id, pe__node_name(primary_node), attr, colocation->score, dependent->id); g_hash_table_iter_init(&iter, dependent->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (void **) &node)) { const char *dependent_value = pcmk__colocation_node_attr(node, attr, dependent); if (pcmk__str_eq(primary_value, dependent_value, pcmk__str_casei)) { node->weight = pcmk__add_scores(node->weight, colocation->score); pe_rsc_trace(colocation->primary, "Added %s score (%s) to %s (now %s)", colocation->id, pcmk_readable_score(colocation->score), pe__node_name(node), pcmk_readable_score(node->weight)); } } } /*! * \brief Update dependent for a colocation with a promotable clone * * \param[in] primary Primary resource in the colocation * \param[in,out] dependent Dependent resource in the colocation * \param[in] colocation Colocation constraint to apply */ void pcmk__update_dependent_with_promotable(const pe_resource_t *primary, pe_resource_t *dependent, const pcmk__colocation_t *colocation) { GList *affected_nodes = NULL; /* Build a list of all nodes where an instance of the primary will be, and * (for optional colocations) update the dependent's allowed node scores for * each one. */ for (GList *iter = primary->children; iter != NULL; iter = iter->next) { pe_resource_t *instance = (pe_resource_t *) iter->data; pe_node_t *node = instance->fns->location(instance, NULL, FALSE); if (node == NULL) { continue; } if (instance->fns->state(instance, FALSE) == colocation->primary_role) { update_dependent_allowed_nodes(dependent, primary, node, colocation); affected_nodes = g_list_prepend(affected_nodes, node); } } /* For mandatory colocations, add the primary's node score to the * dependent's node score for each affected node, and ban the dependent * from all other nodes. * * However, skip this for promoted-with-promoted colocations, otherwise * inactive dependent instances can't start (in the unpromoted role). */ if ((colocation->score >= INFINITY) && ((colocation->dependent_role != RSC_ROLE_PROMOTED) || (colocation->primary_role != RSC_ROLE_PROMOTED))) { pe_rsc_trace(colocation->primary, "Applying %s (mandatory %s with %s) to %s", colocation->id, colocation->dependent->id, colocation->primary->id, dependent->id); node_list_exclude(dependent->allowed_nodes, affected_nodes, TRUE); } g_list_free(affected_nodes); } /*! * \internal * \brief Update dependent priority for colocation with promotable * * \param[in] primary Primary resource in the colocation * \param[in,out] dependent Dependent resource in the colocation * \param[in] colocation Colocation constraint to apply */ void pcmk__update_promotable_dependent_priority(const pe_resource_t *primary, pe_resource_t *dependent, const pcmk__colocation_t *colocation) { pe_resource_t *primary_instance = NULL; // Look for a primary instance where dependent will be primary_instance = pcmk__find_compatible_instance(dependent, primary, colocation->primary_role, false); if (primary_instance != NULL) { // Add primary instance's priority to dependent's int new_priority = pcmk__add_scores(dependent->priority, colocation->score); pe_rsc_trace(colocation->primary, "Applying %s (%s with %s) to %s priority (%s + %s = %s)", colocation->id, colocation->dependent->id, colocation->primary->id, dependent->id, pcmk_readable_score(dependent->priority), pcmk_readable_score(colocation->score), pcmk_readable_score(new_priority)); dependent->priority = new_priority; } else if (colocation->score >= INFINITY) { // Mandatory colocation, but primary won't be here pe_rsc_trace(colocation->primary, "Applying %s (%s with %s) to %s: can't be promoted", colocation->id, colocation->dependent->id, colocation->primary->id, dependent->id); dependent->priority = -INFINITY; } }