diff --git a/lib/pacemaker/pcmk_sched_allocate.c b/lib/pacemaker/pcmk_sched_allocate.c index 79c648587c..112f0d3ad9 100644 --- a/lib/pacemaker/pcmk_sched_allocate.c +++ b/lib/pacemaker/pcmk_sched_allocate.c @@ -1,753 +1,752 @@ /* * Copyright 2004-2022 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU General Public License version 2 * or later (GPLv2+) WITHOUT ANY WARRANTY. */ #include #include #include #include #include #include #include #include #include #include "libpacemaker_private.h" CRM_TRACE_INIT_DATA(pacemaker); /*! * \internal * \brief Do deferred action checks after allocation * * When unpacking the resource history, the scheduler checks for resource * configurations that have changed since an action was run. However, at that * time, bundles using the REMOTE_CONTAINER_HACK don't have their final * parameter information, so instead they add a deferred check to a list. This * function processes one entry in that list. * * \param[in] rsc Resource that action history is for * \param[in] node Node that action history is for * \param[in] rsc_op Action history entry * \param[in] check Type of deferred check to do * \param[in] data_set Working set for cluster */ static void check_params(pe_resource_t *rsc, pe_node_t *node, xmlNode *rsc_op, enum pe_check_parameters check, pe_working_set_t *data_set) { const char *reason = NULL; op_digest_cache_t *digest_data = NULL; switch (check) { case pe_check_active: if (pcmk__check_action_config(rsc, node, rsc_op) && pe_get_failcount(node, rsc, NULL, pe_fc_effective, NULL, data_set)) { reason = "action definition changed"; } break; case pe_check_last_failure: digest_data = rsc_action_digest_cmp(rsc, rsc_op, node, data_set); switch (digest_data->rc) { case RSC_DIGEST_UNKNOWN: crm_trace("Resource %s history entry %s on %s has " "no digest to compare", rsc->id, ID(rsc_op), node->details->id); break; case RSC_DIGEST_MATCH: break; default: reason = "resource parameters have changed"; break; } break; } if (reason != NULL) { pe__clear_failcount(rsc, node, reason, data_set); } } /*! * \internal * \brief Check whether a resource has failcount clearing scheduled on a node * * \param[in] node Node to check * \param[in] rsc Resource to check * * \return true if \p rsc has failcount clearing scheduled on \p node, * otherwise false */ static bool failcount_clear_action_exists(pe_node_t *node, pe_resource_t *rsc) { GList *list = pe__resource_actions(rsc, node, CRM_OP_CLEAR_FAILCOUNT, TRUE); if (list != NULL) { g_list_free(list); return true; } return false; } /*! * \internal * \brief Ban a resource from a node if it reached its failure threshold there * * \param[in] rsc Resource to check failure threshold for * \param[in] node Node to check \p rsc on */ static void check_failure_threshold(pe_resource_t *rsc, pe_node_t *node) { // If this is a collective resource, apply recursively to children instead if (rsc->children != NULL) { g_list_foreach(rsc->children, (GFunc) check_failure_threshold, node); return; } else if (failcount_clear_action_exists(node, rsc)) { /* Don't force the resource away from this node due to a failcount * that's going to be cleared. * * @TODO Failcount clearing can be scheduled in * pcmk__handle_rsc_config_changes() via process_rsc_history(), or in * stage5() via check_params(). This runs well before then, so it cannot * detect those, meaning we might check the migration threshold when we * shouldn't. Worst case, we stop or move the resource, then move it * back in the next transition. */ return; } else { pe_resource_t *failed = NULL; if (pcmk__threshold_reached(rsc, node, &failed)) { resource_location(failed, node, -INFINITY, "__fail_limit__", rsc->cluster); } } } /*! * \internal * \brief If resource has exclusive discovery, ban node if not allowed * * Location constraints have a resource-discovery option that allows users to * specify where probes are done for the affected resource. If this is set to * exclusive, probes will only be done on nodes listed in exclusive constraints. * This function bans the resource from the node if the node is not listed. * * \param[in] rsc Resource to check * \param[in] node Node to check \p rsc on */ static void apply_exclusive_discovery(pe_resource_t *rsc, pe_node_t *node) { if (rsc->exclusive_discover || uber_parent(rsc)->exclusive_discover) { pe_node_t *match = NULL; // If this is a collective resource, apply recursively to children g_list_foreach(rsc->children, (GFunc) apply_exclusive_discovery, node); match = g_hash_table_lookup(rsc->allowed_nodes, node->details->id); if ((match != NULL) && (match->rsc_discover_mode != pe_discover_exclusive)) { match->weight = -INFINITY; } } } /*! * \internal * \brief Apply stickiness to a resource if appropriate * * \param[in] rsc Resource to check for stickiness * \param[in] data_set Cluster working set */ static void apply_stickiness(pe_resource_t *rsc, pe_working_set_t *data_set) { pe_node_t *node = NULL; // If this is a collective resource, apply recursively to children instead if (rsc->children != NULL) { g_list_foreach(rsc->children, (GFunc) apply_stickiness, data_set); return; } /* A resource is sticky if it is managed, has stickiness configured, and is * active on a single node. */ if (!pcmk_is_set(rsc->flags, pe_rsc_managed) || (rsc->stickiness < 1) || !pcmk__list_of_1(rsc->running_on)) { return; } node = rsc->running_on->data; /* In a symmetric cluster, stickiness can always be used. In an * asymmetric cluster, we have to check whether the resource is still * allowed on the node, so we don't keep the resource somewhere it is no * longer explicitly enabled. */ if (!pcmk_is_set(rsc->cluster->flags, pe_flag_symmetric_cluster) && (pe_hash_table_lookup(rsc->allowed_nodes, node->details->id) == NULL)) { pe_rsc_debug(rsc, "Ignoring %s stickiness because the cluster is " "asymmetric and node %s is not explicitly allowed", rsc->id, node->details->uname); return; } pe_rsc_debug(rsc, "Resource %s has %d stickiness on node %s", rsc->id, rsc->stickiness, node->details->uname); resource_location(rsc, node, rsc->stickiness, "stickiness", rsc->cluster); } /*! * \internal * \brief Apply shutdown locks for all resources as appropriate * * \param[in] data_set Cluster working set */ static void apply_shutdown_locks(pe_working_set_t *data_set) { if (pcmk_is_set(data_set->flags, pe_flag_shutdown_lock)) { for (GList *iter = data_set->resources; iter != NULL; iter = iter->next) { pe_resource_t *rsc = (pe_resource_t *) iter->data; rsc->cmds->shutdown_lock(rsc); } } } /*! * \internal * \brief Calculate the number of available nodes in the cluster * * \param[in] data_set Cluster working set */ static void count_available_nodes(pe_working_set_t *data_set) { if (pcmk_is_set(data_set->flags, pe_flag_no_compat)) { return; } // @COMPAT for API backward compatibility only (cluster does not use value) for (GList *iter = data_set->nodes; iter != NULL; iter = iter->next) { pe_node_t *node = (pe_node_t *) iter->data; if ((node != NULL) && (node->weight >= 0) && node->details->online && (node->details->type != node_ping)) { data_set->max_valid_nodes++; } } crm_trace("Online node count: %d", data_set->max_valid_nodes); } /* * \internal * \brief Apply node-specific scheduling criteria * * After the CIB has been unpacked, process node-specific scheduling criteria * including shutdown locks, location constraints, resource stickiness, * migration thresholds, and exclusive resource discovery. */ static void apply_node_criteria(pe_working_set_t *data_set) { crm_trace("Applying node-specific scheduling criteria"); apply_shutdown_locks(data_set); count_available_nodes(data_set); pcmk__apply_locations(data_set); g_list_foreach(data_set->resources, (GFunc) apply_stickiness, data_set); for (GList *node_iter = data_set->nodes; node_iter != NULL; node_iter = node_iter->next) { for (GList *rsc_iter = data_set->resources; rsc_iter != NULL; rsc_iter = rsc_iter->next) { pe_node_t *node = (pe_node_t *) node_iter->data; pe_resource_t *rsc = (pe_resource_t *) rsc_iter->data; check_failure_threshold(rsc, node); apply_exclusive_discovery(rsc, node); } } } /*! * \internal * \brief Allocate resources to nodes * * \param[in] data_set Cluster working set */ static void allocate_resources(pe_working_set_t *data_set) { GList *iter = NULL; crm_trace("Allocating resources to nodes"); if (!pcmk__str_eq(data_set->placement_strategy, "default", pcmk__str_casei)) { pcmk__sort_resources(data_set); } pcmk__show_node_capacities("Original", data_set); if (pcmk_is_set(data_set->flags, pe_flag_have_remote_nodes)) { /* Allocate remote connection resources first (which will also allocate * any colocation dependencies). If the connection is migrating, always * prefer the partial migration target. */ for (iter = data_set->resources; iter != NULL; iter = iter->next) { pe_resource_t *rsc = (pe_resource_t *) iter->data; if (rsc->is_remote_node) { pe_rsc_trace(rsc, "Allocating remote connection resource '%s'", rsc->id); rsc->cmds->allocate(rsc, rsc->partial_migration_target, data_set); } } } /* now do the rest of the resources */ for (iter = data_set->resources; iter != NULL; iter = iter->next) { pe_resource_t *rsc = (pe_resource_t *) iter->data; if (!rsc->is_remote_node) { pe_rsc_trace(rsc, "Allocating %s resource '%s'", crm_element_name(rsc->xml), rsc->id); rsc->cmds->allocate(rsc, NULL, data_set); } } pcmk__show_node_capacities("Remaining", data_set); } /*! * \internal * \brief Schedule fail count clearing on online nodes if resource is orphaned * * \param[in] rsc Resource to check * \param[in] data_set Cluster working set */ static void clear_failcounts_if_orphaned(pe_resource_t *rsc, pe_working_set_t *data_set) { if (!pcmk_is_set(rsc->flags, pe_rsc_orphan)) { return; } crm_trace("Clear fail counts for orphaned resource %s", rsc->id); /* There's no need to recurse into rsc->children because those * should just be unallocated clone instances. */ for (GList *iter = data_set->nodes; iter != NULL; iter = iter->next) { pe_node_t *node = (pe_node_t *) iter->data; pe_action_t *clear_op = NULL; if (!node->details->online) { continue; } if (pe_get_failcount(node, rsc, NULL, pe_fc_effective, NULL, data_set) == 0) { continue; } clear_op = pe__clear_failcount(rsc, node, "it is orphaned", data_set); /* We can't use order_action_then_stop() here because its * pe_order_preserve breaks things */ pcmk__new_ordering(clear_op->rsc, NULL, clear_op, rsc, stop_key(rsc), NULL, pe_order_optional, data_set); } } gboolean stage5(pe_working_set_t * data_set) { GList *gIter = NULL; - allocate_resources(data_set); - // Process deferred action checks pe__foreach_param_check(data_set, check_params); pe__free_param_checks(data_set); if (pcmk_is_set(data_set->flags, pe_flag_startup_probes)) { crm_trace("Calculating needed probes"); pcmk__schedule_probes(data_set); } if (pcmk_is_set(data_set->flags, pe_flag_stop_rsc_orphans)) { g_list_foreach(data_set->resources, (GFunc) clear_failcounts_if_orphaned, data_set); } crm_trace("Creating actions"); for (gIter = data_set->resources; gIter != NULL; gIter = gIter->next) { pe_resource_t *rsc = (pe_resource_t *) gIter->data; rsc->cmds->create_actions(rsc, data_set); } crm_trace("Creating done"); return TRUE; } static gboolean is_managed(const pe_resource_t * rsc) { GList *gIter = rsc->children; if (pcmk_is_set(rsc->flags, pe_rsc_managed)) { return TRUE; } for (; gIter != NULL; gIter = gIter->next) { pe_resource_t *child_rsc = (pe_resource_t *) gIter->data; if (is_managed(child_rsc)) { return TRUE; } } return FALSE; } static gboolean any_managed_resources(pe_working_set_t * data_set) { GList *gIter = data_set->resources; for (; gIter != NULL; gIter = gIter->next) { pe_resource_t *rsc = (pe_resource_t *) gIter->data; if (is_managed(rsc)) { return TRUE; } } return FALSE; } /* * Create dependencies for stonith and shutdown operations */ gboolean stage6(pe_working_set_t * data_set) { pe_action_t *dc_down = NULL; pe_action_t *stonith_op = NULL; gboolean integrity_lost = FALSE; gboolean need_stonith = TRUE; GList *gIter; GList *stonith_ops = NULL; GList *shutdown_ops = NULL; /* Remote ordering constraints need to happen prior to calculating fencing * because it is one more place we can mark nodes as needing fencing. */ pcmk__order_remote_connection_actions(data_set); crm_trace("Processing fencing and shutdown cases"); if (any_managed_resources(data_set) == FALSE) { crm_notice("Delaying fencing operations until there are resources to manage"); need_stonith = FALSE; } /* Check each node for stonith/shutdown */ for (gIter = data_set->nodes; gIter != NULL; gIter = gIter->next) { pe_node_t *node = (pe_node_t *) gIter->data; /* Guest nodes are "fenced" by recovering their container resource, * so handle them separately. */ if (pe__is_guest_node(node)) { if (node->details->remote_requires_reset && need_stonith && pe_can_fence(data_set, node)) { pcmk__fence_guest(node, data_set); } continue; } stonith_op = NULL; if (node->details->unclean && need_stonith && pe_can_fence(data_set, node)) { stonith_op = pe_fence_op(node, NULL, FALSE, "node is unclean", FALSE, data_set); pe_warn("Scheduling Node %s for STONITH", node->details->uname); pcmk__order_vs_fence(stonith_op, data_set); if (node->details->is_dc) { // Remember if the DC is being fenced dc_down = stonith_op; } else { if (!pcmk_is_set(data_set->flags, pe_flag_concurrent_fencing) && (stonith_ops != NULL)) { /* Concurrent fencing is disabled, so order each non-DC * fencing in a chain. If there is any DC fencing or * shutdown, it will be ordered after the last action in the * chain later. */ order_actions((pe_action_t *) stonith_ops->data, stonith_op, pe_order_optional); } // Remember all non-DC fencing actions in a separate list stonith_ops = g_list_prepend(stonith_ops, stonith_op); } } else if (node->details->online && node->details->shutdown && /* TODO define what a shutdown op means for a remote node. * For now we do not send shutdown operations for remote nodes, but * if we can come up with a good use for this in the future, we will. */ pe__is_guest_or_remote_node(node) == FALSE) { pe_action_t *down_op = pcmk__new_shutdown_action(node, data_set); if (node->details->is_dc) { // Remember if the DC is being shut down dc_down = down_op; } else { // Remember non-DC shutdowns for later ordering shutdown_ops = g_list_prepend(shutdown_ops, down_op); } } if (node->details->unclean && stonith_op == NULL) { integrity_lost = TRUE; pe_warn("Node %s is unclean!", node->details->uname); } } if (integrity_lost) { if (!pcmk_is_set(data_set->flags, pe_flag_stonith_enabled)) { pe_warn("YOUR RESOURCES ARE NOW LIKELY COMPROMISED"); pe_err("ENABLE STONITH TO KEEP YOUR RESOURCES SAFE"); } else if (!pcmk_is_set(data_set->flags, pe_flag_have_quorum)) { crm_notice("Cannot fence unclean nodes until quorum is" " attained (or no-quorum-policy is set to ignore)"); } } if (dc_down != NULL) { /* Order any non-DC shutdowns before any DC shutdown, to avoid repeated * DC elections. However, we don't want to order non-DC shutdowns before * a DC *fencing*, because even though we don't want a node that's * shutting down to become DC, the DC fencing could be ordered before a * clone stop that's also ordered before the shutdowns, thus leading to * a graph loop. */ if (pcmk__str_eq(dc_down->task, CRM_OP_SHUTDOWN, pcmk__str_casei)) { for (gIter = shutdown_ops; gIter != NULL; gIter = gIter->next) { pe_action_t *node_stop = (pe_action_t *) gIter->data; crm_debug("Ordering shutdown on %s before %s on DC %s", node_stop->node->details->uname, dc_down->task, dc_down->node->details->uname); order_actions(node_stop, dc_down, pe_order_optional); } } // Order any non-DC fencing before any DC fencing or shutdown if (pcmk_is_set(data_set->flags, pe_flag_concurrent_fencing)) { /* With concurrent fencing, order each non-DC fencing action * separately before any DC fencing or shutdown. */ for (gIter = stonith_ops; gIter != NULL; gIter = gIter->next) { order_actions((pe_action_t *) gIter->data, dc_down, pe_order_optional); } } else if (stonith_ops) { /* Without concurrent fencing, the non-DC fencing actions are * already ordered relative to each other, so we just need to order * the DC fencing after the last action in the chain (which is the * first item in the list). */ order_actions((pe_action_t *) stonith_ops->data, dc_down, pe_order_optional); } } g_list_free(stonith_ops); g_list_free(shutdown_ops); return TRUE; } static void log_resource_details(pe_working_set_t *data_set) { pcmk__output_t *out = data_set->priv; GList *all = NULL; /* We need a list of nodes that we are allowed to output information for. * This is necessary because out->message for all the resource-related * messages expects such a list, due to the `crm_mon --node=` feature. Here, * we just make it a list of all the nodes. */ all = g_list_prepend(all, (gpointer) "*"); for (GList *item = data_set->resources; item != NULL; item = item->next) { pe_resource_t *rsc = (pe_resource_t *) item->data; // Log all resources except inactive orphans if (!pcmk_is_set(rsc->flags, pe_rsc_orphan) || (rsc->role != RSC_ROLE_STOPPED)) { out->message(out, crm_map_element_name(rsc->xml), 0, rsc, all, all); } } g_list_free(all); } static void log_all_actions(pe_working_set_t *data_set) { /* This only ever outputs to the log, so ignore whatever output object was * previously set and just log instead. */ pcmk__output_t *prev_out = data_set->priv; pcmk__output_t *out = pcmk__new_logger(); if (out == NULL) { return; } pcmk__output_set_log_level(out, LOG_NOTICE); data_set->priv = out; out->begin_list(out, NULL, NULL, "Actions"); pcmk__output_actions(data_set); out->end_list(out); out->finish(out, CRM_EX_OK, true, NULL); pcmk__output_free(out); data_set->priv = prev_out; } /*! * \internal * \brief Log all required but unrunnable actions at trace level * * \param[in] data_set Cluster working set */ static void log_unrunnable_actions(pe_working_set_t *data_set) { const uint64_t flags = pe_action_optional|pe_action_runnable|pe_action_pseudo; crm_trace("Required but unrunnable actions:"); for (GList *iter = data_set->actions; iter != NULL; iter = iter->next) { pe_action_t *action = (pe_action_t *) iter->data; if (!pcmk_any_flags_set(action->flags, flags)) { pcmk__log_action("\t", action, true); } } } /*! * \internal * \brief Unpack the CIB for scheduling * * \param[in] cib CIB XML to unpack (may be NULL if previously unpacked) * \param[in] flags Working set flags to set in addition to defaults * \param[in] data_set Cluster working set */ static void unpack_cib(xmlNode *cib, unsigned long long flags, pe_working_set_t *data_set) { if (pcmk_is_set(data_set->flags, pe_flag_have_status)) { crm_trace("Reusing previously calculated cluster status"); pe__set_working_set_flags(data_set, flags); return; } CRM_ASSERT(cib != NULL); crm_trace("Calculating cluster status"); /* This will zero the entire struct without freeing anything first, so * callers should never call pcmk__schedule_actions() with a populated data * set unless pe_flag_have_status is set (i.e. cluster_status() was * previously called, whether directly or via pcmk__schedule_actions()). */ set_working_set_defaults(data_set); pe__set_working_set_flags(data_set, flags); data_set->input = cib; cluster_status(data_set); // Sets pe_flag_have_status } /*! * \internal * \brief Run the scheduler for a given CIB * * \param[in] cib CIB XML to use as scheduler input * \param[in] flags Working set flags to set in addition to defaults * \param[in,out] data_set Cluster working set */ void pcmk__schedule_actions(xmlNode *cib, unsigned long long flags, pe_working_set_t *data_set) { unpack_cib(cib, flags, data_set); pcmk__set_allocation_methods(data_set); pcmk__apply_node_health(data_set); pcmk__unpack_constraints(data_set); if (pcmk_is_set(data_set->flags, pe_flag_check_config)) { return; } if (!pcmk_is_set(data_set->flags, pe_flag_quick_location) && pcmk__is_daemon) { log_resource_details(data_set); } apply_node_criteria(data_set); if (pcmk_is_set(data_set->flags, pe_flag_quick_location)) { return; } pcmk__create_internal_constraints(data_set); pcmk__handle_rsc_config_changes(data_set); - crm_trace("Allocate resources"); + allocate_resources(data_set); + stage5(data_set); crm_trace("Processing fencing and shutdown cases"); stage6(data_set); pcmk__apply_orderings(data_set); log_all_actions(data_set); crm_trace("Create transition graph"); pcmk__create_graph(data_set); if (get_crm_log_level() == LOG_TRACE) { log_unrunnable_actions(data_set); } }