diff --git a/lib/pacemaker/pcmk_sched_allocate.c b/lib/pacemaker/pcmk_sched_allocate.c index fe52dba4d3..57e9942194 100644 --- a/lib/pacemaker/pcmk_sched_allocate.c +++ b/lib/pacemaker/pcmk_sched_allocate.c @@ -1,3073 +1,3072 @@ /* * Copyright 2004-2020 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 CRM_TRACE_INIT_DATA(pacemaker); void set_alloc_actions(pe_working_set_t * data_set); extern void ReloadRsc(resource_t * rsc, node_t *node, pe_working_set_t * data_set); extern gboolean DeleteRsc(resource_t * rsc, node_t * node, gboolean optional, pe_working_set_t * data_set); static void apply_remote_node_ordering(pe_working_set_t *data_set); static enum remote_connection_state get_remote_node_state(pe_node_t *node); enum remote_connection_state { remote_state_unknown = 0, remote_state_alive = 1, remote_state_resting = 2, remote_state_failed = 3, remote_state_stopped = 4 }; static const char * state2text(enum remote_connection_state state) { switch (state) { case remote_state_unknown: return "unknown"; case remote_state_alive: return "alive"; case remote_state_resting: return "resting"; case remote_state_failed: return "failed"; case remote_state_stopped: return "stopped"; } return "impossible"; } resource_alloc_functions_t resource_class_alloc_functions[] = { { pcmk__native_merge_weights, pcmk__native_allocate, native_create_actions, native_create_probe, native_internal_constraints, native_rsc_colocation_lh, native_rsc_colocation_rh, native_rsc_location, native_action_flags, native_update_actions, native_expand, native_append_meta, }, { pcmk__group_merge_weights, pcmk__group_allocate, group_create_actions, native_create_probe, group_internal_constraints, group_rsc_colocation_lh, group_rsc_colocation_rh, group_rsc_location, group_action_flags, group_update_actions, group_expand, group_append_meta, }, { pcmk__native_merge_weights, pcmk__clone_allocate, clone_create_actions, clone_create_probe, clone_internal_constraints, clone_rsc_colocation_lh, clone_rsc_colocation_rh, clone_rsc_location, clone_action_flags, pcmk__multi_update_actions, clone_expand, clone_append_meta, }, { pcmk__native_merge_weights, pcmk__bundle_allocate, pcmk__bundle_create_actions, pcmk__bundle_create_probe, pcmk__bundle_internal_constraints, pcmk__bundle_rsc_colocation_lh, pcmk__bundle_rsc_colocation_rh, pcmk__bundle_rsc_location, pcmk__bundle_action_flags, pcmk__multi_update_actions, pcmk__bundle_expand, pcmk__bundle_append_meta, } }; gboolean update_action_flags(action_t * action, enum pe_action_flags flags, const char *source, int line) { static unsigned long calls = 0; gboolean changed = FALSE; gboolean clear = is_set(flags, pe_action_clear); enum pe_action_flags last = action->flags; if (clear) { action->flags = crm_clear_bit(source, line, action->uuid, action->flags, flags); } else { action->flags = crm_set_bit(source, line, action->uuid, action->flags, flags); } if (last != action->flags) { calls++; changed = TRUE; /* Useful for tracking down _who_ changed a specific flag */ /* CRM_ASSERT(calls != 534); */ clear_bit(flags, pe_action_clear); crm_trace("%s on %s: %sset flags 0x%.6x (was 0x%.6x, now 0x%.6x, %lu, %s)", action->uuid, action->node ? action->node->details->uname : "[none]", clear ? "un-" : "", flags, last, action->flags, calls, source); } return changed; } static gboolean check_rsc_parameters(resource_t * rsc, node_t * node, xmlNode * rsc_entry, gboolean active_here, pe_working_set_t * data_set) { int attr_lpc = 0; gboolean force_restart = FALSE; gboolean delete_resource = FALSE; gboolean changed = FALSE; const char *value = NULL; const char *old_value = NULL; const char *attr_list[] = { XML_ATTR_TYPE, XML_AGENT_ATTR_CLASS, XML_AGENT_ATTR_PROVIDER }; for (; attr_lpc < DIMOF(attr_list); attr_lpc++) { value = crm_element_value(rsc->xml, attr_list[attr_lpc]); old_value = crm_element_value(rsc_entry, attr_list[attr_lpc]); if (value == old_value /* i.e. NULL */ || crm_str_eq(value, old_value, TRUE)) { continue; } changed = TRUE; trigger_unfencing(rsc, node, "Device definition changed", NULL, data_set); if (active_here) { force_restart = TRUE; crm_notice("Forcing restart of %s on %s, %s changed: %s -> %s", rsc->id, node->details->uname, attr_list[attr_lpc], crm_str(old_value), crm_str(value)); } } if (force_restart) { /* make sure the restart happens */ stop_action(rsc, node, FALSE); set_bit(rsc->flags, pe_rsc_start_pending); delete_resource = TRUE; } else if (changed) { delete_resource = TRUE; } return delete_resource; } static void CancelXmlOp(resource_t * rsc, xmlNode * xml_op, node_t * active_node, const char *reason, pe_working_set_t * data_set) { guint interval_ms = 0; action_t *cancel = NULL; const char *task = NULL; const char *call_id = NULL; const char *interval_ms_s = NULL; CRM_CHECK(xml_op != NULL, return); CRM_CHECK(active_node != NULL, return); task = crm_element_value(xml_op, XML_LRM_ATTR_TASK); call_id = crm_element_value(xml_op, XML_LRM_ATTR_CALLID); interval_ms_s = crm_element_value(xml_op, XML_LRM_ATTR_INTERVAL_MS); interval_ms = crm_parse_ms(interval_ms_s); crm_info("Action " CRM_OP_FMT " on %s will be stopped: %s", rsc->id, task, interval_ms, active_node->details->uname, (reason? reason : "unknown")); cancel = pe_cancel_op(rsc, task, interval_ms, active_node, data_set); add_hash_param(cancel->meta, XML_LRM_ATTR_CALLID, call_id); custom_action_order(rsc, stop_key(rsc), NULL, rsc, NULL, cancel, pe_order_optional, data_set); } static gboolean check_action_definition(resource_t * rsc, node_t * active_node, xmlNode * xml_op, pe_working_set_t * data_set) { char *key = NULL; guint interval_ms = 0; const char *interval_ms_s = NULL; const op_digest_cache_t *digest_data = NULL; gboolean did_change = FALSE; const char *task = crm_element_value(xml_op, XML_LRM_ATTR_TASK); const char *digest_secure = NULL; CRM_CHECK(active_node != NULL, return FALSE); interval_ms_s = crm_element_value(xml_op, XML_LRM_ATTR_INTERVAL_MS); interval_ms = crm_parse_ms(interval_ms_s); if (interval_ms > 0) { xmlNode *op_match = NULL; /* we need to reconstruct the key because of the way we used to construct resource IDs */ key = pcmk__op_key(rsc->id, task, interval_ms); pe_rsc_trace(rsc, "Checking parameters for %s", key); op_match = find_rsc_op_entry(rsc, key); if (op_match == NULL && is_set(data_set->flags, pe_flag_stop_action_orphans)) { CancelXmlOp(rsc, xml_op, active_node, "orphan", data_set); free(key); return TRUE; } else if (op_match == NULL) { pe_rsc_debug(rsc, "Orphan action detected: %s on %s", key, active_node->details->uname); free(key); return TRUE; } free(key); key = NULL; } crm_trace("Testing " CRM_OP_FMT " on %s", rsc->id, task, interval_ms, active_node->details->uname); if ((interval_ms == 0) && safe_str_eq(task, RSC_STATUS)) { /* Reload based on the start action not a probe */ task = RSC_START; } else if ((interval_ms == 0) && safe_str_eq(task, RSC_MIGRATED)) { /* Reload based on the start action not a migrate */ task = RSC_START; } else if ((interval_ms == 0) && safe_str_eq(task, RSC_PROMOTE)) { /* Reload based on the start action not a promote */ task = RSC_START; } digest_data = rsc_action_digest_cmp(rsc, xml_op, active_node, data_set); if(is_set(data_set->flags, pe_flag_sanitized)) { digest_secure = crm_element_value(xml_op, XML_LRM_ATTR_SECURE_DIGEST); } if(digest_data->rc != RSC_DIGEST_MATCH && digest_secure && digest_data->digest_secure_calc && strcmp(digest_data->digest_secure_calc, digest_secure) == 0) { if (is_set(data_set->flags, pe_flag_stdout)) { printf("Only 'private' parameters to " CRM_OP_FMT " on %s changed: %s\n", rsc->id, task, interval_ms, active_node->details->uname, crm_element_value(xml_op, XML_ATTR_TRANSITION_MAGIC)); } } else if (digest_data->rc == RSC_DIGEST_RESTART) { /* Changes that force a restart */ pe_action_t *required = NULL; did_change = TRUE; key = pcmk__op_key(rsc->id, task, interval_ms); crm_log_xml_info(digest_data->params_restart, "params:restart"); required = custom_action(rsc, key, task, NULL, TRUE, TRUE, data_set); pe_action_set_flag_reason(__FUNCTION__, __LINE__, required, NULL, "resource definition change", pe_action_optional, TRUE); trigger_unfencing(rsc, active_node, "Device parameters changed", NULL, data_set); } else if ((digest_data->rc == RSC_DIGEST_ALL) || (digest_data->rc == RSC_DIGEST_UNKNOWN)) { /* Changes that can potentially be handled by a reload */ const char *digest_restart = crm_element_value(xml_op, XML_LRM_ATTR_RESTART_DIGEST); did_change = TRUE; trigger_unfencing(rsc, active_node, "Device parameters changed (reload)", NULL, data_set); crm_log_xml_info(digest_data->params_all, "params:reload"); key = pcmk__op_key(rsc->id, task, interval_ms); if (interval_ms > 0) { action_t *op = NULL; #if 0 /* Always reload/restart the entire resource */ ReloadRsc(rsc, active_node, data_set); #else /* Re-sending the recurring op is sufficient - the old one will be cancelled automatically */ op = custom_action(rsc, key, task, active_node, TRUE, TRUE, data_set); set_bit(op->flags, pe_action_reschedule); #endif } else if (digest_restart) { pe_rsc_trace(rsc, "Reloading '%s' action for resource %s", task, rsc->id); /* Reload this resource */ ReloadRsc(rsc, active_node, data_set); free(key); } else { pe_action_t *required = NULL; pe_rsc_trace(rsc, "Resource %s doesn't know how to reload", rsc->id); /* Re-send the start/demote/promote op * Recurring ops will be detected independently */ required = custom_action(rsc, key, task, NULL, TRUE, TRUE, data_set); pe_action_set_flag_reason(__FUNCTION__, __LINE__, required, NULL, "resource definition change", pe_action_optional, TRUE); } } return did_change; } /*! * \internal * \brief Do deferred action checks after allocation * * \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 (check_action_definition(rsc, node, rsc_op, data_set) && 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) { pe__clear_failcount(rsc, node, reason, data_set); } } static void check_actions_for(xmlNode * rsc_entry, resource_t * rsc, node_t * node, pe_working_set_t * data_set) { GListPtr gIter = NULL; int offset = -1; guint interval_ms = 0; int stop_index = 0; int start_index = 0; const char *task = NULL; const char *interval_ms_s = NULL; xmlNode *rsc_op = NULL; GListPtr op_list = NULL; GListPtr sorted_op_list = NULL; CRM_CHECK(node != NULL, return); if (is_set(rsc->flags, pe_rsc_orphan)) { resource_t *parent = uber_parent(rsc); if(parent == NULL || pe_rsc_is_clone(parent) == FALSE || is_set(parent->flags, pe_rsc_unique)) { pe_rsc_trace(rsc, "Skipping param check for %s and deleting: orphan", rsc->id); DeleteRsc(rsc, node, FALSE, data_set); } else { pe_rsc_trace(rsc, "Skipping param check for %s (orphan clone)", rsc->id); } return; } else if (pe_find_node_id(rsc->running_on, node->details->id) == NULL) { if (check_rsc_parameters(rsc, node, rsc_entry, FALSE, data_set)) { DeleteRsc(rsc, node, FALSE, data_set); } pe_rsc_trace(rsc, "Skipping param check for %s: no longer active on %s", rsc->id, node->details->uname); return; } pe_rsc_trace(rsc, "Processing %s on %s", rsc->id, node->details->uname); if (check_rsc_parameters(rsc, node, rsc_entry, TRUE, data_set)) { DeleteRsc(rsc, node, FALSE, data_set); } for (rsc_op = __xml_first_child_element(rsc_entry); rsc_op != NULL; rsc_op = __xml_next_element(rsc_op)) { if (crm_str_eq((const char *)rsc_op->name, XML_LRM_TAG_RSC_OP, TRUE)) { op_list = g_list_prepend(op_list, rsc_op); } } sorted_op_list = g_list_sort(op_list, sort_op_by_callid); calculate_active_ops(sorted_op_list, &start_index, &stop_index); for (gIter = sorted_op_list; gIter != NULL; gIter = gIter->next) { xmlNode *rsc_op = (xmlNode *) gIter->data; offset++; if (start_index < stop_index) { /* stopped */ continue; } else if (offset < start_index) { /* action occurred prior to a start */ continue; } task = crm_element_value(rsc_op, XML_LRM_ATTR_TASK); interval_ms_s = crm_element_value(rsc_op, XML_LRM_ATTR_INTERVAL_MS); interval_ms = crm_parse_ms(interval_ms_s); if ((interval_ms > 0) && (is_set(rsc->flags, pe_rsc_maintenance) || node->details->maintenance)) { // Maintenance mode cancels recurring operations CancelXmlOp(rsc, rsc_op, node, "maintenance mode", data_set); } else if ((interval_ms > 0) || safe_str_eq(task, RSC_STATUS) || safe_str_eq(task, RSC_START) || safe_str_eq(task, RSC_PROMOTE) || safe_str_eq(task, RSC_MIGRATED)) { /* If a resource operation failed, and the operation's definition * has changed, clear any fail count so they can be retried fresh. */ if (pe__bundle_needs_remote_name(rsc)) { /* We haven't allocated resources to nodes yet, so if the * REMOTE_CONTAINER_HACK is used, we may calculate the digest * based on the literal "#uname" value rather than the properly * substituted value. That would mistakenly make the action * definition appear to have been changed. Defer the check until * later in this case. */ pe__add_param_check(rsc_op, rsc, node, pe_check_active, data_set); } else if (check_action_definition(rsc, node, rsc_op, data_set) && pe_get_failcount(node, rsc, NULL, pe_fc_effective, NULL, data_set)) { pe__clear_failcount(rsc, node, "action definition changed", data_set); } } } g_list_free(sorted_op_list); } static GListPtr find_rsc_list(GListPtr result, resource_t * rsc, const char *id, gboolean renamed_clones, gboolean partial, pe_working_set_t * data_set) { GListPtr gIter = NULL; gboolean match = FALSE; if (id == NULL) { return NULL; } if (rsc == NULL) { if (data_set == NULL) { return NULL; } for (gIter = data_set->resources; gIter != NULL; gIter = gIter->next) { pe_resource_t *child = (pe_resource_t *) gIter->data; result = find_rsc_list(result, child, id, renamed_clones, partial, NULL); } return result; } if (partial) { if (strstr(rsc->id, id)) { match = TRUE; } else if (renamed_clones && rsc->clone_name && strstr(rsc->clone_name, id)) { match = TRUE; } } else { if (strcmp(rsc->id, id) == 0) { match = TRUE; } else if (renamed_clones && rsc->clone_name && strcmp(rsc->clone_name, id) == 0) { match = TRUE; } } if (match) { result = g_list_prepend(result, rsc); } if (rsc->children) { gIter = rsc->children; for (; gIter != NULL; gIter = gIter->next) { resource_t *child = (resource_t *) gIter->data; result = find_rsc_list(result, child, id, renamed_clones, partial, NULL); } } return result; } static void check_actions(pe_working_set_t * data_set) { const char *id = NULL; node_t *node = NULL; xmlNode *lrm_rscs = NULL; xmlNode *status = get_object_root(XML_CIB_TAG_STATUS, data_set->input); xmlNode *node_state = NULL; for (node_state = __xml_first_child_element(status); node_state != NULL; node_state = __xml_next_element(node_state)) { if (crm_str_eq((const char *)node_state->name, XML_CIB_TAG_STATE, TRUE)) { id = crm_element_value(node_state, XML_ATTR_ID); lrm_rscs = find_xml_node(node_state, XML_CIB_TAG_LRM, FALSE); lrm_rscs = find_xml_node(lrm_rscs, XML_LRM_TAG_RESOURCES, FALSE); node = pe_find_node_id(data_set->nodes, id); if (node == NULL) { continue; /* Still need to check actions for a maintenance node to cancel existing monitor operations */ } else if (can_run_resources(node) == FALSE && node->details->maintenance == FALSE) { crm_trace("Skipping param check for %s: can't run resources", node->details->uname); continue; } crm_trace("Processing node %s", node->details->uname); if (node->details->online || is_set(data_set->flags, pe_flag_stonith_enabled)) { xmlNode *rsc_entry = NULL; for (rsc_entry = __xml_first_child_element(lrm_rscs); rsc_entry != NULL; rsc_entry = __xml_next_element(rsc_entry)) { if (crm_str_eq((const char *)rsc_entry->name, XML_LRM_TAG_RESOURCE, TRUE)) { if (xml_has_children(rsc_entry)) { GListPtr gIter = NULL; GListPtr result = NULL; const char *rsc_id = ID(rsc_entry); CRM_CHECK(rsc_id != NULL, return); result = find_rsc_list(NULL, NULL, rsc_id, TRUE, FALSE, data_set); for (gIter = result; gIter != NULL; gIter = gIter->next) { resource_t *rsc = (resource_t *) gIter->data; if (rsc->variant != pe_native) { continue; } check_actions_for(rsc_entry, rsc, node, data_set); } g_list_free(result); } } } } } } } static void apply_placement_constraints(pe_working_set_t * data_set) { for (GList *gIter = data_set->placement_constraints; gIter != NULL; gIter = gIter->next) { pe__location_t *cons = gIter->data; cons->rsc_lh->cmds->rsc_location(cons->rsc_lh, cons); } } static gboolean failcount_clear_action_exists(node_t * node, resource_t * rsc) { gboolean rc = FALSE; GList *list = pe__resource_actions(rsc, node, CRM_OP_CLEAR_FAILCOUNT, TRUE); if (list) { rc = TRUE; } g_list_free(list); return rc; } /*! * \internal * \brief Force resource away if failures hit migration threshold * * \param[in,out] rsc Resource to check for failures * \param[in,out] node Node to check for failures * \param[in,out] data_set Cluster working set to update */ static void check_migration_threshold(resource_t *rsc, node_t *node, pe_working_set_t *data_set) { int fail_count, countdown; resource_t *failed; /* Migration threshold of 0 means never force away */ if (rsc->migration_threshold == 0) { return; } // If we're ignoring failures, also ignore the migration threshold if (is_set(rsc->flags, pe_rsc_failure_ignored)) { return; } /* If there are no failures, there's no need to force away */ fail_count = pe_get_failcount(node, rsc, NULL, pe_fc_effective|pe_fc_fillers, NULL, data_set); if (fail_count <= 0) { return; } /* How many more times recovery will be tried on this node */ countdown = QB_MAX(rsc->migration_threshold - fail_count, 0); /* If failed resource has a parent, we'll force the parent away */ failed = rsc; if (is_not_set(rsc->flags, pe_rsc_unique)) { failed = uber_parent(rsc); } if (countdown == 0) { resource_location(failed, node, -INFINITY, "__fail_limit__", data_set); crm_warn("Forcing %s away from %s after %d failures (max=%d)", failed->id, node->details->uname, fail_count, rsc->migration_threshold); } else { crm_info("%s can fail %d more times on %s before being forced off", failed->id, countdown, node->details->uname); } } static void common_apply_stickiness(resource_t * rsc, node_t * node, pe_working_set_t * data_set) { if (rsc->children) { GListPtr gIter = rsc->children; for (; gIter != NULL; gIter = gIter->next) { resource_t *child_rsc = (resource_t *) gIter->data; common_apply_stickiness(child_rsc, node, data_set); } return; } if (is_set(rsc->flags, pe_rsc_managed) && rsc->stickiness != 0 && pcmk__list_of_1(rsc->running_on)) { node_t *current = pe_find_node_id(rsc->running_on, node->details->id); node_t *match = pe_hash_table_lookup(rsc->allowed_nodes, node->details->id); if (current == NULL) { } else if (match != NULL || is_set(data_set->flags, pe_flag_symmetric_cluster)) { resource_t *sticky_rsc = rsc; resource_location(sticky_rsc, node, rsc->stickiness, "stickiness", data_set); pe_rsc_debug(sticky_rsc, "Resource %s: preferring current location" " (node=%s, weight=%d)", sticky_rsc->id, node->details->uname, rsc->stickiness); } else { GHashTableIter iter; node_t *nIter = NULL; pe_rsc_debug(rsc, "Ignoring stickiness for %s: the cluster is asymmetric" " and node %s is not explicitly allowed", rsc->id, node->details->uname); g_hash_table_iter_init(&iter, rsc->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (void **)&nIter)) { crm_err("%s[%s] = %d", rsc->id, nIter->details->uname, nIter->weight); } } } /* Check the migration threshold only if a failcount clear action * has not already been placed for this resource on the node. * There is no sense in potentially forcing the resource from this * node if the failcount is being reset anyway. * * @TODO A clear_failcount operation can be scheduled in stage4() via * check_actions_for(), or in stage5() via check_params(). This runs in * stage2(), 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 next transition. */ if (failcount_clear_action_exists(node, rsc) == FALSE) { check_migration_threshold(rsc, node, data_set); } } void complex_set_cmds(resource_t * rsc) { GListPtr gIter = rsc->children; rsc->cmds = &resource_class_alloc_functions[rsc->variant]; for (; gIter != NULL; gIter = gIter->next) { resource_t *child_rsc = (resource_t *) gIter->data; complex_set_cmds(child_rsc); } } void set_alloc_actions(pe_working_set_t * data_set) { GListPtr gIter = data_set->resources; for (; gIter != NULL; gIter = gIter->next) { resource_t *rsc = (resource_t *) gIter->data; complex_set_cmds(rsc); } } static void calculate_system_health(gpointer gKey, gpointer gValue, gpointer user_data) { const char *key = (const char *)gKey; const char *value = (const char *)gValue; int *system_health = (int *)user_data; if (!gKey || !gValue || !user_data) { return; } if (crm_starts_with(key, "#health")) { int score; /* Convert the value into an integer */ score = char2score(value); /* Add it to the running total */ *system_health = merge_weights(score, *system_health); } } static gboolean apply_system_health(pe_working_set_t * data_set) { GListPtr gIter = NULL; const char *health_strategy = pe_pref(data_set->config_hash, "node-health-strategy"); int base_health = 0; if (health_strategy == NULL || safe_str_eq(health_strategy, "none")) { /* Prevent any accidental health -> score translation */ node_score_red = 0; node_score_yellow = 0; node_score_green = 0; return TRUE; } else if (safe_str_eq(health_strategy, "migrate-on-red")) { /* Resources on nodes which have health values of red are * weighted away from that node. */ node_score_red = -INFINITY; node_score_yellow = 0; node_score_green = 0; } else if (safe_str_eq(health_strategy, "only-green")) { /* Resources on nodes which have health values of red or yellow * are forced away from that node. */ node_score_red = -INFINITY; node_score_yellow = -INFINITY; node_score_green = 0; } else if (safe_str_eq(health_strategy, "progressive")) { /* Same as the above, but use the r/y/g scores provided by the user * Defaults are provided by the pe_prefs table * Also, custom health "base score" can be used */ base_health = crm_parse_int(pe_pref(data_set->config_hash, "node-health-base"), "0"); } else if (safe_str_eq(health_strategy, "custom")) { /* Requires the admin to configure the rsc_location constaints for * processing the stored health scores */ /* TODO: Check for the existence of appropriate node health constraints */ return TRUE; } else { crm_err("Unknown node health strategy: %s", health_strategy); return FALSE; } crm_info("Applying automated node health strategy: %s", health_strategy); for (gIter = data_set->nodes; gIter != NULL; gIter = gIter->next) { int system_health = base_health; node_t *node = (node_t *) gIter->data; /* Search through the node hash table for system health entries. */ g_hash_table_foreach(node->details->attrs, calculate_system_health, &system_health); crm_info(" Node %s has an combined system health of %d", node->details->uname, system_health); /* If the health is non-zero, then create a new rsc2node so that the * weight will be added later on. */ if (system_health != 0) { GListPtr gIter2 = data_set->resources; for (; gIter2 != NULL; gIter2 = gIter2->next) { resource_t *rsc = (resource_t *) gIter2->data; rsc2node_new(health_strategy, rsc, system_health, NULL, node, data_set); } } } return TRUE; } gboolean stage0(pe_working_set_t * data_set) { xmlNode *cib_constraints = get_object_root(XML_CIB_TAG_CONSTRAINTS, data_set->input); if (data_set->input == NULL) { return FALSE; } if (is_set(data_set->flags, pe_flag_have_status) == FALSE) { crm_trace("Calculating status"); cluster_status(data_set); } set_alloc_actions(data_set); apply_system_health(data_set); unpack_constraints(cib_constraints, data_set); return TRUE; } /* * Check nodes for resources started outside of the LRM */ gboolean probe_resources(pe_working_set_t * data_set) { action_t *probe_node_complete = NULL; for (GListPtr gIter = data_set->nodes; gIter != NULL; gIter = gIter->next) { node_t *node = (node_t *) gIter->data; const char *probed = pe_node_attribute_raw(node, CRM_OP_PROBED); if (node->details->online == FALSE) { if (pe__is_remote_node(node) && node->details->remote_rsc && (get_remote_node_state(node) == remote_state_failed)) { pe_fence_node(data_set, node, "the connection is unrecoverable"); } continue; } else if (node->details->unclean) { continue; } else if (node->details->rsc_discovery_enabled == FALSE) { /* resource discovery is disabled for this node */ continue; } if (probed != NULL && crm_is_true(probed) == FALSE) { action_t *probe_op = custom_action(NULL, crm_strdup_printf("%s-%s", CRM_OP_REPROBE, node->details->uname), CRM_OP_REPROBE, node, FALSE, TRUE, data_set); add_hash_param(probe_op->meta, XML_ATTR_TE_NOWAIT, XML_BOOLEAN_TRUE); continue; } for (GListPtr gIter2 = data_set->resources; gIter2 != NULL; gIter2 = gIter2->next) { resource_t *rsc = (resource_t *) gIter2->data; rsc->cmds->create_probe(rsc, node, probe_node_complete, FALSE, data_set); } } return TRUE; } static void rsc_discover_filter(resource_t *rsc, node_t *node) { GListPtr gIter = rsc->children; resource_t *top = uber_parent(rsc); node_t *match; if (rsc->exclusive_discover == FALSE && top->exclusive_discover == FALSE) { return; } for (; gIter != NULL; gIter = gIter->next) { resource_t *child_rsc = (resource_t *) gIter->data; rsc_discover_filter(child_rsc, node); } match = g_hash_table_lookup(rsc->allowed_nodes, node->details->id); if (match && match->rsc_discover_mode != pe_discover_exclusive) { match->weight = -INFINITY; } } static time_t shutdown_time(pe_node_t *node, pe_working_set_t *data_set) { const char *shutdown = pe_node_attribute_raw(node, XML_CIB_ATTR_SHUTDOWN); time_t result = 0; if (shutdown) { errno = 0; result = (time_t) crm_int_helper(shutdown, NULL); if (errno != 0) { result = 0; } } return result? result : get_effective_time(data_set); } static void apply_shutdown_lock(pe_resource_t *rsc, pe_working_set_t *data_set) { const char *class; // Only primitives and (uncloned) groups may be locked if (rsc->variant == pe_group) { for (GList *item = rsc->children; item != NULL; item = item->next) { apply_shutdown_lock((pe_resource_t *) item->data, data_set); } } else if (rsc->variant != pe_native) { return; } // Fence devices and remote connections can't be locked class = crm_element_value(rsc->xml, XML_AGENT_ATTR_CLASS); if ((class == NULL) || !strcmp(class, PCMK_RESOURCE_CLASS_STONITH) || pe__resource_is_remote_conn(rsc, data_set)) { 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, data_set); 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, node->details->uname); } else { rsc->lock_node = node; rsc->lock_time = shutdown_time(node, data_set); } } } if (rsc->lock_node == NULL) { // No lock needed return; } if (data_set->shutdown_lock > 0) { time_t lock_expiration = rsc->lock_time + data_set->shutdown_lock; pe_rsc_info(rsc, "Locking %s to %s due to shutdown (expires @%lld)", rsc->id, rsc->lock_node->details->uname, (long long) lock_expiration); pe__update_recheck_time(++lock_expiration, data_set); } else { pe_rsc_info(rsc, "Locking %s to %s due to shutdown", rsc->id, rsc->lock_node->details->uname); } // If resource is locked to one node, ban it from all other nodes for (GList *item = data_set->nodes; item != NULL; item = item->next) { pe_node_t *node = item->data; if (strcmp(node->details->uname, rsc->lock_node->details->uname)) { resource_location(rsc, node, -CRM_SCORE_INFINITY, XML_CONFIG_ATTR_SHUTDOWN_LOCK, data_set); } } } /* * \internal * \brief Stage 2 of cluster status: apply node-specific criteria * * Count known nodes, and apply location constraints, stickiness, and exclusive * resource discovery. */ gboolean stage2(pe_working_set_t * data_set) { GListPtr gIter = NULL; if (is_set(data_set->flags, pe_flag_shutdown_lock)) { for (gIter = data_set->resources; gIter != NULL; gIter = gIter->next) { apply_shutdown_lock((pe_resource_t *) gIter->data, data_set); } } for (gIter = data_set->nodes; gIter != NULL; gIter = gIter->next) { node_t *node = (node_t *) gIter->data; if (node == NULL) { /* error */ } else if (node->weight >= 0.0 /* global weight */ && node->details->online && node->details->type != node_ping) { data_set->max_valid_nodes++; } } apply_placement_constraints(data_set); gIter = data_set->nodes; for (; gIter != NULL; gIter = gIter->next) { GListPtr gIter2 = NULL; node_t *node = (node_t *) gIter->data; gIter2 = data_set->resources; for (; gIter2 != NULL; gIter2 = gIter2->next) { resource_t *rsc = (resource_t *) gIter2->data; common_apply_stickiness(rsc, node, data_set); rsc_discover_filter(rsc, node); } } return TRUE; } /* * Create internal resource constraints before allocation */ gboolean stage3(pe_working_set_t * data_set) { GListPtr gIter = data_set->resources; for (; gIter != NULL; gIter = gIter->next) { resource_t *rsc = (resource_t *) gIter->data; rsc->cmds->internal_constraints(rsc, data_set); } return TRUE; } /* * Check for orphaned or redefined actions */ gboolean stage4(pe_working_set_t * data_set) { check_actions(data_set); return TRUE; } static void * convert_const_pointer(const void *ptr) { /* Worst function ever */ return (void *)ptr; } static gint sort_rsc_process_order(gconstpointer a, gconstpointer b, gpointer data) { int rc = 0; int r1_weight = -INFINITY; int r2_weight = -INFINITY; const char *reason = "existence"; const GListPtr nodes = (GListPtr) data; const resource_t *resource1 = a; const resource_t *resource2 = b; node_t *r1_node = NULL; node_t *r2_node = NULL; GListPtr gIter = NULL; GHashTable *r1_nodes = NULL; GHashTable *r2_nodes = NULL; if (a == NULL && b == NULL) { goto done; } if (a == NULL) { return 1; } if (b == NULL) { return -1; } reason = "priority"; r1_weight = resource1->priority; r2_weight = resource2->priority; if (r1_weight > r2_weight) { rc = -1; goto done; } if (r1_weight < r2_weight) { rc = 1; goto done; } reason = "no node list"; if (nodes == NULL) { goto done; } r1_nodes = pcmk__native_merge_weights(convert_const_pointer(resource1), resource1->id, NULL, NULL, 1, pe_weights_forward | pe_weights_init); pe__show_node_weights(true, NULL, resource1->id, r1_nodes); r2_nodes = pcmk__native_merge_weights(convert_const_pointer(resource2), resource2->id, NULL, NULL, 1, pe_weights_forward | pe_weights_init); pe__show_node_weights(true, NULL, resource2->id, r2_nodes); /* Current location score */ reason = "current location"; r1_weight = -INFINITY; r2_weight = -INFINITY; if (resource1->running_on) { r1_node = pe__current_node(resource1); r1_node = g_hash_table_lookup(r1_nodes, r1_node->details->id); if (r1_node != NULL) { r1_weight = r1_node->weight; } } if (resource2->running_on) { r2_node = pe__current_node(resource2); r2_node = g_hash_table_lookup(r2_nodes, r2_node->details->id); if (r2_node != NULL) { r2_weight = r2_node->weight; } } if (r1_weight > r2_weight) { rc = -1; goto done; } if (r1_weight < r2_weight) { rc = 1; goto done; } reason = "score"; for (gIter = nodes; gIter != NULL; gIter = gIter->next) { node_t *node = (node_t *) gIter->data; r1_node = NULL; r2_node = NULL; r1_weight = -INFINITY; if (r1_nodes) { r1_node = g_hash_table_lookup(r1_nodes, node->details->id); } if (r1_node) { r1_weight = r1_node->weight; } r2_weight = -INFINITY; if (r2_nodes) { r2_node = g_hash_table_lookup(r2_nodes, node->details->id); } if (r2_node) { r2_weight = r2_node->weight; } if (r1_weight > r2_weight) { rc = -1; goto done; } if (r1_weight < r2_weight) { rc = 1; goto done; } } done: crm_trace("%s (%d) on %s %c %s (%d) on %s: %s", resource1->id, r1_weight, r1_node ? r1_node->details->id : "n/a", rc < 0 ? '>' : rc > 0 ? '<' : '=', resource2->id, r2_weight, r2_node ? r2_node->details->id : "n/a", reason); if (r1_nodes) { g_hash_table_destroy(r1_nodes); } if (r2_nodes) { g_hash_table_destroy(r2_nodes); } return rc; } static void allocate_resources(pe_working_set_t * data_set) { GListPtr gIter = NULL; if (is_set(data_set->flags, pe_flag_have_remote_nodes)) { - /* Force remote connection resources to be allocated first. This - * also forces any colocation dependencies to be allocated as well */ + /* Allocate remote connection resources first (which will also allocate + * any colocation dependencies). If the connection is migrating, always + * prefer the partial migration target. + */ for (gIter = data_set->resources; gIter != NULL; gIter = gIter->next) { resource_t *rsc = (resource_t *) gIter->data; if (rsc->is_remote_node == FALSE) { continue; } - pe_rsc_trace(rsc, "Allocating: %s", rsc->id); - /* For remote node connection resources, always prefer the partial - * migration target during resource allocation, if the rsc is in the - * middle of a migration. - */ + 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 (gIter = data_set->resources; gIter != NULL; gIter = gIter->next) { resource_t *rsc = (resource_t *) gIter->data; if (rsc->is_remote_node == TRUE) { continue; } - pe_rsc_trace(rsc, "Allocating: %s", rsc->id); + pe_rsc_trace(rsc, "Allocating resource '%s'", rsc->id); rsc->cmds->allocate(rsc, NULL, data_set); } } /* We always use pe_order_preserve with these convenience functions to exempt * internally generated constraints from the prohibition of user constraints * involving remote connection resources. * * The start ordering additionally uses pe_order_runnable_left so that the * specified action is not runnable if the start is not runnable. */ static inline void order_start_then_action(resource_t *lh_rsc, action_t *rh_action, enum pe_ordering extra, pe_working_set_t *data_set) { if (lh_rsc && rh_action && data_set) { custom_action_order(lh_rsc, start_key(lh_rsc), NULL, rh_action->rsc, NULL, rh_action, pe_order_preserve | pe_order_runnable_left | extra, data_set); } } static inline void order_action_then_stop(action_t *lh_action, resource_t *rh_rsc, enum pe_ordering extra, pe_working_set_t *data_set) { if (lh_action && rh_rsc && data_set) { custom_action_order(lh_action->rsc, NULL, lh_action, rh_rsc, stop_key(rh_rsc), NULL, pe_order_preserve | extra, data_set); } } // Clear fail counts for orphaned rsc on all online nodes static void cleanup_orphans(resource_t * rsc, pe_working_set_t * data_set) { GListPtr gIter = NULL; for (gIter = data_set->nodes; gIter != NULL; gIter = gIter->next) { node_t *node = (node_t *) gIter->data; if (node->details->online && pe_get_failcount(node, rsc, NULL, pe_fc_effective, NULL, data_set)) { pe_action_t *clear_op = NULL; 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 */ custom_action_order(clear_op->rsc, NULL, clear_op, rsc, stop_key(rsc), NULL, pe_order_optional, data_set); } } } gboolean stage5(pe_working_set_t * data_set) { GListPtr gIter = NULL; int log_prio = show_utilization? LOG_STDOUT : utilization_log_level; if (safe_str_neq(data_set->placement_strategy, "default")) { GListPtr nodes = g_list_copy(data_set->nodes); nodes = sort_nodes_by_weight(nodes, NULL, data_set); data_set->resources = g_list_sort_with_data(data_set->resources, sort_rsc_process_order, nodes); g_list_free(nodes); } gIter = data_set->nodes; for (; gIter != NULL; gIter = gIter->next) { node_t *node = (node_t *) gIter->data; dump_node_capacity(log_prio, "Original", node); } crm_trace("Allocating services"); /* Take (next) highest resource, assign it and create its actions */ allocate_resources(data_set); gIter = data_set->nodes; for (; gIter != NULL; gIter = gIter->next) { node_t *node = (node_t *) gIter->data; dump_node_capacity(log_prio, "Remaining", node); } // Process deferred action checks pe__foreach_param_check(data_set, check_params); pe__free_param_checks(data_set); if (is_set(data_set->flags, pe_flag_startup_probes)) { crm_trace("Calculating needed probes"); /* This code probably needs optimization * ptest -x with 100 nodes, 100 clones and clone-max=100: With probes: ptest[14781]: 2010/09/27_17:56:46 notice: TRACE: do_calculations: pengine.c:258 Calculate cluster status ptest[14781]: 2010/09/27_17:56:46 notice: TRACE: do_calculations: pengine.c:278 Applying placement constraints ptest[14781]: 2010/09/27_17:56:47 notice: TRACE: do_calculations: pengine.c:285 Create internal constraints ptest[14781]: 2010/09/27_17:56:47 notice: TRACE: do_calculations: pengine.c:292 Check actions ptest[14781]: 2010/09/27_17:56:48 notice: TRACE: do_calculations: pengine.c:299 Allocate resources ptest[14781]: 2010/09/27_17:56:48 notice: TRACE: stage5: allocate.c:881 Allocating services ptest[14781]: 2010/09/27_17:56:49 notice: TRACE: stage5: allocate.c:894 Calculating needed probes ptest[14781]: 2010/09/27_17:56:51 notice: TRACE: stage5: allocate.c:899 Creating actions ptest[14781]: 2010/09/27_17:56:52 notice: TRACE: stage5: allocate.c:905 Creating done ptest[14781]: 2010/09/27_17:56:52 notice: TRACE: do_calculations: pengine.c:306 Processing fencing and shutdown cases ptest[14781]: 2010/09/27_17:56:52 notice: TRACE: do_calculations: pengine.c:313 Applying ordering constraints 36s ptest[14781]: 2010/09/27_17:57:28 notice: TRACE: do_calculations: pengine.c:320 Create transition graph Without probes: ptest[14637]: 2010/09/27_17:56:21 notice: TRACE: do_calculations: pengine.c:258 Calculate cluster status ptest[14637]: 2010/09/27_17:56:22 notice: TRACE: do_calculations: pengine.c:278 Applying placement constraints ptest[14637]: 2010/09/27_17:56:22 notice: TRACE: do_calculations: pengine.c:285 Create internal constraints ptest[14637]: 2010/09/27_17:56:22 notice: TRACE: do_calculations: pengine.c:292 Check actions ptest[14637]: 2010/09/27_17:56:23 notice: TRACE: do_calculations: pengine.c:299 Allocate resources ptest[14637]: 2010/09/27_17:56:23 notice: TRACE: stage5: allocate.c:881 Allocating services ptest[14637]: 2010/09/27_17:56:24 notice: TRACE: stage5: allocate.c:899 Creating actions ptest[14637]: 2010/09/27_17:56:25 notice: TRACE: stage5: allocate.c:905 Creating done ptest[14637]: 2010/09/27_17:56:25 notice: TRACE: do_calculations: pengine.c:306 Processing fencing and shutdown cases ptest[14637]: 2010/09/27_17:56:25 notice: TRACE: do_calculations: pengine.c:313 Applying ordering constraints ptest[14637]: 2010/09/27_17:56:25 notice: TRACE: do_calculations: pengine.c:320 Create transition graph */ probe_resources(data_set); } crm_trace("Handle orphans"); if (is_set(data_set->flags, pe_flag_stop_rsc_orphans)) { for (gIter = data_set->resources; gIter != NULL; gIter = gIter->next) { pe_resource_t *rsc = (pe_resource_t *) gIter->data; /* There's no need to recurse into rsc->children because those * should just be unallocated clone instances. */ if (is_set(rsc->flags, pe_rsc_orphan)) { cleanup_orphans(rsc, data_set); } } } crm_trace("Creating actions"); for (gIter = data_set->resources; gIter != NULL; gIter = gIter->next) { resource_t *rsc = (resource_t *) gIter->data; rsc->cmds->create_actions(rsc, data_set); } crm_trace("Creating done"); return TRUE; } static gboolean is_managed(const resource_t * rsc) { GListPtr gIter = rsc->children; if (is_set(rsc->flags, pe_rsc_managed)) { return TRUE; } for (; gIter != NULL; gIter = gIter->next) { resource_t *child_rsc = (resource_t *) gIter->data; if (is_managed(child_rsc)) { return TRUE; } } return FALSE; } static gboolean any_managed_resources(pe_working_set_t * data_set) { GListPtr gIter = data_set->resources; for (; gIter != NULL; gIter = gIter->next) { resource_t *rsc = (resource_t *) gIter->data; if (is_managed(rsc)) { return TRUE; } } return FALSE; } /*! * \internal * \brief Create pseudo-op for guest node fence, and order relative to it * * \param[in] node Guest node to fence * \param[in] data_set Working set of CIB state */ static void fence_guest(pe_node_t *node, pe_working_set_t *data_set) { resource_t *container = node->details->remote_rsc->container; pe_action_t *stop = NULL; pe_action_t *stonith_op = NULL; /* The fence action is just a label; we don't do anything differently for * off vs. reboot. We specify it explicitly, rather than let it default to * cluster's default action, because we are not _initiating_ fencing -- we * are creating a pseudo-event to describe fencing that is already occurring * by other means (container recovery). */ const char *fence_action = "off"; /* Check whether guest's container resource has any explicit stop or * start (the stop may be implied by fencing of the guest's host). */ if (container) { stop = find_first_action(container->actions, NULL, CRMD_ACTION_STOP, NULL); if (find_first_action(container->actions, NULL, CRMD_ACTION_START, NULL)) { fence_action = "reboot"; } } /* Create a fence pseudo-event, so we have an event to order actions * against, and the controller can always detect it. */ stonith_op = pe_fence_op(node, fence_action, FALSE, "guest is unclean", data_set); update_action_flags(stonith_op, pe_action_pseudo | pe_action_runnable, __FUNCTION__, __LINE__); /* We want to imply stops/demotes after the guest is stopped, not wait until * it is restarted, so we always order pseudo-fencing after stop, not start * (even though start might be closer to what is done for a real reboot). */ if(stop && is_set(stop->flags, pe_action_pseudo)) { pe_action_t *parent_stonith_op = pe_fence_op(stop->node, NULL, FALSE, NULL, data_set); crm_info("Implying guest node %s is down (action %d) after %s fencing", node->details->uname, stonith_op->id, stop->node->details->uname); order_actions(parent_stonith_op, stonith_op, pe_order_runnable_left|pe_order_implies_then); } else if (stop) { order_actions(stop, stonith_op, pe_order_runnable_left|pe_order_implies_then); crm_info("Implying guest node %s is down (action %d) " "after container %s is stopped (action %d)", node->details->uname, stonith_op->id, container->id, stop->id); } else { /* If we're fencing the guest node but there's no stop for the guest * resource, we must think the guest is already stopped. However, we may * think so because its resource history was just cleaned. To avoid * unnecessarily considering the guest node down if it's really up, * order the pseudo-fencing after any stop of the connection resource, * which will be ordered after any container (re-)probe. */ stop = find_first_action(node->details->remote_rsc->actions, NULL, RSC_STOP, NULL); if (stop) { order_actions(stop, stonith_op, pe_order_optional); crm_info("Implying guest node %s is down (action %d) " "after connection is stopped (action %d)", node->details->uname, stonith_op->id, stop->id); } else { /* Not sure why we're fencing, but everything must already be * cleanly stopped. */ crm_info("Implying guest node %s is down (action %d) ", node->details->uname, stonith_op->id); } } /* Order/imply other actions relative to pseudo-fence as with real fence */ pcmk__order_vs_fence(stonith_op, data_set); } /* * Create dependencies for stonith and shutdown operations */ gboolean stage6(pe_working_set_t * data_set) { action_t *dc_down = NULL; action_t *stonith_op = NULL; gboolean integrity_lost = FALSE; gboolean need_stonith = TRUE; GListPtr gIter; GListPtr stonith_ops = NULL; GList *shutdown_ops = NULL; /* Remote ordering constraints need to happen prior to calculating fencing * because it is one more place we will mark the node as dirty. * * A nice side effect of doing them early is that apply_*_ordering() can be * simpler because pe_fence_node() has already done some of the work. */ crm_trace("Creating remote ordering constraints"); apply_remote_node_ordering(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) { node_t *node = (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)) { 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", 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 (is_not_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) { action_t *down_op = sched_shutdown_op(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 (is_set(data_set->flags, pe_flag_stonith_enabled) == FALSE) { pe_warn("YOUR RESOURCES ARE NOW LIKELY COMPROMISED"); pe_err("ENABLE STONITH TO KEEP YOUR RESOURCES SAFE"); } else if (is_set(data_set->flags, pe_flag_have_quorum) == FALSE) { 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 (safe_str_eq(dc_down->task, CRM_OP_SHUTDOWN)) { for (gIter = shutdown_ops; gIter != NULL; gIter = gIter->next) { action_t *node_stop = (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 (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; } /* * Determine the sets of independent actions and the correct order for the * actions in each set. * * Mark dependencies of un-runnable actions un-runnable * */ static GListPtr find_actions_by_task(GListPtr actions, resource_t * rsc, const char *original_key) { GListPtr list = NULL; list = find_actions(actions, original_key, NULL); if (list == NULL) { /* we're potentially searching a child of the original resource */ char *key = NULL; char *task = NULL; guint interval_ms = 0; if (parse_op_key(original_key, NULL, &task, &interval_ms)) { key = pcmk__op_key(rsc->id, task, interval_ms); list = find_actions(actions, key, NULL); } else { crm_err("search key: %s", original_key); } free(key); free(task); } return list; } static void rsc_order_then(pe_action_t *lh_action, pe_resource_t *rsc, pe__ordering_t *order) { GListPtr gIter = NULL; GListPtr rh_actions = NULL; action_t *rh_action = NULL; enum pe_ordering type; CRM_CHECK(rsc != NULL, return); CRM_CHECK(order != NULL, return); type = order->type; rh_action = order->rh_action; crm_trace("Processing RH of ordering constraint %d", order->id); if (rh_action != NULL) { rh_actions = g_list_prepend(NULL, rh_action); } else if (rsc != NULL) { rh_actions = find_actions_by_task(rsc->actions, rsc, order->rh_action_task); } if (rh_actions == NULL) { pe_rsc_trace(rsc, "No RH-Side (%s/%s) found for constraint..." " ignoring", rsc->id, order->rh_action_task); if (lh_action) { pe_rsc_trace(rsc, "LH-Side was: %s", lh_action->uuid); } return; } if (lh_action && lh_action->rsc == rsc && is_set(lh_action->flags, pe_action_dangle)) { pe_rsc_trace(rsc, "Detected dangling operation %s -> %s", lh_action->uuid, order->rh_action_task); clear_bit(type, pe_order_implies_then); } gIter = rh_actions; for (; gIter != NULL; gIter = gIter->next) { action_t *rh_action_iter = (action_t *) gIter->data; if (lh_action) { order_actions(lh_action, rh_action_iter, type); } else if (type & pe_order_implies_then) { update_action_flags(rh_action_iter, pe_action_runnable | pe_action_clear, __FUNCTION__, __LINE__); crm_warn("Unrunnable %s 0x%.6x", rh_action_iter->uuid, type); } else { crm_warn("neither %s 0x%.6x", rh_action_iter->uuid, type); } } g_list_free(rh_actions); } static void rsc_order_first(pe_resource_t *lh_rsc, pe__ordering_t *order, pe_working_set_t *data_set) { GListPtr gIter = NULL; GListPtr lh_actions = NULL; action_t *lh_action = order->lh_action; resource_t *rh_rsc = order->rh_rsc; crm_trace("Processing LH of ordering constraint %d", order->id); CRM_ASSERT(lh_rsc != NULL); if (lh_action != NULL) { lh_actions = g_list_prepend(NULL, lh_action); } else { lh_actions = find_actions_by_task(lh_rsc->actions, lh_rsc, order->lh_action_task); } if (lh_actions == NULL && lh_rsc != rh_rsc) { char *key = NULL; char *op_type = NULL; guint interval_ms = 0; parse_op_key(order->lh_action_task, NULL, &op_type, &interval_ms); key = pcmk__op_key(lh_rsc->id, op_type, interval_ms); if (lh_rsc->fns->state(lh_rsc, TRUE) == RSC_ROLE_STOPPED && safe_str_eq(op_type, RSC_STOP)) { free(key); pe_rsc_trace(lh_rsc, "No LH-Side (%s/%s) found for constraint %d with %s - ignoring", lh_rsc->id, order->lh_action_task, order->id, order->rh_action_task); } else if (lh_rsc->fns->state(lh_rsc, TRUE) == RSC_ROLE_SLAVE && safe_str_eq(op_type, RSC_DEMOTE)) { free(key); pe_rsc_trace(lh_rsc, "No LH-Side (%s/%s) found for constraint %d with %s - ignoring", lh_rsc->id, order->lh_action_task, order->id, order->rh_action_task); } else { pe_rsc_trace(lh_rsc, "No LH-Side (%s/%s) found for constraint %d with %s - creating", lh_rsc->id, order->lh_action_task, order->id, order->rh_action_task); lh_action = custom_action(lh_rsc, key, op_type, NULL, TRUE, TRUE, data_set); lh_actions = g_list_prepend(NULL, lh_action); } free(op_type); } gIter = lh_actions; for (; gIter != NULL; gIter = gIter->next) { action_t *lh_action_iter = (action_t *) gIter->data; if (rh_rsc == NULL && order->rh_action) { rh_rsc = order->rh_action->rsc; } if (rh_rsc) { rsc_order_then(lh_action_iter, rh_rsc, order); } else if (order->rh_action) { order_actions(lh_action_iter, order->rh_action, order->type); } } g_list_free(lh_actions); } extern void update_colo_start_chain(pe_action_t *action, pe_working_set_t *data_set); static int is_recurring_action(action_t *action) { const char *interval_ms_s = g_hash_table_lookup(action->meta, XML_LRM_ATTR_INTERVAL_MS); guint interval_ms = crm_parse_ms(interval_ms_s); return (interval_ms > 0); } static void apply_container_ordering(action_t *action, pe_working_set_t *data_set) { /* VMs are also classified as containers for these purposes... in * that they both involve a 'thing' running on a real or remote * cluster node. * * This allows us to be smarter about the type and extent of * recovery actions required in various scenarios */ resource_t *remote_rsc = NULL; resource_t *container = NULL; enum action_tasks task = text2task(action->task); CRM_ASSERT(action->rsc); CRM_ASSERT(action->node); CRM_ASSERT(pe__is_guest_or_remote_node(action->node)); remote_rsc = action->node->details->remote_rsc; CRM_ASSERT(remote_rsc); container = remote_rsc->container; CRM_ASSERT(container); if(is_set(container->flags, pe_rsc_failed)) { pe_fence_node(data_set, action->node, "container failed"); } crm_trace("Order %s action %s relative to %s%s for %s%s", action->task, action->uuid, is_set(remote_rsc->flags, pe_rsc_failed)? "failed " : "", remote_rsc->id, is_set(container->flags, pe_rsc_failed)? "failed " : "", container->id); if (safe_str_eq(action->task, CRMD_ACTION_MIGRATE) || safe_str_eq(action->task, CRMD_ACTION_MIGRATED)) { /* Migration ops map to "no_action", but we need to apply the same * ordering as for stop or demote (see get_router_node()). */ task = stop_rsc; } switch (task) { case start_rsc: case action_promote: /* Force resource recovery if the container is recovered */ order_start_then_action(container, action, pe_order_implies_then, data_set); /* Wait for the connection resource to be up too */ order_start_then_action(remote_rsc, action, pe_order_none, data_set); break; case stop_rsc: case action_demote: if (is_set(container->flags, pe_rsc_failed)) { /* When the container representing a guest node fails, any stop * or demote actions for resources running on the guest node * are implied by the container stopping. This is similar to * how fencing operations work for cluster nodes and remote * nodes. */ } else { /* Ensure the operation happens before the connection is brought * down. * * If we really wanted to, we could order these after the * connection start, IFF the container's current role was * stopped (otherwise we re-introduce an ordering loop when the * connection is restarting). */ order_action_then_stop(action, remote_rsc, pe_order_none, data_set); } break; default: /* Wait for the connection resource to be up */ if (is_recurring_action(action)) { /* In case we ever get the recovery logic wrong, force * recurring monitors to be restarted, even if just * the connection was re-established */ if(task != no_action) { order_start_then_action(remote_rsc, action, pe_order_implies_then, data_set); } } else { order_start_then_action(remote_rsc, action, pe_order_none, data_set); } break; } } static enum remote_connection_state get_remote_node_state(pe_node_t *node) { resource_t *remote_rsc = NULL; node_t *cluster_node = NULL; CRM_ASSERT(node); remote_rsc = node->details->remote_rsc; CRM_ASSERT(remote_rsc); cluster_node = pe__current_node(remote_rsc); /* If the cluster node the remote connection resource resides on * is unclean or went offline, we can't process any operations * on that remote node until after it starts elsewhere. */ if(remote_rsc->next_role == RSC_ROLE_STOPPED || remote_rsc->allocated_to == NULL) { /* The connection resource is not going to run anywhere */ if (cluster_node && cluster_node->details->unclean) { /* The remote connection is failed because its resource is on a * failed node and can't be recovered elsewhere, so we must fence. */ return remote_state_failed; } if (is_not_set(remote_rsc->flags, pe_rsc_failed)) { /* Connection resource is cleanly stopped */ return remote_state_stopped; } /* Connection resource is failed */ if ((remote_rsc->next_role == RSC_ROLE_STOPPED) && remote_rsc->remote_reconnect_ms && node->details->remote_was_fenced && !pe__shutdown_requested(node)) { /* We won't know whether the connection is recoverable until the * reconnect interval expires and we reattempt connection. */ return remote_state_unknown; } /* The remote connection is in a failed state. If there are any * resources known to be active on it (stop) or in an unknown state * (probe), we must assume the worst and fence it. */ return remote_state_failed; } else if (cluster_node == NULL) { /* Connection is recoverable but not currently running anywhere, see if we can recover it first */ return remote_state_unknown; } else if(cluster_node->details->unclean == TRUE || cluster_node->details->online == FALSE) { /* Connection is running on a dead node, see if we can recover it first */ return remote_state_resting; } else if (pcmk__list_of_multiple(remote_rsc->running_on) && remote_rsc->partial_migration_source && remote_rsc->partial_migration_target) { /* We're in the middle of migrating a connection resource, * wait until after the resource migrates before performing * any actions. */ return remote_state_resting; } return remote_state_alive; } /*! * \internal * \brief Order actions on remote node relative to actions for the connection */ static void apply_remote_ordering(action_t *action, pe_working_set_t *data_set) { resource_t *remote_rsc = NULL; enum action_tasks task = text2task(action->task); enum remote_connection_state state = get_remote_node_state(action->node); enum pe_ordering order_opts = pe_order_none; if (action->rsc == NULL) { return; } CRM_ASSERT(action->node); CRM_ASSERT(pe__is_guest_or_remote_node(action->node)); remote_rsc = action->node->details->remote_rsc; CRM_ASSERT(remote_rsc); crm_trace("Order %s action %s relative to %s%s (state: %s)", action->task, action->uuid, is_set(remote_rsc->flags, pe_rsc_failed)? "failed " : "", remote_rsc->id, state2text(state)); if (safe_str_eq(action->task, CRMD_ACTION_MIGRATE) || safe_str_eq(action->task, CRMD_ACTION_MIGRATED)) { /* Migration ops map to "no_action", but we need to apply the same * ordering as for stop or demote (see get_router_node()). */ task = stop_rsc; } switch (task) { case start_rsc: case action_promote: order_opts = pe_order_none; if (state == remote_state_failed) { /* Force recovery, by making this action required */ order_opts |= pe_order_implies_then; } /* Ensure connection is up before running this action */ order_start_then_action(remote_rsc, action, order_opts, data_set); break; case stop_rsc: if(state == remote_state_alive) { order_action_then_stop(action, remote_rsc, pe_order_implies_first, data_set); } else if(state == remote_state_failed) { /* The resource is active on the node, but since we don't have a * valid connection, the only way to stop the resource is by * fencing the node. There is no need to order the stop relative * to the remote connection, since the stop will become implied * by the fencing. */ pe_fence_node(data_set, action->node, "resources are active and the connection is unrecoverable"); } else if(remote_rsc->next_role == RSC_ROLE_STOPPED) { /* State must be remote_state_unknown or remote_state_stopped. * Since the connection is not coming back up in this * transition, stop this resource first. */ order_action_then_stop(action, remote_rsc, pe_order_implies_first, data_set); } else { /* The connection is going to be started somewhere else, so * stop this resource after that completes. */ order_start_then_action(remote_rsc, action, pe_order_none, data_set); } break; case action_demote: /* Only order this demote relative to the connection start if the * connection isn't being torn down. Otherwise, the demote would be * blocked because the connection start would not be allowed. */ if(state == remote_state_resting || state == remote_state_unknown) { order_start_then_action(remote_rsc, action, pe_order_none, data_set); } /* Otherwise we can rely on the stop ordering */ break; default: /* Wait for the connection resource to be up */ if (is_recurring_action(action)) { /* In case we ever get the recovery logic wrong, force * recurring monitors to be restarted, even if just * the connection was re-established */ order_start_then_action(remote_rsc, action, pe_order_implies_then, data_set); } else { node_t *cluster_node = pe__current_node(remote_rsc); if(task == monitor_rsc && state == remote_state_failed) { /* We would only be here if we do not know the * state of the resource on the remote node. * Since we have no way to find out, it is * necessary to fence the node. */ pe_fence_node(data_set, action->node, "resources are in an unknown state and the connection is unrecoverable"); } if(cluster_node && state == remote_state_stopped) { /* The connection is currently up, but is going * down permanently. * * Make sure we check services are actually * stopped _before_ we let the connection get * closed */ order_action_then_stop(action, remote_rsc, pe_order_runnable_left, data_set); } else { order_start_then_action(remote_rsc, action, pe_order_none, data_set); } } break; } } static void apply_remote_node_ordering(pe_working_set_t *data_set) { if (is_set(data_set->flags, pe_flag_have_remote_nodes) == FALSE) { return; } for (GListPtr gIter = data_set->actions; gIter != NULL; gIter = gIter->next) { action_t *action = (action_t *) gIter->data; resource_t *remote = NULL; // We are only interested in resource actions if (action->rsc == NULL) { continue; } /* Special case: If we are clearing the failcount of an actual * remote connection resource, then make sure this happens before * any start of the resource in this transition. */ if (action->rsc->is_remote_node && safe_str_eq(action->task, CRM_OP_CLEAR_FAILCOUNT)) { custom_action_order(action->rsc, NULL, action, action->rsc, pcmk__op_key(action->rsc->id, RSC_START, 0), NULL, pe_order_optional, data_set); continue; } // We are only interested in actions allocated to a node if (action->node == NULL) { continue; } if (!pe__is_guest_or_remote_node(action->node)) { continue; } /* We are only interested in real actions. * * @TODO This is probably wrong; pseudo-actions might be converted to * real actions and vice versa later in update_actions() at the end of * stage7(). */ if (is_set(action->flags, pe_action_pseudo)) { continue; } remote = action->node->details->remote_rsc; if (remote == NULL) { // Orphaned continue; } /* Another special case: if a resource is moving to a Pacemaker Remote * node, order the stop on the original node after any start of the * remote connection. This ensures that if the connection fails to * start, we leave the resource running on the original node. */ if (safe_str_eq(action->task, RSC_START)) { for (GList *item = action->rsc->actions; item != NULL; item = item->next) { pe_action_t *rsc_action = item->data; if ((rsc_action->node->details != action->node->details) && safe_str_eq(rsc_action->task, RSC_STOP)) { custom_action_order(remote, start_key(remote), NULL, action->rsc, NULL, rsc_action, pe_order_optional, data_set); } } } /* The action occurs across a remote connection, so create * ordering constraints that guarantee the action occurs while the node * is active (after start, before stop ... things like that). * * This is somewhat brittle in that we need to make sure the results of * this ordering are compatible with the result of get_router_node(). * It would probably be better to add XML_LRM_ATTR_ROUTER_NODE as part * of this logic rather than action2xml(). */ if (remote->container) { crm_trace("Container ordering for %s", action->uuid); apply_container_ordering(action, data_set); } else { crm_trace("Remote ordering for %s", action->uuid); apply_remote_ordering(action, data_set); } } } static gboolean order_first_probe_unneeded(pe_action_t * probe, pe_action_t * rh_action) { /* No need to probe the resource on the node that is being * unfenced. Otherwise it might introduce transition loop * since probe will be performed after the node is * unfenced. */ if (safe_str_eq(rh_action->task, CRM_OP_FENCE) && probe->node && rh_action->node && probe->node->details == rh_action->node->details) { const char *op = g_hash_table_lookup(rh_action->meta, "stonith_action"); if (safe_str_eq(op, "on")) { return TRUE; } } // Shutdown waits for probe to complete only if it's on the same node if ((safe_str_eq(rh_action->task, CRM_OP_SHUTDOWN)) && probe->node && rh_action->node && probe->node->details != rh_action->node->details) { return TRUE; } return FALSE; } static void order_first_probes_imply_stops(pe_working_set_t * data_set) { GListPtr gIter = NULL; for (gIter = data_set->ordering_constraints; gIter != NULL; gIter = gIter->next) { pe__ordering_t *order = gIter->data; enum pe_ordering order_type = pe_order_optional; pe_resource_t *lh_rsc = order->lh_rsc; pe_resource_t *rh_rsc = order->rh_rsc; pe_action_t *lh_action = order->lh_action; pe_action_t *rh_action = order->rh_action; const char *lh_action_task = order->lh_action_task; const char *rh_action_task = order->rh_action_task; GListPtr probes = NULL; GListPtr rh_actions = NULL; GListPtr pIter = NULL; if (lh_rsc == NULL) { continue; } else if (rh_rsc && lh_rsc == rh_rsc) { continue; } if (lh_action == NULL && lh_action_task == NULL) { continue; } if (rh_action == NULL && rh_action_task == NULL) { continue; } /* Technically probe is expected to return "not running", which could be * the alternative of stop action if the status of the resource is * unknown yet. */ if (lh_action && safe_str_neq(lh_action->task, RSC_STOP)) { continue; } else if (lh_action == NULL && lh_action_task && crm_ends_with(lh_action_task, "_" RSC_STOP "_0") == FALSE) { continue; } /* Do not probe the resource inside of a stopping container. Otherwise * it might introduce transition loop since probe will be performed * after the container starts again. */ if (rh_rsc && lh_rsc->container == rh_rsc) { if (rh_action && safe_str_eq(rh_action->task, RSC_STOP)) { continue; } else if (rh_action == NULL && rh_action_task && crm_ends_with(rh_action_task,"_" RSC_STOP "_0")) { continue; } } if (order->type == pe_order_none) { continue; } // Preserve the order options for future filtering if (is_set(order->type, pe_order_apply_first_non_migratable)) { set_bit(order_type, pe_order_apply_first_non_migratable); } if (is_set(order->type, pe_order_same_node)) { set_bit(order_type, pe_order_same_node); } // Keep the order types for future filtering if (order->type == pe_order_anti_colocation || order->type == pe_order_load) { order_type = order->type; } probes = pe__resource_actions(lh_rsc, NULL, RSC_STATUS, FALSE); if (probes == NULL) { continue; } if (rh_action) { rh_actions = g_list_prepend(rh_actions, rh_action); } else if (rh_rsc && rh_action_task) { rh_actions = find_actions(rh_rsc->actions, rh_action_task, NULL); } if (rh_actions == NULL) { g_list_free(probes); continue; } crm_trace("Processing for LH probe based on ordering constraint %s -> %s" " (id=%d, type=%.6x)", lh_action ? lh_action->uuid : lh_action_task, rh_action ? rh_action->uuid : rh_action_task, order->id, order->type); for (pIter = probes; pIter != NULL; pIter = pIter->next) { pe_action_t *probe = (pe_action_t *) pIter->data; GListPtr rIter = NULL; for (rIter = rh_actions; rIter != NULL; rIter = rIter->next) { pe_action_t *rh_action_iter = (pe_action_t *) rIter->data; if (order_first_probe_unneeded(probe, rh_action_iter)) { continue; } order_actions(probe, rh_action_iter, order_type); } } g_list_free(rh_actions); g_list_free(probes); } } static void order_first_probe_then_restart_repromote(pe_action_t * probe, pe_action_t * after, pe_working_set_t * data_set) { GListPtr gIter = NULL; bool interleave = FALSE; pe_resource_t *compatible_rsc = NULL; if (probe == NULL || probe->rsc == NULL || probe->rsc->variant != pe_native) { return; } if (after == NULL // Avoid running into any possible loop || is_set(after->flags, pe_action_tracking)) { return; } if (safe_str_neq(probe->task, RSC_STATUS)) { return; } pe_set_action_bit(after, pe_action_tracking); crm_trace("Processing based on %s %s -> %s %s", probe->uuid, probe->node ? probe->node->details->uname: "", after->uuid, after->node ? after->node->details->uname : ""); if (after->rsc /* Better not build a dependency directly with a clone/group. * We are going to proceed through the ordering chain and build * dependencies with its children. */ && after->rsc->variant == pe_native && probe->rsc != after->rsc) { GListPtr then_actions = NULL; enum pe_ordering probe_order_type = pe_order_optional; if (safe_str_eq(after->task, RSC_START)) { then_actions = pe__resource_actions(after->rsc, NULL, RSC_STOP, FALSE); } else if (safe_str_eq(after->task, RSC_PROMOTE)) { then_actions = pe__resource_actions(after->rsc, NULL, RSC_DEMOTE, FALSE); } for (gIter = then_actions; gIter != NULL; gIter = gIter->next) { pe_action_t *then = (pe_action_t *) gIter->data; // Skip any pseudo action which for example is implied by fencing if (is_set(then->flags, pe_action_pseudo)) { continue; } order_actions(probe, then, probe_order_type); } g_list_free(then_actions); } if (after->rsc && after->rsc->variant > pe_group) { const char *interleave_s = g_hash_table_lookup(after->rsc->meta, XML_RSC_ATTR_INTERLEAVE); interleave = crm_is_true(interleave_s); if (interleave) { /* For an interleaved clone, we should build a dependency only * with the relevant clone child. */ compatible_rsc = find_compatible_child(probe->rsc, after->rsc, RSC_ROLE_UNKNOWN, FALSE, data_set); } } for (gIter = after->actions_after; gIter != NULL; gIter = gIter->next) { pe_action_wrapper_t *after_wrapper = (pe_action_wrapper_t *) gIter->data; /* pe_order_implies_then is the reason why a required A.start * implies/enforces B.start to be required too, which is the cause of * B.restart/re-promote. * * Not sure about pe_order_implies_then_on_node though. It's now only * used for unfencing case, which tends to introduce transition * loops... */ if (is_not_set(after_wrapper->type, pe_order_implies_then)) { /* The order type between a group/clone and its child such as * B.start-> B_child.start is: * pe_order_implies_first_printed | pe_order_runnable_left * * Proceed through the ordering chain and build dependencies with * its children. */ if (after->rsc == NULL || after->rsc->variant < pe_group || probe->rsc->parent == after->rsc || after_wrapper->action->rsc == NULL || after_wrapper->action->rsc->variant > pe_group || after->rsc != after_wrapper->action->rsc->parent) { continue; } /* Proceed to the children of a group or a non-interleaved clone. * For an interleaved clone, proceed only to the relevant child. */ if (after->rsc->variant > pe_group && interleave == TRUE && (compatible_rsc == NULL || compatible_rsc != after_wrapper->action->rsc)) { continue; } } crm_trace("Proceeding through %s %s -> %s %s (type=0x%.6x)", after->uuid, after->node ? after->node->details->uname: "", after_wrapper->action->uuid, after_wrapper->action->node ? after_wrapper->action->node->details->uname : "", after_wrapper->type); order_first_probe_then_restart_repromote(probe, after_wrapper->action, data_set); } } static void clear_actions_tracking_flag(pe_working_set_t * data_set) { GListPtr gIter = NULL; for (gIter = data_set->actions; gIter != NULL; gIter = gIter->next) { pe_action_t *action = (pe_action_t *) gIter->data; if (is_set(action->flags, pe_action_tracking)) { pe_clear_action_bit(action, pe_action_tracking); } } } static void order_first_rsc_probes(pe_resource_t * rsc, pe_working_set_t * data_set) { GListPtr gIter = NULL; GListPtr probes = NULL; for (gIter = rsc->children; gIter != NULL; gIter = gIter->next) { pe_resource_t * child = (pe_resource_t *) gIter->data; order_first_rsc_probes(child, data_set); } if (rsc->variant != pe_native) { return; } probes = pe__resource_actions(rsc, NULL, RSC_STATUS, FALSE); for (gIter = probes; gIter != NULL; gIter= gIter->next) { pe_action_t *probe = (pe_action_t *) gIter->data; GListPtr aIter = NULL; for (aIter = probe->actions_after; aIter != NULL; aIter = aIter->next) { pe_action_wrapper_t *after_wrapper = (pe_action_wrapper_t *) aIter->data; order_first_probe_then_restart_repromote(probe, after_wrapper->action, data_set); clear_actions_tracking_flag(data_set); } } g_list_free(probes); } static void order_first_probes(pe_working_set_t * data_set) { GListPtr gIter = NULL; for (gIter = data_set->resources; gIter != NULL; gIter = gIter->next) { pe_resource_t *rsc = (pe_resource_t *) gIter->data; order_first_rsc_probes(rsc, data_set); } order_first_probes_imply_stops(data_set); } static void order_then_probes(pe_working_set_t * data_set) { #if 0 GListPtr gIter = NULL; for (gIter = data_set->resources; gIter != NULL; gIter = gIter->next) { resource_t *rsc = (resource_t *) gIter->data; /* Given "A then B", we would prefer to wait for A to be * started before probing B. * * If A was a filesystem on which the binaries and data for B * lived, it would have been useful if the author of B's agent * could assume that A is running before B.monitor will be * called. * * However we can't _only_ probe once A is running, otherwise * we'd not detect the state of B if A could not be started * for some reason. * * In practice however, we cannot even do an opportunistic * version of this because B may be moving: * * B.probe -> B.start * B.probe -> B.stop * B.stop -> B.start * A.stop -> A.start * A.start -> B.probe * * So far so good, but if we add the result of this code: * * B.stop -> A.stop * * Then we get a loop: * * B.probe -> B.stop -> A.stop -> A.start -> B.probe * * We could kill the 'B.probe -> B.stop' dependency, but that * could mean stopping B "too" soon, because B.start must wait * for the probes to complete. * * Another option is to allow it only if A is a non-unique * clone with clone-max == node-max (since we'll never be * moving it). However, we could still be stopping one * instance at the same time as starting another. * The complexity of checking for allowed conditions combined * with the ever narrowing usecase suggests that this code * should remain disabled until someone gets smarter. */ action_t *start = NULL; GListPtr actions = NULL; GListPtr probes = NULL; actions = pe__resource_actions(rsc, NULL, RSC_START, FALSE); if (actions) { start = actions->data; g_list_free(actions); } if(start == NULL) { crm_err("No start action for %s", rsc->id); continue; } probes = pe__resource_actions(rsc, NULL, RSC_STATUS, FALSE); for (actions = start->actions_before; actions != NULL; actions = actions->next) { action_wrapper_t *before = (action_wrapper_t *) actions->data; GListPtr pIter = NULL; action_t *first = before->action; resource_t *first_rsc = first->rsc; if(first->required_runnable_before) { GListPtr clone_actions = NULL; for (clone_actions = first->actions_before; clone_actions != NULL; clone_actions = clone_actions->next) { before = (action_wrapper_t *) clone_actions->data; crm_trace("Testing %s -> %s (%p) for %s", first->uuid, before->action->uuid, before->action->rsc, start->uuid); CRM_ASSERT(before->action->rsc); first_rsc = before->action->rsc; break; } } else if(safe_str_neq(first->task, RSC_START)) { crm_trace("Not a start op %s for %s", first->uuid, start->uuid); } if(first_rsc == NULL) { continue; } else if(uber_parent(first_rsc) == uber_parent(start->rsc)) { crm_trace("Same parent %s for %s", first_rsc->id, start->uuid); continue; } else if(FALSE && pe_rsc_is_clone(uber_parent(first_rsc)) == FALSE) { crm_trace("Not a clone %s for %s", first_rsc->id, start->uuid); continue; } crm_err("Applying %s before %s %d", first->uuid, start->uuid, uber_parent(first_rsc)->variant); for (pIter = probes; pIter != NULL; pIter = pIter->next) { action_t *probe = (action_t *) pIter->data; crm_err("Ordering %s before %s", first->uuid, probe->uuid); order_actions(first, probe, pe_order_optional); } } } #endif } static void order_probes(pe_working_set_t * data_set) { order_first_probes(data_set); order_then_probes(data_set); } gboolean stage7(pe_working_set_t * data_set) { GList *gIter = NULL; crm_trace("Applying ordering constraints"); /* Don't ask me why, but apparently they need to be processed in * the order they were created in... go figure * * Also g_list_append() has horrendous performance characteristics * So we need to use g_list_prepend() and then reverse the list here */ data_set->ordering_constraints = g_list_reverse(data_set->ordering_constraints); for (gIter = data_set->ordering_constraints; gIter != NULL; gIter = gIter->next) { pe__ordering_t *order = gIter->data; resource_t *rsc = order->lh_rsc; crm_trace("Applying ordering constraint: %d", order->id); if (rsc != NULL) { crm_trace("rsc_action-to-*"); rsc_order_first(rsc, order, data_set); continue; } rsc = order->rh_rsc; if (rsc != NULL) { crm_trace("action-to-rsc_action"); rsc_order_then(order->lh_action, rsc, order); } else { crm_trace("action-to-action"); order_actions(order->lh_action, order->rh_action, order->type); } } for (gIter = data_set->actions; gIter != NULL; gIter = gIter->next) { action_t *action = (action_t *) gIter->data; update_colo_start_chain(action, data_set); } crm_trace("Ordering probes"); order_probes(data_set); crm_trace("Updating %d actions", g_list_length(data_set->actions)); for (gIter = data_set->actions; gIter != NULL; gIter = gIter->next) { action_t *action = (action_t *) gIter->data; update_action(action, data_set); } // Check for invalid orderings for (gIter = data_set->actions; gIter != NULL; gIter = gIter->next) { pe_action_t *action = (pe_action_t *) gIter->data; pe_action_wrapper_t *input = NULL; for (GList *input_iter = action->actions_before; input_iter != NULL; input_iter = input_iter->next) { input = (pe_action_wrapper_t *) input_iter->data; if (pcmk__ordering_is_invalid(action, input)) { input->type = pe_order_none; } } } LogNodeActions(data_set, FALSE); for (gIter = data_set->resources; gIter != NULL; gIter = gIter->next) { resource_t *rsc = (resource_t *) gIter->data; LogActions(rsc, data_set, FALSE); } return TRUE; } static int transition_id = -1; /*! * \internal * \brief Log a message after calculating a transition * * \param[in] filename Where transition input is stored */ void pcmk__log_transition_summary(const char *filename) { if (was_processing_error) { crm_err("Calculated transition %d (with errors), saving inputs in %s", transition_id, filename); } else if (was_processing_warning) { crm_warn("Calculated transition %d (with warnings), saving inputs in %s", transition_id, filename); } else { crm_notice("Calculated transition %d, saving inputs in %s", transition_id, filename); } if (crm_config_error) { crm_notice("Configuration errors found during scheduler processing," " please run \"crm_verify -L\" to identify issues"); } } /* * Create a dependency graph to send to the transitioner (via the controller) */ gboolean stage8(pe_working_set_t * data_set) { GListPtr gIter = NULL; const char *value = NULL; transition_id++; crm_trace("Creating transition graph %d.", transition_id); data_set->graph = create_xml_node(NULL, XML_TAG_GRAPH); value = pe_pref(data_set->config_hash, "cluster-delay"); crm_xml_add(data_set->graph, "cluster-delay", value); value = pe_pref(data_set->config_hash, "stonith-timeout"); crm_xml_add(data_set->graph, "stonith-timeout", value); crm_xml_add(data_set->graph, "failed-stop-offset", "INFINITY"); if (is_set(data_set->flags, pe_flag_start_failure_fatal)) { crm_xml_add(data_set->graph, "failed-start-offset", "INFINITY"); } else { crm_xml_add(data_set->graph, "failed-start-offset", "1"); } value = pe_pref(data_set->config_hash, "batch-limit"); crm_xml_add(data_set->graph, "batch-limit", value); crm_xml_add_int(data_set->graph, "transition_id", transition_id); value = pe_pref(data_set->config_hash, "migration-limit"); if (crm_int_helper(value, NULL) > 0) { crm_xml_add(data_set->graph, "migration-limit", value); } if (data_set->recheck_by > 0) { char *recheck_epoch = NULL; recheck_epoch = crm_strdup_printf("%llu", (long long) data_set->recheck_by); crm_xml_add(data_set->graph, "recheck-by", recheck_epoch); free(recheck_epoch); } /* errors... slist_iter(action, action_t, action_list, lpc, if(action->optional == FALSE && action->runnable == FALSE) { print_action("Ignoring", action, TRUE); } ); */ /* The following code will de-duplicate action inputs, so nothing past this * should rely on the action input type flags retaining their original * values. */ gIter = data_set->resources; for (; gIter != NULL; gIter = gIter->next) { resource_t *rsc = (resource_t *) gIter->data; pe_rsc_trace(rsc, "processing actions for rsc=%s", rsc->id); rsc->cmds->expand(rsc, data_set); } crm_log_xml_trace(data_set->graph, "created resource-driven action list"); /* pseudo action to distribute list of nodes with maintenance state update */ add_maintenance_update(data_set); /* catch any non-resource specific actions */ crm_trace("processing non-resource actions"); gIter = data_set->actions; for (; gIter != NULL; gIter = gIter->next) { action_t *action = (action_t *) gIter->data; if (action->rsc && action->node && action->node->details->shutdown && is_not_set(action->rsc->flags, pe_rsc_maintenance) && is_not_set(action->flags, pe_action_optional) && is_not_set(action->flags, pe_action_runnable) && crm_str_eq(action->task, RSC_STOP, TRUE) ) { /* Eventually we should just ignore the 'fence' case * But for now it's the best way to detect (in CTS) when * CIB resource updates are being lost */ if (is_set(data_set->flags, pe_flag_have_quorum) || data_set->no_quorum_policy == no_quorum_ignore) { crm_crit("Cannot %s node '%s' because of %s:%s%s (%s)", action->node->details->unclean ? "fence" : "shut down", action->node->details->uname, action->rsc->id, is_not_set(action->rsc->flags, pe_rsc_managed) ? " unmanaged" : " blocked", is_set(action->rsc->flags, pe_rsc_failed) ? " failed" : "", action->uuid); } } graph_element_from_action(action, data_set); } crm_log_xml_trace(data_set->graph, "created generic action list"); crm_trace("Created transition graph %d.", transition_id); return TRUE; } void LogNodeActions(pe_working_set_t * data_set, gboolean terminal) { GListPtr gIter = NULL; for (gIter = data_set->actions; gIter != NULL; gIter = gIter->next) { char *node_name = NULL; char *task = NULL; action_t *action = (action_t *) gIter->data; if (action->rsc != NULL) { continue; } else if (is_set(action->flags, pe_action_optional)) { continue; } if (pe__is_guest_node(action->node)) { node_name = crm_strdup_printf("%s (resource: %s)", action->node->details->uname, action->node->details->remote_rsc->container->id); } else if(action->node) { node_name = crm_strdup_printf("%s", action->node->details->uname); } if (safe_str_eq(action->task, CRM_OP_SHUTDOWN)) { task = strdup("Shutdown"); } else if (safe_str_eq(action->task, CRM_OP_FENCE)) { const char *op = g_hash_table_lookup(action->meta, "stonith_action"); task = crm_strdup_printf("Fence (%s)", op); } if(task == NULL) { /* Nothing to report */ } else if(terminal && action->reason) { printf(" * %s %s '%s'\n", task, node_name, action->reason); } else if(terminal) { printf(" * %s %s\n", task, node_name); } else if(action->reason) { crm_notice(" * %s %s '%s'\n", task, node_name, action->reason); } else { crm_notice(" * %s %s\n", task, node_name); } free(node_name); free(task); } } diff --git a/lib/pacemaker/pcmk_sched_bundle.c b/lib/pacemaker/pcmk_sched_bundle.c index 112b670847..1d7b2634c7 100644 --- a/lib/pacemaker/pcmk_sched_bundle.c +++ b/lib/pacemaker/pcmk_sched_bundle.c @@ -1,1075 +1,1083 @@ /* * Copyright 2004-2020 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 #define PE__VARIANT_BUNDLE 1 #include static bool is_bundle_node(pe__bundle_variant_data_t *data, pe_node_t *node) { for (GList *gIter = data->replicas; gIter != NULL; gIter = gIter->next) { pe__bundle_replica_t *replica = gIter->data; if (node->details == replica->node->details) { return TRUE; } } return FALSE; } gint sort_clone_instance(gconstpointer a, gconstpointer b, gpointer data_set); void distribute_children(resource_t *rsc, GListPtr children, GListPtr nodes, int max, int per_host_max, pe_working_set_t * data_set); static GList * get_container_list(pe_resource_t *rsc) { GList *containers = NULL; if (rsc->variant == pe_container) { pe__bundle_variant_data_t *data = NULL; get_bundle_variant_data(data, rsc); for (GList *gIter = data->replicas; gIter != NULL; gIter = gIter->next) { pe__bundle_replica_t *replica = gIter->data; containers = g_list_append(containers, replica->container); } } return containers; } static inline GList * get_containers_or_children(pe_resource_t *rsc) { return (rsc->variant == pe_container)? get_container_list(rsc) : rsc->children; } static bool migration_threshold_reached(resource_t *rsc, node_t *node, pe_working_set_t *data_set) { int fail_count, countdown; /* Migration threshold of 0 means never force away */ if (rsc->migration_threshold == 0) { return FALSE; } // If we're ignoring failures, also ignore the migration threshold if (is_set(rsc->flags, pe_rsc_failure_ignored)) { return FALSE; } /* If there are no failures, there's no need to force away */ fail_count = pe_get_failcount(node, rsc, NULL, pe_fc_effective|pe_fc_fillers, NULL, data_set); if (fail_count <= 0) { return FALSE; } /* How many more times recovery will be tried on this node */ countdown = QB_MAX(rsc->migration_threshold - fail_count, 0); if (countdown == 0) { crm_warn("Forcing %s away from %s after %d failures (max=%d)", rsc->id, node->details->uname, fail_count, rsc->migration_threshold); return TRUE; } crm_info("%s can fail %d more times on %s before being forced off", rsc->id, countdown, node->details->uname); return FALSE; } pe_node_t * pcmk__bundle_allocate(pe_resource_t *rsc, pe_node_t *prefer, pe_working_set_t *data_set) { GListPtr containers = NULL; GListPtr nodes = NULL; pe__bundle_variant_data_t *bundle_data = NULL; CRM_CHECK(rsc != NULL, return NULL); get_bundle_variant_data(bundle_data, rsc); set_bit(rsc->flags, pe_rsc_allocating); containers = get_container_list(rsc); pe__show_node_weights(!show_scores, rsc, __FUNCTION__, rsc->allowed_nodes); nodes = g_hash_table_get_values(rsc->allowed_nodes); nodes = sort_nodes_by_weight(nodes, NULL, data_set); containers = g_list_sort_with_data(containers, sort_clone_instance, data_set); distribute_children(rsc, containers, nodes, bundle_data->nreplicas, bundle_data->nreplicas_per_host, data_set); g_list_free(nodes); g_list_free(containers); for (GList *gIter = bundle_data->replicas; gIter != NULL; gIter = gIter->next) { pe__bundle_replica_t *replica = gIter->data; pe_node_t *container_host = NULL; CRM_ASSERT(replica); if (replica->ip) { + pe_rsc_trace(rsc, "Allocating bundle %s IP %s", + rsc->id, replica->ip->id); replica->ip->cmds->allocate(replica->ip, prefer, data_set); } container_host = replica->container->allocated_to; if (replica->remote && pe__is_guest_or_remote_node(container_host)) { /* We need 'nested' connection resources to be on the same * host because pacemaker-remoted only supports a single * active connection */ rsc_colocation_new("child-remote-with-docker-remote", NULL, INFINITY, replica->remote, container_host->details->remote_rsc, NULL, NULL, data_set); } if (replica->remote) { + pe_rsc_trace(rsc, "Allocating bundle %s connection %s", + rsc->id, replica->remote->id); replica->remote->cmds->allocate(replica->remote, prefer, data_set); } // Explicitly allocate replicas' children before bundle child if (replica->child) { 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 (node->details != replica->node->details) { node->weight = -INFINITY; } else if (!migration_threshold_reached(replica->child, node, data_set)) { node->weight = INFINITY; } } set_bit(replica->child->parent->flags, pe_rsc_allocating); + pe_rsc_trace(rsc, "Allocating bundle %s replica child %s", + rsc->id, replica->child->id); replica->child->cmds->allocate(replica->child, replica->node, data_set); clear_bit(replica->child->parent->flags, pe_rsc_allocating); } } if (bundle_data->child) { pe_node_t *node = NULL; GHashTableIter iter; g_hash_table_iter_init(&iter, bundle_data->child->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (gpointer *) & node)) { if (is_bundle_node(bundle_data, node)) { node->weight = 0; } else { node->weight = -INFINITY; } } + pe_rsc_trace(rsc, "Allocating bundle %s child %s", + rsc->id, bundle_data->child->id); bundle_data->child->cmds->allocate(bundle_data->child, prefer, data_set); } clear_bit(rsc->flags, pe_rsc_allocating); clear_bit(rsc->flags, pe_rsc_provisional); return NULL; } void pcmk__bundle_create_actions(pe_resource_t *rsc, pe_working_set_t *data_set) { pe_action_t *action = NULL; GListPtr containers = NULL; pe__bundle_variant_data_t *bundle_data = NULL; CRM_CHECK(rsc != NULL, return); containers = get_container_list(rsc); get_bundle_variant_data(bundle_data, rsc); for (GList *gIter = bundle_data->replicas; gIter != NULL; gIter = gIter->next) { pe__bundle_replica_t *replica = gIter->data; CRM_ASSERT(replica); if (replica->ip) { replica->ip->cmds->create_actions(replica->ip, data_set); } if (replica->container) { replica->container->cmds->create_actions(replica->container, data_set); } if (replica->remote) { replica->remote->cmds->create_actions(replica->remote, data_set); } } clone_create_pseudo_actions(rsc, containers, NULL, NULL, data_set); if (bundle_data->child) { bundle_data->child->cmds->create_actions(bundle_data->child, data_set); if (is_set(bundle_data->child->flags, pe_rsc_promotable)) { /* promote */ action = create_pseudo_resource_op(rsc, RSC_PROMOTE, TRUE, TRUE, data_set); action = create_pseudo_resource_op(rsc, RSC_PROMOTED, TRUE, TRUE, data_set); action->priority = INFINITY; /* demote */ action = create_pseudo_resource_op(rsc, RSC_DEMOTE, TRUE, TRUE, data_set); action = create_pseudo_resource_op(rsc, RSC_DEMOTED, TRUE, TRUE, data_set); action->priority = INFINITY; } } g_list_free(containers); } void pcmk__bundle_internal_constraints(pe_resource_t *rsc, pe_working_set_t *data_set) { pe__bundle_variant_data_t *bundle_data = NULL; CRM_CHECK(rsc != NULL, return); get_bundle_variant_data(bundle_data, rsc); if (bundle_data->child) { new_rsc_order(rsc, RSC_START, bundle_data->child, RSC_START, pe_order_implies_first_printed, data_set); new_rsc_order(rsc, RSC_STOP, bundle_data->child, RSC_STOP, pe_order_implies_first_printed, data_set); if (bundle_data->child->children) { new_rsc_order(bundle_data->child, RSC_STARTED, rsc, RSC_STARTED, pe_order_implies_then_printed, data_set); new_rsc_order(bundle_data->child, RSC_STOPPED, rsc, RSC_STOPPED, pe_order_implies_then_printed, data_set); } else { new_rsc_order(bundle_data->child, RSC_START, rsc, RSC_STARTED, pe_order_implies_then_printed, data_set); new_rsc_order(bundle_data->child, RSC_STOP, rsc, RSC_STOPPED, pe_order_implies_then_printed, data_set); } } for (GList *gIter = bundle_data->replicas; gIter != NULL; gIter = gIter->next) { pe__bundle_replica_t *replica = gIter->data; CRM_ASSERT(replica); CRM_ASSERT(replica->container); replica->container->cmds->internal_constraints(replica->container, data_set); order_start_start(rsc, replica->container, pe_order_runnable_left|pe_order_implies_first_printed); if (replica->child) { order_stop_stop(rsc, replica->child, pe_order_implies_first_printed); } order_stop_stop(rsc, replica->container, pe_order_implies_first_printed); new_rsc_order(replica->container, RSC_START, rsc, RSC_STARTED, pe_order_implies_then_printed, data_set); new_rsc_order(replica->container, RSC_STOP, rsc, RSC_STOPPED, pe_order_implies_then_printed, data_set); if (replica->ip) { replica->ip->cmds->internal_constraints(replica->ip, data_set); // Start ip then container new_rsc_order(replica->ip, RSC_START, replica->container, RSC_START, pe_order_runnable_left|pe_order_preserve, data_set); new_rsc_order(replica->container, RSC_STOP, replica->ip, RSC_STOP, pe_order_implies_first|pe_order_preserve, data_set); rsc_colocation_new("ip-with-docker", NULL, INFINITY, replica->ip, replica->container, NULL, NULL, data_set); } if (replica->remote) { /* 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, data_set); } if (replica->child) { CRM_ASSERT(replica->remote); // "Start remote then child" is implicit in scheduler's remote logic } } if (bundle_data->child) { bundle_data->child->cmds->internal_constraints(bundle_data->child, data_set); if (is_set(bundle_data->child->flags, pe_rsc_promotable)) { promote_demote_constraints(rsc, data_set); /* child demoted before global demoted */ new_rsc_order(bundle_data->child, RSC_DEMOTED, rsc, RSC_DEMOTED, pe_order_implies_then_printed, data_set); /* global demote before child demote */ new_rsc_order(rsc, RSC_DEMOTE, bundle_data->child, RSC_DEMOTE, pe_order_implies_first_printed, data_set); /* child promoted before global promoted */ new_rsc_order(bundle_data->child, RSC_PROMOTED, rsc, RSC_PROMOTED, pe_order_implies_then_printed, data_set); /* global promote before child promote */ new_rsc_order(rsc, RSC_PROMOTE, bundle_data->child, RSC_PROMOTE, pe_order_implies_first_printed, data_set); } } else { // int type = pe_order_optional | pe_order_implies_then | pe_order_restart; // custom_action_order(rsc, pcmk__op_key(rsc->id, RSC_STOP, 0), NULL, // rsc, pcmk__op_key(rsc->id, RSC_START, 0), NULL, pe_order_optional, data_set); } } static pe_resource_t * compatible_replica_for_node(pe_resource_t *rsc_lh, pe_node_t *candidate, pe_resource_t *rsc, enum rsc_role_e filter, gboolean current) { pe__bundle_variant_data_t *bundle_data = NULL; CRM_CHECK(candidate != NULL, return NULL); get_bundle_variant_data(bundle_data, rsc); crm_trace("Looking for compatible child from %s for %s on %s", rsc_lh->id, rsc->id, candidate->details->uname); for (GList *gIter = bundle_data->replicas; gIter != NULL; gIter = gIter->next) { pe__bundle_replica_t *replica = gIter->data; if (is_child_compatible(replica->container, candidate, filter, current)) { crm_trace("Pairing %s with %s on %s", rsc_lh->id, replica->container->id, candidate->details->uname); return replica->container; } } crm_trace("Can't pair %s with %s", rsc_lh->id, rsc->id); return NULL; } static pe_resource_t * compatible_replica(pe_resource_t *rsc_lh, pe_resource_t *rsc, enum rsc_role_e filter, gboolean current, pe_working_set_t *data_set) { GListPtr scratch = NULL; resource_t *pair = NULL; node_t *active_node_lh = NULL; active_node_lh = rsc_lh->fns->location(rsc_lh, NULL, current); if (active_node_lh) { return compatible_replica_for_node(rsc_lh, active_node_lh, rsc, filter, current); } scratch = g_hash_table_get_values(rsc_lh->allowed_nodes); scratch = sort_nodes_by_weight(scratch, NULL, data_set); for (GListPtr gIter = scratch; gIter != NULL; gIter = gIter->next) { node_t *node = (node_t *) gIter->data; pair = compatible_replica_for_node(rsc_lh, node, rsc, filter, current); if (pair) { goto done; } } pe_rsc_debug(rsc, "Can't pair %s with %s", rsc_lh->id, (rsc? rsc->id : "none")); done: g_list_free(scratch); return pair; } void pcmk__bundle_rsc_colocation_lh(pe_resource_t *rsc, pe_resource_t *rsc_rh, rsc_colocation_t *constraint, pe_working_set_t *data_set) { /* -- Never called -- * * Instead we add the colocation constraints to the child and call from there */ CRM_ASSERT(FALSE); } int copies_per_node(resource_t * rsc) { /* Strictly speaking, there should be a 'copies_per_node' addition * to the resource function table and each case would be a * function. However that would be serious overkill to return an * int. In fact, it seems to me that both function tables * could/should be replaced by resources.{c,h} full of * rsc_{some_operation} functions containing a switch as below * which calls out to functions named {variant}_{some_operation} * as needed. */ switch(rsc->variant) { case pe_unknown: return 0; case pe_native: case pe_group: return 1; case pe_clone: { const char *max_clones_node = g_hash_table_lookup(rsc->meta, XML_RSC_ATTR_INCARNATION_NODEMAX); return crm_parse_int(max_clones_node, "1"); } case pe_container: { pe__bundle_variant_data_t *data = NULL; get_bundle_variant_data(data, rsc); return data->nreplicas_per_host; } } return 0; } void pcmk__bundle_rsc_colocation_rh(pe_resource_t *rsc_lh, pe_resource_t *rsc, rsc_colocation_t *constraint, pe_working_set_t *data_set) { GListPtr allocated_rhs = NULL; pe__bundle_variant_data_t *bundle_data = NULL; CRM_CHECK(constraint != NULL, return); CRM_CHECK(rsc_lh != NULL, pe_err("rsc_lh was NULL for %s", constraint->id); return); CRM_CHECK(rsc != NULL, pe_err("rsc was NULL for %s", constraint->id); return); CRM_ASSERT(rsc_lh->variant == pe_native); if (constraint->score == 0) { return; } if (is_set(rsc->flags, pe_rsc_provisional)) { pe_rsc_trace(rsc, "%s is still provisional", rsc->id); return; } else if(constraint->rsc_lh->variant > pe_group) { resource_t *rh_child = compatible_replica(rsc_lh, rsc, RSC_ROLE_UNKNOWN, FALSE, data_set); if (rh_child) { pe_rsc_debug(rsc, "Pairing %s with %s", rsc_lh->id, rh_child->id); rsc_lh->cmds->rsc_colocation_lh(rsc_lh, rh_child, constraint, data_set); } else if (constraint->score >= INFINITY) { crm_notice("Cannot pair %s with instance of %s", rsc_lh->id, rsc->id); assign_node(rsc_lh, NULL, TRUE); } else { pe_rsc_debug(rsc, "Cannot pair %s with instance of %s", rsc_lh->id, rsc->id); } return; } get_bundle_variant_data(bundle_data, rsc); pe_rsc_trace(rsc, "Processing constraint %s: %s -> %s %d", constraint->id, rsc_lh->id, rsc->id, constraint->score); for (GList *gIter = bundle_data->replicas; gIter != NULL; gIter = gIter->next) { pe__bundle_replica_t *replica = gIter->data; if (constraint->score < INFINITY) { replica->container->cmds->rsc_colocation_rh(rsc_lh, replica->container, constraint, data_set); } else { node_t *chosen = replica->container->fns->location(replica->container, NULL, FALSE); if ((chosen == NULL) || is_set_recursive(replica->container, pe_rsc_block, TRUE)) { continue; } if ((constraint->role_rh >= RSC_ROLE_MASTER) && (replica->child == NULL)) { continue; } if ((constraint->role_rh >= RSC_ROLE_MASTER) && (replica->child->next_role < RSC_ROLE_MASTER)) { continue; } pe_rsc_trace(rsc, "Allowing %s: %s %d", constraint->id, chosen->details->uname, chosen->weight); allocated_rhs = g_list_prepend(allocated_rhs, chosen); } } if (constraint->score >= INFINITY) { node_list_exclude(rsc_lh->allowed_nodes, allocated_rhs, FALSE); } g_list_free(allocated_rhs); } enum pe_action_flags pcmk__bundle_action_flags(pe_action_t *action, pe_node_t *node) { GListPtr containers = NULL; enum pe_action_flags flags = 0; pe__bundle_variant_data_t *data = NULL; get_bundle_variant_data(data, action->rsc); if(data->child) { enum action_tasks task = get_complex_task(data->child, action->task, TRUE); switch(task) { case no_action: case action_notify: case action_notified: case action_promote: case action_promoted: case action_demote: case action_demoted: return summary_action_flags(action, data->child->children, node); default: break; } } containers = get_container_list(action->rsc); flags = summary_action_flags(action, containers, node); g_list_free(containers); return flags; } resource_t * find_compatible_child_by_node(resource_t * local_child, node_t * local_node, resource_t * rsc, enum rsc_role_e filter, gboolean current) { GListPtr gIter = NULL; GListPtr children = NULL; if (local_node == NULL) { crm_err("Can't colocate unrunnable child %s with %s", local_child->id, rsc->id); return NULL; } crm_trace("Looking for compatible child from %s for %s on %s", local_child->id, rsc->id, local_node->details->uname); children = get_containers_or_children(rsc); for (gIter = children; gIter != NULL; gIter = gIter->next) { resource_t *child_rsc = (resource_t *) gIter->data; if(is_child_compatible(child_rsc, local_node, filter, current)) { crm_trace("Pairing %s with %s on %s", local_child->id, child_rsc->id, local_node->details->uname); return child_rsc; } } crm_trace("Can't pair %s with %s", local_child->id, rsc->id); if(children != rsc->children) { g_list_free(children); } return NULL; } static pe__bundle_replica_t * replica_for_container(pe_resource_t *rsc, pe_resource_t *container, pe_node_t *node) { if (rsc->variant == pe_container) { pe__bundle_variant_data_t *data = NULL; get_bundle_variant_data(data, rsc); for (GList *gIter = data->replicas; gIter != NULL; gIter = gIter->next) { pe__bundle_replica_t *replica = gIter->data; if (replica->child && (container == replica->container) && (node->details == replica->node->details)) { return replica; } } } return NULL; } static enum pe_graph_flags multi_update_interleave_actions(pe_action_t *first, pe_action_t *then, pe_node_t *node, enum pe_action_flags flags, enum pe_action_flags filter, enum pe_ordering type, pe_working_set_t *data_set) { GListPtr gIter = NULL; GListPtr children = NULL; gboolean current = FALSE; enum pe_graph_flags changed = pe_graph_none; /* Fix this - lazy */ if (crm_ends_with(first->uuid, "_stopped_0") || crm_ends_with(first->uuid, "_demoted_0")) { current = TRUE; } children = get_containers_or_children(then->rsc); for (gIter = children; gIter != NULL; gIter = gIter->next) { pe_resource_t *then_child = gIter->data; pe_resource_t *first_child = find_compatible_child(then_child, first->rsc, RSC_ROLE_UNKNOWN, current, data_set); if (first_child == NULL && current) { crm_trace("Ignore"); } else if (first_child == NULL) { crm_debug("No match found for %s (%d / %s / %s)", then_child->id, current, first->uuid, then->uuid); /* Me no like this hack - but what else can we do? * * If there is no-one active or about to be active * on the same node as then_child, then they must * not be allowed to start */ if (type & (pe_order_runnable_left | pe_order_implies_then) /* Mandatory */ ) { pe_rsc_info(then->rsc, "Inhibiting %s from being active", then_child->id); if(assign_node(then_child, NULL, TRUE)) { changed |= pe_graph_updated_then; } } } else { pe_action_t *first_action = NULL; pe_action_t *then_action = NULL; enum action_tasks task = clone_child_action(first); const char *first_task = task2text(task); pe__bundle_replica_t *first_replica = NULL; pe__bundle_replica_t *then_replica = NULL; first_replica = replica_for_container(first->rsc, first_child, node); if (strstr(first->task, "stop") && first_replica && first_replica->child) { /* Except for 'stopped' we should be looking at the * in-container resource, actions for the child will * happen later and are therefor more likely to align * with the user's intent. */ first_action = find_first_action(first_replica->child->actions, NULL, task2text(task), node); } else { first_action = find_first_action(first_child->actions, NULL, task2text(task), node); } then_replica = replica_for_container(then->rsc, then_child, node); if (strstr(then->task, "mote") && then_replica && then_replica->child) { /* Promote/demote actions will never be found for the * container resource, look in the child instead * * Alternatively treat: * 'XXXX then promote YYYY' as 'XXXX then start container for YYYY', and * 'demote XXXX then stop YYYY' as 'stop container for XXXX then stop YYYY' */ then_action = find_first_action(then_replica->child->actions, NULL, then->task, node); } else { then_action = find_first_action(then_child->actions, NULL, then->task, node); } if (first_action == NULL) { if (is_not_set(first_child->flags, pe_rsc_orphan) && crm_str_eq(first_task, RSC_STOP, TRUE) == FALSE && crm_str_eq(first_task, RSC_DEMOTE, TRUE) == FALSE) { crm_err("Internal error: No action found for %s in %s (first)", first_task, first_child->id); } else { crm_trace("No action found for %s in %s%s (first)", first_task, first_child->id, is_set(first_child->flags, pe_rsc_orphan) ? " (ORPHAN)" : ""); } continue; } /* We're only interested if 'then' is neither stopping nor being demoted */ if (then_action == NULL) { if (is_not_set(then_child->flags, pe_rsc_orphan) && crm_str_eq(then->task, RSC_STOP, TRUE) == FALSE && crm_str_eq(then->task, RSC_DEMOTE, TRUE) == FALSE) { crm_err("Internal error: No action found for %s in %s (then)", then->task, then_child->id); } else { crm_trace("No action found for %s in %s%s (then)", then->task, then_child->id, is_set(then_child->flags, pe_rsc_orphan) ? " (ORPHAN)" : ""); } continue; } if (order_actions(first_action, then_action, type)) { crm_debug("Created constraint for %s (%d) -> %s (%d) %.6x", first_action->uuid, is_set(first_action->flags, pe_action_optional), then_action->uuid, is_set(then_action->flags, pe_action_optional), type); changed |= (pe_graph_updated_first | pe_graph_updated_then); } if(first_action && then_action) { changed |= then_child->cmds->update_actions(first_action, then_action, node, first_child->cmds->action_flags(first_action, node), filter, type, data_set); } else { crm_err("Nothing found either for %s (%p) or %s (%p) %s", first_child->id, first_action, then_child->id, then_action, task2text(task)); } } } if(children != then->rsc->children) { g_list_free(children); } return changed; } static bool can_interleave_actions(pe_action_t *first, pe_action_t *then) { bool interleave = FALSE; resource_t *rsc = NULL; const char *interleave_s = NULL; if(first->rsc == NULL || then->rsc == NULL) { crm_trace("Not interleaving %s with %s (both must be resources)", first->uuid, then->uuid); return FALSE; } else if(first->rsc == then->rsc) { crm_trace("Not interleaving %s with %s (must belong to different resources)", first->uuid, then->uuid); return FALSE; } else if(first->rsc->variant < pe_clone || then->rsc->variant < pe_clone) { crm_trace("Not interleaving %s with %s (both sides must be clones or bundles)", first->uuid, then->uuid); return FALSE; } if (crm_ends_with(then->uuid, "_stop_0") || crm_ends_with(then->uuid, "_demote_0")) { rsc = first->rsc; } else { rsc = then->rsc; } interleave_s = g_hash_table_lookup(rsc->meta, XML_RSC_ATTR_INTERLEAVE); interleave = crm_is_true(interleave_s); crm_trace("Interleave %s -> %s: %s (based on %s)", first->uuid, then->uuid, interleave ? "yes" : "no", rsc->id); return interleave; } enum pe_graph_flags pcmk__multi_update_actions(pe_action_t *first, pe_action_t *then, pe_node_t *node, enum pe_action_flags flags, enum pe_action_flags filter, enum pe_ordering type, pe_working_set_t *data_set) { enum pe_graph_flags changed = pe_graph_none; crm_trace("%s -> %s", first->uuid, then->uuid); if(can_interleave_actions(first, then)) { changed = multi_update_interleave_actions(first, then, node, flags, filter, type, data_set); } else if(then->rsc) { GListPtr gIter = NULL; GListPtr children = NULL; // Handle the 'primitive' ordering case changed |= native_update_actions(first, then, node, flags, filter, type, data_set); // Now any children (or containers in the case of a bundle) children = get_containers_or_children(then->rsc); for (gIter = children; gIter != NULL; gIter = gIter->next) { resource_t *then_child = (resource_t *) gIter->data; enum pe_graph_flags then_child_changed = pe_graph_none; action_t *then_child_action = find_first_action(then_child->actions, NULL, then->task, node); if (then_child_action) { enum pe_action_flags then_child_flags = then_child->cmds->action_flags(then_child_action, node); if (is_set(then_child_flags, pe_action_runnable)) { then_child_changed |= then_child->cmds->update_actions(first, then_child_action, node, flags, filter, type, data_set); } changed |= then_child_changed; if (then_child_changed & pe_graph_updated_then) { for (GListPtr lpc = then_child_action->actions_after; lpc != NULL; lpc = lpc->next) { action_wrapper_t *next = (action_wrapper_t *) lpc->data; update_action(next->action, data_set); } } } } if(children != then->rsc->children) { g_list_free(children); } } return changed; } void pcmk__bundle_rsc_location(pe_resource_t *rsc, pe__location_t *constraint) { pe__bundle_variant_data_t *bundle_data = NULL; get_bundle_variant_data(bundle_data, rsc); pe_rsc_trace(rsc, "Processing location constraint %s for %s", constraint->id, rsc->id); native_rsc_location(rsc, constraint); for (GList *gIter = bundle_data->replicas; gIter != NULL; gIter = gIter->next) { pe__bundle_replica_t *replica = gIter->data; if (replica->container) { replica->container->cmds->rsc_location(replica->container, constraint); } if (replica->ip) { replica->ip->cmds->rsc_location(replica->ip, constraint); } } if (bundle_data->child && ((constraint->role_filter == RSC_ROLE_SLAVE) || (constraint->role_filter == RSC_ROLE_MASTER))) { bundle_data->child->cmds->rsc_location(bundle_data->child, constraint); bundle_data->child->rsc_location = g_list_prepend(bundle_data->child->rsc_location, constraint); } } void pcmk__bundle_expand(pe_resource_t *rsc, pe_working_set_t * data_set) { pe__bundle_variant_data_t *bundle_data = NULL; CRM_CHECK(rsc != NULL, return); get_bundle_variant_data(bundle_data, rsc); if (bundle_data->child) { bundle_data->child->cmds->expand(bundle_data->child, data_set); } for (GList *gIter = bundle_data->replicas; gIter != NULL; gIter = gIter->next) { pe__bundle_replica_t *replica = gIter->data; CRM_ASSERT(replica); if (replica->remote && replica->container && 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("//nvpair[@name='" XML_RSC_ATTR_REMOTE_RA_ADDR "']", replica->remote->xml, LOG_ERR); const char *calculated_addr = NULL; calculated_addr = pe__add_bundle_remote_name(replica->remote, nvpair, "value"); if (calculated_addr) { crm_trace("Set address for bundle connection %s to bundle host %s", replica->remote->id, calculated_addr); g_hash_table_replace(replica->remote->parameters, strdup(XML_RSC_ATTR_REMOTE_RA_ADDR), strdup(calculated_addr)); } else { /* 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). */ crm_info("Unable to determine address for bundle %s remote connection", rsc->id); } } if (replica->ip) { replica->ip->cmds->expand(replica->ip, data_set); } if (replica->container) { replica->container->cmds->expand(replica->container, data_set); } if (replica->remote) { replica->remote->cmds->expand(replica->remote, data_set); } } } gboolean pcmk__bundle_create_probe(pe_resource_t *rsc, pe_node_t *node, pe_action_t *complete, gboolean force, pe_working_set_t * data_set) { bool any_created = FALSE; pe__bundle_variant_data_t *bundle_data = NULL; CRM_CHECK(rsc != NULL, return FALSE); get_bundle_variant_data(bundle_data, rsc); for (GList *gIter = bundle_data->replicas; gIter != NULL; gIter = gIter->next) { pe__bundle_replica_t *replica = gIter->data; CRM_ASSERT(replica); if (replica->ip) { any_created |= replica->ip->cmds->create_probe(replica->ip, node, complete, force, data_set); } if (replica->child && (node->details == replica->node->details)) { any_created |= replica->child->cmds->create_probe(replica->child, node, complete, force, data_set); } if (replica->container) { bool created = replica->container->cmds->create_probe(replica->container, node, complete, force, data_set); if(created) { 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 nreplicas_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 */ for (GList *tIter = bundle_data->replicas; tIter && (bundle_data->nreplicas_per_host == 1); tIter = tIter->next) { pe__bundle_replica_t *other = tIter->data; if ((other != replica) && (other != NULL) && (other->container != NULL)) { custom_action_order(replica->container, pcmk__op_key(replica->container->id, RSC_STATUS, 0), NULL, other->container, pcmk__op_key(other->container->id, RSC_START, 0), NULL, pe_order_optional|pe_order_same_node, data_set); } } } } if (replica->container && replica->remote && replica->remote->cmds->create_probe(replica->remote, node, complete, force, data_set)) { /* 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); action_t *probe = find_first_action(replica->remote->actions, probe_uuid, NULL, node); free(probe_uuid); if (probe) { any_created = TRUE; crm_trace("Ordering %s probe on %s", replica->remote->id, node->details->uname); custom_action_order(replica->container, pcmk__op_key(replica->container->id, RSC_START, 0), NULL, replica->remote, NULL, probe, pe_order_probe, data_set); } } } return any_created; } void pcmk__bundle_append_meta(pe_resource_t *rsc, xmlNode *xml) { } void pcmk__bundle_log_actions(pe_resource_t *rsc, pe_working_set_t *data_set, gboolean terminal) { pe__bundle_variant_data_t *bundle_data = NULL; CRM_CHECK(rsc != NULL, return); get_bundle_variant_data(bundle_data, rsc); for (GList *gIter = bundle_data->replicas; gIter != NULL; gIter = gIter->next) { pe__bundle_replica_t *replica = gIter->data; CRM_ASSERT(replica); if (replica->ip) { LogActions(replica->ip, data_set, terminal); } if (replica->container) { LogActions(replica->container, data_set, terminal); } if (replica->remote) { LogActions(replica->remote, data_set, terminal); } if (replica->child) { LogActions(replica->child, data_set, terminal); } } } diff --git a/lib/pacemaker/pcmk_sched_clone.c b/lib/pacemaker/pcmk_sched_clone.c index 848a86c4e5..ce665aeef8 100644 --- a/lib/pacemaker/pcmk_sched_clone.c +++ b/lib/pacemaker/pcmk_sched_clone.c @@ -1,1503 +1,1503 @@ /* * Copyright 2004-2020 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 #define VARIANT_CLONE 1 #include gint sort_clone_instance(gconstpointer a, gconstpointer b, gpointer data_set); static void append_parent_colocation(resource_t * rsc, resource_t * child, gboolean all); static gint sort_rsc_id(gconstpointer a, gconstpointer b) { const resource_t *resource1 = (const resource_t *)a; const resource_t *resource2 = (const resource_t *)b; long num1, num2; CRM_ASSERT(resource1 != NULL); CRM_ASSERT(resource2 != NULL); /* * Sort clone instances numerically by instance number, so instance :10 * comes after :9. */ num1 = strtol(strrchr(resource1->id, ':') + 1, NULL, 10); num2 = strtol(strrchr(resource2->id, ':') + 1, NULL, 10); if (num1 < num2) { return -1; } else if (num1 > num2) { return 1; } return 0; } static node_t * parent_node_instance(const resource_t * rsc, node_t * node) { node_t *ret = NULL; if (node != NULL && rsc->parent) { ret = pe_hash_table_lookup(rsc->parent->allowed_nodes, node->details->id); } else if(node != NULL) { ret = pe_hash_table_lookup(rsc->allowed_nodes, node->details->id); } return ret; } static gboolean did_fail(const resource_t * rsc) { GListPtr gIter = rsc->children; if (is_set(rsc->flags, pe_rsc_failed)) { return TRUE; } for (; gIter != NULL; gIter = gIter->next) { resource_t *child_rsc = (resource_t *) gIter->data; if (did_fail(child_rsc)) { return TRUE; } } return FALSE; } gint sort_clone_instance(gconstpointer a, gconstpointer b, gpointer data_set) { int rc = 0; node_t *node1 = NULL; node_t *node2 = NULL; node_t *current_node1 = NULL; node_t *current_node2 = NULL; unsigned int nnodes1 = 0; unsigned int nnodes2 = 0; gboolean can1 = TRUE; gboolean can2 = TRUE; const resource_t *resource1 = (const resource_t *)a; const resource_t *resource2 = (const resource_t *)b; CRM_ASSERT(resource1 != NULL); CRM_ASSERT(resource2 != NULL); /* allocation order: * - active instances * - instances running on nodes with the least copies * - active instances on nodes that can't support them or are to be fenced * - failed instances * - inactive instances */ current_node1 = pe__find_active_on(resource1, &nnodes1, NULL); current_node2 = pe__find_active_on(resource2, &nnodes2, NULL); if (nnodes1 && nnodes2) { if (nnodes1 < nnodes2) { crm_trace("%s < %s: running_on", resource1->id, resource2->id); return -1; } else if (nnodes1 > nnodes2) { crm_trace("%s > %s: running_on", resource1->id, resource2->id); return 1; } } node1 = current_node1; node2 = current_node2; if (node1) { node_t *match = pe_hash_table_lookup(resource1->allowed_nodes, node1->details->id); if (match == NULL || match->weight < 0) { crm_trace("%s: current location is unavailable", resource1->id); node1 = NULL; can1 = FALSE; } } if (node2) { node_t *match = pe_hash_table_lookup(resource2->allowed_nodes, node2->details->id); if (match == NULL || match->weight < 0) { crm_trace("%s: current location is unavailable", resource2->id); node2 = NULL; can2 = FALSE; } } if (can1 != can2) { if (can1) { crm_trace("%s < %s: availability of current location", resource1->id, resource2->id); return -1; } crm_trace("%s > %s: availability of current location", resource1->id, resource2->id); return 1; } if (resource1->priority < resource2->priority) { crm_trace("%s < %s: priority", resource1->id, resource2->id); return 1; } else if (resource1->priority > resource2->priority) { crm_trace("%s > %s: priority", resource1->id, resource2->id); return -1; } if (node1 == NULL && node2 == NULL) { crm_trace("%s == %s: not active", resource1->id, resource2->id); return 0; } if (node1 != node2) { if (node1 == NULL) { crm_trace("%s > %s: active", resource1->id, resource2->id); return 1; } else if (node2 == NULL) { crm_trace("%s < %s: active", resource1->id, resource2->id); return -1; } } can1 = can_run_resources(node1); can2 = can_run_resources(node2); if (can1 != can2) { if (can1) { crm_trace("%s < %s: can", resource1->id, resource2->id); return -1; } crm_trace("%s > %s: can", resource1->id, resource2->id); return 1; } node1 = parent_node_instance(resource1, node1); node2 = parent_node_instance(resource2, node2); if (node1 != NULL && node2 == NULL) { crm_trace("%s < %s: not allowed", resource1->id, resource2->id); return -1; } else if (node1 == NULL && node2 != NULL) { crm_trace("%s > %s: not allowed", resource1->id, resource2->id); return 1; } if (node1 == NULL || node2 == NULL) { crm_trace("%s == %s: not allowed", resource1->id, resource2->id); return 0; } if (node1->count < node2->count) { crm_trace("%s < %s: count", resource1->id, resource2->id); return -1; } else if (node1->count > node2->count) { crm_trace("%s > %s: count", resource1->id, resource2->id); return 1; } can1 = did_fail(resource1); can2 = did_fail(resource2); if (can1 != can2) { if (can1) { crm_trace("%s > %s: failed", resource1->id, resource2->id); return 1; } crm_trace("%s < %s: failed", resource1->id, resource2->id); return -1; } if (node1 && node2) { int lpc = 0; int max = 0; node_t *n = NULL; GListPtr gIter = NULL; GListPtr list1 = NULL; GListPtr list2 = NULL; GHashTable *hash1 = g_hash_table_new_full(crm_str_hash, g_str_equal, NULL, free); GHashTable *hash2 = g_hash_table_new_full(crm_str_hash, g_str_equal, NULL, free); n = node_copy(current_node1); g_hash_table_insert(hash1, (gpointer) n->details->id, n); n = node_copy(current_node2); g_hash_table_insert(hash2, (gpointer) n->details->id, n); if(resource1->parent) { for (gIter = resource1->parent->rsc_cons; gIter; gIter = gIter->next) { rsc_colocation_t *constraint = (rsc_colocation_t *) gIter->data; if (constraint->score == 0) { continue; } crm_trace("Applying %s to %s", constraint->id, resource1->id); hash1 = pcmk__native_merge_weights(constraint->rsc_rh, resource1->id, hash1, constraint->node_attribute, constraint->score / (float) INFINITY, 0); } for (gIter = resource1->parent->rsc_cons_lhs; gIter; gIter = gIter->next) { rsc_colocation_t *constraint = (rsc_colocation_t *) gIter->data; if (constraint->score == 0) { continue; } crm_trace("Applying %s to %s", constraint->id, resource1->id); hash1 = pcmk__native_merge_weights(constraint->rsc_lh, resource1->id, hash1, constraint->node_attribute, constraint->score / (float) INFINITY, pe_weights_positive); } } if(resource2->parent) { for (gIter = resource2->parent->rsc_cons; gIter; gIter = gIter->next) { rsc_colocation_t *constraint = (rsc_colocation_t *) gIter->data; crm_trace("Applying %s to %s", constraint->id, resource2->id); hash2 = pcmk__native_merge_weights(constraint->rsc_rh, resource2->id, hash2, constraint->node_attribute, constraint->score / (float) INFINITY, 0); } for (gIter = resource2->parent->rsc_cons_lhs; gIter; gIter = gIter->next) { rsc_colocation_t *constraint = (rsc_colocation_t *) gIter->data; crm_trace("Applying %s to %s", constraint->id, resource2->id); hash2 = pcmk__native_merge_weights(constraint->rsc_lh, resource2->id, hash2, constraint->node_attribute, constraint->score / (float) INFINITY, pe_weights_positive); } } /* Current location score */ node1 = g_hash_table_lookup(hash1, current_node1->details->id); node2 = g_hash_table_lookup(hash2, current_node2->details->id); if (node1->weight < node2->weight) { if (node1->weight < 0) { crm_trace("%s > %s: current score: %d %d", resource1->id, resource2->id, node1->weight, node2->weight); rc = -1; goto out; } else { crm_trace("%s < %s: current score: %d %d", resource1->id, resource2->id, node1->weight, node2->weight); rc = 1; goto out; } } else if (node1->weight > node2->weight) { crm_trace("%s > %s: current score: %d %d", resource1->id, resource2->id, node1->weight, node2->weight); rc = -1; goto out; } /* All location scores */ list1 = g_hash_table_get_values(hash1); list2 = g_hash_table_get_values(hash2); list1 = sort_nodes_by_weight(list1, current_node1, data_set); list2 = sort_nodes_by_weight(list2, current_node2, data_set); max = g_list_length(list1); if (max < g_list_length(list2)) { max = g_list_length(list2); } for (; lpc < max; lpc++) { node1 = g_list_nth_data(list1, lpc); node2 = g_list_nth_data(list2, lpc); if (node1 == NULL) { crm_trace("%s < %s: colocated score NULL", resource1->id, resource2->id); rc = 1; break; } else if (node2 == NULL) { crm_trace("%s > %s: colocated score NULL", resource1->id, resource2->id); rc = -1; break; } if (node1->weight < node2->weight) { crm_trace("%s < %s: colocated score", resource1->id, resource2->id); rc = 1; break; } else if (node1->weight > node2->weight) { crm_trace("%s > %s: colocated score", resource1->id, resource2->id); rc = -1; break; } } /* Order by reverse uname - same as sort_node_weight() does? */ out: g_hash_table_destroy(hash1); /* Free mem */ g_hash_table_destroy(hash2); /* Free mem */ g_list_free(list1); g_list_free(list2); if (rc != 0) { return rc; } } rc = strcmp(resource1->id, resource2->id); crm_trace("%s %c %s: default", resource1->id, rc < 0 ? '<' : '>', resource2->id); return rc; } static node_t * can_run_instance(resource_t * rsc, node_t * node, int limit) { node_t *local_node = NULL; if (node == NULL && rsc->allowed_nodes) { GHashTableIter iter; g_hash_table_iter_init(&iter, rsc->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (void **)&local_node)) { can_run_instance(rsc, local_node, limit); } return NULL; } if (can_run_resources(node) == FALSE) { goto bail; } else if (is_set(rsc->flags, pe_rsc_orphan)) { goto bail; } local_node = parent_node_instance(rsc, node); if (local_node == NULL) { crm_warn("%s cannot run on %s: node not allowed", rsc->id, node->details->uname); goto bail; } else if (local_node->weight < 0) { common_update_score(rsc, node->details->id, local_node->weight); pe_rsc_trace(rsc, "%s cannot run on %s: Parent node weight doesn't allow it.", rsc->id, node->details->uname); } else if (local_node->count < limit) { pe_rsc_trace(rsc, "%s can run on %s (already running %d)", rsc->id, node->details->uname, local_node->count); return local_node; } else { pe_rsc_trace(rsc, "%s cannot run on %s: node full (%d >= %d)", rsc->id, node->details->uname, local_node->count, limit); } bail: if (node) { common_update_score(rsc, node->details->id, -INFINITY); } return NULL; } static pe_node_t * allocate_instance(pe_resource_t *rsc, pe_node_t *prefer, gboolean all_coloc, int limit, pe_working_set_t *data_set) { node_t *chosen = NULL; GHashTable *backup = NULL; CRM_ASSERT(rsc); pe_rsc_trace(rsc, "Checking allocation of %s (preferring %s, using %s parent colocations)", rsc->id, (prefer? prefer->details->uname: "none"), (all_coloc? "all" : "some")); if (is_not_set(rsc->flags, pe_rsc_provisional)) { return rsc->fns->location(rsc, NULL, FALSE); } else if (is_set(rsc->flags, pe_rsc_allocating)) { pe_rsc_debug(rsc, "Dependency loop detected involving %s", rsc->id); return NULL; } /* Only include positive colocation preferences of dependent resources * if not every node will get a copy of the clone */ append_parent_colocation(rsc->parent, rsc, all_coloc); if (prefer) { node_t *local_prefer = g_hash_table_lookup(rsc->allowed_nodes, prefer->details->id); if (local_prefer == NULL || local_prefer->weight < 0) { pe_rsc_trace(rsc, "Not pre-allocating %s to %s - unavailable", rsc->id, prefer->details->uname); return NULL; } } can_run_instance(rsc, NULL, limit); backup = node_hash_dup(rsc->allowed_nodes); + pe_rsc_trace(rsc, "Allocating instance %s", rsc->id); chosen = rsc->cmds->allocate(rsc, prefer, data_set); if (chosen && prefer && (chosen->details != prefer->details)) { crm_info("Not pre-allocating %s to %s because %s is better", rsc->id, prefer->details->uname, chosen->details->uname); g_hash_table_destroy(rsc->allowed_nodes); rsc->allowed_nodes = backup; native_deallocate(rsc); chosen = NULL; backup = NULL; } if (chosen) { pe_node_t *local_node = parent_node_instance(rsc, chosen); if (local_node) { local_node->count++; } else if (is_set(rsc->flags, pe_rsc_managed)) { /* what to do? we can't enforce per-node limits in this case */ crm_config_err("%s not found in %s (list=%d)", chosen->details->id, rsc->parent->id, g_hash_table_size(rsc->parent->allowed_nodes)); } } if(backup) { g_hash_table_destroy(backup); } return chosen; } static void append_parent_colocation(resource_t * rsc, resource_t * child, gboolean all) { GListPtr gIter = NULL; gIter = rsc->rsc_cons; for (; gIter != NULL; gIter = gIter->next) { rsc_colocation_t *cons = (rsc_colocation_t *) gIter->data; if (cons->score == 0) { continue; } if (all || cons->score < 0 || cons->score == INFINITY) { child->rsc_cons = g_list_prepend(child->rsc_cons, cons); } } gIter = rsc->rsc_cons_lhs; for (; gIter != NULL; gIter = gIter->next) { rsc_colocation_t *cons = (rsc_colocation_t *) gIter->data; if (cons->score == 0) { continue; } if (all || cons->score < 0) { child->rsc_cons_lhs = g_list_prepend(child->rsc_cons_lhs, cons); } } } void distribute_children(resource_t *rsc, GListPtr children, GListPtr nodes, int max, int per_host_max, pe_working_set_t * data_set); void distribute_children(resource_t *rsc, GListPtr children, GListPtr nodes, int max, int per_host_max, pe_working_set_t * data_set) { int loop_max = 0; int allocated = 0; int available_nodes = 0; /* count now tracks the number of clones currently allocated */ for(GListPtr nIter = nodes; nIter != NULL; nIter = nIter->next) { pe_node_t *node = nIter->data; node->count = 0; if (can_run_resources(node)) { available_nodes++; } } if(available_nodes) { loop_max = max / available_nodes; } if (loop_max < 1) { loop_max = 1; } pe_rsc_debug(rsc, "Allocating up to %d %s instances to a possible %d nodes (at most %d per host, %d optimal)", max, rsc->id, available_nodes, per_host_max, loop_max); /* Pre-allocate as many instances as we can to their current location */ for (GListPtr gIter = children; gIter != NULL && allocated < max; gIter = gIter->next) { resource_t *child = (resource_t *) gIter->data; if (child->running_on && is_set(child->flags, pe_rsc_provisional) && is_not_set(child->flags, pe_rsc_failed)) { node_t *child_node = pe__current_node(child); node_t *local_node = parent_node_instance(child, child_node); pe_rsc_trace(rsc, "Checking pre-allocation of %s to %s (%d remaining of %d)", child->id, child_node->details->uname, max - allocated, max); if (can_run_resources(child_node) == FALSE || child_node->weight < 0) { pe_rsc_trace(rsc, "Not pre-allocating because %s can not run %s", child_node->details->uname, child->id); } else if(local_node && local_node->count >= loop_max) { pe_rsc_trace(rsc, "Not pre-allocating because %s already allocated optimal instances", child_node->details->uname); } else if (allocate_instance(child, child_node, max < available_nodes, per_host_max, data_set)) { pe_rsc_trace(rsc, "Pre-allocated %s to %s", child->id, child_node->details->uname); allocated++; } } } pe_rsc_trace(rsc, "Done pre-allocating (%d of %d)", allocated, max); for (GListPtr gIter = children; gIter != NULL; gIter = gIter->next) { resource_t *child = (resource_t *) gIter->data; if (child->running_on != NULL) { node_t *child_node = pe__current_node(child); node_t *local_node = parent_node_instance(child, child_node); if (local_node == NULL) { crm_err("%s is running on %s which isn't allowed", child->id, child_node->details->uname); } } if (is_not_set(child->flags, pe_rsc_provisional)) { } else if (allocated >= max) { pe_rsc_debug(rsc, "Child %s not allocated - limit reached %d %d", child->id, allocated, max); resource_location(child, NULL, -INFINITY, "clone:limit_reached", data_set); } else { if (allocate_instance(child, NULL, max < available_nodes, per_host_max, data_set)) { allocated++; } } } pe_rsc_debug(rsc, "Allocated %d %s instances of a possible %d", allocated, rsc->id, max); } pe_node_t * pcmk__clone_allocate(pe_resource_t *rsc, pe_node_t *prefer, pe_working_set_t *data_set) { GListPtr nodes = NULL; clone_variant_data_t *clone_data = NULL; get_clone_variant_data(clone_data, rsc); if (is_not_set(rsc->flags, pe_rsc_provisional)) { return NULL; } else if (is_set(rsc->flags, pe_rsc_allocating)) { pe_rsc_debug(rsc, "Dependency loop detected involving %s", rsc->id); return NULL; } if (is_set(rsc->flags, pe_rsc_promotable)) { apply_master_prefs(rsc); } set_bit(rsc->flags, pe_rsc_allocating); - pe_rsc_trace(rsc, "Processing %s", rsc->id); /* this information is used by sort_clone_instance() when deciding in which * order to allocate clone instances */ for (GListPtr gIter = rsc->rsc_cons; gIter != NULL; gIter = gIter->next) { rsc_colocation_t *constraint = (rsc_colocation_t *) gIter->data; if (constraint->score == 0) { continue; } pe_rsc_trace(rsc, "%s: Allocating %s first", rsc->id, constraint->rsc_rh->id); constraint->rsc_rh->cmds->allocate(constraint->rsc_rh, prefer, data_set); } for (GListPtr gIter = rsc->rsc_cons_lhs; gIter != NULL; gIter = gIter->next) { rsc_colocation_t *constraint = (rsc_colocation_t *) gIter->data; if (constraint->score == 0) { continue; } rsc->allowed_nodes = constraint->rsc_lh->cmds->merge_weights(constraint->rsc_lh, rsc->id, rsc->allowed_nodes, constraint->node_attribute, (float)constraint->score / INFINITY, (pe_weights_rollback | pe_weights_positive)); } pe__show_node_weights(!show_scores, rsc, __FUNCTION__, rsc->allowed_nodes); nodes = g_hash_table_get_values(rsc->allowed_nodes); nodes = sort_nodes_by_weight(nodes, NULL, data_set); rsc->children = g_list_sort_with_data(rsc->children, sort_clone_instance, data_set); distribute_children(rsc, rsc->children, nodes, clone_data->clone_max, clone_data->clone_node_max, data_set); g_list_free(nodes); if (is_set(rsc->flags, pe_rsc_promotable)) { pcmk__set_instance_roles(rsc, data_set); } clear_bit(rsc->flags, pe_rsc_provisional); clear_bit(rsc->flags, pe_rsc_allocating); pe_rsc_trace(rsc, "Done allocating %s", rsc->id); return NULL; } static void clone_update_pseudo_status(resource_t * rsc, gboolean * stopping, gboolean * starting, gboolean * active) { GListPtr gIter = NULL; if (rsc->children) { gIter = rsc->children; for (; gIter != NULL; gIter = gIter->next) { resource_t *child = (resource_t *) gIter->data; clone_update_pseudo_status(child, stopping, starting, active); } return; } CRM_ASSERT(active != NULL); CRM_ASSERT(starting != NULL); CRM_ASSERT(stopping != NULL); if (rsc->running_on) { *active = TRUE; } gIter = rsc->actions; for (; gIter != NULL; gIter = gIter->next) { action_t *action = (action_t *) gIter->data; if (*starting && *stopping) { return; } else if (is_set(action->flags, pe_action_optional)) { pe_rsc_trace(rsc, "Skipping optional: %s", action->uuid); continue; } else if (is_set(action->flags, pe_action_pseudo) == FALSE && is_set(action->flags, pe_action_runnable) == FALSE) { pe_rsc_trace(rsc, "Skipping unrunnable: %s", action->uuid); continue; } else if (safe_str_eq(RSC_STOP, action->task)) { pe_rsc_trace(rsc, "Stopping due to: %s", action->uuid); *stopping = TRUE; } else if (safe_str_eq(RSC_START, action->task)) { if (is_set(action->flags, pe_action_runnable) == FALSE) { pe_rsc_trace(rsc, "Skipping pseudo-op: %s run=%d, pseudo=%d", action->uuid, is_set(action->flags, pe_action_runnable), is_set(action->flags, pe_action_pseudo)); } else { pe_rsc_trace(rsc, "Starting due to: %s", action->uuid); pe_rsc_trace(rsc, "%s run=%d, pseudo=%d", action->uuid, is_set(action->flags, pe_action_runnable), is_set(action->flags, pe_action_pseudo)); *starting = TRUE; } } } } static action_t * find_rsc_action(pe_resource_t *rsc, const char *task, gboolean active_only, GList **list) { action_t *match = NULL; GListPtr possible = NULL; GListPtr active = NULL; possible = pe__resource_actions(rsc, NULL, task, FALSE); if (active_only) { GListPtr gIter = possible; for (; gIter != NULL; gIter = gIter->next) { action_t *op = (action_t *) gIter->data; if (is_set(op->flags, pe_action_optional) == FALSE) { active = g_list_prepend(active, op); } } if (active && pcmk__list_of_1(active)) { match = g_list_nth_data(active, 0); } if (list) { *list = active; active = NULL; } } else if (possible && pcmk__list_of_1(possible)) { match = g_list_nth_data(possible, 0); } if (list) { *list = possible; possible = NULL; } if (possible) { g_list_free(possible); } if (active) { g_list_free(active); } return match; } static void child_ordering_constraints(resource_t * rsc, pe_working_set_t * data_set) { action_t *stop = NULL; action_t *start = NULL; action_t *last_stop = NULL; action_t *last_start = NULL; GListPtr gIter = NULL; gboolean active_only = TRUE; /* change to false to get the old behavior */ clone_variant_data_t *clone_data = NULL; get_clone_variant_data(clone_data, rsc); if (clone_data->ordered == FALSE) { return; } /* we have to maintain a consistent sorted child list when building order constraints */ rsc->children = g_list_sort(rsc->children, sort_rsc_id); for (gIter = rsc->children; gIter != NULL; gIter = gIter->next) { resource_t *child = (resource_t *) gIter->data; stop = find_rsc_action(child, RSC_STOP, active_only, NULL); if (stop) { if (last_stop) { /* child/child relative stop */ order_actions(stop, last_stop, pe_order_optional); } last_stop = stop; } start = find_rsc_action(child, RSC_START, active_only, NULL); if (start) { if (last_start) { /* child/child relative start */ order_actions(last_start, start, pe_order_optional); } last_start = start; } } } void clone_create_actions(resource_t *rsc, pe_working_set_t *data_set) { clone_variant_data_t *clone_data = NULL; get_clone_variant_data(clone_data, rsc); clone_create_pseudo_actions(rsc, rsc->children, &clone_data->start_notify, &clone_data->stop_notify,data_set); child_ordering_constraints(rsc, data_set); if (is_set(rsc->flags, pe_rsc_promotable)) { create_promotable_actions(rsc, data_set); } } void clone_create_pseudo_actions( resource_t * rsc, GListPtr children, notify_data_t **start_notify, notify_data_t **stop_notify, pe_working_set_t * data_set) { gboolean child_active = FALSE; gboolean child_starting = FALSE; gboolean child_stopping = FALSE; gboolean allow_dependent_migrations = TRUE; action_t *stop = NULL; action_t *stopped = NULL; action_t *start = NULL; action_t *started = NULL; pe_rsc_trace(rsc, "Creating actions for %s", rsc->id); for (GListPtr gIter = children; gIter != NULL; gIter = gIter->next) { resource_t *child_rsc = (resource_t *) gIter->data; gboolean starting = FALSE; gboolean stopping = FALSE; child_rsc->cmds->create_actions(child_rsc, data_set); clone_update_pseudo_status(child_rsc, &stopping, &starting, &child_active); if (stopping && starting) { allow_dependent_migrations = FALSE; } child_stopping |= stopping; child_starting |= starting; } /* start */ start = create_pseudo_resource_op(rsc, RSC_START, !child_starting, TRUE, data_set); started = create_pseudo_resource_op(rsc, RSC_STARTED, !child_starting, FALSE, data_set); started->priority = INFINITY; if (child_active || child_starting) { update_action_flags(started, pe_action_runnable, __FUNCTION__, __LINE__); } if (start_notify != NULL && *start_notify == NULL) { *start_notify = create_notification_boundaries(rsc, RSC_START, start, started, data_set); } /* stop */ stop = create_pseudo_resource_op(rsc, RSC_STOP, !child_stopping, TRUE, data_set); stopped = create_pseudo_resource_op(rsc, RSC_STOPPED, !child_stopping, TRUE, data_set); stopped->priority = INFINITY; if (allow_dependent_migrations) { update_action_flags(stop, pe_action_migrate_runnable, __FUNCTION__, __LINE__); } if (stop_notify != NULL && *stop_notify == NULL) { *stop_notify = create_notification_boundaries(rsc, RSC_STOP, stop, stopped, data_set); if (start_notify && *start_notify && *stop_notify) { order_actions((*stop_notify)->post_done, (*start_notify)->pre, pe_order_optional); } } } void clone_internal_constraints(resource_t *rsc, pe_working_set_t *data_set) { resource_t *last_rsc = NULL; GListPtr gIter; clone_variant_data_t *clone_data = NULL; get_clone_variant_data(clone_data, rsc); pe_rsc_trace(rsc, "Internal constraints for %s", rsc->id); new_rsc_order(rsc, RSC_STOPPED, rsc, RSC_START, pe_order_optional, data_set); new_rsc_order(rsc, RSC_START, rsc, RSC_STARTED, pe_order_runnable_left, data_set); new_rsc_order(rsc, RSC_STOP, rsc, RSC_STOPPED, pe_order_runnable_left, data_set); if (is_set(rsc->flags, pe_rsc_promotable)) { new_rsc_order(rsc, RSC_DEMOTED, rsc, RSC_STOP, pe_order_optional, data_set); new_rsc_order(rsc, RSC_STARTED, rsc, RSC_PROMOTE, pe_order_runnable_left, data_set); } if (clone_data->ordered) { /* we have to maintain a consistent sorted child list when building order constraints */ rsc->children = g_list_sort(rsc->children, sort_rsc_id); } for (gIter = rsc->children; gIter != NULL; gIter = gIter->next) { resource_t *child_rsc = (resource_t *) gIter->data; child_rsc->cmds->internal_constraints(child_rsc, data_set); order_start_start(rsc, child_rsc, pe_order_runnable_left | pe_order_implies_first_printed); new_rsc_order(child_rsc, RSC_START, rsc, RSC_STARTED, pe_order_implies_then_printed, data_set); if (clone_data->ordered && last_rsc) { order_start_start(last_rsc, child_rsc, pe_order_optional); } order_stop_stop(rsc, child_rsc, pe_order_implies_first_printed); new_rsc_order(child_rsc, RSC_STOP, rsc, RSC_STOPPED, pe_order_implies_then_printed, data_set); if (clone_data->ordered && last_rsc) { order_stop_stop(child_rsc, last_rsc, pe_order_optional); } last_rsc = child_rsc; } if (is_set(rsc->flags, pe_rsc_promotable)) { promotable_constraints(rsc, data_set); } } bool assign_node(resource_t * rsc, node_t * node, gboolean force) { bool changed = FALSE; if (rsc->children) { for (GListPtr gIter = rsc->children; gIter != NULL; gIter = gIter->next) { resource_t *child_rsc = (resource_t *) gIter->data; changed |= assign_node(child_rsc, node, force); } return changed; } if (rsc->allocated_to != NULL) { changed = true; } native_assign_node(rsc, NULL, node, force); return changed; } gboolean is_child_compatible(resource_t *child_rsc, node_t * local_node, enum rsc_role_e filter, gboolean current) { node_t *node = NULL; enum rsc_role_e next_role = child_rsc->fns->state(child_rsc, current); CRM_CHECK(child_rsc && local_node, return FALSE); if (is_set_recursive(child_rsc, pe_rsc_block, TRUE) == FALSE) { /* We only want instances that haven't failed */ node = child_rsc->fns->location(child_rsc, NULL, current); } if (filter != RSC_ROLE_UNKNOWN && next_role != filter) { crm_trace("Filtered %s", child_rsc->id); return FALSE; } if (node && (node->details == local_node->details)) { return TRUE; } else if (node) { crm_trace("%s - %s vs %s", child_rsc->id, node->details->uname, local_node->details->uname); } else { crm_trace("%s - not allocated %d", child_rsc->id, current); } return FALSE; } pe_resource_t * find_compatible_child(pe_resource_t *local_child, pe_resource_t *rsc, enum rsc_role_e filter, gboolean current, pe_working_set_t *data_set) { resource_t *pair = NULL; GListPtr gIter = NULL; GListPtr scratch = NULL; node_t *local_node = NULL; local_node = local_child->fns->location(local_child, NULL, current); if (local_node) { return find_compatible_child_by_node(local_child, local_node, rsc, filter, current); } scratch = g_hash_table_get_values(local_child->allowed_nodes); scratch = sort_nodes_by_weight(scratch, NULL, data_set); gIter = scratch; for (; gIter != NULL; gIter = gIter->next) { node_t *node = (node_t *) gIter->data; pair = find_compatible_child_by_node(local_child, node, rsc, filter, current); if (pair) { goto done; } } pe_rsc_debug(rsc, "Can't pair %s with %s", local_child->id, rsc->id); done: g_list_free(scratch); return pair; } void clone_rsc_colocation_lh(pe_resource_t *rsc_lh, pe_resource_t *rsc_rh, rsc_colocation_t *constraint, pe_working_set_t *data_set) { /* -- Never called -- * * Instead we add the colocation constraints to the child and call from there */ CRM_ASSERT(FALSE); } void clone_rsc_colocation_rh(pe_resource_t *rsc_lh, pe_resource_t *rsc_rh, rsc_colocation_t *constraint, pe_working_set_t *data_set) { GListPtr gIter = NULL; gboolean do_interleave = FALSE; const char *interleave_s = NULL; CRM_CHECK(constraint != NULL, return); CRM_CHECK(rsc_lh != NULL, pe_err("rsc_lh was NULL for %s", constraint->id); return); CRM_CHECK(rsc_rh != NULL, pe_err("rsc_rh was NULL for %s", constraint->id); return); CRM_CHECK(rsc_lh->variant == pe_native, return); if (constraint->score == 0) { return; } pe_rsc_trace(rsc_rh, "Processing constraint %s: %s -> %s %d", constraint->id, rsc_lh->id, rsc_rh->id, constraint->score); if (is_set(rsc_rh->flags, pe_rsc_promotable)) { if (is_set(rsc_rh->flags, pe_rsc_provisional)) { pe_rsc_trace(rsc_rh, "%s is still provisional", rsc_rh->id); return; } else if (constraint->role_rh == RSC_ROLE_UNKNOWN) { pe_rsc_trace(rsc_rh, "Handling %s as a clone colocation", constraint->id); } else { promotable_colocation_rh(rsc_lh, rsc_rh, constraint, data_set); return; } } /* only the LHS side needs to be labeled as interleave */ interleave_s = g_hash_table_lookup(constraint->rsc_lh->meta, XML_RSC_ATTR_INTERLEAVE); if(crm_is_true(interleave_s) && constraint->rsc_lh->variant > pe_group) { // TODO: Do we actually care about multiple RH copies sharing a LH copy anymore? if (copies_per_node(constraint->rsc_lh) != copies_per_node(constraint->rsc_rh)) { crm_config_err("Cannot interleave %s and %s because" " they do not support the same number of copies per node", constraint->rsc_lh->id, constraint->rsc_rh->id); } else { do_interleave = TRUE; } } if (is_set(rsc_rh->flags, pe_rsc_provisional)) { pe_rsc_trace(rsc_rh, "%s is still provisional", rsc_rh->id); return; } else if (do_interleave) { resource_t *rh_child = NULL; rh_child = find_compatible_child(rsc_lh, rsc_rh, RSC_ROLE_UNKNOWN, FALSE, data_set); if (rh_child) { pe_rsc_debug(rsc_rh, "Pairing %s with %s", rsc_lh->id, rh_child->id); rsc_lh->cmds->rsc_colocation_lh(rsc_lh, rh_child, constraint, data_set); } else if (constraint->score >= INFINITY) { crm_notice("Cannot pair %s with instance of %s", rsc_lh->id, rsc_rh->id); assign_node(rsc_lh, NULL, TRUE); } else { pe_rsc_debug(rsc_rh, "Cannot pair %s with instance of %s", rsc_lh->id, rsc_rh->id); } return; } else if (constraint->score >= INFINITY) { GListPtr rhs = NULL; gIter = rsc_rh->children; for (; gIter != NULL; gIter = gIter->next) { resource_t *child_rsc = (resource_t *) gIter->data; node_t *chosen = child_rsc->fns->location(child_rsc, NULL, FALSE); if (chosen != NULL && is_set_recursive(child_rsc, pe_rsc_block, TRUE) == FALSE) { pe_rsc_trace(rsc_rh, "Allowing %s: %s %d", constraint->id, chosen->details->uname, chosen->weight); rhs = g_list_prepend(rhs, chosen); } } node_list_exclude(rsc_lh->allowed_nodes, rhs, FALSE); g_list_free(rhs); return; } gIter = rsc_rh->children; for (; gIter != NULL; gIter = gIter->next) { resource_t *child_rsc = (resource_t *) gIter->data; child_rsc->cmds->rsc_colocation_rh(rsc_lh, child_rsc, constraint, data_set); } } enum action_tasks clone_child_action(action_t * action) { enum action_tasks result = no_action; resource_t *child = (resource_t *) action->rsc->children->data; if (safe_str_eq(action->task, "notify") || safe_str_eq(action->task, "notified")) { /* Find the action we're notifying about instead */ int stop = 0; char *key = action->uuid; int lpc = strlen(key); for (; lpc > 0; lpc--) { if (key[lpc] == '_' && stop == 0) { stop = lpc; } else if (key[lpc] == '_') { char *task_mutable = NULL; lpc++; task_mutable = strdup(key + lpc); task_mutable[stop - lpc] = 0; crm_trace("Extracted action '%s' from '%s'", task_mutable, key); result = get_complex_task(child, task_mutable, TRUE); free(task_mutable); break; } } } else { result = get_complex_task(child, action->task, TRUE); } return result; } enum pe_action_flags summary_action_flags(action_t * action, GListPtr children, node_t * node) { GListPtr gIter = NULL; gboolean any_runnable = FALSE; gboolean check_runnable = TRUE; enum action_tasks task = clone_child_action(action); enum pe_action_flags flags = (pe_action_optional | pe_action_runnable | pe_action_pseudo); const char *task_s = task2text(task); for (gIter = children; gIter != NULL; gIter = gIter->next) { action_t *child_action = NULL; resource_t *child = (resource_t *) gIter->data; child_action = find_first_action(child->actions, NULL, task_s, child->children ? NULL : node); pe_rsc_trace(action->rsc, "Checking for %s in %s on %s (%s)", task_s, child->id, node ? node->details->uname : "none", child_action?child_action->uuid:"NA"); if (child_action) { enum pe_action_flags child_flags = child->cmds->action_flags(child_action, node); if (is_set(flags, pe_action_optional) && is_set(child_flags, pe_action_optional) == FALSE) { pe_rsc_trace(child, "%s is mandatory because of %s", action->uuid, child_action->uuid); flags = crm_clear_bit(__FUNCTION__, __LINE__, action->rsc->id, flags, pe_action_optional); pe_clear_action_bit(action, pe_action_optional); } if (is_set(child_flags, pe_action_runnable)) { any_runnable = TRUE; } } } if (check_runnable && any_runnable == FALSE) { pe_rsc_trace(action->rsc, "%s is not runnable because no children are", action->uuid); flags = crm_clear_bit(__FUNCTION__, __LINE__, action->rsc->id, flags, pe_action_runnable); if (node == NULL) { pe_clear_action_bit(action, pe_action_runnable); } } return flags; } enum pe_action_flags clone_action_flags(action_t * action, node_t * node) { return summary_action_flags(action, action->rsc->children, node); } void clone_rsc_location(pe_resource_t *rsc, pe__location_t *constraint) { GListPtr gIter = rsc->children; pe_rsc_trace(rsc, "Processing location constraint %s for %s", constraint->id, rsc->id); native_rsc_location(rsc, constraint); for (; gIter != NULL; gIter = gIter->next) { resource_t *child_rsc = (resource_t *) gIter->data; child_rsc->cmds->rsc_location(child_rsc, constraint); } } void clone_expand(resource_t * rsc, pe_working_set_t * data_set) { GListPtr gIter = NULL; clone_variant_data_t *clone_data = NULL; get_clone_variant_data(clone_data, rsc); gIter = rsc->actions; for (; gIter != NULL; gIter = gIter->next) { action_t *op = (action_t *) gIter->data; rsc->cmds->action_flags(op, NULL); } if (clone_data->start_notify) { collect_notification_data(rsc, TRUE, TRUE, clone_data->start_notify); expand_notification_data(rsc, clone_data->start_notify, data_set); create_notifications(rsc, clone_data->start_notify, data_set); } if (clone_data->stop_notify) { collect_notification_data(rsc, TRUE, TRUE, clone_data->stop_notify); expand_notification_data(rsc, clone_data->stop_notify, data_set); create_notifications(rsc, clone_data->stop_notify, data_set); } if (clone_data->promote_notify) { collect_notification_data(rsc, TRUE, TRUE, clone_data->promote_notify); expand_notification_data(rsc, clone_data->promote_notify, data_set); create_notifications(rsc, clone_data->promote_notify, data_set); } if (clone_data->demote_notify) { collect_notification_data(rsc, TRUE, TRUE, clone_data->demote_notify); expand_notification_data(rsc, clone_data->demote_notify, data_set); create_notifications(rsc, clone_data->demote_notify, data_set); } /* Now that the notifcations have been created we can expand the children */ gIter = rsc->children; for (; gIter != NULL; gIter = gIter->next) { resource_t *child_rsc = (resource_t *) gIter->data; child_rsc->cmds->expand(child_rsc, data_set); } native_expand(rsc, data_set); /* The notifications are in the graph now, we can destroy the notify_data */ free_notification_data(clone_data->demote_notify); clone_data->demote_notify = NULL; free_notification_data(clone_data->stop_notify); clone_data->stop_notify = NULL; free_notification_data(clone_data->start_notify); clone_data->start_notify = NULL; free_notification_data(clone_data->promote_notify); clone_data->promote_notify = NULL; } // Check whether a resource or any of its children is known on node static bool rsc_known_on(const pe_resource_t *rsc, const pe_node_t *node) { if (rsc->children) { for (GList *child_iter = rsc->children; child_iter != NULL; child_iter = child_iter->next) { resource_t *child = (resource_t *) child_iter->data; if (rsc_known_on(child, node)) { return TRUE; } } } else if (rsc->known_on) { GHashTableIter iter; 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 (node->details == known_node->details) { return TRUE; } } } return FALSE; } // Look for an instance of clone that is known on node static pe_resource_t * find_instance_on(const pe_resource_t *clone, const pe_node_t *node) { for (GList *gIter = clone->children; gIter != NULL; gIter = gIter->next) { resource_t *child = (resource_t *) gIter->data; if (rsc_known_on(child, node)) { return child; } } return NULL; } // For unique clones, probe each instance separately static gboolean probe_unique_clone(pe_resource_t *rsc, pe_node_t *node, pe_action_t *complete, gboolean force, pe_working_set_t *data_set) { gboolean any_created = FALSE; for (GList *child_iter = rsc->children; child_iter != NULL; child_iter = child_iter->next) { resource_t *child = (resource_t *) child_iter->data; any_created |= child->cmds->create_probe(child, node, complete, force, data_set); } return any_created; } // For anonymous clones, only a single instance needs to be probed static gboolean probe_anonymous_clone(pe_resource_t *rsc, pe_node_t *node, pe_action_t *complete, gboolean force, pe_working_set_t *data_set) { // First, check if we probed an instance on this node last time pe_resource_t *child = find_instance_on(rsc, node); // Otherwise, check if we plan to start an instance on this node if (child == NULL) { for (GList *child_iter = rsc->children; child_iter && !child; child_iter = child_iter->next) { node_t *local_node = NULL; resource_t *child_rsc = (resource_t *) child_iter->data; local_node = child_rsc->fns->location(child_rsc, NULL, FALSE); if (local_node && (local_node->details == node->details)) { child = child_rsc; } } } // Otherwise, use the first clone instance if (child == NULL) { child = rsc->children->data; } return child->cmds->create_probe(child, node, complete, force, data_set); } gboolean clone_create_probe(resource_t * rsc, node_t * node, action_t * complete, gboolean force, pe_working_set_t * data_set) { gboolean any_created = FALSE; CRM_ASSERT(rsc); rsc->children = g_list_sort(rsc->children, sort_rsc_id); if (rsc->children == NULL) { pe_warn("Clone %s has no children", rsc->id); return FALSE; } if (rsc->exclusive_discover) { node_t *allowed = g_hash_table_lookup(rsc->allowed_nodes, node->details->id); if (allowed && allowed->rsc_discover_mode != pe_discover_exclusive) { /* exclusive discover is enabled and this node is not marked * as a node this resource should be discovered on * * remove the node from allowed_nodes so that the * notification contains only nodes that we might ever run * on */ g_hash_table_remove(rsc->allowed_nodes, node->details->id); /* Bit of a shortcut - might as well take it */ return FALSE; } } if (is_set(rsc->flags, pe_rsc_unique)) { any_created = probe_unique_clone(rsc, node, complete, force, data_set); } else { any_created = probe_anonymous_clone(rsc, node, complete, force, data_set); } return any_created; } void clone_append_meta(resource_t * rsc, xmlNode * xml) { char *name = NULL; clone_variant_data_t *clone_data = NULL; get_clone_variant_data(clone_data, rsc); name = crm_meta_name(XML_RSC_ATTR_UNIQUE); crm_xml_add(xml, name, is_set(rsc->flags, pe_rsc_unique) ? "true" : "false"); free(name); name = crm_meta_name(XML_RSC_ATTR_NOTIFY); crm_xml_add(xml, name, is_set(rsc->flags, pe_rsc_notify) ? "true" : "false"); free(name); name = crm_meta_name(XML_RSC_ATTR_INCARNATION_MAX); crm_xml_add_int(xml, name, clone_data->clone_max); free(name); name = crm_meta_name(XML_RSC_ATTR_INCARNATION_NODEMAX); crm_xml_add_int(xml, name, clone_data->clone_node_max); free(name); if (is_set(rsc->flags, pe_rsc_promotable)) { name = crm_meta_name(XML_RSC_ATTR_PROMOTED_MAX); crm_xml_add_int(xml, name, clone_data->promoted_max); free(name); name = crm_meta_name(XML_RSC_ATTR_PROMOTED_NODEMAX); crm_xml_add_int(xml, name, clone_data->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(XML_RSC_ATTR_MASTER_MAX); crm_xml_add_int(xml, name, clone_data->promoted_max); free(name); name = crm_meta_name(XML_RSC_ATTR_MASTER_NODEMAX); crm_xml_add_int(xml, name, clone_data->promoted_node_max); free(name); } } diff --git a/lib/pacemaker/pcmk_sched_group.c b/lib/pacemaker/pcmk_sched_group.c index de01ec92a0..6d8c9d113b 100644 --- a/lib/pacemaker/pcmk_sched_group.c +++ b/lib/pacemaker/pcmk_sched_group.c @@ -1,530 +1,531 @@ /* * Copyright 2004-2019 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 #define VARIANT_GROUP 1 #include pe_node_t * pcmk__group_allocate(pe_resource_t *rsc, pe_node_t *prefer, pe_working_set_t *data_set) { node_t *node = NULL; node_t *group_node = NULL; GListPtr gIter = NULL; group_variant_data_t *group_data = NULL; get_group_variant_data(group_data, rsc); if (is_not_set(rsc->flags, pe_rsc_provisional)) { return rsc->allocated_to; } - pe_rsc_trace(rsc, "Processing %s", rsc->id); if (is_set(rsc->flags, pe_rsc_allocating)) { pe_rsc_debug(rsc, "Dependency loop detected involving %s", rsc->id); return NULL; } if (group_data->first_child == NULL) { // Nothing to allocate clear_bit(rsc->flags, pe_rsc_provisional); return NULL; } set_bit(rsc->flags, pe_rsc_allocating); rsc->role = group_data->first_child->role; group_data->first_child->rsc_cons = g_list_concat(group_data->first_child->rsc_cons, rsc->rsc_cons); rsc->rsc_cons = NULL; group_data->last_child->rsc_cons_lhs = g_list_concat(group_data->last_child->rsc_cons_lhs, rsc->rsc_cons_lhs); rsc->rsc_cons_lhs = NULL; pe__show_node_weights(!show_scores, rsc, __FUNCTION__, rsc->allowed_nodes); gIter = rsc->children; for (; gIter != NULL; gIter = gIter->next) { resource_t *child_rsc = (resource_t *) gIter->data; + pe_rsc_trace(rsc, "Allocating group %s member %s", + rsc->id, child_rsc->id); node = child_rsc->cmds->allocate(child_rsc, prefer, data_set); if (group_node == NULL) { group_node = node; } } rsc->next_role = group_data->first_child->next_role; clear_bit(rsc->flags, pe_rsc_allocating); clear_bit(rsc->flags, pe_rsc_provisional); if (group_data->colocated) { return group_node; } return NULL; } void group_update_pseudo_status(resource_t * parent, resource_t * child); void group_create_actions(resource_t * rsc, pe_working_set_t * data_set) { action_t *op = NULL; const char *value = NULL; GListPtr gIter = rsc->children; pe_rsc_trace(rsc, "Creating actions for %s", rsc->id); for (; gIter != NULL; gIter = gIter->next) { resource_t *child_rsc = (resource_t *) gIter->data; child_rsc->cmds->create_actions(child_rsc, data_set); group_update_pseudo_status(rsc, child_rsc); } op = start_action(rsc, NULL, TRUE /* !group_data->child_starting */ ); set_bit(op->flags, pe_action_pseudo | pe_action_runnable); op = custom_action(rsc, started_key(rsc), RSC_STARTED, NULL, TRUE /* !group_data->child_starting */ , TRUE, data_set); set_bit(op->flags, pe_action_pseudo | pe_action_runnable); op = stop_action(rsc, NULL, TRUE /* !group_data->child_stopping */ ); set_bit(op->flags, pe_action_pseudo | pe_action_runnable); op = custom_action(rsc, stopped_key(rsc), RSC_STOPPED, NULL, TRUE /* !group_data->child_stopping */ , TRUE, data_set); set_bit(op->flags, pe_action_pseudo | pe_action_runnable); value = g_hash_table_lookup(rsc->meta, XML_RSC_ATTR_PROMOTABLE); if (crm_is_true(value)) { op = custom_action(rsc, demote_key(rsc), RSC_DEMOTE, NULL, TRUE, TRUE, data_set); set_bit(op->flags, pe_action_pseudo); set_bit(op->flags, pe_action_runnable); op = custom_action(rsc, demoted_key(rsc), RSC_DEMOTED, NULL, TRUE, TRUE, data_set); set_bit(op->flags, pe_action_pseudo); set_bit(op->flags, pe_action_runnable); op = custom_action(rsc, promote_key(rsc), RSC_PROMOTE, NULL, TRUE, TRUE, data_set); set_bit(op->flags, pe_action_pseudo); set_bit(op->flags, pe_action_runnable); op = custom_action(rsc, promoted_key(rsc), RSC_PROMOTED, NULL, TRUE, TRUE, data_set); set_bit(op->flags, pe_action_pseudo); set_bit(op->flags, pe_action_runnable); } } void group_update_pseudo_status(resource_t * parent, resource_t * child) { GListPtr gIter = child->actions; group_variant_data_t *group_data = NULL; get_group_variant_data(group_data, parent); if (group_data->ordered == FALSE) { /* If this group is not ordered, then leave the meta-actions as optional */ return; } if (group_data->child_stopping && group_data->child_starting) { return; } for (; gIter != NULL; gIter = gIter->next) { action_t *action = (action_t *) gIter->data; if (is_set(action->flags, pe_action_optional)) { continue; } if (safe_str_eq(RSC_STOP, action->task) && is_set(action->flags, pe_action_runnable)) { group_data->child_stopping = TRUE; pe_rsc_trace(action->rsc, "Based on %s the group is stopping", action->uuid); } else if (safe_str_eq(RSC_START, action->task) && is_set(action->flags, pe_action_runnable)) { group_data->child_starting = TRUE; pe_rsc_trace(action->rsc, "Based on %s the group is starting", action->uuid); } } } void group_internal_constraints(resource_t * rsc, pe_working_set_t * data_set) { GListPtr gIter = rsc->children; resource_t *last_rsc = NULL; resource_t *last_active = NULL; resource_t *top = uber_parent(rsc); group_variant_data_t *group_data = NULL; get_group_variant_data(group_data, rsc); new_rsc_order(rsc, RSC_STOPPED, rsc, RSC_START, pe_order_optional, data_set); new_rsc_order(rsc, RSC_START, rsc, RSC_STARTED, pe_order_runnable_left, data_set); new_rsc_order(rsc, RSC_STOP, rsc, RSC_STOPPED, pe_order_runnable_left, data_set); for (; gIter != NULL; gIter = gIter->next) { resource_t *child_rsc = (resource_t *) gIter->data; int stop = pe_order_none; int stopped = pe_order_implies_then_printed; int start = pe_order_implies_then | pe_order_runnable_left; int started = pe_order_runnable_left | pe_order_implies_then | pe_order_implies_then_printed; child_rsc->cmds->internal_constraints(child_rsc, data_set); if (last_rsc == NULL) { if (group_data->ordered) { stop |= pe_order_optional; stopped = pe_order_implies_then; } } else if (group_data->colocated) { rsc_colocation_new("group:internal_colocation", NULL, INFINITY, child_rsc, last_rsc, NULL, NULL, data_set); } if (is_set(top->flags, pe_rsc_promotable)) { new_rsc_order(rsc, RSC_DEMOTE, child_rsc, RSC_DEMOTE, stop | pe_order_implies_first_printed, data_set); new_rsc_order(child_rsc, RSC_DEMOTE, rsc, RSC_DEMOTED, stopped, data_set); new_rsc_order(child_rsc, RSC_PROMOTE, rsc, RSC_PROMOTED, started, data_set); new_rsc_order(rsc, RSC_PROMOTE, child_rsc, RSC_PROMOTE, pe_order_implies_first_printed, data_set); } order_start_start(rsc, child_rsc, pe_order_implies_first_printed); order_stop_stop(rsc, child_rsc, stop | pe_order_implies_first_printed); new_rsc_order(child_rsc, RSC_STOP, rsc, RSC_STOPPED, stopped, data_set); new_rsc_order(child_rsc, RSC_START, rsc, RSC_STARTED, started, data_set); if (group_data->ordered == FALSE) { order_start_start(rsc, child_rsc, start | pe_order_implies_first_printed); if (is_set(top->flags, pe_rsc_promotable)) { new_rsc_order(rsc, RSC_PROMOTE, child_rsc, RSC_PROMOTE, start | pe_order_implies_first_printed, data_set); } } else if (last_rsc != NULL) { child_rsc->restart_type = pe_restart_restart; order_start_start(last_rsc, child_rsc, start); order_stop_stop(child_rsc, last_rsc, pe_order_optional | pe_order_restart); if (is_set(top->flags, pe_rsc_promotable)) { new_rsc_order(last_rsc, RSC_PROMOTE, child_rsc, RSC_PROMOTE, start, data_set); new_rsc_order(child_rsc, RSC_DEMOTE, last_rsc, RSC_DEMOTE, pe_order_optional, data_set); } } else { /* If anyone in the group is starting, then * pe_order_implies_then will cause _everyone_ in the group * to be sent a start action * But this is safe since starting something that is already * started is required to be "safe" */ int flags = pe_order_none; order_start_start(rsc, child_rsc, flags); if (is_set(top->flags, pe_rsc_promotable)) { new_rsc_order(rsc, RSC_PROMOTE, child_rsc, RSC_PROMOTE, flags, data_set); } } /* Look for partially active groups * Make sure they still shut down in sequence */ if (child_rsc->running_on) { if (group_data->ordered && last_rsc && last_rsc->running_on == NULL && last_active && last_active->running_on) { order_stop_stop(child_rsc, last_active, pe_order_optional); } last_active = child_rsc; } last_rsc = child_rsc; } if (group_data->ordered && last_rsc != NULL) { int stop_stop_flags = pe_order_implies_then; int stop_stopped_flags = pe_order_optional; order_stop_stop(rsc, last_rsc, stop_stop_flags); new_rsc_order(last_rsc, RSC_STOP, rsc, RSC_STOPPED, stop_stopped_flags, data_set); if (is_set(top->flags, pe_rsc_promotable)) { new_rsc_order(rsc, RSC_DEMOTE, last_rsc, RSC_DEMOTE, stop_stop_flags, data_set); new_rsc_order(last_rsc, RSC_DEMOTE, rsc, RSC_DEMOTED, stop_stopped_flags, data_set); } } } void group_rsc_colocation_lh(pe_resource_t *rsc_lh, pe_resource_t *rsc_rh, rsc_colocation_t *constraint, pe_working_set_t *data_set) { GListPtr gIter = NULL; group_variant_data_t *group_data = NULL; if (rsc_lh == NULL) { pe_err("rsc_lh was NULL for %s", constraint->id); return; } else if (rsc_rh == NULL) { pe_err("rsc_rh was NULL for %s", constraint->id); return; } if (constraint->score == 0) { return; } gIter = rsc_lh->children; pe_rsc_trace(rsc_lh, "Processing constraints from %s", rsc_lh->id); get_group_variant_data(group_data, rsc_lh); if (group_data->colocated) { group_data->first_child->cmds->rsc_colocation_lh(group_data->first_child, rsc_rh, constraint, data_set); return; } else if (constraint->score >= INFINITY) { crm_config_err("%s: Cannot perform mandatory colocation" " between non-colocated group and %s", rsc_lh->id, rsc_rh->id); return; } for (; gIter != NULL; gIter = gIter->next) { resource_t *child_rsc = (resource_t *) gIter->data; child_rsc->cmds->rsc_colocation_lh(child_rsc, rsc_rh, constraint, data_set); } } void group_rsc_colocation_rh(pe_resource_t *rsc_lh, pe_resource_t *rsc_rh, rsc_colocation_t *constraint, pe_working_set_t *data_set) { GListPtr gIter = rsc_rh->children; group_variant_data_t *group_data = NULL; get_group_variant_data(group_data, rsc_rh); CRM_CHECK(rsc_lh->variant == pe_native, return); if (constraint->score == 0) { return; } pe_rsc_trace(rsc_rh, "Processing RH %s of constraint %s (LH is %s)", rsc_rh->id, constraint->id, rsc_lh->id); if (is_set(rsc_rh->flags, pe_rsc_provisional)) { return; } else if (group_data->colocated && group_data->first_child) { if (constraint->score >= INFINITY) { /* Ensure RHS is _fully_ up before can start LHS */ group_data->last_child->cmds->rsc_colocation_rh(rsc_lh, group_data->last_child, constraint, data_set); } else { /* A partially active RHS is fine */ group_data->first_child->cmds->rsc_colocation_rh(rsc_lh, group_data->first_child, constraint, data_set); } return; } else if (constraint->score >= INFINITY) { crm_config_err("%s: Cannot perform mandatory colocation with" " non-colocated group: %s", rsc_lh->id, rsc_rh->id); return; } for (; gIter != NULL; gIter = gIter->next) { resource_t *child_rsc = (resource_t *) gIter->data; child_rsc->cmds->rsc_colocation_rh(rsc_lh, child_rsc, constraint, data_set); } } enum pe_action_flags group_action_flags(action_t * action, node_t * node) { GListPtr gIter = NULL; enum pe_action_flags flags = (pe_action_optional | pe_action_runnable | pe_action_pseudo); for (gIter = action->rsc->children; gIter != NULL; gIter = gIter->next) { resource_t *child = (resource_t *) gIter->data; enum action_tasks task = get_complex_task(child, action->task, TRUE); const char *task_s = task2text(task); action_t *child_action = find_first_action(child->actions, NULL, task_s, node); if (child_action) { enum pe_action_flags child_flags = child->cmds->action_flags(child_action, node); if (is_set(flags, pe_action_optional) && is_set(child_flags, pe_action_optional) == FALSE) { pe_rsc_trace(action->rsc, "%s is mandatory because of %s", action->uuid, child_action->uuid); clear_bit(flags, pe_action_optional); pe_clear_action_bit(action, pe_action_optional); } if (safe_str_neq(task_s, action->task) && is_set(flags, pe_action_runnable) && is_set(child_flags, pe_action_runnable) == FALSE) { pe_rsc_trace(action->rsc, "%s is not runnable because of %s", action->uuid, child_action->uuid); clear_bit(flags, pe_action_runnable); pe_clear_action_bit(action, pe_action_runnable); } } else if (task != stop_rsc && task != action_demote) { pe_rsc_trace(action->rsc, "%s is not runnable because of %s (not found in %s)", action->uuid, task_s, child->id); clear_bit(flags, pe_action_runnable); } } return flags; } enum pe_graph_flags group_update_actions(pe_action_t *first, pe_action_t *then, pe_node_t *node, enum pe_action_flags flags, enum pe_action_flags filter, enum pe_ordering type, pe_working_set_t *data_set) { GListPtr gIter = then->rsc->children; enum pe_graph_flags changed = pe_graph_none; CRM_ASSERT(then->rsc != NULL); changed |= native_update_actions(first, then, node, flags, filter, type, data_set); for (; gIter != NULL; gIter = gIter->next) { resource_t *child = (resource_t *) gIter->data; action_t *child_action = find_first_action(child->actions, NULL, then->task, node); if (child_action) { changed |= child->cmds->update_actions(first, child_action, node, flags, filter, type, data_set); } } return changed; } void group_rsc_location(pe_resource_t *rsc, pe__location_t *constraint) { GListPtr gIter = rsc->children; GListPtr saved = constraint->node_list_rh; GListPtr zero = node_list_dup(constraint->node_list_rh, TRUE, FALSE); gboolean reset_scores = TRUE; group_variant_data_t *group_data = NULL; get_group_variant_data(group_data, rsc); pe_rsc_debug(rsc, "Processing rsc_location %s for %s", constraint->id, rsc->id); native_rsc_location(rsc, constraint); for (; gIter != NULL; gIter = gIter->next) { resource_t *child_rsc = (resource_t *) gIter->data; child_rsc->cmds->rsc_location(child_rsc, constraint); if (group_data->colocated && reset_scores) { reset_scores = FALSE; constraint->node_list_rh = zero; } } constraint->node_list_rh = saved; g_list_free_full(zero, free); } void group_expand(resource_t * rsc, pe_working_set_t * data_set) { CRM_CHECK(rsc != NULL, return); pe_rsc_trace(rsc, "Processing actions from %s", rsc->id); native_expand(rsc, data_set); for (GListPtr gIter = rsc->children; gIter != NULL; gIter = gIter->next) { resource_t *child_rsc = (resource_t *) gIter->data; child_rsc->cmds->expand(child_rsc, data_set); } } GHashTable * pcmk__group_merge_weights(pe_resource_t *rsc, const char *rhs, GHashTable *nodes, const char *attr, float factor, uint32_t flags) { GListPtr gIter = rsc->rsc_cons_lhs; group_variant_data_t *group_data = NULL; get_group_variant_data(group_data, rsc); if (is_set(rsc->flags, pe_rsc_merging)) { pe_rsc_info(rsc, "Breaking dependency loop with %s at %s", rsc->id, rhs); return nodes; } set_bit(rsc->flags, pe_rsc_merging); nodes = group_data->first_child->cmds->merge_weights(group_data->first_child, rhs, nodes, attr, factor, flags); for (; gIter != NULL; gIter = gIter->next) { rsc_colocation_t *constraint = (rsc_colocation_t *) gIter->data; if (constraint->score == 0) { continue; } nodes = pcmk__native_merge_weights(constraint->rsc_lh, rsc->id, nodes, constraint->node_attribute, constraint->score / (float) INFINITY, flags); } clear_bit(rsc->flags, pe_rsc_merging); return nodes; } void group_append_meta(resource_t * rsc, xmlNode * xml) { } diff --git a/lib/pacemaker/pcmk_sched_native.c b/lib/pacemaker/pcmk_sched_native.c index 5312150ed8..aae6974942 100644 --- a/lib/pacemaker/pcmk_sched_native.c +++ b/lib/pacemaker/pcmk_sched_native.c @@ -1,3389 +1,3391 @@ /* * Copyright 2004-2020 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 // The controller removes the resource from the CIB, making this redundant // #define DELETE_THEN_REFRESH 1 #define INFINITY_HACK (INFINITY * -100) #define VARIANT_NATIVE 1 #include void native_rsc_colocation_rh_must(resource_t * rsc_lh, gboolean update_lh, resource_t * rsc_rh, gboolean update_rh); void native_rsc_colocation_rh_mustnot(resource_t * rsc_lh, gboolean update_lh, resource_t * rsc_rh, gboolean update_rh); static void Recurring(resource_t *rsc, action_t *start, node_t *node, pe_working_set_t *data_set); static void RecurringOp(resource_t *rsc, action_t *start, node_t *node, xmlNode *operation, pe_working_set_t *data_set); static void Recurring_Stopped(resource_t *rsc, action_t *start, node_t *node, pe_working_set_t *data_set); static void RecurringOp_Stopped(resource_t *rsc, action_t *start, node_t *node, xmlNode *operation, pe_working_set_t *data_set); void ReloadRsc(resource_t * rsc, node_t *node, pe_working_set_t * data_set); gboolean DeleteRsc(resource_t * rsc, node_t * node, gboolean optional, pe_working_set_t * data_set); gboolean StopRsc(resource_t * rsc, node_t * next, gboolean optional, pe_working_set_t * data_set); gboolean StartRsc(resource_t * rsc, node_t * next, gboolean optional, pe_working_set_t * data_set); gboolean DemoteRsc(resource_t * rsc, node_t * next, gboolean optional, pe_working_set_t * data_set); gboolean PromoteRsc(resource_t * rsc, node_t * next, gboolean optional, pe_working_set_t * data_set); gboolean RoleError(resource_t * rsc, node_t * next, gboolean optional, pe_working_set_t * data_set); gboolean NullOp(resource_t * rsc, node_t * next, gboolean optional, pe_working_set_t * data_set); /* *INDENT-OFF* */ enum rsc_role_e rsc_state_matrix[RSC_ROLE_MAX][RSC_ROLE_MAX] = { /* Current State */ /* Next State: Unknown Stopped Started Slave Master */ /* Unknown */ { RSC_ROLE_UNKNOWN, RSC_ROLE_STOPPED, RSC_ROLE_STOPPED, RSC_ROLE_STOPPED, RSC_ROLE_STOPPED, }, /* Stopped */ { RSC_ROLE_STOPPED, RSC_ROLE_STOPPED, RSC_ROLE_STARTED, RSC_ROLE_SLAVE, RSC_ROLE_SLAVE, }, /* Started */ { RSC_ROLE_STOPPED, RSC_ROLE_STOPPED, RSC_ROLE_STARTED, RSC_ROLE_SLAVE, RSC_ROLE_MASTER, }, /* Slave */ { RSC_ROLE_STOPPED, RSC_ROLE_STOPPED, RSC_ROLE_STOPPED, RSC_ROLE_SLAVE, RSC_ROLE_MASTER, }, /* Master */ { RSC_ROLE_STOPPED, RSC_ROLE_SLAVE, RSC_ROLE_SLAVE, RSC_ROLE_SLAVE, RSC_ROLE_MASTER, }, }; gboolean (*rsc_action_matrix[RSC_ROLE_MAX][RSC_ROLE_MAX])(resource_t*,node_t*,gboolean,pe_working_set_t*) = { /* Current State */ /* Next State: Unknown Stopped Started Slave Master */ /* Unknown */ { RoleError, StopRsc, RoleError, RoleError, RoleError, }, /* Stopped */ { RoleError, NullOp, StartRsc, StartRsc, RoleError, }, /* Started */ { RoleError, StopRsc, NullOp, NullOp, PromoteRsc, }, /* Slave */ { RoleError, StopRsc, StopRsc, NullOp, PromoteRsc, }, /* Master */ { RoleError, DemoteRsc, DemoteRsc, DemoteRsc, NullOp, }, }; /* *INDENT-ON* */ static gboolean native_choose_node(resource_t * rsc, node_t * prefer, pe_working_set_t * data_set) { GListPtr nodes = NULL; node_t *chosen = NULL; node_t *best = NULL; int multiple = 1; int length = 0; gboolean result = FALSE; process_utilization(rsc, &prefer, data_set); if (is_not_set(rsc->flags, pe_rsc_provisional)) { return rsc->allocated_to ? TRUE : FALSE; } // Sort allowed nodes by weight if (rsc->allowed_nodes) { length = g_hash_table_size(rsc->allowed_nodes); } if (length > 0) { nodes = g_hash_table_get_values(rsc->allowed_nodes); nodes = sort_nodes_by_weight(nodes, pe__current_node(rsc), data_set); // First node in sorted list has the best score best = g_list_nth_data(nodes, 0); } if (prefer && nodes) { 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", prefer->details->uname, rsc->id); /* Favor the preferred node as long as its weight 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 weight is less than INFINITY. */ } else if ((chosen->weight < 0) || (chosen->weight < best->weight)) { pe_rsc_trace(rsc, "Preferred node %s for %s was unsuitable", chosen->details->uname, rsc->id); chosen = NULL; } else if (!can_run_resources(chosen)) { pe_rsc_trace(rsc, "Preferred node %s for %s was unavailable", chosen->details->uname, rsc->id); chosen = NULL; } else { pe_rsc_trace(rsc, "Chose preferred node %s for %s (ignoring %d candidates)", chosen->details->uname, rsc->id, length); } } if ((chosen == NULL) && nodes) { /* Either there is no preferred node, or the preferred node is not * available, but there are other nodes allowed to run the resource. */ chosen = best; pe_rsc_trace(rsc, "Chose node %s for %s from %d candidates", chosen ? chosen->details->uname : "", rsc->id, length); if (!pe_rsc_is_unique_clone(rsc->parent) && chosen && (chosen->weight > 0) && can_run_resources(chosen)) { /* 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 * distribute_children() has already assigned instances to their * running nodes when appropriate, and if we get here, we don't want * remaining unallocated instances to prefer a node that's already * running another instance. */ node_t *running = pe__current_node(rsc); if (running && (can_run_resources(running) == FALSE)) { pe_rsc_trace(rsc, "Current node for %s (%s) can't run resources", rsc->id, running->details->uname); } else if (running) { for (GList *iter = nodes->next; iter; iter = iter->next) { node_t *tmp = (node_t *) iter->data; if (tmp->weight != chosen->weight) { // The nodes are sorted by weight, so no more are equal break; } if (tmp->details == running->details) { // Scores are equal, so prefer the current node chosen = tmp; } multiple++; } } } } if (multiple > 1) { static char score[33]; int log_level = (chosen->weight >= INFINITY)? LOG_WARNING : LOG_INFO; score2char_stack(chosen->weight, score, sizeof(score)); do_crm_log(log_level, "Chose node %s for %s from %d nodes with score %s", chosen->details->uname, rsc->id, multiple, score); } result = native_assign_node(rsc, nodes, chosen, FALSE); g_list_free(nodes); return result; } static int node_list_attr_score(GHashTable * list, const char *attr, const char *value) { GHashTableIter iter; node_t *node = NULL; int best_score = -INFINITY; const char *best_node = NULL; if (attr == NULL) { attr = CRM_ATTR_UNAME; } g_hash_table_iter_init(&iter, list); while (g_hash_table_iter_next(&iter, NULL, (void **)&node)) { int weight = node->weight; if (can_run_resources(node) == FALSE) { weight = -INFINITY; } if (weight > best_score || best_node == NULL) { const char *tmp = pe_node_attribute_raw(node, attr); if (safe_str_eq(value, tmp)) { best_score = weight; best_node = node->details->uname; } } } if (safe_str_neq(attr, CRM_ATTR_UNAME)) { crm_info("Best score for %s=%s was %s with %d", attr, value, best_node ? best_node : "", best_score); } return best_score; } static void node_hash_update(GHashTable * list1, GHashTable * list2, const char *attr, float factor, gboolean only_positive) { int score = 0; int new_score = 0; GHashTableIter iter; node_t *node = NULL; if (attr == NULL) { attr = CRM_ATTR_UNAME; } g_hash_table_iter_init(&iter, list1); while (g_hash_table_iter_next(&iter, NULL, (void **)&node)) { float weight_f = 0; int weight = 0; score = node_list_attr_score(list2, attr, pe_node_attribute_raw(node, attr)); if ((factor < 0) && (score < 0)) { /* Negative preference for a node with a negative score * should not become a positive preference. * * @TODO Consider filtering only if weight == -INFINITY */ crm_trace("%s: Filtering %d + %f * %d (double negative disallowed)", node->details->uname, node->weight, factor, score); continue; } if (node->weight == INFINITY_HACK) { crm_trace("%s: Filtering %d + %f * %d (node marked unusable)", node->details->uname, node->weight, factor, score); continue; } weight_f = factor * score; // Round the number; see http://c-faq.com/fp/round.html weight = (int) ((weight_f < 0)? (weight_f - 0.5) : (weight_f + 0.5)); new_score = merge_weights(weight, node->weight); if (only_positive && (new_score < 0) && (node->weight > 0)) { crm_trace("%s: Filtering %d + %f * %d = %d " "(negative disallowed, marking node unusable)", node->details->uname, node->weight, factor, score, new_score); node->weight = INFINITY_HACK; continue; } if (only_positive && (new_score < 0) && (node->weight == 0)) { crm_trace("%s: Filtering %d + %f * %d = %d (negative disallowed)", node->details->uname, node->weight, factor, score, new_score); continue; } crm_trace("%s: %d + %f * %d = %d", node->details->uname, node->weight, factor, score, new_score); node->weight = new_score; } } GHashTable * node_hash_dup(GHashTable * hash) { /* Hack! */ GListPtr list = g_hash_table_get_values(hash); GHashTable *result = node_hash_from_list(list); g_list_free(list); return result; } /*! * \internal * \brief Incorporate colocation constraint scores into node weights * * \param[in,out] rsc Resource being placed * \param[in] rhs ID of 'with' resource * \param[in,out] nodes Nodes, with scores as of this point * \param[in] attr Colocation attribute (ID by default) * \param[in] factor Incorporate scores multiplied by this factor * \param[in] flags Bitmask of enum pe_weights values * * \return Nodes, with scores modified by this constraint * \note This function assumes ownership of the nodes argument. The caller * should free the returned copy rather than the original. */ GHashTable * pcmk__native_merge_weights(pe_resource_t *rsc, const char *rhs, GHashTable *nodes, const char *attr, float factor, uint32_t flags) { GHashTable *work = NULL; // Avoid infinite recursion if (is_set(rsc->flags, pe_rsc_merging)) { pe_rsc_info(rsc, "%s: Breaking dependency loop at %s", rhs, rsc->id); return nodes; } set_bit(rsc->flags, pe_rsc_merging); if (is_set(flags, pe_weights_init)) { if ((rsc->variant == pe_group) && (rsc->children != NULL)) { GList *last = g_list_last(rsc->children); pe_resource_t *last_rsc = last->data; pe_rsc_trace(rsc, "%s: Merging scores from group %s " "using last member %s (at %.6f)", rhs, rsc->id, last_rsc->id, factor); work = pcmk__native_merge_weights(last_rsc, rhs, NULL, attr, factor, flags); } else { work = node_hash_dup(rsc->allowed_nodes); } clear_bit(flags, pe_weights_init); } else if ((rsc->variant == pe_group) && (rsc->children != NULL)) { pe_rsc_trace(rsc, "%s: Merging scores from %d members of group %s (at %.6f)", rhs, g_list_length(rsc->children), rsc->id, factor); work = node_hash_dup(nodes); for (GList *iter = rsc->children; iter->next != NULL; iter = iter->next) { work = pcmk__native_merge_weights(iter->data, rhs, work, attr, factor, flags); } } else { pe_rsc_trace(rsc, "%s: Merging scores from %s (at %.6f)", rhs, rsc->id, factor); work = node_hash_dup(nodes); node_hash_update(work, rsc->allowed_nodes, attr, factor, is_set(flags, pe_weights_positive)); } if (can_run_any(work)) { GListPtr gIter = NULL; int multiplier = (factor < 0)? -1 : 1; if (is_set(flags, pe_weights_forward)) { gIter = rsc->rsc_cons; pe_rsc_trace(rsc, "Checking additional %d optional '%s with' constraints", g_list_length(gIter), rsc->id); } else if ((rsc->variant == pe_group) && (rsc->children != NULL)) { GList *last = g_list_last(rsc->children); pe_resource_t *last_rsc = last->data; gIter = last_rsc->rsc_cons_lhs; pe_rsc_trace(rsc, "Checking additional %d optional 'with group %s' " "constraints using last member %s", g_list_length(gIter), rsc->id, last_rsc->id); } else { gIter = rsc->rsc_cons_lhs; pe_rsc_trace(rsc, "Checking additional %d optional 'with %s' constraints", g_list_length(gIter), rsc->id); } for (; gIter != NULL; gIter = gIter->next) { pe_resource_t *other = NULL; rsc_colocation_t *constraint = (rsc_colocation_t *) gIter->data; if (constraint->score == 0) { continue; } if (is_set(flags, pe_weights_forward)) { other = constraint->rsc_rh; } else { other = constraint->rsc_lh; } pe_rsc_trace(rsc, "Optionally merging score of '%s' constraint (%s with %s)", constraint->id, constraint->rsc_lh->id, constraint->rsc_rh->id); work = pcmk__native_merge_weights(other, rhs, work, constraint->node_attribute, multiplier * constraint->score / (float) INFINITY, flags|pe_weights_rollback); pe__show_node_weights(true, NULL, rhs, work); } } else if (is_set(flags, pe_weights_rollback)) { pe_rsc_info(rsc, "%s: Rolling back optional scores from %s", rhs, rsc->id); g_hash_table_destroy(work); clear_bit(rsc->flags, pe_rsc_merging); return nodes; } if (is_set(flags, pe_weights_positive)) { node_t *node = NULL; GHashTableIter iter; g_hash_table_iter_init(&iter, work); while (g_hash_table_iter_next(&iter, NULL, (void **)&node)) { if (node->weight == INFINITY_HACK) { node->weight = 1; } } } if (nodes) { g_hash_table_destroy(nodes); } clear_bit(rsc->flags, pe_rsc_merging); return work; } static inline bool node_has_been_unfenced(node_t *node) { const char *unfenced = pe_node_attribute_raw(node, CRM_ATTR_UNFENCED); return unfenced && strcmp("0", unfenced); } static inline bool is_unfence_device(resource_t *rsc, pe_working_set_t *data_set) { return is_set(rsc->flags, pe_rsc_fence_device) && is_set(data_set->flags, pe_flag_enable_unfencing); } pe_node_t * pcmk__native_allocate(pe_resource_t *rsc, pe_node_t *prefer, pe_working_set_t *data_set) { GListPtr gIter = NULL; if (rsc->parent && is_not_set(rsc->parent->flags, pe_rsc_allocating)) { /* never allocate children on their own */ pe_rsc_debug(rsc, "Escalating allocation of %s to its parent: %s", rsc->id, rsc->parent->id); rsc->parent->cmds->allocate(rsc->parent, prefer, data_set); } if (is_not_set(rsc->flags, pe_rsc_provisional)) { return rsc->allocated_to; } if (is_set(rsc->flags, pe_rsc_allocating)) { pe_rsc_debug(rsc, "Dependency loop detected involving %s", rsc->id); return NULL; } set_bit(rsc->flags, pe_rsc_allocating); pe__show_node_weights(true, rsc, "Pre-alloc", rsc->allowed_nodes); for (gIter = rsc->rsc_cons; gIter != NULL; gIter = gIter->next) { rsc_colocation_t *constraint = (rsc_colocation_t *) gIter->data; GHashTable *archive = NULL; resource_t *rsc_rh = constraint->rsc_rh; if (constraint->score == 0) { continue; } - pe_rsc_trace(rsc, "%s: Pre-Processing %s (%s, %d, %s)", - rsc->id, constraint->id, rsc_rh->id, - constraint->score, role2text(constraint->role_lh)); if (constraint->role_lh >= RSC_ROLE_MASTER || (constraint->score < 0 && constraint->score > -INFINITY)) { archive = node_hash_dup(rsc->allowed_nodes); } + + pe_rsc_trace(rsc, + "%s: Allocating %s first (constraint=%s score=%d role=%s)", + rsc->id, rsc_rh->id, constraint->id, + constraint->score, role2text(constraint->role_lh)); rsc_rh->cmds->allocate(rsc_rh, NULL, data_set); rsc->cmds->rsc_colocation_lh(rsc, rsc_rh, constraint, data_set); if (archive && can_run_any(rsc->allowed_nodes) == FALSE) { pe_rsc_info(rsc, "%s: Rolling back scores from %s", rsc->id, rsc_rh->id); g_hash_table_destroy(rsc->allowed_nodes); rsc->allowed_nodes = archive; archive = NULL; } if (archive) { g_hash_table_destroy(archive); } } pe__show_node_weights(true, rsc, "Post-coloc", rsc->allowed_nodes); for (gIter = rsc->rsc_cons_lhs; gIter != NULL; gIter = gIter->next) { rsc_colocation_t *constraint = (rsc_colocation_t *) gIter->data; if (constraint->score == 0) { continue; } pe_rsc_trace(rsc, "Merging score of '%s' constraint (%s with %s)", constraint->id, constraint->rsc_lh->id, constraint->rsc_rh->id); rsc->allowed_nodes = constraint->rsc_lh->cmds->merge_weights(constraint->rsc_lh, rsc->id, rsc->allowed_nodes, constraint->node_attribute, (float)constraint->score / INFINITY, pe_weights_rollback); } if (rsc->next_role == RSC_ROLE_STOPPED) { pe_rsc_trace(rsc, "Making sure %s doesn't get allocated", rsc->id); /* make sure it doesn't come up again */ resource_location(rsc, NULL, -INFINITY, XML_RSC_ATTR_TARGET_ROLE, data_set); } else if(rsc->next_role > rsc->role && is_set(data_set->flags, pe_flag_have_quorum) == FALSE && data_set->no_quorum_policy == no_quorum_freeze) { crm_notice("Resource %s cannot be elevated from %s to %s: no-quorum-policy=freeze", rsc->id, role2text(rsc->role), role2text(rsc->next_role)); rsc->next_role = rsc->role; } pe__show_node_weights(!show_scores, rsc, __FUNCTION__, rsc->allowed_nodes); if (is_set(data_set->flags, pe_flag_stonith_enabled) && is_set(data_set->flags, pe_flag_have_stonith_resource) == FALSE) { clear_bit(rsc->flags, pe_rsc_managed); } if (is_not_set(rsc->flags, pe_rsc_managed)) { const char *reason = NULL; node_t *assign_to = NULL; rsc->next_role = rsc->role; assign_to = pe__current_node(rsc); if (assign_to == NULL) { reason = "inactive"; } else if (rsc->role == RSC_ROLE_MASTER) { reason = "master"; } else if (is_set(rsc->flags, pe_rsc_failed)) { reason = "failed"; } else { reason = "active"; } pe_rsc_info(rsc, "Unmanaged resource %s allocated to %s: %s", rsc->id, (assign_to? assign_to->details->uname : "no node"), reason); native_assign_node(rsc, NULL, assign_to, TRUE); } else if (is_set(data_set->flags, pe_flag_stop_everything)) { pe_rsc_debug(rsc, "Forcing %s to stop", rsc->id); native_assign_node(rsc, NULL, NULL, TRUE); } else if (is_set(rsc->flags, pe_rsc_provisional) && native_choose_node(rsc, prefer, data_set)) { pe_rsc_trace(rsc, "Allocated resource %s to %s", rsc->id, rsc->allocated_to->details->uname); } else if (rsc->allocated_to == NULL) { if (is_not_set(rsc->flags, pe_rsc_orphan)) { pe_rsc_info(rsc, "Resource %s cannot run anywhere", rsc->id); } else if (rsc->running_on != NULL) { pe_rsc_info(rsc, "Stopping orphan resource %s", rsc->id); } } else { pe_rsc_debug(rsc, "Pre-Allocated resource %s to %s", rsc->id, rsc->allocated_to->details->uname); } clear_bit(rsc->flags, pe_rsc_allocating); if (rsc->is_remote_node) { node_t *remote_node = pe_find_node(data_set->nodes, rsc->id); CRM_ASSERT(remote_node != NULL); if (rsc->allocated_to && rsc->next_role != RSC_ROLE_STOPPED) { crm_trace("Setting Pacemaker Remote node %s to ONLINE", remote_node->details->id); remote_node->details->online = TRUE; /* We shouldn't consider an unseen remote-node unclean if we are going * to try and connect to it. Otherwise we get an unnecessary fence */ if (remote_node->details->unseen == TRUE) { remote_node->details->unclean = FALSE; } } else { crm_trace("Setting Pacemaker Remote node %s to SHUTDOWN (next role %s, %sallocated)", remote_node->details->id, role2text(rsc->next_role), (rsc->allocated_to? "" : "un")); remote_node->details->shutdown = TRUE; } } return rsc->allocated_to; } static gboolean is_op_dup(resource_t *rsc, const char *name, guint interval_ms) { gboolean dup = FALSE; const char *id = NULL; const char *value = NULL; xmlNode *operation = NULL; guint interval2_ms = 0; CRM_ASSERT(rsc); for (operation = __xml_first_child_element(rsc->ops_xml); operation != NULL; operation = __xml_next_element(operation)) { if (crm_str_eq((const char *)operation->name, "op", TRUE)) { value = crm_element_value(operation, "name"); if (safe_str_neq(value, name)) { continue; } value = crm_element_value(operation, XML_LRM_ATTR_INTERVAL); interval2_ms = crm_parse_interval_spec(value); if (interval_ms != interval2_ms) { continue; } if (id == NULL) { id = ID(operation); } else { crm_config_err("Operation %s is a duplicate of %s", ID(operation), id); crm_config_err ("Do not use the same (name, interval) combination more than once per resource"); dup = TRUE; } } } return dup; } static bool op_cannot_recur(const char *name) { return safe_str_eq(name, RSC_STOP) || safe_str_eq(name, RSC_START) || safe_str_eq(name, RSC_DEMOTE) || safe_str_eq(name, RSC_PROMOTE); } static void RecurringOp(resource_t * rsc, action_t * start, node_t * node, xmlNode * operation, pe_working_set_t * data_set) { char *key = NULL; const char *name = NULL; const char *role = NULL; const char *interval_spec = NULL; const char *node_uname = node? node->details->uname : "n/a"; guint interval_ms = 0; action_t *mon = NULL; gboolean is_optional = TRUE; GListPtr possible_matches = NULL; CRM_ASSERT(rsc); /* Only process for the operations without role="Stopped" */ role = crm_element_value(operation, "role"); if (role && text2role(role) == RSC_ROLE_STOPPED) { return; } interval_spec = crm_element_value(operation, XML_LRM_ATTR_INTERVAL); interval_ms = crm_parse_interval_spec(interval_spec); if (interval_ms == 0) { return; } name = crm_element_value(operation, "name"); if (is_op_dup(rsc, name, interval_ms)) { crm_trace("Not creating duplicate recurring action %s for %dms %s", ID(operation), interval_ms, name); return; } if (op_cannot_recur(name)) { crm_config_err("Ignoring %s because action '%s' cannot be recurring", ID(operation), name); return; } key = pcmk__op_key(rsc->id, name, interval_ms); if (find_rsc_op_entry(rsc, key) == NULL) { crm_trace("Not creating recurring action %s for disabled resource %s", ID(operation), rsc->id); free(key); return; } pe_rsc_trace(rsc, "Creating recurring action %s for %s in role %s on %s", ID(operation), rsc->id, role2text(rsc->next_role), node_uname); if (start != NULL) { pe_rsc_trace(rsc, "Marking %s %s due to %s", key, is_set(start->flags, pe_action_optional) ? "optional" : "mandatory", start->uuid); is_optional = (rsc->cmds->action_flags(start, NULL) & pe_action_optional); } else { pe_rsc_trace(rsc, "Marking %s optional", key); is_optional = TRUE; } /* start a monitor for an already active resource */ possible_matches = find_actions_exact(rsc->actions, key, node); if (possible_matches == NULL) { is_optional = FALSE; pe_rsc_trace(rsc, "Marking %s mandatory: not active", key); } else { GListPtr gIter = NULL; for (gIter = possible_matches; gIter != NULL; gIter = gIter->next) { action_t *op = (action_t *) gIter->data; if (is_set(op->flags, pe_action_reschedule)) { is_optional = FALSE; break; } } g_list_free(possible_matches); } if ((rsc->next_role == RSC_ROLE_MASTER && role == NULL) || (role != NULL && text2role(role) != rsc->next_role)) { int log_level = LOG_TRACE; const char *result = "Ignoring"; if (is_optional) { char *after_key = NULL; action_t *cancel_op = NULL; // It's running, so cancel it log_level = LOG_INFO; result = "Cancelling"; cancel_op = pe_cancel_op(rsc, name, interval_ms, node, data_set); switch (rsc->role) { case RSC_ROLE_SLAVE: case RSC_ROLE_STARTED: if (rsc->next_role == RSC_ROLE_MASTER) { after_key = promote_key(rsc); } else if (rsc->next_role == RSC_ROLE_STOPPED) { after_key = stop_key(rsc); } break; case RSC_ROLE_MASTER: after_key = demote_key(rsc); break; default: break; } if (after_key) { custom_action_order(rsc, NULL, cancel_op, rsc, after_key, NULL, pe_order_runnable_left, data_set); } } do_crm_log(log_level, "%s action %s (%s vs. %s)", result, key, role ? role : role2text(RSC_ROLE_SLAVE), role2text(rsc->next_role)); free(key); return; } mon = custom_action(rsc, key, name, node, is_optional, TRUE, data_set); key = mon->uuid; if (is_optional) { pe_rsc_trace(rsc, "%s\t %s (optional)", node_uname, mon->uuid); } if (start == NULL || is_set(start->flags, pe_action_runnable) == FALSE) { pe_rsc_debug(rsc, "%s\t %s (cancelled : start un-runnable)", node_uname, mon->uuid); update_action_flags(mon, pe_action_runnable | pe_action_clear, __FUNCTION__, __LINE__); } else if (node == NULL || node->details->online == FALSE || node->details->unclean) { pe_rsc_debug(rsc, "%s\t %s (cancelled : no node available)", node_uname, mon->uuid); update_action_flags(mon, pe_action_runnable | pe_action_clear, __FUNCTION__, __LINE__); } else if (is_set(mon->flags, pe_action_optional) == FALSE) { pe_rsc_info(rsc, " Start recurring %s (%us) for %s on %s", mon->task, interval_ms / 1000, rsc->id, node_uname); } if (rsc->next_role == RSC_ROLE_MASTER) { char *running_master = crm_itoa(PCMK_OCF_RUNNING_MASTER); add_hash_param(mon->meta, XML_ATTR_TE_TARGET_RC, running_master); free(running_master); } if (node == NULL || is_set(rsc->flags, pe_rsc_managed)) { custom_action_order(rsc, start_key(rsc), NULL, NULL, strdup(key), mon, pe_order_implies_then | pe_order_runnable_left, data_set); custom_action_order(rsc, reload_key(rsc), NULL, NULL, strdup(key), mon, pe_order_implies_then | pe_order_runnable_left, data_set); if (rsc->next_role == RSC_ROLE_MASTER) { custom_action_order(rsc, promote_key(rsc), NULL, rsc, NULL, mon, pe_order_optional | pe_order_runnable_left, data_set); } else if (rsc->role == RSC_ROLE_MASTER) { custom_action_order(rsc, demote_key(rsc), NULL, rsc, NULL, mon, pe_order_optional | pe_order_runnable_left, data_set); } } } static void Recurring(resource_t * rsc, action_t * start, node_t * node, pe_working_set_t * data_set) { if (is_not_set(rsc->flags, pe_rsc_maintenance) && (node == NULL || node->details->maintenance == FALSE)) { xmlNode *operation = NULL; for (operation = __xml_first_child_element(rsc->ops_xml); operation != NULL; operation = __xml_next_element(operation)) { if (crm_str_eq((const char *)operation->name, "op", TRUE)) { RecurringOp(rsc, start, node, operation, data_set); } } } } static void RecurringOp_Stopped(resource_t * rsc, action_t * start, node_t * node, xmlNode * operation, pe_working_set_t * data_set) { char *key = NULL; const char *name = NULL; const char *role = NULL; const char *interval_spec = NULL; const char *node_uname = node? node->details->uname : "n/a"; guint interval_ms = 0; GListPtr possible_matches = NULL; GListPtr gIter = NULL; /* Only process for the operations with role="Stopped" */ role = crm_element_value(operation, "role"); if (role == NULL || text2role(role) != RSC_ROLE_STOPPED) { return; } interval_spec = crm_element_value(operation, XML_LRM_ATTR_INTERVAL); interval_ms = crm_parse_interval_spec(interval_spec); if (interval_ms == 0) { return; } name = crm_element_value(operation, "name"); if (is_op_dup(rsc, name, interval_ms)) { crm_trace("Not creating duplicate recurring action %s for %dms %s", ID(operation), interval_ms, name); return; } if (op_cannot_recur(name)) { crm_config_err("Invalid recurring action %s wth name: '%s'", ID(operation), name); return; } key = pcmk__op_key(rsc->id, name, interval_ms); if (find_rsc_op_entry(rsc, key) == NULL) { crm_trace("Not creating recurring action %s for disabled resource %s", ID(operation), rsc->id); free(key); return; } // @TODO add support if (is_set(rsc->flags, pe_rsc_unique) == FALSE) { crm_notice("Ignoring %s (recurring monitors for Stopped role are " "not supported for anonymous clones)", ID(operation)); return; } pe_rsc_trace(rsc, "Creating recurring action %s for %s in role %s on nodes where it should not be running", ID(operation), rsc->id, role2text(rsc->next_role)); /* if the monitor exists on the node where the resource will be running, cancel it */ if (node != NULL) { possible_matches = find_actions_exact(rsc->actions, key, node); if (possible_matches) { action_t *cancel_op = NULL; g_list_free(possible_matches); cancel_op = pe_cancel_op(rsc, name, interval_ms, node, data_set); if (rsc->next_role == RSC_ROLE_STARTED || rsc->next_role == RSC_ROLE_SLAVE) { /* rsc->role == RSC_ROLE_STOPPED: cancel the monitor before start */ /* rsc->role == RSC_ROLE_STARTED: for a migration, cancel the monitor on the target node before start */ custom_action_order(rsc, NULL, cancel_op, rsc, start_key(rsc), NULL, pe_order_runnable_left, data_set); } pe_rsc_info(rsc, "Cancel action %s (%s vs. %s) on %s", key, role, role2text(rsc->next_role), node_uname); } } for (gIter = data_set->nodes; gIter != NULL; gIter = gIter->next) { node_t *stop_node = (node_t *) gIter->data; const char *stop_node_uname = stop_node->details->uname; gboolean is_optional = TRUE; gboolean probe_is_optional = TRUE; gboolean stop_is_optional = TRUE; action_t *stopped_mon = NULL; char *rc_inactive = NULL; GListPtr probe_complete_ops = NULL; GListPtr stop_ops = NULL; GListPtr local_gIter = NULL; if (node && safe_str_eq(stop_node_uname, node_uname)) { continue; } pe_rsc_trace(rsc, "Creating recurring action %s for %s on %s", ID(operation), rsc->id, crm_str(stop_node_uname)); /* start a monitor for an already stopped resource */ possible_matches = find_actions_exact(rsc->actions, key, stop_node); if (possible_matches == NULL) { pe_rsc_trace(rsc, "Marking %s mandatory on %s: not active", key, crm_str(stop_node_uname)); is_optional = FALSE; } else { pe_rsc_trace(rsc, "Marking %s optional on %s: already active", key, crm_str(stop_node_uname)); is_optional = TRUE; g_list_free(possible_matches); } stopped_mon = custom_action(rsc, strdup(key), name, stop_node, is_optional, TRUE, data_set); rc_inactive = crm_itoa(PCMK_OCF_NOT_RUNNING); add_hash_param(stopped_mon->meta, XML_ATTR_TE_TARGET_RC, rc_inactive); free(rc_inactive); if (is_set(rsc->flags, pe_rsc_managed)) { GList *probes = pe__resource_actions(rsc, stop_node, RSC_STATUS, FALSE); GListPtr pIter = NULL; for (pIter = probes; pIter != NULL; pIter = pIter->next) { action_t *probe = (action_t *) pIter->data; order_actions(probe, stopped_mon, pe_order_runnable_left); crm_trace("%s then %s on %s", probe->uuid, stopped_mon->uuid, stop_node->details->uname); } g_list_free(probes); } if (probe_complete_ops) { g_list_free(probe_complete_ops); } stop_ops = pe__resource_actions(rsc, stop_node, RSC_STOP, TRUE); for (local_gIter = stop_ops; local_gIter != NULL; local_gIter = local_gIter->next) { action_t *stop = (action_t *) local_gIter->data; if (is_set(stop->flags, pe_action_optional) == FALSE) { stop_is_optional = FALSE; } if (is_set(stop->flags, pe_action_runnable) == FALSE) { crm_debug("%s\t %s (cancelled : stop un-runnable)", crm_str(stop_node_uname), stopped_mon->uuid); update_action_flags(stopped_mon, pe_action_runnable | pe_action_clear, __FUNCTION__, __LINE__); } if (is_set(rsc->flags, pe_rsc_managed)) { custom_action_order(rsc, stop_key(rsc), stop, NULL, strdup(key), stopped_mon, pe_order_implies_then | pe_order_runnable_left, data_set); } } if (stop_ops) { g_list_free(stop_ops); } if (is_optional == FALSE && probe_is_optional && stop_is_optional && is_set(rsc->flags, pe_rsc_managed) == FALSE) { pe_rsc_trace(rsc, "Marking %s optional on %s due to unmanaged", key, crm_str(stop_node_uname)); update_action_flags(stopped_mon, pe_action_optional, __FUNCTION__, __LINE__); } if (is_set(stopped_mon->flags, pe_action_optional)) { pe_rsc_trace(rsc, "%s\t %s (optional)", crm_str(stop_node_uname), stopped_mon->uuid); } if (stop_node->details->online == FALSE || stop_node->details->unclean) { pe_rsc_debug(rsc, "%s\t %s (cancelled : no node available)", crm_str(stop_node_uname), stopped_mon->uuid); update_action_flags(stopped_mon, pe_action_runnable | pe_action_clear, __FUNCTION__, __LINE__); } if (is_set(stopped_mon->flags, pe_action_runnable) && is_set(stopped_mon->flags, pe_action_optional) == FALSE) { crm_notice(" Start recurring %s (%us) for %s on %s", stopped_mon->task, interval_ms / 1000, rsc->id, crm_str(stop_node_uname)); } } free(key); } static void Recurring_Stopped(resource_t * rsc, action_t * start, node_t * node, pe_working_set_t * data_set) { if (is_not_set(rsc->flags, pe_rsc_maintenance) && (node == NULL || node->details->maintenance == FALSE)) { xmlNode *operation = NULL; for (operation = __xml_first_child_element(rsc->ops_xml); operation != NULL; operation = __xml_next_element(operation)) { if (crm_str_eq((const char *)operation->name, "op", TRUE)) { RecurringOp_Stopped(rsc, start, node, operation, data_set); } } } } static void handle_migration_actions(resource_t * rsc, node_t *current, node_t *chosen, pe_working_set_t * data_set) { action_t *migrate_to = NULL; action_t *migrate_from = NULL; action_t *start = NULL; action_t *stop = NULL; gboolean partial = rsc->partial_migration_target ? TRUE : FALSE; pe_rsc_trace(rsc, "Processing migration actions %s moving from %s to %s . partial migration = %s", rsc->id, current->details->id, chosen->details->id, partial ? "TRUE" : "FALSE"); start = start_action(rsc, chosen, TRUE); stop = stop_action(rsc, current, TRUE); if (partial == FALSE) { migrate_to = custom_action(rsc, pcmk__op_key(rsc->id, RSC_MIGRATE, 0), RSC_MIGRATE, current, TRUE, TRUE, data_set); } migrate_from = custom_action(rsc, pcmk__op_key(rsc->id, RSC_MIGRATED, 0), RSC_MIGRATED, chosen, TRUE, TRUE, data_set); if ((migrate_to && migrate_from) || (migrate_from && partial)) { set_bit(start->flags, pe_action_migrate_runnable); set_bit(stop->flags, pe_action_migrate_runnable); update_action_flags(start, pe_action_pseudo, __FUNCTION__, __LINE__); /* easier than trying to delete it from the graph */ /* order probes before migrations */ if (partial) { set_bit(migrate_from->flags, pe_action_migrate_runnable); migrate_from->needs = start->needs; custom_action_order(rsc, pcmk__op_key(rsc->id, RSC_STATUS, 0), NULL, rsc, pcmk__op_key(rsc->id, RSC_MIGRATED, 0), NULL, pe_order_optional, data_set); } else { set_bit(migrate_from->flags, pe_action_migrate_runnable); set_bit(migrate_to->flags, pe_action_migrate_runnable); migrate_to->needs = start->needs; custom_action_order(rsc, pcmk__op_key(rsc->id, RSC_STATUS, 0), NULL, rsc, pcmk__op_key(rsc->id, RSC_MIGRATE, 0), NULL, pe_order_optional, data_set); custom_action_order(rsc, pcmk__op_key(rsc->id, RSC_MIGRATE, 0), NULL, rsc, pcmk__op_key(rsc->id, RSC_MIGRATED, 0), NULL, pe_order_optional|pe_order_implies_first_migratable, data_set); } custom_action_order(rsc, pcmk__op_key(rsc->id, RSC_MIGRATED, 0), NULL, rsc, pcmk__op_key(rsc->id, RSC_STOP, 0), NULL, pe_order_optional|pe_order_implies_first_migratable, data_set); custom_action_order(rsc, pcmk__op_key(rsc->id, RSC_MIGRATED, 0), NULL, rsc, pcmk__op_key(rsc->id, RSC_START, 0), NULL, pe_order_optional|pe_order_implies_first_migratable|pe_order_pseudo_left, data_set); } if (migrate_to) { add_hash_param(migrate_to->meta, XML_LRM_ATTR_MIGRATE_SOURCE, current->details->uname); add_hash_param(migrate_to->meta, XML_LRM_ATTR_MIGRATE_TARGET, chosen->details->uname); /* Pacemaker Remote connections don't require pending to be recorded in * the CIB. We can reduce CIB writes by not setting PENDING for them. */ if (rsc->is_remote_node == FALSE) { /* migrate_to takes place on the source node, but can * have an effect on the target node depending on how * the agent is written. Because of this, we have to maintain * a record that the migrate_to occurred, in case the source node * loses membership while the migrate_to action is still in-flight. */ add_hash_param(migrate_to->meta, XML_OP_ATTR_PENDING, "true"); } } if (migrate_from) { add_hash_param(migrate_from->meta, XML_LRM_ATTR_MIGRATE_SOURCE, current->details->uname); add_hash_param(migrate_from->meta, XML_LRM_ATTR_MIGRATE_TARGET, chosen->details->uname); } } void native_create_actions(resource_t * rsc, pe_working_set_t * data_set) { action_t *start = NULL; node_t *chosen = NULL; node_t *current = NULL; gboolean need_stop = FALSE; gboolean is_moving = FALSE; gboolean allow_migrate = is_set(rsc->flags, pe_rsc_allow_migrate) ? TRUE : FALSE; GListPtr gIter = NULL; unsigned int num_all_active = 0; unsigned int num_clean_active = 0; bool multiply_active = FALSE; enum rsc_role_e role = RSC_ROLE_UNKNOWN; enum rsc_role_e next_role = RSC_ROLE_UNKNOWN; CRM_ASSERT(rsc); chosen = rsc->allocated_to; if (chosen != NULL && rsc->next_role == RSC_ROLE_UNKNOWN) { rsc->next_role = RSC_ROLE_STARTED; pe_rsc_trace(rsc, "Fixed next_role: unknown -> %s", role2text(rsc->next_role)); } else if (rsc->next_role == RSC_ROLE_UNKNOWN) { rsc->next_role = RSC_ROLE_STOPPED; pe_rsc_trace(rsc, "Fixed next_role: unknown -> %s", role2text(rsc->next_role)); } pe_rsc_trace(rsc, "Processing state transition for %s %p: %s->%s", rsc->id, rsc, role2text(rsc->role), role2text(rsc->next_role)); current = pe__find_active_on(rsc, &num_all_active, &num_clean_active); for (gIter = rsc->dangling_migrations; gIter != NULL; gIter = gIter->next) { node_t *dangling_source = (node_t *) gIter->data; action_t *stop = stop_action(rsc, dangling_source, FALSE); set_bit(stop->flags, pe_action_dangle); pe_rsc_trace(rsc, "Forcing a cleanup of %s on %s", rsc->id, dangling_source->details->uname); if (is_set(data_set->flags, pe_flag_remove_after_stop)) { DeleteRsc(rsc, dangling_source, FALSE, data_set); } } if ((num_all_active == 2) && (num_clean_active == 2) && chosen && rsc->partial_migration_source && rsc->partial_migration_target && (current->details == rsc->partial_migration_source->details) && (chosen->details == rsc->partial_migration_target->details)) { /* The chosen node is still the migration target from a partial * migration. Attempt to continue the migration instead of recovering * by stopping the resource everywhere and starting it on a single node. */ pe_rsc_trace(rsc, "Will attempt to continue with a partial migration to target %s from %s", rsc->partial_migration_target->details->id, rsc->partial_migration_source->details->id); } else if (is_not_set(rsc->flags, pe_rsc_needs_fencing)) { /* 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); } else { multiply_active = (num_all_active > 1); } if (multiply_active) { if (rsc->partial_migration_target && rsc->partial_migration_source) { // Migration was in progress, but we've chosen a different target crm_notice("Resource %s can no longer migrate to %s. Stopping on %s too", rsc->id, rsc->partial_migration_target->details->uname, rsc->partial_migration_source->details->uname); } else { // Resource was incorrectly multiply active pe_proc_err("Resource %s is active on %u nodes (%s)", 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"); } if (rsc->recovery_type == recovery_stop_start) { need_stop = TRUE; } /* If by chance a partial migration is in process, but the migration * target is not chosen still, clear all partial migration data. */ rsc->partial_migration_source = rsc->partial_migration_target = NULL; allow_migrate = FALSE; } if (is_set(rsc->flags, pe_rsc_start_pending)) { start = start_action(rsc, chosen, TRUE); set_bit(start->flags, pe_action_print_always); } if (current && chosen && current->details != chosen->details) { pe_rsc_trace(rsc, "Moving %s", rsc->id); is_moving = TRUE; need_stop = TRUE; } else if (is_set(rsc->flags, pe_rsc_failed)) { pe_rsc_trace(rsc, "Recovering %s", rsc->id); need_stop = TRUE; } else if (is_set(rsc->flags, pe_rsc_block)) { pe_rsc_trace(rsc, "Block %s", rsc->id); need_stop = TRUE; } else if (rsc->role > RSC_ROLE_STARTED && current != NULL && chosen != NULL) { /* Recovery of a promoted resource */ start = start_action(rsc, chosen, TRUE); if (is_set(start->flags, pe_action_optional) == FALSE) { pe_rsc_trace(rsc, "Forced start %s", rsc->id); need_stop = TRUE; } } pe_rsc_trace(rsc, "Creating actions for %s: %s->%s", rsc->id, role2text(rsc->role), role2text(rsc->next_role)); /* Create any additional actions required when bringing resource down and * back up to same level. */ role = rsc->role; while (role != RSC_ROLE_STOPPED) { next_role = rsc_state_matrix[role][RSC_ROLE_STOPPED]; pe_rsc_trace(rsc, "Down: Executing: %s->%s (%s)%s", role2text(role), role2text(next_role), rsc->id, need_stop ? " required" : ""); if (rsc_action_matrix[role][next_role] (rsc, current, !need_stop, data_set) == FALSE) { break; } role = next_role; } while (rsc->role <= rsc->next_role && role != rsc->role && is_not_set(rsc->flags, pe_rsc_block)) { next_role = rsc_state_matrix[role][rsc->role]; pe_rsc_trace(rsc, "Up: Executing: %s->%s (%s)%s", role2text(role), role2text(next_role), rsc->id, need_stop ? " required" : ""); if (rsc_action_matrix[role][next_role] (rsc, chosen, !need_stop, data_set) == FALSE) { break; } role = next_role; } role = rsc->role; /* Required steps from this role to the next */ while (role != rsc->next_role) { next_role = rsc_state_matrix[role][rsc->next_role]; pe_rsc_trace(rsc, "Role: Executing: %s->%s = (%s on %s)", role2text(role), role2text(next_role), rsc->id, chosen?chosen->details->uname:"NA"); if (rsc_action_matrix[role][next_role] (rsc, chosen, FALSE, data_set) == FALSE) { break; } role = next_role; } if(is_set(rsc->flags, pe_rsc_block)) { pe_rsc_trace(rsc, "No monitor additional ops for blocked resource"); } else if (rsc->next_role != RSC_ROLE_STOPPED || is_set(rsc->flags, pe_rsc_managed) == FALSE) { pe_rsc_trace(rsc, "Monitor ops for active resource"); start = start_action(rsc, chosen, TRUE); Recurring(rsc, start, chosen, data_set); Recurring_Stopped(rsc, start, chosen, data_set); } else { pe_rsc_trace(rsc, "Monitor ops for inactive resource"); Recurring_Stopped(rsc, NULL, NULL, data_set); } /* if we are stuck in a partial migration, where the target * of the partial migration no longer matches the chosen target. * A full stop/start is required */ if (rsc->partial_migration_target && (chosen == NULL || rsc->partial_migration_target->details != chosen->details)) { pe_rsc_trace(rsc, "Not allowing partial migration to continue. %s", rsc->id); allow_migrate = FALSE; } else if (is_moving == FALSE || is_not_set(rsc->flags, pe_rsc_managed) || is_set(rsc->flags, pe_rsc_failed) || is_set(rsc->flags, pe_rsc_start_pending) || (current && current->details->unclean) || rsc->next_role < RSC_ROLE_STARTED) { allow_migrate = FALSE; } if (allow_migrate) { handle_migration_actions(rsc, current, chosen, data_set); } } static void rsc_avoids_remote_nodes(resource_t *rsc) { GHashTableIter iter; 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) { 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 * \param[in] data_set Cluster working set * * \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(pe_resource_t *rsc, pe_working_set_t *data_set) { GList *allowed_nodes = NULL; if (rsc->allowed_nodes) { allowed_nodes = g_hash_table_get_values(rsc->allowed_nodes); } if (is_set(data_set->flags, pe_flag_stdout)) { allowed_nodes = g_list_sort(allowed_nodes, sort_node_uname); } return allowed_nodes; } void native_internal_constraints(resource_t * rsc, pe_working_set_t * data_set) { /* This function is on the critical path and worth optimizing as much as possible */ pe_resource_t *top = NULL; GList *allowed_nodes = NULL; bool check_unfencing = FALSE; bool check_utilization = FALSE; if (is_not_set(rsc->flags, pe_rsc_managed)) { pe_rsc_trace(rsc, "Skipping native constraints for unmanaged resource: %s", rsc->id); return; } top = uber_parent(rsc); // Whether resource requires unfencing check_unfencing = is_not_set(rsc->flags, pe_rsc_fence_device) && is_set(data_set->flags, pe_flag_enable_unfencing) && 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) && safe_str_neq(data_set->placement_strategy, "default"); // Order stops before starts (i.e. restart) custom_action_order(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, data_set); // Promotable ordering: demote before stop, start before promote if (is_set(top->flags, pe_rsc_promotable) || (rsc->role > RSC_ROLE_SLAVE)) { custom_action_order(rsc, pcmk__op_key(rsc->id, RSC_DEMOTE, 0), NULL, rsc, pcmk__op_key(rsc->id, RSC_STOP, 0), NULL, pe_order_implies_first_master, data_set); custom_action_order(rsc, pcmk__op_key(rsc->id, RSC_START, 0), NULL, rsc, pcmk__op_key(rsc->id, RSC_PROMOTE, 0), NULL, pe_order_runnable_left, data_set); } // Don't clear resource history if probing on same node custom_action_order(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, data_set); // Certain checks need allowed nodes if (check_unfencing || check_utilization || rsc->container) { allowed_nodes = allowed_nodes_as_list(rsc, data_set); } if (check_unfencing) { /* Check if the node needs to be unfenced first */ for (GList *item = allowed_nodes; item; item = item->next) { pe_node_t *node = item->data; pe_action_t *unfence = pe_fence_op(node, "on", TRUE, NULL, data_set); crm_debug("Ordering any stops of %s before %s, and any starts after", rsc->id, unfence->uuid); /* * It would be more efficient to order clone resources once, * rather than order each instance, but ordering the instance * allows us to avoid unnecessary dependencies that might conflict * with user constraints. * * @TODO: This constraint can still produce a transition loop if the * resource has a stop scheduled on the node being unfenced, and * there is a user ordering constraint to start some other resource * (which will be ordered after the unfence) before stopping this * resource. An example is "start some slow-starting cloned service * before stopping an associated virtual IP that may be moving to * it": * stop this -> unfencing -> start that -> stop this */ custom_action_order(rsc, stop_key(rsc), NULL, NULL, strdup(unfence->uuid), unfence, pe_order_optional|pe_order_same_node, data_set); custom_action_order(NULL, strdup(unfence->uuid), unfence, rsc, start_key(rsc), NULL, pe_order_implies_then_on_node|pe_order_same_node, data_set); } } if (check_utilization) { GListPtr gIter = NULL; pe_rsc_trace(rsc, "Creating utilization constraints for %s - strategy: %s", rsc->id, data_set->placement_strategy); for (gIter = rsc->running_on; gIter != NULL; gIter = gIter->next) { node_t *current = (node_t *) gIter->data; char *load_stopped_task = crm_concat(LOAD_STOPPED, current->details->uname, '_'); action_t *load_stopped = get_pseudo_op(load_stopped_task, data_set); if (load_stopped->node == NULL) { load_stopped->node = node_copy(current); update_action_flags(load_stopped, pe_action_optional | pe_action_clear, __FUNCTION__, __LINE__); } custom_action_order(rsc, stop_key(rsc), NULL, NULL, load_stopped_task, load_stopped, pe_order_load, data_set); } for (GList *item = allowed_nodes; item; item = item->next) { pe_node_t *next = item->data; char *load_stopped_task = crm_concat(LOAD_STOPPED, next->details->uname, '_'); action_t *load_stopped = get_pseudo_op(load_stopped_task, data_set); if (load_stopped->node == NULL) { load_stopped->node = node_copy(next); update_action_flags(load_stopped, pe_action_optional | pe_action_clear, __FUNCTION__, __LINE__); } custom_action_order(NULL, strdup(load_stopped_task), load_stopped, rsc, start_key(rsc), NULL, pe_order_load, data_set); custom_action_order(NULL, strdup(load_stopped_task), load_stopped, rsc, pcmk__op_key(rsc->id, RSC_MIGRATE, 0), NULL, pe_order_load, data_set); free(load_stopped_task); } } if (rsc->container) { resource_t *remote_rsc = NULL; if (rsc->is_remote_node) { // rsc is the implicit remote connection for a guest or bundle node /* Do not allow a guest resource to live on a Pacemaker Remote node, * to avoid nesting remotes. However, allow bundles to run on remote * nodes. */ if (is_not_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. */ new_rsc_order(rsc->container, RSC_STATUS, rsc, RSC_STOP, pe_order_optional, data_set); /* 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(data_set, rsc->container); } if (remote_rsc) { /* 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); custom_action_order(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, data_set); custom_action_order(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, data_set); if (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 */ } rsc_colocation_new("resource-with-container", NULL, score, rsc, rsc->container, NULL, NULL, data_set); } } if (rsc->is_remote_node || is_set(rsc->flags, pe_rsc_fence_device)) { /* don't allow remote nodes to run stonith devices * or remote connection resources.*/ rsc_avoids_remote_nodes(rsc); } g_list_free(allowed_nodes); } void native_rsc_colocation_lh(pe_resource_t *rsc_lh, pe_resource_t *rsc_rh, rsc_colocation_t *constraint, pe_working_set_t *data_set) { if (rsc_lh == NULL) { pe_err("rsc_lh was NULL for %s", constraint->id); return; } else if (constraint->rsc_rh == NULL) { pe_err("rsc_rh was NULL for %s", constraint->id); return; } if (constraint->score == 0) { return; } pe_rsc_trace(rsc_lh, "Processing colocation constraint between %s and %s", rsc_lh->id, rsc_rh->id); rsc_rh->cmds->rsc_colocation_rh(rsc_lh, rsc_rh, constraint, data_set); } enum filter_colocation_res filter_colocation_constraint(resource_t * rsc_lh, resource_t * rsc_rh, rsc_colocation_t * constraint, gboolean preview) { if (constraint->score == 0) { return influence_nothing; } /* rh side must be allocated before we can process constraint */ if (preview == FALSE && is_set(rsc_rh->flags, pe_rsc_provisional)) { return influence_nothing; } if ((constraint->role_lh >= RSC_ROLE_SLAVE) && rsc_lh->parent && is_set(rsc_lh->parent->flags, pe_rsc_promotable) && is_not_set(rsc_lh->flags, pe_rsc_provisional)) { /* LH and RH resources have already been allocated, place the correct * priority on LH rsc for the given promotable clone resource role */ return influence_rsc_priority; } if (preview == FALSE && is_not_set(rsc_lh->flags, pe_rsc_provisional)) { // Log an error if we violated a mandatory colocation constraint const pe_node_t *rh_node = rsc_rh->allocated_to; if (rsc_lh->allocated_to == NULL) { // Dependent resource isn't allocated, so constraint doesn't matter return influence_nothing; } if (constraint->score >= INFINITY) { // Dependent resource must colocate with rh_node if ((rh_node == NULL) || (rh_node->details != rsc_lh->allocated_to->details)) { crm_err("%s must be colocated with %s but is not (%s vs. %s)", rsc_lh->id, rsc_rh->id, rsc_lh->allocated_to->details->uname, (rh_node? rh_node->details->uname : "unallocated")); } } else if (constraint->score <= -INFINITY) { // Dependent resource must anti-colocate with rh_node if ((rh_node != NULL) && (rsc_lh->allocated_to->details == rh_node->details)) { crm_err("%s and %s must be anti-colocated but are allocated " "to the same node (%s)", rsc_lh->id, rsc_rh->id, rh_node->details->uname); } } return influence_nothing; } if (constraint->score > 0 && constraint->role_lh != RSC_ROLE_UNKNOWN && constraint->role_lh != rsc_lh->next_role) { crm_trace("LH: Skipping constraint: \"%s\" state filter nextrole is %s", role2text(constraint->role_lh), role2text(rsc_lh->next_role)); return influence_nothing; } if (constraint->score > 0 && constraint->role_rh != RSC_ROLE_UNKNOWN && constraint->role_rh != rsc_rh->next_role) { crm_trace("RH: Skipping constraint: \"%s\" state filter", role2text(constraint->role_rh)); return FALSE; } if (constraint->score < 0 && constraint->role_lh != RSC_ROLE_UNKNOWN && constraint->role_lh == rsc_lh->next_role) { crm_trace("LH: Skipping negative constraint: \"%s\" state filter", role2text(constraint->role_lh)); return influence_nothing; } if (constraint->score < 0 && constraint->role_rh != RSC_ROLE_UNKNOWN && constraint->role_rh == rsc_rh->next_role) { crm_trace("RH: Skipping negative constraint: \"%s\" state filter", role2text(constraint->role_rh)); return influence_nothing; } return influence_rsc_location; } static void influence_priority(resource_t * rsc_lh, resource_t * rsc_rh, rsc_colocation_t * constraint) { const char *rh_value = NULL; const char *lh_value = NULL; const char *attribute = CRM_ATTR_ID; int score_multiplier = 1; if (constraint->score == 0) { return; } if (!rsc_rh->allocated_to || !rsc_lh->allocated_to) { return; } if (constraint->node_attribute != NULL) { attribute = constraint->node_attribute; } lh_value = pe_node_attribute_raw(rsc_lh->allocated_to, attribute); rh_value = pe_node_attribute_raw(rsc_rh->allocated_to, attribute); if (!safe_str_eq(lh_value, rh_value)) { if(constraint->score == INFINITY && constraint->role_lh == RSC_ROLE_MASTER) { rsc_lh->priority = -INFINITY; } return; } if (constraint->role_rh && (constraint->role_rh != rsc_rh->next_role)) { return; } if (constraint->role_lh == RSC_ROLE_SLAVE) { score_multiplier = -1; } rsc_lh->priority = merge_weights(score_multiplier * constraint->score, rsc_lh->priority); } static void colocation_match(resource_t * rsc_lh, resource_t * rsc_rh, rsc_colocation_t * constraint) { const char *tmp = NULL; const char *value = NULL; const char *attribute = CRM_ATTR_ID; GHashTable *work = NULL; gboolean do_check = FALSE; GHashTableIter iter; node_t *node = NULL; if (constraint->score == 0) { return; } if (constraint->node_attribute != NULL) { attribute = constraint->node_attribute; } if (rsc_rh->allocated_to) { value = pe_node_attribute_raw(rsc_rh->allocated_to, attribute); do_check = TRUE; } else if (constraint->score < 0) { /* nothing to do: * anti-colocation with something that is not running */ return; } work = node_hash_dup(rsc_lh->allowed_nodes); g_hash_table_iter_init(&iter, work); while (g_hash_table_iter_next(&iter, NULL, (void **)&node)) { tmp = pe_node_attribute_raw(node, attribute); if (do_check && safe_str_eq(tmp, value)) { if (constraint->score < INFINITY) { pe_rsc_trace(rsc_lh, "%s: %s.%s += %d", constraint->id, rsc_lh->id, node->details->uname, constraint->score); node->weight = merge_weights(constraint->score, node->weight); } } else if (do_check == FALSE || constraint->score >= INFINITY) { pe_rsc_trace(rsc_lh, "%s: %s.%s -= %d (%s)", constraint->id, rsc_lh->id, node->details->uname, constraint->score, do_check ? "failed" : "unallocated"); node->weight = merge_weights(-constraint->score, node->weight); } } if (can_run_any(work) || constraint->score <= -INFINITY || constraint->score >= INFINITY) { g_hash_table_destroy(rsc_lh->allowed_nodes); rsc_lh->allowed_nodes = work; work = NULL; } else { static char score[33]; score2char_stack(constraint->score, score, sizeof(score)); pe_rsc_info(rsc_lh, "%s: Rolling back scores from %s (%d, %s)", rsc_lh->id, rsc_rh->id, do_check, score); } if (work) { g_hash_table_destroy(work); } } void native_rsc_colocation_rh(pe_resource_t *rsc_lh, pe_resource_t *rsc_rh, rsc_colocation_t *constraint, pe_working_set_t *data_set) { enum filter_colocation_res filter_results; CRM_ASSERT(rsc_lh); CRM_ASSERT(rsc_rh); filter_results = filter_colocation_constraint(rsc_lh, rsc_rh, constraint, FALSE); pe_rsc_trace(rsc_lh, "%sColocating %s with %s (%s, weight=%d, filter=%d)", constraint->score >= 0 ? "" : "Anti-", rsc_lh->id, rsc_rh->id, constraint->id, constraint->score, filter_results); switch (filter_results) { case influence_rsc_priority: influence_priority(rsc_lh, rsc_rh, constraint); break; case influence_rsc_location: pe_rsc_trace(rsc_lh, "%sColocating %s with %s (%s, weight=%d)", constraint->score >= 0 ? "" : "Anti-", rsc_lh->id, rsc_rh->id, constraint->id, constraint->score); colocation_match(rsc_lh, rsc_rh, constraint); break; case influence_nothing: default: return; } } static gboolean filter_rsc_ticket(resource_t * rsc_lh, rsc_ticket_t * rsc_ticket) { if (rsc_ticket->role_lh != RSC_ROLE_UNKNOWN && rsc_ticket->role_lh != rsc_lh->role) { pe_rsc_trace(rsc_lh, "LH: Skipping constraint: \"%s\" state filter", role2text(rsc_ticket->role_lh)); return FALSE; } return TRUE; } void rsc_ticket_constraint(resource_t * rsc_lh, rsc_ticket_t * rsc_ticket, pe_working_set_t * data_set) { if (rsc_ticket == NULL) { pe_err("rsc_ticket was NULL"); return; } if (rsc_lh == NULL) { pe_err("rsc_lh was NULL for %s", rsc_ticket->id); return; } if (rsc_ticket->ticket->granted && rsc_ticket->ticket->standby == FALSE) { return; } if (rsc_lh->children) { GListPtr gIter = rsc_lh->children; pe_rsc_trace(rsc_lh, "Processing ticket dependencies from %s", rsc_lh->id); for (; gIter != NULL; gIter = gIter->next) { resource_t *child_rsc = (resource_t *) gIter->data; rsc_ticket_constraint(child_rsc, rsc_ticket, data_set); } return; } pe_rsc_trace(rsc_lh, "%s: Processing ticket dependency on %s (%s, %s)", rsc_lh->id, rsc_ticket->ticket->id, rsc_ticket->id, role2text(rsc_ticket->role_lh)); if ((rsc_ticket->ticket->granted == FALSE) && (rsc_lh->running_on != NULL)) { GListPtr gIter = NULL; switch (rsc_ticket->loss_policy) { case loss_ticket_stop: resource_location(rsc_lh, NULL, -INFINITY, "__loss_of_ticket__", data_set); break; case loss_ticket_demote: // Promotion score will be set to -INFINITY in promotion_order() if (rsc_ticket->role_lh != RSC_ROLE_MASTER) { resource_location(rsc_lh, NULL, -INFINITY, "__loss_of_ticket__", data_set); } break; case loss_ticket_fence: if (filter_rsc_ticket(rsc_lh, rsc_ticket) == FALSE) { return; } resource_location(rsc_lh, NULL, -INFINITY, "__loss_of_ticket__", data_set); for (gIter = rsc_lh->running_on; gIter != NULL; gIter = gIter->next) { node_t *node = (node_t *) gIter->data; pe_fence_node(data_set, node, "deadman ticket was lost"); } break; case loss_ticket_freeze: if (filter_rsc_ticket(rsc_lh, rsc_ticket) == FALSE) { return; } if (rsc_lh->running_on != NULL) { clear_bit(rsc_lh->flags, pe_rsc_managed); set_bit(rsc_lh->flags, pe_rsc_block); } break; } } else if (rsc_ticket->ticket->granted == FALSE) { if (rsc_ticket->role_lh != RSC_ROLE_MASTER || rsc_ticket->loss_policy == loss_ticket_stop) { resource_location(rsc_lh, NULL, -INFINITY, "__no_ticket__", data_set); } } else if (rsc_ticket->ticket->standby) { if (rsc_ticket->role_lh != RSC_ROLE_MASTER || rsc_ticket->loss_policy == loss_ticket_stop) { resource_location(rsc_lh, NULL, -INFINITY, "__ticket_standby__", data_set); } } } enum pe_action_flags native_action_flags(action_t * action, node_t * node) { return action->flags; } static inline bool is_primitive_action(pe_action_t *action) { return action && action->rsc && (action->rsc->variant == pe_native); } /*! * \internal * \brief Set action bits appropriately when pe_restart_order is used * * \param[in] first 'First' action in an ordering with pe_restart_order * \param[in] then 'Then' action in an ordering with pe_restart_order * \param[in] filter What ordering flags to care about * * \note pe_restart_order is set for "stop resource before starting it" and * "stop later group member before stopping earlier group member" */ static void handle_restart_ordering(pe_action_t *first, pe_action_t *then, enum pe_action_flags filter) { const char *reason = NULL; CRM_ASSERT(is_primitive_action(first)); CRM_ASSERT(is_primitive_action(then)); // We need to update the action in two cases: // ... if 'then' is required if (is_set(filter, pe_action_optional) && is_not_set(then->flags, pe_action_optional)) { reason = "restart"; } /* ... if 'then' is unrunnable start of managed resource (if a resource * should restart but can't start, we still want to stop) */ if (is_set(filter, pe_action_runnable) && is_not_set(then->flags, pe_action_runnable) && is_set(then->rsc->flags, pe_rsc_managed) && safe_str_eq(then->task, RSC_START)) { reason = "stop"; } if (reason == NULL) { return; } pe_rsc_trace(first->rsc, "Handling %s -> %s for %s", first->uuid, then->uuid, reason); // Make 'first' required if it is runnable if (is_set(first->flags, pe_action_runnable)) { pe_action_implies(first, then, pe_action_optional); } // Make 'first' required if 'then' is required if (is_not_set(then->flags, pe_action_optional)) { pe_action_implies(first, then, pe_action_optional); } // Make 'first' unmigratable if 'then' is unmigratable if (is_not_set(then->flags, pe_action_migrate_runnable)) { pe_action_implies(first, then, pe_action_migrate_runnable); } // Make 'then' unrunnable if 'first' is required but unrunnable if (is_not_set(first->flags, pe_action_optional) && is_not_set(first->flags, pe_action_runnable)) { pe_action_implies(then, first, pe_action_runnable); } } enum pe_graph_flags native_update_actions(pe_action_t *first, pe_action_t *then, pe_node_t *node, enum pe_action_flags flags, enum pe_action_flags filter, enum pe_ordering type, pe_working_set_t *data_set) { /* flags == get_action_flags(first, then_node) called from update_action() */ enum pe_graph_flags changed = pe_graph_none; enum pe_action_flags then_flags = then->flags; enum pe_action_flags first_flags = first->flags; crm_trace( "Testing %s on %s (0x%.6x) with %s 0x%.6x", first->uuid, first->node ? first->node->details->uname : "[none]", first->flags, then->uuid, then->flags); if (type & pe_order_asymmetrical) { resource_t *then_rsc = then->rsc; enum rsc_role_e then_rsc_role = then_rsc ? then_rsc->fns->state(then_rsc, TRUE) : 0; if (!then_rsc) { /* ignore */ } else if ((then_rsc_role == RSC_ROLE_STOPPED) && safe_str_eq(then->task, RSC_STOP)) { /* ignore... if 'then' is supposed to be stopped after 'first', but * then is already stopped, there is nothing to be done when non-symmetrical. */ } else if ((then_rsc_role >= RSC_ROLE_STARTED) && safe_str_eq(then->task, RSC_START) && is_set(then->flags, pe_action_optional) && then->node && pcmk__list_of_1(then_rsc->running_on) && then->node->details == ((node_t *) then_rsc->running_on->data)->details) { /* Ignore. If 'then' is supposed to be started after 'first', but * 'then' is already started, there is nothing to be done when * asymmetrical -- unless the start is mandatory, which indicates * the resource is restarting, and the ordering is still needed. */ } else if (!(first->flags & pe_action_runnable)) { /* prevent 'then' action from happening if 'first' is not runnable and * 'then' has not yet occurred. */ pe_action_implies(then, first, pe_action_optional); pe_action_implies(then, first, pe_action_runnable); pe_rsc_trace(then->rsc, "Unset optional and runnable on %s", then->uuid); } else { /* ignore... then is allowed to start/stop if it wants to. */ } } if (type & pe_order_implies_first) { if (is_set(filter, pe_action_optional) && is_not_set(flags /* Should be then_flags? */, pe_action_optional)) { // Needs is_set(first_flags, pe_action_optional) too? pe_rsc_trace(first->rsc, "Unset optional on %s because of %s", first->uuid, then->uuid); pe_action_implies(first, then, pe_action_optional); } if (is_set(flags, pe_action_migrate_runnable) && is_set(then->flags, pe_action_migrate_runnable) == FALSE && is_set(then->flags, pe_action_optional) == FALSE) { pe_rsc_trace(first->rsc, "Unset migrate runnable on %s because of %s", first->uuid, then->uuid); pe_action_implies(first, then, pe_action_migrate_runnable); } } if (type & pe_order_implies_first_master) { if ((filter & pe_action_optional) && ((then->flags & pe_action_optional) == FALSE) && then->rsc && (then->rsc->role == RSC_ROLE_MASTER)) { pe_action_implies(first, then, pe_action_optional); if (is_set(first->flags, pe_action_migrate_runnable) && is_set(then->flags, pe_action_migrate_runnable) == FALSE) { pe_rsc_trace(first->rsc, "Unset migrate runnable on %s because of %s", first->uuid, then->uuid); pe_action_implies(first, then, pe_action_migrate_runnable); } pe_rsc_trace(then->rsc, "Unset optional on %s because of %s", first->uuid, then->uuid); } } if ((type & pe_order_implies_first_migratable) && is_set(filter, pe_action_optional)) { if (((then->flags & pe_action_migrate_runnable) == FALSE) || ((then->flags & pe_action_runnable) == FALSE)) { pe_rsc_trace(then->rsc, "Unset runnable on %s because %s is neither runnable or migratable", first->uuid, then->uuid); pe_action_implies(first, then, pe_action_runnable); } if ((then->flags & pe_action_optional) == 0) { pe_rsc_trace(then->rsc, "Unset optional on %s because %s is not optional", first->uuid, then->uuid); pe_action_implies(first, then, pe_action_optional); } } if ((type & pe_order_pseudo_left) && is_set(filter, pe_action_optional)) { if ((first->flags & pe_action_runnable) == FALSE) { pe_action_implies(then, first, pe_action_migrate_runnable); pe_clear_action_bit(then, pe_action_pseudo); pe_rsc_trace(then->rsc, "Unset pseudo on %s because %s is not runnable", then->uuid, first->uuid); } } if (is_set(type, pe_order_runnable_left) && is_set(filter, pe_action_runnable) && is_set(then->flags, pe_action_runnable) && is_set(flags, pe_action_runnable) == FALSE) { pe_rsc_trace(then->rsc, "Unset runnable on %s because of %s", then->uuid, first->uuid); pe_action_implies(then, first, pe_action_runnable); pe_action_implies(then, first, pe_action_migrate_runnable); } if (is_set(type, pe_order_implies_then) && is_set(filter, pe_action_optional) && is_set(then->flags, pe_action_optional) && is_set(flags, pe_action_optional) == FALSE) { /* in this case, treat migrate_runnable as if first is optional */ if (is_set(first->flags, pe_action_migrate_runnable) == FALSE) { pe_rsc_trace(then->rsc, "Unset optional on %s because of %s", then->uuid, first->uuid); pe_action_implies(then, first, pe_action_optional); } } if (is_set(type, pe_order_restart)) { handle_restart_ordering(first, then, filter); } if (then_flags != then->flags) { changed |= pe_graph_updated_then; pe_rsc_trace(then->rsc, "Then: Flags for %s on %s are now 0x%.6x (was 0x%.6x) because of %s 0x%.6x", then->uuid, then->node ? then->node->details->uname : "[none]", then->flags, then_flags, first->uuid, first->flags); if(then->rsc && then->rsc->parent) { /* "X_stop then X_start" doesn't get handled for cloned groups unless we do this */ update_action(then, data_set); } } if (first_flags != first->flags) { changed |= pe_graph_updated_first; pe_rsc_trace(first->rsc, "First: Flags for %s on %s are now 0x%.6x (was 0x%.6x) because of %s 0x%.6x", first->uuid, first->node ? first->node->details->uname : "[none]", first->flags, first_flags, then->uuid, then->flags); } return changed; } void native_rsc_location(pe_resource_t *rsc, pe__location_t *constraint) { GListPtr gIter = NULL; GHashTableIter iter; node_t *node = NULL; if (constraint == NULL) { pe_err("Constraint is NULL"); return; } else if (rsc == NULL) { pe_err("LHS of rsc_to_node (%s) is NULL", constraint->id); return; } pe_rsc_trace(rsc, "Applying %s (%s) to %s", constraint->id, role2text(constraint->role_filter), rsc->id); /* take "lifetime" into account */ if (constraint->role_filter > RSC_ROLE_UNKNOWN && constraint->role_filter != rsc->next_role) { pe_rsc_debug(rsc, "Constraint (%s) is not active (role : %s vs. %s)", constraint->id, role2text(constraint->role_filter), role2text(rsc->next_role)); return; } if (constraint->node_list_rh == NULL) { pe_rsc_trace(rsc, "RHS of constraint %s is NULL", constraint->id); return; } for (gIter = constraint->node_list_rh; gIter != NULL; gIter = gIter->next) { node_t *node = (node_t *) gIter->data; node_t *other_node = NULL; other_node = (node_t *) pe_hash_table_lookup(rsc->allowed_nodes, node->details->id); if (other_node != NULL) { pe_rsc_trace(rsc, "%s + %s: %d + %d", node->details->uname, other_node->details->uname, node->weight, other_node->weight); other_node->weight = merge_weights(other_node->weight, node->weight); } else { other_node = node_copy(node); pe_rsc_trace(rsc, "%s: %d (insert %d)", other_node->details->uname, other_node->weight, constraint->discover_mode); g_hash_table_insert(rsc->allowed_nodes, (gpointer) other_node->details->id, other_node); } if (other_node->rsc_discover_mode < constraint->discover_mode) { if (constraint->discover_mode == pe_discover_exclusive) { rsc->exclusive_discover = TRUE; } /* exclusive > never > always... always is default */ other_node->rsc_discover_mode = constraint->discover_mode; } } g_hash_table_iter_init(&iter, rsc->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (void **)&node)) { pe_rsc_trace(rsc, "%s + %s : %d", rsc->id, node->details->uname, node->weight); } } void native_expand(resource_t * rsc, pe_working_set_t * data_set) { GListPtr gIter = NULL; CRM_ASSERT(rsc); pe_rsc_trace(rsc, "Processing actions from %s", rsc->id); for (gIter = rsc->actions; gIter != NULL; gIter = gIter->next) { action_t *action = (action_t *) gIter->data; crm_trace("processing action %d for rsc=%s", action->id, rsc->id); graph_element_from_action(action, data_set); } for (gIter = rsc->children; gIter != NULL; gIter = gIter->next) { resource_t *child_rsc = (resource_t *) gIter->data; child_rsc->cmds->expand(child_rsc, data_set); } } #define log_change(a, fmt, args...) do { \ if(a && a->reason && terminal) { \ printf(" * "fmt" \tdue to %s\n", ##args, a->reason); \ } else if(a && a->reason) { \ crm_notice(fmt" \tdue to %s", ##args, a->reason); \ } else if(terminal) { \ printf(" * "fmt"\n", ##args); \ } else { \ crm_notice(fmt, ##args); \ } \ } while(0) #define STOP_SANITY_ASSERT(lineno) do { \ if(current && current->details->unclean) { \ /* It will be a pseudo op */ \ } else if(stop == NULL) { \ crm_err("%s:%d: No stop action exists for %s", __FUNCTION__, lineno, rsc->id); \ CRM_ASSERT(stop != NULL); \ } else if(is_set(stop->flags, pe_action_optional)) { \ crm_err("%s:%d: Action %s is still optional", __FUNCTION__, lineno, stop->uuid); \ CRM_ASSERT(is_not_set(stop->flags, pe_action_optional)); \ } \ } while(0) static int rsc_width = 5; static int detail_width = 5; static void LogAction(const char *change, resource_t *rsc, pe_node_t *origin, pe_node_t *destination, pe_action_t *action, pe_action_t *source, gboolean terminal) { int len = 0; char *reason = NULL; char *details = NULL; bool same_host = FALSE; bool same_role = FALSE; bool need_role = FALSE; CRM_ASSERT(action); CRM_ASSERT(destination != NULL || origin != NULL); if(source == NULL) { source = action; } len = strlen(rsc->id); if(len > rsc_width) { rsc_width = len + 2; } if(rsc->role > RSC_ROLE_STARTED || rsc->next_role > RSC_ROLE_SLAVE) { need_role = TRUE; } if(origin != NULL && destination != NULL && origin->details == destination->details) { same_host = TRUE; } if(rsc->role == rsc->next_role) { same_role = TRUE; } if(need_role && origin == NULL) { /* Promoting from Stopped */ details = crm_strdup_printf("%s -> %s %s", role2text(rsc->role), role2text(rsc->next_role), destination->details->uname); } else if(need_role && destination == NULL) { /* Demoting a Master or Stopping a Slave */ details = crm_strdup_printf("%s %s", role2text(rsc->role), origin->details->uname); } else if(origin == NULL || destination == NULL) { /* Starting or stopping a resource */ details = crm_strdup_printf("%s", origin?origin->details->uname:destination->details->uname); } else if(need_role && same_role && same_host) { /* Recovering or restarting a promotable clone resource */ details = crm_strdup_printf("%s %s", role2text(rsc->role), origin->details->uname); } else if(same_role && same_host) { /* Recovering or Restarting a normal resource */ details = crm_strdup_printf("%s", origin->details->uname); } else if(same_role && need_role) { /* Moving a promotable clone resource */ details = crm_strdup_printf("%s -> %s %s", origin->details->uname, destination->details->uname, role2text(rsc->role)); } else if(same_role) { /* Moving a normal resource */ details = crm_strdup_printf("%s -> %s", origin->details->uname, destination->details->uname); } else if(same_host) { /* Promoting or demoting a promotable clone resource */ details = crm_strdup_printf("%s -> %s %s", role2text(rsc->role), role2text(rsc->next_role), origin->details->uname); } else { /* Moving and promoting/demoting */ details = crm_strdup_printf("%s %s -> %s %s", role2text(rsc->role), origin->details->uname, role2text(rsc->next_role), destination->details->uname); } len = strlen(details); if(len > detail_width) { detail_width = len; } if(source->reason && is_not_set(action->flags, pe_action_runnable)) { reason = crm_strdup_printf(" due to %s (blocked)", source->reason); } else if(source->reason) { reason = crm_strdup_printf(" due to %s", source->reason); } else if(is_not_set(action->flags, pe_action_runnable)) { reason = strdup(" blocked"); } else { reason = strdup(""); } if(terminal) { printf(" * %-8s %-*s ( %*s ) %s\n", change, rsc_width, rsc->id, detail_width, details, reason); } else { crm_notice(" * %-8s %-*s ( %*s ) %s", change, rsc_width, rsc->id, detail_width, details, reason); } free(details); free(reason); } void LogActions(resource_t * rsc, pe_working_set_t * data_set, gboolean terminal) { node_t *next = NULL; node_t *current = NULL; pe_node_t *start_node = NULL; action_t *stop = NULL; action_t *start = NULL; action_t *demote = NULL; action_t *promote = NULL; char *key = NULL; gboolean moving = FALSE; GListPtr possible_matches = NULL; if(rsc->variant == pe_container) { pcmk__bundle_log_actions(rsc, data_set, terminal); return; } if (rsc->children) { GListPtr gIter = NULL; for (gIter = rsc->children; gIter != NULL; gIter = gIter->next) { resource_t *child_rsc = (resource_t *) gIter->data; LogActions(child_rsc, data_set, terminal); } return; } next = rsc->allocated_to; if (rsc->running_on) { current = pe__current_node(rsc); if (rsc->role == RSC_ROLE_STOPPED) { /* * This can occur when resources are being recovered * We fiddle with the current role in native_create_actions() */ rsc->role = RSC_ROLE_STARTED; } } if (current == NULL && is_set(rsc->flags, pe_rsc_orphan)) { /* Don't log stopped orphans */ return; } if (is_not_set(rsc->flags, pe_rsc_managed) || (current == NULL && next == NULL)) { pe_rsc_info(rsc, "Leave %s\t(%s%s)", rsc->id, role2text(rsc->role), is_not_set(rsc->flags, pe_rsc_managed) ? " unmanaged" : ""); return; } if (current != NULL && next != NULL && safe_str_neq(current->details->id, next->details->id)) { moving = TRUE; } possible_matches = pe__resource_actions(rsc, next, RSC_START, FALSE); if (possible_matches) { start = possible_matches->data; g_list_free(possible_matches); } if ((start == NULL) || is_not_set(start->flags, pe_action_runnable)) { start_node = NULL; } else { start_node = current; } possible_matches = pe__resource_actions(rsc, start_node, RSC_STOP, FALSE); if (possible_matches) { stop = possible_matches->data; g_list_free(possible_matches); } possible_matches = pe__resource_actions(rsc, next, RSC_PROMOTE, FALSE); if (possible_matches) { promote = possible_matches->data; g_list_free(possible_matches); } possible_matches = pe__resource_actions(rsc, next, RSC_DEMOTE, FALSE); if (possible_matches) { demote = possible_matches->data; g_list_free(possible_matches); } if (rsc->role == rsc->next_role) { action_t *migrate_op = NULL; possible_matches = pe__resource_actions(rsc, next, RSC_MIGRATED, FALSE); if (possible_matches) { migrate_op = possible_matches->data; } CRM_CHECK(next != NULL,); if (next == NULL) { } else if (migrate_op && is_set(migrate_op->flags, pe_action_runnable) && current) { LogAction("Migrate", rsc, current, next, start, NULL, terminal); } else if (is_set(rsc->flags, pe_rsc_reload)) { LogAction("Reload", rsc, current, next, start, NULL, terminal); } else if (start == NULL || is_set(start->flags, pe_action_optional)) { pe_rsc_info(rsc, "Leave %s\t(%s %s)", rsc->id, role2text(rsc->role), next->details->uname); } else if (start && is_set(start->flags, pe_action_runnable) == FALSE) { LogAction("Stop", rsc, current, NULL, stop, (stop && stop->reason)? stop : start, terminal); STOP_SANITY_ASSERT(__LINE__); } else if (moving && current) { LogAction(is_set(rsc->flags, pe_rsc_failed) ? "Recover" : "Move", rsc, current, next, stop, NULL, terminal); } else if (is_set(rsc->flags, pe_rsc_failed)) { LogAction("Recover", rsc, current, NULL, stop, NULL, terminal); STOP_SANITY_ASSERT(__LINE__); } else { LogAction("Restart", rsc, current, next, start, NULL, terminal); /* STOP_SANITY_ASSERT(__LINE__); False positive for migrate-fail-7 */ } g_list_free(possible_matches); return; } if(stop && (rsc->next_role == RSC_ROLE_STOPPED || (start && is_not_set(start->flags, pe_action_runnable)))) { GListPtr gIter = NULL; key = stop_key(rsc); for (gIter = rsc->running_on; gIter != NULL; gIter = gIter->next) { node_t *node = (node_t *) gIter->data; action_t *stop_op = NULL; possible_matches = find_actions(rsc->actions, key, node); if (possible_matches) { stop_op = possible_matches->data; g_list_free(possible_matches); } if (stop_op && (stop_op->flags & pe_action_runnable)) { STOP_SANITY_ASSERT(__LINE__); } LogAction("Stop", rsc, node, NULL, stop_op, (stop_op && stop_op->reason)? stop_op : start, terminal); } free(key); } else if (stop && is_set(rsc->flags, pe_rsc_failed)) { /* 'stop' may be NULL if the failure was ignored */ LogAction("Recover", rsc, current, next, stop, start, terminal); STOP_SANITY_ASSERT(__LINE__); } else if (moving) { LogAction("Move", rsc, current, next, stop, NULL, terminal); STOP_SANITY_ASSERT(__LINE__); } else if (is_set(rsc->flags, pe_rsc_reload)) { LogAction("Reload", rsc, current, next, start, NULL, terminal); } else if (stop != NULL && is_not_set(stop->flags, pe_action_optional)) { LogAction("Restart", rsc, current, next, start, NULL, terminal); STOP_SANITY_ASSERT(__LINE__); } else if (rsc->role == RSC_ROLE_MASTER) { CRM_LOG_ASSERT(current != NULL); LogAction("Demote", rsc, current, next, demote, NULL, terminal); } else if(rsc->next_role == RSC_ROLE_MASTER) { CRM_LOG_ASSERT(next); LogAction("Promote", rsc, current, next, promote, NULL, terminal); } else if (rsc->role == RSC_ROLE_STOPPED && rsc->next_role > RSC_ROLE_STOPPED) { LogAction("Start", rsc, current, next, start, NULL, terminal); } } gboolean StopRsc(resource_t * rsc, node_t * next, gboolean optional, pe_working_set_t * data_set) { GListPtr gIter = NULL; CRM_ASSERT(rsc); pe_rsc_trace(rsc, "%s", rsc->id); for (gIter = rsc->running_on; gIter != NULL; gIter = gIter->next) { node_t *current = (node_t *) gIter->data; action_t *stop; if (rsc->partial_migration_target) { if (rsc->partial_migration_target->details == current->details) { pe_rsc_trace(rsc, "Filtered %s -> %s %s", current->details->uname, next->details->uname, rsc->id); continue; } else { pe_rsc_trace(rsc, "Forced on %s %s", current->details->uname, rsc->id); optional = FALSE; } } pe_rsc_trace(rsc, "%s on %s", rsc->id, current->details->uname); stop = stop_action(rsc, current, optional); if(rsc->allocated_to == NULL) { pe_action_set_reason(stop, "node availability", TRUE); } if (is_not_set(rsc->flags, pe_rsc_managed)) { update_action_flags(stop, pe_action_runnable | pe_action_clear, __FUNCTION__, __LINE__); } if (is_set(data_set->flags, pe_flag_remove_after_stop)) { DeleteRsc(rsc, current, optional, data_set); } if(is_set(rsc->flags, pe_rsc_needs_unfencing)) { action_t *unfence = pe_fence_op(current, "on", TRUE, NULL, data_set); order_actions(stop, unfence, pe_order_implies_first); if (!node_has_been_unfenced(current)) { pe_proc_err("Stopping %s until %s can be unfenced", rsc->id, current->details->uname); } } } return TRUE; } static void order_after_unfencing(resource_t *rsc, pe_node_t *node, action_t *action, enum pe_ordering order, pe_working_set_t *data_set) { /* When unfencing is in use, we order unfence actions before any probe or * start of resources that require unfencing, and also of fence devices. * * This might seem to violate the principle that fence devices require * only quorum. However, fence agents that unfence often don't have enough * information to even probe or start unless the node is first unfenced. */ if (is_unfence_device(rsc, data_set) || is_set(rsc->flags, pe_rsc_needs_unfencing)) { /* Start with an optional ordering. Requiring unfencing would result in * the node being unfenced, and all its resources being stopped, * whenever a new resource is added -- which would be highly suboptimal. */ action_t *unfence = pe_fence_op(node, "on", TRUE, NULL, data_set); order_actions(unfence, action, order); if (!node_has_been_unfenced(node)) { // But unfencing is required if it has never been done char *reason = crm_strdup_printf("required by %s %s", rsc->id, action->task); trigger_unfencing(NULL, node, reason, NULL, data_set); free(reason); } } } gboolean StartRsc(resource_t * rsc, node_t * next, gboolean optional, pe_working_set_t * data_set) { action_t *start = NULL; CRM_ASSERT(rsc); pe_rsc_trace(rsc, "%s on %s %d %d", rsc->id, next ? next->details->uname : "N/A", optional, next ? next->weight : 0); start = start_action(rsc, next, TRUE); order_after_unfencing(rsc, next, start, pe_order_implies_then, data_set); if (is_set(start->flags, pe_action_runnable) && optional == FALSE) { update_action_flags(start, pe_action_optional | pe_action_clear, __FUNCTION__, __LINE__); } return TRUE; } gboolean PromoteRsc(resource_t * rsc, node_t * next, gboolean optional, pe_working_set_t * data_set) { GListPtr gIter = NULL; gboolean runnable = TRUE; GListPtr action_list = NULL; CRM_ASSERT(rsc); CRM_CHECK(next != NULL, return FALSE); pe_rsc_trace(rsc, "%s on %s", rsc->id, next->details->uname); action_list = pe__resource_actions(rsc, next, RSC_START, TRUE); for (gIter = action_list; gIter != NULL; gIter = gIter->next) { action_t *start = (action_t *) gIter->data; if (is_set(start->flags, pe_action_runnable) == FALSE) { runnable = FALSE; } } g_list_free(action_list); if (runnable) { promote_action(rsc, next, optional); return TRUE; } pe_rsc_debug(rsc, "%s\tPromote %s (canceled)", next->details->uname, rsc->id); action_list = pe__resource_actions(rsc, next, RSC_PROMOTE, TRUE); for (gIter = action_list; gIter != NULL; gIter = gIter->next) { action_t *promote = (action_t *) gIter->data; update_action_flags(promote, pe_action_runnable | pe_action_clear, __FUNCTION__, __LINE__); } g_list_free(action_list); return TRUE; } gboolean DemoteRsc(resource_t * rsc, node_t * next, gboolean optional, pe_working_set_t * data_set) { GListPtr gIter = NULL; CRM_ASSERT(rsc); pe_rsc_trace(rsc, "%s", rsc->id); /* CRM_CHECK(rsc->next_role == RSC_ROLE_SLAVE, return FALSE); */ for (gIter = rsc->running_on; gIter != NULL; gIter = gIter->next) { node_t *current = (node_t *) gIter->data; pe_rsc_trace(rsc, "%s on %s", rsc->id, next ? next->details->uname : "N/A"); demote_action(rsc, current, optional); } return TRUE; } gboolean RoleError(resource_t * rsc, node_t * next, gboolean optional, pe_working_set_t * data_set) { CRM_ASSERT(rsc); crm_err("%s on %s", rsc->id, next ? next->details->uname : "N/A"); CRM_CHECK(FALSE, return FALSE); return FALSE; } gboolean NullOp(resource_t * rsc, node_t * next, gboolean optional, pe_working_set_t * data_set) { CRM_ASSERT(rsc); pe_rsc_trace(rsc, "%s", rsc->id); return FALSE; } gboolean DeleteRsc(resource_t * rsc, node_t * node, gboolean optional, pe_working_set_t * data_set) { if (is_set(rsc->flags, pe_rsc_failed)) { pe_rsc_trace(rsc, "Resource %s not deleted from %s: failed", rsc->id, node->details->uname); return FALSE; } else if (node == NULL) { pe_rsc_trace(rsc, "Resource %s not deleted: NULL node", rsc->id); return FALSE; } else if (node->details->unclean || node->details->online == FALSE) { pe_rsc_trace(rsc, "Resource %s not deleted from %s: unrunnable", rsc->id, node->details->uname); return FALSE; } crm_notice("Removing %s from %s", rsc->id, node->details->uname); delete_action(rsc, node, optional); new_rsc_order(rsc, RSC_STOP, rsc, RSC_DELETE, optional ? pe_order_implies_then : pe_order_optional, data_set); new_rsc_order(rsc, RSC_DELETE, rsc, RSC_START, optional ? pe_order_implies_then : pe_order_optional, data_set); return TRUE; } gboolean native_create_probe(resource_t * rsc, node_t * node, action_t * complete, gboolean force, pe_working_set_t * data_set) { enum pe_ordering flags = pe_order_optional; char *key = NULL; action_t *probe = NULL; node_t *running = NULL; node_t *allowed = NULL; resource_t *top = uber_parent(rsc); static const char *rc_master = NULL; static const char *rc_inactive = NULL; if (rc_inactive == NULL) { rc_inactive = crm_itoa(PCMK_OCF_NOT_RUNNING); rc_master = crm_itoa(PCMK_OCF_RUNNING_MASTER); } CRM_CHECK(node != NULL, return FALSE); if (force == FALSE && is_not_set(data_set->flags, pe_flag_startup_probes)) { pe_rsc_trace(rsc, "Skipping active resource detection for %s", rsc->id); return FALSE; } if (pe__is_guest_or_remote_node(node)) { const char *class = crm_element_value(rsc->xml, XML_AGENT_ATTR_CLASS); if (safe_str_eq(class, PCMK_RESOURCE_CLASS_STONITH)) { pe_rsc_trace(rsc, "Skipping probe for %s on %s because Pacemaker Remote nodes cannot run stonith agents", rsc->id, node->details->id); return FALSE; } else if (pe__is_guest_node(node) && pe__resource_contains_guest_node(data_set, rsc)) { pe_rsc_trace(rsc, "Skipping probe for %s on %s because guest nodes cannot run resources containing guest nodes", rsc->id, node->details->id); return FALSE; } else if (rsc->is_remote_node) { pe_rsc_trace(rsc, "Skipping probe for %s on %s because Pacemaker Remote nodes cannot host remote connections", rsc->id, node->details->id); return FALSE; } } if (rsc->children) { GListPtr gIter = NULL; gboolean any_created = FALSE; for (gIter = rsc->children; gIter != NULL; gIter = gIter->next) { resource_t *child_rsc = (resource_t *) gIter->data; any_created = child_rsc->cmds->create_probe(child_rsc, node, complete, force, data_set) || any_created; } return any_created; } else if ((rsc->container) && (!rsc->is_remote_node)) { pe_rsc_trace(rsc, "Skipping %s: it is within container %s", rsc->id, rsc->container->id); return FALSE; } if (is_set(rsc->flags, pe_rsc_orphan)) { pe_rsc_trace(rsc, "Skipping orphan: %s", rsc->id); return FALSE; } // Check whether resource is already known on node if (!force && g_hash_table_lookup(rsc->known_on, node->details->id)) { pe_rsc_trace(rsc, "Skipping known: %s on %s", rsc->id, node->details->uname); return FALSE; } allowed = g_hash_table_lookup(rsc->allowed_nodes, node->details->id); if (rsc->exclusive_discover || top->exclusive_discover) { if (allowed == NULL) { /* exclusive discover is enabled and this node is not in the allowed list. */ pe_rsc_trace(rsc, "Skipping probe for %s on node %s, A", rsc->id, node->details->id); return FALSE; } else if (allowed->rsc_discover_mode != pe_discover_exclusive) { /* exclusive discover is enabled and this node is not marked * as a node this resource should be discovered on */ pe_rsc_trace(rsc, "Skipping probe for %s on node %s, B", rsc->id, node->details->id); return FALSE; } } if(allowed == NULL && node->rsc_discover_mode == pe_discover_never) { /* If this node was allowed to host this resource it would * have been explicitly added to the 'allowed_nodes' list. * However it wasn't and the node has discovery disabled, so * no need to probe for this resource. */ pe_rsc_trace(rsc, "Skipping probe for %s on node %s, C", rsc->id, node->details->id); return FALSE; } if (allowed && allowed->rsc_discover_mode == pe_discover_never) { /* this resource is marked as not needing to be discovered on this node */ pe_rsc_trace(rsc, "Skipping probe for %s on node %s, discovery mode", rsc->id, node->details->id); return FALSE; } if (pe__is_guest_node(node)) { resource_t *remote = node->details->remote_rsc->container; if(remote->role == RSC_ROLE_STOPPED) { /* If the container is stopped, then we know anything that * might have been inside it is also stopped and there is * no need to probe. * * If we don't know the container's state on the target * either: * * - the container is running, the transition will abort * and we'll end up in a different case next time, or * * - the container is stopped * * Either way there is no need to probe. * */ if(remote->allocated_to && g_hash_table_lookup(remote->known_on, remote->allocated_to->details->id) == NULL) { /* For safety, we order the 'rsc' start after 'remote' * has been probed. * * Using 'top' helps for groups, but we may need to * follow the start's ordering chain backwards. */ custom_action_order(remote, pcmk__op_key(remote->id, RSC_STATUS, 0), NULL, top, pcmk__op_key(top->id, RSC_START, 0), NULL, pe_order_optional, data_set); } pe_rsc_trace(rsc, "Skipping probe for %s on node %s, %s is stopped", rsc->id, node->details->id, remote->id); return FALSE; /* Here we really we want to check if remote->stop is required, * but that information doesn't exist yet */ } else if(node->details->remote_requires_reset || node->details->unclean || is_set(remote->flags, pe_rsc_failed) || remote->next_role == RSC_ROLE_STOPPED || (remote->allocated_to && pe_find_node(remote->running_on, remote->allocated_to->details->uname) == NULL) ) { /* The container is stopping or restarting, don't start * 'rsc' until 'remote' stops as this also implies that * 'rsc' is stopped - avoiding the need to probe */ custom_action_order(remote, pcmk__op_key(remote->id, RSC_STOP, 0), NULL, top, pcmk__op_key(top->id, RSC_START, 0), NULL, pe_order_optional, data_set); pe_rsc_trace(rsc, "Skipping probe for %s on node %s, %s is stopping, restarting or moving", rsc->id, node->details->id, remote->id); return FALSE; /* } else { * The container is running so there is no problem probing it */ } } key = pcmk__op_key(rsc->id, RSC_STATUS, 0); probe = custom_action(rsc, key, RSC_STATUS, node, FALSE, TRUE, data_set); update_action_flags(probe, pe_action_optional | pe_action_clear, __FUNCTION__, __LINE__); order_after_unfencing(rsc, node, probe, pe_order_optional, data_set); /* * We need to know if it's running_on (not just known_on) this node * to correctly determine the target rc. */ running = pe_find_node_id(rsc->running_on, node->details->id); if (running == NULL) { add_hash_param(probe->meta, XML_ATTR_TE_TARGET_RC, rc_inactive); } else if (rsc->role == RSC_ROLE_MASTER) { add_hash_param(probe->meta, XML_ATTR_TE_TARGET_RC, rc_master); } crm_debug("Probing %s on %s (%s) %d %p", rsc->id, node->details->uname, role2text(rsc->role), is_set(probe->flags, pe_action_runnable), rsc->running_on); if (is_unfence_device(rsc, data_set) || !pe_rsc_is_clone(top)) { top = rsc; } else { crm_trace("Probing %s on %s (%s) as %s", rsc->id, node->details->uname, role2text(rsc->role), top->id); } if(is_not_set(probe->flags, pe_action_runnable) && rsc->running_on == NULL) { /* Prevent the start from occurring if rsc isn't active, but * don't cause it to stop if it was active already */ flags |= pe_order_runnable_left; } custom_action_order(rsc, NULL, probe, top, pcmk__op_key(top->id, RSC_START, 0), NULL, flags, data_set); /* Before any reloads, if they exist */ custom_action_order(rsc, NULL, probe, top, reload_key(rsc), NULL, pe_order_optional, data_set); #if 0 // complete is always null currently if (!is_unfence_device(rsc, data_set)) { /* Normally rsc.start depends on probe complete which depends * on rsc.probe. But this can't be the case for fence devices * with unfencing, as it would create graph loops. * * So instead we explicitly order 'rsc.probe then rsc.start' */ order_actions(probe, complete, pe_order_implies_then); } #endif return TRUE; } /*! * \internal * \brief Check whether a resource is known on a particular node * * \param[in] rsc Resource to check * \param[in] node Node to check * * \return TRUE if resource (or parent if an anonymous clone) is known */ static bool rsc_is_known_on(pe_resource_t *rsc, const pe_node_t *node) { if (pe_hash_table_lookup(rsc->known_on, node->details->id)) { return TRUE; } else if ((rsc->variant == pe_native) && pe_rsc_is_anon_clone(rsc->parent) && pe_hash_table_lookup(rsc->parent->known_on, node->details->id)) { /* We check only the parent, not the uber-parent, because we cannot * assume that the resource is known if it is in an anonymously cloned * group (which may be only partially known). */ return TRUE; } return FALSE; } /*! * \internal * \brief Order a resource's start and promote actions relative to fencing * * \param[in] rsc Resource to be ordered * \param[in] stonith_op Fence action * \param[in] data_set Cluster information */ static void native_start_constraints(resource_t * rsc, action_t * stonith_op, pe_working_set_t * data_set) { node_t *target; GListPtr gIter = NULL; CRM_CHECK(stonith_op && stonith_op->node, return); target = stonith_op->node; for (gIter = rsc->actions; gIter != NULL; gIter = gIter->next) { action_t *action = (action_t *) gIter->data; switch (action->needs) { case rsc_req_nothing: // Anything other than start or promote requires nothing break; case rsc_req_stonith: order_actions(stonith_op, action, pe_order_optional); break; case rsc_req_quorum: if (safe_str_eq(action->task, RSC_START) && pe_hash_table_lookup(rsc->allowed_nodes, target->details->id) && !rsc_is_known_on(rsc, target)) { /* If we don't know the status of the resource on the node * we're about to shoot, we have to assume it may be active * there. Order the resource start after the fencing. This * is analogous to waiting for all the probes for a resource * to complete before starting it. * * The most likely explanation is that the DC died and took * its status with it. */ pe_rsc_debug(rsc, "Ordering %s after %s recovery", action->uuid, target->details->uname); order_actions(stonith_op, action, pe_order_optional | pe_order_runnable_left); } break; } } } static void native_stop_constraints(resource_t * rsc, action_t * stonith_op, pe_working_set_t * data_set) { GListPtr gIter = NULL; GListPtr action_list = NULL; bool order_implicit = false; resource_t *top = uber_parent(rsc); pe_action_t *parent_stop = NULL; node_t *target; CRM_CHECK(stonith_op && stonith_op->node, return); target = stonith_op->node; /* Get a list of stop actions potentially implied by the fencing */ action_list = pe__resource_actions(rsc, target, RSC_STOP, FALSE); /* If resource requires fencing, implicit actions must occur after fencing. * * Implied stops and demotes of resources running on guest nodes are always * ordered after fencing, even if the resource does not require fencing, * because guest node "fencing" is actually just a resource stop. */ if (is_set(rsc->flags, pe_rsc_needs_fencing) || pe__is_guest_node(target)) { order_implicit = true; } if (action_list && order_implicit) { parent_stop = find_first_action(top->actions, NULL, RSC_STOP, NULL); } for (gIter = action_list; gIter != NULL; gIter = gIter->next) { action_t *action = (action_t *) gIter->data; // The stop would never complete, so convert it into a pseudo-action. update_action_flags(action, pe_action_pseudo|pe_action_runnable, __FUNCTION__, __LINE__); if (order_implicit) { update_action_flags(action, pe_action_implied_by_stonith, __FUNCTION__, __LINE__); /* Order the stonith before the parent stop (if any). * * Also order the stonith before the resource stop, unless the * resource is inside a bundle -- that would cause a graph loop. * We can rely on the parent stop's ordering instead. * * User constraints must not order a resource in a guest node * relative to the guest node container resource. The * pe_order_preserve flag marks constraints as generated by the * cluster and thus immune to that check (and is irrelevant if * target is not a guest). */ if (!pe_rsc_is_bundled(rsc)) { order_actions(stonith_op, action, pe_order_preserve); } order_actions(stonith_op, parent_stop, pe_order_preserve); } if (is_set(rsc->flags, pe_rsc_failed)) { crm_notice("Stop of failed resource %s is implicit %s %s is fenced", rsc->id, (order_implicit? "after" : "because"), target->details->uname); } else { crm_info("%s is implicit %s %s is fenced", action->uuid, (order_implicit? "after" : "because"), target->details->uname); } if (is_set(rsc->flags, pe_rsc_notify)) { /* Create a second notification that will be delivered * immediately after the node is fenced * * Basic problem: * - C is a clone active on the node to be shot and stopping on another * - R is a resource that depends on C * * + C.stop depends on R.stop * + C.stopped depends on STONITH * + C.notify depends on C.stopped * + C.healthy depends on C.notify * + R.stop depends on C.healthy * * The extra notification here changes * + C.healthy depends on C.notify * into: * + C.healthy depends on C.notify' * + C.notify' depends on STONITH' * thus breaking the loop */ create_secondary_notification(action, rsc, stonith_op, data_set); } /* From Bug #1601, successful fencing must be an input to a failed resources stop action. However given group(rA, rB) running on nodeX and B.stop has failed, A := stop healthy resource (rA.stop) B := stop failed resource (pseudo operation B.stop) C := stonith nodeX A requires B, B requires C, C requires A This loop would prevent the cluster from making progress. This block creates the "C requires A" dependency and therefore must (at least for now) be disabled. Instead, run the block above and treat all resources on nodeX as B would be (marked as a pseudo op depending on the STONITH). TODO: Break the "A requires B" dependency in update_action() and re-enable this block } else if(is_stonith == FALSE) { crm_info("Moving healthy resource %s" " off %s before fencing", rsc->id, node->details->uname); * stop healthy resources before the * stonith op * custom_action_order( rsc, stop_key(rsc), NULL, NULL,strdup(CRM_OP_FENCE),stonith_op, pe_order_optional, data_set); */ } g_list_free(action_list); /* Get a list of demote actions potentially implied by the fencing */ action_list = pe__resource_actions(rsc, target, RSC_DEMOTE, FALSE); for (gIter = action_list; gIter != NULL; gIter = gIter->next) { action_t *action = (action_t *) gIter->data; if (action->node->details->online == FALSE || action->node->details->unclean == TRUE || is_set(rsc->flags, pe_rsc_failed)) { if (is_set(rsc->flags, pe_rsc_failed)) { pe_rsc_info(rsc, "Demote of failed resource %s is implicit after %s is fenced", rsc->id, target->details->uname); } else { pe_rsc_info(rsc, "%s is implicit after %s is fenced", action->uuid, target->details->uname); } /* The demote would never complete and is now implied by the * fencing, so convert it into a pseudo-action. */ update_action_flags(action, pe_action_pseudo|pe_action_runnable, __FUNCTION__, __LINE__); if (pe_rsc_is_bundled(rsc)) { /* Do nothing, let the recovery be ordered after the parent's implied stop */ } else if (order_implicit) { order_actions(stonith_op, action, pe_order_preserve|pe_order_optional); } } } g_list_free(action_list); } void rsc_stonith_ordering(resource_t * rsc, action_t * stonith_op, pe_working_set_t * data_set) { if (rsc->children) { GListPtr gIter = NULL; for (gIter = rsc->children; gIter != NULL; gIter = gIter->next) { resource_t *child_rsc = (resource_t *) gIter->data; rsc_stonith_ordering(child_rsc, stonith_op, data_set); } } else if (is_not_set(rsc->flags, pe_rsc_managed)) { pe_rsc_trace(rsc, "Skipping fencing constraints for unmanaged resource: %s", rsc->id); } else { native_start_constraints(rsc, stonith_op, data_set); native_stop_constraints(rsc, stonith_op, data_set); } } void ReloadRsc(resource_t * rsc, node_t *node, pe_working_set_t * data_set) { GListPtr gIter = NULL; action_t *reload = NULL; if (rsc->children) { for (gIter = rsc->children; gIter != NULL; gIter = gIter->next) { resource_t *child_rsc = (resource_t *) gIter->data; ReloadRsc(child_rsc, node, data_set); } return; } else if (rsc->variant > pe_native) { /* Complex resource with no children */ return; } else if (is_not_set(rsc->flags, pe_rsc_managed)) { pe_rsc_trace(rsc, "%s: unmanaged", rsc->id); return; } else if (is_set(rsc->flags, pe_rsc_failed) || is_set(rsc->flags, pe_rsc_start_pending)) { pe_rsc_trace(rsc, "%s: general resource state: flags=0x%.16llx", rsc->id, rsc->flags); stop_action(rsc, node, FALSE); /* Force a full restart, overkill? */ return; } else if (node == NULL) { pe_rsc_trace(rsc, "%s: not active", rsc->id); return; } pe_rsc_trace(rsc, "Processing %s", rsc->id); set_bit(rsc->flags, pe_rsc_reload); reload = custom_action( rsc, reload_key(rsc), CRMD_ACTION_RELOAD, node, FALSE, TRUE, data_set); pe_action_set_reason(reload, "resource definition change", FALSE); custom_action_order(NULL, NULL, reload, rsc, stop_key(rsc), NULL, pe_order_optional|pe_order_then_cancels_first, data_set); custom_action_order(NULL, NULL, reload, rsc, demote_key(rsc), NULL, pe_order_optional|pe_order_then_cancels_first, data_set); } void native_append_meta(resource_t * rsc, xmlNode * xml) { char *value = g_hash_table_lookup(rsc->meta, XML_RSC_ATTR_INCARNATION); resource_t *parent; if (value) { char *name = NULL; name = crm_meta_name(XML_RSC_ATTR_INCARNATION); crm_xml_add(xml, name, value); free(name); } value = g_hash_table_lookup(rsc->meta, XML_RSC_ATTR_REMOTE_NODE); if (value) { char *name = NULL; name = crm_meta_name(XML_RSC_ATTR_REMOTE_NODE); crm_xml_add(xml, name, value); free(name); } for (parent = rsc; parent != NULL; parent = parent->parent) { if (parent->container) { crm_xml_add(xml, CRM_META"_"XML_RSC_ATTR_CONTAINER, parent->container->id); } } }