diff --git a/include/crm/common/scheduler_internal.h b/include/crm/common/scheduler_internal.h index b83291007b..7b63d2078d 100644 --- a/include/crm/common/scheduler_internal.h +++ b/include/crm/common/scheduler_internal.h @@ -1,79 +1,89 @@ /* * Copyright 2004-2023 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU Lesser General Public License * version 2.1 or later (LGPLv2.1+) WITHOUT ANY WARRANTY. */ #ifndef PCMK__CRM_COMMON_SCHEDULER_INTERNAL__H # define PCMK__CRM_COMMON_SCHEDULER_INTERNAL__H #include #include #include #include #include #include #include #include #ifdef __cplusplus extern "C" { #endif /* Some warnings are too noisy when logged every time a give function is called * (for example, using a deprecated feature). As an alternative, we allow * warnings to be logged once per scheduler sequence (transition). Each of those * warnings needs a flag defined here. */ enum pcmk__sched_warnings { pcmk__wo_blind = (1 << 0), pcmk__wo_restart_type = (1 << 1), pcmk__wo_role_after = (1 << 2), pcmk__wo_poweroff = (1 << 3), pcmk__wo_require_all = (1 << 4), pcmk__wo_order_score = (1 << 5), pcmk__wo_neg_threshold = (1 << 6), pcmk__wo_remove_after = (1 << 7), pcmk__wo_ping_node = (1 << 8), pcmk__wo_order_inst = (1 << 9), pcmk__wo_coloc_inst = (1 << 10), pcmk__wo_group_order = (1 << 11), pcmk__wo_group_coloc = (1 << 12), pcmk__wo_upstart = (1 << 13), pcmk__wo_nagios = (1 << 14), pcmk__wo_set_ordering = (1 << 15), }; enum pcmk__check_parameters { /* Clear fail count if parameters changed for un-expired start or monitor * last_failure. */ pcmk__check_last_failure, /* Clear fail count if parameters changed for start, monitor, promote, or * migrate_from actions for active resources. */ pcmk__check_active, }; // Group of enum pcmk__sched_warnings flags for warnings we want to log once extern uint32_t pcmk__warnings; /*! * \internal * \brief Log a resource-tagged message at info severity * * \param[in] rsc Tag message with this resource's ID * \param[in] fmt... printf(3)-style format and arguments */ #define pcmk__rsc_info(rsc, fmt, args...) \ crm_log_tag(LOG_INFO, ((rsc) == NULL)? "" : (rsc)->id, (fmt), ##args) +/*! + * \internal + * \brief Log a resource-tagged message at debug severity + * + * \param[in] rsc Tag message with this resource's ID + * \param[in] fmt... printf(3)-style format and arguments + */ +#define pcmk__rsc_debug(rsc, fmt, args...) \ + crm_log_tag(LOG_DEBUG, ((rsc) == NULL)? "" : (rsc)->id, (fmt), ##args) + #ifdef __cplusplus } #endif #endif // PCMK__CRM_COMMON_SCHEDULER_INTERNAL__H diff --git a/include/crm/pengine/internal.h b/include/crm/pengine/internal.h index 1df74dae7a..9592293fd6 100644 --- a/include/crm/pengine/internal.h +++ b/include/crm/pengine/internal.h @@ -1,655 +1,654 @@ /* * Copyright 2004-2023 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU Lesser General Public License * version 2.1 or later (LGPLv2.1+) WITHOUT ANY WARRANTY. */ #ifndef PE_INTERNAL__H # define PE_INTERNAL__H # include # include # include # include # include # include # include # include # include # include const char *pe__resource_description(const pcmk_resource_t *rsc, uint32_t show_opts); bool pe__clone_is_ordered(const pcmk_resource_t *clone); int pe__set_clone_flag(pcmk_resource_t *clone, enum pcmk__clone_flags flag); bool pe__clone_flag_is_set(const pcmk_resource_t *clone, uint32_t flags); bool pe__group_flag_is_set(const pcmk_resource_t *group, uint32_t flags); pcmk_resource_t *pe__last_group_member(const pcmk_resource_t *group); -# define pe_rsc_debug(rsc, fmt, args...) crm_log_tag(LOG_DEBUG, rsc ? rsc->id : "", fmt, ##args) # define pe_rsc_trace(rsc, fmt, args...) crm_log_tag(LOG_TRACE, rsc ? rsc->id : "", fmt, ##args) # define pe_err(fmt...) do { \ was_processing_error = TRUE; \ pcmk__config_err(fmt); \ } while (0) # define pe_warn(fmt...) do { \ was_processing_warning = TRUE; \ pcmk__config_warn(fmt); \ } while (0) # define pe_proc_err(fmt...) { was_processing_error = TRUE; crm_err(fmt); } # define pe_proc_warn(fmt...) { was_processing_warning = TRUE; crm_warn(fmt); } #define pe__set_working_set_flags(scheduler, flags_to_set) do { \ (scheduler)->flags = pcmk__set_flags_as(__func__, __LINE__, \ LOG_TRACE, "Scheduler", crm_system_name, \ (scheduler)->flags, (flags_to_set), #flags_to_set); \ } while (0) #define pe__clear_working_set_flags(scheduler, flags_to_clear) do { \ (scheduler)->flags = pcmk__clear_flags_as(__func__, __LINE__, \ LOG_TRACE, "Scheduler", crm_system_name, \ (scheduler)->flags, (flags_to_clear), #flags_to_clear); \ } while (0) #define pe__set_resource_flags(resource, flags_to_set) do { \ (resource)->flags = pcmk__set_flags_as(__func__, __LINE__, \ LOG_TRACE, "Resource", (resource)->id, (resource)->flags, \ (flags_to_set), #flags_to_set); \ } while (0) #define pe__clear_resource_flags(resource, flags_to_clear) do { \ (resource)->flags = pcmk__clear_flags_as(__func__, __LINE__, \ LOG_TRACE, "Resource", (resource)->id, (resource)->flags, \ (flags_to_clear), #flags_to_clear); \ } while (0) #define pe__set_action_flags(action, flags_to_set) do { \ (action)->flags = pcmk__set_flags_as(__func__, __LINE__, \ LOG_TRACE, \ "Action", (action)->uuid, \ (action)->flags, \ (flags_to_set), \ #flags_to_set); \ } while (0) #define pe__clear_action_flags(action, flags_to_clear) do { \ (action)->flags = pcmk__clear_flags_as(__func__, __LINE__, \ LOG_TRACE, \ "Action", (action)->uuid, \ (action)->flags, \ (flags_to_clear), \ #flags_to_clear); \ } while (0) #define pe__set_raw_action_flags(action_flags, action_name, flags_to_set) do { \ action_flags = pcmk__set_flags_as(__func__, __LINE__, \ LOG_TRACE, "Action", action_name, \ (action_flags), \ (flags_to_set), #flags_to_set); \ } while (0) #define pe__clear_raw_action_flags(action_flags, action_name, flags_to_clear) do { \ action_flags = pcmk__clear_flags_as(__func__, __LINE__, \ LOG_TRACE, \ "Action", action_name, \ (action_flags), \ (flags_to_clear), \ #flags_to_clear); \ } while (0) #define pe__set_action_flags_as(function, line, action, flags_to_set) do { \ (action)->flags = pcmk__set_flags_as((function), (line), \ LOG_TRACE, \ "Action", (action)->uuid, \ (action)->flags, \ (flags_to_set), \ #flags_to_set); \ } while (0) #define pe__clear_action_flags_as(function, line, action, flags_to_clear) do { \ (action)->flags = pcmk__clear_flags_as((function), (line), \ LOG_TRACE, \ "Action", (action)->uuid, \ (action)->flags, \ (flags_to_clear), \ #flags_to_clear); \ } while (0) #define pe__set_order_flags(order_flags, flags_to_set) do { \ order_flags = pcmk__set_flags_as(__func__, __LINE__, LOG_TRACE, \ "Ordering", "constraint", \ order_flags, (flags_to_set), \ #flags_to_set); \ } while (0) #define pe__clear_order_flags(order_flags, flags_to_clear) do { \ order_flags = pcmk__clear_flags_as(__func__, __LINE__, LOG_TRACE, \ "Ordering", "constraint", \ order_flags, (flags_to_clear), \ #flags_to_clear); \ } while (0) #define pe_warn_once(pe_wo_bit, fmt...) do { \ if (!pcmk_is_set(pcmk__warnings, pe_wo_bit)) { \ if (pe_wo_bit == pcmk__wo_blind) { \ crm_warn(fmt); \ } else { \ pe_warn(fmt); \ } \ pcmk__warnings = pcmk__set_flags_as(__func__, __LINE__, \ LOG_TRACE, \ "Warn-once", "logging", \ pcmk__warnings, \ (pe_wo_bit), #pe_wo_bit); \ } \ } while (0); const pcmk_resource_t *pe__const_top_resource(const pcmk_resource_t *rsc, bool include_bundle); int pe__clone_max(const pcmk_resource_t *clone); int pe__clone_node_max(const pcmk_resource_t *clone); int pe__clone_promoted_max(const pcmk_resource_t *clone); int pe__clone_promoted_node_max(const pcmk_resource_t *clone); void pe__create_clone_notifications(pcmk_resource_t *clone); void pe__free_clone_notification_data(pcmk_resource_t *clone); void pe__create_clone_notif_pseudo_ops(pcmk_resource_t *clone, pcmk_action_t *start, pcmk_action_t *started, pcmk_action_t *stop, pcmk_action_t *stopped); pcmk_action_t *pe__new_rsc_pseudo_action(pcmk_resource_t *rsc, const char *task, bool optional, bool runnable); void pe__create_promotable_pseudo_ops(pcmk_resource_t *clone, bool any_promoting, bool any_demoting); bool pe_can_fence(const pcmk_scheduler_t *scheduler, const pcmk_node_t *node); void add_hash_param(GHashTable * hash, const char *name, const char *value); char *native_parameter(pcmk_resource_t *rsc, pcmk_node_t *node, gboolean create, const char *name, pcmk_scheduler_t *scheduler); pcmk_node_t *native_location(const pcmk_resource_t *rsc, GList **list, int current); void pe_metadata(pcmk__output_t *out); void verify_pe_options(GHashTable * options); void native_add_running(pcmk_resource_t *rsc, pcmk_node_t *node, pcmk_scheduler_t *scheduler, gboolean failed); gboolean native_unpack(pcmk_resource_t *rsc, pcmk_scheduler_t *scheduler); gboolean group_unpack(pcmk_resource_t *rsc, pcmk_scheduler_t *scheduler); gboolean clone_unpack(pcmk_resource_t *rsc, pcmk_scheduler_t *scheduler); gboolean pe__unpack_bundle(pcmk_resource_t *rsc, pcmk_scheduler_t *scheduler); pcmk_resource_t *native_find_rsc(pcmk_resource_t *rsc, const char *id, const pcmk_node_t *node, int flags); gboolean native_active(pcmk_resource_t *rsc, gboolean all); gboolean group_active(pcmk_resource_t *rsc, gboolean all); gboolean clone_active(pcmk_resource_t *rsc, gboolean all); gboolean pe__bundle_active(pcmk_resource_t *rsc, gboolean all); //! \deprecated This function will be removed in a future release void native_print(pcmk_resource_t *rsc, const char *pre_text, long options, void *print_data); //! \deprecated This function will be removed in a future release void group_print(pcmk_resource_t *rsc, const char *pre_text, long options, void *print_data); //! \deprecated This function will be removed in a future release void clone_print(pcmk_resource_t *rsc, const char *pre_text, long options, void *print_data); //! \deprecated This function will be removed in a future release void pe__print_bundle(pcmk_resource_t *rsc, const char *pre_text, long options, void *print_data); gchar *pcmk__native_output_string(const pcmk_resource_t *rsc, const char *name, const pcmk_node_t *node, uint32_t show_opts, const char *target_role, bool show_nodes); int pe__name_and_nvpairs_xml(pcmk__output_t *out, bool is_list, const char *tag_name , size_t pairs_count, ...); char *pe__node_display_name(pcmk_node_t *node, bool print_detail); // Clone notifications (pe_notif.c) void pe__order_notifs_after_fencing(const pcmk_action_t *action, pcmk_resource_t *rsc, pcmk_action_t *stonith_op); static inline const char * pe__rsc_bool_str(const pcmk_resource_t *rsc, uint64_t rsc_flag) { return pcmk__btoa(pcmk_is_set(rsc->flags, rsc_flag)); } int pe__clone_xml(pcmk__output_t *out, va_list args); int pe__clone_default(pcmk__output_t *out, va_list args); int pe__group_xml(pcmk__output_t *out, va_list args); int pe__group_default(pcmk__output_t *out, va_list args); int pe__bundle_xml(pcmk__output_t *out, va_list args); int pe__bundle_html(pcmk__output_t *out, va_list args); int pe__bundle_text(pcmk__output_t *out, va_list args); int pe__node_html(pcmk__output_t *out, va_list args); int pe__node_text(pcmk__output_t *out, va_list args); int pe__node_xml(pcmk__output_t *out, va_list args); int pe__resource_xml(pcmk__output_t *out, va_list args); int pe__resource_html(pcmk__output_t *out, va_list args); int pe__resource_text(pcmk__output_t *out, va_list args); void native_free(pcmk_resource_t *rsc); void group_free(pcmk_resource_t *rsc); void clone_free(pcmk_resource_t *rsc); void pe__free_bundle(pcmk_resource_t *rsc); enum rsc_role_e native_resource_state(const pcmk_resource_t *rsc, gboolean current); enum rsc_role_e group_resource_state(const pcmk_resource_t *rsc, gboolean current); enum rsc_role_e clone_resource_state(const pcmk_resource_t *rsc, gboolean current); enum rsc_role_e pe__bundle_resource_state(const pcmk_resource_t *rsc, gboolean current); void pe__count_common(pcmk_resource_t *rsc); void pe__count_bundle(pcmk_resource_t *rsc); void common_free(pcmk_resource_t *rsc); pcmk_node_t *pe__copy_node(const pcmk_node_t *this_node); time_t get_effective_time(pcmk_scheduler_t *scheduler); /* Failure handling utilities (from failcounts.c) */ int pe_get_failcount(const pcmk_node_t *node, pcmk_resource_t *rsc, time_t *last_failure, uint32_t flags, const xmlNode *xml_op); pcmk_action_t *pe__clear_failcount(pcmk_resource_t *rsc, const pcmk_node_t *node, const char *reason, pcmk_scheduler_t *scheduler); /* Functions for finding/counting a resource's active nodes */ bool pe__count_active_node(const pcmk_resource_t *rsc, pcmk_node_t *node, pcmk_node_t **active, unsigned int *count_all, unsigned int *count_clean); pcmk_node_t *pe__find_active_requires(const pcmk_resource_t *rsc, unsigned int *count); static inline pcmk_node_t * pe__current_node(const pcmk_resource_t *rsc) { return (rsc == NULL)? NULL : rsc->fns->active_node(rsc, NULL, NULL); } /* Binary like operators for lists of nodes */ GHashTable *pe__node_list2table(const GList *list); pcmk_action_t *get_pseudo_op(const char *name, pcmk_scheduler_t *scheduler); gboolean order_actions(pcmk_action_t *lh_action, pcmk_action_t *rh_action, uint32_t flags); void pe__show_node_scores_as(const char *file, const char *function, int line, bool to_log, const pcmk_resource_t *rsc, const char *comment, GHashTable *nodes, pcmk_scheduler_t *scheduler); #define pe__show_node_scores(level, rsc, text, nodes, scheduler) \ pe__show_node_scores_as(__FILE__, __func__, __LINE__, \ (level), (rsc), (text), (nodes), (scheduler)) GHashTable *pcmk__unpack_action_meta(pcmk_resource_t *rsc, const pcmk_node_t *node, const char *action_name, guint interval_ms, const xmlNode *action_config); GHashTable *pcmk__unpack_action_rsc_params(const xmlNode *action_xml, GHashTable *node_attrs, pcmk_scheduler_t *data_set); xmlNode *pcmk__find_action_config(const pcmk_resource_t *rsc, const char *action_name, guint interval_ms, bool include_disabled); enum rsc_start_requirement pcmk__action_requires(const pcmk_resource_t *rsc, const char *action_name); enum action_fail_response pcmk__parse_on_fail(const pcmk_resource_t *rsc, const char *action_name, guint interval_ms, const char *value); enum rsc_role_e pcmk__role_after_failure(const pcmk_resource_t *rsc, const char *action_name, enum action_fail_response on_fail, GHashTable *meta); pcmk_action_t *custom_action(pcmk_resource_t *rsc, char *key, const char *task, const pcmk_node_t *on_node, gboolean optional, pcmk_scheduler_t *scheduler); # define delete_key(rsc) pcmk__op_key(rsc->id, PCMK_ACTION_DELETE, 0) # define delete_action(rsc, node, optional) custom_action( \ rsc, delete_key(rsc), PCMK_ACTION_DELETE, node, \ optional, rsc->cluster); # define stop_key(rsc) pcmk__op_key(rsc->id, PCMK_ACTION_STOP, 0) # define stop_action(rsc, node, optional) custom_action( \ rsc, stop_key(rsc), PCMK_ACTION_STOP, node, \ optional, rsc->cluster); # define reload_key(rsc) pcmk__op_key(rsc->id, PCMK_ACTION_RELOAD_AGENT, 0) # define start_key(rsc) pcmk__op_key(rsc->id, PCMK_ACTION_START, 0) # define start_action(rsc, node, optional) custom_action( \ rsc, start_key(rsc), PCMK_ACTION_START, node, \ optional, rsc->cluster) # define promote_key(rsc) pcmk__op_key(rsc->id, PCMK_ACTION_PROMOTE, 0) # define promote_action(rsc, node, optional) custom_action( \ rsc, promote_key(rsc), PCMK_ACTION_PROMOTE, node, \ optional, rsc->cluster) # define demote_key(rsc) pcmk__op_key(rsc->id, PCMK_ACTION_DEMOTE, 0) # define demote_action(rsc, node, optional) custom_action( \ rsc, demote_key(rsc), PCMK_ACTION_DEMOTE, node, \ optional, rsc->cluster) extern int pe_get_configured_timeout(pcmk_resource_t *rsc, const char *action, pcmk_scheduler_t *scheduler); pcmk_action_t *find_first_action(const GList *input, const char *uuid, const char *task, const pcmk_node_t *on_node); enum action_tasks get_complex_task(const pcmk_resource_t *rsc, const char *name); GList *find_actions(GList *input, const char *key, const pcmk_node_t *on_node); GList *find_actions_exact(GList *input, const char *key, const pcmk_node_t *on_node); GList *pe__resource_actions(const pcmk_resource_t *rsc, const pcmk_node_t *node, const char *task, bool require_node); extern void pe_free_action(pcmk_action_t *action); void resource_location(pcmk_resource_t *rsc, const pcmk_node_t *node, int score, const char *tag, pcmk_scheduler_t *scheduler); extern int pe__is_newer_op(const xmlNode *xml_a, const xmlNode *xml_b, bool same_node_default); extern gint sort_op_by_callid(gconstpointer a, gconstpointer b); gboolean get_target_role(const pcmk_resource_t *rsc, enum rsc_role_e *role); void pe__set_next_role(pcmk_resource_t *rsc, enum rsc_role_e role, const char *why); pcmk_resource_t *find_clone_instance(const pcmk_resource_t *rsc, const char *sub_id); extern void destroy_ticket(gpointer data); pcmk_ticket_t *ticket_new(const char *ticket_id, pcmk_scheduler_t *scheduler); // Resources for manipulating resource names const char *pe_base_name_end(const char *id); char *clone_strip(const char *last_rsc_id); char *clone_zero(const char *last_rsc_id); static inline bool pe_base_name_eq(const pcmk_resource_t *rsc, const char *id) { if (id && rsc && rsc->id) { // Number of characters in rsc->id before any clone suffix size_t base_len = pe_base_name_end(rsc->id) - rsc->id + 1; return (strlen(id) == base_len) && !strncmp(id, rsc->id, base_len); } return false; } int pe__target_rc_from_xml(const xmlNode *xml_op); gint pe__cmp_node_name(gconstpointer a, gconstpointer b); bool is_set_recursive(const pcmk_resource_t *rsc, long long flag, bool any); pcmk__op_digest_t *pe__calculate_digests(pcmk_resource_t *rsc, const char *task, guint *interval_ms, const pcmk_node_t *node, const xmlNode *xml_op, GHashTable *overrides, bool calc_secure, pcmk_scheduler_t *scheduler); void pe__free_digests(gpointer ptr); pcmk__op_digest_t *rsc_action_digest_cmp(pcmk_resource_t *rsc, const xmlNode *xml_op, pcmk_node_t *node, pcmk_scheduler_t *scheduler); pcmk_action_t *pe_fence_op(pcmk_node_t *node, const char *op, bool optional, const char *reason, bool priority_delay, pcmk_scheduler_t *scheduler); void trigger_unfencing(pcmk_resource_t *rsc, pcmk_node_t *node, const char *reason, pcmk_action_t *dependency, pcmk_scheduler_t *scheduler); char *pe__action2reason(const pcmk_action_t *action, enum pe_action_flags flag); void pe_action_set_reason(pcmk_action_t *action, const char *reason, bool overwrite); void pe__add_action_expected_result(pcmk_action_t *action, int expected_result); void pe__set_resource_flags_recursive(pcmk_resource_t *rsc, uint64_t flags); void pe__clear_resource_flags_recursive(pcmk_resource_t *rsc, uint64_t flags); void pe__clear_resource_flags_on_all(pcmk_scheduler_t *scheduler, uint64_t flag); gboolean add_tag_ref(GHashTable * tags, const char * tag_name, const char * obj_ref); //! \deprecated This function will be removed in a future release void print_rscs_brief(GList *rsc_list, const char * pre_text, long options, void * print_data, gboolean print_all); int pe__rscs_brief_output(pcmk__output_t *out, GList *rsc_list, unsigned int options); void pe_fence_node(pcmk_scheduler_t *scheduler, pcmk_node_t *node, const char *reason, bool priority_delay); pcmk_node_t *pe_create_node(const char *id, const char *uname, const char *type, const char *score, pcmk_scheduler_t *scheduler); //! \deprecated This function will be removed in a future release void common_print(pcmk_resource_t *rsc, const char *pre_text, const char *name, const pcmk_node_t *node, long options, void *print_data); int pe__common_output_text(pcmk__output_t *out, const pcmk_resource_t *rsc, const char *name, const pcmk_node_t *node, unsigned int options); int pe__common_output_html(pcmk__output_t *out, const pcmk_resource_t *rsc, const char *name, const pcmk_node_t *node, unsigned int options); GList *pe__bundle_containers(const pcmk_resource_t *bundle); int pe__bundle_max(const pcmk_resource_t *rsc); bool pe__node_is_bundle_instance(const pcmk_resource_t *bundle, const pcmk_node_t *node); pcmk_resource_t *pe__bundled_resource(const pcmk_resource_t *rsc); const pcmk_resource_t *pe__get_rsc_in_container(const pcmk_resource_t *instance); pcmk_resource_t *pe__first_container(const pcmk_resource_t *bundle); void pe__foreach_bundle_replica(pcmk_resource_t *bundle, bool (*fn)(pcmk__bundle_replica_t *, void *), void *user_data); void pe__foreach_const_bundle_replica(const pcmk_resource_t *bundle, bool (*fn)(const pcmk__bundle_replica_t *, void *), void *user_data); pcmk_resource_t *pe__find_bundle_replica(const pcmk_resource_t *bundle, const pcmk_node_t *node); bool pe__bundle_needs_remote_name(pcmk_resource_t *rsc); const char *pe__add_bundle_remote_name(pcmk_resource_t *rsc, pcmk_scheduler_t *scheduler, xmlNode *xml, const char *field); const char *pe__node_attribute_calculated(const pcmk_node_t *node, const char *name, const pcmk_resource_t *rsc, enum pcmk__rsc_node node_type, bool force_host); const char *pe_node_attribute_raw(const pcmk_node_t *node, const char *name); bool pe__is_universal_clone(const pcmk_resource_t *rsc, const pcmk_scheduler_t *scheduler); void pe__add_param_check(const xmlNode *rsc_op, pcmk_resource_t *rsc, pcmk_node_t *node, enum pcmk__check_parameters, pcmk_scheduler_t *scheduler); void pe__foreach_param_check(pcmk_scheduler_t *scheduler, void (*cb)(pcmk_resource_t*, pcmk_node_t*, const xmlNode*, enum pcmk__check_parameters)); void pe__free_param_checks(pcmk_scheduler_t *scheduler); bool pe__shutdown_requested(const pcmk_node_t *node); void pe__update_recheck_time(time_t recheck, pcmk_scheduler_t *scheduler, const char *reason); /*! * \internal * \brief Register xml formatting message functions. * * \param[in,out] out Output object to register messages with */ void pe__register_messages(pcmk__output_t *out); void pe__unpack_dataset_nvpairs(const xmlNode *xml_obj, const char *set_name, const pe_rule_eval_data_t *rule_data, GHashTable *hash, const char *always_first, gboolean overwrite, pcmk_scheduler_t *scheduler); bool pe__resource_is_disabled(const pcmk_resource_t *rsc); void pe__clear_resource_history(pcmk_resource_t *rsc, const pcmk_node_t *node); GList *pe__rscs_with_tag(pcmk_scheduler_t *scheduler, const char *tag_name); GList *pe__unames_with_tag(pcmk_scheduler_t *scheduler, const char *tag_name); bool pe__rsc_has_tag(pcmk_scheduler_t *scheduler, const char *rsc, const char *tag); bool pe__uname_has_tag(pcmk_scheduler_t *scheduler, const char *node, const char *tag); bool pe__rsc_running_on_only(const pcmk_resource_t *rsc, const pcmk_node_t *node); bool pe__rsc_running_on_any(pcmk_resource_t *rsc, GList *node_list); GList *pe__filter_rsc_list(GList *rscs, GList *filter); GList * pe__build_node_name_list(pcmk_scheduler_t *scheduler, const char *s); GList * pe__build_rsc_list(pcmk_scheduler_t *scheduler, const char *s); bool pcmk__rsc_filtered_by_node(pcmk_resource_t *rsc, GList *only_node); gboolean pe__bundle_is_filtered(const pcmk_resource_t *rsc, GList *only_rsc, gboolean check_parent); gboolean pe__clone_is_filtered(const pcmk_resource_t *rsc, GList *only_rsc, gboolean check_parent); gboolean pe__group_is_filtered(const pcmk_resource_t *rsc, GList *only_rsc, gboolean check_parent); gboolean pe__native_is_filtered(const pcmk_resource_t *rsc, GList *only_rsc, gboolean check_parent); xmlNode *pe__failed_probe_for_rsc(const pcmk_resource_t *rsc, const char *name); const char *pe__clone_child_id(const pcmk_resource_t *rsc); int pe__sum_node_health_scores(const pcmk_node_t *node, int base_health); int pe__node_health(pcmk_node_t *node); static inline enum pcmk__health_strategy pe__health_strategy(pcmk_scheduler_t *scheduler) { return pcmk__parse_health_strategy(pe_pref(scheduler->config_hash, PCMK__OPT_NODE_HEALTH_STRATEGY)); } static inline int pe__health_score(const char *option, pcmk_scheduler_t *scheduler) { return char2score(pe_pref(scheduler->config_hash, option)); } /*! * \internal * \brief Return a string suitable for logging as a node name * * \param[in] node Node to return a node name string for * * \return Node name if available, otherwise node ID if available, * otherwise "unspecified node" if node is NULL or "unidentified node" * if node has neither a name nor ID. */ static inline const char * pe__node_name(const pcmk_node_t *node) { if (node == NULL) { return "unspecified node"; } else if (node->details->uname != NULL) { return node->details->uname; } else if (node->details->id != NULL) { return node->details->id; } else { return "unidentified node"; } } /*! * \internal * \brief Check whether two node objects refer to the same node * * \param[in] node1 First node object to compare * \param[in] node2 Second node object to compare * * \return true if \p node1 and \p node2 refer to the same node */ static inline bool pe__same_node(const pcmk_node_t *node1, const pcmk_node_t *node2) { return (node1 != NULL) && (node2 != NULL) && (node1->details == node2->details); } /*! * \internal * \brief Get the operation key from an action history entry * * \param[in] xml Action history entry * * \return Entry's operation key */ static inline const char * pe__xe_history_key(const xmlNode *xml) { if (xml == NULL) { return NULL; } else { /* @COMPAT Pacemaker <= 1.1.5 did not add the key, and used the ID * instead. Checking for that allows us to process old saved CIBs, * including some regression tests. */ const char *key = crm_element_value(xml, XML_LRM_ATTR_TASK_KEY); return pcmk__str_empty(key)? ID(xml) : key; } } #endif diff --git a/lib/pacemaker/pcmk_sched_actions.c b/lib/pacemaker/pcmk_sched_actions.c index cf58baecb1..7196ce4dff 100644 --- a/lib/pacemaker/pcmk_sched_actions.c +++ b/lib/pacemaker/pcmk_sched_actions.c @@ -1,1921 +1,1921 @@ /* * Copyright 2004-2023 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU General Public License version 2 * or later (GPLv2+) WITHOUT ANY WARRANTY. */ #include #include #include #include #include #include #include #include "libpacemaker_private.h" /*! * \internal * \brief Get the action flags relevant to ordering constraints * * \param[in,out] action Action to check * \param[in] node Node that *other* action in the ordering is on * (used only for clone resource actions) * * \return Action flags that should be used for orderings */ static uint32_t action_flags_for_ordering(pcmk_action_t *action, const pcmk_node_t *node) { bool runnable = false; uint32_t flags; // For non-resource actions, return the action flags if (action->rsc == NULL) { return action->flags; } /* For non-clone resources, or a clone action not assigned to a node, * return the flags as determined by the resource method without a node * specified. */ flags = action->rsc->cmds->action_flags(action, NULL); if ((node == NULL) || !pe_rsc_is_clone(action->rsc)) { return flags; } /* Otherwise (i.e., for clone resource actions on a specific node), first * remember whether the non-node-specific action is runnable. */ runnable = pcmk_is_set(flags, pcmk_action_runnable); // Then recheck the resource method with the node flags = action->rsc->cmds->action_flags(action, node); /* For clones in ordering constraints, the node-specific "runnable" doesn't * matter, just the non-node-specific setting (i.e., is the action runnable * anywhere). * * This applies only to runnable, and only for ordering constraints. This * function shouldn't be used for other types of constraints without * changes. Not very satisfying, but it's logical and appears to work well. */ if (runnable && !pcmk_is_set(flags, pcmk_action_runnable)) { pe__set_raw_action_flags(flags, action->rsc->id, pcmk_action_runnable); } return flags; } /*! * \internal * \brief Get action UUID that should be used with a resource ordering * * When an action is ordered relative to an action for a collective resource * (clone, group, or bundle), it actually needs to be ordered after all * instances of the collective have completed the relevant action (for example, * given "start CLONE then start RSC", RSC must wait until all instances of * CLONE have started). Given the UUID and resource of the first action in an * ordering, this returns the UUID of the action that should actually be used * for ordering (for example, "CLONE_started_0" instead of "CLONE_start_0"). * * \param[in] first_uuid UUID of first action in ordering * \param[in] first_rsc Resource of first action in ordering * * \return Newly allocated copy of UUID to use with ordering * \note It is the caller's responsibility to free the return value. */ static char * action_uuid_for_ordering(const char *first_uuid, const pcmk_resource_t *first_rsc) { guint interval_ms = 0; char *uuid = NULL; char *rid = NULL; char *first_task_str = NULL; enum action_tasks first_task = pcmk_action_unspecified; enum action_tasks remapped_task = pcmk_action_unspecified; // Only non-notify actions for collective resources need remapping if ((strstr(first_uuid, PCMK_ACTION_NOTIFY) != NULL) || (first_rsc->variant < pcmk_rsc_variant_group)) { goto done; } // Only non-recurring actions need remapping CRM_ASSERT(parse_op_key(first_uuid, &rid, &first_task_str, &interval_ms)); if (interval_ms > 0) { goto done; } first_task = text2task(first_task_str); switch (first_task) { case pcmk_action_stop: case pcmk_action_start: case pcmk_action_notify: case pcmk_action_promote: case pcmk_action_demote: remapped_task = first_task + 1; break; case pcmk_action_stopped: case pcmk_action_started: case pcmk_action_notified: case pcmk_action_promoted: case pcmk_action_demoted: remapped_task = first_task; break; case pcmk_action_monitor: case pcmk_action_shutdown: case pcmk_action_fence: break; default: crm_err("Unknown action '%s' in ordering", first_task_str); break; } if (remapped_task != pcmk_action_unspecified) { /* If a clone or bundle has notifications enabled, the ordering will be * relative to when notifications have been sent for the remapped task. */ if (pcmk_is_set(first_rsc->flags, pcmk_rsc_notify) && (pe_rsc_is_clone(first_rsc) || pe_rsc_is_bundled(first_rsc))) { uuid = pcmk__notify_key(rid, "confirmed-post", task2text(remapped_task)); } else { uuid = pcmk__op_key(rid, task2text(remapped_task), 0); } pe_rsc_trace(first_rsc, "Remapped action UUID %s to %s for ordering purposes", first_uuid, uuid); } done: if (uuid == NULL) { uuid = strdup(first_uuid); CRM_ASSERT(uuid != NULL); } free(first_task_str); free(rid); return uuid; } /*! * \internal * \brief Get actual action that should be used with an ordering * * When an action is ordered relative to an action for a collective resource * (clone, group, or bundle), it actually needs to be ordered after all * instances of the collective have completed the relevant action (for example, * given "start CLONE then start RSC", RSC must wait until all instances of * CLONE have started). Given the first action in an ordering, this returns the * the action that should actually be used for ordering (for example, the * started action instead of the start action). * * \param[in] action First action in an ordering * * \return Actual action that should be used for the ordering */ static pcmk_action_t * action_for_ordering(pcmk_action_t *action) { pcmk_action_t *result = action; pcmk_resource_t *rsc = action->rsc; if ((rsc != NULL) && (rsc->variant >= pcmk_rsc_variant_group) && (action->uuid != NULL)) { char *uuid = action_uuid_for_ordering(action->uuid, rsc); result = find_first_action(rsc->actions, uuid, NULL, NULL); if (result == NULL) { crm_warn("Not remapping %s to %s because %s does not have " "remapped action", action->uuid, uuid, rsc->id); result = action; } free(uuid); } return result; } /*! * \internal * \brief Wrapper for update_ordered_actions() method for readability * * \param[in,out] rsc Resource to call method for * \param[in,out] first 'First' action in an ordering * \param[in,out] then 'Then' action in an ordering * \param[in] node If not NULL, limit scope of ordering to this * node (only used when interleaving instances) * \param[in] flags Action flags for \p first for ordering purposes * \param[in] filter Action flags to limit scope of certain updates * (may include pcmk_action_optional to affect only * mandatory actions, and pe_action_runnable to * affect only runnable actions) * \param[in] type Group of enum pcmk__action_relation_flags to apply * \param[in,out] scheduler Scheduler data * * \return Group of enum pcmk__updated flags indicating what was updated */ static inline uint32_t update(pcmk_resource_t *rsc, pcmk_action_t *first, pcmk_action_t *then, const pcmk_node_t *node, uint32_t flags, uint32_t filter, uint32_t type, pcmk_scheduler_t *scheduler) { return rsc->cmds->update_ordered_actions(first, then, node, flags, filter, type, scheduler); } /*! * \internal * \brief Update flags for ordering's actions appropriately for ordering's flags * * \param[in,out] first First action in an ordering * \param[in,out] then Then action in an ordering * \param[in] first_flags Action flags for \p first for ordering purposes * \param[in] then_flags Action flags for \p then for ordering purposes * \param[in,out] order Action wrapper for \p first in ordering * \param[in,out] scheduler Scheduler data * * \return Group of enum pcmk__updated flags */ static uint32_t update_action_for_ordering_flags(pcmk_action_t *first, pcmk_action_t *then, uint32_t first_flags, uint32_t then_flags, pcmk__related_action_t *order, pcmk_scheduler_t *scheduler) { uint32_t changed = pcmk__updated_none; /* The node will only be used for clones. If interleaved, node will be NULL, * otherwise the ordering scope will be limited to the node. Normally, the * whole 'then' clone should restart if 'first' is restarted, so then->node * is needed. */ pcmk_node_t *node = then->node; if (pcmk_is_set(order->type, pcmk__ar_first_implies_same_node_then)) { /* For unfencing, only instances of 'then' on the same node as 'first' * (the unfencing operation) should restart, so reset node to * first->node, at which point this case is handled like a normal * pcmk__ar_first_implies_then. */ pe__clear_order_flags(order->type, pcmk__ar_first_implies_same_node_then); pe__set_order_flags(order->type, pcmk__ar_first_implies_then); node = first->node; pe_rsc_trace(then->rsc, "%s then %s: mapped pcmk__ar_first_implies_same_node_then " "to pcmk__ar_first_implies_then on %s", first->uuid, then->uuid, pe__node_name(node)); } if (pcmk_is_set(order->type, pcmk__ar_first_implies_then)) { if (then->rsc != NULL) { changed |= update(then->rsc, first, then, node, first_flags & pcmk_action_optional, pcmk_action_optional, pcmk__ar_first_implies_then, scheduler); } else if (!pcmk_is_set(first_flags, pcmk_action_optional) && pcmk_is_set(then->flags, pcmk_action_optional)) { pe__clear_action_flags(then, pcmk_action_optional); pcmk__set_updated_flags(changed, first, pcmk__updated_then); } pe_rsc_trace(then->rsc, "%s then %s: %s after pcmk__ar_first_implies_then", first->uuid, then->uuid, (changed? "changed" : "unchanged")); } if (pcmk_is_set(order->type, pcmk__ar_intermediate_stop) && (then->rsc != NULL)) { enum pe_action_flags restart = pcmk_action_optional |pcmk_action_runnable; changed |= update(then->rsc, first, then, node, first_flags, restart, pcmk__ar_intermediate_stop, scheduler); pe_rsc_trace(then->rsc, "%s then %s: %s after pcmk__ar_intermediate_stop", first->uuid, then->uuid, (changed? "changed" : "unchanged")); } if (pcmk_is_set(order->type, pcmk__ar_then_implies_first)) { if (first->rsc != NULL) { changed |= update(first->rsc, first, then, node, first_flags, pcmk_action_optional, pcmk__ar_then_implies_first, scheduler); } else if (!pcmk_is_set(first_flags, pcmk_action_optional) && pcmk_is_set(first->flags, pcmk_action_runnable)) { pe__clear_action_flags(first, pcmk_action_runnable); pcmk__set_updated_flags(changed, first, pcmk__updated_first); } pe_rsc_trace(then->rsc, "%s then %s: %s after pcmk__ar_then_implies_first", first->uuid, then->uuid, (changed? "changed" : "unchanged")); } if (pcmk_is_set(order->type, pcmk__ar_promoted_then_implies_first)) { if (then->rsc != NULL) { changed |= update(then->rsc, first, then, node, first_flags & pcmk_action_optional, pcmk_action_optional, pcmk__ar_promoted_then_implies_first, scheduler); } pe_rsc_trace(then->rsc, "%s then %s: %s after pcmk__ar_promoted_then_implies_first", first->uuid, then->uuid, (changed? "changed" : "unchanged")); } if (pcmk_is_set(order->type, pcmk__ar_min_runnable)) { if (then->rsc != NULL) { changed |= update(then->rsc, first, then, node, first_flags, pcmk_action_runnable, pcmk__ar_min_runnable, scheduler); } else if (pcmk_is_set(first_flags, pcmk_action_runnable)) { // We have another runnable instance of "first" then->runnable_before++; /* Mark "then" as runnable if it requires a certain number of * "before" instances to be runnable, and they now are. */ if ((then->runnable_before >= then->required_runnable_before) && !pcmk_is_set(then->flags, pcmk_action_runnable)) { pe__set_action_flags(then, pcmk_action_runnable); pcmk__set_updated_flags(changed, first, pcmk__updated_then); } } pe_rsc_trace(then->rsc, "%s then %s: %s after pcmk__ar_min_runnable", first->uuid, then->uuid, (changed? "changed" : "unchanged")); } if (pcmk_is_set(order->type, pcmk__ar_nested_remote_probe) && (then->rsc != NULL)) { if (!pcmk_is_set(first_flags, pcmk_action_runnable) && (first->rsc != NULL) && (first->rsc->running_on != NULL)) { pe_rsc_trace(then->rsc, "%s then %s: ignoring because first is stopping", first->uuid, then->uuid); order->type = (enum pe_ordering) pcmk__ar_none; } else { changed |= update(then->rsc, first, then, node, first_flags, pcmk_action_runnable, pcmk__ar_unrunnable_first_blocks, scheduler); } pe_rsc_trace(then->rsc, "%s then %s: %s after pcmk__ar_nested_remote_probe", first->uuid, then->uuid, (changed? "changed" : "unchanged")); } if (pcmk_is_set(order->type, pcmk__ar_unrunnable_first_blocks)) { if (then->rsc != NULL) { changed |= update(then->rsc, first, then, node, first_flags, pcmk_action_runnable, pcmk__ar_unrunnable_first_blocks, scheduler); } else if (!pcmk_is_set(first_flags, pcmk_action_runnable) && pcmk_is_set(then->flags, pcmk_action_runnable)) { pe__clear_action_flags(then, pcmk_action_runnable); pcmk__set_updated_flags(changed, first, pcmk__updated_then); } pe_rsc_trace(then->rsc, "%s then %s: %s after pcmk__ar_unrunnable_first_blocks", first->uuid, then->uuid, (changed? "changed" : "unchanged")); } if (pcmk_is_set(order->type, pcmk__ar_unmigratable_then_blocks)) { if (then->rsc != NULL) { changed |= update(then->rsc, first, then, node, first_flags, pcmk_action_optional, pcmk__ar_unmigratable_then_blocks, scheduler); } pe_rsc_trace(then->rsc, "%s then %s: %s after " "pcmk__ar_unmigratable_then_blocks", first->uuid, then->uuid, (changed? "changed" : "unchanged")); } if (pcmk_is_set(order->type, pcmk__ar_first_else_then)) { if (then->rsc != NULL) { changed |= update(then->rsc, first, then, node, first_flags, pcmk_action_optional, pcmk__ar_first_else_then, scheduler); } pe_rsc_trace(then->rsc, "%s then %s: %s after pcmk__ar_first_else_then", first->uuid, then->uuid, (changed? "changed" : "unchanged")); } if (pcmk_is_set(order->type, pcmk__ar_ordered)) { if (then->rsc != NULL) { changed |= update(then->rsc, first, then, node, first_flags, pcmk_action_runnable, pcmk__ar_ordered, scheduler); } pe_rsc_trace(then->rsc, "%s then %s: %s after pcmk__ar_ordered", first->uuid, then->uuid, (changed? "changed" : "unchanged")); } if (pcmk_is_set(order->type, pcmk__ar_asymmetric)) { if (then->rsc != NULL) { changed |= update(then->rsc, first, then, node, first_flags, pcmk_action_runnable, pcmk__ar_asymmetric, scheduler); } pe_rsc_trace(then->rsc, "%s then %s: %s after pcmk__ar_asymmetric", first->uuid, then->uuid, (changed? "changed" : "unchanged")); } if (pcmk_is_set(first->flags, pcmk_action_runnable) && pcmk_is_set(order->type, pcmk__ar_first_implies_then_graphed) && !pcmk_is_set(first_flags, pcmk_action_optional)) { pe_rsc_trace(then->rsc, "%s will be in graph because %s is required", then->uuid, first->uuid); pe__set_action_flags(then, pcmk_action_always_in_graph); // Don't bother marking 'then' as changed just for this } if (pcmk_is_set(order->type, pcmk__ar_then_implies_first_graphed) && !pcmk_is_set(then_flags, pcmk_action_optional)) { pe_rsc_trace(then->rsc, "%s will be in graph because %s is required", first->uuid, then->uuid); pe__set_action_flags(first, pcmk_action_always_in_graph); // Don't bother marking 'first' as changed just for this } if (pcmk_any_flags_set(order->type, pcmk__ar_first_implies_then |pcmk__ar_then_implies_first |pcmk__ar_intermediate_stop) && (first->rsc != NULL) && !pcmk_is_set(first->rsc->flags, pcmk_rsc_managed) && pcmk_is_set(first->rsc->flags, pcmk_rsc_blocked) && !pcmk_is_set(first->flags, pcmk_action_runnable) && pcmk__str_eq(first->task, PCMK_ACTION_STOP, pcmk__str_none)) { if (pcmk_is_set(then->flags, pcmk_action_runnable)) { pe__clear_action_flags(then, pcmk_action_runnable); pcmk__set_updated_flags(changed, first, pcmk__updated_then); } pe_rsc_trace(then->rsc, "%s then %s: %s after checking whether first " "is blocked, unmanaged, unrunnable stop", first->uuid, then->uuid, (changed? "changed" : "unchanged")); } return changed; } // Convenience macros for logging action properties #define action_type_str(flags) \ (pcmk_is_set((flags), pcmk_action_pseudo)? "pseudo-action" : "action") #define action_optional_str(flags) \ (pcmk_is_set((flags), pcmk_action_optional)? "optional" : "required") #define action_runnable_str(flags) \ (pcmk_is_set((flags), pcmk_action_runnable)? "runnable" : "unrunnable") #define action_node_str(a) \ (((a)->node == NULL)? "no node" : (a)->node->details->uname) /*! * \internal * \brief Update an action's flags for all orderings where it is "then" * * \param[in,out] then Action to update * \param[in,out] scheduler Scheduler data */ void pcmk__update_action_for_orderings(pcmk_action_t *then, pcmk_scheduler_t *scheduler) { GList *lpc = NULL; uint32_t changed = pcmk__updated_none; int last_flags = then->flags; pe_rsc_trace(then->rsc, "Updating %s %s (%s %s) on %s", action_type_str(then->flags), then->uuid, action_optional_str(then->flags), action_runnable_str(then->flags), action_node_str(then)); if (pcmk_is_set(then->flags, pcmk_action_min_runnable)) { /* Initialize current known "runnable before" actions. As * update_action_for_ordering_flags() is called for each of then's * before actions, this number will increment as runnable 'first' * actions are encountered. */ then->runnable_before = 0; if (then->required_runnable_before == 0) { /* @COMPAT This ordering constraint uses the deprecated * "require-all=false" attribute. Treat it like "clone-min=1". */ then->required_runnable_before = 1; } /* The pcmk__ar_min_runnable clause of * update_action_for_ordering_flags() (called below) * will reset runnable if appropriate. */ pe__clear_action_flags(then, pcmk_action_runnable); } for (lpc = then->actions_before; lpc != NULL; lpc = lpc->next) { pcmk__related_action_t *other = lpc->data; pcmk_action_t *first = other->action; pcmk_node_t *then_node = then->node; pcmk_node_t *first_node = first->node; if ((first->rsc != NULL) && (first->rsc->variant == pcmk_rsc_variant_group) && pcmk__str_eq(first->task, PCMK_ACTION_START, pcmk__str_none)) { first_node = first->rsc->fns->location(first->rsc, NULL, FALSE); if (first_node != NULL) { pe_rsc_trace(first->rsc, "Found %s for 'first' %s", pe__node_name(first_node), first->uuid); } } if ((then->rsc != NULL) && (then->rsc->variant == pcmk_rsc_variant_group) && pcmk__str_eq(then->task, PCMK_ACTION_START, pcmk__str_none)) { then_node = then->rsc->fns->location(then->rsc, NULL, FALSE); if (then_node != NULL) { pe_rsc_trace(then->rsc, "Found %s for 'then' %s", pe__node_name(then_node), then->uuid); } } // Disable constraint if it only applies when on same node, but isn't if (pcmk_is_set(other->type, pcmk__ar_if_on_same_node) && (first_node != NULL) && (then_node != NULL) && !pe__same_node(first_node, then_node)) { pe_rsc_trace(then->rsc, "Disabled ordering %s on %s then %s on %s: " "not same node", other->action->uuid, pe__node_name(first_node), then->uuid, pe__node_name(then_node)); other->type = (enum pe_ordering) pcmk__ar_none; continue; } pcmk__clear_updated_flags(changed, then, pcmk__updated_first); if ((first->rsc != NULL) && pcmk_is_set(other->type, pcmk__ar_then_cancels_first) && !pcmk_is_set(then->flags, pcmk_action_optional)) { /* 'then' is required, so we must abandon 'first' * (e.g. a required stop cancels any agent reload). */ pe__set_action_flags(other->action, pcmk_action_optional); if (!strcmp(first->task, PCMK_ACTION_RELOAD_AGENT)) { pe__clear_resource_flags(first->rsc, pcmk_rsc_reload); } } if ((first->rsc != NULL) && (then->rsc != NULL) && (first->rsc != then->rsc) && !is_parent(then->rsc, first->rsc)) { first = action_for_ordering(first); } if (first != other->action) { pe_rsc_trace(then->rsc, "Ordering %s after %s instead of %s", then->uuid, first->uuid, other->action->uuid); } pe_rsc_trace(then->rsc, "%s (%#.6x) then %s (%#.6x): type=%#.6x node=%s", first->uuid, first->flags, then->uuid, then->flags, other->type, action_node_str(first)); if (first == other->action) { /* 'first' was not remapped (e.g. from 'start' to 'running'), which * could mean it is a non-resource action, a primitive resource * action, or already expanded. */ uint32_t first_flags, then_flags; first_flags = action_flags_for_ordering(first, then_node); then_flags = action_flags_for_ordering(then, first_node); changed |= update_action_for_ordering_flags(first, then, first_flags, then_flags, other, scheduler); /* 'first' was for a complex resource (clone, group, etc), * create a new dependency if necessary */ } else if (order_actions(first, then, other->type)) { /* This was the first time 'first' and 'then' were associated, * start again to get the new actions_before list */ pcmk__set_updated_flags(changed, then, pcmk__updated_then); pe_rsc_trace(then->rsc, "Disabled ordering %s then %s in favor of %s then %s", other->action->uuid, then->uuid, first->uuid, then->uuid); other->type = (enum pe_ordering) pcmk__ar_none; } if (pcmk_is_set(changed, pcmk__updated_first)) { crm_trace("Re-processing %s and its 'after' actions " "because it changed", first->uuid); for (GList *lpc2 = first->actions_after; lpc2 != NULL; lpc2 = lpc2->next) { pcmk__related_action_t *other = lpc2->data; pcmk__update_action_for_orderings(other->action, scheduler); } pcmk__update_action_for_orderings(first, scheduler); } } if (pcmk_is_set(then->flags, pcmk_action_min_runnable)) { if (last_flags == then->flags) { pcmk__clear_updated_flags(changed, then, pcmk__updated_then); } else { pcmk__set_updated_flags(changed, then, pcmk__updated_then); } } if (pcmk_is_set(changed, pcmk__updated_then)) { crm_trace("Re-processing %s and its 'after' actions because it changed", then->uuid); if (pcmk_is_set(last_flags, pcmk_action_runnable) && !pcmk_is_set(then->flags, pcmk_action_runnable)) { pcmk__block_colocation_dependents(then); } pcmk__update_action_for_orderings(then, scheduler); for (lpc = then->actions_after; lpc != NULL; lpc = lpc->next) { pcmk__related_action_t *other = lpc->data; pcmk__update_action_for_orderings(other->action, scheduler); } } } static inline bool is_primitive_action(const pcmk_action_t *action) { return (action != NULL) && (action->rsc != NULL) && (action->rsc->variant == pcmk_rsc_variant_primitive); } /*! * \internal * \brief Clear a single action flag and set reason text * * \param[in,out] action Action whose flag should be cleared * \param[in] flag Action flag that should be cleared * \param[in] reason Action that is the reason why flag is being cleared */ #define clear_action_flag_because(action, flag, reason) do { \ if (pcmk_is_set((action)->flags, (flag))) { \ pe__clear_action_flags(action, flag); \ if ((action)->rsc != (reason)->rsc) { \ char *reason_text = pe__action2reason((reason), (flag)); \ pe_action_set_reason((action), reason_text, false); \ free(reason_text); \ } \ } \ } while (0) /*! * \internal * \brief Update actions in an asymmetric ordering * * If the "first" action in an asymmetric ordering is unrunnable, make the * "second" action unrunnable as well, if appropriate. * * \param[in] first 'First' action in an asymmetric ordering * \param[in,out] then 'Then' action in an asymmetric ordering */ static void handle_asymmetric_ordering(const pcmk_action_t *first, pcmk_action_t *then) { /* Only resource actions after an unrunnable 'first' action need updates for * asymmetric ordering. */ if ((then->rsc == NULL) || pcmk_is_set(first->flags, pcmk_action_runnable)) { return; } // Certain optional 'then' actions are unaffected by unrunnable 'first' if (pcmk_is_set(then->flags, pcmk_action_optional)) { enum rsc_role_e then_rsc_role = then->rsc->fns->state(then->rsc, TRUE); if ((then_rsc_role == pcmk_role_stopped) && pcmk__str_eq(then->task, PCMK_ACTION_STOP, pcmk__str_none)) { /* If 'then' should stop after 'first' but is already stopped, the * ordering is irrelevant. */ return; } else if ((then_rsc_role >= pcmk_role_started) && pcmk__str_eq(then->task, PCMK_ACTION_START, pcmk__str_none) && pe__rsc_running_on_only(then->rsc, then->node)) { /* Similarly if 'then' should start after 'first' but is already * started on a single node. */ return; } } // 'First' can't run, so 'then' can't either clear_action_flag_because(then, pcmk_action_optional, first); clear_action_flag_because(then, pcmk_action_runnable, first); } /*! * \internal * \brief Set action bits appropriately when pe_restart_order is used * * \param[in,out] first 'First' action in an ordering with pe_restart_order * \param[in,out] then 'Then' action in an ordering with pe_restart_order * \param[in] filter What action 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(pcmk_action_t *first, pcmk_action_t *then, uint32_t 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 (pcmk_is_set(filter, pcmk_action_optional) && !pcmk_is_set(then->flags, pcmk_action_optional)) { reason = "restart"; } /* ... if 'then' is unrunnable action on same resource (if a resource * should restart but can't start, we still want to stop) */ if (pcmk_is_set(filter, pcmk_action_runnable) && !pcmk_is_set(then->flags, pcmk_action_runnable) && pcmk_is_set(then->rsc->flags, pcmk_rsc_managed) && (first->rsc == then->rsc)) { 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 (pcmk_is_set(first->flags, pcmk_action_runnable)) { clear_action_flag_because(first, pcmk_action_optional, then); } // Make 'first' required if 'then' is required if (!pcmk_is_set(then->flags, pcmk_action_optional)) { clear_action_flag_because(first, pcmk_action_optional, then); } // Make 'first' unmigratable if 'then' is unmigratable if (!pcmk_is_set(then->flags, pcmk_action_migratable)) { clear_action_flag_because(first, pcmk_action_migratable, then); } // Make 'then' unrunnable if 'first' is required but unrunnable if (!pcmk_is_set(first->flags, pcmk_action_optional) && !pcmk_is_set(first->flags, pcmk_action_runnable)) { clear_action_flag_because(then, pcmk_action_runnable, first); } } /*! * \internal * \brief Update two actions according to an ordering between them * * Given information about an ordering of two actions, update the actions' flags * (and runnable_before members if appropriate) as appropriate for the ordering. * Effects may cascade to other orderings involving the actions as well. * * \param[in,out] first 'First' action in an ordering * \param[in,out] then 'Then' action in an ordering * \param[in] node If not NULL, limit scope of ordering to this node * (ignored) * \param[in] flags Action flags for \p first for ordering purposes * \param[in] filter Action flags to limit scope of certain updates (may * include pcmk_action_optional to affect only * mandatory actions, and pcmk_action_runnable to * affect only runnable actions) * \param[in] type Group of enum pcmk__action_relation_flags to apply * \param[in,out] scheduler Scheduler data * * \return Group of enum pcmk__updated flags indicating what was updated */ uint32_t pcmk__update_ordered_actions(pcmk_action_t *first, pcmk_action_t *then, const pcmk_node_t *node, uint32_t flags, uint32_t filter, uint32_t type, pcmk_scheduler_t *scheduler) { uint32_t changed = pcmk__updated_none; uint32_t then_flags = 0U; uint32_t first_flags = 0U; CRM_ASSERT((first != NULL) && (then != NULL) && (scheduler != NULL)); then_flags = then->flags; first_flags = first->flags; if (pcmk_is_set(type, pcmk__ar_asymmetric)) { handle_asymmetric_ordering(first, then); } if (pcmk_is_set(type, pcmk__ar_then_implies_first) && !pcmk_is_set(then_flags, pcmk_action_optional)) { // Then is required, and implies first should be, too if (pcmk_is_set(filter, pcmk_action_optional) && !pcmk_is_set(flags, pcmk_action_optional) && pcmk_is_set(first_flags, pcmk_action_optional)) { clear_action_flag_because(first, pcmk_action_optional, then); } if (pcmk_is_set(flags, pcmk_action_migratable) && !pcmk_is_set(then->flags, pcmk_action_migratable)) { clear_action_flag_because(first, pcmk_action_migratable, then); } } if (pcmk_is_set(type, pcmk__ar_promoted_then_implies_first) && (then->rsc != NULL) && (then->rsc->role == pcmk_role_promoted) && pcmk_is_set(filter, pcmk_action_optional) && !pcmk_is_set(then->flags, pcmk_action_optional)) { clear_action_flag_because(first, pcmk_action_optional, then); if (pcmk_is_set(first->flags, pcmk_action_migratable) && !pcmk_is_set(then->flags, pcmk_action_migratable)) { clear_action_flag_because(first, pcmk_action_migratable, then); } } if (pcmk_is_set(type, pcmk__ar_unmigratable_then_blocks) && pcmk_is_set(filter, pcmk_action_optional)) { if (!pcmk_all_flags_set(then->flags, pcmk_action_migratable |pcmk_action_runnable)) { clear_action_flag_because(first, pcmk_action_runnable, then); } if (!pcmk_is_set(then->flags, pcmk_action_optional)) { clear_action_flag_because(first, pcmk_action_optional, then); } } if (pcmk_is_set(type, pcmk__ar_first_else_then) && pcmk_is_set(filter, pcmk_action_optional) && !pcmk_is_set(first->flags, pcmk_action_runnable)) { clear_action_flag_because(then, pcmk_action_migratable, first); pe__clear_action_flags(then, pcmk_action_pseudo); } if (pcmk_is_set(type, pcmk__ar_unrunnable_first_blocks) && pcmk_is_set(filter, pcmk_action_runnable) && pcmk_is_set(then->flags, pcmk_action_runnable) && !pcmk_is_set(flags, pcmk_action_runnable)) { clear_action_flag_because(then, pcmk_action_runnable, first); clear_action_flag_because(then, pcmk_action_migratable, first); } if (pcmk_is_set(type, pcmk__ar_first_implies_then) && pcmk_is_set(filter, pcmk_action_optional) && pcmk_is_set(then->flags, pcmk_action_optional) && !pcmk_is_set(flags, pcmk_action_optional) && !pcmk_is_set(first->flags, pcmk_action_migratable)) { clear_action_flag_because(then, pcmk_action_optional, first); } if (pcmk_is_set(type, pcmk__ar_intermediate_stop)) { handle_restart_ordering(first, then, filter); } if (then_flags != then->flags) { pcmk__set_updated_flags(changed, first, pcmk__updated_then); pe_rsc_trace(then->rsc, "%s on %s: flags are now %#.6x (was %#.6x) " "because of 'first' %s (%#.6x)", then->uuid, pe__node_name(then->node), then->flags, then_flags, first->uuid, first->flags); if ((then->rsc != NULL) && (then->rsc->parent != NULL)) { // Required to handle "X_stop then X_start" for cloned groups pcmk__update_action_for_orderings(then, scheduler); } } if (first_flags != first->flags) { pcmk__set_updated_flags(changed, first, pcmk__updated_first); pe_rsc_trace(first->rsc, "%s on %s: flags are now %#.6x (was %#.6x) " "because of 'then' %s (%#.6x)", first->uuid, pe__node_name(first->node), first->flags, first_flags, then->uuid, then->flags); } return changed; } /*! * \internal * \brief Trace-log an action (optionally with its dependent actions) * * \param[in] pre_text If not NULL, prefix the log with this plus ": " * \param[in] action Action to log * \param[in] details If true, recursively log dependent actions */ void pcmk__log_action(const char *pre_text, const pcmk_action_t *action, bool details) { const char *node_uname = NULL; const char *node_uuid = NULL; const char *desc = NULL; CRM_CHECK(action != NULL, return); if (!pcmk_is_set(action->flags, pcmk_action_pseudo)) { if (action->node != NULL) { node_uname = action->node->details->uname; node_uuid = action->node->details->id; } else { node_uname = ""; } } switch (text2task(action->task)) { case pcmk_action_fence: case pcmk_action_shutdown: if (pcmk_is_set(action->flags, pcmk_action_pseudo)) { desc = "Pseudo "; } else if (pcmk_is_set(action->flags, pcmk_action_optional)) { desc = "Optional "; } else if (!pcmk_is_set(action->flags, pcmk_action_runnable)) { desc = "!!Non-Startable!! "; } else { desc = "(Provisional) "; } crm_trace("%s%s%sAction %d: %s%s%s%s%s%s", ((pre_text == NULL)? "" : pre_text), ((pre_text == NULL)? "" : ": "), desc, action->id, action->uuid, (node_uname? "\ton " : ""), (node_uname? node_uname : ""), (node_uuid? "\t\t(" : ""), (node_uuid? node_uuid : ""), (node_uuid? ")" : "")); break; default: if (pcmk_is_set(action->flags, pcmk_action_optional)) { desc = "Optional "; } else if (pcmk_is_set(action->flags, pcmk_action_pseudo)) { desc = "Pseudo "; } else if (!pcmk_is_set(action->flags, pcmk_action_runnable)) { desc = "!!Non-Startable!! "; } else { desc = "(Provisional) "; } crm_trace("%s%s%sAction %d: %s %s%s%s%s%s%s", ((pre_text == NULL)? "" : pre_text), ((pre_text == NULL)? "" : ": "), desc, action->id, action->uuid, (action->rsc? action->rsc->id : ""), (node_uname? "\ton " : ""), (node_uname? node_uname : ""), (node_uuid? "\t\t(" : ""), (node_uuid? node_uuid : ""), (node_uuid? ")" : "")); break; } if (details) { const GList *iter = NULL; const pcmk__related_action_t *other = NULL; crm_trace("\t\t====== Preceding Actions"); for (iter = action->actions_before; iter != NULL; iter = iter->next) { other = (const pcmk__related_action_t *) iter->data; pcmk__log_action("\t\t", other->action, false); } crm_trace("\t\t====== Subsequent Actions"); for (iter = action->actions_after; iter != NULL; iter = iter->next) { other = (const pcmk__related_action_t *) iter->data; pcmk__log_action("\t\t", other->action, false); } crm_trace("\t\t====== End"); } else { crm_trace("\t\t(before=%d, after=%d)", g_list_length(action->actions_before), g_list_length(action->actions_after)); } } /*! * \internal * \brief Create a new shutdown action for a node * * \param[in,out] node Node being shut down * * \return Newly created shutdown action for \p node */ pcmk_action_t * pcmk__new_shutdown_action(pcmk_node_t *node) { char *shutdown_id = NULL; pcmk_action_t *shutdown_op = NULL; CRM_ASSERT(node != NULL); shutdown_id = crm_strdup_printf("%s-%s", PCMK_ACTION_DO_SHUTDOWN, node->details->uname); shutdown_op = custom_action(NULL, shutdown_id, PCMK_ACTION_DO_SHUTDOWN, node, FALSE, node->details->data_set); pcmk__order_stops_before_shutdown(node, shutdown_op); add_hash_param(shutdown_op->meta, XML_ATTR_TE_NOWAIT, XML_BOOLEAN_TRUE); return shutdown_op; } /*! * \internal * \brief Calculate and add an operation digest to XML * * Calculate an operation digest, which enables us to later determine when a * restart is needed due to the resource's parameters being changed, and add it * to given XML. * * \param[in] op Operation result from executor * \param[in,out] update XML to add digest to */ static void add_op_digest_to_xml(const lrmd_event_data_t *op, xmlNode *update) { char *digest = NULL; xmlNode *args_xml = NULL; if (op->params == NULL) { return; } args_xml = create_xml_node(NULL, XML_TAG_PARAMS); g_hash_table_foreach(op->params, hash2field, args_xml); pcmk__filter_op_for_digest(args_xml); digest = calculate_operation_digest(args_xml, NULL); crm_xml_add(update, XML_LRM_ATTR_OP_DIGEST, digest); free_xml(args_xml); free(digest); } #define FAKE_TE_ID "xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxx" /*! * \internal * \brief Create XML for resource operation history update * * \param[in,out] parent Parent XML node to add to * \param[in,out] op Operation event data * \param[in] caller_version DC feature set * \param[in] target_rc Expected result of operation * \param[in] node Name of node on which operation was performed * \param[in] origin Arbitrary description of update source * * \return Newly created XML node for history update */ xmlNode * pcmk__create_history_xml(xmlNode *parent, lrmd_event_data_t *op, const char *caller_version, int target_rc, const char *node, const char *origin) { char *key = NULL; char *magic = NULL; char *op_id = NULL; char *op_id_additional = NULL; char *local_user_data = NULL; const char *exit_reason = NULL; xmlNode *xml_op = NULL; const char *task = NULL; CRM_CHECK(op != NULL, return NULL); crm_trace("Creating history XML for %s-interval %s action for %s on %s " "(DC version: %s, origin: %s)", pcmk__readable_interval(op->interval_ms), op->op_type, op->rsc_id, ((node == NULL)? "no node" : node), caller_version, origin); task = op->op_type; /* Record a successful agent reload as a start, and a failed one as a * monitor, to make life easier for the scheduler when determining the * current state. * * @COMPAT We should check "reload" here only if the operation was for a * pre-OCF-1.1 resource agent, but we don't know that here, and we should * only ever get results for actions scheduled by us, so we can reasonably * assume any "reload" is actually a pre-1.1 agent reload. */ if (pcmk__str_any_of(task, PCMK_ACTION_RELOAD, PCMK_ACTION_RELOAD_AGENT, NULL)) { if (op->op_status == PCMK_EXEC_DONE) { task = PCMK_ACTION_START; } else { task = PCMK_ACTION_MONITOR; } } key = pcmk__op_key(op->rsc_id, task, op->interval_ms); if (pcmk__str_eq(task, PCMK_ACTION_NOTIFY, pcmk__str_none)) { const char *n_type = crm_meta_value(op->params, "notify_type"); const char *n_task = crm_meta_value(op->params, "notify_operation"); CRM_LOG_ASSERT(n_type != NULL); CRM_LOG_ASSERT(n_task != NULL); op_id = pcmk__notify_key(op->rsc_id, n_type, n_task); if (op->op_status != PCMK_EXEC_PENDING) { /* Ignore notify errors. * * @TODO It might be better to keep the correct result here, and * ignore it in process_graph_event(). */ lrmd__set_result(op, PCMK_OCF_OK, PCMK_EXEC_DONE, NULL); } /* Migration history is preserved separately, which usually matters for * multiple nodes and is important for future cluster transitions. */ } else if (pcmk__str_any_of(op->op_type, PCMK_ACTION_MIGRATE_TO, PCMK_ACTION_MIGRATE_FROM, NULL)) { op_id = strdup(key); } else if (did_rsc_op_fail(op, target_rc)) { op_id = pcmk__op_key(op->rsc_id, "last_failure", 0); if (op->interval_ms == 0) { // Ensure 'last' gets updated, in case record-pending is true op_id_additional = pcmk__op_key(op->rsc_id, "last", 0); } exit_reason = op->exit_reason; } else if (op->interval_ms > 0) { op_id = strdup(key); } else { op_id = pcmk__op_key(op->rsc_id, "last", 0); } again: xml_op = pcmk__xe_match(parent, XML_LRM_TAG_RSC_OP, XML_ATTR_ID, op_id); if (xml_op == NULL) { xml_op = create_xml_node(parent, XML_LRM_TAG_RSC_OP); } if (op->user_data == NULL) { crm_debug("Generating fake transition key for: " PCMK__OP_FMT " %d from %s", op->rsc_id, op->op_type, op->interval_ms, op->call_id, origin); local_user_data = pcmk__transition_key(-1, op->call_id, target_rc, FAKE_TE_ID); op->user_data = local_user_data; } if (magic == NULL) { magic = crm_strdup_printf("%d:%d;%s", op->op_status, op->rc, (const char *) op->user_data); } crm_xml_add(xml_op, XML_ATTR_ID, op_id); crm_xml_add(xml_op, XML_LRM_ATTR_TASK_KEY, key); crm_xml_add(xml_op, XML_LRM_ATTR_TASK, task); crm_xml_add(xml_op, XML_ATTR_ORIGIN, origin); crm_xml_add(xml_op, XML_ATTR_CRM_VERSION, caller_version); crm_xml_add(xml_op, XML_ATTR_TRANSITION_KEY, op->user_data); crm_xml_add(xml_op, XML_ATTR_TRANSITION_MAGIC, magic); crm_xml_add(xml_op, XML_LRM_ATTR_EXIT_REASON, pcmk__s(exit_reason, "")); crm_xml_add(xml_op, XML_LRM_ATTR_TARGET, node); // For context during triage crm_xml_add_int(xml_op, XML_LRM_ATTR_CALLID, op->call_id); crm_xml_add_int(xml_op, XML_LRM_ATTR_RC, op->rc); crm_xml_add_int(xml_op, XML_LRM_ATTR_OPSTATUS, op->op_status); crm_xml_add_ms(xml_op, XML_LRM_ATTR_INTERVAL_MS, op->interval_ms); if (compare_version("2.1", caller_version) <= 0) { if (op->t_run || op->t_rcchange || op->exec_time || op->queue_time) { crm_trace("Timing data (" PCMK__OP_FMT "): last=%u change=%u exec=%u queue=%u", op->rsc_id, op->op_type, op->interval_ms, op->t_run, op->t_rcchange, op->exec_time, op->queue_time); if ((op->interval_ms != 0) && (op->t_rcchange != 0)) { // Recurring ops may have changed rc after initial run crm_xml_add_ll(xml_op, XML_RSC_OP_LAST_CHANGE, (long long) op->t_rcchange); } else { crm_xml_add_ll(xml_op, XML_RSC_OP_LAST_CHANGE, (long long) op->t_run); } crm_xml_add_int(xml_op, XML_RSC_OP_T_EXEC, op->exec_time); crm_xml_add_int(xml_op, XML_RSC_OP_T_QUEUE, op->queue_time); } } if (pcmk__str_any_of(op->op_type, PCMK_ACTION_MIGRATE_TO, PCMK_ACTION_MIGRATE_FROM, NULL)) { /* * Record migrate_source and migrate_target always for migrate ops. */ const char *name = XML_LRM_ATTR_MIGRATE_SOURCE; crm_xml_add(xml_op, name, crm_meta_value(op->params, name)); name = XML_LRM_ATTR_MIGRATE_TARGET; crm_xml_add(xml_op, name, crm_meta_value(op->params, name)); } add_op_digest_to_xml(op, xml_op); if (op_id_additional) { free(op_id); op_id = op_id_additional; op_id_additional = NULL; goto again; } if (local_user_data) { free(local_user_data); op->user_data = NULL; } free(magic); free(op_id); free(key); return xml_op; } /*! * \internal * \brief Check whether an action shutdown-locks a resource to a node * * If the shutdown-lock cluster property is set, resources will not be recovered * on a different node if cleanly stopped, and may start only on that same node. * This function checks whether that applies to a given action, so that the * transition graph can be marked appropriately. * * \param[in] action Action to check * * \return true if \p action locks its resource to the action's node, * otherwise false */ bool pcmk__action_locks_rsc_to_node(const pcmk_action_t *action) { // Only resource actions taking place on resource's lock node are locked if ((action == NULL) || (action->rsc == NULL) || !pe__same_node(action->node, action->rsc->lock_node)) { return false; } /* During shutdown, only stops are locked (otherwise, another action such as * a demote would cause the controller to clear the lock) */ if (action->node->details->shutdown && (action->task != NULL) && (strcmp(action->task, PCMK_ACTION_STOP) != 0)) { return false; } return true; } /* lowest to highest */ static gint sort_action_id(gconstpointer a, gconstpointer b) { const pcmk__related_action_t *action_wrapper2 = a; const pcmk__related_action_t *action_wrapper1 = b; if (a == NULL) { return 1; } if (b == NULL) { return -1; } if (action_wrapper1->action->id < action_wrapper2->action->id) { return 1; } if (action_wrapper1->action->id > action_wrapper2->action->id) { return -1; } return 0; } /*! * \internal * \brief Remove any duplicate action inputs, merging action flags * * \param[in,out] action Action whose inputs should be checked */ void pcmk__deduplicate_action_inputs(pcmk_action_t *action) { GList *item = NULL; GList *next = NULL; pcmk__related_action_t *last_input = NULL; action->actions_before = g_list_sort(action->actions_before, sort_action_id); for (item = action->actions_before; item != NULL; item = next) { pcmk__related_action_t *input = item->data; next = item->next; if ((last_input != NULL) && (input->action->id == last_input->action->id)) { crm_trace("Input %s (%d) duplicate skipped for action %s (%d)", input->action->uuid, input->action->id, action->uuid, action->id); /* For the purposes of scheduling, the ordering flags no longer * matter, but crm_simulate looks at certain ones when creating a * dot graph. Combining the flags is sufficient for that purpose. */ last_input->type |= input->type; if (input->state == pe_link_dumped) { last_input->state = pe_link_dumped; } free(item->data); action->actions_before = g_list_delete_link(action->actions_before, item); } else { last_input = input; input->state = pe_link_not_dumped; } } } /*! * \internal * \brief Output all scheduled actions * * \param[in,out] scheduler Scheduler data */ void pcmk__output_actions(pcmk_scheduler_t *scheduler) { pcmk__output_t *out = scheduler->priv; // Output node (non-resource) actions for (GList *iter = scheduler->actions; iter != NULL; iter = iter->next) { char *node_name = NULL; char *task = NULL; pcmk_action_t *action = (pcmk_action_t *) iter->data; if (action->rsc != NULL) { continue; // Resource actions will be output later } else if (pcmk_is_set(action->flags, pcmk_action_optional)) { continue; // This action was not scheduled } if (pcmk__str_eq(action->task, PCMK_ACTION_DO_SHUTDOWN, pcmk__str_none)) { task = strdup("Shutdown"); } else if (pcmk__str_eq(action->task, PCMK_ACTION_STONITH, pcmk__str_none)) { const char *op = g_hash_table_lookup(action->meta, "stonith_action"); task = crm_strdup_printf("Fence (%s)", op); } else { continue; // Don't display other node action types } if (pe__is_guest_node(action->node)) { const pcmk_resource_t *remote = action->node->details->remote_rsc; node_name = crm_strdup_printf("%s (resource: %s)", pe__node_name(action->node), remote->container->id); } else if (action->node != NULL) { node_name = crm_strdup_printf("%s", pe__node_name(action->node)); } out->message(out, "node-action", task, node_name, action->reason); free(node_name); free(task); } // Output resource actions for (GList *iter = scheduler->resources; iter != NULL; iter = iter->next) { pcmk_resource_t *rsc = (pcmk_resource_t *) iter->data; rsc->cmds->output_actions(rsc); } } /*! * \internal * \brief Get action name needed to compare digest for configuration changes * * \param[in] task Action name from history * \param[in] interval_ms Action interval (in milliseconds) * * \return Action name whose digest should be compared */ static const char * task_for_digest(const char *task, guint interval_ms) { /* Certain actions need to be compared against the parameters used to start * the resource. */ if ((interval_ms == 0) && pcmk__str_any_of(task, PCMK_ACTION_MONITOR, PCMK_ACTION_MIGRATE_FROM, PCMK_ACTION_PROMOTE, NULL)) { task = PCMK_ACTION_START; } return task; } /*! * \internal * \brief Check whether only sanitized parameters to an action changed * * When collecting CIB files for troubleshooting, crm_report will mask * sensitive resource parameters. If simulations were run using that, affected * resources would appear to need a restart, which would complicate * troubleshooting. To avoid that, we save a "secure digest" of non-sensitive * parameters. This function used that digest to check whether only masked * parameters are different. * * \param[in] xml_op Resource history entry with secure digest * \param[in] digest_data Operation digest information being compared * \param[in] scheduler Scheduler data * * \return true if only sanitized parameters changed, otherwise false */ static bool only_sanitized_changed(const xmlNode *xml_op, const pcmk__op_digest_t *digest_data, const pcmk_scheduler_t *scheduler) { const char *digest_secure = NULL; if (!pcmk_is_set(scheduler->flags, pcmk_sched_sanitized)) { // The scheduler is not being run as a simulation return false; } digest_secure = crm_element_value(xml_op, XML_LRM_ATTR_SECURE_DIGEST); return (digest_data->rc != pcmk__digest_match) && (digest_secure != NULL) && (digest_data->digest_secure_calc != NULL) && (strcmp(digest_data->digest_secure_calc, digest_secure) == 0); } /*! * \internal * \brief Force a restart due to a configuration change * * \param[in,out] rsc Resource that action is for * \param[in] task Name of action whose configuration changed * \param[in] interval_ms Action interval (in milliseconds) * \param[in,out] node Node where resource should be restarted */ static void force_restart(pcmk_resource_t *rsc, const char *task, guint interval_ms, pcmk_node_t *node) { char *key = pcmk__op_key(rsc->id, task, interval_ms); pcmk_action_t *required = custom_action(rsc, key, task, NULL, FALSE, rsc->cluster); pe_action_set_reason(required, "resource definition change", true); trigger_unfencing(rsc, node, "Device parameters changed", NULL, rsc->cluster); } /*! * \internal * \brief Schedule a reload of a resource on a node * * \param[in,out] data Resource to reload * \param[in] user_data Where resource should be reloaded */ static void schedule_reload(gpointer data, gpointer user_data) { pcmk_resource_t *rsc = data; const pcmk_node_t *node = user_data; pcmk_action_t *reload = NULL; // For collective resources, just call recursively for children if (rsc->variant > pcmk_rsc_variant_primitive) { g_list_foreach(rsc->children, schedule_reload, user_data); return; } // Skip the reload in certain situations if ((node == NULL) || !pcmk_is_set(rsc->flags, pcmk_rsc_managed) || pcmk_is_set(rsc->flags, pcmk_rsc_failed)) { pe_rsc_trace(rsc, "Skip reload of %s:%s%s %s", rsc->id, pcmk_is_set(rsc->flags, pcmk_rsc_managed)? "" : " unmanaged", pcmk_is_set(rsc->flags, pcmk_rsc_failed)? " failed" : "", (node == NULL)? "inactive" : node->details->uname); return; } /* If a resource's configuration changed while a start was pending, * force a full restart instead of a reload. */ if (pcmk_is_set(rsc->flags, pcmk_rsc_start_pending)) { pe_rsc_trace(rsc, "%s: preventing agent reload because start pending", rsc->id); custom_action(rsc, stop_key(rsc), PCMK_ACTION_STOP, node, FALSE, rsc->cluster); return; } // Schedule the reload pe__set_resource_flags(rsc, pcmk_rsc_reload); reload = custom_action(rsc, reload_key(rsc), PCMK_ACTION_RELOAD_AGENT, node, FALSE, rsc->cluster); pe_action_set_reason(reload, "resource definition change", FALSE); // Set orderings so that a required stop or demote cancels the reload pcmk__new_ordering(NULL, NULL, reload, rsc, stop_key(rsc), NULL, pcmk__ar_ordered|pcmk__ar_then_cancels_first, rsc->cluster); pcmk__new_ordering(NULL, NULL, reload, rsc, demote_key(rsc), NULL, pcmk__ar_ordered|pcmk__ar_then_cancels_first, rsc->cluster); } /*! * \internal * \brief Handle any configuration change for an action * * Given an action from resource history, if the resource's configuration * changed since the action was done, schedule any actions needed (restart, * reload, unfencing, rescheduling recurring actions, etc.). * * \param[in,out] rsc Resource that action is for * \param[in,out] node Node that action was on * \param[in] xml_op Action XML from resource history * * \return true if action configuration changed, otherwise false */ bool pcmk__check_action_config(pcmk_resource_t *rsc, pcmk_node_t *node, const xmlNode *xml_op) { guint interval_ms = 0; const char *task = NULL; const pcmk__op_digest_t *digest_data = NULL; CRM_CHECK((rsc != NULL) && (node != NULL) && (xml_op != NULL), return false); task = crm_element_value(xml_op, XML_LRM_ATTR_TASK); CRM_CHECK(task != NULL, return false); crm_element_value_ms(xml_op, XML_LRM_ATTR_INTERVAL_MS, &interval_ms); // If this is a recurring action, check whether it has been orphaned if (interval_ms > 0) { if (pcmk__find_action_config(rsc, task, interval_ms, false) != NULL) { pe_rsc_trace(rsc, "%s-interval %s for %s on %s is in configuration", pcmk__readable_interval(interval_ms), task, rsc->id, pe__node_name(node)); } else if (pcmk_is_set(rsc->cluster->flags, pcmk_sched_cancel_removed_actions)) { pcmk__schedule_cancel(rsc, crm_element_value(xml_op, XML_LRM_ATTR_CALLID), task, interval_ms, node, "orphan"); return true; } else { - pe_rsc_debug(rsc, "%s-interval %s for %s on %s is orphaned", - pcmk__readable_interval(interval_ms), task, rsc->id, - pe__node_name(node)); + pcmk__rsc_debug(rsc, "%s-interval %s for %s on %s is orphaned", + pcmk__readable_interval(interval_ms), task, rsc->id, + pe__node_name(node)); return true; } } crm_trace("Checking %s-interval %s for %s on %s for configuration changes", pcmk__readable_interval(interval_ms), task, rsc->id, pe__node_name(node)); task = task_for_digest(task, interval_ms); digest_data = rsc_action_digest_cmp(rsc, xml_op, node, rsc->cluster); if (only_sanitized_changed(xml_op, digest_data, rsc->cluster)) { if (!pcmk__is_daemon && (rsc->cluster->priv != NULL)) { pcmk__output_t *out = rsc->cluster->priv; out->info(out, "Only 'private' parameters to %s-interval %s for %s " "on %s changed: %s", pcmk__readable_interval(interval_ms), task, rsc->id, pe__node_name(node), crm_element_value(xml_op, XML_ATTR_TRANSITION_MAGIC)); } return false; } switch (digest_data->rc) { case pcmk__digest_restart: crm_log_xml_debug(digest_data->params_restart, "params:restart"); force_restart(rsc, task, interval_ms, node); return true; case pcmk__digest_unknown: case pcmk__digest_mismatch: // Changes that can potentially be handled by an agent reload if (interval_ms > 0) { /* Recurring actions aren't reloaded per se, they are just * re-scheduled so the next run uses the new parameters. * The old instance will be cancelled automatically. */ crm_log_xml_debug(digest_data->params_all, "params:reschedule"); pcmk__reschedule_recurring(rsc, task, interval_ms, node); } else if (crm_element_value(xml_op, XML_LRM_ATTR_RESTART_DIGEST) != NULL) { // Agent supports reload, so use it trigger_unfencing(rsc, node, "Device parameters changed (reload)", NULL, rsc->cluster); crm_log_xml_debug(digest_data->params_all, "params:reload"); schedule_reload((gpointer) rsc, (gpointer) node); } else { pe_rsc_trace(rsc, "Restarting %s " "because agent doesn't support reload", rsc->id); crm_log_xml_debug(digest_data->params_restart, "params:restart"); force_restart(rsc, task, interval_ms, node); } return true; default: break; } return false; } /*! * \internal * \brief Create a list of resource's action history entries, sorted by call ID * * \param[in] rsc_entry Resource's status XML * \param[out] start_index Where to store index of start-like action, if any * \param[out] stop_index Where to store index of stop action, if any */ static GList * rsc_history_as_list(const xmlNode *rsc_entry, int *start_index, int *stop_index) { GList *ops = NULL; for (xmlNode *rsc_op = first_named_child(rsc_entry, XML_LRM_TAG_RSC_OP); rsc_op != NULL; rsc_op = crm_next_same_xml(rsc_op)) { ops = g_list_prepend(ops, rsc_op); } ops = g_list_sort(ops, sort_op_by_callid); calculate_active_ops(ops, start_index, stop_index); return ops; } /*! * \internal * \brief Process a resource's action history from the CIB status * * Given a resource's action history, if the resource's configuration * changed since the actions were done, schedule any actions needed (restart, * reload, unfencing, rescheduling recurring actions, clean-up, etc.). * (This also cancels recurring actions for maintenance mode, which is not * entirely related but convenient to do here.) * * \param[in] rsc_entry Resource's status XML * \param[in,out] rsc Resource whose history is being processed * \param[in,out] node Node whose history is being processed */ static void process_rsc_history(const xmlNode *rsc_entry, pcmk_resource_t *rsc, pcmk_node_t *node) { int offset = -1; int stop_index = 0; int start_index = 0; GList *sorted_op_list = NULL; if (pcmk_is_set(rsc->flags, pcmk_rsc_removed)) { if (pe_rsc_is_anon_clone(pe__const_top_resource(rsc, false))) { pe_rsc_trace(rsc, "Skipping configuration check " "for orphaned clone instance %s", rsc->id); } else { pe_rsc_trace(rsc, "Skipping configuration check and scheduling clean-up " "for orphaned resource %s", rsc->id); pcmk__schedule_cleanup(rsc, node, false); } return; } if (pe_find_node_id(rsc->running_on, node->details->id) == NULL) { if (pcmk__rsc_agent_changed(rsc, node, rsc_entry, false)) { pcmk__schedule_cleanup(rsc, node, false); } pe_rsc_trace(rsc, "Skipping configuration check for %s " "because no longer active on %s", rsc->id, pe__node_name(node)); return; } pe_rsc_trace(rsc, "Checking for configuration changes for %s on %s", rsc->id, pe__node_name(node)); if (pcmk__rsc_agent_changed(rsc, node, rsc_entry, true)) { pcmk__schedule_cleanup(rsc, node, false); } sorted_op_list = rsc_history_as_list(rsc_entry, &start_index, &stop_index); if (start_index < stop_index) { return; // Resource is stopped } for (GList *iter = sorted_op_list; iter != NULL; iter = iter->next) { xmlNode *rsc_op = (xmlNode *) iter->data; const char *task = NULL; guint interval_ms = 0; if (++offset < start_index) { // Skip actions that happened before a start continue; } task = crm_element_value(rsc_op, XML_LRM_ATTR_TASK); crm_element_value_ms(rsc_op, XML_LRM_ATTR_INTERVAL_MS, &interval_ms); if ((interval_ms > 0) && (pcmk_is_set(rsc->flags, pcmk_rsc_maintenance) || node->details->maintenance)) { // Maintenance mode cancels recurring operations pcmk__schedule_cancel(rsc, crm_element_value(rsc_op, XML_LRM_ATTR_CALLID), task, interval_ms, node, "maintenance mode"); } else if ((interval_ms > 0) || pcmk__strcase_any_of(task, PCMK_ACTION_MONITOR, PCMK_ACTION_START, PCMK_ACTION_PROMOTE, PCMK_ACTION_MIGRATE_FROM, NULL)) { /* 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 assigned 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, pcmk__check_active, rsc->cluster); } else if (pcmk__check_action_config(rsc, node, rsc_op) && (pe_get_failcount(node, rsc, NULL, pcmk__fc_effective, NULL) != 0)) { pe__clear_failcount(rsc, node, "action definition changed", rsc->cluster); } } } g_list_free(sorted_op_list); } /*! * \internal * \brief Process a node's action history from the CIB status * * Given a node's resource history, if the resource's configuration changed * since the actions were done, schedule any actions needed (restart, * reload, unfencing, rescheduling recurring actions, clean-up, etc.). * (This also cancels recurring actions for maintenance mode, which is not * entirely related but convenient to do here.) * * \param[in,out] node Node whose history is being processed * \param[in] lrm_rscs Node's from CIB status XML */ static void process_node_history(pcmk_node_t *node, const xmlNode *lrm_rscs) { crm_trace("Processing node history for %s", pe__node_name(node)); for (const xmlNode *rsc_entry = first_named_child(lrm_rscs, XML_LRM_TAG_RESOURCE); rsc_entry != NULL; rsc_entry = crm_next_same_xml(rsc_entry)) { if (rsc_entry->children != NULL) { GList *result = pcmk__rscs_matching_id(ID(rsc_entry), node->details->data_set); for (GList *iter = result; iter != NULL; iter = iter->next) { pcmk_resource_t *rsc = (pcmk_resource_t *) iter->data; if (rsc->variant == pcmk_rsc_variant_primitive) { process_rsc_history(rsc_entry, rsc, node); } } g_list_free(result); } } } // XPath to find a node's resource history #define XPATH_NODE_HISTORY "/" XML_TAG_CIB "/" XML_CIB_TAG_STATUS \ "/" XML_CIB_TAG_STATE "[@" XML_ATTR_UNAME "='%s']" \ "/" XML_CIB_TAG_LRM "/" XML_LRM_TAG_RESOURCES /*! * \internal * \brief Process any resource configuration changes in the CIB status * * Go through all nodes' resource history, and if a resource's configuration * changed since its actions were done, schedule any actions needed (restart, * reload, unfencing, rescheduling recurring actions, clean-up, etc.). * (This also cancels recurring actions for maintenance mode, which is not * entirely related but convenient to do here.) * * \param[in,out] scheduler Scheduler data */ void pcmk__handle_rsc_config_changes(pcmk_scheduler_t *scheduler) { crm_trace("Check resource and action configuration for changes"); /* Rather than iterate through the status section, iterate through the nodes * and search for the appropriate status subsection for each. This skips * orphaned nodes and lets us eliminate some cases before searching the XML. */ for (GList *iter = scheduler->nodes; iter != NULL; iter = iter->next) { pcmk_node_t *node = (pcmk_node_t *) iter->data; /* Don't bother checking actions for a node that can't run actions ... * unless it's in maintenance mode, in which case we still need to * cancel any existing recurring monitors. */ if (node->details->maintenance || pcmk__node_available(node, false, false)) { char *xpath = NULL; xmlNode *history = NULL; xpath = crm_strdup_printf(XPATH_NODE_HISTORY, node->details->uname); history = get_xpath_object(xpath, scheduler->input, LOG_NEVER); free(xpath); process_node_history(node, history); } } } diff --git a/lib/pacemaker/pcmk_sched_bundle.c b/lib/pacemaker/pcmk_sched_bundle.c index bb375d776d..802f4d147b 100644 --- a/lib/pacemaker/pcmk_sched_bundle.c +++ b/lib/pacemaker/pcmk_sched_bundle.c @@ -1,1059 +1,1059 @@ /* * Copyright 2004-2023 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU General Public License version 2 * or later (GPLv2+) WITHOUT ANY WARRANTY. */ #include #include #include #include #include "libpacemaker_private.h" struct assign_data { const pcmk_node_t *prefer; bool stop_if_fail; }; /*! * \internal * \brief Assign a single bundle replica's resources (other than container) * * \param[in,out] replica Replica to assign * \param[in] user_data Preferred node, if any * * \return true (to indicate that any further replicas should be processed) */ static bool assign_replica(pcmk__bundle_replica_t *replica, void *user_data) { pcmk_node_t *container_host = NULL; struct assign_data *assign_data = user_data; const pcmk_node_t *prefer = assign_data->prefer; bool stop_if_fail = assign_data->stop_if_fail; const pcmk_resource_t *bundle = pe__const_top_resource(replica->container, true); if (replica->ip != NULL) { pe_rsc_trace(bundle, "Assigning bundle %s IP %s", bundle->id, replica->ip->id); replica->ip->cmds->assign(replica->ip, prefer, stop_if_fail); } container_host = replica->container->allocated_to; if (replica->remote != NULL) { if (pe__is_guest_or_remote_node(container_host)) { /* REMOTE_CONTAINER_HACK: "Nested" connection resources must be on * the same host because Pacemaker Remote only supports a single * active connection. */ pcmk__new_colocation("#replica-remote-with-host-remote", NULL, INFINITY, replica->remote, container_host->details->remote_rsc, NULL, NULL, pcmk__coloc_influence); } pe_rsc_trace(bundle, "Assigning bundle %s connection %s", bundle->id, replica->remote->id); replica->remote->cmds->assign(replica->remote, prefer, stop_if_fail); } if (replica->child != NULL) { pcmk_node_t *node = NULL; GHashTableIter iter; g_hash_table_iter_init(&iter, replica->child->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (gpointer *) &node)) { if (!pe__same_node(node, replica->node)) { node->weight = -INFINITY; } else if (!pcmk__threshold_reached(replica->child, node, NULL)) { node->weight = INFINITY; } } pe__set_resource_flags(replica->child->parent, pcmk_rsc_assigning); pe_rsc_trace(bundle, "Assigning bundle %s replica child %s", bundle->id, replica->child->id); replica->child->cmds->assign(replica->child, replica->node, stop_if_fail); pe__clear_resource_flags(replica->child->parent, pcmk_rsc_assigning); } return true; } /*! * \internal * \brief Assign a bundle resource to a node * * \param[in,out] rsc Resource to assign to a node * \param[in] prefer Node to prefer, if all else is equal * \param[in] stop_if_fail If \c true and a primitive descendant of \p rsc * can't be assigned to a node, set the * descendant's next role to stopped and update * existing actions * * \return Node that \p rsc is assigned to, if assigned entirely to one node * * \note If \p stop_if_fail is \c false, then \c pcmk__unassign_resource() can * completely undo the assignment. A successful assignment can be either * undone or left alone as final. A failed assignment has the same effect * as calling pcmk__unassign_resource(); there are no side effects on * roles or actions. */ pcmk_node_t * pcmk__bundle_assign(pcmk_resource_t *rsc, const pcmk_node_t *prefer, bool stop_if_fail) { GList *containers = NULL; pcmk_resource_t *bundled_resource = NULL; struct assign_data assign_data = { prefer, stop_if_fail }; CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_bundle)); pe_rsc_trace(rsc, "Assigning bundle %s", rsc->id); pe__set_resource_flags(rsc, pcmk_rsc_assigning); pe__show_node_scores(!pcmk_is_set(rsc->cluster->flags, pcmk_sched_output_scores), rsc, __func__, rsc->allowed_nodes, rsc->cluster); // Assign all containers first, so we know what nodes the bundle will be on containers = g_list_sort(pe__bundle_containers(rsc), pcmk__cmp_instance); pcmk__assign_instances(rsc, containers, pe__bundle_max(rsc), rsc->fns->max_per_node(rsc)); g_list_free(containers); // Then assign remaining replica resources pe__foreach_bundle_replica(rsc, assign_replica, (void *) &assign_data); // Finally, assign the bundled resources to each bundle node bundled_resource = pe__bundled_resource(rsc); if (bundled_resource != NULL) { pcmk_node_t *node = NULL; GHashTableIter iter; g_hash_table_iter_init(&iter, bundled_resource->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (gpointer *) & node)) { if (pe__node_is_bundle_instance(rsc, node)) { node->weight = 0; } else { node->weight = -INFINITY; } } bundled_resource->cmds->assign(bundled_resource, prefer, stop_if_fail); } pe__clear_resource_flags(rsc, pcmk_rsc_assigning|pcmk_rsc_unassigned); return NULL; } /*! * \internal * \brief Create actions for a bundle replica's resources (other than child) * * \param[in,out] replica Replica to create actions for * \param[in] user_data Unused * * \return true (to indicate that any further replicas should be processed) */ static bool create_replica_actions(pcmk__bundle_replica_t *replica, void *user_data) { if (replica->ip != NULL) { replica->ip->cmds->create_actions(replica->ip); } if (replica->container != NULL) { replica->container->cmds->create_actions(replica->container); } if (replica->remote != NULL) { replica->remote->cmds->create_actions(replica->remote); } return true; } /*! * \internal * \brief Create all actions needed for a given bundle resource * * \param[in,out] rsc Bundle resource to create actions for */ void pcmk__bundle_create_actions(pcmk_resource_t *rsc) { pcmk_action_t *action = NULL; GList *containers = NULL; pcmk_resource_t *bundled_resource = NULL; CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_bundle)); pe__foreach_bundle_replica(rsc, create_replica_actions, NULL); containers = pe__bundle_containers(rsc); pcmk__create_instance_actions(rsc, containers); g_list_free(containers); bundled_resource = pe__bundled_resource(rsc); if (bundled_resource != NULL) { bundled_resource->cmds->create_actions(bundled_resource); if (pcmk_is_set(bundled_resource->flags, pcmk_rsc_promotable)) { pe__new_rsc_pseudo_action(rsc, PCMK_ACTION_PROMOTE, true, true); action = pe__new_rsc_pseudo_action(rsc, PCMK_ACTION_PROMOTED, true, true); action->priority = INFINITY; pe__new_rsc_pseudo_action(rsc, PCMK_ACTION_DEMOTE, true, true); action = pe__new_rsc_pseudo_action(rsc, PCMK_ACTION_DEMOTED, true, true); action->priority = INFINITY; } } } /*! * \internal * \brief Create internal constraints for a bundle replica's resources * * \param[in,out] replica Replica to create internal constraints for * \param[in,out] user_data Replica's parent bundle * * \return true (to indicate that any further replicas should be processed) */ static bool replica_internal_constraints(pcmk__bundle_replica_t *replica, void *user_data) { pcmk_resource_t *bundle = user_data; replica->container->cmds->internal_constraints(replica->container); // Start bundle -> start replica container pcmk__order_starts(bundle, replica->container, pcmk__ar_unrunnable_first_blocks |pcmk__ar_then_implies_first_graphed); // Stop bundle -> stop replica child and container if (replica->child != NULL) { pcmk__order_stops(bundle, replica->child, pcmk__ar_then_implies_first_graphed); } pcmk__order_stops(bundle, replica->container, pcmk__ar_then_implies_first_graphed); // Start replica container -> bundle is started pcmk__order_resource_actions(replica->container, PCMK_ACTION_START, bundle, PCMK_ACTION_RUNNING, pcmk__ar_first_implies_then_graphed); // Stop replica container -> bundle is stopped pcmk__order_resource_actions(replica->container, PCMK_ACTION_STOP, bundle, PCMK_ACTION_STOPPED, pcmk__ar_first_implies_then_graphed); if (replica->ip != NULL) { replica->ip->cmds->internal_constraints(replica->ip); // Replica IP address -> replica container (symmetric) pcmk__order_starts(replica->ip, replica->container, pcmk__ar_unrunnable_first_blocks |pcmk__ar_guest_allowed); pcmk__order_stops(replica->container, replica->ip, pcmk__ar_then_implies_first|pcmk__ar_guest_allowed); pcmk__new_colocation("#ip-with-container", NULL, INFINITY, replica->ip, replica->container, NULL, NULL, pcmk__coloc_influence); } if (replica->remote != NULL) { /* This handles ordering and colocating remote relative to container * (via "#resource-with-container"). Since IP is also ordered and * colocated relative to the container, we don't need to do anything * explicit here with IP. */ replica->remote->cmds->internal_constraints(replica->remote); } if (replica->child != NULL) { CRM_ASSERT(replica->remote != NULL); // "Start remote then child" is implicit in scheduler's remote logic } return true; } /*! * \internal * \brief Create implicit constraints needed for a bundle resource * * \param[in,out] rsc Bundle resource to create implicit constraints for */ void pcmk__bundle_internal_constraints(pcmk_resource_t *rsc) { pcmk_resource_t *bundled_resource = NULL; CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_bundle)); pe__foreach_bundle_replica(rsc, replica_internal_constraints, rsc); bundled_resource = pe__bundled_resource(rsc); if (bundled_resource == NULL) { return; } // Start bundle -> start bundled clone pcmk__order_resource_actions(rsc, PCMK_ACTION_START, bundled_resource, PCMK_ACTION_START, pcmk__ar_then_implies_first_graphed); // Bundled clone is started -> bundle is started pcmk__order_resource_actions(bundled_resource, PCMK_ACTION_RUNNING, rsc, PCMK_ACTION_RUNNING, pcmk__ar_first_implies_then_graphed); // Stop bundle -> stop bundled clone pcmk__order_resource_actions(rsc, PCMK_ACTION_STOP, bundled_resource, PCMK_ACTION_STOP, pcmk__ar_then_implies_first_graphed); // Bundled clone is stopped -> bundle is stopped pcmk__order_resource_actions(bundled_resource, PCMK_ACTION_STOPPED, rsc, PCMK_ACTION_STOPPED, pcmk__ar_first_implies_then_graphed); bundled_resource->cmds->internal_constraints(bundled_resource); if (!pcmk_is_set(bundled_resource->flags, pcmk_rsc_promotable)) { return; } pcmk__promotable_restart_ordering(rsc); // Demote bundle -> demote bundled clone pcmk__order_resource_actions(rsc, PCMK_ACTION_DEMOTE, bundled_resource, PCMK_ACTION_DEMOTE, pcmk__ar_then_implies_first_graphed); // Bundled clone is demoted -> bundle is demoted pcmk__order_resource_actions(bundled_resource, PCMK_ACTION_DEMOTED, rsc, PCMK_ACTION_DEMOTED, pcmk__ar_first_implies_then_graphed); // Promote bundle -> promote bundled clone pcmk__order_resource_actions(rsc, PCMK_ACTION_PROMOTE, bundled_resource, PCMK_ACTION_PROMOTE, pcmk__ar_then_implies_first_graphed); // Bundled clone is promoted -> bundle is promoted pcmk__order_resource_actions(bundled_resource, PCMK_ACTION_PROMOTED, rsc, PCMK_ACTION_PROMOTED, pcmk__ar_first_implies_then_graphed); } struct match_data { const pcmk_node_t *node; // Node to compare against replica pcmk_resource_t *container; // Replica container corresponding to node }; /*! * \internal * \brief Check whether a replica container is assigned to a given node * * \param[in] replica Replica to check * \param[in,out] user_data struct match_data with node to compare against * * \return true if the replica does not match (to indicate further replicas * should be processed), otherwise false */ static bool match_replica_container(const pcmk__bundle_replica_t *replica, void *user_data) { struct match_data *match_data = user_data; if (pcmk__instance_matches(replica->container, match_data->node, pcmk_role_unknown, false)) { match_data->container = replica->container; return false; // Match found, don't bother searching further replicas } return true; // No match, keep searching } /*! * \internal * \brief Get the host to which a bundle node is assigned * * \param[in] node Possible bundle node to check * * \return Node to which the container for \p node is assigned if \p node is a * bundle node, otherwise \p node itself */ static const pcmk_node_t * get_bundle_node_host(const pcmk_node_t *node) { if (pe__is_bundle_node(node)) { const pcmk_resource_t *container = node->details->remote_rsc->container; return container->fns->location(container, NULL, 0); } return node; } /*! * \internal * \brief Find a bundle container compatible with a dependent resource * * \param[in] dependent Dependent resource in colocation with bundle * \param[in] bundle Bundle that \p dependent is colocated with * * \return A container from \p bundle assigned to the same node as \p dependent * if assigned, otherwise assigned to any of dependent's allowed nodes, * otherwise NULL. */ static pcmk_resource_t * compatible_container(const pcmk_resource_t *dependent, const pcmk_resource_t *bundle) { GList *scratch = NULL; struct match_data match_data = { NULL, NULL }; // If dependent is assigned, only check there match_data.node = dependent->fns->location(dependent, NULL, 0); match_data.node = get_bundle_node_host(match_data.node); if (match_data.node != NULL) { pe__foreach_const_bundle_replica(bundle, match_replica_container, &match_data); return match_data.container; } // Otherwise, check for any of the dependent's allowed nodes scratch = g_hash_table_get_values(dependent->allowed_nodes); scratch = pcmk__sort_nodes(scratch, NULL); for (const GList *iter = scratch; iter != NULL; iter = iter->next) { match_data.node = iter->data; match_data.node = get_bundle_node_host(match_data.node); if (match_data.node == NULL) { continue; } pe__foreach_const_bundle_replica(bundle, match_replica_container, &match_data); if (match_data.container != NULL) { break; } } g_list_free(scratch); return match_data.container; } struct coloc_data { const pcmk__colocation_t *colocation; pcmk_resource_t *dependent; GList *container_hosts; }; /*! * \internal * \brief Apply a colocation score to replica node scores or resource priority * * \param[in] replica Replica of primary bundle resource in colocation * \param[in,out] user_data struct coloc_data for colocation being applied * * \return true (to indicate that any further replicas should be processed) */ static bool replica_apply_coloc_score(const pcmk__bundle_replica_t *replica, void *user_data) { struct coloc_data *coloc_data = user_data; pcmk_node_t *chosen = NULL; if (coloc_data->colocation->score < INFINITY) { replica->container->cmds->apply_coloc_score(coloc_data->dependent, replica->container, coloc_data->colocation, false); return true; } chosen = replica->container->fns->location(replica->container, NULL, 0); if ((chosen == NULL) || is_set_recursive(replica->container, pcmk_rsc_blocked, true)) { return true; } if ((coloc_data->colocation->primary_role >= pcmk_role_promoted) && ((replica->child == NULL) || (replica->child->next_role < pcmk_role_promoted))) { return true; } pe_rsc_trace(pe__const_top_resource(replica->container, true), "Allowing mandatory colocation %s using %s @%d", coloc_data->colocation->id, pe__node_name(chosen), chosen->weight); coloc_data->container_hosts = g_list_prepend(coloc_data->container_hosts, chosen); return true; } /*! * \internal * \brief Apply a colocation's score to node scores or resource priority * * Given a colocation constraint, apply its score to the dependent's * allowed node scores (if we are still placing resources) or priority (if * we are choosing promotable clone instance roles). * * \param[in,out] dependent Dependent resource in colocation * \param[in] primary Primary resource in colocation * \param[in] colocation Colocation constraint to apply * \param[in] for_dependent true if called on behalf of dependent */ void pcmk__bundle_apply_coloc_score(pcmk_resource_t *dependent, const pcmk_resource_t *primary, const pcmk__colocation_t *colocation, bool for_dependent) { struct coloc_data coloc_data = { colocation, dependent, NULL }; /* This should never be called for the bundle itself as a dependent. * Instead, we add its colocation constraints to its containers and bundled * primitive and call the apply_coloc_score() method for them as dependents. */ CRM_ASSERT((primary != NULL) && (primary->variant == pcmk_rsc_variant_bundle) && (dependent != NULL) && (dependent->variant == pcmk_rsc_variant_primitive) && (colocation != NULL) && !for_dependent); if (pcmk_is_set(primary->flags, pcmk_rsc_unassigned)) { pe_rsc_trace(primary, "Skipping applying colocation %s " "because %s is still provisional", colocation->id, primary->id); return; } pe_rsc_trace(primary, "Applying colocation %s (%s with %s at %s)", colocation->id, dependent->id, primary->id, pcmk_readable_score(colocation->score)); /* If the constraint dependent is a clone or bundle, "dependent" here is one * of its instances. Look for a compatible instance of this bundle. */ if (colocation->dependent->variant > pcmk_rsc_variant_group) { const pcmk_resource_t *primary_container = NULL; primary_container = compatible_container(dependent, primary); if (primary_container != NULL) { // Success, we found one - pe_rsc_debug(primary, "Pairing %s with %s", - dependent->id, primary_container->id); + pcmk__rsc_debug(primary, "Pairing %s with %s", + dependent->id, primary_container->id); dependent->cmds->apply_coloc_score(dependent, primary_container, colocation, true); } else if (colocation->score >= INFINITY) { // Failure, and it's fatal crm_notice("%s cannot run because there is no compatible " "instance of %s to colocate with", dependent->id, primary->id); pcmk__assign_resource(dependent, NULL, true, true); } else { // Failure, but we can ignore it - pe_rsc_debug(primary, - "%s cannot be colocated with any instance of %s", - dependent->id, primary->id); + pcmk__rsc_debug(primary, + "%s cannot be colocated with any instance of %s", + dependent->id, primary->id); } return; } pe__foreach_const_bundle_replica(primary, replica_apply_coloc_score, &coloc_data); if (colocation->score >= INFINITY) { pcmk__colocation_intersect_nodes(dependent, primary, colocation, coloc_data.container_hosts, false); } g_list_free(coloc_data.container_hosts); } // Bundle implementation of pcmk_assignment_methods_t:with_this_colocations() void pcmk__with_bundle_colocations(const pcmk_resource_t *rsc, const pcmk_resource_t *orig_rsc, GList **list) { const pcmk_resource_t *bundled_rsc = NULL; CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_bundle) && (orig_rsc != NULL) && (list != NULL)); // The bundle itself and its containers always get its colocations if ((orig_rsc == rsc) || pcmk_is_set(orig_rsc->flags, pcmk_rsc_replica_container)) { pcmk__add_with_this_list(list, rsc->rsc_cons_lhs, orig_rsc); return; } /* The bundled resource gets the colocations if it's promotable and we've * begun choosing roles */ bundled_rsc = pe__bundled_resource(rsc); if ((bundled_rsc == NULL) || !pcmk_is_set(bundled_rsc->flags, pcmk_rsc_promotable) || (pe__const_top_resource(orig_rsc, false) != bundled_rsc)) { return; } if (orig_rsc == bundled_rsc) { if (pe__clone_flag_is_set(orig_rsc, pcmk__clone_promotion_constrained)) { /* orig_rsc is the clone and we're setting roles (or have already * done so) */ pcmk__add_with_this_list(list, rsc->rsc_cons_lhs, orig_rsc); } } else if (!pcmk_is_set(orig_rsc->flags, pcmk_rsc_unassigned)) { /* orig_rsc is an instance and is already assigned. If something * requests colocations for orig_rsc now, it's for setting roles. */ pcmk__add_with_this_list(list, rsc->rsc_cons_lhs, orig_rsc); } } // Bundle implementation of pcmk_assignment_methods_t:this_with_colocations() void pcmk__bundle_with_colocations(const pcmk_resource_t *rsc, const pcmk_resource_t *orig_rsc, GList **list) { const pcmk_resource_t *bundled_rsc = NULL; CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_bundle) && (orig_rsc != NULL) && (list != NULL)); // The bundle itself and its containers always get its colocations if ((orig_rsc == rsc) || pcmk_is_set(orig_rsc->flags, pcmk_rsc_replica_container)) { pcmk__add_this_with_list(list, rsc->rsc_cons, orig_rsc); return; } /* The bundled resource gets the colocations if it's promotable and we've * begun choosing roles */ bundled_rsc = pe__bundled_resource(rsc); if ((bundled_rsc == NULL) || !pcmk_is_set(bundled_rsc->flags, pcmk_rsc_promotable) || (pe__const_top_resource(orig_rsc, false) != bundled_rsc)) { return; } if (orig_rsc == bundled_rsc) { if (pe__clone_flag_is_set(orig_rsc, pcmk__clone_promotion_constrained)) { /* orig_rsc is the clone and we're setting roles (or have already * done so) */ pcmk__add_this_with_list(list, rsc->rsc_cons, orig_rsc); } } else if (!pcmk_is_set(orig_rsc->flags, pcmk_rsc_unassigned)) { /* orig_rsc is an instance and is already assigned. If something * requests colocations for orig_rsc now, it's for setting roles. */ pcmk__add_this_with_list(list, rsc->rsc_cons, orig_rsc); } } /*! * \internal * \brief Return action flags for a given bundle resource action * * \param[in,out] action Bundle resource action to get flags for * \param[in] node If not NULL, limit effects to this node * * \return Flags appropriate to \p action on \p node */ uint32_t pcmk__bundle_action_flags(pcmk_action_t *action, const pcmk_node_t *node) { GList *containers = NULL; uint32_t flags = 0; pcmk_resource_t *bundled_resource = NULL; CRM_ASSERT((action != NULL) && (action->rsc != NULL) && (action->rsc->variant == pcmk_rsc_variant_bundle)); bundled_resource = pe__bundled_resource(action->rsc); if (bundled_resource != NULL) { // Clone actions are done on the bundled clone resource, not container switch (get_complex_task(bundled_resource, action->task)) { case pcmk_action_unspecified: case pcmk_action_notify: case pcmk_action_notified: case pcmk_action_promote: case pcmk_action_promoted: case pcmk_action_demote: case pcmk_action_demoted: return pcmk__collective_action_flags(action, bundled_resource->children, node); default: break; } } containers = pe__bundle_containers(action->rsc); flags = pcmk__collective_action_flags(action, containers, node); g_list_free(containers); return flags; } /*! * \internal * \brief Apply a location constraint to a bundle replica * * \param[in,out] replica Replica to apply constraint to * \param[in,out] user_data Location constraint to apply * * \return true (to indicate that any further replicas should be processed) */ static bool apply_location_to_replica(pcmk__bundle_replica_t *replica, void *user_data) { pcmk__location_t *location = user_data; if (replica->container != NULL) { replica->container->cmds->apply_location(replica->container, location); } if (replica->ip != NULL) { replica->ip->cmds->apply_location(replica->ip, location); } return true; } /*! * \internal * \brief Apply a location constraint to a bundle resource's allowed node scores * * \param[in,out] rsc Bundle resource to apply constraint to * \param[in,out] location Location constraint to apply */ void pcmk__bundle_apply_location(pcmk_resource_t *rsc, pcmk__location_t *location) { pcmk_resource_t *bundled_resource = NULL; CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_bundle) && (location != NULL)); pcmk__apply_location(rsc, location); pe__foreach_bundle_replica(rsc, apply_location_to_replica, location); bundled_resource = pe__bundled_resource(rsc); if ((bundled_resource != NULL) && ((location->role_filter == pcmk_role_unpromoted) || (location->role_filter == pcmk_role_promoted))) { bundled_resource->cmds->apply_location(bundled_resource, location); bundled_resource->rsc_location = g_list_prepend( bundled_resource->rsc_location, location); } } #define XPATH_REMOTE "//nvpair[@name='" XML_RSC_ATTR_REMOTE_RA_ADDR "']" /*! * \internal * \brief Add a bundle replica's actions to transition graph * * \param[in,out] replica Replica to add to graph * \param[in] user_data Bundle that replica belongs to (for logging only) * * \return true (to indicate that any further replicas should be processed) */ static bool add_replica_actions_to_graph(pcmk__bundle_replica_t *replica, void *user_data) { if ((replica->remote != NULL) && (replica->container != NULL) && pe__bundle_needs_remote_name(replica->remote)) { /* REMOTE_CONTAINER_HACK: Allow remote nodes to run containers that * run pacemaker-remoted inside, without needing a separate IP for * the container. This is done by configuring the inner remote's * connection host as the magic string "#uname", then * replacing it with the underlying host when needed. */ xmlNode *nvpair = get_xpath_object(XPATH_REMOTE, replica->remote->xml, LOG_ERR); const char *calculated_addr = NULL; // Replace the value in replica->remote->xml (if appropriate) calculated_addr = pe__add_bundle_remote_name(replica->remote, replica->remote->cluster, nvpair, "value"); if (calculated_addr != NULL) { /* Since this is for the bundle as a resource, and not any * particular action, replace the value in the default * parameters (not evaluated for node). create_graph_action() * will grab it from there to replace it in node-evaluated * parameters. */ GHashTable *params = pe_rsc_params(replica->remote, NULL, replica->remote->cluster); g_hash_table_replace(params, strdup(XML_RSC_ATTR_REMOTE_RA_ADDR), strdup(calculated_addr)); } else { pcmk_resource_t *bundle = user_data; /* The only way to get here is if the remote connection is * neither currently running nor scheduled to run. That means we * won't be doing any operations that require addr (only start * requires it; we additionally use it to compare digests when * unpacking status, promote, and migrate_from history, but * that's already happened by this point). */ pcmk__rsc_info(bundle, "Unable to determine address for bundle %s " "remote connection", bundle->id); } } if (replica->ip != NULL) { replica->ip->cmds->add_actions_to_graph(replica->ip); } if (replica->container != NULL) { replica->container->cmds->add_actions_to_graph(replica->container); } if (replica->remote != NULL) { replica->remote->cmds->add_actions_to_graph(replica->remote); } return true; } /*! * \internal * \brief Add a bundle resource's actions to the transition graph * * \param[in,out] rsc Bundle resource whose actions should be added */ void pcmk__bundle_add_actions_to_graph(pcmk_resource_t *rsc) { pcmk_resource_t *bundled_resource = NULL; CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_bundle)); bundled_resource = pe__bundled_resource(rsc); if (bundled_resource != NULL) { bundled_resource->cmds->add_actions_to_graph(bundled_resource); } pe__foreach_bundle_replica(rsc, add_replica_actions_to_graph, rsc); } struct probe_data { pcmk_resource_t *bundle; // Bundle being probed pcmk_node_t *node; // Node to create probes on bool any_created; // Whether any probes have been created }; /*! * \internal * \brief Order a bundle replica's start after another replica's probe * * \param[in,out] replica Replica to order start for * \param[in,out] user_data Replica with probe to order after * * \return true (to indicate that any further replicas should be processed) */ static bool order_replica_start_after(pcmk__bundle_replica_t *replica, void *user_data) { pcmk__bundle_replica_t *probed_replica = user_data; if ((replica == probed_replica) || (replica->container == NULL)) { return true; } pcmk__new_ordering(probed_replica->container, pcmk__op_key(probed_replica->container->id, PCMK_ACTION_MONITOR, 0), NULL, replica->container, pcmk__op_key(replica->container->id, PCMK_ACTION_START, 0), NULL, pcmk__ar_ordered|pcmk__ar_if_on_same_node, replica->container->cluster); return true; } /*! * \internal * \brief Create probes for a bundle replica's resources * * \param[in,out] replica Replica to create probes for * \param[in,out] user_data struct probe_data * * \return true (to indicate that any further replicas should be processed) */ static bool create_replica_probes(pcmk__bundle_replica_t *replica, void *user_data) { struct probe_data *probe_data = user_data; if ((replica->ip != NULL) && replica->ip->cmds->create_probe(replica->ip, probe_data->node)) { probe_data->any_created = true; } if ((replica->child != NULL) && pe__same_node(probe_data->node, replica->node) && replica->child->cmds->create_probe(replica->child, probe_data->node)) { probe_data->any_created = true; } if ((replica->container != NULL) && replica->container->cmds->create_probe(replica->container, probe_data->node)) { probe_data->any_created = true; /* If we're limited to one replica per host (due to * the lack of an IP range probably), then we don't * want any of our peer containers starting until * we've established that no other copies are already * running. * * Partly this is to ensure that the maximum replicas per host is * observed, but also to ensure that the containers * don't fail to start because the necessary port * mappings (which won't include an IP for uniqueness) * are already taken */ if (probe_data->bundle->fns->max_per_node(probe_data->bundle) == 1) { pe__foreach_bundle_replica(probe_data->bundle, order_replica_start_after, replica); } } if ((replica->container != NULL) && (replica->remote != NULL) && replica->remote->cmds->create_probe(replica->remote, probe_data->node)) { /* Do not probe the remote resource until we know where the container is * running. This is required for REMOTE_CONTAINER_HACK to correctly * probe remote resources. */ char *probe_uuid = pcmk__op_key(replica->remote->id, PCMK_ACTION_MONITOR, 0); pcmk_action_t *probe = find_first_action(replica->remote->actions, probe_uuid, NULL, probe_data->node); free(probe_uuid); if (probe != NULL) { probe_data->any_created = true; pe_rsc_trace(probe_data->bundle, "Ordering %s probe on %s", replica->remote->id, pe__node_name(probe_data->node)); pcmk__new_ordering(replica->container, pcmk__op_key(replica->container->id, PCMK_ACTION_START, 0), NULL, replica->remote, NULL, probe, pcmk__ar_nested_remote_probe, probe_data->bundle->cluster); } } return true; } /*! * \internal * * \brief Schedule any probes needed for a bundle resource on a node * * \param[in,out] rsc Bundle resource to create probes for * \param[in,out] node Node to create probe on * * \return true if any probe was created, otherwise false */ bool pcmk__bundle_create_probe(pcmk_resource_t *rsc, pcmk_node_t *node) { struct probe_data probe_data = { rsc, node, false }; CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_bundle)); pe__foreach_bundle_replica(rsc, create_replica_probes, &probe_data); return probe_data.any_created; } /*! * \internal * \brief Output actions for one bundle replica * * \param[in,out] replica Replica to output actions for * \param[in] user_data Unused * * \return true (to indicate that any further replicas should be processed) */ static bool output_replica_actions(pcmk__bundle_replica_t *replica, void *user_data) { if (replica->ip != NULL) { replica->ip->cmds->output_actions(replica->ip); } if (replica->container != NULL) { replica->container->cmds->output_actions(replica->container); } if (replica->remote != NULL) { replica->remote->cmds->output_actions(replica->remote); } if (replica->child != NULL) { replica->child->cmds->output_actions(replica->child); } return true; } /*! * \internal * \brief Output a summary of scheduled actions for a bundle resource * * \param[in,out] rsc Bundle resource to output actions for */ void pcmk__output_bundle_actions(pcmk_resource_t *rsc) { CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_bundle)); pe__foreach_bundle_replica(rsc, output_replica_actions, NULL); } // Bundle implementation of pcmk_assignment_methods_t:add_utilization() void pcmk__bundle_add_utilization(const pcmk_resource_t *rsc, const pcmk_resource_t *orig_rsc, GList *all_rscs, GHashTable *utilization) { pcmk_resource_t *container = NULL; CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_bundle)); if (!pcmk_is_set(rsc->flags, pcmk_rsc_unassigned)) { return; } /* All bundle replicas are identical, so using the utilization of the first * is sufficient for any. Only the implicit container resource can have * utilization values. */ container = pe__first_container(rsc); if (container != NULL) { container->cmds->add_utilization(container, orig_rsc, all_rscs, utilization); } } // Bundle implementation of pcmk_assignment_methods_t:shutdown_lock() void pcmk__bundle_shutdown_lock(pcmk_resource_t *rsc) { CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_bundle)); // Bundles currently don't support shutdown locks } diff --git a/lib/pacemaker/pcmk_sched_clone.c b/lib/pacemaker/pcmk_sched_clone.c index 2a586a75a4..92ba3b8acc 100644 --- a/lib/pacemaker/pcmk_sched_clone.c +++ b/lib/pacemaker/pcmk_sched_clone.c @@ -1,709 +1,709 @@ /* * Copyright 2004-2023 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU General Public License version 2 * or later (GPLv2+) WITHOUT ANY WARRANTY. */ #include #include #include #include "libpacemaker_private.h" /*! * \internal * \brief Assign a clone resource's instances to nodes * * \param[in,out] rsc Clone resource to assign * \param[in] prefer Node to prefer, if all else is equal * \param[in] stop_if_fail If \c true and a primitive descendant of \p rsc * can't be assigned to a node, set the * descendant's next role to stopped and update * existing actions * * \return NULL (clones are not assigned to a single node) * * \note If \p stop_if_fail is \c false, then \c pcmk__unassign_resource() can * completely undo the assignment. A successful assignment can be either * undone or left alone as final. A failed assignment has the same effect * as calling pcmk__unassign_resource(); there are no side effects on * roles or actions. */ pcmk_node_t * pcmk__clone_assign(pcmk_resource_t *rsc, const pcmk_node_t *prefer, bool stop_if_fail) { GList *colocations = NULL; CRM_ASSERT(pe_rsc_is_clone(rsc)); if (!pcmk_is_set(rsc->flags, pcmk_rsc_unassigned)) { return NULL; // Assignment has already been done } // Detect assignment loops if (pcmk_is_set(rsc->flags, pcmk_rsc_assigning)) { - pe_rsc_debug(rsc, "Breaking assignment loop involving %s", rsc->id); + pcmk__rsc_debug(rsc, "Breaking assignment loop involving %s", rsc->id); return NULL; } pe__set_resource_flags(rsc, pcmk_rsc_assigning); // If this clone is promotable, consider nodes' promotion scores if (pcmk_is_set(rsc->flags, pcmk_rsc_promotable)) { pcmk__add_promotion_scores(rsc); } // If this clone is colocated with any other resources, assign those first colocations = pcmk__this_with_colocations(rsc); for (GList *iter = colocations; iter != NULL; iter = iter->next) { pcmk__colocation_t *constraint = (pcmk__colocation_t *) iter->data; pe_rsc_trace(rsc, "%s: Assigning colocation %s primary %s first", rsc->id, constraint->id, constraint->primary->id); constraint->primary->cmds->assign(constraint->primary, prefer, stop_if_fail); } g_list_free(colocations); // If any resources are colocated with this one, consider their preferences colocations = pcmk__with_this_colocations(rsc); g_list_foreach(colocations, pcmk__add_dependent_scores, rsc); g_list_free(colocations); pe__show_node_scores(!pcmk_is_set(rsc->cluster->flags, pcmk_sched_output_scores), rsc, __func__, rsc->allowed_nodes, rsc->cluster); rsc->children = g_list_sort(rsc->children, pcmk__cmp_instance); pcmk__assign_instances(rsc, rsc->children, pe__clone_max(rsc), pe__clone_node_max(rsc)); if (pcmk_is_set(rsc->flags, pcmk_rsc_promotable)) { pcmk__set_instance_roles(rsc); } pe__clear_resource_flags(rsc, pcmk_rsc_unassigned|pcmk_rsc_assigning); pe_rsc_trace(rsc, "Assigned clone %s", rsc->id); return NULL; } /*! * \internal * \brief Create all actions needed for a given clone resource * * \param[in,out] rsc Clone resource to create actions for */ void pcmk__clone_create_actions(pcmk_resource_t *rsc) { CRM_ASSERT(pe_rsc_is_clone(rsc)); pe_rsc_trace(rsc, "Creating actions for clone %s", rsc->id); pcmk__create_instance_actions(rsc, rsc->children); if (pcmk_is_set(rsc->flags, pcmk_rsc_promotable)) { pcmk__create_promotable_actions(rsc); } } /*! * \internal * \brief Create implicit constraints needed for a clone resource * * \param[in,out] rsc Clone resource to create implicit constraints for */ void pcmk__clone_internal_constraints(pcmk_resource_t *rsc) { bool ordered = false; CRM_ASSERT(pe_rsc_is_clone(rsc)); pe_rsc_trace(rsc, "Creating internal constraints for clone %s", rsc->id); // Restart ordering: Stop -> stopped -> start -> started pcmk__order_resource_actions(rsc, PCMK_ACTION_STOPPED, rsc, PCMK_ACTION_START, pcmk__ar_ordered); pcmk__order_resource_actions(rsc, PCMK_ACTION_START, rsc, PCMK_ACTION_RUNNING, pcmk__ar_unrunnable_first_blocks); pcmk__order_resource_actions(rsc, PCMK_ACTION_STOP, rsc, PCMK_ACTION_STOPPED, pcmk__ar_unrunnable_first_blocks); // Demoted -> stop and started -> promote if (pcmk_is_set(rsc->flags, pcmk_rsc_promotable)) { pcmk__order_resource_actions(rsc, PCMK_ACTION_DEMOTED, rsc, PCMK_ACTION_STOP, pcmk__ar_ordered); pcmk__order_resource_actions(rsc, PCMK_ACTION_RUNNING, rsc, PCMK_ACTION_PROMOTE, pcmk__ar_unrunnable_first_blocks); } ordered = pe__clone_is_ordered(rsc); if (ordered) { /* Ordered clone instances must start and stop by instance number. The * instances might have been previously shuffled for assignment or * promotion purposes, so re-sort them. */ rsc->children = g_list_sort(rsc->children, pcmk__cmp_instance_number); } for (GList *iter = rsc->children; iter != NULL; iter = iter->next) { pcmk_resource_t *instance = (pcmk_resource_t *) iter->data; instance->cmds->internal_constraints(instance); // Start clone -> start instance -> clone started pcmk__order_starts(rsc, instance, pcmk__ar_unrunnable_first_blocks |pcmk__ar_then_implies_first_graphed); pcmk__order_resource_actions(instance, PCMK_ACTION_START, rsc, PCMK_ACTION_RUNNING, pcmk__ar_first_implies_then_graphed); // Stop clone -> stop instance -> clone stopped pcmk__order_stops(rsc, instance, pcmk__ar_then_implies_first_graphed); pcmk__order_resource_actions(instance, PCMK_ACTION_STOP, rsc, PCMK_ACTION_STOPPED, pcmk__ar_first_implies_then_graphed); /* Instances of ordered clones must be started and stopped by instance * number. Since only some instances may be starting or stopping, order * each instance relative to every later instance. */ if (ordered) { for (GList *later = iter->next; later != NULL; later = later->next) { pcmk__order_starts(instance, (pcmk_resource_t *) later->data, pcmk__ar_ordered); pcmk__order_stops((pcmk_resource_t *) later->data, instance, pcmk__ar_ordered); } } } if (pcmk_is_set(rsc->flags, pcmk_rsc_promotable)) { pcmk__order_promotable_instances(rsc); } } /*! * \internal * \brief Check whether colocated resources can be interleaved * * \param[in] colocation Colocation constraint with clone as primary * * \return true if colocated resources can be interleaved, otherwise false */ static bool can_interleave(const pcmk__colocation_t *colocation) { const pcmk_resource_t *dependent = colocation->dependent; // Only colocations between clone or bundle resources use interleaving if (dependent->variant <= pcmk_rsc_variant_group) { return false; } // Only the dependent needs to be marked for interleaving if (!crm_is_true(g_hash_table_lookup(dependent->meta, XML_RSC_ATTR_INTERLEAVE))) { return false; } /* @TODO Do we actually care about multiple primary instances sharing a * dependent instance? */ if (dependent->fns->max_per_node(dependent) != colocation->primary->fns->max_per_node(colocation->primary)) { pcmk__config_err("Cannot interleave %s and %s because they do not " "support the same number of instances per node", dependent->id, colocation->primary->id); return false; } return true; } /*! * \internal * \brief Apply a colocation's score to node scores or resource priority * * Given a colocation constraint, apply its score to the dependent's * allowed node scores (if we are still placing resources) or priority (if * we are choosing promotable clone instance roles). * * \param[in,out] dependent Dependent resource in colocation * \param[in] primary Primary resource in colocation * \param[in] colocation Colocation constraint to apply * \param[in] for_dependent true if called on behalf of dependent */ void pcmk__clone_apply_coloc_score(pcmk_resource_t *dependent, const pcmk_resource_t *primary, const pcmk__colocation_t *colocation, bool for_dependent) { const GList *iter = NULL; /* This should never be called for the clone itself as a dependent. Instead, * we add its colocation constraints to its instances and call the * apply_coloc_score() method for the instances as dependents. */ CRM_ASSERT(!for_dependent); CRM_ASSERT((colocation != NULL) && pe_rsc_is_clone(primary) && (dependent != NULL) && (dependent->variant == pcmk_rsc_variant_primitive)); if (pcmk_is_set(primary->flags, pcmk_rsc_unassigned)) { pe_rsc_trace(primary, "Delaying processing colocation %s " "because cloned primary %s is still provisional", colocation->id, primary->id); return; } pe_rsc_trace(primary, "Processing colocation %s (%s with clone %s @%s)", colocation->id, dependent->id, primary->id, pcmk_readable_score(colocation->score)); // Apply role-specific colocations if (pcmk_is_set(primary->flags, pcmk_rsc_promotable) && (colocation->primary_role != pcmk_role_unknown)) { if (pcmk_is_set(dependent->flags, pcmk_rsc_unassigned)) { // We're assigning the dependent to a node pcmk__update_dependent_with_promotable(primary, dependent, colocation); return; } if (colocation->dependent_role == pcmk_role_promoted) { // We're choosing a role for the dependent pcmk__update_promotable_dependent_priority(primary, dependent, colocation); return; } } // Apply interleaved colocations if (can_interleave(colocation)) { const pcmk_resource_t *primary_instance = NULL; primary_instance = pcmk__find_compatible_instance(dependent, primary, pcmk_role_unknown, false); if (primary_instance != NULL) { - pe_rsc_debug(primary, "Interleaving %s with %s", - dependent->id, primary_instance->id); + pcmk__rsc_debug(primary, "Interleaving %s with %s", + dependent->id, primary_instance->id); dependent->cmds->apply_coloc_score(dependent, primary_instance, colocation, true); } else if (colocation->score >= INFINITY) { crm_notice("%s cannot run because it cannot interleave with " "any instance of %s", dependent->id, primary->id); pcmk__assign_resource(dependent, NULL, true, true); } else { - pe_rsc_debug(primary, - "%s will not colocate with %s " - "because no instance can interleave with it", - dependent->id, primary->id); + pcmk__rsc_debug(primary, + "%s will not colocate with %s " + "because no instance can interleave with it", + dependent->id, primary->id); } return; } // Apply mandatory colocations if (colocation->score >= INFINITY) { GList *primary_nodes = NULL; // Dependent can run only where primary will have unblocked instances for (iter = primary->children; iter != NULL; iter = iter->next) { const pcmk_resource_t *instance = iter->data; pcmk_node_t *chosen = instance->fns->location(instance, NULL, 0); if ((chosen != NULL) && !is_set_recursive(instance, pcmk_rsc_blocked, TRUE)) { pe_rsc_trace(primary, "Allowing %s: %s %d", colocation->id, pe__node_name(chosen), chosen->weight); primary_nodes = g_list_prepend(primary_nodes, chosen); } } pcmk__colocation_intersect_nodes(dependent, primary, colocation, primary_nodes, false); g_list_free(primary_nodes); return; } // Apply optional colocations for (iter = primary->children; iter != NULL; iter = iter->next) { const pcmk_resource_t *instance = iter->data; instance->cmds->apply_coloc_score(dependent, instance, colocation, false); } } // Clone implementation of pcmk_assignment_methods_t:with_this_colocations() void pcmk__with_clone_colocations(const pcmk_resource_t *rsc, const pcmk_resource_t *orig_rsc, GList **list) { CRM_CHECK((rsc != NULL) && (orig_rsc != NULL) && (list != NULL), return); pcmk__add_with_this_list(list, rsc->rsc_cons_lhs, orig_rsc); if (rsc->parent != NULL) { rsc->parent->cmds->with_this_colocations(rsc->parent, orig_rsc, list); } } // Clone implementation of pcmk_assignment_methods_t:this_with_colocations() void pcmk__clone_with_colocations(const pcmk_resource_t *rsc, const pcmk_resource_t *orig_rsc, GList **list) { CRM_CHECK((rsc != NULL) && (orig_rsc != NULL) && (list != NULL), return); pcmk__add_this_with_list(list, rsc->rsc_cons, orig_rsc); if (rsc->parent != NULL) { rsc->parent->cmds->this_with_colocations(rsc->parent, orig_rsc, list); } } /*! * \internal * \brief Return action flags for a given clone resource action * * \param[in,out] action Action to get flags for * \param[in] node If not NULL, limit effects to this node * * \return Flags appropriate to \p action on \p node */ uint32_t pcmk__clone_action_flags(pcmk_action_t *action, const pcmk_node_t *node) { CRM_ASSERT((action != NULL) && pe_rsc_is_clone(action->rsc)); return pcmk__collective_action_flags(action, action->rsc->children, node); } /*! * \internal * \brief Apply a location constraint to a clone resource's allowed node scores * * \param[in,out] rsc Clone resource to apply constraint to * \param[in,out] location Location constraint to apply */ void pcmk__clone_apply_location(pcmk_resource_t *rsc, pcmk__location_t *location) { CRM_CHECK((location != NULL) && pe_rsc_is_clone(rsc), return); pcmk__apply_location(rsc, location); for (GList *iter = rsc->children; iter != NULL; iter = iter->next) { pcmk_resource_t *instance = (pcmk_resource_t *) iter->data; instance->cmds->apply_location(instance, location); } } // GFunc wrapper for calling the action_flags() resource method static void call_action_flags(gpointer data, gpointer user_data) { pcmk_resource_t *rsc = user_data; rsc->cmds->action_flags((pcmk_action_t *) data, NULL); } /*! * \internal * \brief Add a clone resource's actions to the transition graph * * \param[in,out] rsc Resource whose actions should be added */ void pcmk__clone_add_actions_to_graph(pcmk_resource_t *rsc) { CRM_ASSERT(pe_rsc_is_clone(rsc)); g_list_foreach(rsc->actions, call_action_flags, rsc); pe__create_clone_notifications(rsc); for (GList *iter = rsc->children; iter != NULL; iter = iter->next) { pcmk_resource_t *child_rsc = (pcmk_resource_t *) iter->data; child_rsc->cmds->add_actions_to_graph(child_rsc); } pcmk__add_rsc_actions_to_graph(rsc); pe__free_clone_notification_data(rsc); } /*! * \internal * \brief Check whether a resource or any children have been probed on a node * * \param[in] rsc Resource to check * \param[in] node Node to check * * \return true if \p node is in the known_on table of \p rsc or any of its * children, otherwise false */ static bool rsc_probed_on(const pcmk_resource_t *rsc, const pcmk_node_t *node) { if (rsc->children != NULL) { for (GList *child_iter = rsc->children; child_iter != NULL; child_iter = child_iter->next) { pcmk_resource_t *child = (pcmk_resource_t *) child_iter->data; if (rsc_probed_on(child, node)) { return true; } } return false; } if (rsc->known_on != NULL) { GHashTableIter iter; pcmk_node_t *known_node = NULL; g_hash_table_iter_init(&iter, rsc->known_on); while (g_hash_table_iter_next(&iter, NULL, (gpointer *) &known_node)) { if (pe__same_node(node, known_node)) { return true; } } } return false; } /*! * \internal * \brief Find clone instance that has been probed on given node * * \param[in] clone Clone resource to check * \param[in] node Node to check * * \return Instance of \p clone that has been probed on \p node if any, * otherwise NULL */ static pcmk_resource_t * find_probed_instance_on(const pcmk_resource_t *clone, const pcmk_node_t *node) { for (GList *iter = clone->children; iter != NULL; iter = iter->next) { pcmk_resource_t *instance = (pcmk_resource_t *) iter->data; if (rsc_probed_on(instance, node)) { return instance; } } return NULL; } /*! * \internal * \brief Probe an anonymous clone on a node * * \param[in,out] clone Anonymous clone to probe * \param[in,out] node Node to probe \p clone on */ static bool probe_anonymous_clone(pcmk_resource_t *clone, pcmk_node_t *node) { // Check whether we already probed an instance on this node pcmk_resource_t *child = find_probed_instance_on(clone, node); // Otherwise, check if we plan to start an instance on this node for (GList *iter = clone->children; (iter != NULL) && (child == NULL); iter = iter->next) { pcmk_resource_t *instance = (pcmk_resource_t *) iter->data; const pcmk_node_t *instance_node = NULL; instance_node = instance->fns->location(instance, NULL, 0); if (pe__same_node(instance_node, node)) { child = instance; } } // Otherwise, use the first clone instance if (child == NULL) { child = clone->children->data; } // Anonymous clones only need to probe a single instance return child->cmds->create_probe(child, node); } /*! * \internal * \brief Schedule any probes needed for a resource on a node * * \param[in,out] rsc Resource to create probe for * \param[in,out] node Node to create probe on * * \return true if any probe was created, otherwise false */ bool pcmk__clone_create_probe(pcmk_resource_t *rsc, pcmk_node_t *node) { CRM_ASSERT((node != NULL) && pe_rsc_is_clone(rsc)); if (rsc->exclusive_discover) { /* The clone is configured to be probed only where a location constraint * exists with resource-discovery set to exclusive. * * This check is not strictly necessary here since the instance's * create_probe() method would also check, but doing it here is more * efficient (especially for unique clones with a large number of * instances), and affects the CRM_meta_notify_available_uname variable * passed with notify actions. */ pcmk_node_t *allowed = g_hash_table_lookup(rsc->allowed_nodes, node->details->id); if ((allowed == NULL) || (allowed->rsc_discover_mode != pcmk_probe_exclusive)) { /* This node is not marked for resource discovery. Remove it from * allowed_nodes so that notifications contain only nodes that the * clone can possibly run on. */ pe_rsc_trace(rsc, "Skipping probe for %s on %s because resource has " "exclusive discovery but is not allowed on node", rsc->id, pe__node_name(node)); g_hash_table_remove(rsc->allowed_nodes, node->details->id); return false; } } rsc->children = g_list_sort(rsc->children, pcmk__cmp_instance_number); if (pcmk_is_set(rsc->flags, pcmk_rsc_unique)) { return pcmk__probe_resource_list(rsc->children, node); } else { return probe_anonymous_clone(rsc, node); } } /*! * \internal * \brief Add meta-attributes relevant to transition graph actions to XML * * Add clone-specific meta-attributes needed for transition graph actions. * * \param[in] rsc Clone resource whose meta-attributes should be added * \param[in,out] xml Transition graph action attributes XML to add to */ void pcmk__clone_add_graph_meta(const pcmk_resource_t *rsc, xmlNode *xml) { char *name = NULL; CRM_ASSERT(pe_rsc_is_clone(rsc) && (xml != NULL)); name = crm_meta_name(XML_RSC_ATTR_UNIQUE); crm_xml_add(xml, name, pe__rsc_bool_str(rsc, pcmk_rsc_unique)); free(name); name = crm_meta_name(XML_RSC_ATTR_NOTIFY); crm_xml_add(xml, name, pe__rsc_bool_str(rsc, pcmk_rsc_notify)); free(name); name = crm_meta_name(PCMK_META_CLONE_MAX); crm_xml_add_int(xml, name, pe__clone_max(rsc)); free(name); name = crm_meta_name(PCMK_META_CLONE_NODE_MAX); crm_xml_add_int(xml, name, pe__clone_node_max(rsc)); free(name); if (pcmk_is_set(rsc->flags, pcmk_rsc_promotable)) { int promoted_max = pe__clone_promoted_max(rsc); int promoted_node_max = pe__clone_promoted_node_max(rsc); name = crm_meta_name(PCMK_META_PROMOTED_MAX); crm_xml_add_int(xml, name, promoted_max); free(name); name = crm_meta_name(PCMK_META_PROMOTED_NODE_MAX); crm_xml_add_int(xml, name, promoted_node_max); free(name); /* @COMPAT Maintain backward compatibility with resource agents that * expect the old names (deprecated since 2.0.0). */ name = crm_meta_name(PCMK_XA_PROMOTED_MAX_LEGACY); crm_xml_add_int(xml, name, promoted_max); free(name); name = crm_meta_name(PCMK_XA_PROMOTED_NODE_MAX_LEGACY); crm_xml_add_int(xml, name, promoted_node_max); free(name); } } // Clone implementation of pcmk_assignment_methods_t:add_utilization() void pcmk__clone_add_utilization(const pcmk_resource_t *rsc, const pcmk_resource_t *orig_rsc, GList *all_rscs, GHashTable *utilization) { bool existing = false; pcmk_resource_t *child = NULL; CRM_ASSERT(pe_rsc_is_clone(rsc) && (orig_rsc != NULL) && (utilization != NULL)); if (!pcmk_is_set(rsc->flags, pcmk_rsc_unassigned)) { return; } // Look for any child already existing in the list for (GList *iter = rsc->children; iter != NULL; iter = iter->next) { child = (pcmk_resource_t *) iter->data; if (g_list_find(all_rscs, child)) { existing = true; // Keep checking remaining children } else { // If this is a clone of a group, look for group's members for (GList *member_iter = child->children; member_iter != NULL; member_iter = member_iter->next) { pcmk_resource_t *member = (pcmk_resource_t *) member_iter->data; if (g_list_find(all_rscs, member) != NULL) { // Add *child's* utilization, not group member's child->cmds->add_utilization(child, orig_rsc, all_rscs, utilization); existing = true; break; } } } } if (!existing && (rsc->children != NULL)) { // If nothing was found, still add first child's utilization child = (pcmk_resource_t *) rsc->children->data; child->cmds->add_utilization(child, orig_rsc, all_rscs, utilization); } } // Clone implementation of pcmk_assignment_methods_t:shutdown_lock() void pcmk__clone_shutdown_lock(pcmk_resource_t *rsc) { CRM_ASSERT(pe_rsc_is_clone(rsc)); return; // Clones currently don't support shutdown locks } diff --git a/lib/pacemaker/pcmk_sched_fencing.c b/lib/pacemaker/pcmk_sched_fencing.c index fe963040c8..262909cab3 100644 --- a/lib/pacemaker/pcmk_sched_fencing.c +++ b/lib/pacemaker/pcmk_sched_fencing.c @@ -1,500 +1,500 @@ /* * Copyright 2004-2023 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU General Public License version 2 * or later (GPLv2+) WITHOUT ANY WARRANTY. */ #include #include #include #include #include #include "libpacemaker_private.h" /*! * \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(const pcmk_resource_t *rsc, const pcmk_node_t *node) { if (g_hash_table_lookup(rsc->known_on, node->details->id) != NULL) { return TRUE; } else if ((rsc->variant == pcmk_rsc_variant_primitive) && pe_rsc_is_anon_clone(rsc->parent) && (g_hash_table_lookup(rsc->parent->known_on, node->details->id) != NULL)) { /* 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,out] rsc Resource to be ordered * \param[in,out] stonith_op Fence action */ static void order_start_vs_fencing(pcmk_resource_t *rsc, pcmk_action_t *stonith_op) { pcmk_node_t *target; CRM_CHECK(stonith_op && stonith_op->node, return); target = stonith_op->node; for (GList *iter = rsc->actions; iter != NULL; iter = iter->next) { pcmk_action_t *action = iter->data; switch (action->needs) { case pcmk_requires_nothing: // Anything other than start or promote requires nothing break; case pcmk_requires_fencing: order_actions(stonith_op, action, pcmk__ar_ordered); break; case pcmk_requires_quorum: if (pcmk__str_eq(action->task, PCMK_ACTION_START, pcmk__str_none) && (g_hash_table_lookup(rsc->allowed_nodes, target->details->id) != NULL) && !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, pe__node_name(target)); + pcmk__rsc_debug(rsc, "Ordering %s after %s recovery", + action->uuid, pe__node_name(target)); order_actions(stonith_op, action, pcmk__ar_ordered |pcmk__ar_unrunnable_first_blocks); } break; } } } /*! * \internal * \brief Order a resource's stop and demote actions relative to fencing * * \param[in,out] rsc Resource to be ordered * \param[in,out] stonith_op Fence action */ static void order_stop_vs_fencing(pcmk_resource_t *rsc, pcmk_action_t *stonith_op) { GList *iter = NULL; GList *action_list = NULL; bool order_implicit = false; pcmk_resource_t *top = uber_parent(rsc); pcmk_action_t *parent_stop = NULL; pcmk_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, PCMK_ACTION_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 (pcmk_is_set(rsc->flags, pcmk_rsc_needs_fencing) || pe__is_guest_node(target)) { order_implicit = true; } if (action_list && order_implicit) { parent_stop = find_first_action(top->actions, NULL, PCMK_ACTION_STOP, NULL); } for (iter = action_list; iter != NULL; iter = iter->next) { pcmk_action_t *action = iter->data; // The stop would never complete, so convert it into a pseudo-action. pe__set_action_flags(action, pcmk_action_pseudo|pcmk_action_runnable); if (order_implicit) { /* 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 * pcmk__ar_guest_allowed 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, pcmk__ar_guest_allowed); } order_actions(stonith_op, parent_stop, pcmk__ar_guest_allowed); } if (pcmk_is_set(rsc->flags, pcmk_rsc_failed)) { crm_notice("Stop of failed resource %s is implicit %s %s is fenced", rsc->id, (order_implicit? "after" : "because"), pe__node_name(target)); } else { crm_info("%s is implicit %s %s is fenced", action->uuid, (order_implicit? "after" : "because"), pe__node_name(target)); } if (pcmk_is_set(rsc->flags, pcmk_rsc_notify)) { pe__order_notifs_after_fencing(action, rsc, stonith_op); } #if 0 /* It might be a good idea to stop healthy resources on a node about to * be fenced, when possible. * * However, fencing must be done before a failed resource's * (pseudo-)stop action, so that could create a loop. For example, given * a group of A and B running on node N with a failed stop of B: * * fence N -> stop B (pseudo-op) -> stop A -> fence N * * The block below creates the stop A -> fence N ordering and therefore * must (at least for now) be disabled. Instead, run the block above and * treat all resources on N as B would be (i.e., as a pseudo-op after * the fencing). * * @TODO Maybe break the "A requires B" dependency in * pcmk__update_action_for_orderings() and use this block for healthy * resources instead of the above. */ crm_info("Moving healthy resource %s off %s before fencing", rsc->id, pe__node_name(node)); pcmk__new_ordering(rsc, stop_key(rsc), NULL, NULL, strdup(PCMK_ACTION_STONITH), stonith_op, pcmk__ar_ordered, rsc->cluster); #endif } g_list_free(action_list); /* Get a list of demote actions potentially implied by the fencing */ action_list = pe__resource_actions(rsc, target, PCMK_ACTION_DEMOTE, FALSE); for (iter = action_list; iter != NULL; iter = iter->next) { pcmk_action_t *action = iter->data; if (!(action->node->details->online) || action->node->details->unclean || pcmk_is_set(rsc->flags, pcmk_rsc_failed)) { if (pcmk_is_set(rsc->flags, pcmk_rsc_failed)) { pcmk__rsc_info(rsc, "Demote of failed resource %s is implicit " "after %s is fenced", rsc->id, pe__node_name(target)); } else { pcmk__rsc_info(rsc, "%s is implicit after %s is fenced", action->uuid, pe__node_name(target)); } /* The demote would never complete and is now implied by the * fencing, so convert it into a pseudo-action. */ pe__set_action_flags(action, pcmk_action_pseudo|pcmk_action_runnable); if (pe_rsc_is_bundled(rsc)) { // Recovery will be ordered as usual after parent's implied stop } else if (order_implicit) { order_actions(stonith_op, action, pcmk__ar_guest_allowed|pcmk__ar_ordered); } } } g_list_free(action_list); } /*! * \internal * \brief Order resource actions properly relative to fencing * * \param[in,out] rsc Resource whose actions should be ordered * \param[in,out] stonith_op Fencing operation to be ordered against */ static void rsc_stonith_ordering(pcmk_resource_t *rsc, pcmk_action_t *stonith_op) { if (rsc->children) { for (GList *iter = rsc->children; iter != NULL; iter = iter->next) { pcmk_resource_t *child_rsc = iter->data; rsc_stonith_ordering(child_rsc, stonith_op); } } else if (!pcmk_is_set(rsc->flags, pcmk_rsc_managed)) { pe_rsc_trace(rsc, "Skipping fencing constraints for unmanaged resource: %s", rsc->id); } else { order_start_vs_fencing(rsc, stonith_op); order_stop_vs_fencing(rsc, stonith_op); } } /*! * \internal * \brief Order all actions appropriately relative to a fencing operation * * Ensure start operations of affected resources are ordered after fencing, * imply stop and demote operations of affected resources by marking them as * pseudo-actions, etc. * * \param[in,out] stonith_op Fencing operation * \param[in,out] scheduler Scheduler data */ void pcmk__order_vs_fence(pcmk_action_t *stonith_op, pcmk_scheduler_t *scheduler) { CRM_CHECK(stonith_op && scheduler, return); for (GList *r = scheduler->resources; r != NULL; r = r->next) { rsc_stonith_ordering((pcmk_resource_t *) r->data, stonith_op); } } /*! * \internal * \brief Order an action after unfencing * * \param[in] rsc Resource that action is for * \param[in,out] node Node that action is on * \param[in,out] action Action to be ordered after unfencing * \param[in] order Ordering flags */ void pcmk__order_vs_unfence(const pcmk_resource_t *rsc, pcmk_node_t *node, pcmk_action_t *action, enum pcmk__action_relation_flags order) { /* 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 ((pcmk_is_set(rsc->flags, pcmk_rsc_fence_device) && pcmk_is_set(rsc->cluster->flags, pcmk_sched_enable_unfencing)) || pcmk_is_set(rsc->flags, pcmk_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. */ pcmk_action_t *unfence = pe_fence_op(node, PCMK_ACTION_ON, TRUE, NULL, FALSE, node->details->data_set); order_actions(unfence, action, order); if (!pcmk__node_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, node->details->data_set); free(reason); } } } /*! * \internal * \brief Create pseudo-op for guest node fence, and order relative to it * * \param[in,out] node Guest node to fence */ void pcmk__fence_guest(pcmk_node_t *node) { pcmk_resource_t *container = NULL; pcmk_action_t *stop = NULL; pcmk_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 = PCMK_ACTION_OFF; CRM_ASSERT(node != NULL); /* Check whether guest's container resource has any explicit stop or * start (the stop may be implied by fencing of the guest's host). */ container = node->details->remote_rsc->container; if (container) { stop = find_first_action(container->actions, NULL, PCMK_ACTION_STOP, NULL); if (find_first_action(container->actions, NULL, PCMK_ACTION_START, NULL)) { fence_action = PCMK_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", FALSE, node->details->data_set); pe__set_action_flags(stonith_op, pcmk_action_pseudo|pcmk_action_runnable); /* 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 != NULL) && pcmk_is_set(stop->flags, pcmk_action_pseudo)) { pcmk_action_t *parent_stonith_op = pe_fence_op(stop->node, NULL, FALSE, NULL, FALSE, node->details->data_set); crm_info("Implying guest %s is down (action %d) after %s fencing", pe__node_name(node), stonith_op->id, pe__node_name(stop->node)); order_actions(parent_stonith_op, stonith_op, pcmk__ar_unrunnable_first_blocks |pcmk__ar_first_implies_then); } else if (stop) { order_actions(stop, stonith_op, pcmk__ar_unrunnable_first_blocks |pcmk__ar_first_implies_then); crm_info("Implying guest %s is down (action %d) " "after container %s is stopped (action %d)", pe__node_name(node), 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, PCMK_ACTION_STOP, NULL); if (stop) { order_actions(stop, stonith_op, pcmk__ar_ordered); crm_info("Implying guest %s is down (action %d) " "after connection is stopped (action %d)", pe__node_name(node), stonith_op->id, stop->id); } else { /* Not sure why we're fencing, but everything must already be * cleanly stopped. */ crm_info("Implying guest %s is down (action %d) ", pe__node_name(node), stonith_op->id); } } // Order/imply other actions relative to pseudo-fence as with real fence pcmk__order_vs_fence(stonith_op, node->details->data_set); } /*! * \internal * \brief Check whether node has already been unfenced * * \param[in] node Node to check * * \return true if node has a nonzero #node-unfenced attribute (or none), * otherwise false */ bool pcmk__node_unfenced(const pcmk_node_t *node) { const char *unfenced = pe_node_attribute_raw(node, CRM_ATTR_UNFENCED); return !pcmk__str_eq(unfenced, "0", pcmk__str_null_matches); } /*! * \internal * \brief Order a resource's start and stop relative to unfencing of a node * * \param[in,out] data Node that could be unfenced * \param[in,out] user_data Resource to order */ void pcmk__order_restart_vs_unfence(gpointer data, gpointer user_data) { pcmk_node_t *node = (pcmk_node_t *) data; pcmk_resource_t *rsc = (pcmk_resource_t *) user_data; pcmk_action_t *unfence = pe_fence_op(node, PCMK_ACTION_ON, true, NULL, false, rsc->cluster); 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 */ pcmk__new_ordering(rsc, stop_key(rsc), NULL, NULL, strdup(unfence->uuid), unfence, pcmk__ar_ordered|pcmk__ar_if_on_same_node, rsc->cluster); pcmk__new_ordering(NULL, strdup(unfence->uuid), unfence, rsc, start_key(rsc), NULL, pcmk__ar_first_implies_same_node_then |pcmk__ar_if_on_same_node, rsc->cluster); } diff --git a/lib/pacemaker/pcmk_sched_group.c b/lib/pacemaker/pcmk_sched_group.c index acf301cefd..4b250ccf1a 100644 --- a/lib/pacemaker/pcmk_sched_group.c +++ b/lib/pacemaker/pcmk_sched_group.c @@ -1,950 +1,950 @@ /* * Copyright 2004-2023 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU General Public License version 2 * or later (GPLv2+) WITHOUT ANY WARRANTY. */ #include #include #include #include #include "libpacemaker_private.h" /*! * \internal * \brief Assign a group resource to a node * * \param[in,out] rsc Group resource to assign to a node * \param[in] prefer Node to prefer, if all else is equal * \param[in] stop_if_fail If \c true and a child of \p rsc can't be * assigned to a node, set the child's next role to * stopped and update existing actions * * \return Node that \p rsc is assigned to, if assigned entirely to one node * * \note If \p stop_if_fail is \c false, then \c pcmk__unassign_resource() can * completely undo the assignment. A successful assignment can be either * undone or left alone as final. A failed assignment has the same effect * as calling pcmk__unassign_resource(); there are no side effects on * roles or actions. */ pcmk_node_t * pcmk__group_assign(pcmk_resource_t *rsc, const pcmk_node_t *prefer, bool stop_if_fail) { pcmk_node_t *first_assigned_node = NULL; pcmk_resource_t *first_member = NULL; CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_group)); if (!pcmk_is_set(rsc->flags, pcmk_rsc_unassigned)) { return rsc->allocated_to; // Assignment already done } if (pcmk_is_set(rsc->flags, pcmk_rsc_assigning)) { - pe_rsc_debug(rsc, "Assignment dependency loop detected involving %s", - rsc->id); + pcmk__rsc_debug(rsc, "Assignment dependency loop detected involving %s", + rsc->id); return NULL; } if (rsc->children == NULL) { // No members to assign pe__clear_resource_flags(rsc, pcmk_rsc_unassigned); return NULL; } pe__set_resource_flags(rsc, pcmk_rsc_assigning); first_member = (pcmk_resource_t *) rsc->children->data; rsc->role = first_member->role; pe__show_node_scores(!pcmk_is_set(rsc->cluster->flags, pcmk_sched_output_scores), rsc, __func__, rsc->allowed_nodes, rsc->cluster); for (GList *iter = rsc->children; iter != NULL; iter = iter->next) { pcmk_resource_t *member = (pcmk_resource_t *) iter->data; pcmk_node_t *node = NULL; pe_rsc_trace(rsc, "Assigning group %s member %s", rsc->id, member->id); node = member->cmds->assign(member, prefer, stop_if_fail); if (first_assigned_node == NULL) { first_assigned_node = node; } } pe__set_next_role(rsc, first_member->next_role, "first group member"); pe__clear_resource_flags(rsc, pcmk_rsc_assigning|pcmk_rsc_unassigned); if (!pe__group_flag_is_set(rsc, pcmk__group_colocated)) { return NULL; } return first_assigned_node; } /*! * \internal * \brief Create a pseudo-operation for a group as an ordering point * * \param[in,out] group Group resource to create action for * \param[in] action Action name * * \return Newly created pseudo-operation */ static pcmk_action_t * create_group_pseudo_op(pcmk_resource_t *group, const char *action) { pcmk_action_t *op = custom_action(group, pcmk__op_key(group->id, action, 0), action, NULL, TRUE, group->cluster); pe__set_action_flags(op, pcmk_action_pseudo|pcmk_action_runnable); return op; } /*! * \internal * \brief Create all actions needed for a given group resource * * \param[in,out] rsc Group resource to create actions for */ void pcmk__group_create_actions(pcmk_resource_t *rsc) { CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_group)); pe_rsc_trace(rsc, "Creating actions for group %s", rsc->id); // Create actions for individual group members for (GList *iter = rsc->children; iter != NULL; iter = iter->next) { pcmk_resource_t *member = (pcmk_resource_t *) iter->data; member->cmds->create_actions(member); } // Create pseudo-actions for group itself to serve as ordering points create_group_pseudo_op(rsc, PCMK_ACTION_START); create_group_pseudo_op(rsc, PCMK_ACTION_RUNNING); create_group_pseudo_op(rsc, PCMK_ACTION_STOP); create_group_pseudo_op(rsc, PCMK_ACTION_STOPPED); if (crm_is_true(g_hash_table_lookup(rsc->meta, XML_RSC_ATTR_PROMOTABLE))) { create_group_pseudo_op(rsc, PCMK_ACTION_DEMOTE); create_group_pseudo_op(rsc, PCMK_ACTION_DEMOTED); create_group_pseudo_op(rsc, PCMK_ACTION_PROMOTE); create_group_pseudo_op(rsc, PCMK_ACTION_PROMOTED); } } // User data for member_internal_constraints() struct member_data { // These could be derived from member but this avoids some function calls bool ordered; bool colocated; bool promotable; pcmk_resource_t *last_active; pcmk_resource_t *previous_member; }; /*! * \internal * \brief Create implicit constraints needed for a group member * * \param[in,out] data Group member to create implicit constraints for * \param[in,out] user_data Member data (struct member_data *) */ static void member_internal_constraints(gpointer data, gpointer user_data) { pcmk_resource_t *member = (pcmk_resource_t *) data; struct member_data *member_data = (struct member_data *) user_data; // For ordering demote vs demote or stop vs stop uint32_t down_flags = pcmk__ar_then_implies_first_graphed; // For ordering demote vs demoted or stop vs stopped uint32_t post_down_flags = pcmk__ar_first_implies_then_graphed; // Create the individual member's implicit constraints member->cmds->internal_constraints(member); if (member_data->previous_member == NULL) { // This is first member if (member_data->ordered) { pe__set_order_flags(down_flags, pcmk__ar_ordered); post_down_flags = pcmk__ar_first_implies_then; } } else if (member_data->colocated) { uint32_t flags = pcmk__coloc_none; if (pcmk_is_set(member->flags, pcmk_rsc_critical)) { flags |= pcmk__coloc_influence; } // Colocate this member with the previous one pcmk__new_colocation("#group-members", NULL, INFINITY, member, member_data->previous_member, NULL, NULL, flags); } if (member_data->promotable) { // Demote group -> demote member -> group is demoted pcmk__order_resource_actions(member->parent, PCMK_ACTION_DEMOTE, member, PCMK_ACTION_DEMOTE, down_flags); pcmk__order_resource_actions(member, PCMK_ACTION_DEMOTE, member->parent, PCMK_ACTION_DEMOTED, post_down_flags); // Promote group -> promote member -> group is promoted pcmk__order_resource_actions(member, PCMK_ACTION_PROMOTE, member->parent, PCMK_ACTION_PROMOTED, pcmk__ar_unrunnable_first_blocks |pcmk__ar_first_implies_then |pcmk__ar_first_implies_then_graphed); pcmk__order_resource_actions(member->parent, PCMK_ACTION_PROMOTE, member, PCMK_ACTION_PROMOTE, pcmk__ar_then_implies_first_graphed); } // Stop group -> stop member -> group is stopped pcmk__order_stops(member->parent, member, down_flags); pcmk__order_resource_actions(member, PCMK_ACTION_STOP, member->parent, PCMK_ACTION_STOPPED, post_down_flags); // Start group -> start member -> group is started pcmk__order_starts(member->parent, member, pcmk__ar_then_implies_first_graphed); pcmk__order_resource_actions(member, PCMK_ACTION_START, member->parent, PCMK_ACTION_RUNNING, pcmk__ar_unrunnable_first_blocks |pcmk__ar_first_implies_then |pcmk__ar_first_implies_then_graphed); if (!member_data->ordered) { pcmk__order_starts(member->parent, member, pcmk__ar_first_implies_then |pcmk__ar_unrunnable_first_blocks |pcmk__ar_then_implies_first_graphed); if (member_data->promotable) { pcmk__order_resource_actions(member->parent, PCMK_ACTION_PROMOTE, member, PCMK_ACTION_PROMOTE, pcmk__ar_first_implies_then |pcmk__ar_unrunnable_first_blocks |pcmk__ar_then_implies_first_graphed); } } else if (member_data->previous_member == NULL) { pcmk__order_starts(member->parent, member, pcmk__ar_none); if (member_data->promotable) { pcmk__order_resource_actions(member->parent, PCMK_ACTION_PROMOTE, member, PCMK_ACTION_PROMOTE, pcmk__ar_none); } } else { // Order this member relative to the previous one pcmk__order_starts(member_data->previous_member, member, pcmk__ar_first_implies_then |pcmk__ar_unrunnable_first_blocks); pcmk__order_stops(member, member_data->previous_member, pcmk__ar_ordered|pcmk__ar_intermediate_stop); /* In unusual circumstances (such as adding a new member to the middle * of a group with unmanaged later members), this member may be active * while the previous (new) member is inactive. In this situation, the * usual restart orderings will be irrelevant, so we need to order this * member's stop before the previous member's start. */ if ((member->running_on != NULL) && (member_data->previous_member->running_on == NULL)) { pcmk__order_resource_actions(member, PCMK_ACTION_STOP, member_data->previous_member, PCMK_ACTION_START, pcmk__ar_then_implies_first |pcmk__ar_unrunnable_first_blocks); } if (member_data->promotable) { pcmk__order_resource_actions(member_data->previous_member, PCMK_ACTION_PROMOTE, member, PCMK_ACTION_PROMOTE, pcmk__ar_first_implies_then |pcmk__ar_unrunnable_first_blocks); pcmk__order_resource_actions(member, PCMK_ACTION_DEMOTE, member_data->previous_member, PCMK_ACTION_DEMOTE, pcmk__ar_ordered); } } // Make sure partially active groups shut down in sequence if (member->running_on != NULL) { if (member_data->ordered && (member_data->previous_member != NULL) && (member_data->previous_member->running_on == NULL) && (member_data->last_active != NULL) && (member_data->last_active->running_on != NULL)) { pcmk__order_stops(member, member_data->last_active, pcmk__ar_ordered); } member_data->last_active = member; } member_data->previous_member = member; } /*! * \internal * \brief Create implicit constraints needed for a group resource * * \param[in,out] rsc Group resource to create implicit constraints for */ void pcmk__group_internal_constraints(pcmk_resource_t *rsc) { struct member_data member_data = { false, }; const pcmk_resource_t *top = NULL; CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_group)); /* Order group pseudo-actions relative to each other for restarting: * stop group -> group is stopped -> start group -> group is started */ pcmk__order_resource_actions(rsc, PCMK_ACTION_STOP, rsc, PCMK_ACTION_STOPPED, pcmk__ar_unrunnable_first_blocks); pcmk__order_resource_actions(rsc, PCMK_ACTION_STOPPED, rsc, PCMK_ACTION_START, pcmk__ar_ordered); pcmk__order_resource_actions(rsc, PCMK_ACTION_START, rsc, PCMK_ACTION_RUNNING, pcmk__ar_unrunnable_first_blocks); top = pe__const_top_resource(rsc, false); member_data.ordered = pe__group_flag_is_set(rsc, pcmk__group_ordered); member_data.colocated = pe__group_flag_is_set(rsc, pcmk__group_colocated); member_data.promotable = pcmk_is_set(top->flags, pcmk_rsc_promotable); g_list_foreach(rsc->children, member_internal_constraints, &member_data); } /*! * \internal * \brief Apply a colocation's score to node scores or resource priority * * Given a colocation constraint for a group with some other resource, apply the * score to the dependent's allowed node scores (if we are still placing * resources) or priority (if we are choosing promotable clone instance roles). * * \param[in,out] dependent Dependent group resource in colocation * \param[in] primary Primary resource in colocation * \param[in] colocation Colocation constraint to apply */ static void colocate_group_with(pcmk_resource_t *dependent, const pcmk_resource_t *primary, const pcmk__colocation_t *colocation) { pcmk_resource_t *member = NULL; if (dependent->children == NULL) { return; } pe_rsc_trace(primary, "Processing %s (group %s with %s) for dependent", colocation->id, dependent->id, primary->id); if (pe__group_flag_is_set(dependent, pcmk__group_colocated)) { // Colocate first member (internal colocations will handle the rest) member = (pcmk_resource_t *) dependent->children->data; member->cmds->apply_coloc_score(member, primary, colocation, true); return; } if (colocation->score >= INFINITY) { pcmk__config_err("%s: Cannot perform mandatory colocation between " "non-colocated group and %s", dependent->id, primary->id); return; } // Colocate each member individually for (GList *iter = dependent->children; iter != NULL; iter = iter->next) { member = (pcmk_resource_t *) iter->data; member->cmds->apply_coloc_score(member, primary, colocation, true); } } /*! * \internal * \brief Apply a colocation's score to node scores or resource priority * * Given a colocation constraint for some other resource with a group, apply the * score to the dependent's allowed node scores (if we are still placing * resources) or priority (if we are choosing promotable clone instance roles). * * \param[in,out] dependent Dependent resource in colocation * \param[in] primary Primary group resource in colocation * \param[in] colocation Colocation constraint to apply */ static void colocate_with_group(pcmk_resource_t *dependent, const pcmk_resource_t *primary, const pcmk__colocation_t *colocation) { const pcmk_resource_t *member = NULL; pe_rsc_trace(primary, "Processing colocation %s (%s with group %s) for primary", colocation->id, dependent->id, primary->id); if (pcmk_is_set(primary->flags, pcmk_rsc_unassigned)) { return; } if (pe__group_flag_is_set(primary, pcmk__group_colocated)) { if (colocation->score >= INFINITY) { /* For mandatory colocations, the entire group must be assignable * (and in the specified role if any), so apply the colocation based * on the last member. */ member = pe__last_group_member(primary); } else if (primary->children != NULL) { /* For optional colocations, whether the group is partially or fully * up doesn't matter, so apply the colocation based on the first * member. */ member = (pcmk_resource_t *) primary->children->data; } if (member == NULL) { return; // Nothing to colocate with } member->cmds->apply_coloc_score(dependent, member, colocation, false); return; } if (colocation->score >= INFINITY) { pcmk__config_err("%s: Cannot perform mandatory colocation with" " non-colocated group %s", dependent->id, primary->id); return; } // Colocate dependent with each member individually for (const GList *iter = primary->children; iter != NULL; iter = iter->next) { member = iter->data; member->cmds->apply_coloc_score(dependent, member, colocation, false); } } /*! * \internal * \brief Apply a colocation's score to node scores or resource priority * * Given a colocation constraint, apply its score to the dependent's * allowed node scores (if we are still placing resources) or priority (if * we are choosing promotable clone instance roles). * * \param[in,out] dependent Dependent resource in colocation * \param[in] primary Primary resource in colocation * \param[in] colocation Colocation constraint to apply * \param[in] for_dependent true if called on behalf of dependent */ void pcmk__group_apply_coloc_score(pcmk_resource_t *dependent, const pcmk_resource_t *primary, const pcmk__colocation_t *colocation, bool for_dependent) { CRM_ASSERT((dependent != NULL) && (primary != NULL) && (colocation != NULL)); if (for_dependent) { colocate_group_with(dependent, primary, colocation); } else { // Method should only be called for primitive dependents CRM_ASSERT(dependent->variant == pcmk_rsc_variant_primitive); colocate_with_group(dependent, primary, colocation); } } /*! * \internal * \brief Return action flags for a given group resource action * * \param[in,out] action Group action to get flags for * \param[in] node If not NULL, limit effects to this node * * \return Flags appropriate to \p action on \p node */ uint32_t pcmk__group_action_flags(pcmk_action_t *action, const pcmk_node_t *node) { // Default flags for a group action uint32_t flags = pcmk_action_optional |pcmk_action_runnable |pcmk_action_pseudo; CRM_ASSERT(action != NULL); // Update flags considering each member's own flags for same action for (GList *iter = action->rsc->children; iter != NULL; iter = iter->next) { pcmk_resource_t *member = (pcmk_resource_t *) iter->data; // Check whether member has the same action enum action_tasks task = get_complex_task(member, action->task); const char *task_s = task2text(task); pcmk_action_t *member_action = find_first_action(member->actions, NULL, task_s, node); if (member_action != NULL) { uint32_t member_flags = member->cmds->action_flags(member_action, node); // Group action is mandatory if any member action is if (pcmk_is_set(flags, pcmk_action_optional) && !pcmk_is_set(member_flags, pcmk_action_optional)) { pe_rsc_trace(action->rsc, "%s is mandatory because %s is", action->uuid, member_action->uuid); pe__clear_raw_action_flags(flags, "group action", pcmk_action_optional); pe__clear_action_flags(action, pcmk_action_optional); } // Group action is unrunnable if any member action is if (!pcmk__str_eq(task_s, action->task, pcmk__str_none) && pcmk_is_set(flags, pcmk_action_runnable) && !pcmk_is_set(member_flags, pcmk_action_runnable)) { pe_rsc_trace(action->rsc, "%s is unrunnable because %s is", action->uuid, member_action->uuid); pe__clear_raw_action_flags(flags, "group action", pcmk_action_runnable); pe__clear_action_flags(action, pcmk_action_runnable); } /* Group (pseudo-)actions other than stop or demote are unrunnable * unless every member will do it. */ } else if ((task != pcmk_action_stop) && (task != pcmk_action_demote)) { pe_rsc_trace(action->rsc, "%s is not runnable because %s will not %s", action->uuid, member->id, task_s); pe__clear_raw_action_flags(flags, "group action", pcmk_action_runnable); } } return flags; } /*! * \internal * \brief Update two actions according to an ordering between them * * Given information about an ordering of two actions, update the actions' flags * (and runnable_before members if appropriate) as appropriate for the ordering. * Effects may cascade to other orderings involving the actions as well. * * \param[in,out] first 'First' action in an ordering * \param[in,out] then 'Then' action in an ordering * \param[in] node If not NULL, limit scope of ordering to this node * (only used when interleaving instances) * \param[in] flags Action flags for \p first for ordering purposes * \param[in] filter Action flags to limit scope of certain updates (may * include pcmk_action_optional to affect only * mandatory actions, and pcmk_action_runnable to * affect only runnable actions) * \param[in] type Group of enum pcmk__action_relation_flags to apply * \param[in,out] scheduler Scheduler data * * \return Group of enum pcmk__updated flags indicating what was updated */ uint32_t pcmk__group_update_ordered_actions(pcmk_action_t *first, pcmk_action_t *then, const pcmk_node_t *node, uint32_t flags, uint32_t filter, uint32_t type, pcmk_scheduler_t *scheduler) { uint32_t changed = pcmk__updated_none; // Group method can be called only on behalf of "then" action CRM_ASSERT((first != NULL) && (then != NULL) && (then->rsc != NULL) && (scheduler != NULL)); // Update the actions for the group itself changed |= pcmk__update_ordered_actions(first, then, node, flags, filter, type, scheduler); // Update the actions for each group member for (GList *iter = then->rsc->children; iter != NULL; iter = iter->next) { pcmk_resource_t *member = (pcmk_resource_t *) iter->data; pcmk_action_t *member_action = find_first_action(member->actions, NULL, then->task, node); if (member_action != NULL) { changed |= member->cmds->update_ordered_actions(first, member_action, node, flags, filter, type, scheduler); } } return changed; } /*! * \internal * \brief Apply a location constraint to a group's allowed node scores * * \param[in,out] rsc Group resource to apply constraint to * \param[in,out] location Location constraint to apply */ void pcmk__group_apply_location(pcmk_resource_t *rsc, pcmk__location_t *location) { GList *node_list_orig = NULL; GList *node_list_copy = NULL; bool reset_scores = true; CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_group) && (location != NULL)); node_list_orig = location->nodes; node_list_copy = pcmk__copy_node_list(node_list_orig, true); reset_scores = pe__group_flag_is_set(rsc, pcmk__group_colocated); // Apply the constraint for the group itself (updates node scores) pcmk__apply_location(rsc, location); // Apply the constraint for each member for (GList *iter = rsc->children; iter != NULL; iter = iter->next) { pcmk_resource_t *member = (pcmk_resource_t *) iter->data; member->cmds->apply_location(member, location); if (reset_scores) { /* The first member of colocated groups needs to use the original * node scores, but subsequent members should work on a copy, since * the first member's scores already incorporate theirs. */ reset_scores = false; location->nodes = node_list_copy; } } location->nodes = node_list_orig; g_list_free_full(node_list_copy, free); } // Group implementation of pcmk_assignment_methods_t:colocated_resources() GList * pcmk__group_colocated_resources(const pcmk_resource_t *rsc, const pcmk_resource_t *orig_rsc, GList *colocated_rscs) { const pcmk_resource_t *member = NULL; CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_group)); if (orig_rsc == NULL) { orig_rsc = rsc; } if (pe__group_flag_is_set(rsc, pcmk__group_colocated) || pe_rsc_is_clone(rsc->parent)) { /* This group has colocated members and/or is cloned -- either way, * add every child's colocated resources to the list. The first and last * members will include the group's own colocations. */ colocated_rscs = g_list_prepend(colocated_rscs, (gpointer) rsc); for (const GList *iter = rsc->children; iter != NULL; iter = iter->next) { member = (const pcmk_resource_t *) iter->data; colocated_rscs = member->cmds->colocated_resources(member, orig_rsc, colocated_rscs); } } else if (rsc->children != NULL) { /* This group's members are not colocated, and the group is not cloned, * so just add the group's own colocations to the list. */ colocated_rscs = pcmk__colocated_resources(rsc, orig_rsc, colocated_rscs); } return colocated_rscs; } // Group implementation of pcmk_assignment_methods_t:with_this_colocations() void pcmk__with_group_colocations(const pcmk_resource_t *rsc, const pcmk_resource_t *orig_rsc, GList **list) { CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_group) && (orig_rsc != NULL) && (list != NULL)); // Ignore empty groups if (rsc->children == NULL) { return; } /* "With this" colocations are needed only for the group itself and for its * last member. (Previous members will chain via the group internal * colocations.) */ if ((orig_rsc != rsc) && (orig_rsc != pe__last_group_member(rsc))) { return; } pe_rsc_trace(rsc, "Adding 'with %s' colocations to list for %s", rsc->id, orig_rsc->id); // Add the group's own colocations pcmk__add_with_this_list(list, rsc->rsc_cons_lhs, orig_rsc); // If cloned, add any relevant colocations with the clone if (rsc->parent != NULL) { rsc->parent->cmds->with_this_colocations(rsc->parent, orig_rsc, list); } if (!pe__group_flag_is_set(rsc, pcmk__group_colocated)) { // @COMPAT Non-colocated groups are deprecated return; } // Add explicit colocations with the group's (other) children for (const GList *iter = rsc->children; iter != NULL; iter = iter->next) { const pcmk_resource_t *member = iter->data; if (member != orig_rsc) { member->cmds->with_this_colocations(member, orig_rsc, list); } } } // Group implementation of pcmk_assignment_methods_t:this_with_colocations() void pcmk__group_with_colocations(const pcmk_resource_t *rsc, const pcmk_resource_t *orig_rsc, GList **list) { const pcmk_resource_t *member = NULL; CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_group) && (orig_rsc != NULL) && (list != NULL)); // Ignore empty groups if (rsc->children == NULL) { return; } /* "This with" colocations are normally needed only for the group itself and * for its first member. */ if ((rsc == orig_rsc) || (orig_rsc == (const pcmk_resource_t *) rsc->children->data)) { pe_rsc_trace(rsc, "Adding '%s with' colocations to list for %s", rsc->id, orig_rsc->id); // Add the group's own colocations pcmk__add_this_with_list(list, rsc->rsc_cons, orig_rsc); // If cloned, add any relevant colocations involving the clone if (rsc->parent != NULL) { rsc->parent->cmds->this_with_colocations(rsc->parent, orig_rsc, list); } if (!pe__group_flag_is_set(rsc, pcmk__group_colocated)) { // @COMPAT Non-colocated groups are deprecated return; } // Add explicit colocations involving the group's (other) children for (const GList *iter = rsc->children; iter != NULL; iter = iter->next) { member = iter->data; if (member != orig_rsc) { member->cmds->this_with_colocations(member, orig_rsc, list); } } return; } /* Later group members honor the group's colocations indirectly, due to the * internal group colocations that chain everything from the first member. * However, if an earlier group member is unmanaged, this chaining will not * happen, so the group's mandatory colocations must be explicitly added. */ for (const GList *iter = rsc->children; iter != NULL; iter = iter->next) { member = iter->data; if (orig_rsc == member) { break; // We've seen all earlier members, and none are unmanaged } if (!pcmk_is_set(member->flags, pcmk_rsc_managed)) { crm_trace("Adding mandatory '%s with' colocations to list for " "member %s because earlier member %s is unmanaged", rsc->id, orig_rsc->id, member->id); for (const GList *cons_iter = rsc->rsc_cons; cons_iter != NULL; cons_iter = cons_iter->next) { const pcmk__colocation_t *colocation = NULL; colocation = (const pcmk__colocation_t *) cons_iter->data; if (colocation->score == INFINITY) { pcmk__add_this_with(list, colocation, orig_rsc); } } // @TODO Add mandatory (or all?) clone constraints if cloned break; } } } /*! * \internal * \brief Update nodes with scores of colocated resources' nodes * * Given a table of nodes and a resource, update the nodes' scores with the * scores of the best nodes matching the attribute used for each of the * resource's relevant colocations. * * \param[in,out] source_rsc Group resource whose node scores to add * \param[in] target_rsc Resource on whose behalf to update \p *nodes * \param[in] log_id Resource ID for logs (if \c NULL, use * \p source_rsc ID) * \param[in,out] nodes Nodes to update (set initial contents to \c NULL * to copy allowed nodes from \p source_rsc) * \param[in] colocation Original colocation constraint (used to get * configured primary resource's stickiness, and * to get colocation node attribute; if \c NULL, * source_rsc's own matching node scores will * not be added, and \p *nodes must be \c NULL as * well) * \param[in] factor Incorporate scores multiplied by this factor * \param[in] flags Bitmask of enum pcmk__coloc_select values * * \note \c NULL \p target_rsc, \c NULL \p *nodes, \c NULL \p colocation, and * the \c pcmk__coloc_select_this_with flag are used together (and only by * \c cmp_resources()). * \note The caller remains responsible for freeing \p *nodes. * \note This is the group implementation of * \c pcmk_assignment_methods_t:add_colocated_node_scores(). */ void pcmk__group_add_colocated_node_scores(pcmk_resource_t *source_rsc, const pcmk_resource_t *target_rsc, const char *log_id, GHashTable **nodes, const pcmk__colocation_t *colocation, float factor, uint32_t flags) { pcmk_resource_t *member = NULL; CRM_ASSERT((source_rsc != NULL) && (source_rsc->variant == pcmk_rsc_variant_group) && (nodes != NULL) && ((colocation != NULL) || ((target_rsc == NULL) && (*nodes == NULL)))); if (log_id == NULL) { log_id = source_rsc->id; } // Avoid infinite recursion if (pcmk_is_set(source_rsc->flags, pcmk_rsc_updating_nodes)) { pcmk__rsc_info(source_rsc, "%s: Breaking dependency loop at %s", log_id, source_rsc->id); return; } pe__set_resource_flags(source_rsc, pcmk_rsc_updating_nodes); // Ignore empty groups (only possible with schema validation disabled) if (source_rsc->children == NULL) { return; } /* Refer the operation to the first or last member as appropriate. * * cmp_resources() is the only caller that passes a NULL nodes table, * and is also the only caller using pcmk__coloc_select_this_with. * For "this with" colocations, the last member will recursively incorporate * all the other members' "this with" colocations via the internal group * colocations (and via the first member, the group's own colocations). * * For "with this" colocations, the first member works similarly. */ if (*nodes == NULL) { member = pe__last_group_member(source_rsc); } else { member = source_rsc->children->data; } pe_rsc_trace(source_rsc, "%s: Merging scores from group %s using member %s " "(at %.6f)", log_id, source_rsc->id, member->id, factor); member->cmds->add_colocated_node_scores(member, target_rsc, log_id, nodes, colocation, factor, flags); pe__clear_resource_flags(source_rsc, pcmk_rsc_updating_nodes); } // Group implementation of pcmk_assignment_methods_t:add_utilization() void pcmk__group_add_utilization(const pcmk_resource_t *rsc, const pcmk_resource_t *orig_rsc, GList *all_rscs, GHashTable *utilization) { pcmk_resource_t *member = NULL; CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_group) && (orig_rsc != NULL) && (utilization != NULL)); if (!pcmk_is_set(rsc->flags, pcmk_rsc_unassigned)) { return; } pe_rsc_trace(orig_rsc, "%s: Adding group %s as colocated utilization", orig_rsc->id, rsc->id); if (pe__group_flag_is_set(rsc, pcmk__group_colocated) || pe_rsc_is_clone(rsc->parent)) { // Every group member will be on same node, so sum all members for (GList *iter = rsc->children; iter != NULL; iter = iter->next) { member = (pcmk_resource_t *) iter->data; if (pcmk_is_set(member->flags, pcmk_rsc_unassigned) && (g_list_find(all_rscs, member) == NULL)) { member->cmds->add_utilization(member, orig_rsc, all_rscs, utilization); } } } else if (rsc->children != NULL) { // Just add first member's utilization member = (pcmk_resource_t *) rsc->children->data; if ((member != NULL) && pcmk_is_set(member->flags, pcmk_rsc_unassigned) && (g_list_find(all_rscs, member) == NULL)) { member->cmds->add_utilization(member, orig_rsc, all_rscs, utilization); } } } void pcmk__group_shutdown_lock(pcmk_resource_t *rsc) { CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_group)); for (GList *iter = rsc->children; iter != NULL; iter = iter->next) { pcmk_resource_t *member = (pcmk_resource_t *) iter->data; member->cmds->shutdown_lock(member); } } diff --git a/lib/pacemaker/pcmk_sched_instances.c b/lib/pacemaker/pcmk_sched_instances.c index a700234d46..e57c3d8591 100644 --- a/lib/pacemaker/pcmk_sched_instances.c +++ b/lib/pacemaker/pcmk_sched_instances.c @@ -1,1681 +1,1682 @@ /* * Copyright 2004-2023 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU General Public License version 2 * or later (GPLv2+) WITHOUT ANY WARRANTY. */ /* This file is intended for code usable with both clone instances and bundle * replica containers. */ #include #include #include #include "libpacemaker_private.h" /*! * \internal * \brief Check whether a node is allowed to run an instance * * \param[in] instance Clone instance or bundle container to check * \param[in] node Node to check * \param[in] max_per_node Maximum number of instances allowed to run on a node * * \return true if \p node is allowed to run \p instance, otherwise false */ static bool can_run_instance(const pcmk_resource_t *instance, const pcmk_node_t *node, int max_per_node) { pcmk_node_t *allowed_node = NULL; if (pcmk_is_set(instance->flags, pcmk_rsc_removed)) { pe_rsc_trace(instance, "%s cannot run on %s: orphaned", instance->id, pe__node_name(node)); return false; } if (!pcmk__node_available(node, false, false)) { pe_rsc_trace(instance, "%s cannot run on %s: node cannot run resources", instance->id, pe__node_name(node)); return false; } allowed_node = pcmk__top_allowed_node(instance, node); if (allowed_node == NULL) { crm_warn("%s cannot run on %s: node not allowed", instance->id, pe__node_name(node)); return false; } if (allowed_node->weight < 0) { pe_rsc_trace(instance, "%s cannot run on %s: parent score is %s there", instance->id, pe__node_name(node), pcmk_readable_score(allowed_node->weight)); return false; } if (allowed_node->count >= max_per_node) { pe_rsc_trace(instance, "%s cannot run on %s: node already has %d instance%s", instance->id, pe__node_name(node), max_per_node, pcmk__plural_s(max_per_node)); return false; } pe_rsc_trace(instance, "%s can run on %s (%d already running)", instance->id, pe__node_name(node), allowed_node->count); return true; } /*! * \internal * \brief Ban a clone instance or bundle replica from unavailable allowed nodes * * \param[in,out] instance Clone instance or bundle replica to ban * \param[in] max_per_node Maximum instances allowed to run on a node */ static void ban_unavailable_allowed_nodes(pcmk_resource_t *instance, int max_per_node) { if (instance->allowed_nodes != NULL) { GHashTableIter iter; pcmk_node_t *node = NULL; g_hash_table_iter_init(&iter, instance->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (void **) &node)) { if (!can_run_instance(instance, node, max_per_node)) { pe_rsc_trace(instance, "Banning %s from unavailable node %s", instance->id, pe__node_name(node)); node->weight = -INFINITY; for (GList *child_iter = instance->children; child_iter != NULL; child_iter = child_iter->next) { pcmk_resource_t *child = child_iter->data; pcmk_node_t *child_node = NULL; child_node = g_hash_table_lookup(child->allowed_nodes, node->details->id); if (child_node != NULL) { pe_rsc_trace(instance, "Banning %s child %s " "from unavailable node %s", instance->id, child->id, pe__node_name(node)); child_node->weight = -INFINITY; } } } } } } /*! * \internal * \brief Create a hash table with a single node in it * * \param[in] node Node to copy into new table * * \return Newly created hash table containing a copy of \p node * \note The caller is responsible for freeing the result with * g_hash_table_destroy(). */ static GHashTable * new_node_table(pcmk_node_t *node) { GHashTable *table = pcmk__strkey_table(NULL, free); node = pe__copy_node(node); g_hash_table_insert(table, (gpointer) node->details->id, node); return table; } /*! * \internal * \brief Apply a resource's parent's colocation scores to a node table * * \param[in] rsc Resource whose colocations should be applied * \param[in,out] nodes Node table to apply colocations to */ static void apply_parent_colocations(const pcmk_resource_t *rsc, GHashTable **nodes) { GList *colocations = pcmk__this_with_colocations(rsc); for (const GList *iter = colocations; iter != NULL; iter = iter->next) { const pcmk__colocation_t *colocation = iter->data; pcmk_resource_t *other = colocation->primary; float factor = colocation->score / (float) INFINITY; other->cmds->add_colocated_node_scores(other, rsc, rsc->id, nodes, colocation, factor, pcmk__coloc_select_default); } g_list_free(colocations); colocations = pcmk__with_this_colocations(rsc); for (const GList *iter = colocations; iter != NULL; iter = iter->next) { const pcmk__colocation_t *colocation = iter->data; pcmk_resource_t *other = colocation->dependent; float factor = colocation->score / (float) INFINITY; if (!pcmk__colocation_has_influence(colocation, rsc)) { continue; } other->cmds->add_colocated_node_scores(other, rsc, rsc->id, nodes, colocation, factor, pcmk__coloc_select_nonnegative); } g_list_free(colocations); } /*! * \internal * \brief Compare clone or bundle instances based on colocation scores * * Determine the relative order in which two clone or bundle instances should be * assigned to nodes, considering the scores of colocation constraints directly * or indirectly involving them. * * \param[in] instance1 First instance to compare * \param[in] instance2 Second instance to compare * * \return A negative number if \p instance1 should be assigned first, * a positive number if \p instance2 should be assigned first, * or 0 if assignment order doesn't matter */ static int cmp_instance_by_colocation(const pcmk_resource_t *instance1, const pcmk_resource_t *instance2) { int rc = 0; pcmk_node_t *node1 = NULL; pcmk_node_t *node2 = NULL; pcmk_node_t *current_node1 = pe__current_node(instance1); pcmk_node_t *current_node2 = pe__current_node(instance2); GHashTable *colocated_scores1 = NULL; GHashTable *colocated_scores2 = NULL; CRM_ASSERT((instance1 != NULL) && (instance1->parent != NULL) && (instance2 != NULL) && (instance2->parent != NULL) && (current_node1 != NULL) && (current_node2 != NULL)); // Create node tables initialized with each node colocated_scores1 = new_node_table(current_node1); colocated_scores2 = new_node_table(current_node2); // Apply parental colocations apply_parent_colocations(instance1, &colocated_scores1); apply_parent_colocations(instance2, &colocated_scores2); // Find original nodes again, with scores updated for colocations node1 = g_hash_table_lookup(colocated_scores1, current_node1->details->id); node2 = g_hash_table_lookup(colocated_scores2, current_node2->details->id); // Compare nodes by updated scores if (node1->weight < node2->weight) { crm_trace("Assign %s (%d on %s) after %s (%d on %s)", instance1->id, node1->weight, pe__node_name(node1), instance2->id, node2->weight, pe__node_name(node2)); rc = 1; } else if (node1->weight > node2->weight) { crm_trace("Assign %s (%d on %s) before %s (%d on %s)", instance1->id, node1->weight, pe__node_name(node1), instance2->id, node2->weight, pe__node_name(node2)); rc = -1; } g_hash_table_destroy(colocated_scores1); g_hash_table_destroy(colocated_scores2); return rc; } /*! * \internal * \brief Check whether a resource or any of its children are failed * * \param[in] rsc Resource to check * * \return true if \p rsc or any of its children are failed, otherwise false */ static bool did_fail(const pcmk_resource_t *rsc) { if (pcmk_is_set(rsc->flags, pcmk_rsc_failed)) { return true; } for (GList *iter = rsc->children; iter != NULL; iter = iter->next) { if (did_fail((const pcmk_resource_t *) iter->data)) { return true; } } return false; } /*! * \internal * \brief Check whether a node is allowed to run a resource * * \param[in] rsc Resource to check * \param[in,out] node Node to check (will be set NULL if not allowed) * * \return true if *node is either NULL or allowed for \p rsc, otherwise false */ static bool node_is_allowed(const pcmk_resource_t *rsc, pcmk_node_t **node) { if (*node != NULL) { pcmk_node_t *allowed = g_hash_table_lookup(rsc->allowed_nodes, (*node)->details->id); if ((allowed == NULL) || (allowed->weight < 0)) { pe_rsc_trace(rsc, "%s: current location (%s) is unavailable", rsc->id, pe__node_name(*node)); *node = NULL; return false; } } return true; } /*! * \internal * \brief Compare two clone or bundle instances' instance numbers * * \param[in] a First instance to compare * \param[in] b Second instance to compare * * \return A negative number if \p a's instance number is lower, * a positive number if \p b's instance number is lower, * or 0 if their instance numbers are the same */ gint pcmk__cmp_instance_number(gconstpointer a, gconstpointer b) { const pcmk_resource_t *instance1 = (const pcmk_resource_t *) a; const pcmk_resource_t *instance2 = (const pcmk_resource_t *) b; char *div1 = NULL; char *div2 = NULL; CRM_ASSERT((instance1 != NULL) && (instance2 != NULL)); // Clone numbers are after a colon, bundle numbers after a dash div1 = strrchr(instance1->id, ':'); if (div1 == NULL) { div1 = strrchr(instance1->id, '-'); } div2 = strrchr(instance2->id, ':'); if (div2 == NULL) { div2 = strrchr(instance2->id, '-'); } CRM_ASSERT((div1 != NULL) && (div2 != NULL)); return (gint) (strtol(div1 + 1, NULL, 10) - strtol(div2 + 1, NULL, 10)); } /*! * \internal * \brief Compare clone or bundle instances according to assignment order * * Compare two clone or bundle instances according to the order they should be * assigned to nodes, preferring (in order): * * - Active instance that is less multiply active * - Instance that is not active on a disallowed node * - Instance with higher configured priority * - Active instance whose current node can run resources * - Active instance whose parent is allowed on current node * - Active instance whose current node has fewer other instances * - Active instance * - Instance that isn't failed * - Instance whose colocations result in higher score on current node * - Instance with lower ID in lexicographic order * * \param[in] a First instance to compare * \param[in] b Second instance to compare * * \return A negative number if \p a should be assigned first, * a positive number if \p b should be assigned first, * or 0 if assignment order doesn't matter */ gint pcmk__cmp_instance(gconstpointer a, gconstpointer b) { int rc = 0; pcmk_node_t *node1 = NULL; pcmk_node_t *node2 = NULL; unsigned int nnodes1 = 0; unsigned int nnodes2 = 0; bool can1 = true; bool can2 = true; const pcmk_resource_t *instance1 = (const pcmk_resource_t *) a; const pcmk_resource_t *instance2 = (const pcmk_resource_t *) b; CRM_ASSERT((instance1 != NULL) && (instance2 != NULL)); node1 = instance1->fns->active_node(instance1, &nnodes1, NULL); node2 = instance2->fns->active_node(instance2, &nnodes2, NULL); /* If both instances are running and at least one is multiply * active, prefer instance that's running on fewer nodes. */ if ((nnodes1 > 0) && (nnodes2 > 0)) { if (nnodes1 < nnodes2) { crm_trace("Assign %s (active on %d) before %s (active on %d): " "less multiply active", instance1->id, nnodes1, instance2->id, nnodes2); return -1; } else if (nnodes1 > nnodes2) { crm_trace("Assign %s (active on %d) after %s (active on %d): " "more multiply active", instance1->id, nnodes1, instance2->id, nnodes2); return 1; } } /* An instance that is either inactive or active on an allowed node is * preferred over an instance that is active on a no-longer-allowed node. */ can1 = node_is_allowed(instance1, &node1); can2 = node_is_allowed(instance2, &node2); if (can1 && !can2) { crm_trace("Assign %s before %s: not active on a disallowed node", instance1->id, instance2->id); return -1; } else if (!can1 && can2) { crm_trace("Assign %s after %s: active on a disallowed node", instance1->id, instance2->id); return 1; } // Prefer instance with higher configured priority if (instance1->priority > instance2->priority) { crm_trace("Assign %s before %s: priority (%d > %d)", instance1->id, instance2->id, instance1->priority, instance2->priority); return -1; } else if (instance1->priority < instance2->priority) { crm_trace("Assign %s after %s: priority (%d < %d)", instance1->id, instance2->id, instance1->priority, instance2->priority); return 1; } // Prefer active instance if ((node1 == NULL) && (node2 == NULL)) { crm_trace("No assignment preference for %s vs. %s: inactive", instance1->id, instance2->id); return 0; } else if (node1 == NULL) { crm_trace("Assign %s after %s: active", instance1->id, instance2->id); return 1; } else if (node2 == NULL) { crm_trace("Assign %s before %s: active", instance1->id, instance2->id); return -1; } // Prefer instance whose current node can run resources can1 = pcmk__node_available(node1, false, false); can2 = pcmk__node_available(node2, false, false); if (can1 && !can2) { crm_trace("Assign %s before %s: current node can run resources", instance1->id, instance2->id); return -1; } else if (!can1 && can2) { crm_trace("Assign %s after %s: current node can't run resources", instance1->id, instance2->id); return 1; } // Prefer instance whose parent is allowed to run on instance's current node node1 = pcmk__top_allowed_node(instance1, node1); node2 = pcmk__top_allowed_node(instance2, node2); if ((node1 == NULL) && (node2 == NULL)) { crm_trace("No assignment preference for %s vs. %s: " "parent not allowed on either instance's current node", instance1->id, instance2->id); return 0; } else if (node1 == NULL) { crm_trace("Assign %s after %s: parent not allowed on current node", instance1->id, instance2->id); return 1; } else if (node2 == NULL) { crm_trace("Assign %s before %s: parent allowed on current node", instance1->id, instance2->id); return -1; } // Prefer instance whose current node is running fewer other instances if (node1->count < node2->count) { crm_trace("Assign %s before %s: fewer active instances on current node", instance1->id, instance2->id); return -1; } else if (node1->count > node2->count) { crm_trace("Assign %s after %s: more active instances on current node", instance1->id, instance2->id); return 1; } // Prefer instance that isn't failed can1 = did_fail(instance1); can2 = did_fail(instance2); if (!can1 && can2) { crm_trace("Assign %s before %s: not failed", instance1->id, instance2->id); return -1; } else if (can1 && !can2) { crm_trace("Assign %s after %s: failed", instance1->id, instance2->id); return 1; } // Prefer instance with higher cumulative colocation score on current node rc = cmp_instance_by_colocation(instance1, instance2); if (rc != 0) { return rc; } // Prefer instance with lower instance number rc = pcmk__cmp_instance_number(instance1, instance2); if (rc < 0) { crm_trace("Assign %s before %s: instance number", instance1->id, instance2->id); } else if (rc > 0) { crm_trace("Assign %s after %s: instance number", instance1->id, instance2->id); } else { crm_trace("No assignment preference for %s vs. %s", instance1->id, instance2->id); } return rc; } /*! * \internal * \brief Increment the parent's instance count after assigning an instance * * An instance's parent tracks how many instances have been assigned to each * node via its pcmk_node_t:count member. After assigning an instance to a node, * find the corresponding node in the parent's allowed table and increment it. * * \param[in,out] instance Instance whose parent to update * \param[in] assigned_to Node to which the instance was assigned */ static void increment_parent_count(pcmk_resource_t *instance, const pcmk_node_t *assigned_to) { pcmk_node_t *allowed = NULL; if (assigned_to == NULL) { return; } allowed = pcmk__top_allowed_node(instance, assigned_to); if (allowed == NULL) { /* The instance is allowed on the node, but its parent isn't. This * shouldn't be possible if the resource is managed, and we won't be * able to limit the number of instances assigned to the node. */ CRM_LOG_ASSERT(!pcmk_is_set(instance->flags, pcmk_rsc_managed)); } else { allowed->count++; } } /*! * \internal * \brief Assign an instance to a node * * \param[in,out] instance Clone instance or bundle replica container * \param[in] prefer If not NULL, attempt early assignment to this * node, if still the best choice; otherwise, * perform final assignment * \param[in] max_per_node Assign at most this many instances to one node * * \return Node to which \p instance is assigned */ static const pcmk_node_t * assign_instance(pcmk_resource_t *instance, const pcmk_node_t *prefer, int max_per_node) { pcmk_node_t *chosen = NULL; pe_rsc_trace(instance, "Assigning %s (preferring %s)", instance->id, ((prefer == NULL)? "no node" : prefer->details->uname)); if (pcmk_is_set(instance->flags, pcmk_rsc_assigning)) { - pe_rsc_debug(instance, - "Assignment loop detected involving %s colocations", - instance->id); + pcmk__rsc_debug(instance, + "Assignment loop detected involving %s colocations", + instance->id); return NULL; } ban_unavailable_allowed_nodes(instance, max_per_node); // Failed early assignments are reversible (stop_if_fail=false) chosen = instance->cmds->assign(instance, prefer, (prefer == NULL)); increment_parent_count(instance, chosen); return chosen; } /*! * \internal * \brief Try to assign an instance to its current node early * * \param[in] rsc Clone or bundle being assigned (for logs only) * \param[in] instance Clone instance or bundle replica container * \param[in] current Instance's current node * \param[in] max_per_node Maximum number of instances per node * \param[in] available Number of instances still available for assignment * * \return \c true if \p instance was successfully assigned to its current node, * or \c false otherwise */ static bool assign_instance_early(const pcmk_resource_t *rsc, pcmk_resource_t *instance, const pcmk_node_t *current, int max_per_node, int available) { const pcmk_node_t *chosen = NULL; int reserved = 0; pcmk_resource_t *parent = instance->parent; GHashTable *allowed_orig = NULL; GHashTable *allowed_orig_parent = parent->allowed_nodes; const pcmk_node_t *allowed_node = NULL; pe_rsc_trace(instance, "Trying to assign %s to its current node %s", instance->id, pe__node_name(current)); allowed_node = g_hash_table_lookup(instance->allowed_nodes, current->details->id); if (!pcmk__node_available(allowed_node, true, false)) { pcmk__rsc_info(instance, "Not assigning %s to current node %s: unavailable", instance->id, pe__node_name(current)); return false; } /* On each iteration, if instance gets assigned to a node other than its * current one, we reserve one instance for the chosen node, unassign * instance, restore instance's original node tables, and try again. This * way, instances are proportionally assigned to nodes based on preferences, * but shuffling of specific instances is minimized. If a node will be * assigned instances at all, it preferentially receives instances that are * currently active there. * * parent->allowed_nodes tracks the number of instances assigned to each * node. If a node already has max_per_node instances assigned, * ban_unavailable_allowed_nodes() marks it as unavailable. * * In the end, we restore the original parent->allowed_nodes to undo the * changes to counts during tentative assignments. If we successfully * assigned instance to its current node, we increment that node's counter. */ // Back up the allowed node tables of instance and its children recursively pcmk__copy_node_tables(instance, &allowed_orig); // Update instances-per-node counts in a scratch table parent->allowed_nodes = pcmk__copy_node_table(parent->allowed_nodes); while (reserved < available) { chosen = assign_instance(instance, current, max_per_node); if (pe__same_node(chosen, current)) { // Successfully assigned to current node break; } // Assignment updates scores, so restore to original state - pe_rsc_debug(instance, "Rolling back node scores for %s", instance->id); + pcmk__rsc_debug(instance, "Rolling back node scores for %s", + instance->id); pcmk__restore_node_tables(instance, allowed_orig); if (chosen == NULL) { // Assignment failed, so give up pcmk__rsc_info(instance, "Not assigning %s to current node %s: unavailable", instance->id, pe__node_name(current)); pe__set_resource_flags(instance, pcmk_rsc_unassigned); break; } // We prefer more strongly to assign an instance to the chosen node - pe_rsc_debug(instance, - "Not assigning %s to current node %s: %s is better", - instance->id, pe__node_name(current), - pe__node_name(chosen)); + pcmk__rsc_debug(instance, + "Not assigning %s to current node %s: %s is better", + instance->id, pe__node_name(current), + pe__node_name(chosen)); // Reserve one instance for the chosen node and try again if (++reserved >= available) { pcmk__rsc_info(instance, "Not assigning %s to current node %s: " "other assignments are more important", instance->id, pe__node_name(current)); } else { - pe_rsc_debug(instance, - "Reserved an instance of %s for %s. Retrying " - "assignment of %s to %s", - rsc->id, pe__node_name(chosen), instance->id, - pe__node_name(current)); + pcmk__rsc_debug(instance, + "Reserved an instance of %s for %s. Retrying " + "assignment of %s to %s", + rsc->id, pe__node_name(chosen), instance->id, + pe__node_name(current)); } // Clear this assignment (frees chosen); leave instance counts in parent pcmk__unassign_resource(instance); chosen = NULL; } g_hash_table_destroy(allowed_orig); // Restore original instances-per-node counts g_hash_table_destroy(parent->allowed_nodes); parent->allowed_nodes = allowed_orig_parent; if (chosen == NULL) { // Couldn't assign instance to current node return false; } pe_rsc_trace(instance, "Assigned %s to current node %s", instance->id, pe__node_name(current)); increment_parent_count(instance, chosen); return true; } /*! * \internal * \brief Reset the node counts of a resource's allowed nodes to zero * * \param[in,out] rsc Resource to reset * * \return Number of nodes that are available to run resources */ static unsigned int reset_allowed_node_counts(pcmk_resource_t *rsc) { unsigned int available_nodes = 0; pcmk_node_t *node = NULL; GHashTableIter iter; g_hash_table_iter_init(&iter, rsc->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (gpointer *) &node)) { node->count = 0; if (pcmk__node_available(node, false, false)) { available_nodes++; } } return available_nodes; } /*! * \internal * \brief Check whether an instance has a preferred node * * \param[in] instance Clone instance or bundle replica container * \param[in] optimal_per_node Optimal number of instances per node * * \return Instance's current node if still available, otherwise NULL */ static const pcmk_node_t * preferred_node(const pcmk_resource_t *instance, int optimal_per_node) { const pcmk_node_t *node = NULL; const pcmk_node_t *parent_node = NULL; // Check whether instance is active, healthy, and not yet assigned if ((instance->running_on == NULL) || !pcmk_is_set(instance->flags, pcmk_rsc_unassigned) || pcmk_is_set(instance->flags, pcmk_rsc_failed)) { return NULL; } // Check whether instance's current node can run resources node = pe__current_node(instance); if (!pcmk__node_available(node, true, false)) { pe_rsc_trace(instance, "Not assigning %s to %s early (unavailable)", instance->id, pe__node_name(node)); return NULL; } // Check whether node already has optimal number of instances assigned parent_node = pcmk__top_allowed_node(instance, node); if ((parent_node != NULL) && (parent_node->count >= optimal_per_node)) { pe_rsc_trace(instance, "Not assigning %s to %s early " "(optimal instances already assigned)", instance->id, pe__node_name(node)); return NULL; } return node; } /*! * \internal * \brief Assign collective instances to nodes * * \param[in,out] collective Clone or bundle resource being assigned * \param[in,out] instances List of clone instances or bundle containers * \param[in] max_total Maximum instances to assign in total * \param[in] max_per_node Maximum instances to assign to any one node */ void pcmk__assign_instances(pcmk_resource_t *collective, GList *instances, int max_total, int max_per_node) { // Reuse node count to track number of assigned instances unsigned int available_nodes = reset_allowed_node_counts(collective); int optimal_per_node = 0; int assigned = 0; GList *iter = NULL; pcmk_resource_t *instance = NULL; const pcmk_node_t *current = NULL; if (available_nodes > 0) { optimal_per_node = max_total / available_nodes; } if (optimal_per_node < 1) { optimal_per_node = 1; } - pe_rsc_debug(collective, - "Assigning up to %d %s instance%s to up to %u node%s " - "(at most %d per host, %d optimal)", - max_total, collective->id, pcmk__plural_s(max_total), - available_nodes, pcmk__plural_s(available_nodes), - max_per_node, optimal_per_node); + pcmk__rsc_debug(collective, + "Assigning up to %d %s instance%s to up to %u node%s " + "(at most %d per host, %d optimal)", + max_total, collective->id, pcmk__plural_s(max_total), + available_nodes, pcmk__plural_s(available_nodes), + max_per_node, optimal_per_node); // Assign as many instances as possible to their current location for (iter = instances; (iter != NULL) && (assigned < max_total); iter = iter->next) { int available = max_total - assigned; instance = iter->data; if (!pcmk_is_set(instance->flags, pcmk_rsc_unassigned)) { continue; // Already assigned } current = preferred_node(instance, optimal_per_node); if ((current != NULL) && assign_instance_early(collective, instance, current, max_per_node, available)) { assigned++; } } pe_rsc_trace(collective, "Assigned %d of %d instance%s to current node", assigned, max_total, pcmk__plural_s(max_total)); for (iter = instances; iter != NULL; iter = iter->next) { instance = (pcmk_resource_t *) iter->data; if (!pcmk_is_set(instance->flags, pcmk_rsc_unassigned)) { continue; // Already assigned } if (instance->running_on != NULL) { current = pe__current_node(instance); if (pcmk__top_allowed_node(instance, current) == NULL) { const char *unmanaged = ""; if (!pcmk_is_set(instance->flags, pcmk_rsc_managed)) { unmanaged = "Unmanaged resource "; } crm_notice("%s%s is running on %s which is no longer allowed", unmanaged, instance->id, pe__node_name(current)); } } if (assigned >= max_total) { - pe_rsc_debug(collective, - "Not assigning %s because maximum %d instances " - "already assigned", - instance->id, max_total); + pcmk__rsc_debug(collective, + "Not assigning %s because maximum %d instances " + "already assigned", + instance->id, max_total); resource_location(instance, NULL, -INFINITY, "collective_limit_reached", collective->cluster); } else if (assign_instance(instance, NULL, max_per_node) != NULL) { assigned++; } } - pe_rsc_debug(collective, "Assigned %d of %d possible instance%s of %s", - assigned, max_total, pcmk__plural_s(max_total), - collective->id); + pcmk__rsc_debug(collective, "Assigned %d of %d possible instance%s of %s", + assigned, max_total, pcmk__plural_s(max_total), + collective->id); } enum instance_state { instance_starting = (1 << 0), instance_stopping = (1 << 1), /* This indicates that some instance is restarting. It's not the same as * instance_starting|instance_stopping, which would indicate that some * instance is starting, and some instance (not necessarily the same one) is * stopping. */ instance_restarting = (1 << 2), instance_active = (1 << 3), instance_all = instance_starting|instance_stopping |instance_restarting|instance_active, }; /*! * \internal * \brief Check whether an instance is active, starting, and/or stopping * * \param[in] instance Clone instance or bundle replica container * \param[in,out] state Whether any instance is starting, stopping, etc. */ static void check_instance_state(const pcmk_resource_t *instance, uint32_t *state) { const GList *iter = NULL; uint32_t instance_state = 0; // State of just this instance // No need to check further if all conditions have already been detected if (pcmk_all_flags_set(*state, instance_all)) { return; } // If instance is a collective (a cloned group), check its children instead if (instance->variant > pcmk_rsc_variant_primitive) { for (iter = instance->children; (iter != NULL) && !pcmk_all_flags_set(*state, instance_all); iter = iter->next) { check_instance_state((const pcmk_resource_t *) iter->data, state); } return; } // If we get here, instance is a primitive if (instance->running_on != NULL) { instance_state |= instance_active; } // Check each of the instance's actions for runnable start or stop for (iter = instance->actions; (iter != NULL) && !pcmk_all_flags_set(instance_state, instance_starting |instance_stopping); iter = iter->next) { const pcmk_action_t *action = (const pcmk_action_t *) iter->data; const bool optional = pcmk_is_set(action->flags, pcmk_action_optional); if (pcmk__str_eq(PCMK_ACTION_START, action->task, pcmk__str_none)) { if (!optional && pcmk_is_set(action->flags, pcmk_action_runnable)) { pe_rsc_trace(instance, "Instance is starting due to %s", action->uuid); instance_state |= instance_starting; } else { pe_rsc_trace(instance, "%s doesn't affect %s state (%s)", action->uuid, instance->id, (optional? "optional" : "unrunnable")); } } else if (pcmk__str_eq(PCMK_ACTION_STOP, action->task, pcmk__str_none)) { /* Only stop actions can be pseudo-actions for primitives. That * indicates that the node they are on is being fenced, so the stop * is implied rather than actually executed. */ if (!optional && pcmk_any_flags_set(action->flags, pcmk_action_pseudo |pcmk_action_runnable)) { pe_rsc_trace(instance, "Instance is stopping due to %s", action->uuid); instance_state |= instance_stopping; } else { pe_rsc_trace(instance, "%s doesn't affect %s state (%s)", action->uuid, instance->id, (optional? "optional" : "unrunnable")); } } } if (pcmk_all_flags_set(instance_state, instance_starting|instance_stopping)) { instance_state |= instance_restarting; } *state |= instance_state; } /*! * \internal * \brief Create actions for collective resource instances * * \param[in,out] collective Clone or bundle resource to create actions for * \param[in,out] instances List of clone instances or bundle containers */ void pcmk__create_instance_actions(pcmk_resource_t *collective, GList *instances) { uint32_t state = 0; pcmk_action_t *stop = NULL; pcmk_action_t *stopped = NULL; pcmk_action_t *start = NULL; pcmk_action_t *started = NULL; pe_rsc_trace(collective, "Creating collective instance actions for %s", collective->id); // Create actions for each instance appropriate to its variant for (GList *iter = instances; iter != NULL; iter = iter->next) { pcmk_resource_t *instance = (pcmk_resource_t *) iter->data; instance->cmds->create_actions(instance); check_instance_state(instance, &state); } // Create pseudo-actions for rsc start and started start = pe__new_rsc_pseudo_action(collective, PCMK_ACTION_START, !pcmk_is_set(state, instance_starting), true); started = pe__new_rsc_pseudo_action(collective, PCMK_ACTION_RUNNING, !pcmk_is_set(state, instance_starting), false); started->priority = INFINITY; if (pcmk_any_flags_set(state, instance_active|instance_starting)) { pe__set_action_flags(started, pcmk_action_runnable); } // Create pseudo-actions for rsc stop and stopped stop = pe__new_rsc_pseudo_action(collective, PCMK_ACTION_STOP, !pcmk_is_set(state, instance_stopping), true); stopped = pe__new_rsc_pseudo_action(collective, PCMK_ACTION_STOPPED, !pcmk_is_set(state, instance_stopping), true); stopped->priority = INFINITY; if (!pcmk_is_set(state, instance_restarting)) { pe__set_action_flags(stop, pcmk_action_migratable); } if (collective->variant == pcmk_rsc_variant_clone) { pe__create_clone_notif_pseudo_ops(collective, start, started, stop, stopped); } } /*! * \internal * \brief Get a list of clone instances or bundle replica containers * * \param[in] rsc Clone or bundle resource * * \return Clone instances if \p rsc is a clone, or a newly created list of * \p rsc's replica containers if \p rsc is a bundle * \note The caller must call free_instance_list() on the result when the list * is no longer needed. */ static inline GList * get_instance_list(const pcmk_resource_t *rsc) { if (rsc->variant == pcmk_rsc_variant_bundle) { return pe__bundle_containers(rsc); } else { return rsc->children; } } /*! * \internal * \brief Free any memory created by get_instance_list() * * \param[in] rsc Clone or bundle resource passed to get_instance_list() * \param[in,out] list Return value of get_instance_list() for \p rsc */ static inline void free_instance_list(const pcmk_resource_t *rsc, GList *list) { if (list != rsc->children) { g_list_free(list); } } /*! * \internal * \brief Check whether an instance is compatible with a role and node * * \param[in] instance Clone instance or bundle replica container * \param[in] node Instance must match this node * \param[in] role If not pcmk_role_unknown, instance must match this role * \param[in] current If true, compare instance's original node and role, * otherwise compare assigned next node and role * * \return true if \p instance is compatible with \p node and \p role, * otherwise false */ bool pcmk__instance_matches(const pcmk_resource_t *instance, const pcmk_node_t *node, enum rsc_role_e role, bool current) { pcmk_node_t *instance_node = NULL; CRM_CHECK((instance != NULL) && (node != NULL), return false); if ((role != pcmk_role_unknown) && (role != instance->fns->state(instance, current))) { pe_rsc_trace(instance, "%s is not a compatible instance (role is not %s)", instance->id, role2text(role)); return false; } if (!is_set_recursive(instance, pcmk_rsc_blocked, true)) { // We only want instances that haven't failed instance_node = instance->fns->location(instance, NULL, current); } if (instance_node == NULL) { pe_rsc_trace(instance, "%s is not a compatible instance (not assigned to a node)", instance->id); return false; } if (!pe__same_node(instance_node, node)) { pe_rsc_trace(instance, "%s is not a compatible instance (assigned to %s not %s)", instance->id, pe__node_name(instance_node), pe__node_name(node)); return false; } return true; } /*! * \internal * \brief Find an instance that matches a given resource by node and role * * \param[in] match_rsc Resource that instance must match (for logging only) * \param[in] rsc Clone or bundle resource to check for matching instance * \param[in] node Instance must match this node * \param[in] role If not pcmk_role_unknown, instance must match this role * \param[in] current If true, compare instance's original node and role, * otherwise compare assigned next node and role * * \return \p rsc instance matching \p node and \p role if any, otherwise NULL */ static pcmk_resource_t * find_compatible_instance_on_node(const pcmk_resource_t *match_rsc, const pcmk_resource_t *rsc, const pcmk_node_t *node, enum rsc_role_e role, bool current) { GList *instances = NULL; instances = get_instance_list(rsc); for (GList *iter = instances; iter != NULL; iter = iter->next) { pcmk_resource_t *instance = (pcmk_resource_t *) iter->data; if (pcmk__instance_matches(instance, node, role, current)) { pe_rsc_trace(match_rsc, "Found %s %s instance %s compatible with %s on %s", role == pcmk_role_unknown? "matching" : role2text(role), rsc->id, instance->id, match_rsc->id, pe__node_name(node)); free_instance_list(rsc, instances); // Only frees list, not contents return instance; } } free_instance_list(rsc, instances); pe_rsc_trace(match_rsc, "No %s %s instance found compatible with %s on %s", ((role == pcmk_role_unknown)? "matching" : role2text(role)), rsc->id, match_rsc->id, pe__node_name(node)); return NULL; } /*! * \internal * \brief Find a clone instance or bundle container compatible with a resource * * \param[in] match_rsc Resource that instance must match * \param[in] rsc Clone or bundle resource to check for matching instance * \param[in] role If not pcmk_role_unknown, instance must match this role * \param[in] current If true, compare instance's original node and role, * otherwise compare assigned next node and role * * \return Compatible (by \p role and \p match_rsc location) instance of \p rsc * if any, otherwise NULL */ pcmk_resource_t * pcmk__find_compatible_instance(const pcmk_resource_t *match_rsc, const pcmk_resource_t *rsc, enum rsc_role_e role, bool current) { pcmk_resource_t *instance = NULL; GList *nodes = NULL; const pcmk_node_t *node = NULL; // If match_rsc has a node, check only that node node = match_rsc->fns->location(match_rsc, NULL, current); if (node != NULL) { return find_compatible_instance_on_node(match_rsc, rsc, node, role, current); } // Otherwise check for an instance matching any of match_rsc's allowed nodes nodes = pcmk__sort_nodes(g_hash_table_get_values(match_rsc->allowed_nodes), NULL); for (GList *iter = nodes; (iter != NULL) && (instance == NULL); iter = iter->next) { instance = find_compatible_instance_on_node(match_rsc, rsc, (pcmk_node_t *) iter->data, role, current); } if (instance == NULL) { - pe_rsc_debug(rsc, "No %s instance found compatible with %s", - rsc->id, match_rsc->id); + pcmk__rsc_debug(rsc, "No %s instance found compatible with %s", + rsc->id, match_rsc->id); } g_list_free(nodes); return instance; } /*! * \internal * \brief Unassign an instance if mandatory ordering has no interleave match * * \param[in] first 'First' action in an ordering * \param[in] then 'Then' action in an ordering * \param[in,out] then_instance 'Then' instance that has no interleave match * \param[in] type Group of enum pcmk__action_relation_flags * \param[in] current If true, "then" action is stopped or demoted * * \return true if \p then_instance was unassigned, otherwise false */ static bool unassign_if_mandatory(const pcmk_action_t *first, const pcmk_action_t *then, pcmk_resource_t *then_instance, uint32_t type, bool current) { // Allow "then" instance to go down even without an interleave match if (current) { pe_rsc_trace(then->rsc, "%s has no instance to order before stopping " "or demoting %s", first->rsc->id, then_instance->id); /* If the "first" action must be runnable, but there is no "first" * instance, the "then" instance must not be allowed to come up. */ } else if (pcmk_any_flags_set(type, pcmk__ar_unrunnable_first_blocks |pcmk__ar_first_implies_then)) { pcmk__rsc_info(then->rsc, "Inhibiting %s from being active " "because there is no %s instance to interleave", then_instance->id, first->rsc->id); return pcmk__assign_resource(then_instance, NULL, true, true); } return false; } /*! * \internal * \brief Find first matching action for a clone instance or bundle container * * \param[in] action Action in an interleaved ordering * \param[in] instance Clone instance or bundle container being interleaved * \param[in] action_name Action to look for * \param[in] node If not NULL, require action to be on this node * \param[in] for_first If true, \p instance is the 'first' resource in the * ordering, otherwise it is the 'then' resource * * \return First action for \p instance (or in some cases if \p instance is a * bundle container, its containerized resource) that matches * \p action_name and \p node if any, otherwise NULL */ static pcmk_action_t * find_instance_action(const pcmk_action_t *action, const pcmk_resource_t *instance, const char *action_name, const pcmk_node_t *node, bool for_first) { const pcmk_resource_t *rsc = NULL; pcmk_action_t *matching_action = NULL; /* If instance is a bundle container, sometimes we should interleave the * action for the container itself, and sometimes for the containerized * resource. * * For example, given "start bundle A then bundle B", B likely requires the * service inside A's container to be active, rather than just the * container, so we should interleave the action for A's containerized * resource. On the other hand, it's possible B's container itself requires * something from A, so we should interleave the action for B's container. * * Essentially, for 'first', we should use the containerized resource for * everything except stop, and for 'then', we should use the container for * everything except promote and demote (which can only be performed on the * containerized resource). */ if ((for_first && !pcmk__str_any_of(action->task, PCMK_ACTION_STOP, PCMK_ACTION_STOPPED, NULL)) || (!for_first && pcmk__str_any_of(action->task, PCMK_ACTION_PROMOTE, PCMK_ACTION_PROMOTED, PCMK_ACTION_DEMOTE, PCMK_ACTION_DEMOTED, NULL))) { rsc = pe__get_rsc_in_container(instance); } if (rsc == NULL) { rsc = instance; // No containerized resource, use instance itself } else { node = NULL; // Containerized actions are on bundle-created guest } matching_action = find_first_action(rsc->actions, NULL, action_name, node); if (matching_action != NULL) { return matching_action; } if (pcmk_is_set(instance->flags, pcmk_rsc_removed) || pcmk__str_any_of(action_name, PCMK_ACTION_STOP, PCMK_ACTION_DEMOTE, NULL)) { crm_trace("No %s action found for %s%s", action_name, pcmk_is_set(instance->flags, pcmk_rsc_removed)? "orphan " : "", instance->id); } else { crm_err("No %s action found for %s to interleave (bug?)", action_name, instance->id); } return NULL; } /*! * \internal * \brief Get the original action name of a bundle or clone action * * Given an action for a bundle or clone, get the original action name, * mapping notify to the action being notified, and if the instances are * primitives, mapping completion actions to the action that was completed * (for example, stopped to stop). * * \param[in] action Clone or bundle action to check * * \return Original action name for \p action */ static const char * orig_action_name(const pcmk_action_t *action) { // Any instance will do const pcmk_resource_t *instance = action->rsc->children->data; char *action_type = NULL; const char *action_name = action->task; enum action_tasks orig_task = pcmk_action_unspecified; if (pcmk__strcase_any_of(action->task, PCMK_ACTION_NOTIFY, PCMK_ACTION_NOTIFIED, NULL)) { // action->uuid is RSC_(confirmed-){pre,post}_notify_ACTION_INTERVAL CRM_CHECK(parse_op_key(action->uuid, NULL, &action_type, NULL), return task2text(pcmk_action_unspecified)); action_name = strstr(action_type, "_notify_"); CRM_CHECK(action_name != NULL, return task2text(pcmk_action_unspecified)); action_name += strlen("_notify_"); } orig_task = get_complex_task(instance, action_name); free(action_type); return task2text(orig_task); } /*! * \internal * \brief Update two interleaved actions according to an ordering between them * * Given information about an ordering of two interleaved actions, update the * actions' flags (and runnable_before members if appropriate) as appropriate * for the ordering. Effects may cascade to other orderings involving the * actions as well. * * \param[in,out] first 'First' action in an ordering * \param[in,out] then 'Then' action in an ordering * \param[in] node If not NULL, limit scope of ordering to this node * \param[in] filter Action flags to limit scope of certain updates (may * include pcmk_action_optional to affect only * mandatory actions, and pcmk_action_runnable to * affect only runnable actions) * \param[in] type Group of enum pcmk__action_relation_flags to apply * * \return Group of enum pcmk__updated flags indicating what was updated */ static uint32_t update_interleaved_actions(pcmk_action_t *first, pcmk_action_t *then, const pcmk_node_t *node, uint32_t filter, uint32_t type) { GList *instances = NULL; uint32_t changed = pcmk__updated_none; const char *orig_first_task = orig_action_name(first); // Stops and demotes must be interleaved with instance on current node bool current = pcmk__ends_with(first->uuid, "_" PCMK_ACTION_STOPPED "_0") || pcmk__ends_with(first->uuid, "_" PCMK_ACTION_DEMOTED "_0"); // Update the specified actions for each "then" instance individually instances = get_instance_list(then->rsc); for (GList *iter = instances; iter != NULL; iter = iter->next) { pcmk_resource_t *first_instance = NULL; pcmk_resource_t *then_instance = iter->data; pcmk_action_t *first_action = NULL; pcmk_action_t *then_action = NULL; // Find a "first" instance to interleave with this "then" instance first_instance = pcmk__find_compatible_instance(then_instance, first->rsc, pcmk_role_unknown, current); if (first_instance == NULL) { // No instance can be interleaved if (unassign_if_mandatory(first, then, then_instance, type, current)) { pcmk__set_updated_flags(changed, first, pcmk__updated_then); } continue; } first_action = find_instance_action(first, first_instance, orig_first_task, node, true); if (first_action == NULL) { continue; } then_action = find_instance_action(then, then_instance, then->task, node, false); if (then_action == NULL) { continue; } if (order_actions(first_action, then_action, type)) { pcmk__set_updated_flags(changed, first, pcmk__updated_first|pcmk__updated_then); } changed |= then_instance->cmds->update_ordered_actions( first_action, then_action, node, first_instance->cmds->action_flags(first_action, node), filter, type, then->rsc->cluster); } free_instance_list(then->rsc, instances); return changed; } /*! * \internal * \brief Check whether two actions in an ordering can be interleaved * * \param[in] first 'First' action in the ordering * \param[in] then 'Then' action in the ordering * * \return true if \p first and \p then can be interleaved, otherwise false */ static bool can_interleave_actions(const pcmk_action_t *first, const pcmk_action_t *then) { bool interleave = false; pcmk_resource_t *rsc = NULL; if ((first->rsc == NULL) || (then->rsc == NULL)) { crm_trace("Not interleaving %s with %s: not resource actions", first->uuid, then->uuid); return false; } if (first->rsc == then->rsc) { crm_trace("Not interleaving %s with %s: same resource", first->uuid, then->uuid); return false; } if ((first->rsc->variant < pcmk_rsc_variant_clone) || (then->rsc->variant < pcmk_rsc_variant_clone)) { crm_trace("Not interleaving %s with %s: not clones or bundles", first->uuid, then->uuid); return false; } if (pcmk__ends_with(then->uuid, "_stop_0") || pcmk__ends_with(then->uuid, "_demote_0")) { rsc = first->rsc; } else { rsc = then->rsc; } interleave = crm_is_true(g_hash_table_lookup(rsc->meta, XML_RSC_ATTR_INTERLEAVE)); pe_rsc_trace(rsc, "'%s then %s' will %sbe interleaved (based on %s)", first->uuid, then->uuid, (interleave? "" : "not "), rsc->id); return interleave; } /*! * \internal * \brief Update non-interleaved instance actions according to an ordering * * Given information about an ordering of two non-interleaved actions, update * the actions' flags (and runnable_before members if appropriate) as * appropriate for the ordering. Effects may cascade to other orderings * involving the actions as well. * * \param[in,out] instance Clone instance or bundle container * \param[in,out] first "First" action in ordering * \param[in] then "Then" action in ordering (for \p instance's parent) * \param[in] node If not NULL, limit scope of ordering to this node * \param[in] flags Action flags for \p first for ordering purposes * \param[in] filter Action flags to limit scope of certain updates (may * include pcmk_action_optional to affect only * mandatory actions, and pcmk_action_runnable to * affect only runnable actions) * \param[in] type Group of enum pcmk__action_relation_flags to apply * * \return Group of enum pcmk__updated flags indicating what was updated */ static uint32_t update_noninterleaved_actions(pcmk_resource_t *instance, pcmk_action_t *first, const pcmk_action_t *then, const pcmk_node_t *node, uint32_t flags, uint32_t filter, uint32_t type) { pcmk_action_t *instance_action = NULL; uint32_t instance_flags = 0; uint32_t changed = pcmk__updated_none; // Check whether instance has an equivalent of "then" action instance_action = find_first_action(instance->actions, NULL, then->task, node); if (instance_action == NULL) { return changed; } // Check whether action is runnable instance_flags = instance->cmds->action_flags(instance_action, node); if (!pcmk_is_set(instance_flags, pcmk_action_runnable)) { return changed; } // If so, update actions for the instance changed = instance->cmds->update_ordered_actions(first, instance_action, node, flags, filter, type, instance->cluster); // Propagate any changes to later actions if (pcmk_is_set(changed, pcmk__updated_then)) { for (GList *after_iter = instance_action->actions_after; after_iter != NULL; after_iter = after_iter->next) { pcmk__related_action_t *after = after_iter->data; pcmk__update_action_for_orderings(after->action, instance->cluster); } } return changed; } /*! * \internal * \brief Update two actions according to an ordering between them * * Given information about an ordering of two clone or bundle actions, update * the actions' flags (and runnable_before members if appropriate) as * appropriate for the ordering. Effects may cascade to other orderings * involving the actions as well. * * \param[in,out] first 'First' action in an ordering * \param[in,out] then 'Then' action in an ordering * \param[in] node If not NULL, limit scope of ordering to this node * (only used when interleaving instances) * \param[in] flags Action flags for \p first for ordering purposes * \param[in] filter Action flags to limit scope of certain updates (may * include pcmk_action_optional to affect only * mandatory actions, and pcmk_action_runnable to * affect only runnable actions) * \param[in] type Group of enum pcmk__action_relation_flags to apply * \param[in,out] scheduler Scheduler data * * \return Group of enum pcmk__updated flags indicating what was updated */ uint32_t pcmk__instance_update_ordered_actions(pcmk_action_t *first, pcmk_action_t *then, const pcmk_node_t *node, uint32_t flags, uint32_t filter, uint32_t type, pcmk_scheduler_t *scheduler) { CRM_ASSERT((first != NULL) && (then != NULL) && (scheduler != NULL)); if (then->rsc == NULL) { return pcmk__updated_none; } else if (can_interleave_actions(first, then)) { return update_interleaved_actions(first, then, node, filter, type); } else { uint32_t changed = pcmk__updated_none; GList *instances = get_instance_list(then->rsc); // Update actions for the clone or bundle resource itself changed |= pcmk__update_ordered_actions(first, then, node, flags, filter, type, scheduler); // Update the 'then' clone instances or bundle containers individually for (GList *iter = instances; iter != NULL; iter = iter->next) { pcmk_resource_t *instance = iter->data; changed |= update_noninterleaved_actions(instance, first, then, node, flags, filter, type); } free_instance_list(then->rsc, instances); return changed; } } #define pe__clear_action_summary_flags(flags, action, flag) do { \ flags = pcmk__clear_flags_as(__func__, __LINE__, LOG_TRACE, \ "Action summary", action->rsc->id, \ flags, flag, #flag); \ } while (0) /*! * \internal * \brief Return action flags for a given clone or bundle action * * \param[in,out] action Action for a clone or bundle * \param[in] instances Clone instances or bundle containers * \param[in] node If not NULL, limit effects to this node * * \return Flags appropriate to \p action on \p node */ uint32_t pcmk__collective_action_flags(pcmk_action_t *action, const GList *instances, const pcmk_node_t *node) { bool any_runnable = false; const char *action_name = orig_action_name(action); // Set original assumptions (optional and runnable may be cleared below) uint32_t flags = pcmk_action_optional |pcmk_action_runnable |pcmk_action_pseudo; for (const GList *iter = instances; iter != NULL; iter = iter->next) { const pcmk_resource_t *instance = iter->data; const pcmk_node_t *instance_node = NULL; pcmk_action_t *instance_action = NULL; uint32_t instance_flags; // Node is relevant only to primitive instances if (instance->variant == pcmk_rsc_variant_primitive) { instance_node = node; } instance_action = find_first_action(instance->actions, NULL, action_name, instance_node); if (instance_action == NULL) { pe_rsc_trace(action->rsc, "%s has no %s action on %s", instance->id, action_name, pe__node_name(node)); continue; } pe_rsc_trace(action->rsc, "%s has %s for %s on %s", instance->id, instance_action->uuid, action_name, pe__node_name(node)); instance_flags = instance->cmds->action_flags(instance_action, node); // If any instance action is mandatory, so is the collective action if (pcmk_is_set(flags, pcmk_action_optional) && !pcmk_is_set(instance_flags, pcmk_action_optional)) { pe_rsc_trace(instance, "%s is mandatory because %s is", action->uuid, instance_action->uuid); pe__clear_action_summary_flags(flags, action, pcmk_action_optional); pe__clear_action_flags(action, pcmk_action_optional); } // If any instance action is runnable, so is the collective action if (pcmk_is_set(instance_flags, pcmk_action_runnable)) { any_runnable = true; } } if (!any_runnable) { pe_rsc_trace(action->rsc, "%s is not runnable because no instance can run %s", action->uuid, action_name); pe__clear_action_summary_flags(flags, action, pcmk_action_runnable); if (node == NULL) { pe__clear_action_flags(action, pcmk_action_runnable); } } return flags; } diff --git a/lib/pacemaker/pcmk_sched_ordering.c b/lib/pacemaker/pcmk_sched_ordering.c index 0044040130..f3c8d38844 100644 --- a/lib/pacemaker/pcmk_sched_ordering.c +++ b/lib/pacemaker/pcmk_sched_ordering.c @@ -1,1501 +1,1501 @@ /* * Copyright 2004-2023 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU General Public License version 2 * or later (GPLv2+) WITHOUT ANY WARRANTY. */ #include #include // PRIx32 #include #include #include #include #include "libpacemaker_private.h" enum pe_order_kind { pe_order_kind_optional, pe_order_kind_mandatory, pe_order_kind_serialize, }; enum ordering_symmetry { ordering_asymmetric, // the only relation in an asymmetric ordering ordering_symmetric, // the normal relation in a symmetric ordering ordering_symmetric_inverse, // the inverse relation in a symmetric ordering }; #define EXPAND_CONSTRAINT_IDREF(__set, __rsc, __name) do { \ __rsc = pcmk__find_constraint_resource(scheduler->resources, \ __name); \ if (__rsc == NULL) { \ pcmk__config_err("%s: No resource found for %s", __set, __name);\ return pcmk_rc_unpack_error; \ } \ } while (0) static const char * invert_action(const char *action) { if (pcmk__str_eq(action, PCMK_ACTION_START, pcmk__str_none)) { return PCMK_ACTION_STOP; } else if (pcmk__str_eq(action, PCMK_ACTION_STOP, pcmk__str_none)) { return PCMK_ACTION_START; } else if (pcmk__str_eq(action, PCMK_ACTION_PROMOTE, pcmk__str_none)) { return PCMK_ACTION_DEMOTE; } else if (pcmk__str_eq(action, PCMK_ACTION_DEMOTE, pcmk__str_none)) { return PCMK_ACTION_PROMOTE; } else if (pcmk__str_eq(action, PCMK_ACTION_PROMOTED, pcmk__str_none)) { return PCMK_ACTION_DEMOTED; } else if (pcmk__str_eq(action, PCMK_ACTION_DEMOTED, pcmk__str_none)) { return PCMK_ACTION_PROMOTED; } else if (pcmk__str_eq(action, PCMK_ACTION_RUNNING, pcmk__str_none)) { return PCMK_ACTION_STOPPED; } else if (pcmk__str_eq(action, PCMK_ACTION_STOPPED, pcmk__str_none)) { return PCMK_ACTION_RUNNING; } crm_warn("Unknown action '%s' specified in order constraint", action); return NULL; } static enum pe_order_kind get_ordering_type(const xmlNode *xml_obj) { enum pe_order_kind kind_e = pe_order_kind_mandatory; const char *kind = crm_element_value(xml_obj, XML_ORDER_ATTR_KIND); if (kind == NULL) { const char *score = crm_element_value(xml_obj, XML_RULE_ATTR_SCORE); kind_e = pe_order_kind_mandatory; if (score) { // @COMPAT deprecated informally since 1.0.7, formally since 2.0.1 int score_i = char2score(score); if (score_i == 0) { kind_e = pe_order_kind_optional; } pe_warn_once(pcmk__wo_order_score, "Support for 'score' in rsc_order is deprecated " "and will be removed in a future release " "(use 'kind' instead)"); } } else if (pcmk__str_eq(kind, "Mandatory", pcmk__str_none)) { kind_e = pe_order_kind_mandatory; } else if (pcmk__str_eq(kind, "Optional", pcmk__str_none)) { kind_e = pe_order_kind_optional; } else if (pcmk__str_eq(kind, "Serialize", pcmk__str_none)) { kind_e = pe_order_kind_serialize; } else { pcmk__config_err("Resetting '" XML_ORDER_ATTR_KIND "' for constraint " "%s to 'Mandatory' because '%s' is not valid", pcmk__s(ID(xml_obj), "missing ID"), kind); } return kind_e; } /*! * \internal * \brief Get ordering symmetry from XML * * \param[in] xml_obj Ordering XML * \param[in] parent_kind Default ordering kind * \param[in] parent_symmetrical_s Parent element's symmetrical setting, if any * * \retval ordering_symmetric Ordering is symmetric * \retval ordering_asymmetric Ordering is asymmetric */ static enum ordering_symmetry get_ordering_symmetry(const xmlNode *xml_obj, enum pe_order_kind parent_kind, const char *parent_symmetrical_s) { int rc = pcmk_rc_ok; bool symmetric = false; enum pe_order_kind kind = parent_kind; // Default to parent's kind // Check ordering XML for explicit kind if ((crm_element_value(xml_obj, XML_ORDER_ATTR_KIND) != NULL) || (crm_element_value(xml_obj, XML_RULE_ATTR_SCORE) != NULL)) { kind = get_ordering_type(xml_obj); } // Check ordering XML (and parent) for explicit symmetrical setting rc = pcmk__xe_get_bool_attr(xml_obj, XML_CONS_ATTR_SYMMETRICAL, &symmetric); if (rc != pcmk_rc_ok && parent_symmetrical_s != NULL) { symmetric = crm_is_true(parent_symmetrical_s); rc = pcmk_rc_ok; } if (rc == pcmk_rc_ok) { if (symmetric) { if (kind == pe_order_kind_serialize) { pcmk__config_warn("Ignoring " XML_CONS_ATTR_SYMMETRICAL " for '%s' because not valid with " XML_ORDER_ATTR_KIND " of 'Serialize'", ID(xml_obj)); } else { return ordering_symmetric; } } return ordering_asymmetric; } // Use default symmetry if (kind == pe_order_kind_serialize) { return ordering_asymmetric; } return ordering_symmetric; } /*! * \internal * \brief Get ordering flags appropriate to ordering kind * * \param[in] kind Ordering kind * \param[in] first Action name for 'first' action * \param[in] symmetry This ordering's symmetry role * * \return Minimal ordering flags appropriate to \p kind */ static uint32_t ordering_flags_for_kind(enum pe_order_kind kind, const char *first, enum ordering_symmetry symmetry) { uint32_t flags = pcmk__ar_none; // so we trace-log all flags set switch (kind) { case pe_order_kind_optional: pe__set_order_flags(flags, pcmk__ar_ordered); break; case pe_order_kind_serialize: /* This flag is not used anywhere directly but means the relation * will not match an equality comparison against pcmk__ar_none or * pcmk__ar_ordered. */ pe__set_order_flags(flags, pcmk__ar_serialize); break; case pe_order_kind_mandatory: pe__set_order_flags(flags, pcmk__ar_ordered); switch (symmetry) { case ordering_asymmetric: pe__set_order_flags(flags, pcmk__ar_asymmetric); break; case ordering_symmetric: pe__set_order_flags(flags, pcmk__ar_first_implies_then); if (pcmk__strcase_any_of(first, PCMK_ACTION_START, PCMK_ACTION_PROMOTE, NULL)) { pe__set_order_flags(flags, pcmk__ar_unrunnable_first_blocks); } break; case ordering_symmetric_inverse: pe__set_order_flags(flags, pcmk__ar_then_implies_first); break; } break; } return flags; } /*! * \internal * \brief Find resource corresponding to ID specified in ordering * * \param[in] xml Ordering XML * \param[in] resource_attr XML attribute name for resource ID * \param[in] instance_attr XML attribute name for instance number. * This option is deprecated and will be removed in a * future release. * \param[in] scheduler Scheduler data * * \return Resource corresponding to \p id, or NULL if none */ static pcmk_resource_t * get_ordering_resource(const xmlNode *xml, const char *resource_attr, const char *instance_attr, const pcmk_scheduler_t *scheduler) { // @COMPAT: instance_attr and instance_id variables deprecated since 2.1.5 pcmk_resource_t *rsc = NULL; const char *rsc_id = crm_element_value(xml, resource_attr); const char *instance_id = crm_element_value(xml, instance_attr); if (rsc_id == NULL) { pcmk__config_err("Ignoring constraint '%s' without %s", ID(xml), resource_attr); return NULL; } rsc = pcmk__find_constraint_resource(scheduler->resources, rsc_id); if (rsc == NULL) { pcmk__config_err("Ignoring constraint '%s' because resource '%s' " "does not exist", ID(xml), rsc_id); return NULL; } if (instance_id != NULL) { pe_warn_once(pcmk__wo_order_inst, "Support for " XML_ORDER_ATTR_FIRST_INSTANCE " and " XML_ORDER_ATTR_THEN_INSTANCE " is deprecated and will be " "removed in a future release."); if (!pe_rsc_is_clone(rsc)) { pcmk__config_err("Ignoring constraint '%s' because resource '%s' " "is not a clone but instance '%s' was requested", ID(xml), rsc_id, instance_id); return NULL; } rsc = find_clone_instance(rsc, instance_id); if (rsc == NULL) { pcmk__config_err("Ignoring constraint '%s' because resource '%s' " "does not have an instance '%s'", "'%s'", ID(xml), rsc_id, instance_id); return NULL; } } return rsc; } /*! * \internal * \brief Determine minimum number of 'first' instances required in ordering * * \param[in] rsc 'First' resource in ordering * \param[in] xml Ordering XML * * \return Minimum 'first' instances required (or 0 if not applicable) */ static int get_minimum_first_instances(const pcmk_resource_t *rsc, const xmlNode *xml) { const char *clone_min = NULL; bool require_all = false; if (!pe_rsc_is_clone(rsc)) { return 0; } clone_min = g_hash_table_lookup(rsc->meta, PCMK_META_CLONE_MIN); if (clone_min != NULL) { int clone_min_int = 0; pcmk__scan_min_int(clone_min, &clone_min_int, 0); return clone_min_int; } /* @COMPAT 1.1.13: * require-all=false is deprecated equivalent of clone-min=1 */ if (pcmk__xe_get_bool_attr(xml, "require-all", &require_all) != ENODATA) { pe_warn_once(pcmk__wo_require_all, "Support for require-all in ordering constraints " "is deprecated and will be removed in a future release" " (use clone-min clone meta-attribute instead)"); if (!require_all) { return 1; } } return 0; } /*! * \internal * \brief Create orderings for a constraint with clone-min > 0 * * \param[in] id Ordering ID * \param[in,out] rsc_first 'First' resource in ordering (a clone) * \param[in] action_first 'First' action in ordering * \param[in] rsc_then 'Then' resource in ordering * \param[in] action_then 'Then' action in ordering * \param[in] flags Ordering flags * \param[in] clone_min Minimum required instances of 'first' */ static void clone_min_ordering(const char *id, pcmk_resource_t *rsc_first, const char *action_first, pcmk_resource_t *rsc_then, const char *action_then, uint32_t flags, int clone_min) { // Create a pseudo-action for when the minimum instances are active char *task = crm_strdup_printf(PCMK_ACTION_CLONE_ONE_OR_MORE ":%s", id); pcmk_action_t *clone_min_met = get_pseudo_op(task, rsc_first->cluster); free(task); /* Require the pseudo-action to have the required number of actions to be * considered runnable before allowing the pseudo-action to be runnable. */ clone_min_met->required_runnable_before = clone_min; pe__set_action_flags(clone_min_met, pcmk_action_min_runnable); // Order the actions for each clone instance before the pseudo-action for (GList *iter = rsc_first->children; iter != NULL; iter = iter->next) { pcmk_resource_t *child = iter->data; pcmk__new_ordering(child, pcmk__op_key(child->id, action_first, 0), NULL, NULL, NULL, clone_min_met, pcmk__ar_min_runnable |pcmk__ar_first_implies_then_graphed, rsc_first->cluster); } // Order "then" action after the pseudo-action (if runnable) pcmk__new_ordering(NULL, NULL, clone_min_met, rsc_then, pcmk__op_key(rsc_then->id, action_then, 0), NULL, flags|pcmk__ar_unrunnable_first_blocks, rsc_first->cluster); } /*! * \internal * \brief Update ordering flags for restart-type=restart * * \param[in] rsc 'Then' resource in ordering * \param[in] kind Ordering kind * \param[in] flag Ordering flag to set (when applicable) * \param[in,out] flags Ordering flag set to update * * \compat The restart-type resource meta-attribute is deprecated. Eventually, * it will be removed, and pe_restart_ignore will be the only behavior, * at which time this can just be removed entirely. */ #define handle_restart_type(rsc, kind, flag, flags) do { \ if (((kind) == pe_order_kind_optional) \ && ((rsc)->restart_type == pe_restart_restart)) { \ pe__set_order_flags((flags), (flag)); \ } \ } while (0) /*! * \internal * \brief Create new ordering for inverse of symmetric constraint * * \param[in] id Ordering ID (for logging only) * \param[in] kind Ordering kind * \param[in] rsc_first 'First' resource in ordering (a clone) * \param[in] action_first 'First' action in ordering * \param[in,out] rsc_then 'Then' resource in ordering * \param[in] action_then 'Then' action in ordering */ static void inverse_ordering(const char *id, enum pe_order_kind kind, pcmk_resource_t *rsc_first, const char *action_first, pcmk_resource_t *rsc_then, const char *action_then) { action_then = invert_action(action_then); action_first = invert_action(action_first); if ((action_then == NULL) || (action_first == NULL)) { pcmk__config_warn("Cannot invert constraint '%s' " "(please specify inverse manually)", id); } else { uint32_t flags = ordering_flags_for_kind(kind, action_first, ordering_symmetric_inverse); handle_restart_type(rsc_then, kind, pcmk__ar_then_implies_first, flags); pcmk__order_resource_actions(rsc_then, action_then, rsc_first, action_first, flags); } } static void unpack_simple_rsc_order(xmlNode *xml_obj, pcmk_scheduler_t *scheduler) { pcmk_resource_t *rsc_then = NULL; pcmk_resource_t *rsc_first = NULL; int min_required_before = 0; enum pe_order_kind kind = pe_order_kind_mandatory; uint32_t flags = pcmk__ar_none; enum ordering_symmetry symmetry; const char *action_then = NULL; const char *action_first = NULL; const char *id = NULL; CRM_CHECK(xml_obj != NULL, return); id = crm_element_value(xml_obj, XML_ATTR_ID); if (id == NULL) { pcmk__config_err("Ignoring <%s> constraint without " XML_ATTR_ID, xml_obj->name); return; } rsc_first = get_ordering_resource(xml_obj, XML_ORDER_ATTR_FIRST, XML_ORDER_ATTR_FIRST_INSTANCE, scheduler); if (rsc_first == NULL) { return; } rsc_then = get_ordering_resource(xml_obj, XML_ORDER_ATTR_THEN, XML_ORDER_ATTR_THEN_INSTANCE, scheduler); if (rsc_then == NULL) { return; } action_first = crm_element_value(xml_obj, XML_ORDER_ATTR_FIRST_ACTION); if (action_first == NULL) { action_first = PCMK_ACTION_START; } action_then = crm_element_value(xml_obj, XML_ORDER_ATTR_THEN_ACTION); if (action_then == NULL) { action_then = action_first; } kind = get_ordering_type(xml_obj); symmetry = get_ordering_symmetry(xml_obj, kind, NULL); flags = ordering_flags_for_kind(kind, action_first, symmetry); handle_restart_type(rsc_then, kind, pcmk__ar_first_implies_then, flags); /* If there is a minimum number of instances that must be runnable before * the 'then' action is runnable, we use a pseudo-action for convenience: * minimum number of clone instances have runnable actions -> * pseudo-action is runnable -> dependency is runnable. */ min_required_before = get_minimum_first_instances(rsc_first, xml_obj); if (min_required_before > 0) { clone_min_ordering(id, rsc_first, action_first, rsc_then, action_then, flags, min_required_before); } else { pcmk__order_resource_actions(rsc_first, action_first, rsc_then, action_then, flags); } if (symmetry == ordering_symmetric) { inverse_ordering(id, kind, rsc_first, action_first, rsc_then, action_then); } } /*! * \internal * \brief Create a new ordering between two actions * * \param[in,out] first_rsc Resource for 'first' action (if NULL and * \p first_action is a resource action, that * resource will be used) * \param[in,out] first_action_task Action key for 'first' action (if NULL and * \p first_action is not NULL, its UUID will * be used) * \param[in,out] first_action 'first' action (if NULL, \p first_rsc and * \p first_action_task must be set) * * \param[in] then_rsc Resource for 'then' action (if NULL and * \p then_action is a resource action, that * resource will be used) * \param[in,out] then_action_task Action key for 'then' action (if NULL and * \p then_action is not NULL, its UUID will * be used) * \param[in] then_action 'then' action (if NULL, \p then_rsc and * \p then_action_task must be set) * * \param[in] flags Group of enum pcmk__action_relation_flags * \param[in,out] sched Scheduler data to add ordering to * * \note This function takes ownership of first_action_task and * then_action_task, which do not need to be freed by the caller. */ void pcmk__new_ordering(pcmk_resource_t *first_rsc, char *first_action_task, pcmk_action_t *first_action, pcmk_resource_t *then_rsc, char *then_action_task, pcmk_action_t *then_action, uint32_t flags, pcmk_scheduler_t *sched) { pcmk__action_relation_t *order = NULL; // One of action or resource must be specified for each side CRM_CHECK(((first_action != NULL) || (first_rsc != NULL)) && ((then_action != NULL) || (then_rsc != NULL)), free(first_action_task); free(then_action_task); return); if ((first_rsc == NULL) && (first_action != NULL)) { first_rsc = first_action->rsc; } if ((then_rsc == NULL) && (then_action != NULL)) { then_rsc = then_action->rsc; } order = calloc(1, sizeof(pcmk__action_relation_t)); CRM_ASSERT(order != NULL); order->id = sched->order_id++; order->flags = flags; order->rsc1 = first_rsc; order->rsc2 = then_rsc; order->action1 = first_action; order->action2 = then_action; order->task1 = first_action_task; order->task2 = then_action_task; if ((order->task1 == NULL) && (first_action != NULL)) { order->task1 = strdup(first_action->uuid); } if ((order->task2 == NULL) && (then_action != NULL)) { order->task2 = strdup(then_action->uuid); } if ((order->rsc1 == NULL) && (first_action != NULL)) { order->rsc1 = first_action->rsc; } if ((order->rsc2 == NULL) && (then_action != NULL)) { order->rsc2 = then_action->rsc; } pe_rsc_trace(first_rsc, "Created ordering %d for %s then %s", (sched->order_id - 1), pcmk__s(order->task1, "an underspecified action"), pcmk__s(order->task2, "an underspecified action")); sched->ordering_constraints = g_list_prepend(sched->ordering_constraints, order); pcmk__order_migration_equivalents(order); } /*! * \brief Unpack a set in an ordering constraint * * \param[in] set Set XML to unpack * \param[in] parent_kind rsc_order XML "kind" attribute * \param[in] parent_symmetrical_s rsc_order XML "symmetrical" attribute * \param[in,out] scheduler Scheduler data * * \return Standard Pacemaker return code */ static int unpack_order_set(const xmlNode *set, enum pe_order_kind parent_kind, const char *parent_symmetrical_s, pcmk_scheduler_t *scheduler) { GList *set_iter = NULL; GList *resources = NULL; pcmk_resource_t *last = NULL; pcmk_resource_t *resource = NULL; int local_kind = parent_kind; bool sequential = false; uint32_t flags = pcmk__ar_ordered; enum ordering_symmetry symmetry; char *key = NULL; const char *id = ID(set); const char *action = crm_element_value(set, "action"); const char *sequential_s = crm_element_value(set, "sequential"); const char *kind_s = crm_element_value(set, XML_ORDER_ATTR_KIND); if (action == NULL) { action = PCMK_ACTION_START; } if (kind_s) { local_kind = get_ordering_type(set); } if (sequential_s == NULL) { sequential_s = "1"; } sequential = crm_is_true(sequential_s); symmetry = get_ordering_symmetry(set, parent_kind, parent_symmetrical_s); flags = ordering_flags_for_kind(local_kind, action, symmetry); for (const xmlNode *xml_rsc = first_named_child(set, XML_TAG_RESOURCE_REF); xml_rsc != NULL; xml_rsc = crm_next_same_xml(xml_rsc)) { EXPAND_CONSTRAINT_IDREF(id, resource, ID(xml_rsc)); resources = g_list_append(resources, resource); } if (pcmk__list_of_1(resources)) { crm_trace("Single set: %s", id); goto done; } set_iter = resources; while (set_iter != NULL) { resource = (pcmk_resource_t *) set_iter->data; set_iter = set_iter->next; key = pcmk__op_key(resource->id, action, 0); if (local_kind == pe_order_kind_serialize) { /* Serialize before everything that comes after */ for (GList *iter = set_iter; iter != NULL; iter = iter->next) { pcmk_resource_t *then_rsc = iter->data; char *then_key = pcmk__op_key(then_rsc->id, action, 0); pcmk__new_ordering(resource, strdup(key), NULL, then_rsc, then_key, NULL, flags, scheduler); } } else if (sequential) { if (last != NULL) { pcmk__order_resource_actions(last, action, resource, action, flags); } last = resource; } free(key); } if (symmetry == ordering_asymmetric) { goto done; } last = NULL; action = invert_action(action); flags = ordering_flags_for_kind(local_kind, action, ordering_symmetric_inverse); set_iter = resources; while (set_iter != NULL) { resource = (pcmk_resource_t *) set_iter->data; set_iter = set_iter->next; if (sequential) { if (last != NULL) { pcmk__order_resource_actions(resource, action, last, action, flags); } last = resource; } } done: g_list_free(resources); return pcmk_rc_ok; } /*! * \brief Order two resource sets relative to each other * * \param[in] id Ordering ID (for logging) * \param[in] set1 First listed set * \param[in] set2 Second listed set * \param[in] kind Ordering kind * \param[in,out] scheduler Scheduler data * \param[in] symmetry Which ordering symmetry applies to this relation * * \return Standard Pacemaker return code */ static int order_rsc_sets(const char *id, const xmlNode *set1, const xmlNode *set2, enum pe_order_kind kind, pcmk_scheduler_t *scheduler, enum ordering_symmetry symmetry) { const xmlNode *xml_rsc = NULL; const xmlNode *xml_rsc_2 = NULL; pcmk_resource_t *rsc_1 = NULL; pcmk_resource_t *rsc_2 = NULL; const char *action_1 = crm_element_value(set1, "action"); const char *action_2 = crm_element_value(set2, "action"); uint32_t flags = pcmk__ar_none; bool require_all = true; (void) pcmk__xe_get_bool_attr(set1, "require-all", &require_all); if (action_1 == NULL) { action_1 = PCMK_ACTION_START; } if (action_2 == NULL) { action_2 = PCMK_ACTION_START; } if (symmetry == ordering_symmetric_inverse) { action_1 = invert_action(action_1); action_2 = invert_action(action_2); } if (pcmk__str_eq(PCMK_ACTION_STOP, action_1, pcmk__str_none) || pcmk__str_eq(PCMK_ACTION_DEMOTE, action_1, pcmk__str_none)) { /* Assuming: A -> ( B || C) -> D * The one-or-more logic only applies during the start/promote phase. * During shutdown neither B nor can shutdown until D is down, so simply * turn require_all back on. */ require_all = true; } flags = ordering_flags_for_kind(kind, action_1, symmetry); /* If we have an unordered set1, whether it is sequential or not is * irrelevant in regards to set2. */ if (!require_all) { char *task = crm_strdup_printf(PCMK_ACTION_ONE_OR_MORE ":%s", ID(set1)); pcmk_action_t *unordered_action = get_pseudo_op(task, scheduler); free(task); pe__set_action_flags(unordered_action, pcmk_action_min_runnable); for (xml_rsc = first_named_child(set1, XML_TAG_RESOURCE_REF); xml_rsc != NULL; xml_rsc = crm_next_same_xml(xml_rsc)) { EXPAND_CONSTRAINT_IDREF(id, rsc_1, ID(xml_rsc)); /* Add an ordering constraint between every element in set1 and the * pseudo action. If any action in set1 is runnable the pseudo * action will be runnable. */ pcmk__new_ordering(rsc_1, pcmk__op_key(rsc_1->id, action_1, 0), NULL, NULL, NULL, unordered_action, pcmk__ar_min_runnable |pcmk__ar_first_implies_then_graphed, scheduler); } for (xml_rsc_2 = first_named_child(set2, XML_TAG_RESOURCE_REF); xml_rsc_2 != NULL; xml_rsc_2 = crm_next_same_xml(xml_rsc_2)) { EXPAND_CONSTRAINT_IDREF(id, rsc_2, ID(xml_rsc_2)); /* Add an ordering constraint between the pseudo-action and every * element in set2. If the pseudo-action is runnable, every action * in set2 will be runnable. */ pcmk__new_ordering(NULL, NULL, unordered_action, rsc_2, pcmk__op_key(rsc_2->id, action_2, 0), NULL, flags|pcmk__ar_unrunnable_first_blocks, scheduler); } return pcmk_rc_ok; } if (pcmk__xe_attr_is_true(set1, "sequential")) { if (symmetry == ordering_symmetric_inverse) { // Get the first one xml_rsc = first_named_child(set1, XML_TAG_RESOURCE_REF); if (xml_rsc != NULL) { EXPAND_CONSTRAINT_IDREF(id, rsc_1, ID(xml_rsc)); } } else { // Get the last one const char *rid = NULL; for (xml_rsc = first_named_child(set1, XML_TAG_RESOURCE_REF); xml_rsc != NULL; xml_rsc = crm_next_same_xml(xml_rsc)) { rid = ID(xml_rsc); } EXPAND_CONSTRAINT_IDREF(id, rsc_1, rid); } } if (pcmk__xe_attr_is_true(set2, "sequential")) { if (symmetry == ordering_symmetric_inverse) { // Get the last one const char *rid = NULL; for (xml_rsc = first_named_child(set2, XML_TAG_RESOURCE_REF); xml_rsc != NULL; xml_rsc = crm_next_same_xml(xml_rsc)) { rid = ID(xml_rsc); } EXPAND_CONSTRAINT_IDREF(id, rsc_2, rid); } else { // Get the first one xml_rsc = first_named_child(set2, XML_TAG_RESOURCE_REF); if (xml_rsc != NULL) { EXPAND_CONSTRAINT_IDREF(id, rsc_2, ID(xml_rsc)); } } } if ((rsc_1 != NULL) && (rsc_2 != NULL)) { pcmk__order_resource_actions(rsc_1, action_1, rsc_2, action_2, flags); } else if (rsc_1 != NULL) { for (xml_rsc = first_named_child(set2, XML_TAG_RESOURCE_REF); xml_rsc != NULL; xml_rsc = crm_next_same_xml(xml_rsc)) { EXPAND_CONSTRAINT_IDREF(id, rsc_2, ID(xml_rsc)); pcmk__order_resource_actions(rsc_1, action_1, rsc_2, action_2, flags); } } else if (rsc_2 != NULL) { for (xml_rsc = first_named_child(set1, XML_TAG_RESOURCE_REF); xml_rsc != NULL; xml_rsc = crm_next_same_xml(xml_rsc)) { EXPAND_CONSTRAINT_IDREF(id, rsc_1, ID(xml_rsc)); pcmk__order_resource_actions(rsc_1, action_1, rsc_2, action_2, flags); } } else { for (xml_rsc = first_named_child(set1, XML_TAG_RESOURCE_REF); xml_rsc != NULL; xml_rsc = crm_next_same_xml(xml_rsc)) { EXPAND_CONSTRAINT_IDREF(id, rsc_1, ID(xml_rsc)); for (xmlNode *xml_rsc_2 = first_named_child(set2, XML_TAG_RESOURCE_REF); xml_rsc_2 != NULL; xml_rsc_2 = crm_next_same_xml(xml_rsc_2)) { EXPAND_CONSTRAINT_IDREF(id, rsc_2, ID(xml_rsc_2)); pcmk__order_resource_actions(rsc_1, action_1, rsc_2, action_2, flags); } } } return pcmk_rc_ok; } /*! * \internal * \brief If an ordering constraint uses resource tags, expand them * * \param[in,out] xml_obj Ordering constraint XML * \param[out] expanded_xml Equivalent XML with tags expanded * \param[in] scheduler Scheduler data * * \return Standard Pacemaker return code (specifically, pcmk_rc_ok on success, * and pcmk_rc_unpack_error on invalid configuration) */ static int unpack_order_tags(xmlNode *xml_obj, xmlNode **expanded_xml, const pcmk_scheduler_t *scheduler) { const char *id_first = NULL; const char *id_then = NULL; const char *action_first = NULL; const char *action_then = NULL; pcmk_resource_t *rsc_first = NULL; pcmk_resource_t *rsc_then = NULL; pcmk_tag_t *tag_first = NULL; pcmk_tag_t *tag_then = NULL; xmlNode *rsc_set_first = NULL; xmlNode *rsc_set_then = NULL; bool any_sets = false; // Check whether there are any resource sets with template or tag references *expanded_xml = pcmk__expand_tags_in_sets(xml_obj, scheduler); if (*expanded_xml != NULL) { crm_log_xml_trace(*expanded_xml, "Expanded rsc_order"); return pcmk_rc_ok; } id_first = crm_element_value(xml_obj, XML_ORDER_ATTR_FIRST); id_then = crm_element_value(xml_obj, XML_ORDER_ATTR_THEN); if ((id_first == NULL) || (id_then == NULL)) { return pcmk_rc_ok; } if (!pcmk__valid_resource_or_tag(scheduler, id_first, &rsc_first, &tag_first)) { pcmk__config_err("Ignoring constraint '%s' because '%s' is not a " "valid resource or tag", ID(xml_obj), id_first); return pcmk_rc_unpack_error; } if (!pcmk__valid_resource_or_tag(scheduler, id_then, &rsc_then, &tag_then)) { pcmk__config_err("Ignoring constraint '%s' because '%s' is not a " "valid resource or tag", ID(xml_obj), id_then); return pcmk_rc_unpack_error; } if ((rsc_first != NULL) && (rsc_then != NULL)) { // Neither side references a template or tag return pcmk_rc_ok; } action_first = crm_element_value(xml_obj, XML_ORDER_ATTR_FIRST_ACTION); action_then = crm_element_value(xml_obj, XML_ORDER_ATTR_THEN_ACTION); *expanded_xml = copy_xml(xml_obj); // Convert template/tag reference in "first" into constraint resource_set if (!pcmk__tag_to_set(*expanded_xml, &rsc_set_first, XML_ORDER_ATTR_FIRST, true, scheduler)) { free_xml(*expanded_xml); *expanded_xml = NULL; return pcmk_rc_unpack_error; } if (rsc_set_first != NULL) { if (action_first != NULL) { // Move "first-action" into converted resource_set as "action" crm_xml_add(rsc_set_first, "action", action_first); xml_remove_prop(*expanded_xml, XML_ORDER_ATTR_FIRST_ACTION); } any_sets = true; } // Convert template/tag reference in "then" into constraint resource_set if (!pcmk__tag_to_set(*expanded_xml, &rsc_set_then, XML_ORDER_ATTR_THEN, true, scheduler)) { free_xml(*expanded_xml); *expanded_xml = NULL; return pcmk_rc_unpack_error; } if (rsc_set_then != NULL) { if (action_then != NULL) { // Move "then-action" into converted resource_set as "action" crm_xml_add(rsc_set_then, "action", action_then); xml_remove_prop(*expanded_xml, XML_ORDER_ATTR_THEN_ACTION); } any_sets = true; } if (any_sets) { crm_log_xml_trace(*expanded_xml, "Expanded rsc_order"); } else { free_xml(*expanded_xml); *expanded_xml = NULL; } return pcmk_rc_ok; } /*! * \internal * \brief Unpack ordering constraint XML * * \param[in,out] xml_obj Ordering constraint XML to unpack * \param[in,out] scheduler Scheduler data */ void pcmk__unpack_ordering(xmlNode *xml_obj, pcmk_scheduler_t *scheduler) { xmlNode *set = NULL; xmlNode *last = NULL; xmlNode *orig_xml = NULL; xmlNode *expanded_xml = NULL; const char *id = crm_element_value(xml_obj, XML_ATTR_ID); const char *invert = crm_element_value(xml_obj, XML_CONS_ATTR_SYMMETRICAL); enum pe_order_kind kind = get_ordering_type(xml_obj); enum ordering_symmetry symmetry = get_ordering_symmetry(xml_obj, kind, NULL); // Expand any resource tags in the constraint XML if (unpack_order_tags(xml_obj, &expanded_xml, scheduler) != pcmk_rc_ok) { return; } if (expanded_xml != NULL) { orig_xml = xml_obj; xml_obj = expanded_xml; } // If the constraint has resource sets, unpack them for (set = first_named_child(xml_obj, XML_CONS_TAG_RSC_SET); set != NULL; set = crm_next_same_xml(set)) { set = expand_idref(set, scheduler->input); if ((set == NULL) // Configuration error, message already logged || (unpack_order_set(set, kind, invert, scheduler) != pcmk_rc_ok)) { if (expanded_xml != NULL) { free_xml(expanded_xml); } return; } if (last != NULL) { if (order_rsc_sets(id, last, set, kind, scheduler, symmetry) != pcmk_rc_ok) { if (expanded_xml != NULL) { free_xml(expanded_xml); } return; } if ((symmetry == ordering_symmetric) && (order_rsc_sets(id, set, last, kind, scheduler, ordering_symmetric_inverse) != pcmk_rc_ok)) { if (expanded_xml != NULL) { free_xml(expanded_xml); } return; } } last = set; } if (expanded_xml) { free_xml(expanded_xml); xml_obj = orig_xml; } // If the constraint has no resource sets, unpack it as a simple ordering if (last == NULL) { return unpack_simple_rsc_order(xml_obj, scheduler); } } static bool ordering_is_invalid(pcmk_action_t *action, pcmk__related_action_t *input) { /* Prevent user-defined ordering constraints between resources * running in a guest node and the resource that defines that node. */ if (!pcmk_is_set(input->type, pcmk__ar_guest_allowed) && (input->action->rsc != NULL) && pcmk__rsc_corresponds_to_guest(action->rsc, input->action->node)) { crm_warn("Invalid ordering constraint between %s and %s", input->action->rsc->id, action->rsc->id); return true; } /* If there's an order like * "rscB_stop node2"-> "load_stopped_node2" -> "rscA_migrate_to node1" * * then rscA is being migrated from node1 to node2, while rscB is being * migrated from node2 to node1. If there would be a graph loop, * break the order "load_stopped_node2" -> "rscA_migrate_to node1". */ if (((uint32_t) input->type == pcmk__ar_if_on_same_node_or_target) && (action->rsc != NULL) && pcmk__str_eq(action->task, PCMK_ACTION_MIGRATE_TO, pcmk__str_none) && pcmk__graph_has_loop(action, action, input)) { return true; } return false; } void pcmk__disable_invalid_orderings(pcmk_scheduler_t *scheduler) { for (GList *iter = scheduler->actions; iter != NULL; iter = iter->next) { pcmk_action_t *action = (pcmk_action_t *) iter->data; pcmk__related_action_t *input = NULL; for (GList *input_iter = action->actions_before; input_iter != NULL; input_iter = input_iter->next) { input = input_iter->data; if (ordering_is_invalid(action, input)) { input->type = (enum pe_ordering) pcmk__ar_none; } } } } /*! * \internal * \brief Order stops on a node before the node's shutdown * * \param[in,out] node Node being shut down * \param[in] shutdown_op Shutdown action for node */ void pcmk__order_stops_before_shutdown(pcmk_node_t *node, pcmk_action_t *shutdown_op) { for (GList *iter = node->details->data_set->actions; iter != NULL; iter = iter->next) { pcmk_action_t *action = (pcmk_action_t *) iter->data; // Only stops on the node shutting down are relevant if (!pe__same_node(action->node, node) || !pcmk__str_eq(action->task, PCMK_ACTION_STOP, pcmk__str_none)) { continue; } // Resources and nodes in maintenance mode won't be touched if (pcmk_is_set(action->rsc->flags, pcmk_rsc_maintenance)) { pe_rsc_trace(action->rsc, "Not ordering %s before shutdown of %s because " "resource in maintenance mode", action->uuid, pe__node_name(node)); continue; } else if (node->details->maintenance) { pe_rsc_trace(action->rsc, "Not ordering %s before shutdown of %s because " "node in maintenance mode", action->uuid, pe__node_name(node)); continue; } /* Don't touch a resource that is unmanaged or blocked, to avoid * blocking the shutdown (though if another action depends on this one, * we may still end up blocking) */ if (!pcmk_any_flags_set(action->rsc->flags, pcmk_rsc_managed|pcmk_rsc_blocked)) { pe_rsc_trace(action->rsc, "Not ordering %s before shutdown of %s because " "resource is unmanaged or blocked", action->uuid, pe__node_name(node)); continue; } pe_rsc_trace(action->rsc, "Ordering %s before shutdown of %s", action->uuid, pe__node_name(node)); pe__clear_action_flags(action, pcmk_action_optional); pcmk__new_ordering(action->rsc, NULL, action, NULL, strdup(PCMK_ACTION_DO_SHUTDOWN), shutdown_op, pcmk__ar_ordered|pcmk__ar_unrunnable_first_blocks, node->details->data_set); } } /*! * \brief Find resource actions matching directly or as child * * \param[in] rsc Resource to check * \param[in] original_key Action key to search for (possibly referencing * parent of \rsc) * * \return Newly allocated list of matching actions * \note It is the caller's responsibility to free the result with g_list_free() */ static GList * find_actions_by_task(const pcmk_resource_t *rsc, const char *original_key) { // Search under given task key directly GList *list = find_actions(rsc->actions, original_key, NULL); if (list == NULL) { // Search again using this resource's ID 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(rsc->actions, key, NULL); free(key); free(task); } else { crm_err("Invalid operation key (bug?): %s", original_key); } } return list; } /*! * \internal * \brief Order relevant resource actions after a given action * * \param[in,out] first_action Action to order after (or NULL if none runnable) * \param[in] rsc Resource whose actions should be ordered * \param[in,out] order Ordering constraint being applied */ static void order_resource_actions_after(pcmk_action_t *first_action, const pcmk_resource_t *rsc, pcmk__action_relation_t *order) { GList *then_actions = NULL; uint32_t flags = pcmk__ar_none; CRM_CHECK((rsc != NULL) && (order != NULL), return); flags = order->flags; pe_rsc_trace(rsc, "Applying ordering %d for 'then' resource %s", order->id, rsc->id); if (order->action2 != NULL) { then_actions = g_list_prepend(NULL, order->action2); } else { then_actions = find_actions_by_task(rsc, order->task2); } if (then_actions == NULL) { pe_rsc_trace(rsc, "Ignoring ordering %d: no %s actions found for %s", order->id, order->task2, rsc->id); return; } if ((first_action != NULL) && (first_action->rsc == rsc) && pcmk_is_set(first_action->flags, pcmk_action_migration_abort)) { pe_rsc_trace(rsc, "Detected dangling migration ordering (%s then %s %s)", first_action->uuid, order->task2, rsc->id); pe__clear_order_flags(flags, pcmk__ar_first_implies_then); } if ((first_action == NULL) && !pcmk_is_set(flags, pcmk__ar_first_implies_then)) { - pe_rsc_debug(rsc, - "Ignoring ordering %d for %s: No first action found", - order->id, rsc->id); + pcmk__rsc_debug(rsc, + "Ignoring ordering %d for %s: No first action found", + order->id, rsc->id); g_list_free(then_actions); return; } for (GList *iter = then_actions; iter != NULL; iter = iter->next) { pcmk_action_t *then_action_iter = (pcmk_action_t *) iter->data; if (first_action != NULL) { order_actions(first_action, then_action_iter, flags); } else { pe__clear_action_flags(then_action_iter, pcmk_action_runnable); crm_warn("%s of %s is unrunnable because there is no %s of %s " "to order it after", then_action_iter->task, rsc->id, order->task1, order->rsc1->id); } } g_list_free(then_actions); } static void rsc_order_first(pcmk_resource_t *first_rsc, pcmk__action_relation_t *order) { GList *first_actions = NULL; pcmk_action_t *first_action = order->action1; pcmk_resource_t *then_rsc = order->rsc2; CRM_ASSERT(first_rsc != NULL); pe_rsc_trace(first_rsc, "Applying ordering constraint %d (first: %s)", order->id, first_rsc->id); if (first_action != NULL) { first_actions = g_list_prepend(NULL, first_action); } else { first_actions = find_actions_by_task(first_rsc, order->task1); } if ((first_actions == NULL) && (first_rsc == then_rsc)) { pe_rsc_trace(first_rsc, "Ignoring constraint %d: first (%s for %s) not found", order->id, order->task1, first_rsc->id); } else if (first_actions == NULL) { char *key = NULL; char *op_type = NULL; guint interval_ms = 0; parse_op_key(order->task1, NULL, &op_type, &interval_ms); key = pcmk__op_key(first_rsc->id, op_type, interval_ms); if ((first_rsc->fns->state(first_rsc, TRUE) == pcmk_role_stopped) && pcmk__str_eq(op_type, PCMK_ACTION_STOP, pcmk__str_none)) { free(key); pe_rsc_trace(first_rsc, "Ignoring constraint %d: first (%s for %s) not found", order->id, order->task1, first_rsc->id); } else if ((first_rsc->fns->state(first_rsc, TRUE) == pcmk_role_unpromoted) && pcmk__str_eq(op_type, PCMK_ACTION_DEMOTE, pcmk__str_none)) { free(key); pe_rsc_trace(first_rsc, "Ignoring constraint %d: first (%s for %s) not found", order->id, order->task1, first_rsc->id); } else { pe_rsc_trace(first_rsc, "Creating first (%s for %s) for constraint %d ", order->task1, first_rsc->id, order->id); first_action = custom_action(first_rsc, key, op_type, NULL, TRUE, first_rsc->cluster); first_actions = g_list_prepend(NULL, first_action); } free(op_type); } if (then_rsc == NULL) { if (order->action2 == NULL) { pe_rsc_trace(first_rsc, "Ignoring constraint %d: then not found", order->id); return; } then_rsc = order->action2->rsc; } for (GList *iter = first_actions; iter != NULL; iter = iter->next) { first_action = iter->data; if (then_rsc == NULL) { order_actions(first_action, order->action2, order->flags); } else { order_resource_actions_after(first_action, then_rsc, order); } } g_list_free(first_actions); } // GFunc to call pcmk__block_colocation_dependents() static void block_colocation_dependents(gpointer data, gpointer user_data) { pcmk__block_colocation_dependents(data); } // GFunc to call pcmk__update_action_for_orderings() static void update_action_for_orderings(gpointer data, gpointer user_data) { pcmk__update_action_for_orderings((pcmk_action_t *) data, (pcmk_scheduler_t *) user_data); } /*! * \internal * \brief Apply all ordering constraints * * \param[in,out] sched Scheduler data */ void pcmk__apply_orderings(pcmk_scheduler_t *sched) { crm_trace("Applying ordering constraints"); /* Ordering constraints need to be processed in the order they were created. * rsc_order_first() and order_resource_actions_after() require the relevant * actions to already exist in some cases, but rsc_order_first() will create * the 'first' action in certain cases. Thus calling rsc_order_first() can * change the behavior of later-created orderings. * * Also, g_list_append() should be avoided for performance reasons, so we * prepend orderings when creating them and reverse the list here. * * @TODO This is brittle and should be carefully redesigned so that the * order of creation doesn't matter, and the reverse becomes unneeded. */ sched->ordering_constraints = g_list_reverse(sched->ordering_constraints); for (GList *iter = sched->ordering_constraints; iter != NULL; iter = iter->next) { pcmk__action_relation_t *order = iter->data; pcmk_resource_t *rsc = order->rsc1; if (rsc != NULL) { rsc_order_first(rsc, order); continue; } rsc = order->rsc2; if (rsc != NULL) { order_resource_actions_after(order->action1, rsc, order); } else { crm_trace("Applying ordering constraint %d (non-resource actions)", order->id); order_actions(order->action1, order->action2, order->flags); } } g_list_foreach(sched->actions, block_colocation_dependents, NULL); crm_trace("Ordering probes"); pcmk__order_probes(sched); crm_trace("Updating %d actions", g_list_length(sched->actions)); g_list_foreach(sched->actions, update_action_for_orderings, sched); pcmk__disable_invalid_orderings(sched); } /*! * \internal * \brief Order a given action after each action in a given list * * \param[in,out] after "After" action * \param[in,out] list List of "before" actions */ void pcmk__order_after_each(pcmk_action_t *after, GList *list) { const char *after_desc = (after->task == NULL)? after->uuid : after->task; for (GList *iter = list; iter != NULL; iter = iter->next) { pcmk_action_t *before = (pcmk_action_t *) iter->data; const char *before_desc = before->task? before->task : before->uuid; crm_debug("Ordering %s on %s before %s on %s", before_desc, pe__node_name(before->node), after_desc, pe__node_name(after->node)); order_actions(before, after, pcmk__ar_ordered); } } /*! * \internal * \brief Order promotions and demotions for restarts of a clone or bundle * * \param[in,out] rsc Clone or bundle to order */ void pcmk__promotable_restart_ordering(pcmk_resource_t *rsc) { // Order start and promote after all instances are stopped pcmk__order_resource_actions(rsc, PCMK_ACTION_STOPPED, rsc, PCMK_ACTION_START, pcmk__ar_ordered); pcmk__order_resource_actions(rsc, PCMK_ACTION_STOPPED, rsc, PCMK_ACTION_PROMOTE, pcmk__ar_ordered); // Order stop, start, and promote after all instances are demoted pcmk__order_resource_actions(rsc, PCMK_ACTION_DEMOTED, rsc, PCMK_ACTION_STOP, pcmk__ar_ordered); pcmk__order_resource_actions(rsc, PCMK_ACTION_DEMOTED, rsc, PCMK_ACTION_START, pcmk__ar_ordered); pcmk__order_resource_actions(rsc, PCMK_ACTION_DEMOTED, rsc, PCMK_ACTION_PROMOTE, pcmk__ar_ordered); // Order promote after all instances are started pcmk__order_resource_actions(rsc, PCMK_ACTION_RUNNING, rsc, PCMK_ACTION_PROMOTE, pcmk__ar_ordered); // Order demote after all instances are demoted pcmk__order_resource_actions(rsc, PCMK_ACTION_DEMOTE, rsc, PCMK_ACTION_DEMOTED, pcmk__ar_ordered); } diff --git a/lib/pacemaker/pcmk_sched_primitive.c b/lib/pacemaker/pcmk_sched_primitive.c index 37ac358190..76a5c9f100 100644 --- a/lib/pacemaker/pcmk_sched_primitive.c +++ b/lib/pacemaker/pcmk_sched_primitive.c @@ -1,1663 +1,1664 @@ /* * Copyright 2004-2023 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU General Public License version 2 * or later (GPLv2+) WITHOUT ANY WARRANTY. */ #include #include #include // uint8_t, uint32_t #include #include #include "libpacemaker_private.h" static void stop_resource(pcmk_resource_t *rsc, pcmk_node_t *node, bool optional); static void start_resource(pcmk_resource_t *rsc, pcmk_node_t *node, bool optional); static void demote_resource(pcmk_resource_t *rsc, pcmk_node_t *node, bool optional); static void promote_resource(pcmk_resource_t *rsc, pcmk_node_t *node, bool optional); static void assert_role_error(pcmk_resource_t *rsc, pcmk_node_t *node, bool optional); #define RSC_ROLE_MAX (pcmk_role_promoted + 1) static enum rsc_role_e rsc_state_matrix[RSC_ROLE_MAX][RSC_ROLE_MAX] = { /* This array lists the immediate next role when transitioning from one role * to a target role. For example, when going from Stopped to Promoted, the * next role is Unpromoted, because the resource must be started before it * can be promoted. The current state then becomes Started, which is fed * into this array again, giving a next role of Promoted. * * Current role Immediate next role Final target role * ------------ ------------------- ----------------- */ /* Unknown */ { pcmk_role_unknown, /* Unknown */ pcmk_role_stopped, /* Stopped */ pcmk_role_stopped, /* Started */ pcmk_role_stopped, /* Unpromoted */ pcmk_role_stopped, /* Promoted */ }, /* Stopped */ { pcmk_role_stopped, /* Unknown */ pcmk_role_stopped, /* Stopped */ pcmk_role_started, /* Started */ pcmk_role_unpromoted, /* Unpromoted */ pcmk_role_unpromoted, /* Promoted */ }, /* Started */ { pcmk_role_stopped, /* Unknown */ pcmk_role_stopped, /* Stopped */ pcmk_role_started, /* Started */ pcmk_role_unpromoted, /* Unpromoted */ pcmk_role_promoted, /* Promoted */ }, /* Unpromoted */ { pcmk_role_stopped, /* Unknown */ pcmk_role_stopped, /* Stopped */ pcmk_role_stopped, /* Started */ pcmk_role_unpromoted, /* Unpromoted */ pcmk_role_promoted, /* Promoted */ }, /* Promoted */ { pcmk_role_stopped, /* Unknown */ pcmk_role_unpromoted, /* Stopped */ pcmk_role_unpromoted, /* Started */ pcmk_role_unpromoted, /* Unpromoted */ pcmk_role_promoted, /* Promoted */ }, }; /*! * \internal * \brief Function to schedule actions needed for a role change * * \param[in,out] rsc Resource whose role is changing * \param[in,out] node Node where resource will be in its next role * \param[in] optional Whether scheduled actions should be optional */ typedef void (*rsc_transition_fn)(pcmk_resource_t *rsc, pcmk_node_t *node, bool optional); static rsc_transition_fn rsc_action_matrix[RSC_ROLE_MAX][RSC_ROLE_MAX] = { /* This array lists the function needed to transition directly from one role * to another. NULL indicates that nothing is needed. * * Current role Transition function Next role * ------------ ------------------- ---------- */ /* Unknown */ { assert_role_error, /* Unknown */ stop_resource, /* Stopped */ assert_role_error, /* Started */ assert_role_error, /* Unpromoted */ assert_role_error, /* Promoted */ }, /* Stopped */ { assert_role_error, /* Unknown */ NULL, /* Stopped */ start_resource, /* Started */ start_resource, /* Unpromoted */ assert_role_error, /* Promoted */ }, /* Started */ { assert_role_error, /* Unknown */ stop_resource, /* Stopped */ NULL, /* Started */ NULL, /* Unpromoted */ promote_resource, /* Promoted */ }, /* Unpromoted */ { assert_role_error, /* Unknown */ stop_resource, /* Stopped */ stop_resource, /* Started */ NULL, /* Unpromoted */ promote_resource, /* Promoted */ }, /* Promoted */ { assert_role_error, /* Unknown */ demote_resource, /* Stopped */ demote_resource, /* Started */ demote_resource, /* Unpromoted */ NULL, /* Promoted */ }, }; /*! * \internal * \brief Get a list of a resource's allowed nodes sorted by node score * * \param[in] rsc Resource to check * * \return List of allowed nodes sorted by node score */ static GList * sorted_allowed_nodes(const pcmk_resource_t *rsc) { if (rsc->allowed_nodes != NULL) { GList *nodes = g_hash_table_get_values(rsc->allowed_nodes); if (nodes != NULL) { return pcmk__sort_nodes(nodes, pe__current_node(rsc)); } } return NULL; } /*! * \internal * \brief Assign a resource to its best allowed node, if possible * * \param[in,out] rsc Resource to choose a node for * \param[in] prefer If not \c NULL, prefer this node when all else * equal * \param[in] stop_if_fail If \c true and \p rsc can't be assigned to a * node, set next role to stopped and update * existing actions * * \return true if \p rsc could be assigned to a node, otherwise false * * \note If \p stop_if_fail is \c false, then \c pcmk__unassign_resource() can * completely undo the assignment. A successful assignment can be either * undone or left alone as final. A failed assignment has the same effect * as calling pcmk__unassign_resource(); there are no side effects on * roles or actions. */ static bool assign_best_node(pcmk_resource_t *rsc, const pcmk_node_t *prefer, bool stop_if_fail) { GList *nodes = NULL; pcmk_node_t *chosen = NULL; pcmk_node_t *best = NULL; const pcmk_node_t *most_free_node = pcmk__ban_insufficient_capacity(rsc); if (prefer == NULL) { prefer = most_free_node; } if (!pcmk_is_set(rsc->flags, pcmk_rsc_unassigned)) { // We've already finished assignment of resources to nodes return rsc->allocated_to != NULL; } // Sort allowed nodes by score nodes = sorted_allowed_nodes(rsc); if (nodes != NULL) { best = (pcmk_node_t *) nodes->data; // First node has best score } if ((prefer != NULL) && (nodes != NULL)) { // Get the allowed node version of prefer chosen = g_hash_table_lookup(rsc->allowed_nodes, prefer->details->id); if (chosen == NULL) { pe_rsc_trace(rsc, "Preferred node %s for %s was unknown", pe__node_name(prefer), rsc->id); /* Favor the preferred node as long as its score is at least as good as * the best allowed node's. * * An alternative would be to favor the preferred node even if the best * node is better, when the best node's score is less than INFINITY. */ } else if (chosen->weight < best->weight) { pe_rsc_trace(rsc, "Preferred node %s for %s was unsuitable", pe__node_name(chosen), rsc->id); chosen = NULL; } else if (!pcmk__node_available(chosen, true, false)) { pe_rsc_trace(rsc, "Preferred node %s for %s was unavailable", pe__node_name(chosen), rsc->id); chosen = NULL; } else { pe_rsc_trace(rsc, "Chose preferred node %s for %s " "(ignoring %d candidates)", pe__node_name(chosen), rsc->id, g_list_length(nodes)); } } if ((chosen == NULL) && (best != NULL)) { /* Either there is no preferred node, or the preferred node is not * suitable, but another node is allowed to run the resource. */ chosen = best; if (!pe_rsc_is_unique_clone(rsc->parent) && (chosen->weight > 0) // Zero not acceptable && pcmk__node_available(chosen, false, false)) { /* If the resource is already running on a node, prefer that node if * it is just as good as the chosen node. * * We don't do this for unique clone instances, because * pcmk__assign_instances() has already assigned instances to their * running nodes when appropriate, and if we get here, we don't want * remaining unassigned instances to prefer a node that's already * running another instance. */ pcmk_node_t *running = pe__current_node(rsc); if (running == NULL) { // Nothing to do } else if (!pcmk__node_available(running, true, false)) { pe_rsc_trace(rsc, "Current node for %s (%s) can't run resources", rsc->id, pe__node_name(running)); } else { int nodes_with_best_score = 1; for (GList *iter = nodes->next; iter; iter = iter->next) { pcmk_node_t *allowed = (pcmk_node_t *) iter->data; if (allowed->weight != chosen->weight) { // The nodes are sorted by score, so no more are equal break; } if (pe__same_node(allowed, running)) { // Scores are equal, so prefer the current node chosen = allowed; } nodes_with_best_score++; } if (nodes_with_best_score > 1) { uint8_t log_level = LOG_INFO; if (chosen->weight >= INFINITY) { log_level = LOG_WARNING; } do_crm_log(log_level, "Chose %s for %s from %d nodes with score %s", pe__node_name(chosen), rsc->id, nodes_with_best_score, pcmk_readable_score(chosen->weight)); } } } pe_rsc_trace(rsc, "Chose %s for %s from %d candidates", pe__node_name(chosen), rsc->id, g_list_length(nodes)); } pcmk__assign_resource(rsc, chosen, false, stop_if_fail); g_list_free(nodes); return rsc->allocated_to != NULL; } /*! * \internal * \brief Apply a "this with" colocation to a node's allowed node scores * * \param[in,out] colocation Colocation to apply * \param[in,out] rsc Resource being assigned */ static void apply_this_with(pcmk__colocation_t *colocation, pcmk_resource_t *rsc) { GHashTable *archive = NULL; pcmk_resource_t *other = colocation->primary; // In certain cases, we will need to revert the node scores if ((colocation->dependent_role >= pcmk_role_promoted) || ((colocation->score < 0) && (colocation->score > -INFINITY))) { archive = pcmk__copy_node_table(rsc->allowed_nodes); } if (pcmk_is_set(other->flags, pcmk_rsc_unassigned)) { pe_rsc_trace(rsc, "%s: Assigning colocation %s primary %s first" "(score=%d role=%s)", rsc->id, colocation->id, other->id, colocation->score, role2text(colocation->dependent_role)); other->cmds->assign(other, NULL, true); } // Apply the colocation score to this resource's allowed node scores rsc->cmds->apply_coloc_score(rsc, other, colocation, true); if ((archive != NULL) && !pcmk__any_node_available(rsc->allowed_nodes)) { pcmk__rsc_info(rsc, "%s: Reverting scores from colocation with %s " "because no nodes allowed", rsc->id, other->id); g_hash_table_destroy(rsc->allowed_nodes); rsc->allowed_nodes = archive; archive = NULL; } if (archive != NULL) { g_hash_table_destroy(archive); } } /*! * \internal * \brief Update a Pacemaker Remote node once its connection has been assigned * * \param[in] connection Connection resource that has been assigned */ static void remote_connection_assigned(const pcmk_resource_t *connection) { pcmk_node_t *remote_node = pe_find_node(connection->cluster->nodes, connection->id); CRM_CHECK(remote_node != NULL, return); if ((connection->allocated_to != NULL) && (connection->next_role != pcmk_role_stopped)) { crm_trace("Pacemaker Remote node %s will be online", remote_node->details->id); remote_node->details->online = TRUE; if (remote_node->details->unseen) { // Avoid unnecessary fence, since we will attempt connection remote_node->details->unclean = FALSE; } } else { crm_trace("Pacemaker Remote node %s will be shut down " "(%sassigned connection's next role is %s)", remote_node->details->id, ((connection->allocated_to == NULL)? "un" : ""), role2text(connection->next_role)); remote_node->details->shutdown = TRUE; } } /*! * \internal * \brief Assign a primitive resource to a node * * \param[in,out] rsc Resource to assign to a node * \param[in] prefer Node to prefer, if all else is equal * \param[in] stop_if_fail If \c true and \p rsc can't be assigned to a * node, set next role to stopped and update * existing actions * * \return Node that \p rsc is assigned to, if assigned entirely to one node * * \note If \p stop_if_fail is \c false, then \c pcmk__unassign_resource() can * completely undo the assignment. A successful assignment can be either * undone or left alone as final. A failed assignment has the same effect * as calling pcmk__unassign_resource(); there are no side effects on * roles or actions. */ pcmk_node_t * pcmk__primitive_assign(pcmk_resource_t *rsc, const pcmk_node_t *prefer, bool stop_if_fail) { GList *this_with_colocations = NULL; GList *with_this_colocations = NULL; GList *iter = NULL; pcmk__colocation_t *colocation = NULL; CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_primitive)); // Never assign a child without parent being assigned first if ((rsc->parent != NULL) && !pcmk_is_set(rsc->parent->flags, pcmk_rsc_assigning)) { - pe_rsc_debug(rsc, "%s: Assigning parent %s first", - rsc->id, rsc->parent->id); + pcmk__rsc_debug(rsc, "%s: Assigning parent %s first", + rsc->id, rsc->parent->id); rsc->parent->cmds->assign(rsc->parent, prefer, stop_if_fail); } if (!pcmk_is_set(rsc->flags, pcmk_rsc_unassigned)) { // Assignment has already been done const char *node_name = "no node"; if (rsc->allocated_to != NULL) { node_name = pe__node_name(rsc->allocated_to); } - pe_rsc_debug(rsc, "%s: pre-assigned to %s", rsc->id, node_name); + pcmk__rsc_debug(rsc, "%s: pre-assigned to %s", rsc->id, node_name); return rsc->allocated_to; } // Ensure we detect assignment loops if (pcmk_is_set(rsc->flags, pcmk_rsc_assigning)) { - pe_rsc_debug(rsc, "Breaking assignment loop involving %s", rsc->id); + pcmk__rsc_debug(rsc, "Breaking assignment loop involving %s", rsc->id); return NULL; } pe__set_resource_flags(rsc, pcmk_rsc_assigning); pe__show_node_scores(true, rsc, "Pre-assignment", rsc->allowed_nodes, rsc->cluster); this_with_colocations = pcmk__this_with_colocations(rsc); with_this_colocations = pcmk__with_this_colocations(rsc); // Apply mandatory colocations first, to satisfy as many as possible for (iter = this_with_colocations; iter != NULL; iter = iter->next) { colocation = iter->data; if ((colocation->score <= -CRM_SCORE_INFINITY) || (colocation->score >= CRM_SCORE_INFINITY)) { apply_this_with(colocation, rsc); } } for (iter = with_this_colocations; iter != NULL; iter = iter->next) { colocation = iter->data; if ((colocation->score <= -CRM_SCORE_INFINITY) || (colocation->score >= CRM_SCORE_INFINITY)) { pcmk__add_dependent_scores(colocation, rsc); } } pe__show_node_scores(true, rsc, "Mandatory-colocations", rsc->allowed_nodes, rsc->cluster); // Then apply optional colocations for (iter = this_with_colocations; iter != NULL; iter = iter->next) { colocation = iter->data; if ((colocation->score > -CRM_SCORE_INFINITY) && (colocation->score < CRM_SCORE_INFINITY)) { apply_this_with(colocation, rsc); } } for (iter = with_this_colocations; iter != NULL; iter = iter->next) { colocation = iter->data; if ((colocation->score > -CRM_SCORE_INFINITY) && (colocation->score < CRM_SCORE_INFINITY)) { pcmk__add_dependent_scores(colocation, rsc); } } g_list_free(this_with_colocations); g_list_free(with_this_colocations); if (rsc->next_role == pcmk_role_stopped) { pe_rsc_trace(rsc, "Banning %s from all nodes because it will be stopped", rsc->id); resource_location(rsc, NULL, -INFINITY, XML_RSC_ATTR_TARGET_ROLE, rsc->cluster); } else if ((rsc->next_role > rsc->role) && !pcmk_is_set(rsc->cluster->flags, pcmk_sched_quorate) && (rsc->cluster->no_quorum_policy == pcmk_no_quorum_freeze)) { crm_notice("Resource %s cannot be elevated from %s to %s due to " "no-quorum-policy=freeze", rsc->id, role2text(rsc->role), role2text(rsc->next_role)); pe__set_next_role(rsc, rsc->role, "no-quorum-policy=freeze"); } pe__show_node_scores(!pcmk_is_set(rsc->cluster->flags, pcmk_sched_output_scores), rsc, __func__, rsc->allowed_nodes, rsc->cluster); // Unmanage resource if fencing is enabled but no device is configured if (pcmk_is_set(rsc->cluster->flags, pcmk_sched_fencing_enabled) && !pcmk_is_set(rsc->cluster->flags, pcmk_sched_have_fencing)) { pe__clear_resource_flags(rsc, pcmk_rsc_managed); } if (!pcmk_is_set(rsc->flags, pcmk_rsc_managed)) { // Unmanaged resources stay on their current node const char *reason = NULL; pcmk_node_t *assign_to = NULL; pe__set_next_role(rsc, rsc->role, "unmanaged"); assign_to = pe__current_node(rsc); if (assign_to == NULL) { reason = "inactive"; } else if (rsc->role == pcmk_role_promoted) { reason = "promoted"; } else if (pcmk_is_set(rsc->flags, pcmk_rsc_failed)) { reason = "failed"; } else { reason = "active"; } pcmk__rsc_info(rsc, "Unmanaged resource %s assigned to %s: %s", rsc->id, (assign_to? assign_to->details->uname : "no node"), reason); pcmk__assign_resource(rsc, assign_to, true, stop_if_fail); } else if (pcmk_is_set(rsc->cluster->flags, pcmk_sched_stop_all)) { // Must stop at some point, but be consistent with stop_if_fail if (stop_if_fail) { - pe_rsc_debug(rsc, "Forcing %s to stop: stop-all-resources", - rsc->id); + pcmk__rsc_debug(rsc, "Forcing %s to stop: stop-all-resources", + rsc->id); } pcmk__assign_resource(rsc, NULL, true, stop_if_fail); } else if (!assign_best_node(rsc, prefer, stop_if_fail)) { // Assignment failed if (!pcmk_is_set(rsc->flags, pcmk_rsc_removed)) { pcmk__rsc_info(rsc, "Resource %s cannot run anywhere", rsc->id); } else if ((rsc->running_on != NULL) && stop_if_fail) { pcmk__rsc_info(rsc, "Stopping orphan resource %s", rsc->id); } } pe__clear_resource_flags(rsc, pcmk_rsc_assigning); if (rsc->is_remote_node) { remote_connection_assigned(rsc); } return rsc->allocated_to; } /*! * \internal * \brief Schedule actions to bring resource down and back to current role * * \param[in,out] rsc Resource to restart * \param[in,out] current Node that resource should be brought down on * \param[in] need_stop Whether the resource must be stopped * \param[in] need_promote Whether the resource must be promoted * * \return Role that resource would have after scheduled actions are taken */ static void schedule_restart_actions(pcmk_resource_t *rsc, pcmk_node_t *current, bool need_stop, bool need_promote) { enum rsc_role_e role = rsc->role; enum rsc_role_e next_role; rsc_transition_fn fn = NULL; pe__set_resource_flags(rsc, pcmk_rsc_restarting); // Bring resource down to a stop on its current node while (role != pcmk_role_stopped) { next_role = rsc_state_matrix[role][pcmk_role_stopped]; pe_rsc_trace(rsc, "Creating %s action to take %s down from %s to %s", (need_stop? "required" : "optional"), rsc->id, role2text(role), role2text(next_role)); fn = rsc_action_matrix[role][next_role]; if (fn == NULL) { break; } fn(rsc, current, !need_stop); role = next_role; } // Bring resource up to its next role on its next node while ((rsc->role <= rsc->next_role) && (role != rsc->role) && !pcmk_is_set(rsc->flags, pcmk_rsc_blocked)) { bool required = need_stop; next_role = rsc_state_matrix[role][rsc->role]; if ((next_role == pcmk_role_promoted) && need_promote) { required = true; } pe_rsc_trace(rsc, "Creating %s action to take %s up from %s to %s", (required? "required" : "optional"), rsc->id, role2text(role), role2text(next_role)); fn = rsc_action_matrix[role][next_role]; if (fn == NULL) { break; } fn(rsc, rsc->allocated_to, !required); role = next_role; } pe__clear_resource_flags(rsc, pcmk_rsc_restarting); } /*! * \internal * \brief If a resource's next role is not explicitly specified, set a default * * \param[in,out] rsc Resource to set next role for * * \return "explicit" if next role was explicitly set, otherwise "implicit" */ static const char * set_default_next_role(pcmk_resource_t *rsc) { if (rsc->next_role != pcmk_role_unknown) { return "explicit"; } if (rsc->allocated_to == NULL) { pe__set_next_role(rsc, pcmk_role_stopped, "assignment"); } else { pe__set_next_role(rsc, pcmk_role_started, "assignment"); } return "implicit"; } /*! * \internal * \brief Create an action to represent an already pending start * * \param[in,out] rsc Resource to create start action for */ static void create_pending_start(pcmk_resource_t *rsc) { pcmk_action_t *start = NULL; pe_rsc_trace(rsc, "Creating action for %s to represent already pending start", rsc->id); start = start_action(rsc, rsc->allocated_to, TRUE); pe__set_action_flags(start, pcmk_action_always_in_graph); } /*! * \internal * \brief Schedule actions needed to take a resource to its next role * * \param[in,out] rsc Resource to schedule actions for */ static void schedule_role_transition_actions(pcmk_resource_t *rsc) { enum rsc_role_e role = rsc->role; while (role != rsc->next_role) { enum rsc_role_e next_role = rsc_state_matrix[role][rsc->next_role]; rsc_transition_fn fn = NULL; pe_rsc_trace(rsc, "Creating action to take %s from %s to %s (ending at %s)", rsc->id, role2text(role), role2text(next_role), role2text(rsc->next_role)); fn = rsc_action_matrix[role][next_role]; if (fn == NULL) { break; } fn(rsc, rsc->allocated_to, false); role = next_role; } } /*! * \internal * \brief Create all actions needed for a given primitive resource * * \param[in,out] rsc Primitive resource to create actions for */ void pcmk__primitive_create_actions(pcmk_resource_t *rsc) { bool need_stop = false; bool need_promote = false; bool is_moving = false; bool allow_migrate = false; bool multiply_active = false; pcmk_node_t *current = NULL; unsigned int num_all_active = 0; unsigned int num_clean_active = 0; const char *next_role_source = NULL; CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_primitive)); next_role_source = set_default_next_role(rsc); pe_rsc_trace(rsc, "Creating all actions for %s transition from %s to %s " "(%s) on %s", rsc->id, role2text(rsc->role), role2text(rsc->next_role), next_role_source, pe__node_name(rsc->allocated_to)); current = rsc->fns->active_node(rsc, &num_all_active, &num_clean_active); g_list_foreach(rsc->dangling_migrations, pcmk__abort_dangling_migration, rsc); if ((current != NULL) && (rsc->allocated_to != NULL) && !pe__same_node(current, rsc->allocated_to) && (rsc->next_role >= pcmk_role_started)) { pe_rsc_trace(rsc, "Moving %s from %s to %s", rsc->id, pe__node_name(current), pe__node_name(rsc->allocated_to)); is_moving = true; allow_migrate = pcmk__rsc_can_migrate(rsc, current); // This is needed even if migrating (though I'm not sure why ...) need_stop = true; } // Check whether resource is partially migrated and/or multiply active if ((rsc->partial_migration_source != NULL) && (rsc->partial_migration_target != NULL) && allow_migrate && (num_all_active == 2) && pe__same_node(current, rsc->partial_migration_source) && pe__same_node(rsc->allocated_to, rsc->partial_migration_target)) { /* A partial migration is in progress, and the migration target remains * the same as when the migration began. */ pe_rsc_trace(rsc, "Partial migration of %s from %s to %s will continue", rsc->id, pe__node_name(rsc->partial_migration_source), pe__node_name(rsc->partial_migration_target)); } else if ((rsc->partial_migration_source != NULL) || (rsc->partial_migration_target != NULL)) { // A partial migration is in progress but can't be continued if (num_all_active > 2) { // The resource is migrating *and* multiply active! crm_notice("Forcing recovery of %s because it is migrating " "from %s to %s and possibly active elsewhere", rsc->id, pe__node_name(rsc->partial_migration_source), pe__node_name(rsc->partial_migration_target)); } else { // The migration source or target isn't available crm_notice("Forcing recovery of %s because it can no longer " "migrate from %s to %s", rsc->id, pe__node_name(rsc->partial_migration_source), pe__node_name(rsc->partial_migration_target)); } need_stop = true; rsc->partial_migration_source = rsc->partial_migration_target = NULL; allow_migrate = false; } else if (pcmk_is_set(rsc->flags, pcmk_rsc_needs_fencing)) { multiply_active = (num_all_active > 1); } else { /* If a resource has "requires" set to nothing or quorum, don't consider * it active on unclean nodes (similar to how all resources behave when * stonith-enabled is false). We can start such resources elsewhere * before fencing completes, and if we considered the resource active on * the failed node, we would attempt recovery for being active on * multiple nodes. */ multiply_active = (num_clean_active > 1); } if (multiply_active) { const char *class = crm_element_value(rsc->xml, XML_AGENT_ATTR_CLASS); // Resource was (possibly) incorrectly multiply active pe_proc_err("%s resource %s might be active on %u nodes (%s)", pcmk__s(class, "Untyped"), rsc->id, num_all_active, recovery2text(rsc->recovery_type)); crm_notice("See https://wiki.clusterlabs.org/wiki/FAQ" "#Resource_is_Too_Active for more information"); switch (rsc->recovery_type) { case pcmk_multiply_active_restart: need_stop = true; break; case pcmk_multiply_active_unexpected: need_stop = true; // stop_resource() will skip expected node pe__set_resource_flags(rsc, pcmk_rsc_stop_unexpected); break; default: break; } } else { pe__clear_resource_flags(rsc, pcmk_rsc_stop_unexpected); } if (pcmk_is_set(rsc->flags, pcmk_rsc_start_pending)) { create_pending_start(rsc); } if (is_moving) { // Remaining tests are only for resources staying where they are } else if (pcmk_is_set(rsc->flags, pcmk_rsc_failed)) { if (pcmk_is_set(rsc->flags, pcmk_rsc_stop_if_failed)) { need_stop = true; pe_rsc_trace(rsc, "Recovering %s", rsc->id); } else { pe_rsc_trace(rsc, "Recovering %s by demotion", rsc->id); if (rsc->next_role == pcmk_role_promoted) { need_promote = true; } } } else if (pcmk_is_set(rsc->flags, pcmk_rsc_blocked)) { pe_rsc_trace(rsc, "Blocking further actions on %s", rsc->id); need_stop = true; } else if ((rsc->role > pcmk_role_started) && (current != NULL) && (rsc->allocated_to != NULL)) { pcmk_action_t *start = NULL; pe_rsc_trace(rsc, "Creating start action for promoted resource %s", rsc->id); start = start_action(rsc, rsc->allocated_to, TRUE); if (!pcmk_is_set(start->flags, pcmk_action_optional)) { // Recovery of a promoted resource pe_rsc_trace(rsc, "%s restart is required for recovery", rsc->id); need_stop = true; } } // Create any actions needed to bring resource down and back up to same role schedule_restart_actions(rsc, current, need_stop, need_promote); // Create any actions needed to take resource from this role to the next schedule_role_transition_actions(rsc); pcmk__create_recurring_actions(rsc); if (allow_migrate) { pcmk__create_migration_actions(rsc, current); } } /*! * \internal * \brief Ban a resource from any allowed nodes that are Pacemaker Remote nodes * * \param[in] rsc Resource to check */ static void rsc_avoids_remote_nodes(const pcmk_resource_t *rsc) { GHashTableIter iter; pcmk_node_t *node = NULL; g_hash_table_iter_init(&iter, rsc->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (void **) &node)) { if (node->details->remote_rsc != NULL) { node->weight = -INFINITY; } } } /*! * \internal * \brief Return allowed nodes as (possibly sorted) list * * Convert a resource's hash table of allowed nodes to a list. If printing to * stdout, sort the list, to keep action ID numbers consistent for regression * test output (while avoiding the performance hit on a live cluster). * * \param[in] rsc Resource to check for allowed nodes * * \return List of resource's allowed nodes * \note Callers should take care not to rely on the list being sorted. */ static GList * allowed_nodes_as_list(const pcmk_resource_t *rsc) { GList *allowed_nodes = NULL; if (rsc->allowed_nodes) { allowed_nodes = g_hash_table_get_values(rsc->allowed_nodes); } if (!pcmk__is_daemon) { allowed_nodes = g_list_sort(allowed_nodes, pe__cmp_node_name); } return allowed_nodes; } /*! * \internal * \brief Create implicit constraints needed for a primitive resource * * \param[in,out] rsc Primitive resource to create implicit constraints for */ void pcmk__primitive_internal_constraints(pcmk_resource_t *rsc) { GList *allowed_nodes = NULL; bool check_unfencing = false; bool check_utilization = false; CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_primitive)); if (!pcmk_is_set(rsc->flags, pcmk_rsc_managed)) { pe_rsc_trace(rsc, "Skipping implicit constraints for unmanaged resource %s", rsc->id); return; } // Whether resource requires unfencing check_unfencing = !pcmk_is_set(rsc->flags, pcmk_rsc_fence_device) && pcmk_is_set(rsc->cluster->flags, pcmk_sched_enable_unfencing) && pcmk_is_set(rsc->flags, pcmk_rsc_needs_unfencing); // Whether a non-default placement strategy is used check_utilization = (g_hash_table_size(rsc->utilization) > 0) && !pcmk__str_eq(rsc->cluster->placement_strategy, "default", pcmk__str_casei); // Order stops before starts (i.e. restart) pcmk__new_ordering(rsc, pcmk__op_key(rsc->id, PCMK_ACTION_STOP, 0), NULL, rsc, pcmk__op_key(rsc->id, PCMK_ACTION_START, 0), NULL, pcmk__ar_ordered |pcmk__ar_first_implies_then |pcmk__ar_intermediate_stop, rsc->cluster); // Promotable ordering: demote before stop, start before promote if (pcmk_is_set(pe__const_top_resource(rsc, false)->flags, pcmk_rsc_promotable) || (rsc->role > pcmk_role_unpromoted)) { pcmk__new_ordering(rsc, pcmk__op_key(rsc->id, PCMK_ACTION_DEMOTE, 0), NULL, rsc, pcmk__op_key(rsc->id, PCMK_ACTION_STOP, 0), NULL, pcmk__ar_promoted_then_implies_first, rsc->cluster); pcmk__new_ordering(rsc, pcmk__op_key(rsc->id, PCMK_ACTION_START, 0), NULL, rsc, pcmk__op_key(rsc->id, PCMK_ACTION_PROMOTE, 0), NULL, pcmk__ar_unrunnable_first_blocks, rsc->cluster); } // Don't clear resource history if probing on same node pcmk__new_ordering(rsc, pcmk__op_key(rsc->id, PCMK_ACTION_LRM_DELETE, 0), NULL, rsc, pcmk__op_key(rsc->id, PCMK_ACTION_MONITOR, 0), NULL, pcmk__ar_if_on_same_node|pcmk__ar_then_cancels_first, rsc->cluster); // Certain checks need allowed nodes if (check_unfencing || check_utilization || (rsc->container != NULL)) { allowed_nodes = allowed_nodes_as_list(rsc); } if (check_unfencing) { g_list_foreach(allowed_nodes, pcmk__order_restart_vs_unfence, rsc); } if (check_utilization) { pcmk__create_utilization_constraints(rsc, allowed_nodes); } if (rsc->container != NULL) { pcmk_resource_t *remote_rsc = NULL; if (rsc->is_remote_node) { // rsc is the implicit remote connection for a guest or bundle node /* Guest resources are not allowed to run on Pacemaker Remote nodes, * to avoid nesting remotes. However, bundles are allowed. */ if (!pcmk_is_set(rsc->flags, pcmk_rsc_remote_nesting_allowed)) { rsc_avoids_remote_nodes(rsc->container); } /* If someone cleans up a guest or bundle node's container, we will * likely schedule a (re-)probe of the container and recovery of the * connection. Order the connection stop after the container probe, * so that if we detect the container running, we will trigger a new * transition and avoid the unnecessary recovery. */ pcmk__order_resource_actions(rsc->container, PCMK_ACTION_MONITOR, rsc, PCMK_ACTION_STOP, pcmk__ar_ordered); /* A user can specify that a resource must start on a Pacemaker Remote * node by explicitly configuring it with the container=NODENAME * meta-attribute. This is of questionable merit, since location * constraints can accomplish the same thing. But we support it, so here * we check whether a resource (that is not itself a remote connection) * has container set to a remote node or guest node resource. */ } else if (rsc->container->is_remote_node) { remote_rsc = rsc->container; } else { remote_rsc = pe__resource_contains_guest_node(rsc->cluster, rsc->container); } if (remote_rsc != NULL) { /* Force the resource on the Pacemaker Remote node instead of * colocating the resource with the container resource. */ for (GList *item = allowed_nodes; item; item = item->next) { pcmk_node_t *node = item->data; if (node->details->remote_rsc != remote_rsc) { node->weight = -INFINITY; } } } else { /* This resource is either a filler for a container that does NOT * represent a Pacemaker Remote node, or a Pacemaker Remote * connection resource for a guest node or bundle. */ int score; crm_trace("Order and colocate %s relative to its container %s", rsc->id, rsc->container->id); pcmk__new_ordering(rsc->container, pcmk__op_key(rsc->container->id, PCMK_ACTION_START, 0), NULL, rsc, pcmk__op_key(rsc->id, PCMK_ACTION_START, 0), NULL, pcmk__ar_first_implies_then |pcmk__ar_unrunnable_first_blocks, rsc->cluster); pcmk__new_ordering(rsc, pcmk__op_key(rsc->id, PCMK_ACTION_STOP, 0), NULL, rsc->container, pcmk__op_key(rsc->container->id, PCMK_ACTION_STOP, 0), NULL, pcmk__ar_then_implies_first, rsc->cluster); if (pcmk_is_set(rsc->flags, pcmk_rsc_remote_nesting_allowed)) { score = 10000; /* Highly preferred but not essential */ } else { score = INFINITY; /* Force them to run on the same host */ } pcmk__new_colocation("#resource-with-container", NULL, score, rsc, rsc->container, NULL, NULL, pcmk__coloc_influence); } } if (rsc->is_remote_node || pcmk_is_set(rsc->flags, pcmk_rsc_fence_device)) { /* Remote connections and fencing devices are not allowed to run on * Pacemaker Remote nodes */ rsc_avoids_remote_nodes(rsc); } g_list_free(allowed_nodes); } /*! * \internal * \brief Apply a colocation's score to node scores or resource priority * * Given a colocation constraint, apply its score to the dependent's * allowed node scores (if we are still placing resources) or priority (if * we are choosing promotable clone instance roles). * * \param[in,out] dependent Dependent resource in colocation * \param[in] primary Primary resource in colocation * \param[in] colocation Colocation constraint to apply * \param[in] for_dependent true if called on behalf of dependent */ void pcmk__primitive_apply_coloc_score(pcmk_resource_t *dependent, const pcmk_resource_t *primary, const pcmk__colocation_t *colocation, bool for_dependent) { enum pcmk__coloc_affects filter_results; CRM_ASSERT((dependent != NULL) && (primary != NULL) && (colocation != NULL)); if (for_dependent) { // Always process on behalf of primary resource primary->cmds->apply_coloc_score(dependent, primary, colocation, false); return; } filter_results = pcmk__colocation_affects(dependent, primary, colocation, false); pe_rsc_trace(dependent, "%s %s with %s (%s, score=%d, filter=%d)", ((colocation->score > 0)? "Colocating" : "Anti-colocating"), dependent->id, primary->id, colocation->id, colocation->score, filter_results); switch (filter_results) { case pcmk__coloc_affects_role: pcmk__apply_coloc_to_priority(dependent, primary, colocation); break; case pcmk__coloc_affects_location: pcmk__apply_coloc_to_scores(dependent, primary, colocation); break; default: // pcmk__coloc_affects_nothing return; } } /* Primitive implementation of * pcmk_assignment_methods_t:with_this_colocations() */ void pcmk__with_primitive_colocations(const pcmk_resource_t *rsc, const pcmk_resource_t *orig_rsc, GList **list) { CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_primitive) && (list != NULL)); if (rsc == orig_rsc) { /* For the resource itself, add all of its own colocations and relevant * colocations from its parent (if any). */ pcmk__add_with_this_list(list, rsc->rsc_cons_lhs, orig_rsc); if (rsc->parent != NULL) { rsc->parent->cmds->with_this_colocations(rsc->parent, orig_rsc, list); } } else { // For an ancestor, add only explicitly configured constraints for (GList *iter = rsc->rsc_cons_lhs; iter != NULL; iter = iter->next) { pcmk__colocation_t *colocation = iter->data; if (pcmk_is_set(colocation->flags, pcmk__coloc_explicit)) { pcmk__add_with_this(list, colocation, orig_rsc); } } } } /* Primitive implementation of * pcmk_assignment_methods_t:this_with_colocations() */ void pcmk__primitive_with_colocations(const pcmk_resource_t *rsc, const pcmk_resource_t *orig_rsc, GList **list) { CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_primitive) && (list != NULL)); if (rsc == orig_rsc) { /* For the resource itself, add all of its own colocations and relevant * colocations from its parent (if any). */ pcmk__add_this_with_list(list, rsc->rsc_cons, orig_rsc); if (rsc->parent != NULL) { rsc->parent->cmds->this_with_colocations(rsc->parent, orig_rsc, list); } } else { // For an ancestor, add only explicitly configured constraints for (GList *iter = rsc->rsc_cons; iter != NULL; iter = iter->next) { pcmk__colocation_t *colocation = iter->data; if (pcmk_is_set(colocation->flags, pcmk__coloc_explicit)) { pcmk__add_this_with(list, colocation, orig_rsc); } } } } /*! * \internal * \brief Return action flags for a given primitive resource action * * \param[in,out] action Action to get flags for * \param[in] node If not NULL, limit effects to this node (ignored) * * \return Flags appropriate to \p action on \p node */ uint32_t pcmk__primitive_action_flags(pcmk_action_t *action, const pcmk_node_t *node) { CRM_ASSERT(action != NULL); return (uint32_t) action->flags; } /*! * \internal * \brief Check whether a node is a multiply active resource's expected node * * \param[in] rsc Resource to check * \param[in] node Node to check * * \return true if \p rsc is multiply active with multiple-active set to * stop_unexpected, and \p node is the node where it will remain active * \note This assumes that the resource's next role cannot be changed to stopped * after this is called, which should be reasonable if status has already * been unpacked and resources have been assigned to nodes. */ static bool is_expected_node(const pcmk_resource_t *rsc, const pcmk_node_t *node) { return pcmk_all_flags_set(rsc->flags, pcmk_rsc_stop_unexpected|pcmk_rsc_restarting) && (rsc->next_role > pcmk_role_stopped) && pe__same_node(rsc->allocated_to, node); } /*! * \internal * \brief Schedule actions needed to stop a resource wherever it is active * * \param[in,out] rsc Resource being stopped * \param[in] node Node where resource is being stopped (ignored) * \param[in] optional Whether actions should be optional */ static void stop_resource(pcmk_resource_t *rsc, pcmk_node_t *node, bool optional) { for (GList *iter = rsc->running_on; iter != NULL; iter = iter->next) { pcmk_node_t *current = (pcmk_node_t *) iter->data; pcmk_action_t *stop = NULL; if (is_expected_node(rsc, current)) { /* We are scheduling restart actions for a multiply active resource * with multiple-active=stop_unexpected, and this is where it should * not be stopped. */ pe_rsc_trace(rsc, "Skipping stop of multiply active resource %s " "on expected node %s", rsc->id, pe__node_name(current)); continue; } if (rsc->partial_migration_target != NULL) { // Continue migration if node originally was and remains target if (pe__same_node(current, rsc->partial_migration_target) && pe__same_node(current, rsc->allocated_to)) { pe_rsc_trace(rsc, "Skipping stop of %s on %s " "because partial migration there will continue", rsc->id, pe__node_name(current)); continue; } else { pe_rsc_trace(rsc, "Forcing stop of %s on %s " "because migration target changed", rsc->id, pe__node_name(current)); optional = false; } } pe_rsc_trace(rsc, "Scheduling stop of %s on %s", rsc->id, pe__node_name(current)); stop = stop_action(rsc, current, optional); if (rsc->allocated_to == NULL) { pe_action_set_reason(stop, "node availability", true); } else if (pcmk_all_flags_set(rsc->flags, pcmk_rsc_restarting |pcmk_rsc_stop_unexpected)) { /* We are stopping a multiply active resource on a node that is * not its expected node, and we are still scheduling restart * actions, so the stop is for being multiply active. */ pe_action_set_reason(stop, "being multiply active", true); } if (!pcmk_is_set(rsc->flags, pcmk_rsc_managed)) { pe__clear_action_flags(stop, pcmk_action_runnable); } if (pcmk_is_set(rsc->cluster->flags, pcmk_sched_remove_after_stop)) { pcmk__schedule_cleanup(rsc, current, optional); } if (pcmk_is_set(rsc->flags, pcmk_rsc_needs_unfencing)) { pcmk_action_t *unfence = pe_fence_op(current, PCMK_ACTION_ON, true, NULL, false, rsc->cluster); order_actions(stop, unfence, pcmk__ar_then_implies_first); if (!pcmk__node_unfenced(current)) { pe_proc_err("Stopping %s until %s can be unfenced", rsc->id, pe__node_name(current)); } } } } /*! * \internal * \brief Schedule actions needed to start a resource on a node * * \param[in,out] rsc Resource being started * \param[in,out] node Node where resource should be started * \param[in] optional Whether actions should be optional */ static void start_resource(pcmk_resource_t *rsc, pcmk_node_t *node, bool optional) { pcmk_action_t *start = NULL; CRM_ASSERT(node != NULL); pe_rsc_trace(rsc, "Scheduling %s start of %s on %s (score %d)", (optional? "optional" : "required"), rsc->id, pe__node_name(node), node->weight); start = start_action(rsc, node, TRUE); pcmk__order_vs_unfence(rsc, node, start, pcmk__ar_first_implies_then); if (pcmk_is_set(start->flags, pcmk_action_runnable) && !optional) { pe__clear_action_flags(start, pcmk_action_optional); } if (is_expected_node(rsc, node)) { /* This could be a problem if the start becomes necessary for other * reasons later. */ pe_rsc_trace(rsc, "Start of multiply active resouce %s " "on expected node %s will be a pseudo-action", rsc->id, pe__node_name(node)); pe__set_action_flags(start, pcmk_action_pseudo); } } /*! * \internal * \brief Schedule actions needed to promote a resource on a node * * \param[in,out] rsc Resource being promoted * \param[in] node Node where resource should be promoted * \param[in] optional Whether actions should be optional */ static void promote_resource(pcmk_resource_t *rsc, pcmk_node_t *node, bool optional) { GList *iter = NULL; GList *action_list = NULL; bool runnable = true; CRM_ASSERT(node != NULL); // Any start must be runnable for promotion to be runnable action_list = pe__resource_actions(rsc, node, PCMK_ACTION_START, true); for (iter = action_list; iter != NULL; iter = iter->next) { pcmk_action_t *start = (pcmk_action_t *) iter->data; if (!pcmk_is_set(start->flags, pcmk_action_runnable)) { runnable = false; } } g_list_free(action_list); if (runnable) { pcmk_action_t *promote = promote_action(rsc, node, optional); pe_rsc_trace(rsc, "Scheduling %s promotion of %s on %s", (optional? "optional" : "required"), rsc->id, pe__node_name(node)); if (is_expected_node(rsc, node)) { /* This could be a problem if the promote becomes necessary for * other reasons later. */ pe_rsc_trace(rsc, "Promotion of multiply active resouce %s " "on expected node %s will be a pseudo-action", rsc->id, pe__node_name(node)); pe__set_action_flags(promote, pcmk_action_pseudo); } } else { pe_rsc_trace(rsc, "Not promoting %s on %s: start unrunnable", rsc->id, pe__node_name(node)); action_list = pe__resource_actions(rsc, node, PCMK_ACTION_PROMOTE, true); for (iter = action_list; iter != NULL; iter = iter->next) { pcmk_action_t *promote = (pcmk_action_t *) iter->data; pe__clear_action_flags(promote, pcmk_action_runnable); } g_list_free(action_list); } } /*! * \internal * \brief Schedule actions needed to demote a resource wherever it is active * * \param[in,out] rsc Resource being demoted * \param[in] node Node where resource should be demoted (ignored) * \param[in] optional Whether actions should be optional */ static void demote_resource(pcmk_resource_t *rsc, pcmk_node_t *node, bool optional) { /* Since this will only be called for a primitive (possibly as an instance * of a collective resource), the resource is multiply active if it is * running on more than one node, so we want to demote on all of them as * part of recovery, regardless of which one is the desired node. */ for (GList *iter = rsc->running_on; iter != NULL; iter = iter->next) { pcmk_node_t *current = (pcmk_node_t *) iter->data; if (is_expected_node(rsc, current)) { pe_rsc_trace(rsc, "Skipping demote of multiply active resource %s " "on expected node %s", rsc->id, pe__node_name(current)); } else { pe_rsc_trace(rsc, "Scheduling %s demotion of %s on %s", (optional? "optional" : "required"), rsc->id, pe__node_name(current)); demote_action(rsc, current, optional); } } } static void assert_role_error(pcmk_resource_t *rsc, pcmk_node_t *node, bool optional) { CRM_ASSERT(false); } /*! * \internal * \brief Schedule cleanup of a resource * * \param[in,out] rsc Resource to clean up * \param[in] node Node to clean up on * \param[in] optional Whether clean-up should be optional */ void pcmk__schedule_cleanup(pcmk_resource_t *rsc, const pcmk_node_t *node, bool optional) { /* If the cleanup is required, its orderings are optional, because they're * relevant only if both actions are required. Conversely, if the cleanup is * optional, the orderings make the then action required if the first action * becomes required. */ uint32_t flag = optional? pcmk__ar_first_implies_then : pcmk__ar_ordered; CRM_CHECK((rsc != NULL) && (node != NULL), return); if (pcmk_is_set(rsc->flags, pcmk_rsc_failed)) { pe_rsc_trace(rsc, "Skipping clean-up of %s on %s: resource failed", rsc->id, pe__node_name(node)); return; } if (node->details->unclean || !node->details->online) { pe_rsc_trace(rsc, "Skipping clean-up of %s on %s: node unavailable", rsc->id, pe__node_name(node)); return; } crm_notice("Scheduling clean-up of %s on %s", rsc->id, pe__node_name(node)); delete_action(rsc, node, optional); // stop -> clean-up -> start pcmk__order_resource_actions(rsc, PCMK_ACTION_STOP, rsc, PCMK_ACTION_DELETE, flag); pcmk__order_resource_actions(rsc, PCMK_ACTION_DELETE, rsc, PCMK_ACTION_START, flag); } /*! * \internal * \brief Add primitive meta-attributes relevant to graph actions to XML * * \param[in] rsc Primitive resource whose meta-attributes should be added * \param[in,out] xml Transition graph action attributes XML to add to */ void pcmk__primitive_add_graph_meta(const pcmk_resource_t *rsc, xmlNode *xml) { char *name = NULL; char *value = NULL; const pcmk_resource_t *parent = NULL; CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_primitive) && (xml != NULL)); /* Clone instance numbers get set internally as meta-attributes, and are * needed in the transition graph (for example, to tell unique clone * instances apart). */ value = g_hash_table_lookup(rsc->meta, XML_RSC_ATTR_INCARNATION); if (value != NULL) { name = crm_meta_name(XML_RSC_ATTR_INCARNATION); crm_xml_add(xml, name, value); free(name); } // Not sure if this one is really needed ... value = g_hash_table_lookup(rsc->meta, XML_RSC_ATTR_REMOTE_NODE); if (value != NULL) { name = crm_meta_name(XML_RSC_ATTR_REMOTE_NODE); crm_xml_add(xml, name, value); free(name); } /* The container meta-attribute can be set on the primitive itself or one of * its parents (for example, a group inside a container resource), so check * them all, and keep the highest one found. */ for (parent = rsc; parent != NULL; parent = parent->parent) { if (parent->container != NULL) { crm_xml_add(xml, CRM_META "_" XML_RSC_ATTR_CONTAINER, parent->container->id); } } /* Bundle replica children will get their external-ip set internally as a * meta-attribute. The graph action needs it, but under a different naming * convention than other meta-attributes. */ value = g_hash_table_lookup(rsc->meta, "external-ip"); if (value != NULL) { crm_xml_add(xml, "pcmk_external_ip", value); } } // Primitive implementation of pcmk_assignment_methods_t:add_utilization() void pcmk__primitive_add_utilization(const pcmk_resource_t *rsc, const pcmk_resource_t *orig_rsc, GList *all_rscs, GHashTable *utilization) { CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_primitive) && (orig_rsc != NULL) && (utilization != NULL)); if (!pcmk_is_set(rsc->flags, pcmk_rsc_unassigned)) { return; } pe_rsc_trace(orig_rsc, "%s: Adding primitive %s as colocated utilization", orig_rsc->id, rsc->id); pcmk__release_node_capacity(utilization, rsc); } /*! * \internal * \brief Get epoch time of node's shutdown attribute (or now if none) * * \param[in,out] node Node to check * * \return Epoch time corresponding to shutdown attribute if set or now if not */ static time_t shutdown_time(pcmk_node_t *node) { const char *shutdown = pe_node_attribute_raw(node, XML_CIB_ATTR_SHUTDOWN); time_t result = 0; if (shutdown != NULL) { long long result_ll; if (pcmk__scan_ll(shutdown, &result_ll, 0LL) == pcmk_rc_ok) { result = (time_t) result_ll; } } return (result == 0)? get_effective_time(node->details->data_set) : result; } /*! * \internal * \brief Ban a resource from a node if it's not locked to the node * * \param[in] data Node to check * \param[in,out] user_data Resource to check */ static void ban_if_not_locked(gpointer data, gpointer user_data) { const pcmk_node_t *node = (const pcmk_node_t *) data; pcmk_resource_t *rsc = (pcmk_resource_t *) user_data; if (strcmp(node->details->uname, rsc->lock_node->details->uname) != 0) { resource_location(rsc, node, -CRM_SCORE_INFINITY, XML_CONFIG_ATTR_SHUTDOWN_LOCK, rsc->cluster); } } // Primitive implementation of pcmk_assignment_methods_t:shutdown_lock() void pcmk__primitive_shutdown_lock(pcmk_resource_t *rsc) { const char *class = NULL; CRM_ASSERT((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_primitive)); class = crm_element_value(rsc->xml, XML_AGENT_ATTR_CLASS); // Fence devices and remote connections can't be locked if (pcmk__str_eq(class, PCMK_RESOURCE_CLASS_STONITH, pcmk__str_null_matches) || pe__resource_is_remote_conn(rsc)) { 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. */ pcmk__rsc_info(rsc, "Cancelling shutdown lock " "because %s is already active", rsc->id); pe__clear_resource_history(rsc, rsc->lock_node); 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)) { pcmk_node_t *node = rsc->running_on->data; if (node->details->shutdown) { if (node->details->unclean) { - pe_rsc_debug(rsc, "Not locking %s to unclean %s for shutdown", - rsc->id, pe__node_name(node)); + pcmk__rsc_debug(rsc, + "Not locking %s to unclean %s for shutdown", + rsc->id, pe__node_name(node)); } else { rsc->lock_node = node; rsc->lock_time = shutdown_time(node); } } } if (rsc->lock_node == NULL) { // No lock needed return; } if (rsc->cluster->shutdown_lock > 0) { time_t lock_expiration = rsc->lock_time + rsc->cluster->shutdown_lock; pcmk__rsc_info(rsc, "Locking %s to %s due to shutdown (expires @%lld)", rsc->id, pe__node_name(rsc->lock_node), (long long) lock_expiration); pe__update_recheck_time(++lock_expiration, rsc->cluster, "shutdown lock expiration"); } else { pcmk__rsc_info(rsc, "Locking %s to %s due to shutdown", rsc->id, pe__node_name(rsc->lock_node)); } // If resource is locked to one node, ban it from all other nodes g_list_foreach(rsc->cluster->nodes, ban_if_not_locked, rsc); } diff --git a/lib/pacemaker/pcmk_sched_promotable.c b/lib/pacemaker/pcmk_sched_promotable.c index 24c312d801..f49715fdf0 100644 --- a/lib/pacemaker/pcmk_sched_promotable.c +++ b/lib/pacemaker/pcmk_sched_promotable.c @@ -1,1301 +1,1301 @@ /* * Copyright 2004-2023 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU General Public License version 2 * or later (GPLv2+) WITHOUT ANY WARRANTY. */ #include #include #include #include "libpacemaker_private.h" /*! * \internal * \brief Add implicit promotion ordering for a promotable instance * * \param[in,out] clone Clone resource * \param[in,out] child Instance of \p clone being ordered * \param[in,out] last Previous instance ordered (NULL if \p child is first) */ static void order_instance_promotion(pcmk_resource_t *clone, pcmk_resource_t *child, pcmk_resource_t *last) { // "Promote clone" -> promote instance -> "clone promoted" pcmk__order_resource_actions(clone, PCMK_ACTION_PROMOTE, child, PCMK_ACTION_PROMOTE, pcmk__ar_ordered); pcmk__order_resource_actions(child, PCMK_ACTION_PROMOTE, clone, PCMK_ACTION_PROMOTED, pcmk__ar_ordered); // If clone is ordered, order this instance relative to last if ((last != NULL) && pe__clone_is_ordered(clone)) { pcmk__order_resource_actions(last, PCMK_ACTION_PROMOTE, child, PCMK_ACTION_PROMOTE, pcmk__ar_ordered); } } /*! * \internal * \brief Add implicit demotion ordering for a promotable instance * * \param[in,out] clone Clone resource * \param[in,out] child Instance of \p clone being ordered * \param[in] last Previous instance ordered (NULL if \p child is first) */ static void order_instance_demotion(pcmk_resource_t *clone, pcmk_resource_t *child, pcmk_resource_t *last) { // "Demote clone" -> demote instance -> "clone demoted" pcmk__order_resource_actions(clone, PCMK_ACTION_DEMOTE, child, PCMK_ACTION_DEMOTE, pcmk__ar_then_implies_first_graphed); pcmk__order_resource_actions(child, PCMK_ACTION_DEMOTE, clone, PCMK_ACTION_DEMOTED, pcmk__ar_first_implies_then_graphed); // If clone is ordered, order this instance relative to last if ((last != NULL) && pe__clone_is_ordered(clone)) { pcmk__order_resource_actions(child, PCMK_ACTION_DEMOTE, last, PCMK_ACTION_DEMOTE, pcmk__ar_ordered); } } /*! * \internal * \brief Check whether an instance will be promoted or demoted * * \param[in] rsc Instance to check * \param[out] demoting If \p rsc will be demoted, this will be set to true * \param[out] promoting If \p rsc will be promoted, this will be set to true */ static void check_for_role_change(const pcmk_resource_t *rsc, bool *demoting, bool *promoting) { const GList *iter = NULL; // If this is a cloned group, check group members recursively if (rsc->children != NULL) { for (iter = rsc->children; iter != NULL; iter = iter->next) { check_for_role_change((const pcmk_resource_t *) iter->data, demoting, promoting); } return; } for (iter = rsc->actions; iter != NULL; iter = iter->next) { const pcmk_action_t *action = (const pcmk_action_t *) iter->data; if (*promoting && *demoting) { return; } else if (pcmk_is_set(action->flags, pcmk_action_optional)) { continue; } else if (pcmk__str_eq(PCMK_ACTION_DEMOTE, action->task, pcmk__str_none)) { *demoting = true; } else if (pcmk__str_eq(PCMK_ACTION_PROMOTE, action->task, pcmk__str_none)) { *promoting = true; } } } /*! * \internal * \brief Add promoted-role location constraint scores to an instance's priority * * Adjust a promotable clone instance's promotion priority by the scores of any * location constraints in a list that are both limited to the promoted role and * for the node where the instance will be placed. * * \param[in,out] child Promotable clone instance * \param[in] location_constraints List of location constraints to apply * \param[in] chosen Node where \p child will be placed */ static void apply_promoted_locations(pcmk_resource_t *child, const GList *location_constraints, const pcmk_node_t *chosen) { for (const GList *iter = location_constraints; iter; iter = iter->next) { const pcmk__location_t *location = iter->data; const pcmk_node_t *constraint_node = NULL; if (location->role_filter == pcmk_role_promoted) { constraint_node = pe_find_node_id(location->nodes, chosen->details->id); } if (constraint_node != NULL) { int new_priority = pcmk__add_scores(child->priority, constraint_node->weight); pe_rsc_trace(child, "Applying location %s to %s promotion priority on %s: " "%s + %s = %s", location->id, child->id, pe__node_name(constraint_node), pcmk_readable_score(child->priority), pcmk_readable_score(constraint_node->weight), pcmk_readable_score(new_priority)); child->priority = new_priority; } } } /*! * \internal * \brief Get the node that an instance will be promoted on * * \param[in] rsc Promotable clone instance to check * * \return Node that \p rsc will be promoted on, or NULL if none */ static pcmk_node_t * node_to_be_promoted_on(const pcmk_resource_t *rsc) { pcmk_node_t *node = NULL; pcmk_node_t *local_node = NULL; const pcmk_resource_t *parent = NULL; // If this is a cloned group, bail if any group member can't be promoted for (GList *iter = rsc->children; iter != NULL; iter = iter->next) { pcmk_resource_t *child = (pcmk_resource_t *) iter->data; if (node_to_be_promoted_on(child) == NULL) { pe_rsc_trace(rsc, "%s can't be promoted because member %s can't", rsc->id, child->id); return NULL; } } node = rsc->fns->location(rsc, NULL, FALSE); if (node == NULL) { pe_rsc_trace(rsc, "%s can't be promoted because it won't be active", rsc->id); return NULL; } else if (!pcmk_is_set(rsc->flags, pcmk_rsc_managed)) { if (rsc->fns->state(rsc, TRUE) == pcmk_role_promoted) { crm_notice("Unmanaged instance %s will be left promoted on %s", rsc->id, pe__node_name(node)); } else { pe_rsc_trace(rsc, "%s can't be promoted because it is unmanaged", rsc->id); return NULL; } } else if (rsc->priority < 0) { pe_rsc_trace(rsc, "%s can't be promoted because its promotion priority %d " "is negative", rsc->id, rsc->priority); return NULL; } else if (!pcmk__node_available(node, false, true)) { pe_rsc_trace(rsc, "%s can't be promoted because %s can't run resources", rsc->id, pe__node_name(node)); return NULL; } parent = pe__const_top_resource(rsc, false); local_node = g_hash_table_lookup(parent->allowed_nodes, node->details->id); if (local_node == NULL) { /* It should not be possible for the scheduler to have assigned the * instance to a node where its parent is not allowed, but it's good to * have a fail-safe. */ if (pcmk_is_set(rsc->flags, pcmk_rsc_managed)) { crm_warn("%s can't be promoted because %s is not allowed on %s " "(scheduler bug?)", rsc->id, parent->id, pe__node_name(node)); } // else the instance is unmanaged and already promoted return NULL; } else if ((local_node->count >= pe__clone_promoted_node_max(parent)) && pcmk_is_set(rsc->flags, pcmk_rsc_managed)) { pe_rsc_trace(rsc, "%s can't be promoted because %s has " "maximum promoted instances already", rsc->id, pe__node_name(node)); return NULL; } return local_node; } /*! * \internal * \brief Compare two promotable clone instances by promotion priority * * \param[in] a First instance to compare * \param[in] b Second instance to compare * * \return A negative number if \p a has higher promotion priority, * a positive number if \p b has higher promotion priority, * or 0 if promotion priorities are equal */ static gint cmp_promotable_instance(gconstpointer a, gconstpointer b) { const pcmk_resource_t *rsc1 = (const pcmk_resource_t *) a; const pcmk_resource_t *rsc2 = (const pcmk_resource_t *) b; enum rsc_role_e role1 = pcmk_role_unknown; enum rsc_role_e role2 = pcmk_role_unknown; CRM_ASSERT((rsc1 != NULL) && (rsc2 != NULL)); // Check sort index set by pcmk__set_instance_roles() if (rsc1->sort_index > rsc2->sort_index) { pe_rsc_trace(rsc1, "%s has higher promotion priority than %s " "(sort index %d > %d)", rsc1->id, rsc2->id, rsc1->sort_index, rsc2->sort_index); return -1; } else if (rsc1->sort_index < rsc2->sort_index) { pe_rsc_trace(rsc1, "%s has lower promotion priority than %s " "(sort index %d < %d)", rsc1->id, rsc2->id, rsc1->sort_index, rsc2->sort_index); return 1; } // If those are the same, prefer instance whose current role is higher role1 = rsc1->fns->state(rsc1, TRUE); role2 = rsc2->fns->state(rsc2, TRUE); if (role1 > role2) { pe_rsc_trace(rsc1, "%s has higher promotion priority than %s " "(higher current role)", rsc1->id, rsc2->id); return -1; } else if (role1 < role2) { pe_rsc_trace(rsc1, "%s has lower promotion priority than %s " "(lower current role)", rsc1->id, rsc2->id); return 1; } // Finally, do normal clone instance sorting return pcmk__cmp_instance(a, b); } /*! * \internal * \brief Add a promotable clone instance's sort index to its node's score * * Add a promotable clone instance's sort index (which sums its promotion * preferences and scores of relevant location constraints for the promoted * role) to the node score of the instance's assigned node. * * \param[in] data Promotable clone instance * \param[in,out] user_data Clone parent of \p data */ static void add_sort_index_to_node_score(gpointer data, gpointer user_data) { const pcmk_resource_t *child = (const pcmk_resource_t *) data; pcmk_resource_t *clone = (pcmk_resource_t *) user_data; pcmk_node_t *node = NULL; const pcmk_node_t *chosen = NULL; if (child->sort_index < 0) { pe_rsc_trace(clone, "Not adding sort index of %s: negative", child->id); return; } chosen = child->fns->location(child, NULL, FALSE); if (chosen == NULL) { pe_rsc_trace(clone, "Not adding sort index of %s: inactive", child->id); return; } node = g_hash_table_lookup(clone->allowed_nodes, chosen->details->id); CRM_ASSERT(node != NULL); node->weight = pcmk__add_scores(child->sort_index, node->weight); pe_rsc_trace(clone, "Added cumulative priority of %s (%s) to score on %s (now %s)", child->id, pcmk_readable_score(child->sort_index), pe__node_name(node), pcmk_readable_score(node->weight)); } /*! * \internal * \brief Apply colocation to dependent's node scores if for promoted role * * \param[in,out] data Colocation constraint to apply * \param[in,out] user_data Promotable clone that is constraint's dependent */ static void apply_coloc_to_dependent(gpointer data, gpointer user_data) { pcmk__colocation_t *colocation = data; pcmk_resource_t *clone = user_data; pcmk_resource_t *primary = colocation->primary; uint32_t flags = pcmk__coloc_select_default; float factor = colocation->score / (float) INFINITY; if (colocation->dependent_role != pcmk_role_promoted) { return; } if (colocation->score < INFINITY) { flags = pcmk__coloc_select_active; } pe_rsc_trace(clone, "Applying colocation %s (promoted %s with %s) @%s", colocation->id, colocation->dependent->id, colocation->primary->id, pcmk_readable_score(colocation->score)); primary->cmds->add_colocated_node_scores(primary, clone, clone->id, &clone->allowed_nodes, colocation, factor, flags); } /*! * \internal * \brief Apply colocation to primary's node scores if for promoted role * * \param[in,out] data Colocation constraint to apply * \param[in,out] user_data Promotable clone that is constraint's primary */ static void apply_coloc_to_primary(gpointer data, gpointer user_data) { pcmk__colocation_t *colocation = data; pcmk_resource_t *clone = user_data; pcmk_resource_t *dependent = colocation->dependent; const float factor = colocation->score / (float) INFINITY; const uint32_t flags = pcmk__coloc_select_active |pcmk__coloc_select_nonnegative; if ((colocation->primary_role != pcmk_role_promoted) || !pcmk__colocation_has_influence(colocation, NULL)) { return; } pe_rsc_trace(clone, "Applying colocation %s (%s with promoted %s) @%s", colocation->id, colocation->dependent->id, colocation->primary->id, pcmk_readable_score(colocation->score)); dependent->cmds->add_colocated_node_scores(dependent, clone, clone->id, &clone->allowed_nodes, colocation, factor, flags); } /*! * \internal * \brief Set clone instance's sort index to its node's score * * \param[in,out] data Promotable clone instance * \param[in] user_data Parent clone of \p data */ static void set_sort_index_to_node_score(gpointer data, gpointer user_data) { pcmk_resource_t *child = (pcmk_resource_t *) data; const pcmk_resource_t *clone = (const pcmk_resource_t *) user_data; pcmk_node_t *chosen = child->fns->location(child, NULL, FALSE); if (!pcmk_is_set(child->flags, pcmk_rsc_managed) && (child->next_role == pcmk_role_promoted)) { child->sort_index = INFINITY; pe_rsc_trace(clone, "Final sort index for %s is INFINITY (unmanaged promoted)", child->id); } else if ((chosen == NULL) || (child->sort_index < 0)) { pe_rsc_trace(clone, "Final sort index for %s is %d (ignoring node score)", child->id, child->sort_index); } else { const pcmk_node_t *node = g_hash_table_lookup(clone->allowed_nodes, chosen->details->id); CRM_ASSERT(node != NULL); child->sort_index = node->weight; pe_rsc_trace(clone, "Adding scores for %s: final sort index for %s is %d", clone->id, child->id, child->sort_index); } } /*! * \internal * \brief Sort a promotable clone's instances by descending promotion priority * * \param[in,out] clone Promotable clone to sort */ static void sort_promotable_instances(pcmk_resource_t *clone) { GList *colocations = NULL; if (pe__set_clone_flag(clone, pcmk__clone_promotion_constrained) == pcmk_rc_already) { return; } pe__set_resource_flags(clone, pcmk_rsc_updating_nodes); for (GList *iter = clone->children; iter != NULL; iter = iter->next) { pcmk_resource_t *child = (pcmk_resource_t *) iter->data; pe_rsc_trace(clone, "Adding scores for %s: initial sort index for %s is %d", clone->id, child->id, child->sort_index); } pe__show_node_scores(true, clone, "Before", clone->allowed_nodes, clone->cluster); g_list_foreach(clone->children, add_sort_index_to_node_score, clone); colocations = pcmk__this_with_colocations(clone); g_list_foreach(colocations, apply_coloc_to_dependent, clone); g_list_free(colocations); colocations = pcmk__with_this_colocations(clone); g_list_foreach(colocations, apply_coloc_to_primary, clone); g_list_free(colocations); // Ban resource from all nodes if it needs a ticket but doesn't have it pcmk__require_promotion_tickets(clone); pe__show_node_scores(true, clone, "After", clone->allowed_nodes, clone->cluster); // Reset sort indexes to final node scores g_list_foreach(clone->children, set_sort_index_to_node_score, clone); // Finally, sort instances in descending order of promotion priority clone->children = g_list_sort(clone->children, cmp_promotable_instance); pe__clear_resource_flags(clone, pcmk_rsc_updating_nodes); } /*! * \internal * \brief Find the active instance (if any) of an anonymous clone on a node * * \param[in] clone Anonymous clone to check * \param[in] id Instance ID (without instance number) to check * \param[in] node Node to check * * \return */ static pcmk_resource_t * find_active_anon_instance(const pcmk_resource_t *clone, const char *id, const pcmk_node_t *node) { for (GList *iter = clone->children; iter; iter = iter->next) { pcmk_resource_t *child = iter->data; pcmk_resource_t *active = NULL; // Use ->find_rsc() in case this is a cloned group active = clone->fns->find_rsc(child, id, node, pcmk_rsc_match_clone_only |pcmk_rsc_match_current_node); if (active != NULL) { return active; } } return NULL; } /* * \brief Check whether an anonymous clone instance is known on a node * * \param[in] clone Anonymous clone to check * \param[in] id Instance ID (without instance number) to check * \param[in] node Node to check * * \return true if \p id instance of \p clone is known on \p node, * otherwise false */ static bool anonymous_known_on(const pcmk_resource_t *clone, const char *id, const pcmk_node_t *node) { for (GList *iter = clone->children; iter; iter = iter->next) { pcmk_resource_t *child = iter->data; /* Use ->find_rsc() because this might be a cloned group, and knowing * that other members of the group are known here implies nothing. */ child = clone->fns->find_rsc(child, id, NULL, pcmk_rsc_match_clone_only); CRM_LOG_ASSERT(child != NULL); if (child != NULL) { if (g_hash_table_lookup(child->known_on, node->details->id)) { return true; } } } return false; } /*! * \internal * \brief Check whether a node is allowed to run a resource * * \param[in] rsc Resource to check * \param[in] node Node to check * * \return true if \p node is allowed to run \p rsc, otherwise false */ static bool is_allowed(const pcmk_resource_t *rsc, const pcmk_node_t *node) { pcmk_node_t *allowed = g_hash_table_lookup(rsc->allowed_nodes, node->details->id); return (allowed != NULL) && (allowed->weight >= 0); } /*! * \brief Check whether a clone instance's promotion score should be considered * * \param[in] rsc Promotable clone instance to check * \param[in] node Node where score would be applied * * \return true if \p rsc's promotion score should be considered on \p node, * otherwise false */ static bool promotion_score_applies(const pcmk_resource_t *rsc, const pcmk_node_t *node) { char *id = clone_strip(rsc->id); const pcmk_resource_t *parent = pe__const_top_resource(rsc, false); pcmk_resource_t *active = NULL; const char *reason = "allowed"; // Some checks apply only to anonymous clone instances if (!pcmk_is_set(rsc->flags, pcmk_rsc_unique)) { // If instance is active on the node, its score definitely applies active = find_active_anon_instance(parent, id, node); if (active == rsc) { reason = "active"; goto check_allowed; } /* If *no* instance is active on this node, this instance's score will * count if it has been probed on this node. */ if ((active == NULL) && anonymous_known_on(parent, id, node)) { reason = "probed"; goto check_allowed; } } /* If this clone's status is unknown on *all* nodes (e.g. cluster startup), * take all instances' scores into account, to make sure we use any * permanent promotion scores. */ if ((rsc->running_on == NULL) && (g_hash_table_size(rsc->known_on) == 0)) { reason = "none probed"; goto check_allowed; } /* Otherwise, we've probed and/or started the resource *somewhere*, so * consider promotion scores on nodes where we know the status. */ if ((g_hash_table_lookup(rsc->known_on, node->details->id) != NULL) || (pe_find_node_id(rsc->running_on, node->details->id) != NULL)) { reason = "known"; } else { pe_rsc_trace(rsc, "Ignoring %s promotion score (for %s) on %s: not probed", rsc->id, id, pe__node_name(node)); free(id); return false; } check_allowed: if (is_allowed(rsc, node)) { pe_rsc_trace(rsc, "Counting %s promotion score (for %s) on %s: %s", rsc->id, id, pe__node_name(node), reason); free(id); return true; } pe_rsc_trace(rsc, "Ignoring %s promotion score (for %s) on %s: not allowed", rsc->id, id, pe__node_name(node)); free(id); return false; } /*! * \internal * \brief Get the value of a promotion score node attribute * * \param[in] rsc Promotable clone instance to get promotion score for * \param[in] node Node to get promotion score for * \param[in] name Resource name to use in promotion score attribute name * * \return Value of promotion score node attribute for \p rsc on \p node */ static const char * promotion_attr_value(const pcmk_resource_t *rsc, const pcmk_node_t *node, const char *name) { char *attr_name = NULL; const char *attr_value = NULL; enum pcmk__rsc_node node_type = pcmk__rsc_node_assigned; if (pcmk_is_set(rsc->flags, pcmk_rsc_unassigned)) { // Not assigned yet node_type = pcmk__rsc_node_current; } attr_name = pcmk_promotion_score_name(name); attr_value = pe__node_attribute_calculated(node, attr_name, rsc, node_type, false); free(attr_name); return attr_value; } /*! * \internal * \brief Get the promotion score for a clone instance on a node * * \param[in] rsc Promotable clone instance to get score for * \param[in] node Node to get score for * \param[out] is_default If non-NULL, will be set true if no score available * * \return Promotion score for \p rsc on \p node (or 0 if none) */ static int promotion_score(const pcmk_resource_t *rsc, const pcmk_node_t *node, bool *is_default) { char *name = NULL; const char *attr_value = NULL; if (is_default != NULL) { *is_default = true; } CRM_CHECK((rsc != NULL) && (node != NULL), return 0); /* If this is an instance of a cloned group, the promotion score is the sum * of all members' promotion scores. */ if (rsc->children != NULL) { int score = 0; for (const GList *iter = rsc->children; iter != NULL; iter = iter->next) { const pcmk_resource_t *child = (const pcmk_resource_t *) iter->data; bool child_default = false; int child_score = promotion_score(child, node, &child_default); if (!child_default && (is_default != NULL)) { *is_default = false; } score += child_score; } return score; } if (!promotion_score_applies(rsc, node)) { return 0; } /* For the promotion score attribute name, use the name the resource is * known as in resource history, since that's what crm_attribute --promotion * would have used. */ name = (rsc->clone_name == NULL)? rsc->id : rsc->clone_name; attr_value = promotion_attr_value(rsc, node, name); if (attr_value != NULL) { pe_rsc_trace(rsc, "Promotion score for %s on %s = %s", name, pe__node_name(node), pcmk__s(attr_value, "(unset)")); } else if (!pcmk_is_set(rsc->flags, pcmk_rsc_unique)) { /* If we don't have any resource history yet, we won't have clone_name. * In that case, for anonymous clones, try the resource name without * any instance number. */ name = clone_strip(rsc->id); if (strcmp(rsc->id, name) != 0) { attr_value = promotion_attr_value(rsc, node, name); pe_rsc_trace(rsc, "Promotion score for %s on %s (for %s) = %s", name, pe__node_name(node), rsc->id, pcmk__s(attr_value, "(unset)")); } free(name); } if (attr_value == NULL) { return 0; } if (is_default != NULL) { *is_default = false; } return char2score(attr_value); } /*! * \internal * \brief Include promotion scores in instances' node scores and priorities * * \param[in,out] rsc Promotable clone resource to update */ void pcmk__add_promotion_scores(pcmk_resource_t *rsc) { if (pe__set_clone_flag(rsc, pcmk__clone_promotion_added) == pcmk_rc_already) { return; } for (GList *iter = rsc->children; iter != NULL; iter = iter->next) { pcmk_resource_t *child_rsc = (pcmk_resource_t *) iter->data; GHashTableIter iter; pcmk_node_t *node = NULL; int score, new_score; g_hash_table_iter_init(&iter, child_rsc->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (void **) &node)) { if (!pcmk__node_available(node, false, false)) { /* This node will never be promoted, so don't apply the * promotion score, as that may lead to clone shuffling. */ continue; } score = promotion_score(child_rsc, node, NULL); if (score > 0) { new_score = pcmk__add_scores(node->weight, score); if (new_score != node->weight) { // Could remain INFINITY node->weight = new_score; pe_rsc_trace(rsc, "Added %s promotion priority (%s) to score " "on %s (now %s)", child_rsc->id, pcmk_readable_score(score), pe__node_name(node), pcmk_readable_score(new_score)); } } if (score > child_rsc->priority) { pe_rsc_trace(rsc, "Updating %s priority to promotion score (%d->%d)", child_rsc->id, child_rsc->priority, score); child_rsc->priority = score; } } } } /*! * \internal * \brief If a resource's current role is started, change it to unpromoted * * \param[in,out] data Resource to update * \param[in] user_data Ignored */ static void set_current_role_unpromoted(void *data, void *user_data) { pcmk_resource_t *rsc = (pcmk_resource_t *) data; if (rsc->role == pcmk_role_started) { // Promotable clones should use unpromoted role instead of started rsc->role = pcmk_role_unpromoted; } g_list_foreach(rsc->children, set_current_role_unpromoted, NULL); } /*! * \internal * \brief Set a resource's next role to unpromoted (or stopped if unassigned) * * \param[in,out] data Resource to update * \param[in] user_data Ignored */ static void set_next_role_unpromoted(void *data, void *user_data) { pcmk_resource_t *rsc = (pcmk_resource_t *) data; GList *assigned = NULL; rsc->fns->location(rsc, &assigned, FALSE); if (assigned == NULL) { pe__set_next_role(rsc, pcmk_role_stopped, "stopped instance"); } else { pe__set_next_role(rsc, pcmk_role_unpromoted, "unpromoted instance"); g_list_free(assigned); } g_list_foreach(rsc->children, set_next_role_unpromoted, NULL); } /*! * \internal * \brief Set a resource's next role to promoted if not already set * * \param[in,out] data Resource to update * \param[in] user_data Ignored */ static void set_next_role_promoted(void *data, gpointer user_data) { pcmk_resource_t *rsc = (pcmk_resource_t *) data; if (rsc->next_role == pcmk_role_unknown) { pe__set_next_role(rsc, pcmk_role_promoted, "promoted instance"); } g_list_foreach(rsc->children, set_next_role_promoted, NULL); } /*! * \internal * \brief Show instance's promotion score on node where it will be active * * \param[in,out] instance Promotable clone instance to show */ static void show_promotion_score(pcmk_resource_t *instance) { pcmk_node_t *chosen = instance->fns->location(instance, NULL, FALSE); if (pcmk_is_set(instance->cluster->flags, pcmk_sched_output_scores) && !pcmk__is_daemon && (instance->cluster->priv != NULL)) { pcmk__output_t *out = instance->cluster->priv; out->message(out, "promotion-score", instance, chosen, pcmk_readable_score(instance->sort_index)); } else { - pe_rsc_debug(pe__const_top_resource(instance, false), - "%s promotion score on %s: sort=%s priority=%s", - instance->id, - ((chosen == NULL)? "none" : pe__node_name(chosen)), - pcmk_readable_score(instance->sort_index), - pcmk_readable_score(instance->priority)); + pcmk__rsc_debug(pe__const_top_resource(instance, false), + "%s promotion score on %s: sort=%s priority=%s", + instance->id, + ((chosen == NULL)? "none" : pe__node_name(chosen)), + pcmk_readable_score(instance->sort_index), + pcmk_readable_score(instance->priority)); } } /*! * \internal * \brief Set a clone instance's promotion priority * * \param[in,out] data Promotable clone instance to update * \param[in] user_data Instance's parent clone */ static void set_instance_priority(gpointer data, gpointer user_data) { pcmk_resource_t *instance = (pcmk_resource_t *) data; const pcmk_resource_t *clone = (const pcmk_resource_t *) user_data; const pcmk_node_t *chosen = NULL; enum rsc_role_e next_role = pcmk_role_unknown; GList *list = NULL; pe_rsc_trace(clone, "Assigning priority for %s: %s", instance->id, role2text(instance->next_role)); if (instance->fns->state(instance, TRUE) == pcmk_role_started) { set_current_role_unpromoted(instance, NULL); } // Only an instance that will be active can be promoted chosen = instance->fns->location(instance, &list, FALSE); if (pcmk__list_of_multiple(list)) { pcmk__config_err("Cannot promote non-colocated child %s", instance->id); } g_list_free(list); if (chosen == NULL) { return; } next_role = instance->fns->state(instance, FALSE); switch (next_role) { case pcmk_role_started: case pcmk_role_unknown: // Set instance priority to its promotion score (or -1 if none) { bool is_default = false; instance->priority = promotion_score(instance, chosen, &is_default); if (is_default) { /* * Default to -1 if no value is set. This allows * instances eligible for promotion to be specified * based solely on rsc_location constraints, but * prevents any instance from being promoted if neither * a constraint nor a promotion score is present */ instance->priority = -1; } } break; case pcmk_role_unpromoted: case pcmk_role_stopped: // Instance can't be promoted instance->priority = -INFINITY; break; case pcmk_role_promoted: // Nothing needed (re-creating actions after scheduling fencing) break; default: CRM_CHECK(FALSE, crm_err("Unknown resource role %d for %s", next_role, instance->id)); } // Add relevant location constraint scores for promoted role apply_promoted_locations(instance, instance->rsc_location, chosen); apply_promoted_locations(instance, clone->rsc_location, chosen); // Consider instance's role-based colocations with other resources list = pcmk__this_with_colocations(instance); for (GList *iter = list; iter != NULL; iter = iter->next) { pcmk__colocation_t *cons = (pcmk__colocation_t *) iter->data; instance->cmds->apply_coloc_score(instance, cons->primary, cons, true); } g_list_free(list); instance->sort_index = instance->priority; if (next_role == pcmk_role_promoted) { instance->sort_index = INFINITY; } pe_rsc_trace(clone, "Assigning %s priority = %d", instance->id, instance->priority); } /*! * \internal * \brief Set a promotable clone instance's role * * \param[in,out] data Promotable clone instance to update * \param[in,out] user_data Pointer to count of instances chosen for promotion */ static void set_instance_role(gpointer data, gpointer user_data) { pcmk_resource_t *instance = (pcmk_resource_t *) data; int *count = (int *) user_data; const pcmk_resource_t *clone = pe__const_top_resource(instance, false); pcmk_node_t *chosen = NULL; show_promotion_score(instance); if (instance->sort_index < 0) { pe_rsc_trace(clone, "Not supposed to promote instance %s", instance->id); } else if ((*count < pe__clone_promoted_max(instance)) || !pcmk_is_set(clone->flags, pcmk_rsc_managed)) { chosen = node_to_be_promoted_on(instance); } if (chosen == NULL) { set_next_role_unpromoted(instance, NULL); return; } if ((instance->role < pcmk_role_promoted) && !pcmk_is_set(instance->cluster->flags, pcmk_sched_quorate) && (instance->cluster->no_quorum_policy == pcmk_no_quorum_freeze)) { crm_notice("Clone instance %s cannot be promoted without quorum", instance->id); set_next_role_unpromoted(instance, NULL); return; } chosen->count++; pcmk__rsc_info(clone, "Choosing %s (%s) on %s for promotion", instance->id, role2text(instance->role), pe__node_name(chosen)); set_next_role_promoted(instance, NULL); (*count)++; } /*! * \internal * \brief Set roles for all instances of a promotable clone * * \param[in,out] rsc Promotable clone resource to update */ void pcmk__set_instance_roles(pcmk_resource_t *rsc) { int promoted = 0; GHashTableIter iter; pcmk_node_t *node = NULL; // Repurpose count to track the number of promoted instances assigned g_hash_table_iter_init(&iter, rsc->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (void **)&node)) { node->count = 0; } // Set instances' promotion priorities and sort by highest priority first g_list_foreach(rsc->children, set_instance_priority, rsc); sort_promotable_instances(rsc); // Choose the first N eligible instances to be promoted g_list_foreach(rsc->children, set_instance_role, &promoted); pcmk__rsc_info(rsc, "%s: Promoted %d instances of a possible %d", rsc->id, promoted, pe__clone_promoted_max(rsc)); } /*! * * \internal * \brief Create actions for promotable clone instances * * \param[in,out] clone Promotable clone to create actions for * \param[out] any_promoting Will be set true if any instance is promoting * \param[out] any_demoting Will be set true if any instance is demoting */ static void create_promotable_instance_actions(pcmk_resource_t *clone, bool *any_promoting, bool *any_demoting) { for (GList *iter = clone->children; iter != NULL; iter = iter->next) { pcmk_resource_t *instance = (pcmk_resource_t *) iter->data; instance->cmds->create_actions(instance); check_for_role_change(instance, any_demoting, any_promoting); } } /*! * \internal * \brief Reset each promotable instance's resource priority * * Reset the priority of each instance of a promotable clone to the clone's * priority (after promotion actions are scheduled, when instance priorities * were repurposed as promotion scores). * * \param[in,out] clone Promotable clone to reset */ static void reset_instance_priorities(pcmk_resource_t *clone) { for (GList *iter = clone->children; iter != NULL; iter = iter->next) { pcmk_resource_t *instance = (pcmk_resource_t *) iter->data; instance->priority = clone->priority; } } /*! * \internal * \brief Create actions specific to promotable clones * * \param[in,out] clone Promotable clone to create actions for */ void pcmk__create_promotable_actions(pcmk_resource_t *clone) { bool any_promoting = false; bool any_demoting = false; // Create actions for each clone instance individually create_promotable_instance_actions(clone, &any_promoting, &any_demoting); // Create pseudo-actions for clone as a whole pe__create_promotable_pseudo_ops(clone, any_promoting, any_demoting); // Undo our temporary repurposing of resource priority for instances reset_instance_priorities(clone); } /*! * \internal * \brief Create internal orderings for a promotable clone's instances * * \param[in,out] clone Promotable clone instance to order */ void pcmk__order_promotable_instances(pcmk_resource_t *clone) { pcmk_resource_t *previous = NULL; // Needed for ordered clones pcmk__promotable_restart_ordering(clone); for (GList *iter = clone->children; iter != NULL; iter = iter->next) { pcmk_resource_t *instance = (pcmk_resource_t *) iter->data; // Demote before promote pcmk__order_resource_actions(instance, PCMK_ACTION_DEMOTE, instance, PCMK_ACTION_PROMOTE, pcmk__ar_ordered); order_instance_promotion(clone, instance, previous); order_instance_demotion(clone, instance, previous); previous = instance; } } /*! * \internal * \brief Update dependent's allowed nodes for colocation with promotable * * \param[in,out] dependent Dependent resource to update * \param[in] primary Primary resource * \param[in] primary_node Node where an instance of the primary will be * \param[in] colocation Colocation constraint to apply */ static void update_dependent_allowed_nodes(pcmk_resource_t *dependent, const pcmk_resource_t *primary, const pcmk_node_t *primary_node, const pcmk__colocation_t *colocation) { GHashTableIter iter; pcmk_node_t *node = NULL; const char *primary_value = NULL; const char *attr = colocation->node_attribute; if (colocation->score >= INFINITY) { return; // Colocation is mandatory, so allowed node scores don't matter } primary_value = pcmk__colocation_node_attr(primary_node, attr, primary); pe_rsc_trace(colocation->primary, "Applying %s (%s with %s on %s by %s @%d) to %s", colocation->id, colocation->dependent->id, colocation->primary->id, pe__node_name(primary_node), attr, colocation->score, dependent->id); g_hash_table_iter_init(&iter, dependent->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (void **) &node)) { const char *dependent_value = pcmk__colocation_node_attr(node, attr, dependent); if (pcmk__str_eq(primary_value, dependent_value, pcmk__str_casei)) { node->weight = pcmk__add_scores(node->weight, colocation->score); pe_rsc_trace(colocation->primary, "Added %s score (%s) to %s (now %s)", colocation->id, pcmk_readable_score(colocation->score), pe__node_name(node), pcmk_readable_score(node->weight)); } } } /*! * \brief Update dependent for a colocation with a promotable clone * * \param[in] primary Primary resource in the colocation * \param[in,out] dependent Dependent resource in the colocation * \param[in] colocation Colocation constraint to apply */ void pcmk__update_dependent_with_promotable(const pcmk_resource_t *primary, pcmk_resource_t *dependent, const pcmk__colocation_t *colocation) { GList *affected_nodes = NULL; /* Build a list of all nodes where an instance of the primary will be, and * (for optional colocations) update the dependent's allowed node scores for * each one. */ for (GList *iter = primary->children; iter != NULL; iter = iter->next) { pcmk_resource_t *instance = (pcmk_resource_t *) iter->data; pcmk_node_t *node = instance->fns->location(instance, NULL, FALSE); if (node == NULL) { continue; } if (instance->fns->state(instance, FALSE) == colocation->primary_role) { update_dependent_allowed_nodes(dependent, primary, node, colocation); affected_nodes = g_list_prepend(affected_nodes, node); } } /* For mandatory colocations, add the primary's node score to the * dependent's node score for each affected node, and ban the dependent * from all other nodes. * * However, skip this for promoted-with-promoted colocations, otherwise * inactive dependent instances can't start (in the unpromoted role). */ if ((colocation->score >= INFINITY) && ((colocation->dependent_role != pcmk_role_promoted) || (colocation->primary_role != pcmk_role_promoted))) { pe_rsc_trace(colocation->primary, "Applying %s (mandatory %s with %s) to %s", colocation->id, colocation->dependent->id, colocation->primary->id, dependent->id); pcmk__colocation_intersect_nodes(dependent, primary, colocation, affected_nodes, true); } g_list_free(affected_nodes); } /*! * \internal * \brief Update dependent priority for colocation with promotable * * \param[in] primary Primary resource in the colocation * \param[in,out] dependent Dependent resource in the colocation * \param[in] colocation Colocation constraint to apply */ void pcmk__update_promotable_dependent_priority(const pcmk_resource_t *primary, pcmk_resource_t *dependent, const pcmk__colocation_t *colocation) { pcmk_resource_t *primary_instance = NULL; // Look for a primary instance where dependent will be primary_instance = pcmk__find_compatible_instance(dependent, primary, colocation->primary_role, false); if (primary_instance != NULL) { // Add primary instance's priority to dependent's int new_priority = pcmk__add_scores(dependent->priority, colocation->score); pe_rsc_trace(colocation->primary, "Applying %s (%s with %s) to %s priority (%s + %s = %s)", colocation->id, colocation->dependent->id, colocation->primary->id, dependent->id, pcmk_readable_score(dependent->priority), pcmk_readable_score(colocation->score), pcmk_readable_score(new_priority)); dependent->priority = new_priority; } else if (colocation->score >= INFINITY) { // Mandatory colocation, but primary won't be here pe_rsc_trace(colocation->primary, "Applying %s (%s with %s) to %s: can't be promoted", colocation->id, colocation->dependent->id, colocation->primary->id, dependent->id); dependent->priority = -INFINITY; } } diff --git a/lib/pacemaker/pcmk_sched_recurring.c b/lib/pacemaker/pcmk_sched_recurring.c index b8b9b3946c..ebd66c8d3d 100644 --- a/lib/pacemaker/pcmk_sched_recurring.c +++ b/lib/pacemaker/pcmk_sched_recurring.c @@ -1,728 +1,728 @@ /* * Copyright 2004-2023 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU General Public License version 2 * or later (GPLv2+) WITHOUT ANY WARRANTY. */ #include #include #include #include #include #include "libpacemaker_private.h" // Information parsed from an operation history entry in the CIB struct op_history { // XML attributes const char *id; // ID of history entry const char *name; // Action name // Parsed information char *key; // Operation key for action enum rsc_role_e role; // Action role (or pcmk_role_unknown for default) guint interval_ms; // Action interval }; /*! * \internal * \brief Parse an interval from XML * * \param[in] xml XML containing an interval attribute * * \return Interval parsed from XML (or 0 as default) */ static guint xe_interval(const xmlNode *xml) { return crm_parse_interval_spec(crm_element_value(xml, XML_LRM_ATTR_INTERVAL)); } /*! * \internal * \brief Check whether an operation exists multiple times in resource history * * \param[in] rsc Resource with history to search * \param[in] name Name of action to search for * \param[in] interval_ms Interval (in milliseconds) of action to search for * * \return true if an operation with \p name and \p interval_ms exists more than * once in the operation history of \p rsc, otherwise false */ static bool is_op_dup(const pcmk_resource_t *rsc, const char *name, guint interval_ms) { const char *id = NULL; for (xmlNode *op = first_named_child(rsc->ops_xml, "op"); op != NULL; op = crm_next_same_xml(op)) { // Check whether action name and interval match if (!pcmk__str_eq(crm_element_value(op, "name"), name, pcmk__str_none) || (xe_interval(op) != interval_ms)) { continue; } if (ID(op) == NULL) { continue; // Shouldn't be possible } if (id == NULL) { id = ID(op); // First matching op } else { pcmk__config_err("Operation %s is duplicate of %s (do not use " "same name and interval combination more " "than once per resource)", ID(op), id); return true; } } return false; } /*! * \internal * \brief Check whether an action name is one that can be recurring * * \param[in] name Action name to check * * \return true if \p name is an action known to be unsuitable as a recurring * operation, otherwise false * * \note Pacemaker's current philosophy is to allow users to configure recurring * operations except for a short list of actions known not to be suitable * for that (as opposed to allowing only actions known to be suitable, * which includes only monitor). Among other things, this approach allows * users to define their own custom operations and make them recurring, * though that use case is not well tested. */ static bool op_cannot_recur(const char *name) { return pcmk__str_any_of(name, PCMK_ACTION_STOP, PCMK_ACTION_START, PCMK_ACTION_DEMOTE, PCMK_ACTION_PROMOTE, PCMK_ACTION_RELOAD_AGENT, PCMK_ACTION_MIGRATE_TO, PCMK_ACTION_MIGRATE_FROM, NULL); } /*! * \internal * \brief Check whether a resource history entry is for a recurring action * * \param[in] rsc Resource that history entry is for * \param[in] xml XML of resource history entry to check * \param[out] op Where to store parsed info if recurring * * \return true if \p xml is for a recurring action, otherwise false */ static bool is_recurring_history(const pcmk_resource_t *rsc, const xmlNode *xml, struct op_history *op) { const char *role = NULL; op->interval_ms = xe_interval(xml); if (op->interval_ms == 0) { return false; // Not recurring } op->id = ID(xml); if (pcmk__str_empty(op->id)) { pcmk__config_err("Ignoring resource history entry without ID"); return false; // Shouldn't be possible (unless CIB was manually edited) } op->name = crm_element_value(xml, "name"); if (op_cannot_recur(op->name)) { pcmk__config_err("Ignoring %s because %s action cannot be recurring", op->id, pcmk__s(op->name, "unnamed")); return false; } // There should only be one recurring operation per action/interval if (is_op_dup(rsc, op->name, op->interval_ms)) { return false; } // Ensure role is valid if specified role = crm_element_value(xml, "role"); if (role == NULL) { op->role = pcmk_role_unknown; } else { op->role = text2role(role); if (op->role == pcmk_role_unknown) { pcmk__config_err("Ignoring %s because %s is not a valid role", op->id, role); return false; } } // Only actions that are still configured and enabled matter if (pcmk__find_action_config(rsc, op->name, op->interval_ms, false) == NULL) { pe_rsc_trace(rsc, "Ignoring %s (%s-interval %s for %s) because it is " "disabled or no longer in configuration", op->id, pcmk__readable_interval(op->interval_ms), op->name, rsc->id); return false; } op->key = pcmk__op_key(rsc->id, op->name, op->interval_ms); return true; } /*! * \internal * \brief Check whether a recurring action for an active role should be optional * * \param[in] rsc Resource that recurring action is for * \param[in] node Node that \p rsc will be active on (if any) * \param[in] key Operation key for recurring action to check * \param[in,out] start Start action for \p rsc * * \return true if recurring action should be optional, otherwise false */ static bool active_recurring_should_be_optional(const pcmk_resource_t *rsc, const pcmk_node_t *node, const char *key, pcmk_action_t *start) { GList *possible_matches = NULL; if (node == NULL) { // Should only be possible if unmanaged and stopped pe_rsc_trace(rsc, "%s will be mandatory because resource is unmanaged", key); return false; } if (!pcmk_is_set(rsc->cmds->action_flags(start, NULL), pcmk_action_optional)) { pe_rsc_trace(rsc, "%s will be mandatory because %s is", key, start->uuid); return false; } possible_matches = find_actions_exact(rsc->actions, key, node); if (possible_matches == NULL) { pe_rsc_trace(rsc, "%s will be mandatory because it is not active on %s", key, pe__node_name(node)); return false; } for (const GList *iter = possible_matches; iter != NULL; iter = iter->next) { const pcmk_action_t *op = (const pcmk_action_t *) iter->data; if (pcmk_is_set(op->flags, pcmk_action_reschedule)) { pe_rsc_trace(rsc, "%s will be mandatory because " "it needs to be rescheduled", key); g_list_free(possible_matches); return false; } } g_list_free(possible_matches); return true; } /*! * \internal * \brief Create recurring action from resource history entry for an active role * * \param[in,out] rsc Resource that resource history is for * \param[in,out] start Start action for \p rsc on \p node * \param[in] node Node that resource will be active on (if any) * \param[in] op Resource history entry */ static void recurring_op_for_active(pcmk_resource_t *rsc, pcmk_action_t *start, const pcmk_node_t *node, const struct op_history *op) { pcmk_action_t *mon = NULL; bool is_optional = true; const bool is_default_role = (op->role == pcmk_role_unknown); // We're only interested in recurring actions for active roles if (op->role == pcmk_role_stopped) { return; } is_optional = active_recurring_should_be_optional(rsc, node, op->key, start); if ((!is_default_role && (rsc->next_role != op->role)) || (is_default_role && (rsc->next_role == pcmk_role_promoted))) { // Configured monitor role doesn't match role resource will have if (is_optional) { // It's running, so cancel it char *after_key = NULL; pcmk_action_t *cancel_op = pcmk__new_cancel_action(rsc, op->name, op->interval_ms, node); switch (rsc->role) { case pcmk_role_unpromoted: case pcmk_role_started: if (rsc->next_role == pcmk_role_promoted) { after_key = promote_key(rsc); } else if (rsc->next_role == pcmk_role_stopped) { after_key = stop_key(rsc); } break; case pcmk_role_promoted: after_key = demote_key(rsc); break; default: break; } if (after_key) { pcmk__new_ordering(rsc, NULL, cancel_op, rsc, after_key, NULL, pcmk__ar_unrunnable_first_blocks, rsc->cluster); } } do_crm_log((is_optional? LOG_INFO : LOG_TRACE), "%s recurring action %s because %s configured for %s role " "(not %s)", (is_optional? "Cancelling" : "Ignoring"), op->key, op->id, role2text(is_default_role? pcmk_role_unpromoted : op->role), role2text(rsc->next_role)); return; } pe_rsc_trace(rsc, "Creating %s recurring action %s for %s (%s %s on %s)", (is_optional? "optional" : "mandatory"), op->key, op->id, rsc->id, role2text(rsc->next_role), pe__node_name(node)); mon = custom_action(rsc, strdup(op->key), op->name, node, is_optional, rsc->cluster); if (!pcmk_is_set(start->flags, pcmk_action_runnable)) { pe_rsc_trace(rsc, "%s is unrunnable because start is", mon->uuid); pe__clear_action_flags(mon, pcmk_action_runnable); } else if ((node == NULL) || !node->details->online || node->details->unclean) { pe_rsc_trace(rsc, "%s is unrunnable because no node is available", mon->uuid); pe__clear_action_flags(mon, pcmk_action_runnable); } else if (!pcmk_is_set(mon->flags, pcmk_action_optional)) { pcmk__rsc_info(rsc, "Start %s-interval %s for %s on %s", pcmk__readable_interval(op->interval_ms), mon->task, rsc->id, pe__node_name(node)); } if (rsc->next_role == pcmk_role_promoted) { pe__add_action_expected_result(mon, CRM_EX_PROMOTED); } // Order monitor relative to other actions if ((node == NULL) || pcmk_is_set(rsc->flags, pcmk_rsc_managed)) { pcmk__new_ordering(rsc, start_key(rsc), NULL, NULL, strdup(mon->uuid), mon, pcmk__ar_first_implies_then |pcmk__ar_unrunnable_first_blocks, rsc->cluster); pcmk__new_ordering(rsc, reload_key(rsc), NULL, NULL, strdup(mon->uuid), mon, pcmk__ar_first_implies_then |pcmk__ar_unrunnable_first_blocks, rsc->cluster); if (rsc->next_role == pcmk_role_promoted) { pcmk__new_ordering(rsc, promote_key(rsc), NULL, rsc, NULL, mon, pcmk__ar_ordered |pcmk__ar_unrunnable_first_blocks, rsc->cluster); } else if (rsc->role == pcmk_role_promoted) { pcmk__new_ordering(rsc, demote_key(rsc), NULL, rsc, NULL, mon, pcmk__ar_ordered |pcmk__ar_unrunnable_first_blocks, rsc->cluster); } } } /*! * \internal * \brief Cancel a recurring action if running on a node * * \param[in,out] rsc Resource that action is for * \param[in] node Node to cancel action on * \param[in] key Operation key for action * \param[in] name Action name * \param[in] interval_ms Action interval (in milliseconds) */ static void cancel_if_running(pcmk_resource_t *rsc, const pcmk_node_t *node, const char *key, const char *name, guint interval_ms) { GList *possible_matches = find_actions_exact(rsc->actions, key, node); pcmk_action_t *cancel_op = NULL; if (possible_matches == NULL) { return; // Recurring action isn't running on this node } g_list_free(possible_matches); cancel_op = pcmk__new_cancel_action(rsc, name, interval_ms, node); switch (rsc->next_role) { case pcmk_role_started: case pcmk_role_unpromoted: /* Order starts after cancel. If the current role is * stopped, this cancels the monitor before the resource * starts; if the current role is started, then this cancels * the monitor on a migration target before starting there. */ pcmk__new_ordering(rsc, NULL, cancel_op, rsc, start_key(rsc), NULL, pcmk__ar_unrunnable_first_blocks, rsc->cluster); break; default: break; } pcmk__rsc_info(rsc, "Cancelling %s-interval %s action for %s on %s because " "configured for " PCMK__ROLE_STOPPED " role (not %s)", pcmk__readable_interval(interval_ms), name, rsc->id, pe__node_name(node), role2text(rsc->next_role)); } /*! * \internal * \brief Order an action after all probes of a resource on a node * * \param[in,out] rsc Resource to check for probes * \param[in] node Node to check for probes of \p rsc * \param[in,out] action Action to order after probes of \p rsc on \p node */ static void order_after_probes(pcmk_resource_t *rsc, const pcmk_node_t *node, pcmk_action_t *action) { GList *probes = pe__resource_actions(rsc, node, PCMK_ACTION_MONITOR, FALSE); for (GList *iter = probes; iter != NULL; iter = iter->next) { order_actions((pcmk_action_t *) iter->data, action, pcmk__ar_unrunnable_first_blocks); } g_list_free(probes); } /*! * \internal * \brief Order an action after all stops of a resource on a node * * \param[in,out] rsc Resource to check for stops * \param[in] node Node to check for stops of \p rsc * \param[in,out] action Action to order after stops of \p rsc on \p node */ static void order_after_stops(pcmk_resource_t *rsc, const pcmk_node_t *node, pcmk_action_t *action) { GList *stop_ops = pe__resource_actions(rsc, node, PCMK_ACTION_STOP, TRUE); for (GList *iter = stop_ops; iter != NULL; iter = iter->next) { pcmk_action_t *stop = (pcmk_action_t *) iter->data; if (!pcmk_is_set(stop->flags, pcmk_action_optional) && !pcmk_is_set(action->flags, pcmk_action_optional) && !pcmk_is_set(rsc->flags, pcmk_rsc_managed)) { pe_rsc_trace(rsc, "%s optional on %s: unmanaged", action->uuid, pe__node_name(node)); pe__set_action_flags(action, pcmk_action_optional); } if (!pcmk_is_set(stop->flags, pcmk_action_runnable)) { crm_debug("%s unrunnable on %s: stop is unrunnable", action->uuid, pe__node_name(node)); pe__clear_action_flags(action, pcmk_action_runnable); } if (pcmk_is_set(rsc->flags, pcmk_rsc_managed)) { pcmk__new_ordering(rsc, stop_key(rsc), stop, NULL, NULL, action, pcmk__ar_first_implies_then |pcmk__ar_unrunnable_first_blocks, rsc->cluster); } } g_list_free(stop_ops); } /*! * \internal * \brief Create recurring action from resource history entry for inactive role * * \param[in,out] rsc Resource that resource history is for * \param[in] node Node that resource will be active on (if any) * \param[in] op Resource history entry */ static void recurring_op_for_inactive(pcmk_resource_t *rsc, const pcmk_node_t *node, const struct op_history *op) { GList *possible_matches = NULL; // We're only interested in recurring actions for the inactive role if (op->role != pcmk_role_stopped) { return; } if (!pcmk_is_set(rsc->flags, pcmk_rsc_unique)) { crm_notice("Ignoring %s (recurring monitors for " PCMK__ROLE_STOPPED " role are not supported for anonymous clones)", op->id); return; // @TODO add support } pe_rsc_trace(rsc, "Creating recurring action %s for %s on nodes " "where it should not be running", op->id, rsc->id); for (GList *iter = rsc->cluster->nodes; iter != NULL; iter = iter->next) { pcmk_node_t *stop_node = (pcmk_node_t *) iter->data; bool is_optional = true; pcmk_action_t *stopped_mon = NULL; // Cancel action on node where resource will be active if ((node != NULL) && pcmk__str_eq(stop_node->details->uname, node->details->uname, pcmk__str_casei)) { cancel_if_running(rsc, node, op->key, op->name, op->interval_ms); continue; } // Recurring action on this node is optional if it's already active here possible_matches = find_actions_exact(rsc->actions, op->key, stop_node); is_optional = (possible_matches != NULL); g_list_free(possible_matches); pe_rsc_trace(rsc, "Creating %s recurring action %s for %s (%s " PCMK__ROLE_STOPPED " on %s)", (is_optional? "optional" : "mandatory"), op->key, op->id, rsc->id, pe__node_name(stop_node)); stopped_mon = custom_action(rsc, strdup(op->key), op->name, stop_node, is_optional, rsc->cluster); pe__add_action_expected_result(stopped_mon, CRM_EX_NOT_RUNNING); if (pcmk_is_set(rsc->flags, pcmk_rsc_managed)) { order_after_probes(rsc, stop_node, stopped_mon); } /* The recurring action is for the inactive role, so it shouldn't be * performed until the resource is inactive. */ order_after_stops(rsc, stop_node, stopped_mon); if (!stop_node->details->online || stop_node->details->unclean) { - pe_rsc_debug(rsc, "%s unrunnable on %s: node unavailable)", - stopped_mon->uuid, pe__node_name(stop_node)); + pcmk__rsc_debug(rsc, "%s unrunnable on %s: node unavailable)", + stopped_mon->uuid, pe__node_name(stop_node)); pe__clear_action_flags(stopped_mon, pcmk_action_runnable); } if (pcmk_is_set(stopped_mon->flags, pcmk_action_runnable) && !pcmk_is_set(stopped_mon->flags, pcmk_action_optional)) { crm_notice("Start recurring %s-interval %s for " PCMK__ROLE_STOPPED " %s on %s", pcmk__readable_interval(op->interval_ms), stopped_mon->task, rsc->id, pe__node_name(stop_node)); } } } /*! * \internal * \brief Create recurring actions for a resource * * \param[in,out] rsc Resource to create recurring actions for */ void pcmk__create_recurring_actions(pcmk_resource_t *rsc) { pcmk_action_t *start = NULL; if (pcmk_is_set(rsc->flags, pcmk_rsc_blocked)) { pe_rsc_trace(rsc, "Skipping recurring actions for blocked resource %s", rsc->id); return; } if (pcmk_is_set(rsc->flags, pcmk_rsc_maintenance)) { pe_rsc_trace(rsc, "Skipping recurring actions for %s " "in maintenance mode", rsc->id); return; } if (rsc->allocated_to == NULL) { // Recurring actions for active roles not needed } else if (rsc->allocated_to->details->maintenance) { pe_rsc_trace(rsc, "Skipping recurring actions for %s on %s " "in maintenance mode", rsc->id, pe__node_name(rsc->allocated_to)); } else if ((rsc->next_role != pcmk_role_stopped) || !pcmk_is_set(rsc->flags, pcmk_rsc_managed)) { // Recurring actions for active roles needed start = start_action(rsc, rsc->allocated_to, TRUE); } pe_rsc_trace(rsc, "Creating any recurring actions needed for %s", rsc->id); for (xmlNode *op = first_named_child(rsc->ops_xml, "op"); op != NULL; op = crm_next_same_xml(op)) { struct op_history op_history = { NULL, }; if (!is_recurring_history(rsc, op, &op_history)) { continue; } if (start != NULL) { recurring_op_for_active(rsc, start, rsc->allocated_to, &op_history); } recurring_op_for_inactive(rsc, rsc->allocated_to, &op_history); free(op_history.key); } } /*! * \internal * \brief Create an executor cancel action * * \param[in,out] rsc Resource of action to cancel * \param[in] task Name of action to cancel * \param[in] interval_ms Interval of action to cancel * \param[in] node Node of action to cancel * * \return Created op */ pcmk_action_t * pcmk__new_cancel_action(pcmk_resource_t *rsc, const char *task, guint interval_ms, const pcmk_node_t *node) { pcmk_action_t *cancel_op = NULL; char *key = NULL; char *interval_ms_s = NULL; CRM_ASSERT((rsc != NULL) && (task != NULL) && (node != NULL)); // @TODO dangerous if possible to schedule another action with this key key = pcmk__op_key(rsc->id, task, interval_ms); cancel_op = custom_action(rsc, key, PCMK_ACTION_CANCEL, node, FALSE, rsc->cluster); pcmk__str_update(&cancel_op->task, PCMK_ACTION_CANCEL); pcmk__str_update(&cancel_op->cancel_task, task); interval_ms_s = crm_strdup_printf("%u", interval_ms); add_hash_param(cancel_op->meta, XML_LRM_ATTR_TASK, task); add_hash_param(cancel_op->meta, XML_LRM_ATTR_INTERVAL_MS, interval_ms_s); free(interval_ms_s); return cancel_op; } /*! * \internal * \brief Schedule cancellation of a recurring action * * \param[in,out] rsc Resource that action is for * \param[in] call_id Action's call ID from history * \param[in] task Action name * \param[in] interval_ms Action interval * \param[in] node Node that history entry is for * \param[in] reason Short description of why action is cancelled */ void pcmk__schedule_cancel(pcmk_resource_t *rsc, const char *call_id, const char *task, guint interval_ms, const pcmk_node_t *node, const char *reason) { pcmk_action_t *cancel = NULL; CRM_CHECK((rsc != NULL) && (task != NULL) && (node != NULL) && (reason != NULL), return); crm_info("Recurring %s-interval %s for %s will be stopped on %s: %s", pcmk__readable_interval(interval_ms), task, rsc->id, pe__node_name(node), reason); cancel = pcmk__new_cancel_action(rsc, task, interval_ms, node); add_hash_param(cancel->meta, XML_LRM_ATTR_CALLID, call_id); // Cancellations happen after stops pcmk__new_ordering(rsc, stop_key(rsc), NULL, rsc, NULL, cancel, pcmk__ar_ordered, rsc->cluster); } /*! * \internal * \brief Create a recurring action marked as needing rescheduling if active * * \param[in,out] rsc Resource that action is for * \param[in] task Name of action being rescheduled * \param[in] interval_ms Action interval (in milliseconds) * \param[in,out] node Node where action should be rescheduled */ void pcmk__reschedule_recurring(pcmk_resource_t *rsc, const char *task, guint interval_ms, pcmk_node_t *node) { pcmk_action_t *op = NULL; trigger_unfencing(rsc, node, "Device parameters changed (reschedule)", NULL, rsc->cluster); op = custom_action(rsc, pcmk__op_key(rsc->id, task, interval_ms), task, node, TRUE, rsc->cluster); pe__set_action_flags(op, pcmk_action_reschedule); } /*! * \internal * \brief Check whether an action is recurring * * \param[in] action Action to check * * \return true if \p action has a nonzero interval, otherwise false */ bool pcmk__action_is_recurring(const pcmk_action_t *action) { guint interval_ms = 0; if (pcmk__guint_from_hash(action->meta, XML_LRM_ATTR_INTERVAL_MS, 0, &interval_ms) != pcmk_rc_ok) { return false; } return (interval_ms > 0); } diff --git a/lib/pacemaker/pcmk_sched_resource.c b/lib/pacemaker/pcmk_sched_resource.c index 908c43425a..c1a8db25fa 100644 --- a/lib/pacemaker/pcmk_sched_resource.c +++ b/lib/pacemaker/pcmk_sched_resource.c @@ -1,773 +1,773 @@ /* * Copyright 2014-2023 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU General Public License version 2 * or later (GPLv2+) WITHOUT ANY WARRANTY. */ #include #include #include #include #include #include "libpacemaker_private.h" // Resource assignment methods by resource variant static pcmk_assignment_methods_t assignment_methods[] = { { pcmk__primitive_assign, pcmk__primitive_create_actions, pcmk__probe_rsc_on_node, pcmk__primitive_internal_constraints, pcmk__primitive_apply_coloc_score, pcmk__colocated_resources, pcmk__with_primitive_colocations, pcmk__primitive_with_colocations, pcmk__add_colocated_node_scores, pcmk__apply_location, pcmk__primitive_action_flags, pcmk__update_ordered_actions, pcmk__output_resource_actions, pcmk__add_rsc_actions_to_graph, pcmk__primitive_add_graph_meta, pcmk__primitive_add_utilization, pcmk__primitive_shutdown_lock, }, { pcmk__group_assign, pcmk__group_create_actions, pcmk__probe_rsc_on_node, pcmk__group_internal_constraints, pcmk__group_apply_coloc_score, pcmk__group_colocated_resources, pcmk__with_group_colocations, pcmk__group_with_colocations, pcmk__group_add_colocated_node_scores, pcmk__group_apply_location, pcmk__group_action_flags, pcmk__group_update_ordered_actions, pcmk__output_resource_actions, pcmk__add_rsc_actions_to_graph, pcmk__noop_add_graph_meta, pcmk__group_add_utilization, pcmk__group_shutdown_lock, }, { pcmk__clone_assign, pcmk__clone_create_actions, pcmk__clone_create_probe, pcmk__clone_internal_constraints, pcmk__clone_apply_coloc_score, pcmk__colocated_resources, pcmk__with_clone_colocations, pcmk__clone_with_colocations, pcmk__add_colocated_node_scores, pcmk__clone_apply_location, pcmk__clone_action_flags, pcmk__instance_update_ordered_actions, pcmk__output_resource_actions, pcmk__clone_add_actions_to_graph, pcmk__clone_add_graph_meta, pcmk__clone_add_utilization, pcmk__clone_shutdown_lock, }, { pcmk__bundle_assign, pcmk__bundle_create_actions, pcmk__bundle_create_probe, pcmk__bundle_internal_constraints, pcmk__bundle_apply_coloc_score, pcmk__colocated_resources, pcmk__with_bundle_colocations, pcmk__bundle_with_colocations, pcmk__add_colocated_node_scores, pcmk__bundle_apply_location, pcmk__bundle_action_flags, pcmk__instance_update_ordered_actions, pcmk__output_bundle_actions, pcmk__bundle_add_actions_to_graph, pcmk__noop_add_graph_meta, pcmk__bundle_add_utilization, pcmk__bundle_shutdown_lock, } }; /*! * \internal * \brief Check whether a resource's agent standard, provider, or type changed * * \param[in,out] rsc Resource to check * \param[in,out] node Node needing unfencing if agent changed * \param[in] rsc_entry XML with previously known agent information * \param[in] active_on_node Whether \p rsc is active on \p node * * \return true if agent for \p rsc changed, otherwise false */ bool pcmk__rsc_agent_changed(pcmk_resource_t *rsc, pcmk_node_t *node, const xmlNode *rsc_entry, bool active_on_node) { bool changed = false; const char *attr_list[] = { XML_ATTR_TYPE, XML_AGENT_ATTR_CLASS, XML_AGENT_ATTR_PROVIDER }; for (int i = 0; i < PCMK__NELEM(attr_list); i++) { const char *value = crm_element_value(rsc->xml, attr_list[i]); const char *old_value = crm_element_value(rsc_entry, attr_list[i]); if (!pcmk__str_eq(value, old_value, pcmk__str_none)) { changed = true; trigger_unfencing(rsc, node, "Device definition changed", NULL, rsc->cluster); if (active_on_node) { crm_notice("Forcing restart of %s on %s " "because %s changed from '%s' to '%s'", rsc->id, pe__node_name(node), attr_list[i], pcmk__s(old_value, ""), pcmk__s(value, "")); } } } if (changed && active_on_node) { // Make sure the resource is restarted custom_action(rsc, stop_key(rsc), PCMK_ACTION_STOP, node, FALSE, rsc->cluster); pe__set_resource_flags(rsc, pcmk_rsc_start_pending); } return changed; } /*! * \internal * \brief Add resource (and any matching children) to list if it matches ID * * \param[in] result List to add resource to * \param[in] rsc Resource to check * \param[in] id ID to match * * \return (Possibly new) head of list */ static GList * add_rsc_if_matching(GList *result, pcmk_resource_t *rsc, const char *id) { if ((strcmp(rsc->id, id) == 0) || ((rsc->clone_name != NULL) && (strcmp(rsc->clone_name, id) == 0))) { result = g_list_prepend(result, rsc); } for (GList *iter = rsc->children; iter != NULL; iter = iter->next) { pcmk_resource_t *child = (pcmk_resource_t *) iter->data; result = add_rsc_if_matching(result, child, id); } return result; } /*! * \internal * \brief Find all resources matching a given ID by either ID or clone name * * \param[in] id Resource ID to check * \param[in] scheduler Scheduler data * * \return List of all resources that match \p id * \note The caller is responsible for freeing the return value with * g_list_free(). */ GList * pcmk__rscs_matching_id(const char *id, const pcmk_scheduler_t *scheduler) { GList *result = NULL; CRM_CHECK((id != NULL) && (scheduler != NULL), return NULL); for (GList *iter = scheduler->resources; iter != NULL; iter = iter->next) { result = add_rsc_if_matching(result, (pcmk_resource_t *) iter->data, id); } return result; } /*! * \internal * \brief Set the variant-appropriate assignment methods for a resource * * \param[in,out] data Resource to set assignment methods for * \param[in] user_data Ignored */ static void set_assignment_methods_for_rsc(gpointer data, gpointer user_data) { pcmk_resource_t *rsc = data; rsc->cmds = &assignment_methods[rsc->variant]; g_list_foreach(rsc->children, set_assignment_methods_for_rsc, NULL); } /*! * \internal * \brief Set the variant-appropriate assignment methods for all resources * * \param[in,out] scheduler Scheduler data */ void pcmk__set_assignment_methods(pcmk_scheduler_t *scheduler) { g_list_foreach(scheduler->resources, set_assignment_methods_for_rsc, NULL); } /*! * \internal * \brief Wrapper for colocated_resources() method for readability * * \param[in] rsc Resource to add to colocated list * \param[in] orig_rsc Resource originally requested * \param[in,out] list Pointer to list to add to * * \return (Possibly new) head of list */ static inline void add_colocated_resources(const pcmk_resource_t *rsc, const pcmk_resource_t *orig_rsc, GList **list) { *list = rsc->cmds->colocated_resources(rsc, orig_rsc, *list); } // Shared implementation of pcmk_assignment_methods_t:colocated_resources() GList * pcmk__colocated_resources(const pcmk_resource_t *rsc, const pcmk_resource_t *orig_rsc, GList *colocated_rscs) { const GList *iter = NULL; GList *colocations = NULL; if (orig_rsc == NULL) { orig_rsc = rsc; } if ((rsc == NULL) || (g_list_find(colocated_rscs, rsc) != NULL)) { return colocated_rscs; } pe_rsc_trace(orig_rsc, "%s is in colocation chain with %s", rsc->id, orig_rsc->id); colocated_rscs = g_list_prepend(colocated_rscs, (gpointer) rsc); // Follow colocations where this resource is the dependent resource colocations = pcmk__this_with_colocations(rsc); for (iter = colocations; iter != NULL; iter = iter->next) { const pcmk__colocation_t *constraint = iter->data; const pcmk_resource_t *primary = constraint->primary; if (primary == orig_rsc) { continue; // Break colocation loop } if ((constraint->score == INFINITY) && (pcmk__colocation_affects(rsc, primary, constraint, true) == pcmk__coloc_affects_location)) { add_colocated_resources(primary, orig_rsc, &colocated_rscs); } } g_list_free(colocations); // Follow colocations where this resource is the primary resource colocations = pcmk__with_this_colocations(rsc); for (iter = colocations; iter != NULL; iter = iter->next) { const pcmk__colocation_t *constraint = iter->data; const pcmk_resource_t *dependent = constraint->dependent; if (dependent == orig_rsc) { continue; // Break colocation loop } if (pe_rsc_is_clone(rsc) && !pe_rsc_is_clone(dependent)) { continue; // We can't be sure whether dependent will be colocated } if ((constraint->score == INFINITY) && (pcmk__colocation_affects(dependent, rsc, constraint, true) == pcmk__coloc_affects_location)) { add_colocated_resources(dependent, orig_rsc, &colocated_rscs); } } g_list_free(colocations); return colocated_rscs; } // No-op function for variants that don't need to implement add_graph_meta() void pcmk__noop_add_graph_meta(const pcmk_resource_t *rsc, xmlNode *xml) { } /*! * \internal * \brief Output a summary of scheduled actions for a resource * * \param[in,out] rsc Resource to output actions for */ void pcmk__output_resource_actions(pcmk_resource_t *rsc) { pcmk_node_t *next = NULL; pcmk_node_t *current = NULL; pcmk__output_t *out = NULL; CRM_ASSERT(rsc != NULL); out = rsc->cluster->priv; if (rsc->children != NULL) { for (GList *iter = rsc->children; iter != NULL; iter = iter->next) { pcmk_resource_t *child = (pcmk_resource_t *) iter->data; child->cmds->output_actions(child); } return; } next = rsc->allocated_to; if (rsc->running_on) { current = pe__current_node(rsc); if (rsc->role == pcmk_role_stopped) { /* This can occur when resources are being recovered because * the current role can change in pcmk__primitive_create_actions() */ rsc->role = pcmk_role_started; } } if ((current == NULL) && pcmk_is_set(rsc->flags, pcmk_rsc_removed)) { /* Don't log stopped orphans */ return; } out->message(out, "rsc-action", rsc, current, next); } /*! * \internal * \brief Add a resource to a node's list of assigned resources * * \param[in,out] node Node to add resource to * \param[in] rsc Resource to add */ static inline void add_assigned_resource(pcmk_node_t *node, pcmk_resource_t *rsc) { node->details->allocated_rsc = g_list_prepend(node->details->allocated_rsc, rsc); } /*! * \internal * \brief Assign a specified resource (of any variant) to a node * * Assign a specified resource and its children (if any) to a specified node, if * the node can run the resource (or unconditionally, if \p force is true). Mark * the resources as no longer provisional. * * If a resource can't be assigned (or \p node is \c NULL), unassign any * previous assignment. If \p stop_if_fail is \c true, set next role to stopped * and update any existing actions scheduled for the resource. * * \param[in,out] rsc Resource to assign * \param[in,out] node Node to assign \p rsc to * \param[in] force If true, assign to \p node even if unavailable * \param[in] stop_if_fail If \c true and either \p rsc can't be assigned * or \p chosen is \c NULL, set next role to * stopped and update existing actions (if \p rsc * is not a primitive, this applies to its * primitive descendants instead) * * \return \c true if the assignment of \p rsc changed, or \c false otherwise * * \note Assigning a resource to the NULL node using this function is different * from calling pcmk__unassign_resource(), in that it may also update any * actions created for the resource. * \note The \c pcmk_assignment_methods_t:assign() method is preferred, unless * a resource should be assigned to the \c NULL node or every resource in * a tree should be assigned to the same node. * \note If \p stop_if_fail is \c false, then \c pcmk__unassign_resource() can * completely undo the assignment. A successful assignment can be either * undone or left alone as final. A failed assignment has the same effect * as calling pcmk__unassign_resource(); there are no side effects on * roles or actions. */ bool pcmk__assign_resource(pcmk_resource_t *rsc, pcmk_node_t *node, bool force, bool stop_if_fail) { bool changed = false; CRM_ASSERT(rsc != NULL); if (rsc->children != NULL) { for (GList *iter = rsc->children; iter != NULL; iter = iter->next) { pcmk_resource_t *child_rsc = iter->data; changed |= pcmk__assign_resource(child_rsc, node, force, stop_if_fail); } return changed; } // Assigning a primitive if (!force && (node != NULL) && ((node->weight < 0) // Allow graph to assume that guest node connections will come up || (!pcmk__node_available(node, true, false) && !pe__is_guest_node(node)))) { - pe_rsc_debug(rsc, - "All nodes for resource %s are unavailable, unclean or " - "shutting down (%s can%s run resources, with score %s)", - rsc->id, pe__node_name(node), - (pcmk__node_available(node, true, false)? "" : "not"), - pcmk_readable_score(node->weight)); + pcmk__rsc_debug(rsc, + "All nodes for resource %s are unavailable, unclean or " + "shutting down (%s can%s run resources, with score %s)", + rsc->id, pe__node_name(node), + (pcmk__node_available(node, true, false)? "" : "not"), + pcmk_readable_score(node->weight)); if (stop_if_fail) { pe__set_next_role(rsc, pcmk_role_stopped, "node availability"); } node = NULL; } if (rsc->allocated_to != NULL) { changed = !pe__same_node(rsc->allocated_to, node); } else { changed = (node != NULL); } pcmk__unassign_resource(rsc); pe__clear_resource_flags(rsc, pcmk_rsc_unassigned); if (node == NULL) { char *rc_stopped = NULL; - pe_rsc_debug(rsc, "Could not assign %s to a node", rsc->id); + pcmk__rsc_debug(rsc, "Could not assign %s to a node", rsc->id); if (!stop_if_fail) { return changed; } pe__set_next_role(rsc, pcmk_role_stopped, "unable to assign"); for (GList *iter = rsc->actions; iter != NULL; iter = iter->next) { pcmk_action_t *op = (pcmk_action_t *) iter->data; - pe_rsc_debug(rsc, "Updating %s for %s assignment failure", - op->uuid, rsc->id); + pcmk__rsc_debug(rsc, "Updating %s for %s assignment failure", + op->uuid, rsc->id); if (pcmk__str_eq(op->task, PCMK_ACTION_STOP, pcmk__str_none)) { pe__clear_action_flags(op, pcmk_action_optional); } else if (pcmk__str_eq(op->task, PCMK_ACTION_START, pcmk__str_none)) { pe__clear_action_flags(op, pcmk_action_runnable); } else { // Cancel recurring actions, unless for stopped state const char *interval_ms_s = NULL; const char *target_rc_s = NULL; interval_ms_s = g_hash_table_lookup(op->meta, XML_LRM_ATTR_INTERVAL_MS); target_rc_s = g_hash_table_lookup(op->meta, XML_ATTR_TE_TARGET_RC); if (rc_stopped == NULL) { rc_stopped = pcmk__itoa(PCMK_OCF_NOT_RUNNING); } if (!pcmk__str_eq(interval_ms_s, "0", pcmk__str_null_matches) && !pcmk__str_eq(rc_stopped, target_rc_s, pcmk__str_none)) { pe__clear_action_flags(op, pcmk_action_runnable); } } } free(rc_stopped); return changed; } - pe_rsc_debug(rsc, "Assigning %s to %s", rsc->id, pe__node_name(node)); + pcmk__rsc_debug(rsc, "Assigning %s to %s", rsc->id, pe__node_name(node)); rsc->allocated_to = pe__copy_node(node); add_assigned_resource(node, rsc); node->details->num_resources++; node->count++; pcmk__consume_node_capacity(node->details->utilization, rsc); if (pcmk_is_set(rsc->cluster->flags, pcmk_sched_show_utilization)) { pcmk__output_t *out = rsc->cluster->priv; out->message(out, "resource-util", rsc, node, __func__); } return changed; } /*! * \internal * \brief Remove any node assignment from a specified resource and its children * * If a specified resource has been assigned to a node, remove that assignment * and mark the resource as provisional again. * * \param[in,out] rsc Resource to unassign * * \note This function is called recursively on \p rsc and its children. */ void pcmk__unassign_resource(pcmk_resource_t *rsc) { pcmk_node_t *old = rsc->allocated_to; if (old == NULL) { crm_info("Unassigning %s", rsc->id); } else { crm_info("Unassigning %s from %s", rsc->id, pe__node_name(old)); } pe__set_resource_flags(rsc, pcmk_rsc_unassigned); if (rsc->children == NULL) { if (old == NULL) { return; } rsc->allocated_to = NULL; /* We're going to free the pcmk_node_t, but its details member is shared * and will remain, so update that appropriately first. */ old->details->allocated_rsc = g_list_remove(old->details->allocated_rsc, rsc); old->details->num_resources--; pcmk__release_node_capacity(old->details->utilization, rsc); free(old); return; } for (GList *iter = rsc->children; iter != NULL; iter = iter->next) { pcmk__unassign_resource((pcmk_resource_t *) iter->data); } } /*! * \internal * \brief Check whether a resource has reached its migration threshold on a node * * \param[in,out] rsc Resource to check * \param[in] node Node to check * \param[out] failed If threshold has been reached, this will be set to * resource that failed (possibly a parent of \p rsc) * * \return true if the migration threshold has been reached, false otherwise */ bool pcmk__threshold_reached(pcmk_resource_t *rsc, const pcmk_node_t *node, pcmk_resource_t **failed) { int fail_count, remaining_tries; pcmk_resource_t *rsc_to_ban = rsc; // Migration threshold of 0 means never force away if (rsc->migration_threshold == 0) { return false; } // If we're ignoring failures, also ignore the migration threshold if (pcmk_is_set(rsc->flags, pcmk_rsc_ignore_failure)) { return false; } // If there are no failures, there's no need to force away fail_count = pe_get_failcount(node, rsc, NULL, pcmk__fc_effective|pcmk__fc_fillers, NULL); if (fail_count <= 0) { return false; } // If failed resource is anonymous clone instance, we'll force clone away if (!pcmk_is_set(rsc->flags, pcmk_rsc_unique)) { rsc_to_ban = uber_parent(rsc); } // How many more times recovery will be tried on this node remaining_tries = rsc->migration_threshold - fail_count; if (remaining_tries <= 0) { crm_warn("%s cannot run on %s due to reaching migration threshold " "(clean up resource to allow again)" CRM_XS " failures=%d migration-threshold=%d", rsc_to_ban->id, pe__node_name(node), fail_count, rsc->migration_threshold); if (failed != NULL) { *failed = rsc_to_ban; } return true; } crm_info("%s can fail %d more time%s on " "%s before reaching migration threshold (%d)", rsc_to_ban->id, remaining_tries, pcmk__plural_s(remaining_tries), pe__node_name(node), rsc->migration_threshold); return false; } /*! * \internal * \brief Get a node's score * * \param[in] node Node with ID to check * \param[in] nodes List of nodes to look for \p node score in * * \return Node's score, or -INFINITY if not found */ static int get_node_score(const pcmk_node_t *node, GHashTable *nodes) { pcmk_node_t *found_node = NULL; if ((node != NULL) && (nodes != NULL)) { found_node = g_hash_table_lookup(nodes, node->details->id); } return (found_node == NULL)? -INFINITY : found_node->weight; } /*! * \internal * \brief Compare two resources according to which should be assigned first * * \param[in] a First resource to compare * \param[in] b Second resource to compare * \param[in] data Sorted list of all nodes in cluster * * \return -1 if \p a should be assigned before \b, 0 if they are equal, * or +1 if \p a should be assigned after \b */ static gint cmp_resources(gconstpointer a, gconstpointer b, gpointer data) { /* GLib insists that this function require gconstpointer arguments, but we * make a small, temporary change to each argument (setting the * pe_rsc_merging flag) during comparison */ pcmk_resource_t *resource1 = (pcmk_resource_t *) a; pcmk_resource_t *resource2 = (pcmk_resource_t *) b; const GList *nodes = data; int rc = 0; int r1_score = -INFINITY; int r2_score = -INFINITY; pcmk_node_t *r1_node = NULL; pcmk_node_t *r2_node = NULL; GHashTable *r1_nodes = NULL; GHashTable *r2_nodes = NULL; const char *reason = NULL; // Resources with highest priority should be assigned first reason = "priority"; r1_score = resource1->priority; r2_score = resource2->priority; if (r1_score > r2_score) { rc = -1; goto done; } if (r1_score < r2_score) { rc = 1; goto done; } // We need nodes to make any other useful comparisons reason = "no node list"; if (nodes == NULL) { goto done; } // Calculate and log node scores resource1->cmds->add_colocated_node_scores(resource1, NULL, resource1->id, &r1_nodes, NULL, 1, pcmk__coloc_select_this_with); resource2->cmds->add_colocated_node_scores(resource2, NULL, resource2->id, &r2_nodes, NULL, 1, pcmk__coloc_select_this_with); pe__show_node_scores(true, NULL, resource1->id, r1_nodes, resource1->cluster); pe__show_node_scores(true, NULL, resource2->id, r2_nodes, resource2->cluster); // The resource with highest score on its current node goes first reason = "current location"; if (resource1->running_on != NULL) { r1_node = pe__current_node(resource1); } if (resource2->running_on != NULL) { r2_node = pe__current_node(resource2); } r1_score = get_node_score(r1_node, r1_nodes); r2_score = get_node_score(r2_node, r2_nodes); if (r1_score > r2_score) { rc = -1; goto done; } if (r1_score < r2_score) { rc = 1; goto done; } // Otherwise a higher score on any node will do reason = "score"; for (const GList *iter = nodes; iter != NULL; iter = iter->next) { const pcmk_node_t *node = (const pcmk_node_t *) iter->data; r1_score = get_node_score(node, r1_nodes); r2_score = get_node_score(node, r2_nodes); if (r1_score > r2_score) { rc = -1; goto done; } if (r1_score < r2_score) { rc = 1; goto done; } } done: crm_trace("%s (%d)%s%s %c %s (%d)%s%s: %s", resource1->id, r1_score, ((r1_node == NULL)? "" : " on "), ((r1_node == NULL)? "" : r1_node->details->id), ((rc < 0)? '>' : ((rc > 0)? '<' : '=')), resource2->id, r2_score, ((r2_node == NULL)? "" : " on "), ((r2_node == NULL)? "" : r2_node->details->id), reason); if (r1_nodes != NULL) { g_hash_table_destroy(r1_nodes); } if (r2_nodes != NULL) { g_hash_table_destroy(r2_nodes); } return rc; } /*! * \internal * \brief Sort resources in the order they should be assigned to nodes * * \param[in,out] scheduler Scheduler data */ void pcmk__sort_resources(pcmk_scheduler_t *scheduler) { GList *nodes = g_list_copy(scheduler->nodes); nodes = pcmk__sort_nodes(nodes, NULL); scheduler->resources = g_list_sort_with_data(scheduler->resources, cmp_resources, nodes); g_list_free(nodes); } diff --git a/lib/pacemaker/pcmk_sched_utilization.c b/lib/pacemaker/pcmk_sched_utilization.c index 962a94c86a..6442a3a903 100644 --- a/lib/pacemaker/pcmk_sched_utilization.c +++ b/lib/pacemaker/pcmk_sched_utilization.c @@ -1,467 +1,467 @@ /* * Copyright 2014-2023 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU General Public License version 2 * or later (GPLv2+) WITHOUT ANY WARRANTY. */ #include #include #include #include "libpacemaker_private.h" /*! * \internal * \brief Get integer utilization from a string * * \param[in] s String representation of a node utilization value * * \return Integer equivalent of \p s * \todo It would make sense to restrict utilization values to nonnegative * integers, but the documentation just says "integers" and we didn't * restrict them initially, so for backward compatibility, allow any * integer. */ static int utilization_value(const char *s) { int value = 0; if ((s != NULL) && (pcmk__scan_min_int(s, &value, INT_MIN) == EINVAL)) { pe_warn("Using 0 for utilization instead of invalid value '%s'", value); value = 0; } return value; } /* * Functions for comparing node capacities */ struct compare_data { const pcmk_node_t *node1; const pcmk_node_t *node2; bool node2_only; int result; }; /*! * \internal * \brief Compare a single utilization attribute for two nodes * * Compare one utilization attribute for two nodes, decrementing the result if * the first node has greater capacity, and incrementing it if the second node * has greater capacity. * * \param[in] key Utilization attribute name to compare * \param[in] value Utilization attribute value to compare * \param[in,out] user_data Comparison data (as struct compare_data*) */ static void compare_utilization_value(gpointer key, gpointer value, gpointer user_data) { int node1_capacity = 0; int node2_capacity = 0; struct compare_data *data = user_data; const char *node2_value = NULL; if (data->node2_only) { if (g_hash_table_lookup(data->node1->details->utilization, key)) { return; // We've already compared this attribute } } else { node1_capacity = utilization_value((const char *) value); } node2_value = g_hash_table_lookup(data->node2->details->utilization, key); node2_capacity = utilization_value(node2_value); if (node1_capacity > node2_capacity) { data->result--; } else if (node1_capacity < node2_capacity) { data->result++; } } /*! * \internal * \brief Compare utilization capacities of two nodes * * \param[in] node1 First node to compare * \param[in] node2 Second node to compare * * \return Negative integer if node1 has more free capacity, * 0 if the capacities are equal, or a positive integer * if node2 has more free capacity */ int pcmk__compare_node_capacities(const pcmk_node_t *node1, const pcmk_node_t *node2) { struct compare_data data = { .node1 = node1, .node2 = node2, .node2_only = false, .result = 0, }; // Compare utilization values that node1 and maybe node2 have g_hash_table_foreach(node1->details->utilization, compare_utilization_value, &data); // Compare utilization values that only node2 has data.node2_only = true; g_hash_table_foreach(node2->details->utilization, compare_utilization_value, &data); return data.result; } /* * Functions for updating node capacities */ struct calculate_data { GHashTable *current_utilization; bool plus; }; /*! * \internal * \brief Update a single utilization attribute with a new value * * \param[in] key Name of utilization attribute to update * \param[in] value Value to add or substract * \param[in,out] user_data Calculation data (as struct calculate_data *) */ static void update_utilization_value(gpointer key, gpointer value, gpointer user_data) { int result = 0; const char *current = NULL; struct calculate_data *data = user_data; current = g_hash_table_lookup(data->current_utilization, key); if (data->plus) { result = utilization_value(current) + utilization_value(value); } else if (current) { result = utilization_value(current) - utilization_value(value); } g_hash_table_replace(data->current_utilization, strdup(key), pcmk__itoa(result)); } /*! * \internal * \brief Subtract a resource's utilization from node capacity * * \param[in,out] current_utilization Current node utilization attributes * \param[in] rsc Resource with utilization to subtract */ void pcmk__consume_node_capacity(GHashTable *current_utilization, const pcmk_resource_t *rsc) { struct calculate_data data = { .current_utilization = current_utilization, .plus = false, }; g_hash_table_foreach(rsc->utilization, update_utilization_value, &data); } /*! * \internal * \brief Add a resource's utilization to node capacity * * \param[in,out] current_utilization Current node utilization attributes * \param[in] rsc Resource with utilization to add */ void pcmk__release_node_capacity(GHashTable *current_utilization, const pcmk_resource_t *rsc) { struct calculate_data data = { .current_utilization = current_utilization, .plus = true, }; g_hash_table_foreach(rsc->utilization, update_utilization_value, &data); } /* * Functions for checking for sufficient node capacity */ struct capacity_data { const pcmk_node_t *node; const char *rsc_id; bool is_enough; }; /*! * \internal * \brief Check whether a single utilization attribute has sufficient capacity * * \param[in] key Name of utilization attribute to check * \param[in] value Amount of utilization required * \param[in,out] user_data Capacity data (as struct capacity_data *) */ static void check_capacity(gpointer key, gpointer value, gpointer user_data) { int required = 0; int remaining = 0; const char *node_value_s = NULL; struct capacity_data *data = user_data; node_value_s = g_hash_table_lookup(data->node->details->utilization, key); required = utilization_value(value); remaining = utilization_value(node_value_s); if (required > remaining) { crm_debug("Remaining capacity for %s on %s (%d) is insufficient " "for resource %s usage (%d)", (const char *) key, pe__node_name(data->node), remaining, data->rsc_id, required); data->is_enough = false; } } /*! * \internal * \brief Check whether a node has sufficient capacity for a resource * * \param[in] node Node to check * \param[in] rsc_id ID of resource to check (for debug logs only) * \param[in] utilization Required utilization amounts * * \return true if node has sufficient capacity for resource, otherwise false */ static bool have_enough_capacity(const pcmk_node_t *node, const char *rsc_id, GHashTable *utilization) { struct capacity_data data = { .node = node, .rsc_id = rsc_id, .is_enough = true, }; g_hash_table_foreach(utilization, check_capacity, &data); return data.is_enough; } /*! * \internal * \brief Sum the utilization requirements of a list of resources * * \param[in] orig_rsc Resource being assigned (for logging purposes) * \param[in] rscs Resources whose utilization should be summed * * \return Newly allocated hash table with sum of all utilization values * \note It is the caller's responsibility to free the return value using * g_hash_table_destroy(). */ static GHashTable * sum_resource_utilization(const pcmk_resource_t *orig_rsc, GList *rscs) { GHashTable *utilization = pcmk__strkey_table(free, free); for (GList *iter = rscs; iter != NULL; iter = iter->next) { pcmk_resource_t *rsc = (pcmk_resource_t *) iter->data; rsc->cmds->add_utilization(rsc, orig_rsc, rscs, utilization); } return utilization; } /*! * \internal * \brief Ban resource from nodes with insufficient utilization capacity * * \param[in,out] rsc Resource to check * * \return Allowed node for \p rsc with most spare capacity, if there are no * nodes with enough capacity for \p rsc and all its colocated resources */ const pcmk_node_t * pcmk__ban_insufficient_capacity(pcmk_resource_t *rsc) { bool any_capable = false; char *rscs_id = NULL; pcmk_node_t *node = NULL; const pcmk_node_t *most_capable_node = NULL; GList *colocated_rscs = NULL; GHashTable *unassigned_utilization = NULL; GHashTableIter iter; CRM_CHECK(rsc != NULL, return NULL); // The default placement strategy ignores utilization if (pcmk__str_eq(rsc->cluster->placement_strategy, "default", pcmk__str_casei)) { return NULL; } // Check whether any resources are colocated with this one colocated_rscs = rsc->cmds->colocated_resources(rsc, NULL, NULL); if (colocated_rscs == NULL) { return NULL; } rscs_id = crm_strdup_printf("%s and its colocated resources", rsc->id); // If rsc isn't in the list, add it so we include its utilization if (g_list_find(colocated_rscs, rsc) == NULL) { colocated_rscs = g_list_append(colocated_rscs, rsc); } // Sum utilization of colocated resources that haven't been assigned yet unassigned_utilization = sum_resource_utilization(rsc, colocated_rscs); // Check whether any node has enough capacity for all the resources g_hash_table_iter_init(&iter, rsc->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (void **) &node)) { if (!pcmk__node_available(node, true, false)) { continue; } if (have_enough_capacity(node, rscs_id, unassigned_utilization)) { any_capable = true; } // Keep track of node with most free capacity if ((most_capable_node == NULL) || (pcmk__compare_node_capacities(node, most_capable_node) < 0)) { most_capable_node = node; } } if (any_capable) { // If so, ban resource from any node with insufficient capacity g_hash_table_iter_init(&iter, rsc->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (void **) &node)) { if (pcmk__node_available(node, true, false) && !have_enough_capacity(node, rscs_id, unassigned_utilization)) { - pe_rsc_debug(rsc, "%s does not have enough capacity for %s", - pe__node_name(node), rscs_id); + pcmk__rsc_debug(rsc, "%s does not have enough capacity for %s", + pe__node_name(node), rscs_id); resource_location(rsc, node, -INFINITY, "__limit_utilization__", rsc->cluster); } } most_capable_node = NULL; } else { // Otherwise, ban from nodes with insufficient capacity for rsc alone g_hash_table_iter_init(&iter, rsc->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (void **) &node)) { if (pcmk__node_available(node, true, false) && !have_enough_capacity(node, rsc->id, rsc->utilization)) { - pe_rsc_debug(rsc, "%s does not have enough capacity for %s", - pe__node_name(node), rsc->id); + pcmk__rsc_debug(rsc, "%s does not have enough capacity for %s", + pe__node_name(node), rsc->id); resource_location(rsc, node, -INFINITY, "__limit_utilization__", rsc->cluster); } } } g_hash_table_destroy(unassigned_utilization); g_list_free(colocated_rscs); free(rscs_id); pe__show_node_scores(true, rsc, "Post-utilization", rsc->allowed_nodes, rsc->cluster); return most_capable_node; } /*! * \internal * \brief Create a new load_stopped pseudo-op for a node * * \param[in,out] node Node to create op for * * \return Newly created load_stopped op */ static pcmk_action_t * new_load_stopped_op(pcmk_node_t *node) { char *load_stopped_task = crm_strdup_printf(PCMK_ACTION_LOAD_STOPPED "_%s", node->details->uname); pcmk_action_t *load_stopped = get_pseudo_op(load_stopped_task, node->details->data_set); if (load_stopped->node == NULL) { load_stopped->node = pe__copy_node(node); pe__clear_action_flags(load_stopped, pcmk_action_optional); } free(load_stopped_task); return load_stopped; } /*! * \internal * \brief Create utilization-related internal constraints for a resource * * \param[in,out] rsc Resource to create constraints for * \param[in] allowed_nodes List of allowed next nodes for \p rsc */ void pcmk__create_utilization_constraints(pcmk_resource_t *rsc, const GList *allowed_nodes) { const GList *iter = NULL; pcmk_action_t *load_stopped = NULL; pe_rsc_trace(rsc, "Creating utilization constraints for %s - strategy: %s", rsc->id, rsc->cluster->placement_strategy); // "stop rsc then load_stopped" constraints for current nodes for (iter = rsc->running_on; iter != NULL; iter = iter->next) { load_stopped = new_load_stopped_op(iter->data); pcmk__new_ordering(rsc, stop_key(rsc), NULL, NULL, NULL, load_stopped, pcmk__ar_if_on_same_node_or_target, rsc->cluster); } // "load_stopped then start/migrate_to rsc" constraints for allowed nodes for (iter = allowed_nodes; iter; iter = iter->next) { load_stopped = new_load_stopped_op(iter->data); pcmk__new_ordering(NULL, NULL, load_stopped, rsc, start_key(rsc), NULL, pcmk__ar_if_on_same_node_or_target, rsc->cluster); pcmk__new_ordering(NULL, NULL, load_stopped, rsc, pcmk__op_key(rsc->id, PCMK_ACTION_MIGRATE_TO, 0), NULL, pcmk__ar_if_on_same_node_or_target, rsc->cluster); } } /*! * \internal * \brief Output node capacities if enabled * * \param[in] desc Prefix for output * \param[in,out] scheduler Scheduler data */ void pcmk__show_node_capacities(const char *desc, pcmk_scheduler_t *scheduler) { if (!pcmk_is_set(scheduler->flags, pcmk_sched_show_utilization)) { return; } for (const GList *iter = scheduler->nodes; iter != NULL; iter = iter->next) { const pcmk_node_t *node = (const pcmk_node_t *) iter->data; pcmk__output_t *out = scheduler->priv; out->message(out, "node-capacity", node, desc); } } diff --git a/lib/pacemaker/pcmk_scheduler.c b/lib/pacemaker/pcmk_scheduler.c index 86a2489022..9f65616ee3 100644 --- a/lib/pacemaker/pcmk_scheduler.c +++ b/lib/pacemaker/pcmk_scheduler.c @@ -1,820 +1,820 @@ /* * Copyright 2004-2023 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU General Public License version 2 * or later (GPLv2+) WITHOUT ANY WARRANTY. */ #include #include #include #include #include #include #include #include #include #include #include "libpacemaker_private.h" CRM_TRACE_INIT_DATA(pacemaker); /*! * \internal * \brief Do deferred action checks after assignment * * When unpacking the resource history, the scheduler checks for resource * configurations that have changed since an action was run. However, at that * time, bundles using the REMOTE_CONTAINER_HACK don't have their final * parameter information, so instead they add a deferred check to a list. This * function processes one entry in that list. * * \param[in,out] rsc Resource that action history is for * \param[in,out] node Node that action history is for * \param[in] rsc_op Action history entry * \param[in] check Type of deferred check to do */ static void check_params(pcmk_resource_t *rsc, pcmk_node_t *node, const xmlNode *rsc_op, enum pcmk__check_parameters check) { const char *reason = NULL; pcmk__op_digest_t *digest_data = NULL; switch (check) { case pcmk__check_active: if (pcmk__check_action_config(rsc, node, rsc_op) && pe_get_failcount(node, rsc, NULL, pcmk__fc_effective, NULL)) { reason = "action definition changed"; } break; case pcmk__check_last_failure: digest_data = rsc_action_digest_cmp(rsc, rsc_op, node, rsc->cluster); switch (digest_data->rc) { case pcmk__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 pcmk__digest_match: break; default: reason = "resource parameters have changed"; break; } break; } if (reason != NULL) { pe__clear_failcount(rsc, node, reason, rsc->cluster); } } /*! * \internal * \brief Check whether a resource has failcount clearing scheduled on a node * * \param[in] node Node to check * \param[in] rsc Resource to check * * \return true if \p rsc has failcount clearing scheduled on \p node, * otherwise false */ static bool failcount_clear_action_exists(const pcmk_node_t *node, const pcmk_resource_t *rsc) { GList *list = pe__resource_actions(rsc, node, PCMK_ACTION_CLEAR_FAILCOUNT, TRUE); if (list != NULL) { g_list_free(list); return true; } return false; } /*! * \internal * \brief Ban a resource from a node if it reached its failure threshold there * * \param[in,out] data Resource to check failure threshold for * \param[in] user_data Node to check resource on */ static void check_failure_threshold(gpointer data, gpointer user_data) { pcmk_resource_t *rsc = data; const pcmk_node_t *node = user_data; // If this is a collective resource, apply recursively to children instead if (rsc->children != NULL) { g_list_foreach(rsc->children, check_failure_threshold, user_data); return; } if (!failcount_clear_action_exists(node, rsc)) { /* Don't force the resource away from this node due to a failcount * that's going to be cleared. * * @TODO Failcount clearing can be scheduled in * pcmk__handle_rsc_config_changes() via process_rsc_history(), or in * schedule_resource_actions() via check_params(). This runs well before * then, so it cannot detect those, meaning we might check the migration * threshold when we shouldn't. Worst case, we stop or move the * resource, then move it back in the next transition. */ pcmk_resource_t *failed = NULL; if (pcmk__threshold_reached(rsc, node, &failed)) { resource_location(failed, node, -INFINITY, "__fail_limit__", rsc->cluster); } } } /*! * \internal * \brief If resource has exclusive discovery, ban node if not allowed * * Location constraints have a resource-discovery option that allows users to * specify where probes are done for the affected resource. If this is set to * exclusive, probes will only be done on nodes listed in exclusive constraints. * This function bans the resource from the node if the node is not listed. * * \param[in,out] data Resource to check * \param[in] user_data Node to check resource on */ static void apply_exclusive_discovery(gpointer data, gpointer user_data) { pcmk_resource_t *rsc = data; const pcmk_node_t *node = user_data; if (rsc->exclusive_discover || pe__const_top_resource(rsc, false)->exclusive_discover) { pcmk_node_t *match = NULL; // If this is a collective resource, apply recursively to children g_list_foreach(rsc->children, apply_exclusive_discovery, user_data); match = g_hash_table_lookup(rsc->allowed_nodes, node->details->id); if ((match != NULL) && (match->rsc_discover_mode != pcmk_probe_exclusive)) { match->weight = -INFINITY; } } } /*! * \internal * \brief Apply stickiness to a resource if appropriate * * \param[in,out] data Resource to check for stickiness * \param[in] user_data Ignored */ static void apply_stickiness(gpointer data, gpointer user_data) { pcmk_resource_t *rsc = data; pcmk_node_t *node = NULL; // If this is a collective resource, apply recursively to children instead if (rsc->children != NULL) { g_list_foreach(rsc->children, apply_stickiness, NULL); return; } /* A resource is sticky if it is managed, has stickiness configured, and is * active on a single node. */ if (!pcmk_is_set(rsc->flags, pcmk_rsc_managed) || (rsc->stickiness < 1) || !pcmk__list_of_1(rsc->running_on)) { return; } node = rsc->running_on->data; /* In a symmetric cluster, stickiness can always be used. In an * asymmetric cluster, we have to check whether the resource is still * allowed on the node, so we don't keep the resource somewhere it is no * longer explicitly enabled. */ if (!pcmk_is_set(rsc->cluster->flags, pcmk_sched_symmetric_cluster) && (g_hash_table_lookup(rsc->allowed_nodes, node->details->id) == NULL)) { - pe_rsc_debug(rsc, - "Ignoring %s stickiness because the cluster is " - "asymmetric and %s is not explicitly allowed", - rsc->id, pe__node_name(node)); + pcmk__rsc_debug(rsc, + "Ignoring %s stickiness because the cluster is " + "asymmetric and %s is not explicitly allowed", + rsc->id, pe__node_name(node)); return; } - pe_rsc_debug(rsc, "Resource %s has %d stickiness on %s", - rsc->id, rsc->stickiness, pe__node_name(node)); + pcmk__rsc_debug(rsc, "Resource %s has %d stickiness on %s", + rsc->id, rsc->stickiness, pe__node_name(node)); resource_location(rsc, node, rsc->stickiness, "stickiness", rsc->cluster); } /*! * \internal * \brief Apply shutdown locks for all resources as appropriate * * \param[in,out] scheduler Scheduler data */ static void apply_shutdown_locks(pcmk_scheduler_t *scheduler) { if (!pcmk_is_set(scheduler->flags, pcmk_sched_shutdown_lock)) { return; } for (GList *iter = scheduler->resources; iter != NULL; iter = iter->next) { pcmk_resource_t *rsc = (pcmk_resource_t *) iter->data; rsc->cmds->shutdown_lock(rsc); } } /*! * \internal * \brief Calculate the number of available nodes in the cluster * * \param[in,out] scheduler Scheduler data */ static void count_available_nodes(pcmk_scheduler_t *scheduler) { if (pcmk_is_set(scheduler->flags, pcmk_sched_no_compat)) { return; } // @COMPAT for API backward compatibility only (cluster does not use value) for (GList *iter = scheduler->nodes; iter != NULL; iter = iter->next) { pcmk_node_t *node = (pcmk_node_t *) iter->data; if ((node != NULL) && (node->weight >= 0) && node->details->online && (node->details->type != node_ping)) { scheduler->max_valid_nodes++; } } crm_trace("Online node count: %d", scheduler->max_valid_nodes); } /* * \internal * \brief Apply node-specific scheduling criteria * * After the CIB has been unpacked, process node-specific scheduling criteria * including shutdown locks, location constraints, resource stickiness, * migration thresholds, and exclusive resource discovery. */ static void apply_node_criteria(pcmk_scheduler_t *scheduler) { crm_trace("Applying node-specific scheduling criteria"); apply_shutdown_locks(scheduler); count_available_nodes(scheduler); pcmk__apply_locations(scheduler); g_list_foreach(scheduler->resources, apply_stickiness, NULL); for (GList *node_iter = scheduler->nodes; node_iter != NULL; node_iter = node_iter->next) { for (GList *rsc_iter = scheduler->resources; rsc_iter != NULL; rsc_iter = rsc_iter->next) { check_failure_threshold(rsc_iter->data, node_iter->data); apply_exclusive_discovery(rsc_iter->data, node_iter->data); } } } /*! * \internal * \brief Assign resources to nodes * * \param[in,out] scheduler Scheduler data */ static void assign_resources(pcmk_scheduler_t *scheduler) { GList *iter = NULL; crm_trace("Assigning resources to nodes"); if (!pcmk__str_eq(scheduler->placement_strategy, "default", pcmk__str_casei)) { pcmk__sort_resources(scheduler); } pcmk__show_node_capacities("Original", scheduler); if (pcmk_is_set(scheduler->flags, pcmk_sched_have_remote_nodes)) { /* Assign remote connection resources first (which will also assign any * colocation dependencies). If the connection is migrating, always * prefer the partial migration target. */ for (iter = scheduler->resources; iter != NULL; iter = iter->next) { pcmk_resource_t *rsc = (pcmk_resource_t *) iter->data; if (rsc->is_remote_node) { pe_rsc_trace(rsc, "Assigning remote connection resource '%s'", rsc->id); rsc->cmds->assign(rsc, rsc->partial_migration_target, true); } } } /* now do the rest of the resources */ for (iter = scheduler->resources; iter != NULL; iter = iter->next) { pcmk_resource_t *rsc = (pcmk_resource_t *) iter->data; if (!rsc->is_remote_node) { pe_rsc_trace(rsc, "Assigning %s resource '%s'", rsc->xml->name, rsc->id); rsc->cmds->assign(rsc, NULL, true); } } pcmk__show_node_capacities("Remaining", scheduler); } /*! * \internal * \brief Schedule fail count clearing on online nodes if resource is orphaned * * \param[in,out] data Resource to check * \param[in] user_data Ignored */ static void clear_failcounts_if_orphaned(gpointer data, gpointer user_data) { pcmk_resource_t *rsc = data; if (!pcmk_is_set(rsc->flags, pcmk_rsc_removed)) { return; } crm_trace("Clear fail counts for orphaned resource %s", rsc->id); /* There's no need to recurse into rsc->children because those * should just be unassigned clone instances. */ for (GList *iter = rsc->cluster->nodes; iter != NULL; iter = iter->next) { pcmk_node_t *node = (pcmk_node_t *) iter->data; pcmk_action_t *clear_op = NULL; if (!node->details->online) { continue; } if (pe_get_failcount(node, rsc, NULL, pcmk__fc_effective, NULL) == 0) { continue; } clear_op = pe__clear_failcount(rsc, node, "it is orphaned", rsc->cluster); /* We can't use order_action_then_stop() here because its * pcmk__ar_guest_allowed breaks things */ pcmk__new_ordering(clear_op->rsc, NULL, clear_op, rsc, stop_key(rsc), NULL, pcmk__ar_ordered, rsc->cluster); } } /*! * \internal * \brief Schedule any resource actions needed * * \param[in,out] scheduler Scheduler data */ static void schedule_resource_actions(pcmk_scheduler_t *scheduler) { // Process deferred action checks pe__foreach_param_check(scheduler, check_params); pe__free_param_checks(scheduler); if (pcmk_is_set(scheduler->flags, pcmk_sched_probe_resources)) { crm_trace("Scheduling probes"); pcmk__schedule_probes(scheduler); } if (pcmk_is_set(scheduler->flags, pcmk_sched_stop_removed_resources)) { g_list_foreach(scheduler->resources, clear_failcounts_if_orphaned, NULL); } crm_trace("Scheduling resource actions"); for (GList *iter = scheduler->resources; iter != NULL; iter = iter->next) { pcmk_resource_t *rsc = (pcmk_resource_t *) iter->data; rsc->cmds->create_actions(rsc); } } /*! * \internal * \brief Check whether a resource or any of its descendants are managed * * \param[in] rsc Resource to check * * \return true if resource or any descendant is managed, otherwise false */ static bool is_managed(const pcmk_resource_t *rsc) { if (pcmk_is_set(rsc->flags, pcmk_rsc_managed)) { return true; } for (GList *iter = rsc->children; iter != NULL; iter = iter->next) { if (is_managed((pcmk_resource_t *) iter->data)) { return true; } } return false; } /*! * \internal * \brief Check whether any resources in the cluster are managed * * \param[in] scheduler Scheduler data * * \return true if any resource is managed, otherwise false */ static bool any_managed_resources(const pcmk_scheduler_t *scheduler) { for (const GList *iter = scheduler->resources; iter != NULL; iter = iter->next) { if (is_managed((const pcmk_resource_t *) iter->data)) { return true; } } return false; } /*! * \internal * \brief Check whether a node requires fencing * * \param[in] node Node to check * \param[in] have_managed Whether any resource in cluster is managed * * \return true if \p node should be fenced, otherwise false */ static bool needs_fencing(const pcmk_node_t *node, bool have_managed) { return have_managed && node->details->unclean && pe_can_fence(node->details->data_set, node); } /*! * \internal * \brief Check whether a node requires shutdown * * \param[in] node Node to check * * \return true if \p node should be shut down, otherwise false */ static bool needs_shutdown(const pcmk_node_t *node) { if (pe__is_guest_or_remote_node(node)) { /* Do not send shutdown actions for Pacemaker Remote nodes. * @TODO We might come up with a good use for this in the future. */ return false; } return node->details->online && node->details->shutdown; } /*! * \internal * \brief Track and order non-DC fencing * * \param[in,out] list List of existing non-DC fencing actions * \param[in,out] action Fencing action to prepend to \p list * \param[in] scheduler Scheduler data * * \return (Possibly new) head of \p list */ static GList * add_nondc_fencing(GList *list, pcmk_action_t *action, const pcmk_scheduler_t *scheduler) { if (!pcmk_is_set(scheduler->flags, pcmk_sched_concurrent_fencing) && (list != 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((pcmk_action_t *) list->data, action, pcmk__ar_ordered); } return g_list_prepend(list, action); } /*! * \internal * \brief Schedule a node for fencing * * \param[in,out] node Node that requires fencing */ static pcmk_action_t * schedule_fencing(pcmk_node_t *node) { pcmk_action_t *fencing = pe_fence_op(node, NULL, FALSE, "node is unclean", FALSE, node->details->data_set); pe_warn("Scheduling node %s for fencing", pe__node_name(node)); pcmk__order_vs_fence(fencing, node->details->data_set); return fencing; } /*! * \internal * \brief Create and order node fencing and shutdown actions * * \param[in,out] scheduler Scheduler data */ static void schedule_fencing_and_shutdowns(pcmk_scheduler_t *scheduler) { pcmk_action_t *dc_down = NULL; bool integrity_lost = false; bool have_managed = any_managed_resources(scheduler); GList *fencing_ops = NULL; GList *shutdown_ops = NULL; crm_trace("Scheduling fencing and shutdowns as needed"); if (!have_managed) { crm_notice("No fencing will be done until there are resources " "to manage"); } // Check each node for whether it needs fencing or shutdown for (GList *iter = scheduler->nodes; iter != NULL; iter = iter->next) { pcmk_node_t *node = (pcmk_node_t *) iter->data; pcmk_action_t *fencing = NULL; /* 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 && have_managed && pe_can_fence(scheduler, node)) { pcmk__fence_guest(node); } continue; } if (needs_fencing(node, have_managed)) { fencing = schedule_fencing(node); // Track DC and non-DC fence actions separately if (node->details->is_dc) { dc_down = fencing; } else { fencing_ops = add_nondc_fencing(fencing_ops, fencing, scheduler); } } else if (needs_shutdown(node)) { pcmk_action_t *down_op = pcmk__new_shutdown_action(node); // Track DC and non-DC shutdown actions separately if (node->details->is_dc) { dc_down = down_op; } else { shutdown_ops = g_list_prepend(shutdown_ops, down_op); } } if ((fencing == NULL) && node->details->unclean) { integrity_lost = true; pe_warn("Node %s is unclean but cannot be fenced", pe__node_name(node)); } } if (integrity_lost) { if (!pcmk_is_set(scheduler->flags, pcmk_sched_fencing_enabled)) { pe_warn("Resource functionality and data integrity cannot be " "guaranteed (configure, enable, and test fencing to " "correct this)"); } else if (!pcmk_is_set(scheduler->flags, pcmk_sched_quorate)) { crm_notice("Unclean nodes will not be fenced until quorum is " "attained or no-quorum-policy is set to ignore"); } } if (dc_down != NULL) { /* Order any non-DC shutdowns before any DC shutdown, to avoid repeated * DC elections. However, we don't want to order non-DC shutdowns before * a DC *fencing*, because even though we don't want a node that's * shutting down to become DC, the DC fencing could be ordered before a * clone stop that's also ordered before the shutdowns, thus leading to * a graph loop. */ if (pcmk__str_eq(dc_down->task, PCMK_ACTION_DO_SHUTDOWN, pcmk__str_none)) { pcmk__order_after_each(dc_down, shutdown_ops); } // Order any non-DC fencing before any DC fencing or shutdown if (pcmk_is_set(scheduler->flags, pcmk_sched_concurrent_fencing)) { /* With concurrent fencing, order each non-DC fencing action * separately before any DC fencing or shutdown. */ pcmk__order_after_each(dc_down, fencing_ops); } else if (fencing_ops != NULL) { /* 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((pcmk_action_t *) fencing_ops->data, dc_down, pcmk__ar_ordered); } } g_list_free(fencing_ops); g_list_free(shutdown_ops); } static void log_resource_details(pcmk_scheduler_t *scheduler) { pcmk__output_t *out = scheduler->priv; GList *all = NULL; /* Due to the `crm_mon --node=` feature, out->message() for all the * resource-related messages expects a list of nodes that we are allowed to * output information for. Here, we create a wildcard to match all nodes. */ all = g_list_prepend(all, (gpointer) "*"); for (GList *item = scheduler->resources; item != NULL; item = item->next) { pcmk_resource_t *rsc = (pcmk_resource_t *) item->data; // Log all resources except inactive orphans if (!pcmk_is_set(rsc->flags, pcmk_rsc_removed) || (rsc->role != pcmk_role_stopped)) { out->message(out, crm_map_element_name(rsc->xml), 0, rsc, all, all); } } g_list_free(all); } static void log_all_actions(pcmk_scheduler_t *scheduler) { /* This only ever outputs to the log, so ignore whatever output object was * previously set and just log instead. */ pcmk__output_t *prev_out = scheduler->priv; pcmk__output_t *out = NULL; if (pcmk__log_output_new(&out) != pcmk_rc_ok) { return; } pe__register_messages(out); pcmk__register_lib_messages(out); pcmk__output_set_log_level(out, LOG_NOTICE); scheduler->priv = out; out->begin_list(out, NULL, NULL, "Actions"); pcmk__output_actions(scheduler); out->end_list(out); out->finish(out, CRM_EX_OK, true, NULL); pcmk__output_free(out); scheduler->priv = prev_out; } /*! * \internal * \brief Log all required but unrunnable actions at trace level * * \param[in] scheduler Scheduler data */ static void log_unrunnable_actions(const pcmk_scheduler_t *scheduler) { const uint64_t flags = pcmk_action_optional |pcmk_action_runnable |pcmk_action_pseudo; crm_trace("Required but unrunnable actions:"); for (const GList *iter = scheduler->actions; iter != NULL; iter = iter->next) { const pcmk_action_t *action = (const pcmk_action_t *) iter->data; if (!pcmk_any_flags_set(action->flags, flags)) { pcmk__log_action("\t", action, true); } } } /*! * \internal * \brief Unpack the CIB for scheduling * * \param[in,out] cib CIB XML to unpack (may be NULL if already unpacked) * \param[in] flags Scheduler flags to set in addition to defaults * \param[in,out] scheduler Scheduler data */ static void unpack_cib(xmlNode *cib, unsigned long long flags, pcmk_scheduler_t *scheduler) { const char* localhost_save = NULL; if (pcmk_is_set(scheduler->flags, pcmk_sched_have_status)) { crm_trace("Reusing previously calculated cluster status"); pe__set_working_set_flags(scheduler, flags); return; } if (scheduler->localhost) { localhost_save = scheduler->localhost; } CRM_ASSERT(cib != NULL); crm_trace("Calculating cluster status"); /* This will zero the entire struct without freeing anything first, so * callers should never call pcmk__schedule_actions() with a populated data * set unless pcmk_sched_have_status is set (i.e. cluster_status() was * previously called, whether directly or via pcmk__schedule_actions()). */ set_working_set_defaults(scheduler); if (localhost_save) { scheduler->localhost = localhost_save; } pe__set_working_set_flags(scheduler, flags); scheduler->input = cib; cluster_status(scheduler); // Sets pcmk_sched_have_status } /*! * \internal * \brief Run the scheduler for a given CIB * * \param[in,out] cib CIB XML to use as scheduler input * \param[in] flags Scheduler flags to set in addition to defaults * \param[in,out] scheduler Scheduler data */ void pcmk__schedule_actions(xmlNode *cib, unsigned long long flags, pcmk_scheduler_t *scheduler) { unpack_cib(cib, flags, scheduler); pcmk__set_assignment_methods(scheduler); pcmk__apply_node_health(scheduler); pcmk__unpack_constraints(scheduler); if (pcmk_is_set(scheduler->flags, pcmk_sched_validate_only)) { return; } if (!pcmk_is_set(scheduler->flags, pcmk_sched_location_only) && pcmk__is_daemon) { log_resource_details(scheduler); } apply_node_criteria(scheduler); if (pcmk_is_set(scheduler->flags, pcmk_sched_location_only)) { return; } pcmk__create_internal_constraints(scheduler); pcmk__handle_rsc_config_changes(scheduler); assign_resources(scheduler); schedule_resource_actions(scheduler); /* Remote ordering constraints need to happen prior to calculating fencing * because it is one more place we can mark nodes as needing fencing. */ pcmk__order_remote_connection_actions(scheduler); schedule_fencing_and_shutdowns(scheduler); pcmk__apply_orderings(scheduler); log_all_actions(scheduler); pcmk__create_graph(scheduler); if (get_crm_log_level() == LOG_TRACE) { log_unrunnable_actions(scheduler); } } diff --git a/lib/pengine/failcounts.c b/lib/pengine/failcounts.c index 90883a2f12..1c58667848 100644 --- a/lib/pengine/failcounts.c +++ b/lib/pengine/failcounts.c @@ -1,469 +1,469 @@ /* * Copyright 2008-2023 the Pacemaker project contributors * * This source code is licensed under the GNU Lesser General Public License * version 2.1 or later (LGPLv2.1+) WITHOUT ANY WARRANTY. */ #include #include #include #include #include #include #include #include #include static gboolean is_matched_failure(const char *rsc_id, const xmlNode *conf_op_xml, const xmlNode *lrm_op_xml) { gboolean matched = FALSE; const char *conf_op_name = NULL; const char *lrm_op_task = NULL; const char *conf_op_interval_spec = NULL; guint conf_op_interval_ms = 0; guint lrm_op_interval_ms = 0; const char *lrm_op_id = NULL; char *last_failure_key = NULL; if (rsc_id == NULL || conf_op_xml == NULL || lrm_op_xml == NULL) { return FALSE; } // Get name and interval from configured op conf_op_name = crm_element_value(conf_op_xml, "name"); conf_op_interval_spec = crm_element_value(conf_op_xml, XML_LRM_ATTR_INTERVAL); conf_op_interval_ms = crm_parse_interval_spec(conf_op_interval_spec); // Get name and interval from op history entry lrm_op_task = crm_element_value(lrm_op_xml, XML_LRM_ATTR_TASK); crm_element_value_ms(lrm_op_xml, XML_LRM_ATTR_INTERVAL_MS, &lrm_op_interval_ms); if ((conf_op_interval_ms != lrm_op_interval_ms) || !pcmk__str_eq(conf_op_name, lrm_op_task, pcmk__str_casei)) { return FALSE; } lrm_op_id = ID(lrm_op_xml); last_failure_key = pcmk__op_key(rsc_id, "last_failure", 0); if (pcmk__str_eq(last_failure_key, lrm_op_id, pcmk__str_casei)) { matched = TRUE; } else { char *expected_op_key = pcmk__op_key(rsc_id, conf_op_name, conf_op_interval_ms); if (pcmk__str_eq(expected_op_key, lrm_op_id, pcmk__str_casei)) { int rc = 0; int target_rc = pe__target_rc_from_xml(lrm_op_xml); crm_element_value_int(lrm_op_xml, XML_LRM_ATTR_RC, &rc); if (rc != target_rc) { matched = TRUE; } } free(expected_op_key); } free(last_failure_key); return matched; } static gboolean block_failure(const pcmk_node_t *node, pcmk_resource_t *rsc, const xmlNode *xml_op) { char *xml_name = clone_strip(rsc->id); /* @TODO This xpath search occurs after template expansion, but it is unable * to properly detect on-fail in id-ref, operation meta-attributes, or * op_defaults, or evaluate rules. * * Also, on-fail defaults to block (in unpack_operation()) for stop actions * when stonith is disabled. * * Ideally, we'd unpack the operation before this point, and pass in a * meta-attributes table that takes all that into consideration. */ char *xpath = crm_strdup_printf("//" XML_CIB_TAG_RESOURCE "[@" XML_ATTR_ID "='%s']" "//" XML_ATTR_OP "[@" XML_OP_ATTR_ON_FAIL "='block']", xml_name); xmlXPathObject *xpathObj = xpath_search(rsc->xml, xpath); gboolean should_block = FALSE; free(xpath); if (xpathObj) { int max = numXpathResults(xpathObj); int lpc = 0; for (lpc = 0; lpc < max; lpc++) { xmlNode *pref = getXpathResult(xpathObj, lpc); if (xml_op) { should_block = is_matched_failure(xml_name, pref, xml_op); if (should_block) { break; } } else { const char *conf_op_name = NULL; const char *conf_op_interval_spec = NULL; guint conf_op_interval_ms = 0; char *lrm_op_xpath = NULL; xmlXPathObject *lrm_op_xpathObj = NULL; // Get name and interval from configured op conf_op_name = crm_element_value(pref, "name"); conf_op_interval_spec = crm_element_value(pref, XML_LRM_ATTR_INTERVAL); conf_op_interval_ms = crm_parse_interval_spec(conf_op_interval_spec); #define XPATH_FMT "//" XML_CIB_TAG_STATE "[@" XML_ATTR_UNAME "='%s']" \ "//" XML_LRM_TAG_RESOURCE "[@" XML_ATTR_ID "='%s']" \ "/" XML_LRM_TAG_RSC_OP "[@" XML_LRM_ATTR_TASK "='%s']" \ "[@" XML_LRM_ATTR_INTERVAL "='%u']" lrm_op_xpath = crm_strdup_printf(XPATH_FMT, node->details->uname, xml_name, conf_op_name, conf_op_interval_ms); lrm_op_xpathObj = xpath_search(rsc->cluster->input, lrm_op_xpath); free(lrm_op_xpath); if (lrm_op_xpathObj) { int max2 = numXpathResults(lrm_op_xpathObj); int lpc2 = 0; for (lpc2 = 0; lpc2 < max2; lpc2++) { xmlNode *lrm_op_xml = getXpathResult(lrm_op_xpathObj, lpc2); should_block = is_matched_failure(xml_name, pref, lrm_op_xml); if (should_block) { break; } } } freeXpathObject(lrm_op_xpathObj); if (should_block) { break; } } } } free(xml_name); freeXpathObject(xpathObj); return should_block; } /*! * \internal * \brief Get resource name as used in failure-related node attributes * * \param[in] rsc Resource to check * * \return Newly allocated string containing resource's fail name * \note The caller is responsible for freeing the result. */ static inline char * rsc_fail_name(const pcmk_resource_t *rsc) { const char *name = (rsc->clone_name? rsc->clone_name : rsc->id); return pcmk_is_set(rsc->flags, pcmk_rsc_unique)? strdup(name) : clone_strip(name); } /*! * \internal * \brief Compile regular expression to match a failure-related node attribute * * \param[in] prefix Attribute prefix to match * \param[in] rsc_name Resource name to match as used in failure attributes * \param[in] is_legacy Whether DC uses per-resource fail counts * \param[in] is_unique Whether the resource is a globally unique clone * \param[out] re Where to store resulting regular expression * * \return Standard Pacemaker return code * \note Fail attributes are named like PREFIX-RESOURCE#OP_INTERVAL. * The caller is responsible for freeing re with regfree(). */ static int generate_fail_regex(const char *prefix, const char *rsc_name, gboolean is_legacy, gboolean is_unique, regex_t *re) { char *pattern; /* @COMPAT DC < 1.1.17: Fail counts used to be per-resource rather than * per-operation. */ const char *op_pattern = (is_legacy? "" : "#.+_[0-9]+"); /* Ignore instance numbers for anything other than globally unique clones. * Anonymous clone fail counts could contain an instance number if the * clone was initially unique, failed, then was converted to anonymous. * @COMPAT Also, before 1.1.8, anonymous clone fail counts always contained * clone instance numbers. */ const char *instance_pattern = (is_unique? "" : "(:[0-9]+)?"); pattern = crm_strdup_printf("^%s-%s%s%s$", prefix, rsc_name, instance_pattern, op_pattern); if (regcomp(re, pattern, REG_EXTENDED|REG_NOSUB) != 0) { free(pattern); return EINVAL; } free(pattern); return pcmk_rc_ok; } /*! * \internal * \brief Compile regular expressions to match failure-related node attributes * * \param[in] rsc Resource being checked for failures * \param[out] failcount_re Storage for regular expression for fail count * \param[out] lastfailure_re Storage for regular expression for last failure * * \return Standard Pacemaker return code * \note On success, the caller is responsible for freeing the expressions with * regfree(). */ static int generate_fail_regexes(const pcmk_resource_t *rsc, regex_t *failcount_re, regex_t *lastfailure_re) { int rc = pcmk_rc_ok; char *rsc_name = rsc_fail_name(rsc); const char *version = crm_element_value(rsc->cluster->input, XML_ATTR_CRM_VERSION); // @COMPAT Pacemaker <= 1.1.16 used a single fail count per resource gboolean is_legacy = (compare_version(version, "3.0.13") < 0); if (generate_fail_regex(PCMK__FAIL_COUNT_PREFIX, rsc_name, is_legacy, pcmk_is_set(rsc->flags, pcmk_rsc_unique), failcount_re) != pcmk_rc_ok) { rc = EINVAL; } else if (generate_fail_regex(PCMK__LAST_FAILURE_PREFIX, rsc_name, is_legacy, pcmk_is_set(rsc->flags, pcmk_rsc_unique), lastfailure_re) != pcmk_rc_ok) { rc = EINVAL; regfree(failcount_re); } free(rsc_name); return rc; } // Data for fail-count-related iterators struct failcount_data { const pcmk_node_t *node;// Node to check for fail count pcmk_resource_t *rsc; // Resource to check for fail count uint32_t flags; // Fail count flags const xmlNode *xml_op; // History entry for expiration purposes (or NULL) regex_t failcount_re; // Fail count regular expression to match regex_t lastfailure_re; // Last failure regular expression to match int failcount; // Fail count so far time_t last_failure; // Time of most recent failure so far }; /*! * \internal * \brief Update fail count and last failure appropriately for a node attribute * * \param[in] key Node attribute name * \param[in] value Node attribute value * \param[in] user_data Fail count data to update */ static void update_failcount_for_attr(gpointer key, gpointer value, gpointer user_data) { struct failcount_data *fc_data = user_data; // If this is a matching fail count attribute, update fail count if (regexec(&(fc_data->failcount_re), (const char *) key, 0, NULL, 0) == 0) { fc_data->failcount = pcmk__add_scores(fc_data->failcount, char2score(value)); pe_rsc_trace(fc_data->rsc, "Added %s (%s) to %s fail count (now %s)", (const char *) key, (const char *) value, fc_data->rsc->id, pcmk_readable_score(fc_data->failcount)); return; } // If this is a matching last failure attribute, update last failure if (regexec(&(fc_data->lastfailure_re), (const char *) key, 0, NULL, 0) == 0) { long long last_ll; if (pcmk__scan_ll(value, &last_ll, 0LL) == pcmk_rc_ok) { fc_data->last_failure = (time_t) QB_MAX(fc_data->last_failure, last_ll); } } } /*! * \internal * \brief Update fail count and last failure appropriately for a filler resource * * \param[in] data Filler resource * \param[in] user_data Fail count data to update */ static void update_failcount_for_filler(gpointer data, gpointer user_data) { pcmk_resource_t *filler = data; struct failcount_data *fc_data = user_data; time_t filler_last_failure = 0; fc_data->failcount += pe_get_failcount(fc_data->node, filler, &filler_last_failure, fc_data->flags, fc_data->xml_op); fc_data->last_failure = QB_MAX(fc_data->last_failure, filler_last_failure); } /*! * \internal * \brief Get a resource's fail count on a node * * \param[in] node Node to check * \param[in,out] rsc Resource to check * \param[out] last_failure If not NULL, where to set time of most recent * failure of \p rsc on \p node * \param[in] flags Group of enum pcmk__fc_flags * \param[in] xml_op If not NULL, consider only the action in this * history entry when determining whether on-fail * is configured as "blocked", otherwise consider * all actions configured for \p rsc * * \return Fail count for \p rsc on \p node according to \p flags */ int pe_get_failcount(const pcmk_node_t *node, pcmk_resource_t *rsc, time_t *last_failure, uint32_t flags, const xmlNode *xml_op) { struct failcount_data fc_data = { .node = node, .rsc = rsc, .flags = flags, .xml_op = xml_op, .failcount = 0, .last_failure = (time_t) 0, }; // Calculate resource failcount as sum of all matching operation failcounts CRM_CHECK(generate_fail_regexes(rsc, &fc_data.failcount_re, &fc_data.lastfailure_re) == pcmk_rc_ok, return 0); g_hash_table_foreach(node->details->attrs, update_failcount_for_attr, &fc_data); regfree(&(fc_data.failcount_re)); regfree(&(fc_data.lastfailure_re)); // If failure blocks the resource, disregard any failure timeout if ((fc_data.failcount > 0) && (rsc->failure_timeout > 0) && block_failure(node, rsc, xml_op)) { pe_warn("Ignoring failure timeout %d for %s " "because it conflicts with on-fail=block", rsc->failure_timeout, rsc->id); rsc->failure_timeout = 0; } // If all failures have expired, ignore fail count if (pcmk_is_set(flags, pcmk__fc_effective) && (fc_data.failcount > 0) && (fc_data.last_failure > 0) && (rsc->failure_timeout != 0)) { time_t now = get_effective_time(rsc->cluster); if (now > (fc_data.last_failure + rsc->failure_timeout)) { - pe_rsc_debug(rsc, "Failcount for %s on %s expired after %ds", - rsc->id, pe__node_name(node), rsc->failure_timeout); + pcmk__rsc_debug(rsc, "Failcount for %s on %s expired after %ds", + rsc->id, pe__node_name(node), rsc->failure_timeout); fc_data.failcount = 0; } } /* Add the fail count of any filler resources, except that we never want the * fail counts of a bundle container's fillers to count towards the * container's fail count. * * Most importantly, a Pacemaker Remote connection to a bundle container * is a filler of the container, but can reside on a different node than the * container itself. Counting its fail count on its node towards the * container's fail count on that node could lead to attempting to stop the * container on the wrong node. */ if (pcmk_is_set(flags, pcmk__fc_fillers) && (rsc->fillers != NULL) && !pe_rsc_is_bundled(rsc)) { g_list_foreach(rsc->fillers, update_failcount_for_filler, &fc_data); if (fc_data.failcount > 0) { pcmk__rsc_info(rsc, "Container %s and the resources within it " "have failed %s time%s on %s", rsc->id, pcmk_readable_score(fc_data.failcount), pcmk__plural_s(fc_data.failcount), pe__node_name(node)); } } else if (fc_data.failcount > 0) { pcmk__rsc_info(rsc, "%s has failed %s time%s on %s", rsc->id, pcmk_readable_score(fc_data.failcount), pcmk__plural_s(fc_data.failcount), pe__node_name(node)); } if (last_failure != NULL) { if ((fc_data.failcount > 0) && (fc_data.last_failure > 0)) { *last_failure = fc_data.last_failure; } else { *last_failure = 0; } } return fc_data.failcount; } /*! * \brief Schedule a controller operation to clear a fail count * * \param[in,out] rsc Resource with failure * \param[in] node Node failure occurred on * \param[in] reason Readable description why needed (for logging) * \param[in,out] scheduler Scheduler data cluster * * \return Scheduled action */ pcmk_action_t * pe__clear_failcount(pcmk_resource_t *rsc, const pcmk_node_t *node, const char *reason, pcmk_scheduler_t *scheduler) { char *key = NULL; pcmk_action_t *clear = NULL; CRM_CHECK(rsc && node && reason && scheduler, return NULL); key = pcmk__op_key(rsc->id, PCMK_ACTION_CLEAR_FAILCOUNT, 0); clear = custom_action(rsc, key, PCMK_ACTION_CLEAR_FAILCOUNT, node, FALSE, scheduler); add_hash_param(clear->meta, XML_ATTR_TE_NOWAIT, XML_BOOLEAN_TRUE); crm_notice("Clearing failure of %s on %s because %s " CRM_XS " %s", rsc->id, pe__node_name(node), reason, clear->uuid); return clear; } diff --git a/lib/pengine/pe_actions.c b/lib/pengine/pe_actions.c index 86c574c0b0..f9c1ef722e 100644 --- a/lib/pengine/pe_actions.c +++ b/lib/pengine/pe_actions.c @@ -1,1871 +1,1874 @@ /* * Copyright 2004-2023 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU Lesser General Public License * version 2.1 or later (LGPLv2.1+) WITHOUT ANY WARRANTY. */ #include #include #include #include #include #include #include #include #include "pe_status_private.h" static void unpack_operation(pcmk_action_t *action, const xmlNode *xml_obj, guint interval_ms); static void add_singleton(pcmk_scheduler_t *scheduler, pcmk_action_t *action) { if (scheduler->singletons == NULL) { scheduler->singletons = pcmk__strkey_table(NULL, NULL); } g_hash_table_insert(scheduler->singletons, action->uuid, action); } static pcmk_action_t * lookup_singleton(pcmk_scheduler_t *scheduler, const char *action_uuid) { if (scheduler->singletons == NULL) { return NULL; } return g_hash_table_lookup(scheduler->singletons, action_uuid); } /*! * \internal * \brief Find an existing action that matches arguments * * \param[in] key Action key to match * \param[in] rsc Resource to match (if any) * \param[in] node Node to match (if any) * \param[in] scheduler Scheduler data * * \return Existing action that matches arguments (or NULL if none) */ static pcmk_action_t * find_existing_action(const char *key, const pcmk_resource_t *rsc, const pcmk_node_t *node, const pcmk_scheduler_t *scheduler) { GList *matches = NULL; pcmk_action_t *action = NULL; /* When rsc is NULL, it would be quicker to check scheduler->singletons, * but checking all scheduler->actions takes the node into account. */ matches = find_actions(((rsc == NULL)? scheduler->actions : rsc->actions), key, node); if (matches == NULL) { return NULL; } CRM_LOG_ASSERT(!pcmk__list_of_multiple(matches)); action = matches->data; g_list_free(matches); return action; } /*! * \internal * \brief Find the XML configuration corresponding to a specific action key * * \param[in] rsc Resource to find action configuration for * \param[in] key "RSC_ACTION_INTERVAL" of action to find * \param[in] include_disabled If false, do not return disabled actions * * \return XML configuration of desired action if any, otherwise NULL */ static xmlNode * find_exact_action_config(const pcmk_resource_t *rsc, const char *action_name, guint interval_ms, bool include_disabled) { for (xmlNode *operation = first_named_child(rsc->ops_xml, XML_ATTR_OP); operation != NULL; operation = crm_next_same_xml(operation)) { bool enabled = false; const char *config_name = NULL; const char *interval_spec = NULL; // @TODO This does not consider rules, defaults, etc. if (!include_disabled && (pcmk__xe_get_bool_attr(operation, "enabled", &enabled) == pcmk_rc_ok) && !enabled) { continue; } interval_spec = crm_element_value(operation, XML_LRM_ATTR_INTERVAL); if (crm_parse_interval_spec(interval_spec) != interval_ms) { continue; } config_name = crm_element_value(operation, "name"); if (pcmk__str_eq(action_name, config_name, pcmk__str_none)) { return operation; } } return NULL; } /*! * \internal * \brief Find the XML configuration of a resource action * * \param[in] rsc Resource to find action configuration for * \param[in] action_name Action name to search for * \param[in] interval_ms Action interval (in milliseconds) to search for * \param[in] include_disabled If false, do not return disabled actions * * \return XML configuration of desired action if any, otherwise NULL */ xmlNode * pcmk__find_action_config(const pcmk_resource_t *rsc, const char *action_name, guint interval_ms, bool include_disabled) { xmlNode *action_config = NULL; // Try requested action first action_config = find_exact_action_config(rsc, action_name, interval_ms, include_disabled); // For migrate_to and migrate_from actions, retry with "migrate" // @TODO This should be either documented or deprecated if ((action_config == NULL) && pcmk__str_any_of(action_name, PCMK_ACTION_MIGRATE_TO, PCMK_ACTION_MIGRATE_FROM, NULL)) { action_config = find_exact_action_config(rsc, "migrate", 0, include_disabled); } return action_config; } /*! * \internal * \brief Create a new action object * * \param[in] key Action key * \param[in] task Action name * \param[in,out] rsc Resource that action is for (if any) * \param[in] node Node that action is on (if any) * \param[in] optional Whether action should be considered optional * \param[in,out] scheduler Scheduler data * * \return Newly allocated action * \note This function takes ownership of \p key. It is the caller's * responsibility to free the return value with pe_free_action(). */ static pcmk_action_t * new_action(char *key, const char *task, pcmk_resource_t *rsc, const pcmk_node_t *node, bool optional, pcmk_scheduler_t *scheduler) { pcmk_action_t *action = calloc(1, sizeof(pcmk_action_t)); CRM_ASSERT(action != NULL); action->rsc = rsc; action->task = strdup(task); CRM_ASSERT(action->task != NULL); action->uuid = key; if (node) { action->node = pe__copy_node(node); } if (pcmk__str_eq(task, PCMK_ACTION_LRM_DELETE, pcmk__str_casei)) { // Resource history deletion for a node can be done on the DC pe__set_action_flags(action, pcmk_action_on_dc); } pe__set_action_flags(action, pcmk_action_runnable); if (optional) { pe__set_action_flags(action, pcmk_action_optional); } else { pe__clear_action_flags(action, pcmk_action_optional); } if (rsc == NULL) { action->meta = pcmk__strkey_table(free, free); } else { guint interval_ms = 0; parse_op_key(key, NULL, NULL, &interval_ms); action->op_entry = pcmk__find_action_config(rsc, task, interval_ms, true); /* If the given key is for one of the many notification pseudo-actions * (pre_notify_promote, etc.), the actual action name is "notify" */ if ((action->op_entry == NULL) && (strstr(key, "_notify_") != NULL)) { action->op_entry = find_exact_action_config(rsc, PCMK_ACTION_NOTIFY, 0, true); } unpack_operation(action, action->op_entry, interval_ms); } pe_rsc_trace(rsc, "Created %s action %d (%s): %s for %s on %s", (optional? "optional" : "required"), scheduler->action_id, key, task, ((rsc == NULL)? "no resource" : rsc->id), pe__node_name(node)); action->id = scheduler->action_id++; scheduler->actions = g_list_prepend(scheduler->actions, action); if (rsc == NULL) { add_singleton(scheduler, action); } else { rsc->actions = g_list_prepend(rsc->actions, action); } return action; } /*! * \internal * \brief Unpack a resource's action-specific instance parameters * * \param[in] action_xml XML of action's configuration in CIB (if any) * \param[in,out] node_attrs Table of node attributes (for rule evaluation) * \param[in,out] scheduler Cluster working set (for rule evaluation) * * \return Newly allocated hash table of action-specific instance parameters */ GHashTable * pcmk__unpack_action_rsc_params(const xmlNode *action_xml, GHashTable *node_attrs, pcmk_scheduler_t *scheduler) { GHashTable *params = pcmk__strkey_table(free, free); pe_rule_eval_data_t rule_data = { .node_hash = node_attrs, .role = pcmk_role_unknown, .now = scheduler->now, .match_data = NULL, .rsc_data = NULL, .op_data = NULL }; pe__unpack_dataset_nvpairs(action_xml, XML_TAG_ATTR_SETS, &rule_data, params, NULL, FALSE, scheduler); return params; } /*! * \internal * \brief Update an action's optional flag * * \param[in,out] action Action to update * \param[in] optional Requested optional status */ static void update_action_optional(pcmk_action_t *action, gboolean optional) { // Force a non-recurring action to be optional if its resource is unmanaged if ((action->rsc != NULL) && (action->node != NULL) && !pcmk_is_set(action->flags, pcmk_action_pseudo) && !pcmk_is_set(action->rsc->flags, pcmk_rsc_managed) && (g_hash_table_lookup(action->meta, XML_LRM_ATTR_INTERVAL_MS) == NULL)) { - pe_rsc_debug(action->rsc, "%s on %s is optional (%s is unmanaged)", - action->uuid, pe__node_name(action->node), - action->rsc->id); + pcmk__rsc_debug(action->rsc, + "%s on %s is optional (%s is unmanaged)", + action->uuid, pe__node_name(action->node), + action->rsc->id); pe__set_action_flags(action, pcmk_action_optional); // We shouldn't clear runnable here because ... something // Otherwise require the action if requested } else if (!optional) { pe__clear_action_flags(action, pcmk_action_optional); } } static enum pe_quorum_policy effective_quorum_policy(pcmk_resource_t *rsc, pcmk_scheduler_t *scheduler) { enum pe_quorum_policy policy = scheduler->no_quorum_policy; if (pcmk_is_set(scheduler->flags, pcmk_sched_quorate)) { policy = pcmk_no_quorum_ignore; } else if (scheduler->no_quorum_policy == pcmk_no_quorum_demote) { switch (rsc->role) { case pcmk_role_promoted: case pcmk_role_unpromoted: if (rsc->next_role > pcmk_role_unpromoted) { pe__set_next_role(rsc, pcmk_role_unpromoted, "no-quorum-policy=demote"); } policy = pcmk_no_quorum_ignore; break; default: policy = pcmk_no_quorum_stop; break; } } return policy; } /*! * \internal * \brief Update a resource action's runnable flag * * \param[in,out] action Action to update * \param[in,out] scheduler Scheduler data * * \note This may also schedule fencing if a stop is unrunnable. */ static void update_resource_action_runnable(pcmk_action_t *action, pcmk_scheduler_t *scheduler) { if (pcmk_is_set(action->flags, pcmk_action_pseudo)) { return; } if (action->node == NULL) { pe_rsc_trace(action->rsc, "%s is unrunnable (unallocated)", action->uuid); pe__clear_action_flags(action, pcmk_action_runnable); } else if (!pcmk_is_set(action->flags, pcmk_action_on_dc) && !(action->node->details->online) && (!pe__is_guest_node(action->node) || action->node->details->remote_requires_reset)) { pe__clear_action_flags(action, pcmk_action_runnable); do_crm_log(LOG_WARNING, "%s on %s is unrunnable (node is offline)", action->uuid, pe__node_name(action->node)); if (pcmk_is_set(action->rsc->flags, pcmk_rsc_managed) && pcmk__str_eq(action->task, PCMK_ACTION_STOP, pcmk__str_casei) && !(action->node->details->unclean)) { pe_fence_node(scheduler, action->node, "stop is unrunnable", false); } } else if (!pcmk_is_set(action->flags, pcmk_action_on_dc) && action->node->details->pending) { pe__clear_action_flags(action, pcmk_action_runnable); do_crm_log(LOG_WARNING, "Action %s on %s is unrunnable (node is pending)", action->uuid, pe__node_name(action->node)); } else if (action->needs == pcmk_requires_nothing) { pe_action_set_reason(action, NULL, TRUE); if (pe__is_guest_node(action->node) && !pe_can_fence(scheduler, action->node)) { /* An action that requires nothing usually does not require any * fencing in order to be runnable. However, there is an exception: * such an action cannot be completed if it is on a guest node whose * host is unclean and cannot be fenced. */ - pe_rsc_debug(action->rsc, "%s on %s is unrunnable " - "(node's host cannot be fenced)", - action->uuid, pe__node_name(action->node)); + pcmk__rsc_debug(action->rsc, + "%s on %s is unrunnable " + "(node's host cannot be fenced)", + action->uuid, pe__node_name(action->node)); pe__clear_action_flags(action, pcmk_action_runnable); } else { pe_rsc_trace(action->rsc, "%s on %s does not require fencing or quorum", action->uuid, pe__node_name(action->node)); pe__set_action_flags(action, pcmk_action_runnable); } } else { switch (effective_quorum_policy(action->rsc, scheduler)) { case pcmk_no_quorum_stop: - pe_rsc_debug(action->rsc, "%s on %s is unrunnable (no quorum)", - action->uuid, pe__node_name(action->node)); + pcmk__rsc_debug(action->rsc, + "%s on %s is unrunnable (no quorum)", + action->uuid, pe__node_name(action->node)); pe__clear_action_flags(action, pcmk_action_runnable); pe_action_set_reason(action, "no quorum", true); break; case pcmk_no_quorum_freeze: if (!action->rsc->fns->active(action->rsc, TRUE) || (action->rsc->next_role > action->rsc->role)) { - pe_rsc_debug(action->rsc, - "%s on %s is unrunnable (no quorum)", - action->uuid, pe__node_name(action->node)); + pcmk__rsc_debug(action->rsc, + "%s on %s is unrunnable (no quorum)", + action->uuid, pe__node_name(action->node)); pe__clear_action_flags(action, pcmk_action_runnable); pe_action_set_reason(action, "quorum freeze", true); } break; default: //pe_action_set_reason(action, NULL, TRUE); pe__set_action_flags(action, pcmk_action_runnable); break; } } } /*! * \internal * \brief Update a resource object's flags for a new action on it * * \param[in,out] rsc Resource that action is for (if any) * \param[in] action New action */ static void update_resource_flags_for_action(pcmk_resource_t *rsc, const pcmk_action_t *action) { /* @COMPAT pcmk_rsc_starting and pcmk_rsc_stopping are deprecated and unused * within Pacemaker, and will eventually be removed */ if (pcmk__str_eq(action->task, PCMK_ACTION_STOP, pcmk__str_casei)) { pe__set_resource_flags(rsc, pcmk_rsc_stopping); } else if (pcmk__str_eq(action->task, PCMK_ACTION_START, pcmk__str_casei)) { if (pcmk_is_set(action->flags, pcmk_action_runnable)) { pe__set_resource_flags(rsc, pcmk_rsc_starting); } else { pe__clear_resource_flags(rsc, pcmk_rsc_starting); } } } static bool valid_stop_on_fail(const char *value) { return !pcmk__strcase_any_of(value, "standby", "demote", "stop", NULL); } /*! * \internal * \brief Validate (and possibly reset) resource action's on_fail meta-attribute * * \param[in] rsc Resource that action is for * \param[in] action_name Action name * \param[in] action_config Action configuration XML from CIB (if any) * \param[in,out] meta Table of action meta-attributes */ static void validate_on_fail(const pcmk_resource_t *rsc, const char *action_name, const xmlNode *action_config, GHashTable *meta) { const char *name = NULL; const char *role = NULL; const char *interval_spec = NULL; const char *value = g_hash_table_lookup(meta, XML_OP_ATTR_ON_FAIL); char *key = NULL; char *new_value = NULL; // Stop actions can only use certain on-fail values if (pcmk__str_eq(action_name, PCMK_ACTION_STOP, pcmk__str_none) && !valid_stop_on_fail(value)) { pcmk__config_err("Resetting '" XML_OP_ATTR_ON_FAIL "' for %s stop " "action to default value because '%s' is not " "allowed for stop", rsc->id, value); g_hash_table_remove(meta, XML_OP_ATTR_ON_FAIL); return; } /* Demote actions default on-fail to the on-fail value for the first * recurring monitor for the promoted role (if any). */ if (pcmk__str_eq(action_name, PCMK_ACTION_DEMOTE, pcmk__str_none) && (value == NULL)) { /* @TODO This does not consider promote options set in a meta-attribute * block (which may have rules that need to be evaluated) rather than * XML properties. */ for (xmlNode *operation = first_named_child(rsc->ops_xml, XML_ATTR_OP); operation != NULL; operation = crm_next_same_xml(operation)) { bool enabled = false; const char *promote_on_fail = NULL; /* We only care about explicit on-fail (if promote uses default, so * can demote) */ promote_on_fail = crm_element_value(operation, XML_OP_ATTR_ON_FAIL); if (promote_on_fail == NULL) { continue; } // We only care about recurring monitors for the promoted role name = crm_element_value(operation, "name"); role = crm_element_value(operation, "role"); if (!pcmk__str_eq(name, PCMK_ACTION_MONITOR, pcmk__str_none) || !pcmk__strcase_any_of(role, PCMK__ROLE_PROMOTED, PCMK__ROLE_PROMOTED_LEGACY, NULL)) { continue; } interval_spec = crm_element_value(operation, XML_LRM_ATTR_INTERVAL); if (crm_parse_interval_spec(interval_spec) == 0) { continue; } // We only care about enabled monitors if ((pcmk__xe_get_bool_attr(operation, "enabled", &enabled) == pcmk_rc_ok) && !enabled) { continue; } // Demote actions can't default to on-fail="demote" if (pcmk__str_eq(promote_on_fail, "demote", pcmk__str_casei)) { continue; } // Use value from first applicable promote action found key = strdup(XML_OP_ATTR_ON_FAIL); new_value = strdup(promote_on_fail); CRM_ASSERT((key != NULL) && (new_value != NULL)); g_hash_table_insert(meta, key, new_value); } return; } if (pcmk__str_eq(action_name, PCMK_ACTION_LRM_DELETE, pcmk__str_none) && !pcmk__str_eq(value, "ignore", pcmk__str_casei)) { key = strdup(XML_OP_ATTR_ON_FAIL); new_value = strdup("ignore"); CRM_ASSERT((key != NULL) && (new_value != NULL)); g_hash_table_insert(meta, key, new_value); return; } // on-fail="demote" is allowed only for certain actions if (pcmk__str_eq(value, "demote", pcmk__str_casei)) { name = crm_element_value(action_config, "name"); role = crm_element_value(action_config, "role"); interval_spec = crm_element_value(action_config, XML_LRM_ATTR_INTERVAL); if (!pcmk__str_eq(name, PCMK_ACTION_PROMOTE, pcmk__str_none) && (!pcmk__str_eq(name, PCMK_ACTION_MONITOR, pcmk__str_none) || !pcmk__strcase_any_of(role, PCMK__ROLE_PROMOTED, PCMK__ROLE_PROMOTED_LEGACY, NULL) || (crm_parse_interval_spec(interval_spec) == 0))) { pcmk__config_err("Resetting '" XML_OP_ATTR_ON_FAIL "' for %s %s " "action to default value because 'demote' is not " "allowed for it", rsc->id, name); g_hash_table_remove(meta, XML_OP_ATTR_ON_FAIL); return; } } } static int unpack_timeout(const char *value) { int timeout_ms = crm_get_msec(value); if (timeout_ms < 0) { timeout_ms = PCMK_DEFAULT_ACTION_TIMEOUT_MS; } return timeout_ms; } // true if value contains valid, non-NULL interval origin for recurring op static bool unpack_interval_origin(const char *value, const xmlNode *xml_obj, guint interval_ms, const crm_time_t *now, long long *start_delay) { long long result = 0; guint interval_sec = interval_ms / 1000; crm_time_t *origin = NULL; // Ignore unspecified values and non-recurring operations if ((value == NULL) || (interval_ms == 0) || (now == NULL)) { return false; } // Parse interval origin from text origin = crm_time_new(value); if (origin == NULL) { pcmk__config_err("Ignoring '" XML_OP_ATTR_ORIGIN "' for operation " "'%s' because '%s' is not valid", (ID(xml_obj)? ID(xml_obj) : "(missing ID)"), value); return false; } // Get seconds since origin (negative if origin is in the future) result = crm_time_get_seconds(now) - crm_time_get_seconds(origin); crm_time_free(origin); // Calculate seconds from closest interval to now result = result % interval_sec; // Calculate seconds remaining until next interval result = ((result <= 0)? 0 : interval_sec) - result; crm_info("Calculated a start delay of %llds for operation '%s'", result, (ID(xml_obj)? ID(xml_obj) : "(unspecified)")); if (start_delay != NULL) { *start_delay = result * 1000; // milliseconds } return true; } static int unpack_start_delay(const char *value, GHashTable *meta) { int start_delay = 0; if (value != NULL) { start_delay = crm_get_msec(value); if (start_delay < 0) { start_delay = 0; } if (meta) { g_hash_table_replace(meta, strdup(XML_OP_ATTR_START_DELAY), pcmk__itoa(start_delay)); } } return start_delay; } /*! * \internal * \brief Find a resource's most frequent recurring monitor * * \param[in] rsc Resource to check * * \return Operation XML configured for most frequent recurring monitor for * \p rsc (if any) */ static xmlNode * most_frequent_monitor(const pcmk_resource_t *rsc) { guint min_interval_ms = G_MAXUINT; xmlNode *op = NULL; for (xmlNode *operation = first_named_child(rsc->ops_xml, XML_ATTR_OP); operation != NULL; operation = crm_next_same_xml(operation)) { bool enabled = false; guint interval_ms = 0; const char *interval_spec = crm_element_value(operation, XML_LRM_ATTR_INTERVAL); // We only care about enabled recurring monitors if (!pcmk__str_eq(crm_element_value(operation, "name"), PCMK_ACTION_MONITOR, pcmk__str_none)) { continue; } interval_ms = crm_parse_interval_spec(interval_spec); if (interval_ms == 0) { continue; } // @TODO This does not account for rules, defaults, etc. if ((pcmk__xe_get_bool_attr(operation, "enabled", &enabled) == pcmk_rc_ok) && !enabled) { continue; } if (interval_ms < min_interval_ms) { min_interval_ms = interval_ms; op = operation; } } return op; } /*! * \internal * \brief Unpack action meta-attributes * * \param[in,out] rsc Resource that action is for * \param[in] node Node that action is on * \param[in] action_name Action name * \param[in] interval_ms Action interval (in milliseconds) * \param[in] action_config Action XML configuration from CIB (if any) * * Unpack a resource action's meta-attributes (normalizing the interval, * timeout, and start delay values as integer milliseconds) from its CIB XML * configuration (including defaults). * * \return Newly allocated hash table with normalized action meta-attributes */ GHashTable * pcmk__unpack_action_meta(pcmk_resource_t *rsc, const pcmk_node_t *node, const char *action_name, guint interval_ms, const xmlNode *action_config) { GHashTable *meta = NULL; char *name = NULL; char *value = NULL; const char *timeout_spec = NULL; const char *str = NULL; pe_rsc_eval_data_t rsc_rule_data = { .standard = crm_element_value(rsc->xml, XML_AGENT_ATTR_CLASS), .provider = crm_element_value(rsc->xml, XML_AGENT_ATTR_PROVIDER), .agent = crm_element_value(rsc->xml, XML_EXPR_ATTR_TYPE), }; pe_op_eval_data_t op_rule_data = { .op_name = action_name, .interval = interval_ms, }; pe_rule_eval_data_t rule_data = { .node_hash = (node == NULL)? NULL : node->details->attrs, .role = pcmk_role_unknown, .now = rsc->cluster->now, .match_data = NULL, .rsc_data = &rsc_rule_data, .op_data = &op_rule_data, }; meta = pcmk__strkey_table(free, free); // Cluster-wide pe__unpack_dataset_nvpairs(rsc->cluster->op_defaults, XML_TAG_META_SETS, &rule_data, meta, NULL, FALSE, rsc->cluster); // Derive default timeout for probes from recurring monitor timeouts if (pcmk_is_probe(action_name, interval_ms)) { xmlNode *min_interval_mon = most_frequent_monitor(rsc); if (min_interval_mon != NULL) { /* @TODO This does not consider timeouts set in meta_attributes * blocks (which may also have rules that need to be evaluated). */ timeout_spec = crm_element_value(min_interval_mon, XML_ATTR_TIMEOUT); if (timeout_spec != NULL) { pe_rsc_trace(rsc, "Setting default timeout for %s probe to " "most frequent monitor's timeout '%s'", rsc->id, timeout_spec); name = strdup(XML_ATTR_TIMEOUT); value = strdup(timeout_spec); CRM_ASSERT((name != NULL) && (value != NULL)); g_hash_table_insert(meta, name, value); } } } if (action_config != NULL) { // take precedence over defaults pe__unpack_dataset_nvpairs(action_config, XML_TAG_META_SETS, &rule_data, meta, NULL, TRUE, rsc->cluster); /* Anything set as an XML property has highest precedence. * This ensures we use the name and interval from the tag. * (See below for the only exception, fence device start/probe timeout.) */ for (xmlAttrPtr attr = action_config->properties; attr != NULL; attr = attr->next) { name = strdup((const char *) attr->name); value = strdup(pcmk__xml_attr_value(attr)); CRM_ASSERT((name != NULL) && (value != NULL)); g_hash_table_insert(meta, name, value); } } g_hash_table_remove(meta, XML_ATTR_ID); // Normalize interval to milliseconds if (interval_ms > 0) { name = strdup(XML_LRM_ATTR_INTERVAL); CRM_ASSERT(name != NULL); value = crm_strdup_printf("%u", interval_ms); g_hash_table_insert(meta, name, value); } else { g_hash_table_remove(meta, XML_LRM_ATTR_INTERVAL); } /* Timeout order of precedence (highest to lowest): * 1. pcmk_monitor_timeout resource parameter (only for starts and probes * when rsc has pcmk_ra_cap_fence_params; this gets used for recurring * monitors via the executor instead) * 2. timeout configured in (with taking precedence over * ) * 3. timeout configured in * 4. PCMK_DEFAULT_ACTION_TIMEOUT_MS */ // Check for pcmk_monitor_timeout if (pcmk_is_set(pcmk_get_ra_caps(rsc_rule_data.standard), pcmk_ra_cap_fence_params) && (pcmk__str_eq(action_name, PCMK_ACTION_START, pcmk__str_none) || pcmk_is_probe(action_name, interval_ms))) { GHashTable *params = pe_rsc_params(rsc, node, rsc->cluster); timeout_spec = g_hash_table_lookup(params, "pcmk_monitor_timeout"); if (timeout_spec != NULL) { pe_rsc_trace(rsc, "Setting timeout for %s %s to " "pcmk_monitor_timeout (%s)", rsc->id, action_name, timeout_spec); name = strdup(XML_ATTR_TIMEOUT); value = strdup(timeout_spec); CRM_ASSERT((name != NULL) && (value != NULL)); g_hash_table_insert(meta, name, value); } } // Normalize timeout to positive milliseconds name = strdup(XML_ATTR_TIMEOUT); CRM_ASSERT(name != NULL); timeout_spec = g_hash_table_lookup(meta, XML_ATTR_TIMEOUT); g_hash_table_insert(meta, name, pcmk__itoa(unpack_timeout(timeout_spec))); // Ensure on-fail has a valid value validate_on_fail(rsc, action_name, action_config, meta); // Normalize start-delay str = g_hash_table_lookup(meta, XML_OP_ATTR_START_DELAY); if (str != NULL) { unpack_start_delay(str, meta); } else { long long start_delay = 0; str = g_hash_table_lookup(meta, XML_OP_ATTR_ORIGIN); if (unpack_interval_origin(str, action_config, interval_ms, rsc->cluster->now, &start_delay)) { name = strdup(XML_OP_ATTR_START_DELAY); CRM_ASSERT(name != NULL); g_hash_table_insert(meta, name, crm_strdup_printf("%lld", start_delay)); } } return meta; } /*! * \internal * \brief Determine an action's quorum and fencing dependency * * \param[in] rsc Resource that action is for * \param[in] action_name Name of action being unpacked * * \return Quorum and fencing dependency appropriate to action */ enum rsc_start_requirement pcmk__action_requires(const pcmk_resource_t *rsc, const char *action_name) { const char *value = NULL; enum rsc_start_requirement requires = pcmk_requires_nothing; CRM_CHECK((rsc != NULL) && (action_name != NULL), return requires); if (!pcmk__strcase_any_of(action_name, PCMK_ACTION_START, PCMK_ACTION_PROMOTE, NULL)) { value = "nothing (not start or promote)"; } else if (pcmk_is_set(rsc->flags, pcmk_rsc_needs_fencing)) { requires = pcmk_requires_fencing; value = "fencing"; } else if (pcmk_is_set(rsc->flags, pcmk_rsc_needs_quorum)) { requires = pcmk_requires_quorum; value = "quorum"; } else { value = "nothing"; } pe_rsc_trace(rsc, "%s of %s requires %s", action_name, rsc->id, value); return requires; } /*! * \internal * \brief Parse action failure response from a user-provided string * * \param[in] rsc Resource that action is for * \param[in] action_name Name of action * \param[in] interval_ms Action interval (in milliseconds) * \param[in] value User-provided configuration value for on-fail * * \return Action failure response parsed from \p text */ enum action_fail_response pcmk__parse_on_fail(const pcmk_resource_t *rsc, const char *action_name, guint interval_ms, const char *value) { const char *desc = NULL; bool needs_remote_reset = false; enum action_fail_response on_fail = pcmk_on_fail_ignore; if (value == NULL) { // Use default } else if (pcmk__str_eq(value, "block", pcmk__str_casei)) { on_fail = pcmk_on_fail_block; desc = "block"; } else if (pcmk__str_eq(value, "fence", pcmk__str_casei)) { if (pcmk_is_set(rsc->cluster->flags, pcmk_sched_fencing_enabled)) { on_fail = pcmk_on_fail_fence_node; desc = "node fencing"; } else { pcmk__config_err("Resetting '" XML_OP_ATTR_ON_FAIL "' for " "%s of %s to 'stop' because 'fence' is not " "valid when fencing is disabled", action_name, rsc->id); on_fail = pcmk_on_fail_stop; desc = "stop resource"; } } else if (pcmk__str_eq(value, "standby", pcmk__str_casei)) { on_fail = pcmk_on_fail_standby_node; desc = "node standby"; } else if (pcmk__strcase_any_of(value, "ignore", PCMK__VALUE_NOTHING, NULL)) { desc = "ignore"; } else if (pcmk__str_eq(value, "migrate", pcmk__str_casei)) { on_fail = pcmk_on_fail_ban; desc = "force migration"; } else if (pcmk__str_eq(value, "stop", pcmk__str_casei)) { on_fail = pcmk_on_fail_stop; desc = "stop resource"; } else if (pcmk__str_eq(value, "restart", pcmk__str_casei)) { on_fail = pcmk_on_fail_restart; desc = "restart (and possibly migrate)"; } else if (pcmk__str_eq(value, "restart-container", pcmk__str_casei)) { if (rsc->container == NULL) { - pe_rsc_debug(rsc, - "Using default " XML_OP_ATTR_ON_FAIL - " for %s of %s because it does not have a container", - action_name, rsc->id); + pcmk__rsc_debug(rsc, + "Using default " XML_OP_ATTR_ON_FAIL " for %s " + "of %s because it does not have a container", + action_name, rsc->id); } else { on_fail = pcmk_on_fail_restart_container; desc = "restart container (and possibly migrate)"; } } else if (pcmk__str_eq(value, "demote", pcmk__str_casei)) { on_fail = pcmk_on_fail_demote; desc = "demote instance"; } else { pcmk__config_err("Using default '" XML_OP_ATTR_ON_FAIL "' for " "%s of %s because '%s' is not valid", action_name, rsc->id, value); } /* Remote node connections are handled specially. Failures that result * in dropping an active connection must result in fencing. The only * failures that don't are probes and starts. The user can explicitly set * on-fail="fence" to fence after start failures. */ if (pe__resource_is_remote_conn(rsc) && !pcmk_is_probe(action_name, interval_ms) && !pcmk__str_eq(action_name, PCMK_ACTION_START, pcmk__str_none)) { needs_remote_reset = true; if (!pcmk_is_set(rsc->flags, pcmk_rsc_managed)) { desc = NULL; // Force default for unmanaged connections } } if (desc != NULL) { // Explicit value used, default not needed } else if (rsc->container != NULL) { on_fail = pcmk_on_fail_restart_container; desc = "restart container (and possibly migrate) (default)"; } else if (needs_remote_reset) { if (pcmk_is_set(rsc->flags, pcmk_rsc_managed)) { if (pcmk_is_set(rsc->cluster->flags, pcmk_sched_fencing_enabled)) { desc = "fence remote node (default)"; } else { desc = "recover remote node connection (default)"; } on_fail = pcmk_on_fail_reset_remote; } else { on_fail = pcmk_on_fail_stop; desc = "stop unmanaged remote node (enforcing default)"; } } else if (pcmk__str_eq(action_name, PCMK_ACTION_STOP, pcmk__str_none)) { if (pcmk_is_set(rsc->cluster->flags, pcmk_sched_fencing_enabled)) { on_fail = pcmk_on_fail_fence_node; desc = "resource fence (default)"; } else { on_fail = pcmk_on_fail_block; desc = "resource block (default)"; } } else { on_fail = pcmk_on_fail_restart; desc = "restart (and possibly migrate) (default)"; } pe_rsc_trace(rsc, "Failure handling for %s-interval %s of %s: %s", pcmk__readable_interval(interval_ms), action_name, rsc->id, desc); return on_fail; } /*! * \internal * \brief Determine a resource's role after failure of an action * * \param[in] rsc Resource that action is for * \param[in] action_name Action name * \param[in] on_fail Failure handling for action * \param[in] meta Unpacked action meta-attributes * * \return Resource role that results from failure of action */ enum rsc_role_e pcmk__role_after_failure(const pcmk_resource_t *rsc, const char *action_name, enum action_fail_response on_fail, GHashTable *meta) { const char *value = NULL; enum rsc_role_e role = pcmk_role_unknown; // Set default for role after failure specially in certain circumstances switch (on_fail) { case pcmk_on_fail_stop: role = pcmk_role_stopped; break; case pcmk_on_fail_reset_remote: if (rsc->remote_reconnect_ms != 0) { role = pcmk_role_stopped; } break; default: break; } // @COMPAT Check for explicitly configured role (deprecated) value = g_hash_table_lookup(meta, "role_after_failure"); if (value != NULL) { pe_warn_once(pcmk__wo_role_after, "Support for role_after_failure is deprecated " "and will be removed in a future release"); if (role == pcmk_role_unknown) { role = text2role(value); } } if (role == pcmk_role_unknown) { // Use default if (pcmk__str_eq(action_name, PCMK_ACTION_PROMOTE, pcmk__str_none)) { role = pcmk_role_unpromoted; } else { role = pcmk_role_started; } } pe_rsc_trace(rsc, "Role after %s %s failure is: %s", rsc->id, action_name, role2text(role)); return role; } /*! * \internal * \brief Unpack action configuration * * Unpack a resource action's meta-attributes (normalizing the interval, * timeout, and start delay values as integer milliseconds), requirements, and * failure policy from its CIB XML configuration (including defaults). * * \param[in,out] action Resource action to unpack into * \param[in] xml_obj Action configuration XML (NULL for defaults only) * \param[in] interval_ms How frequently to perform the operation */ static void unpack_operation(pcmk_action_t *action, const xmlNode *xml_obj, guint interval_ms) { const char *value = NULL; action->meta = pcmk__unpack_action_meta(action->rsc, action->node, action->task, interval_ms, xml_obj); action->needs = pcmk__action_requires(action->rsc, action->task); value = g_hash_table_lookup(action->meta, XML_OP_ATTR_ON_FAIL); action->on_fail = pcmk__parse_on_fail(action->rsc, action->task, interval_ms, value); action->fail_role = pcmk__role_after_failure(action->rsc, action->task, action->on_fail, action->meta); } /*! * \brief Create or update an action object * * \param[in,out] rsc Resource that action is for (if any) * \param[in,out] key Action key (must be non-NULL) * \param[in] task Action name (must be non-NULL) * \param[in] on_node Node that action is on (if any) * \param[in] optional Whether action should be considered optional * \param[in,out] scheduler Scheduler data * * \return Action object corresponding to arguments (guaranteed not to be * \c NULL) * \note This function takes ownership of (and might free) \p key, and * \p scheduler takes ownership of the returned action (the caller should * not free it). */ pcmk_action_t * custom_action(pcmk_resource_t *rsc, char *key, const char *task, const pcmk_node_t *on_node, gboolean optional, pcmk_scheduler_t *scheduler) { pcmk_action_t *action = NULL; CRM_ASSERT((key != NULL) && (task != NULL) && (scheduler != NULL)); action = find_existing_action(key, rsc, on_node, scheduler); if (action == NULL) { action = new_action(key, task, rsc, on_node, optional, scheduler); } else { free(key); } update_action_optional(action, optional); if (rsc != NULL) { if ((action->node != NULL) && (action->op_entry != NULL) && !pcmk_is_set(action->flags, pcmk_action_attrs_evaluated)) { GHashTable *attrs = action->node->details->attrs; if (action->extra != NULL) { g_hash_table_destroy(action->extra); } action->extra = pcmk__unpack_action_rsc_params(action->op_entry, attrs, scheduler); pe__set_action_flags(action, pcmk_action_attrs_evaluated); } update_resource_action_runnable(action, scheduler); update_resource_flags_for_action(rsc, action); } if (action->extra == NULL) { action->extra = pcmk__strkey_table(free, free); } return action; } pcmk_action_t * get_pseudo_op(const char *name, pcmk_scheduler_t *scheduler) { pcmk_action_t *op = lookup_singleton(scheduler, name); if (op == NULL) { op = custom_action(NULL, strdup(name), name, NULL, TRUE, scheduler); pe__set_action_flags(op, pcmk_action_pseudo|pcmk_action_runnable); } return op; } static GList * find_unfencing_devices(GList *candidates, GList *matches) { for (GList *gIter = candidates; gIter != NULL; gIter = gIter->next) { pcmk_resource_t *candidate = gIter->data; if (candidate->children != NULL) { matches = find_unfencing_devices(candidate->children, matches); } else if (!pcmk_is_set(candidate->flags, pcmk_rsc_fence_device)) { continue; } else if (pcmk_is_set(candidate->flags, pcmk_rsc_needs_unfencing)) { matches = g_list_prepend(matches, candidate); } else if (pcmk__str_eq(g_hash_table_lookup(candidate->meta, PCMK_STONITH_PROVIDES), PCMK__VALUE_UNFENCING, pcmk__str_casei)) { matches = g_list_prepend(matches, candidate); } } return matches; } static int node_priority_fencing_delay(const pcmk_node_t *node, const pcmk_scheduler_t *scheduler) { int member_count = 0; int online_count = 0; int top_priority = 0; int lowest_priority = 0; GList *gIter = NULL; // `priority-fencing-delay` is disabled if (scheduler->priority_fencing_delay <= 0) { return 0; } /* No need to request a delay if the fencing target is not a normal cluster * member, for example if it's a remote node or a guest node. */ if (node->details->type != pcmk_node_variant_cluster) { return 0; } // No need to request a delay if the fencing target is in our partition if (node->details->online) { return 0; } for (gIter = scheduler->nodes; gIter != NULL; gIter = gIter->next) { pcmk_node_t *n = gIter->data; if (n->details->type != pcmk_node_variant_cluster) { continue; } member_count ++; if (n->details->online) { online_count++; } if (member_count == 1 || n->details->priority > top_priority) { top_priority = n->details->priority; } if (member_count == 1 || n->details->priority < lowest_priority) { lowest_priority = n->details->priority; } } // No need to delay if we have more than half of the cluster members if (online_count > member_count / 2) { return 0; } /* All the nodes have equal priority. * Any configured corresponding `pcmk_delay_base/max` will be applied. */ if (lowest_priority == top_priority) { return 0; } if (node->details->priority < top_priority) { return 0; } return scheduler->priority_fencing_delay; } pcmk_action_t * pe_fence_op(pcmk_node_t *node, const char *op, bool optional, const char *reason, bool priority_delay, pcmk_scheduler_t *scheduler) { char *op_key = NULL; pcmk_action_t *stonith_op = NULL; if(op == NULL) { op = scheduler->stonith_action; } op_key = crm_strdup_printf("%s-%s-%s", PCMK_ACTION_STONITH, node->details->uname, op); stonith_op = lookup_singleton(scheduler, op_key); if(stonith_op == NULL) { stonith_op = custom_action(NULL, op_key, PCMK_ACTION_STONITH, node, TRUE, scheduler); add_hash_param(stonith_op->meta, XML_LRM_ATTR_TARGET, node->details->uname); add_hash_param(stonith_op->meta, XML_LRM_ATTR_TARGET_UUID, node->details->id); add_hash_param(stonith_op->meta, "stonith_action", op); if (pcmk_is_set(scheduler->flags, pcmk_sched_enable_unfencing)) { /* Extra work to detect device changes */ GString *digests_all = g_string_sized_new(1024); GString *digests_secure = g_string_sized_new(1024); GList *matches = find_unfencing_devices(scheduler->resources, NULL); char *key = NULL; char *value = NULL; for (GList *gIter = matches; gIter != NULL; gIter = gIter->next) { pcmk_resource_t *match = gIter->data; const char *agent = g_hash_table_lookup(match->meta, XML_ATTR_TYPE); pcmk__op_digest_t *data = NULL; data = pe__compare_fencing_digest(match, agent, node, scheduler); if (data->rc == pcmk__digest_mismatch) { optional = FALSE; crm_notice("Unfencing node %s because the definition of " "%s changed", pe__node_name(node), match->id); if (!pcmk__is_daemon && scheduler->priv != NULL) { pcmk__output_t *out = scheduler->priv; out->info(out, "notice: Unfencing node %s because the " "definition of %s changed", pe__node_name(node), match->id); } } pcmk__g_strcat(digests_all, match->id, ":", agent, ":", data->digest_all_calc, ",", NULL); pcmk__g_strcat(digests_secure, match->id, ":", agent, ":", data->digest_secure_calc, ",", NULL); } key = strdup(XML_OP_ATTR_DIGESTS_ALL); value = strdup((const char *) digests_all->str); CRM_ASSERT((key != NULL) && (value != NULL)); g_hash_table_insert(stonith_op->meta, key, value); g_string_free(digests_all, TRUE); key = strdup(XML_OP_ATTR_DIGESTS_SECURE); value = strdup((const char *) digests_secure->str); CRM_ASSERT((key != NULL) && (value != NULL)); g_hash_table_insert(stonith_op->meta, key, value); g_string_free(digests_secure, TRUE); } } else { free(op_key); } if (scheduler->priority_fencing_delay > 0 /* It's a suitable case where `priority-fencing-delay` applies. * At least add `priority-fencing-delay` field as an indicator. */ && (priority_delay /* The priority delay needs to be recalculated if this function has * been called by schedule_fencing_and_shutdowns() after node * priority has already been calculated by native_add_running(). */ || g_hash_table_lookup(stonith_op->meta, XML_CONFIG_ATTR_PRIORITY_FENCING_DELAY) != NULL)) { /* Add `priority-fencing-delay` to the fencing op even if it's 0 for * the targeting node. So that it takes precedence over any possible * `pcmk_delay_base/max`. */ char *delay_s = pcmk__itoa(node_priority_fencing_delay(node, scheduler)); g_hash_table_insert(stonith_op->meta, strdup(XML_CONFIG_ATTR_PRIORITY_FENCING_DELAY), delay_s); } if(optional == FALSE && pe_can_fence(scheduler, node)) { pe__clear_action_flags(stonith_op, pcmk_action_optional); pe_action_set_reason(stonith_op, reason, false); } else if(reason && stonith_op->reason == NULL) { stonith_op->reason = strdup(reason); } return stonith_op; } void pe_free_action(pcmk_action_t *action) { if (action == NULL) { return; } g_list_free_full(action->actions_before, free); g_list_free_full(action->actions_after, free); if (action->extra) { g_hash_table_destroy(action->extra); } if (action->meta) { g_hash_table_destroy(action->meta); } free(action->cancel_task); free(action->reason); free(action->task); free(action->uuid); free(action->node); free(action); } int pe_get_configured_timeout(pcmk_resource_t *rsc, const char *action, pcmk_scheduler_t *scheduler) { xmlNode *child = NULL; GHashTable *action_meta = NULL; const char *timeout_spec = NULL; int timeout_ms = 0; pe_rule_eval_data_t rule_data = { .node_hash = NULL, .role = pcmk_role_unknown, .now = scheduler->now, .match_data = NULL, .rsc_data = NULL, .op_data = NULL }; for (child = first_named_child(rsc->ops_xml, XML_ATTR_OP); child != NULL; child = crm_next_same_xml(child)) { if (pcmk__str_eq(action, crm_element_value(child, XML_NVPAIR_ATTR_NAME), pcmk__str_casei)) { timeout_spec = crm_element_value(child, XML_ATTR_TIMEOUT); break; } } if (timeout_spec == NULL && scheduler->op_defaults) { action_meta = pcmk__strkey_table(free, free); pe__unpack_dataset_nvpairs(scheduler->op_defaults, XML_TAG_META_SETS, &rule_data, action_meta, NULL, FALSE, scheduler); timeout_spec = g_hash_table_lookup(action_meta, XML_ATTR_TIMEOUT); } // @TODO check meta-attributes // @TODO maybe use min-interval monitor timeout as default for monitors timeout_ms = crm_get_msec(timeout_spec); if (timeout_ms < 0) { timeout_ms = PCMK_DEFAULT_ACTION_TIMEOUT_MS; } if (action_meta != NULL) { g_hash_table_destroy(action_meta); } return timeout_ms; } enum action_tasks get_complex_task(const pcmk_resource_t *rsc, const char *name) { enum action_tasks task = text2task(name); if ((rsc != NULL) && (rsc->variant == pcmk_rsc_variant_primitive)) { switch (task) { case pcmk_action_stopped: case pcmk_action_started: case pcmk_action_demoted: case pcmk_action_promoted: crm_trace("Folding %s back into its atomic counterpart for %s", name, rsc->id); --task; break; default: break; } } return task; } /*! * \internal * \brief Find first matching action in a list * * \param[in] input List of actions to search * \param[in] uuid If not NULL, action must have this UUID * \param[in] task If not NULL, action must have this action name * \param[in] on_node If not NULL, action must be on this node * * \return First action in list that matches criteria, or NULL if none */ pcmk_action_t * find_first_action(const GList *input, const char *uuid, const char *task, const pcmk_node_t *on_node) { CRM_CHECK(uuid || task, return NULL); for (const GList *gIter = input; gIter != NULL; gIter = gIter->next) { pcmk_action_t *action = (pcmk_action_t *) gIter->data; if (uuid != NULL && !pcmk__str_eq(uuid, action->uuid, pcmk__str_casei)) { continue; } else if (task != NULL && !pcmk__str_eq(task, action->task, pcmk__str_casei)) { continue; } else if (on_node == NULL) { return action; } else if (action->node == NULL) { continue; } else if (on_node->details == action->node->details) { return action; } } return NULL; } GList * find_actions(GList *input, const char *key, const pcmk_node_t *on_node) { GList *gIter = input; GList *result = NULL; CRM_CHECK(key != NULL, return NULL); for (; gIter != NULL; gIter = gIter->next) { pcmk_action_t *action = (pcmk_action_t *) gIter->data; if (!pcmk__str_eq(key, action->uuid, pcmk__str_casei)) { continue; } else if (on_node == NULL) { crm_trace("Action %s matches (ignoring node)", key); result = g_list_prepend(result, action); } else if (action->node == NULL) { crm_trace("Action %s matches (unallocated, assigning to %s)", key, pe__node_name(on_node)); action->node = pe__copy_node(on_node); result = g_list_prepend(result, action); } else if (on_node->details == action->node->details) { crm_trace("Action %s on %s matches", key, pe__node_name(on_node)); result = g_list_prepend(result, action); } } return result; } GList * find_actions_exact(GList *input, const char *key, const pcmk_node_t *on_node) { GList *result = NULL; CRM_CHECK(key != NULL, return NULL); if (on_node == NULL) { return NULL; } for (GList *gIter = input; gIter != NULL; gIter = gIter->next) { pcmk_action_t *action = (pcmk_action_t *) gIter->data; if ((action->node != NULL) && pcmk__str_eq(key, action->uuid, pcmk__str_casei) && pcmk__str_eq(on_node->details->id, action->node->details->id, pcmk__str_casei)) { crm_trace("Action %s on %s matches", key, pe__node_name(on_node)); result = g_list_prepend(result, action); } } return result; } /*! * \brief Find all actions of given type for a resource * * \param[in] rsc Resource to search * \param[in] node Find only actions scheduled on this node * \param[in] task Action name to search for * \param[in] require_node If TRUE, NULL node or action node will not match * * \return List of actions found (or NULL if none) * \note If node is not NULL and require_node is FALSE, matching actions * without a node will be assigned to node. */ GList * pe__resource_actions(const pcmk_resource_t *rsc, const pcmk_node_t *node, const char *task, bool require_node) { GList *result = NULL; char *key = pcmk__op_key(rsc->id, task, 0); if (require_node) { result = find_actions_exact(rsc->actions, key, node); } else { result = find_actions(rsc->actions, key, node); } free(key); return result; } /*! * \internal * \brief Create an action reason string based on the action itself * * \param[in] action Action to create reason string for * \param[in] flag Action flag that was cleared * * \return Newly allocated string suitable for use as action reason * \note It is the caller's responsibility to free() the result. */ char * pe__action2reason(const pcmk_action_t *action, enum pe_action_flags flag) { const char *change = NULL; switch (flag) { case pcmk_action_runnable: change = "unrunnable"; break; case pcmk_action_migratable: change = "unmigrateable"; break; case pcmk_action_optional: change = "required"; break; default: // Bug: caller passed unsupported flag CRM_CHECK(change != NULL, change = ""); break; } return crm_strdup_printf("%s%s%s %s", change, (action->rsc == NULL)? "" : " ", (action->rsc == NULL)? "" : action->rsc->id, action->task); } void pe_action_set_reason(pcmk_action_t *action, const char *reason, bool overwrite) { if (action->reason != NULL && overwrite) { pe_rsc_trace(action->rsc, "Changing %s reason from '%s' to '%s'", action->uuid, action->reason, pcmk__s(reason, "(none)")); } else if (action->reason == NULL) { pe_rsc_trace(action->rsc, "Set %s reason to '%s'", action->uuid, pcmk__s(reason, "(none)")); } else { // crm_assert(action->reason != NULL && !overwrite); return; } pcmk__str_update(&action->reason, reason); } /*! * \internal * \brief Create an action to clear a resource's history from CIB * * \param[in,out] rsc Resource to clear * \param[in] node Node to clear history on */ void pe__clear_resource_history(pcmk_resource_t *rsc, const pcmk_node_t *node) { CRM_ASSERT((rsc != NULL) && (node != NULL)); custom_action(rsc, pcmk__op_key(rsc->id, PCMK_ACTION_LRM_DELETE, 0), PCMK_ACTION_LRM_DELETE, node, FALSE, rsc->cluster); } #define sort_return(an_int, why) do { \ free(a_uuid); \ free(b_uuid); \ crm_trace("%s (%d) %c %s (%d) : %s", \ a_xml_id, a_call_id, an_int>0?'>':an_int<0?'<':'=', \ b_xml_id, b_call_id, why); \ return an_int; \ } while(0) int pe__is_newer_op(const xmlNode *xml_a, const xmlNode *xml_b, bool same_node_default) { int a_call_id = -1; int b_call_id = -1; char *a_uuid = NULL; char *b_uuid = NULL; const char *a_xml_id = crm_element_value(xml_a, XML_ATTR_ID); const char *b_xml_id = crm_element_value(xml_b, XML_ATTR_ID); const char *a_node = crm_element_value(xml_a, XML_LRM_ATTR_TARGET); const char *b_node = crm_element_value(xml_b, XML_LRM_ATTR_TARGET); bool same_node = true; /* @COMPAT The on_node attribute was added to last_failure as of 1.1.13 (via * 8b3ca1c) and the other entries as of 1.1.12 (via 0b07b5c). * * In case that any of the lrm_rsc_op entries doesn't have on_node * attribute, we need to explicitly tell whether the two operations are on * the same node. */ if (a_node == NULL || b_node == NULL) { same_node = same_node_default; } else { same_node = pcmk__str_eq(a_node, b_node, pcmk__str_casei); } if (same_node && pcmk__str_eq(a_xml_id, b_xml_id, pcmk__str_none)) { /* We have duplicate lrm_rsc_op entries in the status * section which is unlikely to be a good thing * - we can handle it easily enough, but we need to get * to the bottom of why it's happening. */ pe_err("Duplicate lrm_rsc_op entries named %s", a_xml_id); sort_return(0, "duplicate"); } crm_element_value_int(xml_a, XML_LRM_ATTR_CALLID, &a_call_id); crm_element_value_int(xml_b, XML_LRM_ATTR_CALLID, &b_call_id); if (a_call_id == -1 && b_call_id == -1) { /* both are pending ops so it doesn't matter since * stops are never pending */ sort_return(0, "pending"); } else if (same_node && a_call_id >= 0 && a_call_id < b_call_id) { sort_return(-1, "call id"); } else if (same_node && b_call_id >= 0 && a_call_id > b_call_id) { sort_return(1, "call id"); } else if (a_call_id >= 0 && b_call_id >= 0 && (!same_node || a_call_id == b_call_id)) { /* * The op and last_failed_op are the same * Order on last-rc-change */ time_t last_a = -1; time_t last_b = -1; crm_element_value_epoch(xml_a, XML_RSC_OP_LAST_CHANGE, &last_a); crm_element_value_epoch(xml_b, XML_RSC_OP_LAST_CHANGE, &last_b); crm_trace("rc-change: %lld vs %lld", (long long) last_a, (long long) last_b); if (last_a >= 0 && last_a < last_b) { sort_return(-1, "rc-change"); } else if (last_b >= 0 && last_a > last_b) { sort_return(1, "rc-change"); } sort_return(0, "rc-change"); } else { /* One of the inputs is a pending operation * Attempt to use XML_ATTR_TRANSITION_MAGIC to determine its age relative to the other */ int a_id = -1; int b_id = -1; const char *a_magic = crm_element_value(xml_a, XML_ATTR_TRANSITION_MAGIC); const char *b_magic = crm_element_value(xml_b, XML_ATTR_TRANSITION_MAGIC); CRM_CHECK(a_magic != NULL && b_magic != NULL, sort_return(0, "No magic")); if (!decode_transition_magic(a_magic, &a_uuid, &a_id, NULL, NULL, NULL, NULL)) { sort_return(0, "bad magic a"); } if (!decode_transition_magic(b_magic, &b_uuid, &b_id, NULL, NULL, NULL, NULL)) { sort_return(0, "bad magic b"); } /* try to determine the relative age of the operation... * some pending operations (e.g. a start) may have been superseded * by a subsequent stop * * [a|b]_id == -1 means it's a shutdown operation and _always_ comes last */ if (!pcmk__str_eq(a_uuid, b_uuid, pcmk__str_casei) || a_id == b_id) { /* * some of the logic in here may be redundant... * * if the UUID from the TE doesn't match then one better * be a pending operation. * pending operations don't survive between elections and joins * because we query the LRM directly */ if (b_call_id == -1) { sort_return(-1, "transition + call"); } else if (a_call_id == -1) { sort_return(1, "transition + call"); } } else if ((a_id >= 0 && a_id < b_id) || b_id == -1) { sort_return(-1, "transition"); } else if ((b_id >= 0 && a_id > b_id) || a_id == -1) { sort_return(1, "transition"); } } /* we should never end up here */ CRM_CHECK(FALSE, sort_return(0, "default")); } gint sort_op_by_callid(gconstpointer a, gconstpointer b) { const xmlNode *xml_a = a; const xmlNode *xml_b = b; return pe__is_newer_op(xml_a, xml_b, true); } /*! * \internal * \brief Create a new pseudo-action for a resource * * \param[in,out] rsc Resource to create action for * \param[in] task Action name * \param[in] optional Whether action should be considered optional * \param[in] runnable Whethe action should be considered runnable * * \return New action object corresponding to arguments */ pcmk_action_t * pe__new_rsc_pseudo_action(pcmk_resource_t *rsc, const char *task, bool optional, bool runnable) { pcmk_action_t *action = NULL; CRM_ASSERT((rsc != NULL) && (task != NULL)); action = custom_action(rsc, pcmk__op_key(rsc->id, task, 0), task, NULL, optional, rsc->cluster); pe__set_action_flags(action, pcmk_action_pseudo); if (runnable) { pe__set_action_flags(action, pcmk_action_runnable); } return action; } /*! * \internal * \brief Add the expected result to an action * * \param[in,out] action Action to add expected result to * \param[in] expected_result Expected result to add * * \note This is more efficient than calling add_hash_param(). */ void pe__add_action_expected_result(pcmk_action_t *action, int expected_result) { char *name = NULL; CRM_ASSERT((action != NULL) && (action->meta != NULL)); name = strdup(XML_ATTR_TE_TARGET_RC); CRM_ASSERT (name != NULL); g_hash_table_insert(action->meta, name, pcmk__itoa(expected_result)); } diff --git a/lib/pengine/unpack.c b/lib/pengine/unpack.c index ccc41e4f79..2a00cf106f 100644 --- a/lib/pengine/unpack.c +++ b/lib/pengine/unpack.c @@ -1,5082 +1,5082 @@ /* * Copyright 2004-2023 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU Lesser General Public License * version 2.1 or later (LGPLv2.1+) WITHOUT ANY WARRANTY. */ #include #include #include #include #include #include #include #include #include #include #include #include #include #include CRM_TRACE_INIT_DATA(pe_status); // A (parsed) resource action history entry struct action_history { pcmk_resource_t *rsc; // Resource that history is for pcmk_node_t *node; // Node that history is for xmlNode *xml; // History entry XML // Parsed from entry XML const char *id; // XML ID of history entry const char *key; // Operation key of action const char *task; // Action name const char *exit_reason; // Exit reason given for result guint interval_ms; // Action interval int call_id; // Call ID of action int expected_exit_status; // Expected exit status of action int exit_status; // Actual exit status of action int execution_status; // Execution status of action }; /* This uses pcmk__set_flags_as()/pcmk__clear_flags_as() directly rather than * use pe__set_working_set_flags()/pe__clear_working_set_flags() so that the * flag is stringified more readably in log messages. */ #define set_config_flag(scheduler, option, flag) do { \ const char *scf_value = pe_pref((scheduler)->config_hash, (option)); \ if (scf_value != NULL) { \ if (crm_is_true(scf_value)) { \ (scheduler)->flags = pcmk__set_flags_as(__func__, __LINE__, \ LOG_TRACE, "Scheduler", \ crm_system_name, (scheduler)->flags, \ (flag), #flag); \ } else { \ (scheduler)->flags = pcmk__clear_flags_as(__func__, __LINE__, \ LOG_TRACE, "Scheduler", \ crm_system_name, (scheduler)->flags, \ (flag), #flag); \ } \ } \ } while(0) static void unpack_rsc_op(pcmk_resource_t *rsc, pcmk_node_t *node, xmlNode *xml_op, xmlNode **last_failure, enum action_fail_response *failed); static void determine_remote_online_status(pcmk_scheduler_t *scheduler, pcmk_node_t *this_node); static void add_node_attrs(const xmlNode *xml_obj, pcmk_node_t *node, bool overwrite, pcmk_scheduler_t *scheduler); static void determine_online_status(const xmlNode *node_state, pcmk_node_t *this_node, pcmk_scheduler_t *scheduler); static void unpack_node_lrm(pcmk_node_t *node, const xmlNode *xml, pcmk_scheduler_t *scheduler); static gboolean is_dangling_guest_node(pcmk_node_t *node) { /* we are looking for a remote-node that was supposed to be mapped to a * container resource, but all traces of that container have disappeared * from both the config and the status section. */ if (pe__is_guest_or_remote_node(node) && node->details->remote_rsc && node->details->remote_rsc->container == NULL && pcmk_is_set(node->details->remote_rsc->flags, pcmk_rsc_removed_filler)) { return TRUE; } return FALSE; } /*! * \brief Schedule a fence action for a node * * \param[in,out] scheduler Scheduler data * \param[in,out] node Node to fence * \param[in] reason Text description of why fencing is needed * \param[in] priority_delay Whether to consider `priority-fencing-delay` */ void pe_fence_node(pcmk_scheduler_t *scheduler, pcmk_node_t *node, const char *reason, bool priority_delay) { CRM_CHECK(node, return); /* A guest node is fenced by marking its container as failed */ if (pe__is_guest_node(node)) { pcmk_resource_t *rsc = node->details->remote_rsc->container; if (!pcmk_is_set(rsc->flags, pcmk_rsc_failed)) { if (!pcmk_is_set(rsc->flags, pcmk_rsc_managed)) { crm_notice("Not fencing guest node %s " "(otherwise would because %s): " "its guest resource %s is unmanaged", pe__node_name(node), reason, rsc->id); } else { crm_warn("Guest node %s will be fenced " "(by recovering its guest resource %s): %s", pe__node_name(node), rsc->id, reason); /* We don't mark the node as unclean because that would prevent the * node from running resources. We want to allow it to run resources * in this transition if the recovery succeeds. */ node->details->remote_requires_reset = TRUE; pe__set_resource_flags(rsc, pcmk_rsc_failed|pcmk_rsc_stop_if_failed); } } } else if (is_dangling_guest_node(node)) { crm_info("Cleaning up dangling connection for guest node %s: " "fencing was already done because %s, " "and guest resource no longer exists", pe__node_name(node), reason); pe__set_resource_flags(node->details->remote_rsc, pcmk_rsc_failed|pcmk_rsc_stop_if_failed); } else if (pe__is_remote_node(node)) { pcmk_resource_t *rsc = node->details->remote_rsc; if ((rsc != NULL) && !pcmk_is_set(rsc->flags, pcmk_rsc_managed)) { crm_notice("Not fencing remote node %s " "(otherwise would because %s): connection is unmanaged", pe__node_name(node), reason); } else if(node->details->remote_requires_reset == FALSE) { node->details->remote_requires_reset = TRUE; crm_warn("Remote node %s %s: %s", pe__node_name(node), pe_can_fence(scheduler, node)? "will be fenced" : "is unclean", reason); } node->details->unclean = TRUE; // No need to apply `priority-fencing-delay` for remote nodes pe_fence_op(node, NULL, TRUE, reason, FALSE, scheduler); } else if (node->details->unclean) { crm_trace("Cluster node %s %s because %s", pe__node_name(node), pe_can_fence(scheduler, node)? "would also be fenced" : "also is unclean", reason); } else { crm_warn("Cluster node %s %s: %s", pe__node_name(node), pe_can_fence(scheduler, node)? "will be fenced" : "is unclean", reason); node->details->unclean = TRUE; pe_fence_op(node, NULL, TRUE, reason, priority_delay, scheduler); } } // @TODO xpaths can't handle templates, rules, or id-refs // nvpair with provides or requires set to unfencing #define XPATH_UNFENCING_NVPAIR XML_CIB_TAG_NVPAIR \ "[(@" XML_NVPAIR_ATTR_NAME "='" PCMK_STONITH_PROVIDES "'" \ "or @" XML_NVPAIR_ATTR_NAME "='" XML_RSC_ATTR_REQUIRES "') " \ "and @" XML_NVPAIR_ATTR_VALUE "='" PCMK__VALUE_UNFENCING "']" // unfencing in rsc_defaults or any resource #define XPATH_ENABLE_UNFENCING \ "/" XML_TAG_CIB "/" XML_CIB_TAG_CONFIGURATION "/" XML_CIB_TAG_RESOURCES \ "//" XML_TAG_META_SETS "/" XPATH_UNFENCING_NVPAIR \ "|/" XML_TAG_CIB "/" XML_CIB_TAG_CONFIGURATION "/" XML_CIB_TAG_RSCCONFIG \ "/" XML_TAG_META_SETS "/" XPATH_UNFENCING_NVPAIR static void set_if_xpath(uint64_t flag, const char *xpath, pcmk_scheduler_t *scheduler) { xmlXPathObjectPtr result = NULL; if (!pcmk_is_set(scheduler->flags, flag)) { result = xpath_search(scheduler->input, xpath); if (result && (numXpathResults(result) > 0)) { pe__set_working_set_flags(scheduler, flag); } freeXpathObject(result); } } gboolean unpack_config(xmlNode *config, pcmk_scheduler_t *scheduler) { const char *value = NULL; GHashTable *config_hash = pcmk__strkey_table(free, free); pe_rule_eval_data_t rule_data = { .node_hash = NULL, .role = pcmk_role_unknown, .now = scheduler->now, .match_data = NULL, .rsc_data = NULL, .op_data = NULL }; scheduler->config_hash = config_hash; pe__unpack_dataset_nvpairs(config, XML_CIB_TAG_PROPSET, &rule_data, config_hash, CIB_OPTIONS_FIRST, FALSE, scheduler); verify_pe_options(scheduler->config_hash); set_config_flag(scheduler, "enable-startup-probes", pcmk_sched_probe_resources); if (!pcmk_is_set(scheduler->flags, pcmk_sched_probe_resources)) { crm_info("Startup probes: disabled (dangerous)"); } value = pe_pref(scheduler->config_hash, XML_ATTR_HAVE_WATCHDOG); if (value && crm_is_true(value)) { crm_info("Watchdog-based self-fencing will be performed via SBD if " "fencing is required and stonith-watchdog-timeout is nonzero"); pe__set_working_set_flags(scheduler, pcmk_sched_have_fencing); } /* Set certain flags via xpath here, so they can be used before the relevant * configuration sections are unpacked. */ set_if_xpath(pcmk_sched_enable_unfencing, XPATH_ENABLE_UNFENCING, scheduler); value = pe_pref(scheduler->config_hash, "stonith-timeout"); scheduler->stonith_timeout = (int) crm_parse_interval_spec(value); crm_debug("STONITH timeout: %d", scheduler->stonith_timeout); set_config_flag(scheduler, "stonith-enabled", pcmk_sched_fencing_enabled); if (pcmk_is_set(scheduler->flags, pcmk_sched_fencing_enabled)) { crm_debug("STONITH of failed nodes is enabled"); } else { crm_debug("STONITH of failed nodes is disabled"); } scheduler->stonith_action = pe_pref(scheduler->config_hash, "stonith-action"); if (!strcmp(scheduler->stonith_action, "poweroff")) { pe_warn_once(pcmk__wo_poweroff, "Support for stonith-action of 'poweroff' is deprecated " "and will be removed in a future release (use 'off' instead)"); scheduler->stonith_action = PCMK_ACTION_OFF; } crm_trace("STONITH will %s nodes", scheduler->stonith_action); set_config_flag(scheduler, "concurrent-fencing", pcmk_sched_concurrent_fencing); if (pcmk_is_set(scheduler->flags, pcmk_sched_concurrent_fencing)) { crm_debug("Concurrent fencing is enabled"); } else { crm_debug("Concurrent fencing is disabled"); } value = pe_pref(scheduler->config_hash, XML_CONFIG_ATTR_PRIORITY_FENCING_DELAY); if (value) { scheduler->priority_fencing_delay = crm_parse_interval_spec(value) / 1000; crm_trace("Priority fencing delay is %ds", scheduler->priority_fencing_delay); } set_config_flag(scheduler, "stop-all-resources", pcmk_sched_stop_all); crm_debug("Stop all active resources: %s", pcmk__btoa(pcmk_is_set(scheduler->flags, pcmk_sched_stop_all))); set_config_flag(scheduler, "symmetric-cluster", pcmk_sched_symmetric_cluster); if (pcmk_is_set(scheduler->flags, pcmk_sched_symmetric_cluster)) { crm_debug("Cluster is symmetric" " - resources can run anywhere by default"); } value = pe_pref(scheduler->config_hash, "no-quorum-policy"); if (pcmk__str_eq(value, "ignore", pcmk__str_casei)) { scheduler->no_quorum_policy = pcmk_no_quorum_ignore; } else if (pcmk__str_eq(value, "freeze", pcmk__str_casei)) { scheduler->no_quorum_policy = pcmk_no_quorum_freeze; } else if (pcmk__str_eq(value, "demote", pcmk__str_casei)) { scheduler->no_quorum_policy = pcmk_no_quorum_demote; } else if (pcmk__str_eq(value, "suicide", pcmk__str_casei)) { if (pcmk_is_set(scheduler->flags, pcmk_sched_fencing_enabled)) { int do_panic = 0; crm_element_value_int(scheduler->input, XML_ATTR_QUORUM_PANIC, &do_panic); if (do_panic || pcmk_is_set(scheduler->flags, pcmk_sched_quorate)) { scheduler->no_quorum_policy = pcmk_no_quorum_fence; } else { crm_notice("Resetting no-quorum-policy to 'stop': cluster has never had quorum"); scheduler->no_quorum_policy = pcmk_no_quorum_stop; } } else { pcmk__config_err("Resetting no-quorum-policy to 'stop' because " "fencing is disabled"); scheduler->no_quorum_policy = pcmk_no_quorum_stop; } } else { scheduler->no_quorum_policy = pcmk_no_quorum_stop; } switch (scheduler->no_quorum_policy) { case pcmk_no_quorum_freeze: crm_debug("On loss of quorum: Freeze resources"); break; case pcmk_no_quorum_stop: crm_debug("On loss of quorum: Stop ALL resources"); break; case pcmk_no_quorum_demote: crm_debug("On loss of quorum: " "Demote promotable resources and stop other resources"); break; case pcmk_no_quorum_fence: crm_notice("On loss of quorum: Fence all remaining nodes"); break; case pcmk_no_quorum_ignore: crm_notice("On loss of quorum: Ignore"); break; } set_config_flag(scheduler, "stop-orphan-resources", pcmk_sched_stop_removed_resources); if (pcmk_is_set(scheduler->flags, pcmk_sched_stop_removed_resources)) { crm_trace("Orphan resources are stopped"); } else { crm_trace("Orphan resources are ignored"); } set_config_flag(scheduler, "stop-orphan-actions", pcmk_sched_cancel_removed_actions); if (pcmk_is_set(scheduler->flags, pcmk_sched_cancel_removed_actions)) { crm_trace("Orphan resource actions are stopped"); } else { crm_trace("Orphan resource actions are ignored"); } value = pe_pref(scheduler->config_hash, "remove-after-stop"); if (value != NULL) { if (crm_is_true(value)) { pe__set_working_set_flags(scheduler, pcmk_sched_remove_after_stop); #ifndef PCMK__COMPAT_2_0 pe_warn_once(pcmk__wo_remove_after, "Support for the remove-after-stop cluster property is" " deprecated and will be removed in a future release"); #endif } else { pe__clear_working_set_flags(scheduler, pcmk_sched_remove_after_stop); } } set_config_flag(scheduler, "maintenance-mode", pcmk_sched_in_maintenance); crm_trace("Maintenance mode: %s", pcmk__btoa(pcmk_is_set(scheduler->flags, pcmk_sched_in_maintenance))); set_config_flag(scheduler, "start-failure-is-fatal", pcmk_sched_start_failure_fatal); if (pcmk_is_set(scheduler->flags, pcmk_sched_start_failure_fatal)) { crm_trace("Start failures are always fatal"); } else { crm_trace("Start failures are handled by failcount"); } if (pcmk_is_set(scheduler->flags, pcmk_sched_fencing_enabled)) { set_config_flag(scheduler, "startup-fencing", pcmk_sched_startup_fencing); } if (pcmk_is_set(scheduler->flags, pcmk_sched_startup_fencing)) { crm_trace("Unseen nodes will be fenced"); } else { pe_warn_once(pcmk__wo_blind, "Blind faith: not fencing unseen nodes"); } pe__unpack_node_health_scores(scheduler); scheduler->placement_strategy = pe_pref(scheduler->config_hash, "placement-strategy"); crm_trace("Placement strategy: %s", scheduler->placement_strategy); set_config_flag(scheduler, "shutdown-lock", pcmk_sched_shutdown_lock); if (pcmk_is_set(scheduler->flags, pcmk_sched_shutdown_lock)) { value = pe_pref(scheduler->config_hash, XML_CONFIG_ATTR_SHUTDOWN_LOCK_LIMIT); scheduler->shutdown_lock = crm_parse_interval_spec(value) / 1000; crm_trace("Resources will be locked to nodes that were cleanly " "shut down (locks expire after %s)", pcmk__readable_interval(scheduler->shutdown_lock)); } else { crm_trace("Resources will not be locked to nodes that were cleanly " "shut down"); } value = pe_pref(scheduler->config_hash, XML_CONFIG_ATTR_NODE_PENDING_TIMEOUT); scheduler->node_pending_timeout = crm_parse_interval_spec(value) / 1000; if (scheduler->node_pending_timeout == 0) { crm_trace("Do not fence pending nodes"); } else { crm_trace("Fence pending nodes after %s", pcmk__readable_interval(scheduler->node_pending_timeout * 1000)); } return TRUE; } pcmk_node_t * pe_create_node(const char *id, const char *uname, const char *type, const char *score, pcmk_scheduler_t *scheduler) { pcmk_node_t *new_node = NULL; if (pe_find_node(scheduler->nodes, uname) != NULL) { pcmk__config_warn("More than one node entry has name '%s'", uname); } new_node = calloc(1, sizeof(pcmk_node_t)); if (new_node == NULL) { return NULL; } new_node->weight = char2score(score); new_node->details = calloc(1, sizeof(struct pe_node_shared_s)); if (new_node->details == NULL) { free(new_node); return NULL; } crm_trace("Creating node for entry %s/%s", uname, id); new_node->details->id = id; new_node->details->uname = uname; new_node->details->online = FALSE; new_node->details->shutdown = FALSE; new_node->details->rsc_discovery_enabled = TRUE; new_node->details->running_rsc = NULL; new_node->details->data_set = scheduler; if (pcmk__str_eq(type, "member", pcmk__str_null_matches | pcmk__str_casei)) { new_node->details->type = pcmk_node_variant_cluster; } else if (pcmk__str_eq(type, "remote", pcmk__str_casei)) { new_node->details->type = pcmk_node_variant_remote; pe__set_working_set_flags(scheduler, pcmk_sched_have_remote_nodes); } else { /* @COMPAT 'ping' is the default for backward compatibility, but it * should be changed to 'member' at a compatibility break */ if (!pcmk__str_eq(type, "ping", pcmk__str_casei)) { pcmk__config_warn("Node %s has unrecognized type '%s', " "assuming 'ping'", pcmk__s(uname, "without name"), type); } pe_warn_once(pcmk__wo_ping_node, "Support for nodes of type 'ping' (such as %s) is " "deprecated and will be removed in a future release", pcmk__s(uname, "unnamed node")); new_node->details->type = node_ping; } new_node->details->attrs = pcmk__strkey_table(free, free); if (pe__is_guest_or_remote_node(new_node)) { g_hash_table_insert(new_node->details->attrs, strdup(CRM_ATTR_KIND), strdup("remote")); } else { g_hash_table_insert(new_node->details->attrs, strdup(CRM_ATTR_KIND), strdup("cluster")); } new_node->details->utilization = pcmk__strkey_table(free, free); new_node->details->digest_cache = pcmk__strkey_table(free, pe__free_digests); scheduler->nodes = g_list_insert_sorted(scheduler->nodes, new_node, pe__cmp_node_name); return new_node; } static const char * expand_remote_rsc_meta(xmlNode *xml_obj, xmlNode *parent, pcmk_scheduler_t *data) { xmlNode *attr_set = NULL; xmlNode *attr = NULL; const char *container_id = ID(xml_obj); const char *remote_name = NULL; const char *remote_server = NULL; const char *remote_port = NULL; const char *connect_timeout = "60s"; const char *remote_allow_migrate=NULL; const char *is_managed = NULL; for (attr_set = pcmk__xe_first_child(xml_obj); attr_set != NULL; attr_set = pcmk__xe_next(attr_set)) { if (!pcmk__str_eq((const char *)attr_set->name, XML_TAG_META_SETS, pcmk__str_casei)) { continue; } for (attr = pcmk__xe_first_child(attr_set); attr != NULL; attr = pcmk__xe_next(attr)) { const char *value = crm_element_value(attr, XML_NVPAIR_ATTR_VALUE); const char *name = crm_element_value(attr, XML_NVPAIR_ATTR_NAME); if (pcmk__str_eq(name, XML_RSC_ATTR_REMOTE_NODE, pcmk__str_casei)) { remote_name = value; } else if (pcmk__str_eq(name, "remote-addr", pcmk__str_casei)) { remote_server = value; } else if (pcmk__str_eq(name, "remote-port", pcmk__str_casei)) { remote_port = value; } else if (pcmk__str_eq(name, "remote-connect-timeout", pcmk__str_casei)) { connect_timeout = value; } else if (pcmk__str_eq(name, "remote-allow-migrate", pcmk__str_casei)) { remote_allow_migrate=value; } else if (pcmk__str_eq(name, XML_RSC_ATTR_MANAGED, pcmk__str_casei)) { is_managed = value; } } } if (remote_name == NULL) { return NULL; } if (pe_find_resource(data->resources, remote_name) != NULL) { return NULL; } pe_create_remote_xml(parent, remote_name, container_id, remote_allow_migrate, is_managed, connect_timeout, remote_server, remote_port); return remote_name; } static void handle_startup_fencing(pcmk_scheduler_t *scheduler, pcmk_node_t *new_node) { if ((new_node->details->type == pcmk_node_variant_remote) && (new_node->details->remote_rsc == NULL)) { /* Ignore fencing for remote nodes that don't have a connection resource * associated with them. This happens when remote node entries get left * in the nodes section after the connection resource is removed. */ return; } if (pcmk_is_set(scheduler->flags, pcmk_sched_startup_fencing)) { // All nodes are unclean until we've seen their status entry new_node->details->unclean = TRUE; } else { // Blind faith ... new_node->details->unclean = FALSE; } /* We need to be able to determine if a node's status section * exists or not separate from whether the node is unclean. */ new_node->details->unseen = TRUE; } gboolean unpack_nodes(xmlNode *xml_nodes, pcmk_scheduler_t *scheduler) { xmlNode *xml_obj = NULL; pcmk_node_t *new_node = NULL; const char *id = NULL; const char *uname = NULL; const char *type = NULL; const char *score = NULL; for (xml_obj = pcmk__xe_first_child(xml_nodes); xml_obj != NULL; xml_obj = pcmk__xe_next(xml_obj)) { if (pcmk__str_eq((const char *)xml_obj->name, XML_CIB_TAG_NODE, pcmk__str_none)) { new_node = NULL; id = crm_element_value(xml_obj, XML_ATTR_ID); uname = crm_element_value(xml_obj, XML_ATTR_UNAME); type = crm_element_value(xml_obj, XML_ATTR_TYPE); score = crm_element_value(xml_obj, XML_RULE_ATTR_SCORE); crm_trace("Processing node %s/%s", uname, id); if (id == NULL) { pcmk__config_err("Ignoring <" XML_CIB_TAG_NODE "> entry in configuration without id"); continue; } new_node = pe_create_node(id, uname, type, score, scheduler); if (new_node == NULL) { return FALSE; } handle_startup_fencing(scheduler, new_node); add_node_attrs(xml_obj, new_node, FALSE, scheduler); crm_trace("Done with node %s", crm_element_value(xml_obj, XML_ATTR_UNAME)); } } if (scheduler->localhost && (pe_find_node(scheduler->nodes, scheduler->localhost) == NULL)) { crm_info("Creating a fake local node"); pe_create_node(scheduler->localhost, scheduler->localhost, NULL, 0, scheduler); } return TRUE; } static void setup_container(pcmk_resource_t *rsc, pcmk_scheduler_t *scheduler) { const char *container_id = NULL; if (rsc->children) { g_list_foreach(rsc->children, (GFunc) setup_container, scheduler); return; } container_id = g_hash_table_lookup(rsc->meta, XML_RSC_ATTR_CONTAINER); if (container_id && !pcmk__str_eq(container_id, rsc->id, pcmk__str_casei)) { pcmk_resource_t *container = pe_find_resource(scheduler->resources, container_id); if (container) { rsc->container = container; pe__set_resource_flags(container, pcmk_rsc_has_filler); container->fillers = g_list_append(container->fillers, rsc); pe_rsc_trace(rsc, "Resource %s's container is %s", rsc->id, container_id); } else { pe_err("Resource %s: Unknown resource container (%s)", rsc->id, container_id); } } } gboolean unpack_remote_nodes(xmlNode *xml_resources, pcmk_scheduler_t *scheduler) { xmlNode *xml_obj = NULL; /* Create remote nodes and guest nodes from the resource configuration * before unpacking resources. */ for (xml_obj = pcmk__xe_first_child(xml_resources); xml_obj != NULL; xml_obj = pcmk__xe_next(xml_obj)) { const char *new_node_id = NULL; /* Check for remote nodes, which are defined by ocf:pacemaker:remote * primitives. */ if (xml_contains_remote_node(xml_obj)) { new_node_id = ID(xml_obj); /* The "pe_find_node" check is here to make sure we don't iterate over * an expanded node that has already been added to the node list. */ if (new_node_id && (pe_find_node(scheduler->nodes, new_node_id) == NULL)) { crm_trace("Found remote node %s defined by resource %s", new_node_id, ID(xml_obj)); pe_create_node(new_node_id, new_node_id, "remote", NULL, scheduler); } continue; } /* Check for guest nodes, which are defined by special meta-attributes * of a primitive of any type (for example, VirtualDomain or Xen). */ if (pcmk__str_eq((const char *)xml_obj->name, XML_CIB_TAG_RESOURCE, pcmk__str_none)) { /* This will add an ocf:pacemaker:remote primitive to the * configuration for the guest node's connection, to be unpacked * later. */ new_node_id = expand_remote_rsc_meta(xml_obj, xml_resources, scheduler); if (new_node_id && (pe_find_node(scheduler->nodes, new_node_id) == NULL)) { crm_trace("Found guest node %s in resource %s", new_node_id, ID(xml_obj)); pe_create_node(new_node_id, new_node_id, "remote", NULL, scheduler); } continue; } /* Check for guest nodes inside a group. Clones are currently not * supported as guest nodes. */ if (pcmk__str_eq((const char *)xml_obj->name, XML_CIB_TAG_GROUP, pcmk__str_none)) { xmlNode *xml_obj2 = NULL; for (xml_obj2 = pcmk__xe_first_child(xml_obj); xml_obj2 != NULL; xml_obj2 = pcmk__xe_next(xml_obj2)) { new_node_id = expand_remote_rsc_meta(xml_obj2, xml_resources, scheduler); if (new_node_id && (pe_find_node(scheduler->nodes, new_node_id) == NULL)) { crm_trace("Found guest node %s in resource %s inside group %s", new_node_id, ID(xml_obj2), ID(xml_obj)); pe_create_node(new_node_id, new_node_id, "remote", NULL, scheduler); } } } } return TRUE; } /* Call this after all the nodes and resources have been * unpacked, but before the status section is read. * * A remote node's online status is reflected by the state * of the remote node's connection resource. We need to link * the remote node to this connection resource so we can have * easy access to the connection resource during the scheduler calculations. */ static void link_rsc2remotenode(pcmk_scheduler_t *scheduler, pcmk_resource_t *new_rsc) { pcmk_node_t *remote_node = NULL; if (new_rsc->is_remote_node == FALSE) { return; } if (pcmk_is_set(scheduler->flags, pcmk_sched_location_only)) { /* remote_nodes and remote_resources are not linked in quick location calculations */ return; } remote_node = pe_find_node(scheduler->nodes, new_rsc->id); CRM_CHECK(remote_node != NULL, return); pe_rsc_trace(new_rsc, "Linking remote connection resource %s to %s", new_rsc->id, pe__node_name(remote_node)); remote_node->details->remote_rsc = new_rsc; if (new_rsc->container == NULL) { /* Handle start-up fencing for remote nodes (as opposed to guest nodes) * the same as is done for cluster nodes. */ handle_startup_fencing(scheduler, remote_node); } else { /* pe_create_node() marks the new node as "remote" or "cluster"; now * that we know the node is a guest node, update it correctly. */ g_hash_table_replace(remote_node->details->attrs, strdup(CRM_ATTR_KIND), strdup("container")); } } static void destroy_tag(gpointer data) { pcmk_tag_t *tag = data; if (tag) { free(tag->id); g_list_free_full(tag->refs, free); free(tag); } } /*! * \internal * \brief Parse configuration XML for resource information * * \param[in] xml_resources Top of resource configuration XML * \param[in,out] scheduler Scheduler data * * \return TRUE * * \note unpack_remote_nodes() MUST be called before this, so that the nodes can * be used when pe__unpack_resource() calls resource_location() */ gboolean unpack_resources(const xmlNode *xml_resources, pcmk_scheduler_t *scheduler) { xmlNode *xml_obj = NULL; GList *gIter = NULL; scheduler->template_rsc_sets = pcmk__strkey_table(free, destroy_tag); for (xml_obj = pcmk__xe_first_child(xml_resources); xml_obj != NULL; xml_obj = pcmk__xe_next(xml_obj)) { pcmk_resource_t *new_rsc = NULL; const char *id = ID(xml_obj); if (pcmk__str_empty(id)) { pcmk__config_err("Ignoring <%s> resource without ID", xml_obj->name); continue; } if (pcmk__str_eq((const char *) xml_obj->name, XML_CIB_TAG_RSC_TEMPLATE, pcmk__str_none)) { if (g_hash_table_lookup_extended(scheduler->template_rsc_sets, id, NULL, NULL) == FALSE) { /* Record the template's ID for the knowledge of its existence anyway. */ g_hash_table_insert(scheduler->template_rsc_sets, strdup(id), NULL); } continue; } crm_trace("Unpacking <%s " XML_ATTR_ID "='%s'>", xml_obj->name, id); if (pe__unpack_resource(xml_obj, &new_rsc, NULL, scheduler) == pcmk_rc_ok) { scheduler->resources = g_list_append(scheduler->resources, new_rsc); pe_rsc_trace(new_rsc, "Added resource %s", new_rsc->id); } else { pcmk__config_err("Ignoring <%s> resource '%s' " "because configuration is invalid", xml_obj->name, id); } } for (gIter = scheduler->resources; gIter != NULL; gIter = gIter->next) { pcmk_resource_t *rsc = (pcmk_resource_t *) gIter->data; setup_container(rsc, scheduler); link_rsc2remotenode(scheduler, rsc); } scheduler->resources = g_list_sort(scheduler->resources, pe__cmp_rsc_priority); if (pcmk_is_set(scheduler->flags, pcmk_sched_location_only)) { /* Ignore */ } else if (pcmk_is_set(scheduler->flags, pcmk_sched_fencing_enabled) && !pcmk_is_set(scheduler->flags, pcmk_sched_have_fencing)) { pcmk__config_err("Resource start-up disabled since no STONITH resources have been defined"); pcmk__config_err("Either configure some or disable STONITH with the stonith-enabled option"); pcmk__config_err("NOTE: Clusters with shared data need STONITH to ensure data integrity"); } return TRUE; } gboolean unpack_tags(xmlNode *xml_tags, pcmk_scheduler_t *scheduler) { xmlNode *xml_tag = NULL; scheduler->tags = pcmk__strkey_table(free, destroy_tag); for (xml_tag = pcmk__xe_first_child(xml_tags); xml_tag != NULL; xml_tag = pcmk__xe_next(xml_tag)) { xmlNode *xml_obj_ref = NULL; const char *tag_id = ID(xml_tag); if (!pcmk__str_eq((const char *)xml_tag->name, XML_CIB_TAG_TAG, pcmk__str_none)) { continue; } if (tag_id == NULL) { pcmk__config_err("Ignoring <%s> without " XML_ATTR_ID, (const char *) xml_tag->name); continue; } for (xml_obj_ref = pcmk__xe_first_child(xml_tag); xml_obj_ref != NULL; xml_obj_ref = pcmk__xe_next(xml_obj_ref)) { const char *obj_ref = ID(xml_obj_ref); if (!pcmk__str_eq((const char *)xml_obj_ref->name, XML_CIB_TAG_OBJ_REF, pcmk__str_none)) { continue; } if (obj_ref == NULL) { pcmk__config_err("Ignoring <%s> for tag '%s' without " XML_ATTR_ID, xml_obj_ref->name, tag_id); continue; } if (add_tag_ref(scheduler->tags, tag_id, obj_ref) == FALSE) { return FALSE; } } } return TRUE; } /* The ticket state section: * "/cib/status/tickets/ticket_state" */ static gboolean unpack_ticket_state(xmlNode *xml_ticket, pcmk_scheduler_t *scheduler) { const char *ticket_id = NULL; const char *granted = NULL; const char *last_granted = NULL; const char *standby = NULL; xmlAttrPtr xIter = NULL; pcmk_ticket_t *ticket = NULL; ticket_id = ID(xml_ticket); if (pcmk__str_empty(ticket_id)) { return FALSE; } crm_trace("Processing ticket state for %s", ticket_id); ticket = g_hash_table_lookup(scheduler->tickets, ticket_id); if (ticket == NULL) { ticket = ticket_new(ticket_id, scheduler); if (ticket == NULL) { return FALSE; } } for (xIter = xml_ticket->properties; xIter; xIter = xIter->next) { const char *prop_name = (const char *)xIter->name; const char *prop_value = pcmk__xml_attr_value(xIter); if (pcmk__str_eq(prop_name, XML_ATTR_ID, pcmk__str_none)) { continue; } g_hash_table_replace(ticket->state, strdup(prop_name), strdup(prop_value)); } granted = g_hash_table_lookup(ticket->state, "granted"); if (granted && crm_is_true(granted)) { ticket->granted = TRUE; crm_info("We have ticket '%s'", ticket->id); } else { ticket->granted = FALSE; crm_info("We do not have ticket '%s'", ticket->id); } last_granted = g_hash_table_lookup(ticket->state, "last-granted"); if (last_granted) { long long last_granted_ll; pcmk__scan_ll(last_granted, &last_granted_ll, 0LL); ticket->last_granted = (time_t) last_granted_ll; } standby = g_hash_table_lookup(ticket->state, "standby"); if (standby && crm_is_true(standby)) { ticket->standby = TRUE; if (ticket->granted) { crm_info("Granted ticket '%s' is in standby-mode", ticket->id); } } else { ticket->standby = FALSE; } crm_trace("Done with ticket state for %s", ticket_id); return TRUE; } static gboolean unpack_tickets_state(xmlNode *xml_tickets, pcmk_scheduler_t *scheduler) { xmlNode *xml_obj = NULL; for (xml_obj = pcmk__xe_first_child(xml_tickets); xml_obj != NULL; xml_obj = pcmk__xe_next(xml_obj)) { if (!pcmk__str_eq((const char *)xml_obj->name, XML_CIB_TAG_TICKET_STATE, pcmk__str_none)) { continue; } unpack_ticket_state(xml_obj, scheduler); } return TRUE; } static void unpack_handle_remote_attrs(pcmk_node_t *this_node, const xmlNode *state, pcmk_scheduler_t *scheduler) { const char *resource_discovery_enabled = NULL; const xmlNode *attrs = NULL; pcmk_resource_t *rsc = NULL; if (!pcmk__str_eq((const char *)state->name, XML_CIB_TAG_STATE, pcmk__str_none)) { return; } if ((this_node == NULL) || !pe__is_guest_or_remote_node(this_node)) { return; } crm_trace("Processing Pacemaker Remote node %s", pe__node_name(this_node)); pcmk__scan_min_int(crm_element_value(state, XML_NODE_IS_MAINTENANCE), &(this_node->details->remote_maintenance), 0); rsc = this_node->details->remote_rsc; if (this_node->details->remote_requires_reset == FALSE) { this_node->details->unclean = FALSE; this_node->details->unseen = FALSE; } attrs = find_xml_node(state, XML_TAG_TRANSIENT_NODEATTRS, FALSE); add_node_attrs(attrs, this_node, TRUE, scheduler); if (pe__shutdown_requested(this_node)) { crm_info("%s is shutting down", pe__node_name(this_node)); this_node->details->shutdown = TRUE; } if (crm_is_true(pe_node_attribute_raw(this_node, "standby"))) { crm_info("%s is in standby mode", pe__node_name(this_node)); this_node->details->standby = TRUE; } if (crm_is_true(pe_node_attribute_raw(this_node, "maintenance")) || ((rsc != NULL) && !pcmk_is_set(rsc->flags, pcmk_rsc_managed))) { crm_info("%s is in maintenance mode", pe__node_name(this_node)); this_node->details->maintenance = TRUE; } resource_discovery_enabled = pe_node_attribute_raw(this_node, XML_NODE_ATTR_RSC_DISCOVERY); if (resource_discovery_enabled && !crm_is_true(resource_discovery_enabled)) { if (pe__is_remote_node(this_node) && !pcmk_is_set(scheduler->flags, pcmk_sched_fencing_enabled)) { crm_warn("Ignoring " XML_NODE_ATTR_RSC_DISCOVERY " attribute on Pacemaker Remote node %s" " because fencing is disabled", pe__node_name(this_node)); } else { /* This is either a remote node with fencing enabled, or a guest * node. We don't care whether fencing is enabled when fencing guest * nodes, because they are "fenced" by recovering their containing * resource. */ crm_info("%s has resource discovery disabled", pe__node_name(this_node)); this_node->details->rsc_discovery_enabled = FALSE; } } } /*! * \internal * \brief Unpack a cluster node's transient attributes * * \param[in] state CIB node state XML * \param[in,out] node Cluster node whose attributes are being unpacked * \param[in,out] scheduler Scheduler data */ static void unpack_transient_attributes(const xmlNode *state, pcmk_node_t *node, pcmk_scheduler_t *scheduler) { const char *discovery = NULL; const xmlNode *attrs = find_xml_node(state, XML_TAG_TRANSIENT_NODEATTRS, FALSE); add_node_attrs(attrs, node, TRUE, scheduler); if (crm_is_true(pe_node_attribute_raw(node, "standby"))) { crm_info("%s is in standby mode", pe__node_name(node)); node->details->standby = TRUE; } if (crm_is_true(pe_node_attribute_raw(node, "maintenance"))) { crm_info("%s is in maintenance mode", pe__node_name(node)); node->details->maintenance = TRUE; } discovery = pe_node_attribute_raw(node, XML_NODE_ATTR_RSC_DISCOVERY); if ((discovery != NULL) && !crm_is_true(discovery)) { crm_warn("Ignoring " XML_NODE_ATTR_RSC_DISCOVERY " attribute for %s because disabling resource discovery " "is not allowed for cluster nodes", pe__node_name(node)); } } /*! * \internal * \brief Unpack a node state entry (first pass) * * Unpack one node state entry from status. This unpacks information from the * node_state element itself and node attributes inside it, but not the * resource history inside it. Multiple passes through the status are needed to * fully unpack everything. * * \param[in] state CIB node state XML * \param[in,out] scheduler Scheduler data */ static void unpack_node_state(const xmlNode *state, pcmk_scheduler_t *scheduler) { const char *id = NULL; const char *uname = NULL; pcmk_node_t *this_node = NULL; id = crm_element_value(state, XML_ATTR_ID); if (id == NULL) { crm_warn("Ignoring malformed " XML_CIB_TAG_STATE " entry without " XML_ATTR_ID); return; } uname = crm_element_value(state, XML_ATTR_UNAME); if (uname == NULL) { /* If a joining peer makes the cluster acquire the quorum from corosync * meanwhile it has not joined CPG membership of pacemaker-controld yet, * it's possible that the created node_state entry doesn't have an uname * yet. We should recognize the node as `pending` and wait for it to * join CPG. */ crm_trace("Handling " XML_CIB_TAG_STATE " entry with id=\"%s\" without " XML_ATTR_UNAME, id); } this_node = pe_find_node_any(scheduler->nodes, id, uname); if (this_node == NULL) { pcmk__config_warn("Ignoring recorded node state for id=\"%s\" (%s) " "because it is no longer in the configuration", id, pcmk__s(uname, "uname unknown")); return; } if (pe__is_guest_or_remote_node(this_node)) { /* We can't determine the online status of Pacemaker Remote nodes until * after all resource history has been unpacked. In this first pass, we * do need to mark whether the node has been fenced, as this plays a * role during unpacking cluster node resource state. */ pcmk__scan_min_int(crm_element_value(state, XML_NODE_IS_FENCED), &(this_node->details->remote_was_fenced), 0); return; } unpack_transient_attributes(state, this_node, scheduler); /* Provisionally mark this cluster node as clean. We have at least seen it * in the current cluster's lifetime. */ this_node->details->unclean = FALSE; this_node->details->unseen = FALSE; crm_trace("Determining online status of cluster node %s (id %s)", pe__node_name(this_node), id); determine_online_status(state, this_node, scheduler); if (!pcmk_is_set(scheduler->flags, pcmk_sched_quorate) && this_node->details->online && (scheduler->no_quorum_policy == pcmk_no_quorum_fence)) { /* Everything else should flow from this automatically * (at least until the scheduler becomes able to migrate off * healthy resources) */ pe_fence_node(scheduler, this_node, "cluster does not have quorum", FALSE); } } /*! * \internal * \brief Unpack nodes' resource history as much as possible * * Unpack as many nodes' resource history as possible in one pass through the * status. We need to process Pacemaker Remote nodes' connections/containers * before unpacking their history; the connection/container history will be * in another node's history, so it might take multiple passes to unpack * everything. * * \param[in] status CIB XML status section * \param[in] fence If true, treat any not-yet-unpacked nodes as unseen * \param[in,out] scheduler Scheduler data * * \return Standard Pacemaker return code (specifically pcmk_rc_ok if done, * or EAGAIN if more unpacking remains to be done) */ static int unpack_node_history(const xmlNode *status, bool fence, pcmk_scheduler_t *scheduler) { int rc = pcmk_rc_ok; // Loop through all node_state entries in CIB status for (const xmlNode *state = first_named_child(status, XML_CIB_TAG_STATE); state != NULL; state = crm_next_same_xml(state)) { const char *id = ID(state); const char *uname = crm_element_value(state, XML_ATTR_UNAME); pcmk_node_t *this_node = NULL; if ((id == NULL) || (uname == NULL)) { // Warning already logged in first pass through status section crm_trace("Not unpacking resource history from malformed " XML_CIB_TAG_STATE " without id and/or uname"); continue; } this_node = pe_find_node_any(scheduler->nodes, id, uname); if (this_node == NULL) { // Warning already logged in first pass through status section crm_trace("Not unpacking resource history for node %s because " "no longer in configuration", id); continue; } if (this_node->details->unpacked) { crm_trace("Not unpacking resource history for node %s because " "already unpacked", id); continue; } if (fence) { // We're processing all remaining nodes } else if (pe__is_guest_node(this_node)) { /* We can unpack a guest node's history only after we've unpacked * other resource history to the point that we know that the node's * connection and containing resource are both up. */ pcmk_resource_t *rsc = this_node->details->remote_rsc; if ((rsc == NULL) || (rsc->role != pcmk_role_started) || (rsc->container->role != pcmk_role_started)) { crm_trace("Not unpacking resource history for guest node %s " "because container and connection are not known to " "be up", id); continue; } } else if (pe__is_remote_node(this_node)) { /* We can unpack a remote node's history only after we've unpacked * other resource history to the point that we know that the node's * connection is up, with the exception of when shutdown locks are * in use. */ pcmk_resource_t *rsc = this_node->details->remote_rsc; if ((rsc == NULL) || (!pcmk_is_set(scheduler->flags, pcmk_sched_shutdown_lock) && (rsc->role != pcmk_role_started))) { crm_trace("Not unpacking resource history for remote node %s " "because connection is not known to be up", id); continue; } /* If fencing and shutdown locks are disabled and we're not processing * unseen nodes, then we don't want to unpack offline nodes until online * nodes have been unpacked. This allows us to number active clone * instances first. */ } else if (!pcmk_any_flags_set(scheduler->flags, pcmk_sched_fencing_enabled |pcmk_sched_shutdown_lock) && !this_node->details->online) { crm_trace("Not unpacking resource history for offline " "cluster node %s", id); continue; } if (pe__is_guest_or_remote_node(this_node)) { determine_remote_online_status(scheduler, this_node); unpack_handle_remote_attrs(this_node, state, scheduler); } crm_trace("Unpacking resource history for %snode %s", (fence? "unseen " : ""), id); this_node->details->unpacked = TRUE; unpack_node_lrm(this_node, state, scheduler); rc = EAGAIN; // Other node histories might depend on this one } return rc; } /* remove nodes that are down, stopping */ /* create positive rsc_to_node constraints between resources and the nodes they are running on */ /* anything else? */ gboolean unpack_status(xmlNode *status, pcmk_scheduler_t *scheduler) { xmlNode *state = NULL; crm_trace("Beginning unpack"); if (scheduler->tickets == NULL) { scheduler->tickets = pcmk__strkey_table(free, destroy_ticket); } for (state = pcmk__xe_first_child(status); state != NULL; state = pcmk__xe_next(state)) { if (pcmk__str_eq((const char *)state->name, XML_CIB_TAG_TICKETS, pcmk__str_none)) { unpack_tickets_state((xmlNode *) state, scheduler); } else if (pcmk__str_eq((const char *)state->name, XML_CIB_TAG_STATE, pcmk__str_none)) { unpack_node_state(state, scheduler); } } while (unpack_node_history(status, FALSE, scheduler) == EAGAIN) { crm_trace("Another pass through node resource histories is needed"); } // Now catch any nodes we didn't see unpack_node_history(status, pcmk_is_set(scheduler->flags, pcmk_sched_fencing_enabled), scheduler); /* Now that we know where resources are, we can schedule stops of containers * with failed bundle connections */ if (scheduler->stop_needed != NULL) { for (GList *item = scheduler->stop_needed; item; item = item->next) { pcmk_resource_t *container = item->data; pcmk_node_t *node = pe__current_node(container); if (node) { stop_action(container, node, FALSE); } } g_list_free(scheduler->stop_needed); scheduler->stop_needed = NULL; } /* Now that we know status of all Pacemaker Remote connections and nodes, * we can stop connections for node shutdowns, and check the online status * of remote/guest nodes that didn't have any node history to unpack. */ for (GList *gIter = scheduler->nodes; gIter != NULL; gIter = gIter->next) { pcmk_node_t *this_node = gIter->data; if (!pe__is_guest_or_remote_node(this_node)) { continue; } if (this_node->details->shutdown && (this_node->details->remote_rsc != NULL)) { pe__set_next_role(this_node->details->remote_rsc, pcmk_role_stopped, "remote shutdown"); } if (!this_node->details->unpacked) { determine_remote_online_status(scheduler, this_node); } } return TRUE; } /*! * \internal * \brief Unpack node's time when it became a member at the cluster layer * * \param[in] node_state Node's node_state entry * \param[in,out] scheduler Scheduler data * * \return Epoch time when node became a cluster member * (or scheduler effective time for legacy entries) if a member, * 0 if not a member, or -1 if no valid information available */ static long long unpack_node_member(const xmlNode *node_state, pcmk_scheduler_t *scheduler) { const char *member_time = crm_element_value(node_state, PCMK__XA_IN_CCM); int member = 0; if (member_time == NULL) { return -1LL; } else if (crm_str_to_boolean(member_time, &member) == 1) { /* If in_ccm=0, we'll return 0 here. If in_ccm=1, either the entry was * recorded as a boolean for a DC < 2.1.7, or the node is pending * shutdown and has left the CPG, in which case it was set to 1 to avoid * fencing for node-pending-timeout. * * We return the effective time for in_ccm=1 because what's important to * avoid fencing is that effective time minus this value is less than * the pending node timeout. */ return member? (long long) get_effective_time(scheduler) : 0LL; } else { long long when_member = 0LL; if ((pcmk__scan_ll(member_time, &when_member, 0LL) != pcmk_rc_ok) || (when_member < 0LL)) { crm_warn("Unrecognized value '%s' for " PCMK__XA_IN_CCM " in " XML_CIB_TAG_STATE " entry", member_time); return -1LL; } return when_member; } } /*! * \internal * \brief Unpack node's time when it became online in process group * * \param[in] node_state Node's node_state entry * * \return Epoch time when node became online in process group (or 0 if not * online, or 1 for legacy online entries) */ static long long unpack_node_online(const xmlNode *node_state) { const char *peer_time = crm_element_value(node_state, PCMK__XA_CRMD); // @COMPAT Entries recorded for DCs < 2.1.7 have "online" or "offline" if (pcmk__str_eq(peer_time, OFFLINESTATUS, pcmk__str_casei|pcmk__str_null_matches)) { return 0LL; } else if (pcmk__str_eq(peer_time, ONLINESTATUS, pcmk__str_casei)) { return 1LL; } else { long long when_online = 0LL; if ((pcmk__scan_ll(peer_time, &when_online, 0LL) != pcmk_rc_ok) || (when_online < 0)) { crm_warn("Unrecognized value '%s' for " PCMK__XA_CRMD " in " XML_CIB_TAG_STATE " entry, assuming offline", peer_time); return 0LL; } return when_online; } } /*! * \internal * \brief Unpack node attribute for user-requested fencing * * \param[in] node Node to check * \param[in] node_state Node's node_state entry in CIB status * * \return \c true if fencing has been requested for \p node, otherwise \c false */ static bool unpack_node_terminate(const pcmk_node_t *node, const xmlNode *node_state) { long long value = 0LL; int value_i = 0; const char *value_s = pe_node_attribute_raw(node, PCMK_NODE_ATTR_TERMINATE); // Value may be boolean or an epoch time if (crm_str_to_boolean(value_s, &value_i) == 1) { return (value_i != 0); } if (pcmk__scan_ll(value_s, &value, 0LL) == pcmk_rc_ok) { return (value > 0); } crm_warn("Ignoring unrecognized value '%s' for " PCMK_NODE_ATTR_TERMINATE "node attribute for %s", value_s, pe__node_name(node)); return false; } static gboolean determine_online_status_no_fencing(pcmk_scheduler_t *scheduler, const xmlNode *node_state, pcmk_node_t *this_node) { gboolean online = FALSE; const char *join = crm_element_value(node_state, PCMK__XA_JOIN); const char *exp_state = crm_element_value(node_state, PCMK__XA_EXPECTED); long long when_member = unpack_node_member(node_state, scheduler); long long when_online = unpack_node_online(node_state); if (when_member <= 0) { crm_trace("Node %s is %sdown", pe__node_name(this_node), ((when_member < 0)? "presumed " : "")); } else if (when_online > 0) { if (pcmk__str_eq(join, CRMD_JOINSTATE_MEMBER, pcmk__str_casei)) { online = TRUE; } else { crm_debug("Node %s is not ready to run resources: %s", pe__node_name(this_node), join); } } else if (this_node->details->expected_up == FALSE) { crm_trace("Node %s controller is down: " "member@%lld online@%lld join=%s expected=%s", pe__node_name(this_node), when_member, when_online, pcmk__s(join, ""), pcmk__s(exp_state, "")); } else { /* mark it unclean */ pe_fence_node(scheduler, this_node, "peer is unexpectedly down", FALSE); crm_info("Node %s member@%lld online@%lld join=%s expected=%s", pe__node_name(this_node), when_member, when_online, pcmk__s(join, ""), pcmk__s(exp_state, "")); } return online; } /*! * \internal * \brief Check whether a node has taken too long to join controller group * * \param[in,out] scheduler Scheduler data * \param[in] node Node to check * \param[in] when_member Epoch time when node became a cluster member * \param[in] when_online Epoch time when node joined controller group * * \return true if node has been pending (on the way up) longer than * node-pending-timeout, otherwise false * \note This will also update the cluster's recheck time if appropriate. */ static inline bool pending_too_long(pcmk_scheduler_t *scheduler, const pcmk_node_t *node, long long when_member, long long when_online) { if ((scheduler->node_pending_timeout > 0) && (when_member > 0) && (when_online <= 0)) { // There is a timeout on pending nodes, and node is pending time_t timeout = when_member + scheduler->node_pending_timeout; if (get_effective_time(node->details->data_set) >= timeout) { return true; // Node has timed out } // Node is pending, but still has time pe__update_recheck_time(timeout, scheduler, "pending node timeout"); } return false; } static bool determine_online_status_fencing(pcmk_scheduler_t *scheduler, const xmlNode *node_state, pcmk_node_t *this_node) { bool termination_requested = unpack_node_terminate(this_node, node_state); const char *join = crm_element_value(node_state, PCMK__XA_JOIN); const char *exp_state = crm_element_value(node_state, PCMK__XA_EXPECTED); long long when_member = unpack_node_member(node_state, scheduler); long long when_online = unpack_node_online(node_state); /* - PCMK__XA_JOIN ::= member|down|pending|banned - PCMK__XA_EXPECTED ::= member|down @COMPAT with entries recorded for DCs < 2.1.7 - PCMK__XA_IN_CCM ::= true|false - PCMK__XA_CRMD ::= online|offline Since crm_feature_set 3.18.0 (pacemaker-2.1.7): - PCMK__XA_IN_CCM ::= |0 Since when node has been a cluster member. A value 0 of means the node is not a cluster member. - PCMK__XA_CRMD ::= |0 Since when peer has been online in CPG. A value 0 means the peer is offline in CPG. */ crm_trace("Node %s member@%lld online@%lld join=%s expected=%s%s", pe__node_name(this_node), when_member, when_online, pcmk__s(join, ""), pcmk__s(exp_state, ""), (termination_requested? " (termination requested)" : "")); if (this_node->details->shutdown) { crm_debug("%s is shutting down", pe__node_name(this_node)); /* Slightly different criteria since we can't shut down a dead peer */ return (when_online > 0); } if (when_member < 0) { pe_fence_node(scheduler, this_node, "peer has not been seen by the cluster", FALSE); return false; } if (pcmk__str_eq(join, CRMD_JOINSTATE_NACK, pcmk__str_none)) { pe_fence_node(scheduler, this_node, "peer failed Pacemaker membership criteria", FALSE); } else if (termination_requested) { if ((when_member <= 0) && (when_online <= 0) && pcmk__str_eq(join, CRMD_JOINSTATE_DOWN, pcmk__str_none)) { crm_info("%s was fenced as requested", pe__node_name(this_node)); return false; } pe_fence_node(scheduler, this_node, "fencing was requested", false); } else if (pcmk__str_eq(exp_state, CRMD_JOINSTATE_DOWN, pcmk__str_null_matches)) { if (pending_too_long(scheduler, this_node, when_member, when_online)) { pe_fence_node(scheduler, this_node, "peer pending timed out on joining the process group", FALSE); } else if ((when_member > 0) || (when_online > 0)) { crm_info("- %s is not ready to run resources", pe__node_name(this_node)); this_node->details->standby = TRUE; this_node->details->pending = TRUE; } else { crm_trace("%s is down or still coming up", pe__node_name(this_node)); } } else if (when_member <= 0) { // Consider `priority-fencing-delay` for lost nodes pe_fence_node(scheduler, this_node, "peer is no longer part of the cluster", TRUE); } else if (when_online <= 0) { pe_fence_node(scheduler, this_node, "peer process is no longer available", FALSE); /* Everything is running at this point, now check join state */ } else if (pcmk__str_eq(join, CRMD_JOINSTATE_MEMBER, pcmk__str_none)) { crm_info("%s is active", pe__node_name(this_node)); } else if (pcmk__str_any_of(join, CRMD_JOINSTATE_PENDING, CRMD_JOINSTATE_DOWN, NULL)) { crm_info("%s is not ready to run resources", pe__node_name(this_node)); this_node->details->standby = TRUE; this_node->details->pending = TRUE; } else { pe_fence_node(scheduler, this_node, "peer was in an unknown state", FALSE); } return (when_member > 0); } static void determine_remote_online_status(pcmk_scheduler_t *scheduler, pcmk_node_t *this_node) { pcmk_resource_t *rsc = this_node->details->remote_rsc; pcmk_resource_t *container = NULL; pcmk_node_t *host = NULL; /* If there is a node state entry for a (former) Pacemaker Remote node * but no resource creating that node, the node's connection resource will * be NULL. Consider it an offline remote node in that case. */ if (rsc == NULL) { this_node->details->online = FALSE; goto remote_online_done; } container = rsc->container; if (container && pcmk__list_of_1(rsc->running_on)) { host = rsc->running_on->data; } /* If the resource is currently started, mark it online. */ if (rsc->role == pcmk_role_started) { crm_trace("%s node %s presumed ONLINE because connection resource is started", (container? "Guest" : "Remote"), this_node->details->id); this_node->details->online = TRUE; } /* consider this node shutting down if transitioning start->stop */ if ((rsc->role == pcmk_role_started) && (rsc->next_role == pcmk_role_stopped)) { crm_trace("%s node %s shutting down because connection resource is stopping", (container? "Guest" : "Remote"), this_node->details->id); this_node->details->shutdown = TRUE; } /* Now check all the failure conditions. */ if(container && pcmk_is_set(container->flags, pcmk_rsc_failed)) { crm_trace("Guest node %s UNCLEAN because guest resource failed", this_node->details->id); this_node->details->online = FALSE; this_node->details->remote_requires_reset = TRUE; } else if (pcmk_is_set(rsc->flags, pcmk_rsc_failed)) { crm_trace("%s node %s OFFLINE because connection resource failed", (container? "Guest" : "Remote"), this_node->details->id); this_node->details->online = FALSE; } else if ((rsc->role == pcmk_role_stopped) || ((container != NULL) && (container->role == pcmk_role_stopped))) { crm_trace("%s node %s OFFLINE because its resource is stopped", (container? "Guest" : "Remote"), this_node->details->id); this_node->details->online = FALSE; this_node->details->remote_requires_reset = FALSE; } else if (host && (host->details->online == FALSE) && host->details->unclean) { crm_trace("Guest node %s UNCLEAN because host is unclean", this_node->details->id); this_node->details->online = FALSE; this_node->details->remote_requires_reset = TRUE; } remote_online_done: crm_trace("Remote node %s online=%s", this_node->details->id, this_node->details->online ? "TRUE" : "FALSE"); } static void determine_online_status(const xmlNode *node_state, pcmk_node_t *this_node, pcmk_scheduler_t *scheduler) { gboolean online = FALSE; const char *exp_state = crm_element_value(node_state, PCMK__XA_EXPECTED); CRM_CHECK(this_node != NULL, return); this_node->details->shutdown = FALSE; this_node->details->expected_up = FALSE; if (pe__shutdown_requested(this_node)) { this_node->details->shutdown = TRUE; } else if (pcmk__str_eq(exp_state, CRMD_JOINSTATE_MEMBER, pcmk__str_casei)) { this_node->details->expected_up = TRUE; } if (this_node->details->type == node_ping) { this_node->details->unclean = FALSE; online = FALSE; /* As far as resource management is concerned, * the node is safely offline. * Anyone caught abusing this logic will be shot */ } else if (!pcmk_is_set(scheduler->flags, pcmk_sched_fencing_enabled)) { online = determine_online_status_no_fencing(scheduler, node_state, this_node); } else { online = determine_online_status_fencing(scheduler, node_state, this_node); } if (online) { this_node->details->online = TRUE; } else { /* remove node from contention */ this_node->fixed = TRUE; // @COMPAT deprecated and unused this_node->weight = -INFINITY; } if (online && this_node->details->shutdown) { /* don't run resources here */ this_node->fixed = TRUE; // @COMPAT deprecated and unused this_node->weight = -INFINITY; } if (this_node->details->type == node_ping) { crm_info("%s is not a Pacemaker node", pe__node_name(this_node)); } else if (this_node->details->unclean) { pe_proc_warn("%s is unclean", pe__node_name(this_node)); } else if (this_node->details->online) { crm_info("%s is %s", pe__node_name(this_node), this_node->details->shutdown ? "shutting down" : this_node->details->pending ? "pending" : this_node->details->standby ? "standby" : this_node->details->maintenance ? "maintenance" : "online"); } else { crm_trace("%s is offline", pe__node_name(this_node)); } } /*! * \internal * \brief Find the end of a resource's name, excluding any clone suffix * * \param[in] id Resource ID to check * * \return Pointer to last character of resource's base name */ const char * pe_base_name_end(const char *id) { if (!pcmk__str_empty(id)) { const char *end = id + strlen(id) - 1; for (const char *s = end; s > id; --s) { switch (*s) { case '0': case '1': case '2': case '3': case '4': case '5': case '6': case '7': case '8': case '9': break; case ':': return (s == end)? s : (s - 1); default: return end; } } return end; } return NULL; } /*! * \internal * \brief Get a resource name excluding any clone suffix * * \param[in] last_rsc_id Resource ID to check * * \return Pointer to newly allocated string with resource's base name * \note It is the caller's responsibility to free() the result. * This asserts on error, so callers can assume result is not NULL. */ char * clone_strip(const char *last_rsc_id) { const char *end = pe_base_name_end(last_rsc_id); char *basename = NULL; CRM_ASSERT(end); basename = strndup(last_rsc_id, end - last_rsc_id + 1); CRM_ASSERT(basename); return basename; } /*! * \internal * \brief Get the name of the first instance of a cloned resource * * \param[in] last_rsc_id Resource ID to check * * \return Pointer to newly allocated string with resource's base name plus :0 * \note It is the caller's responsibility to free() the result. * This asserts on error, so callers can assume result is not NULL. */ char * clone_zero(const char *last_rsc_id) { const char *end = pe_base_name_end(last_rsc_id); size_t base_name_len = end - last_rsc_id + 1; char *zero = NULL; CRM_ASSERT(end); zero = calloc(base_name_len + 3, sizeof(char)); CRM_ASSERT(zero); memcpy(zero, last_rsc_id, base_name_len); zero[base_name_len] = ':'; zero[base_name_len + 1] = '0'; return zero; } static pcmk_resource_t * create_fake_resource(const char *rsc_id, const xmlNode *rsc_entry, pcmk_scheduler_t *scheduler) { pcmk_resource_t *rsc = NULL; xmlNode *xml_rsc = create_xml_node(NULL, XML_CIB_TAG_RESOURCE); copy_in_properties(xml_rsc, rsc_entry); crm_xml_add(xml_rsc, XML_ATTR_ID, rsc_id); crm_log_xml_debug(xml_rsc, "Orphan resource"); if (pe__unpack_resource(xml_rsc, &rsc, NULL, scheduler) != pcmk_rc_ok) { return NULL; } if (xml_contains_remote_node(xml_rsc)) { pcmk_node_t *node; crm_debug("Detected orphaned remote node %s", rsc_id); node = pe_find_node(scheduler->nodes, rsc_id); if (node == NULL) { node = pe_create_node(rsc_id, rsc_id, "remote", NULL, scheduler); } link_rsc2remotenode(scheduler, rsc); if (node) { crm_trace("Setting node %s as shutting down due to orphaned connection resource", rsc_id); node->details->shutdown = TRUE; } } if (crm_element_value(rsc_entry, XML_RSC_ATTR_CONTAINER)) { /* This orphaned rsc needs to be mapped to a container. */ crm_trace("Detected orphaned container filler %s", rsc_id); pe__set_resource_flags(rsc, pcmk_rsc_removed_filler); } pe__set_resource_flags(rsc, pcmk_rsc_removed); scheduler->resources = g_list_append(scheduler->resources, rsc); return rsc; } /*! * \internal * \brief Create orphan instance for anonymous clone resource history * * \param[in,out] parent Clone resource that orphan will be added to * \param[in] rsc_id Orphan's resource ID * \param[in] node Where orphan is active (for logging only) * \param[in,out] scheduler Scheduler data * * \return Newly added orphaned instance of \p parent */ static pcmk_resource_t * create_anonymous_orphan(pcmk_resource_t *parent, const char *rsc_id, const pcmk_node_t *node, pcmk_scheduler_t *scheduler) { pcmk_resource_t *top = pe__create_clone_child(parent, scheduler); // find_rsc() because we might be a cloned group pcmk_resource_t *orphan = top->fns->find_rsc(top, rsc_id, NULL, pcmk_rsc_match_clone_only); - pe_rsc_debug(parent, "Created orphan %s for %s: %s on %s", - top->id, parent->id, rsc_id, pe__node_name(node)); + pcmk__rsc_debug(parent, "Created orphan %s for %s: %s on %s", + top->id, parent->id, rsc_id, pe__node_name(node)); return orphan; } /*! * \internal * \brief Check a node for an instance of an anonymous clone * * Return a child instance of the specified anonymous clone, in order of * preference: (1) the instance running on the specified node, if any; * (2) an inactive instance (i.e. within the total of clone-max instances); * (3) a newly created orphan (i.e. clone-max instances are already active). * * \param[in,out] scheduler Scheduler data * \param[in] node Node on which to check for instance * \param[in,out] parent Clone to check * \param[in] rsc_id Name of cloned resource in history (no instance) */ static pcmk_resource_t * find_anonymous_clone(pcmk_scheduler_t *scheduler, const pcmk_node_t *node, pcmk_resource_t *parent, const char *rsc_id) { GList *rIter = NULL; pcmk_resource_t *rsc = NULL; pcmk_resource_t *inactive_instance = NULL; gboolean skip_inactive = FALSE; CRM_ASSERT(parent != NULL); CRM_ASSERT(pe_rsc_is_clone(parent)); CRM_ASSERT(!pcmk_is_set(parent->flags, pcmk_rsc_unique)); // Check for active (or partially active, for cloned groups) instance pe_rsc_trace(parent, "Looking for %s on %s in %s", rsc_id, pe__node_name(node), parent->id); for (rIter = parent->children; rsc == NULL && rIter; rIter = rIter->next) { GList *locations = NULL; pcmk_resource_t *child = rIter->data; /* Check whether this instance is already known to be active or pending * anywhere, at this stage of unpacking. Because this function is called * for a resource before the resource's individual operation history * entries are unpacked, locations will generally not contain the * desired node. * * However, there are three exceptions: * (1) when child is a cloned group and we have already unpacked the * history of another member of the group on the same node; * (2) when we've already unpacked the history of another numbered * instance on the same node (which can happen if globally-unique * was flipped from true to false); and * (3) when we re-run calculations on the same scheduler data as part of * a simulation. */ child->fns->location(child, &locations, 2); if (locations) { /* We should never associate the same numbered anonymous clone * instance with multiple nodes, and clone instances can't migrate, * so there must be only one location, regardless of history. */ CRM_LOG_ASSERT(locations->next == NULL); if (((pcmk_node_t *) locations->data)->details == node->details) { /* This child instance is active on the requested node, so check * for a corresponding configured resource. We use find_rsc() * instead of child because child may be a cloned group, and we * need the particular member corresponding to rsc_id. * * If the history entry is orphaned, rsc will be NULL. */ rsc = parent->fns->find_rsc(child, rsc_id, NULL, pcmk_rsc_match_clone_only); if (rsc) { /* If there are multiple instance history entries for an * anonymous clone in a single node's history (which can * happen if globally-unique is switched from true to * false), we want to consider the instances beyond the * first as orphans, even if there are inactive instance * numbers available. */ if (rsc->running_on) { crm_notice("Active (now-)anonymous clone %s has " "multiple (orphan) instance histories on %s", parent->id, pe__node_name(node)); skip_inactive = TRUE; rsc = NULL; } else { pe_rsc_trace(parent, "Resource %s, active", rsc->id); } } } g_list_free(locations); } else { pe_rsc_trace(parent, "Resource %s, skip inactive", child->id); if (!skip_inactive && !inactive_instance && !pcmk_is_set(child->flags, pcmk_rsc_blocked)) { // Remember one inactive instance in case we don't find active inactive_instance = parent->fns->find_rsc(child, rsc_id, NULL, pcmk_rsc_match_clone_only); /* ... but don't use it if it was already associated with a * pending action on another node */ if (inactive_instance && inactive_instance->pending_node && (inactive_instance->pending_node->details != node->details)) { inactive_instance = NULL; } } } } if ((rsc == NULL) && !skip_inactive && (inactive_instance != NULL)) { pe_rsc_trace(parent, "Resource %s, empty slot", inactive_instance->id); rsc = inactive_instance; } /* If the resource has "requires" set to "quorum" or "nothing", and we don't * have a clone instance for every node, we don't want to consume a valid * instance number for unclean nodes. Such instances may appear to be active * according to the history, but should be considered inactive, so we can * start an instance elsewhere. Treat such instances as orphans. * * An exception is instances running on guest nodes -- since guest node * "fencing" is actually just a resource stop, requires shouldn't apply. * * @TODO Ideally, we'd use an inactive instance number if it is not needed * for any clean instances. However, we don't know that at this point. */ if ((rsc != NULL) && !pcmk_is_set(rsc->flags, pcmk_rsc_needs_fencing) && (!node->details->online || node->details->unclean) && !pe__is_guest_node(node) && !pe__is_universal_clone(parent, scheduler)) { rsc = NULL; } if (rsc == NULL) { rsc = create_anonymous_orphan(parent, rsc_id, node, scheduler); pe_rsc_trace(parent, "Resource %s, orphan", rsc->id); } return rsc; } static pcmk_resource_t * unpack_find_resource(pcmk_scheduler_t *scheduler, const pcmk_node_t *node, const char *rsc_id) { pcmk_resource_t *rsc = NULL; pcmk_resource_t *parent = NULL; crm_trace("looking for %s", rsc_id); rsc = pe_find_resource(scheduler->resources, rsc_id); if (rsc == NULL) { /* If we didn't find the resource by its name in the operation history, * check it again as a clone instance. Even when clone-max=0, we create * a single :0 orphan to match against here. */ char *clone0_id = clone_zero(rsc_id); pcmk_resource_t *clone0 = pe_find_resource(scheduler->resources, clone0_id); if (clone0 && !pcmk_is_set(clone0->flags, pcmk_rsc_unique)) { rsc = clone0; parent = uber_parent(clone0); crm_trace("%s found as %s (%s)", rsc_id, clone0_id, parent->id); } else { crm_trace("%s is not known as %s either (orphan)", rsc_id, clone0_id); } free(clone0_id); } else if (rsc->variant > pcmk_rsc_variant_primitive) { crm_trace("Resource history for %s is orphaned because it is no longer primitive", rsc_id); return NULL; } else { parent = uber_parent(rsc); } if (pe_rsc_is_anon_clone(parent)) { if (pe_rsc_is_bundled(parent)) { rsc = pe__find_bundle_replica(parent->parent, node); } else { char *base = clone_strip(rsc_id); rsc = find_anonymous_clone(scheduler, node, parent, base); free(base); CRM_ASSERT(rsc != NULL); } } if (rsc && !pcmk__str_eq(rsc_id, rsc->id, pcmk__str_casei) && !pcmk__str_eq(rsc_id, rsc->clone_name, pcmk__str_casei)) { pcmk__str_update(&rsc->clone_name, rsc_id); - pe_rsc_debug(rsc, "Internally renamed %s on %s to %s%s", - rsc_id, pe__node_name(node), rsc->id, - (pcmk_is_set(rsc->flags, pcmk_rsc_removed)? " (ORPHAN)" : "")); + pcmk__rsc_debug(rsc, "Internally renamed %s on %s to %s%s", + rsc_id, pe__node_name(node), rsc->id, + pcmk_is_set(rsc->flags, pcmk_rsc_removed)? " (ORPHAN)" : ""); } return rsc; } static pcmk_resource_t * process_orphan_resource(const xmlNode *rsc_entry, const pcmk_node_t *node, pcmk_scheduler_t *scheduler) { pcmk_resource_t *rsc = NULL; const char *rsc_id = crm_element_value(rsc_entry, XML_ATTR_ID); crm_debug("Detected orphan resource %s on %s", rsc_id, pe__node_name(node)); rsc = create_fake_resource(rsc_id, rsc_entry, scheduler); if (rsc == NULL) { return NULL; } if (!pcmk_is_set(scheduler->flags, pcmk_sched_stop_removed_resources)) { pe__clear_resource_flags(rsc, pcmk_rsc_managed); } else { CRM_CHECK(rsc != NULL, return NULL); pe_rsc_trace(rsc, "Added orphan %s", rsc->id); resource_location(rsc, NULL, -INFINITY, "__orphan_do_not_run__", scheduler); } return rsc; } static void process_rsc_state(pcmk_resource_t *rsc, pcmk_node_t *node, enum action_fail_response on_fail) { pcmk_node_t *tmpnode = NULL; char *reason = NULL; enum action_fail_response save_on_fail = pcmk_on_fail_ignore; CRM_ASSERT(rsc); pe_rsc_trace(rsc, "Resource %s is %s on %s: on_fail=%s", rsc->id, role2text(rsc->role), pe__node_name(node), fail2text(on_fail)); /* process current state */ if (rsc->role != pcmk_role_unknown) { pcmk_resource_t *iter = rsc; while (iter) { if (g_hash_table_lookup(iter->known_on, node->details->id) == NULL) { pcmk_node_t *n = pe__copy_node(node); pe_rsc_trace(rsc, "%s%s%s known on %s", rsc->id, ((rsc->clone_name == NULL)? "" : " also known as "), ((rsc->clone_name == NULL)? "" : rsc->clone_name), pe__node_name(n)); g_hash_table_insert(iter->known_on, (gpointer) n->details->id, n); } if (pcmk_is_set(iter->flags, pcmk_rsc_unique)) { break; } iter = iter->parent; } } /* If a managed resource is believed to be running, but node is down ... */ if ((rsc->role > pcmk_role_stopped) && node->details->online == FALSE && node->details->maintenance == FALSE && pcmk_is_set(rsc->flags, pcmk_rsc_managed)) { gboolean should_fence = FALSE; /* If this is a guest node, fence it (regardless of whether fencing is * enabled, because guest node fencing is done by recovery of the * container resource rather than by the fencer). Mark the resource * we're processing as failed. When the guest comes back up, its * operation history in the CIB will be cleared, freeing the affected * resource to run again once we are sure we know its state. */ if (pe__is_guest_node(node)) { pe__set_resource_flags(rsc, pcmk_rsc_failed|pcmk_rsc_stop_if_failed); should_fence = TRUE; } else if (pcmk_is_set(rsc->cluster->flags, pcmk_sched_fencing_enabled)) { if (pe__is_remote_node(node) && node->details->remote_rsc && !pcmk_is_set(node->details->remote_rsc->flags, pcmk_rsc_failed)) { /* Setting unseen means that fencing of the remote node will * occur only if the connection resource is not going to start * somewhere. This allows connection resources on a failed * cluster node to move to another node without requiring the * remote nodes to be fenced as well. */ node->details->unseen = TRUE; reason = crm_strdup_printf("%s is active there (fencing will be" " revoked if remote connection can " "be re-established elsewhere)", rsc->id); } should_fence = TRUE; } if (should_fence) { if (reason == NULL) { reason = crm_strdup_printf("%s is thought to be active there", rsc->id); } pe_fence_node(rsc->cluster, node, reason, FALSE); } free(reason); } /* In order to calculate priority_fencing_delay correctly, save the failure information and pass it to native_add_running(). */ save_on_fail = on_fail; if (node->details->unclean) { /* No extra processing needed * Also allows resources to be started again after a node is shot */ on_fail = pcmk_on_fail_ignore; } switch (on_fail) { case pcmk_on_fail_ignore: /* nothing to do */ break; case pcmk_on_fail_demote: pe__set_resource_flags(rsc, pcmk_rsc_failed); demote_action(rsc, node, FALSE); break; case pcmk_on_fail_fence_node: /* treat it as if it is still running * but also mark the node as unclean */ reason = crm_strdup_printf("%s failed there", rsc->id); pe_fence_node(rsc->cluster, node, reason, FALSE); free(reason); break; case pcmk_on_fail_standby_node: node->details->standby = TRUE; node->details->standby_onfail = TRUE; break; case pcmk_on_fail_block: /* is_managed == FALSE will prevent any * actions being sent for the resource */ pe__clear_resource_flags(rsc, pcmk_rsc_managed); pe__set_resource_flags(rsc, pcmk_rsc_blocked); break; case pcmk_on_fail_ban: /* make sure it comes up somewhere else * or not at all */ resource_location(rsc, node, -INFINITY, "__action_migration_auto__", rsc->cluster); break; case pcmk_on_fail_stop: pe__set_next_role(rsc, pcmk_role_stopped, "on-fail=stop"); break; case pcmk_on_fail_restart: if ((rsc->role != pcmk_role_stopped) && (rsc->role != pcmk_role_unknown)) { pe__set_resource_flags(rsc, pcmk_rsc_failed|pcmk_rsc_stop_if_failed); stop_action(rsc, node, FALSE); } break; case pcmk_on_fail_restart_container: pe__set_resource_flags(rsc, pcmk_rsc_failed|pcmk_rsc_stop_if_failed); if (rsc->container && pe_rsc_is_bundled(rsc)) { /* A bundle's remote connection can run on a different node than * the bundle's container. We don't necessarily know where the * container is running yet, so remember it and add a stop * action for it later. */ rsc->cluster->stop_needed = g_list_prepend(rsc->cluster->stop_needed, rsc->container); } else if (rsc->container) { stop_action(rsc->container, node, FALSE); } else if ((rsc->role != pcmk_role_stopped) && (rsc->role != pcmk_role_unknown)) { stop_action(rsc, node, FALSE); } break; case pcmk_on_fail_reset_remote: pe__set_resource_flags(rsc, pcmk_rsc_failed|pcmk_rsc_stop_if_failed); if (pcmk_is_set(rsc->cluster->flags, pcmk_sched_fencing_enabled)) { tmpnode = NULL; if (rsc->is_remote_node) { tmpnode = pe_find_node(rsc->cluster->nodes, rsc->id); } if (tmpnode && pe__is_remote_node(tmpnode) && tmpnode->details->remote_was_fenced == 0) { /* The remote connection resource failed in a way that * should result in fencing the remote node. */ pe_fence_node(rsc->cluster, tmpnode, "remote connection is unrecoverable", FALSE); } } /* require the stop action regardless if fencing is occurring or not. */ if (rsc->role > pcmk_role_stopped) { stop_action(rsc, node, FALSE); } /* if reconnect delay is in use, prevent the connection from exiting the * "STOPPED" role until the failure is cleared by the delay timeout. */ if (rsc->remote_reconnect_ms) { pe__set_next_role(rsc, pcmk_role_stopped, "remote reset"); } break; } /* ensure a remote-node connection failure forces an unclean remote-node * to be fenced. By setting unseen = FALSE, the remote-node failure will * result in a fencing operation regardless if we're going to attempt to * reconnect to the remote-node in this transition or not. */ if (pcmk_is_set(rsc->flags, pcmk_rsc_failed) && rsc->is_remote_node) { tmpnode = pe_find_node(rsc->cluster->nodes, rsc->id); if (tmpnode && tmpnode->details->unclean) { tmpnode->details->unseen = FALSE; } } if ((rsc->role != pcmk_role_stopped) && (rsc->role != pcmk_role_unknown)) { if (pcmk_is_set(rsc->flags, pcmk_rsc_removed)) { if (pcmk_is_set(rsc->flags, pcmk_rsc_managed)) { pcmk__config_warn("Detected active orphan %s running on %s", rsc->id, pe__node_name(node)); } else { pcmk__config_warn("Resource '%s' must be stopped manually on " "%s because cluster is configured not to " "stop active orphans", rsc->id, pe__node_name(node)); } } native_add_running(rsc, node, rsc->cluster, (save_on_fail != pcmk_on_fail_ignore)); switch (on_fail) { case pcmk_on_fail_ignore: break; case pcmk_on_fail_demote: case pcmk_on_fail_block: pe__set_resource_flags(rsc, pcmk_rsc_failed); break; default: pe__set_resource_flags(rsc, pcmk_rsc_failed|pcmk_rsc_stop_if_failed); break; } } else if (rsc->clone_name && strchr(rsc->clone_name, ':') != NULL) { /* Only do this for older status sections that included instance numbers * Otherwise stopped instances will appear as orphans */ pe_rsc_trace(rsc, "Resetting clone_name %s for %s (stopped)", rsc->clone_name, rsc->id); free(rsc->clone_name); rsc->clone_name = NULL; } else { GList *possible_matches = pe__resource_actions(rsc, node, PCMK_ACTION_STOP, FALSE); GList *gIter = possible_matches; for (; gIter != NULL; gIter = gIter->next) { pcmk_action_t *stop = (pcmk_action_t *) gIter->data; pe__set_action_flags(stop, pcmk_action_optional); } g_list_free(possible_matches); } /* A successful stop after migrate_to on the migration source doesn't make * the partially migrated resource stopped on the migration target. */ if ((rsc->role == pcmk_role_stopped) && rsc->partial_migration_source && rsc->partial_migration_source->details == node->details && rsc->partial_migration_target && rsc->running_on) { rsc->role = pcmk_role_started; } } /* create active recurring operations as optional */ static void process_recurring(pcmk_node_t *node, pcmk_resource_t *rsc, int start_index, int stop_index, GList *sorted_op_list, pcmk_scheduler_t *scheduler) { int counter = -1; const char *task = NULL; const char *status = NULL; GList *gIter = sorted_op_list; CRM_ASSERT(rsc); pe_rsc_trace(rsc, "%s: Start index %d, stop index = %d", rsc->id, start_index, stop_index); for (; gIter != NULL; gIter = gIter->next) { xmlNode *rsc_op = (xmlNode *) gIter->data; guint interval_ms = 0; char *key = NULL; const char *id = ID(rsc_op); counter++; if (node->details->online == FALSE) { pe_rsc_trace(rsc, "Skipping %s on %s: node is offline", rsc->id, pe__node_name(node)); break; /* Need to check if there's a monitor for role="Stopped" */ } else if (start_index < stop_index && counter <= stop_index) { pe_rsc_trace(rsc, "Skipping %s on %s: resource is not active", id, pe__node_name(node)); continue; } else if (counter < start_index) { pe_rsc_trace(rsc, "Skipping %s on %s: old %d", id, pe__node_name(node), counter); continue; } crm_element_value_ms(rsc_op, XML_LRM_ATTR_INTERVAL_MS, &interval_ms); if (interval_ms == 0) { pe_rsc_trace(rsc, "Skipping %s on %s: non-recurring", id, pe__node_name(node)); continue; } status = crm_element_value(rsc_op, XML_LRM_ATTR_OPSTATUS); if (pcmk__str_eq(status, "-1", pcmk__str_casei)) { pe_rsc_trace(rsc, "Skipping %s on %s: status", id, pe__node_name(node)); continue; } task = crm_element_value(rsc_op, XML_LRM_ATTR_TASK); /* create the action */ key = pcmk__op_key(rsc->id, task, interval_ms); pe_rsc_trace(rsc, "Creating %s on %s", key, pe__node_name(node)); custom_action(rsc, key, task, node, TRUE, scheduler); } } void calculate_active_ops(const GList *sorted_op_list, int *start_index, int *stop_index) { int counter = -1; int implied_monitor_start = -1; int implied_clone_start = -1; const char *task = NULL; const char *status = NULL; *stop_index = -1; *start_index = -1; for (const GList *iter = sorted_op_list; iter != NULL; iter = iter->next) { const xmlNode *rsc_op = (const xmlNode *) iter->data; counter++; task = crm_element_value(rsc_op, XML_LRM_ATTR_TASK); status = crm_element_value(rsc_op, XML_LRM_ATTR_OPSTATUS); if (pcmk__str_eq(task, PCMK_ACTION_STOP, pcmk__str_casei) && pcmk__str_eq(status, "0", pcmk__str_casei)) { *stop_index = counter; } else if (pcmk__strcase_any_of(task, PCMK_ACTION_START, PCMK_ACTION_MIGRATE_FROM, NULL)) { *start_index = counter; } else if ((implied_monitor_start <= *stop_index) && pcmk__str_eq(task, PCMK_ACTION_MONITOR, pcmk__str_casei)) { const char *rc = crm_element_value(rsc_op, XML_LRM_ATTR_RC); if (pcmk__strcase_any_of(rc, "0", "8", NULL)) { implied_monitor_start = counter; } } else if (pcmk__strcase_any_of(task, PCMK_ACTION_PROMOTE, PCMK_ACTION_DEMOTE, NULL)) { implied_clone_start = counter; } } if (*start_index == -1) { if (implied_clone_start != -1) { *start_index = implied_clone_start; } else if (implied_monitor_start != -1) { *start_index = implied_monitor_start; } } } // If resource history entry has shutdown lock, remember lock node and time static void unpack_shutdown_lock(const xmlNode *rsc_entry, pcmk_resource_t *rsc, const pcmk_node_t *node, pcmk_scheduler_t *scheduler) { time_t lock_time = 0; // When lock started (i.e. node shutdown time) if ((crm_element_value_epoch(rsc_entry, XML_CONFIG_ATTR_SHUTDOWN_LOCK, &lock_time) == pcmk_ok) && (lock_time != 0)) { if ((scheduler->shutdown_lock > 0) && (get_effective_time(scheduler) > (lock_time + scheduler->shutdown_lock))) { pcmk__rsc_info(rsc, "Shutdown lock for %s on %s expired", rsc->id, pe__node_name(node)); pe__clear_resource_history(rsc, node); } else { /* @COMPAT I don't like breaking const signatures, but * rsc->lock_node should really be const -- we just can't change it * until the next API compatibility break. */ rsc->lock_node = (pcmk_node_t *) node; rsc->lock_time = lock_time; } } } /*! * \internal * \brief Unpack one lrm_resource entry from a node's CIB status * * \param[in,out] node Node whose status is being unpacked * \param[in] rsc_entry lrm_resource XML being unpacked * \param[in,out] scheduler Scheduler data * * \return Resource corresponding to the entry, or NULL if no operation history */ static pcmk_resource_t * unpack_lrm_resource(pcmk_node_t *node, const xmlNode *lrm_resource, pcmk_scheduler_t *scheduler) { GList *gIter = NULL; int stop_index = -1; int start_index = -1; enum rsc_role_e req_role = pcmk_role_unknown; const char *rsc_id = ID(lrm_resource); pcmk_resource_t *rsc = NULL; GList *op_list = NULL; GList *sorted_op_list = NULL; xmlNode *rsc_op = NULL; xmlNode *last_failure = NULL; enum action_fail_response on_fail = pcmk_on_fail_ignore; enum rsc_role_e saved_role = pcmk_role_unknown; if (rsc_id == NULL) { crm_warn("Ignoring malformed " XML_LRM_TAG_RESOURCE " entry without id"); return NULL; } crm_trace("Unpacking " XML_LRM_TAG_RESOURCE " for %s on %s", rsc_id, pe__node_name(node)); // Build a list of individual lrm_rsc_op entries, so we can sort them for (rsc_op = first_named_child(lrm_resource, XML_LRM_TAG_RSC_OP); rsc_op != NULL; rsc_op = crm_next_same_xml(rsc_op)) { op_list = g_list_prepend(op_list, rsc_op); } if (!pcmk_is_set(scheduler->flags, pcmk_sched_shutdown_lock)) { if (op_list == NULL) { // If there are no operations, there is nothing to do return NULL; } } /* find the resource */ rsc = unpack_find_resource(scheduler, node, rsc_id); if (rsc == NULL) { if (op_list == NULL) { // If there are no operations, there is nothing to do return NULL; } else { rsc = process_orphan_resource(lrm_resource, node, scheduler); } } CRM_ASSERT(rsc != NULL); // Check whether the resource is "shutdown-locked" to this node if (pcmk_is_set(scheduler->flags, pcmk_sched_shutdown_lock)) { unpack_shutdown_lock(lrm_resource, rsc, node, scheduler); } /* process operations */ saved_role = rsc->role; rsc->role = pcmk_role_unknown; sorted_op_list = g_list_sort(op_list, sort_op_by_callid); for (gIter = sorted_op_list; gIter != NULL; gIter = gIter->next) { xmlNode *rsc_op = (xmlNode *) gIter->data; unpack_rsc_op(rsc, node, rsc_op, &last_failure, &on_fail); } /* create active recurring operations as optional */ calculate_active_ops(sorted_op_list, &start_index, &stop_index); process_recurring(node, rsc, start_index, stop_index, sorted_op_list, scheduler); /* no need to free the contents */ g_list_free(sorted_op_list); process_rsc_state(rsc, node, on_fail); if (get_target_role(rsc, &req_role)) { if ((rsc->next_role == pcmk_role_unknown) || (req_role < rsc->next_role)) { pe__set_next_role(rsc, req_role, XML_RSC_ATTR_TARGET_ROLE); } else if (req_role > rsc->next_role) { pcmk__rsc_info(rsc, "%s: Not overwriting calculated next role %s" " with requested next role %s", rsc->id, role2text(rsc->next_role), role2text(req_role)); } } if (saved_role > rsc->role) { rsc->role = saved_role; } return rsc; } static void handle_orphaned_container_fillers(const xmlNode *lrm_rsc_list, pcmk_scheduler_t *scheduler) { for (const xmlNode *rsc_entry = pcmk__xe_first_child(lrm_rsc_list); rsc_entry != NULL; rsc_entry = pcmk__xe_next(rsc_entry)) { pcmk_resource_t *rsc; pcmk_resource_t *container; const char *rsc_id; const char *container_id; if (!pcmk__str_eq((const char *)rsc_entry->name, XML_LRM_TAG_RESOURCE, pcmk__str_casei)) { continue; } container_id = crm_element_value(rsc_entry, XML_RSC_ATTR_CONTAINER); rsc_id = crm_element_value(rsc_entry, XML_ATTR_ID); if (container_id == NULL || rsc_id == NULL) { continue; } container = pe_find_resource(scheduler->resources, container_id); if (container == NULL) { continue; } rsc = pe_find_resource(scheduler->resources, rsc_id); if ((rsc == NULL) || (rsc->container != NULL) || !pcmk_is_set(rsc->flags, pcmk_rsc_removed_filler)) { continue; } pe_rsc_trace(rsc, "Mapped container of orphaned resource %s to %s", rsc->id, container_id); rsc->container = container; container->fillers = g_list_append(container->fillers, rsc); } } /*! * \internal * \brief Unpack one node's lrm status section * * \param[in,out] node Node whose status is being unpacked * \param[in] xml CIB node state XML * \param[in,out] scheduler Scheduler data */ static void unpack_node_lrm(pcmk_node_t *node, const xmlNode *xml, pcmk_scheduler_t *scheduler) { bool found_orphaned_container_filler = false; // Drill down to lrm_resources section xml = find_xml_node(xml, XML_CIB_TAG_LRM, FALSE); if (xml == NULL) { return; } xml = find_xml_node(xml, XML_LRM_TAG_RESOURCES, FALSE); if (xml == NULL) { return; } // Unpack each lrm_resource entry for (const xmlNode *rsc_entry = first_named_child(xml, XML_LRM_TAG_RESOURCE); rsc_entry != NULL; rsc_entry = crm_next_same_xml(rsc_entry)) { pcmk_resource_t *rsc = unpack_lrm_resource(node, rsc_entry, scheduler); if ((rsc != NULL) && pcmk_is_set(rsc->flags, pcmk_rsc_removed_filler)) { found_orphaned_container_filler = true; } } /* Now that all resource state has been unpacked for this node, map any * orphaned container fillers to their container resource. */ if (found_orphaned_container_filler) { handle_orphaned_container_fillers(xml, scheduler); } } static void set_active(pcmk_resource_t *rsc) { const pcmk_resource_t *top = pe__const_top_resource(rsc, false); if (top && pcmk_is_set(top->flags, pcmk_rsc_promotable)) { rsc->role = pcmk_role_unpromoted; } else { rsc->role = pcmk_role_started; } } static void set_node_score(gpointer key, gpointer value, gpointer user_data) { pcmk_node_t *node = value; int *score = user_data; node->weight = *score; } #define XPATH_NODE_STATE "/" XML_TAG_CIB "/" XML_CIB_TAG_STATUS \ "/" XML_CIB_TAG_STATE #define SUB_XPATH_LRM_RESOURCE "/" XML_CIB_TAG_LRM \ "/" XML_LRM_TAG_RESOURCES \ "/" XML_LRM_TAG_RESOURCE #define SUB_XPATH_LRM_RSC_OP "/" XML_LRM_TAG_RSC_OP static xmlNode * find_lrm_op(const char *resource, const char *op, const char *node, const char *source, int target_rc, pcmk_scheduler_t *scheduler) { GString *xpath = NULL; xmlNode *xml = NULL; CRM_CHECK((resource != NULL) && (op != NULL) && (node != NULL), return NULL); xpath = g_string_sized_new(256); pcmk__g_strcat(xpath, XPATH_NODE_STATE "[@" XML_ATTR_UNAME "='", node, "']" SUB_XPATH_LRM_RESOURCE "[@" XML_ATTR_ID "='", resource, "']" SUB_XPATH_LRM_RSC_OP "[@" XML_LRM_ATTR_TASK "='", op, "'", NULL); /* Need to check against transition_magic too? */ if ((source != NULL) && (strcmp(op, PCMK_ACTION_MIGRATE_TO) == 0)) { pcmk__g_strcat(xpath, " and @" XML_LRM_ATTR_MIGRATE_TARGET "='", source, "']", NULL); } else if ((source != NULL) && (strcmp(op, PCMK_ACTION_MIGRATE_FROM) == 0)) { pcmk__g_strcat(xpath, " and @" XML_LRM_ATTR_MIGRATE_SOURCE "='", source, "']", NULL); } else { g_string_append_c(xpath, ']'); } xml = get_xpath_object((const char *) xpath->str, scheduler->input, LOG_DEBUG); g_string_free(xpath, TRUE); if (xml && target_rc >= 0) { int rc = PCMK_OCF_UNKNOWN_ERROR; int status = PCMK_EXEC_ERROR; crm_element_value_int(xml, XML_LRM_ATTR_RC, &rc); crm_element_value_int(xml, XML_LRM_ATTR_OPSTATUS, &status); if ((rc != target_rc) || (status != PCMK_EXEC_DONE)) { return NULL; } } return xml; } static xmlNode * find_lrm_resource(const char *rsc_id, const char *node_name, pcmk_scheduler_t *scheduler) { GString *xpath = NULL; xmlNode *xml = NULL; CRM_CHECK((rsc_id != NULL) && (node_name != NULL), return NULL); xpath = g_string_sized_new(256); pcmk__g_strcat(xpath, XPATH_NODE_STATE "[@" XML_ATTR_UNAME "='", node_name, "']" SUB_XPATH_LRM_RESOURCE "[@" XML_ATTR_ID "='", rsc_id, "']", NULL); xml = get_xpath_object((const char *) xpath->str, scheduler->input, LOG_DEBUG); g_string_free(xpath, TRUE); return xml; } /*! * \internal * \brief Check whether a resource has no completed action history on a node * * \param[in,out] rsc Resource to check * \param[in] node_name Node to check * * \return true if \p rsc_id is unknown on \p node_name, otherwise false */ static bool unknown_on_node(pcmk_resource_t *rsc, const char *node_name) { bool result = false; xmlXPathObjectPtr search; GString *xpath = g_string_sized_new(256); pcmk__g_strcat(xpath, XPATH_NODE_STATE "[@" XML_ATTR_UNAME "='", node_name, "']" SUB_XPATH_LRM_RESOURCE "[@" XML_ATTR_ID "='", rsc->id, "']" SUB_XPATH_LRM_RSC_OP "[@" XML_LRM_ATTR_RC "!='193']", NULL); search = xpath_search(rsc->cluster->input, (const char *) xpath->str); result = (numXpathResults(search) == 0); freeXpathObject(search); g_string_free(xpath, TRUE); return result; } /*! * \brief Check whether a probe/monitor indicating the resource was not running * on a node happened after some event * * \param[in] rsc_id Resource being checked * \param[in] node_name Node being checked * \param[in] xml_op Event that monitor is being compared to * \param[in] same_node Whether the operations are on the same node * \param[in,out] scheduler Scheduler data * * \return true if such a monitor happened after event, false otherwise */ static bool monitor_not_running_after(const char *rsc_id, const char *node_name, const xmlNode *xml_op, bool same_node, pcmk_scheduler_t *scheduler) { /* Any probe/monitor operation on the node indicating it was not running * there */ xmlNode *monitor = find_lrm_op(rsc_id, PCMK_ACTION_MONITOR, node_name, NULL, PCMK_OCF_NOT_RUNNING, scheduler); return (monitor && pe__is_newer_op(monitor, xml_op, same_node) > 0); } /*! * \brief Check whether any non-monitor operation on a node happened after some * event * * \param[in] rsc_id Resource being checked * \param[in] node_name Node being checked * \param[in] xml_op Event that non-monitor is being compared to * \param[in] same_node Whether the operations are on the same node * \param[in,out] scheduler Scheduler data * * \return true if such a operation happened after event, false otherwise */ static bool non_monitor_after(const char *rsc_id, const char *node_name, const xmlNode *xml_op, bool same_node, pcmk_scheduler_t *scheduler) { xmlNode *lrm_resource = NULL; lrm_resource = find_lrm_resource(rsc_id, node_name, scheduler); if (lrm_resource == NULL) { return false; } for (xmlNode *op = first_named_child(lrm_resource, XML_LRM_TAG_RSC_OP); op != NULL; op = crm_next_same_xml(op)) { const char * task = NULL; if (op == xml_op) { continue; } task = crm_element_value(op, XML_LRM_ATTR_TASK); if (pcmk__str_any_of(task, PCMK_ACTION_START, PCMK_ACTION_STOP, PCMK_ACTION_MIGRATE_TO, PCMK_ACTION_MIGRATE_FROM, NULL) && pe__is_newer_op(op, xml_op, same_node) > 0) { return true; } } return false; } /*! * \brief Check whether the resource has newer state on a node after a migration * attempt * * \param[in] rsc_id Resource being checked * \param[in] node_name Node being checked * \param[in] migrate_to Any migrate_to event that is being compared to * \param[in] migrate_from Any migrate_from event that is being compared to * \param[in,out] scheduler Scheduler data * * \return true if such a operation happened after event, false otherwise */ static bool newer_state_after_migrate(const char *rsc_id, const char *node_name, const xmlNode *migrate_to, const xmlNode *migrate_from, pcmk_scheduler_t *scheduler) { const xmlNode *xml_op = migrate_to; const char *source = NULL; const char *target = NULL; bool same_node = false; if (migrate_from) { xml_op = migrate_from; } source = crm_element_value(xml_op, XML_LRM_ATTR_MIGRATE_SOURCE); target = crm_element_value(xml_op, XML_LRM_ATTR_MIGRATE_TARGET); /* It's preferred to compare to the migrate event on the same node if * existing, since call ids are more reliable. */ if (pcmk__str_eq(node_name, target, pcmk__str_casei)) { if (migrate_from) { xml_op = migrate_from; same_node = true; } else { xml_op = migrate_to; } } else if (pcmk__str_eq(node_name, source, pcmk__str_casei)) { if (migrate_to) { xml_op = migrate_to; same_node = true; } else { xml_op = migrate_from; } } /* If there's any newer non-monitor operation on the node, or any newer * probe/monitor operation on the node indicating it was not running there, * the migration events potentially no longer matter for the node. */ return non_monitor_after(rsc_id, node_name, xml_op, same_node, scheduler) || monitor_not_running_after(rsc_id, node_name, xml_op, same_node, scheduler); } /*! * \internal * \brief Parse migration source and target node names from history entry * * \param[in] entry Resource history entry for a migration action * \param[in] source_node If not NULL, source must match this node * \param[in] target_node If not NULL, target must match this node * \param[out] source_name Where to store migration source node name * \param[out] target_name Where to store migration target node name * * \return Standard Pacemaker return code */ static int get_migration_node_names(const xmlNode *entry, const pcmk_node_t *source_node, const pcmk_node_t *target_node, const char **source_name, const char **target_name) { *source_name = crm_element_value(entry, XML_LRM_ATTR_MIGRATE_SOURCE); *target_name = crm_element_value(entry, XML_LRM_ATTR_MIGRATE_TARGET); if ((*source_name == NULL) || (*target_name == NULL)) { crm_err("Ignoring resource history entry %s without " XML_LRM_ATTR_MIGRATE_SOURCE " and " XML_LRM_ATTR_MIGRATE_TARGET, ID(entry)); return pcmk_rc_unpack_error; } if ((source_node != NULL) && !pcmk__str_eq(*source_name, source_node->details->uname, pcmk__str_casei|pcmk__str_null_matches)) { crm_err("Ignoring resource history entry %s because " XML_LRM_ATTR_MIGRATE_SOURCE "='%s' does not match %s", ID(entry), *source_name, pe__node_name(source_node)); return pcmk_rc_unpack_error; } if ((target_node != NULL) && !pcmk__str_eq(*target_name, target_node->details->uname, pcmk__str_casei|pcmk__str_null_matches)) { crm_err("Ignoring resource history entry %s because " XML_LRM_ATTR_MIGRATE_TARGET "='%s' does not match %s", ID(entry), *target_name, pe__node_name(target_node)); return pcmk_rc_unpack_error; } return pcmk_rc_ok; } /* * \internal * \brief Add a migration source to a resource's list of dangling migrations * * If the migrate_to and migrate_from actions in a live migration both * succeeded, but there is no stop on the source, the migration is considered * "dangling." Add the source to the resource's dangling migration list, which * will be used to schedule a stop on the source without affecting the target. * * \param[in,out] rsc Resource involved in migration * \param[in] node Migration source */ static void add_dangling_migration(pcmk_resource_t *rsc, const pcmk_node_t *node) { pe_rsc_trace(rsc, "Dangling migration of %s requires stop on %s", rsc->id, pe__node_name(node)); rsc->role = pcmk_role_stopped; rsc->dangling_migrations = g_list_prepend(rsc->dangling_migrations, (gpointer) node); } /*! * \internal * \brief Update resource role etc. after a successful migrate_to action * * \param[in,out] history Parsed action result history */ static void unpack_migrate_to_success(struct action_history *history) { /* A complete migration sequence is: * 1. migrate_to on source node (which succeeded if we get to this function) * 2. migrate_from on target node * 3. stop on source node * * If no migrate_from has happened, the migration is considered to be * "partial". If the migrate_from succeeded but no stop has happened, the * migration is considered to be "dangling". * * If a successful migrate_to and stop have happened on the source node, we * still need to check for a partial migration, due to scenarios (easier to * produce with batch-limit=1) like: * * - A resource is migrating from node1 to node2, and a migrate_to is * initiated for it on node1. * * - node2 goes into standby mode while the migrate_to is pending, which * aborts the transition. * * - Upon completion of the migrate_to, a new transition schedules a stop * on both nodes and a start on node1. * * - If the new transition is aborted for any reason while the resource is * stopping on node1, the transition after that stop completes will see * the migrate_to and stop on the source, but it's still a partial * migration, and the resource must be stopped on node2 because it is * potentially active there due to the migrate_to. * * We also need to take into account that either node's history may be * cleared at any point in the migration process. */ int from_rc = PCMK_OCF_OK; int from_status = PCMK_EXEC_PENDING; pcmk_node_t *target_node = NULL; xmlNode *migrate_from = NULL; const char *source = NULL; const char *target = NULL; bool source_newer_op = false; bool target_newer_state = false; bool active_on_target = false; // Get source and target node names from XML if (get_migration_node_names(history->xml, history->node, NULL, &source, &target) != pcmk_rc_ok) { return; } // Check for newer state on the source source_newer_op = non_monitor_after(history->rsc->id, source, history->xml, true, history->rsc->cluster); // Check for a migrate_from action from this source on the target migrate_from = find_lrm_op(history->rsc->id, PCMK_ACTION_MIGRATE_FROM, target, source, -1, history->rsc->cluster); if (migrate_from != NULL) { if (source_newer_op) { /* There's a newer non-monitor operation on the source and a * migrate_from on the target, so this migrate_to is irrelevant to * the resource's state. */ return; } crm_element_value_int(migrate_from, XML_LRM_ATTR_RC, &from_rc); crm_element_value_int(migrate_from, XML_LRM_ATTR_OPSTATUS, &from_status); } /* If the resource has newer state on both the source and target after the * migration events, this migrate_to is irrelevant to the resource's state. */ target_newer_state = newer_state_after_migrate(history->rsc->id, target, history->xml, migrate_from, history->rsc->cluster); if (source_newer_op && target_newer_state) { return; } /* Check for dangling migration (migrate_from succeeded but stop not done). * We know there's no stop because we already returned if the target has a * migrate_from and the source has any newer non-monitor operation. */ if ((from_rc == PCMK_OCF_OK) && (from_status == PCMK_EXEC_DONE)) { add_dangling_migration(history->rsc, history->node); return; } /* Without newer state, this migrate_to implies the resource is active. * (Clones are not allowed to migrate, so role can't be promoted.) */ history->rsc->role = pcmk_role_started; target_node = pe_find_node(history->rsc->cluster->nodes, target); active_on_target = !target_newer_state && (target_node != NULL) && target_node->details->online; if (from_status != PCMK_EXEC_PENDING) { // migrate_from failed on target if (active_on_target) { native_add_running(history->rsc, target_node, history->rsc->cluster, TRUE); } else { // Mark resource as failed, require recovery, and prevent migration pe__set_resource_flags(history->rsc, pcmk_rsc_failed|pcmk_rsc_stop_if_failed); pe__clear_resource_flags(history->rsc, pcmk_rsc_migratable); } return; } // The migrate_from is pending, complete but erased, or to be scheduled /* If there is no history at all for the resource on an online target, then * it was likely cleaned. Just return, and we'll schedule a probe. Once we * have the probe result, it will be reflected in target_newer_state. */ if ((target_node != NULL) && target_node->details->online && unknown_on_node(history->rsc, target)) { return; } if (active_on_target) { pcmk_node_t *source_node = pe_find_node(history->rsc->cluster->nodes, source); native_add_running(history->rsc, target_node, history->rsc->cluster, FALSE); if ((source_node != NULL) && source_node->details->online) { /* This is a partial migration: the migrate_to completed * successfully on the source, but the migrate_from has not * completed. Remember the source and target; if the newly * chosen target remains the same when we schedule actions * later, we may continue with the migration. */ history->rsc->partial_migration_target = target_node; history->rsc->partial_migration_source = source_node; } } else if (!source_newer_op) { // Mark resource as failed, require recovery, and prevent migration pe__set_resource_flags(history->rsc, pcmk_rsc_failed|pcmk_rsc_stop_if_failed); pe__clear_resource_flags(history->rsc, pcmk_rsc_migratable); } } /*! * \internal * \brief Update resource role etc. after a failed migrate_to action * * \param[in,out] history Parsed action result history */ static void unpack_migrate_to_failure(struct action_history *history) { xmlNode *target_migrate_from = NULL; const char *source = NULL; const char *target = NULL; // Get source and target node names from XML if (get_migration_node_names(history->xml, history->node, NULL, &source, &target) != pcmk_rc_ok) { return; } /* If a migration failed, we have to assume the resource is active. Clones * are not allowed to migrate, so role can't be promoted. */ history->rsc->role = pcmk_role_started; // Check for migrate_from on the target target_migrate_from = find_lrm_op(history->rsc->id, PCMK_ACTION_MIGRATE_FROM, target, source, PCMK_OCF_OK, history->rsc->cluster); if (/* If the resource state is unknown on the target, it will likely be * probed there. * Don't just consider it running there. We will get back here anyway in * case the probe detects it's running there. */ !unknown_on_node(history->rsc, target) /* If the resource has newer state on the target after the migration * events, this migrate_to no longer matters for the target. */ && !newer_state_after_migrate(history->rsc->id, target, history->xml, target_migrate_from, history->rsc->cluster)) { /* The resource has no newer state on the target, so assume it's still * active there. * (if it is up). */ pcmk_node_t *target_node = pe_find_node(history->rsc->cluster->nodes, target); if (target_node && target_node->details->online) { native_add_running(history->rsc, target_node, history->rsc->cluster, FALSE); } } else if (!non_monitor_after(history->rsc->id, source, history->xml, true, history->rsc->cluster)) { /* We know the resource has newer state on the target, but this * migrate_to still matters for the source as long as there's no newer * non-monitor operation there. */ // Mark node as having dangling migration so we can force a stop later history->rsc->dangling_migrations = g_list_prepend(history->rsc->dangling_migrations, (gpointer) history->node); } } /*! * \internal * \brief Update resource role etc. after a failed migrate_from action * * \param[in,out] history Parsed action result history */ static void unpack_migrate_from_failure(struct action_history *history) { xmlNode *source_migrate_to = NULL; const char *source = NULL; const char *target = NULL; // Get source and target node names from XML if (get_migration_node_names(history->xml, NULL, history->node, &source, &target) != pcmk_rc_ok) { return; } /* If a migration failed, we have to assume the resource is active. Clones * are not allowed to migrate, so role can't be promoted. */ history->rsc->role = pcmk_role_started; // Check for a migrate_to on the source source_migrate_to = find_lrm_op(history->rsc->id, PCMK_ACTION_MIGRATE_TO, source, target, PCMK_OCF_OK, history->rsc->cluster); if (/* If the resource state is unknown on the source, it will likely be * probed there. * Don't just consider it running there. We will get back here anyway in * case the probe detects it's running there. */ !unknown_on_node(history->rsc, source) /* If the resource has newer state on the source after the migration * events, this migrate_from no longer matters for the source. */ && !newer_state_after_migrate(history->rsc->id, source, source_migrate_to, history->xml, history->rsc->cluster)) { /* The resource has no newer state on the source, so assume it's still * active there (if it is up). */ pcmk_node_t *source_node = pe_find_node(history->rsc->cluster->nodes, source); if (source_node && source_node->details->online) { native_add_running(history->rsc, source_node, history->rsc->cluster, TRUE); } } } /*! * \internal * \brief Add an action to cluster's list of failed actions * * \param[in,out] history Parsed action result history */ static void record_failed_op(struct action_history *history) { if (!(history->node->details->online)) { return; } for (const xmlNode *xIter = history->rsc->cluster->failed->children; xIter != NULL; xIter = xIter->next) { const char *key = pe__xe_history_key(xIter); const char *uname = crm_element_value(xIter, XML_ATTR_UNAME); if (pcmk__str_eq(history->key, key, pcmk__str_none) && pcmk__str_eq(uname, history->node->details->uname, pcmk__str_casei)) { crm_trace("Skipping duplicate entry %s on %s", history->key, pe__node_name(history->node)); return; } } crm_trace("Adding entry for %s on %s to failed action list", history->key, pe__node_name(history->node)); crm_xml_add(history->xml, XML_ATTR_UNAME, history->node->details->uname); crm_xml_add(history->xml, XML_LRM_ATTR_RSCID, history->rsc->id); add_node_copy(history->rsc->cluster->failed, history->xml); } static char * last_change_str(const xmlNode *xml_op) { time_t when; char *result = NULL; if (crm_element_value_epoch(xml_op, XML_RSC_OP_LAST_CHANGE, &when) == pcmk_ok) { char *when_s = pcmk__epoch2str(&when, 0); const char *p = strchr(when_s, ' '); // Skip day of week to make message shorter if ((p != NULL) && (*(++p) != '\0')) { result = strdup(p); CRM_ASSERT(result != NULL); } free(when_s); } if (result == NULL) { result = strdup("unknown time"); CRM_ASSERT(result != NULL); } return result; } /*! * \internal * \brief Compare two on-fail values * * \param[in] first One on-fail value to compare * \param[in] second The other on-fail value to compare * * \return A negative number if second is more severe than first, zero if they * are equal, or a positive number if first is more severe than second. * \note This is only needed until the action_fail_response values can be * renumbered at the next API compatibility break. */ static int cmp_on_fail(enum action_fail_response first, enum action_fail_response second) { switch (first) { case pcmk_on_fail_demote: switch (second) { case pcmk_on_fail_ignore: return 1; case pcmk_on_fail_demote: return 0; default: return -1; } break; case pcmk_on_fail_reset_remote: switch (second) { case pcmk_on_fail_ignore: case pcmk_on_fail_demote: case pcmk_on_fail_restart: return 1; case pcmk_on_fail_reset_remote: return 0; default: return -1; } break; case pcmk_on_fail_restart_container: switch (second) { case pcmk_on_fail_ignore: case pcmk_on_fail_demote: case pcmk_on_fail_restart: case pcmk_on_fail_reset_remote: return 1; case pcmk_on_fail_restart_container: return 0; default: return -1; } break; default: break; } switch (second) { case pcmk_on_fail_demote: return (first == pcmk_on_fail_ignore)? -1 : 1; case pcmk_on_fail_reset_remote: switch (first) { case pcmk_on_fail_ignore: case pcmk_on_fail_demote: case pcmk_on_fail_restart: return -1; default: return 1; } break; case pcmk_on_fail_restart_container: switch (first) { case pcmk_on_fail_ignore: case pcmk_on_fail_demote: case pcmk_on_fail_restart: case pcmk_on_fail_reset_remote: return -1; default: return 1; } break; default: break; } return first - second; } /*! * \internal * \brief Ban a resource (or its clone if an anonymous instance) from all nodes * * \param[in,out] rsc Resource to ban */ static void ban_from_all_nodes(pcmk_resource_t *rsc) { int score = -INFINITY; pcmk_resource_t *fail_rsc = rsc; if (fail_rsc->parent != NULL) { pcmk_resource_t *parent = uber_parent(fail_rsc); if (pe_rsc_is_anon_clone(parent)) { /* For anonymous clones, if an operation with on-fail=stop fails for * any instance, the entire clone must stop. */ fail_rsc = parent; } } // Ban the resource from all nodes crm_notice("%s will not be started under current conditions", fail_rsc->id); if (fail_rsc->allowed_nodes != NULL) { g_hash_table_destroy(fail_rsc->allowed_nodes); } fail_rsc->allowed_nodes = pe__node_list2table(rsc->cluster->nodes); g_hash_table_foreach(fail_rsc->allowed_nodes, set_node_score, &score); } /*! * \internal * \brief Get configured failure handling and role after failure for an action * * \param[in,out] history Unpacked action history entry * \param[out] on_fail Where to set configured failure handling * \param[out] fail_role Where to set to role after failure */ static void unpack_failure_handling(struct action_history *history, enum action_fail_response *on_fail, enum rsc_role_e *fail_role) { xmlNode *config = pcmk__find_action_config(history->rsc, history->task, history->interval_ms, true); GHashTable *meta = pcmk__unpack_action_meta(history->rsc, history->node, history->task, history->interval_ms, config); const char *on_fail_str = g_hash_table_lookup(meta, XML_OP_ATTR_ON_FAIL); *on_fail = pcmk__parse_on_fail(history->rsc, history->task, history->interval_ms, on_fail_str); *fail_role = pcmk__role_after_failure(history->rsc, history->task, *on_fail, meta); g_hash_table_destroy(meta); } /*! * \internal * \brief Update resource role, failure handling, etc., after a failed action * * \param[in,out] history Parsed action result history * \param[in] config_on_fail Action failure handling from configuration * \param[in] fail_role Resource's role after failure of this action * \param[out] last_failure This will be set to the history XML * \param[in,out] on_fail Actual handling of action result */ static void unpack_rsc_op_failure(struct action_history *history, enum action_fail_response config_on_fail, enum rsc_role_e fail_role, xmlNode **last_failure, enum action_fail_response *on_fail) { bool is_probe = false; char *last_change_s = NULL; *last_failure = history->xml; is_probe = pcmk_xe_is_probe(history->xml); last_change_s = last_change_str(history->xml); if (!pcmk_is_set(history->rsc->cluster->flags, pcmk_sched_symmetric_cluster) && (history->exit_status == PCMK_OCF_NOT_INSTALLED)) { crm_trace("Unexpected result (%s%s%s) was recorded for " "%s of %s on %s at %s " CRM_XS " exit-status=%d id=%s", services_ocf_exitcode_str(history->exit_status), (pcmk__str_empty(history->exit_reason)? "" : ": "), pcmk__s(history->exit_reason, ""), (is_probe? "probe" : history->task), history->rsc->id, pe__node_name(history->node), last_change_s, history->exit_status, history->id); } else { crm_warn("Unexpected result (%s%s%s) was recorded for " "%s of %s on %s at %s " CRM_XS " exit-status=%d id=%s", services_ocf_exitcode_str(history->exit_status), (pcmk__str_empty(history->exit_reason)? "" : ": "), pcmk__s(history->exit_reason, ""), (is_probe? "probe" : history->task), history->rsc->id, pe__node_name(history->node), last_change_s, history->exit_status, history->id); if (is_probe && (history->exit_status != PCMK_OCF_OK) && (history->exit_status != PCMK_OCF_NOT_RUNNING) && (history->exit_status != PCMK_OCF_RUNNING_PROMOTED)) { /* A failed (not just unexpected) probe result could mean the user * didn't know resources will be probed even where they can't run. */ crm_notice("If it is not possible for %s to run on %s, see " "the resource-discovery option for location constraints", history->rsc->id, pe__node_name(history->node)); } record_failed_op(history); } free(last_change_s); if (cmp_on_fail(*on_fail, config_on_fail) < 0) { pe_rsc_trace(history->rsc, "on-fail %s -> %s for %s", fail2text(*on_fail), fail2text(config_on_fail), history->key); *on_fail = config_on_fail; } if (strcmp(history->task, PCMK_ACTION_STOP) == 0) { resource_location(history->rsc, history->node, -INFINITY, "__stop_fail__", history->rsc->cluster); } else if (strcmp(history->task, PCMK_ACTION_MIGRATE_TO) == 0) { unpack_migrate_to_failure(history); } else if (strcmp(history->task, PCMK_ACTION_MIGRATE_FROM) == 0) { unpack_migrate_from_failure(history); } else if (strcmp(history->task, PCMK_ACTION_PROMOTE) == 0) { history->rsc->role = pcmk_role_promoted; } else if (strcmp(history->task, PCMK_ACTION_DEMOTE) == 0) { if (config_on_fail == pcmk_on_fail_block) { history->rsc->role = pcmk_role_promoted; pe__set_next_role(history->rsc, pcmk_role_stopped, "demote with on-fail=block"); } else if (history->exit_status == PCMK_OCF_NOT_RUNNING) { history->rsc->role = pcmk_role_stopped; } else { /* Staying in the promoted role would put the scheduler and * controller into a loop. Setting the role to unpromoted is not * dangerous because the resource will be stopped as part of * recovery, and any promotion will be ordered after that stop. */ history->rsc->role = pcmk_role_unpromoted; } } if (is_probe && (history->exit_status == PCMK_OCF_NOT_INSTALLED)) { /* leave stopped */ pe_rsc_trace(history->rsc, "Leaving %s stopped", history->rsc->id); history->rsc->role = pcmk_role_stopped; } else if (history->rsc->role < pcmk_role_started) { pe_rsc_trace(history->rsc, "Setting %s active", history->rsc->id); set_active(history->rsc); } pe_rsc_trace(history->rsc, "Resource %s: role=%s, unclean=%s, on_fail=%s, fail_role=%s", history->rsc->id, role2text(history->rsc->role), pcmk__btoa(history->node->details->unclean), fail2text(config_on_fail), role2text(fail_role)); if ((fail_role != pcmk_role_started) && (history->rsc->next_role < fail_role)) { pe__set_next_role(history->rsc, fail_role, "failure"); } if (fail_role == pcmk_role_stopped) { ban_from_all_nodes(history->rsc); } } /*! * \internal * \brief Block a resource with a failed action if it cannot be recovered * * If resource action is a failed stop and fencing is not possible, mark the * resource as unmanaged and blocked, since recovery cannot be done. * * \param[in,out] history Parsed action history entry */ static void block_if_unrecoverable(struct action_history *history) { char *last_change_s = NULL; if (strcmp(history->task, PCMK_ACTION_STOP) != 0) { return; // All actions besides stop are always recoverable } if (pe_can_fence(history->node->details->data_set, history->node)) { return; // Failed stops are recoverable via fencing } last_change_s = last_change_str(history->xml); pe_proc_err("No further recovery can be attempted for %s " "because %s on %s failed (%s%s%s) at %s " CRM_XS " rc=%d id=%s", history->rsc->id, history->task, pe__node_name(history->node), services_ocf_exitcode_str(history->exit_status), (pcmk__str_empty(history->exit_reason)? "" : ": "), pcmk__s(history->exit_reason, ""), last_change_s, history->exit_status, history->id); free(last_change_s); pe__clear_resource_flags(history->rsc, pcmk_rsc_managed); pe__set_resource_flags(history->rsc, pcmk_rsc_blocked); } /*! * \internal * \brief Update action history's execution status and why * * \param[in,out] history Parsed action history entry * \param[out] why Where to store reason for update * \param[in] value New value * \param[in] reason Description of why value was changed */ static inline void remap_because(struct action_history *history, const char **why, int value, const char *reason) { if (history->execution_status != value) { history->execution_status = value; *why = reason; } } /*! * \internal * \brief Remap informational monitor results and operation status * * For the monitor results, certain OCF codes are for providing extended information * to the user about services that aren't yet failed but not entirely healthy either. * These must be treated as the "normal" result by Pacemaker. * * For operation status, the action result can be used to determine an appropriate * status for the purposes of responding to the action. The status provided by the * executor is not directly usable since the executor does not know what was expected. * * \param[in,out] history Parsed action history entry * \param[in,out] on_fail What should be done about the result * \param[in] expired Whether result is expired * * \note If the result is remapped and the node is not shutting down or failed, * the operation will be recorded in the scheduler data's list of failed * operations to highlight it for the user. * * \note This may update the resource's current and next role. */ static void remap_operation(struct action_history *history, enum action_fail_response *on_fail, bool expired) { bool is_probe = false; int orig_exit_status = history->exit_status; int orig_exec_status = history->execution_status; const char *why = NULL; const char *task = history->task; // Remap degraded results to their successful counterparts history->exit_status = pcmk__effective_rc(history->exit_status); if (history->exit_status != orig_exit_status) { why = "degraded result"; if (!expired && (!history->node->details->shutdown || history->node->details->online)) { record_failed_op(history); } } if (!pe_rsc_is_bundled(history->rsc) && pcmk_xe_mask_probe_failure(history->xml) && ((history->execution_status != PCMK_EXEC_DONE) || (history->exit_status != PCMK_OCF_NOT_RUNNING))) { history->execution_status = PCMK_EXEC_DONE; history->exit_status = PCMK_OCF_NOT_RUNNING; why = "equivalent probe result"; } /* If the executor reported an execution status of anything but done or * error, consider that final. But for done or error, we know better whether * it should be treated as a failure or not, because we know the expected * result. */ switch (history->execution_status) { case PCMK_EXEC_DONE: case PCMK_EXEC_ERROR: break; // These should be treated as node-fatal case PCMK_EXEC_NO_FENCE_DEVICE: case PCMK_EXEC_NO_SECRETS: remap_because(history, &why, PCMK_EXEC_ERROR_HARD, "node-fatal error"); goto remap_done; default: goto remap_done; } is_probe = pcmk_xe_is_probe(history->xml); if (is_probe) { task = "probe"; } if (history->expected_exit_status < 0) { /* Pre-1.0 Pacemaker versions, and Pacemaker 1.1.6 or earlier with * Heartbeat 2.0.7 or earlier as the cluster layer, did not include the * expected exit status in the transition key, which (along with the * similar case of a corrupted transition key in the CIB) will be * reported to this function as -1. Pacemaker 2.0+ does not support * rolling upgrades from those versions or processing of saved CIB files * from those versions, so we do not need to care much about this case. */ remap_because(history, &why, PCMK_EXEC_ERROR, "obsolete history format"); crm_warn("Expected result not found for %s on %s " "(corrupt or obsolete CIB?)", history->key, pe__node_name(history->node)); } else if (history->exit_status == history->expected_exit_status) { remap_because(history, &why, PCMK_EXEC_DONE, "expected result"); } else { remap_because(history, &why, PCMK_EXEC_ERROR, "unexpected result"); - pe_rsc_debug(history->rsc, - "%s on %s: expected %d (%s), got %d (%s%s%s)", - history->key, pe__node_name(history->node), - history->expected_exit_status, - services_ocf_exitcode_str(history->expected_exit_status), - history->exit_status, - services_ocf_exitcode_str(history->exit_status), - (pcmk__str_empty(history->exit_reason)? "" : ": "), - pcmk__s(history->exit_reason, "")); + pcmk__rsc_debug(history->rsc, + "%s on %s: expected %d (%s), got %d (%s%s%s)", + history->key, pe__node_name(history->node), + history->expected_exit_status, + services_ocf_exitcode_str(history->expected_exit_status), + history->exit_status, + services_ocf_exitcode_str(history->exit_status), + (pcmk__str_empty(history->exit_reason)? "" : ": "), + pcmk__s(history->exit_reason, "")); } switch (history->exit_status) { case PCMK_OCF_OK: if (is_probe && (history->expected_exit_status == PCMK_OCF_NOT_RUNNING)) { char *last_change_s = last_change_str(history->xml); remap_because(history, &why, PCMK_EXEC_DONE, "probe"); pcmk__rsc_info(history->rsc, "Probe found %s active on %s at %s", history->rsc->id, pe__node_name(history->node), last_change_s); free(last_change_s); } break; case PCMK_OCF_NOT_RUNNING: if (is_probe || (history->expected_exit_status == history->exit_status) || !pcmk_is_set(history->rsc->flags, pcmk_rsc_managed)) { /* For probes, recurring monitors for the Stopped role, and * unmanaged resources, "not running" is not considered a * failure. */ remap_because(history, &why, PCMK_EXEC_DONE, "exit status"); history->rsc->role = pcmk_role_stopped; *on_fail = pcmk_on_fail_ignore; pe__set_next_role(history->rsc, pcmk_role_unknown, "not running"); } break; case PCMK_OCF_RUNNING_PROMOTED: if (is_probe && (history->exit_status != history->expected_exit_status)) { char *last_change_s = last_change_str(history->xml); remap_because(history, &why, PCMK_EXEC_DONE, "probe"); pcmk__rsc_info(history->rsc, "Probe found %s active and promoted on %s at %s", history->rsc->id, pe__node_name(history->node), last_change_s); free(last_change_s); } if (!expired || (history->exit_status == history->expected_exit_status)) { history->rsc->role = pcmk_role_promoted; } break; case PCMK_OCF_FAILED_PROMOTED: if (!expired) { history->rsc->role = pcmk_role_promoted; } remap_because(history, &why, PCMK_EXEC_ERROR, "exit status"); break; case PCMK_OCF_NOT_CONFIGURED: remap_because(history, &why, PCMK_EXEC_ERROR_FATAL, "exit status"); break; case PCMK_OCF_UNIMPLEMENT_FEATURE: { guint interval_ms = 0; crm_element_value_ms(history->xml, XML_LRM_ATTR_INTERVAL_MS, &interval_ms); if (interval_ms == 0) { if (!expired) { block_if_unrecoverable(history); } remap_because(history, &why, PCMK_EXEC_ERROR_HARD, "exit status"); } else { remap_because(history, &why, PCMK_EXEC_NOT_SUPPORTED, "exit status"); } } break; case PCMK_OCF_NOT_INSTALLED: case PCMK_OCF_INVALID_PARAM: case PCMK_OCF_INSUFFICIENT_PRIV: if (!expired) { block_if_unrecoverable(history); } remap_because(history, &why, PCMK_EXEC_ERROR_HARD, "exit status"); break; default: if (history->execution_status == PCMK_EXEC_DONE) { char *last_change_s = last_change_str(history->xml); crm_info("Treating unknown exit status %d from %s of %s " "on %s at %s as failure", history->exit_status, task, history->rsc->id, pe__node_name(history->node), last_change_s); remap_because(history, &why, PCMK_EXEC_ERROR, "unknown exit status"); free(last_change_s); } break; } remap_done: if (why != NULL) { pe_rsc_trace(history->rsc, "Remapped %s result from [%s: %s] to [%s: %s] " "because of %s", history->key, pcmk_exec_status_str(orig_exec_status), crm_exit_str(orig_exit_status), pcmk_exec_status_str(history->execution_status), crm_exit_str(history->exit_status), why); } } // return TRUE if start or monitor last failure but parameters changed static bool should_clear_for_param_change(const xmlNode *xml_op, const char *task, pcmk_resource_t *rsc, pcmk_node_t *node) { if (pcmk__str_any_of(task, PCMK_ACTION_START, PCMK_ACTION_MONITOR, NULL)) { if (pe__bundle_needs_remote_name(rsc)) { /* We haven't allocated resources yet, so we can't reliably * substitute addr parameters for the REMOTE_CONTAINER_HACK. * When that's needed, defer the check until later. */ pe__add_param_check(xml_op, rsc, node, pcmk__check_last_failure, rsc->cluster); } else { pcmk__op_digest_t *digest_data = NULL; digest_data = rsc_action_digest_cmp(rsc, xml_op, node, rsc->cluster); switch (digest_data->rc) { case pcmk__digest_unknown: crm_trace("Resource %s history entry %s on %s" " has no digest to compare", rsc->id, pe__xe_history_key(xml_op), node->details->id); break; case pcmk__digest_match: break; default: return TRUE; } } } return FALSE; } // Order action after fencing of remote node, given connection rsc static void order_after_remote_fencing(pcmk_action_t *action, pcmk_resource_t *remote_conn, pcmk_scheduler_t *scheduler) { pcmk_node_t *remote_node = pe_find_node(scheduler->nodes, remote_conn->id); if (remote_node) { pcmk_action_t *fence = pe_fence_op(remote_node, NULL, TRUE, NULL, FALSE, scheduler); order_actions(fence, action, pcmk__ar_first_implies_then); } } static bool should_ignore_failure_timeout(const pcmk_resource_t *rsc, const char *task, guint interval_ms, bool is_last_failure) { /* Clearing failures of recurring monitors has special concerns. The * executor reports only changes in the monitor result, so if the * monitor is still active and still getting the same failure result, * that will go undetected after the failure is cleared. * * Also, the operation history will have the time when the recurring * monitor result changed to the given code, not the time when the * result last happened. * * @TODO We probably should clear such failures only when the failure * timeout has passed since the last occurrence of the failed result. * However we don't record that information. We could maybe approximate * that by clearing only if there is a more recent successful monitor or * stop result, but we don't even have that information at this point * since we are still unpacking the resource's operation history. * * This is especially important for remote connection resources with a * reconnect interval, so in that case, we skip clearing failures * if the remote node hasn't been fenced. */ if (rsc->remote_reconnect_ms && pcmk_is_set(rsc->cluster->flags, pcmk_sched_fencing_enabled) && (interval_ms != 0) && pcmk__str_eq(task, PCMK_ACTION_MONITOR, pcmk__str_casei)) { pcmk_node_t *remote_node = pe_find_node(rsc->cluster->nodes, rsc->id); if (remote_node && !remote_node->details->remote_was_fenced) { if (is_last_failure) { crm_info("Waiting to clear monitor failure for remote node %s" " until fencing has occurred", rsc->id); } return TRUE; } } return FALSE; } /*! * \internal * \brief Check operation age and schedule failure clearing when appropriate * * This function has two distinct purposes. The first is to check whether an * operation history entry is expired (i.e. the resource has a failure timeout, * the entry is older than the timeout, and the resource either has no fail * count or its fail count is entirely older than the timeout). The second is to * schedule fail count clearing when appropriate (i.e. the operation is expired * and either the resource has an expired fail count or the operation is a * last_failure for a remote connection resource with a reconnect interval, * or the operation is a last_failure for a start or monitor operation and the * resource's parameters have changed since the operation). * * \param[in,out] history Parsed action result history * * \return true if operation history entry is expired, otherwise false */ static bool check_operation_expiry(struct action_history *history) { bool expired = false; bool is_last_failure = pcmk__ends_with(history->id, "_last_failure_0"); time_t last_run = 0; int unexpired_fail_count = 0; const char *clear_reason = NULL; if (history->execution_status == PCMK_EXEC_NOT_INSTALLED) { pe_rsc_trace(history->rsc, "Resource history entry %s on %s is not expired: " "Not Installed does not expire", history->id, pe__node_name(history->node)); return false; // "Not installed" must always be cleared manually } if ((history->rsc->failure_timeout > 0) && (crm_element_value_epoch(history->xml, XML_RSC_OP_LAST_CHANGE, &last_run) == 0)) { // Resource has a failure-timeout, and history entry has a timestamp time_t now = get_effective_time(history->rsc->cluster); time_t last_failure = 0; // Is this particular operation history older than the failure timeout? if ((now >= (last_run + history->rsc->failure_timeout)) && !should_ignore_failure_timeout(history->rsc, history->task, history->interval_ms, is_last_failure)) { expired = true; } // Does the resource as a whole have an unexpired fail count? unexpired_fail_count = pe_get_failcount(history->node, history->rsc, &last_failure, pcmk__fc_effective, history->xml); // Update scheduler recheck time according to *last* failure crm_trace("%s@%lld is %sexpired @%lld with unexpired_failures=%d timeout=%ds" " last-failure@%lld", history->id, (long long) last_run, (expired? "" : "not "), (long long) now, unexpired_fail_count, history->rsc->failure_timeout, (long long) last_failure); last_failure += history->rsc->failure_timeout + 1; if (unexpired_fail_count && (now < last_failure)) { pe__update_recheck_time(last_failure, history->rsc->cluster, "fail count expiration"); } } if (expired) { if (pe_get_failcount(history->node, history->rsc, NULL, pcmk__fc_default, history->xml)) { // There is a fail count ignoring timeout if (unexpired_fail_count == 0) { // There is no fail count considering timeout clear_reason = "it expired"; } else { /* This operation is old, but there is an unexpired fail count. * In a properly functioning cluster, this should only be * possible if this operation is not a failure (otherwise the * fail count should be expired too), so this is really just a * failsafe. */ pe_rsc_trace(history->rsc, "Resource history entry %s on %s is not expired: " "Unexpired fail count", history->id, pe__node_name(history->node)); expired = false; } } else if (is_last_failure && (history->rsc->remote_reconnect_ms != 0)) { /* Clear any expired last failure when reconnect interval is set, * even if there is no fail count. */ clear_reason = "reconnect interval is set"; } } if (!expired && is_last_failure && should_clear_for_param_change(history->xml, history->task, history->rsc, history->node)) { clear_reason = "resource parameters have changed"; } if (clear_reason != NULL) { pcmk_action_t *clear_op = NULL; // Schedule clearing of the fail count clear_op = pe__clear_failcount(history->rsc, history->node, clear_reason, history->rsc->cluster); if (pcmk_is_set(history->rsc->cluster->flags, pcmk_sched_fencing_enabled) && (history->rsc->remote_reconnect_ms != 0)) { /* If we're clearing a remote connection due to a reconnect * interval, we want to wait until any scheduled fencing * completes. * * We could limit this to remote_node->details->unclean, but at * this point, that's always true (it won't be reliable until * after unpack_node_history() is done). */ crm_info("Clearing %s failure will wait until any scheduled " "fencing of %s completes", history->task, history->rsc->id); order_after_remote_fencing(clear_op, history->rsc, history->rsc->cluster); } } if (expired && (history->interval_ms == 0) && pcmk__str_eq(history->task, PCMK_ACTION_MONITOR, pcmk__str_none)) { switch (history->exit_status) { case PCMK_OCF_OK: case PCMK_OCF_NOT_RUNNING: case PCMK_OCF_RUNNING_PROMOTED: case PCMK_OCF_DEGRADED: case PCMK_OCF_DEGRADED_PROMOTED: // Don't expire probes that return these values pe_rsc_trace(history->rsc, "Resource history entry %s on %s is not expired: " "Probe result", history->id, pe__node_name(history->node)); expired = false; break; } } return expired; } int pe__target_rc_from_xml(const xmlNode *xml_op) { int target_rc = 0; const char *key = crm_element_value(xml_op, XML_ATTR_TRANSITION_KEY); if (key == NULL) { return -1; } decode_transition_key(key, NULL, NULL, NULL, &target_rc); return target_rc; } /*! * \internal * \brief Update a resource's state for an action result * * \param[in,out] history Parsed action history entry * \param[in] exit_status Exit status to base new state on * \param[in] last_failure Resource's last_failure entry, if known * \param[in,out] on_fail Resource's current failure handling */ static void update_resource_state(struct action_history *history, int exit_status, const xmlNode *last_failure, enum action_fail_response *on_fail) { bool clear_past_failure = false; if ((exit_status == PCMK_OCF_NOT_INSTALLED) || (!pe_rsc_is_bundled(history->rsc) && pcmk_xe_mask_probe_failure(history->xml))) { history->rsc->role = pcmk_role_stopped; } else if (exit_status == PCMK_OCF_NOT_RUNNING) { clear_past_failure = true; } else if (pcmk__str_eq(history->task, PCMK_ACTION_MONITOR, pcmk__str_none)) { if ((last_failure != NULL) && pcmk__str_eq(history->key, pe__xe_history_key(last_failure), pcmk__str_none)) { clear_past_failure = true; } if (history->rsc->role < pcmk_role_started) { set_active(history->rsc); } } else if (pcmk__str_eq(history->task, PCMK_ACTION_START, pcmk__str_none)) { history->rsc->role = pcmk_role_started; clear_past_failure = true; } else if (pcmk__str_eq(history->task, PCMK_ACTION_STOP, pcmk__str_none)) { history->rsc->role = pcmk_role_stopped; clear_past_failure = true; } else if (pcmk__str_eq(history->task, PCMK_ACTION_PROMOTE, pcmk__str_none)) { history->rsc->role = pcmk_role_promoted; clear_past_failure = true; } else if (pcmk__str_eq(history->task, PCMK_ACTION_DEMOTE, pcmk__str_none)) { if (*on_fail == pcmk_on_fail_demote) { // Demote clears an error only if on-fail=demote clear_past_failure = true; } history->rsc->role = pcmk_role_unpromoted; } else if (pcmk__str_eq(history->task, PCMK_ACTION_MIGRATE_FROM, pcmk__str_none)) { history->rsc->role = pcmk_role_started; clear_past_failure = true; } else if (pcmk__str_eq(history->task, PCMK_ACTION_MIGRATE_TO, pcmk__str_none)) { unpack_migrate_to_success(history); } else if (history->rsc->role < pcmk_role_started) { pe_rsc_trace(history->rsc, "%s active on %s", history->rsc->id, pe__node_name(history->node)); set_active(history->rsc); } if (!clear_past_failure) { return; } switch (*on_fail) { case pcmk_on_fail_stop: case pcmk_on_fail_ban: case pcmk_on_fail_standby_node: case pcmk_on_fail_fence_node: pe_rsc_trace(history->rsc, "%s (%s) is not cleared by a completed %s", history->rsc->id, fail2text(*on_fail), history->task); break; case pcmk_on_fail_block: case pcmk_on_fail_ignore: case pcmk_on_fail_demote: case pcmk_on_fail_restart: case pcmk_on_fail_restart_container: *on_fail = pcmk_on_fail_ignore; pe__set_next_role(history->rsc, pcmk_role_unknown, "clear past failures"); break; case pcmk_on_fail_reset_remote: if (history->rsc->remote_reconnect_ms == 0) { /* With no reconnect interval, the connection is allowed to * start again after the remote node is fenced and * completely stopped. (With a reconnect interval, we wait * for the failure to be cleared entirely before attempting * to reconnect.) */ *on_fail = pcmk_on_fail_ignore; pe__set_next_role(history->rsc, pcmk_role_unknown, "clear past failures and reset remote"); } break; } } /*! * \internal * \brief Check whether a given history entry matters for resource state * * \param[in] history Parsed action history entry * * \return true if action can affect resource state, otherwise false */ static inline bool can_affect_state(struct action_history *history) { #if 0 /* @COMPAT It might be better to parse only actions we know we're interested * in, rather than exclude a couple we don't. However that would be a * behavioral change that should be done at a major or minor series release. * Currently, unknown operations can affect whether a resource is considered * active and/or failed. */ return pcmk__str_any_of(history->task, PCMK_ACTION_MONITOR, PCMK_ACTION_START, PCMK_ACTION_STOP, PCMK_ACTION_PROMOTE, PCMK_ACTION_DEMOTE, PCMK_ACTION_MIGRATE_TO, PCMK_ACTION_MIGRATE_FROM, "asyncmon", NULL); #else return !pcmk__str_any_of(history->task, PCMK_ACTION_NOTIFY, PCMK_ACTION_META_DATA, NULL); #endif } /*! * \internal * \brief Unpack execution/exit status and exit reason from a history entry * * \param[in,out] history Action history entry to unpack * * \return Standard Pacemaker return code */ static int unpack_action_result(struct action_history *history) { if ((crm_element_value_int(history->xml, XML_LRM_ATTR_OPSTATUS, &(history->execution_status)) < 0) || (history->execution_status < PCMK_EXEC_PENDING) || (history->execution_status > PCMK_EXEC_MAX) || (history->execution_status == PCMK_EXEC_CANCELLED)) { crm_err("Ignoring resource history entry %s for %s on %s " "with invalid " XML_LRM_ATTR_OPSTATUS " '%s'", history->id, history->rsc->id, pe__node_name(history->node), pcmk__s(crm_element_value(history->xml, XML_LRM_ATTR_OPSTATUS), "")); return pcmk_rc_unpack_error; } if ((crm_element_value_int(history->xml, XML_LRM_ATTR_RC, &(history->exit_status)) < 0) || (history->exit_status < 0) || (history->exit_status > CRM_EX_MAX)) { #if 0 /* @COMPAT We should ignore malformed entries, but since that would * change behavior, it should be done at a major or minor series * release. */ crm_err("Ignoring resource history entry %s for %s on %s " "with invalid " XML_LRM_ATTR_RC " '%s'", history->id, history->rsc->id, pe__node_name(history->node), pcmk__s(crm_element_value(history->xml, XML_LRM_ATTR_RC), "")); return pcmk_rc_unpack_error; #else history->exit_status = CRM_EX_ERROR; #endif } history->exit_reason = crm_element_value(history->xml, XML_LRM_ATTR_EXIT_REASON); return pcmk_rc_ok; } /*! * \internal * \brief Process an action history entry whose result expired * * \param[in,out] history Parsed action history entry * \param[in] orig_exit_status Action exit status before remapping * * \return Standard Pacemaker return code (in particular, pcmk_rc_ok means the * entry needs no further processing) */ static int process_expired_result(struct action_history *history, int orig_exit_status) { if (!pe_rsc_is_bundled(history->rsc) && pcmk_xe_mask_probe_failure(history->xml) && (orig_exit_status != history->expected_exit_status)) { if (history->rsc->role <= pcmk_role_stopped) { history->rsc->role = pcmk_role_unknown; } crm_trace("Ignoring resource history entry %s for probe of %s on %s: " "Masked failure expired", history->id, history->rsc->id, pe__node_name(history->node)); return pcmk_rc_ok; } if (history->exit_status == history->expected_exit_status) { return pcmk_rc_undetermined; // Only failures expire } if (history->interval_ms == 0) { crm_notice("Ignoring resource history entry %s for %s of %s on %s: " "Expired failure", history->id, history->task, history->rsc->id, pe__node_name(history->node)); return pcmk_rc_ok; } if (history->node->details->online && !history->node->details->unclean) { /* Reschedule the recurring action. schedule_cancel() won't work at * this stage, so as a hacky workaround, forcibly change the restart * digest so pcmk__check_action_config() does what we want later. * * @TODO We should skip this if there is a newer successful monitor. * Also, this causes rescheduling only if the history entry * has an op-digest (which the expire-non-blocked-failure * scheduler regression test doesn't, but that may not be a * realistic scenario in production). */ crm_notice("Rescheduling %s-interval %s of %s on %s " "after failure expired", pcmk__readable_interval(history->interval_ms), history->task, history->rsc->id, pe__node_name(history->node)); crm_xml_add(history->xml, XML_LRM_ATTR_RESTART_DIGEST, "calculated-failure-timeout"); return pcmk_rc_ok; } return pcmk_rc_undetermined; } /*! * \internal * \brief Process a masked probe failure * * \param[in,out] history Parsed action history entry * \param[in] orig_exit_status Action exit status before remapping * \param[in] last_failure Resource's last_failure entry, if known * \param[in,out] on_fail Resource's current failure handling */ static void mask_probe_failure(struct action_history *history, int orig_exit_status, const xmlNode *last_failure, enum action_fail_response *on_fail) { pcmk_resource_t *ban_rsc = history->rsc; if (!pcmk_is_set(history->rsc->flags, pcmk_rsc_unique)) { ban_rsc = uber_parent(history->rsc); } crm_notice("Treating probe result '%s' for %s on %s as 'not running'", services_ocf_exitcode_str(orig_exit_status), history->rsc->id, pe__node_name(history->node)); update_resource_state(history, history->expected_exit_status, last_failure, on_fail); crm_xml_add(history->xml, XML_ATTR_UNAME, history->node->details->uname); record_failed_op(history); resource_location(ban_rsc, history->node, -INFINITY, "masked-probe-failure", history->rsc->cluster); } /*! * \internal Check whether a given failure is for a given pending action * * \param[in] history Parsed history entry for pending action * \param[in] last_failure Resource's last_failure entry, if known * * \return true if \p last_failure is failure of pending action in \p history, * otherwise false * \note Both \p history and \p last_failure must come from the same * lrm_resource block, as node and resource are assumed to be the same. */ static bool failure_is_newer(const struct action_history *history, const xmlNode *last_failure) { guint failure_interval_ms = 0U; long long failure_change = 0LL; long long this_change = 0LL; if (last_failure == NULL) { return false; // Resource has no last_failure entry } if (!pcmk__str_eq(history->task, crm_element_value(last_failure, XML_LRM_ATTR_TASK), pcmk__str_none)) { return false; // last_failure is for different action } if ((crm_element_value_ms(last_failure, XML_LRM_ATTR_INTERVAL_MS, &failure_interval_ms) != pcmk_ok) || (history->interval_ms != failure_interval_ms)) { return false; // last_failure is for action with different interval } if ((pcmk__scan_ll(crm_element_value(history->xml, XML_RSC_OP_LAST_CHANGE), &this_change, 0LL) != pcmk_rc_ok) || (pcmk__scan_ll(crm_element_value(last_failure, XML_RSC_OP_LAST_CHANGE), &failure_change, 0LL) != pcmk_rc_ok) || (failure_change < this_change)) { return false; // Failure is not known to be newer } return true; } /*! * \internal * \brief Update a resource's role etc. for a pending action * * \param[in,out] history Parsed history entry for pending action * \param[in] last_failure Resource's last_failure entry, if known */ static void process_pending_action(struct action_history *history, const xmlNode *last_failure) { /* For recurring monitors, a failure is recorded only in RSC_last_failure_0, * and there might be a RSC_monitor_INTERVAL entry with the last successful * or pending result. * * If last_failure contains the failure of the pending recurring monitor * we're processing here, and is newer, the action is no longer pending. * (Pending results have call ID -1, which sorts last, so the last failure * if any should be known.) */ if (failure_is_newer(history, last_failure)) { return; } if (strcmp(history->task, PCMK_ACTION_START) == 0) { pe__set_resource_flags(history->rsc, pcmk_rsc_start_pending); set_active(history->rsc); } else if (strcmp(history->task, PCMK_ACTION_PROMOTE) == 0) { history->rsc->role = pcmk_role_promoted; } else if ((strcmp(history->task, PCMK_ACTION_MIGRATE_TO) == 0) && history->node->details->unclean) { /* A migrate_to action is pending on a unclean source, so force a stop * on the target. */ const char *migrate_target = NULL; pcmk_node_t *target = NULL; migrate_target = crm_element_value(history->xml, XML_LRM_ATTR_MIGRATE_TARGET); target = pe_find_node(history->rsc->cluster->nodes, migrate_target); if (target != NULL) { stop_action(history->rsc, target, FALSE); } } if (history->rsc->pending_task != NULL) { /* There should never be multiple pending actions, but as a failsafe, * just remember the first one processed for display purposes. */ return; } if (pcmk_is_probe(history->task, history->interval_ms)) { /* Pending probes are currently never displayed, even if pending * operations are requested. If we ever want to change that, * enable the below and the corresponding part of * native.c:native_pending_task(). */ #if 0 history->rsc->pending_task = strdup("probe"); history->rsc->pending_node = history->node; #endif } else { history->rsc->pending_task = strdup(history->task); history->rsc->pending_node = history->node; } } static void unpack_rsc_op(pcmk_resource_t *rsc, pcmk_node_t *node, xmlNode *xml_op, xmlNode **last_failure, enum action_fail_response *on_fail) { int old_rc = 0; bool expired = false; pcmk_resource_t *parent = rsc; enum rsc_role_e fail_role = pcmk_role_unknown; enum action_fail_response failure_strategy = pcmk_on_fail_restart; struct action_history history = { .rsc = rsc, .node = node, .xml = xml_op, .execution_status = PCMK_EXEC_UNKNOWN, }; CRM_CHECK(rsc && node && xml_op, return); history.id = ID(xml_op); if (history.id == NULL) { crm_err("Ignoring resource history entry for %s on %s without ID", rsc->id, pe__node_name(node)); return; } // Task and interval history.task = crm_element_value(xml_op, XML_LRM_ATTR_TASK); if (history.task == NULL) { crm_err("Ignoring resource history entry %s for %s on %s without " XML_LRM_ATTR_TASK, history.id, rsc->id, pe__node_name(node)); return; } crm_element_value_ms(xml_op, XML_LRM_ATTR_INTERVAL_MS, &(history.interval_ms)); if (!can_affect_state(&history)) { pe_rsc_trace(rsc, "Ignoring resource history entry %s for %s on %s " "with irrelevant action '%s'", history.id, rsc->id, pe__node_name(node), history.task); return; } if (unpack_action_result(&history) != pcmk_rc_ok) { return; // Error already logged } history.expected_exit_status = pe__target_rc_from_xml(xml_op); history.key = pe__xe_history_key(xml_op); crm_element_value_int(xml_op, XML_LRM_ATTR_CALLID, &(history.call_id)); pe_rsc_trace(rsc, "Unpacking %s (%s call %d on %s): %s (%s)", history.id, history.task, history.call_id, pe__node_name(node), pcmk_exec_status_str(history.execution_status), crm_exit_str(history.exit_status)); if (node->details->unclean) { pe_rsc_trace(rsc, "%s is running on %s, which is unclean (further action " "depends on value of stop's on-fail attribute)", rsc->id, pe__node_name(node)); } expired = check_operation_expiry(&history); old_rc = history.exit_status; remap_operation(&history, on_fail, expired); if (expired && (process_expired_result(&history, old_rc) == pcmk_rc_ok)) { goto done; } if (!pe_rsc_is_bundled(rsc) && pcmk_xe_mask_probe_failure(xml_op)) { mask_probe_failure(&history, old_rc, *last_failure, on_fail); goto done; } if (!pcmk_is_set(rsc->flags, pcmk_rsc_unique)) { parent = uber_parent(rsc); } switch (history.execution_status) { case PCMK_EXEC_PENDING: process_pending_action(&history, *last_failure); goto done; case PCMK_EXEC_DONE: update_resource_state(&history, history.exit_status, *last_failure, on_fail); goto done; case PCMK_EXEC_NOT_INSTALLED: unpack_failure_handling(&history, &failure_strategy, &fail_role); if (failure_strategy == pcmk_on_fail_ignore) { crm_warn("Cannot ignore failed %s of %s on %s: " "Resource agent doesn't exist " CRM_XS " status=%d rc=%d id=%s", history.task, rsc->id, pe__node_name(node), history.execution_status, history.exit_status, history.id); /* Also for printing it as "FAILED" by marking it as * pcmk_rsc_failed later */ *on_fail = pcmk_on_fail_ban; } resource_location(parent, node, -INFINITY, "hard-error", rsc->cluster); unpack_rsc_op_failure(&history, failure_strategy, fail_role, last_failure, on_fail); goto done; case PCMK_EXEC_NOT_CONNECTED: if (pe__is_guest_or_remote_node(node) && pcmk_is_set(node->details->remote_rsc->flags, pcmk_rsc_managed)) { /* We should never get into a situation where a managed remote * connection resource is considered OK but a resource action * behind the connection gets a "not connected" status. But as a * fail-safe in case a bug or unusual circumstances do lead to * that, ensure the remote connection is considered failed. */ pe__set_resource_flags(node->details->remote_rsc, pcmk_rsc_failed|pcmk_rsc_stop_if_failed); } break; // Not done, do error handling case PCMK_EXEC_ERROR: case PCMK_EXEC_ERROR_HARD: case PCMK_EXEC_ERROR_FATAL: case PCMK_EXEC_TIMEOUT: case PCMK_EXEC_NOT_SUPPORTED: case PCMK_EXEC_INVALID: break; // Not done, do error handling default: // No other value should be possible at this point break; } unpack_failure_handling(&history, &failure_strategy, &fail_role); if ((failure_strategy == pcmk_on_fail_ignore) || ((failure_strategy == pcmk_on_fail_restart_container) && (strcmp(history.task, PCMK_ACTION_STOP) == 0))) { char *last_change_s = last_change_str(xml_op); crm_warn("Pretending failed %s (%s%s%s) of %s on %s at %s succeeded " CRM_XS " %s", history.task, services_ocf_exitcode_str(history.exit_status), (pcmk__str_empty(history.exit_reason)? "" : ": "), pcmk__s(history.exit_reason, ""), rsc->id, pe__node_name(node), last_change_s, history.id); free(last_change_s); update_resource_state(&history, history.expected_exit_status, *last_failure, on_fail); crm_xml_add(xml_op, XML_ATTR_UNAME, node->details->uname); pe__set_resource_flags(rsc, pcmk_rsc_ignore_failure); record_failed_op(&history); if ((failure_strategy == pcmk_on_fail_restart_container) && cmp_on_fail(*on_fail, pcmk_on_fail_restart) <= 0) { *on_fail = failure_strategy; } } else { unpack_rsc_op_failure(&history, failure_strategy, fail_role, last_failure, on_fail); if (history.execution_status == PCMK_EXEC_ERROR_HARD) { uint8_t log_level = LOG_ERR; if (history.exit_status == PCMK_OCF_NOT_INSTALLED) { log_level = LOG_NOTICE; } do_crm_log(log_level, "Preventing %s from restarting on %s because " "of hard failure (%s%s%s) " CRM_XS " %s", parent->id, pe__node_name(node), services_ocf_exitcode_str(history.exit_status), (pcmk__str_empty(history.exit_reason)? "" : ": "), pcmk__s(history.exit_reason, ""), history.id); resource_location(parent, node, -INFINITY, "hard-error", rsc->cluster); } else if (history.execution_status == PCMK_EXEC_ERROR_FATAL) { crm_err("Preventing %s from restarting anywhere because " "of fatal failure (%s%s%s) " CRM_XS " %s", parent->id, services_ocf_exitcode_str(history.exit_status), (pcmk__str_empty(history.exit_reason)? "" : ": "), pcmk__s(history.exit_reason, ""), history.id); resource_location(parent, NULL, -INFINITY, "fatal-error", rsc->cluster); } } done: pe_rsc_trace(rsc, "%s role on %s after %s is %s (next %s)", rsc->id, pe__node_name(node), history.id, role2text(rsc->role), role2text(rsc->next_role)); } static void add_node_attrs(const xmlNode *xml_obj, pcmk_node_t *node, bool overwrite, pcmk_scheduler_t *scheduler) { const char *cluster_name = NULL; pe_rule_eval_data_t rule_data = { .node_hash = NULL, .role = pcmk_role_unknown, .now = scheduler->now, .match_data = NULL, .rsc_data = NULL, .op_data = NULL }; g_hash_table_insert(node->details->attrs, strdup(CRM_ATTR_UNAME), strdup(node->details->uname)); g_hash_table_insert(node->details->attrs, strdup(CRM_ATTR_ID), strdup(node->details->id)); if (pcmk__str_eq(node->details->id, scheduler->dc_uuid, pcmk__str_casei)) { scheduler->dc_node = node; node->details->is_dc = TRUE; g_hash_table_insert(node->details->attrs, strdup(CRM_ATTR_IS_DC), strdup(XML_BOOLEAN_TRUE)); } else { g_hash_table_insert(node->details->attrs, strdup(CRM_ATTR_IS_DC), strdup(XML_BOOLEAN_FALSE)); } cluster_name = g_hash_table_lookup(scheduler->config_hash, "cluster-name"); if (cluster_name) { g_hash_table_insert(node->details->attrs, strdup(CRM_ATTR_CLUSTER_NAME), strdup(cluster_name)); } pe__unpack_dataset_nvpairs(xml_obj, XML_TAG_ATTR_SETS, &rule_data, node->details->attrs, NULL, overwrite, scheduler); pe__unpack_dataset_nvpairs(xml_obj, XML_TAG_UTILIZATION, &rule_data, node->details->utilization, NULL, FALSE, scheduler); if (pe_node_attribute_raw(node, CRM_ATTR_SITE_NAME) == NULL) { const char *site_name = pe_node_attribute_raw(node, "site-name"); if (site_name) { g_hash_table_insert(node->details->attrs, strdup(CRM_ATTR_SITE_NAME), strdup(site_name)); } else if (cluster_name) { /* Default to cluster-name if unset */ g_hash_table_insert(node->details->attrs, strdup(CRM_ATTR_SITE_NAME), strdup(cluster_name)); } } } static GList * extract_operations(const char *node, const char *rsc, xmlNode * rsc_entry, gboolean active_filter) { int counter = -1; int stop_index = -1; int start_index = -1; xmlNode *rsc_op = NULL; GList *gIter = NULL; GList *op_list = NULL; GList *sorted_op_list = NULL; /* extract operations */ op_list = NULL; sorted_op_list = NULL; for (rsc_op = pcmk__xe_first_child(rsc_entry); rsc_op != NULL; rsc_op = pcmk__xe_next(rsc_op)) { if (pcmk__str_eq((const char *)rsc_op->name, XML_LRM_TAG_RSC_OP, pcmk__str_none)) { crm_xml_add(rsc_op, "resource", rsc); crm_xml_add(rsc_op, XML_ATTR_UNAME, node); op_list = g_list_prepend(op_list, rsc_op); } } if (op_list == NULL) { /* if there are no operations, there is nothing to do */ return NULL; } sorted_op_list = g_list_sort(op_list, sort_op_by_callid); /* create active recurring operations as optional */ if (active_filter == FALSE) { return sorted_op_list; } op_list = NULL; 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; counter++; if (start_index < stop_index) { crm_trace("Skipping %s: not active", ID(rsc_entry)); break; } else if (counter < start_index) { crm_trace("Skipping %s: old", ID(rsc_op)); continue; } op_list = g_list_append(op_list, rsc_op); } g_list_free(sorted_op_list); return op_list; } GList * find_operations(const char *rsc, const char *node, gboolean active_filter, pcmk_scheduler_t *scheduler) { GList *output = NULL; GList *intermediate = NULL; xmlNode *tmp = NULL; xmlNode *status = find_xml_node(scheduler->input, XML_CIB_TAG_STATUS, TRUE); pcmk_node_t *this_node = NULL; xmlNode *node_state = NULL; for (node_state = pcmk__xe_first_child(status); node_state != NULL; node_state = pcmk__xe_next(node_state)) { if (pcmk__str_eq((const char *)node_state->name, XML_CIB_TAG_STATE, pcmk__str_none)) { const char *uname = crm_element_value(node_state, XML_ATTR_UNAME); if (node != NULL && !pcmk__str_eq(uname, node, pcmk__str_casei)) { continue; } this_node = pe_find_node(scheduler->nodes, uname); if(this_node == NULL) { CRM_LOG_ASSERT(this_node != NULL); continue; } else if (pe__is_guest_or_remote_node(this_node)) { determine_remote_online_status(scheduler, this_node); } else { determine_online_status(node_state, this_node, scheduler); } if (this_node->details->online || pcmk_is_set(scheduler->flags, pcmk_sched_fencing_enabled)) { /* offline nodes run no resources... * unless stonith is enabled in which case we need to * make sure rsc start events happen after the stonith */ xmlNode *lrm_rsc = NULL; tmp = find_xml_node(node_state, XML_CIB_TAG_LRM, FALSE); tmp = find_xml_node(tmp, XML_LRM_TAG_RESOURCES, FALSE); for (lrm_rsc = pcmk__xe_first_child(tmp); lrm_rsc != NULL; lrm_rsc = pcmk__xe_next(lrm_rsc)) { if (pcmk__str_eq((const char *)lrm_rsc->name, XML_LRM_TAG_RESOURCE, pcmk__str_none)) { const char *rsc_id = crm_element_value(lrm_rsc, XML_ATTR_ID); if (rsc != NULL && !pcmk__str_eq(rsc_id, rsc, pcmk__str_casei)) { continue; } intermediate = extract_operations(uname, rsc_id, lrm_rsc, active_filter); output = g_list_concat(output, intermediate); } } } } } return output; }