diff --git a/daemons/controld/controld_remote_ra.c b/daemons/controld/controld_remote_ra.c index df4e0bd6fc..505f1f5669 100644 --- a/daemons/controld/controld_remote_ra.c +++ b/daemons/controld/controld_remote_ra.c @@ -1,1490 +1,1487 @@ /* * Copyright 2013-2024 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 #define REMOTE_LRMD_RA "remote" /* The max start timeout before cmd retry */ #define MAX_START_TIMEOUT_MS 10000 #define cmd_set_flags(cmd, flags_to_set) do { \ (cmd)->status = pcmk__set_flags_as(__func__, __LINE__, LOG_TRACE, \ "Remote command", (cmd)->rsc_id, (cmd)->status, \ (flags_to_set), #flags_to_set); \ } while (0) #define cmd_clear_flags(cmd, flags_to_clear) do { \ (cmd)->status = pcmk__clear_flags_as(__func__, __LINE__, LOG_TRACE, \ "Remote command", (cmd)->rsc_id, (cmd)->status, \ (flags_to_clear), #flags_to_clear); \ } while (0) enum remote_cmd_status { cmd_reported_success = (1 << 0), cmd_cancel = (1 << 1), }; typedef struct remote_ra_cmd_s { /*! the local node the cmd is issued from */ char *owner; /*! the remote node the cmd is executed on */ char *rsc_id; /*! the action to execute */ char *action; /*! some string the client wants us to give it back */ char *userdata; /*! start delay in ms */ int start_delay; /*! timer id used for start delay. */ int delay_id; /*! timeout in ms for cmd */ int timeout; /*! recurring interval in ms */ guint interval_ms; /*! interval timer id */ int interval_id; int monitor_timeout_id; int takeover_timeout_id; /*! action parameters */ lrmd_key_value_t *params; pcmk__action_result_t result; int call_id; time_t start_time; uint32_t status; } remote_ra_cmd_t; #define lrm_remote_set_flags(lrm_state, flags_to_set) do { \ lrm_state_t *lrm = (lrm_state); \ remote_ra_data_t *ra = lrm->remote_ra_data; \ ra->status = pcmk__set_flags_as(__func__, __LINE__, LOG_TRACE, "Remote", \ lrm->node_name, ra->status, \ (flags_to_set), #flags_to_set); \ } while (0) #define lrm_remote_clear_flags(lrm_state, flags_to_clear) do { \ lrm_state_t *lrm = (lrm_state); \ remote_ra_data_t *ra = lrm->remote_ra_data; \ ra->status = pcmk__clear_flags_as(__func__, __LINE__, LOG_TRACE, "Remote", \ lrm->node_name, ra->status, \ (flags_to_clear), #flags_to_clear); \ } while (0) enum remote_status { expect_takeover = (1 << 0), takeover_complete = (1 << 1), remote_active = (1 << 2), /* Maintenance mode is difficult to determine from the controller's context, * so we have it signalled back with the transition from the scheduler. */ remote_in_maint = (1 << 3), /* Similar for whether we are controlling a guest node or remote node. * Fortunately there is a meta-attribute in the transition already and * as the situation doesn't change over time we can use the * resource start for noting down the information for later use when * the attributes aren't at hand. */ controlling_guest = (1 << 4), }; typedef struct remote_ra_data_s { crm_trigger_t *work; remote_ra_cmd_t *cur_cmd; GList *cmds; GList *recurring_cmds; uint32_t status; } remote_ra_data_t; static int handle_remote_ra_start(lrm_state_t * lrm_state, remote_ra_cmd_t * cmd, int timeout_ms); static void handle_remote_ra_stop(lrm_state_t * lrm_state, remote_ra_cmd_t * cmd); static GList *fail_all_monitor_cmds(GList * list); static void free_cmd(gpointer user_data) { remote_ra_cmd_t *cmd = user_data; if (!cmd) { return; } if (cmd->delay_id) { g_source_remove(cmd->delay_id); } if (cmd->interval_id) { g_source_remove(cmd->interval_id); } if (cmd->monitor_timeout_id) { g_source_remove(cmd->monitor_timeout_id); } if (cmd->takeover_timeout_id) { g_source_remove(cmd->takeover_timeout_id); } free(cmd->owner); free(cmd->rsc_id); free(cmd->action); free(cmd->userdata); pcmk__reset_result(&(cmd->result)); lrmd_key_value_freeall(cmd->params); free(cmd); } static int generate_callid(void) { static int remote_ra_callid = 0; remote_ra_callid++; if (remote_ra_callid <= 0) { remote_ra_callid = 1; } return remote_ra_callid; } static gboolean recurring_helper(gpointer data) { remote_ra_cmd_t *cmd = data; lrm_state_t *connection_rsc = NULL; cmd->interval_id = 0; connection_rsc = controld_get_executor_state(cmd->rsc_id, false); if (connection_rsc && connection_rsc->remote_ra_data) { remote_ra_data_t *ra_data = connection_rsc->remote_ra_data; ra_data->recurring_cmds = g_list_remove(ra_data->recurring_cmds, cmd); ra_data->cmds = g_list_append(ra_data->cmds, cmd); mainloop_set_trigger(ra_data->work); } return FALSE; } static gboolean start_delay_helper(gpointer data) { remote_ra_cmd_t *cmd = data; lrm_state_t *connection_rsc = NULL; cmd->delay_id = 0; connection_rsc = controld_get_executor_state(cmd->rsc_id, false); if (connection_rsc && connection_rsc->remote_ra_data) { remote_ra_data_t *ra_data = connection_rsc->remote_ra_data; mainloop_set_trigger(ra_data->work); } return FALSE; } static bool should_purge_attributes(pcmk__node_status_t *node) { pcmk__node_status_t *conn_node = NULL; lrm_state_t *connection_rsc = NULL; if ((node->conn_host == NULL) || (node->name == NULL)) { return true; } /* Get the node that was hosting the remote connection resource from the * peer cache. That's the one we really care about here. */ conn_node = pcmk__get_node(0, node->conn_host, NULL, pcmk__node_search_cluster_member); if (conn_node == NULL) { return true; } /* Check the uptime of connection_rsc. If it hasn't been running long * enough, set purge=true. "Long enough" means it started running earlier * than the timestamp when we noticed it went away in the first place. */ connection_rsc = controld_get_executor_state(node->name, false); if (connection_rsc != NULL) { lrmd_t *lrm = connection_rsc->conn; time_t uptime = lrmd__uptime(lrm); time_t now = time(NULL); /* Add 20s of fuzziness to give corosync a while to notice the remote * host is gone. On various error conditions (failure to get uptime, * peer_lost isn't set) we default to purging. */ if (uptime > 0 && conn_node->peer_lost > 0 && uptime + 20 >= now - conn_node->peer_lost) { return false; } } return true; } static enum controld_section_e section_to_delete(bool purge) { if (pcmk_is_set(controld_globals.flags, controld_shutdown_lock_enabled)) { if (purge) { return controld_section_all_unlocked; } else { return controld_section_lrm_unlocked; } } else { if (purge) { return controld_section_all; } else { return controld_section_lrm; } } } static void purge_remote_node_attrs(int call_opt, pcmk__node_status_t *node) { bool purge = should_purge_attributes(node); enum controld_section_e section = section_to_delete(purge); /* Purge node from attrd's memory */ if (purge) { update_attrd_remote_node_removed(node->name, NULL); } controld_delete_node_state(node->name, section, call_opt); } /*! * \internal * \brief Handle cluster communication related to pacemaker_remote node joining * * \param[in] node_name Name of newly integrated pacemaker_remote node */ static void remote_node_up(const char *node_name) { int call_opt; xmlNode *update, *state; pcmk__node_status_t *node = NULL; lrm_state_t *connection_rsc = NULL; CRM_CHECK(node_name != NULL, return); crm_info("Announcing Pacemaker Remote node %s", node_name); call_opt = crmd_cib_smart_opt(); /* Delete node's CRM_OP_PROBED attribute. Deleting any attribute ensures * that the attribute manager learns the node is remote. Deletion of this * specfic attribute is a holdover from when it had special meaning. * * @COMPAT Find another way to tell attrd that the node is remote, without * risking deletion or overwrite of an arbitrary attribute. Then work on * deprecating CRM_OP_PROBED. */ update_attrd(node_name, CRM_OP_PROBED, NULL, NULL, TRUE); /* Ensure node is in the remote peer cache with member status */ node = pcmk__cluster_lookup_remote_node(node_name); CRM_CHECK((node != NULL) && (node->name != NULL), return); purge_remote_node_attrs(call_opt, node); pcmk__update_peer_state(__func__, node, PCMK_VALUE_MEMBER, 0); /* Apply any start state that we were given from the environment on the * remote node. */ connection_rsc = controld_get_executor_state(node->name, false); if (connection_rsc != NULL) { lrmd_t *lrm = connection_rsc->conn; const char *start_state = lrmd__node_start_state(lrm); if (start_state) { set_join_state(start_state, node->name, node->xml_id, true); } } /* pacemaker_remote nodes don't participate in the membership layer, * so cluster nodes don't automatically get notified when they come and go. * We send a cluster message to the DC, and update the CIB node state entry, * so the DC will get it sooner (via message) or later (via CIB refresh), * and any other interested parties can query the CIB. */ broadcast_remote_state_message(node_name, true); update = pcmk__xe_create(NULL, PCMK_XE_STATUS); state = create_node_state_update(node, node_update_cluster, update, __func__); /* Clear the PCMK__XA_NODE_FENCED flag in the node state. If the node ever * needs to be fenced, this flag will allow various actions to determine * whether the fencing has happened yet. */ crm_xml_add(state, PCMK__XA_NODE_FENCED, "0"); /* TODO: If the remote connection drops, and this (async) CIB update either * failed or has not yet completed, later actions could mistakenly think the * node has already been fenced (if the PCMK__XA_NODE_FENCED attribute was * previously set, because it won't have been cleared). This could prevent * actual fencing or allow recurring monitor failures to be cleared too * soon. Ideally, we wouldn't rely on the CIB for the fenced status. */ controld_update_cib(PCMK_XE_STATUS, update, call_opt, NULL); pcmk__xml_free(update); } enum down_opts { DOWN_KEEP_LRM, DOWN_ERASE_LRM }; /*! * \internal * \brief Handle cluster communication related to pacemaker_remote node leaving * * \param[in] node_name Name of lost node * \param[in] opts Whether to keep or erase LRM history */ static void remote_node_down(const char *node_name, const enum down_opts opts) { xmlNode *update; int call_opt = crmd_cib_smart_opt(); pcmk__node_status_t *node = NULL; /* Purge node from attrd's memory */ update_attrd_remote_node_removed(node_name, NULL); /* Normally, only node attributes should be erased, and the resource history * should be kept until the node comes back up. However, after a successful * fence, we want to clear the history as well, so we don't think resources * are still running on the node. */ if (opts == DOWN_ERASE_LRM) { controld_delete_node_state(node_name, controld_section_all, call_opt); } else { controld_delete_node_state(node_name, controld_section_attrs, call_opt); } /* Ensure node is in the remote peer cache with lost state */ node = pcmk__cluster_lookup_remote_node(node_name); CRM_CHECK(node != NULL, return); pcmk__update_peer_state(__func__, node, PCMK__VALUE_LOST, 0); /* Notify DC */ broadcast_remote_state_message(node_name, false); /* Update CIB node state */ update = pcmk__xe_create(NULL, PCMK_XE_STATUS); create_node_state_update(node, node_update_cluster, update, __func__); controld_update_cib(PCMK_XE_STATUS, update, call_opt, NULL); pcmk__xml_free(update); } /*! * \internal * \brief Handle effects of a remote RA command on node state * * \param[in] cmd Completed remote RA command */ static void check_remote_node_state(const remote_ra_cmd_t *cmd) { /* Only successful actions can change node state */ if (!pcmk__result_ok(&(cmd->result))) { return; } if (pcmk__str_eq(cmd->action, PCMK_ACTION_START, pcmk__str_casei)) { remote_node_up(cmd->rsc_id); } else if (pcmk__str_eq(cmd->action, PCMK_ACTION_MIGRATE_FROM, pcmk__str_casei)) { /* After a successful migration, we don't need to do remote_node_up() * because the DC already knows the node is up, and we don't want to * clear LRM history etc. We do need to add the remote node to this * host's remote peer cache, because (unless it happens to be DC) * it hasn't been tracking the remote node, and other code relies on * the cache to distinguish remote nodes from unseen cluster nodes. */ pcmk__node_status_t *node = pcmk__cluster_lookup_remote_node(cmd->rsc_id); CRM_CHECK(node != NULL, return); pcmk__update_peer_state(__func__, node, PCMK_VALUE_MEMBER, 0); } else if (pcmk__str_eq(cmd->action, PCMK_ACTION_STOP, pcmk__str_casei)) { lrm_state_t *lrm_state = controld_get_executor_state(cmd->rsc_id, false); remote_ra_data_t *ra_data = lrm_state? lrm_state->remote_ra_data : NULL; if (ra_data) { if (!pcmk_is_set(ra_data->status, takeover_complete)) { /* Stop means down if we didn't successfully migrate elsewhere */ remote_node_down(cmd->rsc_id, DOWN_KEEP_LRM); } else if (AM_I_DC == FALSE) { /* Only the connection host and DC track node state, * so if the connection migrated elsewhere and we aren't DC, * un-cache the node, so we don't have stale info */ pcmk__cluster_forget_remote_node(cmd->rsc_id); } } } /* We don't do anything for successful monitors, which is correct for * routine recurring monitors, and for monitors on nodes where the * connection isn't supposed to be (the cluster will stop the connection in * that case). However, if the initial probe finds the connection already * active on the node where we want it, we probably should do * remote_node_up(). Unfortunately, we can't distinguish that case here. * Given that connections have to be initiated by the cluster, the chance of * that should be close to zero. */ } static void report_remote_ra_result(remote_ra_cmd_t * cmd) { lrmd_event_data_t op = { 0, }; check_remote_node_state(cmd); op.type = lrmd_event_exec_complete; op.rsc_id = cmd->rsc_id; op.op_type = cmd->action; op.user_data = cmd->userdata; op.timeout = cmd->timeout; op.interval_ms = cmd->interval_ms; op.t_run = cmd->start_time; op.t_rcchange = cmd->start_time; lrmd__set_result(&op, cmd->result.exit_status, cmd->result.execution_status, cmd->result.exit_reason); if (pcmk_is_set(cmd->status, cmd_reported_success) && !pcmk__result_ok(&(cmd->result))) { op.t_rcchange = time(NULL); /* This edge case will likely never ever occur, but if it does the * result is that a failure will not be processed correctly. This is only * remotely possible because we are able to detect a connection resource's tcp * connection has failed at any moment after start has completed. The actual * recurring operation is just a connectivity ping. * * basically, we are not guaranteed that the first successful monitor op and * a subsequent failed monitor op will not occur in the same timestamp. We have to * make it look like the operations occurred at separate times though. */ if (op.t_rcchange == op.t_run) { op.t_rcchange++; } } if (cmd->params) { lrmd_key_value_t *tmp; op.params = pcmk__strkey_table(free, free); for (tmp = cmd->params; tmp; tmp = tmp->next) { pcmk__insert_dup(op.params, tmp->key, tmp->value); } } op.call_id = cmd->call_id; op.remote_nodename = cmd->owner; lrm_op_callback(&op); if (op.params) { g_hash_table_destroy(op.params); } lrmd__reset_result(&op); } /*! * \internal * \brief Return a remote command's remaining timeout in seconds * * \param[in] cmd Remote command to check * * \return Command's remaining timeout in seconds */ static int remaining_timeout_sec(const remote_ra_cmd_t *cmd) { return pcmk__timeout_ms2s(cmd->timeout) - (time(NULL) - cmd->start_time); } static gboolean retry_start_cmd_cb(gpointer data) { lrm_state_t *lrm_state = data; remote_ra_data_t *ra_data = lrm_state->remote_ra_data; remote_ra_cmd_t *cmd = NULL; int rc = ETIME; int remaining = 0; if (!ra_data || !ra_data->cur_cmd) { return FALSE; } cmd = ra_data->cur_cmd; - if (!pcmk__strcase_any_of(cmd->action, PCMK_ACTION_START, - PCMK_ACTION_MIGRATE_FROM, NULL)) { + if (!pcmk__is_up_action(cmd->action)) { return FALSE; } remaining = remaining_timeout_sec(cmd); if (remaining > 0) { rc = handle_remote_ra_start(lrm_state, cmd, remaining * 1000); } else { pcmk__set_result(&(cmd->result), PCMK_OCF_UNKNOWN_ERROR, PCMK_EXEC_TIMEOUT, "Not enough time remains to retry remote connection"); } if (rc != pcmk_rc_ok) { report_remote_ra_result(cmd); if (ra_data->cmds) { mainloop_set_trigger(ra_data->work); } ra_data->cur_cmd = NULL; free_cmd(cmd); } else { /* wait for connection event */ } return FALSE; } static gboolean connection_takeover_timeout_cb(gpointer data) { lrm_state_t *lrm_state = NULL; remote_ra_cmd_t *cmd = data; crm_info("takeover event timed out for node %s", cmd->rsc_id); cmd->takeover_timeout_id = 0; lrm_state = controld_get_executor_state(cmd->rsc_id, false); handle_remote_ra_stop(lrm_state, cmd); free_cmd(cmd); return FALSE; } static gboolean monitor_timeout_cb(gpointer data) { lrm_state_t *lrm_state = NULL; remote_ra_cmd_t *cmd = data; lrm_state = controld_get_executor_state(cmd->rsc_id, false); crm_info("Timed out waiting for remote poke response from %s%s", cmd->rsc_id, (lrm_state? "" : " (no LRM state)")); cmd->monitor_timeout_id = 0; pcmk__set_result(&(cmd->result), PCMK_OCF_UNKNOWN_ERROR, PCMK_EXEC_TIMEOUT, "Remote executor did not respond"); if (lrm_state && lrm_state->remote_ra_data) { remote_ra_data_t *ra_data = lrm_state->remote_ra_data; if (ra_data->cur_cmd == cmd) { ra_data->cur_cmd = NULL; } if (ra_data->cmds) { mainloop_set_trigger(ra_data->work); } } report_remote_ra_result(cmd); free_cmd(cmd); if(lrm_state) { lrm_state_disconnect(lrm_state); } return FALSE; } static void synthesize_lrmd_success(lrm_state_t *lrm_state, const char *rsc_id, const char *op_type) { lrmd_event_data_t op = { 0, }; if (lrm_state == NULL) { /* if lrm_state not given assume local */ lrm_state = controld_get_executor_state(NULL, false); } pcmk__assert(lrm_state != NULL); op.type = lrmd_event_exec_complete; op.rsc_id = rsc_id; op.op_type = op_type; op.t_run = time(NULL); op.t_rcchange = op.t_run; op.call_id = generate_callid(); lrmd__set_result(&op, PCMK_OCF_OK, PCMK_EXEC_DONE, NULL); process_lrm_event(lrm_state, &op, NULL, NULL); } void remote_lrm_op_callback(lrmd_event_data_t * op) { gboolean cmd_handled = FALSE; lrm_state_t *lrm_state = NULL; remote_ra_data_t *ra_data = NULL; remote_ra_cmd_t *cmd = NULL; CRM_CHECK((op != NULL) && (op->remote_nodename != NULL), return); crm_debug("Processing '%s%s%s' event on remote connection to %s: %s " "(%d) status=%s (%d)", (op->op_type? op->op_type : ""), (op->op_type? " " : ""), lrmd_event_type2str(op->type), op->remote_nodename, crm_exit_str((crm_exit_t) op->rc), op->rc, pcmk_exec_status_str(op->op_status), op->op_status); lrm_state = controld_get_executor_state(op->remote_nodename, false); if (!lrm_state || !lrm_state->remote_ra_data) { crm_debug("No state information found for remote connection event"); return; } ra_data = lrm_state->remote_ra_data; if (op->type == lrmd_event_new_client) { // Another client has connected to the remote daemon if (pcmk_is_set(ra_data->status, expect_takeover)) { // Great, we knew this was coming lrm_remote_clear_flags(lrm_state, expect_takeover); lrm_remote_set_flags(lrm_state, takeover_complete); } else { crm_err("Disconnecting from Pacemaker Remote node %s due to " "unexpected client takeover", op->remote_nodename); /* In this case, lrmd_tls_connection_destroy() will be called under the control of mainloop. */ /* Do not free lrm_state->conn yet. */ /* It'll be freed in the following stop action. */ lrm_state_disconnect_only(lrm_state); } return; } /* filter all EXEC events up */ if (op->type == lrmd_event_exec_complete) { if (pcmk_is_set(ra_data->status, takeover_complete)) { crm_debug("ignoring event, this connection is taken over by another node"); } else { lrm_op_callback(op); } return; } if ((op->type == lrmd_event_disconnect) && (ra_data->cur_cmd == NULL)) { if (!pcmk_is_set(ra_data->status, remote_active)) { crm_debug("Disconnection from Pacemaker Remote node %s complete", lrm_state->node_name); } else if (!remote_ra_is_in_maintenance(lrm_state)) { crm_err("Lost connection to Pacemaker Remote node %s", lrm_state->node_name); ra_data->recurring_cmds = fail_all_monitor_cmds(ra_data->recurring_cmds); ra_data->cmds = fail_all_monitor_cmds(ra_data->cmds); } else { crm_notice("Unmanaged Pacemaker Remote node %s disconnected", lrm_state->node_name); /* Do roughly what a 'stop' on the remote-resource would do */ handle_remote_ra_stop(lrm_state, NULL); remote_node_down(lrm_state->node_name, DOWN_KEEP_LRM); /* now fake the reply of a successful 'stop' */ synthesize_lrmd_success(NULL, lrm_state->node_name, PCMK_ACTION_STOP); } return; } if (!ra_data->cur_cmd) { crm_debug("no event to match"); return; } cmd = ra_data->cur_cmd; /* Start actions and migrate from actions complete after connection * comes back to us. */ - if ((op->type == lrmd_event_connect) - && pcmk__strcase_any_of(cmd->action, PCMK_ACTION_START, - PCMK_ACTION_MIGRATE_FROM, NULL)) { + if ((op->type == lrmd_event_connect) && pcmk__is_up_action(cmd->action)) { if (op->connection_rc < 0) { int remaining = remaining_timeout_sec(cmd); if ((op->connection_rc == -ENOKEY) || (op->connection_rc == -EKEYREJECTED)) { // Hard error, don't retry pcmk__set_result(&(cmd->result), PCMK_OCF_INVALID_PARAM, PCMK_EXEC_ERROR, pcmk_strerror(op->connection_rc)); } else if (remaining > 3) { crm_trace("Rescheduling start (%ds remains before timeout)", remaining); pcmk__create_timer(1000, retry_start_cmd_cb, lrm_state); return; } else { crm_trace("Not enough time before timeout (%ds) " "to reschedule start", remaining); pcmk__format_result(&(cmd->result), PCMK_OCF_UNKNOWN_ERROR, PCMK_EXEC_TIMEOUT, "%s without enough time to retry", pcmk_strerror(op->connection_rc)); } } else { lrm_state_reset_tables(lrm_state, TRUE); pcmk__set_result(&(cmd->result), PCMK_OCF_OK, PCMK_EXEC_DONE, NULL); lrm_remote_set_flags(lrm_state, remote_active); } crm_debug("Remote connection event matched %s action", cmd->action); report_remote_ra_result(cmd); cmd_handled = TRUE; } else if ((op->type == lrmd_event_poke) && pcmk__str_eq(cmd->action, PCMK_ACTION_MONITOR, pcmk__str_casei)) { if (cmd->monitor_timeout_id) { g_source_remove(cmd->monitor_timeout_id); cmd->monitor_timeout_id = 0; } /* Only report success the first time, after that only worry about failures. * For this function, if we get the poke pack, it is always a success. Pokes * only fail if the send fails, or the response times out. */ if (!pcmk_is_set(cmd->status, cmd_reported_success)) { pcmk__set_result(&(cmd->result), PCMK_OCF_OK, PCMK_EXEC_DONE, NULL); report_remote_ra_result(cmd); cmd_set_flags(cmd, cmd_reported_success); } crm_debug("Remote poke event matched %s action", cmd->action); /* success, keep rescheduling if interval is present. */ if (cmd->interval_ms && !pcmk_is_set(cmd->status, cmd_cancel)) { ra_data->recurring_cmds = g_list_append(ra_data->recurring_cmds, cmd); cmd->interval_id = pcmk__create_timer(cmd->interval_ms, recurring_helper, cmd); cmd = NULL; /* prevent free */ } cmd_handled = TRUE; } else if ((op->type == lrmd_event_disconnect) && pcmk__str_eq(cmd->action, PCMK_ACTION_MONITOR, pcmk__str_casei)) { if (pcmk_is_set(ra_data->status, remote_active) && !pcmk_is_set(cmd->status, cmd_cancel)) { pcmk__set_result(&(cmd->result), PCMK_OCF_UNKNOWN_ERROR, PCMK_EXEC_ERROR, "Remote connection unexpectedly dropped " "during monitor"); report_remote_ra_result(cmd); crm_err("Remote connection to %s unexpectedly dropped during monitor", lrm_state->node_name); } cmd_handled = TRUE; } else { crm_debug("Event did not match %s action", ra_data->cur_cmd->action); } if (cmd_handled) { ra_data->cur_cmd = NULL; if (ra_data->cmds) { mainloop_set_trigger(ra_data->work); } free_cmd(cmd); } } static void handle_remote_ra_stop(lrm_state_t * lrm_state, remote_ra_cmd_t * cmd) { remote_ra_data_t *ra_data = NULL; pcmk__assert(lrm_state != NULL); ra_data = lrm_state->remote_ra_data; if (!pcmk_is_set(ra_data->status, takeover_complete)) { /* delete pending ops when ever the remote connection is intentionally stopped */ g_hash_table_remove_all(lrm_state->active_ops); } else { /* we no longer hold the history if this connection has been migrated, * however, we keep metadata cache for future use */ lrm_state_reset_tables(lrm_state, FALSE); } lrm_remote_clear_flags(lrm_state, remote_active); lrm_state_disconnect(lrm_state); if (ra_data->cmds) { g_list_free_full(ra_data->cmds, free_cmd); } if (ra_data->recurring_cmds) { g_list_free_full(ra_data->recurring_cmds, free_cmd); } ra_data->cmds = NULL; ra_data->recurring_cmds = NULL; ra_data->cur_cmd = NULL; if (cmd) { pcmk__set_result(&(cmd->result), PCMK_OCF_OK, PCMK_EXEC_DONE, NULL); report_remote_ra_result(cmd); } } // \return Standard Pacemaker return code static int handle_remote_ra_start(lrm_state_t * lrm_state, remote_ra_cmd_t * cmd, int timeout_ms) { const char *server = NULL; lrmd_key_value_t *tmp = NULL; int port = 0; int timeout_used = timeout_ms > MAX_START_TIMEOUT_MS ? MAX_START_TIMEOUT_MS : timeout_ms; int rc = pcmk_rc_ok; for (tmp = cmd->params; tmp; tmp = tmp->next) { if (pcmk__strcase_any_of(tmp->key, PCMK_REMOTE_RA_ADDR, PCMK_REMOTE_RA_SERVER, NULL)) { server = tmp->value; } else if (pcmk__str_eq(tmp->key, PCMK_REMOTE_RA_PORT, pcmk__str_none)) { port = atoi(tmp->value); } else if (pcmk__str_eq(tmp->key, CRM_META "_" PCMK__META_CONTAINER, pcmk__str_none)) { lrm_remote_set_flags(lrm_state, controlling_guest); } } rc = controld_connect_remote_executor(lrm_state, server, port, timeout_used); if (rc != pcmk_rc_ok) { pcmk__format_result(&(cmd->result), PCMK_OCF_UNKNOWN_ERROR, PCMK_EXEC_ERROR, "Could not connect to Pacemaker Remote node %s: %s", lrm_state->node_name, pcmk_rc_str(rc)); } return rc; } static gboolean handle_remote_ra_exec(gpointer user_data) { int rc = 0; lrm_state_t *lrm_state = user_data; remote_ra_data_t *ra_data = lrm_state->remote_ra_data; remote_ra_cmd_t *cmd; GList *first = NULL; if (ra_data->cur_cmd) { /* still waiting on previous cmd */ return TRUE; } while (ra_data->cmds) { first = ra_data->cmds; cmd = first->data; if (cmd->delay_id) { /* still waiting for start delay timer to trip */ return TRUE; } ra_data->cmds = g_list_remove_link(ra_data->cmds, first); g_list_free_1(first); if (pcmk__str_any_of(cmd->action, PCMK_ACTION_START, PCMK_ACTION_MIGRATE_FROM, NULL)) { lrm_remote_clear_flags(lrm_state, expect_takeover | takeover_complete); if (handle_remote_ra_start(lrm_state, cmd, cmd->timeout) == pcmk_rc_ok) { /* take care of this later when we get async connection result */ crm_debug("Initiated async remote connection, %s action will complete after connect event", cmd->action); ra_data->cur_cmd = cmd; return TRUE; } report_remote_ra_result(cmd); } else if (!strcmp(cmd->action, PCMK_ACTION_MONITOR)) { if (lrm_state_is_connected(lrm_state) == TRUE) { rc = lrm_state_poke_connection(lrm_state); if (rc < 0) { pcmk__set_result(&(cmd->result), PCMK_OCF_UNKNOWN_ERROR, PCMK_EXEC_ERROR, pcmk_strerror(rc)); } } else { rc = -1; pcmk__set_result(&(cmd->result), PCMK_OCF_NOT_RUNNING, PCMK_EXEC_DONE, "Remote connection inactive"); } if (rc == 0) { crm_debug("Poked Pacemaker Remote at node %s, waiting for async response", cmd->rsc_id); ra_data->cur_cmd = cmd; cmd->monitor_timeout_id = pcmk__create_timer(cmd->timeout, monitor_timeout_cb, cmd); return TRUE; } report_remote_ra_result(cmd); } else if (!strcmp(cmd->action, PCMK_ACTION_STOP)) { if (pcmk_is_set(ra_data->status, expect_takeover)) { /* briefly wait on stop for the takeover event to occur. If the * takeover event does not occur during the wait period, that's fine. * It just means that the remote-node's lrm_status section is going to get * cleared which will require all the resources running in the remote-node * to be explicitly re-detected via probe actions. If the takeover does occur * successfully, then we can leave the status section intact. */ cmd->takeover_timeout_id = pcmk__create_timer((cmd->timeout/2), connection_takeover_timeout_cb, cmd); ra_data->cur_cmd = cmd; return TRUE; } handle_remote_ra_stop(lrm_state, cmd); } else if (strcmp(cmd->action, PCMK_ACTION_MIGRATE_TO) == 0) { lrm_remote_clear_flags(lrm_state, takeover_complete); lrm_remote_set_flags(lrm_state, expect_takeover); pcmk__set_result(&(cmd->result), PCMK_OCF_OK, PCMK_EXEC_DONE, NULL); report_remote_ra_result(cmd); } else if (pcmk__str_any_of(cmd->action, PCMK_ACTION_RELOAD, PCMK_ACTION_RELOAD_AGENT, NULL)) { /* Currently the only reloadable parameter is * PCMK_REMOTE_RA_RECONNECT_INTERVAL, which is only used by the * scheduler via the CIB, so reloads are a no-op. * * @COMPAT DC <2.1.0: We only need to check for "reload" in case * we're in a rolling upgrade with a DC scheduling "reload" instead * of "reload-agent". An OCF 1.1 "reload" would be a no-op anyway, * so this would work for that purpose as well. */ pcmk__set_result(&(cmd->result), PCMK_OCF_OK, PCMK_EXEC_DONE, NULL); report_remote_ra_result(cmd); } free_cmd(cmd); } return TRUE; } static void remote_ra_data_init(lrm_state_t * lrm_state) { remote_ra_data_t *ra_data = NULL; if (lrm_state->remote_ra_data) { return; } ra_data = pcmk__assert_alloc(1, sizeof(remote_ra_data_t)); ra_data->work = mainloop_add_trigger(G_PRIORITY_HIGH, handle_remote_ra_exec, lrm_state); lrm_state->remote_ra_data = ra_data; } void remote_ra_cleanup(lrm_state_t * lrm_state) { remote_ra_data_t *ra_data = lrm_state->remote_ra_data; if (!ra_data) { return; } if (ra_data->cmds) { g_list_free_full(ra_data->cmds, free_cmd); } if (ra_data->recurring_cmds) { g_list_free_full(ra_data->recurring_cmds, free_cmd); } mainloop_destroy_trigger(ra_data->work); free(ra_data); lrm_state->remote_ra_data = NULL; } gboolean is_remote_lrmd_ra(const char *agent, const char *provider, const char *id) { if (agent && provider && !strcmp(agent, REMOTE_LRMD_RA) && !strcmp(provider, "pacemaker")) { return TRUE; } return (id != NULL) && (controld_get_executor_state(id, false) != NULL) && !controld_is_local_node(id); } lrmd_rsc_info_t * remote_ra_get_rsc_info(lrm_state_t * lrm_state, const char *rsc_id) { lrmd_rsc_info_t *info = NULL; CRM_CHECK(rsc_id != NULL, return NULL); if (controld_get_executor_state(rsc_id, false) != NULL) { info = pcmk__assert_alloc(1, sizeof(lrmd_rsc_info_t)); info->id = pcmk__str_copy(rsc_id); info->type = pcmk__str_copy(REMOTE_LRMD_RA); info->standard = pcmk__str_copy(PCMK_RESOURCE_CLASS_OCF); info->provider = pcmk__str_copy("pacemaker"); } return info; } static gboolean is_remote_ra_supported_action(const char *action) { return pcmk__str_any_of(action, PCMK_ACTION_START, PCMK_ACTION_STOP, PCMK_ACTION_MONITOR, PCMK_ACTION_MIGRATE_TO, PCMK_ACTION_MIGRATE_FROM, PCMK_ACTION_RELOAD_AGENT, PCMK_ACTION_RELOAD, NULL); } static GList * fail_all_monitor_cmds(GList * list) { GList *rm_list = NULL; remote_ra_cmd_t *cmd = NULL; GList *gIter = NULL; for (gIter = list; gIter != NULL; gIter = gIter->next) { cmd = gIter->data; if ((cmd->interval_ms > 0) && pcmk__str_eq(cmd->action, PCMK_ACTION_MONITOR, pcmk__str_casei)) { rm_list = g_list_append(rm_list, cmd); } } for (gIter = rm_list; gIter != NULL; gIter = gIter->next) { cmd = gIter->data; pcmk__set_result(&(cmd->result), PCMK_OCF_UNKNOWN_ERROR, PCMK_EXEC_ERROR, "Lost connection to remote executor"); crm_trace("Pre-emptively failing %s %s (interval=%u, %s)", cmd->action, cmd->rsc_id, cmd->interval_ms, cmd->userdata); report_remote_ra_result(cmd); list = g_list_remove(list, cmd); free_cmd(cmd); } /* frees only the list data, not the cmds */ g_list_free(rm_list); return list; } static GList * remove_cmd(GList * list, const char *action, guint interval_ms) { remote_ra_cmd_t *cmd = NULL; GList *gIter = NULL; for (gIter = list; gIter != NULL; gIter = gIter->next) { cmd = gIter->data; if ((cmd->interval_ms == interval_ms) && pcmk__str_eq(cmd->action, action, pcmk__str_casei)) { break; } cmd = NULL; } if (cmd) { list = g_list_remove(list, cmd); free_cmd(cmd); } return list; } int remote_ra_cancel(lrm_state_t *lrm_state, const char *rsc_id, const char *action, guint interval_ms) { lrm_state_t *connection_rsc = NULL; remote_ra_data_t *ra_data = NULL; CRM_CHECK(rsc_id != NULL, return -EINVAL); connection_rsc = controld_get_executor_state(rsc_id, false); if (!connection_rsc || !connection_rsc->remote_ra_data) { return -EINVAL; } ra_data = connection_rsc->remote_ra_data; ra_data->cmds = remove_cmd(ra_data->cmds, action, interval_ms); ra_data->recurring_cmds = remove_cmd(ra_data->recurring_cmds, action, interval_ms); if (ra_data->cur_cmd && (ra_data->cur_cmd->interval_ms == interval_ms) && (pcmk__str_eq(ra_data->cur_cmd->action, action, pcmk__str_casei))) { cmd_set_flags(ra_data->cur_cmd, cmd_cancel); } return 0; } static remote_ra_cmd_t * handle_dup_monitor(remote_ra_data_t *ra_data, guint interval_ms, const char *userdata) { GList *gIter = NULL; remote_ra_cmd_t *cmd = NULL; /* there are 3 places a potential duplicate monitor operation * could exist. * 1. recurring_cmds list. where the op is waiting for its next interval * 2. cmds list, where the op is queued to get executed immediately * 3. cur_cmd, which means the monitor op is in flight right now. */ if (interval_ms == 0) { return NULL; } if (ra_data->cur_cmd && !pcmk_is_set(ra_data->cur_cmd->status, cmd_cancel) && (ra_data->cur_cmd->interval_ms == interval_ms) && pcmk__str_eq(ra_data->cur_cmd->action, PCMK_ACTION_MONITOR, pcmk__str_casei)) { cmd = ra_data->cur_cmd; goto handle_dup; } for (gIter = ra_data->recurring_cmds; gIter != NULL; gIter = gIter->next) { cmd = gIter->data; if ((cmd->interval_ms == interval_ms) && pcmk__str_eq(cmd->action, PCMK_ACTION_MONITOR, pcmk__str_casei)) { goto handle_dup; } } for (gIter = ra_data->cmds; gIter != NULL; gIter = gIter->next) { cmd = gIter->data; if ((cmd->interval_ms == interval_ms) && pcmk__str_eq(cmd->action, PCMK_ACTION_MONITOR, pcmk__str_casei)) { goto handle_dup; } } return NULL; handle_dup: crm_trace("merging duplicate monitor cmd " PCMK__OP_FMT, cmd->rsc_id, PCMK_ACTION_MONITOR, interval_ms); /* update the userdata */ if (userdata) { free(cmd->userdata); cmd->userdata = pcmk__str_copy(userdata); } /* if we've already reported success, generate a new call id */ if (pcmk_is_set(cmd->status, cmd_reported_success)) { cmd->start_time = time(NULL); cmd->call_id = generate_callid(); cmd_clear_flags(cmd, cmd_reported_success); } /* if we have an interval_id set, that means we are in the process of * waiting for this cmd's next interval. instead of waiting, cancel * the timer and execute the action immediately */ if (cmd->interval_id) { g_source_remove(cmd->interval_id); cmd->interval_id = 0; recurring_helper(cmd); } return cmd; } /*! * \internal * \brief Execute an action using the (internal) ocf:pacemaker:remote agent * * \param[in] lrm_state Executor state object for remote connection * \param[in] rsc_id Connection resource ID * \param[in] action Action to execute * \param[in] userdata String to copy and pass to execution callback * \param[in] interval_ms Action interval (in milliseconds) * \param[in] timeout_ms Action timeout (in milliseconds) * \param[in] start_delay_ms Delay (in milliseconds) before executing action * \param[in,out] params Connection resource parameters * \param[out] call_id Where to store call ID on success * * \return Standard Pacemaker return code * \note This takes ownership of \p params, which should not be used or freed * after calling this function. */ int controld_execute_remote_agent(const lrm_state_t *lrm_state, const char *rsc_id, const char *action, const char *userdata, guint interval_ms, int timeout_ms, int start_delay_ms, lrmd_key_value_t *params, int *call_id) { lrm_state_t *connection_rsc = NULL; remote_ra_cmd_t *cmd = NULL; remote_ra_data_t *ra_data = NULL; *call_id = 0; CRM_CHECK((lrm_state != NULL) && (rsc_id != NULL) && (action != NULL) && (userdata != NULL) && (call_id != NULL), lrmd_key_value_freeall(params); return EINVAL); if (!is_remote_ra_supported_action(action)) { lrmd_key_value_freeall(params); return EOPNOTSUPP; } connection_rsc = controld_get_executor_state(rsc_id, false); if (connection_rsc == NULL) { lrmd_key_value_freeall(params); return ENOTCONN; } remote_ra_data_init(connection_rsc); ra_data = connection_rsc->remote_ra_data; cmd = handle_dup_monitor(ra_data, interval_ms, userdata); if (cmd) { *call_id = cmd->call_id; lrmd_key_value_freeall(params); return pcmk_rc_ok; } cmd = pcmk__assert_alloc(1, sizeof(remote_ra_cmd_t)); cmd->owner = pcmk__str_copy(lrm_state->node_name); cmd->rsc_id = pcmk__str_copy(rsc_id); cmd->action = pcmk__str_copy(action); cmd->userdata = pcmk__str_copy(userdata); cmd->interval_ms = interval_ms; cmd->timeout = timeout_ms; cmd->start_delay = start_delay_ms; cmd->params = params; cmd->start_time = time(NULL); cmd->call_id = generate_callid(); if (cmd->start_delay) { cmd->delay_id = pcmk__create_timer(cmd->start_delay, start_delay_helper, cmd); } ra_data->cmds = g_list_append(ra_data->cmds, cmd); mainloop_set_trigger(ra_data->work); *call_id = cmd->call_id; return pcmk_rc_ok; } /*! * \internal * \brief Immediately fail all monitors of a remote node, if proxied here * * \param[in] node_name Name of pacemaker_remote node */ void remote_ra_fail(const char *node_name) { lrm_state_t *lrm_state = NULL; CRM_CHECK(node_name != NULL, return); lrm_state = controld_get_executor_state(node_name, false); if (lrm_state && lrm_state_is_connected(lrm_state)) { remote_ra_data_t *ra_data = lrm_state->remote_ra_data; crm_info("Failing monitors on Pacemaker Remote node %s", node_name); ra_data->recurring_cmds = fail_all_monitor_cmds(ra_data->recurring_cmds); ra_data->cmds = fail_all_monitor_cmds(ra_data->cmds); } } /* A guest node fencing implied by host fencing looks like: * * * * * * * */ #define XPATH_PSEUDO_FENCE "/" PCMK__XE_PSEUDO_EVENT \ "[@" PCMK_XA_OPERATION "='stonith']/" PCMK__XE_DOWNED "/" PCMK_XE_NODE /*! * \internal * \brief Check a pseudo-action for Pacemaker Remote node side effects * * \param[in,out] xml XML of pseudo-action to check */ void remote_ra_process_pseudo(xmlNode *xml) { xmlXPathObjectPtr search = xpath_search(xml, XPATH_PSEUDO_FENCE); if (numXpathResults(search) == 1) { xmlNode *result = getXpathResult(search, 0); /* Normally, we handle the necessary side effects of a guest node stop * action when reporting the remote agent's result. However, if the stop * is implied due to fencing, it will be a fencing pseudo-event, and * there won't be a result to report. Handle that case here. * * This will result in a duplicate call to remote_node_down() if the * guest stop was real instead of implied, but that shouldn't hurt. * * There is still one corner case that isn't handled: if a guest node * isn't running any resources when its host is fenced, it will appear * to be cleanly stopped, so there will be no pseudo-fence, and our * peer cache state will be incorrect unless and until the guest is * recovered. */ if (result) { const char *remote = pcmk__xe_id(result); if (remote) { remote_node_down(remote, DOWN_ERASE_LRM); } } } freeXpathObject(search); } static void remote_ra_maintenance(lrm_state_t * lrm_state, gboolean maintenance) { xmlNode *update, *state; int call_opt; pcmk__node_status_t *node = NULL; call_opt = crmd_cib_smart_opt(); node = pcmk__cluster_lookup_remote_node(lrm_state->node_name); CRM_CHECK(node != NULL, return); update = pcmk__xe_create(NULL, PCMK_XE_STATUS); state = create_node_state_update(node, node_update_none, update, __func__); crm_xml_add(state, PCMK__XA_NODE_IN_MAINTENANCE, (maintenance? "1" : "0")); if (controld_update_cib(PCMK_XE_STATUS, update, call_opt, NULL) == pcmk_rc_ok) { /* TODO: still not 100% sure that async update will succeed ... */ if (maintenance) { lrm_remote_set_flags(lrm_state, remote_in_maint); } else { lrm_remote_clear_flags(lrm_state, remote_in_maint); } } pcmk__xml_free(update); } #define XPATH_PSEUDO_MAINTENANCE "//" PCMK__XE_PSEUDO_EVENT \ "[@" PCMK_XA_OPERATION "='" PCMK_ACTION_MAINTENANCE_NODES "']/" \ PCMK__XE_MAINTENANCE /*! * \internal * \brief Check a pseudo-action holding updates for maintenance state * * \param[in,out] xml XML of pseudo-action to check */ void remote_ra_process_maintenance_nodes(xmlNode *xml) { xmlXPathObjectPtr search = xpath_search(xml, XPATH_PSEUDO_MAINTENANCE); if (numXpathResults(search) == 1) { xmlNode *node; int cnt = 0, cnt_remote = 0; for (node = pcmk__xe_first_child(getXpathResult(search, 0), PCMK_XE_NODE, NULL, NULL); node != NULL; node = pcmk__xe_next(node, PCMK_XE_NODE)) { lrm_state_t *lrm_state = NULL; const char *id = pcmk__xe_id(node); cnt++; if (id == NULL) { continue; // Shouldn't be possible } lrm_state = controld_get_executor_state(id, false); if (lrm_state && lrm_state->remote_ra_data && pcmk_is_set(((remote_ra_data_t *) lrm_state->remote_ra_data)->status, remote_active)) { const char *in_maint_s = NULL; int in_maint; cnt_remote++; in_maint_s = crm_element_value(node, PCMK__XA_NODE_IN_MAINTENANCE); pcmk__scan_min_int(in_maint_s, &in_maint, 0); remote_ra_maintenance(lrm_state, in_maint); } } crm_trace("Action holds %d nodes (%d remotes found) adjusting " PCMK_OPT_MAINTENANCE_MODE, cnt, cnt_remote); } freeXpathObject(search); } gboolean remote_ra_is_in_maintenance(lrm_state_t * lrm_state) { remote_ra_data_t *ra_data = lrm_state->remote_ra_data; return pcmk_is_set(ra_data->status, remote_in_maint); } gboolean remote_ra_controlling_guest(lrm_state_t * lrm_state) { remote_ra_data_t *ra_data = lrm_state->remote_ra_data; return pcmk_is_set(ra_data->status, controlling_guest); } diff --git a/include/crm/common/actions_internal.h b/include/crm/common/actions_internal.h index 3d025cd5b4..0f3aa815a7 100644 --- a/include/crm/common/actions_internal.h +++ b/include/crm/common/actions_internal.h @@ -1,269 +1,286 @@ /* * Copyright 2004-2024 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_ACTIONS_INTERNAL__H #define PCMK__CRM_COMMON_ACTIONS_INTERNAL__H #include // bool #include // uint32_t, UINT32_C() #include // guint, GList, GHashTable #include // xmlNode #include // PCMK_ACTION_MONITOR #include // enum rsc_role_e #include // pcmk_resource_t, pcmk_node_t #include // pcmk__str_eq() +#include // pcmk__str_any_of() + #ifdef __cplusplus extern "C" { #endif //! printf-style format to create operation key from resource, action, interval #define PCMK__OP_FMT "%s_%s_%u" /*! * \internal * \brief Set action flags for an action * * \param[in,out] action Action to set flags for * \param[in] flags_to_set Group of enum pcmk__action_flags to set */ #define pcmk__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) /*! * \internal * \brief Clear action flags for an action * * \param[in,out] action Action to clear flags for * \param[in] flags_to_clear Group of enum pcmk__action_flags to clear */ #define pcmk__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) /*! * \internal * \brief Set action flags for a flag group * * \param[in,out] action_flags Flag group to set flags for * \param[in] action_name Name of action being modified (for logging) * \param[in] to_set Group of enum pcmk__action_flags to set */ #define pcmk__set_raw_action_flags(action_flags, action_name, to_set) do { \ action_flags = pcmk__set_flags_as(__func__, __LINE__, \ LOG_TRACE, "Action", action_name, \ (action_flags), \ (to_set), #to_set); \ } while (0) /*! * \internal * \brief Clear action flags for a flag group * * \param[in,out] action_flags Flag group to clear flags for * \param[in] action_name Name of action being modified (for logging) * \param[in] to_clear Group of enum pcmk__action_flags to clear */ #define pcmk__clear_raw_action_flags(action_flags, action_name, to_clear) \ do { \ action_flags = pcmk__clear_flags_as(__func__, __LINE__, LOG_TRACE, \ "Action", action_name, \ (action_flags), \ (to_clear), #to_clear); \ } while (0) // Possible actions (including some pseudo-actions) enum pcmk__action_type { pcmk__action_unspecified = 0, // Unspecified or unknown action pcmk__action_monitor, // Monitor // Each "completed" action must be the regular action plus 1 pcmk__action_stop, // Stop pcmk__action_stopped, // Stop completed pcmk__action_start, // Start pcmk__action_started, // Start completed pcmk__action_notify, // Notify pcmk__action_notified, // Notify completed pcmk__action_promote, // Promote pcmk__action_promoted, // Promoted pcmk__action_demote, // Demote pcmk__action_demoted, // Demoted pcmk__action_shutdown, // Shut down node pcmk__action_fence, // Fence node }; // Action scheduling flags enum pcmk__action_flags { // No action flags set (compare with equality rather than bit set) pcmk__no_action_flags = 0, // Whether action does not require invoking an agent pcmk__action_pseudo = (UINT32_C(1) << 0), // Whether action is runnable pcmk__action_runnable = (UINT32_C(1) << 1), // Whether action should not be executed pcmk__action_optional = (UINT32_C(1) << 2), // Whether action should be added to transition graph even if optional pcmk__action_always_in_graph = (UINT32_C(1) << 3), // Whether operation-specific instance attributes have been unpacked yet pcmk__action_attrs_evaluated = (UINT32_C(1) << 4), // Whether action is allowed to be part of a live migration pcmk__action_migratable = (UINT32_C(1) << 7), // Whether action has been added to transition graph pcmk__action_added_to_graph = (UINT32_C(1) << 8), // Whether action is a stop to abort a dangling migration pcmk__action_migration_abort = (UINT32_C(1) << 11), // Whether action is recurring monitor that must be rescheduled if active pcmk__action_reschedule = (UINT32_C(1) << 13), // Whether action has already been processed by a recursive procedure pcmk__action_detect_loop = (UINT32_C(1) << 14), // Whether action's inputs have been de-duplicated yet pcmk__action_inputs_deduplicated = (UINT32_C(1) << 15), // Whether action can be executed on DC rather than own node pcmk__action_on_dc = (UINT32_C(1) << 16), }; /* Possible responses to a resource action failure * * The order is significant; the values are in order of increasing severity so * that they can be compared with less than and greater than. */ enum pcmk__on_fail { pcmk__on_fail_ignore, // Act as if failure didn't happen pcmk__on_fail_demote, // Demote if promotable, else stop pcmk__on_fail_restart, // Restart resource /* Fence the remote node created by the resource if fencing is enabled, * otherwise attempt to restart the resource (used internally for some * remote connection failures). */ pcmk__on_fail_reset_remote, pcmk__on_fail_restart_container, // Restart resource's container pcmk__on_fail_ban, // Ban resource from current node pcmk__on_fail_block, // Treat resource as unmanaged pcmk__on_fail_stop, // Stop resource and leave stopped pcmk__on_fail_standby_node, // Put resource's node in standby pcmk__on_fail_fence_node, // Fence resource's node }; // What resource needs before it can be recovered from a failed node enum pcmk__requires { pcmk__requires_nothing = 0, // Resource can be recovered immediately pcmk__requires_quorum = 1, // Resource can be recovered if quorate pcmk__requires_fencing = 2, // Resource can be recovered after fencing }; // Implementation of pcmk_action_t struct pcmk__action { int id; // Counter to identify action /* * When the controller aborts a transition graph, it sets an abort priority. * If this priority is higher, the action will still be executed anyway. * Pseudo-actions are always allowed, so this is irrelevant for them. */ int priority; pcmk_resource_t *rsc; // Resource to apply action to, if any pcmk_node_t *node; // Copy of node to execute action on, if any xmlNode *op_entry; // Action XML configuration, if any char *task; // Action name char *uuid; // Action key char *cancel_task; // If task is "cancel", the action being cancelled char *reason; // Readable description of why action is needed uint32_t flags; // Group of enum pcmk__action_flags enum pcmk__requires needs; // Prerequisite for recovery enum pcmk__on_fail on_fail; // Response to failure enum rsc_role_e fail_role; // Resource role if action fails GHashTable *meta; // Meta-attributes relevant to action GHashTable *extra; // Action-specific instance attributes pcmk_scheduler_t *scheduler; // Scheduler data this action is part of /* Current count of runnable instance actions for "first" action in an * ordering dependency with pcmk__ar_min_runnable set. */ int runnable_before; /* * Number of instance actions for "first" action in an ordering dependency * with pcmk__ar_min_runnable set that must be runnable before this action * can be runnable. */ int required_runnable_before; // Actions in a relation with this one (as pcmk__related_action_t *) GList *actions_before; GList *actions_after; }; char *pcmk__op_key(const char *rsc_id, const char *op_type, guint interval_ms); char *pcmk__notify_key(const char *rsc_id, const char *notify_type, const char *op_type); char *pcmk__transition_key(int transition_id, int action_id, int target_rc, const char *node); void pcmk__filter_op_for_digest(xmlNode *param_set); bool pcmk__is_fencing_action(const char *action); enum pcmk__action_type pcmk__parse_action(const char *action_name); const char *pcmk__action_text(enum pcmk__action_type action); const char *pcmk__on_fail_text(enum pcmk__on_fail on_fail); /*! * \internal * \brief Get a human-friendly action name * * \param[in] action_name Actual action name * \param[in] interval_ms Action interval (in milliseconds) * * \return Action name suitable for display */ static inline const char * pcmk__readable_action(const char *action_name, guint interval_ms) { if ((interval_ms == 0) && pcmk__str_eq(action_name, PCMK_ACTION_MONITOR, pcmk__str_none)) { return "probe"; } return action_name; } +/*! + * \internal + * \brief Check whether an action raises a resource's role + * + * \param[in] action Action to check + * + * \return \c true if \p action is "start", "promote", or "migrate_from" + */ +static inline bool +pcmk__is_up_action(const char *action) +{ + return pcmk__str_any_of(action, PCMK_ACTION_START, PCMK_ACTION_PROMOTE, + PCMK_ACTION_MIGRATE_FROM, NULL); +} + #ifdef __cplusplus } #endif #endif // PCMK__CRM_COMMON_ACTIONS_INTERNAL__H diff --git a/lib/pacemaker/pcmk_sched_ordering.c b/lib/pacemaker/pcmk_sched_ordering.c index 5bc2ffc919..185b44fac1 100644 --- a/lib/pacemaker/pcmk_sched_ordering.c +++ b/lib/pacemaker/pcmk_sched_ordering.c @@ -1,1496 +1,1495 @@ /* * Copyright 2004-2024 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->priv->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; } pcmk__config_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, PCMK_XA_KIND); if (kind == NULL) { const char *score = crm_element_value(xml_obj, PCMK_XA_SCORE); kind_e = pe_order_kind_mandatory; if (score) { // @COMPAT deprecated informally since 1.0.7, formally since 2.0.1 int score_i = 0; (void) pcmk_parse_score(score, &score_i, 0); if (score_i == 0) { kind_e = pe_order_kind_optional; } pcmk__warn_once(pcmk__wo_order_score, "Support for '" PCMK_XA_SCORE "' in " PCMK_XE_RSC_ORDER " is deprecated and will be " "removed in a future release " "(use '" PCMK_XA_KIND "' instead)"); } } else if (pcmk__str_eq(kind, PCMK_VALUE_MANDATORY, pcmk__str_none)) { kind_e = pe_order_kind_mandatory; } else if (pcmk__str_eq(kind, PCMK_VALUE_OPTIONAL, pcmk__str_none)) { kind_e = pe_order_kind_optional; } else if (pcmk__str_eq(kind, PCMK_VALUE_SERIALIZE, pcmk__str_none)) { kind_e = pe_order_kind_serialize; } else { pcmk__config_err("Resetting '" PCMK_XA_KIND "' for constraint %s to " "'" PCMK_VALUE_MANDATORY "' because '%s' is not valid", pcmk__s(pcmk__xe_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 \c PCMK_XA_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, PCMK_XA_KIND) != NULL) || (crm_element_value(xml_obj, PCMK_XA_SCORE) != NULL)) { kind = get_ordering_type(xml_obj); } // Check ordering XML (and parent) for explicit PCMK_XA_SYMMETRICAL setting rc = pcmk__xe_get_bool_attr(xml_obj, PCMK_XA_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 " PCMK_XA_SYMMETRICAL " for '%s' because not valid with " PCMK_XA_KIND " of '" PCMK_VALUE_SERIALIZE "'", pcmk__xe_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: pcmk__set_relation_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. */ pcmk__set_relation_flags(flags, pcmk__ar_serialize); break; case pe_order_kind_mandatory: pcmk__set_relation_flags(flags, pcmk__ar_ordered); switch (symmetry) { case ordering_asymmetric: pcmk__set_relation_flags(flags, pcmk__ar_asymmetric); break; case ordering_symmetric: pcmk__set_relation_flags(flags, pcmk__ar_first_implies_then); - if (pcmk__strcase_any_of(first, PCMK_ACTION_START, - PCMK_ACTION_PROMOTE, NULL)) { + if (pcmk__is_up_action(first)) { pcmk__set_relation_flags(flags, pcmk__ar_unrunnable_first_blocks); } break; case ordering_symmetric_inverse: pcmk__set_relation_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] 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 pcmk_scheduler_t *scheduler) { pcmk_resource_t *rsc = NULL; const char *rsc_id = crm_element_value(xml, resource_attr); if (rsc_id == NULL) { pcmk__config_err("Ignoring constraint '%s' without %s", pcmk__xe_id(xml), resource_attr); return NULL; } rsc = pcmk__find_constraint_resource(scheduler->priv->resources, rsc_id); if (rsc == NULL) { pcmk__config_err("Ignoring constraint '%s' because resource '%s' " "does not exist", pcmk__xe_id(xml), rsc_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 (!pcmk__is_clone(rsc)) { return 0; } clone_min = g_hash_table_lookup(rsc->priv->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: * PCMK_XA_REQUIRE_ALL=PCMK_VALUE_FALSE is deprecated equivalent of * PCMK_META_CLONE_MIN=1 */ if (pcmk__xe_get_bool_attr(xml, PCMK_XA_REQUIRE_ALL, &require_all) != ENODATA) { pcmk__warn_once(pcmk__wo_require_all, "Support for " PCMK_XA_REQUIRE_ALL " in ordering " "constraints is deprecated and will be removed in a " "future release (use " PCMK_META_CLONE_MIN " clone " "meta-attribute instead)"); if (!require_all) { return 1; } } return 0; } /*! * \internal * \brief Create orderings for a constraint with \c PCMK_META_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->priv->scheduler); 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; // Order the actions for each clone instance before the pseudo-action for (GList *iter = rsc_first->priv->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->priv->scheduler); } // 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->priv->scheduler); } /*! * \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); 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, PCMK_XA_ID); if (id == NULL) { pcmk__config_err("Ignoring <%s> constraint without " PCMK_XA_ID, xml_obj->name); return; } rsc_first = get_ordering_resource(xml_obj, PCMK_XA_FIRST, scheduler); if (rsc_first == NULL) { return; } rsc_then = get_ordering_resource(xml_obj, PCMK_XA_THEN, scheduler); if (rsc_then == NULL) { return; } action_first = crm_element_value(xml_obj, PCMK_XA_FIRST_ACTION); if (action_first == NULL) { action_first = PCMK_ACTION_START; } action_then = crm_element_value(xml_obj, PCMK_XA_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); /* 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 = pcmk__assert_alloc(1, sizeof(pcmk__action_relation_t)); order->id = sched->priv->next_ordering_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; } pcmk__rsc_trace(first_rsc, "Created ordering %d for %s then %s", (sched->priv->next_ordering_id - 1), pcmk__s(order->task1, "an underspecified action"), pcmk__s(order->task2, "an underspecified action")); sched->priv->ordering_constraints = g_list_prepend(sched->priv->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 \c PCMK_XE_RSC_ORDER XML \c PCMK_XA_KIND * attribute * \param[in] parent_symmetrical_s \c PCMK_XE_RSC_ORDER XML * \c PCMK_XA_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 = pcmk__xe_id(set); const char *action = crm_element_value(set, PCMK_XA_ACTION); const char *sequential_s = crm_element_value(set, PCMK_XA_SEQUENTIAL); const char *kind_s = crm_element_value(set, PCMK_XA_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 = pcmk__xe_first_child(set, PCMK_XE_RESOURCE_REF, NULL, NULL); xml_rsc != NULL; xml_rsc = pcmk__xe_next(xml_rsc, PCMK_XE_RESOURCE_REF)) { EXPAND_CONSTRAINT_IDREF(id, resource, pcmk__xe_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, PCMK_XA_ACTION); const char *action_2 = crm_element_value(set2, PCMK_XA_ACTION); uint32_t flags = pcmk__ar_none; bool require_all = true; (void) pcmk__xe_get_bool_attr(set1, PCMK_XA_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", pcmk__xe_id(set1)); pcmk_action_t *unordered_action = get_pseudo_op(task, scheduler); free(task); unordered_action->required_runnable_before = 1; for (xml_rsc = pcmk__xe_first_child(set1, PCMK_XE_RESOURCE_REF, NULL, NULL); xml_rsc != NULL; xml_rsc = pcmk__xe_next(xml_rsc, PCMK_XE_RESOURCE_REF)) { EXPAND_CONSTRAINT_IDREF(id, rsc_1, pcmk__xe_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 = pcmk__xe_first_child(set2, PCMK_XE_RESOURCE_REF, NULL, NULL); xml_rsc_2 != NULL; xml_rsc_2 = pcmk__xe_next(xml_rsc_2, PCMK_XE_RESOURCE_REF)) { EXPAND_CONSTRAINT_IDREF(id, rsc_2, pcmk__xe_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, PCMK_XA_SEQUENTIAL)) { if (symmetry == ordering_symmetric_inverse) { // Get the first one xml_rsc = pcmk__xe_first_child(set1, PCMK_XE_RESOURCE_REF, NULL, NULL); if (xml_rsc != NULL) { EXPAND_CONSTRAINT_IDREF(id, rsc_1, pcmk__xe_id(xml_rsc)); } } else { // Get the last one const char *rid = NULL; for (xml_rsc = pcmk__xe_first_child(set1, PCMK_XE_RESOURCE_REF, NULL, NULL); xml_rsc != NULL; xml_rsc = pcmk__xe_next(xml_rsc, PCMK_XE_RESOURCE_REF)) { rid = pcmk__xe_id(xml_rsc); } EXPAND_CONSTRAINT_IDREF(id, rsc_1, rid); } } if (pcmk__xe_attr_is_true(set2, PCMK_XA_SEQUENTIAL)) { if (symmetry == ordering_symmetric_inverse) { // Get the last one const char *rid = NULL; for (xml_rsc = pcmk__xe_first_child(set2, PCMK_XE_RESOURCE_REF, NULL, NULL); xml_rsc != NULL; xml_rsc = pcmk__xe_next(xml_rsc, PCMK_XE_RESOURCE_REF)) { rid = pcmk__xe_id(xml_rsc); } EXPAND_CONSTRAINT_IDREF(id, rsc_2, rid); } else { // Get the first one xml_rsc = pcmk__xe_first_child(set2, PCMK_XE_RESOURCE_REF, NULL, NULL); if (xml_rsc != NULL) { EXPAND_CONSTRAINT_IDREF(id, rsc_2, pcmk__xe_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 = pcmk__xe_first_child(set2, PCMK_XE_RESOURCE_REF, NULL, NULL); xml_rsc != NULL; xml_rsc = pcmk__xe_next(xml_rsc, PCMK_XE_RESOURCE_REF)) { EXPAND_CONSTRAINT_IDREF(id, rsc_2, pcmk__xe_id(xml_rsc)); pcmk__order_resource_actions(rsc_1, action_1, rsc_2, action_2, flags); } } else if (rsc_2 != NULL) { for (xml_rsc = pcmk__xe_first_child(set1, PCMK_XE_RESOURCE_REF, NULL, NULL); xml_rsc != NULL; xml_rsc = pcmk__xe_next(xml_rsc, PCMK_XE_RESOURCE_REF)) { EXPAND_CONSTRAINT_IDREF(id, rsc_1, pcmk__xe_id(xml_rsc)); pcmk__order_resource_actions(rsc_1, action_1, rsc_2, action_2, flags); } } else { for (xml_rsc = pcmk__xe_first_child(set1, PCMK_XE_RESOURCE_REF, NULL, NULL); xml_rsc != NULL; xml_rsc = pcmk__xe_next(xml_rsc, PCMK_XE_RESOURCE_REF)) { EXPAND_CONSTRAINT_IDREF(id, rsc_1, pcmk__xe_id(xml_rsc)); for (xmlNode *xml_rsc_2 = pcmk__xe_first_child(set2, PCMK_XE_RESOURCE_REF, NULL, NULL); xml_rsc_2 != NULL; xml_rsc_2 = pcmk__xe_next(xml_rsc_2, PCMK_XE_RESOURCE_REF)) { EXPAND_CONSTRAINT_IDREF(id, rsc_2, pcmk__xe_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__idref_t *tag_first = NULL; pcmk__idref_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 " PCMK_XE_RSC_ORDER); return pcmk_rc_ok; } id_first = crm_element_value(xml_obj, PCMK_XA_FIRST); id_then = crm_element_value(xml_obj, PCMK_XA_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", pcmk__xe_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", pcmk__xe_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, PCMK_XA_FIRST_ACTION); action_then = crm_element_value(xml_obj, PCMK_XA_THEN_ACTION); *expanded_xml = pcmk__xml_copy(NULL, xml_obj); /* Convert template/tag reference in PCMK_XA_FIRST into constraint * PCMK_XE_RESOURCE_SET */ if (!pcmk__tag_to_set(*expanded_xml, &rsc_set_first, PCMK_XA_FIRST, true, scheduler)) { pcmk__xml_free(*expanded_xml); *expanded_xml = NULL; return pcmk_rc_unpack_error; } if (rsc_set_first != NULL) { if (action_first != NULL) { /* Move PCMK_XA_FIRST_ACTION into converted PCMK_XE_RESOURCE_SET as * PCMK_XA_ACTION */ crm_xml_add(rsc_set_first, PCMK_XA_ACTION, action_first); pcmk__xe_remove_attr(*expanded_xml, PCMK_XA_FIRST_ACTION); } any_sets = true; } /* Convert template/tag reference in PCMK_XA_THEN into constraint * PCMK_XE_RESOURCE_SET */ if (!pcmk__tag_to_set(*expanded_xml, &rsc_set_then, PCMK_XA_THEN, true, scheduler)) { pcmk__xml_free(*expanded_xml); *expanded_xml = NULL; return pcmk_rc_unpack_error; } if (rsc_set_then != NULL) { if (action_then != NULL) { /* Move PCMK_XA_THEN_ACTION into converted PCMK_XE_RESOURCE_SET as * PCMK_XA_ACTION */ crm_xml_add(rsc_set_then, PCMK_XA_ACTION, action_then); pcmk__xe_remove_attr(*expanded_xml, PCMK_XA_THEN_ACTION); } any_sets = true; } if (any_sets) { crm_log_xml_trace(*expanded_xml, "Expanded " PCMK_XE_RSC_ORDER); } else { pcmk__xml_free(*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, PCMK_XA_ID); const char *invert = crm_element_value(xml_obj, PCMK_XA_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 = pcmk__xe_first_child(xml_obj, PCMK_XE_RESOURCE_SET, NULL, NULL); set != NULL; set = pcmk__xe_next(set, PCMK_XE_RESOURCE_SET)) { set = pcmk__xe_resolve_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) { pcmk__xml_free(expanded_xml); } return; } if (last != NULL) { if (order_rsc_sets(id, last, set, kind, scheduler, symmetry) != pcmk_rc_ok) { if (expanded_xml != NULL) { pcmk__xml_free(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) { pcmk__xml_free(expanded_xml); } return; } } last = set; } if (expanded_xml) { pcmk__xml_free(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->flags, pcmk__ar_guest_allowed) && (input->action->rsc != NULL) && pcmk__rsc_corresponds_to_guest(action->rsc, input->action->node)) { pcmk__config_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 ((input->flags == 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->priv->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->flags = 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->priv->scheduler->priv->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 (!pcmk__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)) { pcmk__rsc_trace(action->rsc, "Not ordering %s before shutdown of %s because " "resource in maintenance mode", action->uuid, pcmk__node_name(node)); continue; } else if (node->details->maintenance) { pcmk__rsc_trace(action->rsc, "Not ordering %s before shutdown of %s because " "node in maintenance mode", action->uuid, pcmk__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)) { pcmk__rsc_trace(action->rsc, "Not ordering %s before shutdown of %s because " "resource is unmanaged or blocked", action->uuid, pcmk__node_name(node)); continue; } pcmk__rsc_trace(action->rsc, "Ordering %s before shutdown of %s", action->uuid, pcmk__node_name(node)); pcmk__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->priv->scheduler); } } /*! * \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->priv->actions, original_key, NULL); if (list == NULL) { // Search again using this resource's ID char *key = NULL; char *task = NULL; guint interval_ms = 0; CRM_CHECK(parse_op_key(original_key, NULL, &task, &interval_ms), return NULL); key = pcmk__op_key(rsc->id, task, interval_ms); list = find_actions(rsc->priv->actions, key, NULL); free(key); free(task); } 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; pcmk__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) { pcmk__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)) { pcmk__rsc_trace(rsc, "Detected dangling migration ordering (%s then %s %s)", first_action->uuid, order->task2, rsc->id); pcmk__clear_relation_flags(flags, pcmk__ar_first_implies_then); } if ((first_action == NULL) && !pcmk_is_set(flags, pcmk__ar_first_implies_then)) { 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 { pcmk__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; pcmk__assert(first_rsc != NULL); pcmk__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)) { pcmk__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; enum rsc_role_e first_role; parse_op_key(order->task1, NULL, &op_type, &interval_ms); key = pcmk__op_key(first_rsc->id, op_type, interval_ms); first_role = first_rsc->priv->fns->state(first_rsc, TRUE); if ((first_role == pcmk_role_stopped) && pcmk__str_eq(op_type, PCMK_ACTION_STOP, pcmk__str_none)) { free(key); pcmk__rsc_trace(first_rsc, "Ignoring constraint %d: first (%s for %s) " "not found", order->id, order->task1, first_rsc->id); } else if ((first_role == pcmk_role_unpromoted) && pcmk__str_eq(op_type, PCMK_ACTION_DEMOTE, pcmk__str_none)) { free(key); pcmk__rsc_trace(first_rsc, "Ignoring constraint %d: first (%s for %s) " "not found", order->id, order->task1, first_rsc->id); } else { pcmk__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->priv->scheduler); first_actions = g_list_prepend(NULL, first_action); } free(op_type); } if (then_rsc == NULL) { if (order->action2 == NULL) { pcmk__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->priv->ordering_constraints = g_list_reverse(sched->priv->ordering_constraints); for (GList *iter = sched->priv->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->priv->actions, block_colocation_dependents, NULL); crm_trace("Ordering probes"); pcmk__order_probes(sched); crm_trace("Updating %d actions", g_list_length(sched->priv->actions)); g_list_foreach(sched->priv->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, pcmk__node_name(before->node), after_desc, pcmk__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); }