daemons/controld/controld

/* */
This source file includes following definitions.
crmd_ha_msg_filter
node_alive
peer_update_callback
crm_fsa_trigger
   1 /*
   2  * Copyright 2004-2023 the Pacemaker project contributors
   3  *
   4  * The version control history for this file may have further details.
   5  *
   6  * This source code is licensed under the GNU General Public License version 2
   7  * or later (GPLv2+) WITHOUT ANY WARRANTY.
   8  */
   9 
  10 #include <crm_internal.h>
  11 
  12 #include <sys/param.h>
  13 #include <string.h>
  14 
  15 #include <crm/crm.h>
  16 #include <crm/msg_xml.h>
  17 #include <crm/common/xml.h>
  18 #include <crm/cluster.h>
  19 #include <crm/cib.h>
  20 
  21 #include <pacemaker-controld.h>
  22 
  23 /* From join_dc... */
  24 extern gboolean check_join_state(enum crmd_fsa_state cur_state, const char *source);
  25 
  26 void
  27 crmd_ha_msg_filter(xmlNode * msg)
     /*  */
  28 {
  29     if (AM_I_DC) {
  30         const char *sys_from = crm_element_value(msg, F_CRM_SYS_FROM);
  31 
  32         if (pcmk__str_eq(sys_from, CRM_SYSTEM_DC, pcmk__str_casei)) {
  33             const char *from = crm_element_value(msg, F_ORIG);
  34 
  35             if (!pcmk__str_eq(from, controld_globals.our_nodename,
  36                               pcmk__str_casei)) {
  37                 int level = LOG_INFO;
  38                 const char *op = crm_element_value(msg, F_CRM_TASK);
  39 
  40                 /* make sure the election happens NOW */
  41                 if (controld_globals.fsa_state != S_ELECTION) {
  42                     ha_msg_input_t new_input;
  43 
  44                     level = LOG_WARNING;
  45                     new_input.msg = msg;
  46                     register_fsa_error_adv(C_FSA_INTERNAL, I_ELECTION, NULL, &new_input,
  47                                            __func__);
  48                 }
  49 
  50                 do_crm_log(level, "Another DC detected: %s (op=%s)", from, op);
  51                 goto done;
  52             }
  53         }
  54 
  55     } else {
  56         const char *sys_to = crm_element_value(msg, F_CRM_SYS_TO);
  57 
  58         if (pcmk__str_eq(sys_to, CRM_SYSTEM_DC, pcmk__str_casei)) {
  59             return;
  60         }
  61     }
  62 
  63     /* crm_log_xml_trace(msg, "HA[inbound]"); */
  64     route_message(C_HA_MESSAGE, msg);
  65 
  66   done:
  67     controld_trigger_fsa();
  68 }
  69 
  70 /*!
  71  * \internal
  72  * \brief Check whether a node is online
  73  *
  74  * \param[in] node  Node to check
  75  *
  76  * \retval -1 if completely dead
  77  * \retval  0 if partially alive
  78  * \retval  1 if completely alive
  79  */
  80 static int
  81 node_alive(const crm_node_t *node)
     /*  */
  82 {
  83     if (pcmk_is_set(node->flags, crm_remote_node)) {
  84         // Pacemaker Remote nodes can't be partially alive
  85         return pcmk__str_eq(node->state, CRM_NODE_MEMBER, pcmk__str_casei) ? 1: -1;
  86 
  87     } else if (crm_is_peer_active(node)) {
  88         // Completely up cluster node: both cluster member and peer
  89         return 1;
  90 
  91     } else if (!pcmk_is_set(node->processes, crm_get_cluster_proc())
  92                && !pcmk__str_eq(node->state, CRM_NODE_MEMBER, pcmk__str_casei)) {
  93         // Completely down cluster node: neither cluster member nor peer
  94         return -1;
  95     }
  96 
  97     // Partially up cluster node: only cluster member or only peer
  98     return 0;
  99 }
 100 
 101 #define state_text(state) ((state)? (const char *)(state) : "in unknown state")
 102 
 103 void
 104 peer_update_callback(enum crm_status_type type, crm_node_t * node, const void *data)
     /*  */
 105 {
 106     uint32_t old = 0;
 107     bool appeared = FALSE;
 108     bool is_remote = pcmk_is_set(node->flags, crm_remote_node);
 109 
 110     controld_node_pending_timer(node);
 111 
 112     /* The controller waits to receive some information from the membership
 113      * layer before declaring itself operational. If this is being called for a
 114      * cluster node, indicate that we have it.
 115      */
 116     if (!is_remote) {
 117         controld_set_fsa_input_flags(R_PEER_DATA);
 118     }
 119 
 120     if (type == crm_status_processes
 121         && pcmk_is_set(node->processes, crm_get_cluster_proc())
 122         && !AM_I_DC
 123         && !is_remote) {
 124         /*
 125          * This is a hack until we can send to a nodeid and/or we fix node name lookups
 126          * These messages are ignored in crmd_ha_msg_filter()
 127          */
 128         xmlNode *query = create_request(CRM_OP_HELLO, NULL, NULL, CRM_SYSTEM_CRMD, CRM_SYSTEM_CRMD, NULL);
 129 
 130         crm_debug("Sending hello to node %u so that it learns our node name", node->id);
 131         send_cluster_message(node, crm_msg_crmd, query, FALSE);
 132 
 133         free_xml(query);
 134     }
 135 
 136     if (node->uname == NULL) {
 137         return;
 138     }
 139 
 140     switch (type) {
 141         case crm_status_uname:
 142             /* If we've never seen the node, then it also won't be in the status section */
 143             crm_info("%s node %s is now %s",
 144                      (is_remote? "Remote" : "Cluster"),
 145                      node->uname, state_text(node->state));
 146             return;
 147 
 148         case crm_status_nstate:
 149             /* This callback should not be called unless the state actually
 150              * changed, but here's a failsafe just in case.
 151              */
 152             CRM_CHECK(!pcmk__str_eq(data, node->state, pcmk__str_casei),
 153                       return);
 154 
 155             crm_info("%s node %s is now %s (was %s)",
 156                      (is_remote? "Remote" : "Cluster"),
 157                      node->uname, state_text(node->state), state_text(data));
 158 
 159             if (pcmk__str_eq(CRM_NODE_MEMBER, node->state, pcmk__str_casei)) {
 160                 appeared = TRUE;
 161                 if (!is_remote) {
 162                     remove_stonith_cleanup(node->uname);
 163                 }
 164             } else {
 165                 controld_remove_failed_sync_node(node->uname);
 166                 controld_remove_voter(node->uname);
 167             }
 168 
 169             crmd_alert_node_event(node);
 170             break;
 171 
 172         case crm_status_processes:
 173             CRM_CHECK(data != NULL, return);
 174             old = *(const uint32_t *)data;
 175             appeared = pcmk_is_set(node->processes, crm_get_cluster_proc());
 176 
 177             {
 178                 const char *dc_s = controld_globals.dc_name;
 179 
 180                 if ((dc_s == NULL) && AM_I_DC) {
 181                     dc_s = "true";
 182                 }
 183 
 184                 crm_info("Node %s is %s a peer " CRM_XS
 185                          " DC=%s old=%#07x new=%#07x",
 186                          node->uname, (appeared? "now" : "no longer"),
 187                          pcmk__s(dc_s, "<none>"), old, node->processes);
 188             }
 189 
 190             if (!pcmk_is_set((node->processes ^ old), crm_get_cluster_proc())) {
 191                 /* Peer status did not change. This should not be possible,
 192                  * since we don't track process flags other than peer status.
 193                  */
 194                 crm_trace("Process flag %#7x did not change from %#7x to %#7x",
 195                           crm_get_cluster_proc(), old, node->processes);
 196                 return;
 197 
 198             }
 199 
 200             if (!appeared) {
 201                 node->peer_lost = time(NULL);
 202                 controld_remove_failed_sync_node(node->uname);
 203                 controld_remove_voter(node->uname);
 204             }
 205 
 206             if (!pcmk_is_set(controld_globals.fsa_input_register,
 207                              R_CIB_CONNECTED)) {
 208                 crm_trace("Ignoring peer status change because not connected to CIB");
 209                 return;
 210 
 211             } else if (controld_globals.fsa_state == S_STOPPING) {
 212                 crm_trace("Ignoring peer status change because stopping");
 213                 return;
 214             }
 215 
 216             if (!appeared
 217                 && pcmk__str_eq(node->uname, controld_globals.our_nodename,
 218                                 pcmk__str_casei)) {
 219                 /* Did we get evicted? */
 220                 crm_notice("Our peer connection failed");
 221                 register_fsa_input(C_CRMD_STATUS_CALLBACK, I_ERROR, NULL);
 222 
 223             } else if (pcmk__str_eq(node->uname, controld_globals.dc_name,
 224                                     pcmk__str_casei)
 225                        && !crm_is_peer_active(node)) {
 226                 /* Did the DC leave us? */
 227                 crm_notice("Our peer on the DC (%s) is dead",
 228                            controld_globals.dc_name);
 229                 register_fsa_input(C_CRMD_STATUS_CALLBACK, I_ELECTION, NULL);
 230 
 231                 /* @COMPAT DC < 1.1.13: If a DC shuts down normally, we don't
 232                  * want to fence it. Newer DCs will send their shutdown request
 233                  * to all peers, who will update the DC's expected state to
 234                  * down, thus avoiding fencing. We can safely erase the DC's
 235                  * transient attributes when it leaves in that case. However,
 236                  * the only way to avoid fencing older DCs is to leave the
 237                  * transient attributes intact until it rejoins.
 238                  */
 239                 if (compare_version(controld_globals.dc_version, "3.0.9") > 0) {
 240                     controld_delete_node_state(node->uname,
 241                                                controld_section_attrs,
 242                                                cib_scope_local);
 243                 }
 244 
 245             } else if (AM_I_DC
 246                        || pcmk_is_set(controld_globals.flags, controld_dc_left)
 247                        || (controld_globals.dc_name == NULL)) {
 248                 /* This only needs to be done once, so normally the DC should do
 249                  * it. However if there is no DC, every node must do it, since
 250                  * there is no other way to ensure some one node does it.
 251                  */
 252                 if (appeared) {
 253                     te_trigger_stonith_history_sync(FALSE);
 254                 } else {
 255                     controld_delete_node_state(node->uname,
 256                                                controld_section_attrs,
 257                                                cib_scope_local);
 258                 }
 259             }
 260             break;
 261     }
 262 
 263     if (AM_I_DC) {
 264         xmlNode *update = NULL;
 265         int flags = node_update_peer;
 266         int alive = node_alive(node);
 267         pcmk__graph_action_t *down = match_down_event(node->uuid);
 268 
 269         crm_trace("Alive=%d, appeared=%d, down=%d",
 270                   alive, appeared, (down? down->id : -1));
 271 
 272         if (appeared && (alive > 0) && !is_remote) {
 273             register_fsa_input_before(C_FSA_INTERNAL, I_NODE_JOIN, NULL);
 274         }
 275 
 276         if (down) {
 277             const char *task = crm_element_value(down->xml, XML_LRM_ATTR_TASK);
 278 
 279             if (pcmk__str_eq(task, PCMK_ACTION_STONITH, pcmk__str_casei)) {
 280 
 281                 /* tengine_stonith_callback() confirms fence actions */
 282                 crm_trace("Updating CIB %s fencer reported fencing of %s complete",
 283                           (pcmk_is_set(down->flags, pcmk__graph_action_confirmed)? "after" : "before"), node->uname);
 284 
 285             } else if (!appeared && pcmk__str_eq(task, PCMK_ACTION_DO_SHUTDOWN,
 286                                                  pcmk__str_casei)) {
 287 
 288                 // Shutdown actions are immediately confirmed (i.e. no_wait)
 289                 if (!is_remote) {
 290                     flags |= node_update_join | node_update_expected;
 291                     crmd_peer_down(node, FALSE);
 292                     check_join_state(controld_globals.fsa_state, __func__);
 293                 }
 294                 if (alive >= 0) {
 295                     crm_info("%s of peer %s is in progress " CRM_XS " action=%d",
 296                              task, node->uname, down->id);
 297                 } else {
 298                     crm_notice("%s of peer %s is complete " CRM_XS " action=%d",
 299                                task, node->uname, down->id);
 300                     pcmk__update_graph(controld_globals.transition_graph, down);
 301                     trigger_graph();
 302                 }
 303 
 304             } else {
 305                 crm_trace("Node %s is %s, was expected to %s (op %d)",
 306                           node->uname,
 307                           ((alive > 0)? "alive" :
 308                            ((alive < 0)? "dead" : "partially alive")),
 309                           task, down->id);
 310             }
 311 
 312         } else if (appeared == FALSE) {
 313             if ((controld_globals.transition_graph == NULL)
 314                 || (controld_globals.transition_graph->id == -1)) {
 315                 crm_info("Stonith/shutdown of node %s is unknown to the "
 316                          "current DC", node->uname);
 317             } else {
 318                 crm_warn("Stonith/shutdown of node %s was not expected",
 319                          node->uname);
 320             }
 321             if (!is_remote) {
 322                 crm_update_peer_join(__func__, node, crm_join_none);
 323                 check_join_state(controld_globals.fsa_state, __func__);
 324             }
 325             abort_transition(INFINITY, pcmk__graph_restart, "Node failure",
 326                              NULL);
 327             fail_incompletable_actions(controld_globals.transition_graph,
 328                                        node->uuid);
 329 
 330         } else {
 331             crm_trace("Node %s came up, was not expected to be down",
 332                       node->uname);
 333         }
 334 
 335         if (is_remote) {
 336             /* A pacemaker_remote node won't have its cluster status updated
 337              * in the CIB by membership-layer callbacks, so do it here.
 338              */
 339             flags |= node_update_cluster;
 340 
 341             /* Trigger resource placement on newly integrated nodes */
 342             if (appeared) {
 343                 abort_transition(INFINITY, pcmk__graph_restart,
 344                                  "Pacemaker Remote node integrated", NULL);
 345             }
 346         }
 347 
 348         if (!appeared && (type == crm_status_processes)
 349             && (node->when_member > 1)) {
 350             /* The node left CPG but is still a cluster member. Set its
 351              * membership time to 1 to record it in the cluster state as a
 352              * boolean, so we don't fence it due to node-pending-timeout.
 353              */
 354             node->when_member = 1;
 355             flags |= node_update_cluster;
 356             controld_node_pending_timer(node);
 357         }
 358 
 359         /* Update the CIB node state */
 360         update = create_node_state_update(node, flags, NULL, __func__);
 361         if (update == NULL) {
 362             crm_debug("Node state update not yet possible for %s", node->uname);
 363         } else {
 364             fsa_cib_anon_update(XML_CIB_TAG_STATUS, update);
 365         }
 366         free_xml(update);
 367     }
 368 
 369     controld_trigger_fsa();
 370 }
 371 
 372 gboolean
 373 crm_fsa_trigger(gpointer user_data)
     /*  */
 374 {
 375     crm_trace("Invoked (queue len: %d)",
 376               g_list_length(controld_globals.fsa_message_queue));
 377     s_crmd_fsa(C_FSA_INTERNAL);
 378     crm_trace("Exited  (queue len: %d)",
 379               g_list_length(controld_globals.fsa_message_queue));
 380     return TRUE;
 381 }
/* */
root/daemons/controld/controld_callbacks.c

DEFINITIONS
