daemons/controld/controld

/* */
This source file includes following definitions.
controld_shutdown_schedulerd_ipc
save_cib_contents
handle_disconnect
handle_reply
scheduler_event_callback
new_schedulerd_ipc_connection
do_pe_control
controld_sched_timeout
controld_stop_sched_timer
controld_expect_sched_reply
controld_free_sched_timer
do_pe_invoke
force_local_option
do_pe_invoke_callback
   1 /*
   2  * Copyright 2004-2024 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 <unistd.h>  /* pid_t, sleep, ssize_t */
  13 
  14 #include <crm/cib.h>
  15 #include <crm/cluster.h>
  16 #include <crm/common/xml.h>
  17 #include <crm/crm.h>
  18 #include <crm/common/xml_internal.h>
  19 #include <crm/common/ipc.h>
  20 #include <crm/common/ipc_schedulerd.h>
  21 
  22 #include <pacemaker-controld.h>
  23 
  24 static void handle_disconnect(void);
  25 
  26 static pcmk_ipc_api_t *schedulerd_api = NULL;
  27 
  28 /*!
  29  * \internal
  30  * \brief Close any scheduler connection and free associated memory
  31  */
  32 void
  33 controld_shutdown_schedulerd_ipc(void)
     /*  */
  34 {
  35     controld_clear_fsa_input_flags(R_PE_REQUIRED);
  36     pcmk_disconnect_ipc(schedulerd_api);
  37     handle_disconnect();
  38 
  39     pcmk_free_ipc_api(schedulerd_api);
  40     schedulerd_api = NULL;
  41 }
  42 
  43 /*!
  44  * \internal
  45  * \brief Save CIB query result to file, raising FSA error
  46  *
  47  * \param[in] msg        Ignored
  48  * \param[in] call_id    Call ID of CIB query
  49  * \param[in] rc         Return code of CIB query
  50  * \param[in] output     Result of CIB query
  51  * \param[in] user_data  Unique identifier for filename
  52  *
  53  * \note This is intended to be called after a scheduler connection fails.
  54  */
  55 static void
  56 save_cib_contents(xmlNode *msg, int call_id, int rc, xmlNode *output,
     /*  */
  57                   void *user_data)
  58 {
  59     const char *id = user_data;
  60 
  61     register_fsa_error_adv(C_FSA_INTERNAL, I_ERROR, NULL, NULL, __func__);
  62     CRM_CHECK(id != NULL, return);
  63 
  64     if (rc == pcmk_ok) {
  65         char *filename = crm_strdup_printf(PCMK_SCHEDULER_INPUT_DIR "/pe-core-%s.bz2", id);
  66 
  67         if (pcmk__xml_write_file(output, filename, true) != pcmk_rc_ok) {
  68             crm_err("Could not save Cluster Information Base to %s after scheduler crash",
  69                     filename);
  70         } else {
  71             crm_notice("Saved Cluster Information Base to %s after scheduler crash",
  72                        filename);
  73         }
  74         free(filename);
  75     }
  76 }
  77 
  78 /*!
  79  * \internal
  80  * \brief Respond to scheduler connection failure
  81  */
  82 static void
  83 handle_disconnect(void)
     /*  */
  84 {
  85     // If we aren't connected to the scheduler, we can't expect a reply
  86     controld_expect_sched_reply(NULL);
  87 
  88     if (pcmk_is_set(controld_globals.fsa_input_register, R_PE_REQUIRED)) {
  89         int rc = pcmk_ok;
  90         char *uuid_str = crm_generate_uuid();
  91 
  92         crm_crit("Lost connection to the scheduler "
  93                  QB_XS " CIB will be saved to " PCMK_SCHEDULER_INPUT_DIR "/pe-core-%s.bz2",
  94                  uuid_str);
  95 
  96         /*
  97          * The scheduler died...
  98          *
  99          * Save the current CIB so that we have a chance of
 100          * figuring out what killed it.
 101          *
 102          * Delay raising the I_ERROR until the query below completes or
 103          * 5s is up, whichever comes first.
 104          *
 105          */
 106         rc = controld_globals.cib_conn->cmds->query(controld_globals.cib_conn,
 107                                                     NULL, NULL, cib_none);
 108         fsa_register_cib_callback(rc, uuid_str, save_cib_contents);
 109     }
 110 
 111     controld_clear_fsa_input_flags(R_PE_CONNECTED);
 112     controld_trigger_fsa();
 113     return;
 114 }
 115 
 116 static void
 117 handle_reply(pcmk_schedulerd_api_reply_t *reply)
     /*  */
 118 {
 119     const char *msg_ref = NULL;
 120 
 121     if (!AM_I_DC) {
 122         return;
 123     }
 124 
 125     msg_ref = reply->data.graph.reference;
 126 
 127     if (msg_ref == NULL) {
 128         crm_err("%s - Ignoring calculation with no reference", CRM_OP_PECALC);
 129 
 130     } else if (pcmk__str_eq(msg_ref, controld_globals.fsa_pe_ref,
 131                             pcmk__str_none)) {
 132         ha_msg_input_t fsa_input;
 133         xmlNode *crm_data_node;
 134 
 135         controld_stop_sched_timer();
 136 
 137         /* do_te_invoke (which will eventually process the fsa_input we are constructing
 138          * here) requires that fsa_input.xml be non-NULL.  That will only happen if
 139          * copy_ha_msg_input (which is called by register_fsa_input_adv) sees the
 140          * fsa_input.msg that it is expecting. The scheduler's IPC dispatch function
 141          * gave us the values we need, we just need to put them into XML.
 142          *
 143          * The name of the top level element here is irrelevant.  Nothing checks it.
 144          */
 145         fsa_input.msg = pcmk__xe_create(NULL, "dummy-reply");
 146         crm_xml_add(fsa_input.msg, PCMK_XA_REFERENCE, msg_ref);
 147         crm_xml_add(fsa_input.msg, PCMK__XA_CRM_TGRAPH_IN,
 148                     reply->data.graph.input);
 149 
 150         crm_data_node = pcmk__xe_create(fsa_input.msg, PCMK__XE_CRM_XML);
 151         pcmk__xml_copy(crm_data_node, reply->data.graph.tgraph);
 152         register_fsa_input_later(C_IPC_MESSAGE, I_PE_SUCCESS, &fsa_input);
 153 
 154         pcmk__xml_free(fsa_input.msg);
 155 
 156     } else {
 157         crm_info("%s calculation %s is obsolete", CRM_OP_PECALC, msg_ref);
 158     }
 159 }
 160 
 161 static void
 162 scheduler_event_callback(pcmk_ipc_api_t *api, enum pcmk_ipc_event event_type,
     /*  */
 163                          crm_exit_t status, void *event_data, void *user_data)
 164 {
 165     pcmk_schedulerd_api_reply_t *reply = event_data;
 166 
 167     switch (event_type) {
 168         case pcmk_ipc_event_disconnect:
 169             handle_disconnect();
 170             break;
 171 
 172         case pcmk_ipc_event_reply:
 173             handle_reply(reply);
 174             break;
 175 
 176         default:
 177             break;
 178     }
 179 }
 180 
 181 static bool
 182 new_schedulerd_ipc_connection(void)
     /*  */
 183 {
 184     int rc;
 185 
 186     controld_set_fsa_input_flags(R_PE_REQUIRED);
 187 
 188     if (schedulerd_api == NULL) {
 189         rc = pcmk_new_ipc_api(&schedulerd_api, pcmk_ipc_schedulerd);
 190 
 191         if (rc != pcmk_rc_ok) {
 192             crm_err("Error connecting to the scheduler: %s", pcmk_rc_str(rc));
 193             return false;
 194         }
 195     }
 196 
 197     pcmk_register_ipc_callback(schedulerd_api, scheduler_event_callback, NULL);
 198 
 199     rc = pcmk__connect_ipc(schedulerd_api, pcmk_ipc_dispatch_main, 3);
 200     if (rc != pcmk_rc_ok) {
 201         crm_err("Error connecting to %s: %s",
 202                 pcmk_ipc_name(schedulerd_api, true), pcmk_rc_str(rc));
 203         return false;
 204     }
 205 
 206     controld_set_fsa_input_flags(R_PE_CONNECTED);
 207     return true;
 208 }
 209 
 210 static void do_pe_invoke_callback(xmlNode *msg, int call_id, int rc,
 211                                   xmlNode *output, void *user_data);
 212 
 213 /*       A_PE_START, A_PE_STOP, O_PE_RESTART    */
 214 void
 215 do_pe_control(long long action,
     /*  */
 216               enum crmd_fsa_cause cause,
 217               enum crmd_fsa_state cur_state,
 218               enum crmd_fsa_input current_input, fsa_data_t * msg_data)
 219 {
 220     if (pcmk_is_set(action, A_PE_STOP)) {
 221         controld_clear_fsa_input_flags(R_PE_REQUIRED);
 222         pcmk_disconnect_ipc(schedulerd_api);
 223         handle_disconnect();
 224     }
 225     if (pcmk_is_set(action, A_PE_START)
 226         && !pcmk_is_set(controld_globals.fsa_input_register, R_PE_CONNECTED)) {
 227 
 228         if (cur_state == S_STOPPING) {
 229             crm_info("Ignoring request to connect to scheduler while shutting down");
 230 
 231         } else if (!new_schedulerd_ipc_connection()) {
 232             crm_warn("Could not connect to scheduler");
 233             register_fsa_error(C_FSA_INTERNAL, I_FAIL, NULL);
 234         }
 235     }
 236 }
 237 
 238 static int fsa_pe_query = 0;
 239 static mainloop_timer_t *controld_sched_timer = NULL;
 240 
 241 // @TODO Make this a configurable cluster option if there's demand for it
 242 #define SCHED_TIMEOUT_MS (120000)
 243 
 244 /*!
 245  * \internal
 246  * \brief Handle a timeout waiting for scheduler reply
 247  *
 248  * \param[in] user_data  Ignored
 249  *
 250  * \return FALSE (indicating that timer should not be restarted)
 251  */
 252 static gboolean
 253 controld_sched_timeout(gpointer user_data)
     /*  */
 254 {
 255     if (AM_I_DC) {
 256         /* If this node is the DC but can't communicate with the scheduler, just
 257          * exit (and likely get fenced) so this node doesn't interfere with any
 258          * further DC elections.
 259          *
 260          * @TODO We could try something less drastic first, like disconnecting
 261          * and reconnecting to the scheduler, but something is likely going
 262          * seriously wrong, so perhaps it's better to just fail as quickly as
 263          * possible.
 264          */
 265         crmd_exit(CRM_EX_FATAL);
 266     }
 267     return FALSE;
 268 }
 269 
 270 void
 271 controld_stop_sched_timer(void)
     /*  */
 272 {
 273     if ((controld_sched_timer != NULL)
 274         && (controld_globals.fsa_pe_ref != NULL)) {
 275         crm_trace("Stopping timer for scheduler reply %s",
 276                   controld_globals.fsa_pe_ref);
 277     }
 278     mainloop_timer_stop(controld_sched_timer);
 279 }
 280 
 281 /*!
 282  * \internal
 283  * \brief Set the scheduler request currently being waited on
 284  *
 285  * \param[in] ref  Request to expect reply to (or NULL for none)
 286  *
 287  * \note This function takes ownership of \p ref.
 288  */
 289 void
 290 controld_expect_sched_reply(char *ref)
     /*  */
 291 {
 292     if (ref) {
 293         if (controld_sched_timer == NULL) {
 294             controld_sched_timer = mainloop_timer_add("scheduler_reply_timer",
 295                                                       SCHED_TIMEOUT_MS, FALSE,
 296                                                       controld_sched_timeout,
 297                                                       NULL);
 298         }
 299         mainloop_timer_start(controld_sched_timer);
 300     } else {
 301         controld_stop_sched_timer();
 302     }
 303     free(controld_globals.fsa_pe_ref);
 304     controld_globals.fsa_pe_ref = ref;
 305 }
 306 
 307 /*!
 308  * \internal
 309  * \brief Free the scheduler reply timer
 310  */
 311 void
 312 controld_free_sched_timer(void)
     /*  */
 313 {
 314     if (controld_sched_timer != NULL) {
 315         mainloop_timer_del(controld_sched_timer);
 316         controld_sched_timer = NULL;
 317     }
 318 }
 319 
 320 /*       A_PE_INVOKE    */
 321 void
 322 do_pe_invoke(long long action,
     /*  */
 323              enum crmd_fsa_cause cause,
 324              enum crmd_fsa_state cur_state,
 325              enum crmd_fsa_input current_input, fsa_data_t * msg_data)
 326 {
 327     cib_t *cib_conn = controld_globals.cib_conn;
 328 
 329     if (AM_I_DC == FALSE) {
 330         crm_err("Not invoking scheduler because not DC: %s",
 331                 fsa_action2string(action));
 332         return;
 333     }
 334 
 335     if (!pcmk_is_set(controld_globals.fsa_input_register, R_PE_CONNECTED)) {
 336         if (pcmk_is_set(controld_globals.fsa_input_register, R_SHUTDOWN)) {
 337             crm_err("Cannot shut down gracefully without the scheduler");
 338             register_fsa_input_before(C_FSA_INTERNAL, I_TERMINATE, NULL);
 339 
 340         } else {
 341             crm_info("Waiting for the scheduler to connect");
 342             crmd_fsa_stall(FALSE);
 343             controld_set_fsa_action_flags(A_PE_START);
 344             controld_trigger_fsa();
 345         }
 346         return;
 347     }
 348 
 349     if (cur_state != S_POLICY_ENGINE) {
 350         crm_notice("Not invoking scheduler because in state %s",
 351                    fsa_state2string(cur_state));
 352         return;
 353     }
 354     if (!pcmk_is_set(controld_globals.fsa_input_register, R_HAVE_CIB)) {
 355         crm_err("Attempted to invoke scheduler without consistent Cluster Information Base!");
 356 
 357         /* start the join from scratch */
 358         register_fsa_input_before(C_FSA_INTERNAL, I_ELECTION, NULL);
 359         return;
 360     }
 361 
 362     fsa_pe_query = cib_conn->cmds->query(cib_conn, NULL, NULL, cib_none);
 363 
 364     crm_debug("Query %d: Requesting the current CIB: %s", fsa_pe_query,
 365               fsa_state2string(controld_globals.fsa_state));
 366 
 367     controld_expect_sched_reply(NULL);
 368     fsa_register_cib_callback(fsa_pe_query, NULL, do_pe_invoke_callback);
 369 }
 370 
 371 static void
 372 force_local_option(xmlNode *xml, const char *attr_name, const char *attr_value)
     /*  */
 373 {
 374     int max = 0;
 375     int lpc = 0;
 376     const char *xpath_base = NULL;
 377     char *xpath_string = NULL;
 378     xmlXPathObjectPtr xpathObj = NULL;
 379 
 380     xpath_base = pcmk_cib_xpath_for(PCMK_XE_CRM_CONFIG);
 381     if (xpath_base == NULL) {
 382         crm_err(PCMK_XE_CRM_CONFIG " CIB element not known (bug?)");
 383         return;
 384     }
 385 
 386     xpath_string = crm_strdup_printf("%s//%s//nvpair[@name='%s']",
 387                                      xpath_base, PCMK_XE_CLUSTER_PROPERTY_SET,
 388                                      attr_name);
 389     xpathObj = xpath_search(xml, xpath_string);
 390     max = numXpathResults(xpathObj);
 391     free(xpath_string);
 392 
 393     for (lpc = 0; lpc < max; lpc++) {
 394         xmlNode *match = getXpathResult(xpathObj, lpc);
 395         crm_trace("Forcing %s/%s = %s",
 396                   pcmk__xe_id(match), attr_name, attr_value);
 397         crm_xml_add(match, PCMK_XA_VALUE, attr_value);
 398     }
 399 
 400     if(max == 0) {
 401         xmlNode *configuration = NULL;
 402         xmlNode *crm_config = NULL;
 403         xmlNode *cluster_property_set = NULL;
 404 
 405         crm_trace("Creating %s-%s for %s=%s",
 406                   PCMK_VALUE_CIB_BOOTSTRAP_OPTIONS, attr_name, attr_name,
 407                   attr_value);
 408 
 409         configuration = pcmk__xe_first_child(xml, PCMK_XE_CONFIGURATION, NULL,
 410                                              NULL);
 411         if (configuration == NULL) {
 412             configuration = pcmk__xe_create(xml, PCMK_XE_CONFIGURATION);
 413         }
 414 
 415         crm_config = pcmk__xe_first_child(configuration, PCMK_XE_CRM_CONFIG,
 416                                           NULL, NULL);
 417         if (crm_config == NULL) {
 418             crm_config = pcmk__xe_create(configuration, PCMK_XE_CRM_CONFIG);
 419         }
 420 
 421         cluster_property_set =
 422             pcmk__xe_first_child(crm_config, PCMK_XE_CLUSTER_PROPERTY_SET, NULL,
 423                                  NULL);
 424         if (cluster_property_set == NULL) {
 425             cluster_property_set =
 426                 pcmk__xe_create(crm_config, PCMK_XE_CLUSTER_PROPERTY_SET);
 427             crm_xml_add(cluster_property_set, PCMK_XA_ID,
 428                         PCMK_VALUE_CIB_BOOTSTRAP_OPTIONS);
 429         }
 430 
 431         xml = pcmk__xe_create(cluster_property_set, PCMK_XE_NVPAIR);
 432 
 433         pcmk__xe_set_id(xml, "%s-%s",
 434                         PCMK_VALUE_CIB_BOOTSTRAP_OPTIONS, attr_name);
 435         crm_xml_add(xml, PCMK_XA_NAME, attr_name);
 436         crm_xml_add(xml, PCMK_XA_VALUE, attr_value);
 437     }
 438     freeXpathObject(xpathObj);
 439 }
 440 
 441 static void
 442 do_pe_invoke_callback(xmlNode * msg, int call_id, int rc, xmlNode * output, void *user_data)
     /*  */
 443 {
 444     char *ref = NULL;
 445     pid_t watchdog = pcmk__locate_sbd();
 446 
 447     if (rc != pcmk_ok) {
 448         crm_err("Could not retrieve the Cluster Information Base: %s "
 449                 QB_XS " rc=%d call=%d", pcmk_strerror(rc), rc, call_id);
 450         register_fsa_error_adv(C_FSA_INTERNAL, I_ERROR, NULL, NULL, __func__);
 451         return;
 452 
 453     } else if (call_id != fsa_pe_query) {
 454         crm_trace("Skipping superseded CIB query: %d (current=%d)", call_id, fsa_pe_query);
 455         return;
 456 
 457     } else if (!AM_I_DC
 458                || !pcmk_is_set(controld_globals.fsa_input_register,
 459                                R_PE_CONNECTED)) {
 460         crm_debug("No need to invoke the scheduler anymore");
 461         return;
 462 
 463     } else if (controld_globals.fsa_state != S_POLICY_ENGINE) {
 464         crm_debug("Discarding scheduler request in state: %s",
 465                   fsa_state2string(controld_globals.fsa_state));
 466         return;
 467 
 468     /* this callback counts as 1 */
 469     } else if (num_cib_op_callbacks() > 1) {
 470         crm_debug("Re-asking for the CIB: %d other peer updates still pending",
 471                   (num_cib_op_callbacks() - 1));
 472         sleep(1);
 473         controld_set_fsa_action_flags(A_PE_INVOKE);
 474         controld_trigger_fsa();
 475         return;
 476     }
 477 
 478     CRM_LOG_ASSERT(output != NULL);
 479 
 480     /* Refresh the remote node cache and the known node cache when the
 481      * scheduler is invoked */
 482     pcmk__refresh_node_caches_from_cib(output);
 483 
 484     crm_xml_add(output, PCMK_XA_DC_UUID, controld_globals.our_uuid);
 485     pcmk__xe_set_bool_attr(output, PCMK_XA_HAVE_QUORUM,
 486                            pcmk_is_set(controld_globals.flags,
 487                                        controld_has_quorum));
 488 
 489     force_local_option(output, PCMK_OPT_HAVE_WATCHDOG, pcmk__btoa(watchdog));
 490 
 491     if (pcmk_is_set(controld_globals.flags, controld_ever_had_quorum)
 492         && !pcmk__cluster_has_quorum()) {
 493 
 494         crm_xml_add_int(output, PCMK_XA_NO_QUORUM_PANIC, 1);
 495     }
 496 
 497     rc = pcmk_schedulerd_api_graph(schedulerd_api, output, &ref);
 498     if (rc != pcmk_rc_ok) {
 499         free(ref);
 500         crm_err("Could not contact the scheduler: %s " QB_XS " rc=%d",
 501                 pcmk_rc_str(rc), rc);
 502         register_fsa_error_adv(C_FSA_INTERNAL, I_ERROR, NULL, NULL, __func__);
 503     } else {
 504         pcmk__assert(ref != NULL);
 505         controld_expect_sched_reply(ref);
 506         crm_debug("Invoking the scheduler: query=%d, ref=%s, seq=%llu, "
 507                   "quorate=%s",
 508                   fsa_pe_query, controld_globals.fsa_pe_ref,
 509                   controld_globals.peer_seq,
 510                   pcmk__flag_text(controld_globals.flags, controld_has_quorum));
 511     }
 512 }
/* */
root/daemons/controld/controld_schedulerd.c

DEFINITIONS
