daemons/pacemakerd/pcmkd

/* */
This source file includes following definitions.
pcmkd_cluster_connected
check_next_subdaemon
escalate_shutdown
pcmk_child_exit
pcmk_process_exit
pcmk_shutdown_worker
start_child
child_liveness
find_and_track_existing_processes
init_children_processes
pcmk_shutdown
restart_cluster_subdaemons
stop_child
   1 /*
   2  * Copyright 2010-2022 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 #include "pacemakerd.h"
  12 
  13 #include <errno.h>
  14 #include <grp.h>
  15 #include <signal.h>
  16 #include <stdbool.h>
  17 #include <stdlib.h>
  18 #include <string.h>
  19 #include <sys/types.h>
  20 #include <time.h>
  21 #include <unistd.h>
  22 
  23 #include <crm/cluster.h>
  24 #include <crm/msg_xml.h>
  25 
  26 typedef struct pcmk_child_s {
  27     pid_t pid;
  28     int respawn_count;
  29     bool respawn;
  30     const char *name;
  31     const char *uid;
  32     const char *command;
  33     const char *endpoint;  /* IPC server name */
  34     bool needs_cluster;
  35     int check_count;
  36 
  37     /* Anything below here will be dynamically initialized */
  38     bool needs_retry;
  39     bool active_before_startup;
  40 } pcmk_child_t;
  41 
  42 #define PCMK_PROCESS_CHECK_INTERVAL 1
  43 #define PCMK_PROCESS_CHECK_RETRIES  5
  44 #define SHUTDOWN_ESCALATION_PERIOD  180000  /* 3m */
  45 
  46 /* Index into the array below */
  47 #define PCMK_CHILD_CONTROLD  5
  48 
  49 static pcmk_child_t pcmk_children[] = {
  50     {
  51         0, 0, true,  "pacemaker-based", CRM_DAEMON_USER,
  52         CRM_DAEMON_DIR "/pacemaker-based", PCMK__SERVER_BASED_RO,
  53         true
  54     },
  55     {
  56         0, 0, true, "pacemaker-fenced", NULL,
  57         CRM_DAEMON_DIR "/pacemaker-fenced", "stonith-ng",
  58         true
  59     },
  60     {
  61         0, 0, true,  "pacemaker-execd", NULL,
  62         CRM_DAEMON_DIR "/pacemaker-execd", CRM_SYSTEM_LRMD,
  63         false
  64     },
  65     {
  66         0, 0, true, "pacemaker-attrd", CRM_DAEMON_USER,
  67         CRM_DAEMON_DIR "/pacemaker-attrd", T_ATTRD,
  68         true
  69     },
  70     {
  71         0, 0, true, "pacemaker-schedulerd", CRM_DAEMON_USER,
  72         CRM_DAEMON_DIR "/pacemaker-schedulerd", CRM_SYSTEM_PENGINE,
  73         false
  74     },
  75     {
  76         0, 0, true, "pacemaker-controld", CRM_DAEMON_USER,
  77         CRM_DAEMON_DIR "/pacemaker-controld", CRM_SYSTEM_CRMD,
  78         true
  79     },
  80 };
  81 
  82 static char *opts_default[] = { NULL, NULL };
  83 static char *opts_vgrind[] = { NULL, NULL, NULL, NULL, NULL };
  84 
  85 crm_trigger_t *shutdown_trigger = NULL;
  86 crm_trigger_t *startup_trigger = NULL;
  87 time_t subdaemon_check_progress = 0;
  88 
  89 /* When contacted via pacemakerd-api by a client having sbd in
  90  * the name we assume it is sbd-daemon which wants to know
  91  * if pacemakerd shutdown gracefully.
  92  * Thus when everything is shutdown properly pacemakerd
  93  * waits till it has reported the graceful completion of
  94  * shutdown to sbd and just when sbd-client closes the
  95  * connection we can assume that the report has arrived
  96  * properly so that pacemakerd can finally exit.
  97  * Following two variables are used to track that handshake.
  98  */
  99 unsigned int shutdown_complete_state_reported_to = 0;
 100 gboolean shutdown_complete_state_reported_client_closed = FALSE;
 101 
 102 /* state we report when asked via pacemakerd-api status-ping */
 103 const char *pacemakerd_state = XML_PING_ATTR_PACEMAKERDSTATE_INIT;
 104 gboolean running_with_sbd = FALSE; /* local copy */
 105 
 106 GMainLoop *mainloop = NULL;
 107 
 108 static gboolean fatal_error = FALSE;
 109 
 110 static int child_liveness(pcmk_child_t *child);
 111 static gboolean escalate_shutdown(gpointer data);
 112 static int start_child(pcmk_child_t * child);
 113 static void pcmk_child_exit(mainloop_child_t * p, pid_t pid, int core, int signo, int exitcode);
 114 static void pcmk_process_exit(pcmk_child_t * child);
 115 static gboolean pcmk_shutdown_worker(gpointer user_data);
 116 static gboolean stop_child(pcmk_child_t * child, int signal);
 117 
 118 static bool
 119 pcmkd_cluster_connected(void)
     /*  */
 120 {
 121 #if SUPPORT_COROSYNC
 122     return pcmkd_corosync_connected();
 123 #else
 124     return true;
 125 #endif
 126 }
 127 
 128 static gboolean
 129 check_next_subdaemon(gpointer user_data)
     /*  */
 130 {
 131     static int next_child = 0;
 132     int rc = child_liveness(&pcmk_children[next_child]);
 133 
 134     crm_trace("Checked %s[%lld]: %s (%d)",
 135               pcmk_children[next_child].name,
 136               (long long) PCMK__SPECIAL_PID_AS_0(pcmk_children[next_child].pid),
 137               pcmk_rc_str(rc), rc);
 138 
 139     switch (rc) {
 140         case pcmk_rc_ok:
 141             pcmk_children[next_child].check_count = 0;
 142             subdaemon_check_progress = time(NULL);
 143             break;
 144         case pcmk_rc_ipc_pid_only: // This case: it was previously OK
 145             pcmk_children[next_child].check_count++;
 146             if (pcmk_children[next_child].check_count >= PCMK_PROCESS_CHECK_RETRIES) {
 147                 crm_err("%s[%lld] is unresponsive to ipc after %d tries but "
 148                         "we found the pid so have it killed that we can restart",
 149                         pcmk_children[next_child].name,
 150                         (long long) PCMK__SPECIAL_PID_AS_0(
 151                             pcmk_children[next_child].pid),
 152                         pcmk_children[next_child].check_count);
 153                 stop_child(&pcmk_children[next_child], SIGKILL);
 154                 if (pcmk_children[next_child].respawn) {
 155                     /* as long as the respawn-limit isn't reached
 156                        give it another round of check retries
 157                      */
 158                     pcmk_children[next_child].check_count = 0;
 159                 }
 160             } else {
 161                 crm_notice("%s[%lld] is unresponsive to ipc after %d tries",
 162                         pcmk_children[next_child].name,
 163                         (long long) PCMK__SPECIAL_PID_AS_0(
 164                             pcmk_children[next_child].pid),
 165                         pcmk_children[next_child].check_count);
 166                 if (pcmk_children[next_child].respawn) {
 167                     /* as long as the respawn-limit isn't reached
 168                        and we haven't run out of connect retries
 169                        we account this as progress we are willing
 170                        to tell to sbd
 171                      */
 172                     subdaemon_check_progress = time(NULL);
 173                 }
 174             }
 175             /* go to the next child and see if
 176                we can make progress there
 177              */
 178             break;
 179         case pcmk_rc_ipc_unresponsive:
 180             if (!pcmk_children[next_child].respawn) {
 181                 /* if a subdaemon is down and we don't want it
 182                    to be restarted this is a success during
 183                    shutdown. if it isn't restarted anymore
 184                    due to MAX_RESPAWN it is
 185                    rather no success.
 186                  */
 187                 if (pcmk_children[next_child].respawn_count <= MAX_RESPAWN) {
 188                     subdaemon_check_progress = time(NULL);
 189                 }
 190             }
 191             if (!pcmk_children[next_child].active_before_startup) {
 192                 crm_trace("found %s[%lld] missing - signal-handler "
 193                           "will take care of it",
 194                            pcmk_children[next_child].name,
 195                            (long long) PCMK__SPECIAL_PID_AS_0(
 196                             pcmk_children[next_child].pid));
 197                 break;
 198             }
 199             if (pcmk_children[next_child].respawn) {
 200                 crm_err("%s[%lld] terminated",
 201                         pcmk_children[next_child].name,
 202                         (long long) PCMK__SPECIAL_PID_AS_0(
 203                             pcmk_children[next_child].pid));
 204             } else {
 205                 /* orderly shutdown */
 206                 crm_notice("%s[%lld] terminated",
 207                            pcmk_children[next_child].name,
 208                            (long long) PCMK__SPECIAL_PID_AS_0(
 209                                 pcmk_children[next_child].pid));
 210             }
 211             pcmk_process_exit(&(pcmk_children[next_child]));
 212             break;
 213         default:
 214             crm_exit(CRM_EX_FATAL);
 215             break;  /* static analysis/noreturn */
 216     }
 217 
 218     next_child++;
 219     if (next_child >= PCMK__NELEM(pcmk_children)) {
 220         next_child = 0;
 221     }
 222 
 223     return G_SOURCE_CONTINUE;
 224 }
 225 
 226 static gboolean
 227 escalate_shutdown(gpointer data)
     /*  */
 228 {
 229     pcmk_child_t *child = data;
 230 
 231     if (child->pid == PCMK__SPECIAL_PID) {
 232         pcmk_process_exit(child);
 233 
 234     } else if (child->pid != 0) {
 235         /* Use SIGSEGV instead of SIGKILL to create a core so we can see what it was up to */
 236         crm_err("Child %s not terminating in a timely manner, forcing", child->name);
 237         stop_child(child, SIGSEGV);
 238     }
 239     return FALSE;
 240 }
 241 
 242 static void
 243 pcmk_child_exit(mainloop_child_t * p, pid_t pid, int core, int signo, int exitcode)
     /*  */
 244 {
 245     pcmk_child_t *child = mainloop_child_userdata(p);
 246     const char *name = mainloop_child_name(p);
 247 
 248     if (signo) {
 249         do_crm_log(((signo == SIGKILL)? LOG_WARNING : LOG_ERR),
 250                    "%s[%d] terminated with signal %d (%s)%s",
 251                    name, pid, signo, strsignal(signo),
 252                    (core? " and dumped core" : ""));
 253 
 254     } else {
 255         switch(exitcode) {
 256             case CRM_EX_OK:
 257                 crm_info("%s[%d] exited with status %d (%s)",
 258                          name, pid, exitcode, crm_exit_str(exitcode));
 259                 break;
 260 
 261             case CRM_EX_FATAL:
 262                 crm_warn("Shutting cluster down because %s[%d] had fatal failure",
 263                          name, pid);
 264                 child->respawn = false;
 265                 fatal_error = TRUE;
 266                 pcmk_shutdown(SIGTERM);
 267                 break;
 268 
 269             case CRM_EX_PANIC:
 270                 crm_emerg("%s[%d] instructed the machine to reset", name, pid);
 271                 child->respawn = false;
 272                 fatal_error = TRUE;
 273                 pcmk__panic(__func__);
 274                 pcmk_shutdown(SIGTERM);
 275                 break;
 276 
 277             default:
 278                 crm_err("%s[%d] exited with status %d (%s)",
 279                         name, pid, exitcode, crm_exit_str(exitcode));
 280                 break;
 281         }
 282     }
 283 
 284     pcmk_process_exit(child);
 285 }
 286 
 287 static void
 288 pcmk_process_exit(pcmk_child_t * child)
     /*  */
 289 {
 290     child->pid = 0;
 291     child->active_before_startup = false;
 292     child->check_count = 0;
 293 
 294     child->respawn_count += 1;
 295     if (child->respawn_count > MAX_RESPAWN) {
 296         crm_err("Child respawn count exceeded by %s", child->name);
 297         child->respawn = false;
 298     }
 299 
 300     if (shutdown_trigger) {
 301         /* resume step-wise shutdown (returned TRUE yields no parallelizing) */
 302         mainloop_set_trigger(shutdown_trigger);
 303 
 304     } else if (!child->respawn) {
 305         /* nothing to do */
 306 
 307     } else if (crm_is_true(getenv("PCMK_fail_fast"))) {
 308         crm_err("Rebooting system because of %s", child->name);
 309         pcmk__panic(__func__);
 310 
 311     } else if (child_liveness(child) == pcmk_rc_ok) {
 312         crm_warn("One-off suppressing strict respawning of a child process %s,"
 313                  " appears alright per %s IPC end-point",
 314                  child->name, child->endpoint);
 315 
 316     } else if (child->needs_cluster && !pcmkd_cluster_connected()) {
 317         crm_notice("Not respawning %s subdaemon until cluster returns",
 318                    child->name);
 319         child->needs_retry = true;
 320 
 321     } else {
 322         crm_notice("Respawning %s subdaemon after unexpected exit",
 323                    child->name);
 324         start_child(child);
 325     }
 326 }
 327 
 328 static gboolean
 329 pcmk_shutdown_worker(gpointer user_data)
     /*  */
 330 {
 331     static int phase = PCMK__NELEM(pcmk_children) - 1;
 332     static time_t next_log = 0;
 333 
 334     if (phase == PCMK__NELEM(pcmk_children) - 1) {
 335         crm_notice("Shutting down Pacemaker");
 336         pacemakerd_state = XML_PING_ATTR_PACEMAKERDSTATE_SHUTTINGDOWN;
 337     }
 338 
 339     for (; phase >= 0; phase--) {
 340         pcmk_child_t *child = &(pcmk_children[phase]);
 341 
 342         if (child->pid != 0) {
 343             time_t now = time(NULL);
 344 
 345             if (child->respawn) {
 346                 if (child->pid == PCMK__SPECIAL_PID) {
 347                     crm_warn("The process behind %s IPC cannot be"
 348                              " terminated, so either wait the graceful"
 349                              " period of %ld s for its native termination"
 350                              " if it vitally depends on some other daemons"
 351                              " going down in a controlled way already,"
 352                              " or locate and kill the correct %s process"
 353                              " on your own; set PCMK_fail_fast=1 to avoid"
 354                              " this altogether next time around",
 355                              child->name, (long) SHUTDOWN_ESCALATION_PERIOD,
 356                              child->command);
 357                 }
 358                 next_log = now + 30;
 359                 child->respawn = false;
 360                 stop_child(child, SIGTERM);
 361                 if (phase < PCMK_CHILD_CONTROLD) {
 362                     g_timeout_add(SHUTDOWN_ESCALATION_PERIOD,
 363                                   escalate_shutdown, child);
 364                 }
 365 
 366             } else if (now >= next_log) {
 367                 next_log = now + 30;
 368                 crm_notice("Still waiting for %s to terminate "
 369                            CRM_XS " pid=%lld",
 370                            child->name, (long long) child->pid);
 371             }
 372             return TRUE;
 373         }
 374 
 375         /* cleanup */
 376         crm_debug("%s confirmed stopped", child->name);
 377         child->pid = 0;
 378     }
 379 
 380     crm_notice("Shutdown complete");
 381     pacemakerd_state = XML_PING_ATTR_PACEMAKERDSTATE_SHUTDOWNCOMPLETE;
 382     if (!fatal_error && running_with_sbd &&
 383         pcmk__get_sbd_sync_resource_startup() &&
 384         !shutdown_complete_state_reported_client_closed) {
 385         crm_notice("Waiting for SBD to pick up shutdown-complete-state.");
 386         return TRUE;
 387     }
 388 
 389     {
 390         const char *delay = pcmk__env_option(PCMK__ENV_SHUTDOWN_DELAY);
 391         if(delay) {
 392             sync();
 393             pcmk__sleep_ms(crm_get_msec(delay));
 394         }
 395     }
 396 
 397     g_main_loop_quit(mainloop);
 398 
 399     if (fatal_error) {
 400         crm_notice("Shutting down and staying down after fatal error");
 401 #ifdef SUPPORT_COROSYNC
 402         pcmkd_shutdown_corosync();
 403 #endif
 404         crm_exit(CRM_EX_FATAL);
 405     }
 406 
 407     return TRUE;
 408 }
 409 
 410 /* TODO once libqb is taught to juggle with IPC end-points carried over as
 411         bare file descriptor (https://github.com/ClusterLabs/libqb/issues/325)
 412         it shall hand over these descriptors here if/once they are successfully
 413         pre-opened in (presumably) child_liveness(), to avoid any remaining
 414         room for races */
 415  // \return Standard Pacemaker return code
 416 static int
 417 start_child(pcmk_child_t * child)
     /*  */
 418 {
 419     uid_t uid = 0;
 420     gid_t gid = 0;
 421     gboolean use_valgrind = FALSE;
 422     gboolean use_callgrind = FALSE;
 423     const char *env_valgrind = getenv("PCMK_valgrind_enabled");
 424     const char *env_callgrind = getenv("PCMK_callgrind_enabled");
 425 
 426     child->active_before_startup = false;
 427     child->check_count = 0;
 428 
 429     if (child->command == NULL) {
 430         crm_info("Nothing to do for child \"%s\"", child->name);
 431         return pcmk_rc_ok;
 432     }
 433 
 434     if (env_callgrind != NULL && crm_is_true(env_callgrind)) {
 435         use_callgrind = TRUE;
 436         use_valgrind = TRUE;
 437 
 438     } else if (env_callgrind != NULL && strstr(env_callgrind, child->name)) {
 439         use_callgrind = TRUE;
 440         use_valgrind = TRUE;
 441 
 442     } else if (env_valgrind != NULL && crm_is_true(env_valgrind)) {
 443         use_valgrind = TRUE;
 444 
 445     } else if (env_valgrind != NULL && strstr(env_valgrind, child->name)) {
 446         use_valgrind = TRUE;
 447     }
 448 
 449     if (use_valgrind && strlen(VALGRIND_BIN) == 0) {
 450         crm_warn("Cannot enable valgrind for %s:"
 451                  " The location of the valgrind binary is unknown", child->name);
 452         use_valgrind = FALSE;
 453     }
 454 
 455     if (child->uid) {
 456         if (crm_user_lookup(child->uid, &uid, &gid) < 0) {
 457             crm_err("Invalid user (%s) for %s: not found", child->uid, child->name);
 458             return EACCES;
 459         }
 460         crm_info("Using uid=%u and group=%u for process %s", uid, gid, child->name);
 461     }
 462 
 463     child->pid = fork();
 464     CRM_ASSERT(child->pid != -1);
 465 
 466     if (child->pid > 0) {
 467         /* parent */
 468         mainloop_child_add(child->pid, 0, child->name, child, pcmk_child_exit);
 469 
 470         crm_info("Forked child %lld for process %s%s",
 471                  (long long) child->pid, child->name,
 472                  use_valgrind ? " (valgrind enabled: " VALGRIND_BIN ")" : "");
 473         return pcmk_rc_ok;
 474 
 475     } else {
 476         /* Start a new session */
 477         (void)setsid();
 478 
 479         /* Setup the two alternate arg arrays */
 480         opts_vgrind[0] = strdup(VALGRIND_BIN);
 481         if (use_callgrind) {
 482             opts_vgrind[1] = strdup("--tool=callgrind");
 483             opts_vgrind[2] = strdup("--callgrind-out-file=" CRM_STATE_DIR "/callgrind.out.%p");
 484             opts_vgrind[3] = strdup(child->command);
 485             opts_vgrind[4] = NULL;
 486         } else {
 487             opts_vgrind[1] = strdup(child->command);
 488             opts_vgrind[2] = NULL;
 489             opts_vgrind[3] = NULL;
 490             opts_vgrind[4] = NULL;
 491         }
 492         opts_default[0] = strdup(child->command);
 493 
 494         if(gid) {
 495             // Whether we need root group access to talk to cluster layer
 496             bool need_root_group = TRUE;
 497 
 498             if (is_corosync_cluster()) {
 499                 /* Corosync clusters can drop root group access, because we set
 500                  * uidgid.gid.${gid}=1 via CMAP, which allows these processes to
 501                  * connect to corosync.
 502                  */
 503                 need_root_group = FALSE;
 504             }
 505 
 506             // Drop root group access if not needed
 507             if (!need_root_group && (setgid(gid) < 0)) {
 508                 crm_warn("Could not set group to %d: %s", gid, strerror(errno));
 509             }
 510 
 511             /* Initialize supplementary groups to only those always granted to
 512              * the user, plus haclient (so we can access IPC).
 513              */
 514             if (initgroups(child->uid, gid) < 0) {
 515                 crm_err("Cannot initialize groups for %s: %s (%d)",
 516                         child->uid, pcmk_rc_str(errno), errno);
 517             }
 518         }
 519 
 520         if (uid && setuid(uid) < 0) {
 521             crm_warn("Could not set user to %s (id %d): %s",
 522                      child->uid, uid, strerror(errno));
 523         }
 524 
 525         pcmk__close_fds_in_child(true);
 526 
 527         pcmk__open_devnull(O_RDONLY);   // stdin (fd 0)
 528         pcmk__open_devnull(O_WRONLY);   // stdout (fd 1)
 529         pcmk__open_devnull(O_WRONLY);   // stderr (fd 2)
 530 
 531         if (use_valgrind) {
 532             (void)execvp(VALGRIND_BIN, opts_vgrind);
 533         } else {
 534             (void)execvp(child->command, opts_default);
 535         }
 536         crm_crit("Could not execute %s: %s", child->command, strerror(errno));
 537         crm_exit(CRM_EX_FATAL);
 538     }
 539     return pcmk_rc_ok;          /* never reached */
 540 }
 541 
 542 /*!
 543  * \internal
 544  * \brief Check the liveness of the child based on IPC name and PID if tracked
 545  *
 546  * \param[inout] child  Child tracked data
 547  *
 548  * \return Standard Pacemaker return code
 549  *
 550  * \note Return codes of particular interest include pcmk_rc_ipc_unresponsive
 551  *       indicating that no trace of IPC liveness was detected,
 552  *       pcmk_rc_ipc_unauthorized indicating that the IPC endpoint is blocked by
 553  *       an unauthorized process, and pcmk_rc_ipc_pid_only indicating that
 554  *       the child is up by PID but not IPC end-point (possibly starting).
 555  * \note This function doesn't modify any of \p child members but \c pid,
 556  *       and is not actively toying with processes as such but invoking
 557  *       \c stop_child in one particular case (there's for some reason
 558  *       a different authentic holder of the IPC end-point).
 559  */
 560 static int
 561 child_liveness(pcmk_child_t *child)
     /*  */
 562 {
 563     uid_t cl_uid = 0;
 564     gid_t cl_gid = 0;
 565     const uid_t root_uid = 0;
 566     const gid_t root_gid = 0;
 567     const uid_t *ref_uid;
 568     const gid_t *ref_gid;
 569     int rc = pcmk_rc_ipc_unresponsive;
 570     pid_t ipc_pid = 0;
 571 
 572     if (child->endpoint == NULL
 573             && (child->pid <= 0 || child->pid == PCMK__SPECIAL_PID)) {
 574         crm_err("Cannot track child %s for missing both API end-point and PID",
 575                 child->name);
 576         rc = EINVAL; // Misuse of function when child is not trackable
 577 
 578     } else if (child->endpoint != NULL) {
 579         int legacy_rc = pcmk_ok;
 580 
 581         if (child->uid == NULL) {
 582             ref_uid = &root_uid;
 583             ref_gid = &root_gid;
 584         } else {
 585             ref_uid = &cl_uid;
 586             ref_gid = &cl_gid;
 587             legacy_rc = pcmk_daemon_user(&cl_uid, &cl_gid);
 588         }
 589 
 590         if (legacy_rc < 0) {
 591             rc = pcmk_legacy2rc(legacy_rc);
 592             crm_err("Could not find user and group IDs for user %s: %s "
 593                     CRM_XS " rc=%d", CRM_DAEMON_USER, pcmk_rc_str(rc), rc);
 594         } else {
 595             rc = pcmk__ipc_is_authentic_process_active(child->endpoint,
 596                                                        *ref_uid, *ref_gid,
 597                                                        &ipc_pid);
 598             if ((rc == pcmk_rc_ok) || (rc == pcmk_rc_ipc_unresponsive)) {
 599                 if (child->pid <= 0) {
 600                     /* If rc is pcmk_rc_ok, ipc_pid is nonzero and this
 601                      * initializes a new child. If rc is
 602                      * pcmk_rc_ipc_unresponsive, ipc_pid is zero, and we will
 603                      * investigate further.
 604                      */
 605                     child->pid = ipc_pid;
 606                 } else if ((ipc_pid != 0) && (child->pid != ipc_pid)) {
 607                     /* An unexpected (but authorized) process is responding to
 608                      * IPC. Investigate further.
 609                      */
 610                     rc = pcmk_rc_ipc_unresponsive;
 611                 }
 612             }
 613         }
 614     }
 615 
 616     if (rc == pcmk_rc_ipc_unresponsive) {
 617         /* If we get here, a child without IPC is being tracked, no IPC liveness
 618          * has been detected, or IPC liveness has been detected with an
 619          * unexpected (but authorized) process. This is safe on FreeBSD since
 620          * the only change possible from a proper child's PID into "special" PID
 621          * of 1 behind more loosely related process.
 622          */
 623         int ret = pcmk__pid_active(child->pid, child->name);
 624 
 625         if (ipc_pid && ((ret != pcmk_rc_ok)
 626                         || ipc_pid == PCMK__SPECIAL_PID
 627                         || (pcmk__pid_active(ipc_pid,
 628                                              child->name) == pcmk_rc_ok))) {
 629             /* An unexpected (but authorized) process was detected at the IPC
 630              * endpoint, and either it is active, or the child we're tracking is
 631              * not.
 632              */
 633 
 634             if (ret == pcmk_rc_ok) {
 635                 /* The child we're tracking is active. Kill it, and adopt the
 636                  * detected process. This assumes that our children don't fork
 637                  * (thus getting a different PID owning the IPC), but rather the
 638                  * tracking got out of sync because of some means external to
 639                  * Pacemaker, and adopting the detected process is better than
 640                  * killing it and possibly having to spawn a new child.
 641                  */
 642                 /* not possessing IPC, afterall (what about corosync CPG?) */
 643                 stop_child(child, SIGKILL);
 644             }
 645             rc = pcmk_rc_ok;
 646             child->pid = ipc_pid;
 647         } else if (ret == pcmk_rc_ok) {
 648             // Our tracked child's PID was found active, but not its IPC
 649             rc = pcmk_rc_ipc_pid_only;
 650         } else if ((child->pid == 0) && (ret == EINVAL)) {
 651             // FreeBSD can return EINVAL
 652             rc = pcmk_rc_ipc_unresponsive;
 653         } else {
 654             switch (ret) {
 655                 case EACCES:
 656                     rc = pcmk_rc_ipc_unauthorized;
 657                     break;
 658                 case ESRCH:
 659                     rc = pcmk_rc_ipc_unresponsive;
 660                     break;
 661                 default:
 662                     rc = ret;
 663                     break;
 664             }
 665         }
 666     }
 667     return rc;
 668 }
 669 
 670 /*!
 671  * \internal
 672  * \brief Initial one-off check of the pre-existing "child" processes
 673  *
 674  * With "child" process, we mean the subdaemon that defines an API end-point
 675  * (all of them do as of the comment) -- the possible complement is skipped
 676  * as it is deemed it has no such shared resources to cause conflicts about,
 677  * hence it can presumably be started anew without hesitation.
 678  * If that won't hold true in the future, the concept of a shared resource
 679  * will have to be generalized beyond the API end-point.
 680  *
 681  * For boundary cases that the "child" is still starting (IPC end-point is yet
 682  * to be witnessed), or more rarely (practically FreeBSD only), when there's
 683  * a pre-existing "untrackable" authentic process, we give the situation some
 684  * time to possibly unfold in the right direction, meaning that said socket
 685  * will appear or the unattainable process will disappear per the observable
 686  * IPC, respectively.
 687  *
 688  * \return Standard Pacemaker return code
 689  *
 690  * \note Since this gets run at the very start, \c respawn_count fields
 691  *       for particular children get temporarily overloaded with "rounds
 692  *       of waiting" tracking, restored once we are about to finish with
 693  *       success (i.e. returning value >=0) and will remain unrestored
 694  *       otherwise.  One way to suppress liveness detection logic for
 695  *       particular child is to set the said value to a negative number.
 696  */
 697 #define WAIT_TRIES 4  /* together with interleaved sleeps, worst case ~ 1s */
 698 int
 699 find_and_track_existing_processes(void)
     /*  */
 700 {
 701     bool wait_in_progress;
 702     int rc;
 703     size_t i, rounds;
 704 
 705     for (rounds = 1; rounds <= WAIT_TRIES; rounds++) {
 706         wait_in_progress = false;
 707         for (i = 0; i < PCMK__NELEM(pcmk_children); i++) {
 708 
 709             if ((pcmk_children[i].endpoint == NULL)
 710                 || (pcmk_children[i].respawn_count < 0)) {
 711                 continue;
 712             }
 713 
 714             rc = child_liveness(&pcmk_children[i]);
 715             if (rc == pcmk_rc_ipc_unresponsive) {
 716                 /* As a speculation, don't give up if there are more rounds to
 717                  * come for other reasons, but don't artificially wait just
 718                  * because of this, since we would preferably start ASAP.
 719                  */
 720                 continue;
 721             }
 722 
 723             pcmk_children[i].respawn_count = rounds;
 724             switch (rc) {
 725                 case pcmk_rc_ok:
 726                     if (pcmk_children[i].pid == PCMK__SPECIAL_PID) {
 727                         if (crm_is_true(getenv("PCMK_fail_fast"))) {
 728                             crm_crit("Cannot reliably track pre-existing"
 729                                      " authentic process behind %s IPC on this"
 730                                      " platform and PCMK_fail_fast requested",
 731                                      pcmk_children[i].endpoint);
 732                             return EOPNOTSUPP;
 733                         } else if (pcmk_children[i].respawn_count == WAIT_TRIES) {
 734                             crm_notice("Assuming pre-existing authentic, though"
 735                                        " on this platform untrackable, process"
 736                                        " behind %s IPC is stable (was in %d"
 737                                        " previous samples) so rather than"
 738                                        " bailing out (PCMK_fail_fast not"
 739                                        " requested), we just switch to a less"
 740                                        " optimal IPC liveness monitoring"
 741                                        " (not very suitable for heavy load)",
 742                                        pcmk_children[i].name, WAIT_TRIES - 1);
 743                             crm_warn("The process behind %s IPC cannot be"
 744                                      " terminated, so the overall shutdown"
 745                                      " will get delayed implicitly (%ld s),"
 746                                      " which serves as a graceful period for"
 747                                      " its native termination if it vitally"
 748                                      " depends on some other daemons going"
 749                                      " down in a controlled way already",
 750                                      pcmk_children[i].name,
 751                                      (long) SHUTDOWN_ESCALATION_PERIOD);
 752                         } else {
 753                             wait_in_progress = true;
 754                             crm_warn("Cannot reliably track pre-existing"
 755                                      " authentic process behind %s IPC on this"
 756                                      " platform, can still disappear in %d"
 757                                      " attempt(s)", pcmk_children[i].endpoint,
 758                                      WAIT_TRIES - pcmk_children[i].respawn_count);
 759                             continue;
 760                         }
 761                     }
 762                     crm_notice("Tracking existing %s process (pid=%lld)",
 763                                pcmk_children[i].name,
 764                                (long long) PCMK__SPECIAL_PID_AS_0(
 765                                                pcmk_children[i].pid));
 766                     pcmk_children[i].respawn_count = -1;  /* 0~keep watching */
 767                     pcmk_children[i].active_before_startup = true;
 768                     break;
 769                 case pcmk_rc_ipc_pid_only:
 770                     if (pcmk_children[i].respawn_count == WAIT_TRIES) {
 771                         crm_crit("%s IPC end-point for existing authentic"
 772                                  " process %lld did not (re)appear",
 773                                  pcmk_children[i].endpoint,
 774                                  (long long) PCMK__SPECIAL_PID_AS_0(
 775                                                  pcmk_children[i].pid));
 776                         return rc;
 777                     }
 778                     wait_in_progress = true;
 779                     crm_warn("Cannot find %s IPC end-point for existing"
 780                              " authentic process %lld, can still (re)appear"
 781                              " in %d attempts (?)",
 782                              pcmk_children[i].endpoint,
 783                              (long long) PCMK__SPECIAL_PID_AS_0(
 784                                              pcmk_children[i].pid),
 785                              WAIT_TRIES - pcmk_children[i].respawn_count);
 786                     continue;
 787                 default:
 788                     crm_crit("Checked liveness of %s: %s " CRM_XS " rc=%d",
 789                              pcmk_children[i].name, pcmk_rc_str(rc), rc);
 790                     return rc;
 791             }
 792         }
 793         if (!wait_in_progress) {
 794             break;
 795         }
 796         pcmk__sleep_ms(250); // Wait a bit for changes to possibly happen
 797     }
 798     for (i = 0; i < PCMK__NELEM(pcmk_children); i++) {
 799         pcmk_children[i].respawn_count = 0;  /* restore pristine state */
 800     }
 801 
 802     g_timeout_add_seconds(PCMK_PROCESS_CHECK_INTERVAL, check_next_subdaemon,
 803                           NULL);
 804     return pcmk_rc_ok;
 805 }
 806 
 807 gboolean
 808 init_children_processes(void *user_data)
     /*  */
 809 {
 810     /* start any children that have not been detected */
 811     for (int i = 0; i < PCMK__NELEM(pcmk_children); i++) {
 812         if (pcmk_children[i].pid != 0) {
 813             /* we are already tracking it */
 814             continue;
 815         }
 816 
 817         start_child(&(pcmk_children[i]));
 818     }
 819 
 820     /* From this point on, any daemons being started will be due to
 821      * respawning rather than node start.
 822      *
 823      * This may be useful for the daemons to know
 824      */
 825     setenv("PCMK_respawned", "true", 1);
 826     pacemakerd_state = XML_PING_ATTR_PACEMAKERDSTATE_RUNNING;
 827     return TRUE;
 828 }
 829 
 830 void
 831 pcmk_shutdown(int nsig)
     /*  */
 832 {
 833     if (shutdown_trigger == NULL) {
 834         shutdown_trigger = mainloop_add_trigger(G_PRIORITY_HIGH, pcmk_shutdown_worker, NULL);
 835     }
 836     mainloop_set_trigger(shutdown_trigger);
 837 }
 838 
 839 void
 840 restart_cluster_subdaemons(void)
     /*  */
 841 {
 842     for (int i = 0; i < PCMK__NELEM(pcmk_children); i++) {
 843         if (!pcmk_children[i].needs_retry || pcmk_children[i].pid != 0) {
 844             continue;
 845         }
 846 
 847         crm_notice("Respawning cluster-based subdaemon: %s", pcmk_children[i].name);
 848         if (start_child(&pcmk_children[i])) {
 849             pcmk_children[i].needs_retry = false;
 850         }
 851     }
 852 }
 853 
 854 static gboolean
 855 stop_child(pcmk_child_t * child, int signal)
     /*  */
 856 {
 857     if (signal == 0) {
 858         signal = SIGTERM;
 859     }
 860 
 861     /* why to skip PID of 1?
 862        - FreeBSD ~ how untrackable process behind IPC is masqueraded as
 863        - elsewhere: how "init" task is designated; in particular, in systemd
 864          arrangement of socket-based activation, this is pretty real */
 865     if (child->command == NULL || child->pid == PCMK__SPECIAL_PID) {
 866         crm_debug("Nothing to do for child \"%s\" (process %lld)",
 867                   child->name, (long long) PCMK__SPECIAL_PID_AS_0(child->pid));
 868         return TRUE;
 869     }
 870 
 871     if (child->pid <= 0) {
 872         crm_trace("Client %s not running", child->name);
 873         return TRUE;
 874     }
 875 
 876     errno = 0;
 877     if (kill(child->pid, signal) == 0) {
 878         crm_notice("Stopping %s "CRM_XS" sent signal %d to process %lld",
 879                    child->name, signal, (long long) child->pid);
 880 
 881     } else {
 882         crm_err("Could not stop %s (process %lld) with signal %d: %s",
 883                 child->name, (long long) child->pid, signal, strerror(errno));
 884     }
 885 
 886     return TRUE;
 887 }
 888
/* */
root/daemons/pacemakerd/pcmkd_subdaemons.c

DEFINITIONS
