diff --git a/daemons/pacemakerd/pacemakerd.h b/daemons/pacemakerd/pacemakerd.h index 3a4186bcb4..7c541bbf9e 100644 --- a/daemons/pacemakerd/pacemakerd.h +++ b/daemons/pacemakerd/pacemakerd.h @@ -1,34 +1,36 @@ /* * Copyright 2010-2021 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 #define MAX_RESPAWN 100 extern GMainLoop *mainloop; extern struct qb_ipcs_service_handlers mcp_ipc_callbacks; extern const char *pacemakerd_state; extern gboolean running_with_sbd; extern unsigned int shutdown_complete_state_reported_to; extern gboolean shutdown_complete_state_reported_client_closed; extern crm_trigger_t *shutdown_trigger; extern crm_trigger_t *startup_trigger; gboolean mcp_read_config(void); gboolean cluster_connect_cfg(void); void cluster_disconnect_cfg(void); int find_and_track_existing_processes(void); gboolean init_children_processes(void *user_data); +void restart_cluster_subdaemons(void); void pcmk_shutdown(int nsig); void pcmk_handle_ping_request(pcmk__client_t *c, xmlNode *msg, uint32_t id); void pcmk_handle_shutdown_request(pcmk__client_t *c, xmlNode *msg, uint32_t id, uint32_t flags); void pcmkd_shutdown_corosync(void); +bool pcmkd_corosync_connected(void); diff --git a/daemons/pacemakerd/pcmkd_corosync.c b/daemons/pacemakerd/pcmkd_corosync.c index 61ce7a05a6..49b206ed82 100644 --- a/daemons/pacemakerd/pcmkd_corosync.c +++ b/daemons/pacemakerd/pcmkd_corosync.c @@ -1,352 +1,374 @@ /* * Copyright 2010-2021 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 "pacemakerd.h" #include #include /* for calls to stat() */ #include /* For basename() and dirname() */ #include #include /* For getpwname() */ #include #include #include #include #include #include /* for crm_ipc_is_authentic_process */ #include #include /* PCMK__SPECIAL_PID* */ static corosync_cfg_handle_t cfg_handle = 0; static mainloop_timer_t *reconnect_timer = NULL; /* =::=::=::= CFG - Shutdown stuff =::=::=::= */ static void cfg_shutdown_callback(corosync_cfg_handle_t h, corosync_cfg_shutdown_flags_t flags) { crm_info("Corosync wants to shut down: %s", (flags == COROSYNC_CFG_SHUTDOWN_FLAG_IMMEDIATE) ? "immediate" : (flags == COROSYNC_CFG_SHUTDOWN_FLAG_REGARDLESS) ? "forced" : "optional"); /* Never allow corosync to shut down while we're running */ corosync_cfg_replyto_shutdown(h, COROSYNC_CFG_SHUTDOWN_FLAG_NO); } static corosync_cfg_callbacks_t cfg_callbacks = { .corosync_cfg_shutdown_callback = cfg_shutdown_callback, }; static int pcmk_cfg_dispatch(gpointer user_data) { corosync_cfg_handle_t *handle = (corosync_cfg_handle_t *) user_data; cs_error_t rc = corosync_cfg_dispatch(*handle, CS_DISPATCH_ALL); if (rc != CS_OK) { return -1; } return 0; } static void close_cfg(void) { if (cfg_handle != 0) { #ifdef HAVE_COROSYNC_CFG_TRACKSTART /* Ideally, we would call corosync_cfg_trackstop(cfg_handle) here, but a * bug in corosync 3.1.1 and 3.1.2 makes it hang forever. Thankfully, * it's not necessary since we exit immediately after this. */ #endif corosync_cfg_finalize(cfg_handle); cfg_handle = 0; } } static gboolean cluster_reconnect_cb(gpointer data) { if (cluster_connect_cfg()) { mainloop_timer_del(reconnect_timer); reconnect_timer = NULL; crm_notice("Cluster reconnect succeeded"); + mcp_read_config(); + restart_cluster_subdaemons(); + return G_SOURCE_REMOVE; } else { crm_info("Cluster reconnect failed" "(connection will be reattempted once per second)"); } /* * In theory this will continue forever. In practice the CIB connection from * attrd will timeout and shut down Pacemaker when it gets bored. */ - return TRUE; + return G_SOURCE_CONTINUE; } static void cfg_connection_destroy(gpointer user_data) { crm_warn("Lost connection to cluster layer " "(connection will be reattempted once per second)"); corosync_cfg_finalize(cfg_handle); cfg_handle = 0; reconnect_timer = mainloop_timer_add("corosync reconnect", 1000, TRUE, cluster_reconnect_cb, NULL); mainloop_timer_start(reconnect_timer); } void cluster_disconnect_cfg(void) { close_cfg(); if (reconnect_timer != NULL) { /* The mainloop should be gone by this point, so this isn't necessary, * but cleaning up memory should make valgrind happier. */ mainloop_timer_del(reconnect_timer); reconnect_timer = NULL; } } #define cs_repeat(counter, max, code) do { \ code; \ if(rc == CS_ERR_TRY_AGAIN || rc == CS_ERR_QUEUE_FULL) { \ counter++; \ crm_debug("Retrying Corosync operation after %ds", counter); \ sleep(counter); \ } else { \ break; \ } \ } while(counter < max) gboolean cluster_connect_cfg(void) { cs_error_t rc; int fd = -1, retries = 0, rv; uid_t found_uid = 0; gid_t found_gid = 0; pid_t found_pid = 0; uint32_t nodeid; static struct mainloop_fd_callbacks cfg_fd_callbacks = { .dispatch = pcmk_cfg_dispatch, .destroy = cfg_connection_destroy, }; cs_repeat(retries, 30, rc = corosync_cfg_initialize(&cfg_handle, &cfg_callbacks)); if (rc != CS_OK) { crm_crit("Could not connect to Corosync CFG: %s " CRM_XS " rc=%d", cs_strerror(rc), rc); return FALSE; } rc = corosync_cfg_fd_get(cfg_handle, &fd); if (rc != CS_OK) { crm_crit("Could not get Corosync CFG descriptor: %s " CRM_XS " rc=%d", cs_strerror(rc), rc); goto bail; } /* CFG provider run as root (in given user namespace, anyway)? */ if (!(rv = crm_ipc_is_authentic_process(fd, (uid_t) 0,(gid_t) 0, &found_pid, &found_uid, &found_gid))) { crm_crit("Rejecting Corosync CFG provider because process %lld " "is running as uid %lld gid %lld, not root", (long long) PCMK__SPECIAL_PID_AS_0(found_pid), (long long) found_uid, (long long) found_gid); goto bail; } else if (rv < 0) { crm_crit("Could not authenticate Corosync CFG provider: %s " CRM_XS " rc=%d", strerror(-rv), -rv); goto bail; } retries = 0; cs_repeat(retries, 30, rc = corosync_cfg_local_get(cfg_handle, &nodeid)); if (rc != CS_OK) { crm_crit("Could not get local node ID from Corosync: %s " CRM_XS " rc=%d", cs_strerror(rc), rc); goto bail; } crm_debug("Corosync reports local node ID is %lu", (unsigned long) nodeid); #ifdef HAVE_COROSYNC_CFG_TRACKSTART rc = corosync_cfg_trackstart(cfg_handle, 0); if (rc != CS_OK) { crm_crit("Could not enable Corosync CFG shutdown tracker: %s " CRM_XS " rc=%d", cs_strerror(rc), rc); goto bail; } #endif mainloop_add_fd("corosync-cfg", G_PRIORITY_DEFAULT, fd, &cfg_handle, &cfg_fd_callbacks); return TRUE; bail: corosync_cfg_finalize(cfg_handle); return FALSE; } void pcmkd_shutdown_corosync(void) { cs_error_t rc; if (cfg_handle == 0) { crm_warn("Unable to shut down Corosync: No connection"); return; } crm_info("Asking Corosync to shut down"); rc = corosync_cfg_try_shutdown(cfg_handle, COROSYNC_CFG_SHUTDOWN_FLAG_IMMEDIATE); if (rc == CS_OK) { close_cfg(); } else { crm_warn("Corosync shutdown failed: %s " CRM_XS " rc=%d", cs_strerror(rc), rc); } } +bool +pcmkd_corosync_connected(void) +{ + cpg_handle_t local_handle = 0; + cpg_model_v1_data_t cpg_model_info = {CPG_MODEL_V1, NULL, NULL, NULL, 0}; + int fd = -1; + + if (cpg_model_initialize(&local_handle, CPG_MODEL_V1, (cpg_model_data_t *) &cpg_model_info, NULL) != CS_OK) { + return false; + } + + if (cpg_fd_get(local_handle, &fd) != CS_OK) { + return false; + } + + cpg_finalize(local_handle); + + return true; +} /* =::=::=::= Configuration =::=::=::= */ static int get_config_opt(uint64_t unused, cmap_handle_t object_handle, const char *key, char **value, const char *fallback) { int rc = 0, retries = 0; cs_repeat(retries, 5, rc = cmap_get_string(object_handle, key, value)); if (rc != CS_OK) { crm_trace("Search for %s failed %d, defaulting to %s", key, rc, fallback); if (fallback) { *value = strdup(fallback); } else { *value = NULL; } } crm_trace("%s: %s", key, *value); return rc; } gboolean mcp_read_config(void) { cs_error_t rc = CS_OK; int retries = 0; cmap_handle_t local_handle; uint64_t config = 0; int fd = -1; uid_t found_uid = 0; gid_t found_gid = 0; pid_t found_pid = 0; int rv; enum cluster_type_e stack; // There can be only one possibility do { rc = pcmk__init_cmap(&local_handle); if (rc != CS_OK) { retries++; crm_info("Could not connect to Corosync CMAP: %s (retrying in %ds) " CRM_XS " rc=%d", cs_strerror(rc), retries, rc); sleep(retries); } else { break; } } while (retries < 5); if (rc != CS_OK) { crm_crit("Could not connect to Corosync CMAP: %s " CRM_XS " rc=%d", cs_strerror(rc), rc); return FALSE; } rc = cmap_fd_get(local_handle, &fd); if (rc != CS_OK) { crm_crit("Could not get Corosync CMAP descriptor: %s " CRM_XS " rc=%d", cs_strerror(rc), rc); cmap_finalize(local_handle); return FALSE; } /* CMAP provider run as root (in given user namespace, anyway)? */ if (!(rv = crm_ipc_is_authentic_process(fd, (uid_t) 0,(gid_t) 0, &found_pid, &found_uid, &found_gid))) { crm_crit("Rejecting Corosync CMAP provider because process %lld " "is running as uid %lld gid %lld, not root", (long long) PCMK__SPECIAL_PID_AS_0(found_pid), (long long) found_uid, (long long) found_gid); cmap_finalize(local_handle); return FALSE; } else if (rv < 0) { crm_crit("Could not authenticate Corosync CMAP provider: %s " CRM_XS " rc=%d", strerror(-rv), -rv); cmap_finalize(local_handle); return FALSE; } stack = get_cluster_type(); if (stack != pcmk_cluster_corosync) { crm_crit("Expected Corosync cluster layer but detected %s " CRM_XS " stack=%d", name_for_cluster_type(stack), stack); return FALSE; } crm_info("Reading configuration for %s stack", name_for_cluster_type(stack)); pcmk__set_env_option(PCMK__ENV_CLUSTER_TYPE, "corosync"); pcmk__set_env_option(PCMK__ENV_QUORUM_TYPE, "corosync"); // If debug logging is not configured, check whether corosync has it if (pcmk__env_option(PCMK__ENV_DEBUG) == NULL) { char *debug_enabled = NULL; get_config_opt(config, local_handle, "logging.debug", &debug_enabled, "off"); if (crm_is_true(debug_enabled)) { pcmk__set_env_option(PCMK__ENV_DEBUG, "1"); if (get_crm_log_level() < LOG_DEBUG) { set_crm_log_level(LOG_DEBUG); } } else { pcmk__set_env_option(PCMK__ENV_DEBUG, "0"); } free(debug_enabled); } if(local_handle){ gid_t gid = 0; if (pcmk_daemon_user(NULL, &gid) < 0) { crm_warn("Could not authorize group with Corosync " CRM_XS " No group found for user %s", CRM_DAEMON_USER); } else { char key[PATH_MAX]; snprintf(key, PATH_MAX, "uidgid.gid.%u", gid); rc = cmap_set_uint8(local_handle, key, 1); if (rc != CS_OK) { crm_warn("Could not authorize group with Corosync: %s " CRM_XS " group=%u rc=%d", pcmk__cs_err_str(rc), gid, rc); } } } cmap_finalize(local_handle); return TRUE; } diff --git a/daemons/pacemakerd/pcmkd_subdaemons.c b/daemons/pacemakerd/pcmkd_subdaemons.c index 8737a60f99..8ac70370a4 100644 --- a/daemons/pacemakerd/pcmkd_subdaemons.c +++ b/daemons/pacemakerd/pcmkd_subdaemons.c @@ -1,806 +1,840 @@ /* * Copyright 2010-2021 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 "pacemakerd.h" #include #include #include #include #include #include #include #include #include #include #include typedef struct pcmk_child_s { pid_t pid; int respawn_count; bool respawn; const char *name; const char *uid; const char *command; const char *endpoint; /* IPC server name */ bool needs_cluster; + /* Anything below here will be dynamically initialized */ + bool needs_retry; bool active_before_startup; } pcmk_child_t; #define PCMK_PROCESS_CHECK_INTERVAL 5 #define SHUTDOWN_ESCALATION_PERIOD 180000 /* 3m */ /* Index into the array below */ #define PCMK_CHILD_CONTROLD 5 static pcmk_child_t pcmk_children[] = { { 0, 0, true, "pacemaker-based", CRM_DAEMON_USER, CRM_DAEMON_DIR "/pacemaker-based", PCMK__SERVER_BASED_RO, true }, { 0, 0, true, "pacemaker-fenced", NULL, CRM_DAEMON_DIR "/pacemaker-fenced", "stonith-ng", true }, { 0, 0, true, "pacemaker-execd", NULL, CRM_DAEMON_DIR "/pacemaker-execd", CRM_SYSTEM_LRMD, false }, { 0, 0, true, "pacemaker-attrd", CRM_DAEMON_USER, CRM_DAEMON_DIR "/pacemaker-attrd", T_ATTRD, - false + true }, { 0, 0, true, "pacemaker-schedulerd", CRM_DAEMON_USER, CRM_DAEMON_DIR "/pacemaker-schedulerd", CRM_SYSTEM_PENGINE, - true + false }, { 0, 0, true, "pacemaker-controld", CRM_DAEMON_USER, CRM_DAEMON_DIR "/pacemaker-controld", CRM_SYSTEM_CRMD, true }, }; static char *opts_default[] = { NULL, NULL }; static char *opts_vgrind[] = { NULL, NULL, NULL, NULL, NULL }; crm_trigger_t *shutdown_trigger = NULL; crm_trigger_t *startup_trigger = NULL; /* When contacted via pacemakerd-api by a client having sbd in * the name we assume it is sbd-daemon which wants to know * if pacemakerd shutdown gracefully. * Thus when everything is shutdown properly pacemakerd * waits till it has reported the graceful completion of * shutdown to sbd and just when sbd-client closes the * connection we can assume that the report has arrived * properly so that pacemakerd can finally exit. * Following two variables are used to track that handshake. */ unsigned int shutdown_complete_state_reported_to = 0; gboolean shutdown_complete_state_reported_client_closed = FALSE; /* state we report when asked via pacemakerd-api status-ping */ const char *pacemakerd_state = XML_PING_ATTR_PACEMAKERDSTATE_INIT; gboolean running_with_sbd = FALSE; /* local copy */ GMainLoop *mainloop = NULL; static gboolean fatal_error = FALSE; static bool global_keep_tracking = false; static gboolean check_active_before_startup_processes(gpointer user_data); static int child_liveness(pcmk_child_t *child); static gboolean escalate_shutdown(gpointer data); static int start_child(pcmk_child_t * child); static void pcmk_child_exit(mainloop_child_t * p, pid_t pid, int core, int signo, int exitcode); static void pcmk_process_exit(pcmk_child_t * child); static gboolean pcmk_shutdown_worker(gpointer user_data); static gboolean stop_child(pcmk_child_t * child, int signal); +static bool +pcmkd_cluster_connected() +{ +#if SUPPORT_COROSYNC + return pcmkd_corosync_connected(); +#else + return true; +#endif +} + static gboolean check_active_before_startup_processes(gpointer user_data) { gboolean keep_tracking = FALSE; for (int i = 0; i < PCMK__NELEM(pcmk_children); i++) { if (!pcmk_children[i].active_before_startup) { /* we are already tracking it as a child process. */ continue; } else { int rc = child_liveness(&pcmk_children[i]); switch (rc) { case pcmk_rc_ok: break; case pcmk_rc_ipc_unresponsive: case pcmk_rc_ipc_pid_only: // This case: it was previously OK if (pcmk_children[i].respawn) { crm_err("%s[%lld] terminated%s", pcmk_children[i].name, (long long) PCMK__SPECIAL_PID_AS_0(pcmk_children[i].pid), (rc == pcmk_rc_ipc_pid_only)? " as IPC server" : ""); } else { /* orderly shutdown */ crm_notice("%s[%lld] terminated%s", pcmk_children[i].name, (long long) PCMK__SPECIAL_PID_AS_0(pcmk_children[i].pid), (rc == pcmk_rc_ipc_pid_only)? " as IPC server" : ""); } pcmk_process_exit(&(pcmk_children[i])); continue; default: crm_exit(CRM_EX_FATAL); break; /* static analysis/noreturn */ } } /* at least one of the processes found at startup * is still going, so keep this recurring timer around */ keep_tracking = TRUE; } global_keep_tracking = keep_tracking; return keep_tracking; } static gboolean escalate_shutdown(gpointer data) { pcmk_child_t *child = data; if (child->pid == PCMK__SPECIAL_PID) { pcmk_process_exit(child); } else if (child->pid != 0) { /* Use SIGSEGV instead of SIGKILL to create a core so we can see what it was up to */ crm_err("Child %s not terminating in a timely manner, forcing", child->name); stop_child(child, SIGSEGV); } return FALSE; } static void pcmk_child_exit(mainloop_child_t * p, pid_t pid, int core, int signo, int exitcode) { pcmk_child_t *child = mainloop_child_userdata(p); const char *name = mainloop_child_name(p); if (signo) { do_crm_log(((signo == SIGKILL)? LOG_WARNING : LOG_ERR), "%s[%d] terminated with signal %d (core=%d)", name, pid, signo, core); } else { switch(exitcode) { case CRM_EX_OK: crm_info("%s[%d] exited with status %d (%s)", name, pid, exitcode, crm_exit_str(exitcode)); break; case CRM_EX_FATAL: crm_warn("Shutting cluster down because %s[%d] had fatal failure", name, pid); child->respawn = false; fatal_error = TRUE; pcmk_shutdown(SIGTERM); break; case CRM_EX_PANIC: crm_emerg("%s[%d] instructed the machine to reset", name, pid); child->respawn = false; fatal_error = TRUE; pcmk__panic(__func__); pcmk_shutdown(SIGTERM); break; default: crm_err("%s[%d] exited with status %d (%s)", name, pid, exitcode, crm_exit_str(exitcode)); break; } } pcmk_process_exit(child); } static void pcmk_process_exit(pcmk_child_t * child) { child->pid = 0; child->active_before_startup = false; child->respawn_count += 1; if (child->respawn_count > MAX_RESPAWN) { crm_err("Child respawn count exceeded by %s", child->name); child->respawn = false; } if (shutdown_trigger) { /* resume step-wise shutdown (returned TRUE yields no parallelizing) */ mainloop_set_trigger(shutdown_trigger); } else if (!child->respawn) { /* nothing to do */ } else if (crm_is_true(getenv("PCMK_fail_fast"))) { crm_err("Rebooting system because of %s", child->name); pcmk__panic(__func__); } else if (child_liveness(child) == pcmk_rc_ok) { crm_warn("One-off suppressing strict respawning of a child process %s," " appears alright per %s IPC end-point", child->name, child->endpoint); /* need to monitor how it evolves, and start new process if badly */ child->active_before_startup = true; if (!global_keep_tracking) { global_keep_tracking = true; g_timeout_add_seconds(PCMK_PROCESS_CHECK_INTERVAL, check_active_before_startup_processes, NULL); } } else { + if (child->needs_cluster && !pcmkd_cluster_connected()) { + crm_notice("Skipping cluster-based subdaemon %s until cluster returns", + child->name); + child->needs_retry = true; + return; + } + crm_notice("Respawning failed child process: %s", child->name); start_child(child); } } static gboolean pcmk_shutdown_worker(gpointer user_data) { static int phase = PCMK__NELEM(pcmk_children) - 1; static time_t next_log = 0; if (phase == PCMK__NELEM(pcmk_children) - 1) { crm_notice("Shutting down Pacemaker"); pacemakerd_state = XML_PING_ATTR_PACEMAKERDSTATE_SHUTTINGDOWN; } for (; phase >= 0; phase--) { pcmk_child_t *child = &(pcmk_children[phase]); if (child->pid != 0) { time_t now = time(NULL); if (child->respawn) { if (child->pid == PCMK__SPECIAL_PID) { crm_warn("The process behind %s IPC cannot be" " terminated, so either wait the graceful" " period of %ld s for its native termination" " if it vitally depends on some other daemons" " going down in a controlled way already," " or locate and kill the correct %s process" " on your own; set PCMK_fail_fast=1 to avoid" " this altogether next time around", child->name, (long) SHUTDOWN_ESCALATION_PERIOD, child->command); } next_log = now + 30; child->respawn = false; stop_child(child, SIGTERM); if (phase < PCMK_CHILD_CONTROLD) { g_timeout_add(SHUTDOWN_ESCALATION_PERIOD, escalate_shutdown, child); } } else if (now >= next_log) { next_log = now + 30; crm_notice("Still waiting for %s to terminate " CRM_XS " pid=%lld", child->name, (long long) child->pid); } return TRUE; } /* cleanup */ crm_debug("%s confirmed stopped", child->name); child->pid = 0; } crm_notice("Shutdown complete"); pacemakerd_state = XML_PING_ATTR_PACEMAKERDSTATE_SHUTDOWNCOMPLETE; if (!fatal_error && running_with_sbd && pcmk__get_sbd_sync_resource_startup() && !shutdown_complete_state_reported_client_closed) { crm_notice("Waiting for SBD to pick up shutdown-complete-state."); return TRUE; } { const char *delay = pcmk__env_option(PCMK__ENV_SHUTDOWN_DELAY); if(delay) { sync(); pcmk__sleep_ms(crm_get_msec(delay)); } } g_main_loop_quit(mainloop); if (fatal_error) { crm_notice("Shutting down and staying down after fatal error"); #ifdef SUPPORT_COROSYNC pcmkd_shutdown_corosync(); #endif crm_exit(CRM_EX_FATAL); } return TRUE; } /* TODO once libqb is taught to juggle with IPC end-points carried over as bare file descriptor (https://github.com/ClusterLabs/libqb/issues/325) it shall hand over these descriptors here if/once they are successfully pre-opened in (presumably) child_liveness(), to avoid any remaining room for races */ // \return Standard Pacemaker return code static int start_child(pcmk_child_t * child) { uid_t uid = 0; gid_t gid = 0; gboolean use_valgrind = FALSE; gboolean use_callgrind = FALSE; const char *env_valgrind = getenv("PCMK_valgrind_enabled"); const char *env_callgrind = getenv("PCMK_callgrind_enabled"); child->active_before_startup = false; if (child->command == NULL) { crm_info("Nothing to do for child \"%s\"", child->name); return pcmk_rc_ok; } if (env_callgrind != NULL && crm_is_true(env_callgrind)) { use_callgrind = TRUE; use_valgrind = TRUE; } else if (env_callgrind != NULL && strstr(env_callgrind, child->name)) { use_callgrind = TRUE; use_valgrind = TRUE; } else if (env_valgrind != NULL && crm_is_true(env_valgrind)) { use_valgrind = TRUE; } else if (env_valgrind != NULL && strstr(env_valgrind, child->name)) { use_valgrind = TRUE; } if (use_valgrind && strlen(VALGRIND_BIN) == 0) { crm_warn("Cannot enable valgrind for %s:" " The location of the valgrind binary is unknown", child->name); use_valgrind = FALSE; } if (child->uid) { if (crm_user_lookup(child->uid, &uid, &gid) < 0) { crm_err("Invalid user (%s) for %s: not found", child->uid, child->name); return EACCES; } crm_info("Using uid=%u and group=%u for process %s", uid, gid, child->name); } child->pid = fork(); CRM_ASSERT(child->pid != -1); if (child->pid > 0) { /* parent */ mainloop_child_add(child->pid, 0, child->name, child, pcmk_child_exit); crm_info("Forked child %lld for process %s%s", (long long) child->pid, child->name, use_valgrind ? " (valgrind enabled: " VALGRIND_BIN ")" : ""); return pcmk_rc_ok; } else { /* Start a new session */ (void)setsid(); /* Setup the two alternate arg arrays */ opts_vgrind[0] = strdup(VALGRIND_BIN); if (use_callgrind) { opts_vgrind[1] = strdup("--tool=callgrind"); opts_vgrind[2] = strdup("--callgrind-out-file=" CRM_STATE_DIR "/callgrind.out.%p"); opts_vgrind[3] = strdup(child->command); opts_vgrind[4] = NULL; } else { opts_vgrind[1] = strdup(child->command); opts_vgrind[2] = NULL; opts_vgrind[3] = NULL; opts_vgrind[4] = NULL; } opts_default[0] = strdup(child->command); if(gid) { // Whether we need root group access to talk to cluster layer bool need_root_group = TRUE; if (is_corosync_cluster()) { /* Corosync clusters can drop root group access, because we set * uidgid.gid.${gid}=1 via CMAP, which allows these processes to * connect to corosync. */ need_root_group = FALSE; } // Drop root group access if not needed if (!need_root_group && (setgid(gid) < 0)) { crm_warn("Could not set group to %d: %s", gid, strerror(errno)); } /* Initialize supplementary groups to only those always granted to * the user, plus haclient (so we can access IPC). */ if (initgroups(child->uid, gid) < 0) { crm_err("Cannot initialize groups for %s: %s (%d)", child->uid, pcmk_strerror(errno), errno); } } if (uid && setuid(uid) < 0) { crm_warn("Could not set user to %s (id %d): %s", child->uid, uid, strerror(errno)); } pcmk__close_fds_in_child(true); pcmk__open_devnull(O_RDONLY); // stdin (fd 0) pcmk__open_devnull(O_WRONLY); // stdout (fd 1) pcmk__open_devnull(O_WRONLY); // stderr (fd 2) if (use_valgrind) { (void)execvp(VALGRIND_BIN, opts_vgrind); } else { (void)execvp(child->command, opts_default); } crm_crit("Could not execute %s: %s", child->command, strerror(errno)); crm_exit(CRM_EX_FATAL); } return pcmk_rc_ok; /* never reached */ } /*! * \internal * \brief Check the liveness of the child based on IPC name and PID if tracked * * \param[inout] child Child tracked data * * \return Standard Pacemaker return code * * \note Return codes of particular interest include pcmk_rc_ipc_unresponsive * indicating that no trace of IPC liveness was detected, * pcmk_rc_ipc_unauthorized indicating that the IPC endpoint is blocked by * an unauthorized process, and pcmk_rc_ipc_pid_only indicating that * the child is up by PID but not IPC end-point (possibly starting). * \note This function doesn't modify any of \p child members but \c pid, * and is not actively toying with processes as such but invoking * \c stop_child in one particular case (there's for some reason * a different authentic holder of the IPC end-point). */ static int child_liveness(pcmk_child_t *child) { uid_t cl_uid = 0; gid_t cl_gid = 0; const uid_t root_uid = 0; const gid_t root_gid = 0; const uid_t *ref_uid; const gid_t *ref_gid; int rc = pcmk_rc_ipc_unresponsive; pid_t ipc_pid = 0; if (child->endpoint == NULL && (child->pid <= 0 || child->pid == PCMK__SPECIAL_PID)) { crm_err("Cannot track child %s for missing both API end-point and PID", child->name); rc = EINVAL; // Misuse of function when child is not trackable } else if (child->endpoint != NULL) { int legacy_rc = pcmk_ok; if (child->uid == NULL) { ref_uid = &root_uid; ref_gid = &root_gid; } else { ref_uid = &cl_uid; ref_gid = &cl_gid; legacy_rc = pcmk_daemon_user(&cl_uid, &cl_gid); } if (legacy_rc < 0) { rc = pcmk_legacy2rc(legacy_rc); crm_err("Could not find user and group IDs for user %s: %s " CRM_XS " rc=%d", CRM_DAEMON_USER, pcmk_rc_str(rc), rc); } else { rc = pcmk__ipc_is_authentic_process_active(child->endpoint, *ref_uid, *ref_gid, &ipc_pid); if ((rc == pcmk_rc_ok) || (rc == pcmk_rc_ipc_unresponsive)) { if (child->pid <= 0) { /* If rc is pcmk_rc_ok, ipc_pid is nonzero and this * initializes a new child. If rc is * pcmk_rc_ipc_unresponsive, ipc_pid is zero, and we will * investigate further. */ child->pid = ipc_pid; } else if ((ipc_pid != 0) && (child->pid != ipc_pid)) { /* An unexpected (but authorized) process is responding to * IPC. Investigate further. */ rc = pcmk_rc_ipc_unresponsive; } } } } if (rc == pcmk_rc_ipc_unresponsive) { /* If we get here, a child without IPC is being tracked, no IPC liveness * has been detected, or IPC liveness has been detected with an * unexpected (but authorized) process. This is safe on FreeBSD since * the only change possible from a proper child's PID into "special" PID * of 1 behind more loosely related process. */ int ret = pcmk__pid_active(child->pid, child->name); if (ipc_pid && ((ret != pcmk_rc_ok) || ipc_pid == PCMK__SPECIAL_PID || (pcmk__pid_active(ipc_pid, child->name) == pcmk_rc_ok))) { /* An unexpected (but authorized) process was detected at the IPC * endpoint, and either it is active, or the child we're tracking is * not. */ if (ret == pcmk_rc_ok) { /* The child we're tracking is active. Kill it, and adopt the * detected process. This assumes that our children don't fork * (thus getting a different PID owning the IPC), but rather the * tracking got out of sync because of some means external to * Pacemaker, and adopting the detected process is better than * killing it and possibly having to spawn a new child. */ /* not possessing IPC, afterall (what about corosync CPG?) */ stop_child(child, SIGKILL); } rc = pcmk_rc_ok; child->pid = ipc_pid; } else if (ret == pcmk_rc_ok) { // Our tracked child's PID was found active, but not its IPC rc = pcmk_rc_ipc_pid_only; } else if ((child->pid == 0) && (ret == EINVAL)) { // FreeBSD can return EINVAL rc = pcmk_rc_ipc_unresponsive; } else { switch (ret) { case EACCES: rc = pcmk_rc_ipc_unauthorized; break; case ESRCH: rc = pcmk_rc_ipc_unresponsive; break; default: rc = ret; break; } } } return rc; } /*! * \internal * \brief Initial one-off check of the pre-existing "child" processes * * With "child" process, we mean the subdaemon that defines an API end-point * (all of them do as of the comment) -- the possible complement is skipped * as it is deemed it has no such shared resources to cause conflicts about, * hence it can presumably be started anew without hesitation. * If that won't hold true in the future, the concept of a shared resource * will have to be generalized beyond the API end-point. * * For boundary cases that the "child" is still starting (IPC end-point is yet * to be witnessed), or more rarely (practically FreeBSD only), when there's * a pre-existing "untrackable" authentic process, we give the situation some * time to possibly unfold in the right direction, meaning that said socket * will appear or the unattainable process will disappear per the observable * IPC, respectively. * * \return Standard Pacemaker return code * * \note Since this gets run at the very start, \c respawn_count fields * for particular children get temporarily overloaded with "rounds * of waiting" tracking, restored once we are about to finish with * success (i.e. returning value >=0) and will remain unrestored * otherwise. One way to suppress liveness detection logic for * particular child is to set the said value to a negative number. */ #define WAIT_TRIES 4 /* together with interleaved sleeps, worst case ~ 1s */ int find_and_track_existing_processes(void) { bool tracking = false; bool wait_in_progress; int rc; size_t i, rounds; for (rounds = 1; rounds <= WAIT_TRIES; rounds++) { wait_in_progress = false; for (i = 0; i < PCMK__NELEM(pcmk_children); i++) { if ((pcmk_children[i].endpoint == NULL) || (pcmk_children[i].respawn_count < 0)) { continue; } rc = child_liveness(&pcmk_children[i]); if (rc == pcmk_rc_ipc_unresponsive) { /* As a speculation, don't give up if there are more rounds to * come for other reasons, but don't artificially wait just * because of this, since we would preferably start ASAP. */ continue; } pcmk_children[i].respawn_count = rounds; switch (rc) { case pcmk_rc_ok: if (pcmk_children[i].pid == PCMK__SPECIAL_PID) { if (crm_is_true(getenv("PCMK_fail_fast"))) { crm_crit("Cannot reliably track pre-existing" " authentic process behind %s IPC on this" " platform and PCMK_fail_fast requested", pcmk_children[i].endpoint); return EOPNOTSUPP; } else if (pcmk_children[i].respawn_count == WAIT_TRIES) { crm_notice("Assuming pre-existing authentic, though" " on this platform untrackable, process" " behind %s IPC is stable (was in %d" " previous samples) so rather than" " bailing out (PCMK_fail_fast not" " requested), we just switch to a less" " optimal IPC liveness monitoring" " (not very suitable for heavy load)", pcmk_children[i].name, WAIT_TRIES - 1); crm_warn("The process behind %s IPC cannot be" " terminated, so the overall shutdown" " will get delayed implicitly (%ld s)," " which serves as a graceful period for" " its native termination if it vitally" " depends on some other daemons going" " down in a controlled way already", pcmk_children[i].name, (long) SHUTDOWN_ESCALATION_PERIOD); } else { wait_in_progress = true; crm_warn("Cannot reliably track pre-existing" " authentic process behind %s IPC on this" " platform, can still disappear in %d" " attempt(s)", pcmk_children[i].endpoint, WAIT_TRIES - pcmk_children[i].respawn_count); continue; } } crm_notice("Tracking existing %s process (pid=%lld)", pcmk_children[i].name, (long long) PCMK__SPECIAL_PID_AS_0( pcmk_children[i].pid)); pcmk_children[i].respawn_count = -1; /* 0~keep watching */ pcmk_children[i].active_before_startup = true; tracking = true; break; case pcmk_rc_ipc_pid_only: if (pcmk_children[i].respawn_count == WAIT_TRIES) { crm_crit("%s IPC end-point for existing authentic" " process %lld did not (re)appear", pcmk_children[i].endpoint, (long long) PCMK__SPECIAL_PID_AS_0( pcmk_children[i].pid)); return rc; } wait_in_progress = true; crm_warn("Cannot find %s IPC end-point for existing" " authentic process %lld, can still (re)appear" " in %d attempts (?)", pcmk_children[i].endpoint, (long long) PCMK__SPECIAL_PID_AS_0( pcmk_children[i].pid), WAIT_TRIES - pcmk_children[i].respawn_count); continue; default: crm_crit("Checked liveness of %s: %s " CRM_XS " rc=%d", pcmk_children[i].name, pcmk_rc_str(rc), rc); return rc; } } if (!wait_in_progress) { break; } pcmk__sleep_ms(250); // Wait a bit for changes to possibly happen } for (i = 0; i < PCMK__NELEM(pcmk_children); i++) { pcmk_children[i].respawn_count = 0; /* restore pristine state */ } if (tracking) { g_timeout_add_seconds(PCMK_PROCESS_CHECK_INTERVAL, check_active_before_startup_processes, NULL); } return pcmk_rc_ok; } gboolean init_children_processes(void *user_data) { /* start any children that have not been detected */ for (int i = 0; i < PCMK__NELEM(pcmk_children); i++) { if (pcmk_children[i].pid != 0) { /* we are already tracking it */ continue; } start_child(&(pcmk_children[i])); } /* From this point on, any daemons being started will be due to * respawning rather than node start. * * This may be useful for the daemons to know */ setenv("PCMK_respawned", "true", 1); pacemakerd_state = XML_PING_ATTR_PACEMAKERDSTATE_RUNNING; return TRUE; } void pcmk_shutdown(int nsig) { if (shutdown_trigger == NULL) { shutdown_trigger = mainloop_add_trigger(G_PRIORITY_HIGH, pcmk_shutdown_worker, NULL); } mainloop_set_trigger(shutdown_trigger); } +void +restart_cluster_subdaemons(void) +{ + for (int i = 0; i < PCMK__NELEM(pcmk_children); i++) { + if (!pcmk_children[i].needs_retry || pcmk_children[i].pid != 0) { + continue; + } + + crm_notice("Respawning cluster-based subdaemon: %s", pcmk_children[i].name); + if (start_child(&pcmk_children[i])) { + pcmk_children[i].needs_retry = false; + } + } +} + static gboolean stop_child(pcmk_child_t * child, int signal) { if (signal == 0) { signal = SIGTERM; } /* why to skip PID of 1? - FreeBSD ~ how untrackable process behind IPC is masqueraded as - elsewhere: how "init" task is designated; in particular, in systemd arrangement of socket-based activation, this is pretty real */ if (child->command == NULL || child->pid == PCMK__SPECIAL_PID) { crm_debug("Nothing to do for child \"%s\" (process %lld)", child->name, (long long) PCMK__SPECIAL_PID_AS_0(child->pid)); return TRUE; } if (child->pid <= 0) { crm_trace("Client %s not running", child->name); return TRUE; } errno = 0; if (kill(child->pid, signal) == 0) { crm_notice("Stopping %s "CRM_XS" sent signal %d to process %lld", child->name, signal, (long long) child->pid); } else { crm_err("Could not stop %s (process %lld) with signal %d: %s", child->name, (long long) child->pid, signal, strerror(errno)); } return TRUE; }