diff --git a/daemons/pacemakerd/Makefile.am b/daemons/pacemakerd/Makefile.am index 78e7c37fe6..7571a6c1dd 100644 --- a/daemons/pacemakerd/Makefile.am +++ b/daemons/pacemakerd/Makefile.am @@ -1,39 +1,40 @@ # # Copyright 2004-2023 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 $(top_srcdir)/mk/common.mk include $(top_srcdir)/mk/man.mk sbin_PROGRAMS = pacemakerd if BUILD_SYSTEMD systemdsystemunit_DATA = pacemaker.service endif EXTRA_DIST = pacemakerd.8.inc ## SOURCES -noinst_HEADERS = pacemakerd.h +noinst_HEADERS = pacemakerd.h \ + pcmkd_corosync.h pacemakerd_CFLAGS = $(CFLAGS_HARDENED_EXE) pacemakerd_LDFLAGS = $(LDFLAGS_HARDENED_EXE) pacemakerd_LDADD = $(top_builddir)/lib/cluster/libcrmcluster.la pacemakerd_LDADD += $(top_builddir)/lib/common/libcrmcommon.la pacemakerd_LDADD += $(CLUSTERLIBS) pacemakerd_SOURCES = pacemakerd.c if BUILD_CS_SUPPORT pacemakerd_SOURCES += pcmkd_corosync.c endif pacemakerd_SOURCES += pcmkd_messages.c pacemakerd_SOURCES += pcmkd_subdaemons.c CLEANFILES = $(man8_MANS) diff --git a/daemons/pacemakerd/pacemakerd.c b/daemons/pacemakerd/pacemakerd.c index 365b743d21..6a3f20216f 100644 --- a/daemons/pacemakerd/pacemakerd.c +++ b/daemons/pacemakerd/pacemakerd.c @@ -1,485 +1,489 @@ /* * Copyright 2010-2023 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" +#if SUPPORT_COROSYNC +#include "pcmkd_corosync.h" +#endif + #include #include #include #include #include #include #include #include #include #include /* indirectly: CRM_EX_* */ #include #include #include #include #include #include #include #define SUMMARY "pacemakerd - primary Pacemaker daemon that launches and monitors all subsidiary Pacemaker daemons" struct { gboolean features; gboolean foreground; gboolean shutdown; gboolean standby; } options; static pcmk__output_t *out = NULL; static pcmk__supported_format_t formats[] = { PCMK__SUPPORTED_FORMAT_NONE, PCMK__SUPPORTED_FORMAT_TEXT, PCMK__SUPPORTED_FORMAT_XML, { NULL, NULL, NULL } }; PCMK__OUTPUT_ARGS("features") static int pacemakerd_features(pcmk__output_t *out, va_list args) { out->info(out, "Pacemaker %s (Build: %s)\n Supporting v%s: %s", PACEMAKER_VERSION, BUILD_VERSION, CRM_FEATURE_SET, CRM_FEATURES); return pcmk_rc_ok; } PCMK__OUTPUT_ARGS("features") static int pacemakerd_features_xml(pcmk__output_t *out, va_list args) { gchar **feature_list = g_strsplit(CRM_FEATURES, " ", 0); pcmk__output_xml_create_parent(out, "pacemakerd", "version", PACEMAKER_VERSION, "build", BUILD_VERSION, "feature_set", CRM_FEATURE_SET, NULL); out->begin_list(out, NULL, NULL, "features"); for (char **s = feature_list; *s != NULL; s++) { pcmk__output_create_xml_text_node(out, "feature", *s); } out->end_list(out); g_strfreev(feature_list); return pcmk_rc_ok; } static pcmk__message_entry_t fmt_functions[] = { { "features", "default", pacemakerd_features }, { "features", "xml", pacemakerd_features_xml }, { NULL, NULL, NULL } }; static gboolean pid_cb(const gchar *option_name, const gchar *optarg, gpointer data, GError **err) { return TRUE; } static gboolean standby_cb(const gchar *option_name, const gchar *optarg, gpointer data, GError **err) { options.standby = TRUE; pcmk__set_env_option(PCMK__ENV_NODE_START_STATE, "standby", false); return TRUE; } static GOptionEntry entries[] = { { "features", 'F', 0, G_OPTION_ARG_NONE, &options.features, "Display full version and list of features Pacemaker was built with", NULL }, { "foreground", 'f', 0, G_OPTION_ARG_NONE, &options.foreground, "(Ignored) Pacemaker always runs in the foreground", NULL }, { "pid-file", 'p', 0, G_OPTION_ARG_CALLBACK, pid_cb, "(Ignored) Daemon pid file location", "FILE" }, { "shutdown", 'S', 0, G_OPTION_ARG_NONE, &options.shutdown, "Instruct Pacemaker to shutdown on this machine", NULL }, { "standby", 's', G_OPTION_FLAG_NO_ARG, G_OPTION_ARG_CALLBACK, standby_cb, "Start node in standby state", NULL }, { NULL } }; static void pcmk_ignore(int nsig) { crm_info("Ignoring signal %s (%d)", strsignal(nsig), nsig); } static void pcmk_sigquit(int nsig) { pcmk__panic(__func__); } static void pacemakerd_chown(const char *path, uid_t uid, gid_t gid) { int rc = chown(path, uid, gid); if (rc < 0) { crm_warn("Cannot change the ownership of %s to user %s and gid %d: %s", path, CRM_DAEMON_USER, gid, pcmk_rc_str(errno)); } } static void create_pcmk_dirs(void) { uid_t pcmk_uid = 0; gid_t pcmk_gid = 0; const char *dirs[] = { CRM_PACEMAKER_DIR, // core/blackbox/scheduler/CIB files CRM_CORE_DIR, // core files CRM_BLACKBOX_DIR, // blackbox dumps PE_STATE_DIR, // scheduler inputs CRM_CONFIG_DIR, // the Cluster Information Base (CIB) // Don't build CRM_RSCTMP_DIR, pacemaker-execd will do it NULL }; if (pcmk_daemon_user(&pcmk_uid, &pcmk_gid) < 0) { crm_err("Cluster user %s does not exist, aborting Pacemaker startup", CRM_DAEMON_USER); crm_exit(CRM_EX_NOUSER); } // Used by some resource agents if ((mkdir(CRM_STATE_DIR, 0750) < 0) && (errno != EEXIST)) { crm_warn("Could not create directory " CRM_STATE_DIR ": %s", pcmk_rc_str(errno)); } else { pacemakerd_chown(CRM_STATE_DIR, pcmk_uid, pcmk_gid); } for (int i = 0; dirs[i] != NULL; ++i) { int rc = pcmk__build_path(dirs[i], 0750); if (rc != pcmk_rc_ok) { crm_warn("Could not create directory %s: %s", dirs[i], pcmk_rc_str(rc)); } else { pacemakerd_chown(dirs[i], pcmk_uid, pcmk_gid); } } } static void remove_core_file_limit(void) { struct rlimit cores; // Get current limits if (getrlimit(RLIMIT_CORE, &cores) < 0) { crm_notice("Unable to check system core file limits " "(consider ensuring the size is unlimited): %s", strerror(errno)); return; } // Check whether core dumps are disabled if (cores.rlim_max == 0) { if (geteuid() != 0) { // Yes, and there's nothing we can do about it crm_notice("Core dumps are disabled (consider enabling them)"); return; } cores.rlim_max = RLIM_INFINITY; // Yes, but we're root, so enable them } // Raise soft limit to hard limit (if not already done) if (cores.rlim_cur != cores.rlim_max) { cores.rlim_cur = cores.rlim_max; if (setrlimit(RLIMIT_CORE, &cores) < 0) { crm_notice("Unable to raise system limit on core file size " "(consider doing so manually): %s", strerror(errno)); return; } } if (cores.rlim_cur == RLIM_INFINITY) { crm_trace("Core file size is unlimited"); } else { crm_trace("Core file size is limited to %llu bytes", (unsigned long long) cores.rlim_cur); } } static void pacemakerd_event_cb(pcmk_ipc_api_t *pacemakerd_api, enum pcmk_ipc_event event_type, crm_exit_t status, void *event_data, void *user_data) { pcmk_pacemakerd_api_reply_t *reply = event_data; switch (event_type) { case pcmk_ipc_event_reply: break; default: return; } if (status != CRM_EX_OK) { out->err(out, "Bad reply from pacemakerd: %s", crm_exit_str(status)); return; } if (reply->reply_type != pcmk_pacemakerd_reply_shutdown) { out->err(out, "Unknown reply type %d from pacemakerd", reply->reply_type); } } static GOptionContext * build_arg_context(pcmk__common_args_t *args, GOptionGroup **group) { GOptionContext *context = NULL; context = pcmk__build_arg_context(args, "text (default), xml", group, NULL); pcmk__add_main_args(context, entries); return context; } int main(int argc, char **argv) { int rc = pcmk_rc_ok; crm_exit_t exit_code = CRM_EX_OK; GError *error = NULL; GOptionGroup *output_group = NULL; pcmk__common_args_t *args = pcmk__new_common_args(SUMMARY); gchar **processed_args = pcmk__cmdline_preproc(argv, "p"); GOptionContext *context = build_arg_context(args, &output_group); bool old_instance_connected = false; pcmk_ipc_api_t *old_instance = NULL; qb_ipcs_service_t *ipcs = NULL; subdaemon_check_progress = time(NULL); setenv("LC_ALL", "C", 1); // Ensure logs are in a common language crm_log_preinit(NULL, argc, argv); mainloop_add_signal(SIGHUP, pcmk_ignore); mainloop_add_signal(SIGQUIT, pcmk_sigquit); pcmk__register_formats(output_group, formats); if (!g_option_context_parse_strv(context, &processed_args, &error)) { exit_code = CRM_EX_USAGE; goto done; } rc = pcmk__output_new(&out, args->output_ty, args->output_dest, argv); if ((rc != pcmk_rc_ok) || (out == NULL)) { exit_code = CRM_EX_ERROR; g_set_error(&error, PCMK__EXITC_ERROR, exit_code, "Error creating output format %s: %s", args->output_ty, pcmk_rc_str(rc)); goto done; } pcmk__force_args(context, &error, "%s --xml-simple-list", g_get_prgname()); pcmk__register_messages(out, fmt_functions); if (options.features) { out->message(out, "features"); exit_code = CRM_EX_OK; goto done; } if (args->version) { out->version(out, false); goto done; } // @COMPAT Drop at 3.0.0; likely last used in 1.1.24 pcmk__set_env_option(PCMK__ENV_MCP, "true", true); if (options.shutdown) { pcmk__cli_init_logging("pacemakerd", args->verbosity); } else { crm_log_init(NULL, LOG_INFO, TRUE, FALSE, argc, argv, FALSE); } crm_debug("Checking for existing Pacemaker instance"); rc = pcmk_new_ipc_api(&old_instance, pcmk_ipc_pacemakerd); if (old_instance == NULL) { out->err(out, "Could not check for existing pacemakerd: %s", pcmk_rc_str(rc)); exit_code = pcmk_rc2exitc(rc); goto done; } pcmk_register_ipc_callback(old_instance, pacemakerd_event_cb, NULL); rc = pcmk__connect_ipc(old_instance, pcmk_ipc_dispatch_sync, 2); if (rc != pcmk_rc_ok) { crm_debug("No existing %s instance found: %s", pcmk_ipc_name(old_instance, true), pcmk_rc_str(rc)); } old_instance_connected = pcmk_ipc_is_connected(old_instance); if (options.shutdown) { if (old_instance_connected) { rc = pcmk_pacemakerd_api_shutdown(old_instance, crm_system_name); pcmk_dispatch_ipc(old_instance); exit_code = pcmk_rc2exitc(rc); if (exit_code != CRM_EX_OK) { pcmk_free_ipc_api(old_instance); goto done; } /* We get the ACK immediately, and the response right after that, * but it might take a while for pacemakerd to get around to * shutting down. Wait for that to happen (with 30-minute timeout). */ for (int i = 0; i < 900; i++) { if (!pcmk_ipc_is_connected(old_instance)) { exit_code = CRM_EX_OK; pcmk_free_ipc_api(old_instance); goto done; } sleep(2); } exit_code = CRM_EX_TIMEOUT; pcmk_free_ipc_api(old_instance); goto done; } else { out->err(out, "Could not request shutdown " "of existing Pacemaker instance: %s", pcmk_rc_str(rc)); pcmk_free_ipc_api(old_instance); exit_code = CRM_EX_DISCONNECT; goto done; } } else if (old_instance_connected) { pcmk_free_ipc_api(old_instance); crm_err("Aborting start-up because active Pacemaker instance found"); exit_code = CRM_EX_FATAL; goto done; } pcmk_free_ipc_api(old_instance); /* Don't allow any accidental output after this point. */ if (out != NULL) { out->finish(out, exit_code, true, NULL); pcmk__output_free(out); out = NULL; } #ifdef SUPPORT_COROSYNC if (pacemakerd_read_config() == FALSE) { crm_exit(CRM_EX_UNAVAILABLE); } #endif // OCF shell functions and cluster-glue need facility under different name { const char *facility = pcmk__env_option(PCMK__ENV_LOGFACILITY); if (!pcmk__str_eq(facility, PCMK__VALUE_NONE, pcmk__str_casei|pcmk__str_null_matches)) { pcmk__set_env_option("LOGFACILITY", facility, true); } } crm_notice("Starting Pacemaker %s "CRM_XS" build=%s features:%s", PACEMAKER_VERSION, BUILD_VERSION, CRM_FEATURES); mainloop = g_main_loop_new(NULL, FALSE); remove_core_file_limit(); create_pcmk_dirs(); pcmk__serve_pacemakerd_ipc(&ipcs, &pacemakerd_ipc_callbacks); #ifdef SUPPORT_COROSYNC /* Allows us to block shutdown */ if (!cluster_connect_cfg()) { exit_code = CRM_EX_PROTOCOL; goto done; } #endif if (pcmk__locate_sbd() > 0) { running_with_sbd = TRUE; } switch (find_and_track_existing_processes()) { case pcmk_rc_ok: break; case pcmk_rc_ipc_unauthorized: exit_code = CRM_EX_CANTCREAT; goto done; default: exit_code = CRM_EX_FATAL; goto done; }; mainloop_add_signal(SIGTERM, pcmk_shutdown); mainloop_add_signal(SIGINT, pcmk_shutdown); if ((running_with_sbd) && pcmk__get_sbd_sync_resource_startup()) { crm_notice("Waiting for startup-trigger from SBD."); pacemakerd_state = XML_PING_ATTR_PACEMAKERDSTATE_WAITPING; startup_trigger = mainloop_add_trigger(G_PRIORITY_HIGH, init_children_processes, NULL); } else { if (running_with_sbd) { crm_warn("Enabling SBD_SYNC_RESOURCE_STARTUP would (if supported " "by your SBD version) improve reliability of " "interworking between SBD & pacemaker."); } pacemakerd_state = XML_PING_ATTR_PACEMAKERDSTATE_STARTINGDAEMONS; init_children_processes(NULL); } crm_notice("Pacemaker daemon successfully started and accepting connections"); g_main_loop_run(mainloop); if (ipcs) { crm_trace("Closing IPC server"); mainloop_del_ipc_server(ipcs); ipcs = NULL; } g_main_loop_unref(mainloop); #ifdef SUPPORT_COROSYNC cluster_disconnect_cfg(); #endif done: g_strfreev(processed_args); pcmk__free_arg_context(context); pcmk__output_and_clear_error(&error, out); if (out != NULL) { out->finish(out, exit_code, true, NULL); pcmk__output_free(out); } pcmk__unregister_formats(); crm_exit(exit_code); } diff --git a/daemons/pacemakerd/pacemakerd.h b/daemons/pacemakerd/pacemakerd.h index ee6facfb36..51e32b19e8 100644 --- a/daemons/pacemakerd/pacemakerd.h +++ b/daemons/pacemakerd/pacemakerd.h @@ -1,35 +1,27 @@ /* * Copyright 2010-2023 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 pacemakerd_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 unsigned int shutdown_complete_state_reported_to; extern crm_trigger_t *shutdown_trigger; extern crm_trigger_t *startup_trigger; extern time_t subdaemon_check_progress; -gboolean pacemakerd_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 pcmkd_shutdown_corosync(void); -bool pcmkd_corosync_connected(void); +void restart_cluster_subdaemons(void); diff --git a/daemons/pacemakerd/pcmkd_corosync.c b/daemons/pacemakerd/pcmkd_corosync.c index 8a1a8671eb..f96611f603 100644 --- a/daemons/pacemakerd/pcmkd_corosync.c +++ b/daemons/pacemakerd/pcmkd_corosync.c @@ -1,373 +1,374 @@ /* * Copyright 2010-2023 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 "pcmkd_corosync.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"); pacemakerd_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 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 retries = 0; cs_repeat(retries, 30, 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); pcmk__str_update(value, fallback); } crm_trace("%s: %s", key, *value); return rc; } gboolean pacemakerd_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", true); // @COMPAT Drop at 3.0.0; added unused in 1.1.9 pcmk__set_env_option(PCMK__ENV_QUORUM_TYPE, "corosync", true); // 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", true); if (get_crm_log_level() < LOG_DEBUG) { set_crm_log_level(LOG_DEBUG); } } else { pcmk__set_env_option(PCMK__ENV_DEBUG, "0", true); } 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_corosync.h b/daemons/pacemakerd/pcmkd_corosync.h new file mode 100644 index 0000000000..8c4a1e10bc --- /dev/null +++ b/daemons/pacemakerd/pcmkd_corosync.h @@ -0,0 +1,16 @@ +/* + * Copyright 2010-2023 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 + +gboolean cluster_connect_cfg(void); +void cluster_disconnect_cfg(void); +gboolean pacemakerd_read_config(void); +bool pcmkd_corosync_connected(void); +void pcmkd_shutdown_corosync(void); diff --git a/daemons/pacemakerd/pcmkd_subdaemons.c b/daemons/pacemakerd/pcmkd_subdaemons.c index 859037c8f9..5ac361d203 100644 --- a/daemons/pacemakerd/pcmkd_subdaemons.c +++ b/daemons/pacemakerd/pcmkd_subdaemons.c @@ -1,893 +1,897 @@ /* * Copyright 2010-2023 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" +#if SUPPORT_COROSYNC +#include "pcmkd_corosync.h" +#endif + #include #include #include #include #include #include #include #include #include #include #include enum child_daemon_flags { child_none = 0, child_respawn = 1 << 0, child_needs_cluster = 1 << 1, child_needs_retry = 1 << 2, child_active_before_startup = 1 << 3, }; typedef struct pcmk_child_s { pid_t pid; int respawn_count; const char *name; const char *uid; const char *command; const char *endpoint; /* IPC server name */ int check_count; uint32_t flags; } pcmk_child_t; #define PCMK_PROCESS_CHECK_INTERVAL 1 #define PCMK_PROCESS_CHECK_RETRIES 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, "pacemaker-based", CRM_DAEMON_USER, CRM_DAEMON_DIR "/pacemaker-based", PCMK__SERVER_BASED_RO, 0, child_respawn | child_needs_cluster }, { 0, 0, "pacemaker-fenced", NULL, CRM_DAEMON_DIR "/pacemaker-fenced", "stonith-ng", 0, child_respawn | child_needs_cluster }, { 0, 0, "pacemaker-execd", NULL, CRM_DAEMON_DIR "/pacemaker-execd", CRM_SYSTEM_LRMD, 0, child_respawn }, { 0, 0, "pacemaker-attrd", CRM_DAEMON_USER, CRM_DAEMON_DIR "/pacemaker-attrd", T_ATTRD, 0, child_respawn | child_needs_cluster }, { 0, 0, "pacemaker-schedulerd", CRM_DAEMON_USER, CRM_DAEMON_DIR "/pacemaker-schedulerd", CRM_SYSTEM_PENGINE, 0, child_respawn }, { 0, 0, "pacemaker-controld", CRM_DAEMON_USER, CRM_DAEMON_DIR "/pacemaker-controld", CRM_SYSTEM_CRMD, 0, child_respawn | child_needs_cluster }, }; 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; time_t subdaemon_check_progress = 0; // Whether we need root group access to talk to cluster layer static bool need_root_group = true; /* 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 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(void) { #if SUPPORT_COROSYNC return pcmkd_corosync_connected(); #else return true; #endif } static gboolean check_next_subdaemon(gpointer user_data) { static int next_child = 0; int rc = child_liveness(&pcmk_children[next_child]); crm_trace("Checked %s[%lld]: %s (%d)", pcmk_children[next_child].name, (long long) PCMK__SPECIAL_PID_AS_0(pcmk_children[next_child].pid), pcmk_rc_str(rc), rc); switch (rc) { case pcmk_rc_ok: pcmk_children[next_child].check_count = 0; subdaemon_check_progress = time(NULL); break; case pcmk_rc_ipc_pid_only: // This case: it was previously OK pcmk_children[next_child].check_count++; if (pcmk_children[next_child].check_count >= PCMK_PROCESS_CHECK_RETRIES) { crm_err("%s[%lld] is unresponsive to ipc after %d tries but " "we found the pid so have it killed that we can restart", pcmk_children[next_child].name, (long long) PCMK__SPECIAL_PID_AS_0( pcmk_children[next_child].pid), pcmk_children[next_child].check_count); stop_child(&pcmk_children[next_child], SIGKILL); if (pcmk_is_set(pcmk_children[next_child].flags, child_respawn)) { /* as long as the respawn-limit isn't reached give it another round of check retries */ pcmk_children[next_child].check_count = 0; } } else { crm_notice("%s[%lld] is unresponsive to ipc after %d tries", pcmk_children[next_child].name, (long long) PCMK__SPECIAL_PID_AS_0( pcmk_children[next_child].pid), pcmk_children[next_child].check_count); if (pcmk_is_set(pcmk_children[next_child].flags, child_respawn)) { /* as long as the respawn-limit isn't reached and we haven't run out of connect retries we account this as progress we are willing to tell to sbd */ subdaemon_check_progress = time(NULL); } } /* go to the next child and see if we can make progress there */ break; case pcmk_rc_ipc_unresponsive: if (!pcmk_is_set(pcmk_children[next_child].flags, child_respawn)) { /* if a subdaemon is down and we don't want it to be restarted this is a success during shutdown. if it isn't restarted anymore due to MAX_RESPAWN it is rather no success. */ if (pcmk_children[next_child].respawn_count <= MAX_RESPAWN) { subdaemon_check_progress = time(NULL); } } if (!pcmk_is_set(pcmk_children[next_child].flags, child_active_before_startup)) { crm_trace("found %s[%lld] missing - signal-handler " "will take care of it", pcmk_children[next_child].name, (long long) PCMK__SPECIAL_PID_AS_0( pcmk_children[next_child].pid)); break; } if (pcmk_is_set(pcmk_children[next_child].flags, child_respawn)) { crm_err("%s[%lld] terminated", pcmk_children[next_child].name, (long long) PCMK__SPECIAL_PID_AS_0( pcmk_children[next_child].pid)); } else { /* orderly shutdown */ crm_notice("%s[%lld] terminated", pcmk_children[next_child].name, (long long) PCMK__SPECIAL_PID_AS_0( pcmk_children[next_child].pid)); } pcmk_process_exit(&(pcmk_children[next_child])); break; default: crm_exit(CRM_EX_FATAL); break; /* static analysis/noreturn */ } next_child++; if (next_child >= PCMK__NELEM(pcmk_children)) { next_child = 0; } return G_SOURCE_CONTINUE; } 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 (%s)%s", name, pid, signo, strsignal(signo), (core? " and dumped 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->flags &= ~child_respawn; fatal_error = TRUE; pcmk_shutdown(SIGTERM); break; case CRM_EX_PANIC: crm_emerg("%s[%d] instructed the machine to reset", name, pid); child->flags &= ~child_respawn; 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->flags &= ~child_active_before_startup; child->check_count = 0; child->respawn_count += 1; if (child->respawn_count > MAX_RESPAWN) { crm_err("Child respawn count exceeded by %s", child->name); child->flags &= ~child_respawn; } if (shutdown_trigger) { /* resume step-wise shutdown (returned TRUE yields no parallelizing) */ mainloop_set_trigger(shutdown_trigger); } else if (!pcmk_is_set(child->flags, child_respawn)) { /* nothing to do */ } else if (crm_is_true(pcmk__env_option(PCMK__ENV_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); } else if (pcmk_is_set(child->flags, child_needs_cluster) && !pcmkd_cluster_connected()) { crm_notice("Not respawning %s subdaemon until cluster returns", child->name); child->flags |= child_needs_retry; } else { crm_notice("Respawning %s subdaemon after unexpected exit", 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 (pcmk_is_set(child->flags, 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_" PCMK__ENV_FAIL_FAST "=1" " to avoid this altogether next time around", child->name, (long) SHUTDOWN_ESCALATION_PERIOD, child->command); } next_log = now + 30; child->flags &= ~child_respawn; 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 = pcmk__env_option(PCMK__ENV_VALGRIND_ENABLED); const char *env_callgrind = pcmk__env_option(PCMK__ENV_CALLGRIND_ENABLED); child->flags &= ~child_active_before_startup; child->check_count = 0; 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) { // 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_rc_str(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[in,out] 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 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; } // @TODO Functionize more of this to reduce nesting pcmk_children[i].respawn_count = rounds; switch (rc) { case pcmk_rc_ok: if (pcmk_children[i].pid == PCMK__SPECIAL_PID) { if (crm_is_true(pcmk__env_option(PCMK__ENV_FAIL_FAST))) { crm_crit("Cannot reliably track pre-existing" " authentic process behind %s IPC on this" " platform and PCMK_" PCMK__ENV_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_" PCMK__ENV_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].flags |= child_active_before_startup; 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 */ } g_timeout_add_seconds(PCMK_PROCESS_CHECK_INTERVAL, check_next_subdaemon, NULL); return pcmk_rc_ok; } gboolean init_children_processes(void *user_data) { 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; } /* 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 */ pcmk__set_env_option(PCMK__ENV_RESPAWNED, "true", false); 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_is_set(pcmk_children[i].flags, child_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].flags &= ~child_needs_retry; } } } 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; }