diff --git a/daemons/pacemakerd/pacemakerd.c b/daemons/pacemakerd/pacemakerd.c index fdc1d9f52e..d8ff53d0cd 100644 --- a/daemons/pacemakerd/pacemakerd.c +++ b/daemons/pacemakerd/pacemakerd.c @@ -1,1420 +1,1416 @@ /* * Copyright 2010-2019 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 /* indirectly: CRM_EX_* */ #include /* cib_channel_ro */ #include #include #include #include #include #include /* PCMK__SPECIAL_PID*, ... */ #ifdef SUPPORT_COROSYNC #include #endif #include #include static gboolean pcmk_quorate = FALSE; static gboolean fatal_error = FALSE; static GMainLoop *mainloop = NULL; static bool global_keep_tracking = false; #define PCMK_PROCESS_CHECK_INTERVAL 5 static const char *local_name = NULL; static uint32_t local_nodeid = 0; static crm_trigger_t *shutdown_trigger = NULL; static const char *pid_file = PCMK_RUN_DIR "/pacemaker.pid"; typedef struct pcmk_child_s { int pid; long flag; int start_seq; int respawn_count; gboolean respawn; const char *name; const char *uid; const char *command; const char *endpoint; /* IPC server name */ gboolean active_before_startup; } pcmk_child_t; /* Index into the array below */ #define PCMK_CHILD_CONTROLD 3 static pcmk_child_t pcmk_children[] = { { 0, crm_proc_none, 0, 0, FALSE, "none", NULL, NULL }, { 0, crm_proc_execd, 3, 0, TRUE, "pacemaker-execd", NULL, CRM_DAEMON_DIR "/pacemaker-execd", CRM_SYSTEM_LRMD }, { 0, crm_proc_based, 1, 0, TRUE, "pacemaker-based", CRM_DAEMON_USER, CRM_DAEMON_DIR "/pacemaker-based", CIB_CHANNEL_RO }, { 0, crm_proc_controld, 6, 0, TRUE, "pacemaker-controld", CRM_DAEMON_USER, CRM_DAEMON_DIR "/pacemaker-controld", CRM_SYSTEM_CRMD }, { 0, crm_proc_attrd, 4, 0, TRUE, "pacemaker-attrd", CRM_DAEMON_USER, CRM_DAEMON_DIR "/pacemaker-attrd", T_ATTRD }, { 0, crm_proc_schedulerd, 5, 0, TRUE, "pacemaker-schedulerd", CRM_DAEMON_USER, CRM_DAEMON_DIR "/pacemaker-schedulerd", CRM_SYSTEM_PENGINE }, { 0, crm_proc_fenced, 2, 0, TRUE, "pacemaker-fenced", NULL, CRM_DAEMON_DIR "/pacemaker-fenced", "stonith-ng" }, }; static gboolean check_active_before_startup_processes(gpointer user_data); static int pcmk_child_active(pcmk_child_t *child); static gboolean start_child(pcmk_child_t * child); static gboolean update_node_processes(uint32_t id, const char *uname, uint32_t procs); void update_process_clients(crm_client_t *client); static uint32_t get_process_list(void) { int lpc = 0; uint32_t procs = crm_get_cluster_proc(); for (lpc = 0; lpc < SIZEOF(pcmk_children); lpc++) { if (pcmk_children[lpc].pid != 0) { procs |= pcmk_children[lpc].flag; } } return procs; } static void pcmk_process_exit(pcmk_child_t * child) { child->pid = 0; child->active_before_startup = FALSE; /* Broadcast the fact that one of our processes died ASAP * * Try to get some logging of the cause out first though * because we're probably about to get fenced * * Potentially do this only if respawn_count > N * to allow for local recovery */ update_node_processes(local_nodeid, NULL, get_process_list()); 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); /* intended to speed up propagating expected lay-off of the daemons? */ update_node_processes(local_nodeid, NULL, get_process_list()); } 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(__FUNCTION__); } else if (pcmk_child_active(child) == 1) { 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 { crm_notice("Respawning failed child process: %s", child->name); start_child(child); } } static void pcmk_exit_with_cluster(int exitcode) { #ifdef SUPPORT_COROSYNC corosync_cfg_handle_t cfg_handle; cs_error_t err; if (exitcode == CRM_EX_FATAL) { crm_info("Asking Corosync to shut down"); err = corosync_cfg_initialize(&cfg_handle, NULL); if (err != CS_OK) { crm_warn("Unable to open handle to corosync to close it down. err=%d", err); } err = corosync_cfg_try_shutdown(cfg_handle, COROSYNC_CFG_SHUTDOWN_FLAG_IMMEDIATE); if (err != CS_OK) { crm_warn("Corosync shutdown failed. err=%d", err); } corosync_cfg_finalize(cfg_handle); } #endif crm_exit(exitcode); } 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: do_crm_log_always(LOG_EMERG, "%s[%d] instructed the machine to reset", name, pid); child->respawn = FALSE; fatal_error = TRUE; pcmk_panic(__FUNCTION__); 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 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 %d", child->name, signal, child->pid); } else { crm_perror(LOG_ERR, "Could not stop %s (process %d) with signal %d", child->name, child->pid, signal); } return TRUE; } static char *opts_default[] = { NULL, NULL }; static char *opts_vgrind[] = { NULL, NULL, NULL, NULL, NULL }; /* 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) pcmk_child_active, to avoid any remaining room for races */ static gboolean start_child(pcmk_child_t * child) { - int lpc = 0; uid_t uid = 0; gid_t gid = 0; - struct rlimit oflimits; gboolean use_valgrind = FALSE; gboolean use_callgrind = FALSE; const char *devnull = "/dev/null"; 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 TRUE; } 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 FALSE; } 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 %d for process %s%s", child->pid, child->name, use_valgrind ? " (valgrind enabled: " VALGRIND_BIN ")" : ""); update_node_processes(local_nodeid, NULL, get_process_list()); return TRUE; } 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_perror(LOG_ERR, "Could not set group to %d", gid); } /* 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_perror(LOG_ERR, "Could not set user to %d (%s)", uid, child->uid); } - /* Close all open file descriptors */ - getrlimit(RLIMIT_NOFILE, &oflimits); - for (lpc = 0; lpc < oflimits.rlim_cur; lpc++) { - close(lpc); - } + pcmk__close_fds_in_child(true); (void)open(devnull, O_RDONLY); /* Stdin: fd 0 */ (void)open(devnull, O_WRONLY); /* Stdout: fd 1 */ (void)open(devnull, O_WRONLY); /* Stderr: fd 2 */ if (use_valgrind) { (void)execvp(VALGRIND_BIN, opts_vgrind); } else { (void)execvp(child->command, opts_default); } crm_perror(LOG_ERR, "FATAL: Cannot exec %s", child->command); crm_exit(CRM_EX_FATAL); } return TRUE; /* never reached */ } 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) { /* 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; } #define SHUTDOWN_ESCALATION_PERIOD 180000 /* 3m */ static gboolean pcmk_shutdown_worker(gpointer user_data) { static int phase = 0; static time_t next_log = 0; static int max = SIZEOF(pcmk_children); int lpc = 0; if (phase == 0) { crm_notice("Shutting down Pacemaker"); phase = max; } for (; phase > 0; phase--) { /* Don't stop anything with start_seq < 1 */ for (lpc = max - 1; lpc >= 0; lpc--) { pcmk_child_t *child = &(pcmk_children[lpc]); if (phase != child->start_seq) { continue; } if (child->pid) { 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_children[PCMK_CHILD_CONTROLD].start_seq) { 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=%d seq=%d", child->name, child->pid, child->start_seq); } return TRUE; } /* cleanup */ crm_debug("%s confirmed stopped", child->name); child->pid = 0; } } /* send_cluster_id(); */ crm_notice("Shutdown complete"); { const char *delay = daemon_option("shutdown_delay"); if(delay) { sync(); sleep(crm_get_msec(delay) / 1000); } } g_main_loop_quit(mainloop); if (fatal_error) { crm_notice("Shutting down and staying down after fatal error"); pcmk_exit_with_cluster(CRM_EX_FATAL); } return TRUE; } static void pcmk_ignore(int nsig) { crm_info("Ignoring signal %s (%d)", strsignal(nsig), nsig); } static void pcmk_sigquit(int nsig) { pcmk_panic(__FUNCTION__); } 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); } static int32_t pcmk_ipc_accept(qb_ipcs_connection_t * c, uid_t uid, gid_t gid) { crm_trace("Connection %p", c); if (crm_client_new(c, uid, gid) == NULL) { return -EIO; } return 0; } static void pcmk_ipc_created(qb_ipcs_connection_t * c) { crm_trace("Connection %p", c); } /* Exit code means? */ static int32_t pcmk_ipc_dispatch(qb_ipcs_connection_t * qbc, void *data, size_t size) { uint32_t id = 0; uint32_t flags = 0; const char *task = NULL; crm_client_t *c = crm_client_get(qbc); xmlNode *msg = crm_ipcs_recv(c, data, size, &id, &flags); crm_ipcs_send_ack(c, id, flags, "ack", __FUNCTION__, __LINE__); if (msg == NULL) { return 0; } task = crm_element_value(msg, F_CRM_TASK); if (crm_str_eq(task, CRM_OP_QUIT, TRUE)) { /* Time to quit */ crm_notice("Shutting down in response to ticket %s (%s)", crm_element_value(msg, F_CRM_REFERENCE), crm_element_value(msg, F_CRM_ORIGIN)); pcmk_shutdown(15); } else if (crm_str_eq(task, CRM_OP_RM_NODE_CACHE, TRUE)) { /* Send to everyone */ struct iovec *iov; int id = 0; const char *name = NULL; crm_element_value_int(msg, XML_ATTR_ID, &id); name = crm_element_value(msg, XML_ATTR_UNAME); crm_notice("Instructing peers to remove references to node %s/%u", name, id); iov = calloc(1, sizeof(struct iovec)); iov->iov_base = dump_xml_unformatted(msg); iov->iov_len = 1 + strlen(iov->iov_base); send_cpg_iov(iov); } else { update_process_clients(c); } free_xml(msg); return 0; } /* Error code means? */ static int32_t pcmk_ipc_closed(qb_ipcs_connection_t * c) { crm_client_t *client = crm_client_get(c); if (client == NULL) { return 0; } crm_trace("Connection %p", c); crm_client_destroy(client); return 0; } static void pcmk_ipc_destroy(qb_ipcs_connection_t * c) { crm_trace("Connection %p", c); pcmk_ipc_closed(c); } struct qb_ipcs_service_handlers mcp_ipc_callbacks = { .connection_accept = pcmk_ipc_accept, .connection_created = pcmk_ipc_created, .msg_process = pcmk_ipc_dispatch, .connection_closed = pcmk_ipc_closed, .connection_destroyed = pcmk_ipc_destroy }; /*! * \internal * \brief Send an XML message with process list of all known peers to client(s) * * \param[in] client Send message to this client, or all clients if NULL */ void update_process_clients(crm_client_t *client) { GHashTableIter iter; crm_node_t *node = NULL; xmlNode *update = create_xml_node(NULL, "nodes"); if (is_corosync_cluster()) { crm_xml_add_int(update, "quorate", pcmk_quorate); } g_hash_table_iter_init(&iter, crm_peer_cache); while (g_hash_table_iter_next(&iter, NULL, (gpointer *) & node)) { xmlNode *xml = create_xml_node(update, "node"); crm_xml_add_int(xml, "id", node->id); crm_xml_add(xml, "uname", node->uname); crm_xml_add(xml, "state", node->state); crm_xml_add_int(xml, "processes", node->processes); } if(client) { crm_trace("Sending process list to client %s", client->id); crm_ipcs_send(client, 0, update, crm_ipc_server_event); } else { crm_trace("Sending process list to %d clients", crm_hash_table_size(client_connections)); g_hash_table_iter_init(&iter, client_connections); while (g_hash_table_iter_next(&iter, NULL, (gpointer *) & client)) { crm_ipcs_send(client, 0, update, crm_ipc_server_event); } } free_xml(update); } /*! * \internal * \brief Send a CPG message with local node's process list to all peers */ static void update_process_peers(void) { /* Do nothing for corosync-2 based clusters */ struct iovec *iov = calloc(1, sizeof(struct iovec)); CRM_ASSERT(iov); if (local_name) { iov->iov_base = crm_strdup_printf("", local_name, get_process_list()); } else { iov->iov_base = crm_strdup_printf("", get_process_list()); } iov->iov_len = strlen(iov->iov_base) + 1; crm_trace("Sending %s", (char*) iov->iov_base); send_cpg_iov(iov); } /*! * \internal * \brief Update a node's process list, notifying clients and peers if needed * * \param[in] id Node ID of affected node * \param[in] uname Uname of affected node * \param[in] procs Affected node's process list mask * * \return TRUE if the process list changed, FALSE otherwise */ static gboolean update_node_processes(uint32_t id, const char *uname, uint32_t procs) { gboolean changed = FALSE; crm_node_t *node = crm_get_peer(id, uname); if (procs != 0) { if (procs != node->processes) { crm_debug("Node %s now has process list: %.32x (was %.32x)", node->uname, procs, node->processes); node->processes = procs; changed = TRUE; /* If local node's processes have changed, notify clients/peers */ if (id == local_nodeid) { update_process_clients(NULL); update_process_peers(); } } else { crm_trace("Node %s still has process list: %.32x", node->uname, procs); } } return changed; } /* *INDENT-OFF* */ static struct crm_option long_options[] = { /* Top-level Options */ {"help", 0, 0, '?', "\tThis text"}, {"version", 0, 0, '$', "\tVersion information" }, {"verbose", 0, 0, 'V', "\tIncrease debug output"}, {"shutdown", 0, 0, 'S', "\tInstruct Pacemaker to shutdown on this machine"}, {"features", 0, 0, 'F', "\tDisplay the full version and list of features Pacemaker was built with"}, {"-spacer-", 1, 0, '-', "\nAdditional Options:"}, {"foreground", 0, 0, 'f', "\t(Ignored) Pacemaker always runs in the foreground"}, {"pid-file", 1, 0, 'p', "\t(Ignored) Daemon pid file location"}, {"standby", 0, 0, 's', "\tStart node in standby state"}, {NULL, 0, 0, 0} }; /* *INDENT-ON* */ static void mcp_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_strerror(errno)); } } /*! * \internal * \brief Check the liveness of the child based on IPC name and PID if tracked * * \param[inout] child Child tracked data * * \return 0 if no trace of child's liveness detected (while it is detectable * to begin with, at least according to one of the two properties), * 1 if everything is fine, 2 if it's up per PID, but not per IPC * end-point (still starting?), -1 on error, and -2 when the child * (its IPC) blocked with an unauthorized process (log message * emitted in both latter cases) * * \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 pcmk_child_active(pcmk_child_t *child) { static uid_t cl_uid = 0; static 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 ret = 0; 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); ret = -1; /* misuse of the function when child is not trackable */ } else if (child->endpoint != NULL) { ref_uid = (child->uid != NULL) ? &cl_uid : &root_uid; ref_gid = (child->uid != NULL) ? &cl_gid : &root_gid; if (child->uid != NULL && !cl_uid && !cl_gid && crm_user_lookup(CRM_DAEMON_USER, &cl_uid, &cl_gid) < 0) { crm_err("Could not find user and group IDs for user %s", CRM_DAEMON_USER); ret = -1; } else if ((ret = pcmk__ipc_is_authentic_process_active(child->endpoint, *ref_uid, *ref_gid, &ipc_pid)) < 0) { /* game over */ } else if (child->pid <= 0) { /* hit new child to be initialized, or reset to zero and investigate further for ret == 0 */ child->pid = ipc_pid; } else if (ipc_pid && child->pid != ipc_pid) { /* ultimately strange for ret == 1; either way, investigate */ ret = 0; } } if (!ret) { /* when no IPC based liveness detected (incl. if ever a child without IPC is tracked), or detected for a different _authentic_ process; safe on FreeBSD since the only change possible from a proper child's PID into "special" PID of 1 behind more loosely related process */ ret = crm_pid_active(child->pid, child->name); if (ipc_pid && (ret != 1 || ipc_pid == PCMK__SPECIAL_PID || crm_pid_active(ipc_pid, child->name) == 1)) { if (ret == 1) { /* assume there's no forking-while-retaining-IPC-socket involved in the "children's" lifecycle, hence that the tracking got out of sync purely because of some external (esotheric?) forces (user initiated process "refresh" by force? or intentionally racing on start-up, even?), and that switching over to this other detected, authentic instance with an IPC already in possession is a better trade-off than "neutralizing" it first so as to give either the original or possibly a new to-be-spawned daemon process a leeway for operation, which would otherwise have to be carried out */ /* not possessing IPC, afterall (what about corosync CPG?) */ stop_child(child, SIGKILL); } else { ret = 1; } child->pid = ipc_pid; } else if (ret == 1) { ret = 2; /* up per PID, but not per IPC (still starting?) */ } else if (!child->pid && ret == -1) { ret = 0; /* correct -1 on FreeBSD from above back to 0 */ } } return ret; } static gboolean check_active_before_startup_processes(gpointer user_data) { int start_seq = 1, lpc = 0; static int max = SIZEOF(pcmk_children); gboolean keep_tracking = FALSE; for (start_seq = 1; start_seq < max; start_seq++) { for (lpc = 0; lpc < max; lpc++) { if (pcmk_children[lpc].active_before_startup == FALSE) { /* we are already tracking it as a child process. */ continue; } else if (start_seq != pcmk_children[lpc].start_seq) { continue; } else { const char *name = pcmk_children[lpc].name; int ret; switch ((ret = pcmk_child_active(&pcmk_children[lpc]))) { case 1: break; case 0: case 2: /* this very case: it was OK once already */ if (pcmk_children[lpc].respawn == TRUE) { crm_err("%s[%d] terminated%s", name, PCMK__SPECIAL_PID_AS_0(pcmk_children[lpc].pid), ret ? " as IPC server" : ""); } else { /* orderly shutdown */ crm_notice("%s[%d] terminated%s", name, PCMK__SPECIAL_PID_AS_0(pcmk_children[lpc].pid), ret ? " as IPC server" : ""); } pcmk_process_exit(&(pcmk_children[lpc])); continue; default: crm_crit("Unexpected value from pcmk_child_active:" " %d (pid=%lld)", ret, (long long) PCMK__SPECIAL_PID_AS_0( pcmk_children[lpc].pid)); /* fall through */ case -1: case -2: /* message(s) already emitted */ 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; } /*! * \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 0 if no such "child" process found, positive number X when X * "children" detected, -1 on an internal error, -2 when any * would-be-used IPC is blocked with an unauthorized process * * \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 */ static int find_and_track_existing_processes(void) { unsigned tracking = 0U; bool wait_in_progress; int cur; size_t i, rounds; for (rounds = 1; rounds <= WAIT_TRIES; rounds++) { wait_in_progress = false; for (i = 0; i < SIZEOF(pcmk_children); i++) { if (!pcmk_children[i].endpoint || pcmk_children[i].respawn_count < 0 || !(cur = pcmk_child_active(&pcmk_children[i]))) { /* as a speculation, don't give up in the context of pcmk_child_active check 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 (cur) { case 1: 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 -1; } 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++; break; case 2: 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 -1; } 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; case -1: case -2: return cur; /* messages already emitted */ default: crm_crit("Unexpected condition"CRM_XS"cur=%d", cur); return -1; /* unexpected condition */ } } if (!wait_in_progress) { break; } (void) poll(NULL, 0, 250); /* a bit for changes to possibly happen */ } for (i = 0; i < SIZEOF(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 (tracking > INT_MAX) ? INT_MAX : tracking; } static void init_children_processes(void) { int start_seq = 1, lpc = 0; static int max = SIZEOF(pcmk_children); /* start any children that have not been detected */ for (start_seq = 1; start_seq < max; start_seq++) { /* don't start anything with start_seq < 1 */ for (lpc = 0; lpc < max; lpc++) { if (pcmk_children[lpc].pid) { /* we are already tracking it */ continue; } if (start_seq == pcmk_children[lpc].start_seq) { start_child(&(pcmk_children[lpc])); } } } /* 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); } static void mcp_cpg_destroy(gpointer user_data) { crm_crit("Lost connection to cluster layer, shutting down"); crm_exit(CRM_EX_DISCONNECT); } /*! * \internal * \brief Process a CPG message (process list or manual peer cache removal) * * \param[in] handle CPG connection (ignored) * \param[in] groupName CPG group name (ignored) * \param[in] nodeid ID of affected node * \param[in] pid Process ID (ignored) * \param[in] msg CPG XML message * \param[in] msg_len Length of msg in bytes (ignored) */ static void mcp_cpg_deliver(cpg_handle_t handle, const struct cpg_name *groupName, uint32_t nodeid, uint32_t pid, void *msg, size_t msg_len) { xmlNode *xml = string2xml(msg); const char *task = crm_element_value(xml, F_CRM_TASK); crm_trace("Received CPG message (%s): %.200s", (task? task : "process list"), (char*)msg); if (task == NULL) { if (nodeid == local_nodeid) { crm_debug("Ignoring message with local node's process list"); } else { uint32_t procs = 0; const char *uname = crm_element_value(xml, "uname"); crm_element_value_int(xml, "proclist", (int *)&procs); if (update_node_processes(nodeid, uname, procs)) { update_process_clients(NULL); } } } else if (crm_str_eq(task, CRM_OP_RM_NODE_CACHE, TRUE)) { int id = 0; const char *name = NULL; crm_element_value_int(xml, XML_ATTR_ID, &id); name = crm_element_value(xml, XML_ATTR_UNAME); reap_crm_member(id, name); } if (xml != NULL) { free_xml(xml); } } static void mcp_cpg_membership(cpg_handle_t handle, const struct cpg_name *groupName, const struct cpg_address *member_list, size_t member_list_entries, const struct cpg_address *left_list, size_t left_list_entries, const struct cpg_address *joined_list, size_t joined_list_entries) { /* Update peer cache if needed */ pcmk_cpg_membership(handle, groupName, member_list, member_list_entries, left_list, left_list_entries, joined_list, joined_list_entries); /* Always broadcast our own presence after any membership change */ update_process_peers(); } static gboolean mcp_quorum_callback(unsigned long long seq, gboolean quorate) { pcmk_quorate = quorate; return TRUE; } static void mcp_quorum_destroy(gpointer user_data) { crm_info("connection lost"); } int main(int argc, char **argv) { int rc; int flag; int argerr = 0; int option_index = 0; gboolean shutdown = FALSE; uid_t pcmk_uid = 0; gid_t pcmk_gid = 0; struct rlimit cores; crm_ipc_t *old_instance = NULL; qb_ipcs_service_t *ipcs = NULL; static crm_cluster_t cluster; crm_log_preinit(NULL, argc, argv); crm_set_options(NULL, "mode [options]", long_options, "Start/Stop Pacemaker\n"); mainloop_add_signal(SIGHUP, pcmk_ignore); mainloop_add_signal(SIGQUIT, pcmk_sigquit); while (1) { flag = crm_get_option(argc, argv, &option_index); if (flag == -1) break; switch (flag) { case 'V': crm_bump_log_level(argc, argv); break; case 'f': /* Legacy */ break; case 'p': pid_file = optarg; break; case 's': set_daemon_option("node_start_state", "standby"); break; case '$': case '?': crm_help(flag, CRM_EX_OK); break; case 'S': shutdown = TRUE; break; case 'F': printf("Pacemaker %s (Build: %s)\n Supporting v%s: %s\n", PACEMAKER_VERSION, BUILD_VERSION, CRM_FEATURE_SET, CRM_FEATURES); crm_exit(CRM_EX_OK); default: printf("Argument code 0%o (%c) is not (?yet?) supported\n", flag, flag); ++argerr; break; } } if (optind < argc) { printf("non-option ARGV-elements: "); while (optind < argc) printf("%s ", argv[optind++]); printf("\n"); } if (argerr) { crm_help('?', CRM_EX_USAGE); } setenv("LC_ALL", "C", 1); set_daemon_option("mcp", "true"); crm_log_init(NULL, LOG_INFO, TRUE, FALSE, argc, argv, FALSE); crm_debug("Checking for existing Pacemaker instance"); old_instance = crm_ipc_new(CRM_SYSTEM_MCP, 0); (void) crm_ipc_connect(old_instance); if (shutdown) { crm_debug("Shutting down existing Pacemaker instance by request"); while (crm_ipc_connected(old_instance)) { xmlNode *cmd = create_request(CRM_OP_QUIT, NULL, NULL, CRM_SYSTEM_MCP, CRM_SYSTEM_MCP, NULL); crm_debug("."); crm_ipc_send(old_instance, cmd, 0, 0, NULL); free_xml(cmd); sleep(2); } crm_ipc_close(old_instance); crm_ipc_destroy(old_instance); crm_exit(CRM_EX_OK); } else if (crm_ipc_connected(old_instance)) { crm_ipc_close(old_instance); crm_ipc_destroy(old_instance); crm_err("Aborting start-up because active Pacemaker instance found"); crm_exit(CRM_EX_FATAL); } crm_ipc_close(old_instance); crm_ipc_destroy(old_instance); if (mcp_read_config() == FALSE) { crm_notice("Could not obtain corosync config data, exiting"); crm_exit(CRM_EX_UNAVAILABLE); } // OCF shell functions and cluster-glue need facility under different name { const char *facility = daemon_option("logfacility"); if (facility && safe_str_neq(facility, "none")) { setenv("HA_LOGFACILITY", facility, 1); } } crm_notice("Starting Pacemaker %s "CRM_XS" build=%s features:%s", PACEMAKER_VERSION, BUILD_VERSION, CRM_FEATURES); mainloop = g_main_loop_new(NULL, FALSE); rc = getrlimit(RLIMIT_CORE, &cores); if (rc < 0) { crm_perror(LOG_ERR, "Cannot determine current maximum core size."); } else { if (cores.rlim_max == 0 && geteuid() == 0) { cores.rlim_max = RLIM_INFINITY; } else { crm_info("Maximum core file size is: %lu", (unsigned long)cores.rlim_max); } cores.rlim_cur = cores.rlim_max; rc = setrlimit(RLIMIT_CORE, &cores); if (rc < 0) { crm_perror(LOG_ERR, "Core file generation will remain disabled." " Core files are an important diagnostic tool, so" " please consider enabling them by default."); } } if (crm_user_lookup(CRM_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 " CRM_STATE_DIR ": %s", pcmk_strerror(errno)); } else { mcp_chown(CRM_STATE_DIR, pcmk_uid, pcmk_gid); } /* Used to store core/blackbox/scheduler/cib files in */ crm_build_path(CRM_PACEMAKER_DIR, 0750); mcp_chown(CRM_PACEMAKER_DIR, pcmk_uid, pcmk_gid); /* Used to store core files in */ crm_build_path(CRM_CORE_DIR, 0750); mcp_chown(CRM_CORE_DIR, pcmk_uid, pcmk_gid); /* Used to store blackbox dumps in */ crm_build_path(CRM_BLACKBOX_DIR, 0750); mcp_chown(CRM_BLACKBOX_DIR, pcmk_uid, pcmk_gid); // Used to store scheduler inputs in crm_build_path(PE_STATE_DIR, 0750); mcp_chown(PE_STATE_DIR, pcmk_uid, pcmk_gid); /* Used to store the cluster configuration */ crm_build_path(CRM_CONFIG_DIR, 0750); mcp_chown(CRM_CONFIG_DIR, pcmk_uid, pcmk_gid); // Don't build CRM_RSCTMP_DIR, pacemaker-execd will do it ipcs = mainloop_add_ipc_server(CRM_SYSTEM_MCP, QB_IPC_NATIVE, &mcp_ipc_callbacks); if (ipcs == NULL) { crm_err("Couldn't start IPC server"); crm_exit(CRM_EX_OSERR); } /* Allows us to block shutdown */ if (cluster_connect_cfg(&local_nodeid) == FALSE) { crm_err("Couldn't connect to Corosync's CFG service"); crm_exit(CRM_EX_PROTOCOL); } if(pcmk_locate_sbd() > 0) { setenv("PCMK_watchdog", "true", 1); } else { setenv("PCMK_watchdog", "false", 1); } switch (find_and_track_existing_processes()) { case -1: crm_crit("Internal fatality, see the log"); crm_exit(CRM_EX_FATAL); case -2: crm_crit("Blocked by foreign process, kill the offender"); crm_exit(CRM_EX_CANTCREAT); default: break; }; cluster.destroy = mcp_cpg_destroy; cluster.cpg.cpg_deliver_fn = mcp_cpg_deliver; cluster.cpg.cpg_confchg_fn = mcp_cpg_membership; crm_set_autoreap(FALSE); rc = pcmk_ok; if (cluster_connect_cpg(&cluster) == FALSE) { crm_err("Couldn't connect to Corosync's CPG service"); rc = -ENOPROTOOPT; } else if (cluster_connect_quorum(mcp_quorum_callback, mcp_quorum_destroy) == FALSE) { rc = -ENOTCONN; } else { local_name = get_local_node_name(); update_node_processes(local_nodeid, local_name, get_process_list()); mainloop_add_signal(SIGTERM, pcmk_shutdown); mainloop_add_signal(SIGINT, pcmk_shutdown); init_children_processes(); 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); cluster_disconnect_cpg(&cluster); cluster_disconnect_cfg(); crm_exit(crm_errno2exit(rc)); } diff --git a/include/crm/common/internal.h b/include/crm/common/internal.h index b2eec000a6..da2c7d7822 100644 --- a/include/crm/common/internal.h +++ b/include/crm/common/internal.h @@ -1,168 +1,171 @@ /* * Copyright 2015-2019 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU Lesser General Public License * version 2.1 or later (LGPLv2.1+) WITHOUT ANY WARRANTY. */ #ifndef CRM_COMMON_INTERNAL__H #define CRM_COMMON_INTERNAL__H #include /* for gboolean */ #include /* for struct dirent */ #include /* for getpid() */ +#include /* for bool */ #include /* for uid_t and gid_t */ #include /* internal I/O utilities (from io.c) */ char *generate_series_filename(const char *directory, const char *series, int sequence, gboolean bzip); int get_last_sequence(const char *directory, const char *series); void write_last_sequence(const char *directory, const char *series, int sequence, int max); int crm_chown_last_sequence(const char *directory, const char *series, uid_t uid, gid_t gid); bool pcmk__daemon_can_write(const char *dir, const char *file); void crm_sync_directory(const char *name); char *crm_read_contents(const char *filename); int crm_write_sync(int fd, const char *contents); int crm_set_nonblocking(int fd); const char *crm_get_tmpdir(void); +void pcmk__close_fds_in_child(bool); + /* internal procfs utilities (from procfs.c) */ int crm_procfs_process_info(struct dirent *entry, char *name, int *pid); int crm_procfs_pid_of(const char *name); unsigned int crm_procfs_num_cores(void); /* internal XML schema functions (from xml.c) */ void crm_schema_init(void); void crm_schema_cleanup(void); /* internal functions related to process IDs (from pid.c) */ /*! * \internal * \brief Detect if process per PID and optionally exe path (component) exists * * \param[in] pid PID of process assumed alive, disproving of which to try * \param[in] daemon exe path (component) to possibly match with procfs entry * * \return -1 on invalid PID specification, -2 when the calling process has no * (is refused an) ability to (dis)prove the predicate, * 0 if the negation of the predicate is confirmed (check-through-kill * indicates so, or the subsequent check-through-procfs-match on * \p daemon when provided and procfs available at the standard path), * 1 if it cannot be disproved (reliably [modulo race conditions] * when \p daemon provided, procfs available at the standard path * and the calling process has permissions to access the respective * procfs location, less so otherwise, since mere check-through-kill * is exercised without powers to exclude PID recycled in the interim). * * \note This function cannot be used to verify \e authenticity of the process. */ int crm_pid_active(long pid, const char *daemon); long crm_pidfile_inuse(const char *filename, long mypid, const char *daemon); long crm_read_pidfile(const char *filename); int crm_lock_pidfile(const char *filename, const char *name); /* interal functions related to resource operations (from operations.c) */ char *generate_op_key(const char *rsc_id, const char *op_type, guint interval_ms); char *generate_notify_key(const char *rsc_id, const char *notify_type, const char *op_type); char *generate_transition_key(int action, int transition_id, int target_rc, const char *node); void filter_action_parameters(xmlNode *param_set, const char *version); // miscellaneous utilities (from utils.c) const char *pcmk_message_name(const char *name); /* internal generic string functions (from strings.c) */ long long crm_int_helper(const char *text, char **end_text); guint crm_parse_ms(const char *text); bool crm_starts_with(const char *str, const char *prefix); gboolean crm_ends_with(const char *s, const char *match); gboolean crm_ends_with_ext(const char *s, const char *match); char *add_list_element(char *list, const char *value); bool crm_compress_string(const char *data, int length, int max, char **result, unsigned int *result_len); gint crm_alpha_sort(gconstpointer a, gconstpointer b); static inline char * crm_concat(const char *prefix, const char *suffix, char join) { CRM_ASSERT(prefix && suffix); return crm_strdup_printf("%s%c%s", prefix, join, suffix); } static inline int crm_strlen_zero(const char *s) { return !s || *s == '\0'; } static inline char * crm_getpid_s() { return crm_strdup_printf("%lu", (unsigned long) getpid()); } /* convenience functions for failure-related node attributes */ #define CRM_FAIL_COUNT_PREFIX "fail-count" #define CRM_LAST_FAILURE_PREFIX "last-failure" /*! * \internal * \brief Generate a failure-related node attribute name for a resource * * \param[in] prefix Start of attribute name * \param[in] rsc_id Resource name * \param[in] op Operation name * \param[in] interval_ms Operation interval * * \return Newly allocated string with attribute name * * \note Failure attributes are named like PREFIX-RSC#OP_INTERVAL (for example, * "fail-count-myrsc#monitor_30000"). The '#' is used because it is not * a valid character in a resource ID, to reliably distinguish where the * operation name begins. The '_' is used simply to be more comparable to * action labels like "myrsc_monitor_30000". */ static inline char * crm_fail_attr_name(const char *prefix, const char *rsc_id, const char *op, guint interval_ms) { CRM_CHECK(prefix && rsc_id && op, return NULL); return crm_strdup_printf("%s-%s#%s_%u", prefix, rsc_id, op, interval_ms); } static inline char * crm_failcount_name(const char *rsc_id, const char *op, guint interval_ms) { return crm_fail_attr_name(CRM_FAIL_COUNT_PREFIX, rsc_id, op, interval_ms); } static inline char * crm_lastfailure_name(const char *rsc_id, const char *op, guint interval_ms) { return crm_fail_attr_name(CRM_LAST_FAILURE_PREFIX, rsc_id, op, interval_ms); } #endif /* CRM_COMMON_INTERNAL__H */ diff --git a/lib/common/io.c b/lib/common/io.c index fa438ddc61..6cbab0a932 100644 --- a/lib/common/io.c +++ b/lib/common/io.c @@ -1,503 +1,575 @@ /* * Copyright 2004-2019 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU Lesser General Public License * version 2.1 or later (LGPLv2.1+) WITHOUT ANY WARRANTY. */ #include #ifndef _GNU_SOURCE # define _GNU_SOURCE #endif #include #include #include +#include #include #include #include #include #include #include #include #include #include #include /*! * \brief Create a directory, including any parent directories needed * * \param[in] path_c Pathname of the directory to create * \param[in] mode Permissions to be used (with current umask) when creating * * \note This logs errors but does not return them to the caller. */ void crm_build_path(const char *path_c, mode_t mode) { int offset = 1, len = 0; char *path = strdup(path_c); CRM_CHECK(path != NULL, return); for (len = strlen(path); offset < len; offset++) { if (path[offset] == '/') { path[offset] = 0; if (mkdir(path, mode) < 0 && errno != EEXIST) { crm_perror(LOG_ERR, "Could not create directory '%s'", path); break; } path[offset] = '/'; } } if (mkdir(path, mode) < 0 && errno != EEXIST) { crm_perror(LOG_ERR, "Could not create directory '%s'", path); } free(path); } /*! * \internal * \brief Allocate and create a file path using a sequence number * * \param[in] directory Directory that contains the file series * \param[in] series Start of file name * \param[in] sequence Sequence number (MUST be less than 33 digits) * \param[in] bzip Whether to use ".bz2" instead of ".raw" as extension * * \return Newly allocated file path, or NULL on error * \note Caller is responsible for freeing the returned memory */ char * generate_series_filename(const char *directory, const char *series, int sequence, gboolean bzip) { const char *ext = "raw"; CRM_CHECK(directory != NULL, return NULL); CRM_CHECK(series != NULL, return NULL); if (bzip) { ext = "bz2"; } return crm_strdup_printf("%s/%s-%d.%s", directory, series, sequence, ext); } /*! * \internal * \brief Read and return sequence number stored in a file series' .last file * * \param[in] directory Directory that contains the file series * \param[in] series Start of file name * * \return The last sequence number, or 0 on error */ int get_last_sequence(const char *directory, const char *series) { FILE *file_strm = NULL; int start = 0, length = 0, read_len = 0; char *series_file = NULL; char *buffer = NULL; int seq = 0; CRM_CHECK(directory != NULL, return 0); CRM_CHECK(series != NULL, return 0); series_file = crm_strdup_printf("%s/%s.last", directory, series); file_strm = fopen(series_file, "r"); if (file_strm == NULL) { crm_debug("Series file %s does not exist", series_file); free(series_file); return 0; } /* see how big the file is */ start = ftell(file_strm); fseek(file_strm, 0L, SEEK_END); length = ftell(file_strm); fseek(file_strm, 0L, start); CRM_ASSERT(length >= 0); CRM_ASSERT(start == ftell(file_strm)); if (length <= 0) { crm_info("%s was not valid", series_file); free(buffer); buffer = NULL; } else { crm_trace("Reading %d bytes from file", length); buffer = calloc(1, (length + 1)); read_len = fread(buffer, 1, length, file_strm); if (read_len != length) { crm_err("Calculated and read bytes differ: %d vs. %d", length, read_len); free(buffer); buffer = NULL; } } seq = crm_parse_int(buffer, "0"); fclose(file_strm); crm_trace("Found %d in %s", seq, series_file); free(series_file); free(buffer); return seq; } /*! * \internal * \brief Write sequence number to a file series' .last file * * \param[in] directory Directory that contains the file series * \param[in] series Start of file name * \param[in] sequence Sequence number to write * \param[in] max Maximum sequence value, after which sequence is reset to 0 * * \note This function logs some errors but does not return any to the caller */ void write_last_sequence(const char *directory, const char *series, int sequence, int max) { int rc = 0; FILE *file_strm = NULL; char *series_file = NULL; CRM_CHECK(directory != NULL, return); CRM_CHECK(series != NULL, return); if (max == 0) { return; } if (max > 0 && sequence >= max) { sequence = 0; } series_file = crm_strdup_printf("%s/%s.last", directory, series); file_strm = fopen(series_file, "w"); if (file_strm != NULL) { rc = fprintf(file_strm, "%d", sequence); if (rc < 0) { crm_perror(LOG_ERR, "Cannot write to series file %s", series_file); } } else { crm_err("Cannot open series file %s for writing", series_file); } if (file_strm != NULL) { fflush(file_strm); fclose(file_strm); } crm_trace("Wrote %d to %s", sequence, series_file); free(series_file); } /*! * \internal * \brief Change the owner and group of a file series' .last file * * \param[in] dir Directory that contains series * \param[in] uid Uid of desired file owner * \param[in] gid Gid of desired file group * * \return 0 on success, -1 on error (in which case errno will be set) * \note The caller must have the appropriate privileges. */ int crm_chown_last_sequence(const char *directory, const char *series, uid_t uid, gid_t gid) { char *series_file = NULL; int rc; CRM_CHECK((directory != NULL) && (series != NULL), errno = EINVAL; return -1); series_file = crm_strdup_printf("%s/%s.last", directory, series); CRM_CHECK(series_file != NULL, return -1); rc = chown(series_file, uid, gid); free(series_file); return rc; } static bool pcmk__daemon_user_can_write(const char *target_name, struct stat *target_stat) { struct passwd *sys_user = NULL; errno = 0; sys_user = getpwnam(CRM_DAEMON_USER); if (sys_user == NULL) { crm_notice("Could not find user %s: %s", CRM_DAEMON_USER, pcmk_strerror(errno)); return FALSE; } if (target_stat->st_uid != sys_user->pw_uid) { crm_notice("%s is not owned by user %s " CRM_XS " uid %d != %d", target_name, CRM_DAEMON_USER, sys_user->pw_uid, target_stat->st_uid); return FALSE; } if ((target_stat->st_mode & (S_IRUSR | S_IWUSR)) == 0) { crm_notice("%s is not readable and writable by user %s " CRM_XS " st_mode=0%lo", target_name, CRM_DAEMON_USER, (unsigned long) target_stat->st_mode); return FALSE; } return TRUE; } static bool pcmk__daemon_group_can_write(const char *target_name, struct stat *target_stat) { struct group *sys_grp = NULL; errno = 0; sys_grp = getgrnam(CRM_DAEMON_GROUP); if (sys_grp == NULL) { crm_notice("Could not find group %s: %s", CRM_DAEMON_GROUP, pcmk_strerror(errno)); return FALSE; } if (target_stat->st_gid != sys_grp->gr_gid) { crm_notice("%s is not owned by group %s " CRM_XS " uid %d != %d", target_name, CRM_DAEMON_GROUP, sys_grp->gr_gid, target_stat->st_gid); return FALSE; } if ((target_stat->st_mode & (S_IRGRP | S_IWGRP)) == 0) { crm_notice("%s is not readable and writable by group %s " CRM_XS " st_mode=0%lo", target_name, CRM_DAEMON_GROUP, (unsigned long) target_stat->st_mode); return FALSE; } return TRUE; } /*! * \internal * \brief Check whether a directory or file is writable by the cluster daemon * * Return TRUE if either the cluster daemon user or cluster daemon group has * write permission on a specified file or directory. * * \param[in] dir Directory to check (this argument must be specified, and * the directory must exist) * \param[in] file File to check (only the directory will be checked if this * argument is not specified or the file does not exist) * * \return TRUE if target is writable by cluster daemon, FALSE otherwise */ bool pcmk__daemon_can_write(const char *dir, const char *file) { int s_res = 0; struct stat buf; char *full_file = NULL; const char *target = NULL; // Caller must supply directory CRM_ASSERT(dir != NULL); // If file is given, check whether it exists as a regular file if (file != NULL) { full_file = crm_concat(dir, file, '/'); target = full_file; s_res = stat(full_file, &buf); if (s_res < 0) { crm_notice("%s not found: %s", target, pcmk_strerror(errno)); free(full_file); full_file = NULL; target = NULL; } else if (S_ISREG(buf.st_mode) == FALSE) { crm_err("%s must be a regular file " CRM_XS " st_mode=0%lo", target, (unsigned long) buf.st_mode); free(full_file); return FALSE; } } // If file is not given, ensure dir exists as directory if (target == NULL) { target = dir; s_res = stat(dir, &buf); if (s_res < 0) { crm_err("%s not found: %s", dir, pcmk_strerror(errno)); return FALSE; } else if (S_ISDIR(buf.st_mode) == FALSE) { crm_err("%s must be a directory " CRM_XS " st_mode=0%lo", dir, (unsigned long) buf.st_mode); return FALSE; } } if (!pcmk__daemon_user_can_write(target, &buf) && !pcmk__daemon_group_can_write(target, &buf)) { crm_err("%s must be owned and writable by either user %s or group %s " CRM_XS " st_mode=0%lo", target, CRM_DAEMON_USER, CRM_DAEMON_GROUP, (unsigned long) buf.st_mode); free(full_file); return FALSE; } free(full_file); return TRUE; } /*! * \internal * \brief Flush and sync a directory to disk * * \param[in] name Directory to flush and sync * \note This function logs errors but does not return them to the caller */ void crm_sync_directory(const char *name) { int fd; DIR *directory; directory = opendir(name); if (directory == NULL) { crm_perror(LOG_ERR, "Could not open %s for syncing", name); return; } fd = dirfd(directory); if (fd < 0) { crm_perror(LOG_ERR, "Could not obtain file descriptor for %s", name); return; } if (fsync(fd) < 0) { crm_perror(LOG_ERR, "Could not sync %s", name); } if (closedir(directory) < 0) { crm_perror(LOG_ERR, "Could not close %s after fsync", name); } } /*! * \internal * \brief Allocate, read and return the contents of a file * * \param[in] filename Name of file to read * * \return Newly allocated memory with contents of file, or NULL on error * \note On success, the caller is responsible for freeing the returned memory; * on error, errno will be 0 (indicating file was nonexistent or empty) * or one of the errno values set by fopen, ftell, or calloc */ char * crm_read_contents(const char *filename) { char *contents = NULL; FILE *fp; int length, read_len; errno = 0; /* enable caller to distinguish error from empty file */ fp = fopen(filename, "r"); if (fp == NULL) { return NULL; } fseek(fp, 0L, SEEK_END); length = ftell(fp); if (length > 0) { contents = calloc(length + 1, sizeof(char)); if (contents == NULL) { fclose(fp); return NULL; } crm_trace("Reading %d bytes from %s", length, filename); rewind(fp); read_len = fread(contents, 1, length, fp); /* Coverity: False positive */ if (read_len != length) { free(contents); contents = NULL; } } fclose(fp); return contents; } /*! * \internal * \brief Write text to a file, flush and sync it to disk, then close the file * * \param[in] fd File descriptor opened for writing * \param[in] contents String to write to file * * \return 0 on success, -1 on error (in which case errno will be set) */ int crm_write_sync(int fd, const char *contents) { int rc = 0; FILE *fp = fdopen(fd, "w"); if (fp == NULL) { return -1; } if ((contents != NULL) && (fprintf(fp, "%s", contents) < 0)) { rc = -1; } if (fflush(fp) != 0) { rc = -1; } if (fsync(fileno(fp)) < 0) { rc = -1; } fclose(fp); return rc; } /*! * \internal * \brief Set a file descriptor to non-blocking * * \param[in] fd File descriptor to use * * \return pcmk_ok on success, -errno on error */ int crm_set_nonblocking(int fd) { int flag = fcntl(fd, F_GETFL); if (flag < 0) { return -errno; } if (fcntl(fd, F_SETFL, flag | O_NONBLOCK) < 0) { return -errno; } return pcmk_ok; } const char * crm_get_tmpdir() { const char *dir = getenv("TMPDIR"); return (dir && (*dir == '/'))? dir : "/tmp"; } + +/*! + * \internal + * \brief Close open file descriptors + * + * Close all file descriptors (except optionally stdin, stdout, and stderr), + * which is a best practice for a new child process forked for the purpose of + * executing an external program. + * + * \param[in] bool If true, close stdin, stdout, and stderr as well + */ +void +pcmk__close_fds_in_child(bool all) +{ + DIR *dir; + struct rlimit rlim; + rlim_t max_fd; + int min_fd = (all? 0 : (STDERR_FILENO + 1)); + + /* Find the current process's (soft) limit for open files. getrlimit() + * should always work, but have a fallback just in case. + */ + if (getrlimit(RLIMIT_NOFILE, &rlim) == 0) { + max_fd = rlim.rlim_cur - 1; + } else { + long conf_max = sysconf(_SC_OPEN_MAX); + + max_fd = (conf_max > 0)? conf_max : 1024; + } + + /* /proc/self/fd (on Linux) or /dev/fd (on most OSes) contains symlinks to + * all open files for the current process, named as the file descriptor. + * Use this if available, because it's more efficient than a shotgun + * approach to closing descriptors. + */ +#if SUPPORT_PROCFS + dir = opendir("/proc/self/fd"); + if (dir == NULL) { + dir = opendir("/dev/fd"); + } +#else + dir = opendir("/dev/fd"); +#endif + if (dir != NULL) { + struct dirent *entry; + int dir_fd = dirfd(dir); + + while ((entry = readdir(dir)) != NULL) { + int lpc = atoi(entry->d_name); + + /* How could one of these entries be higher than max_fd, you ask? + * It isn't possible in normal operation, but when run under + * valgrind, valgrind can open high-numbered file descriptors for + * its own use that are higher than the process's soft limit. + * These will show up in the fd directory but aren't closable. + */ + if ((lpc >= min_fd) && (lpc <= max_fd) && (lpc != dir_fd)) { + close(lpc); + } + } + closedir(dir); + return; + } + + /* If no fd directory is available, iterate over all possible descriptors. + * This is less efficient due to the overhead of many system calls. + */ + for (int lpc = max_fd; lpc >= min_fd; lpc--) { + close(lpc); + } +} diff --git a/lib/services/services_linux.c b/lib/services/services_linux.c index 464fc5b170..6870273da4 100644 --- a/lib/services/services_linux.c +++ b/lib/services/services_linux.c @@ -1,1126 +1,1099 @@ /* * Copyright 2010-2019 Andrew Beekhof * * This source code is licensed under the GNU Lesser General Public License * version 2.1 or later (LGPLv2.1+) WITHOUT ANY WARRANTY. */ #include #ifndef _GNU_SOURCE # define _GNU_SOURCE #endif #include #include #include #include #include #include #include #include #include #include #ifdef HAVE_SYS_SIGNALFD_H #include #endif #include "crm/crm.h" #include "crm/common/mainloop.h" #include "crm/services.h" #include "services_private.h" #if SUPPORT_CIBSECRETS # include "crm/common/cib_secrets.h" #endif static gboolean svc_read_output(int fd, svc_action_t * op, bool is_stderr) { char *data = NULL; int rc = 0, len = 0; char buf[500]; static const size_t buf_read_len = sizeof(buf) - 1; if (fd < 0) { crm_trace("No fd for %s", op->id); return FALSE; } if (is_stderr && op->stderr_data) { len = strlen(op->stderr_data); data = op->stderr_data; crm_trace("Reading %s stderr into offset %d", op->id, len); } else if (is_stderr == FALSE && op->stdout_data) { len = strlen(op->stdout_data); data = op->stdout_data; crm_trace("Reading %s stdout into offset %d", op->id, len); } else { crm_trace("Reading %s %s into offset %d", op->id, is_stderr?"stderr":"stdout", len); } do { rc = read(fd, buf, buf_read_len); if (rc > 0) { buf[rc] = 0; crm_trace("Got %d chars: %.80s", rc, buf); data = realloc_safe(data, len + rc + 1); len += sprintf(data + len, "%s", buf); } else if (errno != EINTR) { /* error or EOF * Cleanup happens in pipe_done() */ rc = FALSE; break; } } while (rc == buf_read_len || rc < 0); if (is_stderr) { op->stderr_data = data; } else { op->stdout_data = data; } return rc; } static int dispatch_stdout(gpointer userdata) { svc_action_t *op = (svc_action_t *) userdata; return svc_read_output(op->opaque->stdout_fd, op, FALSE); } static int dispatch_stderr(gpointer userdata) { svc_action_t *op = (svc_action_t *) userdata; return svc_read_output(op->opaque->stderr_fd, op, TRUE); } static void pipe_out_done(gpointer user_data) { svc_action_t *op = (svc_action_t *) user_data; crm_trace("%p", op); op->opaque->stdout_gsource = NULL; if (op->opaque->stdout_fd > STDOUT_FILENO) { close(op->opaque->stdout_fd); } op->opaque->stdout_fd = -1; } static void pipe_err_done(gpointer user_data) { svc_action_t *op = (svc_action_t *) user_data; op->opaque->stderr_gsource = NULL; if (op->opaque->stderr_fd > STDERR_FILENO) { close(op->opaque->stderr_fd); } op->opaque->stderr_fd = -1; } static struct mainloop_fd_callbacks stdout_callbacks = { .dispatch = dispatch_stdout, .destroy = pipe_out_done, }; static struct mainloop_fd_callbacks stderr_callbacks = { .dispatch = dispatch_stderr, .destroy = pipe_err_done, }; static void set_ocf_env(const char *key, const char *value, gpointer user_data) { if (setenv(key, value, 1) != 0) { crm_perror(LOG_ERR, "setenv failed for key:%s and value:%s", key, value); } } static void set_ocf_env_with_prefix(gpointer key, gpointer value, gpointer user_data) { char buffer[500]; snprintf(buffer, sizeof(buffer), "OCF_RESKEY_%s", (char *)key); set_ocf_env(buffer, value, user_data); } static void set_alert_env(gpointer key, gpointer value, gpointer user_data) { int rc; if (value != NULL) { rc = setenv(key, value, 1); } else { rc = unsetenv(key); } if (rc < 0) { crm_perror(LOG_ERR, "setenv %s=%s", (char*)key, (value? (char*)value : "")); } else { crm_trace("setenv %s=%s", (char*)key, (value? (char*)value : "")); } } /*! * \internal * \brief Add environment variables suitable for an action * * \param[in] op Action to use */ static void add_action_env_vars(const svc_action_t *op) { void (*env_setter)(gpointer, gpointer, gpointer) = NULL; if (op->agent == NULL) { env_setter = set_alert_env; /* we deal with alert handler */ } else if (safe_str_eq(op->standard, PCMK_RESOURCE_CLASS_OCF)) { env_setter = set_ocf_env_with_prefix; } if (env_setter != NULL && op->params != NULL) { g_hash_table_foreach(op->params, env_setter, NULL); } if (env_setter == NULL || env_setter == set_alert_env) { return; } set_ocf_env("OCF_RA_VERSION_MAJOR", "1", NULL); set_ocf_env("OCF_RA_VERSION_MINOR", "0", NULL); set_ocf_env("OCF_ROOT", OCF_ROOT_DIR, NULL); set_ocf_env("OCF_EXIT_REASON_PREFIX", PCMK_OCF_REASON_PREFIX, NULL); if (op->rsc) { set_ocf_env("OCF_RESOURCE_INSTANCE", op->rsc, NULL); } if (op->agent != NULL) { set_ocf_env("OCF_RESOURCE_TYPE", op->agent, NULL); } /* Notes: this is not added to specification yet. Sept 10,2004 */ if (op->provider != NULL) { set_ocf_env("OCF_RESOURCE_PROVIDER", op->provider, NULL); } } static void pipe_in_single_parameter(gpointer key, gpointer value, gpointer user_data) { svc_action_t *op = user_data; char *buffer = crm_strdup_printf("%s=%s\n", (char *)key, (char *) value); int ret, total = 0, len = strlen(buffer); do { errno = 0; ret = write(op->opaque->stdin_fd, buffer + total, len - total); if (ret > 0) { total += ret; } } while ((errno == EINTR) && (total < len)); free(buffer); } /*! * \internal * \brief Pipe parameters in via stdin for action * * \param[in] op Action to use */ static void pipe_in_action_stdin_parameters(const svc_action_t *op) { crm_debug("sending args"); if (op->params) { g_hash_table_foreach(op->params, pipe_in_single_parameter, (gpointer) op); } } gboolean recurring_action_timer(gpointer data) { svc_action_t *op = data; crm_debug("Scheduling another invocation of %s", op->id); /* Clean out the old result */ free(op->stdout_data); op->stdout_data = NULL; free(op->stderr_data); op->stderr_data = NULL; op->opaque->repeat_timer = 0; services_action_async(op, NULL); return FALSE; } /* Returns FALSE if 'op' should be free'd by the caller */ gboolean operation_finalize(svc_action_t * op) { int recurring = 0; if (op->interval_ms) { if (op->cancel) { op->status = PCMK_LRM_OP_CANCELLED; cancel_recurring_action(op); } else { recurring = 1; op->opaque->repeat_timer = g_timeout_add(op->interval_ms, recurring_action_timer, (void *)op); } } if (op->opaque->callback) { op->opaque->callback(op); } op->pid = 0; services_untrack_op(op); if (!recurring && op->synchronous == FALSE) { /* * If this is a recurring action, do not free explicitly. * It will get freed whenever the action gets cancelled. */ services_action_free(op); return TRUE; } services_action_cleanup(op); return FALSE; } static void operation_finished(mainloop_child_t * p, pid_t pid, int core, int signo, int exitcode) { svc_action_t *op = mainloop_child_userdata(p); char *prefix = crm_strdup_printf("%s:%d", op->id, op->pid); mainloop_clear_child_userdata(p); op->status = PCMK_LRM_OP_DONE; CRM_ASSERT(op->pid == pid); crm_trace("%s %p %p", prefix, op->opaque->stderr_gsource, op->opaque->stdout_gsource); if (op->opaque->stderr_gsource) { /* Make sure we have read everything from the buffer. * Depending on the priority mainloop gives the fd, operation_finished * could occur before all the reads are done. Force the read now.*/ crm_trace("%s dispatching stderr", prefix); dispatch_stderr(op); crm_trace("%s: %p", op->id, op->stderr_data); mainloop_del_fd(op->opaque->stderr_gsource); op->opaque->stderr_gsource = NULL; } if (op->opaque->stdout_gsource) { /* Make sure we have read everything from the buffer. * Depending on the priority mainloop gives the fd, operation_finished * could occur before all the reads are done. Force the read now.*/ crm_trace("%s dispatching stdout", prefix); dispatch_stdout(op); crm_trace("%s: %p", op->id, op->stdout_data); mainloop_del_fd(op->opaque->stdout_gsource); op->opaque->stdout_gsource = NULL; } if (op->opaque->stdin_fd >= 0) { close(op->opaque->stdin_fd); } if (signo) { if (mainloop_child_timeout(p)) { crm_warn("%s - timed out after %dms", prefix, op->timeout); op->status = PCMK_LRM_OP_TIMEOUT; op->rc = PCMK_OCF_TIMEOUT; } else if (op->cancel) { /* If an in-flight recurring operation was killed because it was * cancelled, don't treat that as a failure. */ crm_info("%s - terminated with signal %d", prefix, signo); op->status = PCMK_LRM_OP_CANCELLED; op->rc = PCMK_OCF_OK; } else { crm_warn("%s - terminated with signal %d", prefix, signo); op->status = PCMK_LRM_OP_ERROR; op->rc = PCMK_OCF_SIGNAL; } } else { op->rc = exitcode; crm_debug("%s - exited with rc=%d", prefix, exitcode); } free(prefix); prefix = crm_strdup_printf("%s:%d:stderr", op->id, op->pid); crm_log_output(LOG_NOTICE, prefix, op->stderr_data); free(prefix); prefix = crm_strdup_printf("%s:%d:stdout", op->id, op->pid); crm_log_output(LOG_DEBUG, prefix, op->stdout_data); free(prefix); operation_finalize(op); } /*! * \internal * \brief Set operation rc and status per errno from stat(), fork() or execvp() * * \param[in,out] op Operation to set rc and status for * \param[in] error Value of errno after system call * * \return void */ static void services_handle_exec_error(svc_action_t * op, int error) { int rc_not_installed, rc_insufficient_priv, rc_exec_error; /* Mimic the return codes for each standard as that's what we'll convert back from in get_uniform_rc() */ if (safe_str_eq(op->standard, PCMK_RESOURCE_CLASS_LSB) && safe_str_eq(op->action, "status")) { rc_not_installed = PCMK_LSB_STATUS_NOT_INSTALLED; rc_insufficient_priv = PCMK_LSB_STATUS_INSUFFICIENT_PRIV; rc_exec_error = PCMK_LSB_STATUS_UNKNOWN; #if SUPPORT_NAGIOS } else if (safe_str_eq(op->standard, PCMK_RESOURCE_CLASS_NAGIOS)) { rc_not_installed = NAGIOS_NOT_INSTALLED; rc_insufficient_priv = NAGIOS_INSUFFICIENT_PRIV; rc_exec_error = PCMK_OCF_EXEC_ERROR; #endif } else { rc_not_installed = PCMK_OCF_NOT_INSTALLED; rc_insufficient_priv = PCMK_OCF_INSUFFICIENT_PRIV; rc_exec_error = PCMK_OCF_EXEC_ERROR; } switch (error) { /* see execve(2), stat(2) and fork(2) */ case ENOENT: /* No such file or directory */ case EISDIR: /* Is a directory */ case ENOTDIR: /* Path component is not a directory */ case EINVAL: /* Invalid executable format */ case ENOEXEC: /* Invalid executable format */ op->rc = rc_not_installed; op->status = PCMK_LRM_OP_NOT_INSTALLED; break; case EACCES: /* permission denied (various errors) */ case EPERM: /* permission denied (various errors) */ op->rc = rc_insufficient_priv; op->status = PCMK_LRM_OP_ERROR; break; default: op->rc = rc_exec_error; op->status = PCMK_LRM_OP_ERROR; } } static void action_launch_child(svc_action_t *op) { - int lpc; - DIR *dir; - /* SIGPIPE is ignored (which is different from signal blocking) by the gnutls library. * Depending on the libqb version in use, libqb may set SIGPIPE to be ignored as well. * We do not want this to be inherited by the child process. By resetting this the signal * to the default behavior, we avoid some potential odd problems that occur during OCF * scripts when SIGPIPE is ignored by the environment. */ signal(SIGPIPE, SIG_DFL); #if defined(HAVE_SCHED_SETSCHEDULER) if (sched_getscheduler(0) != SCHED_OTHER) { struct sched_param sp; memset(&sp, 0, sizeof(sp)); sp.sched_priority = 0; if (sched_setscheduler(0, SCHED_OTHER, &sp) == -1) { crm_perror(LOG_ERR, "Could not reset scheduling policy to SCHED_OTHER for %s", op->id); } } #endif if (setpriority(PRIO_PROCESS, 0, 0) == -1) { crm_perror(LOG_ERR, "Could not reset process priority to 0 for %s", op->id); } /* Man: The call setpgrp() is equivalent to setpgid(0,0) * _and_ compiles on BSD variants too * need to investigate if it works the same too. */ setpgid(0, 0); - // Close all file descriptors except stdin/stdout/stderr -#if SUPPORT_PROCFS - dir = opendir("/proc/self/fd"); -#else - dir = opendir("/dev/fd"); -#endif - if (dir == NULL) { /* /proc or /dev/fd not available */ - /* Iterate over all possible fds, might be slow */ - for (lpc = getdtablesize() - 1; lpc > STDERR_FILENO; lpc--) { - close(lpc); - } - } else { - /* Iterate over fds obtained from /proc or /dev/fd */ - struct dirent *entry; - int dir_fd = dirfd(dir); - - while ((entry = readdir(dir)) != NULL) { - lpc = atoi(entry->d_name); - if (lpc > STDERR_FILENO && lpc != dir_fd) { - close(lpc); - } - } - - closedir(dir); - } + pcmk__close_fds_in_child(false); #if SUPPORT_CIBSECRETS if (replace_secret_params(op->rsc, op->params) < 0) { /* replacing secrets failed! */ if (safe_str_eq(op->action,"stop")) { /* don't fail on stop! */ crm_info("proceeding with the stop operation for %s", op->rsc); } else { crm_err("failed to get secrets for %s, " "considering resource not configured", op->rsc); _exit(PCMK_OCF_NOT_CONFIGURED); } } #endif add_action_env_vars(op); /* Become the desired user */ if (op->opaque->uid && (geteuid() == 0)) { // If requested, set effective group if (op->opaque->gid && (setgid(op->opaque->gid) < 0)) { crm_perror(LOG_ERR, "Could not set child group to %d", op->opaque->gid); _exit(PCMK_OCF_NOT_CONFIGURED); } // Erase supplementary group list // (We could do initgroups() if we kept a copy of the username) if (setgroups(0, NULL) < 0) { crm_perror(LOG_ERR, "Could not set child groups"); _exit(PCMK_OCF_NOT_CONFIGURED); } // Set effective user if (setuid(op->opaque->uid) < 0) { crm_perror(LOG_ERR, "setting user to %d", op->opaque->uid); _exit(PCMK_OCF_NOT_CONFIGURED); } } /* execute the RA */ execvp(op->opaque->exec, op->opaque->args); /* Most cases should have been already handled by stat() */ services_handle_exec_error(op, errno); _exit(op->rc); } #ifndef HAVE_SYS_SIGNALFD_H static int sigchld_pipe[2] = { -1, -1 }; static void sigchld_handler() { if ((sigchld_pipe[1] >= 0) && (write(sigchld_pipe[1], "", 1) == -1)) { crm_perror(LOG_TRACE, "Could not poke SIGCHLD self-pipe"); } } #endif static void action_synced_wait(svc_action_t * op, sigset_t *mask) { int status = 0; int timeout = op->timeout; int sfd = -1; time_t start = -1; struct pollfd fds[3]; int wait_rc = 0; #ifdef HAVE_SYS_SIGNALFD_H sfd = signalfd(-1, mask, SFD_NONBLOCK); if (sfd < 0) { crm_perror(LOG_ERR, "signalfd() failed"); } #else sfd = sigchld_pipe[0]; #endif fds[0].fd = op->opaque->stdout_fd; fds[0].events = POLLIN; fds[0].revents = 0; fds[1].fd = op->opaque->stderr_fd; fds[1].events = POLLIN; fds[1].revents = 0; fds[2].fd = sfd; fds[2].events = POLLIN; fds[2].revents = 0; crm_trace("Waiting for %d", op->pid); start = time(NULL); do { int poll_rc = poll(fds, 3, timeout); if (poll_rc > 0) { if (fds[0].revents & POLLIN) { svc_read_output(op->opaque->stdout_fd, op, FALSE); } if (fds[1].revents & POLLIN) { svc_read_output(op->opaque->stderr_fd, op, TRUE); } if (fds[2].revents & POLLIN) { #ifdef HAVE_SYS_SIGNALFD_H struct signalfd_siginfo fdsi; ssize_t s; s = read(sfd, &fdsi, sizeof(struct signalfd_siginfo)); if (s != sizeof(struct signalfd_siginfo)) { crm_perror(LOG_ERR, "Read from signal fd %d failed", sfd); } else if (fdsi.ssi_signo == SIGCHLD) { #else if (1) { /* Clear out the sigchld pipe. */ char ch; while (read(sfd, &ch, 1) == 1) /*omit*/; #endif wait_rc = waitpid(op->pid, &status, WNOHANG); if (wait_rc > 0) { break; } else if (wait_rc < 0){ if (errno == ECHILD) { /* Here, don't dare to kill and bail out... */ break; } else { /* ...otherwise pretend process still runs. */ wait_rc = 0; } crm_perror(LOG_ERR, "waitpid() for %d failed", op->pid); } } } } else if (poll_rc == 0) { timeout = 0; break; } else if (poll_rc < 0) { if (errno != EINTR) { crm_perror(LOG_ERR, "poll() failed"); break; } } timeout = op->timeout - (time(NULL) - start) * 1000; } while ((op->timeout < 0 || timeout > 0)); crm_trace("Child done: %d", op->pid); if (wait_rc <= 0) { op->rc = PCMK_OCF_UNKNOWN_ERROR; if (op->timeout > 0 && timeout <= 0) { op->status = PCMK_LRM_OP_TIMEOUT; crm_warn("%s:%d - timed out after %dms", op->id, op->pid, op->timeout); } else { op->status = PCMK_LRM_OP_ERROR; } /* If only child hasn't been successfully waited for, yet. This is to limit killing wrong target a bit more. */ if (wait_rc == 0 && waitpid(op->pid, &status, WNOHANG) == 0) { if (kill(op->pid, SIGKILL)) { crm_err("kill(%d, KILL) failed: %d", op->pid, errno); } /* Safe to skip WNOHANG here as we sent non-ignorable signal. */ while (waitpid(op->pid, &status, 0) == (pid_t) -1 && errno == EINTR) /*omit*/; } } else if (WIFEXITED(status)) { op->status = PCMK_LRM_OP_DONE; op->rc = WEXITSTATUS(status); crm_info("Managed %s process %d exited with rc=%d", op->id, op->pid, op->rc); } else if (WIFSIGNALED(status)) { int signo = WTERMSIG(status); op->status = PCMK_LRM_OP_ERROR; crm_err("Managed %s process %d exited with signal=%d", op->id, op->pid, signo); } #ifdef WCOREDUMP if (WCOREDUMP(status)) { crm_err("Managed %s process %d dumped core", op->id, op->pid); } #endif svc_read_output(op->opaque->stdout_fd, op, FALSE); svc_read_output(op->opaque->stderr_fd, op, TRUE); close(op->opaque->stdout_fd); close(op->opaque->stderr_fd); if (op->opaque->stdin_fd >= 0) { close(op->opaque->stdin_fd); } #ifdef HAVE_SYS_SIGNALFD_H close(sfd); #endif } /* For an asynchronous 'op', returns FALSE if 'op' should be free'd by the caller */ /* For a synchronous 'op', returns FALSE if 'op' fails */ gboolean services_os_action_execute(svc_action_t * op) { int stdout_fd[2]; int stderr_fd[2]; int stdin_fd[2] = {-1, -1}; int rc; struct stat st; sigset_t *pmask = NULL; #ifdef HAVE_SYS_SIGNALFD_H sigset_t mask; sigset_t old_mask; #define sigchld_cleanup() do { \ if (sigismember(&old_mask, SIGCHLD) == 0) { \ if (sigprocmask(SIG_UNBLOCK, &mask, NULL) < 0) { \ crm_perror(LOG_ERR, "sigprocmask() failed to unblock sigchld"); \ } \ } \ } while (0) #else struct sigaction sa; struct sigaction old_sa; #define sigchld_cleanup() do { \ if (sigaction(SIGCHLD, &old_sa, NULL) < 0) { \ crm_perror(LOG_ERR, "sigaction() failed to remove sigchld handler"); \ } \ close(sigchld_pipe[0]); \ close(sigchld_pipe[1]); \ sigchld_pipe[0] = sigchld_pipe[1] = -1; \ } while(0) #endif /* Fail fast */ if(stat(op->opaque->exec, &st) != 0) { rc = errno; crm_warn("Cannot execute '%s': %s (%d)", op->opaque->exec, pcmk_strerror(rc), rc); services_handle_exec_error(op, rc); if (!op->synchronous) { return operation_finalize(op); } return FALSE; } if (pipe(stdout_fd) < 0) { rc = errno; crm_err("pipe(stdout_fd) failed. '%s': %s (%d)", op->opaque->exec, pcmk_strerror(rc), rc); services_handle_exec_error(op, rc); if (!op->synchronous) { return operation_finalize(op); } return FALSE; } if (pipe(stderr_fd) < 0) { rc = errno; close(stdout_fd[0]); close(stdout_fd[1]); crm_err("pipe(stderr_fd) failed. '%s': %s (%d)", op->opaque->exec, pcmk_strerror(rc), rc); services_handle_exec_error(op, rc); if (!op->synchronous) { return operation_finalize(op); } return FALSE; } if (safe_str_eq(op->standard, PCMK_RESOURCE_CLASS_STONITH)) { if (pipe(stdin_fd) < 0) { rc = errno; close(stdout_fd[0]); close(stdout_fd[1]); close(stderr_fd[0]); close(stderr_fd[1]); crm_err("pipe(stdin_fd) failed. '%s': %s (%d)", op->opaque->exec, pcmk_strerror(rc), rc); services_handle_exec_error(op, rc); if (!op->synchronous) { return operation_finalize(op); } return FALSE; } } if (op->synchronous) { #ifdef HAVE_SYS_SIGNALFD_H sigemptyset(&mask); sigaddset(&mask, SIGCHLD); sigemptyset(&old_mask); if (sigprocmask(SIG_BLOCK, &mask, &old_mask) < 0) { crm_perror(LOG_ERR, "sigprocmask() failed to block sigchld"); } pmask = &mask; #else if(pipe(sigchld_pipe) == -1) { crm_perror(LOG_ERR, "pipe() failed"); } rc = crm_set_nonblocking(sigchld_pipe[0]); if (rc < 0) { crm_warn("Could not set pipe input non-blocking: %s " CRM_XS " rc=%d", pcmk_strerror(rc), rc); } rc = crm_set_nonblocking(sigchld_pipe[1]); if (rc < 0) { crm_warn("Could not set pipe output non-blocking: %s " CRM_XS " rc=%d", pcmk_strerror(rc), rc); } sa.sa_handler = sigchld_handler; sa.sa_flags = 0; sigemptyset(&sa.sa_mask); if (sigaction(SIGCHLD, &sa, &old_sa) < 0) { crm_perror(LOG_ERR, "sigaction() failed to set sigchld handler"); } pmask = NULL; #endif } op->pid = fork(); switch (op->pid) { case -1: rc = errno; close(stdout_fd[0]); close(stdout_fd[1]); close(stderr_fd[0]); close(stderr_fd[1]); if (stdin_fd[0] >= 0) { close(stdin_fd[0]); close(stdin_fd[1]); } crm_err("Could not execute '%s': %s (%d)", op->opaque->exec, pcmk_strerror(rc), rc); services_handle_exec_error(op, rc); if (!op->synchronous) { return operation_finalize(op); } sigchld_cleanup(); return FALSE; case 0: /* Child */ close(stdout_fd[0]); close(stderr_fd[0]); if (stdin_fd[1] >= 0) { close(stdin_fd[1]); } if (STDOUT_FILENO != stdout_fd[1]) { if (dup2(stdout_fd[1], STDOUT_FILENO) != STDOUT_FILENO) { crm_err("dup2() failed (stdout)"); } close(stdout_fd[1]); } if (STDERR_FILENO != stderr_fd[1]) { if (dup2(stderr_fd[1], STDERR_FILENO) != STDERR_FILENO) { crm_err("dup2() failed (stderr)"); } close(stderr_fd[1]); } if ((stdin_fd[0] >= 0) && (STDIN_FILENO != stdin_fd[0])) { if (dup2(stdin_fd[0], STDIN_FILENO) != STDIN_FILENO) { crm_err("dup2() failed (stdin)"); } close(stdin_fd[0]); } if (op->synchronous) { sigchld_cleanup(); } action_launch_child(op); CRM_ASSERT(0); /* action_launch_child is effectively noreturn */ } /* Only the parent reaches here */ close(stdout_fd[1]); close(stderr_fd[1]); if (stdin_fd[0] >= 0) { close(stdin_fd[0]); } op->opaque->stdout_fd = stdout_fd[0]; rc = crm_set_nonblocking(op->opaque->stdout_fd); if (rc < 0) { crm_warn("Could not set child output non-blocking: %s " CRM_XS " rc=%d", pcmk_strerror(rc), rc); } op->opaque->stderr_fd = stderr_fd[0]; rc = crm_set_nonblocking(op->opaque->stderr_fd); if (rc < 0) { crm_warn("Could not set child error output non-blocking: %s " CRM_XS " rc=%d", pcmk_strerror(rc), rc); } op->opaque->stdin_fd = stdin_fd[1]; if (op->opaque->stdin_fd >= 0) { // using buffer behind non-blocking-fd here - that could be improved // as long as no other standard uses stdin_fd assume stonith rc = crm_set_nonblocking(op->opaque->stdin_fd); if (rc < 0) { crm_warn("Could not set child input non-blocking: %s " CRM_XS " fd=%d,rc=%d", pcmk_strerror(rc), op->opaque->stdin_fd, rc); } pipe_in_action_stdin_parameters(op); // as long as we are handling parameters directly in here just close close(op->opaque->stdin_fd); op->opaque->stdin_fd = -1; } // after fds are setup properly and before we plug anything into mainloop if (op->opaque->fork_callback) { op->opaque->fork_callback(op); } if (op->synchronous) { action_synced_wait(op, pmask); sigchld_cleanup(); } else { crm_trace("Async waiting for %d - %s", op->pid, op->opaque->exec); mainloop_child_add_with_flags(op->pid, op->timeout, op->id, op, (op->flags & SVC_ACTION_LEAVE_GROUP) ? mainloop_leave_pid_group : 0, operation_finished); op->opaque->stdout_gsource = mainloop_add_fd(op->id, G_PRIORITY_LOW, op->opaque->stdout_fd, op, &stdout_callbacks); op->opaque->stderr_gsource = mainloop_add_fd(op->id, G_PRIORITY_LOW, op->opaque->stderr_fd, op, &stderr_callbacks); services_add_inflight_op(op); } return TRUE; } GList * services_os_get_directory_list(const char *root, gboolean files, gboolean executable) { GList *list = NULL; struct dirent **namelist; int entries = 0, lpc = 0; char buffer[PATH_MAX]; entries = scandir(root, &namelist, NULL, alphasort); if (entries <= 0) { return list; } for (lpc = 0; lpc < entries; lpc++) { struct stat sb; if ('.' == namelist[lpc]->d_name[0]) { free(namelist[lpc]); continue; } snprintf(buffer, sizeof(buffer), "%s/%s", root, namelist[lpc]->d_name); if (stat(buffer, &sb)) { continue; } if (S_ISDIR(sb.st_mode)) { if (files) { free(namelist[lpc]); continue; } } else if (S_ISREG(sb.st_mode)) { if (files == FALSE) { free(namelist[lpc]); continue; } else if (executable && (sb.st_mode & S_IXUSR) == 0 && (sb.st_mode & S_IXGRP) == 0 && (sb.st_mode & S_IXOTH) == 0) { free(namelist[lpc]); continue; } } list = g_list_append(list, strdup(namelist[lpc]->d_name)); free(namelist[lpc]); } free(namelist); return list; } GList * resources_os_list_ocf_providers(void) { return get_directory_list(OCF_ROOT_DIR "/resource.d", FALSE, TRUE); } GList * resources_os_list_ocf_agents(const char *provider) { GList *gIter = NULL; GList *result = NULL; GList *providers = NULL; if (provider) { char buffer[500]; snprintf(buffer, sizeof(buffer), "%s/resource.d/%s", OCF_ROOT_DIR, provider); return get_directory_list(buffer, TRUE, TRUE); } providers = resources_os_list_ocf_providers(); for (gIter = providers; gIter != NULL; gIter = gIter->next) { GList *tmp1 = result; GList *tmp2 = resources_os_list_ocf_agents(gIter->data); if (tmp2) { result = g_list_concat(tmp1, tmp2); } } g_list_free_full(providers, free); return result; } gboolean services__ocf_agent_exists(const char *provider, const char *agent) { char *buf = NULL; gboolean rc = FALSE; struct stat st; if (provider == NULL || agent == NULL) { return rc; } buf = crm_strdup_printf(OCF_ROOT_DIR "/resource.d/%s/%s", provider, agent); if (stat(buf, &st) == 0) { rc = TRUE; } free(buf); return rc; } #if SUPPORT_NAGIOS GList * resources_os_list_nagios_agents(void) { GList *plugin_list = NULL; GList *result = NULL; GList *gIter = NULL; plugin_list = get_directory_list(NAGIOS_PLUGIN_DIR, TRUE, TRUE); /* Make sure both the plugin and its metadata exist */ for (gIter = plugin_list; gIter != NULL; gIter = gIter->next) { const char *plugin = gIter->data; char *metadata = crm_strdup_printf(NAGIOS_METADATA_DIR "/%s.xml", plugin); struct stat st; if (stat(metadata, &st) == 0) { result = g_list_append(result, strdup(plugin)); } free(metadata); } g_list_free_full(plugin_list, free); return result; } gboolean services__nagios_agent_exists(const char *name) { char *buf = NULL; gboolean rc = FALSE; struct stat st; if (name == NULL) { return rc; } buf = crm_strdup_printf(NAGIOS_PLUGIN_DIR "/%s", name); if (stat(buf, &st) == 0) { rc = TRUE; } free(buf); return rc; } #endif