X-Git-Url: https://git.lttng.org/?p=lttng-tools.git;a=blobdiff_plain;f=src%2Fbin%2Flttng-sessiond%2Fhealth.c;h=9ff6bd1ae1013d9b86e5e178f920ff7286e63114;hp=fafc9095df57e89d88baeede25e4adfa8b634fe9;hb=1f4962443;hpb=67e056446003680180ed4f384217facd90b1a776 diff --git a/src/bin/lttng-sessiond/health.c b/src/bin/lttng-sessiond/health.c index fafc9095d..9ff6bd1ae 100644 --- a/src/bin/lttng-sessiond/health.c +++ b/src/bin/lttng-sessiond/health.c @@ -1,5 +1,6 @@ /* * Copyright (C) 2012 - David Goulet + * Copyright (C) 2018 - Jérémie Galarneau * * This program is free software; you can redistribute it and/or modify it * under the terms of the GNU General Public License, version 2 only, as @@ -15,259 +16,279 @@ * Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA. */ -#define _GNU_SOURCE -#include -#include -#include -#include -#include - -#include -#include +#include "lttng-sessiond.h" +#include "health-sessiond.h" #include -#include - -#include "health.h" - -static struct timespec time_delta = { - .tv_sec = DEFAULT_HEALTH_CHECK_DELTA_S, - .tv_nsec = DEFAULT_HEALTH_CHECK_DELTA_NS, -}; - -/* Define TLS health state. */ -DEFINE_URCU_TLS(struct health_state, health_state); - -/* - * It ensures that TLS memory used for the node and its container structure - * don't get reclaimed after the TLS owner thread exits until we have finished - * using it. - */ -static pthread_mutex_t health_mutex = PTHREAD_MUTEX_INITIALIZER; +#include +#include +#include +#include +#include +#include "utils.h" +#include "thread.h" -static struct health_tls_state_list health_state_list = { - .head = CDS_LIST_HEAD_INIT(health_state_list.head), +struct thread_notifiers { + struct lttng_pipe *quit_pipe; + sem_t ready; }; -/* - * This keeps track of the error state for unregistered thread. A thread - * reporting a health error, normally unregisters and quits. This makes the TLS - * health state not available to the health_check_state() call so on unregister - * we update this global error array so we can keep track of which thread was - * on error if the TLS health state has been removed. - */ -static enum health_flags global_error_state[HEALTH_NUM_TYPE]; - -/* - * Lock health state global list mutex. - */ -static void state_lock(void) +static +void mark_thread_as_ready(struct thread_notifiers *notifiers) { - pthread_mutex_lock(&health_mutex); + DBG("Marking health management thread as ready"); + sem_post(¬ifiers->ready); } -/* - * Unlock health state global list mutex. - */ -static void state_unlock(void) +static +void wait_until_thread_is_ready(struct thread_notifiers *notifiers) { - pthread_mutex_unlock(&health_mutex); + DBG("Waiting for health management thread to be ready"); + sem_wait(¬ifiers->ready); + DBG("Health management thread is ready"); } -/* - * Set time difference in res from time_a and time_b. - */ -static void time_diff(const struct timespec *time_a, - const struct timespec *time_b, struct timespec *res) +static void cleanup_health_management_thread(void *data) { - if (time_a->tv_nsec - time_b->tv_nsec < 0) { - res->tv_sec = time_a->tv_sec - time_b->tv_sec - 1; - res->tv_nsec = 1000000000L + time_a->tv_sec - time_b->tv_sec; - } else { - res->tv_sec = time_a->tv_sec - time_b->tv_sec; - res->tv_nsec = time_a->tv_nsec - time_b->tv_nsec; - } + struct thread_notifiers *notifiers = data; + + lttng_pipe_destroy(notifiers->quit_pipe); + sem_destroy(¬ifiers->ready); + free(notifiers); } /* - * Return true if time_a - time_b > diff, else false. + * Thread managing health check socket. */ -static int time_diff_gt(const struct timespec *time_a, - const struct timespec *time_b, const struct timespec *diff) +static void *thread_manage_health(void *data) { - struct timespec res; + const bool is_root = (getuid() == 0); + int sock = -1, new_sock = -1, ret, i, pollfd, err = -1; + uint32_t revents, nb_fd; + struct lttng_poll_event events; + struct health_comm_msg msg; + struct health_comm_reply reply; + /* Thread-specific quit pipe. */ + struct thread_notifiers *notifiers = data; + const int quit_pipe_read_fd = lttng_pipe_get_readfd( + notifiers->quit_pipe); + + DBG("[thread] Manage health check started"); - time_diff(time_a, time_b, &res); - time_diff(&res, diff, &res); + rcu_register_thread(); - if (res.tv_sec > 0) { - return 1; - } else if (res.tv_sec == 0 && res.tv_nsec > 0) { - return 1; + /* + * Created with a size of two for: + * - client socket + * - thread quit pipe + */ + ret = lttng_poll_create(&events, 2, LTTNG_CLOEXEC); + if (ret < 0) { + goto error; } - return 0; -} + /* Create unix socket */ + sock = lttcomm_create_unix_sock(config.health_unix_sock_path.value); + if (sock < 0) { + ERR("Unable to create health check Unix socket"); + goto error; + } -/* - * Validate health state. Checks for the error flag or health conditions. - * - * Return 0 if health is bad or else 1. - */ -static int validate_state(struct health_state *state) -{ - int retval = 1, ret; - unsigned long current, last; - struct timespec current_time; + if (is_root) { + /* lttng health client socket path permissions */ + gid_t gid; - assert(state); + ret = utils_get_group_id(config.tracing_group_name.value, true, &gid); + if (ret) { + /* Default to root group. */ + gid = 0; + } - last = state->last; - current = uatomic_read(&state->current); + ret = chown(config.health_unix_sock_path.value, 0, gid); + if (ret < 0) { + ERR("Unable to set group on %s", config.health_unix_sock_path.value); + PERROR("chown"); + goto error; + } - ret = clock_gettime(CLOCK_MONOTONIC, ¤t_time); - if (ret < 0) { - PERROR("Error reading time\n"); - /* error */ - retval = 0; - goto end; + ret = chmod(config.health_unix_sock_path.value, + S_IRUSR | S_IWUSR | S_IRGRP | S_IWGRP); + if (ret < 0) { + ERR("Unable to set permissions on %s", config.health_unix_sock_path.value); + PERROR("chmod"); + goto error; + } } /* - * Thread is in bad health if flag HEALTH_ERROR is set. It is also in bad - * health if, after the delta delay has passed, its the progress counter - * has not moved and it has NOT been waiting for a poll() call. + * Set the CLOEXEC flag. Return code is useless because either way, the + * show must go on. */ - if (uatomic_read(&state->flags) & HEALTH_ERROR) { - retval = 0; - goto end; + (void) utils_set_fd_cloexec(sock); + + ret = lttcomm_listen_unix_sock(sock); + if (ret < 0) { + goto error; } - /* - * Initial condition need to update the last counter and sample time, but - * should not check health in this initial case, because we don't know how - * much time has passed. - */ - if (state->last_time.tv_sec == 0 && state->last_time.tv_nsec == 0) { - /* update last counter and last sample time */ - state->last = current; - memcpy(&state->last_time, ¤t_time, sizeof(current_time)); - } else { - if (time_diff_gt(¤t_time, &state->last_time, &time_delta)) { - if (current == last && !HEALTH_IS_IN_POLL(current)) { - /* error */ - retval = 0; + ret = lttng_poll_add(&events, quit_pipe_read_fd, LPOLLIN | LPOLLERR); + if (ret < 0) { + goto error; + } + + /* Add the application registration socket */ + ret = lttng_poll_add(&events, sock, LPOLLIN | LPOLLPRI); + if (ret < 0) { + goto error; + } + + mark_thread_as_ready(notifiers); + while (1) { + DBG("Health check ready"); + + /* Infinite blocking call, waiting for transmission */ +restart: + ret = lttng_poll_wait(&events, -1); + if (ret < 0) { + /* + * Restart interrupted system call. + */ + if (errno == EINTR) { + goto restart; } - /* update last counter and last sample time */ - state->last = current; - memcpy(&state->last_time, ¤t_time, sizeof(current_time)); + goto error; + } - /* On error, stop right now and notify caller. */ - if (retval == 0) { - goto end; + nb_fd = ret; + + for (i = 0; i < nb_fd; i++) { + /* Fetch once the poll data */ + revents = LTTNG_POLL_GETEV(&events, i); + pollfd = LTTNG_POLL_GETFD(&events, i); + + /* Event on the registration socket */ + if (pollfd == sock) { + if (revents & LPOLLIN) { + continue; + } else if (revents & (LPOLLERR | LPOLLHUP | LPOLLRDHUP)) { + ERR("Health socket poll error"); + goto error; + } else { + ERR("Unexpected poll events %u for sock %d", revents, pollfd); + goto error; + } + } else { + /* Event on the thread's quit pipe. */ + err = 0; + goto exit; } } - } -end: - DBG("Health state current %lu, last %lu, ret %d", - current, last, ret); - return retval; -} + new_sock = lttcomm_accept_unix_sock(sock); + if (new_sock < 0) { + goto error; + } -/* - * Check health of a specific health type. Note that if a thread has not yet - * initialize its health subsystem or has quit, it's considered in a good - * state. - * - * Return 0 if health is bad or else 1. - */ -int health_check_state(enum health_type type) -{ - int retval = 1; - struct health_state *state; + /* + * Set the CLOEXEC flag. Return code is useless because either way, the + * show must go on. + */ + (void) utils_set_fd_cloexec(new_sock); + + DBG("Receiving data from client for health..."); + ret = lttcomm_recv_unix_sock(new_sock, (void *)&msg, sizeof(msg)); + if (ret <= 0) { + DBG("Nothing recv() from client... continuing"); + ret = close(new_sock); + if (ret) { + PERROR("close"); + } + continue; + } - assert(type < HEALTH_NUM_TYPE); + rcu_thread_online(); - state_lock(); + memset(&reply, 0, sizeof(reply)); + for (i = 0; i < NR_HEALTH_SESSIOND_TYPES; i++) { + /* + * health_check_state returns 0 if health is + * bad. + */ + if (!health_check_state(health_sessiond, i)) { + reply.ret_code |= 1ULL << i; + } + } - cds_list_for_each_entry(state, &health_state_list.head, node) { - int ret; + DBG2("Health check return value %" PRIx64, reply.ret_code); - if (state->type != type) { - continue; + ret = lttcomm_send_unix_sock(new_sock, (void *) &reply, + sizeof(reply)); + if (ret < 0) { + ERR("Failed to send health data back to client"); } - ret = validate_state(state); - if (!ret) { - retval = 0; - goto end; + /* End of transmission */ + ret = close(new_sock); + if (ret) { + PERROR("close"); } } - /* Check the global state since some state might not be visible anymore. */ - if (global_error_state[type] & HEALTH_ERROR) { - retval = 0; +exit: +error: + if (err) { + ERR("Health error occurred in %s", __func__); + } + DBG("Health check thread dying"); + unlink(config.health_unix_sock_path.value); + if (sock >= 0) { + ret = close(sock); + if (ret) { + PERROR("close"); + } } -end: - state_unlock(); - - DBG("Health check for type %d is %s", (int) type, - (retval == 0) ? "BAD" : "GOOD"); - return retval; + lttng_poll_clean(&events); + rcu_unregister_thread(); + return NULL; } -/* - * Init health state. - */ -void health_register(enum health_type type) +static bool shutdown_health_management_thread(void *data) { - assert(type < HEALTH_NUM_TYPE); - - /* Init TLS state. */ - uatomic_set(&URCU_TLS(health_state).last, 0); - uatomic_set(&URCU_TLS(health_state).last_time.tv_sec, 0); - uatomic_set(&URCU_TLS(health_state).last_time.tv_nsec, 0); - uatomic_set(&URCU_TLS(health_state).current, 0); - uatomic_set(&URCU_TLS(health_state).flags, 0); - uatomic_set(&URCU_TLS(health_state).type, type); - - /* Add it to the global TLS state list. */ - state_lock(); - cds_list_add(&URCU_TLS(health_state).node, &health_state_list.head); - state_unlock(); + struct thread_notifiers *notifiers = data; + const int write_fd = lttng_pipe_get_writefd(notifiers->quit_pipe); + + return notify_thread_pipe(write_fd) == 1; } -/* - * Remove node from global list. - */ -void health_unregister(void) +bool launch_health_management_thread(void) { - state_lock(); - /* - * On error, set the global_error_state since we are about to remove - * the node from the global list. - */ - if (uatomic_read(&URCU_TLS(health_state).flags) & HEALTH_ERROR) { - uatomic_set(&global_error_state[URCU_TLS(health_state).type], - HEALTH_ERROR); + struct thread_notifiers *notifiers; + struct lttng_thread *thread; + + notifiers = zmalloc(sizeof(*notifiers)); + if (!notifiers) { + goto error_alloc; } - cds_list_del(&URCU_TLS(health_state).node); - state_unlock(); -} -/* - * Initiliazie health check subsytem. This should be called before any health - * register occurs. - */ -void health_init(void) -{ - /* - * Get the maximum value between the default delta value and the TCP - * timeout with a safety net of the default health check delta. - */ - time_delta.tv_sec = max_t(unsigned long, - lttcomm_inet_tcp_timeout + DEFAULT_HEALTH_CHECK_DELTA_S, - time_delta.tv_sec); - DBG("Health check time delta in seconds set to %lu", time_delta.tv_sec); + sem_init(¬ifiers->ready, 0, 0); + notifiers->quit_pipe = lttng_pipe_open(FD_CLOEXEC); + if (!notifiers->quit_pipe) { + goto error; + } + thread = lttng_thread_create("Health management", + thread_manage_health, + shutdown_health_management_thread, + cleanup_health_management_thread, + notifiers); + if (!thread) { + goto error; + } + + wait_until_thread_is_ready(notifiers); + lttng_thread_put(thread); + return true; +error: + cleanup_health_management_thread(notifiers); +error_alloc: + return false; }