+ goto error_create;
+ }
+
+ cds_list_for_each_entry_safe(wait_node, tmp_wait_node,
+ &wait_queue->head, head) {
+ assert(wait_node->app);
+ ret = lttng_poll_add(&events, wait_node->app->sock,
+ LPOLLHUP | LPOLLERR);
+ if (ret < 0) {
+ goto error;
+ }
+
+ fd_added = 1;
+ }
+
+ if (!fd_added) {
+ goto end;
+ }
+
+ /*
+ * Poll but don't block so we can quickly identify the faulty events and
+ * clean them afterwards from the wait queue.
+ */
+ ret = lttng_poll_wait(&events, 0);
+ if (ret < 0) {
+ goto error;
+ }
+ nb_fd = ret;
+
+ for (i = 0; i < nb_fd; i++) {
+ /* Get faulty FD. */
+ uint32_t revents = LTTNG_POLL_GETEV(&events, i);
+ int pollfd = LTTNG_POLL_GETFD(&events, i);
+
+ if (!revents) {
+ /* No activity for this FD (poll implementation). */
+ continue;
+ }
+
+ cds_list_for_each_entry_safe(wait_node, tmp_wait_node,
+ &wait_queue->head, head) {
+ if (pollfd == wait_node->app->sock &&
+ (revents & (LPOLLHUP | LPOLLERR))) {
+ cds_list_del(&wait_node->head);
+ wait_queue->count--;
+ ust_app_destroy(wait_node->app);
+ free(wait_node);
+ /*
+ * Silence warning of use-after-free in
+ * cds_list_for_each_entry_safe which uses
+ * __typeof__(*wait_node).
+ */
+ wait_node = NULL;
+ break;
+ } else {
+ ERR("Unexpected poll events %u for sock %d", revents, pollfd);
+ goto error;
+ }
+ }
+ }
+
+ if (nb_fd > 0) {
+ DBG("Wait queue sanitized, %d node were cleaned up", nb_fd);
+ }
+
+end:
+ lttng_poll_clean(&events);
+ return;
+
+error:
+ lttng_poll_clean(&events);
+error_create:
+ ERR("Unable to sanitize wait queue");
+ return;
+}
+
+/*
+ * Dispatch request from the registration threads to the application
+ * communication thread.
+ */
+static void *thread_dispatch_ust_registration(void *data)
+{
+ int ret, err = -1;
+ struct cds_wfcq_node *node;
+ struct ust_command *ust_cmd = NULL;
+ struct ust_reg_wait_node *wait_node = NULL, *tmp_wait_node;
+ struct ust_reg_wait_queue wait_queue = {
+ .count = 0,
+ };
+
+ rcu_register_thread();
+
+ health_register(health_sessiond, HEALTH_SESSIOND_TYPE_APP_REG_DISPATCH);
+
+ if (testpoint(sessiond_thread_app_reg_dispatch)) {
+ goto error_testpoint;
+ }
+
+ health_code_update();
+
+ CDS_INIT_LIST_HEAD(&wait_queue.head);
+
+ DBG("[thread] Dispatch UST command started");
+
+ for (;;) {
+ health_code_update();
+
+ /* Atomically prepare the queue futex */
+ futex_nto1_prepare(&ust_cmd_queue.futex);
+
+ if (CMM_LOAD_SHARED(dispatch_thread_exit)) {
+ break;
+ }
+
+ do {
+ struct ust_app *app = NULL;
+ ust_cmd = NULL;
+
+ /*
+ * Make sure we don't have node(s) that have hung up before receiving
+ * the notify socket. This is to clean the list in order to avoid
+ * memory leaks from notify socket that are never seen.
+ */
+ sanitize_wait_queue(&wait_queue);
+
+ health_code_update();
+ /* Dequeue command for registration */
+ node = cds_wfcq_dequeue_blocking(&ust_cmd_queue.head, &ust_cmd_queue.tail);
+ if (node == NULL) {
+ DBG("Woken up but nothing in the UST command queue");
+ /* Continue thread execution */
+ break;
+ }
+
+ ust_cmd = caa_container_of(node, struct ust_command, node);
+
+ DBG("Dispatching UST registration pid:%d ppid:%d uid:%d"
+ " gid:%d sock:%d name:%s (version %d.%d)",
+ ust_cmd->reg_msg.pid, ust_cmd->reg_msg.ppid,
+ ust_cmd->reg_msg.uid, ust_cmd->reg_msg.gid,
+ ust_cmd->sock, ust_cmd->reg_msg.name,
+ ust_cmd->reg_msg.major, ust_cmd->reg_msg.minor);
+
+ if (ust_cmd->reg_msg.type == USTCTL_SOCKET_CMD) {
+ wait_node = zmalloc(sizeof(*wait_node));
+ if (!wait_node) {
+ PERROR("zmalloc wait_node dispatch");
+ ret = close(ust_cmd->sock);
+ if (ret < 0) {
+ PERROR("close ust sock dispatch %d", ust_cmd->sock);
+ }
+ lttng_fd_put(LTTNG_FD_APPS, 1);
+ free(ust_cmd);
+ goto error;
+ }
+ CDS_INIT_LIST_HEAD(&wait_node->head);
+
+ /* Create application object if socket is CMD. */
+ wait_node->app = ust_app_create(&ust_cmd->reg_msg,
+ ust_cmd->sock);
+ if (!wait_node->app) {
+ ret = close(ust_cmd->sock);
+ if (ret < 0) {
+ PERROR("close ust sock dispatch %d", ust_cmd->sock);
+ }
+ lttng_fd_put(LTTNG_FD_APPS, 1);
+ free(wait_node);
+ free(ust_cmd);
+ continue;
+ }
+ /*
+ * Add application to the wait queue so we can set the notify
+ * socket before putting this object in the global ht.
+ */
+ cds_list_add(&wait_node->head, &wait_queue.head);
+ wait_queue.count++;
+
+ free(ust_cmd);
+ /*
+ * We have to continue here since we don't have the notify
+ * socket and the application MUST be added to the hash table
+ * only at that moment.
+ */
+ continue;
+ } else {
+ /*
+ * Look for the application in the local wait queue and set the
+ * notify socket if found.
+ */
+ cds_list_for_each_entry_safe(wait_node, tmp_wait_node,
+ &wait_queue.head, head) {
+ health_code_update();
+ if (wait_node->app->pid == ust_cmd->reg_msg.pid) {
+ wait_node->app->notify_sock = ust_cmd->sock;
+ cds_list_del(&wait_node->head);
+ wait_queue.count--;
+ app = wait_node->app;
+ free(wait_node);
+ DBG3("UST app notify socket %d is set", ust_cmd->sock);
+ break;
+ }
+ }
+
+ /*
+ * With no application at this stage the received socket is
+ * basically useless so close it before we free the cmd data
+ * structure for good.
+ */
+ if (!app) {
+ ret = close(ust_cmd->sock);
+ if (ret < 0) {
+ PERROR("close ust sock dispatch %d", ust_cmd->sock);
+ }
+ lttng_fd_put(LTTNG_FD_APPS, 1);
+ }
+ free(ust_cmd);
+ }
+
+ if (app) {
+ /*
+ * @session_lock_list
+ *
+ * Lock the global session list so from the register up to the
+ * registration done message, no thread can see the application
+ * and change its state.
+ */
+ session_lock_list();
+ rcu_read_lock();
+
+ /*
+ * Add application to the global hash table. This needs to be
+ * done before the update to the UST registry can locate the
+ * application.
+ */
+ ust_app_add(app);
+
+ /* Set app version. This call will print an error if needed. */
+ (void) ust_app_version(app);
+
+ /* Send notify socket through the notify pipe. */
+ ret = send_socket_to_thread(apps_cmd_notify_pipe[1],
+ app->notify_sock);
+ if (ret < 0) {
+ rcu_read_unlock();
+ session_unlock_list();
+ /*
+ * No notify thread, stop the UST tracing. However, this is
+ * not an internal error of the this thread thus setting
+ * the health error code to a normal exit.
+ */
+ err = 0;
+ goto error;
+ }
+
+ /*
+ * Update newly registered application with the tracing
+ * registry info already enabled information.
+ */
+ update_ust_app(app->sock);
+
+ /*
+ * Don't care about return value. Let the manage apps threads
+ * handle app unregistration upon socket close.
+ */
+ (void) ust_app_register_done(app);
+
+ /*
+ * Even if the application socket has been closed, send the app
+ * to the thread and unregistration will take place at that
+ * place.
+ */
+ ret = send_socket_to_thread(apps_cmd_pipe[1], app->sock);
+ if (ret < 0) {
+ rcu_read_unlock();
+ session_unlock_list();
+ /*
+ * No apps. thread, stop the UST tracing. However, this is
+ * not an internal error of the this thread thus setting
+ * the health error code to a normal exit.
+ */
+ err = 0;
+ goto error;
+ }
+
+ rcu_read_unlock();
+ session_unlock_list();
+ }
+ } while (node != NULL);
+
+ health_poll_entry();
+ /* Futex wait on queue. Blocking call on futex() */
+ futex_nto1_wait(&ust_cmd_queue.futex);
+ health_poll_exit();
+ }
+ /* Normal exit, no error */
+ err = 0;
+
+error:
+ /* Clean up wait queue. */
+ cds_list_for_each_entry_safe(wait_node, tmp_wait_node,
+ &wait_queue.head, head) {
+ cds_list_del(&wait_node->head);
+ wait_queue.count--;
+ free(wait_node);
+ }
+
+ /* Empty command queue. */
+ for (;;) {
+ /* Dequeue command for registration */
+ node = cds_wfcq_dequeue_blocking(&ust_cmd_queue.head, &ust_cmd_queue.tail);
+ if (node == NULL) {
+ break;
+ }
+ ust_cmd = caa_container_of(node, struct ust_command, node);
+ ret = close(ust_cmd->sock);
+ if (ret < 0) {
+ PERROR("close ust sock exit dispatch %d", ust_cmd->sock);
+ }
+ lttng_fd_put(LTTNG_FD_APPS, 1);
+ free(ust_cmd);
+ }
+
+error_testpoint:
+ DBG("Dispatch thread dying");
+ if (err) {
+ health_error();
+ ERR("Health error occurred in %s", __func__);
+ }
+ health_unregister(health_sessiond);
+ rcu_unregister_thread();
+ return NULL;
+}
+
+/*
+ * This thread manage application registration.
+ */
+static void *thread_registration_apps(void *data)
+{
+ int sock = -1, i, ret, pollfd, err = -1;
+ uint32_t revents, nb_fd;
+ struct lttng_poll_event events;
+ /*
+ * Get allocated in this thread, enqueued to a global queue, dequeued and
+ * freed in the manage apps thread.
+ */
+ struct ust_command *ust_cmd = NULL;
+
+ DBG("[thread] Manage application registration started");
+
+ health_register(health_sessiond, HEALTH_SESSIOND_TYPE_APP_REG);
+
+ if (testpoint(sessiond_thread_registration_apps)) {
+ goto error_testpoint;
+ }
+
+ ret = lttcomm_listen_unix_sock(apps_sock);
+ if (ret < 0) {
+ goto error_listen;
+ }
+
+ /*
+ * Pass 2 as size here for the thread quit pipe and apps socket. Nothing
+ * more will be added to this poll set.
+ */
+ ret = sessiond_set_thread_pollset(&events, 2);
+ if (ret < 0) {
+ goto error_create_poll;
+ }
+
+ /* Add the application registration socket */
+ ret = lttng_poll_add(&events, apps_sock, LPOLLIN | LPOLLRDHUP);
+ if (ret < 0) {
+ goto error_poll_add;
+ }
+
+ /* Notify all applications to register */
+ ret = notify_ust_apps(1);
+ if (ret < 0) {
+ ERR("Failed to notify applications or create the wait shared memory.\n"
+ "Execution continues but there might be problem for already\n"
+ "running applications that wishes to register.");
+ }
+
+ while (1) {
+ DBG("Accepting application registration");
+
+ /* Inifinite blocking call, waiting for transmission */
+ restart:
+ health_poll_entry();
+ ret = lttng_poll_wait(&events, -1);
+ health_poll_exit();
+ if (ret < 0) {
+ /*
+ * Restart interrupted system call.
+ */
+ if (errno == EINTR) {
+ goto restart;
+ }
+ goto error;
+ }
+
+ nb_fd = ret;
+
+ for (i = 0; i < nb_fd; i++) {
+ health_code_update();
+
+ /* Fetch once the poll data */
+ revents = LTTNG_POLL_GETEV(&events, i);
+ pollfd = LTTNG_POLL_GETFD(&events, i);
+
+ if (!revents) {
+ /* No activity for this FD (poll implementation). */
+ continue;
+ }
+
+ /* Thread quit pipe has been closed. Killing thread. */
+ ret = sessiond_check_thread_quit_pipe(pollfd, revents);
+ if (ret) {
+ err = 0;
+ goto exit;
+ }
+
+ /* Event on the registration socket */
+ if (pollfd == apps_sock) {
+ if (revents & LPOLLIN) {
+ sock = lttcomm_accept_unix_sock(apps_sock);
+ if (sock < 0) {
+ goto error;
+ }
+
+ /*
+ * Set socket timeout for both receiving and ending.
+ * app_socket_timeout is in seconds, whereas
+ * lttcomm_setsockopt_rcv_timeout and
+ * lttcomm_setsockopt_snd_timeout expect msec as
+ * parameter.
+ */
+ if (app_socket_timeout >= 0) {
+ (void) lttcomm_setsockopt_rcv_timeout(sock,
+ app_socket_timeout * 1000);
+ (void) lttcomm_setsockopt_snd_timeout(sock,
+ app_socket_timeout * 1000);
+ }
+
+ /*
+ * Set the CLOEXEC flag. Return code is useless because
+ * either way, the show must go on.
+ */
+ (void) utils_set_fd_cloexec(sock);
+
+ /* Create UST registration command for enqueuing */
+ ust_cmd = zmalloc(sizeof(struct ust_command));
+ if (ust_cmd == NULL) {
+ PERROR("ust command zmalloc");
+ ret = close(sock);
+ if (ret) {
+ PERROR("close");
+ }
+ goto error;
+ }
+
+ /*
+ * Using message-based transmissions to ensure we don't
+ * have to deal with partially received messages.
+ */
+ ret = lttng_fd_get(LTTNG_FD_APPS, 1);
+ if (ret < 0) {
+ ERR("Exhausted file descriptors allowed for applications.");
+ free(ust_cmd);
+ ret = close(sock);
+ if (ret) {
+ PERROR("close");
+ }
+ sock = -1;
+ continue;
+ }
+
+ health_code_update();
+ ret = ust_app_recv_registration(sock, &ust_cmd->reg_msg);
+ if (ret < 0) {
+ free(ust_cmd);
+ /* Close socket of the application. */
+ ret = close(sock);
+ if (ret) {
+ PERROR("close");
+ }
+ lttng_fd_put(LTTNG_FD_APPS, 1);
+ sock = -1;
+ continue;
+ }
+ health_code_update();
+
+ ust_cmd->sock = sock;
+ sock = -1;
+
+ DBG("UST registration received with pid:%d ppid:%d uid:%d"
+ " gid:%d sock:%d name:%s (version %d.%d)",
+ ust_cmd->reg_msg.pid, ust_cmd->reg_msg.ppid,
+ ust_cmd->reg_msg.uid, ust_cmd->reg_msg.gid,
+ ust_cmd->sock, ust_cmd->reg_msg.name,
+ ust_cmd->reg_msg.major, ust_cmd->reg_msg.minor);
+
+ /*
+ * Lock free enqueue the registration request. The red pill
+ * has been taken! This apps will be part of the *system*.
+ */
+ cds_wfcq_enqueue(&ust_cmd_queue.head, &ust_cmd_queue.tail, &ust_cmd->node);
+
+ /*
+ * Wake the registration queue futex. Implicit memory
+ * barrier with the exchange in cds_wfcq_enqueue.
+ */
+ futex_nto1_wake(&ust_cmd_queue.futex);
+ } else if (revents & (LPOLLERR | LPOLLHUP | LPOLLRDHUP)) {
+ ERR("Register apps socket poll error");
+ goto error;
+ } else {
+ ERR("Unexpected poll events %u for sock %d", revents, pollfd);
+ goto error;
+ }
+ }
+ }
+ }
+
+exit:
+error:
+ /* Notify that the registration thread is gone */
+ notify_ust_apps(0);
+
+ if (apps_sock >= 0) {
+ ret = close(apps_sock);
+ if (ret) {
+ PERROR("close");
+ }
+ }
+ if (sock >= 0) {
+ ret = close(sock);
+ if (ret) {
+ PERROR("close");
+ }
+ lttng_fd_put(LTTNG_FD_APPS, 1);
+ }
+ unlink(apps_unix_sock_path);
+
+error_poll_add:
+ lttng_poll_clean(&events);
+error_listen:
+error_create_poll:
+error_testpoint:
+ DBG("UST Registration thread cleanup complete");
+ if (err) {
+ health_error();
+ ERR("Health error occurred in %s", __func__);
+ }
+ health_unregister(health_sessiond);
+
+ return NULL;
+}
+
+/*
+ * Start the thread_manage_consumer. This must be done after a lttng-consumerd
+ * exec or it will fails.
+ */
+static int spawn_consumer_thread(struct consumer_data *consumer_data)
+{
+ int ret, clock_ret;
+ struct timespec timeout;
+
+ /*
+ * Make sure we set the readiness flag to 0 because we are NOT ready.
+ * This access to consumer_thread_is_ready does not need to be
+ * protected by consumer_data.cond_mutex (yet) since the consumer
+ * management thread has not been started at this point.
+ */
+ consumer_data->consumer_thread_is_ready = 0;
+
+ /* Setup pthread condition */
+ ret = pthread_condattr_init(&consumer_data->condattr);
+ if (ret) {
+ errno = ret;
+ PERROR("pthread_condattr_init consumer data");
+ goto error;
+ }
+
+ /*
+ * Set the monotonic clock in order to make sure we DO NOT jump in time
+ * between the clock_gettime() call and the timedwait call. See bug #324
+ * for a more details and how we noticed it.
+ */
+ ret = pthread_condattr_setclock(&consumer_data->condattr, CLOCK_MONOTONIC);
+ if (ret) {
+ errno = ret;
+ PERROR("pthread_condattr_setclock consumer data");
+ goto error;
+ }
+
+ ret = pthread_cond_init(&consumer_data->cond, &consumer_data->condattr);
+ if (ret) {
+ errno = ret;
+ PERROR("pthread_cond_init consumer data");
+ goto error;
+ }
+
+ ret = pthread_create(&consumer_data->thread, default_pthread_attr(),
+ thread_manage_consumer, consumer_data);
+ if (ret) {
+ errno = ret;
+ PERROR("pthread_create consumer");
+ ret = -1;
+ goto error;
+ }
+
+ /* We are about to wait on a pthread condition */
+ pthread_mutex_lock(&consumer_data->cond_mutex);
+
+ /* Get time for sem_timedwait absolute timeout */
+ clock_ret = lttng_clock_gettime(CLOCK_MONOTONIC, &timeout);
+ /*
+ * Set the timeout for the condition timed wait even if the clock gettime
+ * call fails since we might loop on that call and we want to avoid to
+ * increment the timeout too many times.
+ */
+ timeout.tv_sec += DEFAULT_SEM_WAIT_TIMEOUT;
+
+ /*
+ * The following loop COULD be skipped in some conditions so this is why we
+ * set ret to 0 in order to make sure at least one round of the loop is
+ * done.
+ */
+ ret = 0;
+
+ /*
+ * Loop until the condition is reached or when a timeout is reached. Note
+ * that the pthread_cond_timedwait(P) man page specifies that EINTR can NOT
+ * be returned but the pthread_cond(3), from the glibc-doc, says that it is
+ * possible. This loop does not take any chances and works with both of
+ * them.
+ */
+ while (!consumer_data->consumer_thread_is_ready && ret != ETIMEDOUT) {
+ if (clock_ret < 0) {
+ PERROR("clock_gettime spawn consumer");
+ /* Infinite wait for the consumerd thread to be ready */
+ ret = pthread_cond_wait(&consumer_data->cond,
+ &consumer_data->cond_mutex);
+ } else {
+ ret = pthread_cond_timedwait(&consumer_data->cond,
+ &consumer_data->cond_mutex, &timeout);
+ }
+ }
+
+ /* Release the pthread condition */
+ pthread_mutex_unlock(&consumer_data->cond_mutex);
+
+ if (ret != 0) {
+ errno = ret;
+ if (ret == ETIMEDOUT) {
+ int pth_ret;
+
+ /*
+ * Call has timed out so we kill the kconsumerd_thread and return
+ * an error.
+ */
+ ERR("Condition timed out. The consumer thread was never ready."
+ " Killing it");
+ pth_ret = pthread_cancel(consumer_data->thread);
+ if (pth_ret < 0) {