X-Git-Url: https://git.lttng.org/?p=lttng-tools.git;a=blobdiff_plain;f=src%2Fbin%2Flttng-sessiond%2Frotation-thread.c;h=043993a60484130450dbe3dd75818b72ea9f4b66;hp=1704777a09ece70aef872745fd0ffc0c451ba939;hb=85e17b276759c089bf2ba894c080963cad853afa;hpb=d086f507d02078aed618ab291a0bc4a634958fa3

diff --git a/src/bin/lttng-sessiond/rotation-thread.c b/src/bin/lttng-sessiond/rotation-thread.c
index 1704777a0..043993a60 100644
--- a/src/bin/lttng-sessiond/rotation-thread.c
+++ b/src/bin/lttng-sessiond/rotation-thread.c
@@ -1,5 +1,6 @@
 /*
  * Copyright (C) 2017 - Julien Desfossez <jdesfossez@efficios.com>
+ * Copyright (C) 2018 - Jérémie Galarneau <jeremie.galarneau@efficios.com>
  *
  * 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
@@ -33,6 +34,7 @@
 
 #include <common/kernel-ctl/kernel-ctl.h>
 #include <lttng/notification/channel-internal.h>
+#include <lttng/rotate-internal.h>
 
 #include "rotation-thread.h"
 #include "lttng-sessiond.h"
@@ -40,66 +42,119 @@
 #include "rotate.h"
 #include "cmd.h"
 #include "session.h"
-#include "sessiond-timer.h"
+#include "timer.h"
+#include "notification-thread-commands.h"
 
 #include <urcu.h>
 #include <urcu/list.h>
-#include <urcu/rculfhash.h>
+
+struct lttng_notification_channel *rotate_notification_channel = NULL;
+
+struct rotation_thread {
+	struct lttng_poll_event events;
+};
+
+struct rotation_thread_job {
+	enum rotation_thread_job_type type;
+	struct ltt_session *session;
+	/* List member in struct rotation_thread_timer_queue. */
+	struct cds_list_head head;
+};
 
 /*
- * Store a struct rotation_channel_info for each channel that is currently
- * being rotated by the consumer.
+ * The timer thread enqueues jobs and wakes up the rotation thread.
+ * When the rotation thread wakes up, it empties the queue.
  */
-struct cds_lfht *channel_pending_rotate_ht;
+struct rotation_thread_timer_queue {
+	struct lttng_pipe *event_pipe;
+	struct cds_list_head list;
+	pthread_mutex_t lock;
+};
 
-struct rotation_thread_state {
-	struct lttng_poll_event events;
+struct rotation_thread_handle {
+	struct rotation_thread_timer_queue *rotation_timer_queue;
+	/* Access to the notification thread cmd_queue */
+	struct notification_thread_handle *notification_thread_handle;
 };
 
 static
-void channel_rotation_info_destroy(struct rotation_channel_info *channel_info)
+const char *get_job_type_str(enum rotation_thread_job_type job_type)
 {
-	assert(channel_info);
-	free(channel_info);
+	switch (job_type) {
+	case ROTATION_THREAD_JOB_TYPE_CHECK_PENDING_ROTATION:
+		return "CHECK_PENDING_ROTATION";
+	case ROTATION_THREAD_JOB_TYPE_SCHEDULED_ROTATION:
+		return "SCHEDULED_ROTATION";
+	default:
+		abort();
+	}
 }
 
-static
-int match_channel_info(struct cds_lfht_node *node, const void *key)
+struct rotation_thread_timer_queue *rotation_thread_timer_queue_create(void)
 {
-	struct rotation_channel_key *channel_key = (struct rotation_channel_key *) key;
-	struct rotation_channel_info *channel_info;
+	struct rotation_thread_timer_queue *queue = NULL;
 
-	channel_info = caa_container_of(node, struct rotation_channel_info,
-			rotate_channels_ht_node);
+	queue = zmalloc(sizeof(*queue));
+	if (!queue) {
+		PERROR("Failed to allocate timer rotate queue");
+		goto end;
+	}
 
-	return !!((channel_key->key == channel_info->channel_key.key) &&
-			(channel_key->domain == channel_info->channel_key.domain));
+	queue->event_pipe = lttng_pipe_open(FD_CLOEXEC | O_NONBLOCK);
+	CDS_INIT_LIST_HEAD(&queue->list);
+	pthread_mutex_init(&queue->lock, NULL);
+end:
+	return queue;
 }
 
-static
-struct rotation_channel_info *lookup_channel_pending(uint64_t key,
-		enum lttng_domain_type domain)
+void log_job_destruction(const struct rotation_thread_job *job)
 {
-	struct cds_lfht_iter iter;
-	struct cds_lfht_node *node;
-	struct rotation_channel_info *channel_info = NULL;
-	struct rotation_channel_key channel_key = { .key = key,
-			.domain = domain };
-
-	cds_lfht_lookup(channel_pending_rotate_ht,
-			hash_channel_key(&channel_key),
-			match_channel_info,
-			&channel_key, &iter);
-	node = cds_lfht_iter_get_node(&iter);
-	if (!node) {
-		goto end;
+	enum lttng_error_level log_level;
+	const char *job_type_str = get_job_type_str(job->type);
+
+	switch (job->type) {
+	case ROTATION_THREAD_JOB_TYPE_SCHEDULED_ROTATION:
+		/*
+		 * Not a problem, the scheduled rotation is racing with the teardown
+		 * of the daemon. In this case, the rotation will not happen, which
+		 * is not a problem (or at least, not important enough to delay
+		 * the shutdown of the session daemon).
+		 */
+		log_level = PRINT_DBG;
+		break;
+	case ROTATION_THREAD_JOB_TYPE_CHECK_PENDING_ROTATION:
+		/* This is not expected to happen; warn the user. */
+		log_level = PRINT_WARN;
+		break;
+	default:
+		abort();
 	}
 
-	channel_info = caa_container_of(node, struct rotation_channel_info,
-			rotate_channels_ht_node);
-	cds_lfht_del(channel_pending_rotate_ht, node);
-end:
-	return channel_info;
+	LOG(log_level, "Rotation thread timer queue still contains job of type %s targeting session \"%s\" on destruction",
+			job_type_str, job->session->name);
+}
+
+void rotation_thread_timer_queue_destroy(
+		struct rotation_thread_timer_queue *queue)
+{
+	struct rotation_thread_job *job, *tmp_job;
+
+	if (!queue) {
+		return;
+	}
+
+	lttng_pipe_destroy(queue->event_pipe);
+
+	pthread_mutex_lock(&queue->lock);
+	/* Empty wait queue. */
+	cds_list_for_each_entry_safe(job, tmp_job, &queue->list, head) {
+		log_job_destruction(job);
+		cds_list_del(&job->head);
+		free(job);
+	}
+	pthread_mutex_unlock(&queue->lock);
+	pthread_mutex_destroy(&queue->lock);
+	free(queue);
 }
 
 /*
@@ -108,87 +163,105 @@ end:
 void rotation_thread_handle_destroy(
 		struct rotation_thread_handle *handle)
 {
-	int ret;
+	free(handle);
+}
+
+struct rotation_thread_handle *rotation_thread_handle_create(
+		struct rotation_thread_timer_queue *rotation_timer_queue,
+		struct notification_thread_handle *notification_thread_handle)
+{
+	struct rotation_thread_handle *handle;
 
+	handle = zmalloc(sizeof(*handle));
 	if (!handle) {
 		goto end;
 	}
 
-	if (handle->ust32_consumer >= 0) {
-		ret = close(handle->ust32_consumer);
-		if (ret) {
-			PERROR("close 32-bit consumer channel rotation pipe");
-		}
-	}
-	if (handle->ust64_consumer >= 0) {
-		ret = close(handle->ust64_consumer);
-		if (ret) {
-			PERROR("close 64-bit consumer channel rotation pipe");
-		}
-	}
-	if (handle->kernel_consumer >= 0) {
-		ret = close(handle->kernel_consumer);
-		if (ret) {
-			PERROR("close kernel consumer channel rotation pipe");
+	handle->rotation_timer_queue = rotation_timer_queue;
+	handle->notification_thread_handle = notification_thread_handle;
+
+end:
+	return handle;
+}
+
+/*
+ * Called with the rotation_thread_timer_queue lock held.
+ * Return true if the same timer job already exists in the queue, false if not.
+ */
+static
+bool timer_job_exists(const struct rotation_thread_timer_queue *queue,
+		enum rotation_thread_job_type job_type,
+		struct ltt_session *session)
+{
+	bool exists = false;
+	struct rotation_thread_job *job;
+
+	cds_list_for_each_entry(job, &queue->list, head) {
+		if (job->session == session && job->type == job_type) {
+			exists = true;
+			goto end;
 		}
 	}
-
 end:
-	free(handle);
+	return exists;
 }
 
-struct rotation_thread_handle *rotation_thread_handle_create(
-		struct lttng_pipe *ust32_channel_rotate_pipe,
-		struct lttng_pipe *ust64_channel_rotate_pipe,
-		struct lttng_pipe *kernel_channel_rotate_pipe,
-		int thread_quit_pipe,
-		struct rotation_thread_timer_queue *rotation_timer_queue)
+void rotation_thread_enqueue_job(struct rotation_thread_timer_queue *queue,
+		enum rotation_thread_job_type job_type,
+		struct ltt_session *session)
 {
-	struct rotation_thread_handle *handle;
+	int ret;
+	const char * const dummy = "!";
+	struct rotation_thread_job *job = NULL;
+	const char *job_type_str = get_job_type_str(job_type);
 
-	handle = zmalloc(sizeof(*handle));
-	if (!handle) {
+	pthread_mutex_lock(&queue->lock);
+	if (timer_job_exists(queue, job_type, session)) {
+		/*
+		 * This timer job is already pending, we don't need to add
+		 * it.
+		 */
 		goto end;
 	}
 
-	if (ust32_channel_rotate_pipe) {
-		handle->ust32_consumer =
-				lttng_pipe_release_readfd(
-					ust32_channel_rotate_pipe);
-		if (handle->ust32_consumer < 0) {
-			goto error;
-		}
-	} else {
-		handle->ust32_consumer = -1;
-	}
-	if (ust64_channel_rotate_pipe) {
-		handle->ust64_consumer =
-				lttng_pipe_release_readfd(
-					ust64_channel_rotate_pipe);
-		if (handle->ust64_consumer < 0) {
-			goto error;
-		}
-	} else {
-		handle->ust64_consumer = -1;
+	job = zmalloc(sizeof(struct rotation_thread_job));
+	if (!job) {
+		PERROR("Failed to allocate rotation thread job of type \"%s\" for session \"%s\"",
+				job_type_str, session->name);
+		goto end;
 	}
-	if (kernel_channel_rotate_pipe) {
-		handle->kernel_consumer =
-				lttng_pipe_release_readfd(
-					kernel_channel_rotate_pipe);
-		if (handle->kernel_consumer < 0) {
-			goto error;
+	/* No reason for this to fail as the caller must hold a reference. */
+	(void) session_get(session);
+
+	job->session = session;
+	job->type = job_type;
+	cds_list_add_tail(&job->head, &queue->list);
+
+	ret = lttng_write(lttng_pipe_get_writefd(queue->event_pipe), dummy,
+			1);
+	if (ret < 0) {
+		/*
+		 * We do not want to block in the timer handler, the job has
+		 * been enqueued in the list, the wakeup pipe is probably full,
+		 * the job will be processed when the rotation_thread catches
+		 * up.
+		 */
+		if (errno == EAGAIN || errno == EWOULDBLOCK) {
+			/*
+			 * Not an error, but would be surprising and indicate
+			 * that the rotation thread can't keep up with the
+			 * current load.
+			 */
+			DBG("Wake-up pipe of rotation thread job queue is full");
+			goto end;
 		}
-	} else {
-		handle->kernel_consumer = -1;
+		PERROR("Failed to wake-up the rotation thread after pushing a job of type \"%s\" for session \"%s\"",
+				job_type_str, session->name);
+		goto end;
 	}
-	handle->thread_quit_pipe = thread_quit_pipe;
-	handle->rotation_timer_queue = rotation_timer_queue;
 
 end:
-	return handle;
-error:
-	rotation_thread_handle_destroy(handle);
-	return NULL;
+	pthread_mutex_unlock(&queue->lock);
 }
 
 static
@@ -198,22 +271,12 @@ int init_poll_set(struct lttng_poll_event *poll_set,
 	int ret;
 
 	/*
-	 * Create pollset with size 5:
-	 *	- sessiond quit pipe
-	 *	- sessiond timer pipe,
-	 *	- consumerd (32-bit user space) channel rotate pipe,
-	 *	- consumerd (64-bit user space) channel rotate pipe,
-	 *	- consumerd (kernel) channel rotate pipe,
+	 * Create pollset with size 2:
+	 *	- quit pipe,
+	 *	- rotation thread timer queue pipe,
 	 */
-	ret = lttng_poll_create(poll_set, 5, LTTNG_CLOEXEC);
-	if (ret < 0) {
-		goto end;
-	}
-
-	ret = lttng_poll_add(poll_set, handle->thread_quit_pipe,
-			LPOLLIN | LPOLLERR);
-	if (ret < 0) {
-		ERR("[rotation-thread] Failed to add thread_quit_pipe fd to pollset");
+	ret = sessiond_set_thread_pollset(poll_set, 2);
+	if (ret) {
 		goto error;
 	}
 	ret = lttng_poll_add(poll_set,
@@ -223,28 +286,7 @@ int init_poll_set(struct lttng_poll_event *poll_set,
 		ERR("[rotation-thread] Failed to add rotate_pending fd to pollset");
 		goto error;
 	}
-	ret = lttng_poll_add(poll_set, handle->ust32_consumer,
-			LPOLLIN | LPOLLERR);
-	if (ret < 0) {
-		ERR("[rotation-thread] Failed to add ust-32 channel rotation pipe fd to pollset");
-		goto error;
-	}
-	ret = lttng_poll_add(poll_set, handle->ust64_consumer,
-			LPOLLIN | LPOLLERR);
-	if (ret < 0) {
-		ERR("[rotation-thread] Failed to add ust-64 channel rotation pipe fd to pollset");
-		goto error;
-	}
-	if (handle->kernel_consumer >= 0) {
-		ret = lttng_poll_add(poll_set, handle->kernel_consumer,
-				LPOLLIN | LPOLLERR);
-		if (ret < 0) {
-			ERR("[rotation-thread] Failed to add kernel channel rotation pipe fd to pollset");
-			goto error;
-		}
-	}
 
-end:
 	return ret;
 error:
 	lttng_poll_clean(poll_set);
@@ -252,15 +294,17 @@ error:
 }
 
 static
-void fini_thread_state(struct rotation_thread_state *state)
+void fini_thread_state(struct rotation_thread *state)
 {
 	lttng_poll_clean(&state->events);
-	cds_lfht_destroy(channel_pending_rotate_ht, NULL);
+	if (rotate_notification_channel) {
+		lttng_notification_channel_destroy(rotate_notification_channel);
+	}
 }
 
 static
 int init_thread_state(struct rotation_thread_handle *handle,
-		struct rotation_thread_state *state)
+		struct rotation_thread *state)
 {
 	int ret;
 
@@ -273,113 +317,595 @@ int init_thread_state(struct rotation_thread_handle *handle,
 		goto end;
 	}
 
-	channel_pending_rotate_ht = cds_lfht_new(DEFAULT_HT_SIZE,
-			1, 0, CDS_LFHT_AUTO_RESIZE | CDS_LFHT_ACCOUNTING, NULL);
-	if (!channel_pending_rotate_ht) {
-		ERR("[rotation-thread] Failed to create channel pending rotation hash table");
+	rotate_notification_channel = lttng_notification_channel_create(
+			lttng_session_daemon_notification_endpoint);
+	if (!rotate_notification_channel) {
+		ERR("[rotation-thread] Could not create notification channel");
 		ret = -1;
 		goto end;
 	}
+	ret = lttng_poll_add(&state->events, rotate_notification_channel->socket,
+			LPOLLIN | LPOLLERR);
+	if (ret < 0) {
+		ERR("[rotation-thread] Failed to add notification fd to pollset");
+		goto end;
+	}
 
 end:
 	return ret;
 }
 
 static
-int handle_channel_rotation_pipe(int fd, uint32_t revents,
-		struct rotation_thread_handle *handle,
-		struct rotation_thread_state *state)
+int check_session_rotation_pending_local_on_consumer(
+		const struct ltt_session *session,
+		struct consumer_socket *socket, bool *rotation_completed)
 {
-	int ret = 0;
-	enum lttng_domain_type domain;
-	struct rotation_channel_info *channel_info;
-	struct ltt_session *session = NULL;
-	uint64_t key;
-
-	if (fd == handle->ust32_consumer ||
-			fd == handle->ust64_consumer) {
-		domain = LTTNG_DOMAIN_UST;
-	} else if (fd == handle->kernel_consumer) {
-		domain = LTTNG_DOMAIN_KERNEL;
+	int ret;
+
+	pthread_mutex_lock(socket->lock);
+	DBG("[rotation-thread] Checking for locally pending rotation on the %s consumer for session %s",
+			lttng_consumer_type_str(socket->type),
+			session->name);
+	ret = consumer_check_rotation_pending_local(socket,
+			session->id,
+			session->current_archive_id - 1);
+	pthread_mutex_unlock(socket->lock);
+
+	if (ret == 0) {
+		/* Rotation was completed on this consumer. */
+		DBG("[rotation-thread] Local rotation of trace archive %" PRIu64 " of session \"%s\" was completed on the %s consumer",
+				session->current_archive_id - 1,
+				session->name,
+				lttng_consumer_type_str(socket->type));
+		*rotation_completed = true;
+	} else if (ret == 1) {
+		/* Rotation pending on this consumer. */
+		DBG("[rotation-thread] Local rotation of trace archive %" PRIu64 " of session \"%s\" is pending on the %s consumer",
+				session->current_archive_id - 1,
+				session->name,
+				lttng_consumer_type_str(socket->type));
+		*rotation_completed = false;
+		ret = 0;
 	} else {
-		ERR("[rotation-thread] Unknown channel rotation pipe fd %d",
-				fd);
-		abort();
+		/* Not a fatal error. */
+		ERR("[rotation-thread] Encountered an error when checking if local rotation of trace archive %" PRIu64 " of session \"%s\" is pending on the %s consumer",
+				session->current_archive_id - 1,
+				session->name,
+				lttng_consumer_type_str(socket->type));
+		*rotation_completed = false;
 	}
+	return ret;
+}
 
-	if (revents & (LPOLLERR | LPOLLHUP | LPOLLRDHUP)) {
-		ret = lttng_poll_del(&state->events, fd);
+static
+int check_session_rotation_pending_local(struct ltt_session *session)
+{
+	int ret = 0;
+	struct consumer_socket *socket;
+	struct cds_lfht_iter iter;
+	bool rotation_completed = true;
+
+	/*
+	 * Check for a local pending rotation on all consumers (32-bit
+	 * user space, 64-bit user space, and kernel).
+	 */
+	DBG("[rotation-thread] Checking for pending local rotation on session \"%s\", trace archive %" PRIu64,
+			session->name, session->current_archive_id - 1);
+
+	rcu_read_lock();
+	if (!session->ust_session) {
+		goto skip_ust;
+	}
+	cds_lfht_for_each_entry(session->ust_session->consumer->socks->ht,
+			&iter, socket, node.node) {
+		ret = check_session_rotation_pending_local_on_consumer(session,
+				socket, &rotation_completed);
+		if (ret || !rotation_completed) {
+			goto end;
+		}
+	}
+
+skip_ust:
+	if (!session->kernel_session) {
+		goto skip_kernel;
+	}
+	cds_lfht_for_each_entry(session->kernel_session->consumer->socks->ht,
+				&iter, socket, node.node) {
+		ret = check_session_rotation_pending_local_on_consumer(session,
+				socket, &rotation_completed);
+		if (ret || !rotation_completed) {
+			goto end;
+		}
+	}
+skip_kernel:
+end:
+	rcu_read_unlock();
+
+	if (rotation_completed) {
+		DBG("[rotation-thread] Local rotation of trace archive %" PRIu64 " of session \"%s\" is complete on all consumers",
+				session->current_archive_id - 1,
+				session->name);
+		session->rotation_pending_local = false;
+	}
+	if (ret) {
+		ret = session_reset_rotation_state(session,
+				LTTNG_ROTATION_STATE_ERROR);
 		if (ret) {
-			ERR("[rotation-thread] Failed to remove consumer "
-					"rotation pipe from poll set");
+			ERR("Failed to reset rotation state of session \"%s\"",
+					session->name);
 		}
-		goto end;
 	}
+	return 0;
+}
 
-	do {
-		ret = read(fd, &key, sizeof(key));
-	} while (ret == -1 && errno == EINTR);
-	if (ret != sizeof(key)) {
-		ERR("[rotation-thread] Failed to read from pipe (fd = %i)",
-				fd);
-		ret = -1;
-		goto end;
+static
+int check_session_rotation_pending_relay(struct ltt_session *session)
+{
+	int ret;
+	struct consumer_socket *socket;
+	struct cds_lfht_iter iter;
+	bool rotation_completed = true;
+	const struct consumer_output *output;
+
+	/*
+	 * Check for a pending rotation on any consumer as we only use
+	 * it as a "tunnel" to the relayd.
+	 */
+
+	rcu_read_lock();
+	if (session->ust_session) {
+		cds_lfht_first(session->ust_session->consumer->socks->ht,
+				&iter);
+		output = session->ust_session->consumer;
+	} else {
+		cds_lfht_first(session->kernel_session->consumer->socks->ht,
+				&iter);
+		output = session->kernel_session->consumer;
+	}
+	assert(cds_lfht_iter_get_node(&iter));
+
+	socket = caa_container_of(cds_lfht_iter_get_node(&iter),
+			typeof(*socket), node.node);
+
+	pthread_mutex_lock(socket->lock);
+	DBG("[rotation-thread] Checking for pending relay rotation on session \"%s\", trace archive %" PRIu64 " through the %s consumer",
+			session->name, session->current_archive_id - 1,
+			lttng_consumer_type_str(socket->type));
+	ret = consumer_check_rotation_pending_relay(socket,
+			output,
+			session->id,
+			session->current_archive_id - 1);
+	pthread_mutex_unlock(socket->lock);
+
+	if (ret == 0) {
+		/* Rotation was completed on the relay. */
+		DBG("[rotation-thread] Relay rotation of trace archive %" PRIu64 " of session \"%s\" was completed",
+				session->current_archive_id - 1,
+				session->name);
+	} else if (ret == 1) {
+		/* Rotation pending on relay. */
+		DBG("[rotation-thread] Relay rotation of trace archive %" PRIu64 " of session \"%s\" is pending",
+				session->current_archive_id - 1,
+				session->name);
+		rotation_completed = false;
+	} else {
+		/* Not a fatal error. */
+		ERR("[rotation-thread] Encountered an error when checking if rotation of trace archive %" PRIu64 " of session \"%s\" is pending on the relay",
+				session->current_archive_id - 1,
+				session->name);
+		ret = session_reset_rotation_state(session,
+				LTTNG_ROTATION_STATE_ERROR);
+		if (ret) {
+			ERR("Failed to reset rotation state of session \"%s\"",
+					session->name);
+		}
+		rotation_completed = false;
 	}
 
-	DBG("[rotation-thread] Received notification for chan %" PRIu64
-			", domain %d\n", key, domain);
+	rcu_read_unlock();
 
-	channel_info = lookup_channel_pending(key, domain);
-	if (!channel_info) {
-		ERR("[rotation-thread] Failed to find channel_info (key = %"
-				PRIu64 ")", key);
-		ret = -1;
+	if (rotation_completed) {
+		DBG("[rotation-thread] Rotation of trace archive %" PRIu64 " of session \"%s\" is complete on the relay",
+				session->current_archive_id - 1,
+				session->name);
+		session->rotation_pending_relay = false;
+	}
+	return 0;
+}
+
+/*
+ * Check if the last rotation was completed, called with session lock held.
+ */
+static
+int check_session_rotation_pending(struct ltt_session *session,
+		struct notification_thread_handle *notification_thread_handle)
+{
+	int ret;
+	struct lttng_trace_archive_location *location;
+	time_t now;
+
+	DBG("[rotation-thread] Checking for pending rotation on session \"%s\", trace archive %" PRIu64,
+			session->name, session->current_archive_id - 1);
+
+	/*
+	 * The rotation-pending check timer of a session is launched in
+	 * one-shot mode. If the rotation is incomplete, the rotation
+	 * thread will re-enable the pending-check timer.
+	 *
+	 * The timer thread can't stop the timer itself since it is involved
+	 * in the check for the timer's quiescence.
+	 */
+	ret = timer_session_rotation_pending_check_stop(session);
+	if (ret) {
 		goto end;
 	}
-	rcu_read_lock();
-	session_lock_list();
-	session = session_find_by_id(channel_info->session_id);
-	if (!session) {
+
+	if (session->rotation_pending_local) {
+		/* Updates session->rotation_pending_local as needed. */
+		ret = check_session_rotation_pending_local(session);
+		if (ret) {
+			goto end;
+		}
+
 		/*
-		 * The session may have been destroyed before we had a chance to
-		 * perform this action, return gracefully.
+		 * No need to check for a pending rotation on the relay
+		 * since the rotation is not even completed locally yet.
 		 */
-		DBG("[rotation-thread] Session %" PRIu64 " not found",
-				channel_info->session_id);
-		ret = 0;
-		goto end_unlock_session_list;
+		if (session->rotation_pending_local) {
+			goto end;
+		}
 	}
 
-	session_lock(session);
-	if (--session->nr_chan_rotate_pending == 0) {
-		time_t now = time(NULL);
+	if (session->rotation_pending_relay) {
+		/* Updates session->rotation_pending_relay as needed. */
+		ret = check_session_rotation_pending_relay(session);
+		if (ret) {
+			goto end;
+		}
+
+		if (session->rotation_pending_relay) {
+			goto end;
+		}
+	}
+
+	DBG("[rotation-thread] Rotation of trace archive %" PRIu64 " completed for "
+			"session %s", session->current_archive_id - 1,
+			session->name);
 
-		if (now == (time_t) -1) {
-			session->rotation_status = LTTNG_ROTATION_STATUS_ERROR;
-			ret = LTTNG_ERR_UNK;
-			goto end_unlock_session;
+	/* Rename the completed trace archive's location. */
+	now = time(NULL);
+	if (now == (time_t) -1) {
+		ret = session_reset_rotation_state(session,
+				LTTNG_ROTATION_STATE_ERROR);
+		if (ret) {
+			ERR("Failed to reset rotation state of session \"%s\"",
+					session->name);
 		}
+		ret = LTTNG_ERR_UNK;
+		goto end;
+	}
 
-		ret = rename_complete_chunk(session, now);
+	ret = rename_completed_chunk(session, now);
+	if (ret < 0) {
+		ERR("Failed to rename completed rotation chunk");
+		goto end;
+	}
+	session->last_chunk_start_ts = session->current_chunk_start_ts;
+
+	/*
+	 * Now we can clear the "ONGOING" state in the session. New
+	 * rotations can start now.
+	 */
+	session->rotation_state = LTTNG_ROTATION_STATE_COMPLETED;
+
+	/* Ownership of location is transferred. */
+	location = session_get_trace_archive_location(session);
+	ret = notification_thread_command_session_rotation_completed(
+			notification_thread_handle,
+			session->name,
+			session->uid,
+			session->gid,
+			session->current_archive_id,
+			location);
+	if (ret != LTTNG_OK) {
+		ERR("[rotation-thread] Failed to notify notification thread of completed rotation for session %s",
+				session->name);
+	}
+
+	if (!session->active) {
+		/*
+		 * A stop command was issued during the rotation, it is
+		 * up to the rotation completion check to perform the
+		 * renaming of the last chunk that was produced.
+		 */
+		ret = notification_thread_command_session_rotation_ongoing(
+				notification_thread_handle,
+				session->name,
+				session->uid,
+				session->gid,
+				session->current_archive_id);
+		if (ret != LTTNG_OK) {
+			ERR("[rotation-thread] Failed to notify notification thread of completed rotation for session %s",
+					session->name);
+		}
+
+		ret = rename_active_chunk(session);
 		if (ret < 0) {
-			ERR("Failed to rename completed rotation chunk");
-			goto end_unlock_session;
+			ERR("[rotation-thread] Failed to rename active rotation chunk");
+			goto end;
+		}
+
+		/* Ownership of location is transferred. */
+		location = session_get_trace_archive_location(session);
+		ret = notification_thread_command_session_rotation_completed(
+				notification_thread_handle,
+				session->name,
+				session->uid,
+				session->gid,
+				session->current_archive_id,
+				location);
+		if (ret != LTTNG_OK) {
+			ERR("[rotation-thread] Failed to notify notification thread of completed rotation for session %s",
+					session->name);
 		}
-		session->rotate_pending = false;
-		session->rotation_status = LTTNG_ROTATION_STATUS_COMPLETED;
-		session->last_chunk_start_ts = session->current_chunk_start_ts;
-		DBG("Rotation completed for session %s", session->name);
 	}
 
 	ret = 0;
+end:
+	if (session->rotation_state == LTTNG_ROTATION_STATE_ONGOING) {
+		DBG("[rotation-thread] Rotation of trace archive %" PRIu64 " is still pending for session %s",
+				session->current_archive_id - 1, session->name);
+		ret = timer_session_rotation_pending_check_start(session,
+				DEFAULT_ROTATE_PENDING_TIMER);
+		if (ret) {
+			ERR("Re-enabling rotate pending timer");
+			ret = -1;
+			goto end;
+		}
+	}
 
-end_unlock_session:
-	channel_rotation_info_destroy(channel_info);
-	session_unlock(session);
-end_unlock_session_list:
+	return ret;
+}
+
+/* Call with the session and session_list locks held. */
+static
+int launch_session_rotation(struct ltt_session *session)
+{
+	int ret;
+	struct lttng_rotate_session_return rotation_return;
+
+	DBG("[rotation-thread] Launching scheduled time-based rotation on session \"%s\"",
+			session->name);
+
+	ret = cmd_rotate_session(session, &rotation_return);
+	if (ret == LTTNG_OK) {
+		DBG("[rotation-thread] Scheduled time-based rotation successfully launched on session \"%s\"",
+				session->name);
+	} else {
+		/* Don't consider errors as fatal. */
+		DBG("[rotation-thread] Scheduled time-based rotation aborted for session %s: %s",
+				session->name, lttng_strerror(ret));
+	}
+	return 0;
+}
+
+static
+int run_job(struct rotation_thread_job *job, struct ltt_session *session,
+		struct notification_thread_handle *notification_thread_handle)
+{
+	int ret;
+
+	switch (job->type) {
+	case ROTATION_THREAD_JOB_TYPE_SCHEDULED_ROTATION:
+	        ret = launch_session_rotation(session);
+		break;
+	case ROTATION_THREAD_JOB_TYPE_CHECK_PENDING_ROTATION:
+		ret = check_session_rotation_pending(session,
+				notification_thread_handle);
+		break;
+	default:
+		abort();
+	}
+	return ret;
+}
+
+static
+int handle_job_queue(struct rotation_thread_handle *handle,
+		struct rotation_thread *state,
+		struct rotation_thread_timer_queue *queue)
+{
+	int ret = 0;
+
+	for (;;) {
+		struct ltt_session *session;
+		struct rotation_thread_job *job;
+
+		/* Take the queue lock only to pop an element from the list. */
+		pthread_mutex_lock(&queue->lock);
+		if (cds_list_empty(&queue->list)) {
+			pthread_mutex_unlock(&queue->lock);
+			break;
+		}
+		job = cds_list_first_entry(&queue->list,
+				typeof(*job), head);
+		cds_list_del(&job->head);
+		pthread_mutex_unlock(&queue->lock);
+
+		session_lock_list();
+		session = job->session;
+		if (!session) {
+			DBG("[rotation-thread] Session \"%s\" not found",
+					session->name);
+			/*
+			 * This is a non-fatal error, and we cannot report it to
+			 * the user (timer), so just print the error and
+			 * continue the processing.
+			 *
+			 * While the timer thread will purge pending signals for
+			 * a session on the session's destruction, it is
+			 * possible for a job targeting that session to have
+			 * already been queued before it was destroyed.
+			 */
+			session_unlock_list();
+			free(job);
+			session_put(session);
+			continue;
+		}
+
+		session_lock(session);
+	        ret = run_job(job, session, handle->notification_thread_handle);
+		session_unlock(session);
+		/* Release reference held by the job. */
+		session_put(session);
+		session_unlock_list();
+		free(job);
+		if (ret) {
+			goto end;
+		}
+	}
+
+	ret = 0;
+
+end:
+	return ret;
+}
+
+static
+int handle_condition(const struct lttng_condition *condition,
+		const struct lttng_evaluation *evaluation,
+		struct notification_thread_handle *notification_thread_handle)
+{
+	int ret = 0;
+	const char *condition_session_name = NULL;
+	enum lttng_condition_type condition_type;
+	enum lttng_condition_status condition_status;
+	enum lttng_evaluation_status evaluation_status;
+	uint64_t consumed;
+	struct ltt_session *session;
+
+	condition_type = lttng_condition_get_type(condition);
+
+	if (condition_type != LTTNG_CONDITION_TYPE_SESSION_CONSUMED_SIZE) {
+		ret = -1;
+		ERR("[rotation-thread] Condition type and session usage type are not the same");
+		goto end;
+	}
+
+	/* Fetch info to test */
+	condition_status = lttng_condition_session_consumed_size_get_session_name(
+			condition, &condition_session_name);
+	if (condition_status != LTTNG_CONDITION_STATUS_OK) {
+		ERR("[rotation-thread] Session name could not be fetched");
+		ret = -1;
+		goto end;
+	}
+	evaluation_status = lttng_evaluation_session_consumed_size_get_consumed_size(evaluation,
+			&consumed);
+	if (evaluation_status != LTTNG_EVALUATION_STATUS_OK) {
+		ERR("[rotation-thread] Failed to get evaluation");
+		ret = -1;
+		goto end;
+	}
+
+	session_lock_list();
+	session = session_find_by_name(condition_session_name);
+	if (!session) {
+		ret = -1;
+		session_unlock_list();
+		ERR("[rotation-thread] Session \"%s\" not found",
+				condition_session_name);
+		goto end;
+	}
+	session_lock(session);
 	session_unlock_list();
-	rcu_read_unlock();
+
+	ret = unsubscribe_session_consumed_size_rotation(session,
+			notification_thread_handle);
+	if (ret) {
+		goto end_unlock;
+	}
+
+	ret = cmd_rotate_session(session, NULL);
+	if (ret == -LTTNG_ERR_ROTATION_PENDING) {
+		DBG("Rotate already pending, subscribe to the next threshold value");
+	} else if (ret != LTTNG_OK) {
+		ERR("[rotation-thread] Failed to rotate on size notification with error: %s",
+				lttng_strerror(ret));
+		ret = -1;
+		goto end_unlock;
+	}
+	ret = subscribe_session_consumed_size_rotation(session,
+			consumed + session->rotate_size,
+			notification_thread_handle);
+	if (ret) {
+		ERR("[rotation-thread] Failed to subscribe to session consumed size condition");
+		goto end_unlock;
+	}
+	ret = 0;
+
+end_unlock:
+	session_unlock(session);
+	session_put(session);
+end:
+	return ret;
+}
+
+static
+int handle_notification_channel(int fd,
+		struct rotation_thread_handle *handle,
+		struct rotation_thread *state)
+{
+	int ret;
+	bool notification_pending;
+	struct lttng_notification *notification = NULL;
+	enum lttng_notification_channel_status status;
+	const struct lttng_evaluation *notification_evaluation;
+	const struct lttng_condition *notification_condition;
+
+	status = lttng_notification_channel_has_pending_notification(
+			rotate_notification_channel, &notification_pending);
+	if (status != LTTNG_NOTIFICATION_CHANNEL_STATUS_OK) {
+		ERR("[rotation-thread ]Error occurred while checking for pending notification");
+		ret = -1;
+		goto end;
+	}
+
+	if (!notification_pending) {
+		ret = 0;
+		goto end;
+	}
+
+	/* Receive the next notification. */
+	status = lttng_notification_channel_get_next_notification(
+			rotate_notification_channel,
+			&notification);
+
+	switch (status) {
+	case LTTNG_NOTIFICATION_CHANNEL_STATUS_OK:
+		break;
+	case LTTNG_NOTIFICATION_CHANNEL_STATUS_NOTIFICATIONS_DROPPED:
+		/* Not an error, we will wait for the next one */
+		ret = 0;
+		goto end;;
+	case LTTNG_NOTIFICATION_CHANNEL_STATUS_CLOSED:
+		ERR("Notification channel was closed");
+		ret = -1;
+		goto end;
+	default:
+		/* Unhandled conditions / errors. */
+		ERR("Unknown notification channel status");
+		ret = -1;
+		goto end;
+	}
+
+	notification_condition = lttng_notification_get_condition(notification);
+	notification_evaluation = lttng_notification_get_evaluation(notification);
+
+	ret = handle_condition(notification_condition, notification_evaluation,
+			handle->notification_thread_handle);
+
 end:
+	lttng_notification_destroy(notification);
 	return ret;
 }
 
@@ -387,7 +913,7 @@ void *thread_rotation(void *data)
 {
 	int ret;
 	struct rotation_thread_handle *handle = data;
-	struct rotation_thread_state state;
+	struct rotation_thread thread;
 
 	DBG("[rotation-thread] Started rotation thread");
 
@@ -402,9 +928,9 @@ void *thread_rotation(void *data)
 	health_register(health_sessiond, HEALTH_SESSIOND_TYPE_ROTATION);
 	health_code_update();
 
-	ret = init_thread_state(handle, &state);
+	ret = init_thread_state(handle, &thread);
 	if (ret) {
-		goto end;
+		goto error;
 	}
 
 	/* Ready to handle client connections. */
@@ -415,7 +941,7 @@ void *thread_rotation(void *data)
 
 		health_poll_entry();
 		DBG("[rotation-thread] Entering poll wait");
-		ret = lttng_poll_wait(&state.events, -1);
+		ret = lttng_poll_wait(&thread.events, -1);
 		DBG("[rotation-thread] Poll wait returned (%i)", ret);
 		health_poll_exit();
 		if (ret < 0) {
@@ -431,31 +957,62 @@ void *thread_rotation(void *data)
 
 		fd_count = ret;
 		for (i = 0; i < fd_count; i++) {
-			int fd = LTTNG_POLL_GETFD(&state.events, i);
-			uint32_t revents = LTTNG_POLL_GETEV(&state.events, i);
+			int fd = LTTNG_POLL_GETFD(&thread.events, i);
+			uint32_t revents = LTTNG_POLL_GETEV(&thread.events, i);
 
 			DBG("[rotation-thread] Handling fd (%i) activity (%u)",
 					fd, revents);
 
-			if (fd == handle->thread_quit_pipe) {
-				DBG("[rotation-thread] Quit pipe activity");
-				goto exit;
-			} else if (fd == handle->ust32_consumer ||
-					fd == handle->ust64_consumer ||
-					fd == handle->kernel_consumer) {
-				ret = handle_channel_rotation_pipe(fd,
-						revents, handle, &state);
+			if (revents & LPOLLERR) {
+				ERR("[rotation-thread] Polling returned an error on fd %i", fd);
+				goto error;
+			}
+
+			if (fd == rotate_notification_channel->socket) {
+				ret = handle_notification_channel(fd, handle,
+						&thread);
+				if (ret) {
+					ERR("[rotation-thread] Error occurred while handling activity on notification channel socket");
+					goto error;
+				}
+			} else {
+				/* Job queue or quit pipe activity. */
+				if (fd == lttng_pipe_get_readfd(
+						handle->rotation_timer_queue->event_pipe)) {
+					char buf;
+
+					ret = lttng_read(fd, &buf, 1);
+					if (ret != 1) {
+						ERR("[rotation-thread] Failed to read from wakeup pipe (fd = %i)", fd);
+						ret = -1;
+						goto error;
+					}
+				}
+
+				/*
+				 * The job queue is serviced if there is
+				 * activity on the quit pipe to ensure it is
+				 * flushed and all references held in the queue
+				 * are released.
+				 */
+				ret = handle_job_queue(handle, &thread,
+						handle->rotation_timer_queue);
 				if (ret) {
-					ERR("[rotation-thread] Handle channel rotation pipe");
+					ERR("[rotation-thread] Failed to handle rotation timer pipe event");
 					goto error;
 				}
+
+				if (sessiond_check_thread_quit_pipe(fd, revents)) {
+					DBG("[rotation-thread] Quit pipe activity");
+					goto exit;
+				}
 			}
 		}
 	}
 exit:
 error:
 	DBG("[rotation-thread] Exit");
-	fini_thread_state(&state);
+	fini_thread_state(&thread);
 	health_unregister(health_sessiond);
 	rcu_thread_offline();
 	rcu_unregister_thread();