X-Git-Url: https://git.lttng.org/?a=blobdiff_plain;f=libringbuffer%2Fring_buffer_frontend.c;h=eb4e48629ca3d26a08b3c015b069a0e3c29f7d68;hb=bae63afc923740627352cdd6b0aeab1b0706387e;hp=393aa9f0d30d50331a453d1435f3b0e2eacaf24c;hpb=009769ca153da7ddfaa63643a84d4d19708b4586;p=lttng-ust.git

diff --git a/libringbuffer/ring_buffer_frontend.c b/libringbuffer/ring_buffer_frontend.c
index 393aa9f0..eb4e4862 100644
--- a/libringbuffer/ring_buffer_frontend.c
+++ b/libringbuffer/ring_buffer_frontend.c
@@ -55,7 +55,10 @@
 #include <sys/types.h>
 #include <sys/mman.h>
 #include <sys/stat.h>
+#include <unistd.h>
 #include <fcntl.h>
+#include <signal.h>
+#include <time.h>
 #include <urcu/compiler.h>
 #include <urcu/ref.h>
 #include <urcu/tls-compat.h>
@@ -77,6 +80,11 @@
 /* Print DBG() messages about events lost only every 1048576 hits */
 #define DBG_PRINT_NR_LOST	(1UL << 20)
 
+#define LTTNG_UST_RB_SIG_FLUSH		SIGRTMIN
+#define LTTNG_UST_RB_SIG_READ		SIGRTMIN + 1
+#define LTTNG_UST_RB_SIG_TEARDOWN	SIGRTMIN + 2
+#define CLOCKID		CLOCK_MONOTONIC
+
 /*
  * Use POSIX SHM: shm_open(3) and shm_unlink(3).
  * close(2) to close the fd returned by shm_open.
@@ -105,11 +113,37 @@ struct switch_offsets {
 
 DEFINE_URCU_TLS(unsigned int, lib_ring_buffer_nesting);
 
+/*
+ * wakeup_fd_mutex protects wakeup fd use by timer from concurrent
+ * close.
+ */
+static pthread_mutex_t wakeup_fd_mutex = PTHREAD_MUTEX_INITIALIZER;
+
 static
 void lib_ring_buffer_print_errors(struct channel *chan,
 				struct lttng_ust_lib_ring_buffer *buf, int cpu,
 				struct lttng_ust_shm_handle *handle);
 
+/*
+ * Handle timer teardown race wrt memory free of private data by
+ * ring buffer signals are handled by a single thread, which permits
+ * a synchronization point between handling of each signal.
+ * Protected by the lock within the structure.
+ */
+struct timer_signal_data {
+	pthread_t tid;	/* thread id managing signals */
+	int setup_done;
+	int qs_done;
+	pthread_mutex_t lock;
+};
+
+static struct timer_signal_data timer_signal = {
+	.tid = 0,
+	.setup_done = 0,
+	.qs_done = 0,
+	.lock = PTHREAD_MUTEX_INITIALIZER,
+};
+
 /**
  * lib_ring_buffer_reset - Reset ring buffer to initial values.
  * @buf: Ring buffer.
@@ -241,160 +275,355 @@ free_chanbuf:
 	return ret;
 }
 
-#if 0
-static void switch_buffer_timer(unsigned long data)
+static
+void lib_ring_buffer_channel_switch_timer(int sig, siginfo_t *si, void *uc)
 {
-	struct lttng_ust_lib_ring_buffer *buf = (struct lttng_ust_lib_ring_buffer *)data;
-	struct channel *chan = shmp(handle, buf->backend.chan);
-	const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config;
+	const struct lttng_ust_lib_ring_buffer_config *config;
+	struct lttng_ust_shm_handle *handle;
+	struct channel *chan;
+	int cpu;
+
+	assert(CMM_LOAD_SHARED(timer_signal.tid) == pthread_self());
+
+	chan = si->si_value.sival_ptr;
+	handle = chan->handle;
+	config = &chan->backend.config;
+
+	DBG("Switch timer for channel %p\n", chan);
 
 	/*
 	 * Only flush buffers periodically if readers are active.
 	 */
-	if (uatomic_read(&buf->active_readers))
-		lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE, handle);
-
-	//TODO timers
-	//if (config->alloc == RING_BUFFER_ALLOC_PER_CPU)
-	//	mod_timer_pinned(&buf->switch_timer,
-	//			 jiffies + chan->switch_timer_interval);
-	//else
-	//	mod_timer(&buf->switch_timer,
-	//		  jiffies + chan->switch_timer_interval);
+	pthread_mutex_lock(&wakeup_fd_mutex);
+	if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) {
+		for_each_possible_cpu(cpu) {
+			struct lttng_ust_lib_ring_buffer *buf =
+				shmp(handle, chan->backend.buf[cpu].shmp);
+			if (uatomic_read(&buf->active_readers))
+				lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE,
+					chan->handle);
+		}
+	} else {
+		struct lttng_ust_lib_ring_buffer *buf =
+			shmp(handle, chan->backend.buf[0].shmp);
+
+		if (uatomic_read(&buf->active_readers))
+			lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE,
+				chan->handle);
+	}
+	pthread_mutex_unlock(&wakeup_fd_mutex);
+	return;
 }
-#endif //0
 
-static void lib_ring_buffer_start_switch_timer(struct lttng_ust_lib_ring_buffer *buf,
-			   struct lttng_ust_shm_handle *handle)
+static
+void lib_ring_buffer_channel_do_read(struct channel *chan)
 {
-	struct channel *chan = shmp(handle, buf->backend.chan);
-	//const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config;
+	const struct lttng_ust_lib_ring_buffer_config *config;
+	struct lttng_ust_shm_handle *handle;
+	int cpu;
 
-	if (!chan->switch_timer_interval || buf->switch_timer_enabled)
-		return;
-	//TODO
-	//init_timer(&buf->switch_timer);
-	//buf->switch_timer.function = switch_buffer_timer;
-	//buf->switch_timer.expires = jiffies + chan->switch_timer_interval;
-	//buf->switch_timer.data = (unsigned long)buf;
-	//if (config->alloc == RING_BUFFER_ALLOC_PER_CPU)
-	//	add_timer_on(&buf->switch_timer, buf->backend.cpu);
-	//else
-	//	add_timer(&buf->switch_timer);
-	buf->switch_timer_enabled = 1;
+	handle = chan->handle;
+	config = &chan->backend.config;
+
+	/*
+	 * Only flush buffers periodically if readers are active.
+	 */
+	pthread_mutex_lock(&wakeup_fd_mutex);
+	if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) {
+		for_each_possible_cpu(cpu) {
+			struct lttng_ust_lib_ring_buffer *buf =
+				shmp(handle, chan->backend.buf[cpu].shmp);
+
+			if (uatomic_read(&buf->active_readers)
+			    && lib_ring_buffer_poll_deliver(config, buf,
+					chan, handle)) {
+				lib_ring_buffer_wakeup(buf, handle);
+			}
+		}
+	} else {
+		struct lttng_ust_lib_ring_buffer *buf =
+			shmp(handle, chan->backend.buf[0].shmp);
+
+		if (uatomic_read(&buf->active_readers)
+		    && lib_ring_buffer_poll_deliver(config, buf,
+				chan, handle)) {
+			lib_ring_buffer_wakeup(buf, handle);
+		}
+	}
+	pthread_mutex_unlock(&wakeup_fd_mutex);
 }
 
-static void lib_ring_buffer_stop_switch_timer(struct lttng_ust_lib_ring_buffer *buf,
-			   struct lttng_ust_shm_handle *handle)
+static
+void lib_ring_buffer_channel_read_timer(int sig, siginfo_t *si, void *uc)
 {
-	struct channel *chan = shmp(handle, buf->backend.chan);
+	struct channel *chan;
 
-	if (!chan->switch_timer_interval || !buf->switch_timer_enabled)
-		return;
+	assert(CMM_LOAD_SHARED(timer_signal.tid) == pthread_self());
+	chan = si->si_value.sival_ptr;
+	DBG("Read timer for channel %p\n", chan);
+	lib_ring_buffer_channel_do_read(chan);
+	return;
+}
+
+static
+void rb_setmask(sigset_t *mask)
+{
+	int ret;
 
-	//TODO
-	//del_timer_sync(&buf->switch_timer);
-	buf->switch_timer_enabled = 0;
+	ret = sigemptyset(mask);
+	if (ret) {
+		PERROR("sigemptyset");
+	}
+	ret = sigaddset(mask, LTTNG_UST_RB_SIG_FLUSH);
+	if (ret) {
+		PERROR("sigaddset");
+	}
+	ret = sigaddset(mask, LTTNG_UST_RB_SIG_READ);
+	if (ret) {
+		PERROR("sigaddset");
+	}
+	ret = sigaddset(mask, LTTNG_UST_RB_SIG_TEARDOWN);
+	if (ret) {
+		PERROR("sigaddset");
+	}
+}
+
+static
+void *sig_thread(void *arg)
+{
+	sigset_t mask;
+	siginfo_t info;
+	int signr;
+
+	/* Only self thread will receive signal mask. */
+	rb_setmask(&mask);
+	CMM_STORE_SHARED(timer_signal.tid, pthread_self());
+
+	for (;;) {
+		signr = sigwaitinfo(&mask, &info);
+		if (signr == -1) {
+			if (errno != EINTR)
+				PERROR("sigwaitinfo");
+			continue;
+		}
+		if (signr == LTTNG_UST_RB_SIG_FLUSH) {
+			lib_ring_buffer_channel_switch_timer(info.si_signo,
+					&info, NULL);
+		} else if (signr == LTTNG_UST_RB_SIG_READ) {
+			lib_ring_buffer_channel_read_timer(info.si_signo,
+					&info, NULL);
+		} else if (signr == LTTNG_UST_RB_SIG_TEARDOWN) {
+			cmm_smp_mb();
+			CMM_STORE_SHARED(timer_signal.qs_done, 1);
+			cmm_smp_mb();
+		} else {
+			ERR("Unexptected signal %d\n", info.si_signo);
+		}
+	}
+	return NULL;
 }
 
-#if 0
 /*
- * Polling timer to check the channels for data.
+ * Ensure only a single thread listens on the timer signal.
  */
-static void read_buffer_timer(unsigned long data)
+static
+void lib_ring_buffer_setup_timer_thread(void)
 {
-	struct lttng_ust_lib_ring_buffer *buf = (struct lttng_ust_lib_ring_buffer *)data;
-	struct channel *chan = shmp(handle, buf->backend.chan);
-	const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config;
+	pthread_t thread;
+	int ret;
+
+	pthread_mutex_lock(&timer_signal.lock);
+	if (timer_signal.setup_done)
+		goto end;
+
+	ret = pthread_create(&thread, NULL, &sig_thread, NULL);
+	if (ret) {
+		errno = ret;
+		PERROR("pthread_create");
+	}
+	ret = pthread_detach(thread);
+	if (ret) {
+		errno = ret;
+		PERROR("pthread_detach");
+	}
+	timer_signal.setup_done = 1;
+end:
+	pthread_mutex_unlock(&timer_signal.lock);
+}
+
+/*
+ * Wait for signal-handling thread quiescent state.
+ */
+static
+void lib_ring_buffer_wait_signal_thread_qs(unsigned int signr)
+{
+	sigset_t pending_set;
+	int ret;
+
+	/*
+	 * We need to be the only thread interacting with the thread
+	 * that manages signals for teardown synchronization.
+	 */
+	pthread_mutex_lock(&timer_signal.lock);
+
+	/*
+	 * Ensure we don't have any signal queued for this channel.
+	 */
+	for (;;) {
+		ret = sigemptyset(&pending_set);
+		if (ret == -1) {
+			PERROR("sigemptyset");
+		}
+		ret = sigpending(&pending_set);
+		if (ret == -1) {
+			PERROR("sigpending");
+		}
+		if (!sigismember(&pending_set, signr))
+			break;
+		caa_cpu_relax();
+	}
+
+	/*
+	 * From this point, no new signal handler will be fired that
+	 * would try to access "chan". However, we still need to wait
+	 * for any currently executing handler to complete.
+	 */
+	cmm_smp_mb();
+	CMM_STORE_SHARED(timer_signal.qs_done, 0);
+	cmm_smp_mb();
+
+	/*
+	 * Kill with LTTNG_UST_RB_SIG_TEARDOWN, so signal management
+	 * thread wakes up.
+	 */
+	kill(getpid(), LTTNG_UST_RB_SIG_TEARDOWN);
+
+	while (!CMM_LOAD_SHARED(timer_signal.qs_done))
+		caa_cpu_relax();
+	cmm_smp_mb();
+
+	pthread_mutex_unlock(&timer_signal.lock);
+}
 
-	CHAN_WARN_ON(chan, !buf->backend.allocated);
+static
+void lib_ring_buffer_channel_switch_timer_start(struct channel *chan)
+{
+	struct sigevent sev;
+	struct itimerspec its;
+	int ret;
+
+	if (!chan->switch_timer_interval || chan->switch_timer_enabled)
+		return;
+
+	chan->switch_timer_enabled = 1;
 
-	if (uatomic_read(&buf->active_readers))
-	    && lib_ring_buffer_poll_deliver(config, buf, chan)) {
-		//TODO
-		//wake_up_interruptible(&buf->read_wait);
-		//wake_up_interruptible(&chan->read_wait);
+	lib_ring_buffer_setup_timer_thread();
+
+	sev.sigev_notify = SIGEV_SIGNAL;
+	sev.sigev_signo = LTTNG_UST_RB_SIG_FLUSH;
+	sev.sigev_value.sival_ptr = chan;
+	ret = timer_create(CLOCKID, &sev, &chan->switch_timer);
+	if (ret == -1) {
+		PERROR("timer_create");
 	}
 
-	//TODO
-	//if (config->alloc == RING_BUFFER_ALLOC_PER_CPU)
-	//	mod_timer_pinned(&buf->read_timer,
-	//			 jiffies + chan->read_timer_interval);
-	//else
-	//	mod_timer(&buf->read_timer,
-	//		  jiffies + chan->read_timer_interval);
+	its.it_value.tv_sec = chan->switch_timer_interval / 1000000;
+	its.it_value.tv_nsec = chan->switch_timer_interval % 1000000;
+	its.it_interval.tv_sec = its.it_value.tv_sec;
+	its.it_interval.tv_nsec = its.it_value.tv_nsec;
+
+	ret = timer_settime(chan->switch_timer, 0, &its, NULL);
+	if (ret == -1) {
+		PERROR("timer_settime");
+	}
 }
-#endif //0
 
-static void lib_ring_buffer_start_read_timer(struct lttng_ust_lib_ring_buffer *buf,
-			   struct lttng_ust_shm_handle *handle)
+static
+void lib_ring_buffer_channel_switch_timer_stop(struct channel *chan)
+{
+	int ret;
+
+	if (!chan->switch_timer_interval || !chan->switch_timer_enabled)
+		return;
+
+	ret = timer_delete(chan->switch_timer);
+	if (ret == -1) {
+		PERROR("timer_delete");
+	}
+
+	lib_ring_buffer_wait_signal_thread_qs(LTTNG_UST_RB_SIG_FLUSH);
+
+	chan->switch_timer = 0;
+	chan->switch_timer_enabled = 0;
+}
+
+static
+void lib_ring_buffer_channel_read_timer_start(struct channel *chan)
 {
-	struct channel *chan = shmp(handle, buf->backend.chan);
 	const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config;
+	struct sigevent sev;
+	struct itimerspec its;
+	int ret;
 
 	if (config->wakeup != RING_BUFFER_WAKEUP_BY_TIMER
-	    || !chan->read_timer_interval
-	    || buf->read_timer_enabled)
+			|| !chan->read_timer_interval || chan->read_timer_enabled)
 		return;
 
-	//TODO
-	//init_timer(&buf->read_timer);
-	//buf->read_timer.function = read_buffer_timer;
-	//buf->read_timer.expires = jiffies + chan->read_timer_interval;
-	//buf->read_timer.data = (unsigned long)buf;
-
-	//if (config->alloc == RING_BUFFER_ALLOC_PER_CPU)
-	//	add_timer_on(&buf->read_timer, buf->backend.cpu);
-	//else
-	//	add_timer(&buf->read_timer);
-	buf->read_timer_enabled = 1;
+	chan->read_timer_enabled = 1;
+
+	lib_ring_buffer_setup_timer_thread();
+
+	sev.sigev_notify = SIGEV_SIGNAL;
+	sev.sigev_signo = LTTNG_UST_RB_SIG_READ;
+	sev.sigev_value.sival_ptr = chan;
+	ret = timer_create(CLOCKID, &sev, &chan->read_timer);
+	if (ret == -1) {
+		PERROR("timer_create");
+	}
+
+	its.it_value.tv_sec = chan->read_timer_interval / 1000000;
+	its.it_value.tv_nsec = chan->read_timer_interval % 1000000;
+	its.it_interval.tv_sec = its.it_value.tv_sec;
+	its.it_interval.tv_nsec = its.it_value.tv_nsec;
+
+	ret = timer_settime(chan->read_timer, 0, &its, NULL);
+	if (ret == -1) {
+		PERROR("timer_settime");
+	}
 }
 
-static void lib_ring_buffer_stop_read_timer(struct lttng_ust_lib_ring_buffer *buf,
-			   struct lttng_ust_shm_handle *handle)
+static
+void lib_ring_buffer_channel_read_timer_stop(struct channel *chan)
 {
-	struct channel *chan = shmp(handle, buf->backend.chan);
 	const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config;
+	int ret;
 
 	if (config->wakeup != RING_BUFFER_WAKEUP_BY_TIMER
-	    || !chan->read_timer_interval
-	    || !buf->read_timer_enabled)
+			|| !chan->read_timer_interval || !chan->read_timer_enabled)
 		return;
 
-	//TODO
-	//del_timer_sync(&buf->read_timer);
+	ret = timer_delete(chan->read_timer);
+	if (ret == -1) {
+		PERROR("timer_delete");
+	}
+
 	/*
 	 * do one more check to catch data that has been written in the last
 	 * timer period.
 	 */
-	if (lib_ring_buffer_poll_deliver(config, buf, chan, handle)) {
-		//TODO
-		//wake_up_interruptible(&buf->read_wait);
-		//wake_up_interruptible(&chan->read_wait);
-	}
-	buf->read_timer_enabled = 0;
+	lib_ring_buffer_channel_do_read(chan);
+
+	lib_ring_buffer_wait_signal_thread_qs(LTTNG_UST_RB_SIG_READ);
+
+	chan->read_timer = 0;
+	chan->read_timer_enabled = 0;
 }
 
 static void channel_unregister_notifiers(struct channel *chan,
 			   struct lttng_ust_shm_handle *handle)
 {
-	const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config;
-	int cpu;
-
-	if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) {
-		for_each_possible_cpu(cpu) {
-			struct lttng_ust_lib_ring_buffer *buf = shmp(handle, chan->backend.buf[cpu].shmp);
-
-			lib_ring_buffer_stop_switch_timer(buf, handle);
-			lib_ring_buffer_stop_read_timer(buf, handle);
-		}
-	} else {
-		struct lttng_ust_lib_ring_buffer *buf = shmp(handle, chan->backend.buf[0].shmp);
-
-		lib_ring_buffer_stop_switch_timer(buf, handle);
-		lib_ring_buffer_stop_read_timer(buf, handle);
-	}
-	//channel_backend_unregister_notifiers(&chan->backend);
+	lib_ring_buffer_channel_switch_timer_stop(chan);
+	lib_ring_buffer_channel_read_timer_stop(chan);
 }
 
 static void channel_print_errors(struct channel *chan,
@@ -421,7 +650,6 @@ static void channel_print_errors(struct channel *chan,
 static void channel_free(struct channel *chan,
 		struct lttng_ust_shm_handle *handle)
 {
-	channel_print_errors(chan, handle);
 	channel_backend_free(&chan->backend, handle);
 	/* chan is freed by shm teardown */
 	shm_object_table_destroy(handle->table);
@@ -458,7 +686,7 @@ struct lttng_ust_shm_handle *channel_create(const struct lttng_ust_lib_ring_buff
 		   size_t num_subbuf, unsigned int switch_timer_interval,
 		   unsigned int read_timer_interval)
 {
-	int ret, cpu;
+	int ret;
 	size_t shmsize, chansize;
 	struct channel *chan;
 	struct lttng_ust_shm_handle *handle;
@@ -527,31 +755,14 @@ struct lttng_ust_shm_handle *channel_create(const struct lttng_ust_lib_ring_buff
 	if (ret)
 		goto error_backend_init;
 
+	chan->handle = handle;
 	chan->commit_count_mask = (~0UL >> chan->backend.num_subbuf_order);
-	//TODO
-	//chan->switch_timer_interval = usecs_to_jiffies(switch_timer_interval);
-	//chan->read_timer_interval = usecs_to_jiffies(read_timer_interval);
-	//TODO
-	//init_waitqueue_head(&chan->read_wait);
-	//init_waitqueue_head(&chan->hp_wait);
 
-	if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) {
-		/*
-		 * In case of non-hotplug cpu, if the ring-buffer is allocated
-		 * in early initcall, it will not be notified of secondary cpus.
-		 * In that off case, we need to allocate for all possible cpus.
-		 */
-		for_each_possible_cpu(cpu) {
-			struct lttng_ust_lib_ring_buffer *buf = shmp(handle, chan->backend.buf[cpu].shmp);
-			lib_ring_buffer_start_switch_timer(buf, handle);
-			lib_ring_buffer_start_read_timer(buf, handle);
-		}
-	} else {
-		struct lttng_ust_lib_ring_buffer *buf = shmp(handle, chan->backend.buf[0].shmp);
+	chan->switch_timer_interval = switch_timer_interval;
+	chan->read_timer_interval = read_timer_interval;
+	lib_ring_buffer_channel_switch_timer_start(chan);
+	lib_ring_buffer_channel_read_timer_start(chan);
 
-		lib_ring_buffer_start_switch_timer(buf, handle);
-		lib_ring_buffer_start_read_timer(buf, handle);
-	}
 	return handle;
 
 error_backend_init:
@@ -563,7 +774,8 @@ error_table_alloc:
 }
 
 struct lttng_ust_shm_handle *channel_handle_create(void *data,
-					uint64_t memory_map_size)
+					uint64_t memory_map_size,
+					int wakeup_fd)
 {
 	struct lttng_ust_shm_handle *handle;
 	struct shm_object *object;
@@ -578,7 +790,7 @@ struct lttng_ust_shm_handle *channel_handle_create(void *data,
 		goto error_table_alloc;
 	/* Add channel object */
 	object = shm_object_table_append_mem(handle->table, data,
-			memory_map_size);
+			memory_map_size, wakeup_fd);
 	if (!object)
 		goto error_table_object;
 	/* struct channel is at object 0, offset 0 (hardcoded) */
@@ -639,6 +851,10 @@ void channel_destroy(struct channel *chan, struct lttng_ust_shm_handle *handle,
 		 * switching the buffers.
 		 */
 		channel_unregister_notifiers(chan, handle);
+		/*
+		 * The consumer prints errors.
+		 */
+		channel_print_errors(chan, handle);
 	}
 
 	/*
@@ -674,7 +890,27 @@ struct lttng_ust_lib_ring_buffer *channel_get_ring_buffer(
 	return shmp(handle, chan->backend.buf[cpu].shmp);
 }
 
-int ring_buffer_close_wait_fd(const struct lttng_ust_lib_ring_buffer_config *config,
+int ring_buffer_channel_close_wait_fd(const struct lttng_ust_lib_ring_buffer_config *config,
+			struct channel *chan,
+			struct lttng_ust_shm_handle *handle)
+{
+	struct shm_ref *ref;
+
+	ref = &handle->chan._ref;
+	return shm_close_wait_fd(handle, ref);
+}
+
+int ring_buffer_channel_close_wakeup_fd(const struct lttng_ust_lib_ring_buffer_config *config,
+			struct channel *chan,
+			struct lttng_ust_shm_handle *handle)
+{
+	struct shm_ref *ref;
+
+	ref = &handle->chan._ref;
+	return shm_close_wakeup_fd(handle, ref);
+}
+
+int ring_buffer_stream_close_wait_fd(const struct lttng_ust_lib_ring_buffer_config *config,
 			struct channel *chan,
 			struct lttng_ust_shm_handle *handle,
 			int cpu)
@@ -691,12 +927,13 @@ int ring_buffer_close_wait_fd(const struct lttng_ust_lib_ring_buffer_config *con
 	return shm_close_wait_fd(handle, ref);
 }
 
-int ring_buffer_close_wakeup_fd(const struct lttng_ust_lib_ring_buffer_config *config,
+int ring_buffer_stream_close_wakeup_fd(const struct lttng_ust_lib_ring_buffer_config *config,
 			struct channel *chan,
 			struct lttng_ust_shm_handle *handle,
 			int cpu)
 {
 	struct shm_ref *ref;
+	int ret;
 
 	if (config->alloc == RING_BUFFER_ALLOC_GLOBAL) {
 		cpu = 0;
@@ -705,7 +942,10 @@ int ring_buffer_close_wakeup_fd(const struct lttng_ust_lib_ring_buffer_config *c
 			return -EINVAL;
 	}
 	ref = &chan->backend.buf[cpu].shmp._ref;
-	return shm_close_wakeup_fd(handle, ref);
+	pthread_mutex_lock(&wakeup_fd_mutex);
+	ret = shm_close_wakeup_fd(handle, ref);
+	pthread_mutex_unlock(&wakeup_fd_mutex);
+	return ret;
 }
 
 int lib_ring_buffer_open_read(struct lttng_ust_lib_ring_buffer *buf,
@@ -787,7 +1027,7 @@ nodata:
 }
 
 /**
- * lib_ring_buffer_put_snapshot - move consumed counter forward
+ * lib_ring_buffer_move_consumer - move consumed counter forward
  * @buf: ring buffer
  * @consumed_new: new consumed count value
  */
@@ -867,7 +1107,7 @@ retry:
 	 */
 	if (((commit_count - chan->backend.subbuf_size)
 	     & chan->commit_count_mask)
-	    - (buf_trunc(consumed_cur, chan)
+	    - (buf_trunc(consumed, chan)
 	       >> chan->backend.num_subbuf_order)
 	    != 0)
 		goto nodata;
@@ -876,7 +1116,7 @@ retry:
 	 * Check that we are not about to read the same subbuffer in
 	 * which the writer head is.
 	 */
-	if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_cur, chan)
+	if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed, chan)
 	    == 0)
 		goto nodata;
 
@@ -1093,11 +1333,10 @@ void lib_ring_buffer_switch_old_start(struct lttng_ust_lib_ring_buffer *buf,
 	commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, oldidx)->cc);
 	/* Check if the written buffer has to be delivered */
 	lib_ring_buffer_check_deliver(config, buf, chan, offsets->old,
-				      commit_count, oldidx, handle);
+				      commit_count, oldidx, handle, tsc);
 	lib_ring_buffer_write_commit_counter(config, buf, chan, oldidx,
-					     offsets->old, commit_count,
-					     config->cb.subbuffer_header_size(),
-					     handle);
+			offsets->old + config->cb.subbuffer_header_size(),
+			commit_count, handle);
 }
 
 /*
@@ -1132,10 +1371,9 @@ void lib_ring_buffer_switch_old_end(struct lttng_ust_lib_ring_buffer *buf,
 	v_add(config, padding_size, &shmp_index(handle, buf->commit_hot, oldidx)->cc);
 	commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, oldidx)->cc);
 	lib_ring_buffer_check_deliver(config, buf, chan, offsets->old - 1,
-				      commit_count, oldidx, handle);
+				      commit_count, oldidx, handle, tsc);
 	lib_ring_buffer_write_commit_counter(config, buf, chan, oldidx,
-					     offsets->old, commit_count,
-					     padding_size, handle);
+			offsets->old + padding_size, commit_count, handle);
 }
 
 /*
@@ -1168,18 +1406,19 @@ void lib_ring_buffer_switch_new_start(struct lttng_ust_lib_ring_buffer *buf,
 	commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, beginidx)->cc);
 	/* Check if the written buffer has to be delivered */
 	lib_ring_buffer_check_deliver(config, buf, chan, offsets->begin,
-				      commit_count, beginidx, handle);
+				      commit_count, beginidx, handle, tsc);
 	lib_ring_buffer_write_commit_counter(config, buf, chan, beginidx,
-					     offsets->begin, commit_count,
-					     config->cb.subbuffer_header_size(),
-					     handle);
+			offsets->begin + config->cb.subbuffer_header_size(),
+			commit_count, handle);
 }
 
 /*
  * lib_ring_buffer_switch_new_end: finish switching current subbuffer
  *
- * The only remaining threads could be the ones with pending commits. They will
- * have to do the deliver themselves.
+ * Calls subbuffer_set_data_size() to set the data size of the current
+ * sub-buffer. We do not need to perform check_deliver nor commit here,
+ * since this task will be done by the "commit" of the event for which
+ * we are currently doing the space reservation.
  */
 static
 void lib_ring_buffer_switch_new_end(struct lttng_ust_lib_ring_buffer *buf,
@@ -1189,26 +1428,12 @@ void lib_ring_buffer_switch_new_end(struct lttng_ust_lib_ring_buffer *buf,
 				    struct lttng_ust_shm_handle *handle)
 {
 	const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config;
-	unsigned long endidx = subbuf_index(offsets->end - 1, chan);
-	unsigned long commit_count, padding_size, data_size;
+	unsigned long endidx, data_size;
 
+	endidx = subbuf_index(offsets->end - 1, chan);
 	data_size = subbuf_offset(offsets->end - 1, chan) + 1;
-	padding_size = chan->backend.subbuf_size - data_size;
 	subbuffer_set_data_size(config, &buf->backend, endidx, data_size,
 				handle);
-
-	/*
-	 * Order all writes to buffer before the commit count update that will
-	 * determine that the subbuffer is full.
-	 */
-	cmm_smp_wmb();
-	v_add(config, padding_size, &shmp_index(handle, buf->commit_hot, endidx)->cc);
-	commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, endidx)->cc);
-	lib_ring_buffer_check_deliver(config, buf, chan, offsets->end - 1,
-				  commit_count, endidx, handle);
-	lib_ring_buffer_write_commit_counter(config, buf, chan, endidx,
-					     offsets->end, commit_count,
-					     padding_size, handle);
 }
 
 /*
@@ -1221,10 +1446,11 @@ int lib_ring_buffer_try_switch_slow(enum switch_mode mode,
 				    struct lttng_ust_lib_ring_buffer *buf,
 				    struct channel *chan,
 				    struct switch_offsets *offsets,
-				    uint64_t *tsc)
+				    uint64_t *tsc,
+				    struct lttng_ust_shm_handle *handle)
 {
 	const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config;
-	unsigned long off;
+	unsigned long off, reserve_commit_diff;
 
 	offsets->begin = v_read(config, &buf->offset);
 	offsets->old = offsets->begin;
@@ -1249,36 +1475,69 @@ int lib_ring_buffer_try_switch_slow(enum switch_mode mode,
 	 * timestamps) are visible to the reader. This is required for
 	 * quiescence guarantees for the fusion merge.
 	 */
-	if (mode == SWITCH_FLUSH || off > 0) {
-		if (caa_unlikely(off == 0)) {
-			/*
-			 * A final flush that encounters an empty
-			 * sub-buffer cannot switch buffer if a
-			 * reader is located within this sub-buffer.
-			 * Anyway, the purpose of final flushing of a
-			 * sub-buffer at offset 0 is to handle the case
-			 * of entirely empty stream.
-			 */
-			if (caa_unlikely(subbuf_trunc(offsets->begin, chan)
-					 - subbuf_trunc((unsigned long)
-					     uatomic_read(&buf->consumed), chan)
-					>= chan->backend.buf_size))
-				return -1;
-			/*
-			 * The client does not save any header information.
-			 * Don't switch empty subbuffer on finalize, because it
-			 * is invalid to deliver a completely empty subbuffer.
-			 */
-			if (!config->cb.subbuffer_header_size())
+	if (mode != SWITCH_FLUSH && !off)
+		return -1;	/* we do not have to switch : buffer is empty */
+
+	if (caa_unlikely(off == 0)) {
+		unsigned long sb_index, commit_count;
+
+		/*
+		 * We are performing a SWITCH_FLUSH. At this stage, there are no
+		 * concurrent writes into the buffer.
+		 *
+		 * The client does not save any header information.  Don't
+		 * switch empty subbuffer on finalize, because it is invalid to
+		 * deliver a completely empty subbuffer.
+		 */
+		if (!config->cb.subbuffer_header_size())
+			return -1;
+
+		/* Test new buffer integrity */
+		sb_index = subbuf_index(offsets->begin, chan);
+		commit_count = v_read(config,
+				&shmp_index(handle, buf->commit_cold,
+					sb_index)->cc_sb);
+		reserve_commit_diff =
+		  (buf_trunc(offsets->begin, chan)
+		   >> chan->backend.num_subbuf_order)
+		  - (commit_count & chan->commit_count_mask);
+		if (caa_likely(reserve_commit_diff == 0)) {
+			/* Next subbuffer not being written to. */
+			if (caa_unlikely(config->mode != RING_BUFFER_OVERWRITE &&
+				subbuf_trunc(offsets->begin, chan)
+				 - subbuf_trunc((unsigned long)
+				     uatomic_read(&buf->consumed), chan)
+				>= chan->backend.buf_size)) {
+				/*
+				 * We do not overwrite non consumed buffers
+				 * and we are full : don't switch.
+				 */
 				return -1;
+			} else {
+				/*
+				 * Next subbuffer not being written to, and we
+				 * are either in overwrite mode or the buffer is
+				 * not full. It's safe to write in this new
+				 * subbuffer.
+				 */
+			}
+		} else {
 			/*
-			 * Need to write the subbuffer start header on finalize.
+			 * Next subbuffer reserve offset does not match the
+			 * commit offset. Don't perform switch in
+			 * producer-consumer and overwrite mode.  Caused by
+			 * either a writer OOPS or too many nested writes over a
+			 * reserve/commit pair.
 			 */
-			offsets->switch_old_start = 1;
+			return -1;
 		}
-		offsets->begin = subbuf_align(offsets->begin, chan);
-	} else
-		return -1;	/* we do not have to switch : buffer is empty */
+
+		/*
+		 * Need to write the subbuffer start header on finalize.
+		 */
+		offsets->switch_old_start = 1;
+	}
+	offsets->begin = subbuf_align(offsets->begin, chan);
 	/* Note: old points to the next subbuf at offset 0 */
 	offsets->end = offsets->begin;
 	return 0;
@@ -1308,7 +1567,7 @@ void lib_ring_buffer_switch_slow(struct lttng_ust_lib_ring_buffer *buf, enum swi
 	 */
 	do {
 		if (lib_ring_buffer_try_switch_slow(mode, buf, chan, &offsets,
-						    &tsc))
+						    &tsc, handle))
 			return;	/* Switch not needed */
 	} while (v_cmpxchg(config, &buf->offset, offsets.old, offsets.end)
 		 != offsets.old);
@@ -1358,9 +1617,10 @@ int lib_ring_buffer_try_reserve_slow(struct lttng_ust_lib_ring_buffer *buf,
 {
 	const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config;
 	struct lttng_ust_shm_handle *handle = ctx->handle;
-	unsigned long reserve_commit_diff;
+	unsigned long reserve_commit_diff, offset_cmp;
 
-	offsets->begin = v_read(config, &buf->offset);
+retry:
+	offsets->begin = offset_cmp = v_read(config, &buf->offset);
 	offsets->old = offsets->begin;
 	offsets->switch_new_start = 0;
 	offsets->switch_new_end = 0;
@@ -1392,7 +1652,7 @@ int lib_ring_buffer_try_reserve_slow(struct lttng_ust_lib_ring_buffer *buf,
 		}
 	}
 	if (caa_unlikely(offsets->switch_new_start)) {
-		unsigned long sb_index;
+		unsigned long sb_index, commit_count;
 
 		/*
 		 * We are typically not filling the previous buffer completely.
@@ -1403,12 +1663,32 @@ int lib_ring_buffer_try_reserve_slow(struct lttng_ust_lib_ring_buffer *buf,
 				 + config->cb.subbuffer_header_size();
 		/* Test new buffer integrity */
 		sb_index = subbuf_index(offsets->begin, chan);
+		/*
+		 * Read buf->offset before buf->commit_cold[sb_index].cc_sb.
+		 * lib_ring_buffer_check_deliver() has the matching
+		 * memory barriers required around commit_cold cc_sb
+		 * updates to ensure reserve and commit counter updates
+		 * are not seen reordered when updated by another CPU.
+		 */
+		cmm_smp_rmb();
+		commit_count = v_read(config,
+				&shmp_index(handle, buf->commit_cold,
+					sb_index)->cc_sb);
+		/* Read buf->commit_cold[sb_index].cc_sb before buf->offset. */
+		cmm_smp_rmb();
+		if (caa_unlikely(offset_cmp != v_read(config, &buf->offset))) {
+			/*
+			 * The reserve counter have been concurrently updated
+			 * while we read the commit counter. This means the
+			 * commit counter we read might not match buf->offset
+			 * due to concurrent update. We therefore need to retry.
+			 */
+			goto retry;
+		}
 		reserve_commit_diff =
 		  (buf_trunc(offsets->begin, chan)
 		   >> chan->backend.num_subbuf_order)
-		  - ((unsigned long) v_read(config,
-					    &shmp_index(handle, buf->commit_cold, sb_index)->cc_sb)
-		     & chan->commit_count_mask);
+		  - (commit_count & chan->commit_count_mask);
 		if (caa_likely(reserve_commit_diff == 0)) {
 			/* Next subbuffer not being written to. */
 			if (caa_unlikely(config->mode != RING_BUFFER_OVERWRITE &&
@@ -1443,7 +1723,8 @@ int lib_ring_buffer_try_reserve_slow(struct lttng_ust_lib_ring_buffer *buf,
 
 			/*
 			 * Next subbuffer reserve offset does not match the
-			 * commit offset. Drop record in producer-consumer and
+			 * commit offset, and this did not involve update to the
+			 * reserve counter. Drop record in producer-consumer and
 			 * overwrite mode. Caused by either a writer OOPS or too
 			 * many nested writes over a reserve/commit pair.
 			 */
@@ -1592,3 +1873,20 @@ void lttng_fixup_ringbuffer_tls(void)
 {
 	asm volatile ("" : : "m" (URCU_TLS(lib_ring_buffer_nesting)));
 }
+
+void lib_ringbuffer_signal_init(void)
+{
+	sigset_t mask;
+	int ret;
+
+	/*
+	 * Block signal for entire process, so only our thread processes
+	 * it.
+	 */
+	rb_setmask(&mask);
+	ret = pthread_sigmask(SIG_BLOCK, &mask, NULL);
+	if (ret) {
+		errno = ret;
+		PERROR("pthread_sigmask");
+	}
+}