+ if (stream->shm_fd >= 0 && stream->wait_fd != stream->shm_fd) {
+ ret = close(stream->shm_fd);
+ if (ret) {
+ PERROR("close");
+ }
+ }
+
+ /* Check and cleanup relayd */
+ rcu_read_lock();
+ relayd = consumer_find_relayd(stream->net_seq_idx);
+ if (relayd != NULL) {
+ uatomic_dec(&relayd->refcount);
+ assert(uatomic_read(&relayd->refcount) >= 0);
+
+ /* Closing streams requires to lock the control socket. */
+ pthread_mutex_lock(&relayd->ctrl_sock_mutex);
+ ret = relayd_send_close_stream(&relayd->control_sock,
+ stream->relayd_stream_id, stream->next_net_seq_num - 1);
+ pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
+ if (ret < 0) {
+ DBG("Unable to close stream on the relayd. Continuing");
+ /*
+ * Continue here. There is nothing we can do for the relayd.
+ * Chances are that the relayd has closed the socket so we just
+ * continue cleaning up.
+ */
+ }
+
+ /* Both conditions are met, we destroy the relayd. */
+ if (uatomic_read(&relayd->refcount) == 0 &&
+ uatomic_read(&relayd->destroy_flag)) {
+ destroy_relayd(relayd);
+ }
+ }
+ rcu_read_unlock();
+
+ /* Atomically decrement channel refcount since other threads can use it. */
+ uatomic_dec(&stream->chan->refcount);
+ if (!uatomic_read(&stream->chan->refcount)
+ && !uatomic_read(&stream->chan->nb_init_streams)) {
+ /* Go for channel deletion! */
+ free_chan = stream->chan;
+ }
+
+end:
+ pthread_mutex_unlock(&stream->lock);
+ pthread_mutex_unlock(&consumer_data.lock);
+
+ if (free_chan) {
+ consumer_del_channel(free_chan);
+ }
+
+free_stream:
+ call_rcu(&stream->node.head, consumer_free_stream);
+}
+
+/*
+ * Action done with the metadata stream when adding it to the consumer internal
+ * data structures to handle it.
+ */
+static int consumer_add_metadata_stream(struct lttng_consumer_stream *stream,
+ struct lttng_ht *ht)
+{
+ int ret = 0;
+ struct consumer_relayd_sock_pair *relayd;
+ struct lttng_ht_iter iter;
+ struct lttng_ht_node_ulong *node;
+
+ assert(stream);
+ assert(ht);
+
+ DBG3("Adding metadata stream %d to hash table", stream->wait_fd);
+
+ pthread_mutex_lock(&consumer_data.lock);
+ pthread_mutex_lock(&stream->lock);
+
+ /*
+ * From here, refcounts are updated so be _careful_ when returning an error
+ * after this point.
+ */
+
+ rcu_read_lock();
+
+ /*
+ * Lookup the stream just to make sure it does not exist in our internal
+ * state. This should NEVER happen.
+ */
+ lttng_ht_lookup(ht, (void *)((unsigned long) stream->wait_fd), &iter);
+ node = lttng_ht_iter_get_node_ulong(&iter);
+ assert(!node);
+
+ /* Find relayd and, if one is found, increment refcount. */
+ relayd = consumer_find_relayd(stream->net_seq_idx);
+ if (relayd != NULL) {
+ uatomic_inc(&relayd->refcount);
+ }
+
+ /* Update channel refcount once added without error(s). */
+ uatomic_inc(&stream->chan->refcount);
+
+ /*
+ * When nb_init_streams reaches 0, we don't need to trigger any action in
+ * terms of destroying the associated channel, because the action that
+ * causes the count to become 0 also causes a stream to be added. The
+ * channel deletion will thus be triggered by the following removal of this
+ * stream.
+ */
+ if (uatomic_read(&stream->chan->nb_init_streams) > 0) {
+ uatomic_dec(&stream->chan->nb_init_streams);
+ }
+
+ lttng_ht_add_unique_ulong(ht, &stream->node);
+
+ /*
+ * Add stream to the stream_list_ht of the consumer data. No need to steal
+ * the key since the HT does not use it and we allow to add redundant keys
+ * into this table.
+ */
+ lttng_ht_add_ulong(consumer_data.stream_list_ht, &stream->node_session_id);
+
+ rcu_read_unlock();
+
+ pthread_mutex_unlock(&stream->lock);
+ pthread_mutex_unlock(&consumer_data.lock);
+ return ret;
+}
+
+/*
+ * Delete data stream that are flagged for deletion (endpoint_status).
+ */
+static void validate_endpoint_status_data_stream(void)
+{
+ struct lttng_ht_iter iter;
+ struct lttng_consumer_stream *stream;
+
+ DBG("Consumer delete flagged data stream");
+
+ rcu_read_lock();
+ cds_lfht_for_each_entry(data_ht->ht, &iter.iter, stream, node.node) {
+ /* Validate delete flag of the stream */
+ if (stream->endpoint_status == CONSUMER_ENDPOINT_ACTIVE) {
+ continue;
+ }
+ /* Delete it right now */
+ consumer_del_stream(stream, data_ht);
+ }
+ rcu_read_unlock();
+}
+
+/*
+ * Delete metadata stream that are flagged for deletion (endpoint_status).
+ */
+static void validate_endpoint_status_metadata_stream(
+ struct lttng_poll_event *pollset)
+{
+ struct lttng_ht_iter iter;
+ struct lttng_consumer_stream *stream;
+
+ DBG("Consumer delete flagged metadata stream");
+
+ assert(pollset);
+
+ rcu_read_lock();
+ cds_lfht_for_each_entry(metadata_ht->ht, &iter.iter, stream, node.node) {
+ /* Validate delete flag of the stream */
+ if (stream->endpoint_status == CONSUMER_ENDPOINT_ACTIVE) {
+ continue;
+ }
+ /*
+ * Remove from pollset so the metadata thread can continue without
+ * blocking on a deleted stream.
+ */
+ lttng_poll_del(pollset, stream->wait_fd);
+
+ /* Delete it right now */
+ consumer_del_metadata_stream(stream, metadata_ht);
+ }
+ rcu_read_unlock();
+}
+
+/*
+ * Thread polls on metadata file descriptor and write them on disk or on the
+ * network.
+ */
+void *consumer_thread_metadata_poll(void *data)
+{
+ int ret, i, pollfd;
+ uint32_t revents, nb_fd;
+ struct lttng_consumer_stream *stream = NULL;
+ struct lttng_ht_iter iter;
+ struct lttng_ht_node_ulong *node;
+ struct lttng_poll_event events;
+ struct lttng_consumer_local_data *ctx = data;
+ ssize_t len;
+
+ rcu_register_thread();
+
+ metadata_ht = lttng_ht_new(0, LTTNG_HT_TYPE_ULONG);
+ if (!metadata_ht) {
+ /* ENOMEM at this point. Better to bail out. */
+ goto error;
+ }
+
+ DBG("Thread metadata poll started");
+
+ /* Size is set to 1 for the consumer_metadata pipe */
+ ret = lttng_poll_create(&events, 2, LTTNG_CLOEXEC);
+ if (ret < 0) {
+ ERR("Poll set creation failed");
+ goto end;
+ }
+
+ ret = lttng_poll_add(&events, ctx->consumer_metadata_pipe[0], LPOLLIN);
+ if (ret < 0) {
+ goto end;
+ }
+
+ /* Main loop */
+ DBG("Metadata main loop started");
+
+ while (1) {
+ /* Only the metadata pipe is set */
+ if (LTTNG_POLL_GETNB(&events) == 0 && consumer_quit == 1) {
+ goto end;
+ }
+
+restart:
+ DBG("Metadata poll wait with %d fd(s)", LTTNG_POLL_GETNB(&events));
+ ret = lttng_poll_wait(&events, -1);
+ DBG("Metadata event catched in thread");
+ if (ret < 0) {
+ if (errno == EINTR) {
+ ERR("Poll EINTR catched");
+ goto restart;
+ }
+ goto error;
+ }
+
+ nb_fd = ret;
+
+ /* From here, the event is a metadata wait fd */
+ for (i = 0; i < nb_fd; i++) {
+ revents = LTTNG_POLL_GETEV(&events, i);
+ pollfd = LTTNG_POLL_GETFD(&events, i);
+
+ /* Just don't waste time if no returned events for the fd */
+ if (!revents) {
+ continue;
+ }
+
+ if (pollfd == ctx->consumer_metadata_pipe[0]) {
+ if (revents & (LPOLLERR | LPOLLHUP )) {
+ DBG("Metadata thread pipe hung up");
+ /*
+ * Remove the pipe from the poll set and continue the loop
+ * since their might be data to consume.
+ */
+ lttng_poll_del(&events, ctx->consumer_metadata_pipe[0]);
+ ret = close(ctx->consumer_metadata_pipe[0]);
+ if (ret < 0) {
+ PERROR("close metadata pipe");
+ }
+ continue;
+ } else if (revents & LPOLLIN) {
+ do {
+ /* Get the stream pointer received */
+ ret = read(pollfd, &stream, sizeof(stream));
+ } while (ret < 0 && errno == EINTR);
+ if (ret < 0 ||
+ ret < sizeof(struct lttng_consumer_stream *)) {
+ PERROR("read metadata stream");
+ /*
+ * Let's continue here and hope we can still work
+ * without stopping the consumer. XXX: Should we?
+ */
+ continue;
+ }
+
+ /* A NULL stream means that the state has changed. */
+ if (stream == NULL) {
+ /* Check for deleted streams. */
+ validate_endpoint_status_metadata_stream(&events);
+ continue;
+ }
+
+ DBG("Adding metadata stream %d to poll set",
+ stream->wait_fd);
+
+ ret = consumer_add_metadata_stream(stream, metadata_ht);
+ if (ret) {
+ ERR("Unable to add metadata stream");
+ /* Stream was not setup properly. Continuing. */
+ consumer_del_metadata_stream(stream, NULL);
+ continue;
+ }
+
+ /* Add metadata stream to the global poll events list */
+ lttng_poll_add(&events, stream->wait_fd,
+ LPOLLIN | LPOLLPRI);
+ }
+
+ /* Handle other stream */
+ continue;
+ }
+
+ rcu_read_lock();
+ lttng_ht_lookup(metadata_ht, (void *)((unsigned long) pollfd),
+ &iter);
+ node = lttng_ht_iter_get_node_ulong(&iter);
+ assert(node);
+
+ stream = caa_container_of(node, struct lttng_consumer_stream,
+ node);
+
+ /* Check for error event */
+ if (revents & (LPOLLERR | LPOLLHUP)) {
+ DBG("Metadata fd %d is hup|err.", pollfd);
+ if (!stream->hangup_flush_done
+ && (consumer_data.type == LTTNG_CONSUMER32_UST
+ || consumer_data.type == LTTNG_CONSUMER64_UST)) {
+ DBG("Attempting to flush and consume the UST buffers");
+ lttng_ustconsumer_on_stream_hangup(stream);
+
+ /* We just flushed the stream now read it. */
+ do {
+ len = ctx->on_buffer_ready(stream, ctx);
+ /*
+ * We don't check the return value here since if we get
+ * a negative len, it means an error occured thus we
+ * simply remove it from the poll set and free the
+ * stream.
+ */
+ } while (len > 0);
+ }
+
+ lttng_poll_del(&events, stream->wait_fd);
+ /*
+ * This call update the channel states, closes file descriptors
+ * and securely free the stream.
+ */
+ consumer_del_metadata_stream(stream, metadata_ht);
+ } else if (revents & (LPOLLIN | LPOLLPRI)) {
+ /* Get the data out of the metadata file descriptor */
+ DBG("Metadata available on fd %d", pollfd);
+ assert(stream->wait_fd == pollfd);
+
+ len = ctx->on_buffer_ready(stream, ctx);
+ /* It's ok to have an unavailable sub-buffer */
+ if (len < 0 && len != -EAGAIN && len != -ENODATA) {
+ /* Clean up stream from consumer and free it. */
+ lttng_poll_del(&events, stream->wait_fd);
+ consumer_del_metadata_stream(stream, metadata_ht);
+ } else if (len > 0) {
+ stream->data_read = 1;
+ }
+ }
+
+ /* Release RCU lock for the stream looked up */
+ rcu_read_unlock();
+ }
+ }
+
+error:
+end:
+ DBG("Metadata poll thread exiting");
+ lttng_poll_clean(&events);
+
+ destroy_stream_ht(metadata_ht);
+
+ rcu_unregister_thread();
+ return NULL;
+}
+
+/*
+ * This thread polls the fds in the set to consume the data and write
+ * it to tracefile if necessary.
+ */
+void *consumer_thread_data_poll(void *data)
+{
+ int num_rdy, num_hup, high_prio, ret, i;
+ struct pollfd *pollfd = NULL;
+ /* local view of the streams */
+ struct lttng_consumer_stream **local_stream = NULL, *new_stream = NULL;
+ /* local view of consumer_data.fds_count */
+ int nb_fd = 0;
+ struct lttng_consumer_local_data *ctx = data;
+ ssize_t len;
+
+ rcu_register_thread();
+
+ data_ht = lttng_ht_new(0, LTTNG_HT_TYPE_ULONG);
+ if (data_ht == NULL) {
+ /* ENOMEM at this point. Better to bail out. */
+ goto end;
+ }
+
+ local_stream = zmalloc(sizeof(struct lttng_consumer_stream));
+
+ while (1) {
+ high_prio = 0;
+ num_hup = 0;
+
+ /*
+ * the fds set has been updated, we need to update our
+ * local array as well
+ */
+ pthread_mutex_lock(&consumer_data.lock);
+ if (consumer_data.need_update) {
+ if (pollfd != NULL) {
+ free(pollfd);
+ pollfd = NULL;
+ }
+ if (local_stream != NULL) {
+ free(local_stream);
+ local_stream = NULL;
+ }
+
+ /* allocate for all fds + 1 for the consumer_data_pipe */
+ pollfd = zmalloc((consumer_data.stream_count + 1) * sizeof(struct pollfd));
+ if (pollfd == NULL) {
+ PERROR("pollfd malloc");
+ pthread_mutex_unlock(&consumer_data.lock);
+ goto end;
+ }
+
+ /* allocate for all fds + 1 for the consumer_data_pipe */
+ local_stream = zmalloc((consumer_data.stream_count + 1) *
+ sizeof(struct lttng_consumer_stream));
+ if (local_stream == NULL) {
+ PERROR("local_stream malloc");
+ pthread_mutex_unlock(&consumer_data.lock);
+ goto end;
+ }
+ ret = consumer_update_poll_array(ctx, &pollfd, local_stream,
+ data_ht);
+ if (ret < 0) {
+ ERR("Error in allocating pollfd or local_outfds");
+ lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_POLL_ERROR);
+ pthread_mutex_unlock(&consumer_data.lock);
+ goto end;
+ }
+ nb_fd = ret;
+ consumer_data.need_update = 0;
+ }
+ pthread_mutex_unlock(&consumer_data.lock);
+
+ /* No FDs and consumer_quit, consumer_cleanup the thread */
+ if (nb_fd == 0 && consumer_quit == 1) {
+ goto end;
+ }
+ /* poll on the array of fds */
+ restart:
+ DBG("polling on %d fd", nb_fd + 1);
+ num_rdy = poll(pollfd, nb_fd + 1, -1);
+ DBG("poll num_rdy : %d", num_rdy);
+ if (num_rdy == -1) {
+ /*
+ * Restart interrupted system call.
+ */
+ if (errno == EINTR) {
+ goto restart;
+ }
+ PERROR("Poll error");
+ lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_POLL_ERROR);
+ goto end;
+ } else if (num_rdy == 0) {
+ DBG("Polling thread timed out");
+ goto end;
+ }
+
+ /*
+ * If the consumer_data_pipe triggered poll go directly to the
+ * beginning of the loop to update the array. We want to prioritize
+ * array update over low-priority reads.
+ */
+ if (pollfd[nb_fd].revents & (POLLIN | POLLPRI)) {
+ size_t pipe_readlen;
+
+ DBG("consumer_data_pipe wake up");
+ /* Consume 1 byte of pipe data */
+ do {
+ pipe_readlen = read(ctx->consumer_data_pipe[0], &new_stream,
+ sizeof(new_stream));
+ } while (pipe_readlen == -1 && errno == EINTR);
+ if (pipe_readlen < 0) {
+ PERROR("read consumer data pipe");
+ /* Continue so we can at least handle the current stream(s). */
+ continue;
+ }
+
+ /*
+ * If the stream is NULL, just ignore it. It's also possible that
+ * the sessiond poll thread changed the consumer_quit state and is
+ * waking us up to test it.
+ */
+ if (new_stream == NULL) {
+ validate_endpoint_status_data_stream();
+ continue;
+ }
+
+ ret = consumer_add_stream(new_stream, data_ht);
+ if (ret) {
+ ERR("Consumer add stream %d failed. Continuing",
+ new_stream->key);
+ /*
+ * At this point, if the add_stream fails, it is not in the
+ * hash table thus passing the NULL value here.
+ */
+ consumer_del_stream(new_stream, NULL);
+ }
+
+ /* Continue to update the local streams and handle prio ones */
+ continue;
+ }
+
+ /* Take care of high priority channels first. */
+ for (i = 0; i < nb_fd; i++) {
+ if (local_stream[i] == NULL) {
+ continue;
+ }
+ if (pollfd[i].revents & POLLPRI) {
+ DBG("Urgent read on fd %d", pollfd[i].fd);
+ high_prio = 1;
+ len = ctx->on_buffer_ready(local_stream[i], ctx);
+ /* it's ok to have an unavailable sub-buffer */
+ if (len < 0 && len != -EAGAIN && len != -ENODATA) {
+ /* Clean the stream and free it. */
+ consumer_del_stream(local_stream[i], data_ht);
+ local_stream[i] = NULL;
+ } else if (len > 0) {
+ local_stream[i]->data_read = 1;
+ }
+ }
+ }
+
+ /*
+ * If we read high prio channel in this loop, try again
+ * for more high prio data.
+ */
+ if (high_prio) {
+ continue;
+ }
+
+ /* Take care of low priority channels. */
+ for (i = 0; i < nb_fd; i++) {
+ if (local_stream[i] == NULL) {
+ continue;
+ }
+ if ((pollfd[i].revents & POLLIN) ||
+ local_stream[i]->hangup_flush_done) {
+ DBG("Normal read on fd %d", pollfd[i].fd);
+ len = ctx->on_buffer_ready(local_stream[i], ctx);
+ /* it's ok to have an unavailable sub-buffer */
+ if (len < 0 && len != -EAGAIN && len != -ENODATA) {
+ /* Clean the stream and free it. */
+ consumer_del_stream(local_stream[i], data_ht);
+ local_stream[i] = NULL;
+ } else if (len > 0) {
+ local_stream[i]->data_read = 1;
+ }
+ }
+ }
+
+ /* Handle hangup and errors */