Fix: format string type mismatch
[lttng-tools.git] / src / common / consumer.c
index 25f891af9abf17b3f9e7beb4ed2e75bec671296e..4b657f33249f7a4f777b7b3ff784b379b15f5878 100644 (file)
@@ -28,6 +28,7 @@
 #include <sys/types.h>
 #include <unistd.h>
 #include <inttypes.h>
+#include <signal.h>
 
 #include <common/common.h>
 #include <common/utils.h>
@@ -40,6 +41,7 @@
 #include <common/ust-consumer/ust-consumer.h>
 
 #include "consumer.h"
+#include "consumer-stream.h"
 
 struct lttng_consumer_global_data consumer_data = {
        .stream_count = 0,
@@ -47,6 +49,18 @@ struct lttng_consumer_global_data consumer_data = {
        .type = LTTNG_CONSUMER_UNKNOWN,
 };
 
+enum consumer_channel_action {
+       CONSUMER_CHANNEL_ADD,
+       CONSUMER_CHANNEL_DEL,
+       CONSUMER_CHANNEL_QUIT,
+};
+
+struct consumer_channel_msg {
+       enum consumer_channel_action action;
+       struct lttng_consumer_channel *chan;    /* add */
+       uint64_t key;                           /* del */
+};
+
 /*
  * Flag to inform the polling thread to quit when all fd hung up. Updated by
  * the consumer_thread_receive_fds when it notices that all fds has hung up.
@@ -64,18 +78,60 @@ static struct lttng_ht *metadata_ht;
 static struct lttng_ht *data_ht;
 
 /*
- * Notify a thread pipe to poll back again. This usually means that some global
- * state has changed so we just send back the thread in a poll wait call.
+ * Notify a thread lttng pipe to poll back again. This usually means that some
+ * global state has changed so we just send back the thread in a poll wait
+ * call.
  */
-static void notify_thread_pipe(int wpipe)
+static void notify_thread_lttng_pipe(struct lttng_pipe *pipe)
+{
+       struct lttng_consumer_stream *null_stream = NULL;
+
+       assert(pipe);
+
+       (void) lttng_pipe_write(pipe, &null_stream, sizeof(null_stream));
+}
+
+static void notify_channel_pipe(struct lttng_consumer_local_data *ctx,
+               struct lttng_consumer_channel *chan,
+               uint64_t key,
+               enum consumer_channel_action action)
 {
+       struct consumer_channel_msg msg;
        int ret;
 
+       memset(&msg, 0, sizeof(msg));
+
+       msg.action = action;
+       msg.chan = chan;
+       msg.key = key;
        do {
-               struct lttng_consumer_stream *null_stream = NULL;
+               ret = write(ctx->consumer_channel_pipe[1], &msg, sizeof(msg));
+       } while (ret < 0 && errno == EINTR);
+}
+
+void notify_thread_del_channel(struct lttng_consumer_local_data *ctx,
+               uint64_t key)
+{
+       notify_channel_pipe(ctx, NULL, key, CONSUMER_CHANNEL_DEL);
+}
 
-               ret = write(wpipe, &null_stream, sizeof(null_stream));
+static int read_channel_pipe(struct lttng_consumer_local_data *ctx,
+               struct lttng_consumer_channel **chan,
+               uint64_t *key,
+               enum consumer_channel_action *action)
+{
+       struct consumer_channel_msg msg;
+       int ret;
+
+       do {
+               ret = read(ctx->consumer_channel_pipe[0], &msg, sizeof(msg));
        } while (ret < 0 && errno == EINTR);
+       if (ret > 0) {
+               *action = msg.action;
+               *chan = msg.chan;
+               *key = msg.key;
+       }
+       return ret;
 }
 
 /*
@@ -109,20 +165,20 @@ static struct lttng_consumer_stream *find_stream(uint64_t key,
        return stream;
 }
 
-static void steal_stream_key(int key, struct lttng_ht *ht)
+static void steal_stream_key(uint64_t key, struct lttng_ht *ht)
 {
        struct lttng_consumer_stream *stream;
 
        rcu_read_lock();
        stream = find_stream(key, ht);
        if (stream) {
-               stream->key = -1ULL;
+               stream->key = (uint64_t) -1ULL;
                /*
                 * We don't want the lookup to match, but we still need
                 * to iterate on this stream when iterating over the hash table. Just
                 * change the node key.
                 */
-               stream->node.key = -1ULL;
+               stream->node.key = (uint64_t) -1ULL;
        }
        rcu_read_unlock();
 }
@@ -199,10 +255,8 @@ static void free_relayd_rcu(struct rcu_head *head)
 
 /*
  * Destroy and free relayd socket pair object.
- *
- * This function MUST be called with the consumer_data lock acquired.
  */
-static void destroy_relayd(struct consumer_relayd_sock_pair *relayd)
+void consumer_destroy_relayd(struct consumer_relayd_sock_pair *relayd)
 {
        int ret;
        struct lttng_ht_iter iter;
@@ -232,10 +286,23 @@ void consumer_del_channel(struct lttng_consumer_channel *channel)
 {
        int ret;
        struct lttng_ht_iter iter;
+       struct lttng_consumer_stream *stream, *stmp;
 
        DBG("Consumer delete channel key %" PRIu64, channel->key);
 
        pthread_mutex_lock(&consumer_data.lock);
+       pthread_mutex_lock(&channel->lock);
+
+       /* Delete streams that might have been left in the stream list. */
+       cds_list_for_each_entry_safe(stream, stmp, &channel->streams.head,
+                       send_node) {
+               cds_list_del(&stream->send_node);
+               /*
+                * Once a stream is added to this list, the buffers were created so
+                * we have a guarantee that this call will succeed.
+                */
+               consumer_stream_destroy(stream, NULL);
+       }
 
        switch (consumer_data.type) {
        case LTTNG_CONSUMER_KERNEL:
@@ -258,6 +325,7 @@ void consumer_del_channel(struct lttng_consumer_channel *channel)
 
        call_rcu(&channel->node.head, free_channel_rcu);
 end:
+       pthread_mutex_unlock(&channel->lock);
        pthread_mutex_unlock(&consumer_data.lock);
 }
 
@@ -274,12 +342,12 @@ static void cleanup_relayd_ht(void)
 
        cds_lfht_for_each_entry(consumer_data.relayd_ht->ht, &iter.iter, relayd,
                        node.node) {
-               destroy_relayd(relayd);
+               consumer_destroy_relayd(relayd);
        }
 
-       lttng_ht_destroy(consumer_data.relayd_ht);
-
        rcu_read_unlock();
+
+       lttng_ht_destroy(consumer_data.relayd_ht);
 }
 
 /*
@@ -289,13 +357,13 @@ static void cleanup_relayd_ht(void)
  * It's atomically set without having the stream mutex locked which is fine
  * because we handle the write/read race with a pipe wakeup for each thread.
  */
-static void update_endpoint_status_by_netidx(int net_seq_idx,
+static void update_endpoint_status_by_netidx(uint64_t net_seq_idx,
                enum consumer_endpoint_status status)
 {
        struct lttng_ht_iter iter;
        struct lttng_consumer_stream *stream;
 
-       DBG("Consumer set delete flag on stream by idx %d", net_seq_idx);
+       DBG("Consumer set delete flag on stream by idx %" PRIu64, net_seq_idx);
 
        rcu_read_lock();
 
@@ -328,7 +396,7 @@ static void update_endpoint_status_by_netidx(int net_seq_idx,
 static void cleanup_relayd(struct consumer_relayd_sock_pair *relayd,
                struct lttng_consumer_local_data *ctx)
 {
-       int netidx;
+       uint64_t netidx;
 
        assert(relayd);
 
@@ -341,7 +409,7 @@ static void cleanup_relayd(struct consumer_relayd_sock_pair *relayd,
         * Delete the relayd from the relayd hash table, close the sockets and free
         * the object in a RCU call.
         */
-       destroy_relayd(relayd);
+       consumer_destroy_relayd(relayd);
 
        /* Set inactive endpoint to all streams */
        update_endpoint_status_by_netidx(netidx, CONSUMER_ENDPOINT_INACTIVE);
@@ -353,8 +421,8 @@ static void cleanup_relayd(struct consumer_relayd_sock_pair *relayd,
         * read of this status which happens AFTER receiving this notify.
         */
        if (ctx) {
-               notify_thread_pipe(ctx->consumer_data_pipe[1]);
-               notify_thread_pipe(ctx->consumer_metadata_pipe[1]);
+               notify_thread_lttng_pipe(ctx->consumer_data_pipe);
+               notify_thread_lttng_pipe(ctx->consumer_metadata_pipe);
        }
 }
 
@@ -373,121 +441,33 @@ void consumer_flag_relayd_for_destroy(struct consumer_relayd_sock_pair *relayd)
 
        /* Destroy the relayd if refcount is 0 */
        if (uatomic_read(&relayd->refcount) == 0) {
-               destroy_relayd(relayd);
+               consumer_destroy_relayd(relayd);
        }
 }
 
 /*
- * Remove a stream from the global list protected by a mutex. This
- * function is also responsible for freeing its data structures.
+ * Completly destroy stream from every visiable data structure and the given
+ * hash table if one.
+ *
+ * One this call returns, the stream object is not longer usable nor visible.
  */
 void consumer_del_stream(struct lttng_consumer_stream *stream,
                struct lttng_ht *ht)
 {
-       int ret;
-       struct lttng_ht_iter iter;
-       struct lttng_consumer_channel *free_chan = NULL;
-       struct consumer_relayd_sock_pair *relayd;
-
-       assert(stream);
-
-       DBG("Consumer del stream %d", stream->wait_fd);
-
-       if (ht == NULL) {
-               /* Means the stream was allocated but not successfully added */
-               goto free_stream_rcu;
-       }
-
-       pthread_mutex_lock(&consumer_data.lock);
-       pthread_mutex_lock(&stream->lock);
-
-       switch (consumer_data.type) {
-       case LTTNG_CONSUMER_KERNEL:
-               if (stream->mmap_base != NULL) {
-                       ret = munmap(stream->mmap_base, stream->mmap_len);
-                       if (ret != 0) {
-                               PERROR("munmap");
-                       }
-               }
-               break;
-       case LTTNG_CONSUMER32_UST:
-       case LTTNG_CONSUMER64_UST:
-               lttng_ustconsumer_del_stream(stream);
-               break;
-       default:
-               ERR("Unknown consumer_data type");
-               assert(0);
-               goto end;
-       }
-
-       rcu_read_lock();
-       iter.iter.node = &stream->node.node;
-       ret = lttng_ht_del(ht, &iter);
-       assert(!ret);
-
-       /* Remove node session id from the consumer_data stream ht */
-       iter.iter.node = &stream->node_session_id.node;
-       ret = lttng_ht_del(consumer_data.stream_list_ht, &iter);
-       assert(!ret);
-       rcu_read_unlock();
-
-       assert(consumer_data.stream_count > 0);
-       consumer_data.stream_count--;
-
-       if (stream->out_fd >= 0) {
-               ret = close(stream->out_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();
-
-       uatomic_dec(&stream->chan->refcount);
-       if (!uatomic_read(&stream->chan->refcount)
-                       && !uatomic_read(&stream->chan->nb_init_stream_left)) {
-               free_chan = stream->chan;
-       }
-
-end:
-       consumer_data.need_update = 1;
-       pthread_mutex_unlock(&stream->lock);
-       pthread_mutex_unlock(&consumer_data.lock);
+       consumer_stream_destroy(stream, ht);
+}
 
-       if (free_chan) {
-               consumer_del_channel(free_chan);
-       }
+/*
+ * XXX naming of del vs destroy is all mixed up.
+ */
+void consumer_del_stream_for_data(struct lttng_consumer_stream *stream)
+{
+       consumer_stream_destroy(stream, data_ht);
+}
 
-free_stream_rcu:
-       call_rcu(&stream->node.head, free_stream_rcu);
+void consumer_del_stream_for_metadata(struct lttng_consumer_stream *stream)
+{
+       consumer_stream_destroy(stream, metadata_ht);
 }
 
 struct lttng_consumer_stream *consumer_allocate_stream(uint64_t channel_key,
@@ -496,11 +476,12 @@ struct lttng_consumer_stream *consumer_allocate_stream(uint64_t channel_key,
                const char *channel_name,
                uid_t uid,
                gid_t gid,
-               int relayd_id,
+               uint64_t relayd_id,
                uint64_t session_id,
                int cpu,
                int *alloc_ret,
-               enum consumer_channel_type type)
+               enum consumer_channel_type type,
+               unsigned int monitor)
 {
        int ret;
        struct lttng_consumer_stream *stream;
@@ -522,6 +503,8 @@ struct lttng_consumer_stream *consumer_allocate_stream(uint64_t channel_key,
        stream->gid = gid;
        stream->net_seq_idx = relayd_id;
        stream->session_id = session_id;
+       stream->monitor = monitor;
+       stream->endpoint_status = CONSUMER_ENDPOINT_ACTIVE;
        pthread_mutex_init(&stream->lock, NULL);
 
        /* If channel is the metadata, flag this stream as metadata. */
@@ -542,11 +525,16 @@ struct lttng_consumer_stream *consumer_allocate_stream(uint64_t channel_key,
        /* Key is always the wait_fd for streams. */
        lttng_ht_node_init_u64(&stream->node, stream->key);
 
+       /* Init node per channel id key */
+       lttng_ht_node_init_u64(&stream->node_channel_id, channel_key);
+
        /* Init session id node with the stream session id */
        lttng_ht_node_init_u64(&stream->node_session_id, stream->session_id);
 
-       DBG3("Allocated stream %s (key %" PRIu64 ", relayd_id %" PRIu64 ", session_id %" PRIu64,
-                       stream->name, stream->key, stream->net_seq_idx, stream->session_id);
+       DBG3("Allocated stream %s (key %" PRIu64 ", chan_key %" PRIu64
+                       " relayd_id %" PRIu64 ", session_id %" PRIu64,
+                       stream->name, stream->key, channel_key,
+                       stream->net_seq_idx, stream->session_id);
 
        rcu_read_unlock();
        return stream;
@@ -564,11 +552,10 @@ end:
 /*
  * Add a stream to the global list protected by a mutex.
  */
-static int add_stream(struct lttng_consumer_stream *stream,
-               struct lttng_ht *ht)
+int consumer_add_data_stream(struct lttng_consumer_stream *stream)
 {
+       struct lttng_ht *ht = data_ht;
        int ret = 0;
-       struct consumer_relayd_sock_pair *relayd;
 
        assert(stream);
        assert(ht);
@@ -576,6 +563,8 @@ static int add_stream(struct lttng_consumer_stream *stream,
        DBG3("Adding consumer stream %" PRIu64, stream->key);
 
        pthread_mutex_lock(&consumer_data.lock);
+       pthread_mutex_lock(&stream->chan->lock);
+       pthread_mutex_lock(&stream->chan->timer_lock);
        pthread_mutex_lock(&stream->lock);
        rcu_read_lock();
 
@@ -584,6 +573,9 @@ static int add_stream(struct lttng_consumer_stream *stream,
 
        lttng_ht_add_unique_u64(ht, &stream->node);
 
+       lttng_ht_add_u64(consumer_data.stream_per_chan_id_ht,
+                       &stream->node_channel_id);
+
        /*
         * 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
@@ -591,15 +583,6 @@ static int add_stream(struct lttng_consumer_stream *stream,
         */
        lttng_ht_add_u64(consumer_data.stream_list_ht, &stream->node_session_id);
 
-       /* Check and cleanup relayd */
-       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_stream_left reaches 0, we don't need to trigger any action
         * in terms of destroying the associated channel, because the action that
@@ -608,6 +591,8 @@ static int add_stream(struct lttng_consumer_stream *stream,
         * stream.
         */
        if (uatomic_read(&stream->chan->nb_init_stream_left) > 0) {
+               /* Increment refcount before decrementing nb_init_stream_left */
+               cmm_smp_wmb();
                uatomic_dec(&stream->chan->nb_init_stream_left);
        }
 
@@ -617,11 +602,18 @@ static int add_stream(struct lttng_consumer_stream *stream,
 
        rcu_read_unlock();
        pthread_mutex_unlock(&stream->lock);
+       pthread_mutex_unlock(&stream->chan->timer_lock);
+       pthread_mutex_unlock(&stream->chan->lock);
        pthread_mutex_unlock(&consumer_data.lock);
 
        return ret;
 }
 
+void consumer_del_data_stream(struct lttng_consumer_stream *stream)
+{
+       consumer_del_stream(stream, data_ht);
+}
+
 /*
  * Add relayd socket to global consumer data hashtable. RCU read side lock MUST
  * be acquired before calling this.
@@ -650,12 +642,12 @@ end:
  * Allocate and return a consumer relayd socket.
  */
 struct consumer_relayd_sock_pair *consumer_allocate_relayd_sock_pair(
-               int net_seq_idx)
+               uint64_t net_seq_idx)
 {
        struct consumer_relayd_sock_pair *obj = NULL;
 
-       /* Negative net sequence index is a failure */
-       if (net_seq_idx < 0) {
+       /* net sequence index of -1 is a failure */
+       if (net_seq_idx == (uint64_t) -1ULL) {
                goto error;
        }
 
@@ -668,6 +660,8 @@ struct consumer_relayd_sock_pair *consumer_allocate_relayd_sock_pair(
        obj->net_seq_idx = net_seq_idx;
        obj->refcount = 0;
        obj->destroy_flag = 0;
+       obj->control_sock.sock.fd = -1;
+       obj->data_sock.sock.fd = -1;
        lttng_ht_node_init_u64(&obj->node, obj->net_seq_idx);
        pthread_mutex_init(&obj->ctrl_sock_mutex, NULL);
 
@@ -704,6 +698,67 @@ error:
        return relayd;
 }
 
+/*
+ * Find a relayd and send the stream
+ *
+ * Returns 0 on success, < 0 on error
+ */
+int consumer_send_relayd_stream(struct lttng_consumer_stream *stream,
+               char *path)
+{
+       int ret = 0;
+       struct consumer_relayd_sock_pair *relayd;
+
+       assert(stream);
+       assert(stream->net_seq_idx != -1ULL);
+       assert(path);
+
+       /* The stream is not metadata. Get relayd reference if exists. */
+       rcu_read_lock();
+       relayd = consumer_find_relayd(stream->net_seq_idx);
+       if (relayd != NULL) {
+               /* Add stream on the relayd */
+               pthread_mutex_lock(&relayd->ctrl_sock_mutex);
+               ret = relayd_add_stream(&relayd->control_sock, stream->name,
+                               path, &stream->relayd_stream_id,
+                               stream->chan->tracefile_size, stream->chan->tracefile_count);
+               pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
+               if (ret < 0) {
+                       goto end;
+               }
+               uatomic_inc(&relayd->refcount);
+               stream->sent_to_relayd = 1;
+       } else {
+               ERR("Stream %" PRIu64 " relayd ID %" PRIu64 " unknown. Can't send it.",
+                               stream->key, stream->net_seq_idx);
+               ret = -1;
+               goto end;
+       }
+
+       DBG("Stream %s with key %" PRIu64 " sent to relayd id %" PRIu64,
+                       stream->name, stream->key, stream->net_seq_idx);
+
+end:
+       rcu_read_unlock();
+       return ret;
+}
+
+/*
+ * Find a relayd and close the stream
+ */
+void close_relayd_stream(struct lttng_consumer_stream *stream)
+{
+       struct consumer_relayd_sock_pair *relayd;
+
+       /* The stream is not metadata. Get relayd reference if exists. */
+       rcu_read_lock();
+       relayd = consumer_find_relayd(stream->net_seq_idx);
+       if (relayd) {
+               consumer_stream_relayd_close(stream, relayd);
+       }
+       rcu_read_unlock();
+}
+
 /*
  * Handle stream for relayd transmission if the stream applies for network
  * streaming where the net sequence index is set.
@@ -732,7 +787,7 @@ static int write_relayd_stream_header(struct lttng_consumer_stream *stream,
                }
 
                /* Metadata are always sent on the control socket. */
-               outfd = relayd->control_sock.fd;
+               outfd = relayd->control_sock.sock.fd;
        } else {
                /* Set header with stream information */
                data_hdr.stream_id = htobe64(stream->relayd_stream_id);
@@ -757,7 +812,7 @@ static int write_relayd_stream_header(struct lttng_consumer_stream *stream,
                ++stream->next_net_seq_num;
 
                /* Set to go on data socket */
-               outfd = relayd->data_sock.fd;
+               outfd = relayd->data_sock.sock.fd;
        }
 
 error:
@@ -770,14 +825,18 @@ error:
  *
  * On error, return NULL.
  */
-struct lttng_consumer_channel *consumer_allocate_channel(unsigned long key,
+struct lttng_consumer_channel *consumer_allocate_channel(uint64_t key,
                uint64_t session_id,
                const char *pathname,
                const char *name,
                uid_t uid,
                gid_t gid,
-               int relayd_id,
-               enum lttng_event_output output)
+               uint64_t relayd_id,
+               enum lttng_event_output output,
+               uint64_t tracefile_size,
+               uint64_t tracefile_count,
+               uint64_t session_id_per_pid,
+               unsigned int monitor)
 {
        struct lttng_consumer_channel *channel;
 
@@ -790,10 +849,29 @@ struct lttng_consumer_channel *consumer_allocate_channel(unsigned long key,
        channel->key = key;
        channel->refcount = 0;
        channel->session_id = session_id;
+       channel->session_id_per_pid = session_id_per_pid;
        channel->uid = uid;
        channel->gid = gid;
        channel->relayd_id = relayd_id;
        channel->output = output;
+       channel->tracefile_size = tracefile_size;
+       channel->tracefile_count = tracefile_count;
+       channel->monitor = monitor;
+       pthread_mutex_init(&channel->lock, NULL);
+       pthread_mutex_init(&channel->timer_lock, NULL);
+
+       /*
+        * In monitor mode, the streams associated with the channel will be put in
+        * a special list ONLY owned by this channel. So, the refcount is set to 1
+        * here meaning that the channel itself has streams that are referenced.
+        *
+        * On a channel deletion, once the channel is no longer visible, the
+        * refcount is decremented and checked for a zero value to delete it. With
+        * streams in no monitor mode, it will now be safe to destroy the channel.
+        */
+       if (!channel->monitor) {
+               channel->refcount = 1;
+       }
 
        strncpy(channel->pathname, pathname, sizeof(channel->pathname));
        channel->pathname[sizeof(channel->pathname) - 1] = '\0';
@@ -802,6 +880,9 @@ struct lttng_consumer_channel *consumer_allocate_channel(unsigned long key,
        channel->name[sizeof(channel->name) - 1] = '\0';
 
        lttng_ht_node_init_u64(&channel->node, channel->key);
+
+       channel->wait_fd = -1;
+
        CDS_INIT_LIST_HEAD(&channel->streams.head);
 
        DBG("Allocated channel (key %" PRIu64 ")", channel->key)
@@ -812,24 +893,28 @@ end:
 
 /*
  * Add a channel to the global list protected by a mutex.
+ *
+ * On success 0 is returned else a negative value.
  */
-int consumer_add_channel(struct lttng_consumer_channel *channel)
+int consumer_add_channel(struct lttng_consumer_channel *channel,
+               struct lttng_consumer_local_data *ctx)
 {
        int ret = 0;
        struct lttng_ht_node_u64 *node;
        struct lttng_ht_iter iter;
 
        pthread_mutex_lock(&consumer_data.lock);
+       pthread_mutex_lock(&channel->lock);
+       pthread_mutex_lock(&channel->timer_lock);
        rcu_read_lock();
 
-       lttng_ht_lookup(consumer_data.channel_ht,
-                       &channel->key, &iter);
+       lttng_ht_lookup(consumer_data.channel_ht, &channel->key, &iter);
        node = lttng_ht_iter_get_node_u64(&iter);
        if (node != NULL) {
                /* Channel already exist. Ignore the insertion */
                ERR("Consumer add channel key %" PRIu64 " already exists!",
                        channel->key);
-               ret = -1;
+               ret = -EEXIST;
                goto end;
        }
 
@@ -837,8 +922,14 @@ int consumer_add_channel(struct lttng_consumer_channel *channel)
 
 end:
        rcu_read_unlock();
+       pthread_mutex_unlock(&channel->timer_lock);
+       pthread_mutex_unlock(&channel->lock);
        pthread_mutex_unlock(&consumer_data.lock);
 
+       if (!ret && channel->wait_fd != -1 &&
+                       channel->type == CONSUMER_CHANNEL_TYPE_DATA) {
+               notify_channel_pipe(ctx, channel, -1, CONSUMER_CHANNEL_ADD);
+       }
        return ret;
 }
 
@@ -878,7 +969,12 @@ static int update_poll_array(struct lttng_consumer_local_data *ctx,
                                stream->endpoint_status == CONSUMER_ENDPOINT_INACTIVE) {
                        continue;
                }
-               DBG("Active FD %d", stream->wait_fd);
+               /*
+                * This clobbers way too much the debug output. Uncomment that if you
+                * need it for debugging purposes.
+                *
+                * DBG("Active FD %d", stream->wait_fd);
+                */
                (*pollfd)[i].fd = stream->wait_fd;
                (*pollfd)[i].events = POLLIN | POLLPRI;
                local_stream[i] = stream;
@@ -890,7 +986,7 @@ static int update_poll_array(struct lttng_consumer_local_data *ctx,
         * Insert the consumer_data_pipe at the end of the array and don't
         * increment i so nb_fd is the number of real FD.
         */
-       (*pollfd)[i].fd = ctx->consumer_data_pipe[0];
+       (*pollfd)[i].fd = lttng_pipe_get_readfd(ctx->consumer_data_pipe);
        (*pollfd)[i].events = POLLIN | POLLPRI;
        return i;
 }
@@ -979,6 +1075,8 @@ void lttng_consumer_cleanup(void)
 
        cleanup_relayd_ht();
 
+       lttng_ht_destroy(consumer_data.stream_per_chan_id_ht);
+
        /*
         * This HT contains streams that are freed by either the metadata thread or
         * the data thread so we do *nothing* on the hash table and simply destroy
@@ -1061,9 +1159,9 @@ struct lttng_consumer_local_data *lttng_consumer_create(
                        struct lttng_consumer_local_data *ctx),
                int (*recv_channel)(struct lttng_consumer_channel *channel),
                int (*recv_stream)(struct lttng_consumer_stream *stream),
-               int (*update_stream)(int stream_key, uint32_t state))
+               int (*update_stream)(uint64_t stream_key, uint32_t state))
 {
-       int ret, i;
+       int ret;
        struct lttng_consumer_local_data *ctx;
 
        assert(consumer_data.type == LTTNG_CONSUMER_UNKNOWN ||
@@ -1077,32 +1175,19 @@ struct lttng_consumer_local_data *lttng_consumer_create(
        }
 
        ctx->consumer_error_socket = -1;
+       ctx->consumer_metadata_socket = -1;
+       pthread_mutex_init(&ctx->metadata_socket_lock, NULL);
        /* assign the callbacks */
        ctx->on_buffer_ready = buffer_ready;
        ctx->on_recv_channel = recv_channel;
        ctx->on_recv_stream = recv_stream;
        ctx->on_update_stream = update_stream;
 
-       ret = pipe(ctx->consumer_data_pipe);
-       if (ret < 0) {
-               PERROR("Error creating poll pipe");
+       ctx->consumer_data_pipe = lttng_pipe_open(0);
+       if (!ctx->consumer_data_pipe) {
                goto error_poll_pipe;
        }
 
-       /* set read end of the pipe to non-blocking */
-       ret = fcntl(ctx->consumer_data_pipe[0], F_SETFL, O_NONBLOCK);
-       if (ret < 0) {
-               PERROR("fcntl O_NONBLOCK");
-               goto error_poll_fcntl;
-       }
-
-       /* set write end of the pipe to non-blocking */
-       ret = fcntl(ctx->consumer_data_pipe[1], F_SETFL, O_NONBLOCK);
-       if (ret < 0) {
-               PERROR("fcntl O_NONBLOCK");
-               goto error_poll_fcntl;
-       }
-
        ret = pipe(ctx->consumer_should_quit);
        if (ret < 0) {
                PERROR("Error creating recv pipe");
@@ -1115,8 +1200,14 @@ struct lttng_consumer_local_data *lttng_consumer_create(
                goto error_thread_pipe;
        }
 
-       ret = utils_create_pipe(ctx->consumer_metadata_pipe);
+       ret = pipe(ctx->consumer_channel_pipe);
        if (ret < 0) {
+               PERROR("Error creating channel pipe");
+               goto error_channel_pipe;
+       }
+
+       ctx->consumer_metadata_pipe = lttng_pipe_open(0);
+       if (!ctx->consumer_metadata_pipe) {
                goto error_metadata_pipe;
        }
 
@@ -1128,28 +1219,15 @@ struct lttng_consumer_local_data *lttng_consumer_create(
        return ctx;
 
 error_splice_pipe:
-       utils_close_pipe(ctx->consumer_metadata_pipe);
+       lttng_pipe_destroy(ctx->consumer_metadata_pipe);
 error_metadata_pipe:
+       utils_close_pipe(ctx->consumer_channel_pipe);
+error_channel_pipe:
        utils_close_pipe(ctx->consumer_thread_pipe);
 error_thread_pipe:
-       for (i = 0; i < 2; i++) {
-               int err;
-
-               err = close(ctx->consumer_should_quit[i]);
-               if (err) {
-                       PERROR("close");
-               }
-       }
-error_poll_fcntl:
+       utils_close_pipe(ctx->consumer_should_quit);
 error_quit_pipe:
-       for (i = 0; i < 2; i++) {
-               int err;
-
-               err = close(ctx->consumer_data_pipe[i]);
-               if (err) {
-                       PERROR("close");
-               }
-       }
+       lttng_pipe_destroy(ctx->consumer_data_pipe);
 error_poll_pipe:
        free(ctx);
 error:
@@ -1169,30 +1247,15 @@ void lttng_consumer_destroy(struct lttng_consumer_local_data *ctx)
        if (ret) {
                PERROR("close");
        }
-       ret = close(ctx->consumer_thread_pipe[0]);
-       if (ret) {
-               PERROR("close");
-       }
-       ret = close(ctx->consumer_thread_pipe[1]);
-       if (ret) {
-               PERROR("close");
-       }
-       ret = close(ctx->consumer_data_pipe[0]);
-       if (ret) {
-               PERROR("close");
-       }
-       ret = close(ctx->consumer_data_pipe[1]);
-       if (ret) {
-               PERROR("close");
-       }
-       ret = close(ctx->consumer_should_quit[0]);
-       if (ret) {
-               PERROR("close");
-       }
-       ret = close(ctx->consumer_should_quit[1]);
+       ret = close(ctx->consumer_metadata_socket);
        if (ret) {
                PERROR("close");
        }
+       utils_close_pipe(ctx->consumer_thread_pipe);
+       utils_close_pipe(ctx->consumer_channel_pipe);
+       lttng_pipe_destroy(ctx->consumer_data_pipe);
+       lttng_pipe_destroy(ctx->consumer_metadata_pipe);
+       utils_close_pipe(ctx->consumer_should_quit);
        utils_close_pipe(ctx->consumer_splice_metadata_pipe);
 
        unlink(ctx->consumer_command_sock_path);
@@ -1268,7 +1331,7 @@ ssize_t lttng_consumer_on_read_subbuffer_mmap(
        rcu_read_lock();
 
        /* Flag that the current stream if set for network streaming. */
-       if (stream->net_seq_idx != -1) {
+       if (stream->net_seq_idx != (uint64_t) -1ULL) {
                relayd = consumer_find_relayd(stream->net_seq_idx);
                if (relayd == NULL) {
                        goto end;
@@ -1290,6 +1353,7 @@ ssize_t lttng_consumer_on_read_subbuffer_mmap(
                        goto end;
                }
                ret = lttng_ustctl_get_mmap_read_offset(stream, &mmap_offset);
+
                break;
        default:
                ERR("Unknown consumer_data type");
@@ -1345,6 +1409,26 @@ ssize_t lttng_consumer_on_read_subbuffer_mmap(
        } else {
                /* No streaming, we have to set the len with the full padding */
                len += padding;
+
+               /*
+                * Check if we need to change the tracefile before writing the packet.
+                */
+               if (stream->chan->tracefile_size > 0 &&
+                               (stream->tracefile_size_current + len) >
+                               stream->chan->tracefile_size) {
+                       ret = utils_rotate_stream_file(stream->chan->pathname,
+                                       stream->name, stream->chan->tracefile_size,
+                                       stream->chan->tracefile_count, stream->uid, stream->gid,
+                                       stream->out_fd, &(stream->tracefile_count_current));
+                       if (ret < 0) {
+                               ERR("Rotating output file");
+                               goto end;
+                       }
+                       outfd = stream->out_fd = ret;
+                       /* Reset current size because we just perform a rotation. */
+                       stream->tracefile_size_current = 0;
+               }
+               stream->tracefile_size_current += len;
        }
 
        while (len > 0) {
@@ -1446,7 +1530,7 @@ ssize_t lttng_consumer_on_read_subbuffer_splice(
        rcu_read_lock();
 
        /* Flag that the current stream if set for network streaming. */
-       if (stream->net_seq_idx != -1) {
+       if (stream->net_seq_idx != (uint64_t) -1ULL) {
                relayd = consumer_find_relayd(stream->net_seq_idx);
                if (relayd == NULL) {
                        goto end;
@@ -1507,6 +1591,26 @@ ssize_t lttng_consumer_on_read_subbuffer_splice(
        } else {
                /* No streaming, we have to set the len with the full padding */
                len += padding;
+
+               /*
+                * Check if we need to change the tracefile before writing the packet.
+                */
+               if (stream->chan->tracefile_size > 0 &&
+                               (stream->tracefile_size_current + len) >
+                               stream->chan->tracefile_size) {
+                       ret = utils_rotate_stream_file(stream->chan->pathname,
+                                       stream->name, stream->chan->tracefile_size,
+                                       stream->chan->tracefile_count, stream->uid, stream->gid,
+                                       stream->out_fd, &(stream->tracefile_count_current));
+                       if (ret < 0) {
+                               ERR("Rotating output file");
+                               goto end;
+                       }
+                       outfd = stream->out_fd = ret;
+                       /* Reset current size because we just perform a rotation. */
+                       stream->tracefile_size_current = 0;
+               }
+               stream->tracefile_size_current += len;
        }
 
        while (len > 0) {
@@ -1780,6 +1884,7 @@ void consumer_del_metadata_stream(struct lttng_consumer_stream *stream,
        }
 
        pthread_mutex_lock(&consumer_data.lock);
+       pthread_mutex_lock(&stream->chan->lock);
        pthread_mutex_lock(&stream->lock);
 
        switch (consumer_data.type) {
@@ -1790,9 +1895,22 @@ void consumer_del_metadata_stream(struct lttng_consumer_stream *stream,
                                PERROR("munmap metadata stream");
                        }
                }
+               if (stream->wait_fd >= 0) {
+                       ret = close(stream->wait_fd);
+                       if (ret < 0) {
+                               PERROR("close kernel metadata wait_fd");
+                       }
+               }
                break;
        case LTTNG_CONSUMER32_UST:
        case LTTNG_CONSUMER64_UST:
+               if (stream->monitor) {
+                       /* close the write-side in close_metadata */
+                       ret = close(stream->ust_metadata_poll_pipe[0]);
+                       if (ret < 0) {
+                               PERROR("Close UST metadata read-side poll pipe");
+                       }
+               }
                lttng_ustconsumer_del_stream(stream);
                break;
        default:
@@ -1806,7 +1924,10 @@ void consumer_del_metadata_stream(struct lttng_consumer_stream *stream,
        ret = lttng_ht_del(ht, &iter);
        assert(!ret);
 
-       /* Remove node session id from the consumer_data stream ht */
+       iter.iter.node = &stream->node_channel_id.node;
+       ret = lttng_ht_del(consumer_data.stream_per_chan_id_ht, &iter);
+       assert(!ret);
+
        iter.iter.node = &stream->node_session_id.node;
        ret = lttng_ht_del(consumer_data.stream_list_ht, &iter);
        assert(!ret);
@@ -1843,21 +1964,28 @@ void consumer_del_metadata_stream(struct lttng_consumer_stream *stream,
                /* Both conditions are met, we destroy the relayd. */
                if (uatomic_read(&relayd->refcount) == 0 &&
                                uatomic_read(&relayd->destroy_flag)) {
-                       destroy_relayd(relayd);
+                       consumer_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)
+       if (!uatomic_sub_return(&stream->chan->refcount, 1)
                        && !uatomic_read(&stream->chan->nb_init_stream_left)) {
                /* Go for channel deletion! */
                free_chan = stream->chan;
        }
 
 end:
+       /*
+        * Nullify the stream reference so it is not used after deletion. The
+        * channel lock MUST be acquired before being able to check for
+        * a NULL pointer value.
+        */
+       stream->chan->metadata_stream = NULL;
+
        pthread_mutex_unlock(&stream->lock);
+       pthread_mutex_unlock(&stream->chan->lock);
        pthread_mutex_unlock(&consumer_data.lock);
 
        if (free_chan) {
@@ -1872,11 +2000,10 @@ free_stream_rcu:
  * Action done with the metadata stream when adding it to the consumer internal
  * data structures to handle it.
  */
-static int add_metadata_stream(struct lttng_consumer_stream *stream,
-               struct lttng_ht *ht)
+int consumer_add_metadata_stream(struct lttng_consumer_stream *stream)
 {
+       struct lttng_ht *ht = metadata_ht;
        int ret = 0;
-       struct consumer_relayd_sock_pair *relayd;
        struct lttng_ht_iter iter;
        struct lttng_ht_node_u64 *node;
 
@@ -1886,6 +2013,8 @@ static int add_metadata_stream(struct lttng_consumer_stream *stream,
        DBG3("Adding metadata stream %" PRIu64 " to hash table", stream->key);
 
        pthread_mutex_lock(&consumer_data.lock);
+       pthread_mutex_lock(&stream->chan->lock);
+       pthread_mutex_lock(&stream->chan->timer_lock);
        pthread_mutex_lock(&stream->lock);
 
        /*
@@ -1903,15 +2032,6 @@ static int add_metadata_stream(struct lttng_consumer_stream *stream,
        node = lttng_ht_iter_get_node_u64(&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_stream_left reaches 0, we don't need to trigger any action
         * in terms of destroying the associated channel, because the action that
@@ -1920,11 +2040,16 @@ static int add_metadata_stream(struct lttng_consumer_stream *stream,
         * stream.
         */
        if (uatomic_read(&stream->chan->nb_init_stream_left) > 0) {
+               /* Increment refcount before decrementing nb_init_stream_left */
+               cmm_smp_wmb();
                uatomic_dec(&stream->chan->nb_init_stream_left);
        }
 
        lttng_ht_add_unique_u64(ht, &stream->node);
 
+       lttng_ht_add_unique_u64(consumer_data.stream_per_chan_id_ht,
+               &stream->node_channel_id);
+
        /*
         * 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
@@ -1935,6 +2060,8 @@ static int add_metadata_stream(struct lttng_consumer_stream *stream,
        rcu_read_unlock();
 
        pthread_mutex_unlock(&stream->lock);
+       pthread_mutex_unlock(&stream->chan->lock);
+       pthread_mutex_unlock(&stream->chan->timer_lock);
        pthread_mutex_unlock(&consumer_data.lock);
        return ret;
 }
@@ -2012,7 +2139,7 @@ void *consumer_thread_metadata_poll(void *data)
        metadata_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
        if (!metadata_ht) {
                /* ENOMEM at this point. Better to bail out. */
-               goto error;
+               goto end_ht;
        }
 
        DBG("Thread metadata poll started");
@@ -2021,10 +2148,11 @@ void *consumer_thread_metadata_poll(void *data)
        ret = lttng_poll_create(&events, 2, LTTNG_CLOEXEC);
        if (ret < 0) {
                ERR("Poll set creation failed");
-               goto end;
+               goto end_poll;
        }
 
-       ret = lttng_poll_add(&events, ctx->consumer_metadata_pipe[0], LPOLLIN);
+       ret = lttng_poll_add(&events,
+                       lttng_pipe_get_readfd(ctx->consumer_metadata_pipe), LPOLLIN);
        if (ret < 0) {
                goto end;
        }
@@ -2062,30 +2190,26 @@ restart:
                                continue;
                        }
 
-                       if (pollfd == ctx->consumer_metadata_pipe[0]) {
+                       if (pollfd == lttng_pipe_get_readfd(ctx->consumer_metadata_pipe)) {
                                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");
-                                       }
+                                       lttng_poll_del(&events,
+                                                       lttng_pipe_get_readfd(ctx->consumer_metadata_pipe));
+                                       lttng_pipe_read_close(ctx->consumer_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");
+                                       ssize_t pipe_len;
+
+                                       pipe_len = lttng_pipe_read(ctx->consumer_metadata_pipe,
+                                                       &stream, sizeof(stream));
+                                       if (pipe_len < 0) {
+                                               ERR("read metadata stream, ret: %zd", pipe_len);
                                                /*
-                                                * Let's continue here and hope we can still work
-                                                * without stopping the consumer. XXX: Should we?
+                                                * Continue here to handle the rest of the streams.
                                                 */
                                                continue;
                                        }
@@ -2100,14 +2224,6 @@ restart:
                                        DBG("Adding metadata stream %d to poll set",
                                                        stream->wait_fd);
 
-                                       ret = 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);
@@ -2161,14 +2277,21 @@ restart:
                                DBG("Metadata available on fd %d", pollfd);
                                assert(stream->wait_fd == pollfd);
 
-                               len = ctx->on_buffer_ready(stream, ctx);
+                               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);
+
                                /* 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;
                                }
                        }
 
@@ -2180,10 +2303,11 @@ restart:
 error:
 end:
        DBG("Metadata poll thread exiting");
-       lttng_poll_clean(&events);
 
+       lttng_poll_clean(&events);
+end_poll:
        destroy_stream_ht(metadata_ht);
-
+end_ht:
        rcu_unregister_thread();
        return NULL;
 }
@@ -2211,7 +2335,11 @@ void *consumer_thread_data_poll(void *data)
                goto end;
        }
 
-       local_stream = zmalloc(sizeof(struct lttng_consumer_stream));
+       local_stream = zmalloc(sizeof(struct lttng_consumer_stream *));
+       if (local_stream == NULL) {
+               PERROR("local_stream malloc");
+               goto end;
+       }
 
        while (1) {
                high_prio = 0;
@@ -2239,7 +2367,7 @@ void *consumer_thread_data_poll(void *data)
 
                        /* allocate for all fds + 1 for the consumer_data_pipe */
                        local_stream = zmalloc((consumer_data.stream_count + 1) *
-                                       sizeof(struct lttng_consumer_stream));
+                                       sizeof(struct lttng_consumer_stream *));
                        if (local_stream == NULL) {
                                PERROR("local_stream malloc");
                                pthread_mutex_unlock(&consumer_data.lock);
@@ -2291,13 +2419,10 @@ void *consumer_thread_data_poll(void *data)
                        ssize_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);
+                       pipe_readlen = lttng_pipe_read(ctx->consumer_data_pipe,
+                                       &new_stream, sizeof(new_stream));
                        if (pipe_readlen < 0) {
-                               PERROR("read consumer data pipe");
+                               ERR("Consumer data pipe ret %zd", pipe_readlen);
                                /* Continue so we can at least handle the current stream(s). */
                                continue;
                        }
@@ -2312,17 +2437,6 @@ void *consumer_thread_data_poll(void *data)
                                continue;
                        }
 
-                       ret = add_stream(new_stream, data_ht);
-                       if (ret) {
-                               ERR("Consumer add stream %" PRIu64 " 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;
                }
@@ -2435,17 +2549,322 @@ end:
         * only tracked fd in the poll set. The thread will take care of closing
         * the read side.
         */
-       ret = close(ctx->consumer_metadata_pipe[1]);
+       (void) lttng_pipe_write_close(ctx->consumer_metadata_pipe);
+
+       destroy_data_stream_ht(data_ht);
+
+       rcu_unregister_thread();
+       return NULL;
+}
+
+/*
+ * Close wake-up end of each stream belonging to the channel. This will
+ * allow the poll() on the stream read-side to detect when the
+ * write-side (application) finally closes them.
+ */
+static
+void consumer_close_channel_streams(struct lttng_consumer_channel *channel)
+{
+       struct lttng_ht *ht;
+       struct lttng_consumer_stream *stream;
+       struct lttng_ht_iter iter;
+
+       ht = consumer_data.stream_per_chan_id_ht;
+
+       rcu_read_lock();
+       cds_lfht_for_each_entry_duplicate(ht->ht,
+                       ht->hash_fct(&channel->key, lttng_ht_seed),
+                       ht->match_fct, &channel->key,
+                       &iter.iter, stream, node_channel_id.node) {
+               /*
+                * Protect against teardown with mutex.
+                */
+               pthread_mutex_lock(&stream->lock);
+               if (cds_lfht_is_node_deleted(&stream->node.node)) {
+                       goto next;
+               }
+               switch (consumer_data.type) {
+               case LTTNG_CONSUMER_KERNEL:
+                       break;
+               case LTTNG_CONSUMER32_UST:
+               case LTTNG_CONSUMER64_UST:
+                       /*
+                        * Note: a mutex is taken internally within
+                        * liblttng-ust-ctl to protect timer wakeup_fd
+                        * use from concurrent close.
+                        */
+                       lttng_ustconsumer_close_stream_wakeup(stream);
+                       break;
+               default:
+                       ERR("Unknown consumer_data type");
+                       assert(0);
+               }
+       next:
+               pthread_mutex_unlock(&stream->lock);
+       }
+       rcu_read_unlock();
+}
+
+static void destroy_channel_ht(struct lttng_ht *ht)
+{
+       struct lttng_ht_iter iter;
+       struct lttng_consumer_channel *channel;
+       int ret;
+
+       if (ht == NULL) {
+               return;
+       }
+
+       rcu_read_lock();
+       cds_lfht_for_each_entry(ht->ht, &iter.iter, channel, wait_fd_node.node) {
+               ret = lttng_ht_del(ht, &iter);
+               assert(ret != 0);
+       }
+       rcu_read_unlock();
+
+       lttng_ht_destroy(ht);
+}
+
+/*
+ * This thread polls the channel fds to detect when they are being
+ * closed. It closes all related streams if the channel is detected as
+ * closed. It is currently only used as a shim layer for UST because the
+ * consumerd needs to keep the per-stream wakeup end of pipes open for
+ * periodical flush.
+ */
+void *consumer_thread_channel_poll(void *data)
+{
+       int ret, i, pollfd;
+       uint32_t revents, nb_fd;
+       struct lttng_consumer_channel *chan = NULL;
+       struct lttng_ht_iter iter;
+       struct lttng_ht_node_u64 *node;
+       struct lttng_poll_event events;
+       struct lttng_consumer_local_data *ctx = data;
+       struct lttng_ht *channel_ht;
+
+       rcu_register_thread();
+
+       channel_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
+       if (!channel_ht) {
+               /* ENOMEM at this point. Better to bail out. */
+               goto end_ht;
+       }
+
+       DBG("Thread channel poll started");
+
+       /* Size is set to 1 for the consumer_channel pipe */
+       ret = lttng_poll_create(&events, 2, LTTNG_CLOEXEC);
        if (ret < 0) {
-               PERROR("close data pipe");
+               ERR("Poll set creation failed");
+               goto end_poll;
        }
 
-       destroy_data_stream_ht(data_ht);
+       ret = lttng_poll_add(&events, ctx->consumer_channel_pipe[0], LPOLLIN);
+       if (ret < 0) {
+               goto end;
+       }
+
+       /* Main loop */
+       DBG("Channel main loop started");
+
+       while (1) {
+               /* Only the channel pipe is set */
+               if (LTTNG_POLL_GETNB(&events) == 0 && consumer_quit == 1) {
+                       goto end;
+               }
+
+restart:
+               DBG("Channel poll wait with %d fd(s)", LTTNG_POLL_GETNB(&events));
+               ret = lttng_poll_wait(&events, -1);
+               DBG("Channel event catched in thread");
+               if (ret < 0) {
+                       if (errno == EINTR) {
+                               ERR("Poll EINTR catched");
+                               goto restart;
+                       }
+                       goto end;
+               }
+
+               nb_fd = ret;
+
+               /* From here, the event is a channel 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_channel_pipe[0]) {
+                               if (revents & (LPOLLERR | LPOLLHUP)) {
+                                       DBG("Channel 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_channel_pipe[0]);
+                                       continue;
+                               } else if (revents & LPOLLIN) {
+                                       enum consumer_channel_action action;
+                                       uint64_t key;
+
+                                       ret = read_channel_pipe(ctx, &chan, &key, &action);
+                                       if (ret <= 0) {
+                                               ERR("Error reading channel pipe");
+                                               continue;
+                                       }
+
+                                       switch (action) {
+                                       case CONSUMER_CHANNEL_ADD:
+                                               DBG("Adding channel %d to poll set",
+                                                       chan->wait_fd);
+
+                                               lttng_ht_node_init_u64(&chan->wait_fd_node,
+                                                       chan->wait_fd);
+                                               rcu_read_lock();
+                                               lttng_ht_add_unique_u64(channel_ht,
+                                                               &chan->wait_fd_node);
+                                               rcu_read_unlock();
+                                               /* Add channel to the global poll events list */
+                                               lttng_poll_add(&events, chan->wait_fd,
+                                                               LPOLLIN | LPOLLPRI);
+                                               break;
+                                       case CONSUMER_CHANNEL_DEL:
+                                       {
+                                               struct lttng_consumer_stream *stream, *stmp;
+
+                                               rcu_read_lock();
+                                               chan = consumer_find_channel(key);
+                                               if (!chan) {
+                                                       rcu_read_unlock();
+                                                       ERR("UST consumer get channel key %" PRIu64 " not found for del channel", key);
+                                                       break;
+                                               }
+                                               lttng_poll_del(&events, chan->wait_fd);
+                                               iter.iter.node = &chan->wait_fd_node.node;
+                                               ret = lttng_ht_del(channel_ht, &iter);
+                                               assert(ret == 0);
+                                               consumer_close_channel_streams(chan);
+
+                                               switch (consumer_data.type) {
+                                               case LTTNG_CONSUMER_KERNEL:
+                                                       break;
+                                               case LTTNG_CONSUMER32_UST:
+                                               case LTTNG_CONSUMER64_UST:
+                                                       /* Delete streams that might have been left in the stream list. */
+                                                       cds_list_for_each_entry_safe(stream, stmp, &chan->streams.head,
+                                                                       send_node) {
+                                                               cds_list_del(&stream->send_node);
+                                                               lttng_ustconsumer_del_stream(stream);
+                                                               uatomic_sub(&stream->chan->refcount, 1);
+                                                               assert(&chan->refcount);
+                                                               free(stream);
+                                                       }
+                                                       break;
+                                               default:
+                                                       ERR("Unknown consumer_data type");
+                                                       assert(0);
+                                               }
+
+                                               /*
+                                                * Release our own refcount. Force channel deletion even if
+                                                * streams were not initialized.
+                                                */
+                                               if (!uatomic_sub_return(&chan->refcount, 1)) {
+                                                       consumer_del_channel(chan);
+                                               }
+                                               rcu_read_unlock();
+                                               goto restart;
+                                       }
+                                       case CONSUMER_CHANNEL_QUIT:
+                                               /*
+                                                * Remove the pipe from the poll set and continue the loop
+                                                * since their might be data to consume.
+                                                */
+                                               lttng_poll_del(&events, ctx->consumer_channel_pipe[0]);
+                                               continue;
+                                       default:
+                                               ERR("Unknown action");
+                                               break;
+                                       }
+                               }
+
+                               /* Handle other stream */
+                               continue;
+                       }
+
+                       rcu_read_lock();
+                       {
+                               uint64_t tmp_id = (uint64_t) pollfd;
+
+                               lttng_ht_lookup(channel_ht, &tmp_id, &iter);
+                       }
+                       node = lttng_ht_iter_get_node_u64(&iter);
+                       assert(node);
+
+                       chan = caa_container_of(node, struct lttng_consumer_channel,
+                                       wait_fd_node);
+
+                       /* Check for error event */
+                       if (revents & (LPOLLERR | LPOLLHUP)) {
+                               DBG("Channel fd %d is hup|err.", pollfd);
+
+                               lttng_poll_del(&events, chan->wait_fd);
+                               ret = lttng_ht_del(channel_ht, &iter);
+                               assert(ret == 0);
+                               consumer_close_channel_streams(chan);
+
+                               /* Release our own refcount */
+                               if (!uatomic_sub_return(&chan->refcount, 1)
+                                               && !uatomic_read(&chan->nb_init_stream_left)) {
+                                       consumer_del_channel(chan);
+                               }
+                       }
+
+                       /* Release RCU lock for the channel looked up */
+                       rcu_read_unlock();
+               }
+       }
+
+end:
+       lttng_poll_clean(&events);
+end_poll:
+       destroy_channel_ht(channel_ht);
+end_ht:
+       DBG("Channel poll thread exiting");
        rcu_unregister_thread();
        return NULL;
 }
 
+static int set_metadata_socket(struct lttng_consumer_local_data *ctx,
+               struct pollfd *sockpoll, int client_socket)
+{
+       int ret;
+
+       assert(ctx);
+       assert(sockpoll);
+
+       if (lttng_consumer_poll_socket(sockpoll) < 0) {
+               ret = -1;
+               goto error;
+       }
+       DBG("Metadata connection on client_socket");
+
+       /* Blocking call, waiting for transmission */
+       ctx->consumer_metadata_socket = lttcomm_accept_unix_sock(client_socket);
+       if (ctx->consumer_metadata_socket < 0) {
+               WARN("On accept metadata");
+               ret = -1;
+               goto error;
+       }
+       ret = 0;
+
+error:
+       return ret;
+}
+
 /*
  * This thread listens on the consumerd socket and receives the file
  * descriptors from the session daemon.
@@ -2483,12 +2902,6 @@ void *consumer_thread_sessiond_poll(void *data)
                goto end;
        }
 
-       ret = fcntl(client_socket, F_SETFL, O_NONBLOCK);
-       if (ret < 0) {
-               PERROR("fcntl O_NONBLOCK");
-               goto end;
-       }
-
        /* prepare the FDs to poll : to client socket and the should_quit pipe */
        consumer_sockpoll[0].fd = ctx->consumer_should_quit[0];
        consumer_sockpoll[0].events = POLLIN | POLLPRI;
@@ -2506,9 +2919,13 @@ void *consumer_thread_sessiond_poll(void *data)
                WARN("On accept");
                goto end;
        }
-       ret = fcntl(sock, F_SETFL, O_NONBLOCK);
+
+       /*
+        * Setup metadata socket which is the second socket connection on the
+        * command unix socket.
+        */
+       ret = set_metadata_socket(ctx, consumer_sockpoll, client_socket);
        if (ret < 0) {
-               PERROR("fcntl O_NONBLOCK");
                goto end;
        }
 
@@ -2568,7 +2985,9 @@ end:
         * Notify the data poll thread to poll back again and test the
         * consumer_quit state that we just set so to quit gracefully.
         */
-       notify_thread_pipe(ctx->consumer_data_pipe[1]);
+       notify_thread_lttng_pipe(ctx->consumer_data_pipe);
+
+       notify_channel_pipe(ctx, NULL, -1, CONSUMER_CHANNEL_QUIT);
 
        /* Cleaning up possibly open sockets. */
        if (sock >= 0) {
@@ -2578,7 +2997,7 @@ end:
                }
        }
        if (client_socket >= 0) {
-               ret = close(sock);
+               ret = close(client_socket);
                if (ret < 0) {
                        PERROR("close client_socket sessiond poll");
                }
@@ -2637,6 +3056,7 @@ void lttng_consumer_init(void)
        consumer_data.channel_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
        consumer_data.relayd_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
        consumer_data.stream_list_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
+       consumer_data.stream_per_chan_id_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
 }
 
 /*
@@ -2645,57 +3065,79 @@ void lttng_consumer_init(void)
  * This will create a relayd socket pair and add it to the relayd hash table.
  * The caller MUST acquire a RCU read side lock before calling it.
  */
-int consumer_add_relayd_socket(int net_seq_idx, int sock_type,
+int consumer_add_relayd_socket(uint64_t net_seq_idx, int sock_type,
                struct lttng_consumer_local_data *ctx, int sock,
-               struct pollfd *consumer_sockpoll, struct lttcomm_sock *relayd_sock,
-               unsigned int sessiond_id)
+               struct pollfd *consumer_sockpoll,
+               struct lttcomm_relayd_sock *relayd_sock, uint64_t sessiond_id)
 {
        int fd = -1, ret = -1, relayd_created = 0;
        enum lttng_error_code ret_code = LTTNG_OK;
-       struct consumer_relayd_sock_pair *relayd;
+       struct consumer_relayd_sock_pair *relayd = NULL;
 
-       DBG("Consumer adding relayd socket (idx: %d)", net_seq_idx);
+       assert(ctx);
+       assert(relayd_sock);
 
-       /* First send a status message before receiving the fds. */
-       ret = consumer_send_status_msg(sock, ret_code);
-       if (ret < 0) {
-               /* Somehow, the session daemon is not responding anymore. */
-               goto error;
-       }
+       DBG("Consumer adding relayd socket (idx: %" PRIu64 ")", net_seq_idx);
 
        /* Get relayd reference if exists. */
        relayd = consumer_find_relayd(net_seq_idx);
        if (relayd == NULL) {
+               assert(sock_type == LTTNG_STREAM_CONTROL);
                /* Not found. Allocate one. */
                relayd = consumer_allocate_relayd_sock_pair(net_seq_idx);
                if (relayd == NULL) {
-                       lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_OUTFD_ERROR);
-                       ret = -1;
+                       ret = -ENOMEM;
+                       ret_code = LTTCOMM_CONSUMERD_ENOMEM;
                        goto error;
+               } else {
+                       relayd->sessiond_session_id = sessiond_id;
+                       relayd_created = 1;
                }
-               relayd->sessiond_session_id = (uint64_t) sessiond_id;
-               relayd_created = 1;
+
+               /*
+                * This code path MUST continue to the consumer send status message to
+                * we can notify the session daemon and continue our work without
+                * killing everything.
+                */
+       } else {
+               /*
+                * relayd key should never be found for control socket.
+                */
+               assert(sock_type != LTTNG_STREAM_CONTROL);
+       }
+
+       /* First send a status message before receiving the fds. */
+       ret = consumer_send_status_msg(sock, LTTNG_OK);
+       if (ret < 0) {
+               /* Somehow, the session daemon is not responding anymore. */
+               lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_FATAL);
+               goto error_nosignal;
        }
 
        /* Poll on consumer socket. */
        if (lttng_consumer_poll_socket(consumer_sockpoll) < 0) {
+               lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_POLL_ERROR);
                ret = -EINTR;
-               goto error;
+               goto error_nosignal;
        }
 
        /* Get relayd socket from session daemon */
        ret = lttcomm_recv_fds_unix_sock(sock, &fd, 1);
        if (ret != sizeof(fd)) {
-               lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_ERROR_RECV_FD);
                ret = -1;
                fd = -1;        /* Just in case it gets set with an invalid value. */
-               goto error_close;
-       }
 
-       /* We have the fds without error. Send status back. */
-       ret = consumer_send_status_msg(sock, ret_code);
-       if (ret < 0) {
-               /* Somehow, the session daemon is not responding anymore. */
+               /*
+                * Failing to receive FDs might indicate a major problem such as
+                * reaching a fd limit during the receive where the kernel returns a
+                * MSG_CTRUNC and fails to cleanup the fd in the queue. Any case, we
+                * don't take any chances and stop everything.
+                *
+                * XXX: Feature request #558 will fix that and avoid this possible
+                * issue when reaching the fd limit.
+                */
+               lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_ERROR_RECV_FD);
+               ret_code = LTTCOMM_CONSUMERD_ERROR_RECV_FD;
                goto error;
        }
 
@@ -2703,21 +3145,28 @@ int consumer_add_relayd_socket(int net_seq_idx, int sock_type,
        switch (sock_type) {
        case LTTNG_STREAM_CONTROL:
                /* Copy received lttcomm socket */
-               lttcomm_copy_sock(&relayd->control_sock, relayd_sock);
-               ret = lttcomm_create_sock(&relayd->control_sock);
-               /* Immediately try to close the created socket if valid. */
-               if (relayd->control_sock.fd >= 0) {
-                       if (close(relayd->control_sock.fd)) {
-                               PERROR("close relayd control socket");
-                       }
-               }
+               lttcomm_copy_sock(&relayd->control_sock.sock, &relayd_sock->sock);
+               ret = lttcomm_create_sock(&relayd->control_sock.sock);
                /* Handle create_sock error. */
                if (ret < 0) {
+                       ret_code = LTTCOMM_CONSUMERD_ENOMEM;
                        goto error;
                }
+               /*
+                * Close the socket created internally by
+                * lttcomm_create_sock, so we can replace it by the one
+                * received from sessiond.
+                */
+               if (close(relayd->control_sock.sock.fd)) {
+                       PERROR("close");
+               }
 
                /* Assign new file descriptor */
-               relayd->control_sock.fd = fd;
+               relayd->control_sock.sock.fd = fd;
+               fd = -1;        /* For error path */
+               /* Assign version values. */
+               relayd->control_sock.major = relayd_sock->major;
+               relayd->control_sock.minor = relayd_sock->minor;
 
                /*
                 * Create a session on the relayd and store the returned id. Lock the
@@ -2739,31 +3188,40 @@ int consumer_add_relayd_socket(int net_seq_idx, int sock_type,
                         */
                        (void) relayd_close(&relayd->control_sock);
                        (void) relayd_close(&relayd->data_sock);
+                       ret_code = LTTCOMM_CONSUMERD_RELAYD_FAIL;
                        goto error;
                }
 
                break;
        case LTTNG_STREAM_DATA:
                /* Copy received lttcomm socket */
-               lttcomm_copy_sock(&relayd->data_sock, relayd_sock);
-               ret = lttcomm_create_sock(&relayd->data_sock);
-               /* Immediately try to close the created socket if valid. */
-               if (relayd->data_sock.fd >= 0) {
-                       if (close(relayd->data_sock.fd)) {
-                               PERROR("close relayd data socket");
-                       }
-               }
+               lttcomm_copy_sock(&relayd->data_sock.sock, &relayd_sock->sock);
+               ret = lttcomm_create_sock(&relayd->data_sock.sock);
                /* Handle create_sock error. */
                if (ret < 0) {
+                       ret_code = LTTCOMM_CONSUMERD_ENOMEM;
                        goto error;
                }
+               /*
+                * Close the socket created internally by
+                * lttcomm_create_sock, so we can replace it by the one
+                * received from sessiond.
+                */
+               if (close(relayd->data_sock.sock.fd)) {
+                       PERROR("close");
+               }
 
                /* Assign new file descriptor */
-               relayd->data_sock.fd = fd;
+               relayd->data_sock.sock.fd = fd;
+               fd = -1;        /* for eventual error paths */
+               /* Assign version values. */
+               relayd->data_sock.major = relayd_sock->major;
+               relayd->data_sock.minor = relayd_sock->minor;
                break;
        default:
                ERR("Unknown relayd socket type (%d)", sock_type);
                ret = -1;
+               ret_code = LTTCOMM_CONSUMERD_FATAL;
                goto error;
        }
 
@@ -2771,6 +3229,14 @@ int consumer_add_relayd_socket(int net_seq_idx, int sock_type,
                        sock_type == LTTNG_STREAM_CONTROL ? "control" : "data",
                        relayd->net_seq_idx, fd);
 
+       /* We successfully added the socket. Send status back. */
+       ret = consumer_send_status_msg(sock, ret_code);
+       if (ret < 0) {
+               /* Somehow, the session daemon is not responding anymore. */
+               lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_FATAL);
+               goto error_nosignal;
+       }
+
        /*
         * Add relayd socket pair to consumer data hashtable. If object already
         * exists or on error, the function gracefully returns.
@@ -2781,6 +3247,11 @@ int consumer_add_relayd_socket(int net_seq_idx, int sock_type,
        return 0;
 
 error:
+       if (consumer_send_status_msg(sock, ret_code) < 0) {
+               lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_FATAL);
+       }
+
+error_nosignal:
        /* Close received socket if valid. */
        if (fd >= 0) {
                if (close(fd)) {
@@ -2788,7 +3259,6 @@ error:
                }
        }
 
-error_close:
        if (relayd_created) {
                free(relayd);
        }
@@ -3020,3 +3490,23 @@ int consumer_send_status_channel(int sock,
 
        return lttcomm_send_unix_sock(sock, &msg, sizeof(msg));
 }
+
+/*
+ * Using a maximum stream size with the produced and consumed position of a
+ * stream, computes the new consumed position to be as close as possible to the
+ * maximum possible stream size.
+ *
+ * If maximum stream size is lower than the possible buffer size (produced -
+ * consumed), the consumed_pos given is returned untouched else the new value
+ * is returned.
+ */
+unsigned long consumer_get_consumed_maxsize(unsigned long consumed_pos,
+               unsigned long produced_pos, uint64_t max_stream_size)
+{
+       if (max_stream_size && max_stream_size < (produced_pos - consumed_pos)) {
+               /* Offset from the produced position to get the latest buffers. */
+               return produced_pos - max_stream_size;
+       }
+
+       return consumed_pos;
+}
This page took 0.045311 seconds and 4 git commands to generate.