static int send_stream_to_thread(struct lttng_consumer_stream *stream,
struct lttng_consumer_local_data *ctx)
{
- int ret, stream_pipe;
+ int ret;
+ struct lttng_pipe *stream_pipe;
/* Get the right pipe where the stream will be sent. */
if (stream->metadata_flag) {
- stream_pipe = ctx->consumer_metadata_pipe[1];
+ stream_pipe = ctx->consumer_metadata_pipe;
} else {
- stream_pipe = ctx->consumer_data_pipe[1];
+ stream_pipe = ctx->consumer_data_pipe;
}
- do {
- ret = write(stream_pipe, &stream, sizeof(stream));
- } while (ret < 0 && errno == EINTR);
+ ret = lttng_pipe_write(stream_pipe, &stream, sizeof(stream));
if (ret < 0) {
- PERROR("Consumer write %s stream to pipe %d",
- stream->metadata_flag ? "metadata" : "data", stream_pipe);
+ ERR("Consumer write %s stream to pipe %d",
+ stream->metadata_flag ? "metadata" : "data",
+ lttng_pipe_get_writefd(stream_pipe));
}
return ret;
*/
stream->wait_fd = wait_fd;
+ /*
+ * Increment channel refcount since the channel reference has now been
+ * assigned in the allocation process above.
+ */
+ uatomic_inc(&stream->chan->refcount);
+
/*
* Order is important this is why a list is used. On error, the caller
* should clean this list.
channel->wait_fd = ustctl_channel_get_wait_fd(channel->uchan);
- if (ret < 0) {
- goto error;
- }
-
/* Open all streams for this channel. */
ret = create_ust_streams(channel, ctx);
if (ret < 0) {
DBG("UST consumer writing metadata to channel %s", metadata->name);
+ if (!metadata->metadata_stream) {
+ ret = 0;
+ goto error;
+ }
+
assert(target_offset <= metadata->metadata_cache->max_offset);
ret = ustctl_write_metadata_to_channel(metadata->uchan,
metadata_str + target_offset, len);
DBG("UST consumer flush channel key %" PRIu64, chan_key);
+ rcu_read_lock();
channel = consumer_find_channel(chan_key);
if (!channel) {
ERR("UST consumer flush channel %" PRIu64 " not found", chan_key);
ht = consumer_data.stream_per_chan_id_ht;
/* For each stream of the channel id, flush it. */
- 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) {
ustctl_flush_buffer(stream->ustream, 1);
}
- rcu_read_unlock();
-
error:
+ rcu_read_unlock();
return ret;
}
/*
* Close metadata stream wakeup_fd using the given key to retrieve the channel.
+ * RCU read side lock MUST be acquired before calling this function.
*
* Return 0 on success else an LTTng error code.
*/
static int close_metadata(uint64_t chan_key)
{
- int ret;
+ int ret = 0;
struct lttng_consumer_channel *channel;
DBG("UST consumer close metadata key %" PRIu64, chan_key);
channel = consumer_find_channel(chan_key);
if (!channel) {
- ERR("UST consumer close metadata %" PRIu64 " not found", chan_key);
+ /*
+ * This is possible if the metadata thread has issue a delete because
+ * the endpoint point of the stream hung up. There is no way the
+ * session daemon can know about it thus use a DBG instead of an actual
+ * error.
+ */
+ DBG("UST consumer close metadata %" PRIu64 " not found", chan_key);
ret = LTTNG_ERR_UST_CHAN_NOT_FOUND;
goto error;
}
- ret = ustctl_stream_close_wakeup_fd(channel->metadata_stream->ustream);
- if (ret < 0) {
- ERR("UST consumer unable to close fd of metadata (ret: %d)", ret);
- ret = LTTCOMM_CONSUMERD_ERROR_METADATA;
- goto error;
+ pthread_mutex_lock(&consumer_data.lock);
+
+ if (cds_lfht_is_node_deleted(&channel->node.node)) {
+ goto error_unlock;
}
+
if (channel->switch_timer_enabled == 1) {
DBG("Deleting timer on metadata channel");
consumer_timer_switch_stop(channel);
}
- consumer_metadata_cache_destroy(channel);
+ if (channel->metadata_stream) {
+ ret = ustctl_stream_close_wakeup_fd(channel->metadata_stream->ustream);
+ if (ret < 0) {
+ ERR("UST consumer unable to close fd of metadata (ret: %d)", ret);
+ ret = LTTCOMM_CONSUMERD_ERROR_METADATA;
+ goto error_unlock;
+ }
+ }
+
+error_unlock:
+ pthread_mutex_unlock(&consumer_data.lock);
error:
return ret;
}
goto end_free;
}
+ /*
+ * XXX: The consumer data lock is acquired before calling metadata cache
+ * write which calls push metadata that MUST be protected by the consumer
+ * lock in order to be able to check the validity of the metadata stream of
+ * the channel.
+ *
+ * Note that this will be subject to change to better fine grained locking
+ * and ultimately try to get rid of this global consumer data lock.
+ */
+ pthread_mutex_lock(&consumer_data.lock);
+
pthread_mutex_lock(&channel->metadata_cache->lock);
ret = consumer_metadata_cache_write(channel, offset, len, metadata_str);
if (ret < 0) {
/* Unable to handle metadata. Notify session daemon. */
ret_code = LTTCOMM_CONSUMERD_ERROR_METADATA;
+ /*
+ * Skip metadata flush on write error since the offset and len might
+ * not have been updated which could create an infinite loop below when
+ * waiting for the metadata cache to be flushed.
+ */
+ pthread_mutex_unlock(&channel->metadata_cache->lock);
+ pthread_mutex_unlock(&consumer_data.lock);
+ goto end_free;
}
pthread_mutex_unlock(&channel->metadata_cache->lock);
+ pthread_mutex_unlock(&consumer_data.lock);
while (consumer_metadata_cache_flushed(channel, offset + len)) {
DBG("Waiting for metadata to be flushed");
* The ret value might 0 meaning an orderly shutdown but this is ok
* since the caller handles this.
*/
+ if (ret > 0) {
+ ret = -1;
+ }
return ret;
}
if (msg.cmd_type == LTTNG_CONSUMER_STOP) {
sizeof(is_data_pending));
if (ret < 0) {
DBG("Error when sending the data pending ret code: %d", ret);
+ goto error_fatal;
}
/*
goto end_channel_error;
}
- /*
- * Set refcount to 1 for owner. Below, we will pass
- * ownership to the consumer_thread_channel_poll()
- * thread.
- */
- channel->refcount = 1;
-
/* Build channel attributes from received message. */
attr.subbuf_size = msg.u.ask_channel.subbuf_size;
attr.num_subbuf = msg.u.ask_channel.num_subbuf;
case LTTNG_UST_CHAN_PER_CPU:
channel->type = CONSUMER_CHANNEL_TYPE_DATA;
attr.type = LTTNG_UST_CHAN_PER_CPU;
+ /*
+ * Set refcount to 1 for owner. Below, we will
+ * pass ownership to the
+ * consumer_thread_channel_poll() thread.
+ */
+ channel->refcount = 1;
break;
case LTTNG_UST_CHAN_METADATA:
channel->type = CONSUMER_CHANNEL_TYPE_METADATA;
*/
ret = add_channel(channel, ctx);
if (ret < 0) {
+ if (msg.u.ask_channel.type == LTTNG_UST_CHAN_METADATA) {
+ if (channel->switch_timer_enabled == 1) {
+ consumer_timer_switch_stop(channel);
+ }
+ consumer_metadata_cache_destroy(channel);
+ }
goto end_channel_error;
}
-
/*
* Channel and streams are now created. Inform the session daemon that
* everything went well and should wait to receive the channel and
ret = consumer_send_status_channel(sock, channel);
if (ret < 0) {
/*
- * There is probably a problem on the socket so the poll will get
- * it and clean everything up.
+ * There is probably a problem on the socket.
*/
- goto end_nosignal;
+ goto error_fatal;
}
break;
ret = consumer_send_status_msg(sock, LTTNG_OK);
if (ret < 0) {
/* Somehow, the session daemon is not responding anymore. */
- goto end_nosignal;
+ goto error_fatal;
}
/* Send everything to sessiond. */
case LTTNG_CONSUMER_DESTROY_CHANNEL:
{
uint64_t key = msg.u.destroy_channel.key;
- struct lttng_consumer_channel *channel;
-
- channel = consumer_find_channel(key);
- if (!channel) {
- ERR("UST consumer get channel key %" PRIu64 " not found", key);
- ret_code = LTTNG_ERR_UST_CHAN_NOT_FOUND;
- goto end_msg_sessiond;
- }
-
- destroy_channel(channel);
+ /*
+ * Only called if streams have not been sent to stream
+ * manager thread. However, channel has been sent to
+ * channel manager thread.
+ */
+ notify_thread_del_channel(ctx, key);
goto end_msg_sessiond;
}
case LTTNG_CONSUMER_CLOSE_METADATA:
if (!channel) {
ERR("UST consumer push metadata %" PRIu64 " not found", key);
ret_code = LTTNG_ERR_UST_CHAN_NOT_FOUND;
+ goto end_msg_sessiond;
}
/* Tell session daemon we are ready to receive the metadata. */
/* Wait for more data. */
if (lttng_consumer_poll_socket(consumer_sockpoll) < 0) {
- goto end_nosignal;
+ goto error_fatal;
}
ret = lttng_ustconsumer_recv_metadata(sock, key, offset,
len, channel);
if (ret < 0) {
/* error receiving from sessiond */
- goto end_nosignal;
+ goto error_fatal;
} else {
ret_code = ret;
goto end_msg_sessiond;
* the caller because the session daemon socket management is done
* elsewhere. Returning a negative code or 0 will shutdown the consumer.
*/
- (void) consumer_send_status_msg(sock, ret_code);
+ ret = consumer_send_status_msg(sock, ret_code);
+ if (ret < 0) {
+ goto error_fatal;
+ }
rcu_read_unlock();
return 1;
end_channel_error:
assert(chan);
assert(chan->uchan);
+ if (chan->switch_timer_enabled == 1) {
+ consumer_timer_switch_stop(chan);
+ }
+ consumer_metadata_cache_destroy(chan);
ustctl_destroy_channel(chan->uchan);
}
assert(stream);
assert(stream->ustream);
+ if (stream->chan->switch_timer_enabled == 1) {
+ consumer_timer_switch_stop(stream->chan);
+ }
ustctl_destroy_stream(stream->ustream);
}