Fix: metadata channel leak when using the snapshot tracing mode

While running stress tests involving the snapshot mode, it
becomes apparent that the lttng-consumerd leaks a number of file
descriptors.

To isolate the problem, the test was narrowed down to

* Create a session in snapshot mode
* Enable a userspace channel
* Enable all userspace events
* Start tracing
* Run a traced application
* Stop tracing
* Destroy session

This has shown that 5 file descriptors were leaked on each
iteration of the above.

As the comments in this change indicate, the ownership and
lifetime of metadata channels varies depending on the tracing
mode being used.

In non-snapshot tracing modes, metadata channels are owned by
their respective streams. On destruction of a metadata stream,
consumer_del_channel() is invoked since the stream releases its
ownership of the metadata channel.

However, this relationship between metadata streams and channels
does not exist in snapshot mode; streams are created and
destroyed on every snapshot record. Hence, the
LTTNG_CONSUMER_CLOSE_METADATA command must immediately clean the
metadata channel.

The channel's "monitor" flag is used to determine whether or not
the metadata channel is in "snapshot" mode or not.

Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

diff --git a/src/common/ust-consumer/ust-consumer.c b/src/common/ust-consumer/ust-consumer.c
index c5c2a6548e1bddbcf9772d36fe8266f1fab365be..3350403d9db2ef9bd639cccf48c24f852e6f76e2 100644 (file)
--- a/src/common/ust-consumer/ust-consumer.c
+++ b/src/common/ust-consumer/ust-consumer.c
@@ -820,6 +820,7 @@ static int close_metadata(uint64_t chan_key)
 {
        int ret = 0;
        struct lttng_consumer_channel *channel;
+       unsigned int channel_monitor;
 
        DBG("UST consumer close metadata key %" PRIu64, chan_key);
 
@@ -838,13 +839,48 @@ static int close_metadata(uint64_t chan_key)
 
        pthread_mutex_lock(&consumer_data.lock);
        pthread_mutex_lock(&channel->lock);
-
+       channel_monitor = channel->monitor;
        if (cds_lfht_is_node_deleted(&channel->node.node)) {
                goto error_unlock;
        }
 
        lttng_ustconsumer_close_metadata(channel);
+       pthread_mutex_unlock(&channel->lock);
+       pthread_mutex_unlock(&consumer_data.lock);
 
+       /*
+        * The ownership of a metadata channel depends on the type of
+        * session to which it belongs. In effect, the monitor flag is checked
+        * to determine if this metadata channel is in "snapshot" mode or not.
+        *
+        * In the non-snapshot case, the metadata channel is created along with
+        * a single stream which will remain present until the metadata channel
+        * is destroyed (on the destruction of its session). In this case, the
+        * metadata stream in "monitored" by the metadata poll thread and holds
+        * the ownership of its channel.
+        *
+        * Closing the metadata will cause the metadata stream's "metadata poll
+        * pipe" to be closed. Closing this pipe will wake-up the metadata poll
+        * thread which will teardown the metadata stream which, in return,
+        * deletes the metadata channel.
+        *
+        * In the snapshot case, the metadata stream is created and destroyed
+        * on every snapshot record. Since the channel doesn't have an owner
+        * other than the session daemon, it is safe to destroy it immediately
+        * on reception of the CLOSE_METADATA command.
+        */
+       if (!channel_monitor) {
+               /*
+                * The channel and consumer_data locks must be
+                * released before this call since consumer_del_channel
+                * re-acquires the channel and consumer_data locks to teardown
+                * the channel and queue its reclamation by the "call_rcu"
+                * worker thread.
+                */
+               consumer_del_channel(channel);
+       }
+
+       return ret;
 error_unlock:
        pthread_mutex_unlock(&channel->lock);
        pthread_mutex_unlock(&consumer_data.lock);