Fix: consumer socket lock not held during snapshot record

This missing lock was identified while stress-testing the
snapshot tracing mode.

The "post_mortem" test case would sometimes hang on a
push_metadata() call waiting for a status reply from the
consumer daemon.

This test demonstrated a race that consists in killing an
application and taking a snapshot near-simultaneously.

This causes the app management thread to issue a "push metadata"
command to the consumerd while the lttng client is issuing
a snapshot record command.

Since the snapshot record does not acquire the consumer socket lock,
the "push metadata" and "snapshot" commands end-up mixed-up on
the socket which ultimately causes the "apps management" thread
to wait for a reply forever while holding the socket's lock.

This prevents the client, invoked by the test script, from
completing the "stop" operation on the session.

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

diff --git a/src/bin/lttng-sessiond/consumer.c b/src/bin/lttng-sessiond/consumer.c
index 25194460645d377ad5d0680417629f1f69d2f40b..126b01a448c062e35095ddb97fd1eaf37e5dabb2 100644 (file)
--- a/src/bin/lttng-sessiond/consumer.c
+++ b/src/bin/lttng-sessiond/consumer.c
@@ -1442,7 +1442,9 @@ int consumer_snapshot_channel(struct consumer_socket *socket, uint64_t key,
        }
 
        health_code_update();
+       pthread_mutex_lock(socket->lock);
        ret = consumer_send_msg(socket, &msg);
+       pthread_mutex_unlock(socket->lock);
        if (ret < 0) {
                goto error;
        }