discovery: fix asynchronous agent refreshes pruning

Asynchronous refreshes pruning may occur when the agent async close
function is called, but also when a stream is closed. A flag
'discarding' is used in the CandidateRefresh object to determine
if a refresh is already on the way to be asynchronously freed. A refresh
definitely freed is removed from agent refresh_list.

When the agent async close function is called, it is passed a user
callback that will be invoked when all refreshes have been freed.

This is not exactly how things work currently, because right now, the
callback function is also invoked when all CandidateRefresh objects have
the discarding flag set, with the test 'data->items_to_free==0'. In that
case, clearly not all refreshes have been freed, as some are still there
in the refresh_list with the discarding flag set.

This is probably not what the user expects from the function
nice_agent_close_async(), where the callback is supposed to be invoked
after all refreshes have been freed: and including those asynchronously
pruned by nice_prune_stream_async().

This patch adds a supplementary timeout when closing the agent async,
that waits until the agent refresh list becomes empty.

1 files changed, 9 insertions, 0 deletions

diff --git a/agent/agent.c b/agent/agent.c
index 8d2349a..e6132eb 100644
--- a/agent/agent.c
+++ b/agent/agent.c
@@ -7127,6 +7127,15 @@ on_agent_refreshes_pruned (NiceAgent *agent, gpointer user_data)
 {
   GTask *task = user_data;
 
+  if (agent->refresh_list) {
+    GSource *timeout_source = NULL;
+    agent_timeout_add_with_context (agent, &timeout_source,
+        "Async refresh prune", agent->stun_initial_timeout,
+        on_agent_refreshes_pruned, user_data);
+    g_source_unref (timeout_source);
+    return G_SOURCE_REMOVE;
+  }
+
   /* This is called from a timeout cb with agent lock held */
 
   agent_unlock (agent);