MediaStream audio object graph untangling and clean-ups.

content/renderer/media/webrtc/media_stream_remote_audio_track.cc

Index: content/renderer/media/webrtc/media_stream_remote_audio_track.cc
diff --git a/content/renderer/media/webrtc/media_stream_remote_audio_track.cc b/content/renderer/media/webrtc/media_stream_remote_audio_track.cc
index 17df845f81c491b3c9a0631d3a20835f1b05aa1a..f49891e9fb663241a21e1cfb46db167624f5d89c 100644
--- a/content/renderer/media/webrtc/media_stream_remote_audio_track.cc
+++ b/content/renderer/media/webrtc/media_stream_remote_audio_track.cc
@@ -120,7 +120,8 @@ MediaStreamRemoteAudioTrack::MediaStreamRemoteAudioTrack(
 
 MediaStreamRemoteAudioTrack::~MediaStreamRemoteAudioTrack() {
   DCHECK(main_render_thread_checker_.CalledOnValidThread());
-  source()->RemoveAll(this);
+  // Ensure the track is stopped.
+  Stop();
 }
 
 void MediaStreamRemoteAudioTrack::SetEnabled(bool enabled) {
@@ -141,6 +142,11 @@ void MediaStreamRemoteAudioTrack::SetEnabled(bool enabled) {
 
 void MediaStreamRemoteAudioTrack::Stop() {
   DCHECK(main_render_thread_checker_.CalledOnValidThread());
+
+  MediaStreamAudioTrack::WillStopTrack();

[o1ka, 2016/03/01 14:18:59]

What if some observers rely on the fact that the callbacks are run before/after
the track has actually stopped (next lines)? Since observers subscribe in the
parent class, we probably should explicitly say there what is the expected
sequence of Stop()/WillStopTrack() actions (stop first, notify second, or the
opposite) and make the child classes to adhere to it, so that observers know
what to expect regardless of implementation the parent class is instantiated
with. What do you think?

[miu, 2016/03/02 01:12:50]

Good point, and I made the behavior more precisely defined in the latest patch
set.  No more observers: The point is for the audio flow to be immediately
halted, and then let the subclasses perform any additional steps as needed.

[o1ka, 2016/03/02 16:31:14]

I like it!

+
+  source()->RemoveAll(this);
+
   // Stop means that a track should be stopped permanently. But
   // since there is no proper way of doing that on a remote track, we can
   // at least disable the track. Blink will not call down to the content layer