Gapless part 1 #1288
Gapless part 1 #1288
Conversation
While trying to remove traces of stop calls in core to get gapless working I found we had no way to switch to switch tracks without triggering a play. This change fixes this by changing the backends playback provider API. - play() now _only_ starts playback and does not take any arguments. - prepare_change() has been added, this could have been avoided with a kwarg to change_track(track), but that would break more backends. - core has been updated to call prepare_change+change_track+play as needed. - tests have been updated to handle this change. Longer term I hope to completely rework the playback API in backends, as 99% of our backends only use change_track(track) to translate URIs. So we should make simple case simple, and handle mopidy-spotify / appsrc in some other way.
Note that since this doesn't use play via change_track we have to copy over the tracking code and make sure it is only run for the playing case
Same fix as for core.playback.next and mostly the same caveats.
Only handles the playing case, unlike the previous and next changes. It should also be noted that this is just a temporary state on the road to making this method handle gapless.
There is still quite a bit to be done is this area: - We need to start tracking pending tracks as there is time window when we are decoding a new track but still playing the old one - Currently this breaks seek, next, prev during this time window - The old on_end_of_track code needs to die. - Stream changes need to trigger playing events and switch the pending track to current.
This has been replaced by on_about_to_finish and on_end_of_stream. Tests have been updated to reflect these changes.
- Adds stream changed handler to core - Moves playback started trigger to stream changed - Made about to finish store next track in _pending_tl_track - Set the pending track as current in stream changed - Adds tests for all of this and fixes existing tests
- Pending track should only be triggered by stream_changed if there is one. - Tracklist changed was incorrectly calling stop breaking tests and gapless - Tests have been updated to capture and replay audio events. This should avoid test deadlocks while still using the audio fakes.
Conflicts: mopidy/backend.py mopidy/commands.py mopidy/core/actor.py mopidy/core/playback.py tests/audio/test_actor.py tests/core/test_playback.py tests/local/test_playback.py
Conflicts: mopidy/commands.py mopidy/core/playback.py tests/core/test_playback.py tests/local/test_playback.py
This is no longer needed as the plain send method makes sure to use tell to queue actor message. Which has better performance, and avoids deadlocks. A side effect of this is that assuming you have a core actor running and a dummy audio in use audio events just work.
Conflicts: tests/local/test_playback.py
…ed in This is needed in order to make audio events propagate, to core and trigger async state changes in tests.
| self._pending_tl_track = None | ||
|
|
||
| if self._audio: | ||
| self._audio.set_about_to_finish_callback(self._on_about_to_finish) |
There was a problem hiding this comment.
Won't this cause the audio actor to run core's code and alter core state?
If you make the callback public, you can probably fix this by using self.core.actor_ref.proxy().playback.on_about_to_finish as the callback.
There was a problem hiding this comment.
The call path should be something like the following:
about-to-finish --> audio --> core callback --> backend --> audio
The goal here is to block the about to finish until an audio.set_uri type call has made it out. (This is also one of the reasons I want backends out of this path). The about-to-finish should be coming from some gst internal thread, so our audio code will also run there, and then from there core would probably still be in the same context to how the callback is done. We den call into the backend via actors, after which the backend calls into audio via actors. So we must ensure that we can't deadlock that last call, or else the about-to-finish never returns.
There was a problem hiding this comment.
Not exactly easy to follow what's happening here. Are you confident in how this is implemented now, or do I have a point? :-)
There was a problem hiding this comment.
I'm fairly sure what I described is correct, and it was intentional to do things outside of actors due to deadlocks that I no longer remember the details of. And also with respect to the circular deps between audio and core setup.
There was a problem hiding this comment.
self.core.actor_ref.proxy().playback.on_about_to_finish inside the init actually deadlocks as you are trying to create proxy to yourself while still setting yourself up...
There was a problem hiding this comment.
If I give it the actor, and then let it create the proxy on the fly in audio it works at least. And I've not manged to deadlock it yet. I still want to think this over a bit more and do some more testing.
There was a problem hiding this comment.
You should be able to create the proxy in on_start(). __init__() is executed in the actor's parent before the actor is actually started, thus there is no ActorRef object to create an ActorProxy from.
|
When about to finish gets called we are running in some GStreamer thread. Our audio code then calls the shim core callback which is responsible for transferring our execution to the core actor thread and waiting for the response. From this point we do normal actor calls to the backend(s) which in turn call into the audio actor. Since the initial audio code that was called is outside the actor this should never deadlock due to this loop.
|
There, now the core code that gets called is in the actor. |
|
|
Currently blocked on mopidy-spotify / appsrc deadlocking with these changes. |
This is a major cleanup of core playback internals. Some important things to note:
playing -> playingtype events.play,next,prevare now async. Tests must take care to wait for core futures to finish. Flow is not that these calls, about about to finish set the pending track. Once stream changed comes in we switch. This might also confuse clients.play,next,prevstill break streaming, we need to make all changes async by faking end of stream events. Postponed for now.