Improve AudioTrack recovery from mediaserver death
1. Fix race condition in handling of binder death notifications.
AudioSystem has a mixture of APIs for both ordinary app clients,
and the AudioFlinger and AudioPolicy services within mediaserver.
Due to this mix of uses, it is possible for there to be "surprising"
sequences of calls on the call stack.
Previously, we used a single mutex for all global variables, but
this caused a deadlock. To avoid the deadlock, we unlocked the mutex
during the critical sequence of calls. But this was a a
crucial place where it should have stayed locked;
see Change-Id I315c1c5066f62b05e1c13b04fae1272b5fbce977
Now we use separate mutexes for the AudioFlinger, AudioPolicy, and audio port
related global variables. This allows us to correctly hold each mutex
throughout the atomic region, even when AudioFlinger calls AudioPolicy
via AudioSystem, or vice-versa.
2. AudioSystem::clearAudioConfigCache now clears the IAudioFlinger reference.
3. Make AudioSystem::get_audio_policy_service more like get_audio_flinger.
Bug: 18242291
Change-Id: I9761443d8337df5bf66d4ca2316a9fd0bd11be94
diff --git a/media/libmedia/AudioTrack.cpp b/media/libmedia/AudioTrack.cpp
index 082a5e1..cd493f6 100644
--- a/media/libmedia/AudioTrack.cpp
+++ b/media/libmedia/AudioTrack.cpp
@@ -1828,7 +1828,7 @@
status_t result;
// refresh the audio configuration cache in this process to make sure we get new
- // output parameters in createTrack_l()
+ // output parameters and new IAudioFlinger in createTrack_l()
AudioSystem::clearAudioConfigCache();
if (isOffloadedOrDirect_l()) {