The following revision refers to this bug:
  https://chromium.googlesource.com/chromium/src.git/+/908f0cebd618d41880fdd5f7e7e83e19c39a45c7

commit 908f0cebd618d41880fdd5f7e7e83e19c39a45c7
Author: watk <watk@chromium.org>
Date: Sat Nov 05 01:04:17 2016

Send the video surface destruction message on all Android versions

We previously only sent this on JB to avoid crashes we saw there. It
looks like we're seeing crashes on KK as well, so we have to enable it
there too. It will be enabled for M+ soon to implement #setSurface
support, so that leaves L as the odd one out. For simplicity we'll
just enable it everywhere.

BUG= 662599 

Review-Url: https://codereview.chromium.org/2481583003
Cr-Commit-Position: refs/heads/master@{#430108}

[modify] https://crrev.com/908f0cebd618d41880fdd5f7e7e83e19c39a45c7/content/browser/media/android/browser_surface_view_manager.cc
[modify] https://crrev.com/908f0cebd618d41880fdd5f7e7e83e19c39a45c7/content/browser/media/android/browser_surface_view_manager.h

I think I'm going to leave 1) as-is for now. I don't know of a nice solution that works generally so we just have to choose the least bad solution. Which means we might as well see how what we have works in practice and keep an eye on crash reports.

Aren't we already seeing crash reports for this? It seems like we need to at least wait for the MediaCodec destruction to complete on the GpuMain thread? 

Those crash reports were KK because we weren't sending the message at all on KK. Now we're sending the message on KK.

JB currently has the racy teardown, and it's the one that most reliably crashed if we didn't do the sync teardown. So I'll try to find crash reports that indicate we're losing the race.

But yeah, I think waiting on the GPU main thread will be the best solution. Because that will solve the case where the destruction message comes in while an async release is already in progress, and we need to wait for it.

do you mean to post the codec release off the main thread from surfaceDestroyed, then wait (with timeout) on the main thread for it to complete?

which cases doesn't that cover?

Yeah, that should cover all cases minus where we timeout. But that should be rare. I can try that.

The following revision refers to this bug:
  https://chromium.googlesource.com/chromium/src.git/+/f1243d90b59f3361db000970862413cc12ddb6fa

commit f1243d90b59f3361db000970862413cc12ddb6fa
Author: watk <watk@chromium.org>
Date: Fri Nov 11 23:12:14 2016

Remove delayed error posting from AndroidVideoDecodeAccelerator

Now that we're always notifying AVDA when it loses its output surface
there's no need to delay errors. Errors are always "real" errors now.

BUG= 662599 
TEST=entering and exiting fullscreen still works
CQ_INCLUDE_TRYBOTS=master.tryserver.chromium.linux:linux_optional_gpu_tests_rel;master.tryserver.chromium.mac:mac_optional_gpu_tests_rel;master.tryserver.chromium.win:win_optional_gpu_tests_rel

Review-Url: https://codereview.chromium.org/2487813002
Cr-Commit-Position: refs/heads/master@{#431678}

[modify] https://crrev.com/f1243d90b59f3361db000970862413cc12ddb6fa/media/gpu/android_video_decode_accelerator.cc
[modify] https://crrev.com/f1243d90b59f3361db000970862413cc12ddb6fa/media/gpu/android_video_decode_accelerator.h
[modify] https://crrev.com/f1243d90b59f3361db000970862413cc12ddb6fa/media/gpu/avda_picture_buffer_manager.cc
[modify] https://crrev.com/f1243d90b59f3361db000970862413cc12ddb6fa/media/gpu/avda_state_provider.h

The following revision refers to this bug:
  https://chromium.googlesource.com/chromium/src.git/+/8b00de4de08282a2a667fb6a04c753f0964cc67e

commit 8b00de4de08282a2a667fb6a04c753f0964cc67e
Author: watk <watk@chromium.org>
Date: Wed Nov 23 00:45:41 2016

media: Do a TimedWait() for video surface teardown in AVDA

In OnSurfaceDestroyed() we have to release the MediaCodec synchronously
to avoid crashing or hanging. Previously we did the release
asynchronously because we post MediaCodec release to another thread.
Now that SurfaceDestroyed messages are enabled on all Android API
levels, and we don't have deferred error reporting this change
makes the teardown more robust by waiting for MediaCodecs to be
released before returning from OnSurfaceDestroyed().

* AVDACodecAllocator now tracks the allocation of surfaces to AVDAs.
* AVDACodecAllocator now allocates and releases codecs. When a codec
  with an attached surface is released, it tracks the completion of
  the async release task so that it can wait on it.
* AVDASurfaceTracker is gone. Now OnSurfaceDestroyed messages are
  routed through AVDACodecAllocator.

BUG= 662599 
TEST=new tests, manual fullscreen transitions
CQ_INCLUDE_TRYBOTS=master.tryserver.chromium.linux:linux_optional_gpu_tests_rel;master.tryserver.chromium.mac:mac_optional_gpu_tests_rel;master.tryserver.chromium.win:win_optional_gpu_tests_rel

Review-Url: https://codereview.chromium.org/2508053002
Cr-Commit-Position: refs/heads/master@{#434038}

[modify] https://crrev.com/8b00de4de08282a2a667fb6a04c753f0964cc67e/content/gpu/gpu_child_thread.cc
[modify] https://crrev.com/8b00de4de08282a2a667fb6a04c753f0964cc67e/media/gpu/BUILD.gn
[modify] https://crrev.com/8b00de4de08282a2a667fb6a04c753f0964cc67e/media/gpu/android_video_decode_accelerator.cc
[modify] https://crrev.com/8b00de4de08282a2a667fb6a04c753f0964cc67e/media/gpu/android_video_decode_accelerator.h
[modify] https://crrev.com/8b00de4de08282a2a667fb6a04c753f0964cc67e/media/gpu/avda_codec_allocator.cc
[modify] https://crrev.com/8b00de4de08282a2a667fb6a04c753f0964cc67e/media/gpu/avda_codec_allocator.h
[modify] https://crrev.com/8b00de4de08282a2a667fb6a04c753f0964cc67e/media/gpu/avda_codec_allocator_unittest.cc
[delete] https://crrev.com/50f2386d5fc8b1f32aa4e1a2340e73525a1088d0/media/gpu/avda_surface_tracker.cc
[delete] https://crrev.com/50f2386d5fc8b1f32aa4e1a2340e73525a1088d0/media/gpu/avda_surface_tracker.h