GitLab

The internal class owns AdapterRefs which have to be destructed on the
main thread. As such, the last reference to the internal class had to
be freed on the main thread. In flaky cases, when tasks were posted to
the signaling thread, an operation performed and a task posted back to
the main thread - the main thread would finish before the signaling
thread task and the last reference to the internal class would be
released on the signaling thread, causing DCHECK crashes at
~AdapterRef for being on the wrong thread.

With destructor traits, if we are not already on the main thread we
post to the main thread and delete RTCRtpSenderInternal there.

The TearDown of RTCRtpSenderTest is updated to ensure that any pending
tasks get a chance to execute, in case the signaling thread was not
finished yet or else the destructor posted to the main thread does not
get a chance to execute and the test would flakily leak.

Before this CL: Flake's symptoms could be reproduced by adding a thread
sleep at RTCRtpSenderInternal::GetStatsOnSignalingThread. After this CL:
Unable to repro flake. The same applies for the other flaking tests
based on replaceTrack().

Bug: 827450
Change-Id: Ib594a53042e441e591ccc2c87ae8012bcb4ec75e
Reviewed-on: https://chromium-review.googlesource.com/1015002
Commit-Queue: Henrik Boström <hbos@chromium.org>
Reviewed-by: Wez <wez@chromium.org>
Reviewed-by: Guido Urdaneta <guidou@chromium.org>
Cr-Commit-Position: refs/heads/master@{#553550}