tracing: Stop data sources when system service disconnects
Today, when the SDK disconnects from the system backend, all the active
data source instances will keep being enabled. This has two problems:
* While there will be no memory corruption, because the data source
instance will get a NullTraceWriter(), it's better to stop tracing, to
avoid the overhead.
* On reconnection, if the system service restarts tracing with the same
config, the TracingMuxer will get confused because it will avoid
setting up a data source with the same config (thinking it's already
enabled), but it will fail to start it later.
This commit fixes the problem by stopping the data sources when a
disconnection happens.
Since we're now explicitly stopping data sources, we can avoid
incrementing the generation counter, because:
* It is not necessary (it will be incremented by
StopDataSource_AsyncEnd() after a release-store on instances if
required).
* Incrementing the generation counter before stopping the data sources
exposes an existing TSAN race condition in the new test:
DestroyStoppedTraceWritersForCurrentThread() reads DataSourceState
fields while (much later, but it's still a race condition)
SetupDataSourceImpl writes them.
Change-Id: I82ac05b39b6f80c35aeb4a342c2dd3757d478723
diff --git a/test/test_helper.cc b/test/test_helper.cc
index bc8cfdf..2267dc0 100644
--- a/test/test_helper.cc
+++ b/test/test_helper.cc
@@ -116,6 +116,12 @@
env_cleaner_ = service_thread_.Start();
}
+void TestHelper::RestartService() {
+ PERFETTO_CHECK(mode_ == Mode::kStartDaemons);
+ service_thread_.Stop();
+ service_thread_.Start();
+}
+
FakeProducer* TestHelper::ConnectFakeProducer() {
fake_producer_thread_.Connect();
// This will wait until the service has seen the RegisterDataSource() call
