SurfaceFlinger: Use a lockless stack for binder->tracing thread
Though the critical section is kept very small between the binder
and tracing threads, the tracing thread is not RT, and so we still
run the risk of it being descheduled and blocking one of our binder
threads excessively. In this thread we provide a simple lockless
stack implementation and use it to push pending transactions from
the binder thread to the main thread. We also looked up "layer IDs"
on the binder thread, which is no longer possible without the lock.
To work around this we store the pointer itself in the proto until
it reaches the tracing thread where we overwrite it with the mapped
value.
Bug: 200284593
Test: Existing tests pass
Change-Id: I408dc87ddfe088b68f65601455147a96b870627d
diff --git a/services/surfaceflinger/Tracing/TransactionTracing.h b/services/surfaceflinger/Tracing/TransactionTracing.h
index 95256c4..4c291f9 100644
--- a/services/surfaceflinger/Tracing/TransactionTracing.h
+++ b/services/surfaceflinger/Tracing/TransactionTracing.h
@@ -26,6 +26,7 @@
#include <thread>
#include "RingBuffer.h"
+#include "LocklessStack.h"
#include "TransactionProtoParser.h"
using namespace android::surfaceflinger;
@@ -78,6 +79,7 @@
size_t mBufferSizeInBytes GUARDED_BY(mTraceLock) = CONTINUOUS_TRACING_BUFFER_SIZE;
std::unordered_map<uint64_t, proto::TransactionState> mQueuedTransactions
GUARDED_BY(mTraceLock);
+ LocklessStack<proto::TransactionState> mTransactionQueue;
nsecs_t mStartingTimestamp GUARDED_BY(mTraceLock);
std::vector<proto::LayerCreationArgs> mCreatedLayers GUARDED_BY(mTraceLock);
std::unordered_map<BBinder* /* layerHandle */, int32_t /* layerId */> mLayerHandles
@@ -85,6 +87,9 @@
std::vector<int32_t /* layerId */> mRemovedLayerHandles GUARDED_BY(mTraceLock);
std::map<int32_t /* layerId */, TracingLayerState> mStartingStates GUARDED_BY(mTraceLock);
TransactionProtoParser mProtoParser GUARDED_BY(mTraceLock);
+ // Parses the transaction to proto without holding any tracing locks so we can generate proto
+ // in the binder thread without any contention.
+ TransactionProtoParser mLockfreeProtoParser;
// We do not want main thread to block so main thread will try to acquire mMainThreadLock,
// otherwise will push data to temporary container.