Fix FenceTracker releaseFence

This patch:
* Fixes the release fence when GPU compositing.
* Stores the final release fence in ConsumerBase just
    before releasing the Buffer, which helps ensure
    sync points aren't added unknowningly.
* Makes HWC2 release pending buffers as the first step
    of postCompostion, rather than the last, which should
    allow dequeue to unblock a little earlier and helps
    make sure the previous buffer's release fence has
    been finalized before FenceTracker::addFrame is
    called.
* Fence tracker only sets the release fence once it
    has been finalized so it does not report a release
    fence for a buffer that is still latched.

Test: adb shell /data/nativetest/libgui_test/libgui_test
--gtest_filter=*GetFrameTimestamps*

Change-Id: I27d484bfd48f730bdcea2628f96795c6f4b4df7b

1 files changed, 3 insertions, 1 deletions

diff --git a/libs/gui/ConsumerBase.cpp b/libs/gui/ConsumerBase.cpp
index 3cf3078345..be2b1afd36 100644
--- a/libs/gui/ConsumerBase.cpp
+++ b/libs/gui/ConsumerBase.cpp
@@ -56,7 +56,8 @@ static int32_t createProcessUniqueId() {
 
 ConsumerBase::ConsumerBase(const sp<IGraphicBufferConsumer>& bufferQueue, bool controlledByApp) :
         mAbandoned(false),
-        mConsumer(bufferQueue) {
+        mConsumer(bufferQueue),
+        mPrevFinalReleaseFence(Fence::NO_FENCE) {
     // Choose a name using the PID and a process-unique ID.
     mName = String8::format("unnamed-%d-%d", getpid(), createProcessUniqueId());
 
@@ -366,6 +367,7 @@ status_t ConsumerBase::releaseBufferLocked(
         freeBufferLocked(slot);
     }
 
+    mPrevFinalReleaseFence = mSlots[slot].mFence;
     mSlots[slot].mFence = Fence::NO_FENCE;
 
     return err;