BufferQueueProducer: Throttle EGL frame production.

Throttling was previously controlled by a combination of the
driver and the number of buffers in the queue. This patch makes
a more consistent trade-off, which allows two GPU frames pending
but not three. More buffering could improve throughput in the
case of varying frame times, but this also increases latency.

Bug: 17502897
Change-Id: I4ee68019ca94c635294c5959931a555a6c4ef2df

include/gui/BufferQueueProducer.h

@@ -197,6 +197,12 @@ private:
 
     uint32_t mStickyTransform;
 
+    // This saves the fence from the last queueBuffer, such that the
+    // next queueBuffer call can throttle buffer production. The prior
+    // queueBuffer's fence is not nessessarily available elsewhere,
+    // since the previous buffer might have already been acquired.
+    sp<Fence> mLastQueueBufferFence;
+
 }; // class BufferQueueProducer
 
 } // namespace android

libs/gui/BufferQueueProducer.cpp

@@ -38,7 +38,8 @@ BufferQueueProducer::BufferQueueProducer(const sp<BufferQueueCore>& core) :
     mCore(core),
     mSlots(core->mSlots),
     mConsumerName(),
-    mStickyTransform(0) {}
+    mStickyTransform(0),
+    mLastQueueBufferFence(Fence::NO_FENCE) {}
 
 BufferQueueProducer::~BufferQueueProducer() {}
 
@@ -644,6 +645,15 @@ status_t BufferQueueProducer::queueBuffer(int slot,
         ATRACE_INT(mCore->mConsumerName.string(), mCore->mQueue.size());
     } // Autolock scope
 
+    // Wait without lock held
+    if (mCore->mConnectedApi == NATIVE_WINDOW_API_EGL) {
+        // Waiting here allows for two full buffers to be queued but not a
+        // third. In the event that frames take varying time, this makes a
+        // small trade-off in favor of latency rather than throughput.
+        mLastQueueBufferFence->waitForever("Throttling EGL Production");
+        mLastQueueBufferFence = fence;
+    }
+
     // Call back without lock held
     if (listener != NULL) {
         listener->onFrameAvailable();