BLASTBufferQueue: Cap shadow queue size during sync

While waiting for the transaction commit callback on a previous
sync transaction BLASTBufferQueue will halt buffer processing to
ensure later frames do not arrive at SurfaceFlinger first. These
buffers wait in the shadow queue (tracked by mNumFrameAvailable)
to be acquired later. If we end up in a situation where all the
buffers are either in a sync transaction or in the shadow
queue then dequeue buffer will begin to block. This isn't
ideal, as dequeue buffer blocking can cause UI thread to block, aka
UI thread can block on RenderThread. However completing the
sync transaction (from a previous frame) can also depend on UI
thread, aka RenderThread can now block on UI thread (since
we need the transaction to apply in order to thaw the
shadow queue). In this CL we try and avoid that situation by
only keeping 1 frame in the shadow queue while waiting for the
sync to complete. If a second frame comes in we will acquire
and release the first before acquiring the second.

Bug: 200285149
Change-Id: I5072765e7b94820b3e66c557f5a96172ccef8172
Merged-In: I5072765e7b94820b3e66c557f5a96172ccef8172

1 files changed, 3 insertions, 0 deletions

diff --git a/libs/gui/BLASTBufferQueue.cpp b/libs/gui/BLASTBufferQueue.cpp
index f357c17e28..60c2e2e675 100644
--- a/libs/gui/BLASTBufferQueue.cpp
+++ b/libs/gui/BLASTBufferQueue.cpp
@@ -664,6 +664,9 @@ void BLASTBufferQueue::onFrameAvailable(const BufferItem& item) {
 
     // add to shadow queue
     mNumFrameAvailable++;
+    if (mWaitForTransactionCallback && mNumFrameAvailable == 2) {
+        acquireAndReleaseBuffer();
+    }
     ATRACE_INT(mQueuedBufferTrace.c_str(),
                mNumFrameAvailable + mNumAcquired - mPendingRelease.size());