diff --git a/services/surfaceflinger/Layer.h b/services/surfaceflinger/Layer.h
index 7335be7..76c710a 100644
--- a/services/surfaceflinger/Layer.h
+++ b/services/surfaceflinger/Layer.h
@@ -376,13 +376,20 @@
 #ifdef USE_HWC2
     // -----------------------------------------------------------------------
 
+    void eraseHwcLayer(int32_t hwcId) {
+        mHwcLayers.erase(hwcId);
+
+        Mutex::Autolock lock(mHwcBufferCacheMutex);
+        mHwcBufferCaches.erase(hwcId);
+    }
+
     bool hasHwcLayer(int32_t hwcId) {
         if (mHwcLayers.count(hwcId) == 0) {
             return false;
         }
         if (mHwcLayers[hwcId].layer->isAbandoned()) {
             ALOGI("Erasing abandoned layer %s on %d", mName.string(), hwcId);
-            mHwcLayers.erase(hwcId);
+            eraseHwcLayer(hwcId);
             return false;
         }
         return true;
@@ -398,8 +405,11 @@
     void setHwcLayer(int32_t hwcId, std::shared_ptr<HWC2::Layer>&& layer) {
         if (layer) {
             mHwcLayers[hwcId].layer = layer;
+
+            Mutex::Autolock lock(mHwcBufferCacheMutex);
+            mHwcBufferCaches[hwcId] = HWComposerBufferCache();
         } else {
-            mHwcLayers.erase(hwcId);
+            eraseHwcLayer(hwcId);
         }
     }
 
@@ -489,6 +499,7 @@
     // Interface implementation for SurfaceFlingerConsumer::ContentsChangedListener
     virtual void onFrameAvailable(const BufferItem& item) override;
     virtual void onFrameReplaced(const BufferItem& item) override;
+    virtual void onBuffersReleased() override;
     virtual void onSidebandStreamChanged() override;
 
     void commitTransaction(const State& stateToCommit);
@@ -642,6 +653,7 @@
     FenceTimeline mReleaseTimeline;
 
     // main thread
+    int mActiveBufferSlot = BufferQueue::INVALID_BUFFER_SLOT;
     sp<GraphicBuffer> mActiveBuffer;
     sp<NativeHandle> mSidebandStream;
     Rect mCurrentCrop;
@@ -681,6 +693,12 @@
         gfx::FloatRect sourceCrop;
     };
     std::unordered_map<int32_t, HWCInfo> mHwcLayers;
+
+    // We need one HWComposerBufferCache for each HWC display.  We cannot have
+    // HWComposerBufferCache in HWCInfo because HWCInfo can only be accessed
+    // from the main thread.
+    Mutex mHwcBufferCacheMutex;
+    std::unordered_map<int32_t, HWComposerBufferCache> mHwcBufferCaches;
 #else
     bool mIsGlesComposition;
 #endif