diff --git a/libs/gui/BLASTBufferQueue.cpp b/libs/gui/BLASTBufferQueue.cpp
index 82c9268..f778232 100644
--- a/libs/gui/BLASTBufferQueue.cpp
+++ b/libs/gui/BLASTBufferQueue.cpp
@@ -169,8 +169,6 @@
 
     mNumAcquired = 0;
     mNumFrameAvailable = 0;
-    mPendingReleaseItem.item = BufferItem();
-    mPendingReleaseItem.releaseFence = nullptr;
 }
 
 BLASTBufferQueue::~BLASTBufferQueue() {
@@ -242,7 +240,6 @@
         std::unique_lock _lock{mMutex};
         ATRACE_CALL();
         BQA_LOGV("transactionCallback");
-        mInitialCallbackReceived = true;
 
         if (!stats.empty()) {
             mTransformHint = stats[0].transformHint;
@@ -255,38 +252,20 @@
                                             stats[0].frameEventStats.compositorTiming,
                                             stats[0].latchTime,
                                             stats[0].frameEventStats.dequeueReadyTime);
-        }
-        if (mPendingReleaseItem.item.mGraphicBuffer != nullptr) {
-            if (!stats.empty()) {
-                mPendingReleaseItem.releaseFence = stats[0].previousReleaseFence;
-            } else {
-                BQA_LOGE("Warning: no SurfaceControlStats returned in BLASTBufferQueue callback");
-                mPendingReleaseItem.releaseFence = nullptr;
+            currFrameNumber = stats[0].frameEventStats.frameNumber;
+
+            if (mTransactionCompleteCallback &&
+                currFrameNumber >= mTransactionCompleteFrameNumber) {
+                if (currFrameNumber > mTransactionCompleteFrameNumber) {
+                    BQA_LOGE("transactionCallback received for a newer framenumber=%" PRIu64
+                             " than expected=%" PRIu64,
+                             currFrameNumber, mTransactionCompleteFrameNumber);
+                }
+                transactionCompleteCallback = std::move(mTransactionCompleteCallback);
+                mTransactionCompleteFrameNumber = 0;
             }
-            mBufferItemConsumer->releaseBuffer(mPendingReleaseItem.item,
-                                               mPendingReleaseItem.releaseFence
-                                                       ? mPendingReleaseItem.releaseFence
-                                                       : Fence::NO_FENCE);
-            mNumAcquired--;
-            mPendingReleaseItem.item = BufferItem();
-            mPendingReleaseItem.releaseFence = nullptr;
         }
 
-        if (mSubmitted.empty()) {
-            BQA_LOGE("ERROR: callback with no corresponding submitted buffer item");
-        }
-        mPendingReleaseItem.item = std::move(mSubmitted.front());
-        mSubmitted.pop();
-
-        processNextBufferLocked(false /* useNextTransaction */);
-
-        currFrameNumber = mPendingReleaseItem.item.mFrameNumber;
-        if (mTransactionCompleteCallback && mTransactionCompleteFrameNumber == currFrameNumber) {
-            transactionCompleteCallback = std::move(mTransactionCompleteCallback);
-            mTransactionCompleteFrameNumber = 0;
-        }
-
-        mCallbackCV.notify_all();
         decStrong((void*)transactionCallbackThunk);
     }
 
@@ -295,15 +274,46 @@
     }
 }
 
+// Unlike transactionCallbackThunk the release buffer callback does not extend the life of the
+// BBQ. This is because if the BBQ is destroyed, then the buffers will be released by the client.
+// So we pass in a weak pointer to the BBQ and if it still alive, then we release the buffer.
+// Otherwise, this is a no-op.
+static void releaseBufferCallbackThunk(wp<BLASTBufferQueue> context, uint64_t graphicBufferId,
+                                       const sp<Fence>& releaseFence) {
+    sp<BLASTBufferQueue> blastBufferQueue = context.promote();
+    ALOGV("releaseBufferCallbackThunk graphicBufferId=%" PRIu64 " blastBufferQueue=%s",
+          graphicBufferId, blastBufferQueue ? "alive" : "dead");
+    if (blastBufferQueue) {
+        blastBufferQueue->releaseBufferCallback(graphicBufferId, releaseFence);
+    }
+}
+
+void BLASTBufferQueue::releaseBufferCallback(uint64_t graphicBufferId,
+                                             const sp<Fence>& releaseFence) {
+    ATRACE_CALL();
+    std::unique_lock _lock{mMutex};
+    BQA_LOGV("releaseBufferCallback graphicBufferId=%" PRIu64, graphicBufferId);
+
+    auto it = mSubmitted.find(graphicBufferId);
+    if (it == mSubmitted.end()) {
+        BQA_LOGE("ERROR: releaseBufferCallback without corresponding submitted buffer %" PRIu64,
+                 graphicBufferId);
+        return;
+    }
+
+    mBufferItemConsumer->releaseBuffer(it->second, releaseFence);
+    mSubmitted.erase(it);
+    mNumAcquired--;
+    processNextBufferLocked(false /* useNextTransaction */);
+    mCallbackCV.notify_all();
+}
+
 void BLASTBufferQueue::processNextBufferLocked(bool useNextTransaction) {
     ATRACE_CALL();
-    BQA_LOGV("processNextBufferLocked useNextTransaction=%s", toString(useNextTransaction));
-
     // If the next transaction is set, we want to guarantee the our acquire will not fail, so don't
     // include the extra buffer when checking if we can acquire the next buffer.
     const bool includeExtraAcquire = !useNextTransaction;
     if (mNumFrameAvailable == 0 || maxBuffersAcquired(includeExtraAcquire)) {
-        BQA_LOGV("processNextBufferLocked waiting for frame available or callback");
         mCallbackCV.notify_all();
         return;
     }
@@ -353,7 +363,7 @@
     }
 
     mNumAcquired++;
-    mSubmitted.push(bufferItem);
+    mSubmitted[buffer->getId()] = bufferItem;
 
     bool needsDisconnect = false;
     mBufferItemConsumer->getConnectionEvents(bufferItem.mFrameNumber, &needsDisconnect);
@@ -369,7 +379,10 @@
     mLastBufferScalingMode = bufferItem.mScalingMode;
     mLastAcquiredFrameNumber = bufferItem.mFrameNumber;
 
-    t->setBuffer(mSurfaceControl, buffer);
+    auto releaseBufferCallback =
+            std::bind(releaseBufferCallbackThunk, wp<BLASTBufferQueue>(this) /* callbackContext */,
+                      std::placeholders::_1, std::placeholders::_2);
+    t->setBuffer(mSurfaceControl, buffer, releaseBufferCallback);
     t->setDataspace(mSurfaceControl, static_cast<ui::Dataspace>(bufferItem.mDataSpace));
     t->setHdrMetadata(mSurfaceControl, bufferItem.mHdrMetadata);
     t->setSurfaceDamageRegion(mSurfaceControl, bufferItem.mSurfaceDamage);
@@ -427,9 +440,12 @@
     }
 
     BQA_LOGV("processNextBufferLocked size=%dx%d mFrameNumber=%" PRIu64
-             " applyTransaction=%s mTimestamp=%" PRId64 " mPendingTransactions.size=%d",
+             " applyTransaction=%s mTimestamp=%" PRId64 "%s mPendingTransactions.size=%d"
+             " graphicBufferId=%" PRIu64,
              mSize.width, mSize.height, bufferItem.mFrameNumber, toString(applyTransaction),
-             bufferItem.mTimestamp, static_cast<uint32_t>(mPendingTransactions.size()));
+             bufferItem.mTimestamp, bufferItem.mIsAutoTimestamp ? "(auto)" : "",
+             static_cast<uint32_t>(mPendingTransactions.size()),
+             bufferItem.mGraphicBuffer->getId());
 }
 
 Rect BLASTBufferQueue::computeCrop(const BufferItem& item) {
@@ -444,18 +460,17 @@
     std::unique_lock _lock{mMutex};
 
     const bool nextTransactionSet = mNextTransaction != nullptr;
-    BQA_LOGV("onFrameAvailable framenumber=%" PRIu64 " nextTransactionSet=%s mFlushShadowQueue=%s",
-             item.mFrameNumber, toString(nextTransactionSet), toString(mFlushShadowQueue));
-
-    if (nextTransactionSet || mFlushShadowQueue) {
+    if (nextTransactionSet) {
         while (mNumFrameAvailable > 0 || maxBuffersAcquired(false /* includeExtraAcquire */)) {
             BQA_LOGV("waiting in onFrameAvailable...");
             mCallbackCV.wait(_lock);
         }
     }
-    mFlushShadowQueue = false;
     // add to shadow queue
     mNumFrameAvailable++;
+
+    BQA_LOGV("onFrameAvailable framenumber=%" PRIu64 " nextTransactionSet=%s", item.mFrameNumber,
+             toString(nextTransactionSet));
     processNextBufferLocked(nextTransactionSet /* useNextTransaction */);
 }
 
@@ -514,14 +529,12 @@
 }
 
 // Check if we have acquired the maximum number of buffers.
-// As a special case, we wait for the first callback before acquiring the second buffer so we
-// can ensure the first buffer is presented if multiple buffers are queued in succession.
 // Consumer can acquire an additional buffer if that buffer is not droppable. Set
 // includeExtraAcquire is true to include this buffer to the count. Since this depends on the state
 // of the buffer, the next acquire may return with NO_BUFFER_AVAILABLE.
 bool BLASTBufferQueue::maxBuffersAcquired(bool includeExtraAcquire) const {
     int maxAcquiredBuffers = MAX_ACQUIRED_BUFFERS + (includeExtraAcquire ? 2 : 1);
-    return mNumAcquired == maxAcquiredBuffers || (!mInitialCallbackReceived && mNumAcquired == 1);
+    return mNumAcquired == maxAcquiredBuffers;
 }
 
 class BBQSurface : public Surface {
diff --git a/libs/gui/ITransactionCompletedListener.cpp b/libs/gui/ITransactionCompletedListener.cpp
index 0ded936..9b5be1f 100644
--- a/libs/gui/ITransactionCompletedListener.cpp
+++ b/libs/gui/ITransactionCompletedListener.cpp
@@ -27,7 +27,8 @@
 
 enum class Tag : uint32_t {
     ON_TRANSACTION_COMPLETED = IBinder::FIRST_CALL_TRANSACTION,
-    LAST = ON_TRANSACTION_COMPLETED,
+    ON_RELEASE_BUFFER,
+    LAST = ON_RELEASE_BUFFER,
 };
 
 } // Anonymous namespace
@@ -122,6 +123,7 @@
     for (const auto& data : jankData) {
         SAFE_PARCEL(output->writeParcelable, data);
     }
+    SAFE_PARCEL(output->writeUint64, previousBufferId);
     return NO_ERROR;
 }
 
@@ -144,6 +146,7 @@
         SAFE_PARCEL(input->readParcelable, &data);
         jankData.push_back(data);
     }
+    SAFE_PARCEL(input->readUint64, &previousBufferId);
     return NO_ERROR;
 }
 
@@ -245,6 +248,12 @@
                                          onTransactionCompleted)>(Tag::ON_TRANSACTION_COMPLETED,
                                                                   stats);
     }
+
+    void onReleaseBuffer(uint64_t graphicBufferId, sp<Fence> releaseFence) override {
+        callRemoteAsync<decltype(
+                &ITransactionCompletedListener::onReleaseBuffer)>(Tag::ON_RELEASE_BUFFER,
+                                                                  graphicBufferId, releaseFence);
+    }
 };
 
 // Out-of-line virtual method definitions to trigger vtable emission in this translation unit (see
@@ -263,6 +272,8 @@
         case Tag::ON_TRANSACTION_COMPLETED:
             return callLocalAsync(data, reply,
                                   &ITransactionCompletedListener::onTransactionCompleted);
+        case Tag::ON_RELEASE_BUFFER:
+            return callLocalAsync(data, reply, &ITransactionCompletedListener::onReleaseBuffer);
     }
 }
 
diff --git a/libs/gui/LayerState.cpp b/libs/gui/LayerState.cpp
index 6bb8bf2..7a18ca6 100644
--- a/libs/gui/LayerState.cpp
+++ b/libs/gui/LayerState.cpp
@@ -63,7 +63,8 @@
         fixedTransformHint(ui::Transform::ROT_INVALID),
         frameNumber(0),
         frameTimelineInfo(),
-        autoRefresh(false) {
+        autoRefresh(false),
+        releaseBufferListener(nullptr) {
     matrix.dsdx = matrix.dtdy = 1.0f;
     matrix.dsdy = matrix.dtdx = 0.0f;
     hdrMetadata.validTypes = 0;
@@ -152,6 +153,7 @@
     SAFE_PARCEL(output.writeUint64, frameNumber);
     SAFE_PARCEL(frameTimelineInfo.write, output);
     SAFE_PARCEL(output.writeBool, autoRefresh);
+    SAFE_PARCEL(output.writeStrongBinder, IInterface::asBinder(releaseBufferListener));
 
     SAFE_PARCEL(output.writeUint32, blurRegions.size());
     for (auto region : blurRegions) {
@@ -275,6 +277,12 @@
     SAFE_PARCEL(frameTimelineInfo.read, input);
     SAFE_PARCEL(input.readBool, &autoRefresh);
 
+    tmpBinder = nullptr;
+    SAFE_PARCEL(input.readNullableStrongBinder, &tmpBinder);
+    if (tmpBinder) {
+        releaseBufferListener = checked_interface_cast<ITransactionCompletedListener>(tmpBinder);
+    }
+
     uint32_t numRegions = 0;
     SAFE_PARCEL(input.readUint32, &numRegions);
     blurRegions.clear();
@@ -543,6 +551,13 @@
         what |= eAutoRefreshChanged;
         autoRefresh = other.autoRefresh;
     }
+    if (other.what & eReleaseBufferListenerChanged) {
+        if (releaseBufferListener) {
+            ALOGW("Overriding releaseBufferListener");
+        }
+        what |= eReleaseBufferListenerChanged;
+        releaseBufferListener = other.releaseBufferListener;
+    }
     if (other.what & eStretchChanged) {
         what |= eStretchChanged;
         stretchEffect = other.stretchEffect;
diff --git a/libs/gui/SurfaceComposerClient.cpp b/libs/gui/SurfaceComposerClient.cpp
index c16a98f..11fe490 100644
--- a/libs/gui/SurfaceComposerClient.cpp
+++ b/libs/gui/SurfaceComposerClient.cpp
@@ -195,6 +195,17 @@
     }
 }
 
+void TransactionCompletedListener::setReleaseBufferCallback(uint64_t graphicBufferId,
+                                                            ReleaseBufferCallback listener) {
+    std::scoped_lock<std::mutex> lock(mMutex);
+    mReleaseBufferCallbacks[graphicBufferId] = listener;
+}
+
+void TransactionCompletedListener::removeReleaseBufferCallback(uint64_t graphicBufferId) {
+    std::scoped_lock<std::mutex> lock(mMutex);
+    mReleaseBufferCallbacks.erase(graphicBufferId);
+}
+
 void TransactionCompletedListener::addSurfaceStatsListener(void* context, void* cookie,
         sp<SurfaceControl> surfaceControl, SurfaceStatsCallback listener) {
     std::lock_guard<std::mutex> lock(mMutex);
@@ -275,6 +286,20 @@
                             .surfaceControls[surfaceStats.surfaceControl]
                             ->setTransformHint(surfaceStats.transformHint);
                 }
+                // If there is buffer id set, we look up any pending client release buffer callbacks
+                // and call them. This is a performance optimization when we have a transaction
+                // callback and a release buffer callback happening at the same time to avoid an
+                // additional ipc call from the server.
+                if (surfaceStats.previousBufferId) {
+                    ReleaseBufferCallback callback =
+                            popReleaseBufferCallbackLocked(surfaceStats.previousBufferId);
+                    if (callback) {
+                        callback(surfaceStats.previousBufferId,
+                                 surfaceStats.previousReleaseFence
+                                         ? surfaceStats.previousReleaseFence
+                                         : Fence::NO_FENCE);
+                    }
+                }
             }
 
             callbackFunction(transactionStats.latchTime, transactionStats.presentFence,
@@ -297,6 +322,32 @@
     }
 }
 
+void TransactionCompletedListener::onReleaseBuffer(uint64_t graphicBufferId,
+                                                   sp<Fence> releaseFence) {
+    ReleaseBufferCallback callback;
+    {
+        std::scoped_lock<std::mutex> lock(mMutex);
+        callback = popReleaseBufferCallbackLocked(graphicBufferId);
+    }
+    if (!callback) {
+        ALOGE("Could not call release buffer callback, buffer not found %" PRIu64, graphicBufferId);
+        return;
+    }
+    callback(graphicBufferId, releaseFence);
+}
+
+ReleaseBufferCallback TransactionCompletedListener::popReleaseBufferCallbackLocked(
+        uint64_t graphicBufferId) {
+    ReleaseBufferCallback callback;
+    auto itr = mReleaseBufferCallbacks.find(graphicBufferId);
+    if (itr == mReleaseBufferCallbacks.end()) {
+        return nullptr;
+    }
+    callback = itr->second;
+    mReleaseBufferCallbacks.erase(itr);
+    return callback;
+}
+
 // ---------------------------------------------------------------------------
 
 void removeDeadBufferCallback(void* /*context*/, uint64_t graphicBufferId);
@@ -1219,17 +1270,20 @@
 }
 
 SurfaceComposerClient::Transaction& SurfaceComposerClient::Transaction::setBuffer(
-        const sp<SurfaceControl>& sc, const sp<GraphicBuffer>& buffer) {
+        const sp<SurfaceControl>& sc, const sp<GraphicBuffer>& buffer,
+        ReleaseBufferCallback callback) {
     layer_state_t* s = getLayerState(sc);
     if (!s) {
         mStatus = BAD_INDEX;
         return *this;
     }
+    removeReleaseBufferCallback(s);
     s->what |= layer_state_t::eBufferChanged;
     s->buffer = buffer;
     if (mIsAutoTimestamp) {
         mDesiredPresentTime = systemTime();
     }
+    setReleaseBufferCallback(s, callback);
 
     registerSurfaceControlForCallback(sc);
 
@@ -1237,6 +1291,34 @@
     return *this;
 }
 
+void SurfaceComposerClient::Transaction::removeReleaseBufferCallback(layer_state_t* s) {
+    if (!s->releaseBufferListener) {
+        return;
+    }
+
+    s->what &= ~static_cast<uint64_t>(layer_state_t::eReleaseBufferListenerChanged);
+    s->releaseBufferListener = nullptr;
+    TransactionCompletedListener::getInstance()->removeReleaseBufferCallback(s->buffer->getId());
+}
+
+void SurfaceComposerClient::Transaction::setReleaseBufferCallback(layer_state_t* s,
+                                                                  ReleaseBufferCallback callback) {
+    if (!callback) {
+        return;
+    }
+
+    if (!s->buffer) {
+        ALOGW("Transaction::setReleaseBufferCallback"
+              "ignored trying to set a callback on a null buffer.");
+        return;
+    }
+
+    s->what |= layer_state_t::eReleaseBufferListenerChanged;
+    s->releaseBufferListener = TransactionCompletedListener::getIInstance();
+    auto listener = TransactionCompletedListener::getInstance();
+    listener->setReleaseBufferCallback(s->buffer->getId(), callback);
+}
+
 SurfaceComposerClient::Transaction& SurfaceComposerClient::Transaction::setAcquireFence(
         const sp<SurfaceControl>& sc, const sp<Fence>& fence) {
     layer_state_t* s = getLayerState(sc);
diff --git a/libs/gui/include/gui/BLASTBufferQueue.h b/libs/gui/include/gui/BLASTBufferQueue.h
index dd8e714..0173ffd 100644
--- a/libs/gui/include/gui/BLASTBufferQueue.h
+++ b/libs/gui/include/gui/BLASTBufferQueue.h
@@ -89,13 +89,14 @@
 
     void transactionCallback(nsecs_t latchTime, const sp<Fence>& presentFence,
             const std::vector<SurfaceControlStats>& stats);
+    void releaseBufferCallback(uint64_t graphicBufferId, const sp<Fence>& releaseFence);
     void setNextTransaction(SurfaceComposerClient::Transaction *t);
     void mergeWithNextTransaction(SurfaceComposerClient::Transaction* t, uint64_t frameNumber);
     void setTransactionCompleteCallback(uint64_t frameNumber,
                                         std::function<void(int64_t)>&& transactionCompleteCallback);
 
     void update(const sp<SurfaceControl>& surface, uint32_t width, uint32_t height, int32_t format);
-    void flushShadowQueue() { mFlushShadowQueue = true; }
+    void flushShadowQueue() {}
 
     status_t setFrameRate(float frameRate, int8_t compatibility, bool shouldBeSeamless);
     status_t setFrameTimelineInfo(const FrameTimelineInfo& info);
@@ -132,16 +133,10 @@
 
     int32_t mNumFrameAvailable GUARDED_BY(mMutex);
     int32_t mNumAcquired GUARDED_BY(mMutex);
-    bool mInitialCallbackReceived GUARDED_BY(mMutex) = false;
-    struct PendingReleaseItem {
-        BufferItem item;
-        sp<Fence> releaseFence;
-    };
 
-    std::queue<const BufferItem> mSubmitted GUARDED_BY(mMutex);
-    // Keep a reference to the currently presented buffer so we can release it when the next buffer
-    // is ready to be presented.
-    PendingReleaseItem mPendingReleaseItem GUARDED_BY(mMutex);
+    // Keep a reference to the submitted buffers so we can release when surfaceflinger drops the
+    // buffer or the buffer has been presented and a new buffer is ready to be presented.
+    std::unordered_map<uint64_t /* bufferId */, BufferItem> mSubmitted GUARDED_BY(mMutex);
 
     ui::Size mSize GUARDED_BY(mMutex);
     ui::Size mRequestedSize GUARDED_BY(mMutex);
@@ -157,9 +152,6 @@
     std::vector<std::tuple<uint64_t /* framenumber */, SurfaceComposerClient::Transaction>>
             mPendingTransactions GUARDED_BY(mMutex);
 
-    // If set to true, the next queue buffer will wait until the shadow queue has been processed by
-    // the adapter.
-    bool mFlushShadowQueue = false;
     // Last requested auto refresh state set by the producer. The state indicates that the consumer
     // should acquire the next frame as soon as it can and not wait for a frame to become available.
     // This is only relevant for shared buffer mode.
diff --git a/libs/gui/include/gui/ITransactionCompletedListener.h b/libs/gui/include/gui/ITransactionCompletedListener.h
index cb17cee..098760e 100644
--- a/libs/gui/include/gui/ITransactionCompletedListener.h
+++ b/libs/gui/include/gui/ITransactionCompletedListener.h
@@ -87,13 +87,14 @@
     SurfaceStats() = default;
     SurfaceStats(const sp<IBinder>& sc, nsecs_t time, const sp<Fence>& prevReleaseFence,
                  uint32_t hint, FrameEventHistoryStats frameEventStats,
-                 std::vector<JankData> jankData)
+                 std::vector<JankData> jankData, uint64_t previousBufferId)
           : surfaceControl(sc),
             acquireTime(time),
             previousReleaseFence(prevReleaseFence),
             transformHint(hint),
             eventStats(frameEventStats),
-            jankData(std::move(jankData)) {}
+            jankData(std::move(jankData)),
+            previousBufferId(previousBufferId) {}
 
     sp<IBinder> surfaceControl;
     nsecs_t acquireTime = -1;
@@ -101,6 +102,7 @@
     uint32_t transformHint = 0;
     FrameEventHistoryStats eventStats;
     std::vector<JankData> jankData;
+    uint64_t previousBufferId;
 };
 
 class TransactionStats : public Parcelable {
@@ -139,6 +141,8 @@
     DECLARE_META_INTERFACE(TransactionCompletedListener)
 
     virtual void onTransactionCompleted(ListenerStats stats) = 0;
+
+    virtual void onReleaseBuffer(uint64_t graphicBufferId, sp<Fence> releaseFence) = 0;
 };
 
 class BnTransactionCompletedListener : public SafeBnInterface<ITransactionCompletedListener> {
diff --git a/libs/gui/include/gui/LayerState.h b/libs/gui/include/gui/LayerState.h
index 183ec40..d68a9cf 100644
--- a/libs/gui/include/gui/LayerState.h
+++ b/libs/gui/include/gui/LayerState.h
@@ -100,7 +100,7 @@
         eLayerStackChanged = 0x00000080,
         /* was eCropChanged_legacy, now available 0x00000100, */
         eDeferTransaction_legacy = 0x00000200,
-        /* was ScalingModeChanged, now available 0x00000400, */
+        eReleaseBufferListenerChanged = 0x00000400,
         eShadowRadiusChanged = 0x00000800,
         eReparentChildren = 0x00001000,
         /* was eDetachChildren, now available 0x00002000, */
@@ -248,6 +248,11 @@
 
     // Stretch effect to be applied to this layer
     StretchEffect stretchEffect;
+
+    // Listens to when the buffer is safe to be released. This is used for blast
+    // layers only. The callback includes a release fence as well as the graphic
+    // buffer id to identify the buffer.
+    sp<ITransactionCompletedListener> releaseBufferListener;
 };
 
 struct ComposerState {
diff --git a/libs/gui/include/gui/SurfaceComposerClient.h b/libs/gui/include/gui/SurfaceComposerClient.h
index 88484c0..f29983c 100644
--- a/libs/gui/include/gui/SurfaceComposerClient.h
+++ b/libs/gui/include/gui/SurfaceComposerClient.h
@@ -80,6 +80,9 @@
 using TransactionCompletedCallback =
         std::function<void(nsecs_t /*latchTime*/, const sp<Fence>& /*presentFence*/,
                            const std::vector<SurfaceControlStats>& /*stats*/)>;
+using ReleaseBufferCallback =
+        std::function<void(uint64_t /* graphicsBufferId */, const sp<Fence>& /*releaseFence*/)>;
+
 using SurfaceStatsCallback =
         std::function<void(void* /*context*/, nsecs_t /*latchTime*/,
                            const sp<Fence>& /*presentFence*/,
@@ -388,6 +391,8 @@
 
         void cacheBuffers();
         void registerSurfaceControlForCallback(const sp<SurfaceControl>& sc);
+        void setReleaseBufferCallback(layer_state_t* state, ReleaseBufferCallback callback);
+        void removeReleaseBufferCallback(layer_state_t* state);
 
     public:
         Transaction();
@@ -471,7 +476,8 @@
         Transaction& setTransformToDisplayInverse(const sp<SurfaceControl>& sc,
                                                   bool transformToDisplayInverse);
         Transaction& setFrame(const sp<SurfaceControl>& sc, const Rect& frame);
-        Transaction& setBuffer(const sp<SurfaceControl>& sc, const sp<GraphicBuffer>& buffer);
+        Transaction& setBuffer(const sp<SurfaceControl>& sc, const sp<GraphicBuffer>& buffer,
+                               ReleaseBufferCallback callback = nullptr);
         Transaction& setCachedBuffer(const sp<SurfaceControl>& sc, int32_t bufferId);
         Transaction& setAcquireFence(const sp<SurfaceControl>& sc, const sp<Fence>& fence);
         Transaction& setDataspace(const sp<SurfaceControl>& sc, ui::Dataspace dataspace);
@@ -650,6 +656,8 @@
 
     std::unordered_map<CallbackId, CallbackTranslation> mCallbacks GUARDED_BY(mMutex);
     std::multimap<sp<IBinder>, sp<JankDataListener>> mJankListeners GUARDED_BY(mMutex);
+    std::unordered_map<uint64_t /* graphicsBufferId */, ReleaseBufferCallback>
+            mReleaseBufferCallbacks GUARDED_BY(mMutex);
     std::multimap<sp<IBinder>, SurfaceStatsCallbackEntry>
                 mSurfaceStatsListeners GUARDED_BY(mMutex);
 
@@ -683,8 +691,15 @@
                 SurfaceStatsCallback listener);
     void removeSurfaceStatsListener(void* context, void* cookie);
 
-    // Overrides BnTransactionCompletedListener's onTransactionCompleted
+    void setReleaseBufferCallback(uint64_t /* graphicsBufferId */, ReleaseBufferCallback);
+    void removeReleaseBufferCallback(uint64_t /* graphicsBufferId */);
+
+    // BnTransactionCompletedListener overrides
     void onTransactionCompleted(ListenerStats stats) override;
+    void onReleaseBuffer(uint64_t /* graphicsBufferId */, sp<Fence> releaseFence) override;
+
+private:
+    ReleaseBufferCallback popReleaseBufferCallbackLocked(uint64_t /* graphicsBufferId */);
 };
 
 } // namespace android
diff --git a/libs/gui/tests/BLASTBufferQueue_test.cpp b/libs/gui/tests/BLASTBufferQueue_test.cpp
index 7895e99..fe48d88 100644
--- a/libs/gui/tests/BLASTBufferQueue_test.cpp
+++ b/libs/gui/tests/BLASTBufferQueue_test.cpp
@@ -73,13 +73,34 @@
 
     void waitForCallbacks() {
         std::unique_lock lock{mBlastBufferQueueAdapter->mMutex};
-        while (mBlastBufferQueueAdapter->mSubmitted.size() > 0) {
+        // Wait until all but one of the submitted buffers have been released.
+        while (mBlastBufferQueueAdapter->mSubmitted.size() > 1) {
             mBlastBufferQueueAdapter->mCallbackCV.wait(lock);
         }
     }
 
+    void setTransactionCompleteCallback(int64_t frameNumber) {
+        mBlastBufferQueueAdapter->setTransactionCompleteCallback(frameNumber, [&](int64_t frame) {
+            std::unique_lock lock{mMutex};
+            mLastTransactionCompleteFrameNumber = frame;
+            mCallbackCV.notify_all();
+        });
+    }
+
+    void waitForCallback(int64_t frameNumber) {
+        std::unique_lock lock{mMutex};
+        // Wait until all but one of the submitted buffers have been released.
+        while (mLastTransactionCompleteFrameNumber < frameNumber) {
+            mCallbackCV.wait(lock);
+        }
+    }
+
 private:
     sp<BLASTBufferQueue> mBlastBufferQueueAdapter;
+
+    std::mutex mMutex;
+    std::condition_variable mCallbackCV;
+    int64_t mLastTransactionCompleteFrameNumber = -1;
 };
 
 class BLASTBufferQueueTest : public ::testing::Test {
@@ -128,7 +149,7 @@
         mCaptureArgs.dataspace = ui::Dataspace::V0_SRGB;
     }
 
-    void setUpProducer(BLASTBufferQueueHelper adapter, sp<IGraphicBufferProducer>& producer) {
+    void setUpProducer(BLASTBufferQueueHelper& adapter, sp<IGraphicBufferProducer>& producer) {
         producer = adapter.getIGraphicBufferProducer();
         setUpProducer(producer);
     }
@@ -205,10 +226,10 @@
                     EXPECT_GE(epsilon, abs(g - *(pixel + 1)));
                     EXPECT_GE(epsilon, abs(b - *(pixel + 2)));
                 }
+                ASSERT_EQ(false, ::testing::Test::HasFailure());
             }
         }
         captureBuf->unlock();
-        ASSERT_EQ(false, ::testing::Test::HasFailure());
     }
 
     static status_t captureDisplay(DisplayCaptureArgs& captureArgs,
@@ -315,7 +336,8 @@
 
     nsecs_t desiredPresentTime = systemTime() + nsecs_t(5 * 1e8);
     IGraphicBufferProducer::QueueBufferOutput qbOutput;
-    IGraphicBufferProducer::QueueBufferInput input(desiredPresentTime, false, HAL_DATASPACE_UNKNOWN,
+    IGraphicBufferProducer::QueueBufferInput input(desiredPresentTime, true /* autotimestamp */,
+                                                   HAL_DATASPACE_UNKNOWN,
                                                    Rect(mDisplayWidth, mDisplayHeight),
                                                    NATIVE_WINDOW_SCALING_MODE_FREEZE, 0,
                                                    Fence::NO_FENCE);
@@ -351,7 +373,8 @@
     buf->unlock();
 
     IGraphicBufferProducer::QueueBufferOutput qbOutput;
-    IGraphicBufferProducer::QueueBufferInput input(systemTime(), false, HAL_DATASPACE_UNKNOWN,
+    IGraphicBufferProducer::QueueBufferInput input(systemTime(), true /* autotimestamp */,
+                                                   HAL_DATASPACE_UNKNOWN,
                                                    Rect(mDisplayWidth, mDisplayHeight),
                                                    NATIVE_WINDOW_SCALING_MODE_FREEZE, 0,
                                                    Fence::NO_FENCE);
@@ -396,7 +419,8 @@
                                               nullptr, nullptr);
         ASSERT_EQ(NO_ERROR, ret);
         IGraphicBufferProducer::QueueBufferOutput qbOutput;
-        IGraphicBufferProducer::QueueBufferInput input(systemTime(), false, HAL_DATASPACE_UNKNOWN,
+        IGraphicBufferProducer::QueueBufferInput input(systemTime(), true /* autotimestamp */,
+                                                       HAL_DATASPACE_UNKNOWN,
                                                        Rect(mDisplayWidth, mDisplayHeight),
                                                        NATIVE_WINDOW_SCALING_MODE_FREEZE, 0,
                                                        Fence::NO_FENCE);
@@ -429,7 +453,8 @@
     buf->unlock();
 
     IGraphicBufferProducer::QueueBufferOutput qbOutput;
-    IGraphicBufferProducer::QueueBufferInput input(systemTime(), false, HAL_DATASPACE_UNKNOWN,
+    IGraphicBufferProducer::QueueBufferInput input(systemTime(), true /* autotimestamp */,
+                                                   HAL_DATASPACE_UNKNOWN,
                                                    Rect(mDisplayWidth, mDisplayHeight / 2),
                                                    NATIVE_WINDOW_SCALING_MODE_FREEZE, 0,
                                                    Fence::NO_FENCE);
@@ -486,7 +511,8 @@
     buf->unlock();
 
     IGraphicBufferProducer::QueueBufferOutput qbOutput;
-    IGraphicBufferProducer::QueueBufferInput input(systemTime(), false, HAL_DATASPACE_UNKNOWN,
+    IGraphicBufferProducer::QueueBufferInput input(systemTime(), true /* autotimestamp */,
+                                                   HAL_DATASPACE_UNKNOWN,
                                                    Rect(bufferSideLength, finalCropSideLength),
                                                    NATIVE_WINDOW_SCALING_MODE_SCALE_CROP, 0,
                                                    Fence::NO_FENCE);
@@ -537,7 +563,8 @@
         ASSERT_EQ(IGraphicBufferProducer::BUFFER_NEEDS_REALLOCATION, ret);
         ASSERT_EQ(OK, igbProducer->requestBuffer(slot, &buf));
         IGraphicBufferProducer::QueueBufferOutput qbOutput;
-        IGraphicBufferProducer::QueueBufferInput input(systemTime(), false, HAL_DATASPACE_UNKNOWN,
+        IGraphicBufferProducer::QueueBufferInput input(systemTime(), true /* autotimestamp */,
+                                                       HAL_DATASPACE_UNKNOWN,
                                                        Rect(mDisplayWidth, mDisplayHeight),
                                                        NATIVE_WINDOW_SCALING_MODE_FREEZE, 0,
                                                        Fence::NO_FENCE);
@@ -577,7 +604,7 @@
     sp<IGraphicBufferProducer> slowIgbProducer;
     setUpProducer(slowAdapter, slowIgbProducer);
     nsecs_t presentTimeDelay = std::chrono::nanoseconds(500ms).count();
-    queueBuffer(slowIgbProducer, 0 /* r */, 0 /* g */, 0 /* b */, presentTimeDelay);
+    queueBuffer(slowIgbProducer, 0 /* r */, 255 /* g */, 0 /* b */, presentTimeDelay);
 
     BLASTBufferQueueHelper fastAdapter(bgSurface, mDisplayWidth, mDisplayHeight);
     sp<IGraphicBufferProducer> fastIgbProducer;
@@ -617,7 +644,8 @@
         fillQuadrants(buf);
 
         IGraphicBufferProducer::QueueBufferOutput qbOutput;
-        IGraphicBufferProducer::QueueBufferInput input(systemTime(), false, HAL_DATASPACE_UNKNOWN,
+        IGraphicBufferProducer::QueueBufferInput input(systemTime(), true /* autotimestamp */,
+                                                       HAL_DATASPACE_UNKNOWN,
                                                        Rect(bufWidth, bufHeight),
                                                        NATIVE_WINDOW_SCALING_MODE_SCALE_TO_WINDOW,
                                                        tr, Fence::NO_FENCE);
@@ -838,6 +866,7 @@
     IGraphicBufferProducer::QueueBufferOutput qbOutput;
     nsecs_t requestedPresentTimeA = 0;
     nsecs_t postedTimeA = 0;
+    adapter.setTransactionCompleteCallback(1);
     setUpAndQueueBuffer(igbProducer, &requestedPresentTimeA, &postedTimeA, &qbOutput, true);
     history.applyDelta(qbOutput.frameTimestamps);
 
@@ -848,7 +877,7 @@
     ASSERT_EQ(requestedPresentTimeA, events->requestedPresentTime);
     ASSERT_GE(events->postedTime, postedTimeA);
 
-    adapter.waitForCallbacks();
+    adapter.waitForCallback(1);
 
     // queue another buffer so we query for frame event deltas
     nsecs_t requestedPresentTimeB = 0;