diff --git a/services/surfaceflinger/RegionSamplingThread.cpp b/services/surfaceflinger/RegionSamplingThread.cpp
index e579eb0..1eb865d 100644
--- a/services/surfaceflinger/RegionSamplingThread.cpp
+++ b/services/surfaceflinger/RegionSamplingThread.cpp
@@ -356,9 +356,8 @@
     if (FlagManager::getInstance().single_hop_screenshot() &&
         FlagManager::getInstance().ce_fence_promise()) {
         std::vector<sp<LayerFE>> layerFEs;
-        auto displayState =
-                mFlinger.getDisplayAndLayerSnapshotsFromMainThread(renderAreaBuilder,
-                                                                   getLayerSnapshotsFn, layerFEs);
+        auto displayState = mFlinger.getSnapshotsFromMainThread(renderAreaBuilder,
+                                                                getLayerSnapshotsFn, layerFEs);
         fenceResult =
                 mFlinger.captureScreenshot(renderAreaBuilder, buffer, kRegionSampling, kGrayscale,
                                            kIsProtected, nullptr, displayState, layerFEs)
diff --git a/services/surfaceflinger/SurfaceFlinger.cpp b/services/surfaceflinger/SurfaceFlinger.cpp
index e65efc5..4d64c10 100644
--- a/services/surfaceflinger/SurfaceFlinger.cpp
+++ b/services/surfaceflinger/SurfaceFlinger.cpp
@@ -2907,6 +2907,7 @@
         }
     }
 
+    ATRACE_NAME("postComposition");
     mTimeStats->recordFrameDuration(pacesetterTarget.frameBeginTime().ns(), systemTime());
 
     // Send a power hint after presentation is finished.
@@ -8161,12 +8162,12 @@
 // Accessing display requires mStateLock, and contention for this lock
 // is reduced when grabbed from the main thread, thus also reducing
 // risk of deadlocks.
-std::optional<SurfaceFlinger::OutputCompositionState>
-SurfaceFlinger::getDisplayAndLayerSnapshotsFromMainThread(
+std::optional<SurfaceFlinger::OutputCompositionState> SurfaceFlinger::getSnapshotsFromMainThread(
         RenderAreaBuilderVariant& renderAreaBuilder, GetLayerSnapshotsFunction getLayerSnapshotsFn,
         std::vector<sp<LayerFE>>& layerFEs) {
     return mScheduler
             ->schedule([=, this, &renderAreaBuilder, &layerFEs]() REQUIRES(kMainThreadContext) {
+                ATRACE_NAME("getSnapshotsFromMainThread");
                 auto layers = getLayerSnapshotsFn();
                 for (auto& [layer, layerFE] : layers) {
                     attachReleaseFenceFutureToLayer(layer, layerFE.get(), ui::INVALID_LAYER_STACK);
@@ -8196,8 +8197,7 @@
         FlagManager::getInstance().ce_fence_promise()) {
         std::vector<sp<LayerFE>> layerFEs;
         auto displayState =
-                getDisplayAndLayerSnapshotsFromMainThread(renderAreaBuilder, getLayerSnapshotsFn,
-                                                          layerFEs);
+                getSnapshotsFromMainThread(renderAreaBuilder, getLayerSnapshotsFn, layerFEs);
 
         const bool supportsProtected = getRenderEngine().supportsProtectedContent();
         bool hasProtectedLayer = false;
diff --git a/services/surfaceflinger/SurfaceFlinger.h b/services/surfaceflinger/SurfaceFlinger.h
index 7762bbe..988edfe 100644
--- a/services/surfaceflinger/SurfaceFlinger.h
+++ b/services/surfaceflinger/SurfaceFlinger.h
@@ -896,7 +896,7 @@
 
     using OutputCompositionState = compositionengine::impl::OutputCompositionState;
 
-    std::optional<OutputCompositionState> getDisplayAndLayerSnapshotsFromMainThread(
+    std::optional<OutputCompositionState> getSnapshotsFromMainThread(
             RenderAreaBuilderVariant& renderAreaBuilder,
             GetLayerSnapshotsFunction getLayerSnapshotsFn, std::vector<sp<LayerFE>>& layerFEs);
 
diff --git a/services/surfaceflinger/TransactionCallbackInvoker.cpp b/services/surfaceflinger/TransactionCallbackInvoker.cpp
index 222ae30..8721538 100644
--- a/services/surfaceflinger/TransactionCallbackInvoker.cpp
+++ b/services/surfaceflinger/TransactionCallbackInvoker.cpp
@@ -26,6 +26,7 @@
 #include "TransactionCallbackInvoker.h"
 #include "BackgroundExecutor.h"
 #include "Utils/FenceUtils.h"
+#include "utils/Trace.h"
 
 #include <cinttypes>
 
@@ -164,7 +165,7 @@
 void TransactionCallbackInvoker::sendCallbacks(bool onCommitOnly) {
     // For each listener
     auto completedTransactionsItr = mCompletedTransactions.begin();
-    BackgroundExecutor::Callbacks callbacks;
+    ftl::SmallVector<ListenerStats, 10> listenerStatsToSend;
     while (completedTransactionsItr != mCompletedTransactions.end()) {
         auto& [listener, transactionStatsDeque] = *completedTransactionsItr;
         ListenerStats listenerStats;
@@ -199,10 +200,7 @@
                 // keep it as an IBinder due to consistency reasons: if we
                 // interface_cast at the IPC boundary when reading a Parcel,
                 // we get pointers that compare unequal in the SF process.
-                callbacks.emplace_back([stats = std::move(listenerStats)]() {
-                    interface_cast<ITransactionCompletedListener>(stats.listener)
-                            ->onTransactionCompleted(stats);
-                });
+                listenerStatsToSend.emplace_back(std::move(listenerStats));
             }
         }
         completedTransactionsItr++;
@@ -212,7 +210,14 @@
         mPresentFence.clear();
     }
 
-    BackgroundExecutor::getInstance().sendCallbacks(std::move(callbacks));
+    BackgroundExecutor::getInstance().sendCallbacks(
+            {[listenerStatsToSend = std::move(listenerStatsToSend)]() {
+                ATRACE_NAME("TransactionCallbackInvoker::sendCallbacks");
+                for (auto& stats : listenerStatsToSend) {
+                    interface_cast<ITransactionCompletedListener>(stats.listener)
+                            ->onTransactionCompleted(stats);
+                }
+            }});
 }
 
 // -----------------------------------------------------------------------