diff --git a/include/media/stagefright/MediaCodecSource.h b/include/media/stagefright/MediaCodecSource.h
index cc62786..00b2c71 100644
--- a/include/media/stagefright/MediaCodecSource.h
+++ b/include/media/stagefright/MediaCodecSource.h
@@ -50,6 +50,7 @@
     bool isVideo() const { return mIsVideo; }
     sp<IGraphicBufferProducer> getGraphicBufferProducer();
     void setInputBufferTimeOffset(int64_t timeOffsetUs);
+    int64_t getFirstSampleSystemTimeUs();
 
     // MediaSource
     virtual status_t start(MetaData *params = NULL);
@@ -79,6 +80,7 @@
         kWhatStop,
         kWhatPause,
         kWhatSetInputBufferTimeOffset,
+        kWhatGetFirstSampleSystemTimeUs,
         kWhatStopStalled,
     };
 
@@ -90,6 +92,7 @@
             uint32_t flags = 0);
 
     status_t onStart(MetaData *params);
+    void onPause();
     status_t init();
     status_t initEncoder();
     void releaseEncoder();
@@ -123,6 +126,8 @@
     List<size_t> mAvailEncoderInputIndices;
     List<int64_t> mDecodingTimeQueue; // decoding time (us) for video
     int64_t mInputBufferTimeOffsetUs;
+    int64_t mFirstSampleSystemTimeUs;
+    bool mPausePending;
 
     // audio drift time
     int64_t mFirstSampleTimeUs;
diff --git a/media/libmediaplayerservice/StagefrightRecorder.cpp b/media/libmediaplayerservice/StagefrightRecorder.cpp
index 24ca582..bd1fd7c 100644
--- a/media/libmediaplayerservice/StagefrightRecorder.cpp
+++ b/media/libmediaplayerservice/StagefrightRecorder.cpp
@@ -1807,19 +1807,38 @@
         return OK;
     }
 
-    // 30 ms buffer to avoid timestamp overlap
-    mTotalPausedDurationUs += (systemTime() / 1000) - mPauseStartTimeUs - 30000;
+    int64_t bufferStartTimeUs = 0;
+    bool allSourcesStarted = true;
+    for (const auto &source : { mAudioEncoderSource, mVideoEncoderSource }) {
+        if (source == nullptr) {
+            continue;
+        }
+        int64_t timeUs = source->getFirstSampleSystemTimeUs();
+        if (timeUs < 0) {
+            allSourcesStarted = false;
+        }
+        if (bufferStartTimeUs < timeUs) {
+            bufferStartTimeUs = timeUs;
+        }
+    }
+
+    if (allSourcesStarted) {
+        if (mPauseStartTimeUs < bufferStartTimeUs) {
+            mPauseStartTimeUs = bufferStartTimeUs;
+        }
+        // 30 ms buffer to avoid timestamp overlap
+        mTotalPausedDurationUs += (systemTime() / 1000) - mPauseStartTimeUs - 30000;
+    }
     double timeOffset = -mTotalPausedDurationUs;
     if (mCaptureFpsEnable) {
         timeOffset *= mCaptureFps / mFrameRate;
     }
-    if (mAudioEncoderSource != NULL) {
-        mAudioEncoderSource->setInputBufferTimeOffset((int64_t)timeOffset);
-        mAudioEncoderSource->start();
-    }
-    if (mVideoEncoderSource != NULL) {
-        mVideoEncoderSource->setInputBufferTimeOffset((int64_t)timeOffset);
-        mVideoEncoderSource->start();
+    for (const auto &source : { mAudioEncoderSource, mVideoEncoderSource }) {
+        if (source == nullptr) {
+            continue;
+        }
+        source->setInputBufferTimeOffset((int64_t)timeOffset);
+        source->start();
     }
     mPauseStartTimeUs = 0;
 
diff --git a/media/libstagefright/MediaCodecSource.cpp b/media/libstagefright/MediaCodecSource.cpp
index 5039922..311c745 100644
--- a/media/libstagefright/MediaCodecSource.cpp
+++ b/media/libstagefright/MediaCodecSource.cpp
@@ -342,6 +342,17 @@
     postSynchronouslyAndReturnError(msg);
 }
 
+int64_t MediaCodecSource::getFirstSampleSystemTimeUs() {
+    sp<AMessage> msg = new AMessage(kWhatGetFirstSampleSystemTimeUs, mReflector);
+    sp<AMessage> response;
+    msg->postAndAwaitResponse(&response);
+    int64_t timeUs;
+    if (!response->findInt64("time-us", &timeUs)) {
+        timeUs = -1ll;
+    }
+    return timeUs;
+}
+
 status_t MediaCodecSource::start(MetaData* params) {
     sp<AMessage> msg = new AMessage(kWhatStart, mReflector);
     msg->setObject("meta", params);
@@ -408,6 +419,8 @@
       mEncoderDataSpace(0),
       mGraphicBufferConsumer(consumer),
       mInputBufferTimeOffsetUs(0),
+      mFirstSampleSystemTimeUs(-1ll),
+      mPausePending(false),
       mFirstSampleTimeUs(-1ll),
       mGeneration(0) {
     CHECK(mLooper != NULL);
@@ -646,17 +659,19 @@
 
         if (mbuf != NULL) {
             CHECK(mbuf->meta_data()->findInt64(kKeyTime, &timeUs));
-            timeUs += mInputBufferTimeOffsetUs;
-
-            // Due to the extra delay adjustment at the beginning of start/resume,
-            // the adjusted timeUs may be negative if MediaCodecSource goes into pause
-            // state before feeding any buffers to the encoder. Drop the buffer in this
-            // case.
-            if (timeUs < 0) {
-                mbuf->release();
-                return OK;
+            if (mFirstSampleSystemTimeUs < 0ll) {
+                mFirstSampleSystemTimeUs = systemTime() / 1000;
+                if (mPausePending) {
+                    mPausePending = false;
+                    onPause();
+                    mbuf->release();
+                    mAvailEncoderInputIndices.push_back(bufferIndex);
+                    return OK;
+                }
             }
 
+            timeUs += mInputBufferTimeOffsetUs;
+
             // push decoding time for video, or drift time for audio
             if (mIsVideo) {
                 mDecodingTimeQueue.push_back(timeUs);
@@ -665,7 +680,6 @@
                 if (mFirstSampleTimeUs < 0ll) {
                     mFirstSampleTimeUs = timeUs;
                 }
-
                 int64_t driftTimeUs = 0;
                 if (mbuf->meta_data()->findInt64(kKeyDriftTime, &driftTimeUs)
                         && driftTimeUs) {
@@ -717,6 +731,10 @@
 
     if (mStarted) {
         ALOGI("MediaCodecSource (%s) resuming", mIsVideo ? "video" : "audio");
+        if (mPausePending) {
+            mPausePending = false;
+            return OK;
+        }
         if (mIsVideo) {
             mEncoder->requestIDRFrame();
         }
@@ -763,6 +781,15 @@
     return OK;
 }
 
+void MediaCodecSource::onPause() {
+    if (mFlags & FLAG_USE_SURFACE_INPUT) {
+        suspend();
+    } else {
+        CHECK(mPuller != NULL);
+        mPuller->pause();
+    }
+}
+
 void MediaCodecSource::onMessageReceived(const sp<AMessage> &msg) {
     switch (msg->what()) {
     case kWhatPullerNotify:
@@ -832,25 +859,27 @@
             }
 
             MediaBuffer *mbuf = new MediaBuffer(outbuf->size());
-            memcpy(mbuf->data(), outbuf->data(), outbuf->size());
+            mbuf->add_ref();
 
             if (!(flags & MediaCodec::BUFFER_FLAG_CODECCONFIG)) {
                 if (mIsVideo) {
                     int64_t decodingTimeUs;
                     if (mFlags & FLAG_USE_SURFACE_INPUT) {
+                        if (mFirstSampleSystemTimeUs < 0ll) {
+                            mFirstSampleSystemTimeUs = systemTime() / 1000;
+                            if (mPausePending) {
+                                mPausePending = false;
+                                onPause();
+                                mbuf->release();
+                                break;
+                            }
+                        }
                         // Time offset is not applied at
                         // feedEncoderInputBuffer() in surface input case.
                         timeUs += mInputBufferTimeOffsetUs;
-
-                        // Due to the extra delay adjustment at the beginning of
-                        // start/resume, the adjusted timeUs may be negative if
-                        // MediaCodecSource goes into pause state before feeding
-                        // any buffers to the encoder. Drop the buffer in this case.
-                        if (timeUs < 0) {
-                            mEncoder->releaseOutputBuffer(index);
-                            break;
-                        }
-
+                        // GraphicBufferSource is supposed to discard samples
+                        // queued before start, and offset timeUs by start time
+                        CHECK_GE(timeUs, 0ll);
                         // TODO:
                         // Decoding time for surface source is unavailable,
                         // use presentation time for now. May need to move
@@ -883,8 +912,8 @@
             if (flags & MediaCodec::BUFFER_FLAG_SYNCFRAME) {
                 mbuf->meta_data()->setInt32(kKeyIsSyncFrame, true);
             }
+            memcpy(mbuf->data(), outbuf->data(), outbuf->size());
             mbuf->setObserver(this);
-            mbuf->add_ref();
 
             {
                 Mutexed<Output>::Locked output(mOutput);
@@ -975,11 +1004,10 @@
 
     case kWhatPause:
     {
-        if (mFlags & FLAG_USE_SURFACE_INPUT) {
-            suspend();
+        if (mFirstSampleSystemTimeUs < 0) {
+            mPausePending = true;
         } else {
-            CHECK(mPuller != NULL);
-            mPuller->pause();
+            onPause();
         }
         break;
     }
@@ -994,6 +1022,16 @@
         response->postReply(replyID);
         break;
     }
+    case kWhatGetFirstSampleSystemTimeUs:
+    {
+        sp<AReplyToken> replyID;
+        CHECK(msg->senderAwaitsResponse(&replyID));
+
+        sp<AMessage> response = new AMessage;
+        response->setInt64("time-us", mFirstSampleSystemTimeUs);
+        response->postReply(replyID);
+        break;
+    }
     default:
         TRESPASS();
     }