diff --git a/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.h b/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.h
--- a/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.h
+++ b/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.h
@@ -10,21 +10,45 @@
 #include "FFmpegLibWrapper.h"
 #include "FFmpegDataDecoder.h"
 #include "SimpleMap.h"
+#ifdef MOZ_WAYLAND_USE_VAAPI
+#  include "mozilla/widget/WaylandDMABufSurface.h"
+#  include <list>
+#endif
 
 namespace mozilla {
 
 #ifdef MOZ_WAYLAND_USE_VAAPI
-class VAAPIFrameHolder {
+// When VA-API decoding is running, ffmpeg allocates AVHWFramesContext - a pool
+// of "hardware" frames. Every "hardware" frame (VASurface) is backed
+// by actual piece of GPU memory which holds the decoded image data.
+//
+// The VASurface is wrapped by WaylandDMABufSurface and transferred to
+// rendering queue by WaylandDMABUFSurfaceImage, where TextureClient is
+// created and VASurface is used as a texture there.
+//
+// As there's a limited number of VASurfaces, ffmpeg reuses them to decode
+// next frames ASAP even if they are still attached to WaylandDMABufSurface
+// and used as a texture in our rendering engine.
+//
+// Unfortunately there isn't any obvious way how to mark particular VASurface
+// as used. The best we can do is to hold a reference to particular AVBuffer
+// from decoded AVFrame and AVHWFramesContext which owns the AVBuffer.
+
+class VAAPIFrameHolder final {
  public:
-  VAAPIFrameHolder(FFmpegLibWrapper* aLib, AVBufferRef* aVAAPIDeviceContext,
-                   AVBufferRef* aAVHWFramesContext, AVBufferRef* aHWFrame);
+  VAAPIFrameHolder(FFmpegLibWrapper* aLib, WaylandDMABufSurface* aSurface,
+                   AVCodecContext* aAVCodecContext, AVFrame* aAVFrame);
   ~VAAPIFrameHolder();
 
+  // Check if WaylandDMABufSurface is used by any gecko rendering process
+  // (WebRender or GL compositor) or by WaylandDMABUFSurfaceImage/VideoData.
+  bool IsUsed() const { return mSurface->IsGlobalRefSet(); }
+
  private:
-  FFmpegLibWrapper* mLib;
-  AVBufferRef* mVAAPIDeviceContext;
+  const FFmpegLibWrapper* mLib;
+  const RefPtr<WaylandDMABufSurface> mSurface;
   AVBufferRef* mAVHWFramesContext;
-  AVBufferRef* mHWFrame;
+  AVBufferRef* mHWAVBuffer;
 };
 #endif
 
@@ -97,6 +121,8 @@
 
   MediaResult CreateImageVAAPI(int64_t aOffset, int64_t aPts, int64_t aDuration,
                                MediaDataDecoder::DecodedData& aResults);
+  void ReleaseUnusedVAAPIFrames();
+  void ReleaseAllVAAPIFrames();
 #endif
 
   /**
@@ -112,6 +138,7 @@
   AVBufferRef* mVAAPIDeviceContext;
   const bool mDisableHardwareDecoding;
   VADisplay mDisplay;
+  std::list<UniquePtr<VAAPIFrameHolder>> mFrameHolders;
 #endif
   RefPtr<KnowsCompositor> mImageAllocator;
   RefPtr<ImageContainer> mImageContainer;
diff --git a/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.cpp b/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.cpp
--- a/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.cpp
+++ b/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.cpp
@@ -123,18 +123,27 @@
 }
 
 VAAPIFrameHolder::VAAPIFrameHolder(FFmpegLibWrapper* aLib,
-                                   AVBufferRef* aVAAPIDeviceContext,
-                                   AVBufferRef* aAVHWFramesContext,
-                                   AVBufferRef* aHWFrame)
+                                   WaylandDMABufSurface* aSurface,
+                                   AVCodecContext* aAVCodecContext,
+                                   AVFrame* aAVFrame)
     : mLib(aLib),
-      mVAAPIDeviceContext(mLib->av_buffer_ref(aVAAPIDeviceContext)),
-      mAVHWFramesContext(mLib->av_buffer_ref(aAVHWFramesContext)),
-      mHWFrame(mLib->av_buffer_ref(aHWFrame)){};
+      mSurface(aSurface),
+      mAVHWFramesContext(mLib->av_buffer_ref(aAVCodecContext->hw_frames_ctx)),
+      mHWAVBuffer(mLib->av_buffer_ref(aAVFrame->buf[0])) {
+  FFMPEG_LOG("VAAPIFrameHolder is adding dmabuf surface UID = %d\n",
+             mSurface->GetUID());
+
+  // Create global refcount object to track mSurface usage over
+  // gects rendering engine. We can't release it until it's used
+  // by GL compositor / WebRender.
+  mSurface->GlobalRefCountCreate();
+}
 
 VAAPIFrameHolder::~VAAPIFrameHolder() {
-  mLib->av_buffer_unref(&mHWFrame);
+  FFMPEG_LOG("VAAPIFrameHolder is releasing dmabuf surface UID = %d\n",
+             mSurface->GetUID());
+  mLib->av_buffer_unref(&mHWAVBuffer);
   mLib->av_buffer_unref(&mAVHWFramesContext);
-  mLib->av_buffer_unref(&mVAAPIDeviceContext);
 }
 
 AVCodec* FFmpegVideoDecoder<LIBAV_VER>::FindVAAPICodec() {
@@ -422,6 +431,13 @@
       NS_WARNING("FFmpeg h264 decoder failed to allocate frame.");
       return MediaResult(NS_ERROR_OUT_OF_MEMORY, __func__);
     }
+
+#  ifdef MOZ_WAYLAND_USE_VAAPI
+    if (mVAAPIDeviceContext) {
+      ReleaseUnusedVAAPIFrames();
+    }
+#  endif
+
     res = mLib->avcodec_receive_frame(mCodecContext, mFrame);
     if (res == int(AVERROR_EOF)) {
       return NS_ERROR_DOM_MEDIA_END_OF_STREAM;
@@ -628,9 +644,20 @@
 }
 
 #ifdef MOZ_WAYLAND_USE_VAAPI
-static void VAAPIFrameReleaseCallback(VAAPIFrameHolder* aVAAPIFrameHolder) {
-  auto frameHolder = static_cast<VAAPIFrameHolder*>(aVAAPIFrameHolder);
-  delete frameHolder;
+void FFmpegVideoDecoder<LIBAV_VER>::ReleaseUnusedVAAPIFrames() {
+  std::list<UniquePtr<VAAPIFrameHolder>>::iterator holder =
+      mFrameHolders.begin();
+  while (holder != mFrameHolders.end()) {
+    if (!(*holder)->IsUsed()) {
+      holder = mFrameHolders.erase(holder);
+    } else {
+      holder++;
+    }
+  }
+}
+
+void FFmpegVideoDecoder<LIBAV_VER>::ReleaseAllVAAPIFrames() {
+  mFrameHolders.clear();
 }
 
 MediaResult FFmpegVideoDecoder<LIBAV_VER>::CreateImageVAAPI(
@@ -667,20 +694,20 @@
         RESULT_DETAIL("Unable to allocate WaylandDMABufSurfaceNV12."));
   }
 
+#  ifdef MOZ_LOGGING
+  static int uid = 0;
+  surface->SetUID(++uid);
+  FFMPEG_LOG("Created dmabuf UID = %d HW surface %x\n", uid, surface_id);
+#  endif
+
   surface->SetYUVColorSpace(GetFrameColorSpace());
 
-  // mFrame->buf[0] is a reference to H264 VASurface for this mFrame.
-  // We need create WaylandDMABUFSurfaceImage on top of it,
-  // create EGLImage/Texture on top of it and render it by GL.
+  // Store reference to the decoded HW buffer, see VAAPIFrameHolder struct.
+  auto holder =
+      MakeUnique<VAAPIFrameHolder>(mLib, surface, mCodecContext, mFrame);
+  mFrameHolders.push_back(std::move(holder));
 
-  // FFmpeg tends to reuse the particual VASurface for another frame
-  // even when the mFrame is not released. To keep VASurface as is
-  // we explicitly reference it and keep until WaylandDMABUFSurfaceImage
-  // is live.
-  RefPtr<layers::Image> im = new layers::WaylandDMABUFSurfaceImage(
-      surface, VAAPIFrameReleaseCallback,
-      new VAAPIFrameHolder(mLib, mVAAPIDeviceContext,
-                           mCodecContext->hw_frames_ctx, mFrame->buf[0]));
+  RefPtr<layers::Image> im = new layers::WaylandDMABUFSurfaceImage(surface);
 
   RefPtr<VideoData> vp = VideoData::CreateFromImage(
       mInfo.mDisplay, aOffset, TimeUnit::FromMicroseconds(aPts),
@@ -732,6 +759,7 @@
 void FFmpegVideoDecoder<LIBAV_VER>::ProcessShutdown() {
 #ifdef MOZ_WAYLAND_USE_VAAPI
   if (mVAAPIDeviceContext) {
+    ReleaseAllVAAPIFrames();
     mLib->av_buffer_unref(&mVAAPIDeviceContext);
   }
 #endif