Integrate VDA with WebRTC.

content/renderer/media/rtc_video_decoder.cc

+// Copyright (c) 2013 The Chromium Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+#include "content/renderer/media/rtc_video_decoder.h"
+
+#include "base/bind.h"
+#include "base/logging.h"
+#include "base/memory/ref_counted.h"
+#include "base/message_loop/message_loop_proxy.h"
+#include "base/safe_numerics.h"
+#include "base/stl_util.h"
+#include "base/task_runner_util.h"
+#include "content/child/child_thread.h"
+#include "content/renderer/media/native_handle_impl.h"
+#include "media/base/bind_to_loop.h"
+#include "third_party/webrtc/common_video/interface/texture_video_frame.h"
+#include "third_party/webrtc/system_wrappers/interface/ref_count.h"
+
+namespace content {
+
+const int32 RTCVideoDecoder::ID_LAST = 0x3FFFFFFF;

[Ami GONE FROM CHROMIUM, 2013/07/02 23:36:57]

FWIW these could as well be enums declared here.
I don't care which you use.

+const int32 RTCVideoDecoder::ID_HALF = 0x20000000;
+const int32 RTCVideoDecoder::ID_INVALID = -1;
+
+// Maximum number of concurrent VDA::Decode() operations RVD will maintain.
+// Higher values allow better pipelining in the GPU, but also require more
+// resources.
+static const size_t kMaxInFlightDecodes = 8;
+
+// Size of shared-memory segments we allocate.  Since we reuse them we let them
+// be on the beefy side.
+static const size_t kSharedMemorySegmentBytes = 100 << 10;
+
+// Maximum number of allocated shared-memory segments.
+static const int kMaxNumSharedMemorySegments = 16;
+
+// Maximum number of pending WebRTC buffers that are waiting for the shared
+// memory. 10 seconds for 30 fps.
+static const size_t kMaxNumOfPendingBuffers = 300;
+
+// A shared memory segment and its allocated size. This class has the ownership
+// of |shm|.
+class RTCVideoDecoder::SHMBuffer {
+ public:
+  SHMBuffer(base::SharedMemory* shm, size_t size);
+  ~SHMBuffer();
+  base::SharedMemory* const shm;
+  const size_t size;
+};
+
+RTCVideoDecoder::SHMBuffer::SHMBuffer(base::SharedMemory* shm, size_t size)
+    : shm(shm), size(size) {}
+
+RTCVideoDecoder::SHMBuffer::~SHMBuffer() { shm->Close(); }
+
+// Metadata of a bitstream buffer.
+struct RTCVideoDecoder::BufferData {
+  BufferData(int32 bitstream_buffer_id,
+             uint32_t timestamp,
+             int width,
+             int height,
+             size_t size);
+  ~BufferData();
+  int32 bitstream_buffer_id;
+  uint32_t timestamp;  // in 90KHz
+  uint32_t width;
+  uint32_t height;
+  size_t size;  // buffer size
+};
+
+RTCVideoDecoder::BufferData::BufferData(int32 bitstream_buffer_id,
+                                        uint32_t timestamp,
+                                        int width,
+                                        int height,
+                                        size_t size)
+    : bitstream_buffer_id(bitstream_buffer_id),
+      timestamp(timestamp),
+      width(width),
+      height(height),
+      size(size) {}
+
+RTCVideoDecoder::BufferData::~BufferData() {}
+
+RTCVideoDecoder::RTCVideoDecoder(
+    const scoped_refptr<media::GpuVideoDecoder::Factories>& factories)
+    : weak_factory_(this),
+      factories_(factories),
+      vda_loop_proxy_(factories_->GetMessageLoop()),
+      create_shm_thread_("CreateSHMThread"),
+      decoder_texture_target_(0),
+      next_picture_buffer_id_(0),
+      state_(UNINITIALIZED),
+      decode_complete_callback_(NULL),
+      num_shm_buffers_(0),
+      next_bitstream_buffer_id_(0),
+      reset_bitstream_buffer_id_(0) {
+  // Initialize directly if |vda_loop_proxy_| is the renderer thread.
+  base::WaitableEvent compositor_loop_async_waiter(false, false);
+  if (vda_loop_proxy_->BelongsToCurrentThread()) {
+    Initialize(&compositor_loop_async_waiter);
+    return;
+  }
+  // Post the task if |vda_loop_proxy_| is the compositor thread. Waiting here
+  // is safe because the compositor thread will not be stopped until the
+  // renderer thread shuts down.
+  vda_loop_proxy_->PostTask(FROM_HERE,
+                            base::Bind(&RTCVideoDecoder::Initialize,
+                                       base::Unretained(this),
+                                       &compositor_loop_async_waiter));
+  compositor_loop_async_waiter.Wait();
+}
+
+RTCVideoDecoder::~RTCVideoDecoder() {
+  DVLOG(2) << "~RTCVideoDecoder";
+  factories_->Abort();
+  create_shm_thread_.Stop();
+  // Delete vda and remove |this| from the observer if vda thread is alive.
+  if (vda_loop_proxy_->BelongsToCurrentThread()) {
+    base::MessageLoop::current()->RemoveDestructionObserver(this);
+    DestroyVDA();
+  } else {
+    // VDA should have been destroyed in WillDestroyCurrentMessageLoop.
+    DCHECK(!vda_);
+  }
+
+  // Delete all shared memories.
+  STLDeleteElements(&available_shm_segments_);
+  STLDeleteValues(&bitstream_buffers_in_decoder_);
+  STLDeleteContainerPairFirstPointers(decode_buffers_.begin(),
+                                      decode_buffers_.end());
+  decode_buffers_.clear();
+
+  // Delete WebRTC input buffers.
+  for (std::deque<std::pair<webrtc::EncodedImage, BufferData> >::iterator it =
+           pending_buffers_.begin();
+       it != pending_buffers_.end();
+       ++it) {
+    delete it->first._buffer;
+  }
+}
+
+scoped_ptr<RTCVideoDecoder> RTCVideoDecoder::Create(
+    const scoped_refptr<media::GpuVideoDecoder::Factories>& factories) {
+  scoped_ptr<RTCVideoDecoder> decoder(new RTCVideoDecoder(factories));
+  decoder->vda_.reset(factories->CreateVideoDecodeAccelerator(
+      media::VP8PROFILE_MAIN, decoder.get()));
+  // vda can be NULL if VP8 is not supported.
+  if (decoder->vda_ != NULL) {
+    decoder->state_ = INITIALIZED;
+  } else {
+    factories->GetMessageLoop()->DeleteSoon(FROM_HERE, decoder.release());
+  }
+  return decoder.Pass();
+}
+
+int32_t RTCVideoDecoder::InitDecode(const webrtc::VideoCodec* codecSettings,
+                                    int32_t /*numberOfCores*/) {
+  DVLOG(2) << "InitDecode";
+  DCHECK_EQ(codecSettings->codecType, webrtc::kVideoCodecVP8);
+  if (codecSettings->codecSpecific.VP8.feedbackModeOn) {
+    LOG(ERROR) << "Feedback mode not supported";
+    return WEBRTC_VIDEO_CODEC_ERROR;
+  }
+
+  base::AutoLock auto_lock(lock_);
+  if (state_ == UNINITIALIZED || state_ == DECODE_ERROR) {
+    LOG(ERROR) << "VDA is not initialized. state=" << state_;
+    return WEBRTC_VIDEO_CODEC_UNINITIALIZED;
+  }
+  // Create some shared memory if the queue is empty.
+  if (available_shm_segments_.size() == 0) {
+    create_shm_thread_.message_loop_proxy()
+        ->PostTask(FROM_HERE,
+                   base::Bind(&RTCVideoDecoder::CreateSHM,
+                              weak_this_,
+                              kMaxInFlightDecodes,
+                              kSharedMemorySegmentBytes));
+  }
+  return WEBRTC_VIDEO_CODEC_OK;
+}
+
+int32_t RTCVideoDecoder::Decode(
+    const webrtc::EncodedImage& inputImage,
+    bool missingFrames,
+    const webrtc::RTPFragmentationHeader* /*fragmentation*/,
+    const webrtc::CodecSpecificInfo* /*codecSpecificInfo*/,
+    int64_t /*renderTimeMs*/) {
+  DVLOG(3) << "Decode";
+
+  int bitstream_buffer_id;
+  bool no_pending_buffers;
+  {
+    base::AutoLock auto_lock(lock_);
+    if (state_ == UNINITIALIZED || decode_complete_callback_ == NULL) {
+      LOG(ERROR) << "The decoder has not initialized.";
+      return WEBRTC_VIDEO_CODEC_UNINITIALIZED;
+    }
+    if (state_ == DECODE_ERROR) {
+      LOG(ERROR) << "Decoding error occurred.";
+      return WEBRTC_VIDEO_CODEC_ERROR;
+    }
+    bitstream_buffer_id = next_bitstream_buffer_id_;
+    // Mask against 30 bits, to avoid (undefined) wraparound on signed integer.
+    next_bitstream_buffer_id_ = (next_bitstream_buffer_id_ + 1) & ID_LAST;
+    no_pending_buffers = (pending_buffers_.size() == 0);
+  }
+  if (missingFrames || !inputImage._completeFrame) {
+    DLOG(ERROR) << "Missing or incomplete frames.";
+    // Unlike the SW decoder in libvpx, hw decoder cannot handle broken frames.
+    // Return an error to request a key frame.
+    return WEBRTC_VIDEO_CODEC_ERROR;
+  }
+  if (inputImage._frameType == webrtc::kKeyFrame)
+    frame_size_.SetSize(inputImage._encodedWidth, inputImage._encodedHeight);
+
+  // Create buffer metadata.
+  BufferData buffer_data(bitstream_buffer_id,
+                         inputImage._timeStamp,
+                         frame_size_.width(),
+                         frame_size_.height(),
+                         inputImage._length);
+
+  // If the shared memory is available and there are no pending buffers, send
+  // the buffer for decode. If not, save the buffer in the queue for decode
+  // later.
+  scoped_ptr<SHMBuffer> shm_buffer;
+  if (no_pending_buffers)

[Ami GONE FROM CHROMIUM, 2013/07/02 23:36:57]

setting this under lock but then inspecting it without holding the lock
automatically makes me wonder whether the value would be different if you
reacquired the lock here before inspecting pending_buffers_.

[wuchengli, 2013/07/03 09:09:38]

It should be OK because the only place that increases 
|pending_buffers_| is in the bottom of this function. auto_lock(lock_) has been
moved to the top of this function so this is not a problem anymore.

+    shm_buffer = GetSHM(inputImage._length);
+
+  if (!shm_buffer) {
+    if (SaveToPendingBuffers(inputImage, buffer_data))
+      return WEBRTC_VIDEO_CODEC_OK;
+    else
+      return WEBRTC_VIDEO_CODEC_ERROR;
+  }
+
+  SaveToDecodeBuffers(inputImage, shm_buffer.Pass(), buffer_data);
+  vda_loop_proxy_->PostTask(
+      FROM_HERE, base::Bind(&RTCVideoDecoder::RequestBufferDecode, weak_this_));
+  return WEBRTC_VIDEO_CODEC_OK;
+}
+
+int32_t RTCVideoDecoder::RegisterDecodeCompleteCallback(
+    webrtc::DecodedImageCallback* callback) {
+  DVLOG(2) << "RegisterDecodeCompleteCallback";
+  base::AutoLock auto_lock(lock_);
+  decode_complete_callback_ = callback;
+  return WEBRTC_VIDEO_CODEC_OK;
+}
+
+int32_t RTCVideoDecoder::Release() {
+  DVLOG(2) << "Release";
+  // Do not destroy VDA because the decoder will be recycled by
+  // RTCVideoDecoderFactory. Just reset VDA.
+  return Reset();
+}
+
+int32_t RTCVideoDecoder::Reset() {
+  DVLOG(2) << "Reset";
+  base::AutoLock auto_lock(lock_);
+  if (state_ == UNINITIALIZED) {
+    LOG(ERROR) << "Decoder not initialized.";
+    return WEBRTC_VIDEO_CODEC_UNINITIALIZED;
+  }
+  reset_bitstream_buffer_id_ = next_bitstream_buffer_id_;
+  // If VDA is already resetting, no need to request the reset again.
+  if (state_ != RESETTING) {
+    state_ = RESETTING;
+    vda_loop_proxy_->PostTask(
+        FROM_HERE, base::Bind(&RTCVideoDecoder::ResetInternal, weak_this_));
+  }
+  return WEBRTC_VIDEO_CODEC_OK;
+}
+
+void RTCVideoDecoder::NotifyInitializeDone() {
+  DVLOG(2) << "NotifyInitializeDone";
+  NOTREACHED();
+}
+
+void RTCVideoDecoder::ProvidePictureBuffers(uint32 count,
+                                            const gfx::Size& size,
+                                            uint32 texture_target) {
+  DCHECK(vda_loop_proxy_->BelongsToCurrentThread());
+  DVLOG(3) << "ProvidePictureBuffers. texture_target=" << texture_target;
+
+  if (!vda_)
+    return;
+
+  std::vector<uint32> texture_ids;
+  std::vector<gpu::Mailbox> texture_mailboxes;
+  decoder_texture_target_ = texture_target;
+  // Discards the sync point returned here since PictureReady will imply that
+  // the produce has already happened, and the texture is ready for use.
+  if (!factories_->CreateTextures(count,
+                                  size,
+                                  &texture_ids,
+                                  &texture_mailboxes,
+                                  decoder_texture_target_)) {
+    NotifyError(media::VideoDecodeAccelerator::PLATFORM_FAILURE);
+    return;
+  }
+  DCHECK_EQ(count, texture_ids.size());
+  DCHECK_EQ(count, texture_mailboxes.size());
+
+  std::vector<media::PictureBuffer> picture_buffers;
+  for (size_t i = 0; i < texture_ids.size(); ++i) {
+    picture_buffers.push_back(media::PictureBuffer(
+        next_picture_buffer_id_++, size, texture_ids[i], texture_mailboxes[i]));
+    bool inserted = assigned_picture_buffers_.insert(std::make_pair(
+        picture_buffers.back().id(), picture_buffers.back())).second;
+    DCHECK(inserted);
+  }
+  vda_->AssignPictureBuffers(picture_buffers);
+}
+
+void RTCVideoDecoder::DismissPictureBuffer(int32 id) {
+  DVLOG(3) << "DismissPictureBuffer. id=" << id;
+  DCHECK(vda_loop_proxy_->BelongsToCurrentThread());
+
+  std::map<int32, media::PictureBuffer>::iterator it =
+      assigned_picture_buffers_.find(id);
+  if (it == assigned_picture_buffers_.end()) {
+    NOTREACHED() << "Missing picture buffer: " << id;
+    return;
+  }
+
+  media::PictureBuffer buffer_to_dismiss = it->second;
+  assigned_picture_buffers_.erase(it);
+
+  std::set<int32>::iterator at_display_it =
+      picture_buffers_at_display_.find(id);
+
+  if (at_display_it == picture_buffers_at_display_.end()) {
+    // We can delete the texture immediately as it's not being displayed.
+    factories_->DeleteTexture(buffer_to_dismiss.texture_id());
+  } else {
+    // Texture in display. Postpone deletion until after it's returned to us.
+    bool inserted = dismissed_picture_buffers_
+        .insert(std::make_pair(id, buffer_to_dismiss)).second;
+    DCHECK(inserted);
+  }
+}
+
+void RTCVideoDecoder::PictureReady(const media::Picture& picture) {
+  DVLOG(3) << "PictureReady";
+  DCHECK(vda_loop_proxy_->BelongsToCurrentThread());
+
+  std::map<int32, media::PictureBuffer>::iterator it =
+      assigned_picture_buffers_.find(picture.picture_buffer_id());
+  if (it == assigned_picture_buffers_.end()) {
+    NOTREACHED() << "Missing picture buffer: " << picture.picture_buffer_id();
+    NotifyError(media::VideoDecodeAccelerator::PLATFORM_FAILURE);
+    return;
+  }
+  const media::PictureBuffer& pb = it->second;
+
+  // Create a media::VideoFrame.
+  uint32_t timestamp = 0, width = 0, height = 0;
+  size_t size = 0;
+  GetBufferData(
+      picture.bitstream_buffer_id(), &timestamp, &width, &height, &size);
+  scoped_refptr<media::VideoFrame> frame =
+      CreateVideoFrame(picture, pb, timestamp, width, height, size);
+  bool inserted =
+      picture_buffers_at_display_.insert(picture.picture_buffer_id()).second;
+  DCHECK(inserted);
+
+  // Create a WebRTC video frame.
+  webrtc::RefCountImpl<NativeHandleImpl>* handle =
+      new webrtc::RefCountImpl<NativeHandleImpl>();
+  handle->SetHandle(frame.get());
+  webrtc::TextureVideoFrame decoded_image(width, height, timestamp, 0, handle);
+
+  // Invoke decode callback. WebRTC expects no frame callback after Release.
+  {
+    base::AutoLock auto_lock(lock_);
+    DCHECK(decode_complete_callback_ != NULL);
+    if (IsBufferAfterReset(picture.bitstream_buffer_id(),
+                           reset_bitstream_buffer_id_)) {
+      decode_complete_callback_->Decoded(decoded_image);
+    }
+  }
+}
+
+scoped_refptr<media::VideoFrame> RTCVideoDecoder::CreateVideoFrame(
+    const media::Picture& picture,
+    const media::PictureBuffer& pb,
+    uint32_t timestamp,
+    uint32_t width,
+    uint32_t height,
+    size_t size) {
+  gfx::Rect visible_rect(width, height);
+  gfx::Size natural_size(width, height);
+  DCHECK(decoder_texture_target_);
+  // Convert timestamp from 90KHz to ms.
+  base::TimeDelta timestamp_ms = base::TimeDelta::FromInternalValue(
+      base::checked_numeric_cast<uint64_t>(timestamp) * 1000 / 90);
+  return media::VideoFrame::WrapNativeTexture(
+      new media::VideoFrame::MailboxHolder(
+          pb.texture_mailbox(),
+          0,  // sync_point
+          media::BindToCurrentLoop(
+              base::Bind(&RTCVideoDecoder::ReusePictureBuffer,
+                         weak_this_,
+                         picture.picture_buffer_id()))),
+      decoder_texture_target_,
+      pb.size(),
+      visible_rect,
+      natural_size,
+      timestamp_ms,
+      base::Bind(&media::GpuVideoDecoder::Factories::ReadPixels,
+                 factories_,
+                 pb.texture_id(),
+                 decoder_texture_target_,
+                 natural_size),
+      base::Closure());
+}
+
+void RTCVideoDecoder::NotifyEndOfBitstreamBuffer(int32 id) {
+  DVLOG(3) << "NotifyEndOfBitstreamBuffer. id=" << id;
+  DCHECK(vda_loop_proxy_->BelongsToCurrentThread());
+
+  std::map<int32, SHMBuffer*>::iterator it =
+      bitstream_buffers_in_decoder_.find(id);
+  if (it == bitstream_buffers_in_decoder_.end()) {
+    NotifyError(media::VideoDecodeAccelerator::PLATFORM_FAILURE);
+    NOTREACHED() << "Missing bitstream buffer: " << id;
+    return;
+  }
+
+  PutSHM(scoped_ptr<SHMBuffer>(it->second));
+  bitstream_buffers_in_decoder_.erase(it);
+
+  RequestBufferDecode();
+}
+
+void RTCVideoDecoder::NotifyFlushDone() {
+  DVLOG(3) << "NotifyFlushDone";
+  NOTREACHED() << "Unexpected flush done notification.";
+}
+
+void RTCVideoDecoder::NotifyResetDone() {
+  DCHECK(vda_loop_proxy_->BelongsToCurrentThread());
+  DVLOG(3) << "NotifyResetDone";
+
+  if (!vda_)
+    return;
+
+  input_buffer_data_.clear();
+  {
+    base::AutoLock auto_lock(lock_);
+    state_ = INITIALIZED;
+  }
+  // Send the pending buffers for decoding.
+  RequestBufferDecode();
+}
+
+void RTCVideoDecoder::NotifyError(media::VideoDecodeAccelerator::Error error) {
+  DCHECK(vda_loop_proxy_->BelongsToCurrentThread());
+  if (!vda_)
+    return;
+
+  DLOG(ERROR) << "VDA Error:" << error;
+  DestroyVDA();
+
+  base::AutoLock auto_lock(lock_);
+  state_ = DECODE_ERROR;
+}
+
+void RTCVideoDecoder::WillDestroyCurrentMessageLoop() {
+  DCHECK(vda_loop_proxy_->BelongsToCurrentThread());
+  factories_->Abort();
+  weak_factory_.InvalidateWeakPtrs();
+  DestroyVDA();
+}
+
+void RTCVideoDecoder::Initialize(base::WaitableEvent* waiter) {
+  DCHECK(vda_loop_proxy_->BelongsToCurrentThread());
+  base::MessageLoop::current()->AddDestructionObserver(this);
+  weak_this_ = weak_factory_.GetWeakPtr();
+  create_shm_thread_.Start();
+  waiter->Signal();
+}
+
+void RTCVideoDecoder::RequestBufferDecode() {
+  DCHECK(vda_loop_proxy_->BelongsToCurrentThread());
+  if (!vda_)
+    return;
+
+  MovePendingBuffersToDecodeBuffers();
+
+  while (CanMoreDecodeWorkBeDone()) {
+    // Get a buffer and data from the queue.
+    std::pair<SHMBuffer*, BufferData>* buffer_pair;
+    SHMBuffer* shm_buffer = NULL;
+    BufferData* buffer_data = NULL;
+    {
+      base::AutoLock auto_lock(lock_);
+      // Do not request decode if VDA is resetting.
+      if (decode_buffers_.size() == 0 || state_ == RESETTING)
+        return;
+      buffer_pair = &decode_buffers_.front();
+      decode_buffers_.pop_front();
+      shm_buffer = buffer_pair->first;
+      buffer_data = &buffer_pair->second;
+      // Drop the buffers before Reset or Release is called.
+      if (!IsBufferAfterReset(buffer_data->bitstream_buffer_id,
+                              reset_bitstream_buffer_id_)) {
+        available_shm_segments_.push_back(shm_buffer);
+        continue;
+      }
+    }
+
+    // Create a BitstreamBuffer and send to VDA to decode.
+    media::BitstreamBuffer bitstream_buffer(buffer_data->bitstream_buffer_id,
+                                            shm_buffer->shm->handle(),
+                                            buffer_data->size);
+    bool inserted = bitstream_buffers_in_decoder_
+        .insert(std::make_pair(bitstream_buffer.id(), shm_buffer)).second;
+    DCHECK(inserted);
+    RecordBufferData(*buffer_data);
+    vda_->Decode(bitstream_buffer);
+  }
+}
+
+bool RTCVideoDecoder::CanMoreDecodeWorkBeDone() {
+  return bitstream_buffers_in_decoder_.size() < kMaxInFlightDecodes;
+}
+
+bool RTCVideoDecoder::IsBufferAfterReset(int32 id_buffer, int32 id_reset) {
+  if (id_reset == ID_INVALID)
+    return true;
+  int32 diff = id_buffer - id_reset;
+  if (diff < 0)
+    diff += ID_LAST + 1;
+  return diff < ID_HALF;
+}
+
+void RTCVideoDecoder::SaveToDecodeBuffers(
+    const webrtc::EncodedImage& input_image,
+    scoped_ptr<SHMBuffer> shm_buffer,
+    const BufferData& buffer_data) {
+  memcpy(shm_buffer->shm->memory(), input_image._buffer, input_image._length);
+  std::pair<SHMBuffer*, BufferData> buffer_pair =
+      std::make_pair(shm_buffer.release(), buffer_data);
+
+  // Store the buffer and the metadata to the queue.
+  base::AutoLock auto_lock(lock_);
+  decode_buffers_.push_back(buffer_pair);
+}
+
+bool RTCVideoDecoder::SaveToPendingBuffers(
+    const webrtc::EncodedImage& input_image,
+    const BufferData& buffer_data) {
+  DVLOG(2) << "SaveToPendingBuffers";
+  // Queued too many buffers. Something goes wrong.
+  if (pending_buffers_.size() >= kMaxNumOfPendingBuffers) {
+    LOG(WARNING) << "Too many pending buffers!";
+    return false;
+  }
+
+  // Clone the input image and save it to the queue.
+  uint8_t* buffer = static_cast<uint8_t*>(malloc(input_image._length));
+  // TODO(wuchengli): avoid memcpy. Extend webrtc::VideoDecoder::Decode()
+  // interface to take a non-const ptr to the frame and add a method to the
+  // frame that will swap buffers with another.
+  memcpy(buffer, input_image._buffer, input_image._length);
+  webrtc::EncodedImage encoded_image(
+      buffer, input_image._length, input_image._length);
+  std::pair<webrtc::EncodedImage, BufferData> buffer_pair =
+      std::make_pair(encoded_image, buffer_data);
+
+  base::AutoLock auto_lock(lock_);
+  pending_buffers_.push_back(buffer_pair);
+  return true;
+}
+
+void RTCVideoDecoder::MovePendingBuffersToDecodeBuffers() {
+  while (true) {
+    webrtc::EncodedImage* input_image;
+    BufferData* buffer_data;
+    {
+      base::AutoLock auto_lock(lock_);
+      if (pending_buffers_.size() == 0)
+        return;
+      DVLOG(2) << "MovePendingBuffersToDecodeBuffers"
+               << ". pending_buffers size=" << pending_buffers_.size()
+               << ". decode_buffers_ size=" << decode_buffers_.size()
+               << ". available_shm size=" << available_shm_segments_.size();
+      // Get a pending buffer from the queue.
+      input_image = &pending_buffers_.front().first;
+      buffer_data = &pending_buffers_.front().second;
+
+      // Drop the frame if it comes before Reset or Release.
+      if (!IsBufferAfterReset(buffer_data->bitstream_buffer_id,
+                              reset_bitstream_buffer_id_)) {
+        delete input_image->_buffer;
+        pending_buffers_.pop_front();
+        continue;
+      }
+    }
+
+    // Get shared memory and save it to decode buffers.
+    scoped_ptr<SHMBuffer> shm_buffer = GetSHM(input_image->_length);
+    if (!shm_buffer)
+      return;
+    SaveToDecodeBuffers(*input_image, shm_buffer.Pass(), *buffer_data);

[Ami GONE FROM CHROMIUM, 2013/07/02 23:36:57]

The locking in this function is wonky.  Would things be a lot clearer if
SaveToDecodeBuffers was renamed SaveToDecodeBuffers_Locked, and both
MovePendingBuffersToDecodeBuffers() and RTCVideoDecoder::Decode() just had an
AutoLock at the top of their bodies?

(anytime you acquire & release a lock multiple times in a function, questions
arise as to what invariants you're really ok with changing from under you, and
what invariants' changes will be hard to find & fix bugs)

[wuchengli, 2013/07/03 09:09:38]

Done. My concern was SaveToPendingBuffers and SaveToDecodeBuffers had memcpy. I
wasn't sure holding a lock during 1-8KB memcpy was too long. From the
experiment, the decode latency didn't seem to change.
On 2013/07/02 23:36:57, Ami Fischman wrote:

+    delete input_image->_buffer;
+    base::AutoLock auto_lock(lock_);
+    pending_buffers_.pop_front();
+  }
+}
+
+void RTCVideoDecoder::ResetInternal() {
+  DCHECK(vda_loop_proxy_->BelongsToCurrentThread());
+  DVLOG(2) << "ResetInternal";
+  if (vda_)
+    vda_->Reset();
+}
+
+void RTCVideoDecoder::ReusePictureBuffer(int64 picture_buffer_id,
+                                         uint32 sync_point) {
+  DCHECK(vda_loop_proxy_->BelongsToCurrentThread());
+  DVLOG(3) << "ReusePictureBuffer. id=" << picture_buffer_id;
+
+  if (!vda_)
+    return;
+
+  CHECK(!picture_buffers_at_display_.empty());
+
+  size_t num_erased = picture_buffers_at_display_.erase(picture_buffer_id);
+  DCHECK(num_erased);
+
+  std::map<int32, media::PictureBuffer>::iterator it =
+      assigned_picture_buffers_.find(picture_buffer_id);
+
+  if (it == assigned_picture_buffers_.end()) {
+    // This picture was dismissed while in display, so we postponed deletion.
+    it = dismissed_picture_buffers_.find(picture_buffer_id);
+    DCHECK(it != dismissed_picture_buffers_.end());
+    factories_->DeleteTexture(it->second.texture_id());
+    dismissed_picture_buffers_.erase(it);
+    return;
+  }
+
+  factories_->WaitSyncPoint(sync_point);
+
+  vda_->ReusePictureBuffer(picture_buffer_id);
+}
+
+void RTCVideoDecoder::DestroyTextures() {
+  DCHECK(vda_loop_proxy_->BelongsToCurrentThread());
+  std::map<int32, media::PictureBuffer>::iterator it;
+
+  for (it = assigned_picture_buffers_.begin();
+       it != assigned_picture_buffers_.end();
+       ++it) {
+    factories_->DeleteTexture(it->second.texture_id());
+  }
+  assigned_picture_buffers_.clear();
+
+  for (it = dismissed_picture_buffers_.begin();
+       it != dismissed_picture_buffers_.end();
+       ++it) {
+    factories_->DeleteTexture(it->second.texture_id());
+  }
+  dismissed_picture_buffers_.clear();
+}
+
+void RTCVideoDecoder::DestroyVDA() {
+  DVLOG(2) << "DestroyVDA";
+  DCHECK(vda_loop_proxy_->BelongsToCurrentThread());
+  if (vda_)
+    vda_.release()->Destroy();
+  DestroyTextures();
+}
+
+scoped_ptr<RTCVideoDecoder::SHMBuffer> RTCVideoDecoder::GetSHM(
+    size_t min_size) {
+  // Reuse a SHM if possible.
+  SHMBuffer* ret = NULL;
+  base::AutoLock auto_lock(lock_);
+  if (!available_shm_segments_.empty() &&
+      available_shm_segments_.back()->size >= min_size) {
+    ret = available_shm_segments_.back();
+    available_shm_segments_.pop_back();
+  }
+  // Post to the child thread to create shared memory if SHM cannot be reused
+  // or the queue is almost empty.
+  if (num_shm_buffers_ < kMaxNumSharedMemorySegments &&
+      (ret == NULL || available_shm_segments_.size() <= 1)) {
+    create_shm_thread_.message_loop_proxy()->PostTask(
+        FROM_HERE,
+        base::Bind(&RTCVideoDecoder::CreateSHM, weak_this_, 1, min_size));
+  }
+  return scoped_ptr<SHMBuffer>(ret);
+}
+
+void RTCVideoDecoder::CreateSHM(int number, size_t min_size) {
+  DCHECK(create_shm_thread_.message_loop_proxy()->BelongsToCurrentThread());
+  DVLOG(2) << "CreateSharedMemory. size=" << min_size;
+  int number_to_allocate;
+  {
+    base::AutoLock auto_lock(lock_);
+    number_to_allocate =
+        std::min(kMaxNumSharedMemorySegments - num_shm_buffers_, number);
+  }
+  for (int i = 0; i < number_to_allocate; i++) {
+    size_t size_to_allocate = std::max(min_size, kSharedMemorySegmentBytes);

[Ami GONE FROM CHROMIUM, 2013/07/02 23:36:57]

can go outside loop

[wuchengli, 2013/07/03 09:09:38]

Done.

+    base::SharedMemory* shm = factories_->CreateSharedMemory(size_to_allocate);
+    if (shm != NULL) {
+      {
+        base::AutoLock auto_lock(lock_);
+        num_shm_buffers_++;
+      }
+      PutSHM(scoped_ptr<SHMBuffer>(new SHMBuffer(shm, size_to_allocate)));

[Ami GONE FROM CHROMIUM, 2013/07/02 23:36:57]

Ditto would PutSHM_Locked() make more sense?

[wuchengli, 2013/07/03 09:09:38]

Done.

+      // Kick off the decoding.
+      vda_loop_proxy_->PostTask(
+          FROM_HERE,
+          base::Bind(&RTCVideoDecoder::RequestBufferDecode, weak_this_));
+    }
+  }
+}
+
+void RTCVideoDecoder::PutSHM(scoped_ptr<SHMBuffer> shm_buffer) {
+  base::AutoLock auto_lock(lock_);
+  available_shm_segments_.push_back(shm_buffer.release());
+}
+
+void RTCVideoDecoder::RecordBufferData(const BufferData& buffer_data) {
+  input_buffer_data_.push_front(buffer_data);
+  // Why this value?  Because why not.  avformat.h:MAX_REORDER_DELAY is 16, but
+  // that's too small for some pathological B-frame test videos.  The cost of
+  // using too-high a value is low (192 bits per extra slot).
+  static const size_t kMaxInputBufferDataSize = 128;
+  // Pop from the back of the list, because that's the oldest and least likely
+  // to be useful in the future data.
+  if (input_buffer_data_.size() > kMaxInputBufferDataSize)
+    input_buffer_data_.pop_back();
+}
+
+void RTCVideoDecoder::GetBufferData(int32 bitstream_buffer_id,
+                                    uint32_t* timestamp,
+                                    uint32_t* width,
+                                    uint32_t* height,
+                                    size_t* size) {
+  for (std::list<BufferData>::iterator it = input_buffer_data_.begin();
+       it != input_buffer_data_.end();
+       ++it) {
+    if (it->bitstream_buffer_id != bitstream_buffer_id)
+      continue;
+    *timestamp = it->timestamp;
+    *width = it->width;
+    *height = it->height;
+    return;
+  }
+  NOTREACHED() << "Missing bitstream buffer id: " << bitstream_buffer_id;
+}
+
+}  // namespace content