Quelle  MediaDecoderStateMachine.cpp   Sprache: C

/* -*- Mode: C++; tab-width: 2; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
/* vim:set ts=2 sw=2 sts=2 et cindent: */
/* This Source Code Form is subject to the terms of the Mozilla Public
 * License, v. 2.0. If a copy of the MPL was not distributed with this
 * file, You can obtain one at http://mozilla.org/MPL/2.0/. */

#include "MediaDecoderStateMachine.h"

#include <algorithm>
#include <stdint.h>
#include <utility>

#include "AudioSegment.h"
#include "DOMMediaStream.h"
#include "ImageContainer.h"
#include "MediaDecoder.h"
#include "MediaShutdownManager.h"
#include "MediaTimer.h"
#include "MediaTrackGraph.h"
#include "PerformanceRecorder.h"
#include "ReaderProxy.h"
#include "TimeUnits.h"
#include "VideoSegment.h"
#include "VideoUtils.h"
#include "mediasink/AudioSink.h"
#include "mediasink/AudioSinkWrapper.h"
#include "mediasink/DecodedStream.h"
#include "mediasink/VideoSink.h"
#include "mozilla/Logging.h"
#include "mozilla/MathAlgorithms.h"
#include "mozilla/NotNull.h"
#include "mozilla/Preferences.h"
#include "mozilla/ProfilerLabels.h"
#include "mozilla/ProfilerMarkerTypes.h"
#include "mozilla/ProfilerMarkers.h"
#include "mozilla/SharedThreadPool.h"
#include "mozilla/Sprintf.h"
#include "mozilla/StaticPrefs_media.h"
#include "mozilla/TaskQueue.h"
#include "mozilla/Telemetry.h"
#include "nsIMemoryReporter.h"
#include "nsPrintfCString.h"
#include "nsTArray.h"

namespace mozilla {

using namespace mozilla::media;

#define NS_DispatchToMainThread(...) \
  CompileError_UseAbstractThreadDispatchInstead

// avoid redefined macro in unified build
#undef FMT
#undef LOG
#undef LOGV
#undef LOGW
#undef LOGE
#undef SFMT
#undef SLOG
#undef SLOGW
#undef SLOGE

#define FMT(x, ...) "Decoder=%p " x, mDecoderID, ##__VA_ARGS__
#define LOG(x, ...)                                                         \
  DDMOZ_LOG(gMediaDecoderLog, LogLevel::Debug, "Decoder=%p " x, mDecoderID, \
            ##__VA_ARGS__)
#define LOGV(x, ...)                                                          \
  DDMOZ_LOG(gMediaDecoderLog, LogLevel::Verbose, "Decoder=%p " x, mDecoderID, \
            ##__VA_ARGS__)
#define LOGW(x, ...) NS_WARNING(nsPrintfCString(FMT(x, ##__VA_ARGS__)).get())
#define LOGE(x, ...)                                                   \
  NS_DebugBreak(NS_DEBUG_WARNING,                                      \
                nsPrintfCString(FMT(x, ##__VA_ARGS__)).get(), nullptr, \
                __FILE__, __LINE__)

// Used by StateObject and its sub-classes
#define SFMT(x, ...)                                                     \
  "Decoder=%p state=%s " x, mMaster->mDecoderID, ToStateStr(GetState()), \
      ##__VA_ARGS__
#define SLOG(x, ...)                                                     \
  DDMOZ_LOGEX(mMaster, gMediaDecoderLog, LogLevel::Debug, "state=%s " x, \
              ToStateStr(GetState()), ##__VA_ARGS__)
#define SLOGW(x, ...) NS_WARNING(nsPrintfCString(SFMT(x, ##__VA_ARGS__)).get())
#define SLOGE(x, ...)                                                   \
  NS_DebugBreak(NS_DEBUG_WARNING,                                       \
                nsPrintfCString(SFMT(x, ##__VA_ARGS__)).get(), nullptr, \
                __FILE__, __LINE__)

// Certain constants get stored as member variables and then adjusted by various
// scale factors on a per-decoder basis. We want to make sure to avoid using
// these constants directly, so we put them in a namespace.
namespace detail {

// Resume a suspended video decoder to the current playback position plus this
// time premium for compensating the seeking delay.
static constexpr auto RESUME_VIDEO_PREMIUM = TimeUnit::FromMicroseconds(125000);

static const int64_t AMPLE_AUDIO_USECS = 2000000;

// If more than this much decoded audio is queued, we'll hold off
// decoding more audio.
static constexpr auto AMPLE_AUDIO_THRESHOLD =
    TimeUnit::FromMicroseconds(AMPLE_AUDIO_USECS);

}  // namespace detail

// If we have fewer than LOW_VIDEO_FRAMES decoded frames, and
// we're not "prerolling video", we'll skip the video up to the next keyframe
// which is at or after the current playback position.
static const uint32_t LOW_VIDEO_FRAMES = 2;

// Arbitrary "frame duration" when playing only audio.
static const uint32_t AUDIO_DURATION_USECS = 40000;

namespace detail {

// If we have less than this much buffered data available, we'll consider
// ourselves to be running low on buffered data. We determine how much
// buffered data we have remaining using the reader's GetBuffered()
// implementation.
static const int64_t LOW_BUFFER_THRESHOLD_USECS = 5000000;

static constexpr auto LOW_BUFFER_THRESHOLD =
    TimeUnit::FromMicroseconds(LOW_BUFFER_THRESHOLD_USECS);

// LOW_BUFFER_THRESHOLD_USECS needs to be greater than AMPLE_AUDIO_USECS,
// otherwise the skip-to-keyframe logic can activate when we're running low on
// data.
static_assert(LOW_BUFFER_THRESHOLD_USECS > AMPLE_AUDIO_USECS,
              "LOW_BUFFER_THRESHOLD_USECS is too small");

}  // namespace detail

// Amount of excess data to add in to the "should we buffer" calculation.
static constexpr auto EXHAUSTED_DATA_MARGIN =
    TimeUnit::FromMicroseconds(100000);

static const uint32_t MIN_VIDEO_QUEUE_SIZE = 3;
static const uint32_t MAX_VIDEO_QUEUE_SIZE = 10;
#ifdef MOZ_APPLEMEDIA
static const uint32_t HW_VIDEO_QUEUE_SIZE = 10;
#else
static const uint32_t HW_VIDEO_QUEUE_SIZE = 3;
#endif
static const uint32_t VIDEO_QUEUE_SEND_TO_COMPOSITOR_SIZE = 9999;

static uint32_t sVideoQueueDefaultSize = MAX_VIDEO_QUEUE_SIZE;
static uint32_t sVideoQueueHWAccelSize = HW_VIDEO_QUEUE_SIZE;
static uint32_t sVideoQueueSendToCompositorSize =
    VIDEO_QUEUE_SEND_TO_COMPOSITOR_SIZE;

static void InitVideoQueuePrefs() {
  MOZ_ASSERT(NS_IsMainThread());
  static bool sPrefInit = false;
  if (!sPrefInit) {
    sPrefInit = true;
    sVideoQueueDefaultSize = Preferences::GetUint(
        "media.video-queue.default-size", MAX_VIDEO_QUEUE_SIZE);
    sVideoQueueHWAccelSize = Preferences::GetUint(
        "media.video-queue.hw-accel-size", HW_VIDEO_QUEUE_SIZE);
    sVideoQueueSendToCompositorSize =
        Preferences::GetUint("media.video-queue.send-to-compositor-size",
                             VIDEO_QUEUE_SEND_TO_COMPOSITOR_SIZE);
  }
}

template <typename Type, typename Function>
static void DiscardFramesFromTail(MediaQueue<Type>& aQueue,
                                  const Function&& aTest) {
  while (aQueue.GetSize()) {
    if (aTest(aQueue.PeekBack()->mTime.ToMicroseconds())) {
      RefPtr<Type> releaseMe = aQueue.PopBack();
      continue;
    }
    break;
  }
}

// Delay, in milliseconds, that tabs needs to be in background before video
// decoding is suspended.
static TimeDuration SuspendBackgroundVideoDelay() {
  return TimeDuration::FromMilliseconds(
      StaticPrefs::media_suspend_background_video_delay_ms());
}

class MediaDecoderStateMachine::StateObject {
 public:
  virtual ~StateObject() = default;
  virtual void Exit() {}  // Exit action.
  virtual void Step() {}  // Perform a 'cycle' of this state object.
  virtual State GetState() const = 0;

  // Event handlers for various events.
  virtual void HandleAudioCaptured() {}
  virtual void HandleAudioDecoded(AudioData* aAudio) {
    Crash("Unexpected event!", __func__);
  }
  virtual void HandleVideoDecoded(VideoData* aVideo) {
    Crash("Unexpected event!", __func__);
  }
  virtual void HandleAudioWaited(MediaData::Type aType) {
    Crash("Unexpected event!", __func__);
  }
  virtual void HandleVideoWaited(MediaData::Type aType) {
    Crash("Unexpected event!", __func__);
  }
  virtual void HandleWaitingForAudio() { Crash("Unexpected event!", __func__); }
  virtual void HandleAudioCanceled() { Crash("Unexpected event!", __func__); }
  virtual void HandleEndOfAudio() { Crash("Unexpected event!", __func__); }
  virtual void HandleWaitingForVideo() { Crash("Unexpected event!", __func__); }
  virtual void HandleVideoCanceled() { Crash("Unexpected event!", __func__); }
  virtual void HandleEndOfVideo() { Crash("Unexpected event!", __func__); }

  virtual RefPtr<MediaDecoder::SeekPromise> HandleSeek(
      const SeekTarget& aTarget);

  virtual RefPtr<ShutdownPromise> HandleShutdown();

  virtual void HandleVideoSuspendTimeout() = 0;

  virtual void HandleResumeVideoDecoding(const TimeUnit& aTarget);

  virtual void HandlePlayStateChanged(MediaDecoder::PlayState aPlayState) {}

  virtual void GetDebugInfo(
      dom::MediaDecoderStateMachineDecodingStateDebugInfo& aInfo) {}

  virtual void HandleLoopingChanged() {}

 private:
  template <class S, typename R, typename... As>
  auto ReturnTypeHelper(R (S::*)(As...)) -> R;

  void Crash(const char* aReason, const char* aSite) {
    char buf[1024];
    SprintfLiteral(buf, "%s state=%s callsite=%s", aReason,
                   ToStateStr(GetState()), aSite);
    MOZ_ReportAssertionFailure(buf, __FILE__, __LINE__);
    MOZ_CRASH();
  }

 protected:
  enum class EventVisibility : int8_t { Observable, Suppressed };

  using Master = MediaDecoderStateMachine;
  explicit StateObject(Master* aPtr) : mMaster(aPtr) {}
  TaskQueue* OwnerThread() const { return mMaster->mTaskQueue; }
  ReaderProxy* Reader() const { return mMaster->mReader; }
  const MediaInfo& Info() const { return mMaster->Info(); }
  MediaQueue<AudioData>& AudioQueue() const { return mMaster->mAudioQueue; }
  MediaQueue<VideoData>& VideoQueue() const { return mMaster->mVideoQueue; }

  template <class S, typename... Args, size_t... Indexes>
  auto CallEnterMemberFunction(S* aS, std::tuple<Args...>& aTuple,
                               std::index_sequence<Indexes...>)
      -> decltype(ReturnTypeHelper(&S::Enter)) {
    AUTO_PROFILER_LABEL("StateObject::CallEnterMemberFunction", MEDIA_PLAYBACK);
    return aS->Enter(std::move(std::get<Indexes>(aTuple))...);
  }

  // Note this function will delete the current state object.
  // Don't access members to avoid UAF after this call.
  template <class S, typename... Ts>
  auto SetState(Ts&&... aArgs) -> decltype(ReturnTypeHelper(&S::Enter)) {
    // |aArgs| must be passed by reference to avoid passing MOZ_NON_PARAM class
    // SeekJob by value.  See bug 1287006 and bug 1338374.  But we still *must*
    // copy the parameters, because |Exit()| can modify them.  See bug 1312321.
    // So we 1) pass the parameters by reference, but then 2) immediately copy
    // them into a Tuple to be safe against modification, and finally 3) move
    // the elements of the Tuple into the final function call.
    auto copiedArgs = std::make_tuple(std::forward<Ts>(aArgs)...);

    // Copy mMaster which will reset to null.
    auto* master = mMaster;

    auto* s = new S(master);

    // It's possible to seek again during seeking, otherwise the new state
    // should always be different from the original one.
    MOZ_ASSERT(GetState() != s->GetState() ||
               GetState() == DECODER_STATE_SEEKING_ACCURATE ||
               GetState() == DECODER_STATE_SEEKING_FROMDORMANT ||
               GetState() == DECODER_STATE_SEEKING_NEXTFRAMESEEKING ||
               GetState() == DECODER_STATE_SEEKING_VIDEOONLY);

    SLOG("change state to: %s", ToStateStr(s->GetState()));
    PROFILER_MARKER_TEXT("MDSM::StateChange", MEDIA_PLAYBACK, {},
                         nsPrintfCString("%s", ToStateStr(s->GetState())));

    Exit();

    // Delete the old state asynchronously to avoid UAF if the caller tries to
    // access its members after SetState() returns.
    master->OwnerThread()->DispatchDirectTask(
        NS_NewRunnableFunction("MDSM::StateObject::DeleteOldState",
                               [toDelete = std::move(master->mStateObj)]() {}));
    // Also reset mMaster to catch potentail UAF.
    mMaster = nullptr;

    master->mStateObj.reset(s);
    return CallEnterMemberFunction(s, copiedArgs,
                                   std::index_sequence_for<Ts...>{});
  }

  RefPtr<MediaDecoder::SeekPromise> SetSeekingState(
      SeekJob&& aSeekJob, EventVisibility aVisibility);

  void SetDecodingState();

  // Take a raw pointer in order not to change the life cycle of MDSM.
  // It is guaranteed to be valid by MDSM.
  Master* mMaster;
};

/**
 * Purpose: decode metadata like duration and dimensions of the media resource.
 *
 * Transition to other states when decoding metadata is done:
 *   SHUTDOWN if failing to decode metadata.
 *   DECODING_FIRSTFRAME otherwise.
 */
class MediaDecoderStateMachine::DecodeMetadataState
    : public MediaDecoderStateMachine::StateObject {
 public:
  explicit DecodeMetadataState(Master* aPtr) : StateObject(aPtr) {}

  void Enter() {
    MOZ_ASSERT(!mMaster->mVideoDecodeSuspended);
    MOZ_ASSERT(!mMetadataRequest.Exists());
    SLOG("Dispatching AsyncReadMetadata");

    // We disconnect mMetadataRequest in Exit() so it is fine to capture
    // a raw pointer here.
    Reader()
        ->ReadMetadata()
        ->Then(
            OwnerThread(), __func__,
            [this](MetadataHolder&& aMetadata) {
              OnMetadataRead(std::move(aMetadata));
            },
            [this](const MediaResult& aError) { OnMetadataNotRead(aError); })
        ->Track(mMetadataRequest);
  }

  void Exit() override { mMetadataRequest.DisconnectIfExists(); }

  State GetState() const override { return DECODER_STATE_DECODING_METADATA; }

  RefPtr<MediaDecoder::SeekPromise> HandleSeek(
      const SeekTarget& aTarget) override {
    MOZ_DIAGNOSTIC_CRASH("Can't seek while decoding metadata.");
    return MediaDecoder::SeekPromise::CreateAndReject(true, __func__);
  }

  void HandleVideoSuspendTimeout() override {
    // Do nothing since no decoders are created yet.
  }

  void HandleResumeVideoDecoding(const TimeUnit&) override {
    // We never suspend video decoding in this state.
    MOZ_ASSERT(false, "Shouldn't have suspended video decoding.");
  }

 private:
  void OnMetadataRead(MetadataHolder&& aMetadata);

  void OnMetadataNotRead(const MediaResult& aError) {
    AUTO_PROFILER_LABEL("DecodeMetadataState::OnMetadataNotRead",
                        MEDIA_PLAYBACK);

    mMetadataRequest.Complete();
    SLOGE("Decode metadata failed, shutting down decoder");
    mMaster->DecodeError(aError);
  }

  MozPromiseRequestHolder<MediaFormatReader::MetadataPromise> mMetadataRequest;
};

/**
 * Purpose: release decoder resources to save memory and hardware resources.
 *
 * Transition to:
 *   SEEKING if any seek request or play state changes to PLAYING.
 */
class MediaDecoderStateMachine::DormantState
    : public MediaDecoderStateMachine::StateObject {
 public:
  explicit DormantState(Master* aPtr) : StateObject(aPtr) {}

  void Enter() {
    if (mMaster->IsPlaying()) {
      mMaster->StopPlayback();
    }

    // Calculate the position to seek to when exiting dormant.
    auto t = mMaster->mMediaSink->IsStarted() ? mMaster->GetClock()
                                              : mMaster->GetMediaTime();
    mMaster->AdjustByLooping(t);
    mPendingSeek.mTarget.emplace(t, SeekTarget::Accurate);
    // SeekJob asserts |mTarget.IsValid() == !mPromise.IsEmpty()| so we
    // need to create the promise even it is not used at all.
    // The promise may be used when coming out of DormantState into
    // SeekingState.
    RefPtr<MediaDecoder::SeekPromise> x =
        mPendingSeek.mPromise.Ensure(__func__);

    // Reset the decoding state to ensure that any queued video frames are
    // released and don't consume video memory.
    mMaster->ResetDecode();

    // No need to call StopMediaSink() here.
    // We will do it during seeking when exiting dormant.

    // Ignore WAIT_FOR_DATA since we won't decode in dormant.
    mMaster->mAudioWaitRequest.DisconnectIfExists();
    mMaster->mVideoWaitRequest.DisconnectIfExists();

    MaybeReleaseResources();
  }

  void Exit() override {
    // mPendingSeek is either moved when exiting dormant or
    // should be rejected here before transition to SHUTDOWN.
    mPendingSeek.RejectIfExists(__func__);
  }

  State GetState() const override { return DECODER_STATE_DORMANT; }

  RefPtr<MediaDecoder::SeekPromise> HandleSeek(
      const SeekTarget& aTarget) override;

  void HandleVideoSuspendTimeout() override {
    // Do nothing since we've released decoders in Enter().
  }

  void HandleResumeVideoDecoding(const TimeUnit&) override {
    // Do nothing since we won't resume decoding until exiting dormant.
  }

  void HandlePlayStateChanged(MediaDecoder::PlayState aPlayState) override;

  void HandleAudioDecoded(AudioData*) override { MaybeReleaseResources(); }
  void HandleVideoDecoded(VideoData*) override { MaybeReleaseResources(); }
  void HandleWaitingForAudio() override { MaybeReleaseResources(); }
  void HandleWaitingForVideo() override { MaybeReleaseResources(); }
  void HandleAudioCanceled() override { MaybeReleaseResources(); }
  void HandleVideoCanceled() override { MaybeReleaseResources(); }
  void HandleEndOfAudio() override { MaybeReleaseResources(); }
  void HandleEndOfVideo() override { MaybeReleaseResources(); }

 private:
  void MaybeReleaseResources() {
    if (!mMaster->mAudioDataRequest.Exists() &&
        !mMaster->mVideoDataRequest.Exists()) {
      // Release decoders only when they are idle. Otherwise it might cause
      // decode error later when resetting decoders during seeking.
      mMaster->mReader->ReleaseResources();
    }
  }

  SeekJob mPendingSeek;
};

/**
 * Purpose: decode the 1st audio and video frames to fire the 'loadeddata'
 * event.
 *
 * Transition to:
 *   SHUTDOWN if any decode error.
 *   SEEKING if any seek request.
 *   DECODING/LOOPING_DECODING when the 'loadeddata' event is fired.
 */
class MediaDecoderStateMachine::DecodingFirstFrameState
    : public MediaDecoderStateMachine::StateObject {
 public:
  explicit DecodingFirstFrameState(Master* aPtr) : StateObject(aPtr) {}

  void Enter();

  void Exit() override {
    // mPendingSeek is either moved in MaybeFinishDecodeFirstFrame()
    // or should be rejected here before transition to SHUTDOWN.
    mPendingSeek.RejectIfExists(__func__);
  }

  State GetState() const override { return DECODER_STATE_DECODING_FIRSTFRAME; }

  void HandleAudioDecoded(AudioData* aAudio) override {
    mMaster->PushAudio(aAudio);
    MaybeFinishDecodeFirstFrame();
  }

  void HandleVideoDecoded(VideoData* aVideo) override {
    mMaster->PushVideo(aVideo);
    MaybeFinishDecodeFirstFrame();
  }

  void HandleWaitingForAudio() override {
    mMaster->WaitForData(MediaData::Type::AUDIO_DATA);
  }

  void HandleAudioCanceled() override { mMaster->RequestAudioData(); }

  void HandleEndOfAudio() override {
    AudioQueue().Finish();
    MaybeFinishDecodeFirstFrame();
  }

  void HandleWaitingForVideo() override {
    mMaster->WaitForData(MediaData::Type::VIDEO_DATA);
  }

  void HandleVideoCanceled() override {
    mMaster->RequestVideoData(media::TimeUnit());
  }

  void HandleEndOfVideo() override {
    VideoQueue().Finish();
    MaybeFinishDecodeFirstFrame();
  }

  void HandleAudioWaited(MediaData::Type aType) override {
    mMaster->RequestAudioData();
  }

  void HandleVideoWaited(MediaData::Type aType) override {
    mMaster->RequestVideoData(media::TimeUnit());
  }

  void HandleVideoSuspendTimeout() override {
    // Do nothing for we need to decode the 1st video frame to get the
    // dimensions.
  }

  void HandleResumeVideoDecoding(const TimeUnit&) override {
    // We never suspend video decoding in this state.
    MOZ_ASSERT(false, "Shouldn't have suspended video decoding.");
  }

  RefPtr<MediaDecoder::SeekPromise> HandleSeek(
      const SeekTarget& aTarget) override {
    if (mMaster->mIsMSE) {
      return StateObject::HandleSeek(aTarget);
    }
    // Delay seek request until decoding first frames for non-MSE media.
    SLOG("Not Enough Data to seek at this stage, queuing seek");
    mPendingSeek.RejectIfExists(__func__);
    mPendingSeek.mTarget.emplace(aTarget);
    return mPendingSeek.mPromise.Ensure(__func__);
  }

 private:
  // Notify FirstFrameLoaded if having decoded first frames and
  // transition to SEEKING if there is any pending seek, or DECODING otherwise.
  void MaybeFinishDecodeFirstFrame();

  SeekJob mPendingSeek;
};

/**
 * Purpose: decode audio/video data for playback.
 *
 * Transition to:
 *   DORMANT if playback is paused for a while.
 *   SEEKING if any seek request.
 *   SHUTDOWN if any decode error.
 *   BUFFERING if playback can't continue due to lack of decoded data.
 *   COMPLETED when having decoded all audio/video data.
 *   LOOPING_DECODING when media start seamless looping
 */
class MediaDecoderStateMachine::DecodingState
    : public MediaDecoderStateMachine::StateObject {
 public:
  explicit DecodingState(Master* aPtr)
      : StateObject(aPtr), mDormantTimer(OwnerThread()) {}

  void Enter();

  void Exit() override {
    if (!mDecodeStartTime.IsNull()) {
      TimeDuration decodeDuration = TimeStamp::Now() - mDecodeStartTime;
      SLOG("Exiting DECODING, decoded for %.3lfs", decodeDuration.ToSeconds());
    }
    mDormantTimer.Reset();
    mOnAudioPopped.DisconnectIfExists();
    mOnVideoPopped.DisconnectIfExists();
  }

  void Step() override;

  State GetState() const override { return DECODER_STATE_DECODING; }

  void HandleAudioDecoded(AudioData* aAudio) override {
    mMaster->PushAudio(aAudio);
    DispatchDecodeTasksIfNeeded();
    MaybeStopPrerolling();
  }

  void HandleVideoDecoded(VideoData* aVideo) override {
    // We only do this check when we're not looping, which can be known by
    // checking the queue's offset.
    const auto currentTime = mMaster->GetMediaTime();
    if (aVideo->GetEndTime() < currentTime &&
        VideoQueue().GetOffset() == media::TimeUnit::Zero()) {
      if (!mVideoFirstLateTime) {
        mVideoFirstLateTime = Some(TimeStamp::Now());
      }
      PROFILER_MARKER("Video falling behind", MEDIA_PLAYBACK, {},
                      VideoFallingBehindMarker, aVideo->mTime.ToMicroseconds(),
                      currentTime.ToMicroseconds());
      SLOG("video %" PRId64 " starts being late (current=%" PRId64 ")",
           aVideo->mTime.ToMicroseconds(), currentTime.ToMicroseconds());
    } else {
      mVideoFirstLateTime.reset();
    }
    mMaster->PushVideo(aVideo);
    DispatchDecodeTasksIfNeeded();
    MaybeStopPrerolling();
  }

  void HandleAudioCanceled() override { mMaster->RequestAudioData(); }

  void HandleVideoCanceled() override {
    mMaster->RequestVideoData(mMaster->GetMediaTime(),
                              ShouldRequestNextKeyFrame());
  }

  void HandleEndOfAudio() override;
  void HandleEndOfVideo() override;

  void HandleWaitingForAudio() override {
    mMaster->WaitForData(MediaData::Type::AUDIO_DATA);
    MaybeStopPrerolling();
  }

  void HandleWaitingForVideo() override {
    mMaster->WaitForData(MediaData::Type::VIDEO_DATA);
    MaybeStopPrerolling();
  }

  void HandleAudioWaited(MediaData::Type aType) override {
    mMaster->RequestAudioData();
  }

  void HandleVideoWaited(MediaData::Type aType) override {
    mMaster->RequestVideoData(mMaster->GetMediaTime(),
                              ShouldRequestNextKeyFrame());
  }

  void HandleAudioCaptured() override {
    MaybeStopPrerolling();
    // MediaSink is changed. Schedule Step() to check if we can start playback.
    mMaster->ScheduleStateMachine();
  }

  void HandleVideoSuspendTimeout() override {
    // No video, so nothing to suspend.
    if (!mMaster->HasVideo()) {
      return;
    }

    PROFILER_MARKER_UNTYPED("MDSM::EnterVideoSuspend", MEDIA_PLAYBACK);
    mMaster->mVideoDecodeSuspended = true;
    mMaster->mOnPlaybackEvent.Notify(MediaPlaybackEvent::EnterVideoSuspend);
    Reader()->SetVideoBlankDecode(true);
  }

  void HandlePlayStateChanged(MediaDecoder::PlayState aPlayState) override {
    // Schedule Step() to check if we can start or stop playback.
    mMaster->ScheduleStateMachine();
    if (aPlayState == MediaDecoder::PLAY_STATE_PLAYING) {
      // Try to dispatch decoding tasks for mMinimizePreroll might be reset.
      DispatchDecodeTasksIfNeeded();
    }

    if (aPlayState == MediaDecoder::PLAY_STATE_PAUSED) {
      StartDormantTimer();
      mVideoFirstLateTime.reset();
    } else {
      mDormantTimer.Reset();
    }
  }

  void GetDebugInfo(
      dom::MediaDecoderStateMachineDecodingStateDebugInfo& aInfo) override {
    aInfo.mIsPrerolling = mIsPrerolling;
  }

  void HandleLoopingChanged() override { SetDecodingState(); }

 protected:
  virtual void EnsureAudioDecodeTaskQueued();
  virtual void EnsureVideoDecodeTaskQueued();

  virtual bool ShouldStopPrerolling() const {
    return mIsPrerolling &&
           (DonePrerollingAudio() ||
            IsWaitingData(MediaData::Type::AUDIO_DATA)) &&
           (DonePrerollingVideo() ||
            IsWaitingData(MediaData::Type::VIDEO_DATA));
  }

  virtual bool IsWaitingData(MediaData::Type aType) const {
    if (aType == MediaData::Type::AUDIO_DATA) {
      return mMaster->IsWaitingAudioData();
    }
    MOZ_ASSERT(aType == MediaData::Type::VIDEO_DATA);
    return mMaster->IsWaitingVideoData();
  }

  void MaybeStopPrerolling() {
    if (ShouldStopPrerolling()) {
      mIsPrerolling = false;
      // Check if we can start playback.
      mMaster->ScheduleStateMachine();
    }
  }

  bool ShouldRequestNextKeyFrame() const {
    if (!mVideoFirstLateTime) {
      return false;
    }
    const double elapsedTimeMs =
        (TimeStamp::Now() - *mVideoFirstLateTime).ToMilliseconds();
    const bool rv = elapsedTimeMs >=
                    StaticPrefs::media_decoder_skip_when_video_too_slow_ms();
    if (rv) {
      PROFILER_MARKER_UNTYPED("Skipping to next keyframe", MEDIA_PLAYBACK);
      SLOG(
          "video has been late behind media time for %f ms, should skip to "
          "next key frame",
          elapsedTimeMs);
    }
    return rv;
  }

  virtual bool IsBufferingAllowed() const { return true; }

 private:
  void DispatchDecodeTasksIfNeeded();
  void MaybeStartBuffering();

  // At the start of decoding we want to "preroll" the decode until we've
  // got a few frames decoded before we consider whether decode is falling
  // behind. Otherwise our "we're falling behind" logic will trigger
  // unnecessarily if we start playing as soon as the first sample is
  // decoded. These two fields store how many video frames and audio
  // samples we must consume before are considered to be finished prerolling.
  TimeUnit AudioPrerollThreshold() const {
    return (mMaster->mAmpleAudioThreshold / 2)
        .MultDouble(mMaster->mPlaybackRate);
  }

  uint32_t VideoPrerollFrames() const {
    return std::min(
        static_cast<uint32_t>(
            mMaster->GetAmpleVideoFrames() / 2. * mMaster->mPlaybackRate + 1),
        sVideoQueueDefaultSize);
  }

  bool DonePrerollingAudio() const {
    return !mMaster->IsAudioDecoding() ||
           mMaster->GetDecodedAudioDuration() >= AudioPrerollThreshold();
  }

  bool DonePrerollingVideo() const {
    return !mMaster->IsVideoDecoding() ||
           static_cast<uint32_t>(mMaster->VideoQueue().GetSize()) >=
               VideoPrerollFrames();
  }

  void StartDormantTimer() {
    if (!mMaster->mMediaSeekable) {
      // Don't enter dormant if the media is not seekable because we need to
      // seek when exiting dormant.
      return;
    }

    auto timeout = StaticPrefs::media_dormant_on_pause_timeout_ms();
    if (timeout < 0) {
      // Disabled when timeout is negative.
      return;
    }

    if (timeout == 0) {
      // Enter dormant immediately without scheduling a timer.
      SetState<DormantState>();
      return;
    }

    if (mMaster->mMinimizePreroll) {
      SetState<DormantState>();
      return;
    }

    TimeStamp target =
        TimeStamp::Now() + TimeDuration::FromMilliseconds(timeout);

    mDormantTimer.Ensure(
        target,
        [this]() {
          AUTO_PROFILER_LABEL("DecodingState::StartDormantTimer:SetDormant",
                              MEDIA_PLAYBACK);
          mDormantTimer.CompleteRequest();
          SetState<DormantState>();
        },
        [this]() { mDormantTimer.CompleteRequest(); });
  }

  // Time at which we started decoding.
  TimeStamp mDecodeStartTime;

  // When we start decoding (either for the first time, or after a pause)
  // we may be low on decoded data. We don't want our "low data" logic to
  // kick in and decide that we're low on decoded data because the download
  // can't keep up with the decode, and cause us to pause playback. So we
  // have a "preroll" stage, where we ignore the results of our "low data"
  // logic during the first few frames of our decode. This occurs during
  // playback.
  bool mIsPrerolling = true;

  // Fired when playback is paused for a while to enter dormant.
  DelayedScheduler<TimeStamp> mDormantTimer;

  MediaEventListener mOnAudioPopped;
  MediaEventListener mOnVideoPopped;

  // If video has been later than the media time, this will records when the
  // video started being late. It will be reset once video catches up with the
  // media time.
  Maybe<TimeStamp> mVideoFirstLateTime;
};

/**
 * Purpose: decode audio data for playback when media is in seamless
 * looping, we will adjust media time to make samples time monotonically
 * increasing. All its methods runs on its owner thread (MDSM thread).
 *
 * Transition to:
 *   DORMANT if playback is paused for a while.
 *   SEEKING if any seek request.
 *   SHUTDOWN if any decode error.
 *   BUFFERING if playback can't continue due to lack of decoded data.
 *   COMPLETED when the media resource is closed and no data is available
 *             anymore.
 *   DECODING when media stops seamless looping.
 */
class MediaDecoderStateMachine::LoopingDecodingState
    : public MediaDecoderStateMachine::DecodingState {
 public:
  explicit LoopingDecodingState(Master* aPtr)
      : DecodingState(aPtr),
        mIsReachingAudioEOS(!mMaster->IsAudioDecoding()),
        mIsReachingVideoEOS(!mMaster->IsVideoDecoding()),
        mAudioEndedBeforeEnteringStateWithoutDuration(false),
        mVideoEndedBeforeEnteringStateWithoutDuration(false) {
    MOZ_ASSERT(mMaster->mLooping);
    SLOG(
        "LoopingDecodingState ctor, mIsReachingAudioEOS=%d, "
        "mIsReachingVideoEOS=%d",
        mIsReachingAudioEOS, mIsReachingVideoEOS);
    // If the track has reached EOS and we already have its last data, then we
    // can know its duration. But if playback starts from EOS (due to seeking),
    // the decoded end time would be zero because none of data gets decoded yet.
    if (mIsReachingAudioEOS) {
      if (mMaster->HasLastDecodedData(MediaData::Type::AUDIO_DATA) &&
          !mMaster->mAudioTrackDecodedDuration) {
        mMaster->mAudioTrackDecodedDuration.emplace(
            mMaster->mDecodedAudioEndTime);
        SLOG("determine mAudioTrackDecodedDuration");
      } else {
        mAudioEndedBeforeEnteringStateWithoutDuration = true;
        SLOG("still don't know mAudioTrackDecodedDuration");
      }
    }

    if (mIsReachingVideoEOS) {
      if (mMaster->HasLastDecodedData(MediaData::Type::VIDEO_DATA) &&
          !mMaster->mVideoTrackDecodedDuration) {
        mMaster->mVideoTrackDecodedDuration.emplace(
            mMaster->mDecodedVideoEndTime);
        SLOG("determine mVideoTrackDecodedDuration");
      } else {
        mVideoEndedBeforeEnteringStateWithoutDuration = true;
        SLOG("still don't know mVideoTrackDecodedDuration");
      }
    }

    // We might be able to determine the duration already, let's check.
    if (mIsReachingAudioEOS || mIsReachingVideoEOS) {
      Unused << DetermineOriginalDecodedDurationIfNeeded();
    }

    // If we've looped at least once before, then we need to update queue offset
    // correctly to make the media data time and the clock time consistent.
    // Otherwise, it would cause a/v desync.
    if (mMaster->mOriginalDecodedDuration != media::TimeUnit::Zero()) {
      if (mIsReachingAudioEOS && mMaster->HasAudio()) {
        AudioQueue().SetOffset(AudioQueue().GetOffset() +
                               mMaster->mOriginalDecodedDuration);
      }
      if (mIsReachingVideoEOS && mMaster->HasVideo()) {
        VideoQueue().SetOffset(VideoQueue().GetOffset() +
                               mMaster->mOriginalDecodedDuration);
      }
    }
  }

  void Enter() {
    if (mMaster->HasAudio() && mIsReachingAudioEOS) {
      SLOG("audio has ended, request the data again.");
      RequestDataFromStartPosition(TrackInfo::TrackType::kAudioTrack);
    }
    if (mMaster->HasVideo() && mIsReachingVideoEOS) {
      SLOG("video has ended, request the data again.");
      RequestDataFromStartPosition(TrackInfo::TrackType::kVideoTrack);
    }
    DecodingState::Enter();
  }

  void Exit() override {
    MOZ_DIAGNOSTIC_ASSERT(mMaster->OnTaskQueue());
    SLOG("Leaving looping state, offset [a=%" PRId64 ",v=%" PRId64
         "], endtime [a=%" PRId64 ",v=%" PRId64 "], track duration [a=%" PRId64
         ",v=%" PRId64 "], waiting=%s",
         AudioQueue().GetOffset().ToMicroseconds(),
         VideoQueue().GetOffset().ToMicroseconds(),
         mMaster->mDecodedAudioEndTime.ToMicroseconds(),
         mMaster->mDecodedVideoEndTime.ToMicroseconds(),
         mMaster->mAudioTrackDecodedDuration
             ? mMaster->mAudioTrackDecodedDuration->ToMicroseconds()
             : 0,
         mMaster->mVideoTrackDecodedDuration
             ? mMaster->mVideoTrackDecodedDuration->ToMicroseconds()
             : 0,
         mDataWaitingTimestampAdjustment
             ? MediaData::EnumValueToString(
                   mDataWaitingTimestampAdjustment->mType)
             : "none");
    if (ShouldDiscardLoopedData(MediaData::Type::AUDIO_DATA)) {
      DiscardLoopedData(MediaData::Type::AUDIO_DATA);
    }
    if (ShouldDiscardLoopedData(MediaData::Type::VIDEO_DATA)) {
      DiscardLoopedData(MediaData::Type::VIDEO_DATA);
    }

    if (mMaster->HasAudio() && HasDecodedLastAudioFrame()) {
      SLOG("Mark audio queue as finished");
      mMaster->mAudioDataRequest.DisconnectIfExists();
      mMaster->mAudioWaitRequest.DisconnectIfExists();
      AudioQueue().Finish();
    }
    if (mMaster->HasVideo() && HasDecodedLastVideoFrame()) {
      SLOG("Mark video queue as finished");
      mMaster->mVideoDataRequest.DisconnectIfExists();
      mMaster->mVideoWaitRequest.DisconnectIfExists();
      VideoQueue().Finish();
    }

    // Clear waiting data should be done after marking queue as finished.
    mDataWaitingTimestampAdjustment = nullptr;

    mAudioDataRequest.DisconnectIfExists();
    mVideoDataRequest.DisconnectIfExists();
    mAudioSeekRequest.DisconnectIfExists();
    mVideoSeekRequest.DisconnectIfExists();
    DecodingState::Exit();
  }

  ~LoopingDecodingState() {
    MOZ_DIAGNOSTIC_ASSERT(!mAudioDataRequest.Exists());
    MOZ_DIAGNOSTIC_ASSERT(!mVideoDataRequest.Exists());
    MOZ_DIAGNOSTIC_ASSERT(!mAudioSeekRequest.Exists());
    MOZ_DIAGNOSTIC_ASSERT(!mVideoSeekRequest.Exists());
  }

  State GetState() const override { return DECODER_STATE_LOOPING_DECODING; }

  void HandleAudioDecoded(AudioData* aAudio) override {
    // TODO : check if we need to update mOriginalDecodedDuration

    // After pushing data to the queue, timestamp might be adjusted.
    DecodingState::HandleAudioDecoded(aAudio);
    mMaster->mDecodedAudioEndTime =
        std::max(aAudio->GetEndTime(), mMaster->mDecodedAudioEndTime);
    SLOG("audio sample after time-adjustment [%" PRId64 ",%" PRId64 "]",
         aAudio->mTime.ToMicroseconds(), aAudio->GetEndTime().ToMicroseconds());
  }

  void HandleVideoDecoded(VideoData* aVideo) override {
    // TODO : check if we need to update mOriginalDecodedDuration

    // Here sample still keeps its original timestamp.

    // This indicates there is a shorter audio track, and it's the first time in
    // the looping (audio ends but video is playing) so that we haven't been
    // able to determine the decoded duration. Therefore, we fill the gap
    // between two tracks before video ends. Afterward, this adjustment will be
    // done in `HandleEndOfAudio()`.
    if (mMaster->mOriginalDecodedDuration == media::TimeUnit::Zero() &&
        mMaster->mAudioTrackDecodedDuration &&
        aVideo->GetEndTime() > *mMaster->mAudioTrackDecodedDuration) {
      media::TimeUnit gap;
      // First time we fill gap between the video frame to the last audio.
      if (auto prevVideo = VideoQueue().PeekBack();
          prevVideo &&
          prevVideo->GetEndTime() < *mMaster->mAudioTrackDecodedDuration) {
        gap =
            aVideo->GetEndTime().ToBase(*mMaster->mAudioTrackDecodedDuration) -
            *mMaster->mAudioTrackDecodedDuration;
      }
      // Then fill the gap for all following videos.
      else {
        gap = aVideo->mDuration.ToBase(*mMaster->mAudioTrackDecodedDuration);
      }
      SLOG("Longer video %" PRId64 "%s (audio-durtaion=%" PRId64
           "%s), insert silence to fill the gap %" PRId64 "%s",
           aVideo->GetEndTime().ToMicroseconds(),
           aVideo->GetEndTime().ToString().get(),
           mMaster->mAudioTrackDecodedDuration->ToMicroseconds(),
           mMaster->mAudioTrackDecodedDuration->ToString().get(),
           gap.ToMicroseconds(), gap.ToString().get());
      PushFakeAudioDataIfNeeded(gap);
    }

    // After pushing data to the queue, timestamp might be adjusted.
    DecodingState::HandleVideoDecoded(aVideo);
    mMaster->mDecodedVideoEndTime =
        std::max(aVideo->GetEndTime(), mMaster->mDecodedVideoEndTime);
    SLOG("video sample after time-adjustment [%" PRId64 ",%" PRId64 "]",
         aVideo->mTime.ToMicroseconds(), aVideo->GetEndTime().ToMicroseconds());
  }

  void HandleEndOfAudio() override {
    mIsReachingAudioEOS = true;
    if (!mMaster->mAudioTrackDecodedDuration &&
        mMaster->HasLastDecodedData(MediaData::Type::AUDIO_DATA)) {
      mMaster->mAudioTrackDecodedDuration.emplace(
          mMaster->mDecodedAudioEndTime);
    }
    if (DetermineOriginalDecodedDurationIfNeeded()) {
      AudioQueue().SetOffset(AudioQueue().GetOffset() +
                             mMaster->mOriginalDecodedDuration);
    }

    // This indicates that the audio track is shorter than the video track, so
    // we need to add some silence to fill the gap.
    if (mMaster->mAudioTrackDecodedDuration &&
        mMaster->mOriginalDecodedDuration >
            *mMaster->mAudioTrackDecodedDuration) {
      MOZ_ASSERT(mMaster->HasVideo());
      MOZ_ASSERT(mMaster->mVideoTrackDecodedDuration);
      MOZ_ASSERT(mMaster->mOriginalDecodedDuration ==
                 *mMaster->mVideoTrackDecodedDuration);
      auto gap = mMaster->mOriginalDecodedDuration.ToBase(
                     *mMaster->mAudioTrackDecodedDuration) -
                 *mMaster->mAudioTrackDecodedDuration;
      SLOG(
          "Audio track is shorter than the original decoded duration "
          "(a=%" PRId64 "%s, t=%" PRId64
          "%s), insert silence to fill the gap %" PRId64 "%s",
          mMaster->mAudioTrackDecodedDuration->ToMicroseconds(),
          mMaster->mAudioTrackDecodedDuration->ToString().get(),
          mMaster->mOriginalDecodedDuration.ToMicroseconds(),
          mMaster->mOriginalDecodedDuration.ToString().get(),
          gap.ToMicroseconds(), gap.ToString().get());
      PushFakeAudioDataIfNeeded(gap);
    }

    SLOG(
        "received audio EOS when seamless looping, starts seeking, "
        "audioLoopingOffset=[%" PRId64 "], mAudioTrackDecodedDuration=[%" PRId64
        "]",
        AudioQueue().GetOffset().ToMicroseconds(),
        mMaster->mAudioTrackDecodedDuration->ToMicroseconds());
    if (!IsRequestingDataFromStartPosition(MediaData::Type::AUDIO_DATA)) {
      RequestDataFromStartPosition(TrackInfo::TrackType::kAudioTrack);
    }
    ProcessSamplesWaitingAdjustmentIfAny();
  }

  void HandleEndOfVideo() override {
    mIsReachingVideoEOS = true;
    if (!mMaster->mVideoTrackDecodedDuration &&
        mMaster->HasLastDecodedData(MediaData::Type::VIDEO_DATA)) {
      mMaster->mVideoTrackDecodedDuration.emplace(
          mMaster->mDecodedVideoEndTime);
    }
    if (DetermineOriginalDecodedDurationIfNeeded()) {
      VideoQueue().SetOffset(VideoQueue().GetOffset() +
                             mMaster->mOriginalDecodedDuration);
    }

    SLOG(
        "received video EOS when seamless looping, starts seeking, "
        "videoLoopingOffset=[%" PRId64 "], mVideoTrackDecodedDuration=[%" PRId64
        "]",
        VideoQueue().GetOffset().ToMicroseconds(),
        mMaster->mVideoTrackDecodedDuration->ToMicroseconds());
    if (!IsRequestingDataFromStartPosition(MediaData::Type::VIDEO_DATA)) {
      RequestDataFromStartPosition(TrackInfo::TrackType::kVideoTrack);
    }
    ProcessSamplesWaitingAdjustmentIfAny();
  }

 private:
  void RequestDataFromStartPosition(TrackInfo::TrackType aType) {
    MOZ_DIAGNOSTIC_ASSERT(aType == TrackInfo::TrackType::kAudioTrack ||
                          aType == TrackInfo::TrackType::kVideoTrack);

    const bool isAudio = aType == TrackInfo::TrackType::kAudioTrack;
    MOZ_ASSERT_IF(isAudio, mMaster->HasAudio());
    MOZ_ASSERT_IF(!isAudio, mMaster->HasVideo());

    if (IsReaderSeeking()) {
      MOZ_ASSERT(!mPendingSeekingType);
      mPendingSeekingType = Some(aType);
      SLOG("Delay %s seeking until the reader finishes current seeking",
           isAudio ? "audio" : "video");
      return;
    }

    auto& seekRequest = isAudio ? mAudioSeekRequest : mVideoSeekRequest;
    Reader()->ResetDecode(aType);
    Reader()
        ->Seek(SeekTarget(media::TimeUnit::Zero(), SeekTarget::Type::Accurate,
                          isAudio ? SeekTarget::Track::AudioOnly
                                  : SeekTarget::Track::VideoOnly))
        ->Then(
            OwnerThread(), __func__,
            [this, isAudio, master = RefPtr{mMaster}]() mutable -> void {
              AUTO_PROFILER_LABEL(
                  nsPrintfCString(
                      "LoopingDecodingState::RequestDataFromStartPosition(%s)::"
                      "SeekResolved",
                      isAudio ? "audio" : "video")
                      .get(),
                  MEDIA_PLAYBACK);
              if (auto& state = master->mStateObj;
                  state &&
                  state->GetState() != DECODER_STATE_LOOPING_DECODING) {
                MOZ_RELEASE_ASSERT(false, "This shouldn't happen!");
                return;
              }
              if (isAudio) {
                mAudioSeekRequest.Complete();
              } else {
                mVideoSeekRequest.Complete();
              }
              SLOG(
                  "seeking completed, start to request first %s sample "
                  "(queued=%zu, decoder-queued=%zu)",
                  isAudio ? "audio" : "video",
                  isAudio ? AudioQueue().GetSize() : VideoQueue().GetSize(),
                  isAudio ? Reader()->SizeOfAudioQueueInFrames()
                          : Reader()->SizeOfVideoQueueInFrames());
              if (isAudio) {
                RequestAudioDataFromReaderAfterEOS();
              } else {
                RequestVideoDataFromReaderAfterEOS();
              }
              if (mPendingSeekingType) {
                auto seekingType = *mPendingSeekingType;
                mPendingSeekingType.reset();
                SLOG("Perform pending %s seeking", TrackTypeToStr(seekingType));
                RequestDataFromStartPosition(seekingType);
              }
            },
            [this, isAudio, master = RefPtr{mMaster}](
                const SeekRejectValue& aReject) mutable -> void {
              AUTO_PROFILER_LABEL(
                  nsPrintfCString("LoopingDecodingState::"
                                  "RequestDataFromStartPosition(%s)::"
                                  "SeekRejected",
                                  isAudio ? "audio" : "video")
                      .get(),
                  MEDIA_PLAYBACK);
              if (auto& state = master->mStateObj;
                  state &&
                  state->GetState() != DECODER_STATE_LOOPING_DECODING) {
                MOZ_RELEASE_ASSERT(false, "This shouldn't happen!");
                return;
              }
              if (isAudio) {
                mAudioSeekRequest.Complete();
              } else {
                mVideoSeekRequest.Complete();
              }
              HandleError(aReject.mError, isAudio);
            })
        ->Track(seekRequest);
  }

  void RequestAudioDataFromReaderAfterEOS() {
    MOZ_ASSERT(mMaster->HasAudio());
    Reader()
        ->RequestAudioData()
        ->Then(
            OwnerThread(), __func__,
            [this, master = RefPtr{mMaster}](const RefPtr<AudioData>& aAudio) {
              AUTO_PROFILER_LABEL(
                  "LoopingDecodingState::"
                  "RequestAudioDataFromReader::"
                  "RequestDataResolved",
                  MEDIA_PLAYBACK);
              if (auto& state = master->mStateObj;
                  state &&
                  state->GetState() != DECODER_STATE_LOOPING_DECODING) {
                MOZ_RELEASE_ASSERT(false, "This shouldn't happen!");
                return;
              }
              mIsReachingAudioEOS = false;
              mAudioDataRequest.Complete();
              SLOG(
                  "got audio decoded sample "
                  "[%" PRId64 ",%" PRId64 "]",
                  aAudio->mTime.ToMicroseconds(),
                  aAudio->GetEndTime().ToMicroseconds());
              if (ShouldPutDataOnWaiting(MediaData::Type::AUDIO_DATA)) {
                SLOG(
                    "decoded audio sample needs to wait for timestamp "
                    "adjustment after EOS");
                PutDataOnWaiting(aAudio);
                return;
              }
              HandleAudioDecoded(aAudio);
              ProcessSamplesWaitingAdjustmentIfAny();
            },
            [this, master = RefPtr{mMaster}](const MediaResult& aError) {
              AUTO_PROFILER_LABEL(
                  "LoopingDecodingState::"
                  "RequestAudioDataFromReader::"
                  "RequestDataRejected",
                  MEDIA_PLAYBACK);
              if (auto& state = master->mStateObj;
                  state &&
                  state->GetState() != DECODER_STATE_LOOPING_DECODING) {
                MOZ_RELEASE_ASSERT(false, "This shouldn't happen!");
                return;
              }
              mAudioDataRequest.Complete();
              HandleError(aError, true /* isAudio */);
            })
        ->Track(mAudioDataRequest);
  }

  void RequestVideoDataFromReaderAfterEOS() {
    MOZ_ASSERT(mMaster->HasVideo());
    Reader()
        ->RequestVideoData(media::TimeUnit(),
                           false /* aRequestNextVideoKeyFrame */)
        ->Then(
            OwnerThread(), __func__,
            [this, master = RefPtr{mMaster}](const RefPtr<VideoData>& aVideo) {
              AUTO_PROFILER_LABEL(
                  "LoopingDecodingState::"
                  "RequestVideoDataFromReaderAfterEOS()::"
                  "RequestDataResolved",
                  MEDIA_PLAYBACK);
              if (auto& state = master->mStateObj;
                  state &&
                  state->GetState() != DECODER_STATE_LOOPING_DECODING) {
                MOZ_RELEASE_ASSERT(false, "This shouldn't happen!");
                return;
              }
              mIsReachingVideoEOS = false;
              mVideoDataRequest.Complete();
              SLOG(
                  "got video decoded sample "
                  "[%" PRId64 ",%" PRId64 "]",
                  aVideo->mTime.ToMicroseconds(),
                  aVideo->GetEndTime().ToMicroseconds());
              if (ShouldPutDataOnWaiting(MediaData::Type::VIDEO_DATA)) {
                SLOG(
                    "decoded video sample needs to wait for timestamp "
                    "adjustment after EOS");
                PutDataOnWaiting(aVideo);
                return;
              }
              mMaster->mBypassingSkipToNextKeyFrameCheck = true;
              HandleVideoDecoded(aVideo);
              ProcessSamplesWaitingAdjustmentIfAny();
            },
            [this, master = RefPtr{mMaster}](const MediaResult& aError) {
              AUTO_PROFILER_LABEL(
                  "LoopingDecodingState::"
                  "RequestVideoDataFromReaderAfterEOS()::"
                  "RequestDataRejected",
                  MEDIA_PLAYBACK);
              if (auto& state = master->mStateObj;
                  state &&
                  state->GetState() != DECODER_STATE_LOOPING_DECODING) {
                MOZ_RELEASE_ASSERT(false, "This shouldn't happen!");
                return;
              }
              mVideoDataRequest.Complete();
              HandleError(aError, false /* isAudio */);
            })
        ->Track(mVideoDataRequest);
  }

  void HandleError(const MediaResult& aError, bool aIsAudio);

  bool ShouldRequestData(MediaData::Type aType) const {
    MOZ_DIAGNOSTIC_ASSERT(aType == MediaData::Type::AUDIO_DATA ||
                          aType == MediaData::Type::VIDEO_DATA);

    if (aType == MediaData::Type::AUDIO_DATA &&
        (mAudioSeekRequest.Exists() || mAudioDataRequest.Exists() ||
         IsDataWaitingForTimestampAdjustment(MediaData::Type::AUDIO_DATA) ||
         mMaster->IsWaitingAudioData())) {
      return false;
    }
    if (aType == MediaData::Type::VIDEO_DATA &&
        (mVideoSeekRequest.Exists() || mVideoDataRequest.Exists() ||
         IsDataWaitingForTimestampAdjustment(MediaData::Type::VIDEO_DATA) ||
         mMaster->IsWaitingVideoData())) {
      return false;
    }
    return true;
  }

  void HandleAudioCanceled() override {
    if (ShouldRequestData(MediaData::Type::AUDIO_DATA)) {
      mMaster->RequestAudioData();
    }
  }

  void HandleAudioWaited(MediaData::Type aType) override {
    if (ShouldRequestData(MediaData::Type::AUDIO_DATA)) {
      mMaster->RequestAudioData();
    }
  }

  void HandleVideoCanceled() override {
    if (ShouldRequestData(MediaData::Type::VIDEO_DATA)) {
      mMaster->RequestVideoData(mMaster->GetMediaTime(),
                                ShouldRequestNextKeyFrame());
    };
  }

  void HandleVideoWaited(MediaData::Type aType) override {
    if (ShouldRequestData(MediaData::Type::VIDEO_DATA)) {
      mMaster->RequestVideoData(mMaster->GetMediaTime(),
                                ShouldRequestNextKeyFrame());
    };
  }

  void EnsureAudioDecodeTaskQueued() override {
    if (!ShouldRequestData(MediaData::Type::AUDIO_DATA)) {
      return;
    }
    DecodingState::EnsureAudioDecodeTaskQueued();
  }

  void EnsureVideoDecodeTaskQueued() override {
    if (!ShouldRequestData(MediaData::Type::VIDEO_DATA)) {
      return;
    }
    DecodingState::EnsureVideoDecodeTaskQueued();
  }

  bool DetermineOriginalDecodedDurationIfNeeded() {
    // Duration would only need to be set once, unless we get more data which is
    // larger than the duration. That can happen on MSE (reopen stream).
    if (mMaster->mOriginalDecodedDuration != media::TimeUnit::Zero()) {
      return true;
    }

    // Single track situations
    if (mMaster->HasAudio() && !mMaster->HasVideo() &&
        mMaster->mAudioTrackDecodedDuration) {
      mMaster->mOriginalDecodedDuration = *mMaster->mAudioTrackDecodedDuration;
      SLOG("audio only, duration=%" PRId64,
           mMaster->mOriginalDecodedDuration.ToMicroseconds());
      return true;
    }
    if (mMaster->HasVideo() && !mMaster->HasAudio() &&
        mMaster->mVideoTrackDecodedDuration) {
      mMaster->mOriginalDecodedDuration = *mMaster->mVideoTrackDecodedDuration;
      SLOG("video only, duration=%" PRId64,
           mMaster->mOriginalDecodedDuration.ToMicroseconds());
      return true;
    }
    // Two tracks situation
    if (mMaster->HasAudio() && mMaster->HasVideo()) {
      // Both tracks have ended so that we can check which track is longer.
      if (mMaster->mAudioTrackDecodedDuration &&
          mMaster->mVideoTrackDecodedDuration) {
        mMaster->mOriginalDecodedDuration =
            std::max(*mMaster->mVideoTrackDecodedDuration,
                     *mMaster->mAudioTrackDecodedDuration);
        SLOG("Both tracks ended, original duration=%" PRId64 " (a=%" PRId64
             ", v=%" PRId64 ")",
             mMaster->mOriginalDecodedDuration.ToMicroseconds(),
             mMaster->mAudioTrackDecodedDuration->ToMicroseconds(),
             mMaster->mVideoTrackDecodedDuration->ToMicroseconds());
        return true;
      }
      // When entering the state, video has ended but audio hasn't, which means
      // audio is longer.
      if (mMaster->mAudioTrackDecodedDuration &&
          mVideoEndedBeforeEnteringStateWithoutDuration) {
        mMaster->mOriginalDecodedDuration =
            *mMaster->mAudioTrackDecodedDuration;
        mVideoEndedBeforeEnteringStateWithoutDuration = false;
        SLOG("audio is longer, duration=%" PRId64,
             mMaster->mOriginalDecodedDuration.ToMicroseconds());
        return true;
      }
      // When entering the state, audio has ended but video hasn't, which means
      // video is longer.
      if (mMaster->mVideoTrackDecodedDuration &&
          mAudioEndedBeforeEnteringStateWithoutDuration) {
        mMaster->mOriginalDecodedDuration =
            *mMaster->mVideoTrackDecodedDuration;
        mAudioEndedBeforeEnteringStateWithoutDuration = false;
        SLOG("video is longer, duration=%" PRId64,
             mMaster->mOriginalDecodedDuration.ToMicroseconds());
        return true;
      }
      SLOG("Still waiting for another track ends...");
      MOZ_ASSERT(!mMaster->mAudioTrackDecodedDuration ||
                 !mMaster->mVideoTrackDecodedDuration);
    }
    SLOG("can't determine the original decoded duration yet");
    MOZ_ASSERT(mMaster->mOriginalDecodedDuration == media::TimeUnit::Zero());
    return false;
  }

  void ProcessSamplesWaitingAdjustmentIfAny() {
    if (!mDataWaitingTimestampAdjustment) {
      return;
    }

    RefPtr<MediaData> data = mDataWaitingTimestampAdjustment;
    mDataWaitingTimestampAdjustment = nullptr;
    const bool isAudio = data->mType == MediaData::Type::AUDIO_DATA;
    SLOG("process %s sample waiting for timestamp adjustment",
         isAudio ? "audio" : "video");
    if (isAudio) {
      // Waiting sample is for next round of looping, so the queue offset
      // shouldn't be zero. This happens when the track has reached EOS before
      // entering the state (and looping never happens before). Same for below
      // video case.
      if (AudioQueue().GetOffset() == media::TimeUnit::Zero()) {
        AudioQueue().SetOffset(mMaster->mOriginalDecodedDuration);
      }
      HandleAudioDecoded(data->As<AudioData>());
    } else {
      MOZ_DIAGNOSTIC_ASSERT(data->mType == MediaData::Type::VIDEO_DATA);
      if (VideoQueue().GetOffset() == media::TimeUnit::Zero()) {
        VideoQueue().SetOffset(mMaster->mOriginalDecodedDuration);
      }
      HandleVideoDecoded(data->As<VideoData>());
    }
  }

  bool IsDataWaitingForTimestampAdjustment(MediaData::Type aType) const {
    return mDataWaitingTimestampAdjustment &&
           mDataWaitingTimestampAdjustment->mType == aType;
  }

  bool ShouldPutDataOnWaiting(MediaData::Type aType) const {
    // If another track is already waiting, this track shouldn't be waiting.
    // This case only happens when both tracks reached EOS before entering the
    // looping decoding state, so we don't know the decoded duration yet (used
    // to adjust timestamp) But this is fine, because both tracks will start
    // from 0 so we don't need to adjust them now.
    if (mDataWaitingTimestampAdjustment &&
        !IsDataWaitingForTimestampAdjustment(aType)) {
      return false;
    }

    // Only have one track, no need to wait.
    if ((aType == MediaData::Type::AUDIO_DATA && !mMaster->HasVideo()) ||
        (aType == MediaData::Type::VIDEO_DATA && !mMaster->HasAudio())) {
      return false;
    }

    // We don't know the duration yet, so we can't calculate the looping offset.
    return mMaster->mOriginalDecodedDuration == media::TimeUnit::Zero();
  }

  void PutDataOnWaiting(MediaData* aData) {
    MOZ_ASSERT(!mDataWaitingTimestampAdjustment);
    mDataWaitingTimestampAdjustment = aData;
    SLOG("put %s [%" PRId64 ",%" PRId64 "] on waiting",
         MediaData::EnumValueToString(aData->mType),
         aData->mTime.ToMicroseconds(), aData->GetEndTime().ToMicroseconds());
    MaybeStopPrerolling();
  }

  bool ShouldDiscardLoopedData(MediaData::Type aType) const {
    if (!mMaster->mMediaSink->IsStarted()) {
      return false;
    }

    MOZ_DIAGNOSTIC_ASSERT(aType == MediaData::Type::AUDIO_DATA ||
                          aType == MediaData::Type::VIDEO_DATA);
    const bool isAudio = aType == MediaData::Type::AUDIO_DATA;
    if (isAudio && !mMaster->HasAudio()) {
      return false;
    }
    if (!isAudio && !mMaster->HasVideo()) {
      return false;
    }

    /**
     * If media cancels looping, we should check whether there is media data
     * whose time is later than EOS. If so, we should discard them because we
     * won't have a chance to play them.
     *
     *    playback                     last decoded
     *    position          EOS        data time
     *   ----|---------------|------------|---------> (Increasing timeline)
     *    mCurrent         looping      mMaster's
     *    ClockTime        offset      mDecodedXXXEndTime
     *
     */
    const auto offset =
        isAudio ? AudioQueue().GetOffset() : VideoQueue().GetOffset();
    const auto endTime =
        isAudio ? mMaster->mDecodedAudioEndTime : mMaster->mDecodedVideoEndTime;
    const auto clockTime = mMaster->GetClock();
    return (offset != media::TimeUnit::Zero() && clockTime < offset &&
            offset < endTime);
  }

  void DiscardLoopedData(MediaData::Type aType) {
    MOZ_DIAGNOSTIC_ASSERT(aType == MediaData::Type::AUDIO_DATA ||
                          aType == MediaData::Type::VIDEO_DATA);
    const bool isAudio = aType == MediaData::Type::AUDIO_DATA;
    const auto offset =
        isAudio ? AudioQueue().GetOffset() : VideoQueue().GetOffset();
    if (offset == media::TimeUnit::Zero()) {
      return;
    }

    SLOG("Discard %s frames after the time=%" PRId64,
         isAudio ? "audio" : "video", offset.ToMicroseconds());
    if (isAudio) {
      DiscardFramesFromTail(AudioQueue(), [&](int64_t aSampleTime) {
        return aSampleTime > offset.ToMicroseconds();
      });
    } else {
      DiscardFramesFromTail(VideoQueue(), [&](int64_t aSampleTime) {
        return aSampleTime > offset.ToMicroseconds();
      });
    }
  }

  void PushFakeAudioDataIfNeeded(const media::TimeUnit& aDuration) {
    MOZ_ASSERT(Info().HasAudio());

    const auto& audioInfo = Info().mAudio;
    CheckedInt64 frames = aDuration.ToTicksAtRate(audioInfo.mRate);
    if (!frames.isValid() || !audioInfo.mChannels || !audioInfo.mRate) {
      NS_WARNING("Can't create fake audio, invalid frames/channel/rate?");
      return;
    }

    if (!frames.value()) {
      NS_WARNING(nsPrintfCString("Duration (%s) too short, no frame needed",
                                 aDuration.ToString().get())
                     .get());
      return;
    }

    // If we can get the last sample, use its frame. Otherwise, use common 1024.
    int64_t typicalPacketFrameCount = 1024;
    if (RefPtr<AudioData> audio = AudioQueue().PeekBack()) {
      typicalPacketFrameCount = audio->Frames();
    }

    media::TimeUnit totalDuration = TimeUnit::Zero(audioInfo.mRate);
    // Generate fake audio in a smaller size of audio chunk.
    while (frames.value()) {
      int64_t packetFrameCount =
          std::min(frames.value(), typicalPacketFrameCount);
      frames -= packetFrameCount;
      AlignedAudioBuffer samples(packetFrameCount * audioInfo.mChannels);
      if (!samples) {
        NS_WARNING("Can't create audio buffer, OOM?");
        return;
      }
      // `mDecodedAudioEndTime` is adjusted time, and we want unadjusted time
      // otherwise the time would be adjusted twice when pushing sample into the
      // media queue.
      media::TimeUnit startTime = mMaster->mDecodedAudioEndTime;
      if (AudioQueue().GetOffset() != media::TimeUnit::Zero()) {
        startTime -= AudioQueue().GetOffset();
      }
      RefPtr<AudioData> data(new AudioData(0, startTime, std::move(samples),
                                           audioInfo.mChannels,
                                           audioInfo.mRate));
      SLOG("Created fake audio data (duration=%s, frame-left=%" PRId64 ")",
           data->mDuration.ToString().get(), frames.value());
      totalDuration += data->mDuration;
      HandleAudioDecoded(data);
    }
    SLOG("Pushed fake silence audio data in total duration=%" PRId64 "%s",
         totalDuration.ToMicroseconds(), totalDuration.ToString().get());
  }

  bool HasDecodedLastAudioFrame() const {
    // when we're going to leave looping state and have got EOS before, we
    // should mark audio queue as ended because we have got all data we need.
    return mAudioDataRequest.Exists() || mAudioSeekRequest.Exists() ||
           ShouldDiscardLoopedData(MediaData::Type::AUDIO_DATA) ||
           IsDataWaitingForTimestampAdjustment(MediaData::Type::AUDIO_DATA) ||
           mIsReachingAudioEOS;
  }

  bool HasDecodedLastVideoFrame() const {
    // when we're going to leave looping state and have got EOS before, we
    // should mark video queue as ended because we have got all data we need.
    return mVideoDataRequest.Exists() || mVideoSeekRequest.Exists() ||
           ShouldDiscardLoopedData(MediaData::Type::VIDEO_DATA) ||
           IsDataWaitingForTimestampAdjustment(MediaData::Type::VIDEO_DATA) ||
           mIsReachingVideoEOS;
  }

  bool ShouldStopPrerolling() const override {
    // These checks is used to handle the media queue aren't opened correctly
    // because they've been close before entering the looping state. Therefore,
    // we need to preroll data in order to let new data to reopen the queue
    // automatically. Otherwise, playback can't start successfully.
    bool isWaitingForNewData = false;
    if (mMaster->HasAudio()) {
      isWaitingForNewData |= (mIsReachingAudioEOS && AudioQueue().IsFinished());
    }
    if (mMaster->HasVideo()) {
      isWaitingForNewData |= (mIsReachingVideoEOS && VideoQueue().IsFinished());
    }
    return !isWaitingForNewData && DecodingState::ShouldStopPrerolling();
  }

  bool IsReaderSeeking() const {
    return mAudioSeekRequest.Exists() || mVideoSeekRequest.Exists();
  }

  bool IsWaitingData(MediaData::Type aType) const override {
    if (aType == MediaData::Type::AUDIO_DATA) {
      return mMaster->IsWaitingAudioData() ||
             IsDataWaitingForTimestampAdjustment(MediaData::Type::AUDIO_DATA);
    }
    MOZ_DIAGNOSTIC_ASSERT(aType == MediaData::Type::VIDEO_DATA);
    return mMaster->IsWaitingVideoData() ||
           IsDataWaitingForTimestampAdjustment(MediaData::Type::VIDEO_DATA);
  }

  bool IsRequestingDataFromStartPosition(MediaData::Type aType) const {
    MOZ_DIAGNOSTIC_ASSERT(aType == MediaData::Type::AUDIO_DATA ||
                          aType == MediaData::Type::VIDEO_DATA);
    if (aType == MediaData::Type::AUDIO_DATA) {
      return mAudioSeekRequest.Exists() || mAudioDataRequest.Exists();
    }
    return mVideoSeekRequest.Exists() || mVideoDataRequest.Exists();
  }

  bool IsBufferingAllowed() const override {
    return !mIsReachingAudioEOS && !mIsReachingVideoEOS;
  }

  bool mIsReachingAudioEOS;
  bool mIsReachingVideoEOS;

  /**
   * If we have both tracks which have different length, when one track ends
   * first, we can't adjust new data from that track if another longer track
   * hasn't ended yet. The adjusted timestamp needs to be based off the longer
   * track's last data's timestamp, because otherwise it would cause a deviation
   * and eventually a/v unsync. Those sample needs to be stored and we will
   * adjust their timestamp later.
   *
   * Following graph explains the situation in details.
   * o : decoded data with timestamp adjusted or no adjustment (not looping yet)
   * x : decoded data without timestamp adjustment.
   * - : stop decoding and nothing happens
   * EOS : the track reaches to the end. We now know the offset of the track.
   *
   * Timeline ----------------------------------->
   * Track1 :  o EOS x  -  -  o
   * Track2 :  o  o  o EOS o  o
   *
   * Before reaching track2's EOS, we can't adjust samples from track1 because
   * track2 might have longer duration than track1. The sample X would be
   * stored in `mDataWaitingTimestampAdjustment` and we would also stop decoding
   * for track1.
   *
   * After reaching track2's EOS, now we know another track's offset, and the
   * larger one would be used for `mOriginalDecodedDuration`. Once that duration
   * has been determined, we will no longer need to put samples on waiting
   * because we already know how to adjust timestamp.
   */
  RefPtr<MediaData> mDataWaitingTimestampAdjustment;

  MozPromiseRequestHolder<MediaFormatReader::SeekPromise> mAudioSeekRequest;
  MozPromiseRequestHolder<MediaFormatReader::SeekPromise> mVideoSeekRequest;
  MozPromiseRequestHolder<AudioDataPromise> mAudioDataRequest;
  MozPromiseRequestHolder<VideoDataPromise> mVideoDataRequest;

  // The media format reader only allows seeking a track at a time, if we're
  // already in seeking, then delay the new seek until the current one finishes.
--> --------------------

--> maximum size reached

--> --------------------