Android N的Audio系统（五）

• AudioFlinger 回放录制线程

AndioFlinger 作为 Android 的音频系统引擎，重任之一是负责输入输出流设备的管理及音频流数据的处理传输，这是由回放线程（PlaybackThread 及其派生的子类）和录制线程（RecordThread）进行的，我们简单看看回放线程和录制线程类关系：

ThreadBase：PlaybackThread 和 RecordThread 的基类

RecordThread：录制线程类，由 ThreadBase 派生

PlaybackThread：回放线程基类，同由 ThreadBase 派生

MixerThread：混音回放线程类，由 PlaybackThread 派生，负责处理标识为 AUDIO_OUTPUT_FLAG_PRIMARY、AUDIO_OUTPUT_FLAG_FAST、AUDIO_OUTPUT_FLAG_DEEP_BUFFER 的音频流，MixerThread 可以把多个音轨的数据混音后再输出

DirectOutputThread：直输回放线程类，由 PlaybackThread 派生，负责处理标识为 AUDIO_OUTPUT_FLAG_DIRECT 的音频流，这种音频流数据不需要软件混音，直接输出到音频设备即可

DuplicatingThread：复制回放线程类，由 MixerThread 派生，负责复制音频流数据到其他输出设备，使用场景如主声卡设备、蓝牙耳机设备、USB 声卡设备同时输出

OffloadThread：硬解回放线程类，由 DirectOutputThread 派生，负责处理标识为 AUDIO_OUTPUT_FLAG_COMPRESS_OFFLOAD 的音频流，这种音频流未经软件解码的（一般是 MP3、AAC 等格式的数据），需要输出到硬件解码器，由硬件解码器解码成 PCM 数据

PlaybackThread 中有个极为重要的函数 threadLoop()，当 PlaybackThread 被强引用时，threadLoop() 会真正运行起来进入循环主体，处理音频流数据相关事务，threadLoop() 大致流程如下（以 MixerThread 为例）：

bool AudioFlinger::PlaybackThread::threadLoop()
{
    Vector< sp > tracksToRemove;

    mStandbyTimeNs = systemTime();
    nsecs_t lastWriteFinished = -1; // time last server write completed
    int64_t lastFramesWritten = -1; // track changes in timestamp server frames written

    // MIXER
    nsecs_t lastWarning = 0;

    // DUPLICATING
    // FIXME could this be made local to while loop?
    writeFrames = 0;

    int lastGeneration = 0;

    cacheParameters_l();
    mSleepTimeUs = mIdleSleepTimeUs;

    if (mType == MIXER) {
        sleepTimeShift = 0;
    }

    CpuStats cpuStats;
    const String8 myName(String8::format("thread %p type %d TID %d", this, mType, gettid()));

    acquireWakeLock();

#ifdef SRS_PROCESSING
    String8 bt_param = String8("bluetooth_enabled=0");
    POSTPRO_PATCH_PARAMS_SET(bt_param);
    if (mType == MIXER) {
        POSTPRO_PATCH_OUTPROC_PLAY_INIT(this, myName);
    } else if (mType == OFFLOAD) {
        POSTPRO_PATCH_OUTPROC_DIRECT_INIT(this, myName);
        POSTPRO_PATCH_OUTPROC_PLAY_ROUTE_BY_VALUE(this, mOutDevice);
    } else if (mType == DIRECT) {
        POSTPRO_PATCH_OUTPROC_DIRECT_INIT(this, myName);
        POSTPRO_PATCH_OUTPROC_PLAY_ROUTE_BY_VALUE(this, mOutDevice);
    }
#endif

    // mNBLogWriter->log can only be called while thread mutex mLock is held.
    // So if you need to log when mutex is unlocked, set logString to a non-NULL string,
    // and then that string will be logged at the next convenient opportunity.
    const char *logString = NULL;

    checkSilentMode_l();

    while (!exitPending())
    {
        cpuStats.sample(myName);

        Vector< sp > effectChains;

        { // scope for mLock

            Mutex::Autolock _l(mLock);

            processConfigEvents_l();

            if (logString != NULL) {
                mNBLogWriter->logTimestamp();
                mNBLogWriter->log(logString);
                logString = NULL;
            }

            // Gather the framesReleased counters for all active tracks,
            // and associate with the sink frames written out.  We need
            // this to convert the sink timestamp to the track timestamp.
            bool kernelLocationUpdate = false;
            if (mNormalSink != 0) {
                // Note: The DuplicatingThread may not have a mNormalSink.
                // We always fetch the timestamp here because often the downstream
                // sink will block while writing.
                ExtendedTimestamp timestamp; // use private copy to fetch
                (void) mNormalSink->getTimestamp(timestamp);

                // We keep track of the last valid kernel position in case we are in underrun
                // and the normal mixer period is the same as the fast mixer period, or there
                // is some error from the HAL.
                if (mTimestamp.mTimeNs[ExtendedTimestamp::LOCATION_KERNEL] >= 0) {
                    mTimestamp.mPosition[ExtendedTimestamp::LOCATION_KERNEL_LASTKERNELOK] =
                            mTimestamp.mPosition[ExtendedTimestamp::LOCATION_KERNEL];
                    mTimestamp.mTimeNs[ExtendedTimestamp::LOCATION_KERNEL_LASTKERNELOK] =
                            mTimestamp.mTimeNs[ExtendedTimestamp::LOCATION_KERNEL];

                    mTimestamp.mPosition[ExtendedTimestamp::LOCATION_SERVER_LASTKERNELOK] =
                            mTimestamp.mPosition[ExtendedTimestamp::LOCATION_SERVER];
                    mTimestamp.mTimeNs[ExtendedTimestamp::LOCATION_SERVER_LASTKERNELOK] =
                            mTimestamp.mTimeNs[ExtendedTimestamp::LOCATION_SERVER];
                }

                if (timestamp.mTimeNs[ExtendedTimestamp::LOCATION_KERNEL] >= 0) {
                    kernelLocationUpdate = true;
                } else {
                    ALOGVV("getTimestamp error - no valid kernel position");
                }

                // copy over kernel info
                mTimestamp.mPosition[ExtendedTimestamp::LOCATION_KERNEL] =
                        timestamp.mPosition[ExtendedTimestamp::LOCATION_KERNEL]
                        + mSuspendedFrames; // add frames discarded when suspended
                mTimestamp.mTimeNs[ExtendedTimestamp::LOCATION_KERNEL] =
                        timestamp.mTimeNs[ExtendedTimestamp::LOCATION_KERNEL];
            }
            // mFramesWritten for non-offloaded tracks are contiguous
            // even after standby() is called. This is useful for the track frame
            // to sink frame mapping.
            bool serverLocationUpdate = false;
            if (mFramesWritten != lastFramesWritten) {
                serverLocationUpdate = true;
                lastFramesWritten = mFramesWritten;
            }
            // Only update timestamps if there is a meaningful change.
            // Either the kernel timestamp must be valid or we have written something.
            if (kernelLocationUpdate || serverLocationUpdate) {
                if (serverLocationUpdate) {
                    // use the time before we called the HAL write - it is a bit more accurate
                    // to when the server last read data than the current time here.
                    //
                    // If we haven't written anything, mLastWriteTime will be -1
                    // and we use systemTime().
                    mTimestamp.mPosition[ExtendedTimestamp::LOCATION_SERVER] = mFramesWritten;
                    mTimestamp.mTimeNs[ExtendedTimestamp::LOCATION_SERVER] = mLastWriteTime == -1
                            ? systemTime() : mLastWriteTime;
                }
                const size_t size = mActiveTracks.size();
                for (size_t i = 0; i < size; ++i) {
                    sp t = mActiveTracks[i].promote();
                    if (t != 0 && !t->isFastTrack()) {
                        t->updateTrackFrameInfo(
                                t->mAudioTrackServerProxy->framesReleased(),
                                mFramesWritten,
                                mTimestamp);
                    }
                }
            }

            saveOutputTracks();
            if (mSignalPending) {
                // A signal was raised while we were unlocked
                mSignalPending = false;
            } else if (waitingAsyncCallback_l()) {
                if (exitPending()) {
                    break;
                }
                bool released = false;
                if (!keepWakeLock()) {
                    releaseWakeLock_l();
                    released = true;
                    mWakeLockUids.clear();
                    mActiveTracksGeneration++;
                }
                ALOGV("wait async completion");
                mWaitWorkCV.wait(mLock);
                ALOGV("async completion/wake");
                if (released) {
                    acquireWakeLock_l();
                }
                mStandbyTimeNs = systemTime() + mStandbyDelayNs;
                mSleepTimeUs = 0;

                continue;
            }
            if ((!mActiveTracks.size() && systemTime() > mStandbyTimeNs) ||
                                   isSuspended()) {
                // put audio hardware into standby after short delay
                if (shouldStandby_l()) {

                    threadLoop_standby();

                    mStandby = true;
                }

                if (!mActiveTracks.size() && mConfigEvents.isEmpty()) {
                    // we're about to wait, flush the binder command buffer
                    IPCThreadState::self()->flushCommands();

                    clearOutputTracks();

                    if (exitPending()) {
                        break;
                    }

                    releaseWakeLock_l();
                    mWakeLockUids.clear();
                    mActiveTracksGeneration++;
                    // wait until we have something to do...
                    ALOGV("%s going to sleep", myName.string());
                    mWaitWorkCV.wait(mLock);
                    ALOGV("%s waking up", myName.string());
                    acquireWakeLock_l();

                    mMixerStatus = MIXER_IDLE;
                    mMixerStatusIgnoringFastTracks = MIXER_IDLE;
                    mBytesWritten = 0;
                    mBytesRemaining = 0;
                    checkSilentMode_l();

                    mStandbyTimeNs = systemTime() + mStandbyDelayNs;
                    mSleepTimeUs = mIdleSleepTimeUs;
                    if (mType == MIXER) {
                        sleepTimeShift = 0;
                    }

                    continue;
                }
            }
            // mMixerStatusIgnoringFastTracks is also updated internally
            mMixerStatus = prepareTracks_l(&tracksToRemove);
#ifdef DOLBY_ENABLE // DOLBY_DAP_PREGAIN
            // If there are no active tracks, then reset volume to zero for this thread.
            if (mMixerStatus == MIXER_IDLE) {
                ALOGV("EffectDapController: Reset volumes to zeros for threadType = %d flags = %d", mType, mOutput->flags);
                EffectDapController::instance()->updatePregain(mType, mId, mOutput->flags, 0);
            }
#endif // DOLBY_END

            // compare with previously applied list
            if (lastGeneration != mActiveTracksGeneration) {
                // update wakelock
                updateWakeLockUids_l(mWakeLockUids);
                lastGeneration = mActiveTracksGeneration;
            }

            // prevent any changes in effect chain list and in each effect chain
            // during mixing and effect process as the audio buffers could be deleted
            // or modified if an effect is created or deleted
            lockEffectChains_l(effectChains);
        } // mLock scope ends

        if (mBytesRemaining == 0) {
            mCurrentWriteLength = 0;
            if (mMixerStatus == MIXER_TRACKS_READY) {
                // threadLoop_mix() sets mCurrentWriteLength
                threadLoop_mix();
            } else if ((mMixerStatus != MIXER_DRAIN_TRACK)
                        && (mMixerStatus != MIXER_DRAIN_ALL)) {
                // threadLoop_sleepTime sets mSleepTimeUs to 0 if data
                // must be written to HAL
                threadLoop_sleepTime();
                if (mSleepTimeUs == 0) {
                    mCurrentWriteLength = mSinkBufferSize;
                }
            }
            // Either threadLoop_mix() or threadLoop_sleepTime() should have set
            // mMixerBuffer with data if mMixerBufferValid is true and mSleepTimeUs == 0.
            // Merge mMixerBuffer data into mEffectBuffer (if any effects are valid)
            // or mSinkBuffer (if there are no effects).
            //
            // This is done pre-effects computation; if effects change to
            // support higher precision, this needs to move.
            //
            // mMixerBufferValid is only set true by MixerThread::prepareTracks_l().
            // TODO use mSleepTimeUs == 0 as an additional condition.
            if (mMixerBufferValid) {
                void *buffer = mEffectBufferValid ? mEffectBuffer : mSinkBuffer;
                audio_format_t format = mEffectBufferValid ? mEffectBufferFormat : mFormat;

                // mono blend occurs for mixer threads only (not direct or offloaded)
                // and is handled here if we're going directly to the sink.
                if (requireMonoBlend() && !mEffectBufferValid) {
                    mono_blend(mMixerBuffer, mMixerBufferFormat, mChannelCount, mNormalFrameCount,
                               true /*limit*/);
                }

                memcpy_by_audio_format(buffer, format, mMixerBuffer, mMixerBufferFormat,
                        mNormalFrameCount * mChannelCount);
            }

            mBytesRemaining = mCurrentWriteLength;
            if (isSuspended()) {
                // Simulate write to HAL when suspended (e.g. BT SCO phone call).
                mSleepTimeUs = suspendSleepTimeUs(); // assumes full buffer.
                const size_t framesRemaining = mBytesRemaining / mFrameSize;
                mBytesWritten += mBytesRemaining;
                mFramesWritten += framesRemaining;
                mSuspendedFrames += framesRemaining; // to adjust kernel HAL position
                mBytesRemaining = 0;
            }

            // only process effects if we're going to write
            if (mSleepTimeUs == 0 && mType != OFFLOAD &&
                !(mType == DIRECT && mIsDirectPcm)) {
                for (size_t i = 0; i < effectChains.size(); i ++) {
                    effectChains[i]->process_l();
                }
            }
        }
        // Process effect chains for offloaded thread even if no audio
        // was read from audio track: process only updates effect state
        // and thus does have to be synchronized with audio writes but may have
        // to be called while waiting for async write callback
        if ((mType == OFFLOAD) || (mType == DIRECT && mIsDirectPcm)) {
            for (size_t i = 0; i < effectChains.size(); i ++) {
                effectChains[i]->process_l();
            }
        }

#ifdef SRS_PROCESSING
            // Offload thread
            if (mType == OFFLOAD) {
                char buffer[2];
                POSTPRO_PATCH_OUTPROC_DIRECT_SAMPLES(this, AUDIO_FORMAT_PCM_16_BIT, (int16_t *) buffer, 2, 48000, 2);
            }
#endif

        // Only if the Effects buffer is enabled and there is data in the
        // Effects buffer (buffer valid), we need to
        // copy into the sink buffer.
        // TODO use mSleepTimeUs == 0 as an additional condition.
        if (mEffectBufferValid) {
            //ALOGV("writing effect buffer to sink buffer format %#x", mFormat);

            if (requireMonoBlend()) {
                mono_blend(mEffectBuffer, mEffectBufferFormat, mChannelCount, mNormalFrameCount,
                           true /*limit*/);
            }

            memcpy_by_audio_format(mSinkBuffer, mFormat, mEffectBuffer, mEffectBufferFormat,
                    mNormalFrameCount * mChannelCount);
        }

        // enable changes in effect chain
        unlockEffectChains(effectChains);

        if (!waitingAsyncCallback()) {
            // mSleepTimeUs == 0 means we must write to audio hardware
            if (mSleepTimeUs == 0) {
                ssize_t ret = 0;
#ifdef SRS_PROCESSING
                if (mType == MIXER && mMixerStatus == MIXER_TRACKS_READY) {
                    POSTPRO_PATCH_OUTPROC_PLAY_SAMPLES(this, mFormat, mSinkBuffer, mSinkBufferSize, mSampleRate, mChannelCount);
                }
#endif
                // We save lastWriteFinished here, as previousLastWriteFinished,
                // for throttling. On thread start, previousLastWriteFinished will be
                // set to -1, which properly results in no throttling after the first write.
                nsecs_t previousLastWriteFinished = lastWriteFinished;
                nsecs_t delta = 0;
                if (mBytesRemaining) {
                    // FIXME rewrite to reduce number of system calls
                    mLastWriteTime = systemTime();  // also used for dumpsys
                    ret = threadLoop_write();
                    lastWriteFinished = systemTime();
                    delta = lastWriteFinished - mLastWriteTime;
                    if (ret < 0) {
                        mBytesRemaining = 0;
                    } else {
                        mBytesWritten += ret;
                        mBytesRemaining -= ret;
                        mFramesWritten += ret / mFrameSize;
                    }
                } else if ((mMixerStatus == MIXER_DRAIN_TRACK) ||
                        (mMixerStatus == MIXER_DRAIN_ALL)) {
                    threadLoop_drain();
                }
                if (mType == MIXER && !mStandby) {
                    // write blocked detection
                    if (delta > maxPeriod) {
                        mNumDelayedWrites++;
                        if ((lastWriteFinished - lastWarning) > kWarningThrottleNs) {
                            ATRACE_NAME("underrun");
                            ALOGW("write blocked for %llu msecs, %d delayed writes, thread %p",
                                    (unsigned long long) ns2ms(delta), mNumDelayedWrites, this);
                            lastWarning = lastWriteFinished;
                        }
                    }

                    if (mThreadThrottle
                            && mMixerStatus == MIXER_TRACKS_READY // we are mixing (active tracks)
                            && ret > 0) {                         // we wrote something
                        // Limit MixerThread data processing to no more than twice the
                        // expected processing rate.
                        //
                        // This helps prevent underruns with NuPlayer and other applications
                        // which may set up buffers that are close to the minimum size, or use
                        // deep buffers, and rely on a double-buffering sleep strategy to fill.
                        //
                        // The throttle smooths out sudden large data drains from the device,
                        // e.g. when it comes out of standby, which often causes problems with
                        // (1) mixer threads without a fast mixer (which has its own warm-up)
                        // (2) minimum buffer sized tracks (even if the track is full,
                        //     the app won't fill fast enough to handle the sudden draw).
                        //
                        // Total time spent in last processing cycle equals time spent in
                        // 1. threadLoop_write, as well as time spent in
                        // 2. threadLoop_mix (significant for heavy mixing, especially
                        //                    on low tier processors)

                        // it's OK if deltaMs is an overestimate.
                        const int32_t deltaMs =
                                (lastWriteFinished - previousLastWriteFinished) / 1000000;
                        const int32_t throttleMs = mHalfBufferMs - deltaMs;
                        if ((signed)mHalfBufferMs >= throttleMs && throttleMs > 0) {
                            usleep(throttleMs * 1000);
                            // notify of throttle start on verbose log
                            ALOGV_IF(mThreadThrottleEndMs == mThreadThrottleTimeMs,
                                    "mixer(%p) throttle begin:"
                                    " ret(%zd) deltaMs(%d) requires sleep %d ms",
                                    this, ret, deltaMs, throttleMs);
                            mThreadThrottleTimeMs += throttleMs;
                            // Throttle must be attributed to the previous mixer loop's write time
                            // to allow back-to-back throttling.
                            lastWriteFinished += throttleMs * 1000000;
                        } else {
                            uint32_t diff = mThreadThrottleTimeMs - mThreadThrottleEndMs;
                            if (diff > 0) {
                                // notify of throttle end on debug log
                                // but prevent spamming for bluetooth
                                ALOGD_IF(!audio_is_a2dp_out_device(outDevice()),
                                        "mixer(%p) throttle end: throttle time(%u)", this, diff);
                                mThreadThrottleEndMs = mThreadThrottleTimeMs;
                            }
                        }
                    }
                }

            } else {
                ATRACE_BEGIN("sleep");
                Mutex::Autolock _l(mLock);
                if (!mSignalPending && mConfigEvents.isEmpty() && !exitPending()) {
                    mWaitWorkCV.waitRelative(mLock, microseconds((nsecs_t)mSleepTimeUs));
                }
                ATRACE_END();
            }
        }

        // Finally let go of removed track(s), without the lock held
        // since we can't guarantee the destructors won't acquire that
        // same lock.  This will also mutate and push a new fast mixer state.
        threadLoop_removeTracks(tracksToRemove);
        tracksToRemove.clear();

        // FIXME I don't understand the need for this here;
        //       it was in the original code but maybe the
        //       assignment in saveOutputTracks() makes this unnecessary?
        clearOutputTracks();

        // Effect chains will be actually deleted here if they were removed from
        // mEffectChains list during mixing or effects processing
        effectChains.clear();

        // FIXME Note that the above .clear() is no longer necessary since effectChains
        // is now local to this block, but will keep it for now (at least until merge done).
    }

    threadLoop_exit();

    if (!mStandby) {
        threadLoop_standby();
        mStandby = true;
    }

#ifdef SRS_PROCESSING
    if (mType == MIXER) {
        POSTPRO_PATCH_OUTPROC_PLAY_EXIT(this, myName);
    } else if (mType == OFFLOAD) {
        POSTPRO_PATCH_OUTPROC_DIRECT_EXIT(this, myName);
    } else if (mType == DIRECT) {
        POSTPRO_PATCH_OUTPROC_DIRECT_EXIT(this, myName);
    }
#endif

    releaseWakeLock();
    mWakeLockUids.clear();
    mActiveTracksGeneration++;

    ALOGV("Thread %p type %d exiting", this, mType);
    return false;
}

threadLoop() 循环的条件是 exitPending() 返回 false，如果想要 PlaybackThread 结束循环，则可以调用 requestExit() 来请求退出；

processConfigEvents_l() ：处理配置事件；当有配置改变的事件发生时，需要调用 sendConfigEvent_l() 来通知 PlaybackThread，这样 PlaybackThread 才能及时处理配置事件；常见的配置事件是切换音频通路；

检查此时此刻是否符合 standby 条件，比如当前并没有 ACTIVE 状态的 Track（mActiveTracks.size() = 0），那么调用 threadLoop_standby() 关闭音频硬件设备以节省能耗；

prepareTracks_l()： 准备音频流和混音器，该函数非常复杂，这里不详细分析了，仅列一下流程要点：

    遍历 mActiveTracks，逐个处理 mActiveTracks 上的 Track，检查该 Track 是否为 ACTIVE 状态；

    如果 Track 设置是 ACTIVE 状态，则再检查该 Track 的数据是否准备就绪了；

    根据音频流的音量值、格式、声道数、音轨的采样率、硬件设备的采样率，配置好混音器参数；

    如果 Track 的状态是 PAUSED 或 STOPPED，则把该 Track 添加到 tracksToRemove 向量中；

threadLoop_mix()：读取所有置了 ACTIVE 状态的音频流数据，混音器开始处理这些数据；

threadLoop_write()： 把混音器处理后的数据写到输出流设备；

threadLoop_removeTracks()： 把 tracksToRemove 上的所有 Track 从 mActiveTracks 中移除出来；这样下一次循环时就不会处理这些 Track 了。

这里说说 PlaybackThread 与输出流设备的关系：PlaybackThread 实例与输出流设备是一一对应的，比方说 OffloadThread 只会将音频数据输出到 compress_offload 设备中，MixerThread(with FastMixer) 只会将音频数据输出到 low_latency 设备中。

从 Audio HAL 中，我们通常看到如下 4 种输出流设备，分别对应着不同的播放场景：

primary_out：主输出流设备，用于铃声类声音输出，对应着标识为 AUDIO_OUTPUT_FLAG_PRIMARY 的音频流和一个 MixerThread 回放线程实例

low_latency：低延迟输出流设备，用于按键音、游戏背景音等对时延要求高的声音输出，对应着标识为 AUDIO_OUTPUT_FLAG_FAST 的音频流和一个 MixerThread 回放线程实例

deep_buffer：音乐音轨输出流设备，用于音乐等对时延要求不高的声音输出，对应着标识为 AUDIO_OUTPUT_FLAG_DEEP_BUFFER 的音频流和一个 MixerThread 回放线程实例

compress_offload：硬解输出流设备，用于需要硬件解码的数据输出，对应着标识为 AUDIO_OUTPUT_FLAG_COMPRESS_OFFLOAD 的音频流和一个 OffloadThread 回放线程实例

其中 primary_out 设备是必须声明支持的，而且系统启动时就已经打开 primary_out 设备并创建好对应的 MixerThread 实例。其他类型的输出流设备并非必须声明支持的，主要是看硬件上有无这个能力。

可能有人产生这样的疑问：既然 primary_out 设备一直保持打开，那么能耗岂不是很大？这里阐释一个概念：输出流设备属于逻辑设备，并不是硬件设备。所以即使输出流设备一直保持打开，只要硬件设备不工作，那么就不会影响能耗。那么硬件设备什么时候才会打开呢？答案是 PlaybackThread 将音频数据写入到输出流设备时。

下图简单描述 AudioTrack、PlaybackThread、输出流设备三者的对应关系：

我们可以这么说：输出流设备决定了它对应的 PlaybackThread 是什么类型。怎么理解呢？意思是说：只有支持了该类型的输出流设备，那么该类型的 PlaybackThread 才有可能被创建。举个例子：只有硬件上具备硬件解码器，系统才建立 compress_offload 设备，然后播放 mp3 格式的音乐文件时，才会创建 OffloadThread 把数据输出到 compress_offload 设备上；反之，如果硬件上并不具备硬件解码器，系统则不应该建立 compress_offload 设备，那么播放 mp3 格式的音乐文件时，通过 MixerThread 把数据输出到其他输出流设备上。

那么有无可能出现这种情况：底层并不支持 compress_offload 设备，但偏偏有个标识为 AUDIO_OUTPUT_FLAG_COMPRESS_OFFLOAD 的音频流送到 AudioFlinger 了呢？这是不可能的。系统启动时，会检查并保存输入输出流设备的支持信息；播放器在播放 mp3 文件时，首先看 compress_offload 设备是否支持了，如果支持，那么不进行软件解码，直接把数据标识为 AUDIO_OUTPUT_FLAG_COMPRESS_OFFLOAD；如果不支持，那么先进行软件解码，然后把解码好的数据标识为 AUDIO_OUTPUT_FLAG_DEEP_BUFFER，前提是 deep_buffer 设备是支持了的；如果 deep_buffer 设备也不支持，那么把数据标识为 AUDIO_OUTPUT_FLAG_PRIMARY。

系统启动时，就已经打开 primary_out、low_latency、deep_buffer 这三种输出流设备，并创建对应的 MixerThread 了；而此时 DirectOutputThread 与 OffloadThread 不会被创建，直到标识为 AUDIO_OUTPUT_FLAG_DIRECT/AUDIO_OUTPUT_FLAG_COMPRESS_OFFLOAD 的音频流需要输出时，才开始创建 DirectOutputThread/OffloadThread 和打开 direct_out/compress_offload 设备。这一点请参考如下代码，注释非常清晰：

AudioPolicyManager::AudioPolicyManager(AudioPolicyClientInterface *clientInterface)
    :
#ifdef AUDIO_POLICY_TEST
    Thread(false),
#endif //AUDIO_POLICY_TEST
    mLimitRingtoneVolume(false), mLastVoiceVolume(-1.0f),
    mA2dpSuspended(false),
    mAudioPortGeneration(1),
    mBeaconMuteRefCount(0),
    mBeaconPlayingRefCount(0),
    mBeaconMuted(false),
    mTtsOutputAvailable(false),
    mMasterMono(false)
{
    mUidCached = getuid();
    mpClientInterface = clientInterface;

    // TODO: remove when legacy conf file is removed. true on devices that use DRC on the
    // DEVICE_CATEGORY_SPEAKER path to boost soft sounds, used to adjust volume curves accordingly.
    // Note: remove also speaker_drc_enabled from global configuration of XML config file.
    bool speakerDrcEnabled = false;

#ifdef USE_XML_AUDIO_POLICY_CONF
    mVolumeCurves = new VolumeCurvesCollection();
    AudioPolicyConfig config(mHwModules, mAvailableOutputDevices, mAvailableInputDevices,
                             mDefaultOutputDevice, speakerDrcEnabled,
                             static_cast<VolumeCurvesCollection *>(mVolumeCurves));
    PolicySerializer serializer;
    if (serializer.deserialize(AUDIO_POLICY_XML_CONFIG_FILE, config) != NO_ERROR) {
#else
    mVolumeCurves = new StreamDescriptorCollection();
    AudioPolicyConfig config(mHwModules, mAvailableOutputDevices, mAvailableInputDevices,
                             mDefaultOutputDevice, speakerDrcEnabled);
    if ((ConfigParsingUtils::loadConfig(AUDIO_POLICY_VENDOR_CONFIG_FILE, config) != NO_ERROR) &&
            (ConfigParsingUtils::loadConfig(AUDIO_POLICY_CONFIG_FILE, config) != NO_ERROR)) {
#endif
        ALOGE("could not load audio policy configuration file, setting defaults");
        config.setDefault();
    }
    // must be done after reading the policy (since conditionned by Speaker Drc Enabling)
    mVolumeCurves->initializeVolumeCurves(speakerDrcEnabled);

    // Once policy config has been parsed, retrieve an instance of the engine and initialize it.
    audio_policy::EngineInstance *engineInstance = audio_policy::EngineInstance::getInstance();
    if (!engineInstance) {
        ALOGE("%s:  Could not get an instance of policy engine", __FUNCTION__);
        return;
    }
    // Retrieve the Policy Manager Interface
    mEngine = engineInstance->queryInterface<AudioPolicyManagerInterface>();
    if (mEngine == NULL) {
        ALOGE("%s: Failed to get Policy Engine Interface", __FUNCTION__);
        return;
    }
    mEngine->setObserver(this);
    status_t status = mEngine->initCheck();
    (void) status;
    ALOG_ASSERT(status == NO_ERROR, "Policy engine not initialized(err=%d)", status);

    // mAvailableOutputDevices and mAvailableInputDevices now contain all attached devices
    // open all output streams needed to access attached devices
    audio_devices_t outputDeviceTypes = mAvailableOutputDevices.types();
    audio_devices_t inputDeviceTypes = mAvailableInputDevices.types() & ~AUDIO_DEVICE_BIT_IN;
    for (size_t i = 0; i < mHwModules.size(); i++) {
        mHwModules[i]->mHandle = mpClientInterface->loadHwModule(mHwModules[i]->getName());
        if (mHwModules[i]->mHandle == 0) {
            ALOGW("could not open HW module %s", mHwModules[i]->getName());
            continue;
        }
        // open all output streams needed to access attached devices
        // except for direct output streams that are only opened when they are actually
        // required by an app.
        // This also validates mAvailableOutputDevices list
        for (size_t j = 0; j < mHwModules[i]->mOutputProfiles.size(); j++)
        {
            const sp<IOProfile> outProfile = mHwModules[i]->mOutputProfiles[j];

            if (!outProfile->hasSupportedDevices()) {
                ALOGW("Output profile contains no device on module %s", mHwModules[i]->getName());
                continue;
            }
            if ((outProfile->getFlags() & AUDIO_OUTPUT_FLAG_TTS) != 0) {
                mTtsOutputAvailable = true;
            }

            if ((outProfile->getFlags() & AUDIO_OUTPUT_FLAG_DIRECT) != 0) {
                continue;
            }
            audio_devices_t profileType = outProfile->getSupportedDevicesType();
            if ((profileType & mDefaultOutputDevice->type()) != AUDIO_DEVICE_NONE) {
                profileType = mDefaultOutputDevice->type();
            } else {
                // chose first device present in profile's SupportedDevices also part of
                // outputDeviceTypes
                profileType = outProfile->getSupportedDeviceForType(outputDeviceTypes);
            }
            if ((profileType & outputDeviceTypes) == 0) {
                continue;
            }
            sp<SwAudioOutputDescriptor> outputDesc = new SwAudioOutputDescriptor(outProfile,
                                                                                 mpClientInterface);
            const DeviceVector &supportedDevices = outProfile->getSupportedDevices();
            const DeviceVector &devicesForType = supportedDevices.getDevicesFromType(profileType);
            String8 address = devicesForType.size() > 0 ? devicesForType.itemAt(0)->mAddress
                    : String8("");

            outputDesc->mDevice = profileType;
            audio_config_t config = AUDIO_CONFIG_INITIALIZER;
            config.sample_rate = outputDesc->mSamplingRate;
            config.channel_mask = outputDesc->mChannelMask;
            config.format = outputDesc->mFormat;
            audio_io_handle_t output = AUDIO_IO_HANDLE_NONE;
            status_t status = mpClientInterface->openOutput(outProfile->getModuleHandle(),
                                                            &output,
                                                            &config,
                                                            &outputDesc->mDevice,
                                                            address,
                                                            &outputDesc->mLatency,
                                                            outputDesc->mFlags);

            if (status != NO_ERROR) {
                ALOGW("Cannot open output stream for device %08x on hw module %s",
                      outputDesc->mDevice,
                      mHwModules[i]->getName());
            } else {
                outputDesc->mSamplingRate = config.sample_rate;
                outputDesc->mChannelMask = config.channel_mask;
                outputDesc->mFormat = config.format;

                for (size_t k = 0; k  < supportedDevices.size(); k++) {
                    ssize_t index = mAvailableOutputDevices.indexOf(supportedDevices[k]);
                    // give a valid ID to an attached device once confirmed it is reachable
                    if (index >= 0 && !mAvailableOutputDevices[index]->isAttached()) {
                        mAvailableOutputDevices[index]->attach(mHwModules[i]);
                    }
                }
                if (mPrimaryOutput == 0 &&
                        outProfile->getFlags() & AUDIO_OUTPUT_FLAG_PRIMARY) {
                    mPrimaryOutput = outputDesc;
                }
                addOutput(output, outputDesc);
                setOutputDevice(outputDesc,
                                outputDesc->mDevice,
                                true,
                                0,
                                NULL,
                                address.string());
            }
        }
        // open input streams needed to access attached devices to validate
        // mAvailableInputDevices list
        for (size_t j = 0; j < mHwModules[i]->mInputProfiles.size(); j++)
        {
            const sp<IOProfile> inProfile = mHwModules[i]->mInputProfiles[j];

            if (!inProfile->hasSupportedDevices()) {
                ALOGW("Input profile contains no device on module %s", mHwModules[i]->getName());
                continue;
            }
            // chose first device present in profile's SupportedDevices also part of
            // inputDeviceTypes
            audio_devices_t profileType = inProfile->getSupportedDeviceForType(inputDeviceTypes);

            if ((profileType & inputDeviceTypes) == 0) {
                continue;
            }
            sp<AudioInputDescriptor> inputDesc =
                    new AudioInputDescriptor(inProfile);

            inputDesc->mDevice = profileType;

            // find the address
            DeviceVector inputDevices = mAvailableInputDevices.getDevicesFromType(profileType);
            //   the inputs vector must be of size 1, but we don't want to crash here
            String8 address = inputDevices.size() > 0 ? inputDevices.itemAt(0)->mAddress
                    : String8("");
            ALOGV("  for input device 0x%x using address %s", profileType, address.string());
            ALOGE_IF(inputDevices.size() == 0, "Input device list is empty!");

            audio_config_t config = AUDIO_CONFIG_INITIALIZER;
            config.sample_rate = inputDesc->mSamplingRate;
            config.channel_mask = inputDesc->mChannelMask;
            config.format = inputDesc->mFormat;
            audio_io_handle_t input = AUDIO_IO_HANDLE_NONE;
            status_t status = mpClientInterface->openInput(inProfile->getModuleHandle(),
                                                           &input,
                                                           &config,
                                                           &inputDesc->mDevice,
                                                           address,
                                                           AUDIO_SOURCE_MIC,
                                                           AUDIO_INPUT_FLAG_NONE);

            if (status == NO_ERROR) {
                const DeviceVector &supportedDevices = inProfile->getSupportedDevices();
                for (size_t k = 0; k  < supportedDevices.size(); k++) {
                    ssize_t index =  mAvailableInputDevices.indexOf(supportedDevices[k]);
                    // give a valid ID to an attached device once confirmed it is reachable
                    if (index >= 0) {
                        sp<DeviceDescriptor> devDesc = mAvailableInputDevices[index];
                        if (!devDesc->isAttached()) {
                            devDesc->attach(mHwModules[i]);
                            devDesc->importAudioPort(inProfile);
                        }
                    }
                }
                mpClientInterface->closeInput(input);
            } else {
                ALOGW("Cannot open input stream for device %08x on hw module %s",
                      inputDesc->mDevice,
                      mHwModules[i]->getName());
            }
        }
    }
    // make sure all attached devices have been allocated a unique ID
    for (size_t i = 0; i  < mAvailableOutputDevices.size();) {
        if (!mAvailableOutputDevices[i]->isAttached()) {
            ALOGW("Output device %08x unreachable", mAvailableOutputDevices[i]->type());
            mAvailableOutputDevices.remove(mAvailableOutputDevices[i]);
            continue;
        }
        // The device is now validated and can be appended to the available devices of the engine
        mEngine->setDeviceConnectionState(mAvailableOutputDevices[i],
                                          AUDIO_POLICY_DEVICE_STATE_AVAILABLE);
        i++;
    }
    for (size_t i = 0; i  < mAvailableInputDevices.size();) {
        if (!mAvailableInputDevices[i]->isAttached()) {
            ALOGW("Input device %08x unreachable", mAvailableInputDevices[i]->type());
            mAvailableInputDevices.remove(mAvailableInputDevices[i]);
            continue;
        }
        // The device is now validated and can be appended to the available devices of the engine
        mEngine->setDeviceConnectionState(mAvailableInputDevices[i],
                                          AUDIO_POLICY_DEVICE_STATE_AVAILABLE);
        i++;
    }
    // make sure default device is reachable
    if (mDefaultOutputDevice == 0 || mAvailableOutputDevices.indexOf(mDefaultOutputDevice) < 0) {
        ALOGE("Default device %08x is unreachable", mDefaultOutputDevice->type());
    }

    ALOGE_IF((mPrimaryOutput == 0), "Failed to open primary output");

    updateDevicesAndOutputs();

#ifdef AUDIO_POLICY_TEST
    if (mPrimaryOutput != 0) {
        AudioParameter outputCmd = AudioParameter();
        outputCmd.addInt(String8("set_id"), 0);
        mpClientInterface->setParameters(mPrimaryOutput->mIoHandle, outputCmd.toString());

        mTestDevice = AUDIO_DEVICE_OUT_SPEAKER;
        mTestSamplingRate = 44100;
        mTestFormat = AUDIO_FORMAT_PCM_16_BIT;
        mTestChannels =  AUDIO_CHANNEL_OUT_STEREO;
        mTestLatencyMs = 0;
        mCurOutput = 0;
        mDirectOutput = false;
        for (int i = 0; i < NUM_TEST_OUTPUTS; i++) {
            mTestOutputs[i] = 0;
        }

        const size_t SIZE = 256;
        char buffer[SIZE];
        snprintf(buffer, SIZE, "AudioPolicyManagerTest");
        run(buffer, ANDROID_PRIORITY_AUDIO);
    }
#endif //AUDIO_POLICY_TEST
}

其中 mpClientInterface->openOutput() 最终会调用到 AudioFlinger::openOutput()：打开输出流设备，并创建 PlaybackThread 对象：

AudioFlinger OffloadThread

系统启动时，就已经打开 primary_out、low_latency、deep_buffer 这三种输出流设备，并创建对应的 MixerThread 了；而此时 DirectOutputThread 与 OffloadThread 不会被创建，直到标识为 AUDIO_OUTPUT_FLAG_DIRECT/AUDIO_OUTPUT_FLAG_COMPRESS_OFFLOAD 的音频流需要输出时，才开始创建 DirectOutputThread/OffloadThread 和打开 direct_out/compress_offload 设备。

这里不知大家有无疑问：为什么 DirectOutputThread 与 OffloadThread 会被单独对待？DirectOuputThread 使用率较低，尚可以理解，但 OffloadThread 的使用率还是很高的，为什么不让 OffloadThread/compress_offload 设备也进入待命状态呢？

要回答这个问题：我们首先得明白 compress_offload 设备是什么东东，与其他输出流设备有什么不同。先看个图：

红色的是 Offload 音频流，它与其他音频流有什么本质的不同？Offload 音频流是未经 NuPlayerDecoder 进行解码的（NuPlayerDecoder 设置了 Passthrough 模式），所以必须把这些音频流数据送到 DSP，让 DSP 对其解码，解码后的 PCM 数据再送到 Codec 输出。

compress_offload 设备，说白了，就是驱动 DSP 解码数据、Codec 输出声音。 而 DSP 要解码数据，首先得知道数据的编码信息，如编码器 codec_id、采样率 sample_rate、声道数 channel_mask、比特率 bit_rate 等信息，由于 DSP 并没有实现 DataSource/Parser 部件，不能自己解析数据的编码信息，所以得有“人”告诉它，这个“人”无疑是 compress_offload 设备。

AudioTrack 构造函数有个 offloadInfo 的参数，参数原型定义如下：

NuPlayer DataSource/Parser 解析 mp3、flac 等文件得到数据编码信息，并在构造 AudioTrack 实例时作为参数传入，AudioFlinger 将基于这些编码信息打开 compress_offload 设备。

到这里，大家明白了吗？每个 mp3/flac 文件的编码信息可能是不一样的，比如 a.mp3 文件的编码信息是 mp3&44.1KHZ&16bit… ，而 b.flac 文件的编码信息是 flac&48KHz&24bit…； 播放 a.mp3 时，AudioFlinger 打开一个配置为 mp3&44.1KHz&16bit… 的 compress_offload 设备，接着播放 b.flac，就需要关闭之前的 compress_offload 设备，重新打开一个配置为 flac&48KHz&24bit… 的 compress_offload 设备。所以系统不会提前打开 compress_offload 设备，只有等到播放 mp3、flac 时取到明确的数据编码信息，才基于这些编码信息打开 compress_offload 设备。

编码信息包含很多条目，切换音源时，是否编码信息有一点点不一样，都需要重新打开 compress_offload 设备呢？不能运行时更新信息到 DSP 吗？其实 stagefright 和 compress_offload 是支持运行期更新某些信息的，也就是无缝切换，至于是哪些信息，依赖于 DSP 算法实现；有兴趣深入的可以参考 sendMetaDataToHal() 和 compress_set_gapless_metadata()。