Android輸入事件從讀取到分發三：InputDispatcherThread線程分發事件的過程

分析完事件的讀取後，東忙西忙，不知不覺已過去了快五個月了…也不是說沒有時間把這部分源碼分析完，而是實在是分析不下去，所以轉戰到其餘地方去了。然而這一塊始終是心頭那不捨的惦記，通過一段時間的沉澱，又參考了網上不少大神的文章，而後再來閱讀源碼，漸漸感受到能看出點眉目了。於是事不宜遲，趕忙作個記錄吧（注：分析使用的Android源碼版本爲6.0）。 
前面兩篇文章分析輸入事件的讀取，經過分析，發現時間的讀取是在EventHub類中實現的，EventHub類的getEevent方法中使用epoll_wait來監聽驅動程序上報的事件，而在InputReaderThread的threadLoop方法中調用InputReader的loopOnce方法來不斷的調用EventHub的getEvent方法來一直監聽事件的到來，getEvent是一個阻塞的方法，當沒有事件的時候，epoll_wait方法就會是線程休眠，有事件了就會喚醒線程。

從事件讀取線程到事件分發線程的轉移過程
既然有個InputReaderThread線程，那麼有個InputDispatcherThread就不奇怪了，一個用來讀取，一個用來分發，分工合做，相互配合，這樣才能高效的完成事件的讀取與分發。 
咱們先梳理下思路：

從事件讀取線程到事件分發線程的轉移時序圖： 

從時序圖中咱們能夠看到兩個線程交互的過程。InputReaderThread會喚醒InputDispatcherThread來分發事件。 
接下來，咱們經過源碼的追蹤的方式，具體看看時序圖中涉及的代碼的具體實現。

跟着時序圖，咱們代碼從InputReaderThread開始看起：

InputReaderThread::InputReaderThread(const sp<InputReaderInterface>& reader) :
        Thread(/*canCallJava*/ true), mReader(reader) {
}

InputReaderThread::~InputReaderThread() {
}

bool InputReaderThread::threadLoop() {
    mReader->loopOnce();
    return true;
}

構造函數和析構函數都是空的，循環函數簡單的就一句。mReader就是當前的InputReader對象，所以loopOnce以下：

void InputReader::loopOnce() {
    int32_t oldGeneration;
    int32_t timeoutMillis;
    bool inputDevicesChanged = false;
    Vector<InputDeviceInfo> inputDevices;
    { // acquire lock
        AutoMutex _l(mLock);

        oldGeneration = mGeneration;
        timeoutMillis = -1;

       ...
    //獲取事件
    size_t count = mEventHub->getEvents(timeoutMillis, mEventBuffer, EVENT_BUFFER_SIZE);

    { // acquire lock
        AutoMutex _l(mLock);
        mReaderIsAliveCondition.broadcast();

        if (count) {
            //處理事件
            processEventsLocked(mEventBuffer, count);
        }
        ...
        if (oldGeneration != mGeneration) {
            inputDevicesChanged = true;
            getInputDevicesLocked(inputDevices);
        }
    } // release lock

    // Send out a message that the describes the changed input devices.
    if (inputDevicesChanged) {
        mPolicy->notifyInputDevicesChanged(inputDevices);
    }
    ...
    mQueuedListener->flush();
}

只列出了比較重要的一些代碼。這個函數做以下事情： 
1.使用mEventHub->getEvents獲取輸入事件 
2.使用processEventsLocked處理事件 
getEvents方法以前也已經有介紹過，它就是使用epoll來監聽驅動程序上報的事件的。因此咱們從processEventsLocked方法看起：

void InputReader::processEventsLocked(const RawEvent* rawEvents, size_t count) {
    for (const RawEvent* rawEvent = rawEvents; count;) {
        int32_t type = rawEvent->type;
        size_t batchSize = 1;
        if (type < EventHubInterface::FIRST_SYNTHETIC_EVENT) {
            int32_t deviceId = rawEvent->deviceId;
            while (batchSize < count) {
                if (rawEvent[batchSize].type >= EventHubInterface::FIRST_SYNTHETIC_EVENT
                        || rawEvent[batchSize].deviceId != deviceId) {
                    break;
                }
                batchSize += 1;
            }
#if DEBUG_RAW_EVENTS
            ALOGD("BatchSize: %d Count: %d", batchSize, count);
#endif
            processEventsForDeviceLocked(deviceId, rawEvent, batchSize);
        } else {
            switch (rawEvent->type) {
            case EventHubInterface::DEVICE_ADDED:
                addDeviceLocked(rawEvent->when, rawEvent->deviceId);
                break;
            case EventHubInterface::DEVICE_REMOVED:
                removeDeviceLocked(rawEvent->when, rawEvent->deviceId);
                break;
            case EventHubInterface::FINISHED_DEVICE_SCAN:
                handleConfigurationChangedLocked(rawEvent->when);
                break;
            default:
                ALOG_ASSERT(false); // can't happen
                break;
            }
        }
        count -= batchSize;
        rawEvent += batchSize;
    }
}

這個方法會對事件類型進行判斷。咱們假定用戶按了一下返回按鍵，因此輸入的是一個按鍵事件。而這個函數判斷的時間類型的定義以下：

    enum {
        // Sent when a device is added.
        DEVICE_ADDED = 0x10000000,
        // Sent when a device is removed.
        DEVICE_REMOVED = 0x20000000,
        // Sent when all added/removed devices from the most recent scan have been reported.
        // This event is always sent at least once.
        FINISHED_DEVICE_SCAN = 0x30000000,

        FIRST_SYNTHETIC_EVENT = DEVICE_ADDED,
    };

調用device->process因一部處理，device是InputDevice的實例，所以看看InputDevice下的process方法：

void InputDevice::process(const RawEvent* rawEvents, size_t count) {
    // Process all of the events in order for each mapper.
    // We cannot simply ask each mapper to process them in bulk because mappers may
    // have side-effects that must be interleaved.  For example, joystick movement events and
    // gamepad button presses are handled by different mappers but they should be dispatched
    // in the order received.
    size_t numMappers = mMappers.size();
    for (const RawEvent* rawEvent = rawEvents; count--; rawEvent++) {
#if DEBUG_RAW_EVENTS
        ALOGD("Input event: device=%d type=0x%04x code=0x%04x value=0x%08x when=%lld",
                rawEvent->deviceId, rawEvent->type, rawEvent->code, rawEvent->value,
                rawEvent->when);
#endif

        if (mDropUntilNextSync) {
            if (rawEvent->type == EV_SYN && rawEvent->code == SYN_REPORT) {
                mDropUntilNextSync = false;
#if DEBUG_RAW_EVENTS
                ALOGD("Recovered from input event buffer overrun.");
#endif
            } else {
#if DEBUG_RAW_EVENTS
                ALOGD("Dropped input event while waiting for next input sync.");
#endif
            }
        } else if (rawEvent->type == EV_SYN && rawEvent->code == SYN_DROPPED) {
            ALOGI("Detected input event buffer overrun for device %s.", getName().string());
            mDropUntilNextSync = true;
            reset(rawEvent->when);
        } else {
            for (size_t i = 0; i < numMappers; i++) {
                InputMapper* mapper = mMappers[i];
                mapper->process(rawEvent);
            }
        }
    }
}

爲了更好的理解這段代碼，咱們列出事件類型的定義：

/* Events */

#define EV_SYN          0x00
#define EV_KEY          0x01
#define EV_REL          0x02
#define EV_ABS          0x03
#define EV_MSC          0x04
#define EV_LED          0x11
#define EV_SND          0x12
#define EV_REP          0x14
#define EV_FF           0x15
#define EV_PWR          0x16
#define EV_FF_STATUS        0x17
#define EV_MAX          0x1f

所以咱們的按鍵時間應該是0x01,因此這段代碼執行的下面這個for循環：

            for (size_t i = 0; i < numMappers; i++) {
                InputMapper* mapper = mMappers[i];
                mapper->process(rawEvent);
            }

for循環遍歷mMappers數組，分別調用每個mapper的process方法。這個數組包含了全部的輸入事件的類型。咱們按下的按鍵，所以會調用按鍵類型的Mapper來處理。若是想搞清楚這一塊，能夠看下mMappers數組的構造過程，在createDeviceLocked方法中實現，

InputDevice* InputReader::createDeviceLocked(int32_t deviceId, int32_t controllerNumber,
        const InputDeviceIdentifier& identifier, uint32_t classes) {
    InputDevice* device = new InputDevice(&mContext, deviceId, bumpGenerationLocked(),
            controllerNumber, identifier, classes);

    // External devices.
    if (classes & INPUT_DEVICE_CLASS_EXTERNAL) {
        device->setExternal(true);
    }

    // Devices with mics.
    if (classes & INPUT_DEVICE_CLASS_MIC) {
        device->setMic(true);
    }

    // Switch-like devices.
    if (classes & INPUT_DEVICE_CLASS_SWITCH) {
        device->addMapper(new SwitchInputMapper(device));
    }

    // Vibrator-like devices.
    if (classes & INPUT_DEVICE_CLASS_VIBRATOR) {
        device->addMapper(new VibratorInputMapper(device));
    }

    // Keyboard-like devices.
    uint32_t keyboardSource = 0;
    int32_t keyboardType = AINPUT_KEYBOARD_TYPE_NON_ALPHABETIC;
    if (classes & INPUT_DEVICE_CLASS_KEYBOARD) {
        keyboardSource |= AINPUT_SOURCE_KEYBOARD;
    }
    if (classes & INPUT_DEVICE_CLASS_ALPHAKEY) {
        keyboardType = AINPUT_KEYBOARD_TYPE_ALPHABETIC;
    }
    if (classes & INPUT_DEVICE_CLASS_DPAD) {
        keyboardSource |= AINPUT_SOURCE_DPAD;
    }
    if (classes & INPUT_DEVICE_CLASS_GAMEPAD) {
        keyboardSource |= AINPUT_SOURCE_GAMEPAD;
    }

    if (keyboardSource != 0) {
        device->addMapper(new KeyboardInputMapper(device, keyboardSource, keyboardType));
    }

    // Cursor-like devices.
    if (classes & INPUT_DEVICE_CLASS_CURSOR) {
        device->addMapper(new CursorInputMapper(device));
    }

    // Touchscreens and touchpad devices.
    if (classes & INPUT_DEVICE_CLASS_TOUCH_MT) {
        device->addMapper(new MultiTouchInputMapper(device));
    } else if (classes & INPUT_DEVICE_CLASS_TOUCH) {
        device->addMapper(new SingleTouchInputMapper(device));
    }

    // Joystick-like devices.
    if (classes & INPUT_DEVICE_CLASS_JOYSTICK) {
        device->addMapper(new JoystickInputMapper(device));
    }

    // External stylus-like devices.
    if (classes & INPUT_DEVICE_CLASS_EXTERNAL_STYLUS) {
        device->addMapper(new ExternalStylusInputMapper(device));
    }

    return device;
}

由於咱們事件是按鍵事件，因此，其餘的mapper的process什麼都不作，最終是KeyboardInputMapper的process方法。

void KeyboardInputMapper::process(const RawEvent* rawEvent) {
    switch (rawEvent->type) {
    case EV_KEY: {
        int32_t scanCode = rawEvent->code;
        int32_t usageCode = mCurrentHidUsage;
        mCurrentHidUsage = 0;

        if (isKeyboardOrGamepadKey(scanCode)) {
            int32_t keyCode;
            uint32_t flags;
            if (getEventHub()->mapKey(getDeviceId(), scanCode, usageCode, &keyCode, &flags)) {
                keyCode = AKEYCODE_UNKNOWN;
                flags = 0;
            }
            processKey(rawEvent->when, rawEvent->value != 0, keyCode, scanCode, flags);
        }
        break;
    }
    case EV_MSC: {
        if (rawEvent->code == MSC_SCAN) {
            mCurrentHidUsage = rawEvent->value;
        }
        break;
    }
    case EV_SYN: {
        if (rawEvent->code == SYN_REPORT) {
            mCurrentHidUsage = 0;
        }
    }
    }
}

傳入的事件類型爲按鍵事件，所以會調用processKey方法進一步處理：

void KeyboardInputMapper::processKey(nsecs_t when, bool down, int32_t keyCode,
        int32_t scanCode, uint32_t policyFlags) {

    if (down) {
        // Rotate key codes according to orientation if needed.
        if (mParameters.orientationAware && mParameters.hasAssociatedDisplay) {
            keyCode = rotateKeyCode(keyCode, mOrientation);
        }

        // Add key down.
        ssize_t keyDownIndex = findKeyDown(scanCode);
        if (keyDownIndex >= 0) {
            // key repeat, be sure to use same keycode as before in case of rotation
            keyCode = mKeyDowns.itemAt(keyDownIndex).keyCode;
        } else {
            // key down
            if ((policyFlags & POLICY_FLAG_VIRTUAL)
                    && mContext->shouldDropVirtualKey(when,
                            getDevice(), keyCode, scanCode)) {
                return;
            }
            if (policyFlags & POLICY_FLAG_GESTURE) {
                mDevice->cancelTouch(when);
            }

            mKeyDowns.push();
            KeyDown& keyDown = mKeyDowns.editTop();
            keyDown.keyCode = keyCode;
            keyDown.scanCode = scanCode;
        }

        mDownTime = when;
    } else {
        // Remove key down.
        ssize_t keyDownIndex = findKeyDown(scanCode);
        if (keyDownIndex >= 0) {
            // key up, be sure to use same keycode as before in case of rotation
            keyCode = mKeyDowns.itemAt(keyDownIndex).keyCode;
            mKeyDowns.removeAt(size_t(keyDownIndex));
        } else {
            // key was not actually down
            ALOGI("Dropping key up from device %s because the key was not down.  "
                    "keyCode=%d, scanCode=%d",
                    getDeviceName().string(), keyCode, scanCode);
            return;
        }
    }

    int32_t oldMetaState = mMetaState;
    int32_t newMetaState = updateMetaState(keyCode, down, oldMetaState);
    bool metaStateChanged = oldMetaState != newMetaState;
    if (metaStateChanged) {
        mMetaState = newMetaState;
        updateLedState(false);
    }

    nsecs_t downTime = mDownTime;

    // Key down on external an keyboard should wake the device.
    // We don't do this for internal keyboards to prevent them from waking up in your pocket.
    // For internal keyboards, the key layout file should specify the policy flags for
    // each wake key individually.
    // TODO: Use the input device configuration to control this behavior more finely.
    if (down && getDevice()->isExternal()) {
// MStar Android Patch Begin
#ifdef ENABLE_STR
        char value[PROPERTY_VALUE_MAX];
        property_get("mstar.str.suspending", value, "0");
        if (atoi(value) == 0) {
            policyFlags |= POLICY_FLAG_WAKE;
        }
#else
        policyFlags |= POLICY_FLAG_WAKE;
#endif
// MStar Android Patch End
    }

    if (mParameters.handlesKeyRepeat) {
        policyFlags |= POLICY_FLAG_DISABLE_KEY_REPEAT;
    }

    if (metaStateChanged) {
        getContext()->updateGlobalMetaState();
    }

    if (down && !isMetaKey(keyCode)) {
        getContext()->fadePointer();
    }

    NotifyKeyArgs args(when, getDeviceId(), mSource, policyFlags,
            down ? AKEY_EVENT_ACTION_DOWN : AKEY_EVENT_ACTION_UP,
            AKEY_EVENT_FLAG_FROM_SYSTEM, keyCode, scanCode, newMetaState, downTime);
    getListener()->notifyKey(&args);
}

因爲能力有限，我沒法關注太多的細節，這個方法的最後調用了 getListener()->notifyKey(&args);其實就是調用了InputDispatcher的notifyKey方法。爲何這麼說呢？這個方法展開看一下就明白了。

InputListenerInterface* InputReader::ContextImpl::getListener() {
    return mReader->mQueuedListener.get();
}

InpurReader的構造函數中建立了mQueuedListener變量。

InputReader::InputReader(const sp<EventHubInterface>& eventHub,
        const sp<InputReaderPolicyInterface>& policy,
        const sp<InputListenerInterface>& listener) :
        mContext(this), mEventHub(eventHub), mPolicy(policy),
        mGlobalMetaState(0), mGeneration(1),
        mDisableVirtualKeysTimeout(LLONG_MIN), mNextTimeout(LLONG_MAX),
        mConfigurationChangesToRefresh(0) {
    mQueuedListener = new QueuedInputListener(listener);

    { // acquire lock
        AutoMutex _l(mLock);

        refreshConfigurationLocked(0);
        updateGlobalMetaStateLocked();
    } // release lock
}

QueuedInputListener構造函數中傳入了一個listener,這個listener是否是InputDispatcher呢？ 
咱們以前就分析過，InputReader是在InputManager中建立的，看看InputManager得構造函數：

InputManager::InputManager(
        const sp<EventHubInterface>& eventHub,
        const sp<InputReaderPolicyInterface>& readerPolicy,
        const sp<InputDispatcherPolicyInterface>& dispatcherPolicy) {
    mDispatcher = new InputDispatcher(dispatcherPolicy);
    mReader = new InputReader(eventHub, readerPolicy, mDispatcher);
    initialize();
}

真相大白了，listener就是mDispatcher。mDispatcher是InputDispatcher的實例。 
所以，下一步就進入到了InputDispatcher的notifyKey方法了：

void InputDispatcher::notifyKey(const NotifyKeyArgs* args) {
#if DEBUG_INBOUND_EVENT_DETAILS
    ALOGD("notifyKey - eventTime=%lld, deviceId=%d, source=0x%x, policyFlags=0x%x, action=0x%x, "
            "flags=0x%x, keyCode=0x%x, scanCode=0x%x, metaState=0x%x, downTime=%lld",
            args->eventTime, args->deviceId, args->source, args->policyFlags,
            args->action, args->flags, args->keyCode, args->scanCode,
            args->metaState, args->downTime);
#endif
    if (!validateKeyEvent(args->action)) {
        return;
    }

    uint32_t policyFlags = args->policyFlags;
    int32_t flags = args->flags;
    int32_t metaState = args->metaState;
    if ((policyFlags & POLICY_FLAG_VIRTUAL) || (flags & AKEY_EVENT_FLAG_VIRTUAL_HARD_KEY)) {
        policyFlags |= POLICY_FLAG_VIRTUAL;
        flags |= AKEY_EVENT_FLAG_VIRTUAL_HARD_KEY;
    }
    if (policyFlags & POLICY_FLAG_FUNCTION) {
        metaState |= AMETA_FUNCTION_ON;
    }

    policyFlags |= POLICY_FLAG_TRUSTED;

    int32_t keyCode = args->keyCode;
    if (metaState & AMETA_META_ON && args->action == AKEY_EVENT_ACTION_DOWN) {
        int32_t newKeyCode = AKEYCODE_UNKNOWN;
        if (keyCode == AKEYCODE_DEL) {
            newKeyCode = AKEYCODE_BACK;
        } else if (keyCode == AKEYCODE_ENTER) {
            newKeyCode = AKEYCODE_HOME;
        }
        if (newKeyCode != AKEYCODE_UNKNOWN) {
            AutoMutex _l(mLock);
            struct KeyReplacement replacement = {keyCode, args->deviceId};
            mReplacedKeys.add(replacement, newKeyCode);
            keyCode = newKeyCode;
            metaState &= ~AMETA_META_ON;
        }
    } else if (args->action == AKEY_EVENT_ACTION_UP) {
        // In order to maintain a consistent stream of up and down events, check to see if the key
        // going up is one we've replaced in a down event and haven't yet replaced in an up event,
        // even if the modifier was released between the down and the up events.
        AutoMutex _l(mLock);
        struct KeyReplacement replacement = {keyCode, args->deviceId};
        ssize_t index = mReplacedKeys.indexOfKey(replacement);
        if (index >= 0) {
            keyCode = mReplacedKeys.valueAt(index);
            mReplacedKeys.removeItemsAt(index);
            metaState &= ~AMETA_META_ON;
        }
    }

    KeyEvent event;
    event.initialize(args->deviceId, args->source, args->action,
            flags, keyCode, args->scanCode, metaState, 0,
            args->downTime, args->eventTime);

    mPolicy->interceptKeyBeforeQueueing(&event, /*byref*/ policyFlags);

    bool needWake;
    { // acquire lock
        mLock.lock();

        if (shouldSendKeyToInputFilterLocked(args)) {
            mLock.unlock();

            policyFlags |= POLICY_FLAG_FILTERED;
            if (!mPolicy->filterInputEvent(&event, policyFlags)) {
                return; // event was consumed by the filter
            }

            mLock.lock();
        }

        int32_t repeatCount = 0;
        KeyEntry* newEntry = new KeyEntry(args->eventTime,
                args->deviceId, args->source, policyFlags,
                args->action, flags, keyCode, args->scanCode,
                metaState, repeatCount, args->downTime);

        needWake = enqueueInboundEventLocked(newEntry);
        mLock.unlock();
    } // release lock

    if (needWake) {
        mLooper->wake();
    }
}

這份函數作了三件事情： 
1.把時間封裝成了KeyEvent類型的對象。 
2.攔截事件。有些事件是必需要攔截的，這類事件不須要發送到window中去，好比關機鍵，須要系統來處理。攔截的主要代碼：

 mPolicy->interceptKeyBeforeQueueing(&event, /*byref*/ policyFlags);
1
這部分也很重要，可是咱們暫不展開，先知道這裏攔截了事件，由於這裏的攔截最終會調用PhoneWindowManager的interceptKeyBeforeQueueing等方法，解開了我PhoneWindowManager中相關方法何時調用的不清楚的迷惑。咱們先放一放，以後在展開。 
3.喚醒InputDispatcherThread線程：

 mLooper->wake();
1
這個線程喚醒意味着事件的分發終於轉入到InputDispatcherThread了，InputReaderThread線程完成了它的使命。爲何這行代碼能喚醒InputDispatcherThread呢？這還的從InputDispatcherThread提及：

InputDispatcherThread::InputDispatcherThread(const sp<InputDispatcherInterface>& dispatcher) :
        Thread(/*canCallJava*/ true), mDispatcher(dispatcher) {
}

InputDispatcherThread::~InputDispatcherThread() {
}

bool InputDispatcherThread::threadLoop() {
    mDispatcher->dispatchOnce();
    return true;
}

這個線程和InputReaderThread線程相似，調用InputDispatcher的dispatchOnce作一次事件分發：

void InputDispatcher::dispatchOnce() {
    nsecs_t nextWakeupTime = LONG_LONG_MAX;
    { // acquire lock
        AutoMutex _l(mLock);
        mDispatcherIsAliveCondition.broadcast();

        // Run a dispatch loop if there are no pending commands.
        // The dispatch loop might enqueue commands to run afterwards.
        if (!haveCommandsLocked()) {
            dispatchOnceInnerLocked(&nextWakeupTime);
        }

        // Run all pending commands if there are any.
        // If any commands were run then force the next poll to wake up immediately.
        if (runCommandsLockedInterruptible()) {
            nextWakeupTime = LONG_LONG_MIN;
        }
    } // release lock

    // Wait for callback or timeout or wake.  (make sure we round up, not down)
    nsecs_t currentTime = now();
    int timeoutMillis = toMillisecondTimeoutDelay(currentTime, nextWakeupTime);
    mLooper->pollOnce(timeoutMillis);
}

能夠看到它內部使用Looper,調用mLooper->pollOnce(timeoutMillis)後，若是沒有數據能夠讀取，會致使線程休眠，調用mLooper->wake能夠喚醒它，喚醒後dispatchOnceInnerLocked來分發事件。至此，咱們繪製的事件讀取線程到事件分發線程的轉移時序圖就走到了盡頭。下一部分究竟是怎麼把分發到window的呢？因而咱們進入到了事件發送的第二個階段。

InputDispatcherThread中的事件發送
咱們先看一下這個階段的時序圖： 

 
事件的發送和接收必定是你情我願的事情，強求不來的。可是，咱們目前已知在分析事件的發送，沒有關注事件的接收端，所以，真正的發送流程中有不少東西理解不來，不過不要緊，咱們先理清發送的流程，等咱們理解了事件的接收端之後，再反過頭來看這部分就豁然開朗了。所以，這以後的代碼咱們主要是追蹤事件發送的流程，而不苛求必定要理解透徹這部分代碼。 
咱們從上一個時序圖結束的位置開始追蹤吧，dispatchOnceInnerLocked函數定義以下：

void InputDispatcher::dispatchOnceInnerLocked(nsecs_t* nextWakeupTime) {
    nsecs_t currentTime = now();

    // Reset the key repeat timer whenever normal dispatch is suspended while the
    // device is in a non-interactive state.  This is to ensure that we abort a key
    // repeat if the device is just coming out of sleep.
    if (!mDispatchEnabled) {
        resetKeyRepeatLocked();
    }

    // If dispatching is frozen, do not process timeouts or try to deliver any new events.
    if (mDispatchFrozen) {
#if DEBUG_FOCUS
        ALOGD("Dispatch frozen.  Waiting some more.");
#endif
        return;
    }

    // Optimize latency of app switches.
    // Essentially we start a short timeout when an app switch key (HOME / ENDCALL) has
    // been pressed.  When it expires, we preempt dispatch and drop all other pending events.
    bool isAppSwitchDue = mAppSwitchDueTime <= currentTime;
    if (mAppSwitchDueTime < *nextWakeupTime) {
        *nextWakeupTime = mAppSwitchDueTime;
    }

    // Ready to start a new event.
    // If we don't already have a pending event, go grab one.
    if (! mPendingEvent) {
        if (mInboundQueue.isEmpty()) {
            if (isAppSwitchDue) {
                // The inbound queue is empty so the app switch key we were waiting
                // for will never arrive.  Stop waiting for it.
                resetPendingAppSwitchLocked(false);
                isAppSwitchDue = false;
            }

            // Synthesize a key repeat if appropriate.
            if (mKeyRepeatState.lastKeyEntry) {
                if (currentTime >= mKeyRepeatState.nextRepeatTime) {
                    mPendingEvent = synthesizeKeyRepeatLocked(currentTime);
                } else {
                    if (mKeyRepeatState.nextRepeatTime < *nextWakeupTime) {
                        *nextWakeupTime = mKeyRepeatState.nextRepeatTime;
                    }
                }
            }

            // Nothing to do if there is no pending event.
            if (!mPendingEvent) {
                return;
            }
        } else {
            // Inbound queue has at least one entry.
            mPendingEvent = mInboundQueue.dequeueAtHead();
            traceInboundQueueLengthLocked();
        }

        // Poke user activity for this event.
        if (mPendingEvent->policyFlags & POLICY_FLAG_PASS_TO_USER) {
            pokeUserActivityLocked(mPendingEvent);
        }

        // Get ready to dispatch the event.
        resetANRTimeoutsLocked();
    }

    // Now we have an event to dispatch.
    // All events are eventually dequeued and processed this way, even if we intend to drop them.
    ALOG_ASSERT(mPendingEvent != NULL);
    bool done = false;
    DropReason dropReason = DROP_REASON_NOT_DROPPED;
    if (!(mPendingEvent->policyFlags & POLICY_FLAG_PASS_TO_USER)) {
        dropReason = DROP_REASON_POLICY;
    } else if (!mDispatchEnabled) {
        dropReason = DROP_REASON_DISABLED;
    }

    if (mNextUnblockedEvent == mPendingEvent) {
        mNextUnblockedEvent = NULL;
    }

    switch (mPendingEvent->type) {
    case EventEntry::TYPE_CONFIGURATION_CHANGED: {
        ConfigurationChangedEntry* typedEntry =
                static_cast<ConfigurationChangedEntry*>(mPendingEvent);
        done = dispatchConfigurationChangedLocked(currentTime, typedEntry);
        dropReason = DROP_REASON_NOT_DROPPED; // configuration changes are never dropped
        break;
    }

    case EventEntry::TYPE_DEVICE_RESET: {
        DeviceResetEntry* typedEntry =
                static_cast<DeviceResetEntry*>(mPendingEvent);
        done = dispatchDeviceResetLocked(currentTime, typedEntry);
        dropReason = DROP_REASON_NOT_DROPPED; // device resets are never dropped
        break;
    }

    case EventEntry::TYPE_KEY: {
        KeyEntry* typedEntry = static_cast<KeyEntry*>(mPendingEvent);
        if (isAppSwitchDue) {
            if (isAppSwitchKeyEventLocked(typedEntry)) {
                resetPendingAppSwitchLocked(true);
                isAppSwitchDue = false;
            } else if (dropReason == DROP_REASON_NOT_DROPPED) {
                dropReason = DROP_REASON_APP_SWITCH;
            }
        }
        if (dropReason == DROP_REASON_NOT_DROPPED
                && isStaleEventLocked(currentTime, typedEntry)) {
            dropReason = DROP_REASON_STALE;
        }
        if (dropReason == DROP_REASON_NOT_DROPPED && mNextUnblockedEvent) {
            dropReason = DROP_REASON_BLOCKED;
        }
        done = dispatchKeyLocked(currentTime, typedEntry, &dropReason, nextWakeupTime);
        break;
    }

    case EventEntry::TYPE_MOTION: {
        MotionEntry* typedEntry = static_cast<MotionEntry*>(mPendingEvent);
        if (dropReason == DROP_REASON_NOT_DROPPED && isAppSwitchDue) {
            dropReason = DROP_REASON_APP_SWITCH;
        }
        if (dropReason == DROP_REASON_NOT_DROPPED
                && isStaleEventLocked(currentTime, typedEntry)) {
            dropReason = DROP_REASON_STALE;
        }
        if (dropReason == DROP_REASON_NOT_DROPPED && mNextUnblockedEvent) {
            dropReason = DROP_REASON_BLOCKED;
        }
        done = dispatchMotionLocked(currentTime, typedEntry,
                &dropReason, nextWakeupTime);
        break;
    }

    default:
        ALOG_ASSERT(false);
        break;
    }

    if (done) {
        if (dropReason != DROP_REASON_NOT_DROPPED) {
            dropInboundEventLocked(mPendingEvent, dropReason);
        }
        mLastDropReason = dropReason;

        releasePendingEventLocked();
        *nextWakeupTime = LONG_LONG_MIN;  // force next poll to wake up immediately
    }
}

咱們的時間類型爲TYPE_KEY，因此會調用到dispatchKeyLocked方法：

bool InputDispatcher::dispatchKeyLocked(nsecs_t currentTime, KeyEntry* entry,
        DropReason* dropReason, nsecs_t* nextWakeupTime) {
    // Preprocessing.
...

    // Dispatch the key.
    dispatchEventLocked(currentTime, entry, inputTargets);
    return true;
}

省略了不少類容。個人目的是理清框架，因此想理解細節的就請繞道吧。再看看dispatchEventLocked方法的實現：

void InputDispatcher::dispatchEventLocked(nsecs_t currentTime,
        EventEntry* eventEntry, const Vector<InputTarget>& inputTargets) {
#if DEBUG_DISPATCH_CYCLE
    ALOGD("dispatchEventToCurrentInputTargets");
#endif

    ALOG_ASSERT(eventEntry->dispatchInProgress); // should already have been set to true

    pokeUserActivityLocked(eventEntry);

    for (size_t i = 0; i < inputTargets.size(); i++) {
        const InputTarget& inputTarget = inputTargets.itemAt(i);

        ssize_t connectionIndex = getConnectionIndexLocked(inputTarget.inputChannel);
        if (connectionIndex >= 0) {
            sp<Connection> connection = mConnectionsByFd.valueAt(connectionIndex);
            prepareDispatchCycleLocked(currentTime, connection, eventEntry, &inputTarget);
        } else {
#if DEBUG_FOCUS
            ALOGD("Dropping event delivery to target with channel '%s' because it "
                    "is no longer registered with the input dispatcher.",
                    inputTarget.inputChannel->getName().string());
#endif
        }
    }
}

調用prepareDispatchCycleLocked方法進一步處理：

void InputDispatcher::prepareDispatchCycleLocked(nsecs_t currentTime,
        const sp<Connection>& connection, EventEntry* eventEntry, const InputTarget* inputTarget) {
    ...
    // Skip this event if the connection status is not normal.
    // We don't want to enqueue additional outbound events if the connection is broken.
    if (connection->status != Connection::STATUS_NORMAL) {
        return;
    }

    // Split a motion event if needed.
    if (inputTarget->flags & InputTarget::FLAG_SPLIT) {
        ALOG_ASSERT(eventEntry->type == EventEntry::TYPE_MOTION);

        MotionEntry* originalMotionEntry = static_cast<MotionEntry*>(eventEntry);
        if (inputTarget->pointerIds.count() != originalMotionEntry->pointerCount) {
            MotionEntry* splitMotionEntry = splitMotionEvent(
                    originalMotionEntry, inputTarget->pointerIds);
            if (!splitMotionEntry) {
                return; // split event was dropped
            }
#if DEBUG_FOCUS
            ALOGD("channel '%s' ~ Split motion event.",
                    connection->getInputChannelName());
            logOutboundMotionDetailsLocked("  ", splitMotionEntry);
#endif
            enqueueDispatchEntriesLocked(currentTime, connection,
                    splitMotionEntry, inputTarget);
            splitMotionEntry->release();
            return;
        }
    }

這個函數檢查connect狀態是否是正常，什麼鏈接狀態？ 
又調用到enqueueDispatchEntriesLocked方法了：

void InputDispatcher::enqueueDispatchEntriesLocked(nsecs_t currentTime,
        const sp<Connection>& connection, EventEntry* eventEntry, const InputTarget* inputTarget) {
    bool wasEmpty = connection->outboundQueue.isEmpty();

    // Enqueue dispatch entries for the requested modes.
    enqueueDispatchEntryLocked(connection, eventEntry, inputTarget,
            InputTarget::FLAG_DISPATCH_AS_HOVER_EXIT);
    enqueueDispatchEntryLocked(connection, eventEntry, inputTarget,
            InputTarget::FLAG_DISPATCH_AS_OUTSIDE);
    enqueueDispatchEntryLocked(connection, eventEntry, inputTarget,
            InputTarget::FLAG_DISPATCH_AS_HOVER_ENTER);
    enqueueDispatchEntryLocked(connection, eventEntry, inputTarget,
            InputTarget::FLAG_DISPATCH_AS_IS);
    enqueueDispatchEntryLocked(connection, eventEntry, inputTarget,
            InputTarget::FLAG_DISPATCH_AS_SLIPPERY_EXIT);
    enqueueDispatchEntryLocked(connection, eventEntry, inputTarget,
            InputTarget::FLAG_DISPATCH_AS_SLIPPERY_ENTER);

    // If the outbound queue was previously empty, start the dispatch cycle going.
    if (wasEmpty && !connection->outboundQueue.isEmpty()) {
        startDispatchCycleLocked(currentTime, connection);
    }
}

舍繁求簡抓骨幹，繼續追蹤，代碼進入到startDispatchCycleLocked

void InputDispatcher::startDispatchCycleLocked(nsecs_t currentTime,
        const sp<Connection>& connection) {
#if DEBUG_DISPATCH_CYCLE
    ALOGD("channel '%s' ~ startDispatchCycle",
            connection->getInputChannelName());
#endif

    while (connection->status == Connection::STATUS_NORMAL
            && !connection->outboundQueue.isEmpty()) {
        DispatchEntry* dispatchEntry = connection->outboundQueue.head;
        dispatchEntry->deliveryTime = currentTime;

        // Publish the event.
        status_t status;
        EventEntry* eventEntry = dispatchEntry->eventEntry;
        switch (eventEntry->type) {
        case EventEntry::TYPE_KEY: {
            KeyEntry* keyEntry = static_cast<KeyEntry*>(eventEntry);

            // Publish the key event.
            status = connection->inputPublisher.publishKeyEvent(dispatchEntry->seq,
                    keyEntry->deviceId, keyEntry->source,
                    dispatchEntry->resolvedAction, dispatchEntry->resolvedFlags,
                    keyEntry->keyCode, keyEntry->scanCode,
                    keyEntry->metaState, keyEntry->repeatCount, keyEntry->downTime,
                    keyEntry->eventTime);
            break;
        }

        case EventEntry::TYPE_MOTION: {
            MotionEntry* motionEntry = static_cast<MotionEntry*>(eventEntry);

            PointerCoords scaledCoords[MAX_POINTERS];
            const PointerCoords* usingCoords = motionEntry->pointerCoords;

            // Set the X and Y offset depending on the input source.
            float xOffset, yOffset, scaleFactor;
            if ((motionEntry->source & AINPUT_SOURCE_CLASS_POINTER)
                    && !(dispatchEntry->targetFlags & InputTarget::FLAG_ZERO_COORDS)) {
                scaleFactor = dispatchEntry->scaleFactor;
                xOffset = dispatchEntry->xOffset * scaleFactor;
                yOffset = dispatchEntry->yOffset * scaleFactor;
                if (scaleFactor != 1.0f) {
                    for (uint32_t i = 0; i < motionEntry->pointerCount; i++) {
                        scaledCoords[i] = motionEntry->pointerCoords[i];
                        scaledCoords[i].scale(scaleFactor);
                    }
                    usingCoords = scaledCoords;
                }
            } else {
                xOffset = 0.0f;
                yOffset = 0.0f;
                scaleFactor = 1.0f;

                // We don't want the dispatch target to know.
                if (dispatchEntry->targetFlags & InputTarget::FLAG_ZERO_COORDS) {
                    for (uint32_t i = 0; i < motionEntry->pointerCount; i++) {
                        scaledCoords[i].clear();
                    }
                    usingCoords = scaledCoords;
                }
            }

            // Publish the motion event.
            status = connection->inputPublisher.publishMotionEvent(dispatchEntry->seq,
                    motionEntry->deviceId, motionEntry->source,
                    dispatchEntry->resolvedAction, motionEntry->actionButton,
                    dispatchEntry->resolvedFlags, motionEntry->edgeFlags,
                    motionEntry->metaState, motionEntry->buttonState,
                    xOffset, yOffset, motionEntry->xPrecision, motionEntry->yPrecision,
                    motionEntry->downTime, motionEntry->eventTime,
                    motionEntry->pointerCount, motionEntry->pointerProperties,
                    usingCoords);
            break;
        }

        default:
            ALOG_ASSERT(false);
            return;
        }

        // Check the result.
        if (status) {
            if (status == WOULD_BLOCK) {
                if (connection->waitQueue.isEmpty()) {
                    ALOGE("channel '%s' ~ Could not publish event because the pipe is full. "
                            "This is unexpected because the wait queue is empty, so the pipe "
                            "should be empty and we shouldn't have any problems writing an "
                            "event to it, status=%d", connection->getInputChannelName(), status);
                    abortBrokenDispatchCycleLocked(currentTime, connection, true /*notify*/);
                } else {
                    // Pipe is full and we are waiting for the app to finish process some events
                    // before sending more events to it.
#if DEBUG_DISPATCH_CYCLE
                    ALOGD("channel '%s' ~ Could not publish event because the pipe is full, "
                            "waiting for the application to catch up",
                            connection->getInputChannelName());
#endif
                    connection->inputPublisherBlocked = true;
                }
            } else {
                ALOGE("channel '%s' ~ Could not publish event due to an unexpected error, "
                        "status=%d", connection->getInputChannelName(), status);
                abortBrokenDispatchCycleLocked(currentTime, connection, true /*notify*/);
            }
            return;
        }

        // Re-enqueue the event on the wait queue.
        connection->outboundQueue.dequeue(dispatchEntry);
        traceOutboundQueueLengthLocked(connection);
        connection->waitQueue.enqueueAtTail(dispatchEntry);
        traceWaitQueueLengthLocked(connection);
    }
}

只關注咱們的TYPE_KEY類型的事件： 
connection->inputPublisher.publishKeyEvent：

status_t InputPublisher::publishKeyEvent(
        uint32_t seq,
        int32_t deviceId,
        int32_t source,
        int32_t action,
        int32_t flags,
        int32_t keyCode,
        int32_t scanCode,
        int32_t metaState,
        int32_t repeatCount,
        nsecs_t downTime,
        nsecs_t eventTime) {
#if DEBUG_TRANSPORT_ACTIONS
    ALOGD("channel '%s' publisher ~ publishKeyEvent: seq=%u, deviceId=%d, source=0x%x, "
            "action=0x%x, flags=0x%x, keyCode=%d, scanCode=%d, metaState=0x%x, repeatCount=%d,"
            "downTime=%lld, eventTime=%lld",
            mChannel->getName().string(), seq,
            deviceId, source, action, flags, keyCode, scanCode, metaState, repeatCount,
            downTime, eventTime);
#endif

    if (!seq) {
        ALOGE("Attempted to publish a key event with sequence number 0.");
        return BAD_VALUE;
    }

    InputMessage msg;
    msg.header.type = InputMessage::TYPE_KEY;
    msg.body.key.seq = seq;
    msg.body.key.deviceId = deviceId;
    msg.body.key.source = source;
    msg.body.key.action = action;
    msg.body.key.flags = flags;
    msg.body.key.keyCode = keyCode;
    msg.body.key.scanCode = scanCode;
    msg.body.key.metaState = metaState;
    msg.body.key.repeatCount = repeatCount;
    msg.body.key.downTime = downTime;
    msg.body.key.eventTime = eventTime;
    return mChannel->sendMessage(&msg);
}

繼續追蹤代碼流程：調用到了InputChannel的sendMessage方法了。

status_t InputChannel::sendMessage(const InputMessage* msg) {
    size_t msgLength = msg->size();
    ssize_t nWrite;
    do {
        nWrite = ::send(mFd, msg, msgLength, MSG_DONTWAIT | MSG_NOSIGNAL);
    } while (nWrite == -1 && errno == EINTR);

    if (nWrite < 0) {
        int error = errno;
#if DEBUG_CHANNEL_MESSAGES
        ALOGD("channel '%s' ~ error sending message of type %d, errno=%d", mName.string(),
                msg->header.type, error);
#endif
        if (error == EAGAIN || error == EWOULDBLOCK) {
            return WOULD_BLOCK;
        }
        if (error == EPIPE || error == ENOTCONN || error == ECONNREFUSED || error == ECONNRESET) {
            return DEAD_OBJECT;
        }
        return -error;
    }

    if (size_t(nWrite) != msgLength) {
#if DEBUG_CHANNEL_MESSAGES
        ALOGD("channel '%s' ~ error sending message type %d, send was incomplete",
                mName.string(), msg->header.type);
#endif
        return DEAD_OBJECT;
    }

#if DEBUG_CHANNEL_MESSAGES
    ALOGD("channel '%s' ~ sent message of type %d", mName.string(), msg->header.type);
#endif
    return OK;
}

這個方法中，重點是下面這行代碼：

::send(mFd, msg, msgLength, MSG_DONTWAIT | MSG_NOSIGNAL);
1
InputChannel的本質是linux本地套接字。linux本地套接字能夠用於進程間通訊，InputChannel的openInputChannelPair方法中使用了socketpair函數建立了Linux本地套接字，socketpair會返回兩個文件描述符，持有這兩個文件描述符的進程就能夠進行進程間的通訊。

status_t InputChannel::openInputChannelPair(const String8& name,
        sp<InputChannel>& outServerChannel, sp<InputChannel>& outClientChannel) {
    int sockets[2];
    if (socketpair(AF_UNIX, SOCK_SEQPACKET, 0, sockets)) {
        status_t result = -errno;
        ALOGE("channel '%s' ~ Could not create socket pair.  errno=%d",
                name.string(), errno);
        outServerChannel.clear();
        outClientChannel.clear();
        return result;
    }

    int bufferSize = SOCKET_BUFFER_SIZE;
    setsockopt(sockets[0], SOL_SOCKET, SO_SNDBUF, &bufferSize, sizeof(bufferSize));
    setsockopt(sockets[0], SOL_SOCKET, SO_RCVBUF, &bufferSize, sizeof(bufferSize));
    setsockopt(sockets[1], SOL_SOCKET, SO_SNDBUF, &bufferSize, sizeof(bufferSize));
    setsockopt(sockets[1], SOL_SOCKET, SO_RCVBUF, &bufferSize, sizeof(bufferSize));

    String8 serverChannelName = name;
    serverChannelName.append(" (server)");
    outServerChannel = new InputChannel(serverChannelName, sockets[0]);

    String8 clientChannelName = name;     clientChannelName.append(" (client)");     outClientChannel = new InputChannel(clientChannelName, sockets[1]);     return OK; } 這個方法的最後建立了兩個InputChannel，分別傳入一個文件描述符。所以持有這兩個InputChannel實例的進程就能夠作進程間的通訊。  當咱們調用到InputChannel的sendMessage方法時，其實已是在和另外一端通訊了，另外一端收到的消息就是咱們的按鍵事件了。咱們消息分發到此就到了盡頭，另外一端究竟是誰在接收？咱們須要另作分析，我不太喜歡寫特別長的博客，讀起來太費勁，所以，預知是誰在接收事件，請見下回分解。 ---------------------