AndroidQ 图形系统（4）queueBuffer函数分析
作者：互联网
上一篇文章分析了dequeueBuffer函数的过程，本篇接着分析queueBuffer函数，当我们需要绘制图像时，调用dequeueBuffer函数获取到可用的GraphicBuffer之后就可以开始绘制了，最常见的绘制操作就是Android上层View的draw方法了，其他还有OpenGL ES、 mediaserver 视频解码器都可以作为图形数据的来源。
当对GraphicBuffer的绘制操作完成之后就需要调用queueBuffer函数将这块buffer放入BufferQueue队列中并通过回调通知消费者使用这块buffer，图形数据最常见的消耗方就是SurfaceFlinger，该系统服务会消耗当前可见的 Surface（核心是GraphicBuffer），合成到显示部分。
具体实的合成流程是SurfaceFlinger使用 OpenGL和Hardware Composer来合成一组 Surface。
其他 OpenGL ES 应用也可以消耗图像流，例如相机应用会消耗相机预览图像流。非 GL 应用也可以是使用方，例如 ImageReader 类。（引自Android官网）。
我们从Surface.cpp的queueBuffer函数开始看，这个函数比较长我们分为三部分来看。
Surface::queueBuffer
int Surface::queueBuffer(android_native_buffer_t* buffer, int fenceFd) {
......
int i = getSlotFromBufferLocked(buffer);
if (i < 0) {
if (fenceFd >= 0) {
close(fenceFd);
}
return i;
}
if (mSharedBufferSlot == i && mSharedBufferHasBeenQueued) {
if (fenceFd >= 0) {
close(fenceFd);
}
return OK;
}
// Make sure the crop rectangle is entirely inside the buffer.
Rect crop(Rect::EMPTY_RECT);
mCrop.intersect(Rect(buffer->width, buffer->height), &crop);
sp<Fence> fence(fenceFd >= 0 ? new Fence(fenceFd) : Fence::NO_FENCE);
IGraphicBufferProducer::QueueBufferOutput output;
IGraphicBufferProducer::QueueBufferInput input(timestamp, isAutoTimestamp,
static_cast<android_dataspace>(mDataSpace), crop, mScalingMode,
mTransform ^ mStickyTransform, fence, mStickyTransform,
mEnableFrameTimestamps);
// we should send HDR metadata as needed if this becomes a bottleneck
input.setHdrMetadata(mHdrMetadata);
ALOGE("dongjiao...mConnectedToCpu = :%d,INVALID_RECT = :%d",mConnectedToCpu,(mDirtyRegion.bounds() == Rect::INVALID_RECT));
if (mConnectedToCpu || mDirtyRegion.bounds() == Rect::INVALID_RECT) {
input.setSurfaceDamage(Region::INVALID_REGION);
} else {
// Here we do two things:
// 1) The surface damage was specified using the OpenGL ES convention of
// the origin being in the bottom-left corner. Here we flip to the
// convention that the rest of the system uses (top-left corner) by
// subtracting all top/bottom coordinates from the buffer height.
// 2) If the buffer is coming in rotated (for example, because the EGL
// implementation is reacting to the transform hint coming back from
// SurfaceFlinger), the surface damage needs to be rotated the
// opposite direction, since it was generated assuming an unrotated
// buffer (the app doesn't know that the EGL implementation is
// reacting to the transform hint behind its back). The
// transformations in the switch statement below apply those
// complementary rotations (e.g., if 90 degrees, rotate 270 degrees).
int width = buffer->width;
int height = buffer->height;
bool rotated90 = (mTransform ^ mStickyTransform) &
NATIVE_WINDOW_TRANSFORM_ROT_90;
if (rotated90) {
std::swap(width, height);
}
Region flippedRegion;
for (auto rect : mDirtyRegion) {
int left = rect.left;
int right = rect.right;
int top = height - rect.bottom; // Flip from OpenGL convention
int bottom = height - rect.top; // Flip from OpenGL convention
switch (mTransform ^ mStickyTransform) {
case NATIVE_WINDOW_TRANSFORM_ROT_90: {
// Rotate 270 degrees
Rect flippedRect{top, width - right, bottom, width - left};
flippedRegion.orSelf(flippedRect);
break;
}
case NATIVE_WINDOW_TRANSFORM_ROT_180: {
// Rotate 180 degrees
Rect flippedRect{width - right, height - bottom,
width - left, height - top};
flippedRegion.orSelf(flippedRect);
break;
}
case NATIVE_WINDOW_TRANSFORM_ROT_270: {
// Rotate 90 degrees
Rect flippedRect{height - bottom, left,
height - top, right};
flippedRegion.orSelf(flippedRect);
break;
}
default: {
Rect flippedRect{left, top, right, bottom};
flippedRegion.orSelf(flippedRect);
break;
}
}
}
input.setSurfaceDamage(flippedRegion);
}
.....
}
首先第一部分，queueBuffer的第一个参数buffer就是当前绘制完成的GraphicBuffer，第二个参数fenceFd是一种用作同步的工具。
- 首先来看函数
getSlotFromBufferLocked：
int Surface::getSlotFromBufferLocked(
android_native_buffer_t* buffer) const {
for (int i = 0; i < NUM_BUFFER_SLOTS; i++) {
if (mSlots[i].buffer != nullptr &&
mSlots[i].buffer->handle == buffer->handle) {
return i;
}
}
ALOGE("getSlotFromBufferLocked: unknown buffer: %p", buffer->handle);
return BAD_VALUE;
}
这个函数很简单，就是找到buffer在mSlots中的下标，mSlots为所有BufferSlot集合，得到下标之后进行一些合法性判断，如果是mSharedBufferSlot并且mSharedBufferHasBeenQueued代码当前使用的是共享buffer模式则直接返回OK，不太了解共享buffer模式，就跳过了。
- 接着定义一个空的
Rect crop矩形，mCrop也是一个Rect对象，用于buffer的裁剪，我们看它的intersect函数：
bool Rect::intersect(const Rect& with, Rect* result) const {
result->left = max(left, with.left);
result->top = max(top, with.top);
result->right = min(right, with.right);
result->bottom = min(bottom, with.bottom);
return !(result->isEmpty());
}
这个函数用于获取两个矩形相交的部分，并将相交部分赋值给result，再看看调用intersect传递的参数：
mCrop.intersect(Rect(buffer->width, buffer->height), &crop);
其实就是将mCrop和Rect(buffer->width, buffer->height)相交部分赋值给crop。
-
接着创建了两个
QueueBufferOutput对象，input作为输入参数传递到SurfaceFlinger进程，input里面封装的这些信息很多都是外部应用程序设置的，如mDataSpace，mScalingMode，mStickyTransform这些都会应用到最终的显示，而output会作为SurfaceFlinger进程的返回值。 -
接着有一个判断，
mConnectedToCpu的含义是：CPU生产者(在CPU上运行并生成图形数据的代码)连接到BufferQueue，可以通过函数native_window_api_connect传入参数NATIVE_WINDOW_API_CPU设置CPU生产者，但现在所用的图形数据更多来自OpenGL ES、 mediaserver 视频解码器和camera API，对应连接的API为NATIVE_WINDOW_API_EGL，NATIVE_WINDOW_API_MEDIA，
NATIVE_WINDOW_API_CAMERA，所以mConnectedToCpu大多数为false。
mDirtyRegion：脏区域，需要更新图形数据的区域，如果没有需要更新的脏区域则设置无效脏区域。
否则就根据buffer宽高以及旋转角度按一定规则创建flippedRegion，并设置给input传递到SurfaceFlinger。
setSurfaceDamage设置脏区域，对于非脏区域，会直接从之前的buffer中复制，仅仅更新指定的脏区域部分，提高效率。
我们可以看到所有的图形数据都被封装在了input中。
Surface::queueBuffer第一部分分析完了，接着看第二部分：
int Surface::queueBuffer(android_native_buffer_t* buffer, int fenceFd) {
......
status_t err = mGraphicBufferProducer->queueBuffer(i, input, &output);
......
return err;
}
这部分就是本篇的核心，调用生产者queueBuffer函数，传递的依然是BufferSlot下标，到了BufferQueueProducer中通过下标就能找到绑定的GraphicBuffer。
BufferQueueProducer::queueBuffer
status_t BufferQueueProducer::queueBuffer(int slot,
const QueueBufferInput &input, QueueBufferOutput *output) {
ATRACE_CALL();
ATRACE_BUFFER_INDEX(slot);
int64_t requestedPresentTimestamp;
bool isAutoTimestamp;
android_dataspace dataSpace;
Rect crop(Rect::EMPTY_RECT);
int scalingMode;
uint32_t transform;
uint32_t stickyTransform;
sp<Fence> acquireFence;
bool getFrameTimestamps = false;
input.deflate(&requestedPresentTimestamp, &isAutoTimestamp, &dataSpace,
&crop, &scalingMode, &transform, &acquireFence, &stickyTransform,
&getFrameTimestamps);
const Region& surfaceDamage = input.getSurfaceDamage();
const HdrMetadata& hdrMetadata = input.getHdrMetadata();
if (acquireFence == nullptr) {
BQ_LOGE("queueBuffer: fence is NULL");
return BAD_VALUE;
}
auto acquireFenceTime = std::make_shared<FenceTime>(acquireFence);
switch (scalingMode) {
case NATIVE_WINDOW_SCALING_MODE_FREEZE:
case NATIVE_WINDOW_SCALING_MODE_SCALE_TO_WINDOW:
case NATIVE_WINDOW_SCALING_MODE_SCALE_CROP:
case NATIVE_WINDOW_SCALING_MODE_NO_SCALE_CROP:
break;
default:
BQ_LOGE("queueBuffer: unknown scaling mode %d", scalingMode);
return BAD_VALUE;
}
sp<IConsumerListener> frameAvailableListener;
sp<IConsumerListener> frameReplacedListener;
int callbackTicket = 0;
uint64_t currentFrameNumber = 0;
BufferItem item;
{ // Autolock scope
std::lock_guard<std::mutex> lock(mCore->mMutex);
if (mCore->mIsAbandoned) {
BQ_LOGE("queueBuffer: BufferQueue has been abandoned");
return NO_INIT;
}
if (mCore->mConnectedApi == BufferQueueCore::NO_CONNECTED_API) {
BQ_LOGE("queueBuffer: BufferQueue has no connected producer");
return NO_INIT;
}
if (slot < 0 || slot >= BufferQueueDefs::NUM_BUFFER_SLOTS) {
BQ_LOGE("queueBuffer: slot index %d out of range [0, %d)",
slot, BufferQueueDefs::NUM_BUFFER_SLOTS);
return BAD_VALUE;
} else if (!mSlots[slot].mBufferState.isDequeued()) {
BQ_LOGE("queueBuffer: slot %d is not owned by the producer "
"(state = %s)", slot, mSlots[slot].mBufferState.string());
return BAD_VALUE;
} else if (!mSlots[slot].mRequestBufferCalled) {
BQ_LOGE("queueBuffer: slot %d was queued without requesting "
"a buffer", slot);
return BAD_VALUE;
}
// If shared buffer mode has just been enabled, cache the slot of the
// first buffer that is queued and mark it as the shared buffer.
if (mCore->mSharedBufferMode && mCore->mSharedBufferSlot ==
BufferQueueCore::INVALID_BUFFER_SLOT) {
mCore->mSharedBufferSlot = slot;
mSlots[slot].mBufferState.mShared = true;
}
BQ_LOGV("queueBuffer: slot=%d/%" PRIu64 " time=%" PRIu64 " dataSpace=%d"
" validHdrMetadataTypes=0x%x crop=[%d,%d,%d,%d] transform=%#x scale=%s",
slot, mCore->mFrameCounter + 1, requestedPresentTimestamp, dataSpace,
hdrMetadata.validTypes, crop.left, crop.top, crop.right, crop.bottom,
transform,
BufferItem::scalingModeName(static_cast<uint32_t>(scalingMode)));
const sp<GraphicBuffer>& graphicBuffer(mSlots[slot].mGraphicBuffer);
Rect bufferRect(graphicBuffer->getWidth(), graphicBuffer->getHeight());
Rect croppedRect(Rect::EMPTY_RECT);
crop.intersect(bufferRect, &croppedRect);
if (croppedRect != crop) {
BQ_LOGE("queueBuffer: crop rect is not contained within the "
"buffer in slot %d", slot);
return BAD_VALUE;
}
// Override UNKNOWN dataspace with consumer default
if (dataSpace == HAL_DATASPACE_UNKNOWN) {
dataSpace = mCore->mDefaultBufferDataSpace;
}
mSlots[slot].mFence = acquireFence;
mSlots[slot].mBufferState.queue();
// Increment the frame counter and store a local version of it
// for use outside the lock on mCore->mMutex.
++mCore->mFrameCounter;
currentFrameNumber = mCore->mFrameCounter;
mSlots[slot].mFrameNumber = currentFrameNumber;
item.mAcquireCalled = mSlots[slot].mAcquireCalled;
item.mGraphicBuffer = mSlots[slot].mGraphicBuffer;
item.mCrop = crop;
item.mTransform = transform &
~static_cast<uint32_t>(NATIVE_WINDOW_TRANSFORM_INVERSE_DISPLAY);
item.mTransformToDisplayInverse =
(transform & NATIVE_WINDOW_TRANSFORM_INVERSE_DISPLAY) != 0;
item.mScalingMode = static_cast<uint32_t>(scalingMode);
item.mTimestamp = requestedPresentTimestamp;
item.mIsAutoTimestamp = isAutoTimestamp;
item.mDataSpace = dataSpace;
item.mHdrMetadata = hdrMetadata;
item.mFrameNumber = currentFrameNumber;
item.mSlot = slot;
item.mFence = acquireFence;
item.mFenceTime = acquireFenceTime;
item.mIsDroppable = mCore->mAsyncMode ||
(mConsumerIsSurfaceFlinger && mCore->mQueueBufferCanDrop) ||
(mCore->mLegacyBufferDrop && mCore->mQueueBufferCanDrop) ||
(mCore->mSharedBufferMode && mCore->mSharedBufferSlot == slot);
item.mSurfaceDamage = surfaceDamage;
item.mQueuedBuffer = true;
item.mAutoRefresh = mCore->mSharedBufferMode && mCore->mAutoRefresh;
item.mApi = mCore->mConnectedApi;
mStickyTransform = stickyTransform;
// Cache the shared buffer data so that the BufferItem can be recreated.
if (mCore->mSharedBufferMode) {
mCore->mSharedBufferCache.crop = crop;
mCore->mSharedBufferCache.transform = transform;
mCore->mSharedBufferCache.scalingMode = static_cast<uint32_t>(
scalingMode);
mCore->mSharedBufferCache.dataspace = dataSpace;
}
output->bufferReplaced = false;
if (mCore->mQueue.empty()) {
// When the queue is empty, we can ignore mDequeueBufferCannotBlock
// and simply queue this buffer
mCore->mQueue.push_back(item);
frameAvailableListener = mCore->mConsumerListener;
} else {
// When the queue is not empty, we need to look at the last buffer
// in the queue to see if we need to replace it
const BufferItem& last = mCore->mQueue.itemAt(
mCore->mQueue.size() - 1);
if (last.mIsDroppable) {
if (!last.mIsStale) {
mSlots[last.mSlot].mBufferState.freeQueued();
// After leaving shared buffer mode, the shared buffer will
// still be around. Mark it as no longer shared if this
// operation causes it to be free.
if (!mCore->mSharedBufferMode &&
mSlots[last.mSlot].mBufferState.isFree()) {
mSlots[last.mSlot].mBufferState.mShared = false;
}
// Don't put the shared buffer on the free list.
if (!mSlots[last.mSlot].mBufferState.isShared()) {
mCore->mActiveBuffers.erase(last.mSlot);
mCore->mFreeBuffers.push_back(last.mSlot);
output->bufferReplaced = true;
}
}
// Make sure to merge the damage rect from the frame we're about
// to drop into the new frame's damage rect.
if (last.mSurfaceDamage.bounds() == Rect::INVALID_RECT ||
item.mSurfaceDamage.bounds() == Rect::INVALID_RECT) {
item.mSurfaceDamage = Region::INVALID_REGION;
} else {
item.mSurfaceDamage |= last.mSurfaceDamage;
}
// Overwrite the droppable buffer with the incoming one
mCore->mQueue.editItemAt(mCore->mQueue.size() - 1) = item;
frameReplacedListener = mCore->mConsumerListener;
} else {
mCore->mQueue.push_back(item);
frameAvailableListener = mCore->mConsumerListener;
}
}
......
}
这个函数也比较多，同样分部分来看，首先看第一部分：
- 定义了一堆变量，用来保存
Surface传递过来的input里面封装的buffer信息。 - 接着需要检测
scalingMode，缩放模式，当buffer内容和屏幕不成比例时采用什么方式处理，scalingMode可以通过函数native_window_set_scaling_mode进行设置，如果没有设置直接返回BAD_VALUE。 - 又进行一堆检测：
BufferQueue是否被弃用，buffer是否连接到BufferQueue，BufferSlot下标是否合法，当前这个BufferSlot状态是否是DEQUEUE，mRequestBufferCalled是否为true（即此BufferSlot是否调用了requestBuffer函数）。 - 接着又创建了三个对象，
graphicBuffer（此次queue的具体buffer），bufferRect（根据当前graphicBuffer创建的矩形区域），croppedRect（裁剪区域，值为crop和bufferRect相交部分，其实这里的crop.intersect(bufferRect, &croppedRect)和前面Surface里面mCrop.intersect(Rect(buffer->width, buffer->height), &crop)会得到同样的裁剪区域，只要裁剪的部分包含在当前buffer之内）。 - 接着将当前
BufferSlot状态修改为QUEUE，并修改一些状态值。 - 将
BufferSlot以及GraphicBuffer的信息进一步封装到BufferItem，而BufferItem又存储在BufferQueueCore的mQueue中，所以一般我们说的生产者-消费者模型中的buffer队列，更具体的话就是mQueue。如果mQueue为空就直接将BufferItem添加进去，并将BufferQueueCore的mConsumerListener给到frameAvailableListener，从AndroidQ 图形系统（2）生产者-消费者模型知道，mConsumerListener是在ConsumerBase的构造函数中传递的BufferQueue::ProxyConsumerListener。
对于mQueue非空的情况，首先取到mQueue尾部的BufferItem，根据mIsDroppable值查看是否需要丢弃，什么情况下需要丢弃？有很多种情况，常见就是同步模式即mAsyncMode为true，至于需要丢弃的情况怎么处理的我这里不去看了，对于不需要丢弃的情况总是将BufferItem放入mQueue的尾部。
此部分已经分析完了，主要做的事情就是上面这六点。
接着看剩下的部分：
status_t BufferQueueProducer::queueBuffer(int slot,
const QueueBufferInput &input, QueueBufferOutput *output) {
......
mCore->mBufferHasBeenQueued = true;
mCore->mDequeueCondition.notify_all();
mCore->mLastQueuedSlot = slot;
output->width = mCore->mDefaultWidth;
output->height = mCore->mDefaultHeight;
output->transformHint = mCore->mTransformHint;
output->numPendingBuffers = static_cast<uint32_t>(mCore->mQueue.size());
output->nextFrameNumber = mCore->mFrameCounter + 1;
ATRACE_INT(mCore->mConsumerName.string(),
static_cast<int32_t>(mCore->mQueue.size()));
mCore->mOccupancyTracker.registerOccupancyChange(mCore->mQueue.size());
// Take a ticket for the callback functions
callbackTicket = mNextCallbackTicket++;
VALIDATE_CONSISTENCY();
} // Autolock scope
// It is okay not to clear the GraphicBuffer when the consumer is SurfaceFlinger because
// it is guaranteed that the BufferQueue is inside SurfaceFlinger's process and
// there will be no Binder call
if (!mConsumerIsSurfaceFlinger) {
item.mGraphicBuffer.clear();
}
// Call back without the main BufferQueue lock held, but with the callback
// lock held so we can ensure that callbacks occur in order
int connectedApi;
sp<Fence> lastQueuedFence;
{ // scope for the lock
std::unique_lock<std::mutex> lock(mCallbackMutex);
while (callbackTicket != mCurrentCallbackTicket) {
mCallbackCondition.wait(lock);
}
if (frameAvailableListener != nullptr) {
frameAvailableListener->onFrameAvailable(item);
} else if (frameReplacedListener != nullptr) {
frameReplacedListener->onFrameReplaced(item);
}
connectedApi = mCore->mConnectedApi;
lastQueuedFence = std::move(mLastQueueBufferFence);
mLastQueueBufferFence = std::move(acquireFence);
mLastQueuedCrop = item.mCrop;
mLastQueuedTransform = item.mTransform;
++mCurrentCallbackTicket;
mCallbackCondition.notify_all();
}
// Update and get FrameEventHistory.
nsecs_t postedTime = systemTime(SYSTEM_TIME_MONOTONIC);
NewFrameEventsEntry newFrameEventsEntry = {
currentFrameNumber,
postedTime,
requestedPresentTimestamp,
std::move(acquireFenceTime)
};
addAndGetFrameTimestamps(&newFrameEventsEntry,
getFrameTimestamps ? &output->frameTimestamps : nullptr);
// Wait without lock held
if (connectedApi == NATIVE_WINDOW_API_EGL) {
// Waiting here allows for two full buffers to be queued but not a
// third. In the event that frames take varying time, this makes a
// small trade-off in favor of latency rather than throughput.
lastQueuedFence->waitForever("Throttling EGL Production");
}
return NO_ERROR;
}
这部分主要做了如下事情：
mBufferHasBeenQueued一个状态值，代表buffer已经queued，mDequeueCondition是C++条件变量用作等待/唤醒，这里调用notify_all会唤醒调用了wait的线程，即在dequeueBuffer函数中因为tooManyBuffers而陷入等待状态的线程，当一块buffer被queued之后就可以继续dequeue了，output最终会返回给Surface进程，mConsumerIsSurfaceFlinger表示消费者是否SurfaceFlinger进程，默认值为true。frameAvailableListener不为空则调用onFrameAvailable回调函数通知有buffer可以进行消费，这个函数会经过一系列调用，最终的实现类是BufferQueueLayer（一种最常用的Layer），最后如果连接BufferQueue的方式为NATIVE_WINDOW_API_EGL，会通过Fence陷入等待，这里会一直等，知道上一帧绘制完成，图形系统的生产消费非常依赖Fence同步机制，Android引入三重缓冲之后，GPU，CPU，显示硬件可以各持有一个buffer工作，例如当CPU持有的buffer数据已经处理完了，但此时GUP，显示硬件还在使用buffer，CPU并不能立马申请下一个buffer，需要等待，我这里对Fence不了解，所以文中出现的和Fence同步相关的代码就略过了。
到此queueBuffer函数已经分析完毕，代码看着多，逻辑比较简单，就是将Surface进程数据传过来，然后封装成BufferItem，放入BufferQueueCore的mQueue中，再通过frameAvailableListener通知消费者去消费。
接着回到Surface看剩下的代码，
int Surface::queueBuffer(android_native_buffer_t* buffer, int fenceFd) {
......
if (err != OK) {
ALOGE("queueBuffer: error queuing buffer to SurfaceTexture, %d", err);
}
if (mEnableFrameTimestamps) {
mFrameEventHistory->applyDelta(output.frameTimestamps);
// Update timestamps with the local acquire fence.
// The consumer doesn't send it back to prevent us from having two
// file descriptors of the same fence.
mFrameEventHistory->updateAcquireFence(mNextFrameNumber,
std::make_shared<FenceTime>(fence));
// Cache timestamps of signaled fences so we can close their file
// descriptors.
mFrameEventHistory->updateSignalTimes();
}
mLastFrameNumber = mNextFrameNumber;
mDefaultWidth = output.width;
mDefaultHeight = output.height;
mNextFrameNumber = output.nextFrameNumber;
// Ignore transform hint if sticky transform is set or transform to display inverse flag is
// set.
if (mStickyTransform == 0 && !transformToDisplayInverse()) {
mTransformHint = output.transformHint;
}
mConsumerRunningBehind = (output.numPendingBuffers >= 2);
if (!mConnectedToCpu) {
// Clear surface damage back to full-buffer
mDirtyRegion = Region::INVALID_REGION;
}
if (mSharedBufferMode && mAutoRefresh && mSharedBufferSlot == i) {
mSharedBufferHasBeenQueued = true;
}
mQueueBufferCondition.broadcast();
if (CC_UNLIKELY(atrace_is_tag_enabled(ATRACE_TAG_GRAPHICS))) {
static FenceMonitor gpuCompletionThread("GPU completion");
gpuCompletionThread.queueFence(fence);
}
return err;
这后面部分其实没什么好说的了，会更新一些变量，清空脏区域，通过条件变量mQueueBufferCondition唤醒等待下一帧的线程，处理一下Fence。
标签：slot,图形系统,item,buffer,AndroidQ,mCore,queueBuffer,Rect 来源： https://blog.csdn.net/qq_34211365/article/details/106588430