前文Android匿名共享内存（Ashmem）原理分析了匿名共享内存，它最主要的作用就是View视图绘制，Android视图是按照一帧一帧显示到屏幕的，而每一帧都会占用一定的存储空间，通过Ashmem机制APP与SurfaceFlinger共享绘图数据，提高图形处理性能，本文就看Android是怎么利用Ashmem分配及绘制的：
View视图内存的分配
前文Window添加流程中描述了：在添加窗口的时候，WMS会为APP分配一个WindowState，以标识当前窗口并用于窗口管理，同时向SurfaceFlinger端请求分配Layer抽象图层，在SurfaceFlinger分配Layer的时候创建了两个比较关键的Binder对象，用于填充WMS端Surface，一个是sp
- 分配的时机：什么时候分配
- 分配的手段：如何分配
- 传递的方式：如何跨进程传递
Surface被抽象成一块画布，只要拥有Surface就可以绘图，其根本原理就是Surface握有可以绘图的一块内存，这块内存是APP端在需要的时候，通过sp
sp<SurfaceControl> SurfaceComposerClient::createSurface(
const String8& name, uint32_t w, uint32_t h, PixelFormat format, uint32_t flags){
sp<SurfaceControl> sur;
...
if (mStatus == NO_ERROR) {
sp<IBinder> handle;
sp<IGraphicBufferProducer> gbp;
<!--关键点1 获取图层的关键信息handle, gbp-->
status_t err = mClient->createSurface(name, w, h, format, flags,
&handle, &gbp);
<!--关键点2 根据返回的图层关键信息 创建SurfaceControl对象-->
if (err == NO_ERROR) {
sur = new SurfaceControl(this, handle, gbp);
}
}
return sur;
}
看关键点1，这里其实就是建立了一个sp
status_t SurfaceFlinger::createNormalLayer(const sp<Client>& client,
const String8& name, uint32_t w, uint32_t h, uint32_t flags, PixelFormat& format,
sp<IBinder>* handle, sp<IGraphicBufferProducer>* gbp, sp<Layer>* outLayer){
...
<!--关键点 1 -->
*outLayer = new Layer(this, client, name, w, h, flags);
status_t err = (*outLayer)->setBuffers(w, h, format, flags);
<!--关键点 2-->
if (err == NO_ERROR) {
*handle = (*outLayer)->getHandle();
*gbp = (*outLayer)->getProducer();
}
return err;
}
void Layer::onFirstRef() {
sp<IGraphicBufferProducer> producer;
sp<IGraphicBufferConsumer> consumer;
<!--创建producer与consumer-->
BufferQueue::createBufferQueue(&producer, &consumer);
mProducer = new MonitoredProducer(producer, mFlinger);
mSurfaceFlingerConsumer = new SurfaceFlingerConsumer(consumer, mTextureName,
this);
...
}
void BufferQueue::createBufferQueue(sp<IGraphicBufferProducer>* outProducer,
sp<IGraphicBufferConsumer>* outConsumer,
const sp<IGraphicBufferAlloc>& allocator) {
sp<BufferQueueCore> core(new BufferQueueCore(allocator));
sp<IGraphicBufferProducer> producer(new BufferQueueProducer(core));
sp<IGraphicBufferConsumer> consumer(new BufferQueueConsumer(core));
*outProducer = producer;
*outConsumer = consumer;
}
从上面两个函数可以很清楚的看到Producer/Consumer的模型原样，也就说每个图层Layer都有自己的producer/ consumer,sp
class BufferQueueProducer : public BnGraphicBufferProducer,
private IBinder::DeathRecipient {}
在APP端是
class BpGraphicBufferProducer : public BpInterface<IGraphicBufferProducer>{}
IGraphicBufferProducer Binder实体在SurfaceFlinger中创建后，打包到Surface对象，并通过binder通信传递给APP端，APP段通过反序列化将其恢复出来，如下：
status_t Surface::readFromParcel(const Parcel* parcel, bool nameAlreadyRead) {
if (parcel == nullptr) return BAD_VALUE;
status_t res = OK;
if (!nameAlreadyRead) {
name = readMaybeEmptyString16(parcel);
// Discard this for now
int isSingleBuffered;
res = parcel->readInt32(&isSingleBuffered);
if (res != OK) {
return res;
}
}
sp<IBinder> binder;
res = parcel->readStrongBinder(&binder);
if (res != OK) return res;
<!--interface_cast会将其转换成BpGraphicBufferProducer-->
graphicBufferProducer = interface_cast<IGraphicBufferProducer>(binder);
return OK;
}
自此，APP端就获得了申请内存的句柄BpGraphicBufferProducer，它真正发挥作用是在第一次绘图时，看一下ViewRootImpl中的draw
private boolean drawSoftware(Surface surface, AttachInfo attachInfo, int xoff, int yoff,
boolean scalingRequired, Rect dirty) {
final Canvas canvas;
try {
final int left = dirty.left;
final int top = dirty.top;
final int right = dirty.right;
final int bottom = dirty.bottom;
<!--关键点1 获取绘图内存-->
canvas = mSurface.lockCanvas(dirty);
try {
try {
<!--关键点2 绘图-->
mView.draw(canvas);
}
} finally {
try {
<!--关键点 3 绘图结束 ，通知surfacefling混排，更新显示界面-->
surface.unlockCanvasAndPost(canvas);
} catch (IllegalArgumentException e) {}
先看关键点1，内存的分配时机其实就在这里，直接进入到native层
static jlong nativeLockCanvas(JNIEnv* env, jclass clazz,
jlong nativeObject, jobject canvasObj, jobject dirtyRectObj) {
sp<Surface> surface(reinterpret_cast<Surface *>(nativeObject));
...
status_t err = surface->lock(&outBuffer, dirtyRectPtr);
...
sp<Surface> lockedSurface(surface);
lockedSurface->incStrong(&sRefBaseOwner);
return (jlong) lockedSurface.get();
}
surface.cpp的lock会进一步调用dequeueBuffer函数来请求分配内存：
int Surface::dequeueBuffer(android_native_buffer_t** buffer, int* fenceFd) {
...
int buf = -1;
sp<Fence> fence;
nsecs_t now = systemTime();
<!--申请buffer,并获得标识符-->
status_t result = mGraphicBufferProducer->dequeueBuffer(&buf, &fence,
reqWidth, reqHeight, reqFormat, reqUsage);
...
if ((result & IGraphicBufferProducer::BUFFER_NEEDS_REALLOCATION) || gbuf == 0) {
<!--申请的内存是在surfaceflinger进程中，Surface通过调用requestBuffer将图形缓冲区映射到Surface所在进程-->
result = mGraphicBufferProducer->requestBuffer(buf, &gbuf);
...
}
最终会调用BpGraphicBufferProducer的dequeueBuffer向服务端请求分配内存，这里用到了匿名共享内存的知识，在Linux中一切都是文件，共享内存也看成一个文件。分配成功之后，需要跨进程传递tmpfs临时文件的描述符fd。先看下申请的逻辑：
class BpGraphicBufferProducer : public BpInterface<IGraphicBufferProducer>{
virtual status_t dequeueBuffer(int *buf, sp<Fence>* fence, bool async,
uint32_t w, uint32_t h, uint32_t format, uint32_t usage) {
Parcel data, reply;
data.writeInterfaceToken(IGraphicBufferProducer::getInterfaceDescriptor());
data.writeInt32(async);
data.writeInt32(w);
data.writeInt32(h);
data.writeInt32(format);
data.writeInt32(usage);
//通过BpBinder将要什么的buffer的相关参数保存到data，发送给BBinder
status_t result = remote()->transact(DEQUEUE_BUFFER, data, &reply);
if (result != NO_ERROR) {
return result;
}
//BBinder给BpBinder返回了一个int，并不是缓冲区的内存
*buf = reply.readInt32();
bool nonNull = reply.readInt32();
if (nonNull) {
*fence = new Fence();
reply.read(**fence);
}
result = reply.readInt32();
return result;
}
}
在client侧，也就是BpGraphicBufferProducer侧，通过DEQUEUE_BUFFER后核心只返回了一个*buf = reply.readInt32();其实是数组mSlots的下标，在BufferQueue中有个和mSlots对应的数组，也是32个，一一对应，
status_t BnGraphicBufferProducer::onTransact(
uint32_t code, const Parcel& data, Parcel* reply, uint32_t flags)
{
case DEQUEUE_BUFFER: {
CHECK_INTERFACE(IGraphicBufferProducer, data, reply);
bool async = data.readInt32();
uint32_t w = data.readInt32();
uint32_t h = data.readInt32();
uint32_t format = data.readInt32();
uint32_t usage = data.readInt32();
int buf;
sp<Fence> fence;
//调用BufferQueue的dequeueBuffer
//也返回一个int的buf
int result = dequeueBuffer(&buf, &fence, async, w, h, format, usage);
//将buf和fence写入parcel，通过binder传给client
reply->writeInt32(buf);
reply->writeInt32(fence != NULL);
if (fence != NULL) {
reply->write(*fence);
}
reply->writeInt32(result);
return NO_ERROR;
}
可以看到BnGraphicBufferProducer端获取到长宽及格式，之后利用BufferQueueProducer的dequeueBuffer来申请内存，内存可能已经申请，也可能未申请，未申请，则直接申请新内存，每个surface可以对应32块内存：
status_t BufferQueueProducer::dequeueBuffer(int *outSlot,
sp<android::Fence> *outFence, uint32_t width, uint32_t height,
PixelFormat format, uint32_t usage) {
...
sp<GraphicBuffer> graphicBuffer(mCore->mAllocator->createGraphicBuffer(
width, height, format, usage,
{mConsumerName.string(), mConsumerName.size()}, &error));
mCore其实就是上面的BufferQueueCore，mCore->mAllocator = new GraphicBufferAlloc()，最终会利用GraphicBufferAlloc对象分配共享内存:
sp<GraphicBuffer> GraphicBufferAlloc::createGraphicBuffer(uint32_t width,
uint32_t height, PixelFormat format, uint32_t usage,
std::string requestorName, status_t* error) {
<!--直接new新建-->
sp<GraphicBuffer> graphicBuffer(new GraphicBuffer(
width, height, format, usage, std::move(requestorName)));
status_t err = graphicBuffer->initCheck();
return graphicBuffer;
}
从上面看到，直接new GraphicBuffer新建图像内存，
GraphicBuffer::GraphicBuffer(uint32_t inWidth, uint32_t inHeight,
PixelFormat inFormat, uint32_t inUsage, std::string requestorName)
: BASE(), mOwner(ownData), mBufferMapper(GraphicBufferMapper::get()),
mInitCheck(NO_ERROR), mId(getUniqueId()), mGenerationNumber(0){
...
handle = NULL;
mInitCheck = initSize(inWidth, inHeight, inFormat, inUsage,
std::move(requestorName));
}
status_t GraphicBuffer::initSize(uint32_t inWidth, uint32_t inHeight,
PixelFormat inFormat, uint32_t inUsage, std::string requestorName)
{
GraphicBufferAllocator& allocator = GraphicBufferAllocator::get();
uint32_t outStride = 0;
<!--请求分配内存-->
status_t err = allocator.allocate(inWidth, inHeight, inFormat, inUsage,
&handle, &outStride, mId, std::move(requestorName));
if (err == NO_ERROR) {
width = static_cast<int>(inWidth);
height = static_cast<int>(inHeight);
format = inFormat;
usage = static_cast<int>(inUsage);
stride = static_cast<int>(outStride);
}
return err;
}
status_t GraphicBufferAllocator::allocate(uint32_t width, uint32_t height,
PixelFormat format, uint32_t usage, buffer_handle_t* handle,
uint32_t* stride, uint64_t graphicBufferId, std::string requestorName)
{
...
auto descriptor = mDevice->createDescriptor();
auto error = descriptor->setDimensions(width, height);
error = descriptor->setFormat(static_cast<android_pixel_format_t>(format));
error = descriptor->setProducerUsage(
static_cast<gralloc1_producer_usage_t>(usage));
error = descriptor->setConsumerUsage(
static_cast<gralloc1_consumer_usage_t>(usage));
<!--这里的device就是抽象的硬件设备-->
error = mDevice->allocate(descriptor, graphicBufferId, handle);
error = mDevice->getStride(*handle, stride);
...
return NO_ERROR;
}
上面代码的mDevice就是利用hw_get_module及gralloc1_open获取到的硬件抽象层device，hw_get_module装载HAL模块，会加载相应的.so文件gralloc.default.so，它实现位于 hardware/libhardware/modules/gralloc.cpp中，最后将device映射的函数操作加载进来。这里我们关心的是allocate函数，先分析普通图形缓冲区的分配，它最终会调用gralloc_alloc_buffer()利用匿名共享内存进行分配，之前的文章Android匿名共享内存（Ashmem）原理分析了Android是如何通过匿名共享内存进行通信的，这里就直接用了：
static int gralloc_alloc_buffer(alloc_device_t* dev,
size_t size, int usage, buffer_handle_t* pHandle)
{
int err = 0;
int fd = -1;
size = roundUpToPageSize(size);
// 创建共享内存，并且设定名字跟size
fd = ashmem_create_region("gralloc-buffer", size);
if (err == 0) {
private_handle_t* hnd = new private_handle_t(fd, size, 0);
gralloc_module_t* module = reinterpret_cast<gralloc_module_t*>(
dev->common.module);
// 执行mmap，将内存映射到自己的进程
err = mapBuffer(module, hnd);
if (err == 0) {
*pHandle = hnd;
}
}
return err;
}
mapBuffer会进一步调用ashmem的驱动，在tmpfs新建文件，同时开辟虚拟内存，
int mapBuffer(gralloc_module_t const* module,
private_handle_t* hnd)
{
void* vaddr;
// vaddr有个毛用？
return gralloc_map(module, hnd, &vaddr);
}
static int gralloc_map(gralloc_module_t const* module,
buffer_handle_t handle,
void** vaddr)
{
private_handle_t* hnd = (private_handle_t*)handle;
if (!(hnd->flags & private_handle_t::PRIV_FLAGS_FRAMEBUFFER)) {
size_t size = hnd->size;
void* mappedAddress = mmap(0, size,
PROT_READ|PROT_WRITE, MAP_SHARED, hnd->fd, 0);
if (mappedAddress == MAP_FAILED) {
return -errno;
}
hnd->base = intptr_t(mappedAddress) + hnd->offset;
}
*vaddr = (void*)hnd->base;
return 0;
}
View绘制内存的传递
分配之后，会继续利用BpGraphicBufferProducer的requestBuffer，申请将共享内存给映射到当前进程：
virtual status_t requestBuffer(int bufferIdx, sp<GraphicBuffer>* buf) {
Parcel data, reply;
data.writeInterfaceToken(IGraphicBufferProducer::getInterfaceDescriptor());
data.writeInt32(bufferIdx);
status_t result =remote()->transact(REQUEST_BUFFER, data, &reply);
if (result != NO_ERROR) {
return result;
}
bool nonNull = reply.readInt32();
if (nonNull) {
*buf = new GraphicBuffer();
reply.read(**buf);
}
result = reply.readInt32();
return result;
}
private_handle_t对象用来抽象图形缓冲区，其中存储着与共享内存对应tmpfs文件的fd，GraphicBuffer对象会通过序列化，将这个fd会利用Binder通信传递给App进程，APP端获取到fd之后，便可以同mmap将共享内存映射到自己的进程空间，进而进行图形绘制。等到APP端对GraphicBuffer的反序列化的时候，会将共享内存mmap到当前进程空间：
status_t Parcel::read(Flattenable& val) const
{
// size
const size_t len = this->readInt32();
const size_t fd_count = this->readInt32();
// payload
void const* buf = this->readInplace(PAD_SIZE(len));
if (buf == NULL)
return BAD_VALUE;
int* fds = NULL;
if (fd_count) {
fds = new int[fd_count];
}
status_t err = NO_ERROR;
for (size_t i=0 ; i<fd_count && err==NO_ERROR ; i++) {
fds[i] = dup(this->readFileDescriptor());
if (fds[i] < 0) err = BAD_VALUE;
}
if (err == NO_ERROR) {
err = val.unflatten(buf, len, fds, fd_count);
}
if (fd_count) {
delete [] fds;
}
return err;
}
进而调用GraphicBuffer::unflatten：
status_t GraphicBuffer::unflatten(void const* buffer, size_t size,
int fds[], size_t count)
{
...
mOwner = ownHandle;
<!--将共享内存映射当前内存空间-->
if (handle != 0) {
status_t err = mBufferMapper.registerBuffer(handle);
}
return NO_ERROR;
}
mBufferMapper.registerBuffer函数对应gralloc_register_buffer
struct private_module_t HAL_MODULE_INFO_SYM = {
.base = {
.common = {
.tag = HARDWARE_MODULE_TAG,
.version_major = 1,
.version_minor = 0,
.id = GRALLOC_HARDWARE_MODULE_ID,
.name = "Graphics Memory Allocator Module",
.author = "The Android Open Source Project",
.methods = &gralloc_module_methods
},
.registerBuffer = gralloc_register_buffer,
.unregisterBuffer = gralloc_unregister_buffer,
.lock = gralloc_lock,
.unlock = gralloc_unlock,
},
.framebuffer = 0,
.flags = 0,
.numBuffers = 0,
.bufferMask = 0,
.lock = PTHREAD_MUTEX_INITIALIZER,
.currentBuffer = 0,
};
最后会调用gralloc_register_buffer，通过mmap真正将tmpfs文件映射到进程空间：
static int gralloc_register_buffer(gralloc_module_t const* module,
buffer_handle_t handle)
{
...
if (cb->ashmemSize > 0 && cb->mappedPid != getpid()) {
void *vaddr;
<!--mmap-->
int err = map_buffer(cb, &vaddr);
cb->mappedPid = getpid();
}
return 0;
}
终于我们用到tmpfs中文件对应的描述符fd0->cb->fd
static int map_buffer(cb_handle_t *cb, void **vaddr)
{
if (cb->fd < 0 || cb->ashmemSize <= 0) {
return -EINVAL;
}
void *addr = mmap(0, cb->ashmemSize, PROT_READ | PROT_WRITE,
MAP_SHARED, cb->fd, 0);
cb->ashmemBase = intptr_t(addr);
cb->ashmemBasePid = getpid();
*vaddr = addr;
return 0;
}
到这里内存传递成功，App端就可以应用这块内存进行图形绘制了。
View绘制内存的使用
关于内存的使用，我们回到之前的Surface lock函数，内存经过反序列化，拿到内存地址后，会封装一个ANativeWindow_Buffer返回给上层调用：
status_t Surface::lock(
ANativeWindow_Buffer* outBuffer, ARect* inOutDirtyBounds)
{
...
void* vaddr;
<!--lock获取地址-->
status_t res = backBuffer->lock(
GRALLOC_USAGE_SW_READ_OFTEN | GRALLOC_USAGE_SW_WRITE_OFTEN,
newDirtyRegion.bounds(), &vaddr);
if (res != 0) {
err = INVALID_OPERATION;
} else {
mLockedBuffer = backBuffer;
outBuffer->width = backBuffer->width;
outBuffer->height = backBuffer->height;
outBuffer->stride = backBuffer->stride;
outBuffer->format = backBuffer->format;
<!--关键点 设置虚拟内存的地址-->
outBuffer->bits = vaddr;
}
}
return err;
}
ANativeWindow_Buffer的数据结构如下，其中bits字段与虚拟内存地址对应，
typedef struct ANativeWindow_Buffer {
// The number of pixels that are show horizontally.
int32_t width;
// The number of pixels that are shown vertically.
int32_t height;
// The number of *pixels* that a line in the buffer takes in
// memory. This may be >= width.
int32_t stride;
// The format of the buffer. One of WINDOW_FORMAT_*
int32_t format;
// The actual bits.
void* bits;
// Do not touch.
uint32_t reserved[6];
} ANativeWindow_Buffer;
如何使用，看下Canvas的draw
static void nativeLockCanvas(JNIEnv* env, jclass clazz,
jint nativeObject, jobject canvasObj, jobject dirtyRectObj) {
sp<Surface> surface(reinterpret_cast<Surface *>(nativeObject));
...
status_t err = surface->lock(&outBuffer, &dirtyBounds);
...
<!--SkBitmap-->
SkBitmap bitmap;
ssize_t bpr = outBuffer.stride * bytesPerPixel(outBuffer.format);
<!--为SkBitmap填充配置-->
bitmap.setConfig(convertPixelFormat(outBuffer.format), outBuffer.width, outBuffer.height, bpr);
<!--为SkBitmap填充格式-->
if (outBuffer.format == PIXEL_FORMAT_RGBX_8888) {
bitmap.setIsOpaque(true);
}
<!--为SkBitmap填充内存-->
if (outBuffer.width > 0 && outBuffer.height > 0) {
bitmap.setPixels(outBuffer.bits);
} else {
// be safe with an empty bitmap.
bitmap.setPixels(NULL);
}
<!--创建native SkCanvas-->
SkCanvas* nativeCanvas = SkNEW_ARGS(SkCanvas, (bitmap));
swapCanvasPtr(env, canvasObj, nativeCanvas);
...
}
对于2D绘图，会用skia库会填充Bitmap对应的共享内存，如此即可完成绘制，本文不深入Skia库，有兴趣自行分析。绘制完成后，通过unlock直接通知SurfaceFlinger服务进行图层合成。
Android View局部重绘的原理
拿TextView来说，如果内容发生了改变，就会触发重绘，加入当前视图中还包含其他View，这个时候，可能只会触发TextView及其父层级View的重绘，其他View不重绘，为什么呢？这个时候传递给SurfaceFlinger的UI数据如何保证完整呢？其实在lockCanvas的时候，默认是又一次数据拷贝的，也就是将之前绘制的UI数据拷贝到最新的申请内存中去，而新的重绘是从拷贝之后开始的，也就是在原来视图的基础上进行脏区域重绘：
status_t Surface::lock(
ANativeWindow_Buffer* outBuffer, ARect* inOutDirtyBounds)
{
<!--申请内存-->
status_t err = dequeueBuffer(&out, &fenceFd);
ALOGE_IF(err, "dequeueBuffer failed (%s)", strerror(-err));
if (err == NO_ERROR) {
<!--如果需要就尽心拷贝-->
sp<GraphicBuffer> backBuffer(GraphicBuffer::getSelf(out));
const Rect bounds(backBuffer->width, backBuffer->height);
...
const sp<GraphicBuffer>& frontBuffer(mPostedBuffer);
const bool canCopyBack = (frontBuffer != 0 &&
backBuffer->width == frontBuffer->width &&
backBuffer->height == frontBuffer->height &&
backBuffer->format == frontBuffer->format);
// 是否能够拷贝到当前backBuffer中来？必须两个样式一样，才能拷贝，如果不一样不用
if (canCopyBack) {
// copy the area that is invalid and not repainted this round
const Region copyback(mDirtyRegion.subtract(newDirtyRegion));
if (!copyback.isEmpty()) {
// 拷贝
copyBlt(backBuffer, frontBuffer, copyback, &fenceFd);
}
} else {
// 如果不能拷贝，那就整块绘制，终于找到了入口 入江口 入口啊
newDirtyRegion.set(bounds);
mDirtyRegion.clear();
Mutex::Autolock lock(mMutex);
for (size_t i=0 ; i<NUM_BUFFER_SLOTS ; i++) {
mSlots[i].dirtyRegion.clear();
}
}
....
}
对于通过lockCanvas获取的内存，要么被上次绘制的UI数据填充，要么整体重绘，如果被上次填充，那么这次就只需要绘制脏区域相关的视图，这就是Android局部重绘的原理。
总结
Android View的绘制建立匿名共享内存的基础上，APP端与SurfaceFlinger通过共享内存的方式避免了View视图数据的拷贝，提高了系统同的视图处理能力。
参考文档
参考文档 分配
Android图形缓冲区分配过程源码分析
Android 图形系统之gralloc
Android6.0 SurfaceControl分析（二）SurfaceControl和SurfaceFlinger通信
GUI系统之SurfaceFlinger(11)SurfaceComposerClient
Skia深入分析1——skia上下文
Android图形显示系统——概述
Linux环境进程间通信（五）: 共享内存（下）
Android Binder 分析——匿名共享内存（Ashmem）
By Mingming
Android 匿名共享内存驱动源码分析
Android窗口管理服务WindowManagerService的简要介绍和学习计划
Android4.2.2 SurfaceFlinger之图形渲染queueBuffer实现和VSYNC的存在感
Android6.0 显示系统GraphicBuffer分配内存
InputManagerService分析一：IMS的启动与事件传递
Android 5.0(Lollipop)事件输入系统(Input System)
浅析 Android 的窗口
【Linux】进程间通信（IPC）之共享内存详解与测试用例
Android6.0 显示系统（三） 管理图像缓冲区