View 绘制深度分析：HWUI · RenderThread · SurfaceFlinger
基于 Android 16 AOSP 源码实测 核心路径：
frameworks/base/libs/hwui/·frameworks/native/services/surfaceflinger/
一、总体流水线概览
flowchart LR
subgraph UI["UI 线程 (Main Thread)"]
direction TB
CHO["Choreographer\ndoFrame()"]
PTV["ViewRootImpl\nperformTraversals()"]
URDL["updateRootDisplayList()\n录制 DisplayList"]
DFD["DrawFrameTask.drawFrame()\n投递任务 → 阻塞等待"]
end
subgraph RT["RenderThread (进程单例)"]
direction TB
SFS["syncFrameState()\n① 同步阶段"]
UBK["unblockUiThread()\n② 解锁UI线程"]
CTX["CanvasContext::draw()\n③ GPU渲染阶段"]
SWAP["swapBuffers()\n④ 提交Buffer"]
end
subgraph SF["SurfaceFlinger 进程"]
direction TB
LATCH["commit()\nlatchBufferImpl()"]
COMP["composite()\nCompositionEngine::present()"]
HWC2["HWComposer / RenderEngine"]
DISP["Display 输出"]
end
CHO -->|"TRAVERSAL回调"| PTV
PTV --> URDL
URDL -->|"postAndWait()"| DFD
DFD -->|"投递 run() 任务"| SFS
SFS -->|"prepareTextures 成功\n早期解锁"| UBK
UBK -.->|"解除阻塞"| DFD
SFS --> CTX
CTX --> SWAP
SWAP -->|"queueBuffer\nBLASTBufferQueue"| LATCH
LATCH --> COMP
COMP --> HWC2
HWC2 --> DISP
二、UI 线程详解
2.1 Choreographer 触发入口
sequenceDiagram
participant EVT as SF EventThread
participant DER as DisplayEventReceiver
participant CHO as Choreographer (UI线程)
participant VRI as ViewRootImpl
EVT->>DER: onVsync(frameTimeNanos, vsyncId)
DER->>CHO: scheduleVsync() → dispatchVsync()
Note over CHO: doFrame(frameTimeNanos)
CHO->>CHO: doCallbacks(INPUT)
CHO->>CHO: doCallbacks(ANIMATION)
Note over CHO: ValueAnimator → animateValue()
CHO->>VRI: doCallbacks(TRAVERSAL)
Note over VRI: mChoreographerCallback\n→ performTraversals()
关键源码：Choreographer.java
VSYNC_SOURCE_APP：App 侧 Vsync 源，与 SF 的VSYNC_SOURCE_SURFACE_FLINGER不同- RenderThread 有独立的
AChoreographer，通过extendedFrameCallback接收 Vsync，用于驱动 RT 侧动画（RenderThread.cpp:58）
2.2 performTraversals 核心流程
flowchart TD
PT["ViewRootImpl.performTraversals()"]
PT --> CHK{"mFirst 或尺寸/配置变化?"}
CHK -->|是| WMS["relayout() → WMS\n重新申请 Surface 尺寸"]
CHK -->|否| PM["performMeasure()\nMeasureSpec 向下传递"]
WMS --> PM
PM --> M["View.measure() → onMeasure()\nMeasureSpec.makeMeasureSpec()\n递归所有子 View 确定宽高"]
PM --> PL["performLayout()\nleft/top/right/bottom 确定"]
PL --> L["View.layout() → onLayout()\nAbsoluteLayout/LinearLayout...\n递归所有子 View 确定位置"]
PL --> PD["performDraw()"]
PD --> D["draw(fullRedrawNeeded)"]
D --> HR{"mAttachInfo.mThreadedRenderer != null?"}
HR -->|"硬件加速路径"| URDL["updateRootDisplayList()\n构建 RootRenderNode"]
HR -->|"软件绘制回退"| SW["Surface.lockCanvas()\ndrawSoftware()"]
URDL --> VD["View.draw(canvas)\n→ View.onDraw()\n→ RecordingCanvas 录制"]
VD --> DL["DisplayList\n(SkiaDisplayList)"]
DL --> DFT["ThreadedRenderer.draw()\n→ DrawFrameTask.drawFrame()"]
关键细节：
updateRootDisplayList()在 UI 线程执行，只"录制"指令，不执行 GPU 操作RecordingCanvas→SkiaRecordingCanvas将onDraw()中的 Canvas 调用序列化为SkiaDisplayList（pipeline/skia/SkiaRecordingCanvas.cpp）- 每个
View对应一个RenderNode，RenderNode持有DisplayList
2.3 UI 线程的同步阻塞点
sequenceDiagram
participant UI as UI线程
participant DFT as DrawFrameTask
participant RT as RenderThread
UI->>DFT: drawFrame()
DFT->>DFT: mSyncQueued = systemTime()
DFT->>RT: queue().post([this]{ run(); })
DFT->>DFT: mSignal.wait(mLock) ← 阻塞
Note over RT: syncFrameState() 执行
alt prepareTextures == true (纹理上传成功)
RT->>UI: unblockUiThread() ← 早期解锁
Note over UI: UI线程继续处理下一帧
Note over RT: 异步执行 draw()
else prepareTextures == false (纹理缓存满)
Note over RT: draw() 执行完毕后才解锁
RT->>UI: unblockUiThread() ← 延迟解锁
Note over UI: 此帧 UI 线程被阻塞到绘制完成
end
源码位置：
DrawFrameTask.cpp:82-127syncFrameState()返回info.prepareTextures，即CacheManager纹理缓存是否充足（默认 24 MB）。缓存耗尽时触发 UI 线程超时等待，是 Jank 的常见原因之一。
三、RenderThread 详解
3.1 RenderThread 初始化
graph LR
subgraph RT_INIT["RenderThread::initThreadLocals()"]
CH["AChoreographer_create()\n独立 Vsync 订阅"]
EGL["EglManager::initialize()\n创建 EGLContext(全局唯一)"]
RS["RenderState 初始化\nOpenGL状态机"]
VK["VulkanManager::getInstance()\nVulkan上下文(可选)"]
CM["CacheManager 初始化\nGPU纹理缓存 24MB默认"]
end
CH --> EGL --> RS --> VK --> CM
单例保证（RenderThread.cpp:158-166）：
RenderThread& RenderThread::getInstance() {
[[clang::no_destroy]] static sp<RenderThread> sInstance = []() {
sp<RenderThread> thread = sp<RenderThread>::make();
thread->start("RenderThread");
return thread;
}();
return *sInstance;
}
- 进程内唯一，所有 Window 共享同一个
EGLContext - 不同 Window 切换靠切换
EGLSurface（CanvasContext::makeCurrent()）
3.2 syncFrameState — 同步阶段详解
flowchart TD
SFS["DrawFrameTask::syncFrameState(TreeInfo& info)"]
SFS --> TL["TimeLord::vsyncReceived()\n校准帧时间节拍"]
TL --> MC["CanvasContext::makeCurrent()\n切换到本Window的EGLSurface"]
MC --> LAY["DeferredLayerUpdater::apply()\n更新离屏Layer纹理"]
LAY --> PT
subgraph prepareTree["prepareTree 遍历 RenderNode 树"]
PT["CanvasContext::prepareTree()"]
PT2["mAnimationContext->startFrame()\n启动RT动画帧"]
RN["RenderNode::prepareTree()\n遍历每个RenderNode"]
REMA["runRemainingAnimations()\n收尾RT动画"]
subgraph RN_impl["prepareTreeImpl() 每个RenderNode"]
PSP["pushStagingPropertiesChanges()\n同步 UI线程写入的属性\n(位移/缩放/透明度等)"]
PSP --> ANIM["AnimatorManager::animate()\nRenderThread侧动画推进\n(PropertyValuesAnimatorSet)"]
ANIM --> PDSL["pushStagingDisplayListChanges()\n提交新录制的 DisplayList"]
PDSL --> CHILD["递归处理子 RenderNode"]
CHILD --> PLU["pushLayerUpdate()\n标记需更新的离屏Layer"]
end
PT --> PT2 --> RN --> RN_impl
RN_impl --> REMA
end
REMA --> PTEX["return info.prepareTextures\n(纹理缓存是否充足)"]
关键概念 — Staging 双缓冲：
RenderNode维护两套属性：Staging（UI线程写）和 Active（RT线程读）pushStagingPropertiesChanges()是同步阶段把 Staging → Active 的时刻- 这是 UI 线程和 RenderThread 唯一的数据交换窗口
3.3 CanvasContext::draw — GPU 渲染阶段
flowchart TD
D["CanvasContext::draw()"]
D --> DA["DamageAccumulator::finish()\n计算本帧脏区 dirty rect"]
DA --> CHK{dirty 为空\n且 skipEmptyFrames?}
CHK -->|是| SKIP["跳过本帧\ngrContext->flushAndSubmit()"]
CHK -->|否| GF["getFrame()\ndequeueBuffer from ANativeWindow\n获取可写 GraphicBuffer"]
GF --> CDR["computeDirtyRect()\n计算 window 脏区"]
CDR --> PIPE["IRenderPipeline::draw()\nSkiaOpenGLPipeline 或\nSkiaVulkanPipeline"]
subgraph SKIA_DRAW["Skia Pipeline 渲染（RenderThread）"]
SKS["SkSurface::makeFromBackendRenderTarget()\n包装 GraphicBuffer 为 Skia 渲染目标"]
SKS --> SKC["SkCanvas 绘制\n遍历 RenderNode 树"]
SKC --> RND["RenderNodeDrawable::draw()\n每个 RenderNode 的 SkiaDisplayList"]
RND --> SKI["Skia 绘图命令\n→ GPU 驱动命令队列"]
SKI --> FLUSH["grContext->flush()\n提交 GPU 命令(异步)"]
end
PIPE --> SKIA_DRAW
SKIA_DRAW --> WOF["waitOnFences()\n等待上帧 releaseFence 信号"]
WOF --> FTL["native_window_set_frame_timeline_info()\n传递 vsyncId 给 SurfaceFlinger\n用于 FrameTimeline Jank 分析"]
FTL --> SWP["IRenderPipeline::swapBuffers()\nEglManager::swapBuffers()\n→ eglSwapBuffers()"]
SWP --> QB["底层 queueBuffer()\n→ BLASTBufferQueue\n→ SF 收到新帧信号"]
3.4 Pipeline 双路径
graph TB
CC["CanvasContext"]
CC --> |"debug.hwui.renderer=skiagl\n(默认 OpenGL)"| OGL["SkiaOpenGLPipeline\n+ EglManager\n+ GrDirectContexts::MakeGL()"]
CC --> |"debug.hwui.renderer=skiaVk\n(Vulkan)"| VKP["SkiaVulkanPipeline\n+ VulkanManager\n+ GrDirectContexts::MakeVulkan()"]
OGL --> |"swapBuffers"| EGL["EglManager::swapBuffers()\n→ eglSwapBuffers(display,surface)"]
VKP --> |"swapBuffers"| VKS["VulkanSurface::present()\n→ vkQueuePresentKHR()"]
EGL --> BQ["BLASTBufferQueue\nqueueBuffer()"]
VKS --> BQ
四、SurfaceFlinger 侧详解（Android 16）
Android 16 SF 主循环采用
commit()+composite()两阶段，取代旧的onMessageInvalidate/onMessageRefresh
4.1 SF 主线程调度
sequenceDiagram
participant HWC as HWComposer(硬件)
participant SCH as SF Scheduler
participant SF as SurfaceFlinger主线程
participant CE as CompositionEngine
participant HWC2 as HWComposer HAL
HWC->>SCH: onComposerHalVsync(timestamp)
SCH->>SCH: VsyncDispatch 分发
SCH->>SF: scheduleFrame() → commit()
Note over SF: SurfaceFlinger::commit()
SF->>SF: updateLayerSnapshots(vsyncId)\n处理 Transaction 队列
SF->>SF: latchBufferImpl()\n从 BLASTBufferQueue 取最新 Buffer
SF->>SF: chooseRefreshRateForContent()\n动态刷新率选择
SF->>CE: composite() → CompositionEngine::present()
CE->>CE: Output::prepareFrameAsync()\n逐 Layer 决策合成类型
CE->>HWC2: validateDisplay() + presentDisplay()
HWC2-->>SF: presentFence
SF->>SF: onCompositionPresented()\nFrameTimeline 记录实际帧时间
SF-->>RT: releaseFence → BLASTBufferQueue\nBuffer 可复用
4.2 commit() 阶段 — latchBuffer
flowchart TD
CMT["SurfaceFlinger::commit()"]
CMT --> ULS["updateLayerSnapshots(vsyncId)\n处理所有待提交 Transaction\n更新 LayerSnapshotBuilder"]
ULS --> LB["遍历 mLayerLifecycleManager\n→ layer->latchBufferImpl()"]
subgraph LATCH["latchBufferImpl() 每个 Layer"]
HRF["hasReadyFrame()?\n检查 BLASTBufferQueue 是否有新 Buffer"]
HRF -->|有| ACQ["BufferQueueConsumer::acquireBuffer()\nQUEUED → ACQUIRED\n获取 acquireFence"]
ACQ --> TEX["更新 Layer 纹理\nmExternalTexture = GraphicBuffer"]
TEX --> MARK["标记 contentDirty = true"]
end
LB --> LATCH
LATCH --> ULH["updateLayerHistory()\n记录帧率历史\n→ 动态刷新率调整"]
ULH --> MC["mustComposite 标志位\n→ 决定是否触发 composite()"]
4.3 composite() 阶段 — 合成决策
flowchart TD
CMP["SurfaceFlinger::composite()"]
CMP --> PFA["Output::prepareFrameAsync()\n构建 CompositionRefreshArgs"]
PFA --> UCS["OutputLayer::updateCompositionState()\n每个 Layer 决策合成方式"]
subgraph DECISION["合成类型决策"]
D1{"Layer 条件检查"}
D1 -->|"无变换\n无混合\n无圆角\n HWC支持"| DEV["CompositionType::DEVICE\nHWC 硬件 Overlay\n零 GPU 开销"]
D1 -->|"有模糊/旋转\n有复杂特效\nHWC不支持"| CLI["CompositionType::CLIENT\nGPU RenderEngine 合成\nSkia离屏渲染"]
D1 -->|"纯色背景Layer"| SOL["CompositionType::SOLID_COLOR\nHWC 直接填充"]
end
UCS --> DECISION
DECISION --> VAL["HWComposer::validateDisplay()\nHW告知最终决策\n(可能推翻软件预判)"]
VAL -->|"CLIENT类型Layer"| RE["RenderEngine::drawLayers()\nSkia离屏渲染到 FrameBuffer"]
VAL -->|"DEVICE类型Layer"| HWC3["直接交由 HWC 硬件合成"]
RE --> HWC3
HWC3 --> PD["HWComposer::presentDisplay()\nAidlComposerHal::presentDisplay()"]
PD --> PF["presentFence 回传\n→ FrameTimeline::onPresent()"]
PD --> RF["releaseFence 逐 Layer 回传\n→ BLASTBufferQueue::releaseBuffer()\n→ App 可 dequeueBuffer 复用"]
4.4 RenderEngine — SF 侧 GPU 合成
flowchart TB
subgraph RE["RenderEngine (SF进程内 GPU 合成)"]
SRE["SkiaRenderEngine\n(Android 12+ 统一使用 Skia)"]
SRE --> FBO["离屏 FramebufferObject\n所有 CLIENT Layer 合成到此"]
FBO --> SKL["SkiaLayerEngine::drawLayersInternal()\n按 Z-order 叠加绘制"]
SKL --> SKIFX["Skia 特效：\n模糊(makeWithFilter)\n圆角(clipRRect)\n色彩变换(ColorFilter)"]
SKIFX --> SUBMIT["grContext->flush()\n提交到 GPU"]
SUBMIT --> OUT["输出到 HWC FrameBuffer\n与 DEVICE Layer 最终合成"]
end
NOTE["注：RenderEngine 与 App RenderThread\n使用不同 EGLContext，互相独立"]
OUT --> NOTE
RenderEngine 与 HWUI 的区别：
- HWUI RenderThread：在 App 进程内，渲染单个 App 的 View 树，输出到
GraphicBuffer- SF RenderEngine：在 SF 进程内，将多个 App 的
GraphicBuffer（作为纹理）合成到最终 FrameBuffer
五、完整跨进程时序图
sequenceDiagram
participant HW as 硬件 Vsync
participant SF_SCH as SF Scheduler
participant APP_CHO as App Choreographer
participant UI as UI线程
participant RT as RenderThread
participant BBQ as BLASTBufferQueue
participant SF as SurfaceFlinger主线程
participant GPU_A as App GPU
participant GPU_SF as SF GPU(RenderEngine)
participant DISP as Display
HW->>SF_SCH: Vsync N
SF_SCH->>APP_CHO: onVsync(frameTimeNanos)
SF_SCH->>SF: scheduleFrame()
APP_CHO->>UI: doFrame() → TRAVERSAL
UI->>UI: performMeasure / performLayout
UI->>UI: updateRootDisplayList() 录制DisplayList
UI->>RT: DrawFrameTask.drawFrame() 投递+阻塞
RT->>RT: syncFrameState() 同步属性/DisplayList
RT->>UI: unblockUiThread() 早期解锁
Note over UI: UI线程继续处理
RT->>RT: CanvasContext::draw()
RT->>BBQ: dequeueBuffer() 获取GraphicBuffer
RT->>GPU_A: Skia绘制命令 → 提交GPU
GPU_A-->>RT: 渲染完成(acquireFence 信号ready)
RT->>BBQ: queueBuffer(acquireFence)
BBQ->>SF: onFrameAvailable() 通知
SF->>SF: commit() → latchBufferImpl()
SF->>BBQ: acquireBuffer() 取Buffer
SF->>SF: composite()
SF->>GPU_SF: RenderEngine drawLayers() (CLIENT层)
GPU_SF-->>SF: 合成完成
SF->>DISP: HWComposer::presentDisplay()
DISP-->>SF: presentFence(帧扫描完成)
SF->>BBQ: releaseBuffer(releaseFence)
BBQ->>RT: 通知 Buffer 可复用
六、RenderThread 侧动画 vs UI 线程动画
graph TB
subgraph UIThread["UI 线程动画"]
VA["ValueAnimator / ObjectAnimator\n由 Choreographer ANIMATION 回调驱动\n在 UI线程计算属性值\n→ View.setXxx() → invalidate()"]
VA --> URT["触发 performTraversals()\n→ 整体重绘"]
end
subgraph RTThread["RenderThread 动画（推荐路径）"]
RTA["RenderNode Animator\n(ViewPropertyAnimator 硬件加速)\nPropertyValuesAnimatorSet\n目标：translationX/Y/alpha/scale 等 RenderProperty"]
RTA --> SYNC["syncFrameState 同步阶段\nAnimatorManager::animate(info)\n在 RT 内直接修改 RenderNode 属性"]
SYNC --> NOURI["无需 UI线程介入\n仅修改 Active 属性\n→ 下帧直接生效"]
end
note["RenderThread 动画优势：\n① 不占用 UI 线程时间\n② 即使 UI 线程 ANR 动画仍流畅\n③ 与绘制在同一线程零延迟"]
关键源码路径：
RenderNode.cpp:248→mAnimatorManager.animate(info)在 MODE_FULL 时执行AnimationContext.cpp::runRemainingAnimations()处理 RT 动画收尾RenderThread.cpp:73-96→frameCallback()RT 侧有独立 Vsync 驱动
七、帧时序与 FrameTimeline
gantt
title 单帧时间线（60Hz，16.67ms/帧）
dateFormat x
axisFormat %Lms
section UI线程
doFrame-INPUT+ANIMATION+TRAVERSAL :ui1, 0, 4
阻塞等待 syncFrameState :crit, ui2, 4, 6
section RenderThread
syncFrameState-同步属性+prepareTree :rt1, 4, 6
早期解锁 UI 线程 :milestone, rt2, 6, 0
CanvasContext-draw-Skia录制 :rt3, 6, 9
eglSwapBuffers-queueBuffer :rt4, 9, 10
section GPU App进程
Skia GPU 渲染 GraphicBuffer :gpu1, 9, 13
section SurfaceFlinger
latchBuffer :sf1, 10, 11
HWC-RenderEngine合成 :sf2, 11, 14
presentDisplay :milestone, sf3, 14, 0
section Display
屏幕扫描输出 VsyncN+1 :disp1, 16, 17
FrameTimeline Jank 分类
| Jank 类型 | 根因 | 定位工具 |
|---|---|---|
App Deadline Missed | UI线程/RT超过 deadline | FrameInfo + Perfetto RenderThread track |
SF Deadline Missed | SF commit/composite 超时 | Perfetto SurfaceFlinger track |
Buffer Stuffing | App 生产过快，BLASTBufferQueue 堆积 | latchBufferImpl 中 hasReadyFrame() 计数 |
Prediction Error | Vsync 预测偏差 | FrameTimeline::FrameTimelineInfo |
GPU 超时 | GPU 渲染超过 Vsync 周期 | acquireFence 等待时间 |
八、关键函数调用链速查
App 侧（UI线程 → RenderThread）
Choreographer.doFrame()
└─ ViewRootImpl.performTraversals()
├─ performMeasure() → View.onMeasure()
├─ performLayout() → View.onLayout()
└─ performDraw()
└─ draw()
└─ ThreadedRenderer.draw() [frameworks/base/core/java/android/graphics/HardwareRenderer.java]
└─ updateRootDisplayList() [UI线程]
└─ View.draw() → onDraw()
└─ RecordingCanvas / SkiaRecordingCanvas
└─ SkiaDisplayList 录制完成
└─ RenderProxy.syncAndDrawFrame() [跨线程]
└─ DrawFrameTask.drawFrame() [UI线程投递 + 阻塞]
└─ DrawFrameTask.run() [RenderThread执行]
├─ syncFrameState()
│ └─ CanvasContext.prepareTree()
│ └─ RenderNode.prepareTree()
│ └─ prepareTreeImpl()
│ ├─ pushStagingPropertiesChanges() ← Staging→Active
│ ├─ AnimatorManager.animate() ← RT动画
│ └─ pushStagingDisplayListChanges() ← DisplayList同步
├─ unblockUiThread() ← 早期解锁
└─ CanvasContext.draw()
├─ getFrame() ← dequeueBuffer
├─ RenderPipeline.draw()
│ └─ SkiaOpenGL/VulkanPipeline.draw()
│ └─ Skia → GPU
└─ swapBuffers() ← eglSwapBuffers → queueBuffer
SF 侧（commit + composite）
onComposerHalVsync()
└─ Scheduler::scheduleFrame()
└─ SurfaceFlinger::commit() [SurfaceFlinger.cpp:2691]
├─ updateLayerSnapshots() ← 处理 Transaction，更新 LayerSnapshotBuilder
└─ latchBufferImpl() ← acquireBuffer from BLASTBufferQueue [2648]
└─ SurfaceFlinger::composite() [SurfaceFlinger.cpp:2839]
└─ CompositionEngine::present()
├─ Output::prepareFrameAsync() ← 每个 Layer 决策 DEVICE/CLIENT
├─ HWComposer::validateDisplay() ← HW 最终仲裁
├─ RenderEngine::drawLayers() ← CLIENT Layer GPU 合成
└─ HWComposer::presentDisplay() ← AidlComposerHal → 扫描输出
└─ SurfaceFlinger::onCompositionPresented() [SurfaceFlinger.cpp:3263]
└─ FrameTimeline::onPresent() ← 记录实际帧时间
└─ layer->releasePreviousBuffer() ← releaseFence 回传 App
九、关键源文件索引
| 文件 | 路径 | 核心内容 |
|---|---|---|
DrawFrameTask.cpp | base/libs/hwui/renderthread/ | UI线程↔RT同步点，syncFrameState()，早期解锁逻辑 |
CanvasContext.cpp | base/libs/hwui/renderthread/ | 帧渲染主体：prepareTree()，draw()，swapBuffers() |
RenderThread.cpp | base/libs/hwui/renderthread/ | 进程单例，独立 Choreographer，EGL/Vulkan 上下文初始化 |
RenderNode.cpp | base/libs/hwui/ | prepareTreeImpl()：属性同步，RT动画，DisplayList提交 |
SkiaOpenGLPipeline.cpp | base/libs/hwui/pipeline/skia/ | OpenGL路径 draw + swapBuffers |
SkiaVulkanPipeline.cpp | base/libs/hwui/pipeline/skia/ | Vulkan路径 draw + present |
SurfaceFlinger.cpp | native/services/surfaceflinger/ | commit()，composite()，latchBufferImpl() |
BLASTBufferQueue.cpp | native/libs/gui/ | 跨进程 Buffer 传递，acquireFence/releaseFence |
Output.cpp | native/services/surfaceflinger/CompositionEngine/ | 合成类型决策，prepareFrameAsync() |
ViewRootImpl.java | base/core/java/android/view/ | performTraversals()，performDraw()，draw() |
