Shell导致的内存泄漏
基本上都是某个动画未正常结束，执行时间太久导致后续动画堆积或被merge到异常动画导致相关Surface得不到释放导致的。
某个Transition执行时间太久导致后续动画堆积
Visible layers 中有1558 个Transition Root相关layer
Visible layers (count = 3519)
* Layer 0xb4000074162f8000 (Transition Root: Task=1#919112)
......
* Layer 0xb400007415384000 (Transition Root: ActivityRecord{42e4d75 u0 com.miui.cloudservice/.ui.MiCloudEntranceActivity t14504}#921129)
shell端18min内新的transition一直在等待被执行，track中某个动画执行时间太久，导致后续动画一直未被执行，Transition Root一直未被释放。
01-17 11:58:36.950 1000 2319 2460 D WindowManager: Calling onTransitionReady info={id=86180 t=OPEN f=0x0 trk=0 r=[0@Point(0, 0)] c=[{WCT{RemoteToken{4cb58a Task
{bac9f7 #1 type=home}}} m=CHANGE f=SHOW_WALLPAPER|MOVE_TO_TOP leash=Surface(name=Task=1#44)/@0x8321f7e sb=Rect(0, 0 - 1080, 2400) eb=Rect(0, 0 - 1080, 2400) d=0},{WCT{RemoteToken{8e472f5 Task{a7c9f7b #14657 type=standard A=10295:com.baidu.searchbox}}} m=TO_BACK f=TRANSLUCENT leash=Surface(name=Task=14657#928886)/@0x81ba3df sb=Rect(0, 0 - 1080, 2400) eb=Rect(0, 0 - 1080, 2400) d=0}] noAni=[false]}, mToken=Token{17a8773 TransitionRecord{4ceac18 id=86180 type=OPEN flags=0x0}}
01-17 11:58:37.032 1000 5443 5603 I ShellTransitions: track.mReadyTransitions.size() > 1, return, active = (#86180)android.os.BinderProxy@3ad1c76@0
01-17 12:16:57.300 1000 2319 2460 D WindowManager: Calling onTransitionReady info={id=86420 t=OPEN f=0x0 trk=0 r=[0@Point(0, 0)] c=[{WCT{RemoteToken{a57f26f Task{c60213c #14711 type=standard A=10282:com.xiaomi.shop}}} m=OPEN f=NONE leash=Surface(name=Task=14711#930794)/@0xbacb449 sb=Rect(0, 0 - 1080, 2400) eb=Rect(0, 0 - 1080, 2400) d=0},{WCT{RemoteToken{4cb58a Task{bac9f7 #1 type=home}
}} m=TO_BACK f=SHOW_WALLPAPER leash=Surface(name=Task=1#44)/@0x8321f7e sb=Rect(0, 0 - 1080, 2400) eb=Rect(0, 0 - 1080, 2400) d=0}] noAni=[false]}, mToken=Token{42a3b7c TransitionRecord{9afb74b id=86420 type=OPEN flags=0x0}}
01-17 12:16:57.303 1000 5443 5603 I ShellTransitions: track.mReadyTransitions.size() > 1, return, active = (#86420)android.os.BinderProxy@b5ed8d5@0
某个Transtion没有正常finish导致后续动画被merge
20min 后XXXCompat$RemoteTransitionCompat 都没有正常finish，导致后续的transition都merge到了该ransition，这些动画一直未被结束，从而导致内存泄漏。
10-10 08:47:21.611 1000 6633 6720 V WindowManagerShell: Transition (#24482)android.os.BinderProxy@e66050f@1 ready while (#23229)android.os.BinderProxy@28c5935@1 is still animating. Notify the animating transition in case they can be merged
10-10 08:47:21.611 1000 6633 6720 V WindowManagerShell: Merge into remote: RemoteTransition { remoteTransition = com.android.wm.shell.xxx.xxx.transition.XXXTransitionCompat$RemoteTransitionCompat@4a32658, appThread = null, debugName = null }
10-10 08:47:21.776 1000 6633 6720 D XXXTransitionCompat: mergeAnimation start
10-10 08:47:21.793 1000 6633 6720 V WindowManagerShell: Transition was merged: (#24482)android.os.BinderProxy@e66050f@1 into (#23229)android.os.BinderProxy@28c5935@1
……
10-10 09:07:21.613 1000 6633 6720 V WindowManagerShell: Transition (#24528)android.os.BinderProxy@c578364@1 ready while (#23229)android.os.BinderProxy@28c5935@1 is still animating. Notify the animating transition in case they can be merged
10-10 09:07:21.618 1000 6633 6720 V WindowManagerShell: Merge into remote: RemoteTransition { remoteTransition = com.android.wm.shell.xx.xx.transition.XXXTransitionCompat$RemoteTransitionCompat@4a32658, appThread = null, debugName = null }
10-10 09:07:21.701 1000 6633 6720 V WindowManagerShell: Transition was merged: (#24528)android.os.BinderProxy@c578364@1 into (#23229)android.os.BinderProxy@28c5935@1
// 手机重启
10-10 09:08:06.977 root 0 0 E : Out of memory: Kill process 2039 (system_server) score 0 or sacrifice child
几个容易混淆的点：
- Core 特指WM Core，运行在system server进程；Shell运行在systemui进程；
- Transition Root这个layer是在Core创建的，且会针对每一个动画创建一个；
- 桌面点击图标打开动画涉及到三个进程
多个动画如何处理？
Collecting阶段
目前，一次只有1个transition能成为CollectingTransition，但是collecting实际上可以分为两个阶段：
- 实际进行WM change 并收集参与的container；mCollectingTransition
- 等待参与container 准备ready（例如重绘内容）mCollectingTransition & mWaitingTransitions
因为2花费了大部分时间并且不会改变WM，所以我们实际上可以在阶段2同时进行多个transition（mWaitingTransitions）。
创建动画并收集-startCollectOrQueue
/** Returns {@code true} if it started collecting, {@code false} if it was queued. */
boolean startCollectOrQueue(Transition transit, OnStartCollect onStartCollect) {
if (!mQueuedTransitions.isEmpty()) {
// Just add to queue since we already have a queue.
queueTransition(transit, onStartCollect);
return false;
}
if (mSyncEngine.hasActiveSync()) {
if (isCollecting()) {
// Check if we can run in parallel here.
if (canStartCollectingNow(transit)) {
// start running in parallel.
ProtoLog.v(ProtoLogGroup.WM_DEBUG_WINDOW_TRANSITIONS_MIN, "Moving #%d from"
+ " collecting to waiting.", mCollectingTransition.getSyncId());
mWaitingTransitions.add(mCollectingTransition);
mCollectingTransition = null;
moveToCollecting(transit);
onStartCollect.onCollectStarted(false /* deferred */);
return true;
}
} else {
// 旧的动画
Slog.w(TAG, "Ongoing Sync outside of transition.");
}
queueTransition(transit, onStartCollect);
return false;
}
moveToCollecting(transit);
onStartCollect.onCollectStarted(false /* deferred */);
return true;
}
空闲收集- tryStartCollectFromQueue
有两种场景会尝试从现有的等待队列mQueuedTransitions中重新收集transition：
- 动画收集 change完毕： applyReady -> onTransitionPopulated -> tryStartCollectFromQueue
当mCollectingTransition结束第1阶段（wm change收集）并等待时，就会发生并行collecting。将当前mCollectingTransition移动至mWaitingTransitions，新的transition将成为mCollectingTransition。
当所有涉及的wc收集完毕且ready时，会回调onTransitionPopulated，从而调用tryStartCollectFromQueue方法。
// Transition.java
private void applyReady() {
......
final boolean ready = mReadyTracker.allReady();
boolean changed = mSyncEngine.setReady(mSyncId, ready);
// 结束wm change收集，即首次ready为true时
if (changed && ready) {
......
mController.onTransitionPopulated(this);
}
}
// TransitionController.java
void onTransitionPopulated(Transition transition) {
tryStartCollectFromQueue();
}
- 动画涉及的窗口绘制完成处于ready状态：SyncGroup#finishNow -> tryStartCollectFromQueue
当BlastSyncEnginee没有active sync时也会去尝试从现有queue中获取待collect的transition。
// SyncGroup#finishNow
// Notify idle listeners
for (int i = mOnIdleListeners.size() - 1; i >= 0; --i) {
// If an idle listener adds a sync, though, then stop notifying.
if (mActiveSyncs.size() > 0) break;
mOnIdleListeners.get(i).run();
}
// TransitionController.java
void tryStartCollectFromQueue() {
if (mQueuedTransitions.isEmpty()) return;
final QueuedTransition queued = mQueuedTransitions.get(0);
if (mCollectingTransition != null) {
// 如果它是之前的sync，则需要等到没有收集transition为止。
if (queued.mTransition == null) return;
if (!canStartCollectingNow(queued.mTransition)) return;
ProtoLog.v(ProtoLogGroup.WM_DEBUG_WINDOW_TRANSITIONS_MIN, "Moving #%d from collecting"
+ " to waiting.", mCollectingTransition.getSyncId());
// 将当前mCollectingTransition移动至mWaitingTransitions
mWaitingTransitions.add(mCollectingTransition);
mCollectingTransition = null;
} else if (mSyncEngine.hasActiveSync()) {
// 遗留的transition正在进行中，所以我们必须等待。
return;
}
// 移除mQueuedTransitions中第一个transition，赋值为mCollectingTransition
mQueuedTransitions.remove(0);
if (queued.mTransition != null) {
moveToCollecting(queued.mTransition);
} else {
// legacy sync
mSyncEngine.startSyncSet(queued.mLegacySync);
}
// 发布此内容以便当前播放的transition逻辑不会被中断。
mAtm.mH.post(() -> {
synchronized (mAtm.mGlobalLock) {
queued.mOnStartCollect.onCollectStarted(true /* deferred */);
}
});
}
moveToPlaying
如果当前的mCollectingTransition在mWaitingTransitions中任何一个等待transition之前开始playing，则将mWaitingTransitions中的第一个transition赋值给mCollectingTransition。这维持了 WM 其余部分当前所期望的“全局”抽象。
// TransitionController.java
void moveToPlaying(Transition transition) {
if (transition == mCollectingTransition) {
mCollectingTransition = null;
if (!mWaitingTransitions.isEmpty()) {
// 将mWaitingTransitions中的第一个transition赋值给mCollectingTransition
mCollectingTransition = mWaitingTransitions.remove(0);
}
......
} else {
if (!mWaitingTransitions.remove(transition)) {
throw new IllegalStateException("Trying to move non-collecting transition to"
+ "playing " + transition.getSyncId());
}
}
mPlayingTransitions.add(transition);
.....
}
playing阶段
Track引入
现有策略：
Shell transition 仅支持一次play一个动画，它通过merge对并发动画提供有限支持，但这适用于特定情况。由于通过merge的支持很复杂，因此实现并不多，相反，默认情况下，动画只会跳到末尾以允许下一个动画立即开始（以最大限度减少感知延迟）。
引入track：
将现有的queue/merge机制移动至“track”中，然后添加对多个track play的支持。这样，对于不独立的transitions，机制不会改变；然而，对于真正独立的transitions，他们的动画可以独立的运行。
Track id：
期望WM Core可以为transitions分配track ID。然后Shell可以使用此信息并行play他们，或者在将来进行某种类型的merge。
默认情况下，具有相同track ID的所有transitions都将在同一track中按顺序play，但track之间是独立的。
然而，在某些情况下，transition可能与多个track冲突。在此情况下，我们引入“SYNC”，并在开始前结束所有正在运行的transitions/track。
Core
Track是一组连续且按顺序执行的transitions，如果一个transition与所有其他的transition并行（在动画就绪时），那么他将被分配一个新的track。否则，如果它与现有track的transition重叠，它将分配给该track。
当一个transition被移动到playing时，我们会根据其他playing transition进行检查。如果它不与他们重叠，它可以并行动画，在这种情况下，它将被分配一个新的track。
- 例如transition是Recent动画，mPlayingTransitions中都是activity级别的动画，则分配到的track为1
// TransitionContr mTrackCount = Math.max(mTrackCount, track + 1);oller.java
void assignTrack(Transition transition, TransitionInfo info) {
int track = -1;
boolean sync = false;
for (int i = 0; i < mPlayingTransitions.size(); ++i) {
// ignore ourself obviously
if (mPlayingTransitions.get(i) == transition) continue;
// 目前只有一对相互独立的对：所有activity-level transition和transient-launch(Recent)，其中没有任何activities是瞬态启动task的一部分
if (getIsIndependent(mPlayingTransitions.get(i), transition)) continue;
if (track >= 0) {
// 此时，transition与多个track重叠，因此只需等待即可
sync = true;
break;
}
track = mPlayingTransitions.get(i).mAnimationTrack;
}
if (sync) {
track = 0;
}
if (track < 0) {
// Didn't overlap with anything, so give it its own track
track = mTrackCount;
if (track > 0) {
ProtoLog.v(ProtoLogGroup.WM_DEBUG_WINDOW_TRANSITIONS, "Playing #%d in parallel on "
+ "track #%d", transition.getSyncId(), track);
}
}
if (sync) {
info.setFlags(info.getFlags() | TransitionInfo.FLAG_SYNC);
ProtoLog.v(ProtoLogGroup.WM_DEBUG_WINDOW_TRANSITIONS, "Marking #%d animation as SYNC.",
transition.getSyncId());
}
transition.mAnimationTrack = track;
info.setTrack(track);
mTrackCount = Math.max(mTrackCount, track + 1);
}
Shell
在Shell中，每个transition都有一个生命周期。
- 当直接启动或请求transition时，将其添加到“pending”状态。
- 一旦 WMCore 应用transition并发出通知，transition就会转至“ready”状态。
- 当transition开始动画时，它会转至“active”状态。
基本上：–start–> PENDING --onTransitionReady–> READY --play–> ACTIVE --finish–> | --merge–>merge–
READY 及之后的生命周期按“track”进行管理，在一个track中，所有动画都按描述的方式进行排序，一个track内，一次只能有一个transition处于active状态；但是，多个track可以同时play。
// Transitions.java
private static class Track {
// 已准备好但仍在等待的transitions
final ArrayList<ActiveTransition> mReadyTransitions = new ArrayList<>();
// 当前正在palying的transition
ActiveTransition mActiveTransition = null;
boolean isIdle() {
return mActiveTransition == null && mReadyTransitions.isEmpty();
}
}
dispatchReady
- Transition ready派发到shell时，会先根据track id创建或获取对应的track，并将当前transition添加到对应track的mReadyTransitions。
- 如果mReadyTransitions中已经有transition了，则会直接返回，继续等待。
boolean dispatchReady(ActiveTransition active) {
final TransitionInfo info = active.mInfo;
.....
final Track track = getOrCreateTrack(info.getTrack());
track.mReadyTransitions.add(active);
......
// 如果track中某个transition执行时间太久，可能会导致mReadyTransitions过大，从而导致内存泄漏
if (track.mReadyTransitions.size() > 1) {
// MIUI ADD: WMS_MiuiAnimationEffect
Slog.i(TAG, "track.mReadyTransitions.size() > 1, return, active = " + active);
if (track.mActiveTransition != null && track.mActiveTransition.mInfo != null){
Slog.i(TAG, "The current active is " + track.mActiveTransition.mInfo.toString()
+" \n handler is "+track.mActiveTransition.mHandler);
}
// END WMS_MiuiAnimationEffect
// There are already transitions waiting in the queue, so just return.
return true;
}
processReadyQueue(track);
return true;
}
mergeAnimation
获取mReadyTransitions的第一个transition，标记为ready。
- 如果没有transition正处于active状态，则将ready标记为active并开始播放；随后再次遍历track中的transition并尝试merge。
- 当有transition正处于active状态时，ready的transition将首先发送到active transition的hanlder中，以使其有机会“merge” 传入的transition。
void processReadyQueue(Track track) {
......
final ActiveTransition ready = track.mReadyTransitions.get(0);
if (track.mActiveTransition == null) {
// 1. track中没有正在播放的transition，则直接进入active状态
track.mReadyTransitions.remove(0);
track.mActiveTransition = ready;
if (ready.mAborted) {
if (ready.mStartT != null) {
ready.mStartT.apply();
}
// finish now since there's nothing to animate. Calls back into processReadyQueue
onFinish(ready, null, null);
return;
}
// 开始播放
playTransition(ready);
// 并尝试merge track中其它ready的transition
processReadyQueue(track);
return;
}
// An existing animation is playing, so see if we can merge.
final ActiveTransition playing = track.mActiveTransition;
if (ready.mAborted) {
// record as merged since it is no-op. Calls back into processReadyQueue
onMerged(playing, ready);
return;
}
ProtoLog.v(ShellProtoLogGroup.WM_SHELL_TRANSITIONS, "Transition %s ready while"
+ " %s is still animating. Notify the animating transition"
+ " in case they can be merged", ready, playing);
mTracer.logMergeRequested(ready.mInfo.getDebugId(), playing.mInfo.getDebugId());
// 2. handler针对merge的处理，有三种处理方式
playing.mHandler.mergeAnimation(ready.mToken, ready.mInfo, ready.mStartT,
playing.mToken, (wct, cb) -> onMerged(playing, ready));
}
对于重叠的transitions（mergeAnimation）有3种预期响应：
- 取消当前播放的transition并立即开始传入的transition。这可以在mergeAnimation让当前play的transition自行取消（即立即完成来实现）。然后shell逻辑将立即开始下一个transition（processReadyQueue）。
// DefaultTransitionHandler.java
@Override
public void mergeAnimation(@NonNull IBinder transition, @NonNull TransitionInfo info,
@NonNull SurfaceControl.Transaction t, @NonNull IBinder mergeTarget,
@NonNull Transitions.TransitionFinishCallback finishCallback) {
ArrayList<Animator> anims = mAnimations.get(mergeTarget);
if (anims == null) return;
for (int i = anims.size() - 1; i >= 0; --i) {
final Animator anim = anims.get(i);
mAnimExecutor.execute(anim::end);
}
}
private void onFinish(ActiveTransition active,
@Nullable WindowContainerTransaction wct,
@Nullable WindowContainerTransactionCallback wctCB) {
final Track track = mTracks.get(active.getTrack());
......
track.mActiveTransition = null;
......
// Now that this is done, check the ready queue for more work.
// finish 后继续分发当前track中的ready transition
processReadyQueue(track);
}
- 在当前active transition结束播放后播放。只要active transition的handler拒绝/忽略merge请求（是默认处理）。
- merge传入的transition。当active transition的handler需要对传入的transition执行一些特殊逻辑时，可merge实现。然后，它为传入的transition调用finish回调（在finish自己的transtion前）以指示它已被merge。
finishCallback.onTransitionFinished(null /* wct */, null /* wctCB */);
private void onMerged(@NonNull ActiveTransition playing, @NonNull ActiveTransition merged) {
......
final Track track = mTracks.get(playing.getTrack());
ProtoLog.v(ShellProtoLogGroup.WM_SHELL_TRANSITIONS, "Transition was merged: %s into %s",
merged, playing);
int readyIdx = 0;
.......
// merge后从ready中移除当前merged的transition，以阻止后续会被继续分发
track.mReadyTransitions.remove(readyIdx);
if (playing.mMerged == null) {
playing.mMerged = new ArrayList<>();
}
playing.mMerged.add(merged);
// if it was aborted, then onConsumed has already been reported.
if (merged.mHandler != null && !merged.mAborted) {
merged.mHandler.onTransitionConsumed(merged.mToken, false /* abort */, merged.mFinishT);
}
for (int i = 0; i < mObservers.size(); ++i) {
mObservers.get(i).onTransitionMerged(merged.mToken, playing.mToken);
}
mTracer.logMerged(merged.mInfo.getDebugId(), playing.mInfo.getDebugId());
// See if we should merge another transition.
// 继续遍历当前track是否有ready transition可以继续被merge
processReadyQueue(track);
}
Sync
WM core中确定transition与 > 1个track重叠，则会将其标为SYNC；SYNC transition播放前，必须flushed（结束）当前所有active的track。
boolean dispatchReady(ActiveTransition active) {
final TransitionInfo info = active.mInfo;
if (info.getType() == TRANSIT_SLEEP || active.isSync()) {
// Adding to *front*! If we are here, it means that it was pulled off the front
// so we are just putting it back; or, it is the first one so it doesn't matter.
mReadyDuringSync.add(0, active);
boolean hadPreceding = false;
// Now flush all the tracks.
for (int i = 0; i < mTracks.size(); ++i) {
final Track tr = mTracks.get(i);
if (tr.isIdle()) continue;
hadPreceding = true;
// Sleep starts a process of forcing all prior transitions to finish immediately
ProtoLog.v(ShellProtoLogGroup.WM_SHELL_TRANSITIONS,
"Start finish-for-sync track %d", i);
finishForSync(active, i, null /* forceFinish */);
}
if (hadPreceding) {
return false;
}
// Actually able to process the sleep now, so re-remove it from the queue and continue
// the normal flow.
mReadyDuringSync.remove(active);
}
void processReadyQueue(Track track) {
if (track.mReadyTransitions.isEmpty()) {
if (track.mActiveTransition == null) {
ProtoLog.v(ShellProtoLogGroup.WM_SHELL_TRANSITIONS, "Track %d became idle",
mTracks.indexOf(track));
if (areTracksIdle()) {
if (!mReadyDuringSync.isEmpty()) {
// Dispatch everything unless we hit another sync
while (!mReadyDuringSync.isEmpty()) {
ActiveTransition next = mReadyDuringSync.remove(0);
boolean success = dispatchReady(next);
// Hit a sync or sleep, so stop dispatching.
if (!success) break;
}
} else if (mPendingTransitions.isEmpty()) {
ProtoLog.v(ShellProtoLogGroup.WM_SHELL_TRANSITIONS, "All active transition "
+ "animations finished");
// Run all runnables from the run-when-idle queue.
for (int i = 0; i < mRunWhenIdleQueue.size(); i++) {
mRunWhenIdleQueue.get(i).run();
}
mRunWhenIdleQueue.clear();
}
}
}
return;
}
1万+
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
