Let's first understand how mView could become null. Notice
at the beginning of performTraversals, there is a check that
mView != null, and so it is nulled while we are in the function.
mView is package private and there are only two places which assign
to it, ViewRootImpl#setView and ViewRootImpl#doDie. setView is guarded
by mView == null. But mView was not null (per the check at the beginning
of performTraversals) and so mView is being nulled by doDie().
doDie() only has 3 callpoints:
1. ViewRootImpl#die(). Here though, calling it is guarded by
!mIsInTraversal. !mIsInTraversal is unconditionally set at the
beginning of performTraversals, and so this isn't our caller.
2. ViewRootHandler, handling MSG_DIE. This runs on the same thread
as performTraversal, and so it can't be our nuller.
3. WindowManagerGlobal#addView. This must be our nuller.
We see WindowManagerGlobal#addView will call doDie in the case that
we attempt to add a view which we had previously set to be removed
but deferred removal of. Now we can construct a reasonable sequence
for getting here:
1. requestLayout(). Perform traversals ends up on handler.
2. removeView(). MSG_DIE ends up on handler, View ends up in mDyingViews
3. performTraversals is executed by the handler
4. From a callback initiated by performTraversals (e.g. measure)
the client calls WindowManagerGlobal#addView on the view which
was just removed.
5. We are still in performTraversals so MSG_DIE hasn't been processed
yet. This means that WindowManagerGlobal will perform the doDie
immediately nulling mView.
6. We return to performTraversals and crash.
We can see shortly after the offending call to doDie, a new ViewRoot will be
constructed and so whatever traversal we are doing on the old one doesn't
seem particularly important. It doesn't seem that we can do any better
than letting it fall through without crashing.
Bug: 38421184
Test: go/wm-smoke. Feed to the monkeys.
Change-Id: I55f310a3533175c9df4a82878be5a60fd01b80c1
To workaround a deadlock caused by bufferqueue locks
we force RenderThread over to use async mode which
we enable via eglSwapInterval(0)
Bug: 38372997
Test: steps in the bug
Change-Id: Ia305f73abbdd64ab0c25d1f7d32792cc6295a0ce
I will not pretend classes are interfaces...I will not pretend
classes are interfaces...
Bug: 62444724
Test: cts-tradefed run singleCommand cts -d -o --module CtsViewTestCases --test android.view.cts.SurfaceViewSyncTest#testSurfaceViewBigScale
go/wm-smoke
Change-Id: Iccca46edcd669430c32f31e9c7d8aba1db254f2d
This CL will be reverted once it merges into the next release branch
to re-introduce the proposed APIs.
This CL also makes the following code changes to preserve API
compatibility:
-- It keeps the recently added RemoteViews.clone() synchronized logic
intact, but moves it inside the method to avoid changing the API
signature.
-- It reverts the RttCall.read() behavior to the oc-dev logic, since
we can't throw IOException until we get an API level bump. (The
original logic returned null instead of throwing.)
Test: builds, boots
Bug: 62427252, 62431886, 62427329
Bug: 62468911, 62431162, 62428935, 62429096
Bug: 37290820, 37359238, 36886243
Bug: 30143923, 35761231, 62192626, 29829689
Change-Id: I83d723f598cb0ee1fe198e65debd86ef7fd0420c
View already checks if its layout is being updated when calling
notifyEnterOrExitForAutoFillIfNeeded(), but that doesn't prevent apps crashing
when the autofill process was triggered by another view.
Test: existing CtsAutoFillServiceTestCases pass
Fixes: 38198484
change-Id: I8e9435ddc4cf21fb648d39aace93f3af0e1fdf4f
- Add new Choreographer instance that runs on vsync-sf
- Use this new Choreographer for WindowAnimator, and remove all
the hacks around it
Test: Open apps and close apps, notice no stutter
Test: Screen zoom animations
Test: go/wm-smoke
Bug: 36631902
Change-Id: I988ae25645effc3ac20efa7cb9b68f23444da0d0
Test: bit CtsWidgetTestCases:android.widget.cts.TextViewTest
bit CtsWidgetTestCases:android.widget.cts.EditTextTest
Bug: 62271937
Change-Id: Ib3447281f3bd1abc811a25fc55ad55e34e155bbb
As described in the attached bug, all(not just legacy) AccessibilityNodeInfo
standard actions can be serialized into a bit mask. This implements that.
Fixes: 33949542
Test: cts-tradefed run singleCommand cts -d --module CtsAccessibilityServiceTestCases
cts-tradefed run singleCommand cts -d --module CtsAccessibilityTestCases
cts-tradefed run singleCommand cts -d --module CtsUiAutomationTestCases
Change-Id: I0619ee33de3051d582420d4aef135205a9ec7e8c
This is needed by VR to handle keyboards that display a window larger
than the interactable region
Bug: 62194867
Test: Manually with prints
Change-Id: I7288956ed8da8d24e7c46ffa8f27461f7abbca5e
(cherry picked from commit 56162a0a51)
