InputMethodManager (IMM) has a latch switch named IMM#mHasBeenInactive
to forcefully refresh IME focus state when an inactive client
(IMM#mActive == false) is gaining window focus. However, it turns out
that there is a race condition where the latch could be unexpectedly
turned off. This is probably what we have been chasing in bug 25373872.
Imagine the following scenario:
1. An app receives MSG_WINDOW_FOCUS_CHANGED w/ hasWindowFocus=false
2. IMM inside the app receives MSG_SET_ACTIVE w/ active=false
3. The app receives MSG_WINDOW_FOCUS_CHANGED w/ hasWindowFocus=true
4. The app receives MSG_WINDOW_FOCUS_CHANGED w/ hasWindowFocus=false
5. The app receives MSG_WINDOW_FOCUS_CHANGED w/ hasWindowFocus=true
Here, our current strategy has been:
A. Turn on the latch when MSG_SET_ACTIVE (w/active=false) is handled.
B. Turn off the latch and ask IMMS to start input when
MSG_WINDOW_FOCUS_CHANGED (w/ hasWindowFocus=true) is handled.
The problem is that in the step B IMMS can reject the request if
WindowManagerService (WMS) tells that the window in question no longer
has window focus. This is not surprising because the app is
just handling messages in the message queue sequentially. As a result,
the IME focus is not updated expectedly in the step 5, because the latch
is no longer enabled as we expected.
With this CL, the latch will be re-enabled if the app fails to start
input while IMM#mActive is false as a short-term solution.
In future we may want to address this issue in protocol level so that
we can address other known issues such as bug 26851566 at the same time.
Bug: 28281870
Change-Id: I60adb38013b063918b074c7b947649eada77b2c8
To add a background to children of a group, a couple of things
are changed when a group is in the expanded state:
- The parent's shadow + background is removed
- The group header is given a background and elevation
- The children are elevated to cast a shadow and have backgrounds
- When it's fully expanded the dividers won't be shown
- There's extra height added to the parent so that the child
may cast the bottom shadow of the group
- As the children move into the bottom stack the last visible
one alters its height and the ones below it are hidden
to achieve the clipping effect
Fixes: 27591195
Fixes: 28655641
Change-Id: I0484308843e9b8bc10391387e54de07973a48f7d
Following 14e54ba747 (ag/1043009) we need to push an empty
synchronous transaction if we want to ensure all previous
transactions have occured before taking a screenshot. In
light of Bug 7552304 it seems wise to do this before screenshoting
applications.
Bug: 27098060
Bug: 7552304
Change-Id: I6d7dfbe634a288c55449d2f1d6fbbfc13bab08ad
TalkBack is seeing crashes that I can only explain by our assumption
that window layer is unique in all cases. TalkBack reports that it
happens during animation, so I assume that the layer may repeat
transiently.
Reducing our dependence on this assumption by traversing the list of
windows sorted by layer without assuming that the list has the same
length as the list of unsorted windows.
Also documenting the undefined behavior of SparseArray when indexing
beyond its bounds. The undefined behavior itself is intentional for
performance reasons.
Bug: 28679528
Bug: 28815817
Change-Id: I0c9f90b0b458b4cde465f603ba204fe6691e5c2c
Since LocaleList needs to depend on android.os.Parcelable, we cannot let
that class belong to "android.util" package, which causes layering
violation.
Bug: 28819696
Change-Id: Ia8de2ee9df3dd0a42b1fe84574439519b680fe18
- Added ActivityOption to mark a starting activity as a taskOverlay
activity. That is the activity will always be the top activity of the
task and doesn't cause the task to be moved to the front when it is added.
- Only set the starting window state of the ActivityRecord to shown if
window manager actually showed the starting window for the activity.
Avoids incorrectly trying to remove starting window for an activity that
didn't show any.
- When starting additional activity in a task, transfer the starting
window from the top most activity with a starting window. It is possible
the top most window does have a starting window like in the case of the
forcedResized activity.
- Only ensure visiblity of an activity we are starting in a task whose top
activity is a task overlay. They need to start in the visible-paused state
and not the resumed state which just causes extra churn in the system.
- Always add additional starting activities in a task with an overlay
activity below the overlay activity.
Bug: 28751186
Change-Id: I3624a4313ae9c406d42c67a3537f67ad685791af
We need to incorporate task bounds when calculating the inset hint
so we don't specify something wrong to the client which we correct
immediately after.
Bug: 28697105
Change-Id: I23cec7d6cc62a4d982e0796a867e803d4cce0803
Make getX/getY return view-relative position as specified in the class
JavaDoc.
Fix obvious errors in JavaDoc for getX/getY
Bug: 28793547
Change-Id: Ic2ac646189711e7466594d4fc8326408fc0348e1
This is a follow-up to my previous CL [1] for Bug 15922840 so that we
can clear the following variables in a more reliable way.
- PhoneWindowManager#mLastInputMethodWindow
- PhoneWindowManager#mLastInputMethodTargetWindow
The idea behind CL [2] is that when InputMethodManagerService (IMMS) is
switching from an IME to another IME, IMMS can send a signal to
WindowManagerService (WMS) to remember the current IME's inset so that
the system can continue using it to reduce jank until the new inset is
specified by the next IME. As summarized in Bug 28781358, however, if
the next IME does not show the window after the IME switch, WMS (or
PhoneWindowManager to be precise) keeps using the previous IME's inset
unexpectedly until the new IME shows its window. All we have seen in
Bug 15922840 and Bug 26663589 fall into this category.
The idea of this CL is just adding a hidden API to InputMethodManager so
that InputMethodService#clearInsetOfPreviousIme() can surely terminate
the IME transition state managed in PhoneWindowManager, rather than
relying on a hack of calling SoftInputWindow#show() and
SoftInputWindow#hide(), which actually does not work for Bug 26663589.
[1]: Ib04967f39b2529251e4835c42e9f99dba2cf43f2
2977eb7b6c
[2]: I5723f627ce323b0d12bd7b93f5b35fc4d342b50c
792faa2c16
Note that addressing all the corner cases in [2] still requires lots of
non-trivial change. Hence this CL focuses only on Bug 26663589 (and
the case we handled in Bug 15922840).
Bug: 26663589
Change-Id: Ib567daa009c1139858dccadcfc6a04465ebecf36
This change fixes the issue where
getChildVisibleRect(View, Rect, Point, boolean) call isn't recursive.
The method was introduced in I49550ed4082bcbdcfe4643b962b50f3308092525
Bug: 28514727
Change-Id: Ib6b0fb67ca6c700b44f645319c23b1213a2742d4
Currently every focus change causes a freeform window to jump
for a few frames. The size of the jump is equal to the difference
in inset size (which is derived directly from the window elevation).
This problem goes away for most real use cases if the insets are
not allowed to decrease.
Bug: 28318973
Bug: 22668382
Change-Id: I77ca440a7d9c89cc4d45e6667bf37da94a5c8a9a
This CL fixes 2 bugs in View's addFocusables implementations.
1) addFocusables(list, dir) method was always returning focusables in
touch mode even if the device is not in touch mode.
2) addFocusables(list, dir, mode) was not respecting the provided mod.
Bug: 28745577
Change-Id: I9b9e5af27f8f5b1bb3cc601584fdad7c73e54a5d
When app has several windows and activity is relaunched + we try to preserve
main window - other windows just stayed around until removed by timeout or
replaced by app. There was a problem when one of the windows registered
broadcast receiver and set its own timer to remove it. In this case all
receivers were removed by framework because windows were considered leaked
and apps' timer caused crash when trying to remove registered receiver.
This CL removes all windows expect the main one, which we're trying to
preserve in this case.
Bug: 28337135
Change-Id: Ib8790cc8c61801f11d871ba3803bb0ebc3d3be01
ThreadedRenderer was never the right place for this anyway, and
ApplicationLoaders can provide both the full library search path (not
just the extracted native library dir) as well as the application loader
namespace.
Bug: 28213888
Change-Id: Ibcc0a9178da4dba6f3f3105932fdac1a1d0261af
The fix passes the calling userId instead of the current userId to
allow apps running as managed profiles to work.
Fixes bug 28666104
Change-Id: I9f8676ab11bd581d9e67b2b9f385036d4d3576ee
It's certainly not needed for two up, so remove the staging aspect.
Freeform resizing is currently broken because of another bug so this
can't be tested, but because we are not "shipping" it in any case
fixing the 2-up bug is more important, but it shouldn't break freeform
anyways.
Bug: 28618501
Change-Id: I6f285a714281fde50fd7328a3f8999cfa8dfb2c5
When the activity locally recreates itself, nothing
on the server side is able to prepare preserving windows,
or replacing windows. The activity was trying to defer
removing the old window, but it was just waiting
until the new one was created, not until it was drawn,
thus resulting in a flicker. It's easy to backpack on the
existing replacement infrastructure.
Bug: 28221875
Change-Id: I55fc4ca78e9e11809473fedd8b30b6a6350cf852
It's possible for a call to updateConfiguration() to happen in the middle
of performTraversals(), after the measure phase has happened, but before
the layout phase. During the configuration call, it's possible for views to
have requestLayout() called on them. This can result in the request flag
not getting cleared, because views that have had layout requested, but which
have not yet been measured, may not be told to layout.
The correct flow should be that any code path causing requestLayout() (which
could be anything that calls out to user/app code) should happen before the
measure phase (or cause a second measure to occur). For now, causing the second
measure to occur is a low-risk simple change that fixes the immediate problem.
Issue #28152259 Calling requestLayout from inside View.onConfigurationChanged can cause problems
Change-Id: I3b532eeacc3784d8d21193d01ddd7fa15ac0684e
This CL allows getChildVisibleRect to optionally always call the
view's parent. The previous version attempted to optimize the call
by not calling further up the view heirarchy when the rect isn't
visible in the current view.
The call is hidden and the previous behaviour is preserved to limit
the bits of code that this change affects.
Bug: 28514727
Change-Id: I49550ed4082bcbdcfe4643b962b50f3308092525
If the insets change, "mWidth/mHeight" won't change
as it's based on the window frame (not the surface size),
we need to track when the insets change and call
HardwareRenderer.setup with the new values.
Bug: 28257246
Bug: 28368990
Change-Id: Ida304b57c4671d010d1cf7b370674c9453841c97
The current mechanism to sync InputMethodService#mIsFullscreen to
InputMethodManager#mFullscreenMode is really fragile because
1. Currently the state change is notified via
InputConnection#reportFullscreenMode(), where InputConnection is
designed to be valid only while the IME has input focus to the
target widget.
2. In favor of performance InputMethodService (IMS) calls
InputConnection#reportFullscreenMode() only when #mIsFullscreen
changed. If InputConnection#reportFullscreenMode() failed, there
is no recovery mechanism.
3. Screen oriantation change is likely to cause Window/View focus
state change in the target application, which is likely to
invalidate the current InputConnection.
What our previous workaround [1] did for Bug 21455064 was actually
relaxing the rule 1 only for InputConnection#reportFullscreenMode().
However, my another CL [2] made the lifetime check of InputConnection a
bit more strict again, which revived the issue as Bug 28157836.
Probably a long-term fix would be to stop using InputConnection to sync
that boolean state between IMS and the application. However, it's too
late to do such a refactoring in N, hence this CL relaxes the rule 1
again keeping it as secure as possible.
The idea is that we allow InputConnection#reportFullscreenMode() to
update InputMethodManager#mFullscreenMode regardless of whether
InputConnection is active or not, as long as the InputConnection is
bound to the curent IME. Doing this as a short-term solution is
supporsed to not introduce any new risk because the active IME is
already able to mess up the InputMethodManager#mFullscreenMode by
calling InputConnection#reportFullscreenMode() on any other active
InputConnection. Bug 28406127 will track the long-term solution.
[1]: Id10315efc41d86407ccfb0a2d3956bcd7c0909b8
da589dffdd
[2]: If2a03bc84d318775fd4a197fa43acde086eda442
aaa38c9f1a
Bug: 28157836
Change-Id: Iba184245a01a3b340f006bc4e415d304de3c2696
Since the field might be null, we can't just read and write the
object directly. Use Parcel's convenience methods to do so safely
instead.
Bug: 28427070
Change-Id: I6460c9cb43dc6da97d5fd9edeaa78bdaaf105446
Clarifying region used for magnification as "magnificationRegion",
both in the public API and in the code. There's been significant
confusion about what "magnfifiedRegion" means. Removing
"availableRegion" from everywhere except where it's required, as
that region was identical to magnified/magnification region.
Trying to shut down magnification was a complex situation where
animations in progress and new magnification requests were tricky to
handle correctly. It was not possible to guarantee that the
magnification callbacks were unregistered consistently. There were
at least two situations that led to phone restarts:
1. If a triple tap was detected between unregistering the callbacks
and shutting down the input filter. In this case the magnification
request would go through.
2. If an animation had just started when magnification was turned
off, so the current magnification was 1.0 but the animator was
about to change it. In this case the callbacks would be unregistered,
and then the animator would start changing the magnification.
This change makes registering and unregistering magnification atomic.
It also makes MagnificationController stick around indefinitely once it
is created, registering and unregistering as needed to support
magnification gestures and services that control magnification. Services
that merely query the status of magnification no longer register for
callbacks.
One part of shutting down is turning off the animation and guaranteeing
that it won't try to make further changes. Adding a flag to
SpecAnimationBridge and a lock in that class so we can guarantee that
nothing happens when we aren't registered for magnification callbacks.
Also reconfiguring all accessibility options when a service stops to
make sure that only the features required by the current configuration
are enabled.
Bug: 27497138
Bug: 27821103
Change-Id: If697cbd34b117d82c8eee1ba7d0254089ee4241d
