The input system needs to know about the window that has
focus, even if it is on a secondary display. So now we
send it the list of all windows and indicate which display
they are on. We filter the list of windows as necessary
when delivering touch events.
To keep things simple, monitor input channels and input
filters are not supported except on the main display.
We also do not pass the display id to applications; it is
only used inside the input system for now.
Properly scale touch coordinates based on the viewport.
This will be needed to ensure that touch works on external
display as well as when the internal display is being used
to simulate a different resolution.
Change-Id: I1815579a52fcc852c519b5391fc7ab8767c2dc59
Some ANR bugs are very difficult to localize because by the time
the bug report has been captured, the relevant information is
gone. Work around this by capturing a log of the input dispatcher's
state at the exact time of the ANR before anything has changed
and include this information in the bug report.
Also fixed a nit related to some format strings that had
improper field widths specified.
Bug: 6680398
Change-Id: I5323bf18ec5e47a767cd053209753cc46852bf4c
When an ANR occurs, log the associated reason.
When an event takes too long to process (currently more than 2 seconds)
log basic information about the event including how long it actually
took.
Dump the contents of the inbound, outbound and wait queues as part
of dumpsys input.
Bug: 6574842
Change-Id: I9ab754c320f609cb86fe266c469a61e7032dfed6
Prevents ANR on hover or scroll events.
Prevents ANR in the case where one finger touches a window which
is subsequently removed then a second finger comes along and tries
to touch something.
Change-Id: I985ccb70cd86659992328f14f8a705f3f05ea2e5
If a fallback key is generated using a key plus a modifier,
then it's possible we might get a different fallback key
generated if the modifier has changed. PhoneWindowManager
needs to remember which fallback is last generated for a
given key code so that it can apply the same fallback action.
When generating cancellation events, it's important to have
preserved the policyFlags of the original event. Otherwise
we may not dispatch the cancellation properly. For example,
some actions are not performed if the POLICY_FLAG_TRUSTED
is not specified.
Remember the metaState associated with a key event so we can
include it when canceled.
Tell the policy when a fallback is being cancelled so that it
can clean up its state.
After a SEARCH shortcut is invoked, clear the flag indicating
that a shortcut is pending. This is to prevent SEARCH from
getting stuck down in the case where we might forget to send
the up. (Shouldn't happen anymore after the prior fixes.)
Bug: 5616255
Change-Id: I68f0a9679c7af464eaf31c099f2aa50b53fecf1f
The system bar uses input event injection to inject BACK keys
into the application. If the receiving application ANRs, we
used to clear the touch state unconditionally. Doing so would
prevent the system bar from receiving the ACTION_UP event so
the back button would continue to appear pressed until pressed
again.
Now we are more careful to only remove the specific ANR'd
window from the touch state. Other windows should continue
to receive touch events as usual.
Change-Id: If86bfc323e2c7aed82ca1334bc67da649953168f
On reflection, only key events need to be serialized.
This is part of a series of changes to improve input system pipelining.
Bug: 5963420
Change-Id: I028b4eac97497d012036cb60ffbac4cb22d3966c
To support this feature, the input dispatcher now allows input
events to be acknowledged out-of-order. As a result, the
consumer can choose to defer handling an input event from one
device (because it is building a big batch) while continuing
to handle input events from other devices.
The InputEventReceiver now sends a notification when a batch
is pending. The ViewRoot handles this notification by scheduling
a draw on the next sync. When the draw happens, the InputEventReceiver
is instructed to consume all pending batched input events, the
input event queue is fully processed (as much as possible),
and then the ViewRoot performs traversals as usual.
With these changes in place, the input dispatch latency is
consistently less than one frame as long as the application itself
isn't stalled. Input events are delivered to the application
as soon as possible and are handled as soon as possible. In practice,
it is no longer possible for an application to build up a huge
backlog of touch events.
This is part of a series of changes to improve input system pipelining.
Bug: 5963420
Change-Id: I42c01117eca78f12d66d49a736c1c122346ccd1d
Minor tweak to the dispatcher to handle as many finished signals
in a receive callback as possible instead of going back to
the Looper and waiting for the next poll() to hit the callback
again.
This is part of a series of changes to improve input system pipelining.
Bug: 5963420
Change-Id: I8471107371693e21ce8ce7cca1e8d79ba4ca2351
This change enables the input dispatcher to send multiple touch
events to an application without waiting for them to be acknowledged.
Essentially this is a variation on the old streaming optimization
but it is much more comprehensive. Event dispatch will stall as
soon as 0.5sec of unacknowledged events are accumulated or a
focused event (such as a key event) needs to be delivered.
Streaming input events makes a tremendous difference in application
performance. The next step will be to enable batching of input
events on the client side once again.
This is part of a series of changes to improve input system pipelining.
Bug: 5963420
Change-Id: I025df90c06165d719fcca7f63eed322a5cce4a78
The information gathered here will no longer be valid once we
start dispatching multiple events at a time to the same connection.
Moreover, we are more concerned with end-to-end latency, which we
can measure with sufficiently high accuracy in other ways.
This is part of a series of changes to improve input system pipelining.
Bug: 5963420
Change-Id: I49a0c9876b64af56b40e96e0d98c45f325da2a73
The dispatcher no longer needs to track which connections are
active except perhaps for diagnostic purposes, so we might as well
remove this code.
This is part of a series of changes to improve input system pipelining.
Bug: 5963420
Change-Id: Ibadc830b7b792a59b9244d0a6e85f320c4947109
Since we no longer stream events to the application, we don't need to
keep the current list of input targets around longer than it takes
to begin the dispatch cycle.
This is part of a series of changes to improve input system pipelining.
Bug: 5963420
Change-Id: I5824b04e564c8a4dec991598292441e46c331905
Since we will not longer be modifying events in place, we don't need
to use an ashmem region for input. Simplified the code to instead
use a socket of type SOCK_SEQPACKET.
This is part of a series of changes to improve input system pipelining.
Bug: 5963420
Change-Id: I05909075ed8b61b93900913e44c6db84857340d8
Don't worry, these features will be moving to a different part of
the pipeline. We're just getting them out of the way for now so
we can make deeper changes to how the input dispatcher works.
This is part of a series of changes to improve input system pipelining.
Bug: 5963420
Change-Id: If97de923c8165776882eef23f5204cc448dad0fc
Try harder to test for liveness. There are situations where
the lock might not be held but the input system is stuck in
a callback into the window manager policy that has hung.
Bug: 5094994
Change-Id: Iff88655512a5dc8bbb4615be65f4115e975c020b
Fixed a problem where the key up for the ALT or META key was not
delivered to the task switcher dialog because it was deemed
to be inconsistent with the window's observed state. Consequently
the dialog would not be dismissed when the key was released.
Moved global hotkey handling for META+* shortcuts and ALT/META-TAB
into the window manager policy's interceptKeyBeforeDispatching
method. This change prevents applications from hijacking these
keys.
The original idea was that these shortcuts would be handled only
if the application did not handle them itself. That way certain
applications, such as remote desktop tools, could deliberately
override some of these less important system shortcuts.
Unfortunately, that does make the behavior inconsistent across
applications. What's more, bugs in the onKeyDown handler of
applications can cause the shortcuts to not work at all, for
no good reason.
Perhaps we can add an opt-in feature later to enable specific
applications to repurpose these keys when it makes sense.
Bug: 5720358
Change-Id: I22bf17606d12dbea6549c60d20763e6608576cf7
Bug: 5011907
Introduce a 150ms delay in handling volume down keys
while waiting to see if a power key will follow.
Don't trigger the screenshot chord if both volume up and
volume down are pressed together.
Don't trigger the long-press power menu if volume keys are
also pressed.
Require the user to press both keys in the chord within
the debounce time and continue long-pressing them in order
to trigger the screenshot action.
Change-Id: I248968d37b73c09d6d08e7f62667c443eba32da0
Bug: 5156144
Input channels could leak or simply live longer than they should
in some cases.
1. Monitor channels (used by the pointer location overlay) are never
unregistered, so they would leak.
Added code to handle failures in the receive callback by closing
the input channel.
2. The DragState held onto its input window and application handles
even after the input channel was disposed.
Added code to null these handles out when they are no longer needed.
3. Input channels previously used as input event targets would stick
around until the targets were cleared (usually on the next
event).
Added code to detect when the input dispatcher is in
an idle state and to proactively clear the targets then
to ensure that resources are released promptly.
4. Native input window handles held onto the input channel even
after the input window was removed from the input dispatcher.
Consequently, the input channel would not be disposed until
the input window handle itself was freed. Since the input
window handle is held from managed code, this meant that the
window's input channel could stick around until the next GC.
Refactored the input window handle to separate the properties
(info) and identify (handle) state into different objects.
Then modified the dispatcher to release the properties (info)
when no longer needed, including the input channel.
7. The pointer location overlay does not actually use its
standard input channel, only the monitor input channel.
Added INPUT_FEATURE_NO_INPUT_CHANNEL to allow windows to
request that they not be provided with an input channel
at all.
Improved some of the error handling logic to emit the status
code as part of the exception message.
Change-Id: I01988d4391a70c6678c8b0e936ca051af680b1a5
Bug: 5049148
Finished stylus support, including support for indirect stylus
and mouse tools.
Added TILT axis. When stylus tilt X/Y is available, it is transformed
into an orientation and tilt inclination which is a more convenient
representation and a simpler extension to the exiting API.
Touch devices now only report touch data using a single input
source. Previously touch devices in pointer mode would report
both absolute touch pad data and cooked pointer gestures.
Now we just pick one. The touch device switches modes as needed
when the focused application enables/disables pointer gestures.
This change greatly simplifies the code and reduces the load
on the input dispatcher.
Fixed an incorrect assumption that the value of ABS_(MT_)DISTANCE
would be zero whenever the stylus was in direct contact. It appears
that the correct way to determine whether the stylus is in direct
contact (rather than hovering) is by checking for a non-zero
reported pressure.
Added code to read the initial state of tool buttons and axis values
when the input devices are initialized or reset. This fixes
problems where the input mapper state might have the wrong initial
state.
Moved responsibility for cancelling pending inputs (keys down,
touches, etc.) to the InputDispatcher by sending it a device reset
notification. This frees the InputReader from having to synthesize
events during reset, which was cumbersome and somewhat brittle
to begin with.
Consolidated more of the common accumulator logic from
SingleTouchInputMapper and MultiTouchInputMapper into
TouchInputMapper.
Improved the PointerLocation output.
Change-Id: I595d3647f7fd7cb1e3eff8b3c76b85043b5fe2f0
There turn out to be two distinct bugs leading to runtime restarts.
The first, dating from at least Android 3.1, is that following certain kinds
of app crashes we properly clean up the drag-state bookkeeping, but aren't
prepared in the case of the drag-target timeout clock firing with a now-
null drag state in effect. We now catch that edge condition and don't NPE
(and note that there was already similar code around the separate timeout
when an app is *starting* the drag process).
The second bug is that some new-in-ICS code in the input channel management
wasn't prepared for certain cases where the current touch window could have
become unusable and its input channel torn down summarily in the case of the
aforesaid app crash during drag. The code now makes sure that there really
is an input channel that needs to be flushed / cancelled prior to attempting
that operation.
Fixes bug 5173534
Change-Id: Idaae158ecfb4b93456ab1425769b669962893c00
Bug: 5161854
The InputDispatcher contains code to recover from ANRs more quickly
when the user touches a different application. When triggered,
it sets mNextUnblockedEvent to the next event it should dispatch
(to the next application). All prior events are dropped.
This change fixes a bug where mNextUnblockedEvent could potentially
never be reset back to NULL if the inbound queue was drained
for some reason.
Change-Id: I21be2f02c9fc770bd260ea414db8bb184df89ada
Bug: 5064702
Introduced the concept of an InputListener to further decouple
the InputReader from the InputDispatcher. The InputListener
exposes just the minimum interface that the InputReader needs
to communicate with the outside world. The InputReader
passes arguments to the InputListener by reference, which makes
it easy to queue them up.
Consolidated all of the InputReader locks into one simple global
Mutex. The reason this wasn't done before was due to potential
re-entrance in outbound calls to the InputDispatcher. To fix this,
the InputReader now queues up all of the events it wants to send
using a QueuedInputListener, then flushes them outside of the
critical section after all of the event processing is finished.
Removing all of the InputMapper locks greatly simplifies the
implementation.
Added tests for new stylus features such as buttons, tool types,
and hovering.
Added some helpers to BitSet32 to handle common code patterns
like finding the first marked bit and clearing it.
Fixed a bug in VelocityTracker where the wrong pointer trace
could get cleared when handling ACTION_POINTER_DOWN. Oops.
Changed PointerCoords so it no longer stores useless zero
axis values. Removed editAxisValue because it is not very
useful when all zero value axes are absent and therefore
cannot be edited in place.
Added dispatch of stylus hover events.
Added support for distance and tool types.
Change-Id: I4cf14d134fcb1db7d10be5f2af7b37deef8f8468
This fixes an issue where an ANR can occur in a newly focused
application through no fault of its own, simply because
the previous app took a little while to transfer focus to
the new app.
Bug: 4584620
Change-Id: If3227eb68d92a09a108e9de7f0afcbd9a5dbdead
This change moves the cached window and application input state
into the handle objects themselves. It simplifies the dispatcher
somewhat because it no longer needs to fix up references to
transient InputWindow objects each time the window list is updated.
This change will also make it easier to optimize setInputWindows
to avoid doing a lot of redundant data copying. In principle, only
the modified fields need to be updated. However, for now we
continue to update all fields in unison as before.
It turns out that the input dispatcher was inappropriately retaining
pointers to InputWindow objects within the mWindows InputWindow
vector. This vector is copy-on-write so it is possible and the
item pointers to change if an editing operation is performed on
the vector when it does not exclusively own the underlying
SharedBuffer. This bug was uncovered by a previous change that
replaced calls to clear() and appendVector() with a simple use
of operator= which caused the buffer to be shared. Consequently
after editItemAt was called (which it shouldn't have, actually)
the buffer was copied and the cached InputWindow pointers became
invalid. Oops. This change fixes the problem.
Change-Id: I0a259339a6015fcf9113dc4081a6875e047fd425
This fixes a problem where touches can get stuck because the
driver and the framework have different ideas of what the
initial slot index is. The framework assumed it was slot 0
but it could in principle be any slot, such as slot 1. When
that happened, the framework would start tracking the first
touch as slot 0, but it might never receive an "up" for that slot.
Change-Id: Idaffc4534b275d66b9d4360987b28dc2d0f63218
