- Reduce the amount that we ask processes to GC after a significant
operation occurs, but introducing a minimum time between GCs and
using this in various ways to schedule them.
- Don't spam all of the processes with onLowMemory(). Now deliver
these using the same gc facility, so we do the processes one at a
time, and don't allow the same process to get this call more than
once a minute.
- Increase the time a service must run before we will reset its
restart delay to 30 minutes (from 10).
- Increase the restart delay multiplication factor from 2 to 4.
- Ensure that we don't restart more than one service every 10 seconds
(unless some external event causes a service's process to be started
for some other reason of course).
- Increase the amount of time that a service must run before we
decide to lower it to a background process.
And some other things:
- Catch IllegalArgumentException in ViewRoot like we do for no
resources to avoid the system process crashing.
- Fix a number of places where we were missing breaks between the
activity manager's message dispatch func(!!).
- Fix reason printed for processes in the background.
- Print the list of processing waiting to GC.
This adds some new debugging code to make it easier to see why a process is at
a certain oom_adj level -- for example telling you that a certain other process
has a binding to a certain one of its services. This has helped a lot in
identifying cases where processes are holding references to other processes that
they don't need and thus not allowing the system to get memory it needs.
Also fix a few problems with leaking entries on the service restarting and
service stopping lists.
Re-arrange various things to ensure that the search dialog is told about system windows being
closed before it is told about the navigation back to home.
Create a new package setting object for updated system apps rather than moving
around the same setting. This updates the resource, code and version correctly.
For updating system packages, disable the package first which removes the entry
from internal structures, create a new package setting, add it to list of user id's
then rest of installation steps, kill the process if needed via ActivityManager
then add this setting if everything was successful. This also fixes issues with
updating values prematurely.
When a new version of system package is available via OTA, just physically remove
entries for pkg. Note that the component and other info will be eventually updated
later on when scanning the package.
Also move certificate verification slightly ahead before scanning packages.
Some null checks
New api's in ActivityManager to kill an application pkg before finishing installation
This was previously hard-coded to 0. Now set to the value of the
"ro.opengles.version" system property. The default value is
ConfigurationInfo.GL_ES_VERSION_UNDEFINED.
Instead of a list, we now just have a single boolean indicating whether an
application is density aware, and this set set to true by default as of
Donut.
This adds a new API with the activity manager to find out about movement between
activities. For my sanity, the old IActivityWatcher is now renamed to
IActivityController, and the new activity movement interface is named
IActivityWatcher.
This changes the search manager itself to use the new API to manage its state.
Note that there are still problems when going back to the search dialog after
it was hidden -- the suggestions window no longer appears until you explicitly
dismiss and re-show it.
This is the problem where various things are listening for broadcasts
(such as battery status, PIN/PUK/Network) that an application can send
to cause harm to the system.
Solving this is tricky because many of these broadcasts are sticky,
and I have never figured out how to do permissions with sticky
broadcasts in a sane way. So instead, I am going to punt on the
general problem and just brute force it:
There is new a way for system components to declare specific
broadcast actions to be protected, which means that only the system
and the phone can send them. This is good enough for now. None
of it is exposed in the public API so we can make something a little
less stupid in the future if we ever need to.
This adds a new <path-permission> tag you can use inside of a <provide>
to define additional path-based permissions that broaden the global
read and write permissions. The initial use for this will be global
search, so that a content provider that is protected by permissions
can make a part of itself available to global search under another
permission. This addresses the issue with global search not being able
to request permissions it would need of providers it doesn't know
about at build time.
The system now keeps a tag of the last version (just an arbitrary string)
that the setup wizard was run for. If this is different than the current
one in the setup wizard, then setup is launched at boot.
This introduces a new intent action for the part of the setup wizard that
gets run for an ungrade, which the system uses to find its current version
tag for comparing against what was last stored. It is up to the launched
setup activity update the stored setting to reflect its current value,
once it is happy.
It turns out this was not a problem in the resource code at all. Rather,
the system process has a cache of pre-loaded attributes it uses to avoid
continually reloading things as it needs them. Well it turns out this
cache wasn't flushed after a package was uninstalled or a configuration
changed, so you could re-install an app where you change its style resources
so its theme now points to one that is inconsistent in the cache.
This is mostly a problem for developers, where they continually install
new versions of an app where resources have changed. This could possibly
show up when updating an app on a normal phone, although the problem would
eventually correct itself since this cache uses weak references.
Anyway, the cache is now reworked to be flushed appropriately.
This change also includes an update to aapt to be able to dump the
contents of bags in resources.
BackupManager now no longer tries to use a null service binder if it's used
early during the boot process. ActivityManagerService no longer tries to
dereference null pointers if bind/unbind semantics get out of step due to things
being run too early.
The am command is now the one that takes care of opening the target file,
handling the opened file descriptor to the process that will be profiled.
This allows you to send profile data to anywhere the shell can access, and
avoids any problems coming up from the target process trying to open the
file.
- Fix a bug where targetSdkVersion could not be set if minSdkVersion. Stupid, stupid.
Also make sure to fail if minSdkVersion is for a code name. Really stupid.
- Change the API for resize compatibility mode to be a bit in the flags field, instead
of a separate boolean.
- Implement delayed dexopting, to avoid the looong full dexopt during boot. This is
only enabled for "eng" builds. When in this mode, the activity manager will make
sure that a dexopt has been done before loading an .apk into a process, and will
try to avoid displaying ANRs if they are due to the dexopt causing some operation
to take longer than it normally would (though I make no guarantees about this
totally working).
- Add API to Context to get the ApplicationInfo for its package, for easy access to
things like targetSdkVersion.
This introduces a new Uri form of Intent with an "intent:" scheme, and a
corresponding update to the parser to handle these, so that the browser
can use this generic facility for starting activities based on the links
that are clicked and allow for web pages to link to arbitrary intents.
There is also a new "package" field on Intent which allows you to limit
the components it finds to a given package. This replaces the new method
that was added to PackageManger for doing this when resolving activities,
and implements it for all Intent queries against the package manager.
Create a new public IntentSender class that can be used by PackageManager instead.
This new class uses IIntentSender internally and can only be created by PendingIntent for now.
Provide a new getIntentSender api in PendingIntent to create an instance of this class.
Move IIntentSender and IIntentReceiver from android.app to android.content
Change imports of IIntentSender and IIntentReceiver to reflect the new package name
The PackageManager api has been named as freeStorageWithIntent and will be renamed as freeStorage
once the older api(which has been deprecated) will be removed shortly.
We weren't bumping the oom_adj of processes receiving a registered broadcast. Previously
this wasn't a problem, because those processes are allowed to have their oom_adj
managed by whatever else is in them and if the registered receiver goes away no problem.
But now this is also controlling the scheduling class, so we need to bump them up.
This is probably good anyway, since the developer has the same assurance of their process
not being killing in the middle of registered receivers like it had always been for
manifest receivers.
Also fixed a small issue where we were not recomputing the oom_adh after finished with
a broadcast.
There was old code that would kill the system process in some cases when there
was a bad activity token. This is really no longer used, except in a few
places where it allows apps to kill the system. So just get rid of it and
make the world a better place.
* Put in some permission enforcement around agent connection notification
and full-backup scheduling.
* Full backup now applies to any package, not just backup participants who
have declared their own android:backupAgent
* The process of running the backup operation on the set of apps who have
been queued for it is now done in a separate thread, with a notification
mechanism from the main Backup Manager service to pass along new-agent
binding knowledge. There's no longer one do-backup message on the primary
Handler per target application.
* The new backup thread sets up the desired transport now and passes
along the newly backed-up data to it for each backup target. Two
transports have been defined so far, GoogleTransport and AdbTransport;
both are stubs at present.
Note that at present the backup data output file seems to be properly
created, but after doBackup() is called on the test app's agent it's
still zero size.
* Refactored Compatibility code
* Added CompatibilityInfo class
* Removed getApplicationScale from Context
* Added Resources#getCompatibilityInfo so that RootView can get the compatibility info w/o going through Context
* Expandable support
* Added expandable tag under manifest
* Old application w/o expandable is given the default screen size ([320, 480] x density).
* The non-expandable window is centered.
Track the foreground CPU time of an activity so that we can tell if apps are
spending more time in the background compared to foreground.
Update power profile values for screen backlight and GPS.
Fix some javadoc bugs (milliseconds vs. microseconds).
Backups will be handled by launching the application in a special
mode under which no activities or services will be started, only
the BackupAgent subclass named in the app's android:backupAgent
manifest property. This takes the place of the BackupService class
used earlier during development.
In the cases of *full* backup or restore, an application that does
not supply its own BackupAgent will be launched in a restricted
manner; in particular, it will be using the default Application
class rather than any manifest-declared one. This ensures that the
app is not running any code that may try to manipulate its data
while the backup system reads/writes its data set.
The algorithm for this is currently very simple: all persistent processes are
always in the normal scheduling group, all other processes are normal if their
oom_adj is as good or better than VISIBLE, otherwise they are in the background
group.
Note that this currently results in a fair number of log messages about not
being able to change the group, since the system process does not have
permission to do so. Once a kernel fix is in, these will go away and the code
will start working.
If an installerPackageName was specified when the app was installed,
looks for a receiver of ACTION_APP_ERROR in that package. If found,
this is the bug report receiver and the crash/ANR dialog will get a
"Report" button. If pressed, a bug report will be delivered.
Now we have a 5-second time after home is pressed, during which
only the home app (and the status bar) can switch to another app.
After that time, any start activity requests that occurred will
be executed, to allow things like alarms to be displayed. Also
if during that time the user launches another app, the pending
starts will be executed without resuming their activities and
the one they started placed at the top and executed.
Issue 1837610 Adding a Widget before Running the Associated App Causes a Force Close
We were not retrieving the shared libraries of an application when deliving a
broadcast to an explicit component.
Issue 1753079 loading class path of instrumented app into instrumentation may load wrong path when instrumented app shares process with other apps:
We were using the ApplicationInfo that was used to originally create the process, not the one that the
instrumentation is against.
This introduces a new class in the base platform for performing a clean
shutdown (which was copied from the classes in the policies). It
includes new features to send a shutdown broadcast for applications
to do cleanup, and ot have the activity manager pause the current
activity before proceeding with the shutdown. These facilities are
also use to write at the most recent stat files for sync, battery
and user activity.
There are three major classes of changes here:
- Avoid writing lines where their values are often empty, false, or some other typical thing.
- Use partial writes to the PrintWriter to avoid creating temporary strings.
- Use StringBuilder where we need to generate real String objects (and where possible cache the result).
My original implementation was computing averages and medians. Now we do binning, as requested. So much simpler, too! In addition, it fixes a bug where when hoping across activities we were only accounting for the last activity as the total time; now we count the time from the start of the initial activity.
This also includes some reduction and optimization of the activity manager dumpsys output.
The usage stats service now collects per-activity launch time stats. There are a number of fixes and improvements to its statistics management and collection; it now operates its calendar in GMT and ensures that for checkin purposes it always reports one day and only one complete day to the checkin service.
Also change the checkin option from "-c" to "--checkin" since it is really a special thing.
