When I cleaned up how we maintained the lifecycle of the tracker with a
service, I broke most tracking of the service restart state. (Since at
that point the service is no longer associated with a process, so I
must clean up the tracker state). This change introduces a new special
case for interacting with a service tracker to explicitly tell it when
a service is being restarted. It also fixes how we update the process
state when services are attached to it, so it goes in and out of the
restarting state correctly.
In addition:
- Maybe fix issue #11224000 (APR: Dependent processes not getting added
to LRU list). We were not clearing ServiceRecord.app when bringing
down a service, so if for some reason there were still connections to
it at that point (which could happen for example for non-create bindings),
then we would so it when updating the LRU state of that client process.
- dumpsys procstats's package argument can now be a package or process
name, and we will dump all relevent information we can find about that
name.
- Generally improved the quality of the dumpsys procstats output with its
various options.
- Fixed a bug in ActivityManager.dumpPackageState() where it would hang if
the service was dumping too much, added meminfo to the set of things
dumped, and tweaked command line options to include more data.
- Added some more cleaning code to ActiveServices.killServices() to make
sure we clean out any restarting ServiceRecord entries when a process is
being force stopped.
- Re-arranged ActiveServices.killServices() to do the main killing of the
service first, to avoid some wtf() calls that could happen when removing
connections.
Bug: 11223338
Bug: 11224000
Change-Id: I5db28561c2c78aa43561e52256ff92c02311c56f
The android package is now a special case, not being added to the package list
when creating a multi-process component. There is no need, since this package
is actually the framework itself which must be loaded in every process.
Also cleaned up some of the procstats dump output to help see what is going
on here.
Change-Id: If65d35ecd562f3154bdebfded69c454af6ce8c96
The main problem here was a mistake when turning a single process
structure to a multi-package-process structure with a common
process. When we cloned the original process state, if there were
any services already created for the process for that package, they
would be left with their process pointer still referencing the
original now common process instead of the package-specific process,
allowing the active counts to get bad. Now we switch any of those
processes over to the new package-specific process.
There was also another smaller issue with how ServiceRecord is
associated with a ServiceState -- we could be waiting for an
old ServiceRecord to be destroyed while at the same time creating
a new ServiceRecord for that same service class. These would share
the same ServiceState, so when the old record finally finished
destroying itself it would trample over whatever the new service
is doing.
This is fixed by changing the model to instead of using an "active"
reference count, we have an object identifying the current owner
of the ServiceState. Then when the old ServiceRecord is cleaning
up, we know if it is still the owner at that point.
Also some other small things along the way -- new Log.wtfStack()
method that is convenient, new suite of Slog.wtf methods, fixed
some services to use Slog.wtf when catching exceptions being
returned to the caller so that we actually know about them.
Change-Id: I75674ce38050b6423fd3c6f43d1be172b470741f
Not dealing with the case where there is a null list.
Also fixed some bugs I found while looking at this:
- When resetting the stats, we would use a newly computed time stamp
for the total durations rather than the one we used to reset the
proc/service entries. This would result in them being able to be
slightly > 100%.
- There was a bug in how we split a single process state into its
per-package representation, where we would but the cloned process
state into the new package's entry (instead of properly for its
own package entry), to be immediately overwritten by the new
process state we make for that package. This could result in
bad data for processes that have multiple packages.
- There was a bug in resetting service stats, where we wouldn't
update the overall run timestamp, allowing that time to sometimes
be > 100%.
- There was a bug in computing pss data for processes with multiple
packages, where the pss data was not distributed across all of the
activity per-package process states.
- There was a bug in computing the zram information that would cause
it to compute the wrong value, and then never be displayed.
Finally a little code refactoring so that ProcessState and ServiceState
can now share a common implementation for the table of duration values.
Change-Id: I5e0f4e9107829b81f395dad9419c33257b4f8902
...caused runtime restart
There were some situations where the package list could be set
with process stats when it shouldn't. Not sure if this is causing
the problem, since there is no repro.
Also some improvements to debug output -- new commands to clear
all stats, print full details of stats, and print a one-day
summary (which should match what the UI shows).
Change-Id: I9581db4059d7bb094f79f2fe06c1ccff3e1a4e74
...while setting up a new user from settings.
The delayed service start stuff was too aggressive -- it would
allow a process to be killed between the an onReceive() that calls
startService() and that service being started. This means that
apps that set up global state that they expect to remain set up
during that time could be lost.
This is the first part of a fix, which tightens up when we allow
services to be delayed. Now we will immediately start the service
as long as it currently as a process running that is not in the
cached state. (Previously we would delay if the process was in
the receiver state.)
This unfortunately means that our service start delay is much
less effective. To address that, there will be a follow-on change
to tie broadcast delivery into this to see if we can delay the
finish of a broadcast as long as there are background services
starting in that process.
Change-Id: I2bba2295d10699ee3479375bbe87114b2cbb0826
Haven't found the underlying cause, but this will give us more
information when we get into the bad state.
Change-Id: I9aebd3a025a7c0d931f43098461b64ee3c220746
We now keep track of which process and service states are actively
in use, and remove any that are not in use during a commit. The
activity manager needed to be tweaked to report this data, and ensure
it does not try to operate on one of these structures when not in
use.
Also some other fixes:
- We now keep track of process names associated with services, for
display in the UI.
- Keep track of total run time for each service, also for UI.
- The parceled format is more efficient, not storing duplicates of
process/package names, and writing times as ints when possible.
- Reduced commit period from 1 day to 12 hours, so that our UI can
be a little closer at its attempt to display the stats over 1 day.
Change-Id: Ifeda0ffe963a7b49d8eb2a3f6923f3a5e71a4e43
There was a bug in counting the number of starting services
in a process that would cause it to count too many (it would
increment at any state change while the service is started, not
just when starting/stopping).
Also reduce dumpsys output -- only print summaries for old
data. There is probably no utility in printing the long details
of all that data.
Change-Id: I1c1e901b0214c01eb7d071f23166fc6f3702ca67
Also fix a bug where, when parceling the stats, we were
computing the final duration values too late. We need to
do that before we write the long table.
Change-Id: Idb6c1ed95417448c56973fe5866bfb3570e525f4
ProcessStats is now called ProcessCpuTracker.
ProcessTracker is now ProcessStatsService, and its inner State
class is broken out into a separate top-level ProcessStats class.
This ProcessStats is moved to the framework, so we will be able
to use it elsewhere.
Change-Id: I6a127bcb835b6b474b72647c0b99b82c2137e5c5
