Implement issue #10691475: Kill cached processes if about to...
...be uncached and too large
When the device is in a low RAM state, when we go to pull a cached
process out to use for some background operation, we can now kill
the current process if we consider its size to be too large.
Note that the current implementation for killing processes is to
just use the same killUnneededProcessLocked() method that we already
have for other things like too many cached processes. This is a
little wrong here, though, because in this case we are at the
point where the caller is actually looking for a process to use.
This current code is not actually removing or cleaning up the
process, so we still need to return the now killed ProcessRecord
and let things fall out from there, which typically means the caller
trying to make an IPC on it and failing and falling into its "oh
no the process died unexpectedly" path. All code using this
*should* be able to handle this correctly, anyway, since processes
really can be killed at any time.
At some point we may to make this implementation cleaner, where it
actually tears down the process right in the call and returns a
null ProcessRecord. That is very dangerous however (we'd need to
go through all paths into this to make sure they are going to be
okay with process state changing on them like that), and I'm not
sure it is really worthwhile. This intention is that killing
processes like this is unusual, due to processes being too large,
and anyway as I wrote all of our incoming code paths must already
be able to handle the process being killed at this point and one
could argue this is just another way to excercise those code paths.
Really, the main negative to this is that we will often have spam
in the log with exceptions about processes dying unexpectedly.
If that is the only issue, we could just add some conditions to
quiet that up at in this case.
We don't want to compute the size of the process each time we try
to evaluate it here (it takes 10s or ms to do so), so there is now
a new field associated with the process to give us the last pss
size we computed for it while it was in the cached state.
To be able to have better cached pss data when we now need it, the
timing for computing process pss has been tuned to use a much
shorter delay for the situations when the process has first switch
into a new state. This may result in us having a fair amount more
pss data overall, which is good, as long as it doesn't cause us to
be computing pss excessively and burning cpu.
Procstats now also has new state to keep track of the number of
times each process has been killed by this new system, along with
the min, avg, max pss of all the times it has happened. This has
slightly changed the checkin format to include this additional data
at the end of pkgkills/prockills lines.
Other changes here:
- Fixed a problem where GPU RAM was not being seen when dumping
the full RAM details of a process. This was because in that
case the system would ask the process to compute its own MemInfo,
which it returned, but the process doesn't have permission to
access the files containing the GPU RAM data. So now the system
always computes the MemInfo and hands it to the app.
- Improved broadcast delays to not apply the delay if the next receiver
of the broadcast is going to run in the same process as the last
one. A situation I was seeing was an application that had two
receivers, one of which started a service; we are better off letting
the second receiver run while the service is running.
- Changed the alarm manager's TIME_TICK broadcast to be a foreground
broadcast. This really should have been anyway (it is supposed to
go out even minute, on the minute, very accurately, for UI elements
to update), and is even more important now that we are doing more
things to delay background broadcasts.
- Reworked how we maintain the LRU process list. It is now divided
into the two parts, the top always containing the processes holding
activities. This better matches the semantics we want (always try
to keep those around modulated by the LRU order we interleave with
other cached processes), and we now know whether a process is being
moved on the LRU list because of an activity operation so we can
only change the order of these activity processes when user operations
happen. Further, this just makes that common code path a lot simpler
and gets rid of all the old complexity that doesn't make sense any
more.
Change-Id: I04933ec3931b96db70b2b6ac109c071698e124eb
12 files changed