SystemProperties.get() is not particularly fast,
especially if a string is returned. Since ro.* values
are unable to be changed, there's no need to
continously re-query them. Cache the value at
static init time to trivially fix this.
Test: refactoring CL.
Change-Id: Iccb021d3cb2ba3a4a1d0048ddec6811bb7409eec
(cherry picked from commit aa67f684ff)
Add option '--streaming' to 'am start' and 'am profile' commands.
If the option is given, the output of method trace profiling
will be streamed into the specified file, so the output is no
longer limited by the buffer size.
Test: m -j48 test-art-host;
m -j48 ART_TEST_TRACE=true ART_TEST_TRACE_STREAM=true test-art-host;
I also tested manually. Tried all 8 combinations of
sampling/instrumention
streaming/non-streaming
'am start --start-profiler' / 'am profile start'
The output files are all in expected shape.
Bug: 33300765
Merged-In: I8a5136a1c7330c8260b7c6c8da63f42a73aee275
Change-Id: I8a5136a1c7330c8260b7c6c8da63f42a73aee275
The EGL Init thread spends a lot of time loading shared libraries,
which involves reading pages and adjusting VM pages. Previously it was
a low priority thread, which meant that when it blocked while holding
a VM lock, it might take a while to be scheduled again once ready, so
it would hold the VM lock longer than necessary, blocking other
threads waiting on soft faults, etc. Classic priority inversion.
Leaving the thread at normal priority causes it to interfere *less*
with critical-path Activity launch work.
Bug: 34611670
Test: go/platform-startup, specifically:
./tradefed.sh run google/template/local --template:map test
google/test/performance/app-launch-perf
--post-boot-command "service call persistent_data_block 6 i32 1"
--post-boot-command "service call persistent_data_block 7 i32 1"
--apps Clock alarmclock --trial-launch --launch-iteration 10
--launch-order cyclic --drop-cache --trace-directory sdcard
--trace-iteration 2 --simple-perf-cmd "simpleperf stat -a"
--report-metrics all --force-skip-system-props --alt-dir `pwd`
Change-Id: I17647c657da64d15f40f8f2b97ee3400e1f0be7f
Preloading EGL in Zygote was originally a memory footprint
optimization, but it turns out to be an important app startup time
optimization as well. Preloading EGL in Zygote is incompatible with
updatable graphics drivers, but we don't want to do it on-demand as
part of drawing the first frame either, since that increases
first-frame latency unacceptably.
This change removes Zygote preload, and instead loads EGL on a
low-priority background thread immediately after choosing which
graphics driver to use. This means it is usually done well before
drawing the first frame, without significantly disrupting other
activity launch work.
Test: observe systrace of Calculator launch on bullhead
Bug: 34404021
Change-Id: I6a0f6b90ade21848a10d51ddae62c936f70151b5
Merged-In: I887aa09bd35b088b16f53a89838a0c7c98f15761
Preloading EGL in Zygote was originally a memory footprint
optimization, but it turns out to be an important app startup time
optimization as well. Preloading EGL in Zygote is incompatible with
updatable graphics drivers, but we don't want to do it on-demand as
part of drawing the first frame either, since that increases
first-frame latency unacceptably.
This change removes Zygote preload, and instead loads EGL on a
low-priority background thread immediately after choosing which
graphics driver to use. This means it is usually done well before
drawing the first frame, without significantly disrupting other
activity launch work.
Test: observe systrace of Calculator launch on bullhead
Bug: 34404021
Change-Id: I887aa09bd35b088b16f53a89838a0c7c98f15761
This reverts commit 9f7a0acd2d.
Bug: 34350643
Bug: 34396955
Test: installPackage() doesn't cause deadlock if it doesn't call
processPackage() prior to that.
Change-Id: I1da055c86a3326ae341da1b4d5dc79ab4be256fb
If an Exception thrown by a Binder call implements the Parcelable
interface, then parcel it and rethrow back at the caller. There is
strict requirement that these Parcelable exceptions must be defined
by the system (as determined by checking the ClassLoader). We prefix
the Parcelable contents with a length so that native code can skip
over the blobs.
Define a new ParcelableException class that can be used to transport
exceptions that cannot be modified to add Parcelable behavior, and
switch ExceptionUtils to use this new class for sending IOExceptions.
Test: builds, boots, wrapped exceptions work
Bug: 33749182
Change-Id: I1352ea1566ddf01120d9d0e819ba6f70fc407e11
The GraphicsEnvironment class is given information during application
start, and makes it available to EGL/GLES/Vulkan loaders that don't
have easy access to the VM or to the application Context. Currently
only the driver path is handled, but the existing support for setting
library paths (for Vulkan extensions) and cache directory information
should move here.
Bug: 33531483
Test: various apps w/ and w/o driver package installed
Change-Id: I5820d3d1301d5461e10706f551b268c54d4f8926
(cherry picked from commit b12249b671)
The GraphicsEnvironment class is given information during application
start, and makes it available to EGL/GLES/Vulkan loaders that don't
have easy access to the VM or to the application Context. Currently
only the driver path is handled, but the existing support for setting
library paths (for Vulkan extensions) and cache directory information
should move here.
Bug: 33531483
Change-Id: I4e4e7fb21f1bcc67122e9173514af5f18c063991
Merged-In: I5820d3d1301d5461e10706f551b268c54d4f8926
Commit 794c8b0b3f fixed the race condition
when requesting data wipes via uncrypt. We have similar issue with
RecoverySystem.installPackage(). It first requests to set up the BCB,
then triggers a reboot. These two steps should finish atomically.
This CL switches to calling
RecoverySystemService.rebootRecoveryWithCommand(), which guards the two
steps with synchronized blocks.
Bug: 34239871
Test: Having two apps: one calls RecoverySystem.cancelScheduledUpdate()
continuously, and the other calls RecoverySystem.installPackage()
just once. The install request should not be cancelled by the
other.
Change-Id: I5ec56fcaa70eae7c33e3cc8e6cfc7472b935ce4e
To support reading embedded buffers that can be
nullptr (currently only in empty hidl_vec).
Bug: 34255213
Test: hidl_test_java
Change-Id: I72028f580b7863b6bfeb31a5c0f43deed36dfd64
This change saves and loads a different brightness setting when the user
goes in and out of VR Mode.
Bug: 30984614
Change-Id: If3c3e81b592e0c6fd037e5783559683e5cb58379
When "getPrimaryStorageSize" provides a path
to "readLong", the option that the path
doesn't exist is expected, since it tries
all paths from "INTERNAL_STORAGE_SIZE_PATHS"
until there is success.
This patch makes us catch the "FileNotFoundException"
and "NumberFormatException" seperately.
For the above reason a "FileNotFoundException"
is now treated as an information only.
The "NumberFormatException" and other exceptions
are now treated as error since those are not
expected to happen.
Change-Id: I5316f9c3108e36c31b27dc5df8bf8ac4d4257629
Signed-off-by: Alex Naidis <alex.naidis@linux.com>
Now we're getting somewhere! This CL starts measuring disk usage
using quotactl(), which is almost instant and has much lower impact
on flash memory lifetime.
We now grant the per-app cache GID to every launched app, and the
ContextImpl logic that creates cache directories matches the logic
down in installd.
Test: builds, boots, quota stats match manual stats
Bug: 27948817
Change-Id: Ie269a2958ce0e1c17cb74dbfecc791a5c12922cf
This is mostly copied over from binder's existing
death recipient support. The implementation keeps
a list of registered recipients, both for being
able to map a native recipient back to the corresponding
Java recipient, as well as being able to unregister
recipients correctly.
Test: mma, hidl_test_java
Bug: 31632518
Change-Id: Id313fd248be6925056c4ade8298fe5fb04e007cc
This change saves and loads a different brightness setting when the user
goes in and out of VR Mode.
Bug: 30984614
Merged-In: Ie5578bbd6ea346f0eb34fe4abbfd604a5d7c0c93
Change-Id: Ie5578bbd6ea346f0eb34fe4abbfd604a5d7c0c93
When an app is debuggable, check whether a script called "wrap.sh" exists
in the app's native library directory. If so, start the app using the
invoke-with functionality over the script. Weaken the invoke-with check
on the zygote side to allow the functionality for debuggable apps.
The goal of the functionality is to make malloc debug, strace and other
similar tools available for NDK based application developers.
Bug: 33668201
Test: manual - debug malloc can be enabled using the new feature
Change-Id: Ia4bec0854cf4dc08446f1671494200f54ef366ee
Add "--invoke-with" to the zygote connection protocol. It was
already understood as an argument by the zygote.
Bug: 33668201
Test: m
Change-Id: I59095f2ac542aadff78a7ff1dded86cf5f192707
Since N, DNS servers are not set by setDnsServersForNetwork, they
they are set by setDnsConfigurationForNetwork.
Test: compiles
Bug: 30944031
Change-Id: I3a3edb9ffe9d5f9652af0b6637eb89b1b044a270
It looks like one operation was done out of order and some of the times used in
the calculations were leading to incorrect results.
BUG: 31023263
Test: bit FrameworksCoreTests:com.android.internal.os.BatteryStatsDurationTimerTest
Change-Id: I417cc28c5a55748067b6c7f682a66fe3dbc09f09
(cherry picked from commit 47db5a8bf7)
