Since vendor has a way to override the group cpu/schedtune setup, we
cannot assume the group will always return valid data. This CL let
get_sched_policy to fallback to cpuset if no valid data found in
cpu/schedtune cgroup. In longer term, we should find a way to cache the
group or app's process state in framework other than relying on reading
cgroup back.
Test: /data/nativetest64/libcutils_test/libcutils_test
Bug: 210066228
Signed-off-by: Wei Wang <wvw@google.com>
Merged-In: I8b4396365a7fc2d93e3a22746195585c140eef3c
Change-Id: I8b4396365a7fc2d93e3a22746195585c140eef3c
(cherry picked from commit c8c0b5415c14af56277f5e543ebff5fbba7fb7b2)
Using WriteStringToFile will hold kernfs_mutex which is a big lock,
so use WriteStringToFd instead. Besides, also support fd cache for it.
Bug: 206970384
Test: build pass
Change-Id: Id79f9e1095f52079393c58edb9a4d526f4cc6b5e
Add new CgroupGetControllerFromPath function to retrieve the name of the
cgroup using a file path. If the file does not belong to any cgroup, the
function returns false.
Bug: 191283136
Test: build and boot
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
Change-Id: Ic17a474cb25a80a3339b33ed8bc27b07af053abb
UserLifecycleTests test disables all Little cores in the course of the
test, which causes attempts to add a process into /dev/cpuset/restricted
cpuset cgroup to fail with ENOSPC error code, indicating that a process
is joining a cpuset cgroup with no online cpus. Current libprocessgroup
implementation will log an error on each such occurrence, which spams
the logs and makes it hard to analyze test results. Because this
situation does not happen in production environment (we do not offline
cpus), we can prevent flooding the logs by identifying this case,
logging an appropriate error one time and ignore all later similar errors.
Bug: 158766131
Test: adb shell "echo 0 > /sys/devices/system/cpu/cpu[0-3]/online"
Test: start some apps, observe libprocessgroup errors in the logcat
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
Change-Id: Ia91d8839d86787569c255481bde077be51c43d93
Create a new group for dex2oat in cpu cgroup, which is dedicated for
dex2oat processes. Also modify task profiles for this change.
Bug: 201223712
Test: dex2oat group created
Change-Id: Ic61f4b8a64d01c03549b680970805e12b9ce4fcc
Since we are adding SCHED_SP_SYSTEM into framework, the policy is needed
on older releases as well.
Bug: 201428466
Test: Build
Signed-off-by: Wei Wang <wvw@google.com>
Change-Id: I0784820bfec84fd8ef98f2feb2caee084d0a14a2
When creating uid/pid hierarchy, cgroup.subtree_control should be set at
every level of that hierarchy except for the leaf level.
Bug: 195149205
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
Change-Id: Iedc2e859715b31db62158c85016708f252db2b70
Since installation time is also critical and sometime we want a finer
control on its behavior for vendor. Instead of set dex2oat to background
group, add a new task profile.
The default policy is to mirror SP_BACKGROUND to be compatible with
legacy devices.
Test: Boot
Bug: 188947181
Signed-off-by: Wei Wang <wvw@google.com>
Change-Id: I7d5e113d1c368205f712eadad27775de226131ce
(cherry picked from commit 3745ce7aa8)
When a new process is launched it ensures that all files under its uid/
and uid/pid hierarchy are accessible by the user/group of that process.
If the directory already exists that means the access permissions have
been already set before, therefore we do not need to reset them again.
This also avoids a race between two processes in the same uid with one
process being launched and walking the uid/ directory while the other
process is being killed and changing the content of that directory. In
such a race the process walking uid/ might find the uid/pid directory of
the process being killed but by the time it tries to set its permissions
the directory might be removed because the process got killed. The
change eliminates the possibility of this race.
Bug: 192421915
Bug: 192512069
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
Change-Id: I182298c36f6b0b4580ab59e440bd3aea16f5fbfe
Merged-In: I182298c36f6b0b4580ab59e440bd3aea16f5fbfe
In some rare cases, race happens between 2 processes in the same uid.
1. Process A is dying
2. system_server calls RemoveProcessGroup() for A
3. Zygote forks Process B with the same uid of A
4. system_server calls MkdirAndChown(uid) for B
5. system_server calls MkdirAndChown(uid, pid) for B
As 2 & 4/5 belong to different threads, 2 might happens before or after
step 4/5, or even in the middle of 4/5. In such a case, 4 or 5 will
fail, leaving process B in wrong (Zygote) group.
The uid dir is only created when the corresponding apps have been
launched at least once. It's reasonable to assume one of them is going
to be launched again. Deleting and recreating the uid dir just slows
down applaunch.
Introducing a new lock in libprocessgroup can also solve the race issue.
But that will slow down the applaunch further.
Therefore, reusing the uid dir is an optimized way to solve the race.
Ignore-AOSP-First: Freezer is not a public feature yet
Bug: 192512069
Bug: 168907513
Test: Kill corresponding apps and check the uid cgroupfs dir
Merged-In: I2e91088f21f45e4eda6c709a4af65ace7e135801
Change-Id: I2e91088f21f45e4eda6c709a4af65ace7e135801
In some platforms, setting a different cpuset (or remove it) and/or
uclamp setting is beneficial to both performance and power. Adding 2
separate profiles for vendor to override surfaceflinger scheduling
policies.
Bug: 170507315
Test: build and boot
Signed-off-by: Wei Wang <wvw@google.com>
Change-Id: Ib313fd67c6ccd701109c5683d8dbab9ecd1580d8
(cherry picked from commit 6d6b71e582)
The inner profile should print error message if needed.
Bug: 178070122
Test: Build
Signed-off-by: Wei Wang <wvw@google.com>
Change-Id: I2d9d54e21fb064678feccdbae453a398806d51d6
Issue a warning about missing cpu/schedtune controller only if both of
them are missing.
Bug: 185437398
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
Change-Id: I3a9d3c9a8c91c8d2c5346bcb431bb0407c64a811
Remove the vestigial llndk_library modules and replace them with
properties in the llndk clause of the implementation cc_library.
Bug: 170784825
Test: m checkbuild
Test: compare out/soong/build.ninja
Change-Id: Ie3a1bffcf29bb1b6747f7f708826c61bd43ba5a1
Memory cgroup controller is not used by high-end Android devices and is
disabled in Android kernels 5.10+, however Android Go devices still
require this controller.
Similarly schedtune controller is being replaced with cpu controller's
utilclamp feature, however some devices are using older kernels which
only support schedtune.
To prevent spamming logcat with errors we need a way to mark a
controller optional in such cases without removing it from
cgroups.json file.
Introduce "Optional" attribute to cgroup controller's cgroups.json
description to accomodate these cases. When an optional controller
fails to mount due to lack of support in the kernel it will log an
informational message without generating further errors.
Bug: 181866655
Test: boot and observe libprocessgroup messages in logcat
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
Change-Id: Ib5bcfdd8d338704f3f9d8b326d8ddd88b67915a1
this removes the (temporary) hard-coded path to a cgroup.freeze file
to be used for checking the presence of the cgroup v2 uid/pid hierarchy
for the app freezer. From now on, only the new API must be used (see the
implementation under CacheAppOptimizer).
Bug: 180056337
Test: built without errors
Change-Id: Ifd0bce81a540a86c0c74d3e52fbe88fb731a02e1
configure the cgroup v2 hierarchy for recovery mode, and create uid/pid
groups with attributes following the container cgroup directory.
Bug: 168907513
Test: verified correct pid migration in normal and recovery modes
Change-Id: Idc8b96b4db075383a6a2e523c241b0bc632c7030
With the new uid/pid hierarchy the search path looks into uid_0, which
is always present.
Bug: 168907513
Test: verified that the cgroup v2 hierarchy works correctly in normal
mode
Test: verified that the cgroup v2 hierarchy works correctly in recovery
mode
Change-Id: If33594e4891128148da64bed47e68fa74667013d
This reverts commit 088924af2a.
Bug: 168907513
Test: verified correct function of the cgroup v2 hierarchy in normal and
recovery mode
Change-Id: I6e9d21ebe832326ed5a5b2c356fe8363c1546a80
This path can be used for clients to verify freezer support on a
specific platform.
Bug: 168058155
Test: verified correct functioning with ActivityManager
Change-Id: Ie15feed274ccbbe31228ee98c50eab3f4dd479a7
This reverts commit 7707909ba2.
Bug: 179006802
Test: verified correct functioning of the freezer after reverting
Change-Id: I848146a18a0879010abf5b00e7081f2771059064
a1a0497984
Bug: 151660495
Test: verified proper boot in regular mode and proper working of adb in
recovery
Change-Id: I1276796e982fee932cdea7eb145f20b3f1b3463d
cc_test_host now all run in presubmit without any need for explicit TEST_MAPPING
Change-Id: I379ef754b4cb645976daedcd04eae58f2eb32a75
Test: presubmit
Bug: 177450457
The build system will soon require that symbol files describing APIs
across updatable components (e.g. across APEXes or across sytem/vendor)
have the ".map.txt" suffix. This will trigger the NDK API review in the
gerrit so that backwards incompatible changes are prevented.
Bug: N/A
Test: m nothing
Change-Id: I114563782f8e796365ca141eb5d22d2e203a7f55
If there are known fail pattern, we could disable printing logs to
reduce log spam. Use "LogFailures": true or "LogFailures": false.
Default will be true if not specified.
Bug: 178070122
Test: function works
Change-Id: Ia4081a10daf738294937f459f006fbcc3171481b
The following API was added for the API level 30 and beyond. Currently,
its existence is tested using the null check which is done regardless of
the min sdk version of the compilation unit. (which in turn required us
to mark the API symbol weak regardless of the min sdk version.)
* AGroupController_getFlags
Now, we have a better way of testing the API availability;
__builtin_available. The null check is replaced with the call to the
compiler-provided macro which determines if the code is running in a
version of OS where the API is known to exist.
Bug: 150860940
Bug: 134795810
Test: m
Change-Id: I5d17fc877c553c402a62485b61af574a54658a76
__ANDROID_API__ guards are removed in favor of __INTRODUCED_IN macros.
Currently, __INTRODUCED_IN macro does nothing for these headers (it's
meaningful only to the headers processed by versioner which are limited
to binic headers). The plan is to make the macros to tag the declaration
with the availability attribute. Then, when the min_sdk_version of a
caller is set to an API level that is older than the API level of the
APIs, the compiler will provide them as weak symbols and enforce that
calling the APIs are guarded with a runtime check.
For now, these guards are preventing from making a build system change
to let __ANDROID_API__ track the min_sdk_version property instead of the
sdk_version property. With the build system change, __ANDROID_API__ will
suddenly drop for the native modules where min_sdk_version <
sdk_version, which is the case when the modules are included in APEXes.
As a result, some new APIs will be unavailable at build-time. Dropping
the hand-written guards fixes the problem.
Bug: 163288375
Test: m
Change-Id: If1cc6b9af410f536abe6d777c22711209fa76530
This reverts commit 1414a0d4d8.
Reason for revert: reversions of CLs from http://b/174776875 removed dependencies for this CL, and made the freezer non-functional. This CL has to be reverted as well
Bug:174776875
Change-Id: Idbd9532374bb661330b3c7546bbdf086a046aba3
vendor_available modules were available to product modules.
However, not all vendor_available modules are required to be
available to product modules. Some modules want to be available only
to product modules but not vendor modules.
To cover the requirement, we separate product_available from
vendor_available.
vendor_available will not provide product available module.
Rename libcutils/include_vndk to libcutils/include_outside_system as
it provides the headers to the modules outside the system partition
including /product.
Bug: 150902910
Test: build
Change-Id: I791b5a6a1dc99442065debfce6fa0f54a335fcea
Migrate the freezer controller as an attribute for the uid/pid hierarchy
Bug: 168907513
Test: manually verified that the freezer behavior was as expected after
applying the enablers in ActivityManager
Signed-off-by: Marco Ballesio <balejs@google.com>
Change-Id: Ica2c4bfd20190fb2f2dc0a530eb47e1a8427c522
Current per-API level task profiles replace the latest task profiles
rather than overriding the diffs. This poses an issue when a new
feature is added which is applicable to older devices as well we
have to change all per-API level profiles. Change the per-API task
profile support to load the latest task profiles first and then override
with per-API task profiles, similar to how vendor task profiles override
system task profiles. This minimizes the changes we need to do to the
per-API level task profiles going forward and is more consistent with
how vendor task profiles work.
Bug: 170507876
Test: boot with per-API task profiles
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
Change-Id: I08cc4b374cede7363cf96a30ac96971d49bb7847
Enable the uid/pid hierarchy for all groups when using cgroup v2. Mount
the hierarchy under the cgroup v2 root. Make sure that all files under
the hierarchy are accessible by the system user.
Test: booted the device, tested the freezer cgroup, manually verified
the working of the freezer from logs and by checking statuses of
processes.
Bug: 168907513
Test: Booted the device, verified no regressions on process group access
Change-Id: I73f3e767d377902af6e12facb503b9136fb39e08
When task profiles changes happen, devices released under older API
levels might have to use the same profiles as before. This might be
due to for missing kernel features or some other reasons. Add support
for per-API task profiles to accommodate this scenario. With this
mechanism when ro.product.first_api_level is non-empty, the system
looks for /system/etc/profiles/task_profiles_<api level>.json file
and uses it if it exists. If ro.product.first_api_level is not defined
or per-API task profiles file does not exist the system falls back to
the default /system/etc/profiles/task_profiles.json file.
As before, these task profiles can be competely overwritten using
/vendor/etc/task_profiles.json if needed.
Bug: 172066799
Test: boot with per-API task profiles
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
Change-Id: I1671b341b3a0fb5e99d0b0788315c61088326b09
Put tasks into a subgroup so that we can set things up e.g. soft cpu
bandwidth and/or uclamp.
Bug: 170507876
Bug: 171740453
Test: boot and check cgroup
Change-Id: Iaeb081bc4f214aa23eb281e0c28a81ea581bb4d1
There may be some use cases that it needs to write string to a
file. This patch support write "Value" to "FilePath", where
both parameters could use special string "<uid>" and "<pid>" to
represent actual uid and pid.
Bug: 170507963
Test: function works
Change-Id: I543846f523518a9bcb3dd1b3437163a1b3157d95
Instead of assuming a module with the .llndk suffix exists, add an
llndk_stubs property to every cc_library module that has a
corresponding llndk_library. Also rename the llndk_library to have
an explicit .llndk suffix.
Bug: 170784825
Test: no changes to build.ninja (excluding comments) or Android-${TARGET_PRODUCT}.mk
Change-Id: Ifba79a1ae64a67a9d7393dac2fb012cd8af8e149
cgroups v2 support introduces new fields in the json format. Adapt the
proto file accordingly
Bug: 154548692
Test: atest libprocessgroup_proto_test --
Change-Id: I40f8757a8f4e6a0b839caa7faa976dfebf3aac98
remove cgroup v1 freezer entries from init.rc, add a new cgroup v2
controller and modify plists to properly interact with it.
Bug: 154548692
Test: manually verified the the cgroup v1 freezer controller isn't
created and a new controller for cgroup v2 is created under the correct
sysfs directory.
Change-Id: I1b811300ade486f88fdbd157255a7f37750cc54d
create new profiles to allowing thawing and freezing back the freezer
cgroup
Bug: 151225245
Test: Manually verified that using the SetTaskProfiles method on the
profiles thaws and freezes back the freezer cgroup.
Change-Id: I7f3e193ebe79b49c1f6ac52b6138ff4ec26fc570
In the current implementation vendor profile attributes do not override
system ones and instead generate a warning. Fix that by overriding existing
attribute if a new definition is found.
Bug: 160318642
Test: add vendor attributes and confirm no warnings
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
Change-Id: I71a2ee4d4b3c585e7c9a01b791e973390d409cbc
for a first implementation the cgroup v2 freezer controller will be used in a
way similar to cgroup v1, that is a single child group will hold all frozen
processes. Some adjustments are needed for the new structure.
- Add support for cgroup v2 syntax under procfs.
- Separate creation of a directory with ownership/mode changes to allow changes
after mounting the cgroup kernfs root.
- Allow the creation of sub-groups under a cgroup v2 hierarchy.
Bug: 154548692
Test: manually verified that a proper cgroup v2 hierarchy is created and
accessible
Change-Id: I9af59e8214acaead3f520a94c95e75394c0df948
Modules contributing mainline modules (APK/APEX) should set
min_sdk_version as well as apex_available.
For now setting min_sdk_version doesn't change build outputs.
But build-time checks will be added soon.
Bug: 152655956
Test: m
Change-Id: Idaa27ce24cc13d13145eaef85a08b9033922c7bc
The marked library(ies) were available to the APEXes via the
hand-written whitelist in build/soong/apex/apex.go. Trying to remove the
whitelist by adding apex_available property to the Android.bp of the
libraries.
Bug: 150999716
Test: m
Change-Id: Id08058f3105463558c6c82298344a33af0efb2d3
Because we cache file descriptors associated with cgroup "tasks" file it
should not be used with SetProcessProfiles API which operates on entire
processes rather than tasks. Change SetProcessProfiles API to prevent
cache fd usage, modify ExecuteForProcess to not attempt to use cached
fd. Also fix unconditional calls to EnableResourceCaching from
ExecuteForTask which should be called only when SetTaskProfiles is used
with use_fd_cache set to true.
Bug: 149524788
Change-Id: I880efaf8217a4dd7ccfbb4fb167b2295cefc057a
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
task_profiles.json currently uses attributes of a defunct version of
uclamp, which was used for testing.
Let's use the attributes that actually got merged upstream instead.
Bug: 143710295
Change-Id: Ic6a14f52e14efeefc91f6e6bf580c15e2eded82e
Signed-off-by: Quentin Perret <qperret@google.com>
ACgroupController_getFlags was reverted due to LLNDK breakage, however
it allows detection of cgroups that failed to mount in a more efficient
way. Revive the function as a weakly linked symbol to allow for it to
be missing in case older LLNDK library is being used with the new
system software. This effectively reverts the commit
aa1d54f0cc "Remove ACgroupController_getFlags to fix API breakage"'
except it declares ACgroupController_getFlags function as weak and
targets it for API level 30. If LLNKD library does not contain
ACgroupController_getFlags the behavior falls back to the current
way of identifying cgroups that failed to mount.
Test: build and verify correct operation with a missing cgroup
Change-Id: I9158ef53aba97972d41d71dd3396ac43796a7004
Merged-In: I9158ef53aba97972d41d71dd3396ac43796a7004
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
Currently when vendor overrides a profile the profile object is being
replaced with a new one. However the old profile might have been
referenced by an aggregate profile and with such profile replacement
the aggregate profile is left referencing a stale object. Fix this by
replacing the content of the old profile with the content from the new
one instead of replacing the object itself.
Bug: 148311066
Test: override profiles referenced in aggregate profile and verify
Test: correct replacement
Change-Id: Iabddbf3580455e5263fedad6665cf52fb323e50a
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
In the current state, schedboost_enabled() is true if and only if
schedtune is in use. As a result, all tests conditioned by
schedboost_enabled() will be skipped on devices using uclamp since it is
and extension of the CPU controller.
Fix this by making schedboost_enabled() return true if either schedtune
or the CPU controller is enabled.
Bug: 44953631
Change-Id: Idaadf252c9cf411a176180ab8988d559ca8a1332
Signed-off-by: Quentin Perret <qperret@google.com>
~2007 a change was added that would allow oneshot services to
daemonize by not killing their process group. This was a hack at the
time, and should certainly not be needed now. I've resisted removing
the behavior however, as it hadn't caused any issues.
Recently, it was detected that the cgroups that these processes belong
to, would exist forever and therefore leak memory. Instead of simply
removing the cgroups when empty, this provides a good opportunity to
do the right thing and fix this behavior once and for all.
The new (correct) behavior only happens for devices with vendor images
built for Android R or later. Init will log a warning to dmesg when
it detects this difference in behavior has occurred.
Bug: 144545923
Test: boot CF/Coral and see no difference in behavior.
Test: boot CF with a service that daemonizes and see the warning.
Change-Id: I333a2e25a541ec0114ac50ab8ae7f1ea3f055447
Adds the freezer cgroup to process groups
Change-Id: Ib7a8dbe776ff156ff3827b9a659365384f3e6ac8
Bug: 143308662
Test: manual - frozen processes aren't scheduled for execution until
unfrozen
Expose API to return name of task profile for cpuset/sched policy so
that libprocessgroup clients using SetTaskProfiles directly don't have
to maintain the mapping. This reduces the risk of inconsistency and saves
memory.
Bug: 139521784
Test: atest libcutils_test:SchedPolicy
Change-Id: I414312a038613913fb6a827bdcefceb3dec21264
To support setting multiple profiles with one call. The json format
is as below example.
"AggregateProfiles": [
...
{
"Name": "SCHED_SP_BACKGROUND",
"Profiles": [ "HighEnergySaving", "LowIoPriority", "TimerSlackHigh" ]
},
...
}
Bug: 139521784
Test: SetProfile works as expected
Change-Id: Ibe14ed57d5169cafcbcbbdb054df3ed171a2f6a2
This reverts commit e8678cf883.
With b/131214885 resolved, we should reland the change for bg sched group.
If a thread explicitly calls set_sched_policy to set a thread to background, it should not be latency sensitive.
This would help background I/O impacting foreground I/O e.g. b/65727988
Test: build and boot
Bug: 140209629
Change-Id: I218db084fb733cad57af8a5eae926818a6a2cdb8
CgroupController::IsUsable is using a global static variable to store the
existence of the controller. That means the first controller existence
check would affect all other controllers. Fix this by making this variable
to be a member of CgroupController class so that each controller can check
for its existence independently of other controllers.
Fixes: aa1d54f0cc ("Remove ACgroupController_getFlags to fix API breakage")
Bug: 136020193
Test: adb shell cat /proc/$pid/task/*/cgroup" prints "cpuset:/top-app"
Test: for new launched activity process
Change-Id: I4741a9126ea494122d5b2b1a0c4d7252bff6025c
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
A process can give up the permission to set cgroup. If we still
keep the fd that was cached before losing the permission, when
the process sets scheduling group, it will write to the cached
fd without checking if is accessible and lead to sepolicy denied.
Bug: 123043091
Test: Build and boot.
Test: A new process from zygote set cgroup and drop fd cache, and
then specializes to app domain. There is no sepolicy denied
when the process creates new thread.
(android::thread_data_t::trampoline)
Change-Id: I285ee91424ea965ea9c670fc0f6662948e3e2ce5
Since ACgroupController_getFlags was introduced after LLNDK freeze it
causes media crashes due to missing symbol. Remove the new function and
re-implement cgroup controller detection to not require flags field but
instead check for its existence on the first access.
Bug: 135049992
Test: libcutils_test with and without CONFIG_CPUSETS enabled
Change-Id: I0220d6a926884dc22a7424d7d0a980c379c6f4eb
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
