The source code for barman
, gatord
, and related tools.
The barman
subdirectory contains the sources for the barman embedded agent which
can be used to collect performance data within an embedded environment as per the
Arm Streamline Target Setup Guide for Bare-metal Applications.
The rest of this document refers to gatord
and related tools.
daemon
, andnotify
are provided under GPL-2.0-only. See daemon/COPYING, and notify/COPYING respectively.annotate
,barman
andstreamline_me.py
are provided under the BSD-3-Clause license. See annotate/LICENSE.
This project contains code from other projects listed below. The original license text is included in those source files.
libsensors
source code in daemon/libsensors licensed under LGPL-2.1-or-laterperf_event.h
from Linux userspace kernel headers in daemon/k licensed under GPL-2.0-only WITH Linux-syscall-note
The pre-built gatord
shipped with Streamline uses musl. For musl license
information see the COPYRIGHT file shipped with Streamline, or
https://git.musl-libc.org/cgit/musl/tree/COPYRIGHT
Contributions are accepted under the same license as the associated subproject with developer sign-off as described in Contributing.
Instructions on setting up Arm Streamline on the target.
A target agent (gator) is required to run on the Arm Linux target in order for Arm Streamline to operate. Gator requires Linux kernel version 3.4 or later.
A Linux development environment with cross compiling tools is most likely required, depending on what is already created and provided.
Please see release notes for information about changes in this release.
Gator uses the Linux Perf API (perf_event_open) for most of its data collection. Additionally it will use ftrace tracepoints and some other common features such as debugfs/sysfs.
Most users will not need to make any changes to their kernel configuration (and in many cases they cannot) as most recent Android devices and Linux distributions correctly configure their kernel with the required options.
If you are a system integrator, or compiling your own kernel, refer to the section [Kernel configuration options].
Streamline provides pre-built binaries for aarch64 and armv7a-hardfloat Linux and Android. This gator daemon should work in most cases so building the gator daemon is only required if a non-standard configuration is required.
To improve portablility gatord is statically compiled against musl libc from http://www.musl-libc.org/download.html instead of glibc. The gator daemon will work correctly with either glibc or musl.
Building gatord has the following requirements:
- C++17 supporting compiler.
- CMake (3.16 or later).
- GCC or Clang compiler able to target the appropriate target architecture.
- GCC or Clang compiler able to target the host architecture if cross compiling.
- For Android, the Android NDK (LTS r21e, r23b are tested and known to work).
- A Linux build environment (other CMake compatible enironments may work but are not tested).
- Additionally, vcpkg depends on various unix tools being installed.
-- A minimal build environment for linux can be achieved on Ubuntu 20.04
with:
sudo apt-get install ninja-build cmake gcc g++ g++-aarch64-linux-gnu curl zip unzip tar pkg-config git
The most convenient option is to use the provided build-android.sh
script.
./build-android.sh -h
Prints a summary of the available configuration options.
In most cases it should be possible to run:
./build-android.sh
which will compile gatord for Android targetting aarch64 devices.
Using the configuration options, it is possible to change the minimum SDK level, architecture, CMake binary to use, CMake generator, build directory, NDK path.
For simple configurations, the most convenient option is to use the provided
build-linux.sh
script. This allows selection of one of a few predefined configurations:
- Building using the host native Clang toolchain.
- Building using the host native GCC toolchain.
- Building using GCC targetting aarch64 or armv7a against the glibc or musl based libc.
./build-linux.sh -h
Prints a summary of the available configuration options.
When natively compiling it should be possible to run:
./build-linux.sh
Otherwise the typical use is to pass a profile option using -p
.
Since the build is CMake based, it is possible to invoke cmake directly. This option requires some understanding of vcpkg and cmake tools.
Please note, the section on telemetry that vcpkg collects by default. The provided
build-android.sh
and build-linux.sh
scripts disable this by default.
It is possible to build with out using vcpkg at all, by passing -DENABLE_VCPKG=OFF
to the cmake build, but it will be necessary to provide precompiled versions of
any dependencies.
- Copy gatord into the target's filesystem.
- Ensure gatord has execute permissions:
chmod +x gatord
- The daemon must be run with root privileges:
sudo su
gatord &
This configuration requires Linux 3.4 or later with a correctly configured kernel.
- Copy gatord into the target's filesystem.
- Ensure gatord has execute permissions:
chmod +x gatord
- Run the daemon:
./gatord &
This configuration provides a reduced set of software only CPU counters such as CPU utilization and process statistics, as well as Mali hardware counters on supported Mali platforms.
To check the perf PMUs support by your kernel, run
ls /sys/bus/event_source/devices/
If you see something like ARMv7_Cortex_A##
this indicates A## support. If you
see CCI_400
this indicates CCI-400 support. If you see ccn
, it indicates
CCN support.
CCN requires a perf driver to work. The necessary perf driver has been merged into Linux 3.17 but can be backported to previous versions (see https://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/diff/?id=a33b0daab73a0e08cc04459dd44b0121a8e8f81b and later bugfixes)
Recommended compiler settings:
-g
: Debug information, such as line numbers, needed for best analysis results.-fno-inline
: Speed improvement when processing the image files and most accurate analysis results.-fno-omit-frame-pointer
: Arm EABI frame pointers allow recording of the call stack with each sample taken when in Arm state (i.e. not-mthumb
).-marm
: This option is required for ARMv7 and earlier if your compiler is configured with--with-mode=thumb
, otherwise call stack unwinding will not work.
For Android ART, passing --no-strip-symbols
to dex2oat will result in
function names but not line numbers to be included in the dex files. This can be
done by running setprop dalvik.vm.dex2oat-flags --no-strip-symbols
on the device and then regenerating the dex files.
Gator supports reading arbitrary /dev
, /sys
and /proc
files 10 times a
second. It will either interpret the file contents as a number or use a POSIX
extended regex to extract the number, see events-Filesystem.xml
for
examples.
The following options are required for correct functioning of Gator.
menuconfig options (depending on the kernel version, the location of these configuration settings within menuconfig may differ)
- General Setup
- Timers subsystem
- [*] High Resolution Timer Support (enables CONFIG_HIGH_RES_TIMERS)
- Kernel Performance Events And Counters
- [*] Kernel performance events and counters (enables CONFIG_PERF_EVENTS)
- [*] Profiling Support (enables CONFIG_PROFILING)
- Timers subsystem
- Kernel Features
- [*] Use local timer interrupts (only required for SMP and for version before Linux 3.12, enables CONFIG_LOCAL_TIMERS)
- [*] Enable hardware performance counter support for perf events (enables CONFIG_HW_PERF_EVENTS)
- CPU Power Management
- CPU Frequency scaling
- [*] CPU Frequency scaling (enables CONFIG_CPU_FREQ)
- CPU Frequency scaling
- Kernel hacking
- [*] Compile the kernel with debug info (optional, enables CONFIG_DEBUG_INFO)
- [*] Tracers
- [*] Trace process context switches and events (#)
(#) The "Trace process context switches and events" is not the only option that enables tracing (CONFIG_GENERIC_TRACER or CONFIG_TRACING as well as CONFIG_CONTEXT_SWITCH_TRACER) and may not be visible in menuconfig as an option if other trace configurations are enabled. Other trace configurations being enabled is sufficient to turn on tracing.
The configuration options:
- CONFIG_MODULES and MODULE_UNLOAD (not needed if the gator driver is built into the kernel)
- CONFIG_GENERIC_TRACER or CONFIG_TRACING
- CONFIG_CONTEXT_SWITCH_TRACER
- CONFIG_PROFILING
- CONFIG_HIGH_RES_TIMERS
- CONFIG_LOCAL_TIMERS (for SMP systems and kernel versions before 3.12)
- CONFIG_PERF_EVENTS and CONFIG_HW_PERF_EVENTS (kernel versions 3.0 and greater)
- CONFIG_DEBUG_INFO (optional, used for analyzing the kernel)
- CONFIG_CPU_FREQ (optional, provides frequency setting of the CPU)
These may be verified on a running system using /proc/config.gz
(if this file exists) by running zcat /proc/config.gz | grep <option>
.
For example, confirming that CONFIG_PROFILING is enabled
> zcat /proc/config.gz | grep CONFIG_PROFILING
CONFIG_PROFILING=y
If a device tree is used it must include the pmu bindings, see Documentation/devicetree/bindings/arm/pmu.txt for details.
Kernels with CONFIG_CPU_PM
enabled may produce invalid results on kernel
versions prior to 4.6. The problem manifests as counters not showing any data, large
spikes and non-sensible values for counters (e.g. Cycle Counter reading as very
high).
This issue stems from the fact that the kernel PMU driver does not save/restore
state when the CPU is powered down/up. This issue is fixed in 4.6 so to resolve the
issue either upgrade to a later kernel, or apply the fix to an older kernel.
The patch for 4.6 that resolves the issue is found here
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=da4e4f18afe0f3729d68f3785c5802f786d36e34 -
this patch has been tested as applying cleanly to 4.4 kernel and it may be
possible to back port it to other versions as well.
Users of this patch may also need to apply
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=cbcc72e037b8a3eb1fad3c1ae22021df21c97a51
as well.
There is a bug in some Linux kernels where an Oops may occur when a core is offlined (user space gator only). The fix was merged into mainline in 3.14-rc5, see http://git.kernel.org/tip/e3703f8cdfcf39c25c4338c3ad8e68891cca3731, and has been backported to older kernels (3.4.83, 3.10.33, 3.12.14 and 3.13.6).
CPU PMU: CPUx reading wrong counter -1
in dmesg. To work around,
update to the latest Linux kernel.
Scheduler switch resolutions are on exact millisecond boundaries. To work around, update to the latest Linux kernel.
There is a bug in some Linux kernels where perf misidentifies the CPU type. To see
if you are affected by this, run ls /sys/bus/event_source/devices/
and
verify the listed processor type matches what is expected. For example, an A9 should
show the following.
# ls /sys/bus/event_source/devices/
ARMv7_Cortex_A9 breakpoint software tracepoint
To work around the issue try upgrading to a later kernel.
On some versions of Android, annotations may not work unless SELinux is disabled by
running
# setenforce 0
Some targets do not correctly emit uevents when cores go on/offline. This will cause
CPU Activity with user space gator to be either 0% or 100% on a given core and the
Heat Map may show a large number of unresolved processes. There is no user accessible
workaround. To test for this run
# ./gatord -d | grep uevent
When cores go on/offline with user space gator something similar to the following
should be emitted
INFO: read(UEvent.cpp:61): uevent: offline@/devices/system/cpu/cpu1
INFO: read(UEvent.cpp:61): uevent: online@/devices/system/cpu/cpu1
The cores that are on/offline can be checked by running
# cat /sys/devices/system/cpu/cpu*/online
This issue affects a given target if the on/offline cores shown by the cat command
change but no cpu uevent is emitted.
On some older versions of Android, the following issue may occur when starting gatord when using ndk-build
# ./gatord
[1] + Stopped (signal) ./gatord
#
[1] Segmentation fault ./gatord
#
Starting with Android-L only position independent executables (pie) are supported, but some older versions of Android do not support them. To avoid this issue, modify Android.mk and remove the references to pie.
CONFIG_DEBUG_INFO must be enabled, see "Kernel configuration" section above.
Use vmlinux as the image for debug symbols in Streamline.
Drivers may be profiled using this method by statically linking the driver into the kernel image or adding the driver as an image to Streamline.