|
OpenCSD - CoreSight Trace Decode Library 1.4.0
|
Using CoreSight trace and perf with OpenCSD for AutoFDO.
Feedback directed optimization (FDO, also know as profile guided optimization - PGO) uses a profile of a program's execution to guide the optmizations performed by the compiler. Traditionally, this involves building an instrumented version of the program, which records a profile of execution as it runs. The instrumentation adds significant runtime overhead, possibly changing the behaviour of the program and it may not be possible to run the instrumented program in a production environment (e.g. where performance criteria must be met).
AutoFDO uses facilities in the hardware to sample the behaviour of the program in the production environment and generate the execution profile. An improved profile can be obtained by including the branch history (i.e. a record of the last branches taken) when generating an instruction samples. On Arm systems, the ETM can be used to generate such records.
The process can be broken down into the following steps:
This article describes how to enable ETM trace on Arm targets running Linux and use the ETM trace to generate AutoFDO profiles and compile an optimized program.
Debug and trace of Arm targets is provided by CoreSight. This consists of a set of components that allow access to debug logic, record (trace) the execution of a processor and route this data through the system, collecting it into a store.
To record the execution of a processor, we require the following components:
Each Arm SoC design may have a different layout (topology) of components. This topology is described to the OS drivers by the platform's devicetree or (in future) ACPI firmware.
For application profiling, we need to store several megabytes of data within the system, so will use ETR with the capture tool (perf) periodically draining the buffer to a file.
Even though we have a large capture buffer, the ETM can still generate a lot of data very quickly - typically an ETM will generate ~1 bit of data per instruction (depending on the workload), which results in 256Mbytes per second for a core running at 2GHz. This leads to problems storing and decoding such large volumes of data. AutoFDO uses samples of program execution, so we can avoid this problem by using the ETM's features to only record small slices of execution - e.g. collect ~5000 cycles of data every 50M cycles. This reduces the data rate to a manageable level - a few megabytes per minute. This technique is known as 'strobing'.
To collect ETM trace, the CoreSight drivers must be included in the kernel. Some of the driver support is not yet included in the mainline kernel and many targets are using older kernels. To enable CoreSight trace on these targets, Arm have provided backports of the latest CoreSight drivers and ETM strobing patch at:
https://gitlab.arm.com/linux-arm/linux-coresight-backports
This repository can be cloned with:
You can include these backports in your kernel by either merging the appropriate branch using git or generating patches (using git format-patch).
For 5.x based kernel onwards, the only patch which needs to be applied is the one enabling strobing - etm4x: Enable strobing of ETM.
For 4.9 based kernels, use the coresight-4.9-etr-etm_strobe branch:
or
For 4.14 based kernels, use the coresight-4.14-etm_strobe branch:
or
The CoreSight trace drivers must also be enabled in the kernel configuration. This can be done using the configuration menu (make menuconfig), selecting Kernel hacking / arm64 Debugging /CoreSight Tracing Support and enabling all options, or by setting the following in the configuration file:
Compile the kernel for your target in the usual way, e.g.
Each target may have a different layout of CoreSight components. To collect trace into a sink, the kernel drivers need to know which other devices need to be configured to route data from the source to the sink. This is described in the devicetree (and in future, the ACPI tables). The device tree will define which CoreSight devices are present in the system, where they are located and how they are connected together. The devicetree for some platforms includes a description of the platform's CoreSight components, but in other cases you may have to ask the platform/SoC vendor to supply it or create it yourself (see Appendix: Describing CoreSight in Devicetree).
Once the target has been booted with the devicetree describing the CoreSight devices, you should find the devices in sysfs:
The naming convention for etm devices can be different according to the kernel version you're using. For more information about the naming scheme, please check out the Linux Kernel Documentation
If /sys/bus/coresight/devices/ is empty, you may want to check out your Kernel configuration to make sure your .config file is including CoreSight dependencies, such as the clock.
The perf tool is used to capture execution trace, configuring the trace sources to generate trace, routing the data to the sink and collecting the data from the sink.
Arm recommends to use the perf version corresponding to the kernel running on the target. This can be built from the same kernel sources with
When specifying CORESIGHT=1, perf will be built using the installed OpenCSD library. If you are cross compiling, then additional setup is required to ensure the build process links against the correct version of the library.
If the post-processing (perf inject) of the captured data is not being done on the target, then the OpenCSD library is not required for this build of perf.
Trace is captured by collecting the cs_etm event from perf. The sink to collect data into is specified as a parameter of this event. Trace can also be restricted to user space or kernel space with 'u' or 'k' parameters. For example:
Will record the userspace execution of '/bin/ls' using tmc_etr0 as sink.
You can trace a single-threaded program in two different ways:
--per-thread, and in this case the CoreSight subsystem will record only a trace relative to the given program.--per-thread, and in this case CPU-wide tracing will be enabled. In this scenario the trace will contain both the target program trace and other workloads that were executing on the same CPUperf is also used to convert the execution trace an instruction profile. This requires a different build of perf, using the version of perf from Linux v4.17 or later, as the trace processing code isn't included in the driver backports. Trace decode is provided by the OpenCSD library (https://github.com/Linaro/OpenCSD), v0.9.1 or later. This is packaged for debian testing (install the libopencsd0, libopencsd-dev packages) or can be compiled from source and installed.
The autoFDO tool https://github.com/google/autofdo is used to convert the instruction profiles to source profiles for the GCC and clang/llvm compilers.
Once trace collection using perf is working, we can now use it to profile an application.
The application must be compiled to include sufficient debug information to map instructions back to source lines. For GCC, use the -g1 or -gmlt options. For clang/llvm, also add the -fdebug-info-for-profiling option.
perf identifies the active program or library using the build identifier stored in the elf file. This should be added at link time with the compiler flag -Wl,--build-id=sha1.
The next step is to record the execution trace of the application using the perf tool. The ETM strobing should be configured before running the perf tool. There are two parameters:
For example, a typical configuration is to use a window size of 5000 cycles and a period of 10000 - this will collect 5000 cycles of trace every 50M cycles. With these proof-of-concept patches, the strobe parameters are configured via sysfs - each ETM will have strobe_window and strobe_period parameters in /sys/bus/coresight/devices/<sink> and these values will have to be written to each (In a future version, this will be integrated into the drivers and perf tool). The set_strobing.sh script in this directory [<opencsd>/decoder/tests/auto-fdo] automates this process.
To collect trace from an application using ETM strobing, run:
The raw trace can be examined using the perf report command:
Perf needs to be built from your linux kernel version souce code repository against the OpenCSD library in order to be able to properly read ETM-gathered samples and post-process them. If running perf report produces an error like:
or
You are probably using a perf version which is not using this library: please make sure to install this project in your system by either compiling it from Source Code from v0.9.1 or later and compile perf using this library. Otherwise, this project is packaged for debian (install the libopencsd0, libopencsd-dev packages).
For example:
The execution trace is then converted to an instruction profile using the perf build with trace decode support. This may be done on a different machine than that which collected the trace (e.g. when cross compiling for an embedded target). The perf inject command decodes the execution trace and generates periodic instruction samples, with branch histories:
!! Careful: if you are using a device different than the one used to collect the profiling data, you'll need to run perf buildid-cache as described below.
The --itrace option configures the instruction sample behaviour:
i100000i generates an instruction sample every 100000 instructions (only instruction count periods are currently supported, future versions may support time or cycle count periods)l includes the branch histories on each sampleb generates a sample on each branch (not used here)Perf requires the original program binaries to decode the execution trace. If running the inject command on a different system than the trace was captured on, then the binary and any shared libraries must be added to perf's cache with:
perf report can also be used to show the instruction samples:
The instruction samples produced by perf inject is then passed to the autofdo tool to generate source level profiles for the compiler. For clang/LLVM:
And for GCC:
The profiles can be viewed with:
Or, for GCC:
The profile produced by the above steps can then be passed to the compiler to optimize the next build of the program.
For GCC, use the -fauto-profile option:
For Clang, use the -fprofile-sample-use option:
The basic commands to run an application and create a compiler profile are:
Use create_gcov for gcc.
sudo ./set_strobing.sh 5000 10000 perf record -e cs_etm/@tmc_etr0/u --per-thread -- <your app>.If you specify
-N, --no-buildid-cache, perf will just take care of recording the target binary and nothing will be copied.perf archive which saves all the found libraries in a tar (internally, it looks into perf.data file and performs a lookup using perf-buildid-list –with-hits)scp to copy perf.data and the .tar file generated from perf archive.tar xvf perf_data.tar.bz2 -C ~/.debug to populate the buildid-cachea. `perf buildid-list -i perf.data` gives you the list of dynamic libraries buildids whose trace has been recorded and saved in perf.data. b. `perf buildid-cache --list` lists the dynamic libraries in the buildid cache that will be used by `perf inject`.Make sure the output of (a) and (b) overlaps as in buildid value for those binaries you are interested into optimizing with afdo.
perf inject -i perf.data -o inj.data --itrace=i100000il will check for the dynamic libraries using the buildid inside the buildid-cache and post-process the trace.create_llvm_prof -binary=/path/to/binary -profile=inj.data -out=program.llvmprof takes the output from perf-inject and tranforms it into a format that the compiler can read.clang -O2 -fprofile-sample-use=program.llvmprof -o program program.c to make clang use the produced profile.-fprofile-sample-accurate flag, which will penalize all the callsites without corresponding profile, marking them as cold.If you are using the same host for both building the binary to be traced and re-building it with afdo:
--no-buildid-cache when recordingYou can easily add a dynamic library manually into the build-id cache by running:
perf buildid-cache --add <path/to/library/or/binary> -vvv
You can easily check what is currently contained in you buildid-cache by running:
perf buildid-cache --list
You can check the buildid of a given binary/dynamic library:
file <path/to/dynamic/library>
Each component has an entry in the device tree that describes its:
compatible field defines which driver to usereg defines the component's address and size on the busclocks and clock-names fields state which clock provides the apb_pclk clock.port and ports field link the component to ports of other componentsTo create the device tree, some information about the platform is required:
This information can be found in the SoC's reference manual or you may need to ask the platform/SoC vendor to supply it.
An ETMv4 source is declared with a section like this:
This describes an ETMv4 attached to core A72_0, located at 0x22040000, with its output linked to port 0 of a funnel. The funnel is described with:
This describes a funnel located at 0x220c0000, receiving data from 2 ETMs and sending the merged data to another funnel. We continue describing components with similar blocks until we reach the sink (an ETR):
Full descriptions of the properties of each component can be found in the Linux source at Documentation/devicetree/bindings/arm/coresight.txt. The Arm Juno platform's devicetree (arch/arm64/boot/dts/arm) provides an example description of CoreSight description.
Many systems include a TPIU for off-chip trace. While this isn't required for self-hosted trace, it should still be included in the devicetree. This allows the drivers to access it to ensure it is put into a disabled state, otherwise it may limit the trace bandwidth causing data loss.
1.9.4