This repository contains the software artifact needed for paper "Incremental CFG Patching for Binary Rewriting".
We use x86 as the example.
git clone https://github.com/mxz297/Incremental-CFG-Patching-ASPLOS21-AE.git
cd Incremental-CFG-Patching-ASPLOS21-AE/
export AEROOT=`pwd`
cd setup
./build-x86.sh
There should be a file named spec-config-paths.txt, which contains necessary paths needed for running SPEC CPU 2017.
We use Spack (https://github.com/spack/spack) to build software dependencies. Spack by default will build packages under /tmp. If /tmp does not have sufficient space, you can change the build directory for Spack by following this section of Spack's documentation (https://spack.readthedocs.io/en/latest/configuration.html). Search for build_stage in your config.yaml file and add path that contains sufficient space.
Suppose the SPEC CPU 2017 benchmark is available in $AEROOT/spec2017/spec_cpu2017. We expect a shrc file in the directory, for which we can source to setup necessary environments for running SPEC CPU 2017.
First, after source shrc, test if you have all necessary library dependencies available on your system
runcpu
There should be no error.
Second, copy the provided config file to the proper directory and edit the config file:
cp $AEROOT/spec2017/x86/code-patching-x86.cfg $AEROOT/spec2017/spec_cpu2017/config
vim $AEROOT/spec2017/spec_cpu2017/config
The most important editing here is to fill in the paths needed for running experiments. Search for ASPLOS21AEEDIT and there should be two matches. One of them is to change the number of concurrent threads for running the experiments and it should look like:
# ASPLOS21AEEDIT: number of OpenMP threads
intspeed,fpspeed:
threads = 8 # EDIT to change number of OpenMP threads (see above)
The other match is to setup the paths and it should look like:
#------- Compilers ------------------------------------------------------------
default:
# ASPLOS21AEEDIT: Copy and paste genearted path
%define gcc_dir /projects/spack/opt/spack/linux-rhel7-x86_64/gcc-4.8.5/gcc-7.3.0-qrjpi76aeo4bysagruwwfii6oneh56lj
%define boost_lib /home/xm13/spack/opt/spack/linux-rhel7-power9le/gcc-6.4.0/boost-1.73.0-2uqm5ox4miwer5tqwuutcss4wbsxhltd/lib
%define elfutils_lib /home/xm13/spack/opt/spack/linux-rhel7-power9le/gcc-6.4.0/elfutils-0.179-duvfsqtjwguwyo77czviqtitmqlozmoa/lib
%define tbb_lib /home/xm13/spack/opt/spack/linux-rhel7-power9le/gcc-6.4.0/intel-tbb-2020.2-tpdywwttxoj4dbsinjdbxlffbtovf3nd/lib
%define libunwind_lib /home/xm13/spack/opt/spack/linux-rhel7-power9le/gcc-6.4.0/libunwind-1.4.0-q6cg7tx62x62cjpe4lzw3ljlqlodcbdo/lib64
%define dyninst_lib /home/xm13/dyninstapi/install/lib
%define dyninst_mutator /home/xm13/projects/code-patching-work-repo/BlockTrampoline/BlockTrampoline
%define dyninst_release_lib /home/xm13/dyninst-master/install/lib
%define dyninst_release_mutator /home/xm13/dyninst-master/FuncReloc
Now open file $AEROOT/setup/spec-config-paths.txt. You should be able to just copy and paste paths from spec-config-paths.txt to the SPEC config file.
We should be ready to run the experiments:
cd $AEROOT/spec2017/x86
./run_spec.sh $AEROOT/spec2017/spec_cpu2017 10
The run_spec.sh script takes two input the parameters. The first one is the root directory path to the SPEC CPU 2017 benchmark. The second one is the number of iterations to run each benchmark. The results reported in our paper are based on 10 iterations, which can take a week to finish. You can reduce the number of iterations or increase the number of concurrent threads for running the benchmarks.
Finally, we parse the result files from SPEC:
cd $AEROOT/spec2017
./run_result.sh $AEROOT/spec2017/spec_cpu2017/result
Script run_result.sh takes one parameter, which is the result directory of SPEC CPU 2017. The script will first print the overhead columns and then the coverage columns.
We use the rewriting tool BlockTrampoline to rewrite libxul.so, which is available at $AEROOT/setup/BlockTrampoline.
We provide two modes to rewrite the libxul.so.
Firefox (version 80.0) is usually shipped with the latest Ubuntu 18.04 dist. To install it manually, one can visit here and choose the version 80.0 for this evaluation.
To rewrite libxul.so,
cd $AEROOT/firefox
./run_inst.sh /lib64/firefox/libxul.so
Script run_inst.sh takes one parameter, which is the path to Firefox's libxul.so. It will generate libxul.so.jt. This step may print some messages to stderr such as:
ERROR: jump table relocation twice for address 635e234, old value 5157fd4, new value 5159d7e.
These error messages can be ignored.
We use two web-browser-base benchmarks. Please exercise with cautions when replacing original binaries (e.g., libxul.so and docker). Always prepare a backup for the evaluation.
Benchmark Installation.
wget http://google.github.io/latency-benchmark/latency-benchmark-linux.zip
unzip latency-benchmark-linux.zip
First, just run Web Latency Benchmark without any instrumentation to verify that Firefox is working properly.
./latency-benchmark
Then, replace the original libxul.so with the generated libxul.so.jt and rerun the benchmark:
source env.sh
./latency-benchmark
The script env.sh will setup necessary environment for running the instrumented version.
The `Image loading jank' cannot be measured in the stable versions of Firefox (79.0 and 80.0) by Web Latency Benchmark.
The benchmark result will be displayed in the web broswer. We do not have scripts to capture numbers from the web broswer. The overhead numbers are calculated manually.
Benchmark Installation.
Type https://browserbench.org/JetStream/ in Firefox search box.
Run Jetstream2.
Click the Start Test button. The benchmark will automatically run for 120 times and display the final results.
Docker Installation guide can be found at here.
To rewrite docker,
cd $AEROOT/docker
./run_inst.sh /usr/bin/docker
Script run_inst.sh takes one parameter, which is the path to docker executable. It will generate docker.inst. We provide a run_docker.sh script, which contains several basic docker commands as correctness test:
source env.sh
./run_docker.sh