ros-clang-instrumentation

Towards Static Analysis and Instrumentation for Software using Robot Operating System

This is the code for a class project for CSCI-610, Advanced Program Analysis and Verification, by Prof. William G. J. Halfond, Fall 2016.

The goal of the project is to use the static analysis capabilities of clang to instrument C++ ROS nodes for performance and data-flow analysis across ROS nodes.

Installation

Install Ubuntu 16.04

Update the system & install dependencies

sudo apt update
sudo apt upgrade
sudo apt install git libyaml-cpp-dev

Install ROS (taken from http://wiki.ros.org/kinetic/Installation/Ubuntu)

sudo sh -c 'echo "deb http://packages.ros.org/ros/ubuntu $(lsb_release -sc) main" > /etc/apt/sources.list.d/ros-latest.list'
sudo apt-key adv --keyserver hkp://ha.pool.sks-keyservers.net:80 --recv-key 421C365BD9FF1F717815A3895523BAEEB01FA116
sudo apt-get update
sudo apt-get install ros-kinetic-ros-base
sudo rosdep init
rosdep update
echo "source /opt/ros/kinetic/setup.bash" >> ~/.bashrc
source ~/.bashrc

Install clang (& fix cmake issues on Ubuntu 16.04, see iovisor/bcc#492)

sudo apt install -y llvm-3.8-dev libclang-3.8-dev clang
sudo mkdir -p /usr/lib/llvm-3.8/share/llvm
sudo ln -s /usr/share/llvm-3.8/cmake /usr/lib/llvm-3.8/share/llvm/cmake
sudo sed -i -e '/get_filename_component(LLVM_INSTALL_PREFIX/ {s|^|#|}' -e '/^# Compute the installation prefix/i set(LLVM_INSTALL_PREFIX "/usr/lib/llvm-3.8")' /usr/lib/llvm-3.8/share/llvm/cmake/LLVMConfig.cmake
sudo sed -i '/_IMPORT_CHECK_TARGETS Polly/ {s|^|#|}' /usr/lib/llvm-3.8/share/llvm/cmake/LLVMExports-relwithdebinfo.cmake
sudo sed -i '/_IMPORT_CHECK_TARGETS sancov/ {s|^|#|}' /usr/lib/llvm-3.8/share/llvm/cmake/LLVMExports-relwithdebinfo.cmake
sudo ln -s /usr/lib/x86_64-linux-gnu/libLLVM-3.8.so.1 /usr/lib/llvm-3.8/lib/

Clone this repository

git clone https://github.com/whoenig/ros-clang-instrumentation.git

Build clang instrumentation

cd ros-clang-instrumentation
cd llvm-passes
mkdir build
cd build
cmake ..
make

Testing

There are three test cases provided.

Test1

This test is the default publisher/subscriber examples (see http://wiki.ros.org/ROS/Tutorials/WritingPublisherSubscriber(c%2B%2B) ).

Building the example with instrumentation

cd ros-clang-instrumentation/ros_ws
ROS_INSTRUMENTATION_CONFIG_FILE=${PWD}/src/test1/test/config.yaml ./catkin_make_with_instrumentation.sh
source devel/setup.bash

Running the example
```
roslaunch test1 test.launch
```
Use Ctrl+C to exit the example.

Viewing the results

rosrun ros_instrumentation results.py src/test1/test/config.yaml test1

Expected output

Topic: chatter
  Latencies [us]:
    [[ 276.]
     [ 177.]
     [ 357.]
     [ 186.]
     [ 195.]
     [ 195.]]
  Iteration 0:
    talker.cpp:24
    talker.cpp:36
    talker.cpp:43
    talker.cpp:62
    talker.cpp:70
    talker.cpp:98
    talker.cpp:101
    talker.cpp:102
    talker.cpp:102

    Latency: 276.0 us
    chatterCallback(boost::shared_ptr<std_msgs::String_<std::allocator<void> > const> const&)

Where the latencies should be the same order of magnitude (below 1000 us) and the path should match exactly for each iteration.

Test2

This test uses one publisher and one subscriber in the same executable, testing the non-networking code path of the ROS message passing.

Building the example with instrumentation

cd ros-clang-instrumentation/ros_ws
ROS_INSTRUMENTATION_CONFIG_FILE=${PWD}/src/test2/test/config.yaml ./catkin_make_with_instrumentation.sh
source devel/setup.bash

Running the example
```
roslaunch test2 test.launch
```
Use Ctrl+C to exit the example.

Viewing the results

rosrun ros_instrumentation results.py src/test2/test/config.yaml test2

Expected output

Topic: chatter
  Latencies [us]:
    [[ 100199.]
     [ 100089.]
     [ 100097.]
     [ 102300.]
     [ 100040.]
     [ 100173.]
     [ 100016.]
     [ 100248.]]
  Iteration 0:
    talkerAndListener.cpp:14
    talkerAndListener.cpp:26
    talkerAndListener.cpp:33
    talkerAndListener.cpp:52
    talkerAndListener.cpp:53
    talkerAndListener.cpp:61
    talkerAndListener.cpp:88
    talkerAndListener.cpp:91
    talkerAndListener.cpp:92
    talkerAndListener.cpp:92

    Latency: 100199.0 us
    chatterCallback(boost::shared_ptr<std_msgs::String_<std::allocator<void> > const> const&)

Where the latencies should be the same order of magnitude (about 100000 us) and the path should match exactly for each iteration. The latency is high in this case because ros::spinOnce() is only called every 100ms (and the subscriber callbacks are not threaded in roscpp).

Test3

This test uses one publisher, a repeater, and one subscriber, creating a pipeline of 3 ROS nodes. The repeater has two different code paths, one being slow and one being fast, emulated by the following code:

if (counter % 2 == 0) {
  // emulate slow computation
  std::this_thread::sleep_for(std::chrono::milliseconds(50));
} else {
  std::this_thread::sleep_for(std::chrono::milliseconds(1));
}

Building the example with instrumentation

cd ros-clang-instrumentation/ros_ws
ROS_INSTRUMENTATION_CONFIG_FILE=${PWD}/src/test3/test/config.yaml ./catkin_make_with_instrumentation.sh
source devel/setup.bash

Running the example
```
roslaunch test3 test.launch
```
Use Ctrl+C to exit the example.

Viewing the results

rosrun ros_instrumentation results.py src/test3/test/config.yaml test3

Expected output

Topic: chatter2
  Latencies [us]:
    [[   284.  50244.    263.]
     [   177.   1119.    106.]
     [   274.  50165.    188.]
     [   226.   1112.    101.]]
  Iteration 0:
    talker.cpp:7
    talker.cpp:9
    talker.cpp:11
    talker.cpp:12
    talker.cpp:13
    talker.cpp:17
    talker.cpp:40
    talker.cpp:43

    Latency: 284.0 us
    chatter2Callback(boost::shared_ptr<geometry_msgs::Point_<std::allocator<void> > const> const&)

    Latency: 50244.0 us
    repeater.cpp:9
    repeater.cpp:14
    repeater.cpp:20

    Latency: 263.0 us
    chatter3Callback(boost::shared_ptr<geometry_msgs::Vector3_<std::allocator<void> > const> const&)

  Iteration 1:
    talker.cpp:7
    talker.cpp:9
    talker.cpp:11
    talker.cpp:12
    talker.cpp:13
    talker.cpp:17
    talker.cpp:40
    talker.cpp:43

    Latency: 177.0 us
    chatter2Callback(boost::shared_ptr<geometry_msgs::Point_<std::allocator<void> > const> const&)

    Latency: 1119.0 us
    repeater.cpp:9
    repeater.cpp:16
    repeater.cpp:20

    Latency: 106.0 us
    chatter3Callback(boost::shared_ptr<geometry_msgs::Vector3_<std::allocator<void> > const> const&)

Where the latencies should be the same order of magnitude. The middle column for the repeater should alternate in each iteration between about 50000 us and 1000 us. The path for each iteration should change to either include repeater.cpp:14 or repeater.cpp:16.

Limitations

Only ROS C++ nodes are supported (or any other language which can be compiled using LLVM).
The path computation uses strongly connected components (SCC) and the Ball Larus path profiling approach. Loops will be part of a single SCC.
The path profiler ignores exception-based control-flow.