PolyTracker

PolyTracker is a tool originally created for the Automated Lexical Annotation and Navigation of Parsers, a backronym devised solely for the purpose of referring to it as The ALAN Parsers Project. However, it has evolved into a general purpose tool for efficiently performing data-flow and control-flow analysis of programs. PolyTracker is an LLVM pass that instruments programs to track which bytes of an input file are operated on by which functions. It outputs a database containing the data-flow information, as well as a runtime trace. PolyTracker also provides a Python library for interacting with and analyzing its output, as well as an interactive Python REPL.

PolyTracker can be used in conjunction with PolyFile to automatically determine the semantic purpose of the functions in a parser. It also has an experimental feature capable of generating a context free grammar representing the language accepted by a parser.

Unlike dynamic instrumentation alternatives like Taintgrind, PolyTracker imposes negligible performance overhead for almost all inputs, and is capable of tracking every byte of input at once. PolyTracker started as a fork of the LLVM DataFlowSanitizer and takes much inspiration from the Angora Fuzzer. However, unlike the Angora system, PolyTracker is able to track the entire provenance of a taint. In February of 2021, the LLVM DataFlowSanitizer added a new feature for tracking taint provenance called origin tracking. However, it is only able to track at most 16 taints at once, while PolyTracker can track up to 2³¹-1.

This README serves as the general usage guide for installing PolyTracker and compiling/instrumenting binaries. For programmatically interacting with or extending PolyTracker through its Python API, as well as for interacting with runtime traces produced from instrumented code, consult the Python documentation.

Quickstart

PolyTracker is controlled via a Python script called polytracker. You can install it by running

pip3 install polytracker

PolyTracker requires a very particular system environment to run, so almost all users are likely to run it in a containerized environment. Luckily, polytracker makes this easy. All you need to do is have docker installed, then run:

polytracker docker pull

and

polytracker docker run

The latter command will mount the current working directory into the PolyTracker Docker container, and allow you to build and run instrumented programs.

The polytracker control script—which you can run from either your host system or from inside the Docker container—has a variety of commands, both for instrumenting programs as well as analyzing the resulting artifacts. For example, you can explore the dataflows in the execution, reconstruct the instrumented program's control flow graph, and even extract a context free grammar matching the inputs accepted by the program. You can explore these commands by running

polytracker --help

The polytracker script is also a REPL, if run with no command line arguments:

$ polytracker
PolyTracker (4.0.0)
https://github.com/trailofbits/polytracker
Type "help" or "commands"
>>> commands

Instrumenting a simple C/C++ program

PolyTracker also comes with a build command. This command allows the user to run any build command in a Blight instrumented environment. This will produce a blight_journal.jsonl file that records all commands run during the build. If you have a C/C++ target, you can instrument it by invoking polytracker build and passing your build command:

polytracker build gcc -g -o my_binary my_source.c

To instrument a build target, use the instrument-targets command. By default the command will use the a blight_journal.jsonl in your current working directory to build an instrumented version of your build target. The instrumented build target will be built using the same flags as the original build target.

polytracker instrument-targets my_binary

build also supports more complex programs that use a build system like autotiools or CMake:

polytracker build cmake .. -DCMAKE_BUILD_TYPE=Release
polytracker build ninja
# or
polytracker build ./configure
polytracker build make

Then run instrument-targets on any targets of the build:

$ polytracker instrument-targets a.bin b.so

Then a.instrumented.bin and b.instrumented.so will be the instrumented versions. See the Dockerfiles in the examples directory for examples of how real-world programs can be instrumented.

Running and Analyzing an Instrumented Program

The instrumented software will write its output to the path specified in POLYDB, or polytracker.tdag if omitted. This is a binary file that can be operated on by running:

from polytracker import PolyTrackerTrace, taint_dag

trace = PolyTrackerTrace.load("polytracker.tdag")
tdfile = trace.tdfile

first_node = list(tdfile.nodes)[0]
print(f"First node affects control flow: {first_node.affects_control_flow}")

# Operate on all Range nodes
for index, node in enumerate(tdfile.nodes):
  if isinstance(node, taint_dag.TDRangeNode):
    print(f"Node {index}: first {node.first}, last {node.last}")

# Access taint forest
tdforest = trace.taint_forest
n1 = tdforest.get_node(1)
print(
  f"Forest node {n1.label}. Parent labels: {n1.parent_labels}, "
  f"source: {n1.source.path if n1.source is not None else None}, "
  f"affects control flow: {n1.affected_control_flow}"
)

You can also run an instrumented binary directly from the REPL:

$ polytracker
PolyTracker (4.0.0)
https://github.com/trailofbits/polytracker
Type "help" or "commands"
>>> trace = run_trace("path_to_binary", "path_to_input_file")

This will automatically run the instrumented binary in a Docker container, if necessary.

> ⚠️ If running PolyTracker inside Docker or a VM: PolyTracker can be > very slow if running in a virtualized environment and either the input file > or, especially, the output database are located in a directory mapped or > mounted from the host OS. This is particularly true when running PolyTracker > in Docker from a macOS host. The solution is to write the database to a path > inside of the container/VM and then copy it out to the host system at the very > end.

The Python API documentation is available here.

Runtime Parameters and Instrumentation Tuning

At runtime, PolyTracker instrumentation looks for a number of configuration parameters specified through environment variables. This allows one to modify instrumentation parameters without needing to recompile the binary.

Environment Variables

PolyTracker accepts configuration parameters in the form of environment variables to avoid recompiling target programs. The current environment variables PolyTracker supports is:

POLYDB: A path to which to save the output database (default is polytracker.tdag)

WLLVM_ARTIFACT_STORE: Provides a path to an existing directory to store artifact/manifest for all build targets

POLYTRACKER_TAINT_ARGV: Set to '1' to use argv as a taint source.

Polytracker will set its configuration parameters in the following order:

1. If a parameter is specified via an environment variable, use that value
2. Else if a default value for the parameter exists, use the default
3. Else throw an error

ABI Lists

DFSan uses ABI lists to determine what functions it should automatically instrument, what functions it should ignore, and what custom function wrappers exist. See the dfsan documentation for more information.

Creating custom ignore lists from pre-built libraries

Attempting to build large software projects can be time consuming, especially older/unsupported ones. It's even more time consuming to try and modify the build system such that it supports changes, like dfsan's/our instrumentation.

There is a script located in polytracker/scripts that you can run on any ELF library and it will output a list of functions to ignore. We use this when we do not want to track information going through a specific library like libpng, or other sub components of a program. The Dockerfile-listgen.demo exists to build common open source libraries so we can create these lists.

This script is a slightly tweaked version of what DataFlowSanitizer has, which focuses on ignoring system libraries. The original script can be found in dfsan_rt.

Building the Examples

Check out this Git repository. From the root, either build the base PolyTracker Docker image:

pip3 install -e ".[dev]" && polytracker docker rebuild

or pull the latest prebuilt version from DockerHub:

docker pull trailofbits/polytracker:latest

For a demo of PolyTracker running on the MuPDF parser run this command:

docker build -t trailofbits/polytracker-demo-mupdf -f examples/pdf/Dockerfile-mupdf.demo .

mutool_track will be build in /polytracker/the_klondike/mupdf/build/debug. Running mutool_track will output polytracker.tdag which contains the information provided by the taint analysis.

For a demo of PolyTracker running on Poppler utils version 0.84.0 run this command:

docker build -t trailofbits/polytracker-demo-poppler -f examples/pdf/Dockerfile-poppler.demo .

All the poppler utils will be located in /polytracker/the_klondike/poppler-0.84.0/build/utils.

$ cd /polytracker/the_klondike/poppler-0.84.0/build/utils
$ ./pdfinfo_track some_pdf.pdf

Building PolyTracker from Source

The compilation process for both PolyTracker LLVM and PolyTracker is rather fickle, since it involves juggling both instrumented and non-instrumented versions of standard library bitcode. We highly recommend using our pre-built and tested Docker container if at all possible. Installing the PolyTracker Python package on your host system will allow you to seamlessly interact with the prebuilt Docker container. Otherwise, to install PolyTracker natively, we recommend first replicating the install process from the polytracker-llvm Dockerfile, followed by replicating the install process from the PolyTracker Dockerfile.

Build Dependencies

• PolyTracker LLVM. PolyTracker is built atop its own fork of LLVM, polytracker-llvm. This fork modifies the DataFlow Sanitizer to use increased label sizes (to allow for tracking orders of magnitude more taints), as well as alternative data structures to store them. We have investigated up-streaming our changes into LLVM proper, but there has been little interest.
• CMake
• Ninja (ninja-build on Ubuntu)

Runtime Dependencies

The following tools are required to test and run PolyTracker:

• Python 3.7+ and pip (apt-get -y install python3.7 python3-pip). These are used for both seamlessly interacting with the Docker container (if necessary), as well as post-processing and analyzing the artifacts produced from runtime traces.
• gllvm (go get github.com/SRI-CSL/gllvm/cmd/...) is used to create whole program bitcode archives and to extract bitcode from targets.

Building on Apple silicon:

Prebuilt Docker images for polytracker-llvm are only available for amd64. Users with arm64 systems will have to build the image locally and then change polytracker's Dockerfile to point to it:

$ mkdir repos && cd repos
$ git clone https://github.com/trailofbits/polytracker
$ git clone https://github.com/trailofbits/polytracker-llvm
$ cd polytracker-llvm
$ DOCKER_BUILDKIT=1 docker build -t trailofbits/polytracker-llvm .
$ cd ../polytracker
$ ## Replace the first line of the Dockerfile with "FROM trailofbits/polytracker-llvm:latest" (no quotes)
$ docker build -t trailofbits/polytracker .

Current Status and Known Issues

PolyTracker currently only runs on Linux, because that is the only system supported by the DataFlow Santizer. This limitation is just due to a lack of support for semantics for other OSes system calls, which could be added in the future. However, this means that running PolyTracker on a non-Linux system will require Docker to be installed.

Taints will not propagate through dynamically loaded libraries unless those libraries were compiled from source using PolyTracker, or there is specific support for the library calls implemented in PolyTracker. There is currently support for propagating taint through the majority of uninstrumented C standard library calls. To be clear, programs that use uninstrumented functions will still run normally, however, operations performed by unsupported library calls will not propagate taint. We are currently working on adding robust support for C++ programs, but currently the best results will be from C programs.

If there are issues with Docker, try performing a system prune and build with --no-cache for both PolyTracker and whatever demo you are trying to run.

The worst case performance of PolyTracker is exercised when a single byte in memory is simultaneously tainted by a large number of input bytes from the source file. This is most common when instrumenting compression and cryptographic algorithms that have large block sizes. There are a number of mitigations for this behavior currently being researched and developed.

License and Acknowledgements

This research was developed by Trail of Bits with funding from the Defense Advanced Research Projects Agency (DARPA) under the SafeDocs program as a subcontractor to Galois. It is licensed under the Apache 2.0 license. © 2019, Trail of Bits.

Maintainers

Evan Sultanik
Henrik Brodin
Marek Surovič
Facundo Tuesca

firstname.lastname@trailofbits.com