Skip to content

Latest commit

 

History

History
104 lines (72 loc) · 4.69 KB

native_extension_fuzzing.md

File metadata and controls

104 lines (72 loc) · 4.69 KB
Raw

Using Instrumentation with Atheris and Native Extensions

In order for native fuzzing to be effective, your native extensions must be built with Clang, using the argument -fsanitize=fuzzer-no-link (and optionally other sanitizers).

Step 1: Compiling your Extension

While this doesn't work for every extension and build system, if you are using setuptools, you can typically compile a sanitized extension like this:

CC="/usr/bin/clang" CFLAGS="-fsanitize=address,fuzzer-no-link" CXX="/usr/bin/clang++" CXXFLAGS="-fsanitize=address,fuzzer-no-link" pip3 install .

Here, address means Address Sanitizer. You can also use undefined for the Undefined Behavior Sanitizer.

Step 2: Use an external libFuzzer

At runtime, you may get an error about missing symbols. These are libFuzzer symbols, and you must ensure they are available in the global namespace. Here are the two options:

Option A: Sanitizer+libFuzzer preloads

If you can use this option, we recommend it; it is significantly easier than option #2. (However, this option is not yet supported on Mac). When Atheris is installed, it attempts to generate custom ASan and UBSan shared libraries that have libFuzzer linked in. You can find these libraries in the directory returned by this command:

python -c "import atheris; print(atheris.path())"

These files will be called:

  • asan_with_fuzzer.so
  • ubsan_with_fuzzer.so
  • ubsan_cxx_with_fuzzer.so

If these files are present, it means Atheris succesfully generated the files at installation time, and you can use this option. Simply LD_PRELOAD the right .so file, and you're good to go. Here's a complete example:

LD_PRELOAD="$(python -c "import atheris; print(atheris.path())")/asan_with_fuzzer.so" python ./my_fuzzer.py

Option 2: Linking libFuzzer into Python

This option doesn't rely on these custom shared libraries; instead, it involves building a modified CPython. We provide a script and patch file that attempts to do this for Python 3.8.6 in the third_party directory.

cd third_party
./build_modified_libfuzzer.sh

This will clone CPython, check out version 3.8.6, apply the patch file, find libFuzzer, and build. It will not install; you can either make install or just use ./python directly from that directory.

If your new Python is missing certain libraries, you may need to install some prerequisites using apt install (or your platform's equivalent). See regular Python build documentation for help.

We provide a patch file for CPython 3.8.6. Other nearby versions can likely be patched in a similar manner.

If you have issues building a modified CPython, or wish to provide patches for other versions, please open an issue or provide a PR.

Once libFuzzer is linked into Python, you can LD_PRELOAD a normal sanitizer, rather than having to LD_PRELOAD a special one generated by Atheris. You can generally find this sanitizer by running clang -print-search-dirs and looking in the descendants of the first "libraries" dir. You also don't need to LD_PRELOAD at all if not using a sanitizer.

The correct libFuzzer

Building Atheris from source requires a recent version of libFuzzer, but for most reasonable versions, it can perform an in-place upgrade. The correct version (upgraded if needed) is written to the site-packages directory adjacent to where Atheris is installed. You can find it in the directory returned by this command:

python3 -c "import atheris; print(atheris.path())"

The build_modified_libfuzzer.sh script uses the libFuzzer found there by default.

Why this is necessary

Certain code coverage symbols exported by libFuzzer are also exported by ASan and UBSan. Normally, this isn't a problem, because ASan/UBSan export them as weak symbols - libFuzzer's symbols take precedence. However, when ASan/UBSan are preloaded and libFuzzer is loaded as part of a shared library (Atheris), the weak symbols are loaded first. This causes code coverage information to be sent to ASan/UBSan, not libFuzzer.

By linking libFuzzer into Python directly, or by linking libFuzzer into ASan/UBSan, the dynamic linker can correctly select the strong symbols from libFuzzer rather than the weak symbols from ASan/UBSan.

Leak detection

Python is known to leak certain data, such as at interpreter initialization time. You should disable leak detection, for example with -detect_leaks=0.

What if I'm not using a Sanitizer?

While we recommend that you use a sanitizer when fuzzing native code, it's not mandatory. If you'd like to use Atheris to fuzz native code without a sanitizer, you should still build your extension with -fsanitize=fuzzer-no-link, and still LD_PRELOAD asan_with_fuzzer.so. Just remove the address entries from the build command.