Fuzzing

Each fuzzing target can be built with multiple engines. Zstd provides a fuzz corpus for each target that can be downloaded with the command:

make corpora

It will download each corpus into ./corpora/TARGET.

fuzz.py

fuzz.py is a helper script for building and running fuzzers. Run ./fuzz.py -h for the commands and run ./fuzz.py COMMAND -h for command specific help.

Generating Data

fuzz.py provides a utility to generate seed data for each fuzzer.

make -C ../tests decodecorpus
./fuzz.py gen TARGET

By default it outputs 100 samples, each at most 8KB into corpora/TARGET-seed, but that can be configured with the --number, --max-size-log and --seed flags.

Build

It respects the usual build environment variables CC, CFLAGS, etc. The environment variables can be overridden with the corresponding flags --cc, --cflags, etc. The specific fuzzing engine is selected with LIB_FUZZING_ENGINE or --lib-fuzzing-engine, the default is libregression.a. Alternatively, you can use Clang's built in fuzzing engine with --enable-fuzzer. It has flags that can easily set up sanitizers --enable-{a,ub,m}san, and coverage instrumentation --enable-coverage. It sets sane defaults which can be overridden with flags --debug, --enable-ubsan-pointer-overflow, etc. Run ./fuzz.py build -h for help.

Running Fuzzers

./fuzz.py can run libfuzzer, afl, and regression tests. See the help of the relevant command for options. Flags not parsed by fuzz.py are passed to the fuzzing engine. The command used to run the fuzzer is printed for debugging.

Here's a helpful command to fuzz each target across all cores, stopping only if a bug is found:

for target in $(./fuzz.py list); do
    ./fuzz.py libfuzzer $target -jobs=10 -workers=10 -max_total_time=1000 || break;
done

Alternatively, you can fuzz all targets in parallel, using one core per target:

python3 ./fuzz.py list | xargs -P$(python3 ./fuzz.py list | wc -l) -I__ sh -c "python3 ./fuzz.py libfuzzer __ 2>&1 | tee __.log"

Either way, to double-check that no crashes were found, run ls corpora/*crash. If any crashes were found, you can use the hashes to reproduce them.

LibFuzzer

# Build the fuzz targets
./fuzz.py build all --enable-fuzzer --enable-asan --enable-ubsan --cc clang --cxx clang++
# OR equivalently
CC=clang CXX=clang++ ./fuzz.py build all --enable-fuzzer --enable-asan --enable-ubsan
# Run the fuzzer
./fuzz.py libfuzzer TARGET <libfuzzer args like -jobs=4>

where TARGET could be simple_decompress, stream_round_trip, etc.

MSAN

Fuzzing with libFuzzer and MSAN is as easy as:

CC=clang CXX=clang++ ./fuzz.py build all --enable-fuzzer --enable-msan
./fuzz.py libfuzzer TARGET <libfuzzer args>

fuzz.py respects the environment variables / flags MSAN_EXTRA_CPPFLAGS, MSAN_EXTRA_CFLAGS, MSAN_EXTRA_CXXFLAGS, MSAN_EXTRA_LDFLAGS to easily pass the extra parameters only for MSAN.

AFL

The default LIB_FUZZING_ENGINE is libregression.a, which produces a binary that AFL can use.

# Build the fuzz targets
CC=afl-clang CXX=afl-clang++ ./fuzz.py build all --enable-asan --enable-ubsan
# Run the fuzzer without a memory limit because of ASAN
./fuzz.py afl TARGET -m none

Regression Testing

The regression test supports the all target to run all the fuzzers in one command.

CC=clang CXX=clang++ ./fuzz.py build all --enable-asan --enable-ubsan
./fuzz.py regression all
CC=clang CXX=clang++ ./fuzz.py build all --enable-msan
./fuzz.py regression all

Fuzzing a custom sequence producer plugin

Sequence producer plugin authors can use the zstd fuzzers to stress-test their code. See the documentation in fuzz_third_party_seq_prod.h for details.

3.7 KiB Raw Blame History