json/tests/fuzzing.md

Fuzz testing

Each parser of the library (JSON, BJData, BON8, BSON, CBOR, MessagePack, and UBJSON) can be fuzz tested. Currently, libFuzzer and afl++ are supported.

Additionally, parse_json_view_fuzzer (tests/src/fuzzer-parse_json_view.cpp) cross-checks json_document/json_view (the zero-copy, read-only view declared in json_view.hpp) against basic_json on the same JSON text: it asserts that json_document::accept agrees with json::accept, that an accepted input materializes to the same value json::parse produces, and that a rejected input makes both parsers throw with an identical what(). It checks this for a std::string input (borrowed, with a NUL after the last byte) and for an exact-size std::vector<std::uint8_t> (borrowed, with nothing after the last byte), and for the ignore_comments and ignore_trailing_commas options, which are taken from the low bits of the first input byte (the byte stays part of the text). It takes plain JSON text, so it reuses the corpus_json corpus rather than a format of its own.

json_view_image_fuzzer (tests/src/fuzzer-json_view_image.cpp) tests the images of json_document (save() and load()). It uses each input twice: as an image, which load() must either reject with parse_error.116 or read safely (with image_check::full, the document must also serialize to the JSON it reads as), and as a JSON text, whose image must load and serialize to the same text. A corpus of images can be made from JSON files with a small program that calls json_document::parse(text).save(); plain JSON files work as well.

What the fuzzers check

Each fuzzer driver (tests/src/fuzzer-parse_*.cpp) parses its input twice: once with allow_exceptions = false and once with exceptions. Both calls must agree. Where parsing with exceptions fails, the call without exceptions must return a discarded value (or throw the same kind of non-parse error), and it must never throw a parse_error. Where parsing succeeds, both calls must return the same value. The drivers then serialize the value, parse the result back, and check that nothing was lost.

Finally, each driver reads the same input through the other input adapters: a std::istringstream, a non-contiguous iterator range (std::list<std::uint8_t>), and, on POSIX systems, a FILE* from fmemopen. The pointer range takes the fast paths for contiguous input (bulk string scanning, multi-byte loads, lazy position tracking), while the other adapters read byte by byte. Every adapter must give the same value, or the same exception with the same id, message, and byte position. The JSON driver checks this for parse and accept, each with and without ignore_comments. The shared code is check_input_adapters() in tests/src/fuzzer_common.hpp.

The drivers check all of this with assert, so they refuse to build with NDEBUG.

Corpus creation

For most effective fuzzing, a corpus should be provided. A corpus is a directory with some simple input files that cover several features of the parser and is hence a good starting point for mutations.

TEST_DATA_VERSION=3.2.0
wget https://github.com/nlohmann/json_test_data/archive/refs/tags/v$TEST_DATA_VERSION.zip
unzip v$TEST_DATA_VERSION.zip
rm v$TEST_DATA_VERSION.zip
for FORMAT in json bjdata bon8 bson cbor msgpack ubjson
do
  rm -fr corpus_$FORMAT
  mkdir corpus_$FORMAT
  find json_test_data-$TEST_DATA_VERSION -size -5k -name "*.$FORMAT" -exec cp "{}" "corpus_$FORMAT" \;
done
rm -fr json_test_data-$TEST_DATA_VERSION

The generated corpus can be used with both libFuzzer and afl++. The remainder of this documentation assumes the corpus directories have been created in the tests directory.

libFuzzer

To use libFuzzer, you need to pass -fsanitize=fuzzer as FUZZER_ENGINE. In the tests directory, call

make fuzzers FUZZER_ENGINE="-fsanitize=fuzzer"

This creates a fuzz tester binary for each parser that supports these command line options.

In case your default compiler is not a Clang compiler that includes libFuzzer (Clang 6.0 or later), you need to set the CXX variable accordingly. Note the compiler provided by Xcode (AppleClang) does not contain libFuzzer. Please install Clang via Homebrew calling brew install llvm and add CXX=$(brew --prefix llvm)/bin/clang to the make call:

make fuzzers FUZZER_ENGINE="-fsanitize=fuzzer" CXX=$(brew --prefix llvm)/bin/clang

Then pass the corpus directory as command-line argument (assuming it is located in tests):

./parse_cbor_fuzzer corpus_cbor

The fuzzer should be able to run indefinitely without crashing. In case of a crash, the tested input is dumped into a file starting with crash-.

To also detect memory leaks, build with AddressSanitizer (FUZZER_ENGINE="-fsanitize=fuzzer,address"): libFuzzer then runs LeakSanitizer by default (-detect_leaks=1). LeakSanitizer is not available with Apple Clang on macOS.

afl++

To use afl++, you need to pass -fsanitize=fuzzer as FUZZER_ENGINE. It will be replaced by a libAFLDriver.a to re-use the same code written for libFuzzer with afl++. Furthermore, set afl-clang-fast++ as compiler.

CXX=afl-clang-fast++ make fuzzers FUZZER_ENGINE="-fsanitize=fuzzer" 

Then the fuzzer is called like this in the tests directory:

afl-fuzz -i corpus_cbor -o out  -- ./parse_cbor_fuzzer 

The fuzzer should be able to run indefinitely without crashing. In case of a crash, the tested input is written to the directory out.

OSS-Fuzz

The library is further fuzz-tested 24/7 by Google's OSS-Fuzz project. It uses the same fuzzers target as above and also relies on the FUZZER_ENGINE variable. See the used build script for more information. Its default address sanitizer includes LeakSanitizer, so OSS-Fuzz and the CIFuzz workflow (.github/workflows/cifuzz.yml) report memory leaks, too.

In case the build at OSS-Fuzz fails, an issue will be created automatically.

Handling OSS-Fuzz reports

OSS-Fuzz files the crashes it finds in its own issue tracker, not on GitHub. So that each report can be traced to the change that fixed it, and each fix to the report it answers, fixes follow these conventions:

• Reference the OSS-Fuzz issue in the pull request, next to any GitHub issue it closes, as OSS-Fuzz: <id> (for example, OSS-Fuzz: 563659413), and in the commit message. The ID alone does not disclose the crash. If the report was triaged into a GitHub issue, link the OSS-Fuzz issue there too.
• Turn the reproducer into a unit test. Download the testcase from the OSS-Fuzz report, reduce it if possible, and add it as a regression test to the unit test of the affected format (e.g., tests/src/unit-bjdata.cpp), with a comment naming the OSS-Fuzz issue. This way the input is checked by every CI run rather than only by OSS-Fuzz, and it stays covered even if OSS-Fuzz later closes the report as not reproducible.
• Keep the fuzzer drivers and the unit tests in sync. The round-trip checks of the BJData, BON8, BSON, CBOR, MessagePack and UBJSON drivers are also run on a fixed corpus in the unit tests (see tests/src/round_trip_corpus.hpp and the "round-trip invariants" test cases), so a regression shows up in CI first. When a driver's checks change, change the unit tests with them.
• Record in the report whether the bug shipped. OSS-Fuzz asks whether a crash was a short-lived regression or affects a released version; answer it when the fix is merged, as it decides whether the fix needs a release note or a security advisory (see the security policy).

After the fix is merged, OSS-Fuzz re-runs the reproducer on its next build and marks the report as verified and closed. If it does not, the fix is incomplete.