1 files changed, 72 insertions, 0 deletions
diff --git a/docs/fuzzing.md b/docs/fuzzing.md
new file mode 100644
index 0000000..2bf4ccc
--- /dev/null
+++ b/docs/fuzzing.md
@@ -0,0 +1,72 @@
+# Fuzzing
+
+libgit2 is currently using [libFuzzer](https://libfuzzer.info) to perform
+automated fuzz testing. libFuzzer only works with clang.
+
+## Prerequisites for building fuzz targets:
+
+1. All the prerequisites for [building libgit2](https://github.com/libgit2/libgit2).
+2. A recent version of clang. 6.0 is preferred. [pre-build Debian/Ubuntu
+   packages](https://github.com/libgit2/libgit2)
+
+## Build
+
+1. Create a build directory beneath the libgit2 source directory, and change
+   into it: `mkdir build && cd build`
+2. Choose one sanitizers to add. The currently supported sanitizers are
+   [`address`](https://clang.llvm.org/docs/AddressSanitizer.html),
+   [`undefined`](https://clang.llvm.org/docs/UndefinedBehaviorSanitizer.html),
+   and [`leak`/`address,leak`](https://clang.llvm.org/docs/LeakSanitizer.html).
+3. Create the cmake build environment and configure the build with the
+   sanitizer chosen: `CC=/usr/bin/clang-6.0 CFLAGS="-fsanitize=address" cmake
+   -DBUILD_TESTS=OFF -DBUILD_FUZZERS=ON -DCMAKE_BUILD_TYPE=RelWithDebInfo ..`.
+   Note that building the fuzzer targets is incompatible with the
+   tests and examples.
+4. Build libgit2: `cmake --build .`
+5. Exit the cmake build environment: `cd ..`
+
+## Run the fuzz targets
+
+1. `ASAN_SYMBOLIZER_PATH=/usr/bin/llvm-symbolize
+   LSAN_OPTIONS=allocator_may_return_null=1
+   ASAN_OPTIONS=allocator_may_return_null=1 ./build/fuzzers/packfile_fuzzer
+   fuzzers/corpora/packfile/`
+
+The `LSAN_OPTIONS` and `ASAN_OPTIONS` are there to allow `malloc(3)` to return
+`NULL`, which is expected if a huge chunk of memory is allocated. The
+`LLVM_PROFILE_FILE` environment string can also be added to override the path
+where libFuzzer will write the coverage report.
+
+## Get coverage
+
+In order to get coverage information, you need to add the "-fcoverage-mapping"
+and "-fprofile-instr-generate CFLAGS, and then run the fuzz target with
+`-runs=0`. That will produce a file called `default.profraw` (this behavior can
+be overridden by setting the `LLVM_PROFILE_FILE="yourfile.profraw"` environment
+variable).
+
+1. `llvm-profdata-6.0 merge -sparse default.profraw -o
+   fuzz_packfile_raw.profdata` transforms the data from a sparse representation
+   into a format that can be used by the other tools.
+2. `llvm-cov-6.0 report ./build/fuzz/fuzz_packfile_raw
+   -instr-profile=fuzz_packfile_raw.profdata` shows a high-level per-file
+   coverage report.
+3. `llvm-cov-6.0 show ./build/fuzz/fuzz_packfile_raw
+   -instr-profile=fuzz_packfile_raw.profdata [source file]` shows a line-by-line
+   coverage analysis of all the codebase (or a single source file).
+
+## Standalone mode
+
+In order to ensure that there are no regresions, each fuzzer target can be run
+in a standalone mode. This can be done by passing `-DUSE_STANDALONE_FUZZERS=ON`.
+This makes it compatible with gcc. This does not use the fuzzing engine, but
+just invokes every file in the chosen corpus.
+
+In order to get full coverage, though, you might want to also enable one of the
+sanitizers. You might need a recent version of clang to get full support.
+
+## References
+
+* [libFuzzer](https://llvm.org/docs/LibFuzzer.html) documentation.
+* [Source-based Code
+  Coverage](https://clang.llvm.org/docs/SourceBasedCodeCoverage.html).