fuzz-target and fuzz-package to the library's BUILD.gnfuzztest to the library's rules.mk.fx set core.x64 --fuzz-with asan, fx build, and boot.fx fuzz list to display fuzzers.fx fuzz start <fuzzer> to start a fuzzer.fx fuzz check <fuzzer> to see if the fuzzer found crashes.fx fuzz repro <fuzzer> [crash] to replay a crash.A: Fuzzing or fuzz testing is style of testing that stochastically generates inputs to targeted interfaces in order to automatically find defects and/or vulnerabilities. In this document, a distinction will be made between two components of a fuzzer: the fuzzing engine, which produces context-free inputs, and the fuzzing target, which submits those inputs to a specific interface.
Among the various styles of fuzzing, coverage-based fuzzing has been shown to yield a particularly high number of bugs for the effort involved. In coverage-based fuzzing, the code under test is instrumented for coverage. The fuzzing engine can observe when inputs increase the overall code coverage and use those inputs as the basis for generating further inputs. This group of “seed” inputs is collectively referred to as a corpus.
A: LibFuzzer is an in-process fuzzing engine integrated within LLVM as a compiler runtime. Compiler runtimes are libraries that are invoked by hooks that compiler adds to the code it builds. Other examples include sanitizers such as ASan, which detects certain overflows and memory corruptions. LibFuzzer uses these sanitizers both for coverage data provided by sanitizer-common, as well as to detect when inputs trigger a defect.
A: LibFuzzer can be used to make a coverage-based fuzzer by combining it with a sanitized library and the implementation of the fuzz target function:
extern "C" int LLVMFuzzerTestOneInput(const uint8_t *Data, size_t Size) { // Use the data to do something interesting with your API return 0; }
Optionally, you can also add an initial corpus. Without it, libFuzzer will start from an empty fuzzer and will (eventually) learn how to make appropriate inputs on its own.
LibFuzzer then be able to generate, submit, and monitor inputs to the library code: 
Developer-provided components are in green.
A: Coverage based fuzzing works best when fuzzing targets resemble unit tests. If your code is already organized to make it easy to unit test, you can add targets for each of the interfaces being tested., e.g. something like:
// std::string json = ...; Metadata metadata; EXPECT_TRUE(metadata.Parse(json));
becomes:
extern "C" int LLVMFuzzerTestOneInput(const uint8_t *Data, size_t Size) { std::string json(static_cast<const char *>(Data), Size); metadata.Parse(json); return 0; }
With a corpus of JSON inputs, Data may be close to what the Metadata object expects to parse. If not, the fuzzer will eventually discover what inputs are meaningful to it through random mutations, trial and error, and code coverage data.
A: It is easy to map portions of the provided Data to “plain old data” (POD) types. The data can also be sliced into variable length arrays. More complex objects can almost always be (eventually) built out of POD types and variable arrays. If Size isn‘t long enough for your needs, you can simply return 0. The fuzzer will quickly learn that inputs below that length aren’t interesting and will stop generating them.
uint32_t flags; char name[MAX_NAME_LEN]; if (Size < sizeof(flags)) { return 0; } memcpy(&flags, Data, sizeof(flags)); Data += sizeof(flags); Size -= sizeof(flags); size_t name_len; if (Size < sizeof(name_len)) { return 0; } memcpy(&name_len, Data, sizeof(name_len)); Data += sizeof(name_len); Size -= sizeof(name_len); name_len %= sizeof(name_len) - 1; if (Size < name_len) { return 0; } memcpy(name, Data, name_len); Data += name_len; Size -= name_len; name[name_len] = '\0'; Parser parser(name, flags); parser.Parse(Data, Size);
NOTE: A small library to make this easier is under development.
In some cases, you may have expensive set-up operations that you would like to do once. The libfuzzer documentation has tips on how to do startup initialization. Be aware though that such state will be carried over from iteration to iteration. This can be useful as it may expose new bugs that depend on the library's persisted state, but it may also make bugs harder to reproduce when they depend on a sequence of inputs rather than a single one.
A: In general, an in-process coverage-based fuzzer, iterations should be short and focused. The more focused a fuzz target is, the faster libfuzzer will be able to find “interesting” inputs that increase code coverage.
At the same time, becoming too focused can lead to a proliferation of fuzz targets. Consider the example of a routine that parses incoming requests. The parser may recognize dozens of different request types, so developing a separate fuzz target for each may be cumbersome. An alternative in this case may be to develop a single fuzzer, and include examples of the different requests in the initial corpus. In this way the single fuzz target can still bypass a large amount of shallow fuzzing by being guided towards the interesting inputs.
NOTE: Currently, libFuzzer can be used in Fuchsia to fuzz C/C++ code. Additional language support is planned.
A: There's many other fuzzing engines out there:
A: First, create your fuzz target. It‘s recommended that the fuzzer’s target is clear from file name. If the library code already has a directory for unit tests, you should use a similar directory for your fuzzing targets. If not, make sure the file's name clearly reflects it is a fuzz test. In general, use naming and location to make the fuzzing target easy to find and its purpose clear.
Example: A fuzz target for //garnet/lib/cmx might be located at //garnet/lib/cmx/cmx_fuzztest.cc, to match //garnet/lib/cmx/cmx_unittest.cc.
Libfuzzer already provides tips on writing the fuzz target itself.
Next, add the build instructions to the library's BUILD.gn file. Adding an import to //build/fuzzing/fuzzer.gni will provide two templates:
The fuzz-target template is used to build the fuzzer executable. Given a fuzzing target as a source and a the library under test as a dependency, it will provided the correct compiler flags to link against the fuzzing engine:
fuzz_target("cowsay_simple_fuzzer") { sources = [ "cowsay_fuzztest.cpp" ] deps = [ ":cowsay_sources" ] }
It also enables the FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION build macro. If the software under test needs fuzzing-specific modifications, they can be wrapped in a preprocessor conditional on this macro, e.g.:
#ifdef FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION srand(++global_counter); rand_int = rand(); #else zx_cprng_draw(&rand_int, size_of(rand_int)); #endif
This can be useful to allow either more deterministic fuzzing and/or deeper coverage.
The fuzz-target template also allows you include additional inputs to control the fuzzer:
fuzz_target("cowsay_simple_fuzzer") { sources = [ "cowsay_fuzztest.cpp" ] deps = [ ":cowsay_sources" ] corpora = [ "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" ] dictionary = "test_data/various_moos.dict" options = "test_data/fuzzer.opts" }
The merge, jobs, dict, and artifact_prefix are set automatically when using the fuzz tool described below, and do not need to be specified unless they differ from the defaults.
The fuzz-package template bundles fuzz-targets into Fuchsia packages in the same way that normal packages bundle binaries. In addition, the fuzz-package template provides a way to specify which sanitizers can be used with the given fuzzing targets:
fuzz_package("cowsay_fuzzers") { targets = [ ":cowsay_simple_fuzzer" ] sanitizers = [ "asan", "ubsan" ] }
Once this is defined, a package manifest needs to reference this package the usual way. If the package already has a package manifest under packages/tests for the appropriate layer, it may be added there:
{ "imports": [ "garnet/packages/prod/boringssl", "garnet/packages/prod/run" ], "packages": [ "//third_party/boringssl:boringssl_tests", "//third_party/boringssl:boringssl_fuzzers" ] }
Note that the fuzzing framework needs run to be able to run fuzz targets.
IMPORTANT: The build system will build the fuzz targets only if it is explicitly told to instrument them for fuzzing with an appropriate sanitizer. The easiest way to achieve this is using the --fuzz-with <sanitizer> flag with fx set, e.g:
$ fx set core.x64 --fuzz-with asan --with //bundles:tests $ fx build
This will build all the fuzz targets in fuzz packages that support ASan and are referenced by the Garnet's test package manifests. If you prefer, you can add the select_variant statement directly to your args.gn file, e.g.:
select_variant = [ { variant = "asan-fuzzer" target_type = [ "fuzzed_executable" ] }, ]
A: It depends on what you're trying to fuzz:
//garnet/tests/zircon/.Otherwise, the library interfaces are not exposed by the SDK, and the fuzz target needs to reside in Zircon as well. Write the target and save it with a file name that contains both the target and the word “fuzz” and place it in the appropriate system/utest directory, e.g. //zircon/system/utest/hid-parser/hid-parser-fuzztest.cpp.
Add a fuzz test declaration in the corresponding rules.mk; e.g.:
MODULE := $(LOCAL_DIR).fuzztest MODULE_TYPE := fuzztest MODULE_NAME := hid_parser_fuzztest # Matches fuzz_target in BUILD.gn MODULE_SRCS = $(LOCAL_DIR)/hid-parser-fuzztest.cpp MODULE_LIBS = system/ulib/c system/ulib/fdio
NOTE: There are open issues around how Garnet drivers are built and linked into Zircon. The following is a temporary workaround until those issues are resolved.
The resulting fuzz tests can be run directly (look under /boot/test/fuzz), but to use the full fuzzing tools, it is easier to run the Zircon fuzzers as part of a Fuchsia build. To accomplish this, first add the Zircon fuzzer to the special zircon_fuzzers package defined in //garnet/tests/zircon/BUILD.gn. You can use this to add seed corpora, dictionaries, and/or an options file as normal in a GN fuzz target, e.g:
fuzz_target("hid_parser_fuzztest") { # Matches MODULE_NAME in rules.mk corpora = [ "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" ] dictionary = "//zircon/system/utest/hid-parser/hid-parser-fuzztest.dict" options = "//zircon/system/utest/hid-parser/hid-parser-fuzztest.options" } fuzz_package("zircon_fuzzers") { omit_binaries = true sanitizers = [ "asan" ] targets = [ "hid_parser_fuzztest", ... ] }
Build this package as you would any other fuzz package, and finally use the fx fuzz zbi command to include the correct dependencies for you Zircon fuzzer in a Fuchsia build:
$ fx set core.x64 --fuzz-with asan --with //garnet/packages/tests:zircon $ fx build $ fx fuzz zbi
Boot normally. fx fuzz list should detect the Zircon fuzzers.
NOTE: This workflow will build the Zircon fuzzers with ASan. Other sanitizers, like UBSan, are not yet supported in Zircon.
NOTE: This workflow will build the Zircon fuzzers with SanCov instrumentation only. Zircon has an open issue which prevents full coverage instrumentation from being enabled. The fuzzers will work, albeit with reduced corpus precision.
Finally, if needed you can run the fuzz tool (without the fx) directly from the Zircon command line.
A: The fuzzer binary can be started directly, using the normal libFuzzer options, if you prefer. However, it is easier to use the fuzz tool, which understands where to look for fuzzing related files and knows various common options. Try one or more of the following:
$ fuzz help # to see the command usage $ fuzz list # to see available fuzzing targets. $ fuzz start [package]/[target] # to start a fuzzer.
package and target match those reported by fuzz list, and may be abbreviated. For commands that accept a single fuzzer, e.g. start, the abbreviated name must uniquely identify exactly one fuzzer.
A small devshell wrapper also exists as fx fuzz. This tool will use SSH to run the fuzz tool, and background long running tasks, e.g.:
fx fuzz list returns the list of fuzzers immediatelyfx fuzz start [package]/[target] starts a fuzzer in the background and returns.If you do prefer to run the fuzzer directly, note that the fuzz tool echoes the command it is invoking on startup. Look for the line starting with +.
A: Explicit inputs can be passed to libFuzzer to execute without mutation.
If using the fuzz tool, the “artifact prefix” will be /data/fuzzing/<package>/<target>. All inputs that the fuzzer finds “interesting” will be saved as <artifact_prefix>/<type>-<sha1>, e.g. if “foo” caused the Zircon fuzztest bar-fuzztest to consume an excessive amount of memory (as determined by rss_limit_mb), the file /data/fuzzing/zircon_fuzzers/bar-fuzztest/oom-8843d7f92416211de9ebb963ff4ce28125932878 will contain “foo”.
The easiest way to get a list of crashes found is fuzz check [package]/[target].
The easiest way to reproduce a crash is to run fuzz repro [package]/[target] [crash]. As with package and target, crash may be abbreviated, e.g. fuzz bar oom will run the zircon-fuzzers/bar-fuzztest with every input that caused excessive memory usage, but none that caused a crash, leaked memory, timed out, etc.
As with fuzz start, the fuzzer will echo the command it is invoking, prefixed by +. This can be useful if you need to manually reproduce the bug with modified parameters.
A: File them, then fix them!
NOTE: The bug tracker is currently only open to Googlers.
When filing bugs, please use both the custom found-by-fuzzing label, as well as the custom Sec-TriageMe label. This will help the security team see where fuzzers are being used and stay aware of any critical issues they are finding.
As with other potential security issues, bugs should be filed in the project of the code under test (and not in the security project). Conversely, if you encounter problems or shortcomings in the fuzzing framework itself, please do open bugs or feature requests in the security project with the label fuzzing.
As with all potential security issues, don‘t wait for triage to begin fixing the bug! Once fixed, don’t forget to link to the bug in the commit message. This may also be a good time to consider minimizing and uploading your corpus at the same time (see the next section).
A: When you first begin fuzzing a new target, the fuzzer may crash very quickly. Typically, fuzzing has a large initial spike of defects found followed by a long tail. Fixing these initial, shallow defects will allow your fuzzer to reach deeper and deeper into the code. Eventually your fuzzer will run for several hours (e.g. overnight) without crashing. At this point, you will want to save the corpus:
First, minimize the corpus using fx fuzz merge [package]/[target]. This will instruct the fuzzer to use the current corpus instead of looking for new inputs. It will run repeatedly to find the most compact set of inputs with the same coverage as the current corpus. As this may take a long time, you may want to provide a max_total_time option (perhaps again running overnight).
Once complete, you can upload the corpus using fx fuzz store [package]/[target]. This will store the corpus in CIPD. Each fuzz target will have its own corpus CIPD package, organized by fuzz package. Take note of the corpus digest that is printed when the corpus is stored; it is used as the package version.
Add the digest from the previous step to the fuzz-target in the appropriate BUILD.gn file, and submit the resulting CL. This creates a clear association between the state of the corpus and the state of the source code at various points in time.
When deploying your fuzzer to a different instance of Fuchsia, you can install the latest version of the target's corpus in CIPD using fx fuzz fetch [package]/[target]. To use a specific version of the corpus, you may add the version digest, e.g. fx fuzz fetch [package]/[target] digest.
A: Yes! The corpora list in the fuzz-target template can contain either digests or source paths. In general, git isn't the best way to store fuzzing corpora, which typically contain many small binary files that may change dramatically over time. Still, if you have a fuzzing corpus already, perhaps as part of a third party project, you can provide the source path to it, e.g. corpora = [ "//third_party/boringssl/src/fuzz/bn_div_corpus" ].
Running fx fuzz fetch [package]/[target] will pull the corpus from the source location and install it. Subsequently running fx fuzz store [package]/[target] will store this corpus in CIPD. If you need to submit this corpus back upstream, you can save a local copy by explicitly setting the local “staging” directory, e.g. fx fuzz -s <source-path> store [package]/[target].
A: Yes, although the extra tooling of fx fuzz is not currently supported. This means you can build host fuzzers with the GN templates, but you'll need to manually run them, reproduce the bugs they find, and manage their corpus data.
If your fuzzers don't have Fuchsia dependencies, you can build host versions simply by setting fuzz_host=true in the fuzz_package:
fuzz_package("overnet_fuzzers") { targets = [ "packet_protocol:packet_protocol_fuzzer" ] sanitizers = [ "asan" ] fuzz_host = true }
You will also need to edit the package manifest and add an explicit label for the host fuzzers. The fuzz_package template will automatically name the target host_${target_name}:
"labels": [ "//src/connectivity/overnet/lib:host_overnet_fuzzers" ]
Upon building, the host fuzzers can be found in in the host variant output directory, e.g. //out/x64/host_x64-asan-fuzzer.
A: Once crashes begin to become infrequent, it may be because almost all the bugs have been fixed, but it may also be because the fuzzer isn't reaching new code that still has bugs. Code coverage information is needed to determine the quality of the fuzzer. Use source-based code coverage to see what your current corpus reaches.
If coverage in a certain area is low, there are a few options:
FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION build macro described above.The “run, merge, measure, improve” steps can be repeated for as many iterations as you feel are needed to create a quality fuzzer. Once ready, you'll need to upload your corpus and update the GN fuzz target in the appropriate project. At this point, others will be able use your fuzzer, This includes ClusterFuzz which will automatically find new fuzzers and continuously fuzz them, updating their corpora, filing bugs for crashes, and closing them when fixed.
A: As you can see from the various notes in this document, there's still plenty more to do!
We will continue to work on these features and others, and update this document accordingly as they become available.