libFuzzer – a library for coverage-guided fuzz testing.

Introduction

LibFuzzer is in-process, coverage-guided, evolutionary fuzzing engine.

LibFuzzer is linked with the library under test, and feeds fuzzed inputs to the library via a specific fuzzing entrypoint (aka “target function”); the fuzzer then tracks which areas of the code are reached, and generates mutations on the corpus of input data in order to maximize the code coverage. The code coverage information for libFuzzer is provided by LLVM’s SanitizerCoverage instrumentation.

Contact: libfuzzer(#)googlegroups.com

Versions

LibFuzzer is under active development so you will need the current (or at least a very recent) version of the Clang compiler.

(If building Clang from trunk is too time-consuming or difficult, then the Clang binaries that the Chromium developers build are likely to be fairly recent:

mkdir TMP_CLANG
cd TMP_CLANG
git clone https://chromium.googlesource.com/chromium/src/tools/clang
cd ..
TMP_CLANG/clang/scripts/update.py

This installs the Clang binary as ./third_party/llvm-build/Release+Asserts/bin/clang)

The libFuzzer code resides in the LLVM repository, and requires a recent Clang compiler to build (and is used to fuzz various parts of LLVM itself). However the fuzzer itself does not (and should not) depend on any part of LLVM infrastructure and can be used for other projects without requiring the rest of LLVM.

Getting Started

Fuzz Target

The first step in using libFuzzer on a library is to implement a fuzz target – a function that accepts an array of bytes and does something interesting with these bytes using the API under test. Like this:

// fuzz_target.cc
extern "C" int LLVMFuzzerTestOneInput(const uint8_t *Data, size_t Size) {
  DoSomethingInterestingWithMyAPI(Data, Size);
  return 0;  // Non-zero return values are reserved for future use.
}

Note that this fuzz target does not depend on libFuzzer in any way and so it is possible and even desirable to use it with other fuzzing engines e.g. AFL and/or Radamsa.

Some important things to remember about fuzz targets:

  • The fuzzing engine will execute the fuzz target many times with different inputs in the same process.
  • It must tolerate any kind of input (empty, huge, malformed, etc).
  • It must not exit() on any input.
  • It may use threads but ideally all threads should be joined at the end of the function.
  • It must be as deterministic as possible. Non-determinism (e.g. random decisions not based on the input bytes) will make fuzzing inefficient.
  • It must be fast. Try avoiding cubic or greater complexity, logging, or excessive memory consumption.
  • Ideally, it should not modify any global state (although that’s not strict).

Building

Next, build the libFuzzer library as a static archive, without any sanitizer options. Note that the libFuzzer library contains the main() function:

svn co http://llvm.org/svn/llvm-project/llvm/trunk/lib/Fuzzer  # or git clone https://chromium.googlesource.com/chromium/llvm-project/llvm/lib/Fuzzer
./Fuzzer/build.sh  # Produces libFuzzer.a

Then build the fuzzing target function and the library under test using the SanitizerCoverage option, which instruments the code so that the fuzzer can retrieve code coverage information (to guide the fuzzing). Linking with the libFuzzer code then gives a fuzzer executable.

You should also enable one or more of the sanitizers, which help to expose latent bugs by making incorrect behavior generate errors at runtime:

  • AddressSanitizer (ASAN) detects memory access errors. Use -fsanitize=address.
  • UndefinedBehaviorSanitizer (UBSAN) detects the use of various features of C/C++ that are explicitly listed as resulting in undefined behavior. Use -fsanitize=undefined -fno-sanitize-recover=undefined or any individual UBSAN check, e.g. -fsanitize=signed-integer-overflow -fno-sanitize-recover=undefined. You may combine ASAN and UBSAN in one build.
  • MemorySanitizer (MSAN) detects uninitialized reads: code whose behavior relies on memory contents that have not been initialized to a specific value. Use -fsanitize=memory. MSAN can not be combined with other sanirizers and should be used as a seprate build.

Finally, link with libFuzzer.a:

clang -fsanitize-coverage=trace-pc-guard -fsanitize=address your_lib.cc fuzz_target.cc libFuzzer.a -o my_fuzzer

Corpus

Coverage-guided fuzzers like libFuzzer rely on a corpus of sample inputs for the code under test. This corpus should ideally be seeded with a varied collection of valid and invalid inputs for the code under test; for example, for a graphics library the initial corpus might hold a variety of different small PNG/JPG/GIF files. The fuzzer generates random mutations based around the sample inputs in the current corpus. If a mutation triggers execution of a previously-uncovered path in the code under test, then that mutation is saved to the corpus for future variations.

LibFuzzer will work without any initial seeds, but will be less efficient if the library under test accepts complex, structured inputs.

The corpus can also act as a sanity/regression check, to confirm that the fuzzing entrypoint still works and that all of the sample inputs run through the code under test without problems.

If you have a large corpus (either generated by fuzzing or acquired by other means) you may want to minimize it while still preserving the full coverage. One way to do that is to use the -merge=1 flag:

mkdir NEW_CORPUS_DIR  # Store minimized corpus here.
./my_fuzzer -merge=1 NEW_CORPUS_DIR FULL_CORPUS_DIR

You may use the same flag to add more interesting items to an existing corpus. Only the inputs that trigger new coverage will be added to the first corpus.

./my_fuzzer -merge=1 CURRENT_CORPUS_DIR NEW_POTENTIALLY_INTERESTING_INPUTS_DIR

Running

To run the fuzzer, first create a Corpus directory that holds the initial “seed” sample inputs:

mkdir CORPUS_DIR
cp /some/input/samples/* CORPUS_DIR

Then run the fuzzer on the corpus directory:

./my_fuzzer CORPUS_DIR  # -max_len=1000 -jobs=20 ...

As the fuzzer discovers new interesting test cases (i.e. test cases that trigger coverage of new paths through the code under test), those test cases will be added to the corpus directory.

By default, the fuzzing process will continue indefinitely – at least until a bug is found. Any crashes or sanitizer failures will be reported as usual, stopping the fuzzing process, and the particular input that triggered the bug will be written to disk (typically as crash-<sha1>, leak-<sha1>, or timeout-<sha1>).

Parallel Fuzzing

Each libFuzzer process is single-threaded, unless the library under test starts its own threads. However, it is possible to run multiple libFuzzer processes in parallel with a shared corpus directory; this has the advantage that any new inputs found by one fuzzer process will be available to the other fuzzer processes (unless you disable this with the -reload=0 option).

This is primarily controlled by the -jobs=N option, which indicates that that N fuzzing jobs should be run to completion (i.e. until a bug is found or time/iteration limits are reached). These jobs will be run across a set of worker processes, by default using half of the available CPU cores; the count of worker processes can be overridden by the -workers=N option. For example, running with -jobs=30 on a 12-core machine would run 6 workers by default, with each worker averaging 5 bugs by completion of the entire process.

Options

To run the fuzzer, pass zero or more corpus directories as command line arguments. The fuzzer will read test inputs from each of these corpus directories, and any new test inputs that are generated will be written back to the first corpus directory:

./fuzzer [-flag1=val1 [-flag2=val2 ...] ] [dir1 [dir2 ...] ]

If a list of files (rather than directories) are passed to the fuzzer program, then it will re-run those files as test inputs but will not perform any fuzzing. In this mode the fuzzer binary can be used as a regression test (e.g. on a continuous integration system) to check the target function and saved inputs still work.

The most important command line options are:

-help
Print help message.
-seed
Random seed. If 0 (the default), the seed is generated.
-runs
Number of individual test runs, -1 (the default) to run indefinitely.
-max_len
Maximum length of a test input. If 0 (the default), libFuzzer tries to guess a good value based on the corpus (and reports it).
-timeout
Timeout in seconds, default 1200. If an input takes longer than this timeout, the process is treated as a failure case.
-rss_limit_mb
Memory usage limit in Mb, default 2048. Use 0 to disable the limit. If an input requires more than this amount of RSS memory to execute, the process is treated as a failure case. The limit is checked in a separate thread every second. If running w/o ASAN/MSAN, you may use ‘ulimit -v’ instead.
-timeout_exitcode
Exit code (default 77) used if libFuzzer reports a timeout.
-error_exitcode
Exit code (default 77) used if libFuzzer itself (not a sanitizer) reports a bug (leak, OOM, etc).
-max_total_time
If positive, indicates the maximum total time in seconds to run the fuzzer. If 0 (the default), run indefinitely.
-merge
If set to 1, any corpus inputs from the 2nd, 3rd etc. corpus directories that trigger new code coverage will be merged into the first corpus directory. Defaults to 0. This flag can be used to minimize a corpus.
-minimize_crash
If 1, minimizes the provided crash input. Use with -runs=N or -max_total_time=N to limit the number of attempts.
-reload
If set to 1 (the default), the corpus directory is re-read periodically to check for new inputs; this allows detection of new inputs that were discovered by other fuzzing processes.
-jobs
Number of fuzzing jobs to run to completion. Default value is 0, which runs a single fuzzing process until completion. If the value is >= 1, then this number of jobs performing fuzzing are run, in a collection of parallel separate worker processes; each such worker process has its stdout/stderr redirected to fuzz-<JOB>.log.
-workers
Number of simultaneous worker processes to run the fuzzing jobs to completion in. If 0 (the default), min(jobs, NumberOfCpuCores()/2) is used.
-dict
Provide a dictionary of input keywords; see Dictionaries.
-use_counters
Use coverage counters to generate approximate counts of how often code blocks are hit; defaults to 1.
-use_value_profile
Use value profile to guide corpus expansion; defaults to 0.
-only_ascii
If 1, generate only ASCII (isprint``+``isspace) inputs. Defaults to 0.
-artifact_prefix
Provide a prefix to use when saving fuzzing artifacts (crash, timeout, or slow inputs) as $(artifact_prefix)file. Defaults to empty.
-exact_artifact_path
Ignored if empty (the default). If non-empty, write the single artifact on failure (crash, timeout) as $(exact_artifact_path). This overrides -artifact_prefix and will not use checksum in the file name. Do not use the same path for several parallel processes.
-print_pcs
If 1, print out newly covered PCs. Defaults to 0.
-print_final_stats
If 1, print statistics at exit. Defaults to 0.
-detect_leaks
If 1 (default) and if LeakSanitizer is enabled try to detect memory leaks during fuzzing (i.e. not only at shut down).
-close_fd_mask

Indicate output streams to close at startup. Be careful, this will remove diagnostic output from target code (e.g. messages on assert failure).

  • 0 (default): close neither stdout nor stderr
  • 1 : close stdout
  • 2 : close stderr
  • 3 : close both stdout and stderr.

For the full list of flags run the fuzzer binary with -help=1.

Output

During operation the fuzzer prints information to stderr, for example:

INFO: Seed: 1523017872
INFO: Loaded 1 modules (16 guards): [0x744e60, 0x744ea0),
INFO: -max_len is not provided, using 64
INFO: A corpus is not provided, starting from an empty corpus
#0    READ units: 1
#1    INITED cov: 3 ft: 2 corp: 1/1b exec/s: 0 rss: 24Mb
#3811 NEW    cov: 4 ft: 3 corp: 2/2b exec/s: 0 rss: 25Mb L: 1 MS: 5 ChangeBit-ChangeByte-ChangeBit-ShuffleBytes-ChangeByte-
#3827 NEW    cov: 5 ft: 4 corp: 3/4b exec/s: 0 rss: 25Mb L: 2 MS: 1 CopyPart-
#3963 NEW    cov: 6 ft: 5 corp: 4/6b exec/s: 0 rss: 25Mb L: 2 MS: 2 ShuffleBytes-ChangeBit-
#4167 NEW    cov: 7 ft: 6 corp: 5/9b exec/s: 0 rss: 25Mb L: 3 MS: 1 InsertByte-
...

The early parts of the output include information about the fuzzer options and configuration, including the current random seed (in the Seed: line; this can be overridden with the -seed=N flag).

Further output lines have the form of an event code and statistics. The possible event codes are:

READ
The fuzzer has read in all of the provided input samples from the corpus directories.
INITED
The fuzzer has completed initialization, which includes running each of the initial input samples through the code under test.
NEW
The fuzzer has created a test input that covers new areas of the code under test. This input will be saved to the primary corpus directory.
pulse
The fuzzer has generated 2n inputs (generated periodically to reassure the user that the fuzzer is still working).
DONE
The fuzzer has completed operation because it has reached the specified iteration limit (-runs) or time limit (-max_total_time).
MIN<n>
The fuzzer is minimizing the combination of input corpus directories into a single unified corpus (due to the -merge command line option).
RELOAD
The fuzzer is performing a periodic reload of inputs from the corpus directory; this allows it to discover any inputs discovered by other fuzzer processes (see Parallel Fuzzing).

Each output line also reports the following statistics (when non-zero):

cov:
Total number of code blocks or edges covered by the executing the current corpus.
ft:
libFuzzer uses different signals to evaluate the code coverage: edge coverage, edge counters, value profiles, indirect caller/callee pairs, etc. These signals combined are called features (ft:).
corp:
Number of entries in the current in-memory test corpus and its size in bytes.
exec/s:
Number of fuzzer iterations per second.
rss:
Current memory consumption.

For NEW events, the output line also includes information about the mutation operation that produced the new input:

L:
Size of the new input in bytes.
MS: <n> <operations>
Count and list of the mutation operations used to generate the input.

Examples

Toy example

A simple function that does something interesting if it receives the input “HI!”:

cat << EOF > test_fuzzer.cc
#include <stdint.h>
#include <stddef.h>
extern "C" int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
  if (size > 0 && data[0] == 'H')
    if (size > 1 && data[1] == 'I')
       if (size > 2 && data[2] == '!')
       __builtin_trap();
  return 0;
}
EOF
# Build test_fuzzer.cc with asan and link against libFuzzer.a
clang++ -fsanitize=address -fsanitize-coverage=trace-pc-guard test_fuzzer.cc libFuzzer.a
# Run the fuzzer with no corpus.
./a.out

You should get an error pretty quickly:

INFO: Seed: 1523017872
INFO: Loaded 1 modules (16 guards): [0x744e60, 0x744ea0),
INFO: -max_len is not provided, using 64
INFO: A corpus is not provided, starting from an empty corpus
#0    READ units: 1
#1    INITED cov: 3 ft: 2 corp: 1/1b exec/s: 0 rss: 24Mb
#3811 NEW    cov: 4 ft: 3 corp: 2/2b exec/s: 0 rss: 25Mb L: 1 MS: 5 ChangeBit-ChangeByte-ChangeBit-ShuffleBytes-ChangeByte-
#3827 NEW    cov: 5 ft: 4 corp: 3/4b exec/s: 0 rss: 25Mb L: 2 MS: 1 CopyPart-
#3963 NEW    cov: 6 ft: 5 corp: 4/6b exec/s: 0 rss: 25Mb L: 2 MS: 2 ShuffleBytes-ChangeBit-
#4167 NEW    cov: 7 ft: 6 corp: 5/9b exec/s: 0 rss: 25Mb L: 3 MS: 1 InsertByte-
==31511== ERROR: libFuzzer: deadly signal
...
artifact_prefix='./'; Test unit written to ./crash-b13e8756b13a00cf168300179061fb4b91fefbed

More examples

Examples of real-life fuzz targets and the bugs they find can be found at http://tutorial.libfuzzer.info. Among other things you can learn how to detect Heartbleed in one second.

Advanced features

Dictionaries

LibFuzzer supports user-supplied dictionaries with input language keywords or other interesting byte sequences (e.g. multi-byte magic values). Use -dict=DICTIONARY_FILE. For some input languages using a dictionary may significantly improve the search speed. The dictionary syntax is similar to that used by AFL for its -x option:

# Lines starting with '#' and empty lines are ignored.

# Adds "blah" (w/o quotes) to the dictionary.
kw1="blah"
# Use \\ for backslash and \" for quotes.
kw2="\"ac\\dc\""
# Use \xAB for hex values
kw3="\xF7\xF8"
# the name of the keyword followed by '=' may be omitted:
"foo\x0Abar"

Value Profile

EXPERIMENTAL. With an additional compiler flag -fsanitize-coverage=trace-cmp (see SanitizerCoverageTraceDataFlow) and extra run-time flag -use_value_profile=1 the fuzzer will collect value profiles for the parameters of compare instructions and treat some new values as new coverage.

The current imlpementation does roughly the following:

  • The compiler instruments all CMP instructions with a callback that receives both CMP arguments.
  • The callback computes (caller_pc&4095) | (popcnt(Arg1 ^ Arg2) << 12) and uses this value to set a bit in a bitset.
  • Every new observed bit in the bitset is treated as new coverage.

This feature has a potential to discover many interesting inputs, but there are two downsides. First, the extra instrumentation may bring up to 2x additional slowdown. Second, the corpus may grow by several times.

Fuzzer-friendly build mode

Sometimes the code under test is not fuzzing-friendly. Examples:

  • The target code uses a PRNG seeded e.g. by system time and thus two consequent invocations may potentially execute different code paths even if the end result will be the same. This will cause a fuzzer to treat two similar inputs as significantly different and it will blow up the test corpus. E.g. libxml uses rand() inside its hash table.
  • The target code uses checksums to protect from invalid inputs. E.g. png checks CRC for every chunk.

In many cases it makes sense to build a special fuzzing-friendly build with certain fuzzing-unfriendly features disabled. We propose to use a common build macro for all such cases for consistency: FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION.

void MyInitPRNG() {
#ifdef FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION
  // In fuzzing mode the behavior of the code should be deterministic.
  srand(0);
#else
  srand(time(0));
#endif
}

AFL compatibility

LibFuzzer can be used together with AFL on the same test corpus. Both fuzzers expect the test corpus to reside in a directory, one file per input. You can run both fuzzers on the same corpus, one after another:

./afl-fuzz -i testcase_dir -o findings_dir /path/to/program @@
./llvm-fuzz testcase_dir findings_dir  # Will write new tests to testcase_dir

Periodically restart both fuzzers so that they can use each other’s findings. Currently, there is no simple way to run both fuzzing engines in parallel while sharing the same corpus dir.

You may also use AFL on your target function LLVMFuzzerTestOneInput: see an example here.

How good is my fuzzer?

Once you implement your target function LLVMFuzzerTestOneInput and fuzz it to death, you will want to know whether the function or the corpus can be improved further. One easy to use metric is, of course, code coverage. You can get the coverage for your corpus like this:

ASAN_OPTIONS=coverage=1 ./fuzzer CORPUS_DIR -runs=0

This will run all tests in the CORPUS_DIR but will not perform any fuzzing. At the end of the process it will dump a single .sancov file with coverage information. See SanitizerCoverage for details on querying the file using the sancov tool.

You may also use other ways to visualize coverage, e.g. using Clang coverage, but those will require you to rebuild the code with different compiler flags.

User-supplied mutators

LibFuzzer allows to use custom (user-supplied) mutators, see FuzzerInterface.h

Startup initialization

If the library being tested needs to be initialized, there are several options.

The simplest way is to have a statically initialized global object inside LLVMFuzzerTestOneInput (or in global scope if that works for you):

extern "C" int LLVMFuzzerTestOneInput(const uint8_t *Data, size_t Size) {
  static bool Initialized = DoInitialization();
  ...

Alternatively, you may define an optional init function and it will receive the program arguments that you can read and modify. Do this only if you realy need to access argv/argc.

extern "C" int LLVMFuzzerInitialize(int *argc, char ***argv) {
 ReadAndMaybeModify(argc, argv);
 return 0;
}

Leaks

Binaries built with AddressSanitizer or LeakSanitizer will try to detect memory leaks at the process shutdown. For in-process fuzzing this is inconvenient since the fuzzer needs to report a leak with a reproducer as soon as the leaky mutation is found. However, running full leak detection after every mutation is expensive.

By default (-detect_leaks=1) libFuzzer will count the number of malloc and free calls when executing every mutation. If the numbers don’t match (which by itself doesn’t mean there is a leak) libFuzzer will invoke the more expensive LeakSanitizer pass and if the actual leak is found, it will be reported with the reproducer and the process will exit.

If your target has massive leaks and the leak detection is disabled you will eventually run out of RAM (see the -rss_limit_mb flag).

Developing libFuzzer

Building libFuzzer as a part of LLVM project and running its test requires fresh clang as the host compiler and special CMake configuration:

cmake -GNinja  -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DLLVM_USE_SANITIZER=Address -DLLVM_USE_SANITIZE_COVERAGE=YES -DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_ASSERTIONS=ON /path/to/llvm
ninja check-fuzzer

Fuzzing components of LLVM

To build any of the LLVM fuzz targets use the build instructions above.

clang-format-fuzzer

The inputs are random pieces of C++-like text.

ninja clang-format-fuzzer
mkdir CORPUS_DIR
./bin/clang-format-fuzzer CORPUS_DIR

Optionally build other kinds of binaries (ASan+Debug, MSan, UBSan, etc).

Tracking bug: https://llvm.org/bugs/show_bug.cgi?id=23052

clang-fuzzer

The behavior is very similar to clang-format-fuzzer.

Tracking bug: https://llvm.org/bugs/show_bug.cgi?id=23057

llvm-mc-fuzzer

This tool fuzzes the MC layer. Currently it is only able to fuzz the disassembler but it is hoped that assembly, and round-trip verification will be added in future.

When run in dissassembly mode, the inputs are opcodes to be disassembled. The fuzzer will consume as many instructions as possible and will stop when it finds an invalid instruction or runs out of data.

Please note that the command line interface differs slightly from that of other fuzzers. The fuzzer arguments should follow --fuzzer-args and should have a single dash, while other arguments control the operation mode and target in a similar manner to llvm-mc and should have two dashes. For example:

llvm-mc-fuzzer --triple=aarch64-linux-gnu --disassemble --fuzzer-args -max_len=4 -jobs=10

Buildbot

A buildbot continuously runs the above fuzzers for LLVM components, with results shown at http://lab.llvm.org:8011/builders/sanitizer-x86_64-linux-fuzzer .

FAQ

Q. Why doesn’t libFuzzer use any of the LLVM support?

There are two reasons.

First, we want this library to be used outside of the LLVM without users having to build the rest of LLVM. This may sound unconvincing for many LLVM folks, but in practice the need for building the whole LLVM frightens many potential users – and we want more users to use this code.

Second, there is a subtle technical reason not to rely on the rest of LLVM, or any other large body of code (maybe not even STL). When coverage instrumentation is enabled, it will also instrument the LLVM support code which will blow up the coverage set of the process (since the fuzzer is in-process). In other words, by using more external dependencies we will slow down the fuzzer while the main reason for it to exist is extreme speed.

Q. What about Windows then? The fuzzer contains code that does not build on Windows.

Volunteers are welcome.

Q. When libFuzzer is not a good solution for a problem?

  • If the test inputs are validated by the target library and the validator asserts/crashes on invalid inputs, in-process fuzzing is not applicable.
  • Bugs in the target library may accumulate without being detected. E.g. a memory corruption that goes undetected at first and then leads to a crash while testing another input. This is why it is highly recommended to run this in-process fuzzer with all sanitizers to detect most bugs on the spot.
  • It is harder to protect the in-process fuzzer from excessive memory consumption and infinite loops in the target library (still possible).
  • The target library should not have significant global state that is not reset between the runs.
  • Many interesting target libraries are not designed in a way that supports the in-process fuzzer interface (e.g. require a file path instead of a byte array).
  • If a single test run takes a considerable fraction of a second (or more) the speed benefit from the in-process fuzzer is negligible.
  • If the target library runs persistent threads (that outlive execution of one test) the fuzzing results will be unreliable.

Q. So, what exactly this Fuzzer is good for?

This Fuzzer might be a good choice for testing libraries that have relatively small inputs, each input takes < 10ms to run, and the library code is not expected to crash on invalid inputs. Examples: regular expression matchers, text or binary format parsers, compression, network, crypto.