xref: /llvm-project/llvm/docs/LibFuzzer.rst (revision faea104ef17b00d320017d2cd0d10a89277dc48a)
1=======================================================
2libFuzzer – a library for coverage-guided fuzz testing.
3=======================================================
4.. contents::
5   :local:
6   :depth: 1
7
8Introduction
9============
10
11LibFuzzer is an in-process, coverage-guided, evolutionary fuzzing engine.
12
13LibFuzzer is linked with the library under test, and feeds fuzzed inputs to the
14library via a specific fuzzing entrypoint (aka "target function"); the fuzzer
15then tracks which areas of the code are reached, and generates mutations on the
16corpus of input data in order to maximize the code coverage.
17The code coverage
18information for libFuzzer is provided by LLVM's SanitizerCoverage_
19instrumentation.
20
21Contact: libfuzzer(#)googlegroups.com
22
23Status
24======
25
26The original authors of libFuzzer have stopped active work on it and switched
27to working on another fuzzing engine, Centipede_. LibFuzzer is still fully
28supported in that important bugs will get fixed. However, please do not expect
29major new features or code reviews, other than for bug fixes.
30
31Versions
32========
33
34LibFuzzer requires a matching version of Clang.
35
36
37Getting Started
38===============
39
40.. contents::
41   :local:
42   :depth: 1
43
44Fuzz Target
45-----------
46
47The first step in using libFuzzer on a library is to implement a
48*fuzz target* -- a function that accepts an array of bytes and
49does something interesting with these bytes using the API under test.
50Like this:
51
52.. code-block:: c++
53
54  // fuzz_target.cc
55  extern "C" int LLVMFuzzerTestOneInput(const uint8_t *Data, size_t Size) {
56    DoSomethingInterestingWithMyAPI(Data, Size);
57    return 0;  // Values other than 0 and -1 are reserved for future use.
58  }
59
60Note that this fuzz target does not depend on libFuzzer in any way
61and so it is possible and even desirable to use it with other fuzzing engines
62e.g. AFL_ and/or Radamsa_.
63
64Some important things to remember about fuzz targets:
65
66* The fuzzing engine will execute the fuzz target many times with different inputs in the same process.
67* It must tolerate any kind of input (empty, huge, malformed, etc).
68* It must not `exit()` on any input.
69* It may use threads but ideally all threads should be joined at the end of the function.
70* It must be as deterministic as possible. Non-determinism (e.g. random decisions not based on the input bytes) will make fuzzing inefficient.
71* It must be fast. Try avoiding cubic or greater complexity, logging, or excessive memory consumption.
72* Ideally, it should not modify any global state (although that's not strict).
73* Usually, the narrower the target the better. E.g. if your target can parse several data formats, split it into several targets, one per format.
74
75
76Fuzzer Usage
77------------
78
79Recent versions of Clang (starting from 6.0) include libFuzzer, and no extra installation is necessary.
80
81In order to build your fuzzer binary, use the `-fsanitize=fuzzer` flag during the
82compilation and linking. In most cases you may want to combine libFuzzer with
83AddressSanitizer_ (ASAN), UndefinedBehaviorSanitizer_ (UBSAN), or both.  You can
84also build with MemorySanitizer_ (MSAN), but support is experimental::
85
86   clang -g -O1 -fsanitize=fuzzer                         mytarget.c # Builds the fuzz target w/o sanitizers
87   clang -g -O1 -fsanitize=fuzzer,address                 mytarget.c # Builds the fuzz target with ASAN
88   clang -g -O1 -fsanitize=fuzzer,signed-integer-overflow mytarget.c # Builds the fuzz target with a part of UBSAN
89   clang -g -O1 -fsanitize=fuzzer,memory                  mytarget.c # Builds the fuzz target with MSAN
90
91This will perform the necessary instrumentation, as well as linking with the libFuzzer library.
92Note that ``-fsanitize=fuzzer`` links in the libFuzzer's ``main()`` symbol.
93
94If modifying ``CFLAGS`` of a large project, which also compiles executables
95requiring their own ``main`` symbol, it may be desirable to request just the
96instrumentation without linking::
97
98   clang -fsanitize=fuzzer-no-link mytarget.c
99
100Then libFuzzer can be linked to the desired driver by passing in
101``-fsanitize=fuzzer`` during the linking stage.
102
103.. _libfuzzer-corpus:
104
105Corpus
106------
107
108Coverage-guided fuzzers like libFuzzer rely on a corpus of sample inputs for the
109code under test.  This corpus should ideally be seeded with a varied collection
110of valid and invalid inputs for the code under test; for example, for a graphics
111library the initial corpus might hold a variety of different small PNG/JPG/GIF
112files.  The fuzzer generates random mutations based around the sample inputs in
113the current corpus.  If a mutation triggers execution of a previously-uncovered
114path in the code under test, then that mutation is saved to the corpus for
115future variations.
116
117LibFuzzer will work without any initial seeds, but will be less
118efficient if the library under test accepts complex,
119structured inputs.
120
121The corpus can also act as a sanity/regression check, to confirm that the
122fuzzing entrypoint still works and that all of the sample inputs run through
123the code under test without problems.
124
125If you have a large corpus (either generated by fuzzing or acquired by other means)
126you may want to minimize it while still preserving the full coverage. One way to do that
127is to use the `-merge=1` flag:
128
129.. code-block:: console
130
131  mkdir NEW_CORPUS_DIR  # Store minimized corpus here.
132  ./my_fuzzer -merge=1 NEW_CORPUS_DIR FULL_CORPUS_DIR
133
134You may use the same flag to add more interesting items to an existing corpus.
135Only the inputs that trigger new coverage will be added to the first corpus.
136
137.. code-block:: console
138
139  ./my_fuzzer -merge=1 CURRENT_CORPUS_DIR NEW_POTENTIALLY_INTERESTING_INPUTS_DIR
140
141Running
142-------
143
144To run the fuzzer, first create a Corpus_ directory that holds the
145initial "seed" sample inputs:
146
147.. code-block:: console
148
149  mkdir CORPUS_DIR
150  cp /some/input/samples/* CORPUS_DIR
151
152Then run the fuzzer on the corpus directory:
153
154.. code-block:: console
155
156  ./my_fuzzer CORPUS_DIR  # -max_len=1000 -jobs=20 ...
157
158As the fuzzer discovers new interesting test cases (i.e. test cases that
159trigger coverage of new paths through the code under test), those test cases
160will be added to the corpus directory.
161
162By default, the fuzzing process will continue indefinitely – at least until
163a bug is found.  Any crashes or sanitizer failures will be reported as usual,
164stopping the fuzzing process, and the particular input that triggered the bug
165will be written to disk (typically as ``crash-<sha1>``, ``leak-<sha1>``,
166or ``timeout-<sha1>``).
167
168
169Parallel Fuzzing
170----------------
171
172Each libFuzzer process is single-threaded, unless the library under test starts
173its own threads.  However, it is possible to run multiple libFuzzer processes in
174parallel with a shared corpus directory; this has the advantage that any new
175inputs found by one fuzzer process will be available to the other fuzzer
176processes (unless you disable this with the ``-reload=0`` option).
177
178This is primarily controlled by the ``-jobs=N`` option, which indicates that
179that `N` fuzzing jobs should be run to completion (i.e. until a bug is found or
180time/iteration limits are reached).  These jobs will be run across a set of
181worker processes, by default using half of the available CPU cores; the count of
182worker processes can be overridden by the ``-workers=N`` option.  For example,
183running with ``-jobs=30`` on a 12-core machine would run 6 workers by default,
184with each worker averaging 5 bugs by completion of the entire process.
185
186Fork mode
187---------
188
189**Experimental** mode ``-fork=N`` (where ``N`` is the number of parallel jobs)
190enables oom-, timeout-, and crash-resistant
191fuzzing with separate processes (using ``fork-exec``, not just ``fork``).
192
193The top libFuzzer process will not do any fuzzing itself, but will
194spawn up to ``N`` concurrent child processes providing them
195small random subsets of the corpus. After a child exits, the top process
196merges the corpus generated by the child back to the main corpus.
197
198Related flags:
199
200``-ignore_ooms``
201  True by default. If an OOM happens during fuzzing in one of the child processes,
202  the reproducer is saved on disk, and fuzzing continues.
203``-ignore_timeouts``
204  True by default, same as ``-ignore_ooms``, but for timeouts.
205``-ignore_crashes``
206  False by default, same as ``-ignore_ooms``, but for all other crashes.
207
208The plan is to eventually replace ``-jobs=N`` and ``-workers=N`` with ``-fork=N``.
209
210Resuming merge
211--------------
212
213Merging large corpora may be time consuming, and it is often desirable to do it
214on preemptable VMs, where the process may be killed at any time.
215In order to seamlessly resume the merge, use the ``-merge_control_file`` flag
216and use ``killall -SIGUSR1 /path/to/fuzzer/binary`` to stop the merge gracefully. Example:
217
218.. code-block:: console
219
220  % rm -f SomeLocalPath
221  % ./my_fuzzer CORPUS1 CORPUS2 -merge=1 -merge_control_file=SomeLocalPath
222  ...
223  MERGE-INNER: using the control file 'SomeLocalPath'
224  ...
225  # While this is running, do `killall -SIGUSR1 my_fuzzer` in another console
226  ==9015== INFO: libFuzzer: exiting as requested
227
228  # This will leave the file SomeLocalPath with the partial state of the merge.
229  # Now, you can continue the merge by executing the same command. The merge
230  # will continue from where it has been interrupted.
231  % ./my_fuzzer CORPUS1 CORPUS2 -merge=1 -merge_control_file=SomeLocalPath
232  ...
233  MERGE-OUTER: non-empty control file provided: 'SomeLocalPath'
234  MERGE-OUTER: control file ok, 32 files total, first not processed file 20
235  ...
236
237Options
238=======
239
240To run the fuzzer, pass zero or more corpus directories as command line
241arguments.  The fuzzer will read test inputs from each of these corpus
242directories, and any new test inputs that are generated will be written
243back to the first corpus directory:
244
245.. code-block:: console
246
247  ./fuzzer [-flag1=val1 [-flag2=val2 ...] ] [dir1 [dir2 ...] ]
248
249If a list of files (rather than directories) are passed to the fuzzer program,
250then it will re-run those files as test inputs but will not perform any fuzzing.
251In this mode the fuzzer binary can be used as a regression test (e.g. on a
252continuous integration system) to check the target function and saved inputs
253still work.
254
255The most important command line options are:
256
257``-help``
258  Print help message (``-help=1``).
259``-seed``
260  Random seed. If 0 (the default), the seed is generated.
261``-runs``
262  Number of individual test runs, -1 (the default) to run indefinitely.
263``-max_len``
264  Maximum length of a test input. If 0 (the default), libFuzzer tries to guess
265  a good value based on the corpus (and reports it).
266``-len_control``
267  Try generating small inputs first, then try larger inputs over time.
268  Specifies the rate at which the length limit is increased (smaller == faster).
269  Default is 100. If 0, immediately try inputs with size up to max_len.
270``-timeout``
271  Timeout in seconds, default 1200. If an input takes longer than this timeout,
272  the process is treated as a failure case.
273``-rss_limit_mb``
274  Memory usage limit in Mb, default 2048. Use 0 to disable the limit.
275  If an input requires more than this amount of RSS memory to execute,
276  the process is treated as a failure case.
277  The limit is checked in a separate thread every second.
278  If running w/o ASAN/MSAN, you may use 'ulimit -v' instead.
279``-malloc_limit_mb``
280  If non-zero, the fuzzer will exit if the target tries to allocate this
281  number of Mb with one malloc call.
282  If zero (default) same limit as rss_limit_mb is applied.
283``-timeout_exitcode``
284  Exit code (default 77) used if libFuzzer reports a timeout.
285``-error_exitcode``
286  Exit code (default 77) used if libFuzzer itself (not a sanitizer) reports a bug (leak, OOM, etc).
287``-max_total_time``
288  If positive, indicates the maximum total time in seconds to run the fuzzer.
289  If 0 (the default), run indefinitely.
290``-merge``
291  If set to 1, any corpus inputs from the 2nd, 3rd etc. corpus directories
292  that trigger new code coverage will be merged into the first corpus
293  directory.  Defaults to 0. This flag can be used to minimize a corpus.
294``-merge_control_file``
295  Specify a control file used for the merge process.
296  If a merge process gets killed it tries to leave this file in a state
297  suitable for resuming the merge. By default a temporary file will be used.
298``-minimize_crash``
299  If 1, minimizes the provided crash input.
300  Use with -runs=N or -max_total_time=N to limit the number of attempts.
301``-reload``
302  If set to 1 (the default), the corpus directory is re-read periodically to
303  check for new inputs; this allows detection of new inputs that were discovered
304  by other fuzzing processes.
305``-jobs``
306  Number of fuzzing jobs to run to completion. Default value is 0, which runs a
307  single fuzzing process until completion.  If the value is >= 1, then this
308  number of jobs performing fuzzing are run, in a collection of parallel
309  separate worker processes; each such worker process has its
310  ``stdout``/``stderr`` redirected to ``fuzz-<JOB>.log``.
311``-workers``
312  Number of simultaneous worker processes to run the fuzzing jobs to completion
313  in. If 0 (the default), ``min(jobs, NumberOfCpuCores()/2)`` is used.
314``-dict``
315  Provide a dictionary of input keywords; see Dictionaries_.
316``-use_counters``
317  Use `coverage counters`_ to generate approximate counts of how often code
318  blocks are hit; defaults to 1.
319``-reduce_inputs``
320  Try to reduce the size of inputs while preserving their full feature sets;
321  defaults to 1.
322``-use_value_profile``
323  Use `value profile`_ to guide corpus expansion; defaults to 0.
324``-only_ascii``
325  If 1, generate only ASCII (``isprint``+``isspace``) inputs. Defaults to 0.
326``-artifact_prefix``
327  Provide a prefix to use when saving fuzzing artifacts (crash, timeout, or
328  slow inputs) as ``$(artifact_prefix)file``.  Defaults to empty.
329``-exact_artifact_path``
330  Ignored if empty (the default).  If non-empty, write the single artifact on
331  failure (crash, timeout) as ``$(exact_artifact_path)``. This overrides
332  ``-artifact_prefix`` and will not use checksum in the file name. Do not use
333  the same path for several parallel processes.
334``-print_pcs``
335  If 1, print out newly covered PCs. Defaults to 0.
336``-print_final_stats``
337  If 1, print statistics at exit.  Defaults to 0.
338``-detect_leaks``
339  If 1 (default) and if LeakSanitizer is enabled
340  try to detect memory leaks during fuzzing (i.e. not only at shut down).
341``-close_fd_mask``
342  Indicate output streams to close at startup. Be careful, this will
343  remove diagnostic output from target code (e.g. messages on assert failure).
344
345   - 0 (default): close neither ``stdout`` nor ``stderr``
346   - 1 : close ``stdout``
347   - 2 : close ``stderr``
348   - 3 : close both ``stdout`` and ``stderr``.
349
350For the full list of flags run the fuzzer binary with ``-help=1``.
351
352Output
353======
354
355During operation the fuzzer prints information to ``stderr``, for example::
356
357  INFO: Seed: 1523017872
358  INFO: Loaded 1 modules (16 guards): [0x744e60, 0x744ea0),
359  INFO: -max_len is not provided, using 64
360  INFO: A corpus is not provided, starting from an empty corpus
361  #0	READ units: 1
362  #1	INITED cov: 3 ft: 2 corp: 1/1b exec/s: 0 rss: 24Mb
363  #3811	NEW    cov: 4 ft: 3 corp: 2/2b exec/s: 0 rss: 25Mb L: 1 MS: 5 ChangeBit-ChangeByte-ChangeBit-ShuffleBytes-ChangeByte-
364  #3827	NEW    cov: 5 ft: 4 corp: 3/4b exec/s: 0 rss: 25Mb L: 2 MS: 1 CopyPart-
365  #3963	NEW    cov: 6 ft: 5 corp: 4/6b exec/s: 0 rss: 25Mb L: 2 MS: 2 ShuffleBytes-ChangeBit-
366  #4167	NEW    cov: 7 ft: 6 corp: 5/9b exec/s: 0 rss: 25Mb L: 3 MS: 1 InsertByte-
367  ...
368
369The early parts of the output include information about the fuzzer options and
370configuration, including the current random seed (in the ``Seed:`` line; this
371can be overridden with the ``-seed=N`` flag).
372
373Further output lines have the form of an event code and statistics.  The
374possible event codes are:
375
376``READ``
377  The fuzzer has read in all of the provided input samples from the corpus
378  directories.
379``INITED``
380  The fuzzer has completed initialization, which includes running each of
381  the initial input samples through the code under test.
382``NEW``
383  The fuzzer has created a test input that covers new areas of the code
384  under test.  This input will be saved to the primary corpus directory.
385``REDUCE``
386  The fuzzer has found a better (smaller) input that triggers previously
387  discovered features (set ``-reduce_inputs=0`` to disable).
388``pulse``
389  The fuzzer has generated 2\ :sup:`n` inputs (generated periodically to reassure
390  the user that the fuzzer is still working).
391``DONE``
392  The fuzzer has completed operation because it has reached the specified
393  iteration limit (``-runs``) or time limit (``-max_total_time``).
394``RELOAD``
395  The fuzzer is performing a periodic reload of inputs from the corpus
396  directory; this allows it to discover any inputs discovered by other
397  fuzzer processes (see `Parallel Fuzzing`_).
398
399Each output line also reports the following statistics (when non-zero):
400
401``cov:``
402  Total number of code blocks or edges covered by executing the current corpus.
403``ft:``
404  libFuzzer uses different signals to evaluate the code coverage:
405  edge coverage, edge counters, value profiles, indirect caller/callee pairs, etc.
406  These signals combined are called *features* (`ft:`).
407``corp:``
408  Number of entries in the current in-memory test corpus and its size in bytes.
409``lim:``
410  Current limit on the length of new entries in the corpus.  Increases over time
411  until the max length (``-max_len``) is reached.
412``exec/s:``
413  Number of fuzzer iterations per second.
414``rss:``
415  Current memory consumption.
416
417For ``NEW`` and ``REDUCE`` events, the output line also includes information
418about the mutation operation that produced the new input:
419
420``L:``
421  Size of the new input in bytes.
422``MS: <n> <operations>``
423  Count and list of the mutation operations used to generate the input.
424
425
426Examples
427========
428.. contents::
429   :local:
430   :depth: 1
431
432Toy example
433-----------
434
435A simple function that does something interesting if it receives the input
436"HI!"::
437
438  cat << EOF > test_fuzzer.cc
439  #include <stdint.h>
440  #include <stddef.h>
441  extern "C" int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
442    if (size > 0 && data[0] == 'H')
443      if (size > 1 && data[1] == 'I')
444         if (size > 2 && data[2] == '!')
445         __builtin_trap();
446    return 0;
447  }
448  EOF
449  # Build test_fuzzer.cc with asan and link against libFuzzer.
450  clang++ -fsanitize=address,fuzzer test_fuzzer.cc
451  # Run the fuzzer with no corpus.
452  ./a.out
453
454You should get an error pretty quickly::
455
456  INFO: Seed: 1523017872
457  INFO: Loaded 1 modules (16 guards): [0x744e60, 0x744ea0),
458  INFO: -max_len is not provided, using 64
459  INFO: A corpus is not provided, starting from an empty corpus
460  #0	READ units: 1
461  #1	INITED cov: 3 ft: 2 corp: 1/1b exec/s: 0 rss: 24Mb
462  #3811	NEW    cov: 4 ft: 3 corp: 2/2b exec/s: 0 rss: 25Mb L: 1 MS: 5 ChangeBit-ChangeByte-ChangeBit-ShuffleBytes-ChangeByte-
463  #3827	NEW    cov: 5 ft: 4 corp: 3/4b exec/s: 0 rss: 25Mb L: 2 MS: 1 CopyPart-
464  #3963	NEW    cov: 6 ft: 5 corp: 4/6b exec/s: 0 rss: 25Mb L: 2 MS: 2 ShuffleBytes-ChangeBit-
465  #4167	NEW    cov: 7 ft: 6 corp: 5/9b exec/s: 0 rss: 25Mb L: 3 MS: 1 InsertByte-
466  ==31511== ERROR: libFuzzer: deadly signal
467  ...
468  artifact_prefix='./'; Test unit written to ./crash-b13e8756b13a00cf168300179061fb4b91fefbed
469
470
471More examples
472-------------
473
474Examples of real-life fuzz targets and the bugs they find can be found
475at http://tutorial.libfuzzer.info. Among other things you can learn how
476to detect Heartbleed_ in one second.
477
478
479Advanced features
480=================
481.. contents::
482   :local:
483   :depth: 1
484
485Dictionaries
486------------
487LibFuzzer supports user-supplied dictionaries with input language keywords
488or other interesting byte sequences (e.g. multi-byte magic values).
489Use ``-dict=DICTIONARY_FILE``. For some input languages using a dictionary
490may significantly improve the search speed.
491The dictionary syntax is similar to that used by AFL_ for its ``-x`` option::
492
493  # Lines starting with '#' and empty lines are ignored.
494
495  # Adds "blah" (w/o quotes) to the dictionary.
496  kw1="blah"
497  # Use \\ for backslash and \" for quotes.
498  kw2="\"ac\\dc\""
499  # Use \xAB for hex values
500  kw3="\xF7\xF8"
501  # the name of the keyword followed by '=' may be omitted:
502  "foo\x0Abar"
503
504
505
506Tracing CMP instructions
507------------------------
508
509With an additional compiler flag ``-fsanitize-coverage=trace-cmp``
510(on by default as part of ``-fsanitize=fuzzer``, see SanitizerCoverageTraceDataFlow_)
511libFuzzer will intercept CMP instructions and guide mutations based
512on the arguments of intercepted CMP instructions. This may slow down
513the fuzzing but is very likely to improve the results.
514
515Value Profile
516-------------
517
518With  ``-fsanitize-coverage=trace-cmp`` (default with ``-fsanitize=fuzzer``)
519and extra run-time flag ``-use_value_profile=1`` the fuzzer will
520collect value profiles for the parameters of compare instructions
521and treat some new values as new coverage.
522
523The current implementation does roughly the following:
524
525* The compiler instruments all CMP instructions with a callback that receives both CMP arguments.
526* The callback computes `(caller_pc&4095) | (popcnt(Arg1 ^ Arg2) << 12)` and uses this value to set a bit in a bitset.
527* Every new observed bit in the bitset is treated as new coverage.
528
529
530This feature has a potential to discover many interesting inputs,
531but there are two downsides.
532First, the extra instrumentation may bring up to 2x additional slowdown.
533Second, the corpus may grow by several times.
534
535Fuzzer-friendly build mode
536---------------------------
537Sometimes the code under test is not fuzzing-friendly. Examples:
538
539  - The target code uses a PRNG seeded e.g. by system time and
540    thus two consequent invocations may potentially execute different code paths
541    even if the end result will be the same. This will cause a fuzzer to treat
542    two similar inputs as significantly different and it will blow up the test corpus.
543    E.g. libxml uses ``rand()`` inside its hash table.
544  - The target code uses checksums to protect from invalid inputs.
545    E.g. png checks CRC for every chunk.
546
547In many cases it makes sense to build a special fuzzing-friendly build
548with certain fuzzing-unfriendly features disabled. We propose to use a common build macro
549for all such cases for consistency: ``FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION``.
550
551.. code-block:: c++
552
553  void MyInitPRNG() {
554  #ifdef FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION
555    // In fuzzing mode the behavior of the code should be deterministic.
556    srand(0);
557  #else
558    srand(time(0));
559  #endif
560  }
561
562
563
564AFL compatibility
565-----------------
566LibFuzzer can be used together with AFL_ on the same test corpus.
567Both fuzzers expect the test corpus to reside in a directory, one file per input.
568You can run both fuzzers on the same corpus, one after another:
569
570.. code-block:: console
571
572  ./afl-fuzz -i testcase_dir -o findings_dir /path/to/program @@
573  ./llvm-fuzz testcase_dir findings_dir  # Will write new tests to testcase_dir
574
575Periodically restart both fuzzers so that they can use each other's findings.
576Currently, there is no simple way to run both fuzzing engines in parallel while sharing the same corpus dir.
577
578You may also use AFL on your target function ``LLVMFuzzerTestOneInput``:
579see an example `here <https://github.com/llvm/llvm-project/tree/main/compiler-rt/lib/fuzzer/afl>`__.
580
581How good is my fuzzer?
582----------------------
583
584Once you implement your target function ``LLVMFuzzerTestOneInput`` and fuzz it to death,
585you will want to know whether the function or the corpus can be improved further.
586One easy to use metric is, of course, code coverage.
587
588We recommend to use
589`Clang Coverage <https://clang.llvm.org/docs/SourceBasedCodeCoverage.html>`_,
590to visualize and study your code coverage
591(`example <https://github.com/google/fuzzer-test-suite/blob/master/tutorial/libFuzzerTutorial.md#visualizing-coverage>`_).
592
593
594User-supplied mutators
595----------------------
596
597LibFuzzer allows to use custom (user-supplied) mutators, see
598`Structure-Aware Fuzzing <https://github.com/google/fuzzing/blob/master/docs/structure-aware-fuzzing.md>`_
599for more details.
600
601Startup initialization
602----------------------
603If the library being tested needs to be initialized, there are several options.
604
605The simplest way is to have a statically initialized global object inside
606`LLVMFuzzerTestOneInput` (or in global scope if that works for you):
607
608.. code-block:: c++
609
610  extern "C" int LLVMFuzzerTestOneInput(const uint8_t *Data, size_t Size) {
611    static bool Initialized = DoInitialization();
612    ...
613
614Alternatively, you may define an optional init function and it will receive
615the program arguments that you can read and modify. Do this **only** if you
616really need to access ``argv``/``argc``.
617
618.. code-block:: c++
619
620   extern "C" int LLVMFuzzerInitialize(int *argc, char ***argv) {
621    ReadAndMaybeModify(argc, argv);
622    return 0;
623   }
624
625Using libFuzzer as a library
626----------------------------
627If the code being fuzzed must provide its own `main`, it's possible to
628invoke libFuzzer as a library. Be sure to pass ``-fsanitize=fuzzer-no-link``
629during compilation, and link your binary against the no-main version of
630libFuzzer. On Linux installations, this is typically located at:
631
632.. code-block:: bash
633
634  /usr/lib/<llvm-version>/lib/clang/<clang-version>/lib/linux/libclang_rt.fuzzer_no_main-<architecture>.a
635
636If building libFuzzer from source, this is located at the following path
637in the build output directory:
638
639.. code-block:: bash
640
641  lib/linux/libclang_rt.fuzzer_no_main-<architecture>.a
642
643From here, the code can do whatever setup it requires, and when it's ready
644to start fuzzing, it can call `LLVMFuzzerRunDriver`, passing in the program
645arguments and a callback. This callback is invoked just like
646`LLVMFuzzerTestOneInput`, and has the same signature.
647
648.. code-block:: c++
649
650  extern "C" int LLVMFuzzerRunDriver(int *argc, char ***argv,
651                    int (*UserCb)(const uint8_t *Data, size_t Size));
652
653
654Rejecting unwanted inputs
655-------------------------
656
657It may be desirable to reject some inputs, i.e. to not add them to the corpus.
658
659For example, when fuzzing an API consisting of parsing and other logic,
660one may want to allow only those inputs into the corpus that parse successfully.
661
662If the fuzz target returns -1 on a given input,
663libFuzzer will not add that input top the corpus, regardless of what coverage
664it triggers.
665
666
667.. code-block:: c++
668
669  extern "C" int LLVMFuzzerTestOneInput(const uint8_t *Data, size_t Size) {
670    if (auto *Obj = ParseMe(Data, Size)) {
671      Obj->DoSomethingInteresting();
672      return 0;  // Accept. The input may be added to the corpus.
673    }
674    return -1;  // Reject; The input will not be added to the corpus.
675  }
676
677Leaks
678-----
679
680Binaries built with AddressSanitizer_ or LeakSanitizer_ will try to detect
681memory leaks at the process shutdown.
682For in-process fuzzing this is inconvenient
683since the fuzzer needs to report a leak with a reproducer as soon as the leaky
684mutation is found. However, running full leak detection after every mutation
685is expensive.
686
687By default (``-detect_leaks=1``) libFuzzer will count the number of
688``malloc`` and ``free`` calls when executing every mutation.
689If the numbers don't match (which by itself doesn't mean there is a leak)
690libFuzzer will invoke the more expensive LeakSanitizer_
691pass and if the actual leak is found, it will be reported with the reproducer
692and the process will exit.
693
694If your target has massive leaks and the leak detection is disabled
695you will eventually run out of RAM (see the ``-rss_limit_mb`` flag).
696
697
698Developing libFuzzer
699====================
700
701LibFuzzer is built as a part of LLVM project by default on macos and Linux.
702Users of other operating systems can explicitly request compilation using
703``-DCOMPILER_RT_BUILD_LIBFUZZER=ON`` flag.
704Tests are run using ``check-fuzzer`` target from the build directory
705which was configured with ``-DCOMPILER_RT_INCLUDE_TESTS=ON`` flag.
706
707.. code-block:: console
708
709    ninja check-fuzzer
710
711
712FAQ
713=========================
714
715Q. Why doesn't libFuzzer use any of the LLVM support?
716-----------------------------------------------------
717
718There are two reasons.
719
720First, we want this library to be used outside of the LLVM without users having to
721build the rest of LLVM. This may sound unconvincing for many LLVM folks,
722but in practice the need for building the whole LLVM frightens many potential
723users -- and we want more users to use this code.
724
725Second, there is a subtle technical reason not to rely on the rest of LLVM, or
726any other large body of code (maybe not even STL). When coverage instrumentation
727is enabled, it will also instrument the LLVM support code which will blow up the
728coverage set of the process (since the fuzzer is in-process). In other words, by
729using more external dependencies we will slow down the fuzzer while the main
730reason for it to exist is extreme speed.
731
732Q. Does libFuzzer Support Windows?
733------------------------------------------------------------------------------------
734
735Yes, libFuzzer now supports Windows. Initial support was added in r341082.
736Any build of Clang 9 supports it. You can download a build of Clang for Windows
737that has libFuzzer from
738`LLVM Snapshot Builds <https://llvm.org/builds/>`_.
739
740Using libFuzzer on Windows without ASAN is unsupported. Building fuzzers with the
741``/MD`` (dynamic runtime library) compile option is unsupported. Support for these
742may be added in the future. Linking fuzzers with the ``/INCREMENTAL`` link option
743(or the ``/DEBUG`` option which implies it) is also unsupported.
744
745Send any questions or comments to the mailing list: libfuzzer(#)googlegroups.com
746
747Q. When libFuzzer is not a good solution for a problem?
748---------------------------------------------------------
749
750* If the test inputs are validated by the target library and the validator
751  asserts/crashes on invalid inputs, in-process fuzzing is not applicable.
752* Bugs in the target library may accumulate without being detected. E.g. a memory
753  corruption that goes undetected at first and then leads to a crash while
754  testing another input. This is why it is highly recommended to run this
755  in-process fuzzer with all sanitizers to detect most bugs on the spot.
756* It is harder to protect the in-process fuzzer from excessive memory
757  consumption and infinite loops in the target library (still possible).
758* The target library should not have significant global state that is not
759  reset between the runs.
760* Many interesting target libraries are not designed in a way that supports
761  the in-process fuzzer interface (e.g. require a file path instead of a
762  byte array).
763* If a single test run takes a considerable fraction of a second (or
764  more) the speed benefit from the in-process fuzzer is negligible.
765* If the target library runs persistent threads (that outlive
766  execution of one test) the fuzzing results will be unreliable.
767
768Q. So, what exactly this Fuzzer is good for?
769--------------------------------------------
770
771This Fuzzer might be a good choice for testing libraries that have relatively
772small inputs, each input takes < 10ms to run, and the library code is not expected
773to crash on invalid inputs.
774Examples: regular expression matchers, text or binary format parsers, compression,
775network, crypto.
776
777Q. LibFuzzer crashes on my complicated fuzz target (but works fine for me on smaller targets).
778----------------------------------------------------------------------------------------------
779
780Check if your fuzz target uses ``dlclose``.
781Currently, libFuzzer doesn't support targets that call ``dlclose``,
782this may be fixed in future.
783
784
785Trophies
786========
787* Thousands of bugs found on OSS-Fuzz:  https://opensource.googleblog.com/2017/05/oss-fuzz-five-months-later-and.html
788
789* GLIBC: https://sourceware.org/glibc/wiki/FuzzingLibc
790
791* MUSL LIBC: `[1] <http://git.musl-libc.org/cgit/musl/commit/?id=39dfd58417ef642307d90306e1c7e50aaec5a35c>`__ `[2] <http://www.openwall.com/lists/oss-security/2015/03/30/3>`__
792
793* `pugixml <https://github.com/zeux/pugixml/issues/39>`_
794
795* PCRE: Search for "LLVM fuzzer" in http://vcs.pcre.org/pcre2/code/trunk/ChangeLog?view=markup;
796  also in `bugzilla <https://bugs.exim.org/buglist.cgi?bug_status=__all__&content=libfuzzer&no_redirect=1&order=Importance&product=PCRE&query_format=specific>`_
797
798* `ICU <http://bugs.icu-project.org/trac/ticket/11838>`_
799
800* `Freetype <https://savannah.nongnu.org/search/?words=LibFuzzer&type_of_search=bugs&Search=Search&exact=1#options>`_
801
802* `Harfbuzz <https://github.com/behdad/harfbuzz/issues/139>`_
803
804* `SQLite <http://www3.sqlite.org/cgi/src/info/088009efdd56160b>`_
805
806* `Python <http://bugs.python.org/issue25388>`_
807
808* OpenSSL/BoringSSL: `[1] <https://boringssl.googlesource.com/boringssl/+/cb852981cd61733a7a1ae4fd8755b7ff950e857d>`_ `[2] <https://openssl.org/news/secadv/20160301.txt>`_ `[3] <https://boringssl.googlesource.com/boringssl/+/2b07fa4b22198ac02e0cee8f37f3337c3dba91bc>`_ `[4] <https://boringssl.googlesource.com/boringssl/+/6b6e0b20893e2be0e68af605a60ffa2cbb0ffa64>`_  `[5] <https://github.com/openssl/openssl/pull/931/commits/dd5ac557f052cc2b7f718ac44a8cb7ac6f77dca8>`_ `[6] <https://github.com/openssl/openssl/pull/931/commits/19b5b9194071d1d84e38ac9a952e715afbc85a81>`_
809
810* `Libxml2
811  <https://bugzilla.gnome.org/buglist.cgi?bug_status=__all__&content=libFuzzer&list_id=68957&order=Importance&product=libxml2&query_format=specific>`_ and `[HT206167] <https://support.apple.com/en-gb/HT206167>`_ (CVE-2015-5312, CVE-2015-7500, CVE-2015-7942)
812
813* `Linux Kernel's BPF verifier <https://github.com/iovisor/bpf-fuzzer>`_
814
815* `Linux Kernel's Crypto code <https://www.spinics.net/lists/stable/msg199712.html>`_
816
817* Capstone: `[1] <https://github.com/aquynh/capstone/issues/600>`__ `[2] <https://github.com/aquynh/capstone/commit/6b88d1d51eadf7175a8f8a11b690684443b11359>`__
818
819* file:`[1] <http://bugs.gw.com/view.php?id=550>`__  `[2] <http://bugs.gw.com/view.php?id=551>`__  `[3] <http://bugs.gw.com/view.php?id=553>`__  `[4] <http://bugs.gw.com/view.php?id=554>`__
820
821* Radare2: `[1] <https://github.com/revskills?tab=contributions&from=2016-04-09>`__
822
823* gRPC: `[1] <https://github.com/grpc/grpc/pull/6071/commits/df04c1f7f6aec6e95722ec0b023a6b29b6ea871c>`__ `[2] <https://github.com/grpc/grpc/pull/6071/commits/22a3dfd95468daa0db7245a4e8e6679a52847579>`__ `[3] <https://github.com/grpc/grpc/pull/6071/commits/9cac2a12d9e181d130841092e9d40fa3309d7aa7>`__ `[4] <https://github.com/grpc/grpc/pull/6012/commits/82a91c91d01ce9b999c8821ed13515883468e203>`__ `[5] <https://github.com/grpc/grpc/pull/6202/commits/2e3e0039b30edaf89fb93bfb2c1d0909098519fa>`__ `[6] <https://github.com/grpc/grpc/pull/6106/files>`__
824
825* WOFF2: `[1] <https://github.com/google/woff2/commit/a15a8ab>`__
826
827* LLVM: `Clang <https://bugs.llvm.org/show_bug.cgi?id=23057>`_, `Clang-format <https://bugs.llvm.org/show_bug.cgi?id=23052>`_, `libc++ <https://bugs.llvm.org/show_bug.cgi?id=24411>`_, `llvm-as <https://bugs.llvm.org/show_bug.cgi?id=24639>`_, `Demangler <https://bugs.chromium.org/p/chromium/issues/detail?id=606626>`_, Disassembler: http://reviews.llvm.org/rL247405, http://reviews.llvm.org/rL247414, http://reviews.llvm.org/rL247416, http://reviews.llvm.org/rL247417, http://reviews.llvm.org/rL247420, http://reviews.llvm.org/rL247422.
828
829* Tensorflow: `[1] <https://da-data.blogspot.com/2017/01/finding-bugs-in-tensorflow-with.html>`__
830
831* Ffmpeg: `[1] <https://github.com/FFmpeg/FFmpeg/commit/c92f55847a3d9cd12db60bfcd0831ff7f089c37c>`__  `[2] <https://github.com/FFmpeg/FFmpeg/commit/25ab1a65f3acb5ec67b53fb7a2463a7368f1ad16>`__  `[3] <https://github.com/FFmpeg/FFmpeg/commit/85d23e5cbc9ad6835eef870a5b4247de78febe56>`__ `[4] <https://github.com/FFmpeg/FFmpeg/commit/04bd1b38ee6b8df410d0ab8d4949546b6c4af26a>`__
832
833* `Wireshark <https://bugs.wireshark.org/bugzilla/buglist.cgi?bug_status=UNCONFIRMED&bug_status=CONFIRMED&bug_status=IN_PROGRESS&bug_status=INCOMPLETE&bug_status=RESOLVED&bug_status=VERIFIED&f0=OP&f1=OP&f2=product&f3=component&f4=alias&f5=short_desc&f7=content&f8=CP&f9=CP&j1=OR&o2=substring&o3=substring&o4=substring&o5=substring&o6=substring&o7=matches&order=bug_id%20DESC&query_format=advanced&v2=libfuzzer&v3=libfuzzer&v4=libfuzzer&v5=libfuzzer&v6=libfuzzer&v7=%22libfuzzer%22>`_
834
835* `QEMU <https://researchcenter.paloaltonetworks.com/2017/09/unit42-palo-alto-networks-discovers-new-qemu-vulnerability/>`_
836
837.. _pcre2: http://www.pcre.org/
838.. _AFL: http://lcamtuf.coredump.cx/afl/
839.. _Radamsa: https://github.com/aoh/radamsa
840.. _SanitizerCoverage: https://clang.llvm.org/docs/SanitizerCoverage.html
841.. _SanitizerCoverageTraceDataFlow: https://clang.llvm.org/docs/SanitizerCoverage.html#tracing-data-flow
842.. _AddressSanitizer: https://clang.llvm.org/docs/AddressSanitizer.html
843.. _LeakSanitizer: https://clang.llvm.org/docs/LeakSanitizer.html
844.. _Heartbleed: http://en.wikipedia.org/wiki/Heartbleed
845.. _FuzzerInterface.h: https://github.com/llvm/llvm-project/blob/main/compiler-rt/lib/fuzzer/FuzzerInterface.h
846.. _3.7.0: https://llvm.org/releases/3.7.0/docs/LibFuzzer.html
847.. _building Clang from trunk: https://clang.llvm.org/get_started.html
848.. _MemorySanitizer: https://clang.llvm.org/docs/MemorySanitizer.html
849.. _UndefinedBehaviorSanitizer: https://clang.llvm.org/docs/UndefinedBehaviorSanitizer.html
850.. _`coverage counters`: https://clang.llvm.org/docs/SanitizerCoverage.html#coverage-counters
851.. _`value profile`: #value-profile
852.. _`caller-callee pairs`: https://clang.llvm.org/docs/SanitizerCoverage.html#caller-callee-coverage
853.. _BoringSSL: https://boringssl.googlesource.com/boringssl/
854.. _Centipede: https://github.com/google/centipede
855
856