diff options
Diffstat (limited to 'doc/developer/fuzzing.rst')
-rw-r--r-- | doc/developer/fuzzing.rst | 164 |
1 files changed, 164 insertions, 0 deletions
diff --git a/doc/developer/fuzzing.rst b/doc/developer/fuzzing.rst new file mode 100644 index 0000000..8a33187 --- /dev/null +++ b/doc/developer/fuzzing.rst @@ -0,0 +1,164 @@ +.. _fuzzing: + +Fuzzing +======= + +This page describes the fuzzing targets and supported fuzzers available in FRR +and how to use them. Familiarity with fuzzing techniques and tools is assumed. + +Overview +-------- + +It is well known that networked applications tend to be difficult to fuzz on +their network-facing attack surfaces. Approaches involving actual network +transmission tend to be slow and are subject to intermediate devices and +networking stacks which tend to drop fuzzed packets, especially if the fuzzing +surface covers IP itself. Some time was spent on fuzzing FRR this way with some +mediocre results but attention quickly turned towards skipping the actual +networking and instead adding fuzzing targets directly in the packet processing +code for use by more traditional in- and out-of-process fuzzers. Results from +this approach have been very fruitful. + +The patches to add fuzzing targets are kept in a separate git branch. Typically +it is better to keep them in the main branch so they are kept up to date and do +not need to be constantly synchronized with the main codebase. Unfortunately, +changes to FRR to support fuzzing necessarily extend far beyond the +entrypoints. Checksums must be disarmed, interactions with the kernel must be +skipped, sockets and files must be avoided, desired under/overflows must be +marked, etc. There are the usual ``LD_PRELOAD`` libraries to emulate these +things (preeny et al) but FRR is a very kernel-reliant program and these +libraries tend to create annoying problems when used with FRR for whatever +reason. Keeping this code in the main codebase is cluttering, difficult to work +with / around, and runs the risk of accidentally introducing bugs even if +``#ifdef``'d out. Consequently it's in a separate branch that is rebased on +``master`` from time to time. + + +Code +---- + +The git branch with fuzzing targets is located here: + +https://github.com/FRRouting/frr/tree/fuzz + +To build libFuzzer targets, pass ``--enable-libfuzzer`` to ``configure``. +To build AFL targets, compile with ``afl-clang`` as usual. + +Fuzzing with sanitizers is strongly recommended, especially ASAN, which you can +enable by passing ``--enable-address-sanitizer`` to ``configure``. + +Suggested UBSAN flags: ``-fsanitize-recover=unsigned-integer-overflow,implicit-conversion -fsanitize=unsigned-integer-overflow,implicit-conversion,nullability-arg,nullability-assign,nullability-return`` +Recommended cflags: ``-Wno-all -g3 -O3 -funroll-loops`` + +Design +------ + +All fuzzing targets have support for libFuzzer and AFL. This is done by writing +the target as a libFuzzer entrypoint (``LLVMFuzzerTestOneInput()``) and calling +it from the AFL entrypoint in ``main()``. New targets should use this rule. + +When adding AFL entrypoints, it's a good idea to use AFL persistent mode for +better performance. Grep ``bgpd/bgp_main.c`` for ``__AFL_INIT()`` for an +example of how to do this in FRR. Typically it involves moving all internal +daemon setup into a setup function. Then this setup function is called exactly +once for the lifetime of the process. In ``LLVMFuzzerTestOneInput()`` this +means you need to call it at the start of the function protected by a static +boolean that is set to true, since that function is your entrypoint. You also +need to call it prior to ``__AFL_INIT()`` in ``main()`` because ``main()`` is +your entrypoint in the AFL case. + +Adding support to daemons +^^^^^^^^^^^^^^^^^^^^^^^^^ + +This section describes how to add entrypoints to daemons that do not have any +yet. + +Because libFuzzer has its own ``main()`` function, when adding fuzzing support +to a daemon that doesn't have any targets already, ``main()`` needs to be +``#ifdef``'d out like so: + +.. code:: c + + #ifndef FUZZING_LIBFUZZER + + int main(int argc, char **argv) + { + ... + } + + #endif /* FUZZING_LIBFUZZER */ + + +The ``FUZZING_LIBFUZZER`` macro is set by ``--enable-libfuzzer``. + +Because libFuzzer can only be linked into daemons that have +``LLVMFuzzerTestOneInput()`` implemented, we can't pass ``-fsanitize=fuzzer`` +to all daemons in ``AM_CFLAGS``. It needs to go into a variable specific to +each daemon. Since it can be thought of as a kind of sanitizer, for daemons +that have libFuzzer support there are now individual flags variables for those +daemons named ``DAEMON_SAN_FLAGS`` (e.g. ``BGPD_SAN_FLAGS``, +``ZEBRA_SAN_FLAGS``). This variable has the contents of the generic +``SAN_FLAGS`` plus any fuzzing-related flags. It is used in daemons' +``subdir.am`` in place of ``SAN_FLAGS``. Daemons that don't support libFuzzer +still use ``SAN_FLAGS``. If you want to add fuzzing support to a daemon you +need to do this flag variable conversion; look at ``configure.ac`` for +examples, it is fairly straightforward. Remember to update ``subdir.am`` to use +the new variable. + +Do note that when fuzzing is enabled, ``SAN_FLAGS`` gains +``-fsanitize=fuzzer-no-link``; the result is that all daemons are instrumented +for fuzzing but only the ones with ``LLVMFuzzerTestOneInput()`` actually get +linked with libFuzzer. + + +Targets +------- + +A given daemon can have lots of different paths that are interesting to fuzz. +There's not really a great way to handle this, most fuzzers assume the program +has one entrypoint. The approach taken in FRR for multiple entrypoints is to +control which path is taken within ``LLVMFuzzerTestOneInput()`` using +``#ifdef`` and passing whatever controlling macro definition you want. Take a +look at that function for the daemon you're interested in fuzzing, pick the +target, add ``#define MY_TARGET 1`` somewhere before the ``#ifdef`` switch, +recompile. + +.. list-table:: Fuzzing Targets + + * - Daemon + - Target + - Fuzzers + * - bgpd + - packet parser + - libfuzzer, afl + * - ospfd + - packet parser + - libfuzzer, afl + * - pimd + - packet parser + - libfuzzer, afl + * - vrrpd + - packet parser + - libfuzzer, afl + * - vrrpd + - zapi parser + - libfuzzer, afl + * - zebra + - netlink + - libfuzzer, afl + * - zebra + - zserv / zapi + - libfuzzer, afl + + +Fuzzer Notes +------------ + +Some interesting seed corpuses for various daemons are available `here +<https://github.com/qlyoung/frr-fuzz/tree/master/samples>`_. + +For libFuzzer, you need to pass ``-rss_limit_mb=0`` if you are fuzzing with +ASAN enabled, as you should. + +For AFL, afl++ is strongly recommended; afl proper isn't really maintained +anymore. |