You can not select more than 25 topics
Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
BagPipeOuO
e231dae4cc
|
2 months ago | |
---|---|---|
.. | ||
Makefile | 2 months ago | |
README.llvm | 2 months ago | |
afl-clang-fast.c | 2 months ago | |
afl-llvm-pass.so.cc | 2 months ago | |
afl-llvm-rt.o.c | 2 months ago |
README.llvm
============================================ Fast LLVM-based instrumentation for afl-fuzz ============================================ (See ../docs/README for the general instruction manual.) 1) Introduction --------------- The code in this directory allows you to instrument programs for AFL using true compiler-level instrumentation, instead of the more crude assembly-level rewriting approach taken by afl-gcc and afl-clang. This has several interesting properties: - The compiler can make many optimizations that are hard to pull off when manually inserting assembly. As a result, some slow, CPU-bound programs will run up to around 2x faster. The gains are less pronounced for fast binaries, where the speed is limited chiefly by the cost of creating new processes. In such cases, the gain will probably stay within 10%. - The instrumentation is CPU-independent. At least in principle, you should be able to rely on it to fuzz programs on non-x86 architectures (after building afl-fuzz with AFL_NO_X86=1). - The instrumentation can cope a bit better with multi-threaded targets. - Because the feature relies on the internals of LLVM, it is clang-specific and will *not* work with GCC. Once this implementation is shown to be sufficiently robust and portable, it will probably replace afl-clang. For now, it can be built separately and co-exists with the original code. The idea and much of the implementation comes from Laszlo Szekeres. 2) How to use ------------- In order to leverage this mechanism, you need to have clang installed on your system. You should also make sure that the llvm-config tool is in your path (or pointed to via LLVM_CONFIG in the environment). Unfortunately, some systems that do have clang come without llvm-config or the LLVM development headers; one example of this is FreeBSD. FreeBSD users will also run into problems with clang being built statically and not being able to load modules (you'll see "Service unavailable" when loading afl-llvm-pass.so). To solve all your problems, you can grab pre-built binaries for your OS from: http://llvm.org/releases/download.html ...and then put the bin/ directory from the tarball at the beginning of your $PATH when compiling the feature and building packages later on. You don't need to be root for that. To build the instrumentation itself, type 'make'. This will generate binaries called afl-clang-fast and afl-clang-fast++ in the parent directory. Once this is done, you can instrument third-party code in a way similar to the standard operating mode of AFL, e.g.: CC=/path/to/afl/afl-clang-fast ./configure [...options...] make Be sure to also include CXX set to afl-clang-fast++ for C++ code. The tool honors roughly the same environmental variables as afl-gcc (see ../docs/env_variables.txt). This includes AFL_INST_RATIO, AFL_USE_ASAN, AFL_HARDEN, and AFL_DONT_OPTIMIZE. Note: if you want the LLVM helper to be installed on your system for all users, you need to build it before issuing 'make install' in the parent directory. 3) Gotchas, feedback, bugs -------------------------- This is an early-stage mechanism, so field reports are welcome. You can send bug reports to <afl-users@googlegroups.com>. 4) Bonus feature #1: deferred instrumentation --------------------------------------------- AFL tries to optimize performance by executing the targeted binary just once, stopping it just before main(), and then cloning this "master" process to get a steady supply of targets to fuzz. Although this approach eliminates much of the OS-, linker- and libc-level costs of executing the program, it does not always help with binaries that perform other time-consuming initialization steps - say, parsing a large config file before getting to the fuzzed data. In such cases, it's beneficial to initialize the forkserver a bit later, once most of the initialization work is already done, but before the binary attempts to read the fuzzed input and parse it; in some cases, this can offer a 10x+ performance gain. You can implement delayed initialization in LLVM mode in a fairly simple way. First, find a suitable location in the code where the delayed cloning can take place. This needs to be done with *extreme* care to avoid breaking the binary. In particular, the program will probably malfunction if you select a location after: - The creation of any vital threads or child processes - since the forkserver can't clone them easily. - The initialization of timers via setitimer() or equivalent calls. - The creation of temporary files, network sockets, offset-sensitive file descriptors, and similar shared-state resources - but only provided that their state meaningfully influences the behavior of the program later on. - Any access to the fuzzed input, including reading the metadata about its size. With the location selected, add this code in the appropriate spot: #ifdef __AFL_HAVE_MANUAL_CONTROL __AFL_INIT(); #endif You don't need the #ifdef guards, but including them ensures that the program will keep working normally when compiled with a tool other than afl-clang-fast. Finally, recompile the program with afl-clang-fast (afl-gcc or afl-clang will *not* generate a deferred-initialization binary) - and you should be all set! 5) Bonus feature #2: persistent mode ------------------------------------ Some libraries provide APIs that are stateless, or whose state can be reset in between processing different input files. When such a reset is performed, a single long-lived process can be reused to try out multiple test cases, eliminating the need for repeated fork() calls and the associated OS overhead. The basic structure of the program that does this would be: while (__AFL_LOOP(1000)) { /* Read input data. */ /* Call library code to be fuzzed. */ /* Reset state. */ } /* Exit normally */ The numerical value specified within the loop controls the maximum number of iterations before AFL will restart the process from scratch. This minimizes the impact of memory leaks and similar glitches; 1000 is a good starting point, and going much higher increases the likelihood of hiccups without giving you any real performance benefits. A more detailed template is shown in ../experimental/persistent_demo/. Similarly to the previous mode, the feature works only with afl-clang-fast; #ifdef guards can be used to suppress it when using other compilers. Note that as with the previous mode, the feature is easy to misuse; if you do not fully reset the critical state, you may end up with false positives or waste a whole lot of CPU power doing nothing useful at all. Be particularly wary of memory leaks and of the state of file descriptors. PS. Because there are task switches still involved, the mode isn't as fast as "pure" in-process fuzzing offered, say, by LLVM's LibFuzzer; but it is a lot faster than the normal fork() model, and compared to in-process fuzzing, should be a lot more robust. 6) Bonus feature #3: new 'trace-pc-guard' mode ---------------------------------------------- Recent versions of LLVM are shipping with a built-in execution tracing feature that provides AFL with the necessary tracing data without the need to post-process the assembly or install any compiler plugins. See: http://clang.llvm.org/docs/SanitizerCoverage.html#tracing-pcs-with-guards If you have a sufficiently recent compiler and want to give it a try, build afl-clang-fast this way: AFL_TRACE_PC=1 make clean all Note that this mode is currently about 20% slower than "vanilla" afl-clang-fast, and about 5-10% slower than afl-clang. This is likely because the instrumentation is not inlined, and instead involves a function call. On systems that support it, compiling your target with -flto should help.