My afl++ fuzzer setup

Use two build directories: .build-asan and .build-fuzz-asan.

• .build-fuzz-asan contains the fuzzer config, fuzzer input and outpur dirs
• .build-asan contains the normal asan config, and scripts to verify and debug the crashes collected in .build-fuzz-asan.

failures are automatically seperated into crashes, hangs and leaks, see fuzzlib.sh

fuzzer targets are injson, indxf, rw, outdxf, outjson, svg, bmp.

fuzz output directories are numbered, and mainrained in the makefile-fuzz and the crashes*.sh scripts.

To verify against regressions either run the binaries against all old crashes with make -f makefile-fuzz or run the scripts seperately, like ./crashes20.sh

Usually fuzzing is done on seperate machines, rsync the results (hangs and crashes) to your local machine regularly.

Get afl++ at https://github.com/AFLplusplus/AFLplusplus and be sure to make -C llvm_mode, the clang-devel package is needed for this.

I usually run the fuzzer jobs on the big machine with -M m1, -S s1 and -S s2 -R arguments, first the master, then a default slave, and a second slave with radare. The forth CPU is left alone. On the laptop I run the fuzzer not with the performance CPU governor.

The current setup does

export ASAN_OPTIONS="abort_on_error=1:detect_leaks=0:symbolize=0:allocator_may_return_null=1"
export AFL_PATH=/usr/local/lib/afl
export AFL_LLVM_WHITELIST=`pwd`/afl-whitelist-<mode>
export AFL_LLVM_INSTRIM=1
#export AFL_LLVM_INSTRIM_LOOPHEAD=1
export AFL_LLVM_INSTRIM_SKIPSINGLEBLOCK=1
export AFL_LLVM_LAF_SPLIT_SWITCHES=1
AFL_USE_ASAN=1 make -j4 -s -C src
AFL_USE_ASAN=1 make -j4 -s -C examples dwgfuzz

but this varies monthly. LTO needs clang-11. The PCGUARD, NGRAM, CTX and CmpLog instrumentations are also worthwhile and supported. laf_intel esp. ALL is extremely heavy in such a big library.