7e295da81c
We're getting some very large inputs (~500KB) on OSS-Fuzz, which slot down performance a lot. Let's try limiting this, starting with a still fairly large value of 64KB. Also remove the max_execution_time limit, so that slow test cases cause a genuine libfuzzer timeout and we may investigate them.
Fuzzing SAPI for PHP
The following ./configure options can be used to enable the fuzzing SAPI, as well as all availablefuzzers. If you don't build the exif/json/mbstring extensions, fuzzers for these extensions will not be built.
./configure \
--enable-fuzzer \
--with-pic \
--enable-debug-assertions \
--enable-exif \
--enable-json \
--enable-mbstring
The --with-pic option is required to avoid a linking failure. The --enable-debug-assertions option can be used to enable debug assertions despite the use of a release build.
You will need a recent version of clang that supports the -fsanitize=fuzzer-no-link option.
When running make it creates these binaries in sapi/fuzzer/:
php-fuzz-parser: Fuzzing language parser and compilerphp-fuzz-unserialize: Fuzzing unserialize() functionphp-fuzz-json: Fuzzing JSON parser (requires --enable-json)php-fuzz-exif: Fuzzingexif_read_data()function (requires --enable-exif)php-fuzz-mbstring: fuzzingmb_ereg[i]()(requires --enable-mbstring)
Some fuzzers have a seed corpus in sapi/fuzzer/corpus. You can use it as follows:
cp -r sapi/fuzzer/corpus/exif ./my-exif-corpus
sapi/fuzzer/php-fuzz-exif ./my-exif-corpus
For the unserialize fuzzer, a dictionary of internal classes should be generated first:
sapi/cli/php sapi/fuzzer/generate_unserialize_dict.php
cp -r sapi/fuzzer/corpus/unserialize ./my-unserialize-corpus
sapi/fuzzer/php-fuzz-unserialize -dict=$PWD/sapi/fuzzer/dict/unserialize ./my-unserialize-corpus
For the parser fuzzer, a corpus may be generated from Zend test files:
sapi/cli/php sapi/fuzzer/generate_parser_corpus.php
mkdir ./my-parser-corpus
sapi/fuzzer/php-fuzz-parser -merge=1 ./my-parser-corpus sapi/fuzzer/corpus/parser
sapi/fuzzer/php-fuzz-parser -only_ascii=1 ./my-parser-corpus
For the mbstring fuzzer, you may want to build the libonig dependency with instrumentation. At this time, libonig is not clean under ubsan, so only the fuzzer and address sanitizers may be used.
git clone https://github.com/kkos/oniguruma.git
pushd oniguruma
autoreconf -vfi
./configure CC=clang CFLAGS="-fsanitize=fuzzer-no-link,address -O2 -g"
make
popd
export ONIG_CFLAGS="-I$PWD/oniguruma/src"
export ONIG_LIBS="-L$PWD/oniguruma/src/.libs -l:libonig.a"
This will link an instrumented libonig statically into the PHP binary.