* Integrate tensorflow with OSSfuzz
* Don't install already installed dependencies
* Add environment variable to disable logging.
Some fuzzers are outputting too much information due to logging and
that's not ok when fuzzing. Use TF_CPP_MIN_LOG_LEVEL envvar to disable
printing of the log messages.
* Remove the environment variable as it is not used at runtime and we have a workaround in the pipeline
* Workaround bazel bug (bazelbuild/bazel#6697)
* Don't compile with -O2, instead remove line causing bazel bug
* Downgrade Bazel to use the last version that TF supports
* Speed-up building by building fuzzers in parallel
* Add corpora and dictionaries
The cpython patch is derived from the work done in
https://bugs.python.org/issue35214 by Greg Smith. It's required to
unblock msan builds on Python 2.7. Once upstream cpython 3 is fixed to
support msan builds, we should probably move to fuzzing under Python 3
instead of Python 2.
I'll wait to enable msan builds until the associated fixes in hg core
land, since I'm not sure when that will happen.
* Add support for tags field in project.yaml (#1974)
Also add "sundew" label to libxml2_xml_reader_for_file_fuzzer.
These add additional labels in filed bugs.
Whether string_string_checks will be turned on or not is
being discussed in https://github.com/google/oss-fuzz/issues/542.
Until it's settled, it'd probably be better to remove the option
from ASAN_OPTIONS altogether so as not to confuse anybody.
Reduce VPX_MAX_ALLOCABLE_MEMORY in msan builds as it has a higher overhead causing more frequent OOMs in this configuration. This matches the limits in libaom.
Remove the build directory to ensure dot files are deleted.
Skip 'make clean' because the build directory was just cleaned. This removes
a lot of log spam.
* libvpx: Add project
Added projects/libvpx
* libvpx: Use local ivf_read_frame
ivf_read_frame inside libvpx results in lot of prints due to invalid
frame size. This clutters output prints.
* libvpx: Do not return error for incomplete frame read in read_frame
* libvpx: Initialize threads using 32nd byte instead of 1st
In most cases, the first byte in corpus of *.ivf files is 'D'.
So using first byte results in same thread count in most cases.
Using 32nd byte in the data (one of the bytes that signals size of
the frame) will help in testing for different thread configurations.
* libvpx: Updated configure options and removed redundant cflags
Removed generic-gnu as target
Added --disable-webm-io as that is not needed here
Added --enable-debug to enable asserts
Removed redundant cflags and cxxflags
* libvpx: Removed threaded mode from build.sh
vpx_dec_fuzzer.cc in libvpx now tests both single and multi-thread
configurations using a single binary.
* libvpx: Removed vpx_dec_fuzzer.cc and README.md
Removed vpx_dec_fuzzer.cc and README.md from projects/libvpx
vpx_dec_fuzzer.cc is now part of libvpx
Add a check for the recently added routines to combine crc32 checksums. This
change adds more code coverage, however the checksum_fuzzer still does not pass
the complexity check: (the threshold is at 100)
oss-fuzz$ build/out/zlib/checksum_fuzzer -runs=10 2>&1 | grep -Po "INFO: Loaded [[:digit:]]+ module.*\(.*(counters|guards)\):[[:space:]]+\K[[:digit:]]+"
92
Max Moroz