def minimize_corpus(self, target_path, arguments, input_dirs, output_dir,
reproducers_dir, max_time):
"""Optional (but recommended): run corpus minimization.
Args:
target_path: Path to the target.
arguments: Additional arguments needed for corpus minimization.
input_dirs: Input corpora.
output_dir: Output directory to place minimized corpus.
reproducers_dir: The directory to put reproducers in when crashes are
found.
max_time: Maximum allowed time for the minimization.
Returns:
A Result object.
"""
runner = libfuzzer.get_runner(target_path)
launcher.set_sanitizer_options(target_path)
merge_tmp_dir = self._create_temp_corpus_dir('merge-workdir')
merge_result = runner.merge([output_dir] + input_dirs,
merge_timeout=max_time,
tmp_dir=merge_tmp_dir,
additional_args=arguments,
artifact_prefix=reproducers_dir)
if merge_result.timed_out:
raise MergeError('Merging new testcases timed out')
if merge_result.return_code != 0:
raise MergeError('Merging new testcases failed')
# TODO(ochang): Get crashes found during merge.
return engine.FuzzResult(merge_result.output, merge_result.command, [],
{}, merge_result.time_executed)
def fuzz(self,
fuzz_timeout,
additional_args,
unused_additional_args=None,
unused_extra_env=None):
"""This is where actual syzkaller fuzzing is done."""
additional_args = copy.copy(additional_args)
fuzz_result = self.run_and_wait(additional_args, timeout=fuzz_timeout)
log_lines = utils.decode_to_unicode(fuzz_result.output).splitlines()
fuzz_result.output = None
crash_testcase_file_path = self.get_testcase_path(log_lines)
#TODO(hzawawy): remove once syzkaller code is completed.
if not crash_testcase_file_path and fuzz_result.return_code:
crash_testcase_file_path = self._create_empty_testcase_file()
fuzz_logs = '\n'.join(log_lines)
# TODO(hzawawy): Parse stats information and add them to FuzzResult
parsed_stats = []
crashes = []
if crash_testcase_file_path:
#TODO(hzawawy): add repro arguments
reproduce_arguments = []
actual_duration = int(fuzz_result.time_executed)
# Write the new testcase.
# Copy crash testcase contents into the main testcase path.
crashes.append(
engine.Crash(crash_testcase_file_path, fuzz_logs,
reproduce_arguments, actual_duration))
return engine.FuzzResult(fuzz_logs, fuzz_result.command, crashes,
parsed_stats, fuzz_result.time_executed)
def engine_fuzz(engine_impl, target_name, sync_corpus_directory,
testcase_directory):
"""Run engine fuzzer on untrusted worker."""
request = untrusted_runner_pb2.EngineFuzzRequest(
engine=engine_impl.name,
target_name=target_name,
sync_corpus_directory=file_host.rebase_to_worker_root(
sync_corpus_directory),
testcase_directory=file_host.rebase_to_worker_root(testcase_directory))
response = host.stub().EngineFuzz(request)
crashes = [
engine.Crash(
input_path=file_host.rebase_to_host_root(crash.input_path),
stacktrace=crash.stacktrace,
reproduce_args=crash.reproduce_args,
crash_time=crash.crash_time) for crash in response.crashes
]
unpacked_stats = _unpack_values(response.stats)
unpacked_strategies = _unpack_values(response.strategies)
result = engine.FuzzResult(
logs=response.logs,
command=list(response.command),
crashes=crashes,
stats=unpacked_stats,
time_executed=response.time_executed)
file_host.pull_testcases_from_worker()
return result, dict(response.fuzzer_metadata), unpacked_strategies
def fuzz(self, target_path, options, reproducers_dir, max_time):
"""Run a fuzz session.
Args:
target_path: Path to the target.
options: The FuzzOptions object returned by prepare().
reproducers_dir: The directory to put reproducers in when crashes
are found.
max_time: Maximum allowed time for the fuzzing to run.
Returns:
A FuzzResult object.
"""
runner = _get_runner()
arguments = _DEFAULT_ARGUMENTS[:]
arguments.extend(options.arguments)
arguments.extend([
"--input",
options.corpus_dir,
"--workspace",
reproducers_dir,
"--run_time",
str(max_time),
"--",
target_path,
])
fuzz_result = runner.run_and_wait(
additional_args=arguments, timeout=max_time + _CLEAN_EXIT_SECS)
log_lines = fuzz_result.output.splitlines()
sanitizer_stacktrace = _find_sanitizer_stacktrace(reproducers_dir)
crashes = []
stats = None
for line in log_lines:
reproducer_path = _get_reproducer_path(line)
if reproducer_path:
crashes.append(
engine.Crash(
reproducer_path,
sanitizer_stacktrace,
[],
int(fuzz_result.time_executed),
))
continue
stats = _get_stats(line)
if stats is None:
stats = {}
return engine.FuzzResult(
fuzz_result.output,
fuzz_result.command,
crashes,
stats,
fuzz_result.time_executed,
)
def fuzz(self, target_path, options, reproducers_dir, max_time):
"""Run a fuzzing session.
Args:
target_path: Path to the fuzzer script or binary.
options: The FuzzOptions object returned by prepare().
reproducers_dir: The directory to put reproducers in when crashes
are found.
max_time: Maximum allowed time for the fuzzing to run.
Returns:
A FuzzResult object.
"""
# For blackbox fuzzers, |target_path| supplies the path to the fuzzer script
# rather than a target in the build archive.
fuzzer_path = target_path
os.chmod(fuzzer_path, 0o775)
app_path = environment.get_value('APP_PATH')
app_args = testcase_manager.get_command_line_for_application(
get_arguments_only=True).strip()
corpus_dir = options.corpus_dir
command_line_args = _get_arguments(app_path, app_args)
command_line_args.append(f'--input_dir={corpus_dir}')
result = _run_with_interpreter_if_needed(fuzzer_path,
command_line_args, max_time)
crashes = []
for testcase_path in os.listdir(reproducers_dir):
if not testcase_path.startswith(TESTCASE_PREFIX):
continue
output_path = OUTPUT_PREFIX + testcase_path[len(TESTCASE_PREFIX):]
absolute_output_path = os.path.join(reproducers_dir, output_path)
# If no output was written for a test case, skip it.
if not os.path.exists(absolute_output_path):
continue
with open(absolute_output_path, 'r', errors='replace') as handle:
output = handle.read()
# Filter obviously non-crashing test cases. Crashes still follow the
# normal flow in fuzz task to ensure that the state should not be ignored
# for other reasons, but we don't want to log every test case for the
# fuzzers that don't do their own crash processing.
state = stack_analyzer.get_crash_data(output)
if not state.crash_type:
continue
full_testcase_path = os.path.join(reproducers_dir, testcase_path)
crash = engine.Crash(full_testcase_path, output, options.arguments,
int(result.time_executed))
crashes.append(crash)
# TODO(mbarbella): Support stats.
stats = {}
return engine.FuzzResult(result.output, result.command, crashes, stats,
result.time_executed)
def fuzz(self,
fuzz_timeout,
additional_args,
unused_additional_args=None,
unused_extra_env=None):
"""This is where actual syzkaller fuzzing is done.
Args:
fuzz_timeout: The maximum time in seconds that fuzz job is allowed
to run for.
additional_args: A sequence of additional arguments to be passed to
the executable.
"""
logs.log('Running Syzkaller.')
additional_args = copy.copy(additional_args)
fuzz_result = self.run_and_wait(additional_args, timeout=fuzz_timeout)
logs.log('Syzkaller stopped, fuzzing timed out: {}'.format(
fuzz_result.time_executed))
fuzz_logs = ''
crashes = []
parsed_stats = {}
visited = set()
for subdir, _, files in os.walk(get_work_dir()):
for file in files:
# Each crash typically have 2 files: reportN and logN. Similar crashes
# are grouped together in subfolders. unique_crash puts together the
# subfolder name and reportN.
unique_crash = os.path.join(subdir, file)
if fnmatch.fnmatch(file,
'report*') and unique_crash not in visited:
visited.add(unique_crash)
log_lines = utils.read_data_from_file(
os.path.join(subdir,
file), eval_data=False).decode('utf-8')
fuzz_logs += log_lines + '\n'
# Since each crash (report file) has a corresponding log file
# that contains the syscalls that caused the crash. This file is
# located in the same subfolder and has the same number.
# E.g. ./439c37d288d4f26a33a6c7e5c57a97791453a447/report15 and
# ./439c37d288d4f26a33a6c7e5c57a97791453a447/log15.
crash_testcase_file_path = os.path.join(
subdir, 'log' + file[len('report'):])
# TODO(hzawawy): Parse stats information and add them to FuzzResult.
if crash_testcase_file_path:
reproduce_arguments = [unique_crash]
actual_duration = int(fuzz_result.time_executed)
# Write the new testcase.
# Copy crash testcase contents into the main testcase path.
crashes.append(
engine.Crash(crash_testcase_file_path, log_lines,
reproduce_arguments, actual_duration))
return engine.FuzzResult(fuzz_logs, fuzz_result.command, crashes,
parsed_stats, fuzz_result.time_executed)
def minimize_corpus(self, target_path, arguments, input_dirs, output_dir,
reproducers_dir, max_time):
"""Optional (but recommended): run corpus minimization.
Args:
target_path: Path to the target.
arguments: Additional arguments needed for corpus minimization.
input_dirs: Input corpora.
output_dir: Output directory to place minimized corpus.
reproducers_dir: The directory to put reproducers in when crashes are
found.
max_time: Maximum allowed time for the minimization.
Returns:
A Result object.
Raises:
TimeoutError: If the corpus minimization exceeds max_time.
Error: If the merge failed in some other way.
"""
runner = libfuzzer.get_runner(target_path)
libfuzzer.set_sanitizer_options(target_path)
merge_tmp_dir = self._create_temp_corpus_dir('merge-workdir')
result = runner.merge([output_dir] + input_dirs,
merge_timeout=max_time,
tmp_dir=merge_tmp_dir,
additional_args=arguments,
artifact_prefix=reproducers_dir,
merge_control_file=getattr(
self, '_merge_control_file', None))
if result.timed_out:
raise TimeoutError('Merging new testcases timed out\n' +
result.output)
if result.return_code != 0:
raise MergeError('Merging new testcases failed: ' + result.output)
merge_output = result.output
merge_stats = stats.parse_stats_from_merge_log(
merge_output.splitlines())
# TODO(ochang): Get crashes found during merge.
return engine.FuzzResult(merge_output, result.command, [], merge_stats,
result.time_executed)
def engine_fuzz(engine_impl, target_name, sync_corpus_directory,
testcase_directory):
"""Run engine fuzzer on untrusted worker."""
request = untrusted_runner_pb2.EngineFuzzRequest(
engine=engine_impl.name,
target_name=target_name,
sync_corpus_directory=file_host.rebase_to_worker_root(
sync_corpus_directory),
testcase_directory=file_host.rebase_to_worker_root(testcase_directory))
response = host.stub().EngineFuzz(request)
crashes = [
engine.Crash(
input_path=file_host.rebase_to_host_root(crash.input_path),
stacktrace=utils.decode_to_unicode(crash.stacktrace),
reproduce_args=crash.reproduce_args,
crash_time=crash.crash_time) for crash in response.crashes
]
unpacked_stats = {}
for key, packed_value in six.iteritems(response.stats):
if packed_value.Is(wrappers_pb2.DoubleValue.DESCRIPTOR):
value = wrappers_pb2.DoubleValue()
elif packed_value.Is(wrappers_pb2.Int32Value.DESCRIPTOR):
value = wrappers_pb2.Int32Value()
elif packed_value.Is(wrappers_pb2.StringValue.DESCRIPTOR):
value = wrappers_pb2.StringValue()
else:
raise ValueError('Unknown stat type for ' + key)
packed_value.Unpack(value)
unpacked_stats[key] = value.value
result = engine.FuzzResult(
logs=utils.decode_to_unicode(response.logs),
command=list(response.command),
crashes=crashes,
stats=unpacked_stats,
time_executed=response.time_executed)
file_host.pull_testcases_from_worker()
return result, dict(response.fuzzer_metadata)
def fuzz(self, target_path, options, reproducers_dir, max_time):
"""Run a fuzz session.
Args:
target_path: Path to the target.
options: The FuzzOptions object returned by prepare().
reproducers_dir: The directory to put reproducers in when crashes
are found.
max_time: Maximum allowed time for the fuzzing to run.
Returns:
A FuzzResult object.
"""
profiler.start_if_needed('libfuzzer_fuzz')
runner = libfuzzer.get_runner(target_path)
launcher.set_sanitizer_options(target_path)
# Directory to place new units.
new_corpus_dir = self._create_temp_corpus_dir('new')
corpus_directories = [new_corpus_dir] + options.fuzz_corpus_dirs
fuzz_timeout = launcher.get_fuzz_timeout(
options.is_mutations_run, total_timeout=max_time)
fuzz_result = runner.fuzz(
corpus_directories,
fuzz_timeout=fuzz_timeout,
additional_args=options.arguments,
artifact_prefix=reproducers_dir,
extra_env=options.extra_env)
log_lines = utils.decode_to_unicode(fuzz_result.output).splitlines()
# Output can be large, so save some memory by removing reference to the
# original output which is no longer needed.
fuzz_result.output = None
# Check if we crashed, and get the crash testcase path.
crash_testcase_file_path = runner.get_testcase_path(log_lines)
# Parse stats information based on libFuzzer output.
parsed_stats = launcher.parse_log_stats(log_lines)
# Extend parsed stats by additional performance features.
parsed_stats.update(
stats.parse_performance_features(
log_lines,
options.strategies,
options.arguments,
include_strategies=False))
# Set some initial stat overrides.
timeout_limit = fuzzer_utils.extract_argument(
options.arguments, constants.TIMEOUT_FLAG, remove=False)
expected_duration = runner.get_max_total_time(fuzz_timeout)
actual_duration = int(fuzz_result.time_executed)
fuzzing_time_percent = 100 * actual_duration / float(expected_duration)
parsed_stats.update({
'timeout_limit': int(timeout_limit),
'expected_duration': expected_duration,
'actual_duration': actual_duration,
'fuzzing_time_percent': fuzzing_time_percent,
})
# Remove fuzzing arguments before merge and dictionary analysis step.
arguments = options.arguments[:]
launcher.remove_fuzzing_arguments(arguments)
self._merge_new_units(target_path, options.corpus_dir, new_corpus_dir,
options.fuzz_corpus_dirs, arguments, parsed_stats)
fuzz_logs = '\n'.join(log_lines)
crashes = []
if crash_testcase_file_path:
# Write the new testcase.
# Copy crash testcase contents into the main testcase path.
crashes.append(
engine.Crash(crash_testcase_file_path, fuzz_logs, arguments,
actual_duration))
project_qualified_fuzzer_name = (
data_types.fuzz_target_project_qualified_name(
utils.current_project(), os.path.basename(target_path)))
launcher.analyze_and_update_recommended_dictionary(
runner, project_qualified_fuzzer_name, log_lines, options.corpus_dir,
arguments)
return engine.FuzzResult(fuzz_logs, fuzz_result.command, crashes,
parsed_stats, fuzz_result.time_executed)
def _minimize_corpus_two_step(self, target_path, arguments,
existing_corpus_dirs, new_corpus_dir,
output_corpus_dir, reproducers_dir,
max_time):
"""Optional (but recommended): run corpus minimization.
Args:
target_path: Path to the target.
arguments: Additional arguments needed for corpus minimization.
existing_corpus_dirs: Input corpora that existed before the fuzzing run.
new_corpus_dir: Input corpus that was generated during the fuzzing run.
Must have at least one new file.
output_corpus_dir: Output directory to place minimized corpus.
reproducers_dir: The directory to put reproducers in when crashes are
found.
max_time: Maximum allowed time for the minimization.
Returns:
A Result object.
"""
if not _is_multistep_merge_supported(target_path):
# Fallback to the old single step merge. It does not support incremental
# stats and provides only `edge_coverage` and `feature_coverage` stats.
logs.log(
'Old version of libFuzzer is used. Using single step merge.')
return self.minimize_corpus(
target_path, arguments,
existing_corpus_dirs + [new_corpus_dir], output_corpus_dir,
reproducers_dir, max_time)
# The dir where merge control file is located must persist for both merge
# steps. The second step re-uses the MCF produced during the first step.
merge_control_file_dir = self._create_temp_corpus_dir('mcf_tmp_dir')
self._merge_control_file = os.path.join(merge_control_file_dir, 'MCF')
# Two step merge process to obtain accurate stats for the new corpus units.
# See https://reviews.llvm.org/D66107 for a more detailed description.
merge_stats = {}
# Step 1. Use only existing corpus and collect "initial" stats.
result_1 = self.minimize_corpus(target_path, arguments,
existing_corpus_dirs,
output_corpus_dir, reproducers_dir,
max_time)
merge_stats['initial_edge_coverage'] = result_1.stats['edge_coverage']
merge_stats['initial_feature_coverage'] = result_1.stats[
'feature_coverage']
# Clear the output dir as it does not have any new units at this point.
engine_common.recreate_directory(output_corpus_dir)
# Adjust the time limit for the time we spent on the first merge step.
max_time -= result_1.time_executed
if max_time <= 0:
raise TimeoutError('Merging new testcases timed out\n' +
result_1.logs)
# Step 2. Process the new corpus units as well.
result_2 = self.minimize_corpus(
target_path, arguments, existing_corpus_dirs + [new_corpus_dir],
output_corpus_dir, reproducers_dir, max_time)
merge_stats['edge_coverage'] = result_2.stats['edge_coverage']
merge_stats['feature_coverage'] = result_2.stats['feature_coverage']
# Diff the stats to obtain accurate values for the new corpus units.
merge_stats['new_edges'] = (merge_stats['edge_coverage'] -
merge_stats['initial_edge_coverage'])
merge_stats['new_features'] = (merge_stats['feature_coverage'] -
merge_stats['initial_feature_coverage'])
output = result_1.logs + '\n\n' + result_2.logs
if (merge_stats['new_edges'] < 0 or merge_stats['new_features'] < 0):
logs.log_error('Two step merge failed.',
merge_stats=merge_stats,
output=output)
merge_stats['new_edges'] = 0
merge_stats['new_features'] = 0
self._merge_control_file = None
# TODO(ochang): Get crashes found during merge.
return engine.FuzzResult(
output, result_2.command, [], merge_stats,
result_1.time_executed + result_2.time_executed)
def fuzz(self, target_path, options, reproducers_dir, max_time):
"""Run a fuzz session.
Args:
target_path: Path to the target.
options: The FuzzOptions object returned by prepare().
reproducers_dir: The directory to put reproducers in when crashes
are found.
max_time: Maximum allowed time for the fuzzing to run.
Returns:
A FuzzResult object.
"""
profiler.start_if_needed('libfuzzer_fuzz')
runner = libfuzzer.get_runner(target_path)
libfuzzer.set_sanitizer_options(target_path, fuzz_options=options)
# Directory to place new units.
new_corpus_dir = self._create_temp_corpus_dir('new')
corpus_directories = [new_corpus_dir] + options.fuzz_corpus_dirs
fuzz_timeout = libfuzzer.get_fuzz_timeout(options.is_mutations_run,
total_timeout=max_time)
fuzz_result = runner.fuzz(corpus_directories,
fuzz_timeout=fuzz_timeout,
additional_args=options.arguments,
artifact_prefix=reproducers_dir,
extra_env=options.extra_env)
project_qualified_fuzzer_name = _project_qualified_fuzzer_name(
target_path)
dict_error_match = DICT_PARSING_FAILED_REGEX.search(fuzz_result.output)
if dict_error_match:
logs.log_error(
'Dictionary parsing failed (target={target}, line={line}).'.
format(target=project_qualified_fuzzer_name,
line=dict_error_match.group(1)),
engine_output=fuzz_result.output)
elif (not environment.get_value('USE_MINIJAIL') and
fuzz_result.return_code == constants.LIBFUZZER_ERROR_EXITCODE):
# Minijail returns 1 if the exit code is nonzero.
# Otherwise: we can assume that a return code of 1 means that libFuzzer
# itself ran into an error.
logs.log_error(ENGINE_ERROR_MESSAGE + ' (target={target}).'.format(
target=project_qualified_fuzzer_name),
engine_output=fuzz_result.output)
log_lines = fuzz_result.output.splitlines()
# Output can be large, so save some memory by removing reference to the
# original output which is no longer needed.
fuzz_result.output = None
# Check if we crashed, and get the crash testcase path.
crash_testcase_file_path = runner.get_testcase_path(log_lines)
# If we exited with a non-zero return code with no crash file in output from
# libFuzzer, this is most likely a startup crash. Use an empty testcase to
# to store it as a crash.
if not crash_testcase_file_path and fuzz_result.return_code:
crash_testcase_file_path = self._create_empty_testcase_file(
reproducers_dir)
# Parse stats information based on libFuzzer output.
parsed_stats = libfuzzer.parse_log_stats(log_lines)
# Extend parsed stats by additional performance features.
parsed_stats.update(
stats.parse_performance_features(log_lines, options.strategies,
options.arguments))
# Set some initial stat overrides.
timeout_limit = fuzzer_utils.extract_argument(options.arguments,
constants.TIMEOUT_FLAG,
remove=False)
expected_duration = runner.get_max_total_time(fuzz_timeout)
actual_duration = int(fuzz_result.time_executed)
fuzzing_time_percent = 100 * actual_duration / float(expected_duration)
parsed_stats.update({
'timeout_limit': int(timeout_limit),
'expected_duration': expected_duration,
'actual_duration': actual_duration,
'fuzzing_time_percent': fuzzing_time_percent,
})
# Remove fuzzing arguments before merge and dictionary analysis step.
merge_arguments = options.arguments[:]
libfuzzer.remove_fuzzing_arguments(merge_arguments, is_merge=True)
self._merge_new_units(target_path, options.corpus_dir, new_corpus_dir,
options.fuzz_corpus_dirs, merge_arguments,
parsed_stats)
fuzz_logs = '\n'.join(log_lines)
crashes = []
if crash_testcase_file_path:
reproduce_arguments = options.arguments[:]
libfuzzer.remove_fuzzing_arguments(reproduce_arguments)
# Use higher timeout for reproduction.
libfuzzer.fix_timeout_argument_for_reproduction(
reproduce_arguments)
# Write the new testcase.
# Copy crash testcase contents into the main testcase path.
crashes.append(
engine.Crash(crash_testcase_file_path, fuzz_logs,
reproduce_arguments, actual_duration))
libfuzzer.analyze_and_update_recommended_dictionary(
runner, project_qualified_fuzzer_name, log_lines,
options.corpus_dir, merge_arguments)
return engine.FuzzResult(fuzz_logs, fuzz_result.command, crashes,
parsed_stats, fuzz_result.time_executed)
