def cleanse_testcase(runner, testcase_file_path, cleanse_to, cleanse_timeout,
arguments, use_minijail):
"""Cleanse testcase."""
remove_fuzzing_arguments(arguments)
# Write in-progress cleanse testcases to temp dir.
if use_minijail:
arguments.append(constants.TMP_ARTIFACT_PREFIX_ARGUMENT)
else:
cleanse_temp_dir = os.path.join(fuzzer_utils.get_temp_dir(),
'cleanse_temp')
engine_common.recreate_directory(cleanse_temp_dir)
arguments.append('%s%s/' %
(constants.ARTIFACT_PREFIX_FLAG, cleanse_temp_dir))
# Call the fuzzer to cleanse.
result = runner.cleanse_crash(testcase_file_path,
cleanse_to,
cleanse_timeout,
additional_args=arguments)
print(
'Running command:',
get_printable_command(result.command, runner.executable_path,
use_minijail))
print(result.output)
def mocked_fuzz(runner):
"""Mocked version of AflRunner.fuzz."""
fuzz_args = runner.generate_afl_args()
runner._fuzz_args = fuzz_args # pylint: disable=protected-access
engine_common.recreate_directory(runner.afl_output.output_directory)
runner._fuzzer_stderr = '' # pylint: disable=protected-access
# Create the queue directory within AFL's output directory.
queue = runner.afl_output.queue
engine_common.recreate_directory(queue)
new_corpus_dir = os.path.join(DATA_DIRECTORY, 'merge_new_corpus')
for filename in os.listdir(new_corpus_dir):
src = os.path.join(new_corpus_dir, filename)
dst = os.path.join(queue, filename)
shutil.copy(src, dst)
return new_process.ProcessResult(
command=[], return_code=0, output='', time_executed=1)
def main(argv):
"""Run libFuzzer as specified by argv."""
atexit.register(fuzzer_utils.cleanup)
# Initialize variables.
arguments = argv[1:]
testcase_file_path = arguments.pop(0)
target_name = arguments.pop(0)
fuzzer_name = data_types.fuzz_target_project_qualified_name(
utils.current_project(), target_name)
# Initialize log handler.
logs.configure(
'run_fuzzer', {
'fuzzer': fuzzer_name,
'engine': 'libFuzzer',
'job_name': environment.get_value('JOB_NAME')
})
profiler.start_if_needed('libfuzzer_launcher')
# Make sure that the fuzzer binary exists.
build_directory = environment.get_value('BUILD_DIR')
fuzzer_path = engine_common.find_fuzzer_path(build_directory, target_name)
if not fuzzer_path:
# This is an expected case when doing regression testing with old builds
# that do not have that fuzz target. It can also happen when a host sends a
# message to an untrusted worker that just restarted and lost information on
# build directory.
logs.log_warn('Could not find fuzz target %s.' % target_name)
return
# Install signal handler.
signal.signal(signal.SIGTERM, engine_common.signal_term_handler)
# Set up temp dir.
engine_common.recreate_directory(fuzzer_utils.get_temp_dir())
# Setup minijail if needed.
use_minijail = environment.get_value('USE_MINIJAIL')
runner = libfuzzer.get_runner(fuzzer_path,
temp_dir=fuzzer_utils.get_temp_dir())
if use_minijail:
minijail_chroot = runner.chroot
else:
minijail_chroot = None
# Get corpus directory.
corpus_directory = environment.get_value('FUZZ_CORPUS_DIR')
# Add common arguments which are necessary to be used for every run.
arguments = expand_with_common_arguments(arguments)
# Add sanitizer options to environment that were specified in the .options
# file and options that this script requires.
set_sanitizer_options(fuzzer_path)
# Minimize test argument.
minimize_to = fuzzer_utils.extract_argument(arguments,
MINIMIZE_TO_ARGUMENT)
minimize_timeout = fuzzer_utils.extract_argument(
arguments, MINIMIZE_TIMEOUT_ARGUMENT)
if minimize_to and minimize_timeout:
minimize_testcase(runner, testcase_file_path, minimize_to,
int(minimize_timeout), arguments, use_minijail)
return
# Cleanse argument.
cleanse_to = fuzzer_utils.extract_argument(arguments, CLEANSE_TO_ARGUMENT)
cleanse_timeout = fuzzer_utils.extract_argument(arguments,
CLEANSE_TIMEOUT_ARGUMENT)
if cleanse_to and cleanse_timeout:
cleanse_testcase(runner, testcase_file_path, cleanse_to,
int(cleanse_timeout), arguments, use_minijail)
return
# If we don't have a corpus, then that means this is not a fuzzing run.
if not corpus_directory:
load_testcase_if_exists(runner, testcase_file_path, fuzzer_name,
use_minijail, arguments)
return
# We don't have a crash testcase, fuzz.
# Check dict argument to make sure that it's valid.
dict_argument = fuzzer_utils.extract_argument(arguments,
constants.DICT_FLAG,
remove=False)
if dict_argument and not os.path.exists(dict_argument):
logs.log_error('Invalid dict %s for %s.' %
(dict_argument, fuzzer_name))
fuzzer_utils.extract_argument(arguments, constants.DICT_FLAG)
# If there's no dict argument, check for %target_binary_name%.dict file.
if (not fuzzer_utils.extract_argument(
arguments, constants.DICT_FLAG, remove=False)):
default_dict_path = dictionary_manager.get_default_dictionary_path(
fuzzer_path)
if os.path.exists(default_dict_path):
arguments.append(constants.DICT_FLAG + default_dict_path)
fuzzing_strategies = []
# Select a generator to use for existing testcase mutations.
generator = _select_generator()
is_mutations_run = generator != Generator.NONE
# Timeout for fuzzer run.
fuzz_timeout = get_fuzz_timeout(is_mutations_run)
# Set up scratch directory for writing new units.
new_testcases_directory = create_corpus_directory('new')
# Get list of corpus directories.
corpus_directories = get_corpus_directories(corpus_directory,
new_testcases_directory,
fuzzer_path,
fuzzing_strategies,
minijail_chroot)
# Bind corpus directories in minijail.
if use_minijail:
artifact_prefix = constants.ARTIFACT_PREFIX_FLAG + '/'
else:
artifact_prefix = '%s%s/' % (constants.ARTIFACT_PREFIX_FLAG,
os.path.abspath(
os.path.dirname(testcase_file_path)))
# Generate new testcase mutations using radamsa, etc.
if is_mutations_run:
new_testcase_mutations_directory = generate_new_testcase_mutations(
corpus_directory, fuzzer_name, generator, fuzzing_strategies)
corpus_directories.append(new_testcase_mutations_directory)
if use_minijail:
bind_corpus_dirs(minijail_chroot,
[new_testcase_mutations_directory])
max_len_argument = fuzzer_utils.extract_argument(arguments,
constants.MAX_LEN_FLAG,
remove=False)
if not max_len_argument and do_random_max_length():
max_length = random.SystemRandom().randint(1, MAX_VALUE_FOR_MAX_LENGTH)
arguments.append('%s%d' % (constants.MAX_LEN_FLAG, max_length))
fuzzing_strategies.append(strategy.RANDOM_MAX_LENGTH_STRATEGY)
if do_recommended_dictionary():
if add_recommended_dictionary(arguments, fuzzer_name, fuzzer_path):
fuzzing_strategies.append(strategy.RECOMMENDED_DICTIONARY_STRATEGY)
if do_value_profile():
arguments.append(constants.VALUE_PROFILE_ARGUMENT)
fuzzing_strategies.append(strategy.VALUE_PROFILE_STRATEGY)
if do_fork():
max_fuzz_threads = environment.get_value('MAX_FUZZ_THREADS', 1)
num_fuzz_processes = max(
1,
multiprocessing.cpu_count() // max_fuzz_threads)
arguments.append('%s%d' % (constants.FORK_FLAG, num_fuzz_processes))
fuzzing_strategies.append('%s_%d' %
(strategy.FORK_STRATEGY, num_fuzz_processes))
extra_env = {}
if do_mutator_plugin():
if use_mutator_plugin(target_name, extra_env, minijail_chroot):
fuzzing_strategies.append(strategy.MUTATOR_PLUGIN_STRATEGY)
# Execute the fuzzer binary with original arguments.
fuzz_result = runner.fuzz(corpus_directories,
fuzz_timeout=fuzz_timeout,
additional_args=arguments + [artifact_prefix],
extra_env=extra_env)
if (not 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, 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 = None
for line in log_lines:
match = re.match(CRASH_TESTCASE_REGEX, line)
if match:
crash_testcase_file_path = match.group(1)
break
if crash_testcase_file_path:
# Write the new testcase.
if use_minijail:
# Convert chroot relative path to host path. Remove the leading '/' before
# joining.
crash_testcase_file_path = os.path.join(
minijail_chroot.directory, crash_testcase_file_path[1:])
# Copy crash testcase contents into the main testcase path.
shutil.move(crash_testcase_file_path, testcase_file_path)
# Print the command output.
log_header_format = ('Command: %s\n' 'Bot: %s\n' 'Time ran: %f\n')
bot_name = environment.get_value('BOT_NAME', '')
command = fuzz_result.command
if use_minijail:
# Remove minijail prefix.
command = engine_common.strip_minijail_command(command, fuzzer_path)
print(log_header_format % (engine_common.get_command_quoted(command),
bot_name, fuzz_result.time_executed))
# Parse stats information based on libFuzzer output.
parsed_stats = parse_log_stats(log_lines)
# Extend parsed stats by additional performance features.
parsed_stats.update(
stats.parse_performance_features(log_lines, fuzzing_strategies,
arguments))
# Set some initial stat overrides.
timeout_limit = fuzzer_utils.extract_argument(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)
stat_overrides = {
'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.
remove_fuzzing_arguments(arguments)
# Make a decision on whether merge step is needed at all. If there are no
# new units added by libFuzzer run, then no need to do merge at all.
new_units_added = shell.get_directory_file_count(new_testcases_directory)
merge_error = None
if new_units_added:
# Merge the new units with the initial corpus.
if corpus_directory not in corpus_directories:
corpus_directories.append(corpus_directory)
# If this times out, it's possible that we will miss some units. However, if
# we're taking >10 minutes to load/merge the corpus something is going very
# wrong and we probably don't want to make things worse by adding units
# anyway.
merge_tmp_dir = None
if not use_minijail:
merge_tmp_dir = os.path.join(fuzzer_utils.get_temp_dir(),
'merge_workdir')
engine_common.recreate_directory(merge_tmp_dir)
old_corpus_len = shell.get_directory_file_count(corpus_directory)
merge_directory = create_merge_directory()
corpus_directories.insert(0, merge_directory)
if use_minijail:
bind_corpus_dirs(minijail_chroot, [merge_directory])
merge_result = runner.merge(
corpus_directories,
merge_timeout=engine_common.get_merge_timeout(
DEFAULT_MERGE_TIMEOUT),
tmp_dir=merge_tmp_dir,
additional_args=arguments)
move_mergeable_units(merge_directory, corpus_directory)
new_corpus_len = shell.get_directory_file_count(corpus_directory)
new_units_added = 0
merge_error = None
if merge_result.timed_out:
merge_error = 'Merging new testcases timed out:'
elif merge_result.return_code != 0:
merge_error = 'Merging new testcases failed:'
else:
new_units_added = new_corpus_len - old_corpus_len
stat_overrides['new_units_added'] = new_units_added
if merge_result.output:
stat_overrides.update(
stats.parse_stats_from_merge_log(
merge_result.output.splitlines()))
else:
stat_overrides['new_units_added'] = 0
logs.log('Skipped corpus merge since no new units added by fuzzing.')
# Get corpus size after merge. This removes the duplicate units that were
# created during this fuzzing session.
stat_overrides['corpus_size'] = shell.get_directory_file_count(
corpus_directory)
# Delete all corpus directories except for the main one. These were temporary
# directories to store new testcase mutations and have already been merged to
# main corpus directory.
if corpus_directory in corpus_directories:
corpus_directories.remove(corpus_directory)
for directory in corpus_directories:
shutil.rmtree(directory, ignore_errors=True)
if use_minijail:
unbind_corpus_dirs(minijail_chroot, corpus_directories)
# Apply overridden stats to the parsed stats prior to dumping.
parsed_stats.update(stat_overrides)
# Dump stats data for further uploading to BigQuery.
engine_common.dump_big_query_data(parsed_stats, testcase_file_path,
LIBFUZZER_PREFIX, fuzzer_name, command)
# Add custom crash state based on fuzzer name (if needed).
add_custom_crash_state_if_needed(fuzzer_name, log_lines, parsed_stats)
for line in log_lines:
print(line)
# Add fuzzing strategies used.
engine_common.print_fuzzing_strategies(fuzzing_strategies)
# Add merge error (if any).
if merge_error:
print(data_types.CRASH_STACKTRACE_END_MARKER)
print(merge_error)
print(
'Command:',
get_printable_command(merge_result.command, fuzzer_path,
use_minijail))
print(merge_result.output)
analyze_and_update_recommended_dictionary(runner, fuzzer_name, log_lines,
corpus_directory, arguments)
# Close minijail chroot.
if use_minijail:
minijail_chroot.close()
# Record the stats to make them easily searchable in stackdriver.
if new_units_added:
logs.log('New units added to corpus: %d.' % new_units_added,
stats=parsed_stats)
else:
logs.log('No new units found.', stats=parsed_stats)
def create_corpus_directory(name):
"""Create a corpus directory with a give name in temp directory and return its
full path."""
new_corpus_directory = os.path.join(fuzzer_utils.get_temp_dir(), name)
engine_common.recreate_directory(new_corpus_directory)
return new_corpus_directory
def _create_temp_corpus_dir(self, name):
"""Create temporary corpus directory."""
new_corpus_directory = os.path.join(fuzzer_utils.get_temp_dir(), name)
engine_common.recreate_directory(new_corpus_directory)
return new_corpus_directory
def main(argv):
"""Run libFuzzer as specified by argv."""
atexit.register(fuzzer_utils.cleanup)
# Initialize variables.
arguments = argv[1:]
testcase_file_path = arguments.pop(0)
target_name = environment.get_value('FUZZ_TARGET')
if arguments and arguments[0] == target_name:
# Pop legacy fuzz target argument.
arguments.pop(0)
fuzzer_name = data_types.fuzz_target_project_qualified_name(
utils.current_project(), target_name)
# Initialize log handler.
logs.configure(
'run_fuzzer', {
'fuzzer': fuzzer_name,
'engine': 'libFuzzer',
'job_name': environment.get_value('JOB_NAME')
})
profiler.start_if_needed('libfuzzer_launcher')
# Make sure that the fuzzer binary exists.
build_directory = environment.get_value('BUILD_DIR')
fuzzer_path = engine_common.find_fuzzer_path(build_directory, target_name)
if not fuzzer_path:
return
# Install signal handler.
signal.signal(signal.SIGTERM, engine_common.signal_term_handler)
# Set up temp dir.
engine_common.recreate_directory(fuzzer_utils.get_temp_dir())
# Setup minijail if needed.
use_minijail = environment.get_value('USE_MINIJAIL')
runner = libfuzzer.get_runner(
fuzzer_path, temp_dir=fuzzer_utils.get_temp_dir())
if use_minijail:
minijail_chroot = runner.chroot
else:
minijail_chroot = None
# Get corpus directory.
corpus_directory = environment.get_value('FUZZ_CORPUS_DIR')
# Add common arguments which are necessary to be used for every run.
arguments = expand_with_common_arguments(arguments)
# Add sanitizer options to environment that were specified in the .options
# file and options that this script requires.
set_sanitizer_options(fuzzer_path)
# If we don't have a corpus, then that means this is not a fuzzing run.
if not corpus_directory:
load_testcase_if_exists(runner, testcase_file_path, fuzzer_name,
use_minijail, arguments)
return
# We don't have a crash testcase, fuzz.
# Check dict argument to make sure that it's valid.
dict_argument = fuzzer_utils.extract_argument(
arguments, constants.DICT_FLAG, remove=False)
if dict_argument and not os.path.exists(dict_argument):
logs.log_error('Invalid dict %s for %s.' % (dict_argument, fuzzer_name))
fuzzer_utils.extract_argument(arguments, constants.DICT_FLAG)
# If there's no dict argument, check for %target_binary_name%.dict file.
if (not fuzzer_utils.extract_argument(
arguments, constants.DICT_FLAG, remove=False)):
default_dict_path = dictionary_manager.get_default_dictionary_path(
fuzzer_path)
if os.path.exists(default_dict_path):
arguments.append(constants.DICT_FLAG + default_dict_path)
# Set up scratch directory for writing new units.
new_testcases_directory = create_corpus_directory('new')
# Strategy pool is the list of strategies that we attempt to enable, whereas
# fuzzing strategies is the list of strategies that are enabled. (e.g. if
# mutator is selected in the pool, but not available for a given target, it
# would not be added to fuzzing strategies.)
strategy_pool = strategy_selection.generate_weighted_strategy_pool(
strategy_list=strategy.LIBFUZZER_STRATEGY_LIST,
use_generator=True,
engine_name='libFuzzer')
strategy_info = pick_strategies(
strategy_pool,
fuzzer_path,
corpus_directory,
arguments,
minijail_chroot=minijail_chroot)
arguments.extend(strategy_info.arguments)
# Timeout for fuzzer run.
fuzz_timeout = get_fuzz_timeout(strategy_info.is_mutations_run)
# Get list of corpus directories.
# TODO(flowerhack): Implement this to handle corpus sync'ing.
if environment.platform() == 'FUCHSIA':
corpus_directories = []
else:
corpus_directories = get_corpus_directories(
corpus_directory,
new_testcases_directory,
fuzzer_path,
strategy_info.fuzzing_strategies,
strategy_pool,
minijail_chroot=minijail_chroot,
allow_corpus_subset=not strategy_info.use_dataflow_tracing)
corpus_directories.extend(strategy_info.additional_corpus_dirs)
artifact_prefix = os.path.abspath(os.path.dirname(testcase_file_path))
# Execute the fuzzer binary with original arguments.
fuzz_result = runner.fuzz(
corpus_directories,
fuzz_timeout=fuzz_timeout,
artifact_prefix=artifact_prefix,
additional_args=arguments,
extra_env=strategy_info.extra_env)
if (not 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, 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 crash_testcase_file_path:
# Copy crash testcase contents into the main testcase path.
shutil.move(crash_testcase_file_path, testcase_file_path)
# Print the command output.
bot_name = environment.get_value('BOT_NAME', '')
command = fuzz_result.command
if use_minijail:
# Remove minijail prefix.
command = engine_common.strip_minijail_command(command, fuzzer_path)
print(engine_common.get_log_header(command, bot_name,
fuzz_result.time_executed))
# Parse stats information based on libFuzzer output.
parsed_stats = parse_log_stats(log_lines)
# Extend parsed stats by additional performance features.
parsed_stats.update(
stats.parse_performance_features(
log_lines, strategy_info.fuzzing_strategies, arguments))
# Set some initial stat overrides.
timeout_limit = fuzzer_utils.extract_argument(
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)
stat_overrides = {
'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.
remove_fuzzing_arguments(arguments)
# Make a decision on whether merge step is needed at all. If there are no
# new units added by libFuzzer run, then no need to do merge at all.
new_units_added = shell.get_directory_file_count(new_testcases_directory)
merge_error = None
if new_units_added:
# Merge the new units with the initial corpus.
if corpus_directory not in corpus_directories:
corpus_directories.append(corpus_directory)
# If this times out, it's possible that we will miss some units. However, if
# we're taking >10 minutes to load/merge the corpus something is going very
# wrong and we probably don't want to make things worse by adding units
# anyway.
merge_tmp_dir = None
if not use_minijail:
merge_tmp_dir = os.path.join(fuzzer_utils.get_temp_dir(), 'merge_workdir')
engine_common.recreate_directory(merge_tmp_dir)
old_corpus_len = shell.get_directory_file_count(corpus_directory)
merge_directory = create_merge_directory()
corpus_directories.insert(0, merge_directory)
if use_minijail:
bind_corpus_dirs(minijail_chroot, [merge_directory])
merge_result = runner.merge(
corpus_directories,
merge_timeout=engine_common.get_merge_timeout(DEFAULT_MERGE_TIMEOUT),
tmp_dir=merge_tmp_dir,
additional_args=arguments)
move_mergeable_units(merge_directory, corpus_directory)
new_corpus_len = shell.get_directory_file_count(corpus_directory)
new_units_added = 0
merge_error = None
if merge_result.timed_out:
merge_error = 'Merging new testcases timed out:'
elif merge_result.return_code != 0:
merge_error = 'Merging new testcases failed:'
else:
new_units_added = new_corpus_len - old_corpus_len
stat_overrides['new_units_added'] = new_units_added
if merge_result.output:
stat_overrides.update(
stats.parse_stats_from_merge_log(merge_result.output.splitlines()))
else:
stat_overrides['new_units_added'] = 0
logs.log('Skipped corpus merge since no new units added by fuzzing.')
# Get corpus size after merge. This removes the duplicate units that were
# created during this fuzzing session.
# TODO(flowerhack): Remove this workaround once we can handle corpus sync.
if environment.platform() != 'FUCHSIA':
stat_overrides['corpus_size'] = shell.get_directory_file_count(
corpus_directory)
# Delete all corpus directories except for the main one. These were temporary
# directories to store new testcase mutations and have already been merged to
# main corpus directory.
if corpus_directory in corpus_directories:
corpus_directories.remove(corpus_directory)
for directory in corpus_directories:
shutil.rmtree(directory, ignore_errors=True)
if use_minijail:
unbind_corpus_dirs(minijail_chroot, corpus_directories)
# Apply overridden stats to the parsed stats prior to dumping.
parsed_stats.update(stat_overrides)
# Dump stats data for further uploading to BigQuery.
engine_common.dump_big_query_data(parsed_stats, testcase_file_path, command)
# Add custom crash state based on fuzzer name (if needed).
add_custom_crash_state_if_needed(fuzzer_name, log_lines, parsed_stats)
for line in log_lines:
print(line)
# Add fuzzing strategies used.
print(engine_common.format_fuzzing_strategies(
strategy_info.fuzzing_strategies))
# Add merge error (if any).
if merge_error:
print(data_types.CRASH_STACKTRACE_END_MARKER)
print(merge_error)
print('Command:',
get_printable_command(merge_result.command, fuzzer_path,
use_minijail))
print(merge_result.output)
analyze_and_update_recommended_dictionary(runner, fuzzer_name, log_lines,
corpus_directory, arguments)
# Close minijail chroot.
if use_minijail:
minijail_chroot.close()
# Record the stats to make them easily searchable in stackdriver.
if new_units_added:
logs.log(
'New units added to corpus: %d.' % new_units_added, stats=parsed_stats)
else:
logs.log('No new units found.', stats=parsed_stats)
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 create_new_if_needed(self):
"""Checks if any inputs are too large for AFL. If not then does nothing.
Otherwise creates a temporary input directory and copies the non-oversized
inputs.
"""
# TODO(metzman): Get rid of this approach where a new corpus is created.
# Instead use an approach that modifies the input corpus permanently so that
# it doesn't have to be fixed every time by AFL.
# TODO(metzman): Copy testcases in subdirectories so AFL can use them, even
# when there are no oversized files.
corpus_file_paths = list_full_file_paths_recursive(
self.input_directory)
usable_files_and_sizes = [
(path, os.path.getsize(path)) for path in corpus_file_paths
if os.path.getsize(path) < constants.MAX_FILE_BYTES
]
num_files = len(usable_files_and_sizes)
if self.strategies.use_corpus_subset:
num_files = min(num_files, self.strategies.corpus_subset_size)
self.strategies.use_corpus_subset = (
self.strategies.corpus_subset_size == num_files)
if num_files == len(corpus_file_paths):
# Nothing to do here: using full corpus and all files are appropriately
# sized.
return None
# Save the original input directory.
self.original_input_directory = self.input_directory
# Make a new directory that we can modify.
self.input_directory = os.path.join(fuzzer_utils.get_temp_dir(),
'afl_input_dir')
engine_common.recreate_directory(self.input_directory)
copied_size = 0
for src_path, src_size in usable_files_and_sizes:
if not num_files:
break
num_files -= 1
copied_size += src_size
if copied_size > self.MAX_COPIED_CORPUS_SIZE:
break
filename = os.path.basename(src_path)
dst_path = os.path.join(self.input_directory, filename)
# TODO(metzman): Ask Michal to allow skipping of oversized inputs
# automatically. Just copy the small enough, files, we can't use soft
# links because of AFL and we can't use hard links because they do not
# work across devices.
shutil.copy(src_path, dst_path)
num_files = len(os.listdir(self.input_directory))
num_files_original = len(os.listdir(self.original_input_directory))
logs.log((
'Temporary input directory contains %d files. Original contains %d.'
% (num_files, num_files_original)))
return self.input_directory
def __init__(self):
self.output_directory = os.path.join(fuzzer_utils.get_temp_dir(),
'afl_output_dir')
engine_common.recreate_directory(self.output_directory)
def main(argv):
"""Run afl as specified by argv."""
atexit.register(fuzzer_utils.cleanup)
# Initialize variables.
_, testcase_file_path, target_name = argv[:3]
input_directory = environment.get_value('FUZZ_CORPUS_DIR')
fuzzer_name = data_types.fuzz_target_project_qualified_name(
utils.current_project(), target_name)
# Initialize log handler.
logs.configure(
'run_fuzzer', {
'fuzzer': fuzzer_name,
'engine': 'afl',
'job_name': environment.get_value('JOB_NAME')
})
build_directory = environment.get_value('BUILD_DIR')
fuzzer_path = engine_common.find_fuzzer_path(build_directory, target_name)
if not fuzzer_path:
# This is an expected case when doing regression testing with old builds
# that do not have that fuzz target. It can also happen when a host sends a
# message to an untrusted worker that just restarted and lost information on
# build directory.
logs.log_warn('Could not find fuzz target %s.' % target_name)
return
# Install signal handler.
signal.signal(signal.SIGTERM, engine_common.signal_term_handler)
# Set up temp dir.
engine_common.recreate_directory(fuzzer_utils.get_temp_dir())
config = AflConfig.from_target_path(fuzzer_path)
runner = AflRunner(fuzzer_path, config, testcase_file_path,
input_directory)
# Add *SAN_OPTIONS overrides from .options file.
engine_common.process_sanitizer_options_overrides(fuzzer_path)
# If we don't have a corpus, then that means this is not a fuzzing run.
if not input_directory:
load_testcase_if_exists(runner, testcase_file_path)
return
# Make sure afl won't exit because of bad sanitizer options.
set_additional_sanitizer_options_for_afl_fuzz()
# Execute afl-fuzz on the fuzzing target.
fuzz_result = runner.fuzz()
# Print info for the fuzzer logs.
command = fuzz_result.command
print('Command: {0}\n'
'Bot: {1}\n'
'Time ran: {2}\n').format(engine_common.get_command_quoted(command),
BOT_NAME, fuzz_result.time_executed)
print fuzz_result.output
runner.strategies.print_strategies()
if fuzz_result.return_code:
# If AFL returned a non-zero return code quit now without getting stats,
# since they would be meaningless.
print runner.fuzzer_stderr
return
stats_getter = stats.StatsGetter(runner.afl_output.stats_path,
config.dict_path)
try:
new_units_generated, new_units_added, corpus_size = (
runner.libfuzzerize_corpus())
stats_getter.set_stats(fuzz_result.time_executed, new_units_generated,
new_units_added, corpus_size, runner.strategies,
runner.fuzzer_stderr, fuzz_result.output)
engine_common.dump_big_query_data(stats_getter.stats,
testcase_file_path, AFL_PREFIX,
fuzzer_name, command)
finally:
print runner.fuzzer_stderr
# Whenever new units are added to corpus, record the stats to make them
# easily searchable in stackdriver.
if new_units_added:
logs.log('New units added to corpus: %d.' % new_units_added,
stats=stats_getter.stats)