def run(self, input_directory, output_directory, no_of_files):
"""Run the fuzzer to generate testcases."""
fuzzer_binary_name, fuzzer_path = self._get_fuzzer_binary_name_and_path(
)
project_qualified_name = data_types.fuzz_target_project_qualified_name(
utils.current_project(), fuzzer_binary_name)
arguments = self.generate_arguments(fuzzer_path)
corpus_directory = get_corpus_directory(input_directory,
project_qualified_name)
# Create fuzz testcases.
for i in range(no_of_files):
# Contents of testcase file don't matter at this point. Need to create
# something non-null so that it is not ignored.
testcase_file_path = os.path.join(
output_directory, '%s%d' % (testcase_manager.FUZZ_PREFIX, i))
utils.write_data_to_file(' ', testcase_file_path)
# Write the flags file containing command line for running launcher
# script.
flags_file_path = os.path.join(
output_directory, '%s%d' % (testcase_manager.FLAGS_PREFIX, i))
flags = ['%TESTCASE%', fuzzer_binary_name]
if arguments:
flags.append(arguments)
flags_file_content = ' '.join(flags)
utils.write_data_to_file(flags_file_content, flags_file_path)
output = 'Generated %d testcase for fuzzer %s.\n' % (
no_of_files, fuzzer_binary_name)
output += 'metadata::fuzzer_binary_name: %s\n' % fuzzer_binary_name
issue_owners = engine_common.get_issue_owners(fuzzer_path)
if issue_owners:
output += 'metadata::issue_owners: %s\n' % ','.join(issue_owners)
issue_labels = engine_common.get_issue_labels(fuzzer_path)
if issue_labels:
output += 'metadata::issue_labels: %s\n' % ','.join(issue_labels)
issue_components = engine_common.get_issue_components(fuzzer_path)
if issue_components:
output += 'metadata::issue_components: %s\n' % ','.join(
issue_components)
# Update *SAN_OPTIONS in current environment from .options file. This
# environment is used in fuzz task later for deriving the environment
# string in |get_environment_settings_as_string| and embedding this as
# part of stacktrace.
engine_common.process_sanitizer_options_overrides(fuzzer_path)
return BuiltinFuzzerResult(output=output,
corpus_directory=corpus_directory)
def test_overridden(self):
"""Test that env variable PROJECT_NAME is used for current project."""
os.environ['PROJECT_NAME'] = 'other-project'
self.assertEqual('other-project', utils.current_project())
def test_default(self):
"""Test that it returns default project with no environment changes."""
self.assertEqual('test-project', utils.current_project())
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 run(self, input_directory, output_directory, no_of_files):
"""Run the fuzzer to generate testcases."""
build_directory = environment.get_value('BUILD_DIR')
if not build_directory:
raise BuiltinFuzzerException(
'BUILD_DIR environment variable is not set.')
fuzzers = fuzzers_utils.get_fuzz_targets(build_directory)
if not fuzzers:
raise BuiltinFuzzerException(
'No fuzzer binaries found in |BUILD_DIR| directory.')
fuzzer_binary_name = environment.get_value('FUZZ_TARGET')
if fuzzer_binary_name:
fuzzer_path = _get_fuzzer_path(fuzzers, fuzzer_binary_name)
else:
fuzzer_path = random.SystemRandom().choice(fuzzers)
fuzzer_binary_name = os.path.basename(fuzzer_path)
project_qualified_name = data_types.fuzz_target_project_qualified_name(
utils.current_project(), fuzzer_binary_name)
corpus_directory = os.path.join(input_directory,
project_qualified_name)
if environment.is_trusted_host():
from bot.untrusted_runner import file_host
corpus_directory = file_host.rebase_to_worker_root(
corpus_directory)
arguments = self.generate_arguments(fuzzer_path)
# Create corpus directory if it does not exist already.
if environment.is_trusted_host():
from bot.untrusted_runner import file_host
file_host.create_directory(corpus_directory,
create_intermediates=True)
else:
if not os.path.exists(corpus_directory):
os.mkdir(corpus_directory)
# Create fuzz testcases.
for i in range(no_of_files):
# Contents of testcase file don't matter at this point. Need to create
# something non-null so that it is not ignored.
testcase_file_path = os.path.join(output_directory,
'%s%d' % (tests.FUZZ_PREFIX, i))
utils.write_data_to_file(' ', testcase_file_path)
# Write the flags file containing command line for running launcher
# script.
flags_file_path = os.path.join(output_directory,
'%s%d' % (tests.FLAGS_PREFIX, i))
flags = ['%TESTCASE%', fuzzer_binary_name]
if arguments:
flags.append(arguments)
flags_file_content = ' '.join(flags)
utils.write_data_to_file(flags_file_content, flags_file_path)
output = 'Generated %d testcase for fuzzer %s.\n' % (
no_of_files, fuzzer_binary_name)
output += 'metadata::fuzzer_binary_name: %s\n' % fuzzer_binary_name
issue_owners = engine_common.get_issue_owners(fuzzer_path)
if issue_owners:
output += 'metadata::issue_owners: %s\n' % ','.join(issue_owners)
issue_labels = engine_common.get_issue_labels(fuzzer_path)
if issue_labels:
output += 'metadata::issue_labels: %s\n' % ','.join(issue_labels)
# Update *SAN_OPTIONS in current environment from .options file. This
# environment is used in fuzz task later for deriving the environment
# string in |get_environment_settings_as_string| and embedding this as
# part of stacktrace.
engine_common.process_sanitizer_options_overrides(fuzzer_path)
return BuiltinFuzzerResult(output=output,
corpus_directory=corpus_directory)
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 Result object.
"""
profiler.start_if_needed('libfuzzer_fuzz')
runner = libfuzzer.get_runner(target_path)
launcher.set_sanitizer_options(target_path)
artifact_prefix = self._artifact_prefix(
os.path.abspath(reproducers_dir))
# 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],
extra_env=options.extra_env)
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(launcher.CRASH_TESTCASE_REGEX, line)
if match:
crash_testcase_file_path = match.group(1)
break
# 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)
# Add custom crash state based on fuzzer name (if needed).
project_qualified_fuzzer_name = (
data_types.fuzz_target_project_qualified_name(
utils.current_project(), os.path.basename(target_path)))
launcher.add_custom_crash_state_if_needed(
project_qualified_fuzzer_name, log_lines, 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))
launcher.analyze_and_update_recommended_dictionary(
runner, project_qualified_fuzzer_name, log_lines,
options.corpus_dir, arguments)
return engine.Result(fuzz_logs, fuzz_result.command, crashes,
parsed_stats, fuzz_result.time_executed)
def get_project_qualified_fuzzer_name(target_path):
"""Return project qualified fuzzer name for a given target path."""
return data_types.fuzz_target_project_qualified_name(
utils.current_project(), os.path.basename(target_path))
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 pick_strategies(strategy_pool,
fuzzer_path,
corpus_directory,
existing_arguments,
minijail_chroot=None):
"""Pick strategies."""
build_directory = environment.get_value('BUILD_DIR')
target_name = os.path.basename(fuzzer_path)
project_qualified_fuzzer_name = data_types.fuzz_target_project_qualified_name(
utils.current_project(), target_name)
fuzzing_strategies = []
arguments = []
additional_corpus_dirs = []
# Select a generator to attempt to use for existing testcase mutations.
candidate_generator = engine_common.select_generator(strategy_pool,
fuzzer_path)
is_mutations_run = candidate_generator != engine_common.Generator.NONE
# Depends on the presense of DFSan instrumented build.
dataflow_build_dir = environment.get_value('DATAFLOW_BUILD_DIR')
use_dataflow_tracing = (
dataflow_build_dir and
strategy_pool.do_strategy(strategy.DATAFLOW_TRACING_STRATEGY))
if use_dataflow_tracing:
dataflow_binary_path = os.path.join(
dataflow_build_dir, os.path.relpath(fuzzer_path, build_directory))
if os.path.exists(dataflow_binary_path):
arguments.append(
'%s%s' % (constants.COLLECT_DATA_FLOW_FLAG, dataflow_binary_path))
fuzzing_strategies.append(strategy.DATAFLOW_TRACING_STRATEGY.name)
else:
logs.log_error(
'Fuzz target is not found in dataflow build, skiping strategy.')
use_dataflow_tracing = False
# Generate new testcase mutations using radamsa, etc.
if is_mutations_run:
new_testcase_mutations_directory = create_corpus_directory('mutations')
generator_used = engine_common.generate_new_testcase_mutations(
corpus_directory, new_testcase_mutations_directory,
project_qualified_fuzzer_name, candidate_generator)
# Add the used generator strategy to our fuzzing strategies list.
if generator_used:
if candidate_generator == engine_common.Generator.RADAMSA:
fuzzing_strategies.append(
strategy.CORPUS_MUTATION_RADAMSA_STRATEGY.name)
elif candidate_generator == engine_common.Generator.ML_RNN:
fuzzing_strategies.append(strategy.CORPUS_MUTATION_ML_RNN_STRATEGY.name)
additional_corpus_dirs.append(new_testcase_mutations_directory)
if minijail_chroot:
bind_corpus_dirs(minijail_chroot, [new_testcase_mutations_directory])
if strategy_pool.do_strategy(strategy.RANDOM_MAX_LENGTH_STRATEGY):
max_len_argument = fuzzer_utils.extract_argument(
existing_arguments, constants.MAX_LEN_FLAG, remove=False)
if not max_len_argument:
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.name)
if (strategy_pool.do_strategy(strategy.RECOMMENDED_DICTIONARY_STRATEGY) and
add_recommended_dictionary(arguments, project_qualified_fuzzer_name,
fuzzer_path)):
fuzzing_strategies.append(strategy.RECOMMENDED_DICTIONARY_STRATEGY.name)
if strategy_pool.do_strategy(strategy.VALUE_PROFILE_STRATEGY):
arguments.append(constants.VALUE_PROFILE_ARGUMENT)
fuzzing_strategies.append(strategy.VALUE_PROFILE_STRATEGY.name)
# DataFlow Tracing requires fork mode, always use it with DFT strategy.
if use_dataflow_tracing or strategy_pool.do_strategy(strategy.FORK_STRATEGY):
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.name, num_fuzz_processes))
extra_env = {}
if (strategy_pool.do_strategy(strategy.MUTATOR_PLUGIN_STRATEGY) and
use_mutator_plugin(target_name, extra_env, minijail_chroot)):
fuzzing_strategies.append(strategy.MUTATOR_PLUGIN_STRATEGY.name)
return StrategyInfo(fuzzing_strategies, arguments, additional_corpus_dirs,
extra_env, use_dataflow_tracing, is_mutations_run)
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)
# 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)
if (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,
engine_output=fuzz_result.output)
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 = 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,
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[:]
libfuzzer.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:
# Use higher timeout for reproduction.
reproduce_arguments = arguments[:]
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))
project_qualified_fuzzer_name = (
data_types.fuzz_target_project_qualified_name(
utils.current_project(), os.path.basename(target_path)))
libfuzzer.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 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)
