def conduct_trial(self):
"""Conduct the benchmarking trial."""
self.initialize_directories()
log_file = os.path.join(self.results_dir, 'fuzzer-log.txt')
logs.info('Starting trial.')
max_total_time = environment.get('MAX_TOTAL_TIME')
args = (max_total_time, log_file)
fuzz_thread = threading.Thread(target=run_fuzzer, args=args)
fuzz_thread.start()
if environment.get('FUZZ_OUTSIDE_EXPERIMENT'):
# Hack so that the fuzz_thread has some time to fail if something is
# wrong. Without this we will sleep for a long time before checking
# if the fuzz thread is alive.
time.sleep(5)
while fuzz_thread.is_alive():
self.sleep_until_next_sync()
self.do_sync()
self.cycle += 1
logs.info('Doing final sync.')
self.do_sync(final_sync=True)
fuzz_thread.join()
def _unpack_clusterfuzz_seed_corpus(fuzz_target_path, corpus_directory):
"""If a clusterfuzz seed corpus archive is available, unpack it into the
corpus directory if it exists. Copied from unpack_seed_corpus in
engine_common.py in ClusterFuzz.
"""
seed_corpus_archive_path = get_clusterfuzz_seed_corpus_path(
fuzz_target_path)
if not seed_corpus_archive_path:
return
with zipfile.ZipFile(seed_corpus_archive_path) as zip_file:
# Unpack seed corpus recursively into the root of the main corpus
# directory.
idx = 0
for seed_corpus_file in zip_file.infolist():
if seed_corpus_file.filename.endswith('/'):
# Ignore directories.
continue
# Allow callers to opt-out of unpacking large files.
if seed_corpus_file.file_size > CORPUS_ELEMENT_BYTES_LIMIT:
continue
output_filename = '%016d' % idx
output_file_path = os.path.join(corpus_directory, output_filename)
zip_file.extract(seed_corpus_file, output_file_path)
idx += 1
logs.info('Unarchived %d files from seed corpus %s.', idx,
seed_corpus_archive_path)
def measure_loop(experiment: str, max_total_time: int):
"""Continuously measure trials for |experiment|."""
logs.initialize(default_extras={
'component': 'dispatcher',
'subcomponent': 'measurer',
})
logs.info('Start measure_loop.')
with multiprocessing.Pool() as pool, multiprocessing.Manager() as manager:
set_up_coverage_binaries(pool, experiment)
# Using Multiprocessing.Queue will fail with a complaint about
# inheriting queue.
q = manager.Queue() # pytype: disable=attribute-error
while True:
try:
# Get whether all trials have ended before we measure to prevent
# races.
all_trials_ended = scheduler.all_trials_ended(experiment)
if not measure_all_trials(experiment, max_total_time, pool, q):
# We didn't measure any trials.
if all_trials_ended:
# There are no trials producing snapshots to measure.
# Given that we couldn't measure any snapshots, we won't
# be able to measure any the future, so stop now.
break
except Exception: # pylint: disable=broad-except
logger.error('Error occurred during measuring.')
time.sleep(FAIL_WAIT_SECONDS)
logger.info('Finished measuring.')
def _clean_seed_corpus(seed_corpus_dir):
"""Prepares |seed_corpus_dir| for the trial. This ensures that it can be
used by AFL which is picky about the seed corpus. Moves seed corpus files
from sub-directories into the corpus directory root. Also, deletes any files
that exceed the 1 MB limit. If the NO_SEEDS env var is specified than the
seed corpus files are deleted."""
if not os.path.exists(seed_corpus_dir):
return
if environment.get('NO_SEEDS'):
logs.info('NO_SEEDS specified, deleting seed corpus files.')
shutil.rmtree(seed_corpus_dir)
os.mkdir(seed_corpus_dir)
return
failed_to_move_files = []
for root, _, files in os.walk(seed_corpus_dir):
for filename in files:
file_path = os.path.join(root, filename)
if os.path.getsize(file_path) > CORPUS_ELEMENT_BYTES_LIMIT:
os.remove(file_path)
logs.warning('Removed seed file %s as it exceeds 1 Mb limit.',
file_path)
continue
sha1sum = utils.file_hash(file_path)
new_file_path = os.path.join(seed_corpus_dir, sha1sum)
try:
shutil.move(file_path, new_file_path)
except OSError:
failed_to_move_files.append((file_path, new_file_path))
if failed_to_move_files:
logs.error('Failed to move seed corpus files: %s', failed_to_move_files)
def build_measurer(benchmark: str) -> bool:
"""Do a coverage build for a benchmark."""
try:
logger.info('Building measurer for benchmark: %s.', benchmark)
buildlib.build_coverage(benchmark)
docker_name = benchmark_utils.get_docker_name(benchmark)
archive_name = 'coverage-build-%s.tar.gz' % docker_name
coverage_binaries_dir = build_utils.get_coverage_binaries_dir()
benchmark_coverage_binary_dir = coverage_binaries_dir / benchmark
os.mkdir(benchmark_coverage_binary_dir)
cloud_bucket_archive_path = exp_path.gcs(coverage_binaries_dir /
archive_name)
gsutil.cp(cloud_bucket_archive_path,
str(benchmark_coverage_binary_dir),
parallel=False,
write_to_stdout=False)
archive_path = benchmark_coverage_binary_dir / archive_name
tar = tarfile.open(archive_path, 'r:gz')
tar.extractall(benchmark_coverage_binary_dir)
os.remove(archive_path)
logs.info('Done building measurer for benchmark: %s.', benchmark)
return True
except Exception: # pylint: disable=broad-except
logger.error('Failed to build measurer for %s.', benchmark)
return False
def main():
"""Copies files needed to integrate an OSS-Fuzz benchmark and creates the
benchmark's benchmark.yaml file."""
parser = argparse.ArgumentParser(description='Integrate a new benchmark.')
parser.add_argument('-p',
'--project',
help='Project for benchmark. Example: "zlib"',
required=True)
parser.add_argument(
'-f',
'--fuzz-target',
help='Fuzz target for benchmark. Example: "zlib_uncompress_fuzzer"',
required=True)
parser.add_argument(
'-n',
'--benchmark-name',
help='Benchmark name. Defaults to <project>_<fuzz_target>',
required=False)
parser.add_argument('-c', '--commit', help='Project commit hash.')
parser.add_argument(
'-d',
'--date',
help='Date of the commit. Example: 2019-10-19T09:07:25+01:00')
logs.initialize()
args = parser.parse_args()
benchmark = integrate_benchmark(args.project, args.fuzz_target,
args.benchmark_name, args.commit,
args.date)
logs.info('Successfully integrated benchmark: %s.', benchmark)
logs.info('Please run "make test-run-afl-%s" to test integration.',
benchmark)
return 0
def run_diff_experiment(dry_run):
"""Run a diff experiment. This is an experiment that runs only on
fuzzers that have changed since the last experiment."""
fuzzers = experiment_changes.get_fuzzers_changed_since_last()
logs.info('Running experiment with fuzzers: %s.', ' '.join(fuzzers))
fuzzer_configs = fuzzer_utils.get_fuzzer_configs(fuzzers=fuzzers)
return _run_experiment(fuzzer_configs, dry_run)
def create_instance(instance_name: str,
instance_type: InstanceType,
config: dict,
startup_script: str = None,
preemptible: bool = False,
**kwargs) -> bool:
"""Creates a GCE instance with name, |instance_name|, type, |instance_type|
and with optionally provided and |startup_script|."""
if experiment_utils.is_local_experiment():
return run_local_instance(startup_script)
command = [
'gcloud',
'compute',
'instances',
'create',
instance_name,
'--image-family=cos-stable',
'--image-project=cos-cloud',
'--zone=%s' % config['cloud_compute_zone'],
'--scopes=cloud-platform',
]
if instance_type == InstanceType.DISPATCHER:
command.extend([
'--machine-type=%s' % DISPATCHER_MACHINE_TYPE,
'--boot-disk-size=%s' % DISPATCHER_BOOT_DISK_SIZE,
'--boot-disk-type=%s' % DISPATCHER_BOOT_DISK_TYPE,
])
else:
machine_type = config['runner_machine_type']
if machine_type is not None:
command.append('--machine-type=%s' % machine_type)
else:
# Do this to support KLEE experiments.
command.append([
'--custom-memory=%s' % config['runner_memory'],
'--custom-cpu=%s' % config['runner_num_cpu_cores']
])
command.extend([
'--no-address',
'--boot-disk-size=%s' % RUNNER_BOOT_DISK_SIZE,
])
if preemptible:
command.append('--preemptible')
if startup_script:
command.extend(
['--metadata-from-file', 'startup-script=' + startup_script])
result = new_process.execute(command, expect_zero=False, **kwargs)
if result.retcode == 0:
return True
logs.info('Failed to create instance. Command: %s failed. Output: %s',
command, result.output)
return False
def _run_experiment(experiment_name, fuzzers, dry_run=False):
"""Run an experiment named |experiment_name| on |fuzzer_configs| and shut it
down once it terminates."""
logs.info('Starting experiment: %s.', experiment_name)
if dry_run:
logs.info('Dry run. Not actually running experiment.')
return
run_experiment.start_experiment(experiment_name, EXPERIMENT_CONFIG_FILE,
BENCHMARKS, fuzzers)
def experiment_main():
"""Do a trial as part of an experiment."""
logs.info('Doing trial as part of experiment.')
try:
runner = TrialRunner()
runner.conduct_trial()
except Exception as error: # pylint: disable=broad-except
logs.error('Error doing trial.')
raise error
def replace_base_builder(benchmark_dir, commit_date):
"""Replace the parent image of the Dockerfile in |benchmark_dir|,
base-builder (latest), with a version of base-builder that is likely to
build the project as it was on |commit_date| without issue."""
base_builder_repo = _load_base_builder_docker_repo() # pylint: disable=protected-access
if base_builder_repo:
base_builder_digest = base_builder_repo.find_digest(commit_date)
logs.info('Using base-builder with digest %s.', base_builder_digest)
_replace_base_builder_digest( # pylint: disable=protected-access
os.path.join(benchmark_dir, 'Dockerfile'), base_builder_digest)
def build_measurer(benchmark: str) -> bool:
"""Do a coverage build for a benchmark."""
try:
logger.info('Building measurer for benchmark: %s.', benchmark)
buildlib.build_coverage(benchmark)
logs.info('Done building measurer for benchmark: %s.', benchmark)
return True
except Exception: # pylint: disable=broad-except
logger.error('Failed to build measurer for %s.', benchmark)
return False
def dispatcher_main():
"""Do the experiment and report results."""
logs.info('Starting experiment.')
# Set this here because we get failures if we do it in measurer for some
# reason.
multiprocessing.set_start_method('spawn')
db_utils.initialize()
if experiment_utils.is_local_experiment():
models.Base.metadata.create_all(db_utils.engine)
experiment_config_file_path = _get_config_file_path()
experiment = Experiment(experiment_config_file_path)
_initialize_experiment_in_db(experiment.config)
trials = build_images_for_trials(experiment.fuzzers, experiment.benchmarks,
experiment.num_trials,
experiment.preemptible,
experiment.concurrent_builds)
_initialize_trials_in_db(trials)
create_work_subdirs(['experiment-folders', 'measurement-folders'])
# Start measurer and scheduler in seperate threads/processes.
scheduler_loop_thread = threading.Thread(target=scheduler.schedule_loop,
args=(experiment.config, ))
scheduler_loop_thread.start()
measurer_main_process = multiprocessing.Process(
target=measure_manager.measure_main, args=(experiment.config, ))
measurer_main_process.start()
is_complete = False
while True:
time.sleep(LOOP_WAIT_SECONDS)
if not scheduler_loop_thread.is_alive():
is_complete = not measurer_main_process.is_alive()
# Generate periodic output reports.
reporter.output_report(experiment.config,
in_progress=not is_complete,
coverage_report=is_complete)
if is_complete:
# Experiment is complete, bail out.
break
scheduler_loop_thread.join()
measurer_main_process.join()
_record_experiment_time_ended(experiment.experiment_name)
logs.info('Experiment ended.')
def start(self):
"""Start the experiment on the dispatcher."""
logs.info('Started dispatcher with instance name: %s',
self.instance_name)
with tempfile.NamedTemporaryFile(dir=os.getcwd(),
mode='w') as startup_script:
self.write_startup_script(startup_script)
gcloud.create_instance(self.instance_name,
gcloud.InstanceType.DISPATCHER,
self.config,
startup_script=startup_script.name)
def build_fuzzer_benchmark(fuzzer: str, benchmark: str) -> bool:
"""Wrapper around buildlib.build_fuzzer_benchmark that logs and catches
exceptions."""
logger.info('Building benchmark: %s, fuzzer: %s.', benchmark, fuzzer)
try:
buildlib.build_fuzzer_benchmark(fuzzer, benchmark)
except subprocess.CalledProcessError:
logger.error('Failed to build benchmark: %s, fuzzer: %s.', benchmark,
fuzzer)
return False
logs.info('Done building benchmark: %s, fuzzer: %s.', benchmark, fuzzer)
return True
def main():
"""Creates or gets an already created service account key and saves it to
the provided path."""
logs.initialize()
try:
keyfile = sys.argv[1]
get_or_create_key(sys.argv[2], keyfile)
logs.info('Saved key to %s.', keyfile)
except Exception: # pylint: disable=broad-except
logs.error('Failed to get or create key.')
return 1
return 0
def robust_begin_gcloud_ssh(instance_name: str, zone: str):
"""Try to SSH into an instance, |instance_name| in |zone| that might not be
ready."""
for _ in range(10):
result = ssh(instance_name,
zone=zone,
command='echo ping',
expect_zero=False)
if result.retcode == 0:
return
logs.info('GCP instance isn\'t ready yet. Rerunning SSH momentarily.')
time.sleep(5)
raise Exception('Couldn\'t SSH to instance.')
def dispatcher_main():
"""Do the experiment and report results."""
logs.info('Starting experiment.')
# Set this here because we get failures if we do it in measurer for some
# reason.
multiprocessing.set_start_method('spawn')
db_utils.initialize()
if os.getenv('LOCAL_EXPERIMENT'):
models.Base.metadata.create_all(db_utils.engine)
experiment_config_file_path = os.path.join(fuzzer_config_utils.get_dir(),
'experiment.yaml')
experiment = Experiment(experiment_config_file_path)
preemptible = experiment.preemptible
trials = build_images_for_trials(experiment.fuzzers, experiment.benchmarks,
experiment.num_trials, preemptible)
_initialize_experiment_in_db(experiment.experiment_name,
experiment.git_hash, trials)
create_work_subdirs(['experiment-folders', 'measurement-folders'])
# Start measurer and scheduler in seperate threads/processes.
scheduler_loop_thread = threading.Thread(target=scheduler.schedule_loop,
args=(experiment.config, ))
scheduler_loop_thread.start()
max_total_time = experiment.config['max_total_time']
measurer_loop_process = multiprocessing.Process(
target=measurer.measure_loop,
args=(experiment.experiment_name, max_total_time))
measurer_loop_process.start()
is_complete = False
while True:
time.sleep(LOOP_WAIT_SECONDS)
if not scheduler_loop_thread.is_alive():
is_complete = not measurer_loop_process.is_alive()
# Generate periodic output reports.
reporter.output_report(experiment.web_bucket,
in_progress=not is_complete)
if is_complete:
# Experiment is complete, bail out.
break
logs.info('Dispatcher finished.')
scheduler_loop_thread.join()
measurer_loop_process.join()
def dispatcher_main():
"""Do the experiment and report results."""
logs.info('Starting experiment.')
# Set this here because we get failures if we do it in measurer for some
# reason.
multiprocessing.set_start_method('spawn')
db_utils.initialize()
if os.getenv('LOCAL_EXPERIMENT'):
models.Base.metadata.create_all(db_utils.engine)
builder.build_base_images()
experiment_config_file_path = os.path.join(fuzzer_config_utils.get_dir(),
'experiment.yaml')
experiment = Experiment(experiment_config_file_path)
# When building, we only care about the underlying fuzzer rather than the
# display name that we use to identify a specific configuration.
unique_fuzzers = list({
fuzzer_config_utils.get_underlying_fuzzer_name(f)
for f in experiment.fuzzers
})
builder.build_all_fuzzer_benchmarks(unique_fuzzers, experiment.benchmarks)
create_work_subdirs(['experiment-folders', 'measurement-folders'])
# Start measurer and scheduler in threads.
scheduler_loop_thread = threading.Thread(target=scheduler.schedule_loop,
args=(experiment.config, ))
scheduler_loop_thread.start()
measurer_loop_thread = multiprocessing.Process(
target=measurer.measure_loop,
args=(
experiment.config['experiment'],
experiment.config['max_total_time'],
))
measurer_loop_thread.start()
while True:
time.sleep(LOOP_WAIT_SECONDS)
is_complete = (not scheduler_loop_thread.is_alive()
and not measurer_loop_thread.is_alive())
# Generate periodic output reports.
reporter.output_report(experiment.web_bucket,
in_progress=not is_complete)
if is_complete:
# Experiment is complete, bail out.
break
def process_crashes(self, cycle):
"""Process and store crashes."""
if not os.listdir(self.crashes_dir):
logs.info('No crashes found for cycle %d.', cycle)
return []
logs.info('Saving crash files crashes for cycle %d.', cycle)
self.save_crash_files(cycle)
logs.info('Processing crashes for cycle %d.', cycle)
app_binary = coverage_utils.get_coverage_binary(self.benchmark)
crash_metadata = run_crashes.do_crashes_run(app_binary,
self.crashes_dir)
logs.info('Crash metadata: %s', crash_metadata)
crashes = []
for crash_key in crash_metadata:
crash = crash_metadata[crash_key]
crashes.append(
models.Crash(crash_key=crash_key,
crash_testcase=crash.crash_testcase,
crash_type=crash.crash_type,
crash_address=crash.crash_address,
crash_state=crash.crash_state,
crash_stacktrace=crash.crash_stacktrace))
return crashes
def build_fuzzer_benchmark(fuzzer: str, benchmark: str) -> bool:
"""Builds |benchmark| for |fuzzer|."""
logger.info('Building benchmark: %s, fuzzer: %s.', benchmark, fuzzer)
try:
if benchmark_utils.is_oss_fuzz(benchmark):
gcb_build_oss_fuzz_project_fuzzer(benchmark, fuzzer)
else:
gcb_build_benchmark_fuzzer(benchmark, fuzzer)
except subprocess.CalledProcessError:
logger.error('Failed to build benchmark: %s, fuzzer: %s.', benchmark,
fuzzer)
return False
logs.info('Done building benchmark: %s, fuzzer: %s.', benchmark, fuzzer)
return True
def retry_build_loop(build_func: Callable, inputs: List[Tuple],
num_concurrent_builds: int) -> List:
"""Calls |build_func| in parallel on |inputs|. Repeat on failures up to
|NUM_BUILD_RETRIES| times. Returns the list of inputs that |build_func| was
called successfully on."""
successes = []
logs.info('Concurrent builds: %d.', num_concurrent_builds)
with mp_pool.ThreadPool(num_concurrent_builds) as pool:
for _ in range(NUM_BUILD_RETRIES):
logs.info('Building using (%s): %s', build_func, inputs)
results = pool.starmap(build_func, inputs)
curr_successes, curr_failures = split_successes_and_failures(
inputs, results)
logs.info('Build successes: %s', curr_successes)
successes.extend(curr_successes)
if not curr_failures:
break
logs.error('Build failures: %s', curr_failures)
inputs = curr_failures
sleep_interval = random.uniform(1, BUILD_FAIL_WAIT)
logs.info('Sleeping for %d secs before retrying.', sleep_interval)
time.sleep(sleep_interval)
return successes
def handle_retry(num_try, exception=None):
"""Handle retry."""
if (exception is None or
isinstance(exception, exception_type)) and num_try < tries:
logs.info('Retrying on %s failed with %s. Retrying again.',
function_with_type,
sys.exc_info()[1])
sleep(get_delay(num_try, delay, backoff))
return True
logs.error('Retrying on %s failed with %s. Raise.',
function_with_type,
sys.exc_info()[1])
return False
def run_requested_experiment(dry_run):
"""Run the oldest requested experiment that hasn't been run yet in
experiment-requests.yaml."""
requested_experiments = _get_requested_experiments()
# TODO(metzman): Look into supporting benchmarks as an optional parameter so
# that people can add fuzzers that don't support everything.
if PAUSE_SERVICE_KEYWORD in requested_experiments:
# Check if automated experiment service is paused.
logs.warning('Pause service requested, not running experiment.')
return None
requested_experiment = None
for experiment_config in reversed(requested_experiments):
experiment_name = _get_experiment_name(experiment_config)
with db_utils.session_scope() as session:
is_new_experiment = session.query(models.Experiment).filter(
models.Experiment.name == experiment_name).first() is None
if is_new_experiment:
requested_experiment = experiment_config
break
if requested_experiment is None:
logs.info('No new experiment to run. Exiting.')
return None
experiment_name = _get_experiment_name(requested_experiment)
if not validate_experiment_requests([requested_experiment]):
logs.error('Requested experiment: %s in %s is not valid.',
requested_experiment, REQUESTED_EXPERIMENTS_PATH)
return None
fuzzers = requested_experiment['fuzzers']
benchmark_type = requested_experiment.get('type')
if benchmark_type == benchmark_utils.BenchmarkType.BUG.value:
benchmarks = benchmark_utils.get_bug_benchmarks()
else:
benchmarks = (benchmark_utils.get_oss_fuzz_coverage_benchmarks() +
benchmark_utils.get_standard_coverage_benchmarks())
logs.info('Running experiment: %s with fuzzers: %s.', experiment_name,
' '.join(fuzzers))
description = _get_description(requested_experiment)
oss_fuzz_corpus = _use_oss_fuzz_corpus(requested_experiment)
return _run_experiment(experiment_name, fuzzers, benchmarks, description,
oss_fuzz_corpus, dry_run)
def archive_crashes(self, cycle):
"""Archive this cycle's crashes into cloud bucket."""
if not os.listdir(self.crashes_dir):
logs.info('No crashes found for cycle %d.', cycle)
return
logs.info('Archiving crashes for cycle %d.', cycle)
crashes_archive_name = experiment_utils.get_crashes_archive_name(cycle)
archive = os.path.join(os.path.dirname(self.crashes_dir),
crashes_archive_name)
with tarfile.open(archive, 'w:gz') as tar:
tar.add(self.crashes_dir,
arcname=os.path.basename(self.crashes_dir))
gcs_path = exp_path.gcs(
posixpath.join(self.trial_dir, 'crashes', crashes_archive_name))
gsutil.cp(archive, gcs_path)
os.remove(archive)
def _unpack_clusterfuzz_seed_corpus(fuzz_target_path, corpus_directory):
"""If a clusterfuzz seed corpus archive is available, unpack it into the
corpus directory if it exists. Copied from unpack_seed_corpus in
engine_common.py in ClusterFuzz.
"""
oss_fuzz_corpus = environment.get('OSS_FUZZ_CORPUS')
if oss_fuzz_corpus:
benchmark = environment.get('BENCHMARK')
corpus_archive_filename = f'{benchmark}.zip'
oss_fuzz_corpus_archive_path = posixpath.join(
experiment_utils.get_oss_fuzz_corpora_filestore_path(),
corpus_archive_filename)
seed_corpus_archive_path = posixpath.join(FUZZ_TARGET_DIR,
corpus_archive_filename)
filestore_utils.cp(oss_fuzz_corpus_archive_path,
seed_corpus_archive_path)
else:
seed_corpus_archive_path = get_clusterfuzz_seed_corpus_path(
fuzz_target_path)
if not seed_corpus_archive_path:
return
with zipfile.ZipFile(seed_corpus_archive_path) as zip_file:
# Unpack seed corpus recursively into the root of the main corpus
# directory.
idx = 0
for seed_corpus_file in zip_file.infolist():
if seed_corpus_file.filename.endswith('/'):
# Ignore directories.
continue
# Allow callers to opt-out of unpacking large files.
if seed_corpus_file.file_size > CORPUS_ELEMENT_BYTES_LIMIT:
continue
output_filename = '%016d' % idx
output_file_path = os.path.join(corpus_directory, output_filename)
zip_file.extract(seed_corpus_file, output_file_path)
idx += 1
logs.info('Unarchived %d files from seed corpus %s.', idx,
seed_corpus_archive_path)
def _run_experiment( # pylint: disable=too-many-arguments
experiment_name,
fuzzers,
benchmarks,
description,
oss_fuzz_corpus,
dry_run=False):
"""Run an experiment named |experiment_name| on |fuzzer_configs| and shut it
down once it terminates."""
logs.info('Starting experiment: %s.', experiment_name)
if dry_run:
logs.info('Dry run. Not actually running experiment.')
return
run_experiment.start_experiment(experiment_name,
EXPERIMENT_CONFIG_FILE,
benchmarks,
fuzzers,
description=description,
oss_fuzz_corpus=oss_fuzz_corpus)
def conduct_trial(self):
"""Conduct the benchmarking trial."""
self.initialize_directories()
log_file = os.path.join(self.results_dir, 'fuzzer-log.txt')
logs.info('Starting trial.')
max_total_time = environment.get('MAX_TOTAL_TIME')
args = (max_total_time, log_file)
thread = threading.Thread(target=run_fuzzer, args=args)
thread.start()
while thread.is_alive():
self.sleep_until_next_sync()
self.do_sync()
self.cycle += 1
logs.info('Doing final sync.')
self.do_sync(final_sync=True)
thread.join()
def run_requested_experiment(dry_run):
"""Run the oldest requested experiment that hasn't been run yet in
experiment-requests.yaml."""
requested_experiments = _get_requested_experiments()
# TODO(metzman): Look into supporting benchmarks as an optional parameter so
# that people can add fuzzers that don't support everything.
if PAUSE_SERVICE_KEYWORD in requested_experiments:
# Check if automated experiment service is paused.
logs.warning('Pause service requested, not running experiment.')
return None
requested_experiment = None
for experiment_config in reversed(requested_experiments):
experiment_name = _get_experiment_name(experiment_config)
is_new_experiment = db_utils.query(models.Experiment).filter(
models.Experiment.name == experiment_name).first() is None
if is_new_experiment:
requested_experiment = experiment_config
break
if requested_experiment is None:
logs.info('No new experiment to run. Exiting.')
return None
experiment_name = _get_experiment_name(requested_experiment)
if not validate_experiment_requests([requested_experiment]):
logs.error('Requested experiment: %s in %s is not valid.',
requested_experiment, REQUESTED_EXPERIMENTS_PATH)
return None
fuzzers = requested_experiment['fuzzers']
logs.info('Running experiment: %s with fuzzers: %s.', experiment_name,
' '.join(fuzzers))
fuzzer_configs = fuzzer_utils.get_fuzzer_configs(fuzzers=fuzzers)
return _run_experiment(experiment_name, fuzzer_configs, dry_run)
def schedule(experiment_config: dict, queue):
"""Schedule measurer workers. This cannot be called before
initialize_measurers."""
logger.info('Scheduling measurer workers.')
# TODO(metzman): This method doesn't seem to correctly take into account
# jobs that are running (the API provided by rq doesn't work intuitively).
# That is OK for now since scheduling only happens while nothing is being
# measured but this should be fixed.
jobs = queue_utils.get_all_jobs(queue)
counts = collections.defaultdict(int)
for job in jobs:
counts[job.get_status(refresh=False)] += 1
num_instances_needed = counts['queued'] + counts['started']
num_instances_needed = min(num_instances_needed, MAX_INSTANCES_PER_GROUP)
logger.info('Scheduling %d workers.', num_instances_needed)
instance_group_name = get_instance_group_name(
experiment_config['experiment'])
project = experiment_config['cloud_project']
zone = experiment_config['cloud_compute_zone']
num_instances = gce.get_instance_group_size(instance_group_name, project,
zone)
# TODO(metzman): Use autoscaling as it probably can deal with quotas more
# easily.
if not num_instances_needed:
# Can't go below 1 instance per group.
logs.info('num_instances_needed = 0, resizing to 1.')
num_instances_needed = 1
if num_instances_needed != num_instances:
# TODO(metzman): Add some limits so always have some measurers but not
# too many.
gce.resize_instance_group(num_instances_needed, instance_group_name,
project, zone)
