v8/tools/testrunner/local/pool.py
Michael Achenbach ae2ef7d234 [test] Drain queues asynchroneously when terminating workers
Joining a queue-using process can deadlock if the child process is
about to write to the queue, but the parent process wants to join the
child. To fix this, we now drain elements from a separate thread of
the main process.

Bug: v8:13113
Change-Id: Ic279e66ab84eb89a4034ff1f2c025eb850b65013
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/3891116
Commit-Queue: Michael Achenbach <machenbach@chromium.org>
Reviewed-by: Alexander Schulze <alexschulze@chromium.org>
Cr-Commit-Position: refs/heads/main@{#83177}
2022-09-14 05:58:47 +00:00

402 lines
12 KiB
Python

#!/usr/bin/env python3
# Copyright 2014 the V8 project authors. All rights reserved.
# Use of this source code is governed by a BSD-style license that can be
# found in the LICENSE file.
import collections
import logging
import os
import signal
import threading
import traceback
from contextlib import contextmanager
from multiprocessing import Process, Queue
from queue import Empty
def setup_testing():
"""For testing only: Use threading under the hood instead of multiprocessing
to make coverage work.
"""
global Queue
global Process
del Queue
del Process
from queue import Queue
from threading import Thread as Process
# Monkeypatch os.kill and add fake pid property on Thread.
os.kill = lambda *args: None
Process.pid = property(lambda self: None)
class AbortException(Exception):
"""Indicates early abort on SIGINT, SIGTERM or internal hard timeout."""
pass
class NormalResult():
def __init__(self, result):
self.result = result
self.exception = None
class ExceptionResult():
def __init__(self, exception):
self.exception = exception
class MaybeResult():
def __init__(self, heartbeat, value):
self.heartbeat = heartbeat
self.value = value
@staticmethod
def create_heartbeat():
return MaybeResult(True, None)
@staticmethod
def create_result(value):
return MaybeResult(False, value)
def Worker(fn, work_queue, done_queue,
process_context_fn=None, process_context_args=None):
"""Worker to be run in a child process.
The worker stops when the poison pill "STOP" is reached.
"""
# Install a default signal handler for SIGTERM that stops the processing
# loop below on the next occasion. The job function "fn" is supposed to
# register their own handler to avoid blocking, but still chain to this
# handler on SIGTERM to terminate the loop quickly.
stop = [False]
def handler(signum, frame):
stop[0] = True
signal.signal(signal.SIGTERM, handler)
try:
kwargs = {}
if process_context_fn and process_context_args is not None:
kwargs.update(process_context=process_context_fn(*process_context_args))
for args in iter(work_queue.get, "STOP"):
if stop[0]:
# SIGINT, SIGTERM or internal hard timeout caught outside the execution
# of "fn".
break
try:
done_queue.put(NormalResult(fn(*args, **kwargs)))
except AbortException:
# SIGINT, SIGTERM or internal hard timeout caught during execution of
# "fn".
break
except Exception as e:
logging.exception('Unhandled error during worker execution.')
done_queue.put(ExceptionResult(e))
except KeyboardInterrupt:
assert False, 'Unreachable'
@contextmanager
def without_sig():
int_handler = signal.signal(signal.SIGINT, signal.SIG_IGN)
term_handler = signal.signal(signal.SIGTERM, signal.SIG_IGN)
try:
yield
finally:
signal.signal(signal.SIGINT, int_handler)
signal.signal(signal.SIGTERM, term_handler)
@contextmanager
def drain_queue_async(queue):
"""Drains a queue in a background thread until the wrapped code unblocks.
This can be used to unblock joining a child process that might still write
to the queue. The join should be wrapped by this context manager.
"""
keep_running = True
def empty_queue():
elem_count = 0
while keep_running:
try:
while True:
queue.get(True, 0.1)
elem_count += 1
if elem_count < 200:
logging.info('Drained an element from queue.')
except Empty:
pass
except:
logging.exception('Error draining queue.')
emptier = threading.Thread(target=empty_queue)
emptier.start()
yield
keep_running = False
emptier.join()
class ContextPool():
def __init__(self):
self.abort_now = False
def init(self, num_workers, heartbeat_timeout=1, notify_function=None):
"""
Delayed initialization. At context creation time we have no access to the
below described parameters.
Args:
num_workers: Number of worker processes to run in parallel.
heartbeat_timeout: Timeout in seconds for waiting for results. Each time
the timeout is reached, a heartbeat is signalled and timeout is reset.
notify_function: Callable called to signal some events like termination. The
event name is passed as string.
"""
pass
def add_jobs(self, jobs):
pass
def results(self, requirement):
pass
def abort(self):
self.abort_now = True
ProcessContext = collections.namedtuple('ProcessContext', ['result_reduction'])
class DefaultExecutionPool(ContextPool):
"""Distributes tasks to a number of worker processes.
New tasks can be added dynamically even after the workers have been started.
Requirement: Tasks can only be added from the parent process, e.g. while
consuming the results generator."""
# Factor to calculate the maximum number of items in the work/done queue.
# Necessary to not overflow the queue's pipe if a keyboard interrupt happens.
BUFFER_FACTOR = 4
def __init__(self, os_context=None):
super(DefaultExecutionPool, self).__init__()
self.os_context = os_context
self.processes = []
self.terminated = False
# Invariant: processing_count >= #work_queue + #done_queue. It is greater
# when a worker takes an item from the work_queue and before the result is
# submitted to the done_queue. It is equal when no worker is working,
# e.g. when all workers have finished, and when no results are processed.
# Count is only accessed by the parent process. Only the parent process is
# allowed to remove items from the done_queue and to add items to the
# work_queue.
self.processing_count = 0
# Disable sigint and sigterm to prevent subprocesses from capturing the
# signals.
with without_sig():
self.work_queue = Queue()
self.done_queue = Queue()
def init(self, num_workers=1, heartbeat_timeout=1, notify_function=None):
"""
Args:
num_workers: Number of worker processes to run in parallel.
heartbeat_timeout: Timeout in seconds for waiting for results. Each time
the timeout is reached, a heartbeat is signalled and timeout is reset.
notify_function: Callable called to signal some events like termination. The
event name is passed as string.
"""
self.num_workers = num_workers
self.heartbeat_timeout = heartbeat_timeout
self.notify = notify_function or (lambda x: x)
def add_jobs(self, jobs):
self.add(jobs)
def results(self, requirement):
return self.imap_unordered(
fn=run_job,
gen=[],
process_context_fn=ProcessContext,
process_context_args=[requirement],
)
def imap_unordered(self, fn, gen,
process_context_fn=None, process_context_args=None):
"""Maps function "fn" to items in generator "gen" on the worker processes
in an arbitrary order. The items are expected to be lists of arguments to
the function. Returns a results iterator. A result value of type
MaybeResult either indicates a heartbeat of the runner, i.e. indicating
that the runner is still waiting for the result to be computed, or it wraps
the real result.
Args:
process_context_fn: Function executed once by each worker. Expected to
return a process-context object. If present, this object is passed
as additional argument to each call to fn.
process_context_args: List of arguments for the invocation of
process_context_fn. All arguments will be pickled and sent beyond the
process boundary.
"""
if self.terminated:
return
try:
internal_error = False
gen = iter(gen)
self.advance = self._advance_more
# Disable sigint and sigterm to prevent subprocesses from capturing the
# signals.
with without_sig():
for w in range(self.num_workers):
p = Process(target=Worker, args=(fn,
self.work_queue,
self.done_queue,
process_context_fn,
process_context_args))
p.start()
self.processes.append(p)
self.advance(gen)
while self.processing_count > 0:
while True:
try:
# Read from result queue in a responsive fashion. If available,
# this will return a normal result immediately or a heartbeat on
# heartbeat timeout (default 1 second).
result = self._get_result_from_queue()
except:
# TODO(machenbach): Handle a few known types of internal errors
# gracefully, e.g. missing test files.
logging.exception('Internal error in a worker process.')
internal_error = True
continue
finally:
if self.abort_now:
# SIGINT, SIGTERM or internal hard timeout.
return
yield result
break
self.advance(gen)
except KeyboardInterrupt:
assert False, 'Unreachable'
except Exception:
logging.exception('Unhandled error during pool execution.')
finally:
self._terminate()
if internal_error:
raise Exception('Internal error in a worker process.')
def _advance_more(self, gen):
while self.processing_count < self.num_workers * self.BUFFER_FACTOR:
try:
self.work_queue.put(next(gen))
self.processing_count += 1
except StopIteration:
self.advance = self._advance_empty
break
def _advance_empty(self, gen):
pass
def add(self, args):
"""Adds an item to the work queue. Can be called dynamically while
processing the results from imap_unordered."""
assert not self.terminated
self.work_queue.put(args)
self.processing_count += 1
def abort(self):
"""Schedules abort on next queue read.
This is safe to call when handling SIGINT, SIGTERM or when an internal
hard timeout is reached.
"""
self.abort_now = True
def _terminate_processes(self):
for p in self.processes:
self.os_context.terminate_process(p)
def _terminate(self):
"""Terminates execution and cleans up the queues.
If abort() was called before termination, this also terminates the
subprocesses and doesn't wait for ongoing tests.
"""
if self.terminated:
return
self.terminated = True
# Drain out work queue from tests
try:
while True:
self.work_queue.get(True, 0.1)
except Empty:
pass
# Make sure all processes stop
for _ in self.processes:
# During normal tear down the workers block on get(). Feed a poison pill
# per worker to make them stop.
self.work_queue.put("STOP")
# Send a SIGTERM to all workers. They will gracefully terminate their
# processing loop and if the signal is caught during job execution they
# will try to terminate the ongoing test processes quickly.
if self.abort_now:
self._terminate_processes()
self.notify("Joining workers")
with drain_queue_async(self.done_queue):
for p in self.processes:
p.join()
self.notify("Pool terminated")
def _get_result_from_queue(self):
"""Attempts to get the next result from the queue.
Returns: A wrapped result if one was available within heartbeat timeout,
a heartbeat result otherwise.
Raises:
Exception: If an exception occured when processing the task on the
worker side, it is reraised here.
"""
while True:
try:
result = self.done_queue.get(timeout=self.heartbeat_timeout)
self.processing_count -= 1
if result.exception:
raise result.exception
return MaybeResult.create_result(result.result)
except Empty:
return MaybeResult.create_heartbeat()
class SingleThreadedExecutionPool(ContextPool):
def __init__(self):
super(SingleThreadedExecutionPool, self).__init__()
self.work_queue = []
def add_jobs(self, jobs):
self.work_queue.extend(jobs)
def results(self, requirement):
while self.work_queue and not self.abort_now:
job = self.work_queue.pop()
yield MaybeResult.create_result(job.run(ProcessContext(requirement)))
# Global function for multiprocessing, because pickling a static method doesn't
# work on Windows.
def run_job(job, process_context):
return job.run(process_context)