v8/tools/testrunner/local/pool.py
Michael Achenbach 4a2d9b1b20 [test] Abort testing immediately on interrupt and sigterm
This is a partial revert of:
https://crrev.com/c/890938 and https://crrev.com/c/893982

Before this CL, the test runner blocked on ongoing tests in order to
process their results after an internal timeout. However, the logic
required for this feature was overly complicated and prevented an
acceptable implementation for fast aborts. Furthermore, also the fuzzers
suffered from timeouts on swarming due to hanging tests.

Instead, we now abort immediately on internal timeout (used on
fuzzers), SIGINT (Ctrl-C) and SIGTERM. Ongoing tests are immediately
terminated and their results are disregarded. On SIGTERM and SIGINT,
we return with non-zero exit codes, and zero on internal timeout.

This will also properly return json output, when the external hard
timeout is reached on swarming (causes SIGTERM).

TBR=sergiyb@chromium.org

Bug: v8:7423, chromium:813065
Change-Id: Ib20f835f58a0970693bdd3b21dc5d766d8e115d8
Reviewed-on: https://chromium-review.googlesource.com/924852
Reviewed-by: Michael Achenbach <machenbach@chromium.org>
Commit-Queue: Michael Achenbach <machenbach@chromium.org>
Cr-Commit-Position: refs/heads/master@{#51399}
2018-02-20 18:02:25 +00:00

282 lines
8.8 KiB
Python

#!/usr/bin/env python
# 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.
from Queue import Empty
from contextlib import contextmanager
from multiprocessing import Process, Queue
import os
import signal
import time
import traceback
from . import command
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 threading Queue to look like multiprocessing Queue.
Queue.cancel_join_thread = lambda self: None
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.
"""
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"):
try:
done_queue.put(NormalResult(fn(*args, **kwargs)))
except command.AbortException:
# SIGINT, SIGTERM or internal hard timeout.
break
except Exception, e:
traceback.print_exc()
print(">>> EXCEPTION: %s" % e)
done_queue.put(ExceptionResult(e))
# When we reach here on normal tear down, all items have been pulled from
# the done_queue before and this should have no effect. On fast abort, it's
# possible that a fast worker left items on the done_queue in memory, which
# will never be pulled. This call purges those to avoid a deadlock.
done_queue.cancel_join_thread()
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)
class Pool():
"""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, num_workers, heartbeat_timeout=1):
self.num_workers = num_workers
self.processes = []
self.terminated = False
self.abort_now = 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
self.heartbeat_timeout = heartbeat_timeout
# Disable sigint and sigterm to prevent subprocesses from capturing the
# signals.
with without_sig():
self.work_queue = Queue()
self.done_queue = Queue()
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 xrange(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.
internal_error = True
continue
if self.abort_now:
# SIGINT, SIGTERM or internal hard timeout.
return
yield result
break
self.advance(gen)
except KeyboardInterrupt:
assert False, 'Unreachable'
except Exception as e:
traceback.print_exc()
print(">>> EXCEPTION: %s" % e)
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(gen.next())
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(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")
if self.abort_now:
for p in self.processes:
os.kill(p.pid, signal.SIGTERM)
for p in self.processes:
p.join()
# Drain the queues to prevent stderr chatter when queues are garbage
# collected.
try:
while True: self.work_queue.get(False)
except:
pass
try:
while True: self.done_queue.get(False)
except:
pass
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()