Distinctly separate and harden mp spawning

It's clear now that special attention is needed to handle the case where
a spawned `multiprocessing` proc is started but then the parent is
cancelled before the child can connect back; in this case we need to be
sure to kill the near-zombie child asap. This may end up being the
solution to other resiliency issues seen around mp with nested process
trees too. More testing is needed to be sure.

Relates to    
mp_teardown_hardening
Tyler Goodlet 2021-04-28 22:11:59 -04:00
parent af93b8532a
commit 607c48f1ac
1 changed files with 152 additions and 74 deletions
tractor

View File

@ -2,14 +2,13 @@
Machinery for actor process spawning using multiple backends.
"""
import sys
import inspect
import multiprocessing as mp
import platform
from typing import Any, Dict, Optional
import trio
from trio_typing import TaskStatus
from async_generator import aclosing, asynccontextmanager
from async_generator import asynccontextmanager
try:
from multiprocessing import semaphore_tracker # type: ignore
@ -128,7 +127,9 @@ async def cancel_on_completion(
Should only be called for actors spawned with `run_in_actor()`.
"""
with trio.CancelScope() as cs:
task_status.started(cs)
# if this call errors we store the exception for later
# in ``errors`` which will be reraised inside
# a MultiError and we still send out a cancel request
@ -138,6 +139,7 @@ async def cancel_on_completion(
log.warning(
f"Cancelling {portal.channel.uid} after error {result}"
)
else:
log.info(
f"Cancelling {portal.channel.uid} gracefully "
@ -202,7 +204,7 @@ async def spawn_subactor(
async def new_proc(
name: str,
actor_nursery: 'ActorNursery', # type: ignore
actor_nursery: 'ActorNursery', # type: ignore # noqa
subactor: Actor,
errors: Dict[Tuple[str, str], Exception],
# passed through to actor main
@ -221,8 +223,8 @@ async def new_proc(
# mark the new actor with the global spawn method
subactor._spawn_method = _spawn_method
async with trio.open_nursery() as nursery:
if use_trio_run_in_process or _spawn_method == 'trio':
if use_trio_run_in_process or _spawn_method == 'trio':
async with trio.open_nursery() as nursery:
async with spawn_subactor(
subactor,
parent_addr,
@ -261,7 +263,11 @@ async def new_proc(
if portal in actor_nursery._cancel_after_result_on_exit:
cancel_scope = await nursery.start(
cancel_on_completion, portal, subactor, errors)
cancel_on_completion,
portal,
subactor,
errors
)
# Wait for proc termination but **dont' yet** call
# ``trio.Process.__aexit__()`` (it tears down stdio
@ -275,65 +281,107 @@ async def new_proc(
# no actor zombies allowed
# with trio.CancelScope(shield=True):
await proc.wait()
else:
# `multiprocessing`
assert _ctx
start_method = _ctx.get_start_method()
if start_method == 'forkserver':
# XXX do our hackery on the stdlib to avoid multiple
# forkservers (one at each subproc layer).
fs = forkserver._forkserver
curr_actor = current_actor()
if is_main_process() and not curr_actor._forkserver_info:
# if we're the "main" process start the forkserver
# only once and pass its ipc info to downstream
# children
# forkserver.set_forkserver_preload(enable_modules)
forkserver.ensure_running()
fs_info = (
fs._forkserver_address,
fs._forkserver_alive_fd,
getattr(fs, '_forkserver_pid', None),
getattr(
resource_tracker._resource_tracker, '_pid', None),
resource_tracker._resource_tracker._fd,
)
else:
assert curr_actor._forkserver_info
fs_info = (
fs._forkserver_address,
fs._forkserver_alive_fd,
fs._forkserver_pid,
resource_tracker._resource_tracker._pid,
resource_tracker._resource_tracker._fd,
) = curr_actor._forkserver_info
log.debug(f"Joined {proc}")
# pop child entry to indicate we are no longer managing this subactor
subactor, proc, portal = actor_nursery._children.pop(subactor.uid)
# cancel result waiter that may have been spawned in
# tandem if not done already
if cancel_scope:
log.warning(
f"Cancelling existing result waiter task for {subactor.uid}")
cancel_scope.cancel()
else:
# `multiprocessing`
# async with trio.open_nursery() as nursery:
await mp_new_proc(
name=name,
actor_nursery=actor_nursery,
subactor=subactor,
errors=errors,
# passed through to actor main
bind_addr=bind_addr,
parent_addr=parent_addr,
_runtime_vars=_runtime_vars,
task_status=task_status,
)
async def mp_new_proc(
name: str,
actor_nursery: 'ActorNursery', # type: ignore # noqa
subactor: Actor,
errors: Dict[Tuple[str, str], Exception],
# passed through to actor main
bind_addr: Tuple[str, int],
parent_addr: Tuple[str, int],
_runtime_vars: Dict[str, Any], # serialized and sent to _child
*,
use_trio_run_in_process: bool = False,
task_status: TaskStatus[Portal] = trio.TASK_STATUS_IGNORED
) -> None:
async with trio.open_nursery() as nursery:
assert _ctx
start_method = _ctx.get_start_method()
if start_method == 'forkserver':
# XXX do our hackery on the stdlib to avoid multiple
# forkservers (one at each subproc layer).
fs = forkserver._forkserver
curr_actor = current_actor()
if is_main_process() and not curr_actor._forkserver_info:
# if we're the "main" process start the forkserver
# only once and pass its ipc info to downstream
# children
# forkserver.set_forkserver_preload(enable_modules)
forkserver.ensure_running()
fs_info = (
fs._forkserver_address,
fs._forkserver_alive_fd,
getattr(fs, '_forkserver_pid', None),
getattr(
resource_tracker._resource_tracker, '_pid', None),
resource_tracker._resource_tracker._fd,
)
else:
fs_info = (None, None, None, None, None)
assert curr_actor._forkserver_info
fs_info = (
fs._forkserver_address,
fs._forkserver_alive_fd,
fs._forkserver_pid,
resource_tracker._resource_tracker._pid,
resource_tracker._resource_tracker._fd,
) = curr_actor._forkserver_info
else:
fs_info = (None, None, None, None, None)
proc = _ctx.Process( # type: ignore
target=_mp_main,
args=(
subactor,
bind_addr,
fs_info,
start_method,
parent_addr,
),
# daemon=True,
name=name,
)
# `multiprocessing` only (since no async interface):
# register the process before start in case we get a cancel
# request before the actor has fully spawned - then we can wait
# for it to fully come up before sending a cancel request
actor_nursery._children[subactor.uid] = (subactor, proc, None)
proc = _ctx.Process( # type: ignore
target=_mp_main,
args=(
subactor,
bind_addr,
fs_info,
start_method,
parent_addr,
),
# daemon=True,
name=name,
)
# `multiprocessing` only (since no async interface):
# register the process before start in case we get a cancel
# request before the actor has fully spawned - then we can wait
# for it to fully come up before sending a cancel request
actor_nursery._children[subactor.uid] = (subactor, proc, None)
proc.start()
if not proc.is_alive():
raise ActorFailure("Couldn't start sub-actor?")
proc.start()
if not proc.is_alive():
raise ActorFailure("Couldn't start sub-actor?")
log.info(f"Started {proc}")
log.info(f"Started {proc}")
try:
# wait for actor to spawn and connect back to us
# channel should have handshake completed by the
# local actor by the time we get a ref to it
@ -356,23 +404,53 @@ async def new_proc(
# awaited and reported upwards to the supervisor.
await actor_nursery._join_procs.wait()
finally:
# XXX: in the case we were cancelled before the sub-proc
# registered itself back we must be sure to try and clean
# any process we may have started.
reaping_cancelled = False
cancel_scope = None
if portal in actor_nursery._cancel_after_result_on_exit:
cancel_scope = await nursery.start(
cancel_on_completion, portal, subactor, errors)
try:
# async with trio.open_nursery() as n:
# n.cancel_scope.shield = True
cancel_scope = await nursery.start(
cancel_on_completion,
portal,
subactor,
errors
)
except trio.Cancelled:
# if the reaping task was cancelled we may have hit
# a race where the subproc disconnected before we
# could send it a message to cancel (classic 2 generals)
# in that case, wait shortly then kill the process.
reaping_cancelled = True
if proc.is_alive():
with trio.move_on_after(0.1) as cs:
cs.shield = True
await proc_waiter(proc)
if cs.cancelled_caught:
proc.terminate()
if not reaping_cancelled:
if proc.is_alive():
await proc_waiter(proc)
# TODO: timeout block here?
if proc.is_alive():
await proc_waiter(proc)
proc.join()
# This is again common logic for all backends:
log.debug(f"Joined {proc}")
# pop child entry to indicate we are no longer managing this subactor
subactor, proc, portal = actor_nursery._children.pop(subactor.uid)
log.debug(f"Joined {proc}")
# pop child entry to indicate we are no longer managing this subactor
subactor, proc, portal = actor_nursery._children.pop(subactor.uid)
# cancel result waiter that may have been spawned in
# tandem if not done already
if cancel_scope:
log.warning(
f"Cancelling existing result waiter task for {subactor.uid}")
cancel_scope.cancel()
# cancel result waiter that may have been spawned in
# tandem if not done already
if cancel_scope:
log.warning(
f"Cancelling existing result waiter task for {subactor.uid}")
cancel_scope.cancel()