jc211
/
tractor
forked from goodboy/tractor
1
0
Fork 0
Code Pull Requests Activity

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 #84 #89 #134 #146
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
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

226
tractor/_spawn.py
View File

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