weakref spewing exceptions during finalization when combined with multiprocessing

[ambv]

BPO	29519
Nosy	@pitrou, @methane, @ambv
PRs	[2.7] bpo-29519: weakref spewing exceptions during interp finalization #2958
Files	mod1.py mod2.py

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields:

assignee = 'https://github.com/ambv'
closed_at = <Date 2017-02-10.08:22:49.348>
created_at = <Date 2017-02-10.04:35:06.699>
labels = ['3.7', 'type-bug', 'library']
title = 'weakref spewing exceptions during finalization when combined with multiprocessing'
updated_at = <Date 2017-12-03.16:18:39.687>
user = 'https://github.com/ambv'

bugs.python.org fields:

activity = <Date 2017-12-03.16:18:39.687>
actor = 'Warwick Chapman'
assignee = 'lukasz.langa'
closed = True
closed_date = <Date 2017-02-10.08:22:49.348>
closer = 'lukasz.langa'
components = ['Library (Lib)']
creation = <Date 2017-02-10.04:35:06.699>
creator = 'lukasz.langa'
dependencies = []
files = ['46617', '46618']
hgrepos = []
issue_num = 29519
keywords = ['3.5regression', '3.6regression']
message_count = 12.0
messages = ['287475', '287479', '287482', '287483', '287487', '287488', '287489', '299550', '299572', '307487', '307516', '307518']
nosy_count = 5.0
nosy_names = ['pitrou', 'methane', 'lukasz.langa', 'python-dev', 'Warwick Chapman']
pr_nums = ['2958']
priority = 'normal'
resolution = 'fixed'
stage = 'resolved'
status = 'closed'
superseder = None
type = 'behavior'
url = 'https://bugs.python.org/issue29519'
versions = ['Python 2.7', 'Python 3.5', 'Python 3.6', 'Python 3.7']

[ambv]

Antoine, bpo-28427 introduces a regression. When used with multiprocessing, the WeakValueDictionary in multiprocessing.util._afterfork_registry causes the remove() to be invoked during atexit and then sys.meta_path is None, lots of things is None, including the global _remove_dead_weakref.

In effect, I'm getting spew like this:

Exception ignored in: <function WeakValueDictionary.__init__.<locals>.remove at 0x7fb2b905e2f0>
Traceback (most recent call last):
  File "/usr/local/fbcode/gcc-4.9-glibc-2.20-fb/lib/python3.6/weakref.py", line 112, in remove
TypeError: 'NoneType' object is not callable
Exception ignored in: <function WeakValueDictionary.__init__.<locals>.remove at 0x7fb2b905e2f0>
Traceback (most recent call last):
  File "/usr/local/fbcode/gcc-4.9-glibc-2.20-fb/lib/python3.6/weakref.py", line 112, in remove
TypeError: 'NoneType' object is not callable
Exception ignored in: <function WeakValueDictionary.__init__.<locals>.remove at 0x7fb2b905e2f0>
Traceback (most recent call last):
  File "/usr/local/fbcode/gcc-4.9-glibc-2.20-fb/lib/python3.6/weakref.py", line 112, in remove
TypeError: 'NoneType' object is not callable
Exception ignored in: <function WeakValueDictionary.__init__.<locals>.remove at 0x7fb2b905e2f0>
Traceback (most recent call last):
  File "/usr/local/fbcode/gcc-4.9-glibc-2.20-fb/lib/python3.6/weakref.py", line 112, in remove
TypeError: 'NoneType' object is not callable

When debugged, this 'NoneType' is _remove_dead_weakref (weakref.py:117).

I'm working on a smaller repro, unfortunately this happens as part of a rather large multiprocessing app so this might take a while. Just wanted to let you know, maybe you'll know right away what the problem is.

[ambv]

I have a pretty minimal repro. You'll have to download both mod1.py and mod2.py and execute python3.6 mod1.py. You'll see:

Exception ignored in: <function WeakValueDictionary.__init__.<locals>.remove at 0x7fcb56b09400>
Traceback (most recent call last):
  File "/usr/local/fbcode/gcc-4.9-glibc-2.20-fb/lib/python3.6/weakref.py", line 117, in remove
TypeError: 'NoneType' object is not callable

I can reproduce this both on CentOS 7 and macOS Sierra. I don't quite understand what makes this particular combination of instructions in the modules trigger the problem, I'd appreciate some explanation. AFAICT any additional simplification to the two modules makes it stop producing the spew.

Note: the spew doesn't happen with 3.6.0, only with master because it's directly caused by the missing global after the change in bpo-28427.

[python-dev]

New changeset 2cb530243943 by Łukasz Langa in branch '3.5':
Fix bpo-29519: weakref spewing exceptions during interp finalization
https://hg.python.org/cpython/rev/2cb530243943

New changeset c5267272e66a by Łukasz Langa in branch '3.6':
Merge 3.5 (fix bpo-29519)
https://hg.python.org/cpython/rev/c5267272e66a

New changeset e91ec62da088 by Łukasz Langa in branch 'default':
Merge 3.6 (fix bpo-29519)
https://hg.python.org/cpython/rev/e91ec62da088

[ambv]

Given that the weakref behavior during shutdown with the repro was identical in 3.6.0, as well as in 3.5.2 and earlier, I patched this by binding the name to a local, therefore making it usable also during finalization.

[python-dev]

New changeset 6c961c5a5396ef125dbbc47c060272a5d12c1646 by Łukasz Langa in branch 'master':
Fix bpo-29519: weakref spewing exceptions during interp finalization
6c961c5

New changeset 39279a07e8fc7bc14dcb948d970c12dbbf65ff40 by Łukasz Langa in branch 'master':
Merge 3.5 (fix bpo-29519)
39279a0

New changeset 8d6fcedf53048ff53cde91f7a55c1583456fa15e by Łukasz Langa in branch 'master':
Merge 3.6 (fix bpo-29519)
8d6fced

[python-dev]

New changeset 6c961c5a5396ef125dbbc47c060272a5d12c1646 by Łukasz Langa in branch '3.5':
Fix bpo-29519: weakref spewing exceptions during interp finalization
6c961c5

[python-dev]

New changeset 6c961c5a5396ef125dbbc47c060272a5d12c1646 by Łukasz Langa in branch '3.6':
Fix bpo-29519: weakref spewing exceptions during interp finalization
6c961c5

New changeset 39279a07e8fc7bc14dcb948d970c12dbbf65ff40 by Łukasz Langa in branch '3.6':
Merge 3.5 (fix bpo-29519)
39279a0

[methane]

gsutil shows massive "Exception ... ignored" message from this regression.
https://twitter.com/minimum2scp/status/890021408482316289

I think we should backport this fix to Python 2.7.14.