Repository navigation
Difference in ressurrection behavior with __del__ in py2 vs. py3 #76558
Description
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 = None
closed_at = <Date 2017-12-19.19:00:56.948>
created_at = <Date 2017-12-19.16:29:59.034>
labels = ['type-bug', '3.7', 'docs']
title = 'Difference in ressurrection behavior with __del__ in py2 vs. py3'
updated_at = <Date 2020-04-09.15:25:33.816>
user = 'https://github.com/EricCousineau-TRI'bugs.python.org fields:
activity = <Date 2020-04-09.15:25:33.816>
actor = 'Eric Cousineau'
assignee = 'docs@python'
closed = True
closed_date = <Date 2017-12-19.19:00:56.948>
closer = 'pitrou'
components = ['Documentation']
creation = <Date 2017-12-19.16:29:59.034>
creator = 'Eric Cousineau'
dependencies = []
files = ['47339']
hgrepos = []
issue_num = 32377
keywords = ['patch']
message_count = 9.0
messages = ['308660', '308661', '308662', '308664', '308665', '308667', '308671', '366057', '366068']
nosy_count = 3.0
nosy_names = ['pitrou', 'docs@python', 'Eric Cousineau']
pr_nums = ['4927', '4929']
priority = 'normal'
resolution = 'fixed'
stage = 'resolved'
status = 'closed'
superseder = None
type = 'behavior'
url = 'https://bugs.python.org/issue32377'
versions = ['Python 3.6', 'Python 3.7']
Activity
EricCousineau-TRI commentedon Dec 19, 2017
Due to how
PyObject_CallFinalizeris written in python3,__del__will only ever be called once.In my use case, I am experimenting with a feature in
pybind11to prevent slicing with Python class instances that inherit from pybind11-C++ base classes, which involves detecting when an instance loses all reference in Python (Py_REFCNT(...) == 0) but still has reference in C++ (shared_ptr::count() > 0), and reviving the Python portion when this situation happens.In python2, I could do this without a hitch, as a resurrected object could have its
__del__method called multiple times (throughtp_deallocI believe?). But in python3, the object is marked with_PyGC_SET_FINALIZED(...), thus preventing__del__from being called again.It'd be nice to either (a) somehow allow
__del__to be called naturally without too much fuss or, at the least, (b) have this reflected in the documentation:https://docs.python.org/3/reference/datamodel.html#object.__del__
See attached
revive_test. Example execution:pitrou commentedon Dec 19, 2017
Thanks for the report. Apparently I forgot to update that piece of documentation when PEP-442 was implemented.
EricCousineau-TRI commentedon Dec 19, 2017
You're welcome, and thank you for the prompt response!
I will say that it feels a tad odd to only have
tp_finalizebe called once for the entire lifetime of the object, while still having the option of it being resurrected.Is there any way to somehow "un-mark" the object to enable this workflow that I would like to have?
My current hack is to call
_PyGC_SET_FINALIZED(self, 0)- may I ask if there is a simpler way to do this?pitrou commentedon Dec 19, 2017
Le 19/12/2017 à 17:46, Eric Cousineau a écrit :
Well... perhaps you could create another PyObject (it's just a wrapper,
right?) since the old one doesn't have any outside references to it
remaining.
Note that calling __del__ only once is also how PyPy works:
http://doc.pypy.org/en/latest/cpython_differences.html#differences-related-to-garbage-collection-strategies
If there is some demand we could expose a higher-level spelling of
_PyGC_SET_FINALIZED(self, 0).pitrou commentedon Dec 19, 2017
I've proposed a documentation improvement in #4927 . Please chime in if you see have issues with it.
pitrou commentedon Dec 19, 2017
New changeset 4b96593 by Antoine Pitrou in branch 'master':
bpo-32377: improve __del__ docs and fix mention about resurrection (bpo-4927)
4b96593
pitrou commentedon Dec 19, 2017
New changeset dc5770b by Antoine Pitrou (Miss Islington (bot)) in branch '3.6':
bpo-32377: improve __del__ docs and fix mention about resurrection (GH-4927) (bpo-4929)
dc5770b
EricCousineau-TRI commentedon Apr 9, 2020
Super late response, but for this part:
In certain cases, yes, that would be the case. I do have unittests that
(hackily) check
id(o)before and after resurrection by using aweakref,so I could relax that contract.
However, in other cases, the "wrapper" object in Python is actually a C++
base class that has been derived from in Python. If ownership of the C++
instance (which should more-or-less own the Python portion) goes solely to C++,
then the Python portion could be garbage collected, and you get a sort-of
slicing effect:
pybind/pybind11#1145
Granted, within this web of issues, one alternative is to always extend the
lifetime of the Python object by the C++ portion. However, that creates a
(opaque?) reference cycle between C++ and CPython, so I am a bit afraid of what
that would do. I would prefer to make memory management conservative if
possible.
Or rather: It's something cool to try out, but I have not yet had time to dig
in :(
EricCousineau-TRI commentedon Apr 9, 2020
See also bpo-40240: "Expose public spelling of _PyGC_FINALIZED and _PyGC_SET_FINALIZED?"