_py_abc Python implementation of abc is not thread-safe

[colesbury]

Bug report

The update to the invalidation counter is not thread-safe and can lose updates in some Python implementations:

Failures seen on:

• Python 3.14t
• Python 3.9
• pypy3.10
• pypy3.11

But not on Python 3.10-3.14 with GIL due to limited GIL switch opportunities.

cpython/Lib/_py_abc.py

Lines 54 to 70 in 05e89c3

	def register(cls, subclass):
	"""Register a virtual subclass of an ABC.
	Returns the subclass, to allow usage as a class decorator.
	"""
	if not isinstance(subclass, type):
	raise TypeError("Can only register classes")
	if issubclass(subclass, cls):
	return subclass # Already a subclass
	# Subtle: test for cycles *after* testing for "already a subclass";
	# this means we allow X.register(X) and interpret it as a no-op.
	if issubclass(cls, subclass):
	# This would create a cycle, which is bad for the algorithm below
	raise RuntimeError("Refusing to create an inheritance cycle")
	cls._abc_registry.add(subclass)
	ABCMeta._abc_invalidation_counter += 1 # Invalidate negative cache
	return subclass

For example, consider the following repro, adapted from test_abc.test_registration_basics:

import _py_abc as abc # Use Python implementation of ABCs!!
import threading
import os
import sys

sys.setswitchinterval(1e-6)

N = 5

def run(b):
    b.wait()

    class A(metaclass=abc.ABCMeta):
        pass
    A.register(int)
    if not isinstance(42, A):
        print("Oops!")
        os._exit(1)

def main():
    for _ in range(10000):
        threads = []
        b = threading.Barrier(N)
        for _ in range(N):
            t = threading.Thread(target=run, args=(b,))
            threads.append(t)
            t.start()
        for t in threads:
            t.join()


if __name__ == "__main__":
    main()

Linked PRs

• gh-130095: Mark _py_abc tests as not thread-safe #130131

[colesbury]

I'm not really sure what to do here. We could add a threading.Lock() to protect _abc_invalidation_counter, but I'm not sure we want to import the threading module in _py_abc.py.

cc @cfbolz in case you have thoughts on this.

[picnixz]

Not sure if this is relevant: https://github.com/python/cpython/actions/runs/13335219139/job/37248893358?pr=129175 (I don't know if the test is testing the Python or the C implementation).

EDIT: I think this is the reproducer you had. But now we caught it in the CI

[colesbury]

Thanks @picnixz. I put up a PR to avoid running them with --parallel-threads for now until we figure out if and how to make _py_abc thread-safe.

[cfbolz]

yeah, importing threading in _py_abc leads to a circular import problem if _abc is not available. What about using itertools.count? that's supposed to be thread-safe, isn't it?

[cfbolz]

hm, doesn't quite work, because itertools.count has no way to read the current value, only a way to increase the count and get the next value.

[sergey-miryanov]

If I'm right, the problem comes from three bytecode operations: LOAD_ATTR, BINARY_OP, and STORE_ATTR. I think there are two ways to fix this: using a specialized operation for atomic increment/update or the same built-in function. Current main branch has only three uses of _Py_atomic_add_uint64 - qsbr, _asynciomodule and _abc, so there are no existing ways to do atomic update from python. I believe the quickest way to add private built-in for this, but the right one to add atomic ops.

Sorry if I'm obvious.

[colesbury]

The pure Python companion modules like _py_abc aren't really used by CPython in normal circumstances outside of tests. They are primarily used by other Python implementations like PyPy. I don't think it makes much sense to consider CPython specific changes to bytecode.

https://peps.python.org/pep-0399/

[sergey-miryanov]

Thanks for the link - it is useful for my understanding.
I’m stick to atomic ops, but I’m sure you have much more and deeper understanding.

[cfbolz]

@colesbury if you are specifically concerned about pypy, we don't use _py_abc.py anymore either.

[colesbury]

Thanks for the clarification. I also noticed a failure when I substituted abc for _py_abc with pypy311, so there might be a thread-safety issue there as well.

If neither pypy nor CPython use _py_abc outside of tests, I'm not going to spend much more time on this.