Reference leak in code object with reference loop

[serhiy-storchaka]

If the code object creates a reference loop, this causes a reference leak. For example (see :

def f():
    return 1234.5

code = f.__code__
a = []
code = code.replace(co_consts=code.co_consts + (a,))
a.append(code)

(See

cpython/Lib/test/test_marshal.py

Lines 354 to 360 in d61fcf8

	def test_reference_loop_code(self):
	def f():
	return 1234.5
	code = f.__code__
	a = []
	code = code.replace(co_consts=code.co_consts + (a,))
	a.append(code)

added in #148698)

This is because in the GIL-enabled build the code object does not support GC and does not have tp_traverse.

Linked PRs

• gh-148722: Make code objects support cyclic GC #151155

[markshannon]

Supporting cycles in code objects adds a lot of complexity to marshal and extra work for the GC to do, which is going to slow down everything.

I think it would be better to warn that introducing a cycle with code_object.replace will leak memory.
code_object.replace is a bit of an odd thing, and almost all code objects live as long as the interpreter anyway.
Alternatively, and this is a lot more complex, we could reject calls to code_object.replace that create cycles.

[serhiy-storchaka]

At first glance #151155 looks correct to me. Does it add too much overhead in the GIL-enabled build?

[serhiy-storchaka]

I am fine with just adding a tiny warn in the documentation.