In 2.7 and 3.3: >>> import _tkinter >>> _tkinter.Tcl_Obj() Traceback (most recent call last): File "<stdin>", line 1, in <module> TypeError: cannot create '_tkinter.Tcl_Obj' instances >>> _tkinter.TkttType() Traceback (most recent call last): File "<stdin>", line 1, in <module> TypeError: cannot create 'tktimertoken' instances In 3.4+: >>> import _tkinter >>> _tkinter.Tcl_Obj() Segmentation fault (core dumped) And the same for _tkinter.TkttType. Looks as this is a result of issue15721.

With 3.4.3, I get the equivalent on Windows (a 'python has stopped working' box).

Probably a PyType_Ready() missing.

PyType_Ready() is called inside PyType_FromSpec().

Before issue15721 the tp_new was NULL, and this prevented from creating instances. After issue15721 the tp_new is inherited from object. An instance is created, but all internal pointers are NULLs. Following repr() dereferences NULL.

There is special exception -- tp_new is not inherited for static types whose base class is 'object' (see a comment in inherit_special() in Objects/typeobject.c#4569 for explanation). After converting them to heap types they became inherit tp_new from object. Proposed patch adds workarounds for types in _tkinter and curses.panel modules.

The news part of the patch did not apply -- best not to include such in posted patches. After recompiliing 3.5 on Windows 10, the two tkinter test cases give the 3.3 TypeErrors instead of crashing. test_tcl also runs. curses does not run on Windows. I do not know enough internals to really understand #15721 or why you call the fix a workaround, but I do understand 'exception' (OK) versus 'crash' (forbidden). Unless you think restoring the previous situation (tp_new = NUll) can cause a worse problem, I would apply this soon, within a week, so not forgotten before the June update releases. The tcl test addition is correct. The fix can be improved later if you think of something better. I initially wondered why _tkinter has classes that cannot be instantiated (from Python). Then I discovered that this is normal, and that the exception is standard. >>> type(iter([]))() Traceback (most recent call last): File "<stdin>", line 1, in <module> TypeError: cannot create 'list_iterator' instances Is list_iterator.tp_new NULL? Or does it raise instead of crash by other means?

I called the fix a workaround because other solution may be to make PyType_FromSpec() not setting tp_new in such cases. I expect that this is common mistake when convert a static class to a heap class with PyType_FromSpec(). At least we should add warnings in C API documentation and PEP 384. Or change the behavior of PyType_FromSpec() if it is appropriate. But this is different issue. If PyType_FromSpec() will be changed, proposed fix can be reverted.

I noticed that you reported the crash and asked for a fix or reversion on #15721 6 months ago, with no response. Hence, lets not wait longer. Looking again, I see that there were 3 rounds of patches in 2012/2013. If you want to follow up, I suggest closing that issue and open a fresh issue with with alternatives and your recommendation.

New changeset cd25508c62fc by Serhiy Storchaka in branch '3.5': Issue #23815: Fixed crashes related to directly created instances of types in https://hg.python.org/cpython/rev/cd25508c62fc New changeset 1c6326e81c33 by Serhiy Storchaka in branch 'default': Issue #23815: Fixed crashes related to directly created instances of types in https://hg.python.org/cpython/rev/1c6326e81c33

After looking closer I found that making PyType_FromSpec() not inheriting tp_new for static types whose base class is 'object' is not an option, because this breaks examples in the documentation, and likely break some third-party code. Opened new issue26979 for documenting the catch. Thank you for your review Terry.