Support transfer of ownership of Python subclasses of pybind11 C++ classes with `shared_ptr<>`, avoid slicing

[EricCousineau-TRI]

Update (2020-12-22): Simplest reproduction can be seen in #1333

---

If a Python subclass of a pybind11 C++ class, with holder_type = shared_ptr<T>, has a shared_ptr<> that exists in C++ when its Python ref_count() goes to zero, Python will destroy the Python extension, and leave the C++ Alias type (e.g. PyBase, using PYBIND11_OVERLOAD*) functionality bare (and actually aliased / truncated affected by object slicing).

This could be prevented if a custom __del__, PyTypeObject::tp_del, or PyTypeObject::tp_dealloc is installed such that the instance could be resurrected if holder.use_count() > 1 for the given value-holder, and the ownership is released to C++.
(__del__ is the easiest, as Python explicitly mentions that you can increase the refcount of a destructing object and save it from death.)

This is related to #1132.

[ppt-adsk]

Hello,

I am part of the Autodesk Maya team. We are using pybind11 for an upcoming project, and are hitting this exact problem (Python subclass of a pybind11 C++ class, holder_type = shared_ptr, shared_ptr in C++ with Python ref_count of 0, resulting in an object without the Python derived functionality).

We are not familiar with the details of your proposal, but are extremely interested in a solution to the problem, as for us it currently involves kludgey Python-side referencing of objects to keep their ref_count above 0). Thanks for trying to move this forward.

[EricCousineau-TRI]

Thanks for posting your use case!

I have tried to explain how the main portion of the WIP PR would work here:
https://github.com/EricCousineau-TRI/pybind11/blob/feature/py_move_shared-wip/docs/advanced/classes.rst#virtual-inheritance-and-lifetime
Please let me know if you have any questions about that, and please do let me know if you happen to find any better solutions.

[cgerum]

I also ran into this Problem. For now i think I can work around this problem. But I would by Interested in an easier solution. Your proposal seems fine with me.

Can I be of any assistance moving this forward?

[EricCousineau-TRI]

Can I be of any assistance moving this forward?

Most certainly! If you're up to it, could you try using this branch, and provide your feedback?
https://github.com/RobotLocomotion/pybind11/tree/drake

It incorporates a working prototype of this feature (with a few other features). If this implementation is something that is actually useful for you, it can help make a better case for upstreaming the feature.

If you diff the docs against the merge-base of the upstream master, you'll see what it introduces:
RobotLocomotion/pybind11@upstream...RobotLocomotion:drake#diff-47ee721743764a0358552e687404eb8e

This code is presently being used as part of an online course related to the software we are developing, and it seems to be treating us pretty well:
http://underactuated.csail.mit.edu/Spring2018/

EDIT: Note that the changes I've made to this fork breaks the AppVeyor Windows builds. Since I don't have a Windows machine setup and since we aren't using Windows for our applications, I haven't tried delving in to fixing this. Do you use Windows?

[EricCousineau-TRI]

Closing this in lieu of #1333 for tracking purposes.

(This issue has scope creep from bad wording hehe 😬)