Leaked types and shared_ptr

[chrisrichardson]

I'm trying to understand why I get a lot of leaked type messages on exit from nanobind. I have some code that looks a bit like this:

#include <nanobind/nanobind.h>
#include <nanobind/stl/shared_ptr.h>

namespace nb = nanobind;

class Owl
{
public:
  Owl(){};
};

class Barn
{
public:
  Barn(std::shared_ptr<Owl> owl) : _owl(owl){};

  std::shared_ptr<Owl> _owl;
};

NB_MODULE(example, m)
{

  nb::class_<Owl>(m, "Owl").def(nb::init<>());

  nb::class_<Barn>(m, "Barn")
      .def(nb::init<std::shared_ptr<Owl>>())
      .def_ro("owl", &Barn::_owl);
}        

Now doing in Python:

owl = Owl()

w = Barn(owl)

will trigger a leaked type warning.

 - leaked type "example.Owl"

Looking at nanobind/stl/shared_ptr.h it seems like py_deleter is being called on Owl after the interpreter has exited, so the refcount is not being decremented. Is this expected behaviour? Is there anything I can do better to prevent it?

[wjakob]

Older versions of Python are sometimes a bit leaky especially on the global scope. Does the issue happen if you wrap

owl = Owl()

w = Barn(owl)

into a function and then call that function?

[wjakob]

You can disable the leak warning by the way, refer to the documentation for details. In pybind11, you simply did not know about leaks, so it's a benefit for developers.

[chrisrichardson]

Yes, wrapping in a function gets rid of the problem. Maybe it will be better with Python 3.12 (I have 3.10 on this machine).
I know we can disable warnings (and have done it for Release builds)...
Thanks for the quick reply.