pybind11::array argument leaks reference count of dtype singleton

[pv]

Issue description

Calling a function with a const pybind11::array& argument appears to leak refcounts to some Python object. However, it's not the object passed in nor there is a memory leak, so possibly the refcount of some global singleton is not decremented properly.

Numpy itself has a known refcount leak (numpy/numpy#8503), but only with dtype=object arrays --- which are not involved in the case below. Unless, pybind11 uses numpy object arrays internally?
The leaking refcount seems to be that of a dtype singleton.

Reproducible example code

Python debug build is needed for sys.gettotalrefcount(), but you can see the dtype leak also on normal Python:

set -e
PYTHON=python3.7-dbg

cat <<__EOF__ > refleak.cxx
#include <pybind11/pybind11.h>
#include <pybind11/numpy.h>

/* calling this leaks sys.gettotalrefcount() */
void doit(const pybind11::array& a)
{
    return;
}

PYBIND11_MODULE(refleak, m) {
    m.def("doit", doit);
}
__EOF__

cat <<__EOF__ > run.py
import refleak
import numpy as np
import sys

if not hasattr(sys, 'gettotalrefcount'):
    sys.gettotalrefcount = lambda: float('nan')

x = np.array([1])
dtype = np.dtype(np.float_)

before_dtype = sys.getrefcount(dtype)
before = sys.gettotalrefcount()
for j in range(10000):
    refleak.doit(x)
after = sys.gettotalrefcount()
after_dtype = sys.getrefcount(dtype)

import pybind11, numpy
print("numpy:", numpy.__version__)
print("pybind11:", pybind11.__version__)
print("leaked (dtype):", after_dtype - before_dtype, "refs")
print("leaked (total):", after - before, "refs")
print("OK" if after < before + 5000 else "FAIL")
__EOF__

cat <<__EOF__ > setup.py
from setuptools import Extension, setup
import numpy, pybind11
include_dirs = [numpy.get_include(), pybind11.get_include(True), pybind11.get_include(False)]
ext_modules = [Extension("refleak", ["refleak.cxx"], include_dirs=include_dirs)]
setup(name='foo', ext_modules=ext_modules)
__EOF__

$PYTHON setup.py build_ext -i -g
$PYTHON run.py

gives here (same for pybind11 2.3.0 and 4a3464f; numpy 1.16.4 or current master)

numpy: 1.16.4
pybind11: 2.3.0
leaked (dtype): 10000 refs
leaked (total): 10004 refs
FAIL

[pv]

Should this be reinterpret_steal?

pybind11/include/pybind11/numpy.h

Line 1047 in 4a3464f

return reinterpret_borrow<pybind11::dtype>(ptr);

PyArray_DescrFromType returns a new reference. Changing this to steal fixes the refcount leak.

[wjakob]

Yes, that sounds plausible. Can you make a PR?