BUG: Segmentation fault when unpickling an empty ndarray with a none-zero dimension

[wuciting]

Describe the issue:

There is a Segmentation fault when unpickling an empty ndarray with a none-zero dimension.
The first dimension 93049 may depend on the platform. If you cannot reproduce the error, please increase the first dimension.
I think it has something to do with the page allocation.

Reproduce the code example:

import pickle

import numpy as np

a = np.array([]).reshape(93049, 0)
print(a.shape)
s = pickle.dumps(a)
print(s)
b = pickle.loads(s)
print(b)

Error message:

Segmentation fault (core dumped)

NumPy/Python version information:

1.22.1 3.8.11 (default, Aug 3 2021, 15:09:35)
[GCC 7.5.0]

[seberg]

I am very confused why our tests appear to not catch this at least when running in valgrind (at least I assume that they should). I am pretty sure I know the issue, but it should be a fairly straight forward fix if anyone wants to go bug hunting with gdb or valgrind. (A test that actually triggers this reliably would be good though, ideally even covering a few other branches.)

EDIT: silly me, you need the empty array, and we likely don't have a test with empty arrays.

[doctormartin67]

I traced the problem down using gdb and it seems the issue comes from numpy. The seg fault occurs
at numpy/core/src/multiarray/methods.c:2208:
memcpy(PyArray_DATA(self), datastr, num);
When I take a look at the PyArray_DATA function it casts self to a PyArrayObject_fields struct pointer to access it's data field.
The comments above the declaration of this struct says that accessing its field is deprecated and that NPY_NO_DEPRECATED_API should be defined to avoid this issue. Not sure why it isn't though. This is as far I can could go without studying numpy in more depths.

[seberg]

Yes, that is all correct. The "deprecated" stuff isn't a worry, it only applies to users outside of NumPy anyway (and even then isn't problematic).
The code works for non-empty arrays, so the problem is just the num.

[seberg]

@doctormartin67 thanks for looking into this, @alexdesiqueira had just started on this yesterday and opened a PR now, sorry about that bad timing.

[doctormartin67]

@seberg No worries, always have fun taking a look at seg faults in gdb

[charris]

@seberg This is fixed, correct?