Pytest tests cannot load global config that definitely exists even though real use of the library can

[beamerblvd]

While working on and writing tests for #1466, I discovered that Config.get_global_config raised an IOError in my test. Confused as to why that didn't break other tests, I searched for use of this in the tests and found this:

def test_global_config() -> None:
    try:
        assert Config.get_global_config() is not None
    except IOError:
        # There is no user config
        pass


def test_system_config() -> None:
    try:
        assert Config.get_system_config() is not None
    except IOError:
        # There is no system config
        pass

I changed these tests to this so that these situations result in an official "skip" with information about why it was skipped, instead of just silently fake-passing:

def test_global_config() -> None:
    try:
        assert Config.get_global_config() is not None
    except IOError as e:
        settings = Settings()
        pytest.skip(f'Unavailable for testing with home dir = {settings.homedir}: {e}')


def test_system_config() -> None:
    try:
        assert Config.get_system_config() is not None
    except IOError as e:
        pytest.skip(f'Unavailable for testing: {e}')

But the resulting pytest output puzzled me:

test/test_config.py::test_global_config SKIPPED (Unavailable for testing with home dir = /Users/nicholas: the global file '.gitconfig' doesn't exist:)
test/test_config.py::test_system_config SKIPPED (Unavailable for testing: the system file 'gitconfig' doesn't exist:)

The first reason it puzzled me is that /Users/nicholas/.gitconfig definitely exists:

$ ls -al /Users/nicholas/.gitconfig
-rw-r--r--  1 nicholas  staff  366 Apr  9 17:39 /Users/nicholas/.gitconfig

The second reason it puzzled me is that I don't encounter this error if I simply use the library in a Python shell prompt:

>>> c = pygit2.Config.get_global_config()
>>> c["push.gpgSign"]
'if-asked'

Now, it's correct that my system config file does not exist:

>>> pygit2.Config.get_system_config()
Traceback (most recent call last):
  File "<python-input-8>", line 1, in <module>
    pygit2.Config.get_system_config()
    ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
  File "/Users/nicholas/Development/OSS/libgit2/pygit2/.venv/lib/python3.14/site-packages/pygit2/config.py", line 309, in get_system_config
    check_error(err)
           ~~~~^^^^^
  File "/Users/nicholas/Development/OSS/libgit2/pygit2/.venv/lib/python3.14/site-packages/pygit2/config.py", line 300, in _from_found_config
  File "/Users/nicholas/Development/OSS/libgit2/pygit2/.venv/lib/python3.14/site-packages/pygit2/errors.py", line 56, in check_error
    raise IOError(message)
OSError: the system file 'gitconfig' doesn't exist: No such file or directory

But note that that error message is slightly different from the skip message of the test: "No such file or directory" is missing in the skip message. Weird?

If I now manually create /etc/gitconfig:

>>> c = pygit2.Config.get_system_config()
>>> c["bogus.foo"]
'bar'

But, low and behold, that test, also, still skips with the same partial message about not finding the system config file.

So, these two functions do work ... just not in tests. I've been unable to figure out why. My best guess is that mypy is somehow sandboxed so that it can't read these files? But that's just a wild guess. I can't confirm it.

[beamerblvd]

Still investigating this. Looks like the ": " followed by "No such file or directory" is coming from these lines in libgit2. This indicates that the "No such file or directory" is coming from the OS (via strerror), but in the pytest case, strerror is returning nothing (or so it seems).

There could be three reasons for this:

• Something set errno to 0 between when the error occurred and git_error_vset used errno (not clear how that could have happened, but it's possible ... it wasn't another thread, because errno is thread-safe).
• strerror itself encountered an error (and so it set errno and returned a pointer to an empty string).
• strerror was called in another thread between when it returns at this point and when git_str_put copies the string at the resulting pointer, which would result in undefined behavior because strerror is not thread-safe and subsequent calls modify the same pointer previously returned (libgit2 should not be using this and should instead be using strerror_r).

strerror doesn't return an empty string for unknown error codes. It returns "Unknown error nnn." So it's not an unknown error code.