avoid using __file__ in pytest_plugin_registered as can be wrong on Windows

[woutdenolf]

closes #11816

Alternative fix to #11821.

The plugin name is equal to the path for conftest plugins (suggested in #11816 (comment)). Use that instead of plugin.__file__.

Test with the h5py CI that was broken by the issue: h5py/h5py#2370

Another way of testing this MR:

python -m pip install "ewokscore==0.7.1" matplotlib ipykernel^
    git+https://github.com/woutdenolf/pytest.git@fix_missing_fixture_issue --no-cache
python -m pytest -v --pyargs ewokscore.tests

[woutdenolf]

@bluetech Is that what you meant in #11816 (comment)?

[bluetech]

Yep, exactly, this LGTM.

Test failures seem unrelated. "hypothesis.errors.HypothesisSideeffectWarning: Slow code in plugin: avoid lazy evaluation of text() at import time!" never seen this before. Unless the get_name slowness is actually the culprit... I will re-run the jobs later.

As for optimizing the get_name call, I can do it in a separate PR, unless you want to do it here already. The best approach I think is to add a plugin_name parameter to the pytest_plugin_registered hook.

[woutdenolf]

The best approach I think is to add a plugin_name parameter to the pytest_plugin_registered hook.

No idea what this means so if you could handle it in another MR, that would be great.

[bluetech]

The Hypothesis issue seems to reproduce when re-running the jobs, though some jobs pass. The (existing, seems to be unrelated to this PR) test is:

@hypothesis.given(strategies.text() | strategies.binary())
@hypothesis.settings(
    deadline=400.0
)  # very close to std deadline and CI boxes are not reliable in CPU power
def test_idval_hypothesis(self, value) -> None:
    escaped = IdMaker([], [], None, None, None, None, None)._idval(value, "a", 6)
    assert isinstance(escaped, str)
    escaped.encode("ascii")

The error is:

______________________ TestMetafunc.test_idval_hypothesis ______________________

self = <metafunc.TestMetafunc object at 0x7f5c52dfd810>

    @hypothesis.given(strategies.text() | strategies.binary())
>   @hypothesis.settings(
        deadline=400.0
    )  # very close to std deadline and CI boxes are not reliable in CPU power

/home/runner/work/pytest/pytest/testing/python/metafunc.py:296: 
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
/home/runner/work/pytest/pytest/.tox/py311-coverage/lib/python3.11/site-packages/hypothesis/core.py:669: in process_arguments_to_given
    s.validate()
/home/runner/work/pytest/pytest/.tox/py311-coverage/lib/python3.11/site-packages/hypothesis/strategies/_internal/strategies.py:417: in validate
    self.do_validate()
/home/runner/work/pytest/pytest/.tox/py311-coverage/lib/python3.11/site-packages/hypothesis/strategies/_internal/strategies.py:684: in do_validate
    for e in self.element_strategies:
/home/runner/work/pytest/pytest/.tox/py311-coverage/lib/python3.11/site-packages/hypothesis/strategies/_internal/strategies.py:659: in element_strategies
    if not arg.is_empty:
/home/runner/work/pytest/pytest/.tox/py311-coverage/lib/python3.11/site-packages/hypothesis/strategies/_internal/strategies.py:136: in accept
    recur(self)
/home/runner/work/pytest/pytest/.tox/py311-coverage/lib/python3.11/site-packages/hypothesis/strategies/_internal/strategies.py:132: in recur
    mapping[strat] = getattr(strat, calculation)(recur)
/home/runner/work/pytest/pytest/.tox/py311-coverage/lib/python3.11/site-packages/hypothesis/strategies/_internal/lazy.py:89: in calc_is_empty
    return recur(self.wrapped_strategy)
/home/runner/work/pytest/pytest/.tox/py311-coverage/lib/python3.11/site-packages/hypothesis/strategies/_internal/lazy.py:104: in wrapped_strategy
    check_sideeffect_during_initialization("lazy evaluation of {!r}", self)
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 

what = 'lazy evaluation of {!r}', extra = '', fmt_args = (text(),)
msg = 'lazy evaluation of text()'

    def check_sideeffect_during_initialization(
        what: str, *fmt_args: object, extra: str = ""
    ) -> None:
        """Called from locations that should not be executed during initialization, for example
        touching disk or materializing lazy/deferred strategies from plugins. If initialization
        is in progress, a warning is emitted.
    
        Note that computing the repr can take nontrivial time or memory, so we avoid doing so
        unless (and until) we're actually emitting the warning.
        """
        global _first_postinit_what
        # This is not a particularly hot path, but neither is it doing productive work, so we want to
        # minimize the cost by returning immediately. The drawback is that we require
        # notice_initialization_restarted() to be called if in_initialization changes away from zero.
        if _first_postinit_what is not None:
            return
        elif _hypothesis_globals.in_initialization:
            # Note: -Werror is insufficient under pytest, as doesn't take effect until
            # test session start.
            msg = what.format(*fmt_args)
>           warnings.warn(
                f"Slow code in plugin: avoid {msg} at import time!  Set PYTHONWARNINGS=error "
                "to get a traceback and show which plugin is responsible." + extra,
                HypothesisSideeffectWarning,
                stacklevel=3,
            )
E           hypothesis.errors.HypothesisSideeffectWarning: Slow code in plugin: avoid lazy evaluation of text() at import time!  Set PYTHONWARNINGS=error to get a traceback and show which plugin is responsible.

/home/runner/work/pytest/pytest/.tox/py311-coverage/lib/python3.11/site-packages/hypothesis/configuration.py:67: HypothesisSideeffectWarning

@Zac-HD Maybe you have an idea? The test seems run of the mill to me and I don't see any import-time evaluation going on but I don't have much experience with Hypothesis.

[Zac-HD]

We just added that warning last night, and I suspect a pytest-specific interaction... I suggest globally ignoring HypothesisSideeffectWarning for now to unblock this PR and we'll look into a better fix upstream.

[bluetech]

Added the change to the hook. Did it in this PR to make backporting easier.