Semaphore race condition

I unexpectedly got a ‘WouldBlock’ error while not using any *_nowait() calls. I’m using the latest version of the master branch. The problem goes away if I replace https://github.com/agronholm/anyio/blob/master/src/anyio/_core/_synchronization.py#L305 with await self.acquire(), but I’m not sure if that’s actually the best solution. I think the problem is that another task increments the semaphore (by calling acquire()) before the Event.wait()-handling coroutine is scheduled to run.

Relevant part of the traceback:

    async with self._pool.read_transaction() as tx:
/usr/lib/python3.8/contextlib.py:171: in __aenter__
    return await self.gen.__anext__()
tests/test_replicate.py:39: in read_transaction
    async with self.read_semaphore, self._transaction() as conn:
venv/lib/python3.8/site-packages/anyio/_core/_synchronization.py:282: in __aenter__
    await self.acquire()
venv/lib/python3.8/site-packages/anyio/_core/_synchronization.py:306: in acquire
    self.acquire_nowait()
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 

self = <anyio.Semaphore object at 0x7f434698fb50>

    def acquire_nowait(self) -> None:
        """
        Acquire the underlying lock, without blocking.
    
        :raises ~WouldBlock: if the operation would block
    
        """
        if self._value == 0:
>           raise WouldBlock
E           anyio.WouldBlock

venv/lib/python3.8/site-packages/anyio/_core/_synchronization.py:316: WouldBlock

[marten-de-vries]

Yes, it does!

I have a test that reliably hits this issue, and after installing your branch it passes, but sadly it’s an integration test that isn’t easy to reduce to something that could be integrated in anyio’s test suite.

[marten-de-vries]

I can confirm the final fix also works. Thanks @smurfix and @agronholm!