In Python 3.10 it is not possible to instantiate an asyncio.Condition that wraps an asyncio.Lock without raising a "loop argument must agree with lock" exception.

This code raises that exception:

    asyncio.Condition(asyncio.Lock())

This worked in previous Python versions.

Note that the error only occurs if an event loop is running. Here's a simple script that replicates the problem:

    import asyncio

    # This runs without an exception:
    print(asyncio.Condition(asyncio.Lock()))

    # This does not work:
    async def example():
        print(asyncio.Condition(asyncio.Lock()))

    # This raises "ValueError: loop argument must agree with lock":
    asyncio.run(example())

I ran across this issue while trying to use the https://pypi.org/project/janus/ locking library with Python 3.10 - see my issue on their tracker here: aio-libs/janus#358

It looks like the relevant test is here:

    def test_explicit_lock(self):
        lock = asyncio.Lock()
        cond = asyncio.Condition(lock)

        self.assertIs(cond._lock, lock)
        self.assertIs(cond._loop, lock._loop)

But... that test doesn't appear to run inside an event loop, so it's not covering the behaviour described in this issue.

Issue confirmed. The problem is that Condition.__init__ compares its own loop with Lock's internal _loop attribute. That attribute is set to None unless the lock waited to be acquired.

uriyyo, Andrew, Yury, since it's pretty likely that Lock objects will now have _loop set to None, I think the check for "loop agreement" in Condition is kind of useless. So I propose to simply remove it. It's the only such check in all of asyncio.

If you insist on keeping the loop check, it should special-case _loop is None.

Simon, thanks for your detailed report! As a backwards-compatible workaround till we get a fix in, your user code can do the following:

>>> l = asyncio.Lock()
>>> getattr(l, '_get_loop', lambda: None)()
<_UnixSelectorEventLoop running=True closed=False debug=False>

You can use such lock without issues now:

>>> asyncio.Condition(l)
<asyncio.locks.Condition object at 0x10c05bee0 [unlocked]>

Alternatively, if the above disgusts you and you only want to trigger public APIs, you can do this dance:

>>> l = asyncio.Lock()
>>> await l.acquire()  # first acquire will just work
True
>>> try:
...   # second acquire will block so we time it out
...   await asyncio.wait_for(l.acquire(), 0.1)
... except asyncio.TimeoutError:
...   pass
...
>>> l.release()

Now the lock is fully initialized and we can use it:

>> c = asyncio.Condition(l)

Both workarounds should be compatible with Python 3.7+ asyncio.

New changeset 1a78924 by Joongi Kim in branch 'main':
bpo-45416: Fix use of asyncio.Condition() with explicit Lock objects (GH-28850)
1a78924

New changeset 164dddf by Miss Islington (bot) in branch '3.10':
bpo-45416: Fix use of asyncio.Condition() with explicit Lock objects (GH-28850)
164dddf

Thanks, guys!