Example (minimal version of python-trio/trio#425):

-----

async def open_file():
    pass

async def main():
    async with open_file():  # Should be 'async with await open_file()'
        pass

coro = main()
coro.send(None)

Here we accidentally left out an 'await' on the call to 'open_file', so the 'async with' tries to look up 'CoroutineType.__aexit__', which obviously doesn't exist, and the program crashes with an AttributeError("__aexit__"). Yet weirdly, this doesn't trigger a warning about 'open_file' being unawaited. It should!

Yury's theory: maybe BEFORE_ASYNC_WITH's error-handling path is forgetting to DECREF the object.

Yury's theory: maybe BEFORE_ASYNC_WITH's error-handling path is forgetting to DECREF the object.

Nope, that doesn't seem to be it. This version prints "refcount: 2" twice, *and* prints a proper "was never awaited" warning:

-----

import sys

async def open_file():
    pass

async def main():
    open_file_coro = open_file()
    print("refcount:", sys.getrefcount(open_file_coro))

try:
    async with open_file_coro:
        pass
except:
    pass

    print("refcount:", sys.getrefcount(open_file_coro))

coro = main()
try:
    coro.send(None)
except:
    pass

Looking at the ceval code, I think Yury's theory is plausible, and we may also be leaving the interpreter's internal stack in a dubious state. Things then get cleaned up if you wrap the async with in a try/except or try/finally:

==============

>>> async def try_main():
...     try:
...         async with open_file():
...             pass
...     finally:
...         pass
... 
>>> try_main().send(None)
sys:1: RuntimeWarning: coroutine 'open_file' was never awaited
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "<stdin>", line 3, in try_main
AttributeError: __aexit__

==============

Unfortunately for that theory, adding braces and "Py_DECREF(POP());" to the relevant error handling branch *doesn't* fix the problem.

I also found another way to provoke similar misbehaviour without async with:

==========

>>> async def open_file():
...     pass
... 
>>> open_file()
<coroutine object open_file at 0x7f92fe19c548>
>>> _
<coroutine object open_file at 0x7f92fe19c548>
>>> 1
__main__:1: RuntimeWarning: coroutine 'open_file' was never awaited
1
>>> open_file()
<coroutine object open_file at 0x7f92fe19c548>
>>> del _
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
NameError: name '_' is not defined
>>> _
<coroutine object open_file at 0x7f92fe19c548>
>>> 1
1
>>> _
1
>>>

==========

Changing

  async def main():
    async with open_file():
      pass

to

  async def main():
    c = open_file()
    async with c:
      pass

also makes it print the warning :)

Also I've made a test out of this snippet and running tests in refleak mode shows that there's indeed no refleak here.

The difference between these two functions is two extra opcodes: STORE_FAST/LOAD_FAST before BEFORE_ASYNC_WITH. With them we have a warning.

So refactoring it into "c = open_file(); async with c" just prolongs the life of the coroutine so that it's GCed outside of WITH_CLEANUP_START/WITH_CLEANUP_FINISH block. Something weird is going on in that block.

Ah, we should never really get to WITH_CLEANUP_START; the exception should be raised in BEFORE_ASYNC_WITH

So the problem was that _PyGen_Finalize wasn't issuing any warnings if there's any error set in the current tstate. And in Nathaniel's case, the current error was an AttributeError('__aexit__').

This check is weird, because right before raising the warning, we call PyErr_Fetch to temporarily reset the current exception if any, specifically to raise the warning :)

The PR just removes the check. Unless I'm missing something this should fix the issue.

Well, I feel silly then: bpo-32605

Ah, and in my REPL example, the NameError was pending when the internal result storage was getting set back to None.

I'm not sure I even knew the "Don't complain when an exception is pending" check existed, so it would have taken me a long time to find that.

I knew about that "if", but it never fully registered to me why it's there and what it is protecting us from ;) So I had to debug half of ceval loop before I stumbled upon it again ;)

New changeset 2a2270d by Yury Selivanov in branch 'master':
bpo-32703: Fix coroutine resource warning in case where there's an error (GH-5410)
2a2270d

Merged; closing this issue.