BusError in uvloop within loop extract_stack for nested async_generators

提供されたスタックレポートを参照し、uvloop のデバッグモードを有効にした状態で、Python 3.12.4 および macOS 上のネストされた async_generator ケースを再現します。extract_stack 周辺の uvloop/handles/handle.pyx と uvloop/cbhandles.pyx を調査し、Bus Error の原因を特定して、デバッグモードを無効にせずにクラッシュが解消されたことを確認します。

• uvloop version:
• 0.19.0 (tested 0.17.0 to 0.19.0)
• Python version: 3.12.4
• Platform: OSX Ventura 13.4/ Darwin 22.5.0
• Can you reproduce the bug with PYTHONASYNCIODEBUG in env?:
• This bug only occurs if -X dev or PYTHONDEVMODE=1 or PYTHONASYNCIODEBUG=1 or loop.set_debug(True)
• Does uvloop behave differently from vanilla asyncio? How?:

• Vanilla asyncio and uvloop with loop.set_debug(False) do not crash in nested async_generator .asend calls

I have a pattern called DeferredExecutor which basically allows one to write:

class Foo:
        @database(transaction=True, iterable=False)
        async def save(self, *args, **kwargs):
               # /snip
              rows = yield DeferredQuery("some sql")
              assert isinstance(rows, tuple)
              yield Return(...)

And have it be wrapped by a decorator that will call the async_generator, run the deferred queries against a DB and .asend( the result back into the generator, which will throw an StopAsyncIteration at the conclusion of the generator, and allows me to catch it and end the generators execution.

I am getting my database access object library to run on Python 3.12 (jumping from Python 3.7).

However, on Python 3.12, a particular integration test that uses a particularly reduced implementation:

class Tenant:
    @database(iterable=False, transaction=True)
    async def invite_users(cls, *tenant_user_pairs, instance=None):
        invites = (instance or cls).create_invites_for(*tenant_user_pairs)
        gen_event_loop = Invite.save.raw_iterable(*invites)
        result = None
        while True:
            result = await gen_event_loop.asend(result)
            if isinstance(result, Return):
                break
            result = yield result
        yield result # StopAsyncIteration gets thrown here

causes uvloop to crash with a Bus Error on OSX but only when loop.set_debug(True) (which is autoset via the environment variables or interpreter devmode)

Regular asyncio, debug or not, works just fine. If I patch the EventLoopPolicy to.set_debug(False) like:

def patch_uvloop():
    if uvloop_version_info < (99, 99, 99) and platform.system() == 'Darwin':
        cls = uvloop.EventLoopPolicy
        if hasattr(cls, '__patched__'):
            return cls
        class EventLoopPolicy(cls):
            __patched__ = True
            #  ARJ: so uvloop will crash the entire interpreter sometimes
            # and the crash originates on OSX. Disable debug mode
            # (which keeps it from touching uvloop's extract_stack() which 
            # explodes badly)

            def _loop_factory(self):
                loop = super()._loop_factory()
                loop.set_debug(False)
                return loop
        uvloop.EventLoopPolicy = EventLoopPolicy

asyncio.set_event_loop_policy(uvloop.EventLoopPolicy())

Then the crashes never happen.

I do have lldb/gdb and the patience to recompile things with more debugging information (provided I know -how-) but a careful inspection of the following stack trace shows that it explodes after going into the extract_stack function of uvloop:

• https://github.com/MagicStack/uvloop/blob/6c770dc3fbdd281d15c2ad46588c139696f9269c/uvloop/handles/handle.pyx#L118-L119
• https://github.com/MagicStack/uvloop/blob/6c770dc3fbdd281d15c2ad46588c139696f9269c/uvloop/cbhandles.pyx#L14-L15

This does remind me of https://github.com/python/cpython/issues/94694

I wonder if the extract_stack function is tolerant of encountering negative numbers in the offset extraction of the trace.

Stack report:
stack.txt