When we spawn a task, we inject a magic local variable into the top stack frame's local namespace, as part of our tricky control-C magic:
|
coro.cr_frame.f_locals.setdefault(LOCALS_KEY_KI_PROTECTION_ENABLED, system_task) |
However, this doesn't work for async def functions defined in C or Cython, because they don't have a Python frame. So we a simple workaround, we detect this case and inject a no-op Python frame where we can insert our magic local (cf #550):
|
if not hasattr(coro, "cr_frame"): |
|
# This async function is implemented in C or Cython |
|
async def python_wrapper(orig_coro: Awaitable[RetT]) -> RetT: |
|
return await orig_coro |
However, the bug came back at some point with newer versions of Cython:
| File "/root/.pyenv/versions/3.11.6/lib/python3.11/site-packages/trio/_core/_run.py", line 1599, in spawn_impl
| coro.cr_frame.f_locals.setdefault(LOCALS_KEY_KI_PROTECTION_ENABLED, system_task)
| ^^^^^^^^^^^^^^^^^^^^^^
| AttributeError: 'NoneType' object has no attribute 'f_locals'
From the error message, it appears that Cython has added a cr_frame attribute to its fake coroutine objects... just it's None. So it's still not really there, but it's enough to sneak past our hasattr(coro, "cr_frame") test.
I think the fix here is just to replace the if not hasattr(coro, "cr_frame"): with if getattr(coro, "cr_frame", None) is None:. But it would be good to also run a test for this in CI. We do have a test for the interpreter's built-in C coroutines (test_async_function_implemented_in_C) and I originally assumed that would be enough, but obviously it wasn't, so I guess we need to do the legwork to install cython in CI and build a tiny Cython module.
Edit by CoolCat467:
Things to do:
When we spawn a task, we inject a magic local variable into the top stack frame's local namespace, as part of our tricky control-C magic:
trio/src/trio/_core/_run.py
Line 1753 in d4ce2f9
However, this doesn't work for
async deffunctions defined in C or Cython, because they don't have a Python frame. So we a simple workaround, we detect this case and inject a no-op Python frame where we can insert our magic local (cf #550):trio/src/trio/_core/_run.py
Lines 1747 to 1750 in d4ce2f9
However, the bug came back at some point with newer versions of Cython:
From the error message, it appears that Cython has added a
cr_frameattribute to its fake coroutine objects... just it'sNone. So it's still not really there, but it's enough to sneak past ourhasattr(coro, "cr_frame")test.I think the fix here is just to replace the
if not hasattr(coro, "cr_frame"):withif getattr(coro, "cr_frame", None) is None:. But it would be good to also run a test for this in CI. We do have a test for the interpreter's built-in C coroutines (test_async_function_implemented_in_C) and I originally assumed that would be enough, but obviously it wasn't, so I guess we need to do the legwork to install cython in CI and build a tiny Cython module.Edit by CoolCat467:
Things to do: