[issue40696] "await" hangs in Python3.9.0b1.

[Chris Jerdonek]

Chris Jerdonek  added the comment:

Also, I just want to point out one thing about _PyErr_SetObject(). I believe it 
can detect cycles of arbitrary length (which is what the while loop is for). 
It's just that it can only detect cycles that involve the first node. So for 
things to fail with _PyErr_SetObject(), someone would need to mess with 
exceptions further down the chain. So I *think* hangs should be pretty unlikely 
with the narrower fix.

--

___
Python tracker 

___
___
Python-bugs-list mailing list
Unsubscribe: 
https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com

[issue40696] "await" hangs in Python3.9.0b1.

[Chris Jerdonek]

Chris Jerdonek  added the comment:

I just posted a draft PR that implements the narrower fix:
https://github.com/python/cpython/pull/20287
I confirmed that the Django test passes with it. I also included two regression 
tests: one using only generators, and one more like the Django test that awaits 
a task.

My solution was to update the exception context in gen_send_ex() using 
_PyErr_SetObject() instead of _PyErr_ChainExceptions() -- because 
_PyErr_SetObject() does the cycle detection we've been discussing, and 
_PyErr_ChainExceptions() doesn't.

While _PyErr_SetObject()'s cycle detection isn't complete in that it can't 
detect cycles that begin further down the chain, it's good enough for this case.

--

___
Python tracker 

___
___
Python-bugs-list mailing list
Unsubscribe: 
https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com

[issue40696] "await" hangs in Python3.9.0b1.

[Chris Jerdonek]

Change by Chris Jerdonek :


--
keywords: +patch
pull_requests: +19561
stage:  -> patch review
pull_request: https://github.com/python/cpython/pull/20287

___
Python tracker 

___
___
Python-bugs-list mailing list
Unsubscribe: 
https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com

[issue40696] "await" hangs in Python3.9.0b1.

[Chris Jerdonek]

Chris Jerdonek  added the comment:

Okay, I'll keep it one issue then. Someone else is still welcome to work on the 
more general issue.

Note that there is some chance the narrower fix should happen independent of 
the more general fix. This is because _PyErr_ChainExceptions() (which is the 
call I added for the gen.throw() case) doesn't call the code path that we're 
discussing. Thus, cycles could still wind up being introduced at that call site.

--

___
Python tracker 

___
___
Python-bugs-list mailing list
Unsubscribe: 
https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com

[issue40696] "await" hangs in Python3.9.0b1.

[Yury Selivanov]

Yury Selivanov  added the comment:

> If someone else agrees, I can create a new issue.

I'd keep this one issue, but really up to you. I don't think I have time in the 
next few days to work on what I proposed but would be happy to brainstorm / 
review PRs.

--

___
Python tracker 

___
___
Python-bugs-list mailing list
Unsubscribe: 
https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com

[issue40696] "await" hangs in Python3.9.0b1.

[Chris Jerdonek]

Chris Jerdonek  added the comment:

>From a process perspective, I think we should probably pursue two PR's for 
>this: one for the general issue that affects all Python versions (what Yury is 
>talking about), and something narrower that addresses the 3.9.0b1 case that 
>came up here. I'd like to focus on the latter first. Someone else is welcome 
>to work on the more general issue while I'm doing that.

I'm not 100% sure if the more general issue should be a new bpo issue or not. 
I'm leaning towards separate because it is bigger and there are different 
decisions to be made around backporting, etc, but we should decide that now. If 
someone else agrees, I can create a new issue.

--

___
Python tracker 

___
___
Python-bugs-list mailing list
Unsubscribe: 
https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com

[issue40696] "await" hangs in Python3.9.0b1.

[Yury Selivanov]

Yury Selivanov  added the comment:

Just a note, __context__ cycles can theoretically be longer than 2 nodes. I've 
encountered cycles like `exc.__context__.__context__.__context__ is exc` a few 
times in my life, typically resulting from some weird third-party libraries.

The only solution is to use a `set()` collection to track already visited 
exceptions.

To make it fast I propose to modify the code to:

1. Do a fast traverse with a regular while loop without tracking (no set())

2. If the number of iterations is longer than 100 and there's still no top 
context found -- go to (3)

3. Do a slower implementation with set()

--

___
Python tracker 

___
___
Python-bugs-list mailing list
Unsubscribe: 
https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com

[issue40696] "await" hangs in Python3.9.0b1.

[Chris Jerdonek]

Chris Jerdonek  added the comment:

I don't think that would be a real solution because it looks like that would 
cause the while loop always to loop at most once (which would defeat its 
purpose) -- because the loop ends with "o = context":

while ((context = PyException_GetContext(o))) {
Py_DECREF(context);
if (context == value) {
PyException_SetContext(o, NULL);
break;
}
o = context;
}

--

___
Python tracker 

___
___
Python-bugs-list mailing list
Unsubscribe: 
https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com

[issue40696] "await" hangs in Python3.9.0b1.

[Batuhan Taskaya]

Batuhan Taskaya  added the comment:

> Is this a separate bug? So maybe the issue is that the new code is letting 
> things get into this state. Some of my changes added new chaining in various 
> places, so that would fit (but still investigating).

Looks like there isn't a recursion guard on 
https://github.com/python/cpython/blob/e572c7f6dbe5397153803eab256e4a4ca3384f80/Python/errors.c#L143-L154

I'm not sure if this would be the real solution but, for this case, it works 
diff --git a/Python/errors.c b/Python/errors.c
index 3b42c1120b..ba3df483e2 100644
--- a/Python/errors.c
+++ b/Python/errors.c
@@ -141,8 +141,8 @@ _PyErr_SetObject(PyThreadState *tstate, PyObject 
*exception, PyObject *value)
usually very short. Sensitive readers may try
to inline the call to PyException_GetContext. */
 if (exc_value != value) {
-PyObject *o = exc_value, *context;
-while ((context = PyException_GetContext(o))) {
+PyObject *o = exc_value, *context = NULL;
+while (o != context && (context = PyException_GetContext(o))) {
 Py_DECREF(context);
 if (context == value) {
 PyException_SetContext(o, NULL);
(END)

--
nosy: +BTaskaya

___
Python tracker 

___
___
Python-bugs-list mailing list
Unsubscribe: 
https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com

[issue40696] "await" hangs in Python3.9.0b1.

[Chris Jerdonek]

Chris Jerdonek  added the comment:

The Django details might not matter so much at this point, but to add to 
something I said above: It might not only be process_exception_by_middleware() 
as I mentioned, but also asgiref's sync_to_async() function. In that function, 
you can see an already active exception being re-raised (here the exc_info 
comes from sys.exc_info()):

# If we have an exception, run the function inside the except block
# after raising it so exc_info is correctly populated.
if exc_info[1]:
try:
raise exc_info[1]
except:
return func(*args, **kwargs)
else:
return func(*args, **kwargs)

https://github.com/django/asgiref/blob/edd0570a4f6e46f0948afa5ef197a192bb95b7b7/asgiref/sync.py#L306-L314

--

___
Python tracker 

___
___
Python-bugs-list mailing list
Unsubscribe: 
https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com

[issue40696] "await" hangs in Python3.9.0b1.

[Chris Jerdonek]

Chris Jerdonek  added the comment:

To start out sharing what I found in the Django code:

Here inside BaseHandler._get_response_async():
https://github.com/django/django/blob/3460ea49e839fd6bb924c48eaa1cd3d6dc888035/django/core/handlers/base.py#L226-L232

try:
response = await wrapped_callback(request, *callback_args,
  **callback_kwargs)
except Exception as e:
response = await sync_to_async(  # This line hangs.
self.process_exception_by_middleware,
thread_sensitive=True,
)(e, request)

you can see an exception being handled, which is then passed to 
process_exception_by_middleware(). Process_exception_by_middleware() can wind 
up re-raising that same exception, which causes __context__ to be set 
circularly inside the except block:
https://github.com/django/django/blob/3460ea49e839fd6bb924c48eaa1cd3d6dc888035/django/core/handlers/base.py#L323-L332

If you boil this down, you get the following as a simple reproducer. This 
doesn't hang, but you can tell the difference by comparing exc2 to 
exc2.__context as indicated below:

import asyncio

async def process_exc(exc):
raise exc

async def run():
try:
raise RuntimeError
except Exception as exc:
task = asyncio.create_task(process_exc(exc))
try:
await task
except BaseException as exc2:
# Prints True in 3.9.0b1 and False in 3.9.0a6.
print(exc2 is exc2.__context__)

loop = asyncio.new_event_loop()
try:
loop.run_until_complete(run())
finally:
loop.close()

The cause is probably the following PR, which enabled exception chaining for 
gen.throw() in the yield from case:
https://github.com/python/cpython/pull/19858
So the answer might be to do some cycle detection when chaining the exception, 
which apparently _PyErr_ChainExceptions() doesn't do.

--

___
Python tracker 

___
___
Python-bugs-list mailing list
Unsubscribe: 
https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com

[issue40696] "await" hangs in Python3.9.0b1.

[Chris Jerdonek]

Chris Jerdonek  added the comment:

FWIW, I found that the following hangs, but it also hangs on earlier versions 
of Python:

import traceback

try:
raise RuntimeError
except Exception as exc:
print(f'handling: {exc!r}')
exc.__context__ = exc
print('printing traceback')
print(traceback.format_exc())  # hangs

Is this a separate bug? So maybe the issue is that the new code is letting 
things get into this state. Some of my changes added new chaining in various 
places, so that would fit (but still investigating).

--

___
Python tracker 

___
___
Python-bugs-list mailing list
Unsubscribe: 
https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com

[issue40696] "await" hangs in Python3.9.0b1.

[Emmanuel Arias]

Change by Emmanuel Arias :


--
nosy: +eamanu

___
Python tracker 

___
___
Python-bugs-list mailing list
Unsubscribe: 
https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com

[issue40696] "await" hangs in Python3.9.0b1.

[Chris Jerdonek]

Chris Jerdonek  added the comment:

I'm getting close to tracking this down. There is a certain point in the code 
path of the Django test where `exc is exc.__context__` becomes True. I'm 
guessing this is what's causing the hang. I'll try to get a simple reproducer 
(there is a lot of Django code to sort through).

--

___
Python tracker 

___
___
Python-bugs-list mailing list
Unsubscribe: 
https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com

[issue40696] "await" hangs in Python3.9.0b1.

[Pablo Galindo Salgado]

Change by Pablo Galindo Salgado :


--
nosy: +aeros

___
Python tracker 

___
___
Python-bugs-list mailing list
Unsubscribe: 
https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com

[issue40696] "await" hangs in Python3.9.0b1.

[Chris Jerdonek]

Change by Chris Jerdonek :


--
nosy: +chris.jerdonek

___
Python tracker 

___
___
Python-bugs-list mailing list
Unsubscribe: 
https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com

[issue40696] "await" hangs in Python3.9.0b1.

[Mariusz Felisiak]

New submission from Mariusz Felisiak :

We noticed a behavior change in Python3.9.0b1 (it works properly in 
Python3.9.0a6). One of our tests 
`handlers.tests.AsyncHandlerRequestTests.test_suspiciousop_in_view_returns_400`[1]
 hangs on `await`. `/suspicious/` is a view that raises a custom exception 
`SuspiciousOperation`.

[1] 
https://github.com/django/django/blob/8328811f048fed0dd22573224def8c65410c9f2e/tests/handlers/tests.py#L258-L260

--
components: asyncio
messages: 369429
nosy: asvetlov, carltongibson, felixxm, yselivanov
priority: normal
severity: normal
status: open
title: "await" hangs in Python3.9.0b1.
versions: Python 3.9

___
Python tracker 

___
___
Python-bugs-list mailing list
Unsubscribe: 
https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com