Example 1:

-----------

def f():
    try:
        raise KeyError
    except Exception:
        yield

gen = f()
gen.send(None)
gen.throw(ValueError)

Output:
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "<stdin>", line 5, in f
ValueError

Expected output:

Something involving the string "During handling of the above exception, another exception occurred", and a traceback for the original KeyError.

Example 2:

-----------

import sys

def f():
    try:
        raise KeyError
    except Exception:
        print(sys.exc_info())
        try:
            yield
        except Exception:
            pass
        print(sys.exc_info())
        raise

gen = f()
gen.send(None)
gen.throw(ValueError)

Output:

(<class 'KeyError'>, KeyError(), <traceback object at 0x7f67ce3c3f88>)
(None, None, None)
Traceback (most recent call last):
  File "/tmp/foo.py", line 17, in <module>
    gen.throw(ValueError)
  File "/tmp/foo.py", line 13, in f
    raise
RuntimeError: No active exception to reraise

Expected output: certainly not that :-)

This seems to happen because normally, generators save the current exc_info when yielding, and then restore it when re-entered. But, if we re-enter through 'throw' (throwflag is true), this is disabled:

This check seems to have been added in ae5f2f4 as a fix for:

https://bugs.python.org/issue7173

which had to do with a nasty situation involving a generator object that was part of a reference cycle: the gc sometimes would free the objects stored in the generator's saved exc_info, and then try to clean up the generator by throwing in a GeneratorExit.

AFAICT this situation shouldn't be possible anymore thanks to PEP-442, which makes it so that finalizers are run before any part of the cycle is freed. And in any case it certainly doesn't justify breaking code like the examples above.

(Note: the examples use generators for simplicity, but of course the way I noticed this was that I had some async/await code where exceptions were mysteriously disappearing instead of showing up in __context__ and couldn't figure out why. It's likely that more people will run into this in the future as async/await becomes more widely used. As a workaround for now I'll probably modify my coroutine runner so that it never uses 'throw'.)

The second example seems like the original complaint in bpo-25612. That spawned a separate bug related to the first situation, which was later closed: bpo-25683.

Yury do you agree this is a bug? Do you have any ideas on how to fix it?

Guido,

The second example ("No active exception to reraise") was an actual long open bug. It was recently fixed by Mark Shannon in [1].

I think we should be able to fix the first case (missing __context__) although it's not critical. I'll look into it after the context PEP and few other open asyncio issues.

[1] #1773

I made a naive attempt at addressing this issue here:
#19811

The code has diverged significantly since Nathaniel first linked to a specific line above. However, what I came up with is simple, and seems to work. But I could very well be missing some things.

It would be great if this could be fixed in some way.

New changeset 2514a63 by Chris Jerdonek in branch 'master':
bpo-29587: Enable implicit exception chaining with gen.throw() (GH-19811)
2514a63

Example 1 is now fixed too by Chris Jerdonek's PR 19811. Thanks Chris!

Oh, no! Buildbot failures. Help?!

New changeset 3c7f9db by Victor Stinner in branch 'master':
Revert "bpo-29587: Enable implicit exception chaining with gen.throw() (GH-19811)" (bpo-19821)
3c7f9db

bpo-29587: Enable implicit exception chaining with gen.throw() (GH-19811)
2514a63

Sorry, I had to revert this change since it broke like 41+ buildbots:
https://pythondev.readthedocs.io/ci.html#revert-on-fail

Almost all CIs passed on the PR (except of the Ubuntu job of Azure Pipelines). That's unusual. No idea why the bug occurs only on some platforms and why it wasn't seen before.

The revert is an opportunity to get more time to investigate the issue, without having the pressure to have to fix the CI "ASAP".

Sorry, I had to revert this change since it broke like 41+ buildbots: (...)

If someone wants to investigate, you can find examples of failed buildbot builds at:
#19811

Oh, no! Buildbot failures. Help?!

Yes, I'm sorry. There's something not so simple going on that I haven't yet reproduced locally. Will reply on the tracker.

Whatever the resolution of this is, it seems to me that this sort of behavior change should not be introduced at this stage of 3.7.x's life. Whether it should go into 3.8.x should be Łukasz's call once the final change is in master and has stabilized.

IMO this is a 3.9 fix.

Could it be running out of memory due to excessive chaining of tracebacks?

(And yes, it's 3.9 only.)

Okay, I was able to get the tests passing on the other platforms.

I'm still not sure why Mac and Fedora behaved differently with my original PR. I was able to come up with a simple script that isolated the behavior difference (posted in the PR comments). It's very strange. Maybe it signals an issue elsewhere in the Python code base.

By the way, I just posted a new issue about exception chaining getting suppressed, but this time in an asyncio setting with ensure_future():
https://bugs.python.org/issue40466

I first noticed this issue two or three years ago and raised it on the async-sig list. It was suggested there that this was a special case of the current issue. Having written code to fix the current issue, though, the ensure_future() issue still exists. So I filed a separate issue.

New changeset 0204726 by Chris Jerdonek in branch 'master':
bpo-29587: Update gen.throw() to chain exceptions (bpo-19823)
0204726

Okay, thanks everyone. I think I'll take a look at bpo-40466 next.

Wow, thanks all! FWIW I agree that ideally things should work with a NULL
value...

--Guido (mobile)

FYI, I just created a PR to add one more test to the prior fix, to test a slightly different aspect.

Whereas the test in the first PR checks that exc.__context__ is set after gen.throw() is finished, the new test checks that exc.__context__ is available also from within the generator (while the generator is running). I'm adding this test partly because this behavior is different from the "awaiting on a task" case, which I'm working on next.

Both of these tests were failing prior to the fix, which I confirmed.

FYI, I was able to finish addressing the other exception-chaining cases (the ones covered by the other issue) with a simple change to the fix for this PR.

I moved the call to _PyErr_ChainExceptions() a bit deeper: into gen_send_ex() before the call to _PyEval_EvalFrame(), and behind an if (exc) check.

So I think the tracebacks should improve a lot.

Okay, I was able to remove the NULL value check and get things working with a NULL exception value. I just posted a second follow-on PR that does this (which Guido approved - thanks, Guido!):
#19877

I wanted to tell you what I found since it raises some questions.

To remove that check I had to add the following new check prior to calling _PyErr_ChainExceptions() with the exception info, as a replacement:
gen->gi_exc_state.exc_type != Py_None
Without doing this, code like the following would crash e.g. on Fedora 32 (this is the crash that was happening in the first version of my PR, reduced down):

    def g():
        try:
            raise KeyError
        except KeyError:
            pass

    try:
        yield
    except Exception:
        # Crash happens here e.g. on Fedora 32 but not Mac.
        raise RuntimeError

    gen = g()
    gen.send(None)
    gen.throw(ValueError)

This raises two questions for me:

First, I thought exc_type could only ever be NULL or an exception class. So I'm wondering if this points to a problem elsewhere in the code base.

Second, I don't know why it would crash on Fedora but not Mac. (On Mac, you instead see the following exception chained beneath the ValueError:

TypeError: print_exception(): Exception expected for value, NoneType found )

New changeset 21893fb by Chris Jerdonek in branch 'master':
bpo-29587: allow chaining NULL exceptions in _gen_throw() (GH-19877)
21893fb

Since none of you are subscribed to the other issue except for Nathaniel, I thought I'd let you know I also posted a PR to fix another issue he filed similar to this one (but this time about the stack trace being incorrect for gen.throw() in certain circumstances): https://bugs.python.org/issue29590

That issue is the second of the two issues he cited back in 2017 about gen.throw() being broken:
https://vorpus.org/~njs/misc/trio-language-summit-2017.pdf

tl; dr I wrote PR 19902 to remove the "XXX" comment.

--

Commit 21893fb adds the following code:

    /* XXX It seems like we shouldn't have to check not equal to Py_None
       here because exc_type should only ever be a class.  But not including
       this check was causing crashes on certain tests e.g. on Fedora. */
    if (gen->gi_exc_state.exc_type && gen->gi_exc_state.exc_type != Py_None) { ... }

I don't think that you should mention "Fedora" as a platform impacted by this issue: all platforms should be affected. It's just unclear why the issue was first seen on Fedora.

gen_send_ex() copies tstate->exc_info to gen->gi_exc_state.

tstate->exc_info->exc_type is set at least in the following 3 places in CPython code base:

• ceval.c: POP_EXCEPT opcode: exc_info->exc_type = POP();
• ceval.c: UNWIND_EXCEPT_HANDLER() macro: exc_info->exc_type = POP();
• errors.c: PyErr_SetExcInfo()

I saw in practice POP_EXCEPT and UNWIND_EXCEPT_HANDLER() setting exc_info->exc_type to Py_None (I didn't test PyErr_SetExcInfo()).

_PyErr_GetTopmostException() also handles exc_info->exc_type == Py_None case:

_PyErr_StackItem *
_PyErr_GetTopmostException(PyThreadState *tstate)
{
    _PyErr_StackItem *exc_info = tstate->exc_info;
    while ((exc_info->exc_type == NULL || exc_info->exc_type == Py_None) &&
           exc_info->previous_item != NULL)
    {
        exc_info = exc_info->previous_item;
    }
    return exc_info;
}

--

So exc_type=None is not a bug, but it's done on purpose.

If you don't want to get None in genobject.c, we should modify all places which set tstate->exc_info->exc_type. Problem: the structure currently exposed in the public C API (bpo-40429), and I wouldn't be surprised if Cython or greenlet modify tstate->exc_info directly.

New changeset b0be6b3 by Victor Stinner in branch 'master':
bpo-29587: _PyErr_ChainExceptions() checks exception (GH-19902)
b0be6b3

Would someone be able to approve / take a look at this small PR addressing another aspect of this issue?
#19858
I'm hoping it can be part of 3.9, and the deadline is just a week away.

It builds on the already merged PR to address an "Example 3" of this issue (the "yield from" case as opposed to the "yield" case, which the merged PR addressed):

Example 3:

def f():
    yield

def g():
    try:
        raise KeyError('a')
    except Exception:
        yield from f()

gen = g()
gen.send(None)
gen.throw(ValueError)

Output without PR:

Traceback (most recent call last):
  File "/.../test-example3.py", line 12, in <module>
    gen.throw(ValueError)
  File "/.../test-example3.py", line 8, in g
    yield from f()
  File "/.../test-example3.py", line 2, in f
    yield
ValueError

Output with PR:

Traceback (most recent call last):
  File "/.../test-example3.py", line 6, in g
    raise KeyError('a')
KeyError: 'a'

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "/.../test-example3.py", line 12, in <module>
    gen.throw(ValueError)
  File "/.../test-example3.py", line 8, in g
    yield from f()
  File "/.../test-example3.py", line 2, in f
    yield
ValueError

New changeset 75cd8e4 by Chris Jerdonek in branch 'master':
bpo-29587: Make gen.throw() chain exceptions with yield from (GH-19858)
75cd8e4

New changeset d7184d3 by Chris Jerdonek in branch 'master':
bpo-29587: Add another test for the gen.throw() fix. (GH-19859)
d7184d3