USB disconnects with asyncio program that has exception in task

[jepler]

CircuitPython version

Adafruit CircuitPython 8.0.0-alpha.1-96-g3bff36685-dirty on 2022-08-06; Raspberry Pi Pico with rp2040

Code/REPL

# SPDX-FileCopyrightText: 2022 Dan Halbert for Adafruit Industries
#
# SPDX-License-Identifier: MIT

import asyncio
import board
import digitalio

async def blink(pin, interval, count):  # Don't forget the async!
    with digitalio.DigitalInOut(pin) as led:
        led.switch_to_output(value=False)
        for _ in range(count):
            led.value = True
            await asyncio.sleep(interval)  # Don't forget the await!
            led.value = False
            await asyncio.sleep(interval)  # Don't forget the await!

async def crash():
    1/0

async def main():  # Don't forget the async!
    led_task = asyncio.create_task(blink(board.LED, 0.25, 10))
    crash_task = asyncio.create_task(crash())
    await asyncio.gather(led_task, crash_task)  # Don't forget the await!
    print("done")


asyncio.run(main())

Behavior

A traceback is printed, the LED blinks 10 times beause the led_task keeps running, and then the device becomes unresponsive (eventual USB disconnect):

Auto-reload is on. Simply save files over USB to run them or enter REPL to disable.
code.py output:
Traceback (most recent call last):
  File "/lib/asyncio/core.py", line 214, in run_until_complete
  File "code.py", line 19, in crash
ZeroDivisionError: division by zero

[tio 11:47:17] Disconnected

Note that before the 10 blinks have completed, it's possible to ctrl-c and get to a working repl.

Take care when using this code; you may need to know how to start your device in safe mode so that it doesn't just end up in a cycle of freezing.

Description

Situations like this frequently occur during the development of asyncio programs, so it'd be nice if the whole device didn't crash.

Additional information

No response

[jepler]

asyncio version is

__version__ = "0.5.12"

no difference if return_exceptions=True is specified.

[jepler]

Using the pure-Python implementation of Task and TaskQueue, the problem does not occur.

[Neradoc]

I was curious because I have been testing cancellling tasks (catching asyncio.CancelledError after a few trials and errors) and did not notice that issue. Then again it was with never-ending loops. Anyway I did some tests and...

This causes the issue:

async def crash():
    raise Exception("oops")

But this doesn't:

async def crash():
    raise BaseException("oops")

Instead it leads to:

Traceback (most recent call last):
  File "code.py", line 102, in <module>
  File "asyncio/core.py", line 235, in run
  File "asyncio/core.py", line 189, in run_until_complete
  File "code.py", line 93, in crash
BaseException: oops

Code done running.

Press any key to enter the REPL. Use CTRL-D to reload.

[jepler]

When the problem occurs, task_iternext is called while the task "is done" but data is mp_rom_none.

STATIC mp_obj_t task_iternext(mp_obj_t self_in) {
    mp_obj_task_t *self = MP_OBJ_TO_PTR(self_in);
    if (TASK_IS_DONE(self)) {
        // Task finished, raise return value to caller so it can continue.
        nlr_raise(self->data);

This value (a special constant 0x6, which is not a valid pointer) ends up being checked as follows in py/vm.c:

            if (mp_obj_is_subclass_fast(MP_OBJ_FROM_PTR(((mp_obj_base_t*)nlr.ret_val)->type), MP_OBJ_FROM_PTR(&mp_type_StopIteration))) {

however, it's not valid to cast 0x6 to an mp_obj_base_t.

Looking at differences in the Python level between using native task and not using it, I saw that when using native task I saw that the line in gather ts[i] = await ts[i] never returns in the native task case. Instead, the exception is printed on the terminal and .. I'm not sure what happens. Bad things.

I tried "mixing and matching" the native and python-coded Task and TaskQueue, but this combination leads to a different failure. While all crashes in the core are in some sense interesting, this is probably not worth pursuing further:

Assertion 'node->child == NULL && node->next == NULL' failed, at file pairheap.h:88

when using native TaskQueue but Python Task.

[jepler]

It looks like we need some form of 90aaf2d in our asyncio library.

[jepler]

With other changes, this test segfaults on Unix, which may give good light to test by:

$ MICROPYPATH=frozen/Adafruit_CircuitPython_asyncio/:frozen/Adafruit_CircuitPython_Ticks/ ./ports/unix/micropython-coverage asyncio_crash1.py 
Traceback (most recent call last):
  File "frozen/Adafruit_CircuitPython_asyncio//asyncio/core.py", line 241, in run_until_complete
  File "asyncio_crash1.py", line 13, in crash
ZeroDivisionError: division by zero
Segmentation fault

# SPDX-FileCopyrightText: 2022 Dan Halbert for Adafruit Industries
#
# SPDX-License-Identifier: MIT

import asyncio

async def blink(pin, interval, count):  # Don't forget the async!
    for _ in range(count):
        await asyncio.sleep(interval)  # Don't forget the await!
        await asyncio.sleep(interval)  # Don't forget the await!

async def crash():
    1/0

async def main():  # Don't forget the async!
    led_task = asyncio.create_task(blink(None, 0.25, 2))
    crash_task = asyncio.create_task(crash())
    await asyncio.gather(led_task, crash_task)  # Don't forget the await!
    print("done")


asyncio.run(main())

[jepler]

Also relevant is the core change: micropython#8929

This would have prevented the problem from being a hardfault / segmentation fault / hang.