Repository navigation
DetailedList - no smooth scroll and crash after 30s - macOS #549
Description
Expected Behavior
Fluent scrolling, no jitter, no crash
Current Behavior
- jitter when scrolling up and down.
- app crashes after 30-60s of scrolling with this error:
/Users/Jonas/.local/share/virtualenvs/toga-Z1tvi9lv/bin/python -m detailedlist
Fatal Python error: deallocating None
Current thread 0x00007fffb2b9b380 (most recent call first):
File "/Users/Jonas/Documents/Programming/PyBee/toga/src/cocoa/toga_cocoa/widgets/internal/refresh.py", line 172 in scrollWheel_
File "/Users/Jonas/Documents/Programming/PyBee/rubicon-objc/rubicon/objc/runtime.py", line 1032 in _objc_method
File "/Users/Jonas/Documents/Programming/PyBee/rubicon-objc/rubicon/objc/runtime.py", line 829 in __call__
File "/Users/Jonas/Documents/Programming/PyBee/rubicon-objc/rubicon/objc/runtime.py", line 907 in __call__
File "/Users/Jonas/Documents/Programming/PyBee/rubicon-objc/rubicon/objc/eventloop.py", line 621 in start
File "/Users/Jonas/Documents/Programming/PyBee/rubicon-objc/rubicon/objc/eventloop.py", line 381 in run
File "/Users/Jonas/Documents/Programming/PyBee/rubicon-objc/rubicon/objc/eventloop.py", line 425 in run_forever
File "/Users/Jonas/Documents/Programming/PyBee/toga/src/cocoa/toga_cocoa/app.py", line 171 in main_loop
File "/Users/Jonas/Documents/Programming/PyBee/toga/src/core/toga/app.py", line 169 in main_loop
File "/Users/Jonas/Documents/Programming/PyBee/toga/examples/detailedlist/detailedlist/__main__.py", line 4 in <module>
File "/usr/local/Cellar/python3/3.6.3/Frameworks/Python.framework/Versions/3.6/lib/python3.6/runpy.py", line 85 in _run_code
File "/usr/local/Cellar/python3/3.6.3/Frameworks/Python.framework/Versions/3.6/lib/python3.6/runpy.py", line 193 in _run_module_as_main
Process finished with exit code 134 (interrupted by signal 6: SIGABRT)Steps to reproduce
- run the
detailedlist.pyexample - scroll up and down for 30-60s.
Your Environment
-
Python Version (3.6.5)
-
Operating System and Version (select from the following and list the specific version number; if your OS is not listed, list that as well)
- macOS - version: 10.13.5 (High Sierra)
- Linux - distro: - version:
- Windows - version:
- Other - name: - version:
-
Toga Target (the type of app you are trying to generate)
- android
- cocoa
- django
- gtk
- iOS
- tvOS
- watchOS
- winforms
- win32
- Other (please specify)
Activity
nehararora commentedon May 6, 2019
The place where the crash happens seems to vary - I've reproduced a couple so far:
lschumm commentedon May 8, 2019
reproed again
lschumm commentedon May 8, 2019
the crash appears to be due to Objective-C deallocating the
Nonesingleton.freakboy3742 commentedon May 10, 2019
@dgelessus If you happen to have some spare time and are looking for a gnarly problem to debug, you might enjoy this one :-)
@lschumm and I spend a few hours trying to narrow it down at the PyCon sprints, but ran out of time. As best as we could work out, there are conditions where invoking send_super() will cause Objective C to decrement the Python reference count to the None singleton. We were able to observe it reliably in the override of [NSScrollView scrollWheel:event] that exists in toga_cocoa/widgets/internal/refresh.py; calls to super that include a complex object (like an NSEvent) exhibit the problem, but references to structs (like NSRect) or supers with no arguments don't (at least - not as far as we could observe).
dgelessus commentedon May 10, 2019
Turns out that the root cause of the crash is a bug in
ctypesitself, which I've reported now (bpo-36880). Whenever actypescallback with restypepy_objectreturnsNone, the refcount ofNoneis decremented once too much.In this example, the bug is triggered a few times for each scroll operation. If this happens often enough,
None's refcount decreases to zero - this is never supposed to happen, so Python crashes.The code that triggers the bug is in rubicon-objc here. In this case we can probably work around the bug by returning an object other than
Nonefrom the callback - I'll try to work on a patch for Rubicon soon, so that we don't have to wait for the bug to be fixed in CPython.Add workaround for bpo-36880 to ctypes_patch
freakboy3742 commentedon May 25, 2019
Thanks to the awesome debugging work by @lschumm and @dgelessus, this bug has now been squashed; Rubicon 0.3.1 includes the code necessary to work around this problem.