Race conditions in Tkinter with non-threaded Tcl

[native-api]

BPO	33257
Nosy	@terryjreedy, @roseman, @serhiy-storchaka, @native-api
PRs	bpo-33257: Fix race conditions for non-threaded Tcl #6444 [2.7] bpo-33257: Fix race conditions for non-threaded Tcl (GH-6444) #6972 bpo-33479: Remove unqualified tkinter threadsafe claim. #6990
Files	TkinterCrash2-2.py: Py2 version, crashing TkinterCrash3-2-2.py: Py3 version, working TkinterHanders.py TkinterHanders3.py 0001-Build-non-threaded-debug-Tcl.zip: For testing only! Build master (win64) with nonthreaded debug Tcl 0001-build-2.7-with-threaded-Tcl.patch: For testing only! Build 2.7 (win64) with threaded Tcl test_threads_tk.py: reproduce hang when creating multiple Tk's

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields:

assignee = None
closed_at = None
created_at = <Date 2018-04-10.16:41:09.432>
labels = ['3.8', '3.7', 'expert-tkinter', 'type-crash']
title = 'Race conditions in Tkinter with non-threaded Tcl'
updated_at = <Date 2018-10-27.23:10:44.251>
user = 'https://github.com/native-api'

bugs.python.org fields:

activity = <Date 2018-10-27.23:10:44.251>
actor = 'terry.reedy'
assignee = 'none'
closed = False
closed_date = None
closer = None
components = ['Tkinter']
creation = <Date 2018-04-10.16:41:09.432>
creator = 'Ivan.Pozdeev'
dependencies = []
files = ['47529', '47530', '47536', '47561', '47565', '47566', '47620']
hgrepos = []
issue_num = 33257
keywords = ['patch']
message_count = 29.0
messages = ['315172', '315271', '315283', '315285', '315286', '315304', '315878', '315899', '316087', '316089', '316092', '316093', '316095', '316096', '316097', '316098', '316099', '316100', '316103', '316105', '316106', '316241', '316265', '316418', '316451', '316689', '316921', '317875', '328662']
nosy_count = 5.0
nosy_names = ['terry.reedy', 'markroseman', 'PythonInTheGrass', 'serhiy.storchaka', 'Ivan.Pozdeev']
pr_nums = ['6444', '6972', '6990']
priority = 'normal'
resolution = None
stage = 'patch review'
status = 'open'
superseder = None
type = 'crash'
url = 'https://bugs.python.org/issue33257'
versions = ['Python 2.7', 'Python 3.6', 'Python 3.7', 'Python 3.8']

[native-api]

(Marked only 2.7 as affected but this would affect any branch if built with nonthreaded Tcl.)

When running the attached TkinterCrash2-2.py repeatedly with 2.7.14 and 2.7 head, win7 x64, two kinds of errors pop up randomly:

1. Crashes and freezes.
2. Exceptions on the console like:

Exception in thread Thread-14:
Traceback (most recent call last):
  File "C:\Users\Sasha\Documents\cpython\lib\threading.py", line 801, in __boots
trap_inner
    self.run()
  File "../tkt/TkinterCrash2-2.py", line 50, in run
    self.deliverToqueue((self.target, z, y))
  File "../tkt/TkinterCrash2-2.py", line 133, in arrival_122
    new_yz[1])
  File "C:\Users\Sasha\Documents\cpython\lib\lib-tk\Tkinter.py", line 2328, in create_line
    return self._create('line', args, kw)
  File "C:\Users\Sasha\Documents\cpython\lib\lib-tk\Tkinter.py", line 2310, in _create
    *(args + self._options(cnf, kw))))
ValueError: invalid literal for int() with base 10: 'None'

Running the same code with the minimal required changes (attached as TkinterCrash3-2-2.py) under 3.6 (same platform) goes without any errors.

Diagnostics showed:

1. Under debug Python, a crash became an MSVC double-free assertion error. The stacktrace is: Tkapp_Call->Tkapp_CallDeallocArgs->Tcl_DecrRefCount->TclFreeObj->free

2. The exceptions are caused by a <canvas> create line Tk call randomly returning "None" (a string) instead of an integer -- which it should never do according to its doc. Since it happens inconsistently, this also suggests a race condition.

3. In Tkapp_Call and SetVar, Tcl lock isn't held when creating/destroying Tcl objects for the call. The underlying fns modify the global state (free objects list, reference counters), so it should be held.
   These two are the only such places in the code.

I've fixed this, will file the PR shortly. Holding the lock when calling Tkapp_CallDeallocArgs in Tkapp_Call eliminated the crashes and freezes, locking objects' creation in the same fn decreased the number of exceptions greatly, and locking creation in SetVar, too, eliminated them completely.

I did not check if objects created with AsObj in SetVar need to be disposed of like in Tkapp_Call.

Also have no idea how to autotest the fix. If someone does, I'm all ears.

[native-api]

My best idea for a test as of now is to run the example ~20 times in a loop (or continuously for a comparable amount of time & threads) and catch and register any exceptions in all threads as per https://stackoverflow.com/questions/18349614/check-if-a-python-thread-threw-an-exception .

[terryjreedy]

I believe I have seen this exact error message within the last few months, and I found one example on SO. https://stackoverflow.com/questions/21209124/weird-error-that-comes-up-during-python-program-run

[native-api]

@terry.reedy Apparently, this bug has gone unnoticed for years (to be precise, since b5bfb9f in 2002),
and there are all kinds of weird rumors circulating about whether tkinter is thread-safe or not, and what can be run where ( see e.g. https://stackoverflow.com/search?q=tkinter+thread+safe ).
Having run into it in my project, I set out to get to the bottom of this ( RedFantom/mtTkinter#5 ) ...and looks like I did.

[native-api]

Wait a second... I think I noticed another similar bug.

In Tkapp_CallProc. Same case, Tkapp_CallArgs and Tkapp_CallDeallocArgs are called when not holding both locks.

The attached example doesn't use Python event handlers, so it didn't manifest itself in my tests.

[native-api]

Attached reproducing example for event handlers. If launching more than one EventThread, it abort()'s immediately.

[native-api]

Finally dug up the reason but unsure how to fix it.

Here's the trace. The format is `$TID: $FUNCTION [some of the args]'

0x1A34: Tkapp_Call(_object *, _object *)
0x1A34: PythonCmd(void *, Tcl_Interp *, int, const char * *) 0x000000000230f430 0x0000000002d297d0 "47566760dummy_handler" 0x0000000003a28580 "5" 0x0000000003a29140 "??" 0x0000000003a29140 "??"
0x1D7C: Tkapp_Call(_object *, _object *)
0x1D7C: PythonCmd(void *, Tcl_Interp *, int, const char * *) 0x000000000230f430 0x0000000002d297d0 "47566760dummy_handler" 0x0000000003a28580 "5" 0x0000000003a29140 "??" 0x0000000003a29140 "??"
0x1A34: PythonCmd(void *, Tcl_Interp *, int, const char * *) setting result
0x1A34: PythonCmd(void *, Tcl_Interp *, int, const char * *) exit: 0
TclStackFree: incorrect freePtr (0000000002BA6430 != 0000000002BA6A20). Call out of sequence?python_d.exe has triggered a breakpoint.

A second PythonCmd is started when the 1st one is still in progress, then the 1st one returns, the Tcl interpreter tries to unwind the topmost stack frame, and it's the wrong one.

Denying other PythonCmd is out of question (a nested cmd deadlocks). Waiting for them to finish, either (would wait for potentially unlimited time).

Look like the only way is to rearrange Tcl stack frames so that the right one is on top. (Since Tkinter's "interpreter calls" (sequences of Tcl calls done while holding the lock) are supposed to be independent from one another, it doesn't really matter which order the frames are in.)

I seriously doubt that's possible with just the public interface though.

[native-api]

Look like the only way is to rearrange Tcl stack frames so that the right one is on top.

Scratch that. Tkinter allows to execute entire scripts where the order of the frames is important.

[native-api]

Here are the possible courses of action to fix this.

All require design changes, thus can be shot down on a whim. So I'll ask for advice and approval at python-dev first, to be sure that my efforts won't be scrapped.

First, some terminology I'll be using:

• A "(Tcl interpreter) call" (in quotes) is a sequence of interdependent calls into Tcl that a Tkinter routine makes in the course of its action.
• Outstanding "calls" are those from other threads that have not started yet and are currently waiting for the Tcl lock.
• A "downstream" "call" is a call made by a running PythonCmd's Python payload.
• Each call to PythonCmd has a "start" -- before releasing the Tcl lock -- and an "end" -- when it reacquires it to Tcl_SetObjResult() and return
• A PythonCmd is "active" when it's before its "start" and "end"

Now, the fix options proper:

• At the "end", rearrange the Tcl interpreter's stack frames so that the current PythonCmd's one is on top (effectively, make it look as if independent "calls" use independent Tcl stacks. Theoretically possible.)
  • Confirmed this to be impossible with Tcl's public API
• Prohibit other `PythonCmd's while one is already "active", except its "downstream" ones. Options:
  • At the "end", the old PythonCmd checks if there are others "calls" on top of it (will only happen if such "call" has another "active" PythonCmd on top: all other Tkinter routines hold the Tcl lock for their entire duration after the 33257 fix), wait if there are.
    • Problematic: this wait algorithm is biased against the old PythonCmd. Will lead to a potentially infinite wait if there are many threads making Tcl calls that result in PythonCmds.
  • Hold the Tcl lock for the entire duration of a PythonCmd. To allow "downstream" calls, make the Tcl lock reentrant.
    • Requires redesign of about two dozen functions in thread_*.c and pycond.h: there are only non-reentrant locks currently.
    • Behaviour change: wouldn't allow outstanding "calls" while a PythonCmd is "active". This change is transparent to Python code though.
  • Reuse the threaded versions of Tkapp_* routines: bind a nonthreaded interpreter to a thread and forward calls to it as events like in the threaded case
    • Problematic: visible behaviour change: tk.call()s would now hang without a running mainloop(). Currently, only event handlers don't fire.

Rejected options:

• The new PythonCmd somehow checks at "start" if another one is "active" and waits if it's not a descendant of it.
  • The new PythonCmd's Tcl interpreter frame has already been allocated when it gets control, so the old one cannot return before the new one in any case.

[gvanrossum]

So this issue doesn't occur when linked with a thread-aware Tcl/Tk, right? Maybe we should just make sure that's the only configuration we ensure?

[Vano]

So this issue doesn't occur when linked with a thread-aware Tcl/Tk,
right? Maybe we should just make sure that's the only configuration we
ensure?

This would break compatibility, including some usage patterns (see the
"Reuse the threaded versions" option above).
Can this actually be considered for anything short of the next major
release?
While official Windows releases use a pocket version, POSIX ones would
probably link against whatever a distribution provides.