[vstinner]

Fix handle leaks in test_tcl when the test is run with -R 3:3. Call explicitly Tcl_FinalizeThread() in test_returning_deeply_nested() to clean up explicitly per-thread state. The test is run in a thread because of the @run_with_limited_c_stack() decorator.

On Windows, Tcl_CreateInterp() creates an event in the Tcl notifier implementation. The event is stored in per-thread state which is only cleaned up at exit.

Add _tkinter.finalize_thread() function.

• Issue: test_tcl leaked [1, 1, 1] handles on AMD64 Windows Server 2025 Refleaks 3.x #161353

[vstinner]

See #161353 (comment) for the longer rationale.

@serhiy-storchaka @terryjreedy @gpshead: Would you mind to review my fix?

[vstinner]

This fix should be backported to 3.15 once #161071 will be merged.

[vstinner]

An alternative would be to register an atexit callback with threading._register_atexit() to call _tkinter.finalize_thread(). But I don't understand well if Tcl/Tk already cleans the thread state automatically when a thread exits, or only when the process exits.

I mean, I don't understand well if the handle leak (one Windows event) only impacts tests, or call impact real Python code.

[encukou]

!buildbot AMD64.Windows.Server.2025.Refleaks.

[bedevere-bot]

🤖 New build scheduled with the buildbot fleet by @encukou for commit bc80f6a 🤖

Results will be shown at:

https://buildbot.python.org/all/#/grid?branch=refs%2Fpull%2F161368%2Fmerge

The command will test the builders whose names match following regular expression: AMD64.Windows.Server.2025.Refleaks.

The builders matched are:

• AMD64 Windows Server 2025 Refleaks PR

[serhiy-storchaka]

_tkinter.finalize_thread() is a large footgun. It does far more than close that event handle, and the damage is silent until later. I will try to find more focused fix, but I afraid that the only feasible solution is to change the test.

[serhiy-storchaka]

Well, the simplest fix is to skip the test in the leak-hunting mode.

The leak is still here, but it is small, and once per thread.

[vstinner]

_tkinter.finalize_thread() is a large footgun. It does far more than close that event handle, and the damage is silent until later.

In my PR, _tkinter.finalize_thread() is called when the thread exits, so I'm not sure of what you mean by "later"?

Do you mean executing more Tcl/Tk code after calling _tkinter.finalize_thread()?

Would you prefer to rename the function to _tkinter._finalize_thread() (add an underscore prefix) to highlight that it should only be used in tests, or if you really understand what you're doing?

[serhiy-storchaka]

Adding an underscore and changing the docstring will help. But skipping the test is simpler.

[vstinner]

An alternative would be to register an atexit callback with threading._register_atexit() to call _tkinter.finalize_thread().

Oh wait, I misunderstood threading._register_atexit(): it's not per-thread! In fact, I'm looking for #58281 which was no implemented yet.

Adding an underscore and changing the docstring will help. But skipping the test is simpler.

I added an underscore. How do you want to change the docstring? The function is documented at https://www.tcl-lang.org/man/tcl8.2.3/TclLib/Exit.htm as:

Tcl_ExitThread is used to terminate the current thread and invoke per-thread exit handlers. This finalization is done by Tcl_FinalizeThread, which you can call if you just want to clean up per-thread state and invoke the thread exit handlers. Tcl_Finalize calls Tcl_FinalizeThread for the current thread automatically.

I wasn't sure about the "invoke the thread exit handlers" part so I omitted it.

[vstinner]

Adding an underscore and changing the docstring will help. But skipping the test is simpler.

I prefer to not skip the test. It's useful to run it on Refleak buildbots. For example, it catched this leak :-)

The buildbot is broken for 6 days. That's way longer than what's expected on a Tier-1 platform: https://peps.python.org/pep-0011/#tier-1. So I enabled auto-merge to fix the buildbot regression.

IMO this change only hides a real bug. When a Python thread calls Tcl_CreateInterp(), we should call Tcl_FinalizeThread() when the Python thread exits, especially if it's not the main thread. But I leave this as an exercice to someone else more motivated than me to fix the issue :-)