fix finalizing and atexit timing by youknowone · Pull Request #6626 · RustPython/RustPython

🏁 Script executed:

# First, let's understand the context of this code
# Find where the finalizing flag is set and where atexit runs

cd crates/vm/src/vm
grep -n "finalizing" mod.rs | head -20

Repository: RustPython/RustPython

Length of output: 184

🏁 Script executed:

# Look for atexit-related code in the vm
grep -r "atexit" crates/vm/src --include="*.rs" | head -20

Repository: RustPython/RustPython

Length of output: 1449

🏁 Script executed:

# Find where finalizing is set to understand the order of operations
grep -r "finalizing.*=" crates/ --include="*.rs" -B2 -A2 | head -40

Repository: RustPython/RustPython

Length of output: 1454

🏁 Script executed:

# Look at the vm initialization and shutdown sequence
grep -n "fn run_unraisable\|fn finalize\|fn exit" crates/vm/src/vm/mod.rs | head -20

Repository: RustPython/RustPython

Length of output: 168

🏁 Script executed:

# Look at the finalization sequence in interpreter.rs
grep -n "finalizing\|atexit" crates/vm/src/vm/interpreter.rs | head -30

Repository: RustPython/RustPython

Length of output: 435

🏁 Script executed:

# Get more context around the finalization and atexit calls
sed -n '1,300p' crates/vm/src/vm/interpreter.rs | tail -100

Repository: RustPython/RustPython

Length of output: 3393

🏁 Script executed:

# Look at the full shutdown/exit handling sequence
rg "fn.*exit|fn.*finali" crates/vm/src/vm/interpreter.rs -A 10 | head -50

Repository: RustPython/RustPython

Length of output: 389

🌐 Web query:

CPython Py_FinalizeEx atexit exceptions handling unraisable hook

💡 Result:

Short summary — how CPython handles exceptions during Py_FinalizeEx / atexit and the unraisable hook

• Py_FinalizeEx runs Python-level shutdown which includes running atexit handlers. Exceptions raised in atexit handlers produce a printed traceback (to sys.stderr) and the last non-SystemExit exception raised is re-raised after all handlers have run. [1][2]

• C-level exit handlers (Py_AtExit) are separate: they are called by Py_FinalizeEx but run after Python internal finalization; they must not call Python APIs. Py_AtExit handlers are for C code, not Python-level cleanup. [3]

• sys.unraisablehook (added in Python 3.8) is the configurable hook used for "unraisable exceptions" (e.g., exceptions in del, weakref callbacks, GC callbacks). By default Python logs these to sys.stderr; you can override sys.unraisablehook to customize handling. [4]

• Important caveat: some "unraisable" errors occur very late in interpreter finalization (after core modules like sys and stderr have been cleared). In that case the custom sys.unraisablehook cannot be called and the error may go unreported (PyErr_WriteUnraisable may be a no‑op if stderr/sys are gone). This limitation is documented and tracked in the issue thread that introduced sys.unraisablehook. [4][5]

Practical implications

• Do not call Py_FinalizeEx recursively or from code that may run during shutdown (atexit handlers, del, etc.). The docs explicitly warn against that. [3]
• If you need reliable reporting of errors during shutdown, avoid relying on very-late finalizers; prefer earlier, explicit cleanup or handle errors before finalization. [3][1][4]

Sources
[1] atexit — Exit handlers (Python docs).
[2] atexit behavior / examples (Python docs, PyMotW).
[3] Initialization, Finalization, and Threads / Py_AtExit (CPython C-API docs).
[4] Issue and PR discussion and docs for sys.unraisablehook (bpo-36829 / sys docs / Victor Stinner blog).
[5] Python tracker discussion about very-late unraisable exceptions during finalization.