[Free-threading] `assert(!_Py_IsImmortal(op))` failure when BRC merge queue races with `sys.intern`

[AnnaAr321]

Bug report

Bug description:

In the free-threaded build, there is a race condition where an object can be queued for Biased Reference Counting (BRC) merging, but become immortal before the queue is processed. This causes a fatal C-level assertion failure:

assert.h assertion failed at Objects/object.c:467 in Py_ssize_t _Py_ExplicitMergeRefcount(PyObject *, Py_ssize_t): !_Py_IsImmortal(op)

*** Check failure stack trace: ***
    @     _Py_ExplicitMergeRefcount
    @     merge_queued_objects
    @     _Py_brc_merge_refcounts
    @     _Py_HandlePending
    @     _PyEval_EvalFrameDefault

Reproduction. Fails when running under ASAN/TSAN and the free-threading build:

import threading
import sys
def test_race():
    # 1 million strings will trap Thread A inside map() for milliseconds,
    # giving Thread B plenty of time to queue references to strings that 
    # Thread A hasn't interned yet.
    strings = ["race_string_massive_" + str(j) for j in range(1000000)]
    
    # Give Thread B a copy of the list.
    shared = list(strings)
    
    def thread_b_func():
        # Thread B clears its list, reducing the shared refcount to 0.
        # Since Thread B is not the owner, they are placed in Thread A's merge queue.
        shared.clear()
    tb = threading.Thread(target=thread_b_func)
    tb.start()
    # Thread A executes entirely in C for milliseconds.
    # Strings are queued by Thread B, AND THEN made immortal by sys.intern!
    list(map(sys.intern, strings))
    tb.join()
if __name__ == "__main__":
    test_race()

CPython versions tested on:

3.14

Operating systems tested on:

Linux

Linked PRs

• gh-152276: Fix assert(!_Py_IsImmortal(op)) failure when BRC merge queue races with sys.intern #152277

[puhhztufefhuzdsadf]

import threading
import sys
def test_race():
# 1 million strings will trap Thread A inside map() for milliseconds,
# giving Thread B plenty of time to queue references to strings that
# Thread A hasn't interned yet.
strings = ["race_string_massive_" + str(j) for j in range(1000000)]

# Give Thread B a copy of the list.
shared = list(strings)

def thread_b_func():
    # Thread B clears its list, reducing the shared refcount to 0.
    # Since Thread B is not the owner, they are placed in Thread A's merge queue.
    shared.clear()
tb = threading.Thread(target=thread_b_func)
tb.start()
# Thread A executes entirely in C for milliseconds.
# Strings are queued by Thread B, AND THEN made immortal by sys.intern!
list(map(sys.intern, strings))
tb.join()

if name == "main":
test_race()