Bug report

Bug description:

I suspect there might be a bug in the garbage collector under free threading. If I run this script

import gc
import threading
import time

def worker():
    while not stop.is_set():
        _ = [dict(x=i) for i in range(1000)]

stop = threading.Event()
threads = [threading.Thread(target=worker) for _ in range(8)]
for t in threads:
    t.start()
time.sleep(2)
stop.set()
for t in threads:
    t.join()

gen0 = gc.get_count()[0]
print(f"gen0 = {gen0}")
assert gen0 >= 0

the assertion fails.

We observe in our full application that gen0 is quite negative, less than -10000 at times. The net effect of this is that garbage collection runs less frequently than it otherwise might, resulting in large GC pauses in our application (~3s). This doesn't seem to result in any functional problems aside from much longer stop-the-world pauses than ideal.

This seems to repro even if I set PYTHON_GIL=1 but not against an older version of python (3.10.12). I'm thus not completely sure if this is intended or not. It seems semantically ok for gen0 to go negative, like if deallocations exceed allocations that doesn't seem wrong, but is it supposed to?

I have reproduced this with 3.14t and cpython built against main.

CPython versions tested on:

3.14

Operating systems tested on:

Linux

Linked PRs

• gh-142048: Reset allocation counts in gc thread-local buffers #142051