gh-143249: Fix buffer leak when overlapped operation fails to start (OS-Windows)

[hyongtao-code]

On Windows, some overlapped I/O operations leak buffers when they fail to start, notably in WSASendTo(), WSARecvFrom(), and WSARecvFromInto().

This change ensures that buffers are released on failure by clearing the overlapped state, following the same pattern used in other overlapped error paths. This issue is similar in nature to commit 5485085, which fixed related overlapped cleanup problems.

Python script for tracking leaks

import sys
import _overlapped

def total():
    return sys.gettotalrefcount()

if not hasattr(sys, "gettotalrefcount"):
    raise SystemExit("need a debug build (Py_DEBUG) for gettotalrefcount")

before = total()

N = 20000
STEP = 1000

for i in range(1, N + 1):
    ov = _overlapped.Overlapped()
    buf = memoryview(bytearray(4096))
    try:
        ov.WSASendTo(0x1234, buf, 0, ("127.0.0.1", 1))
    except OSError:
        pass

    if i % STEP == 0:
        now = total()
        print(f"{i:6d}: totalrefcount delta = {now - before:+d}")

after = total()
print(f"done: N={N}, totalrefcount delta = {after - before:+d}")

Result without the patch

d:\MyCode\cpython\PCbuild\amd64>python_d.exe py_overlapped_leak.py
  1000: totalrefcount delta = +4007
  2000: totalrefcount delta = +8009
  3000: totalrefcount delta = +11906
  4000: totalrefcount delta = +15906
  5000: totalrefcount delta = +19906
......
 19000: totalrefcount delta = +75906
 20000: totalrefcount delta = +79906
done: N=20000, totalrefcount delta = +79905

Result with the patch

d:\MyCode\cpython\PCbuild\amd64>python_d.exe py_overlapped_leak.py
  1000: totalrefcount delta = +10
  2000: totalrefcount delta = +10
  3000: totalrefcount delta = +10
  4000: totalrefcount delta = +10
  5000: totalrefcount delta = +10
......
 19000: totalrefcount delta = +10
 20000: totalrefcount delta = +10
done: N=20000, totalrefcount delta = +9

• Issue: Overlapped I/O leaks buffers when operation fails to start #143249

[picnixz]

You'll need a NEWS entry for that and some tests as well. Please add them so that we can run them under the refleak bot as well.

cc @zooba

[picnixz]

I was very confused with this, I thought it was tp_clear but it's explicitly not tp_clear so this looks like the correct fix.

wrong button

[hyongtao-code]

and some tests as well

Thanks for the review. I will add a suitable test case.

[picnixz]

Do we really need this? I think you can just run -R : from the CLI to check if three is a refleak. It's more robust as well.

[hyongtao-code]

Thanks for the guidance — that makes sense now. I’ve updated the test case, and here are the results from my local run.

Without the patch

d:\MyCode\cpython>PCbuild\amd64\python_d.exe -m test -R 3:3 test_asyncio.test_windows_utils
Using random seed: 514712329
0:00:00 Run 1 test sequentially in a single process
0:00:00 [1/1] test_asyncio.test_windows_utils
beginning 6 repetitions. Showing number of leaks (. for 0 or less, X for 10 or more)
123:456
XX5 555
test_asyncio.test_windows_utils leaked [5, 5, 5] references, sum=15
0:00:01 [1/1/1] test_asyncio.test_windows_utils failed (reference leak)

== Tests result: FAILURE ==

1 test failed:
    test_asyncio.test_windows_utils

Total duration: 1.2 sec
Total tests: run=6
Total test files: run=1/1 failed=1
Result: FAILURE

with the patch

d:\MyCode\cpython>PCbuild\amd64\python_d.exe -m test -R 3:3 test_asyncio.test_windows_utils
Using random seed: 3834115587
0:00:00 Run 1 test sequentially in a single process
0:00:00 [1/1] test_asyncio.test_windows_utils
beginning 6 repetitions. Showing number of leaks (. for 0 or less, X for 10 or more)
123:456
XX. ...
0:00:01 [1/1] test_asyncio.test_windows_utils passed

== Tests result: SUCCESS ==

1 test OK.

Total duration: 1.2 sec
Total tests: run=6
Total test files: run=1/1
Result: SUCCESS

[picnixz]

Suggested change

	ov.WSARecvFromInto(0x1234, buf, len(buf), 0)
	ov.WSARecvFromInto(0x1234, buf, len(buf), 0)