asyncio: awaiting StreamWriter.wait_closed() in an except block rewrites the in-flight exception's traceback in place (same exception object shared between read futures and the close waiter)

Bug description

While debugging a network client I found that the logged tracebacks for connection failures pointed at the cleanup code (writer.close() / await writer.wait_closed()) instead of the read call that actually failed. It turns out that a very common cleanup idiom silently rewrites the caught exception's __traceback__ in place.

When a stream transport dies (e.g. peer sends TCP RST), StreamReaderProtocol.connection_lost sets one and the same exception object on both the reader (and thus any pending read future) and the stream's _closed waiter:

https://github.com/python/cpython/blob/v3.14.3/Lib/asyncio/streams.py#L260-L271

def connection_lost(self, exc):
    reader = self._stream_reader
    if reader is not None:
        if exc is None:
            reader.feed_eof()
        else:
            reader.set_exception(exc)
    if not self._closed.done():
        if exc is None:
            self._closed.set_result(None)
        else:
            self._closed.set_exception(exc)

Since bpo-45924 / gh-90082 (the fix for traceback accumulation on repeated Future.result() calls), a future snapshots the exception's traceback at set_exception time and restores it with with_traceback() on every re-raise:

https://github.com/python/cpython/blob/v3.14.3/Lib/asyncio/futures.py#L208

raise self._exception.with_traceback(self._exception_tb)

with_traceback() mutates the exception object. That is harmless when the exception belongs to a single future, but here the same object is owned by two futures, so awaiting the second future (the close waiter) rewrites the traceback of the exception the user is currently handling — even if that second raise is caught/suppressed.

Consequence: the widely used cleanup idiom (the close() + wait_closed() sequence recommended by the asyncio.StreamWriter docs), e.g.

try:
    data = await reader.readexactly(n)      # raises ConnectionResetError E with the true traceback
except Exception:
    writer.close()
    with contextlib.suppress(OSError):
        await writer.wait_closed()          # re-raises the SAME object E; suppressed, but
                                            # Future.result() has already rewritten E.__traceback__
    raise                                    # propagates E with a misleading traceback

produces a final traceback that:

1. points at the wrong raise site — the top-most application frame is the await writer.wait_closed() line, not the readexactly() call that actually failed;
2. is missing intermediate frames — the frames added while E propagated out of readexactly() / _wait_for_data() are gone, replaced by wait_closed()'s frames;
3. has no During handling of the above exception... chain — it is the same object, so no __context__ is attached and nothing hints that a second raise happened.

The result is a log/traceback.format_exc() output that claims the failure happened during cleanup, erasing the actual failing operation. This is quite misleading when debugging production failures.

Reproducer

import asyncio
import socket
import struct
import threading
import traceback
from contextlib import suppress


def fmt(tb):
    return " <- ".join(f"{f.name}:{f.lineno}" for f in traceback.extract_tb(tb))


port_holder = []


def server():
    # Accept one connection, read the request, then close with SO_LINGER(0)
    # so the client gets a TCP RST -> ConnectionResetError.
    srv = socket.socket()
    srv.bind(("127.0.0.1", 0))
    port_holder.append(srv.getsockname()[1])
    srv.listen(1)
    conn, _ = srv.accept()
    conn.recv(4)
    conn.setsockopt(socket.SOL_SOCKET, socket.SO_LINGER, struct.pack("ii", 1, 0))
    conn.close()
    srv.close()


async def main():
    threading.Thread(target=server, daemon=True).start()
    while not port_holder:
        await asyncio.sleep(0.01)
    reader, writer = await asyncio.open_connection("127.0.0.1", port_holder[0])
    try:
        writer.write(b"REQ1")
        await writer.drain()
        await reader.readexactly(4)  # <- the real raise site
    except Exception as e:
        print("BEFORE wait_closed:", fmt(e.__traceback__))
        writer.close()
        try:
            await writer.wait_closed()
        except Exception as e2:
            print("wait_closed raised the identical object:", e2 is e)
        print("AFTER  wait_closed:", fmt(e.__traceback__))
        raise  # propagates e with the rewritten traceback


with suppress(ConnectionResetError):
    asyncio.run(main())

Output on 3.14.3 (3.12.4 and 3.13.5 differ only in stdlib line numbers):

BEFORE wait_closed: main:38 <- readexactly:769 <- _wait_for_data:539 <- _read_ready__data_received:1009
wait_closed raised the identical object: True
AFTER  wait_closed: main:43 <- wait_closed:358 <- _read_ready__data_received:1009

Before wait_closed(), the traceback correctly shows readexactly → _wait_for_data. After it, those frames are gone: the traceback now claims the exception surfaced at the await writer.wait_closed() line (main:43), and readexactly/_wait_for_data have been replaced by wait_closed. The exception object raised by wait_closed() is id-identical to the one being handled.

Expected behavior

Handling (even suppressing) the exception raised by await writer.wait_closed() should not mutate the traceback of the in-flight exception being handled in the except block. Either the traceback restore should not mutate a shared exception object in place, or the two waiters should not share one exception object.

Notes

• The mechanism is the interaction of two individually reasonable behaviors: (a) connection_lost sharing one exception object across the read future and the close waiter (Lib/asyncio/streams.py), and (b) the anti-accumulation traceback snapshot/restore from gh-90082 (Lib/asyncio/futures.py, both the Python and C implementations), which restores via with_traceback() and therefore writes to the shared object.
• gh-154791 is related (it is about the C future clearing its stored traceback after the first result() call) but distinct: this report is about in-place mutation of an exception object that is shared between two futures, observable from an unrelated except block.
• The same pattern presumably affects any place where one exception object is set on multiple futures, not just streams.

CPython versions tested on

3.12.4, 3.13.5, 3.14.3 (identical behavior on all three)

Operating systems tested on

macOS

Linked PRs

• gh-156286

まず再現プログラムを実行し、その後 Lib/asyncio/streams.py と Lib/asyncio/futures.py を読み、レポートで言及されている対応する C の Future 実装も確認してください。connection_lost が例外を共有する方法と、Future.result() がその traceback を復元する方法を追跡してください。cleanup handling が元の実行中の例外の traceback を変更しなくなり、実証された動作に対する回帰テストのカバレッジがあることが完了の条件です。