Activity

1. aisk

   added

   type-bugAn unexpected behavior, bug, or errorAn unexpected behavior, bug, or error

   topic-asyncio

   stdlibStandard Library Python modules in the Lib/ directoryStandard Library Python modules in the Lib/ directory

   on Aug 25, 2026
2. github-project-automation

   added this to asyncioon Aug 25, 2026
3. github-project-automation

   moved this to Todo in asyncioon Aug 25, 2026
4. aisk

   added

   OS-windows

   on Aug 25, 2026
5. Chris0Jeky commented on Nov 5, 2026

   

   Chris0Jeky

   on Nov 5, 2026

   More actions

   A related case on the same path: accept() itself failing with a transient OSError also closes the listening socket for good.

   On the proactor event loop, BaseProactorEventLoop._start_serving() handles any OSError from self._proactor.accept(sock) (or from f.result()) by calling the exception handler ("Accept failed on a socket") and then sock.close() on the listening socket. When the error is transient, such as WinError 10055 (WSAENOBUFS, "the system lacked sufficient buffer space or because a queue was full"), which a Windows host can return briefly under memory pressure, the server never accepts again: the process stays alive and keeps running its other work, but the port refuses every new connection. We hit this on a long-running aiohttp server on a machine under heavy memory pressure.

   Seen on CPython 3.12.10 and 3.14.3 (Windows, default proactor loop); the code path is the same in both.

   Minimal repro (Windows; injects one WinError 10055 from accept, then tries a client connection):

   import asyncio, socket
   
   async def main():
       loop = asyncio.get_running_loop()
       proactor = loop._proactor
       orig, fired = proactor.accept, []
   
       def flaky(sock):
           if not fired:
               fired.append(1)
               raise OSError(10055, "no buffer space", None, 10055)  # WSAENOBUFS, once
           return orig(sock)
   
       proactor.accept = flaky
       loop.set_exception_handler(lambda l, ctx: print("handler:", ctx["message"]))
   
       async def handle(reader, writer):
           writer.write(b"ok"); await writer.drain(); writer.close()
   
       lsock = socket.socket(); lsock.bind(("127.0.0.1", 0)); port = lsock.getsockname()[1]
       server = await asyncio.start_server(handle, sock=lsock)
       await asyncio.sleep(0.5)
       try:
           r, w = await asyncio.wait_for(asyncio.open_connection("127.0.0.1", port), 3)
           print("connected:", await r.read(2)); w.close()
       except OSError as e:
           print("connection refused after one transient accept error:", e)
       server.close()
   
   asyncio.run(main())

   Output on 3.12.10:

   handler: Accept failed on a socket
   connection refused after one transient accept error: [WinError 1225] The remote computer refused the network connection
   

   Retrying after a short backoff instead of closing the socket (keeping the sock.fileno() == -1 case as is) keeps the server alive in this repro. Related: gh-93821 (ERROR_NETNAME_DELETED closing the serving socket).
6. tonghuaroot

   added a commit that references this issue on Nov 6, 2026

   Merge branch 'main' into pythongh-156365-proactor-accept-transport-error

   002fc17