[zero-context]

Independent, narrower alternative to #131009 (the _ServerState enum
redesign) — flagging that up front since #131009 is already open and
maintainer-preferred for this issue. This PR is a small, local guard
that fixes the specific reproduced crash; it doesn't replace the
broader design #131009 is doing.

Server.close() sets self._sockets = None synchronously, but
_accept_connection2() runs as a Task, which always defers its first
step by at least one event-loop iteration. If close() lands in that
gap, _attach()'s assert self._sockets is not None fires.

Reproduced deterministically with real sockets (no mocking): closing
the server after exactly 3 asyncio.sleep(0) yields following a client
connect crashes 300/300 times on current main; every other yield
count, 0/300. The same trigger also produces a second, previously
undocumented crash: a TypeError: 'NoneType' object is not iterable in
_wakeup(), fired during garbage collection when the never-fully-
attached transport's __del__ calls _detach() a second time. That one
doesn't show up in loop.call_exception_handler() output at all —
__del__ failures are reported through a separate path — so it's easy
to miss even while actively looking at this bug. Both crashes are fixed
by the same change, confirmed 300/300 → 0/300 for the first and 300 → 0
occurrences for the second.

--- a/Lib/asyncio/selector_events.py
+++ b/Lib/asyncio/selector_events.py
@@ -218,6 +218,16 @@
         protocol = None
         transport = None
         try:
+            # gh-109564: create_task() defers this coroutine's first
+            # step by at least one loop iteration, so the server can
+            # legitimately close() in the gap between "connection
+            # accepted" and "this task actually runs". Attaching to an
+            # already-closed server crashes deep inside transport
+            # construction (Server._attach()'s assert); check up front
+            # instead and drop the already-accepted connection cleanly.
+            if server is not None and server._sockets is None:
+                conn.close()
+                return
             protocol = protocol_factory()
             waiter = self.create_future()
             if sslcontext:

Also includes two new regression tests in test_selector_events.py
(deterministic, mock-based — no real-socket timing dependency, so
they're portable to CI) and a Misc/NEWS.d entry.

Validated on current main: the new tests fail without the fix and
pass with it; the full test_asyncio suite passes (2,811 tests, 0
failures, same skip set before and after); Tools/patchcheck/patchcheck.py
runs clean.

Why this doesn't implement the #131009 design, and one thing it doesn't cover

• Checked both linked PRs before writing this: gh-109564: Add None check for asyncio.Server._wakeup #126660 (closed,
  unmerged, targeted the _wakeup() double-call symptom specifically)
  and GH-109564: add asyncio.Server state machine #131009 (open, the maintainer-preferred _ServerState enum,
  still awaiting core review). This PR is a narrower, local guard, not
  a replacement for GH-109564: add asyncio.Server state machine #131009's broader design.
• The secondary concern from the issue thread — connection_made()
  firing asynchronously after close() has already returned — isn't
  addressed here.

[python-cla-bot]

All commit authors signed the Contributor License Agreement.

[github-actions]

This PR is stale because it has been open for 90 days with no activity.

[asvetlov]

Superseded by #160851 (the latter fixes both selector and proactor event loops)
Thanks for the PR, @zero-context!