asgi: decrement nr_conns on server-initiated connection close (#3661)

[apoorvdarshan]

Summary

Fixes #3661. On the ASGI worker, nr_conns is never decremented for server-initiated connection closes, so it grows with worker age and every graceful stop (SIGTERM, --max-requests autorestart) blocks for the full graceful_timeout — even on an idle worker. This hits any setup with Connection: close / HTTP/1.0-per-request (e.g. nginx proxying to the worker).

Root cause

ASGIProtocol used a single _closed flag for two different jobs:

• _close_transport() sets _closed = True to mark the transport closed;
• connection_lost() used if self._closed: return as its idempotency guard.

A client-initiated close calls connection_lost() directly (_closed is False) → the count is decremented. A server-initiated close runs _close_transport() first, setting _closed = True; when asyncio then calls connection_lost(), the guard returns early and skips self.worker.nr_conns -= 1 (and the rest of the cleanup).

Fix

Introduce a dedicated _conn_lost_handled flag for connection_lost()'s idempotency guard, independent of the transport's _closed state. The cleanup (decrementing nr_conns, disconnect signalling, task cancellation) now runs exactly once regardless of who initiated the close.

Tests

• Added test_server_initiated_close_still_decrements_nr_conns: _close_transport() then connection_lost() must leave nr_conns == 0. It fails on master (stays 1) and passes here.
• tests/test_asgi_disconnect.py (14) and tests/test_asgi_worker.py (32) pass; the full ASGI suite is green (628 passed, 121 skipped); pycodestyle clean.

Happy to add a changelog entry if you'd like one.

Disclosure: prepared with AI assistance; reviewed and verified locally.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

[benoitc]

Reopening to trigger CI, which never fired on this PR. Thanks for the fix, the approach here is the right one.

[benoitc]

Thanks, your analysis and fix were the right ones. Superseded by #3686, which carries your commit over unchanged and still under your authorship, and adds coverage for three gaps: the keepalive close path leaking as well, the cleanup steps the early return also skipped (keepalive timer, reader EOF, body receiver disconnect signal), and the reported symptom itself, that _shutdown() burns the full graceful_timeout.

Closing here in favour of #3686.