pollfd: route the fds-removal io stop through _lws_event_loop_ops_io()

[saghul]

Third one found while implementing the parallel-connect event lib ops (d31f2d830 "event-loop: HE races") in a custom event lib driving lws off a libuv loop.

__remove_wsi_socket_from_fds() calls event_loop_ops->io() directly, bypassing the parallel routing that _lws_event_loop_ops_io() performs:

	if (context->event_loop_ops->io)
		context->event_loop_ops->io(wsi, LWS_EV_STOP | LWS_EV_READ |
							       LWS_EV_WRITE);

lws_remove_parallel_fd_safely() points wsi->desc and position_in_fds_table at the racer it is removing and then calls __remove_wsi_socket_from_fds(), so this raw call reaches the event lib as "stop everything on this wsi". Event libs resolve their watcher from the wsi — lws's own libuv elops_io_uv() uses wsi_to_priv_uv(wsi)->w_read, i.e. the primary — so it stops the primary socket's watcher, when the racer's is the one being torn down. The lws_plat_delete_socket_from_fds() call right after does go through _lws_event_loop_ops_io() and correctly stops the racer, so the net effect is: racer stopped (intended) and primary stopped (not intended).

The primary is typically still connecting when a racer is removed (e.g. an unroutable AAAA racer failing first), and nothing re-enables its POLLOUT afterwards — _lws_change_pollfd() short-circuits on pa->prev_events == pa->events, and from lws's point of view the primary's fds entry never changed — so the connect attempt hangs until it times out. Symptom on our side was Timed out waiting SSL on hosts with more than one address.

This routes it the same way the two lws_plat_{insert,delete}_socket_into_fds() call sites were routed by 'event-loop: HE races', so the stop lands on the socket wsi->desc actually refers to.

Note this one is not needed by an event lib that resolves its watcher by fd rather than by wsi (which is what we ended up doing), but it does affect the in-tree libuv/libev/libevent/glib/sdevent/uloop libs, which all key off the wsi.

[saghul]

Related happy-eyeballs / parallel-connect fixes from the same investigation, all independent of each other: #3648 (racing fd gets the POLLOUT), #3649 (primary retired when a racer wins), #3650 (fds-removal io stop routed through the parallel ops).

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud