Post-merge review of #12790 demonstrated a real leak: the resume-from-pause
pardon cleared banked misses outright, so a host whose sweep stalls once every
three ticks reset the budget forever and a dead socket was never reaped. The
reviewer ran 300 sweeps against a permanently dead peer with a >1.5x gap every
third tick and observed zero terminate() calls.
Pre-#12790 that required a stall on *every* tick; the counter widened the
pathological window 3x, and the failure mode is permanent non-reaping — the
MAX_WS_CONNECTIONS leak the reaper exists to prevent.
A stalled tick now charges no miss, which is all the original rationale needed
(the client had no chance to answer that probe), but no longer forgives the
misses already banked. A live client still clears its own count by answering
the probe that is still sent on the stalled tick.
The tolerance test's pause case is rewritten to assert the new contract rather
than the old forgive-everything one, and a new test pins the leak directly: a
host stalling every third tick must still reap a dead socket.
Co-authored-by: Jinwoo-H <Jinwoo-H@users.noreply.github.com>
Jinwoo HongandJinwoo-H
ff01fad4ad