lunarmodules / copas / 30297691869

85%

Build Type

push

github

Committed by Tieske

Commit Message

fix(io): stale timeout waiters can block a reused socket's real waiter

A socket's read/write/connect timeout resumed the timed-out coroutine
and dropped the socket from the select set, but never cleared that
coroutine's slot in the socket's FIFO wait queue -- that queue was only
ever drained by the normal readiness path, which the timer bypasses.

If the caller kept the socket after a timeout (eg. returned it to a
connection pool) instead of closing it, the stale entry lingered. A
later coroutine waiting on the same socket queued behind it. When the
socket did become ready, the dispatcher popped and resumed the stale
(often already-finished) coroutine instead of the real waiter, dropped
the socket from the select set, and left the real waiter parked
forever with nothing left to wake it. Fixes #208.

Root cause was the FIFO itself: nothing in copas ever legitimately
waits more than one coroutine deep on a socket per direction -- a
socket is only ever driven by one owner at a time. Replaced the FIFO
with a single-slot claim/release pair on the socket set, and moved the
claim to happen before a coroutine yields (inside copas.receive/send/
receivefrom/receivepartial/connect/dohandshake) rather than after, in
the dispatcher. A conflicting second waiter now gets an immediate
"Operation already in progress" back as a normal return value -- same
shape as "timeout" or "closed" -- instead of being silently abandoned
with no way to observe the failure. This also means the dispatcher
never needs to reach into an unrelated coroutine's pending timer: each
coroutine cancels its own via the existing sto_timeout() path, exactly
like every other early-return branch already does.

Adds regression tests covering both directions (duplicate_read_waiter_
errors, duplicate_write_waiter_errors) in tests/tcptimeout.lua.

Coverage Stats

55 of 74 new or added lines in 1 file covered. (74.32%)

97 existing lines in 5 files now uncovered.

1471 of 1730 relevant lines covered (85.03%)

53656.51 hits per line