prisma-risk / tsoracle / 26346282296

95%

push

github

fix(driver-paxos): mint a fresh apply-shutdown Notify per start (#232)

StandaloneHost reused a single Arc<Notify> for apply-task shutdown across
its whole lifetime, and stop() signalled it with notify_one(). notify_one
stores a permit when no task is parked on notified(), so a stop() with no
running apply task (stop-before-start, double-stop, or stop after the task
already exited) left a stale permit behind. The next start() spawned an
apply task that consumed the permit on its first select! turn and broke out
of the loop immediately. The runner kept deciding entries, but with no apply
task to drain them the in-memory high-water never advanced and reads waiting
on the apply notifier (current_high_water/submit_advance) hung indefinitely.

Mint a fresh Notify in each start() and publish it through an Option that
stop() takes. A per-start Notify confines every permit to the task it was
minted for, and the Option lets stop() signal nothing when no task is live,
so no permit can survive to poison a later start(). This preserves the
existing notify_one-over-notify_waiters behaviour that handles a stop() while
the apply task is mid-drain.

Add a regression test that poisons each host with a pre-start stop(), then
asserts a decided Advance still folds into every node's high-water.

6 of 6 new or added lines in 1 file covered. (100.0%)

10210 of 10779 relevant lines covered (94.72%)

1497615.96 hits per line