Commit Graph

3 Commits

Author SHA1 Message Date
Joseph Doherty 8a941de4cd fix(abcip): keep a sync Dispose, but one that never blocks on the teardown
F10 follow-through: removing IDisposable outright broke a synchronous
drv.Dispose() call site outside this lane (Runtime.Tests
AbCipAlmaNativeAlarmEndToEndTests). The sync Dispose() is back, but it starts
the same teardown as DisposeAsync and returns instead of parking the caller on
DisposeAsync().AsTask().GetAwaiter().GetResult(); a background fault is
logged. The runtime still takes the IAsyncDisposable path (preferred, never
both).

Refs #737
Claude-Session: https://claude.ai/code/session_01Ub9hGmYLQvn9XETzw7CMsT
2026-10-22 14:34:53 -04:00
Joseph Doherty 48b0173ae9 fix(abcip): a reader racing a re-init neither leaks a runtime nor degrades
Found by the H13 stress test (200 re-inits under a concurrent reader): a read
that resolved the device just before the table swap could cold-create a
runtime into the retired state after its teardown had disposed it (leaked
native handle / lease), and could hit a per-tag SemaphoreSlim the teardown had
disposed, which the generic catch reported as a comms fault (Degraded).

- DeviceState.Retire marks the state retired before disposing handles; each
  handle is removed-then-disposed exactly once. A creator re-checks the flag
  after its TryAdd (fenced, Dekker-paired with Retire) and disposes the orphan.
- A failure on a retired device is the teardown's doing: BadCommunicationError
  for that one call, no evict, no health signal.
- Per-tag/RMW gates are no longer disposed (async-only waits hold no handle),
  matching the ConnectGate rationale.

Refs #737
Claude-Session: https://claude.ai/code/session_01Ub9hGmYLQvn9XETzw7CMsT
2026-10-22 14:17:16 -04:00
Joseph Doherty 20cb2a52d6 fix(abcip): build-and-swap device tables, volatile health, async-only dispose
H13: _devices/_devicesByRoutingKey are published by reference swap under a
lifecycle lock and never mutated after publish, so the retained poll engine,
discovery and the actor's GetHostStatuses pull never read a table mid-Clear.
Every retire (init's own release, shutdown, dispose) takes the table it
replaces, so ShutdownAsync racing DisposeAsync tears each state down once, and
a generation check makes an InitializeAsync overtaken by a shutdown publish
nothing (it throws OperationCanceledException) instead of resurrecting probe
loops behind the shutdown. Probe join is one bounded 5 s wait independent of
the caller token.

H21: _health is volatile and read-modify-write updates go through a CAS.

F10: the sync Dispose() (sync-over-async) is removed; DisposeAsync is the
complete, idempotent teardown the runtime calls after ShutdownAsync.

Refs #737
Claude-Session: https://claude.ai/code/session_01Ub9hGmYLQvn9XETzw7CMsT
2026-10-22 14:02:00 -04:00