fix(storage): shut down dedicated runtime on engine drop to break self-cycle leak

[detail-app]

Detail bug report: View on Detail

Bug

A disk cache built with a dedicated runtime via with_spawner(Spawner::Runtime(...)) (the path used by the README quickstart and examples/hybrid_full.rs) leaks the entire dedicated tokio::Runtime — worker threads, pending tasks, channels, and buffers — when the cache is dropped in a long-running process.

Root cause: the block engine spawns its flusher/reclaimer workers as tasks on that dedicated runtime, and each worker future captures a Spawner clone (Arc<BackgroundShutdownRuntime>). This forms a self-cycle — Arc<BackgroundShutdownRuntime> → Runtime → spawned task future → Spawner(Arc<…>) — that keeps the Arc strong count above zero forever. Since Runtime::shutdown_background() was only ever called from BackgroundShutdownRuntime::drop (which fires on the last Arc clone), it never ran: the runtime was never torn down. The default Spawner::Handle path was unaffected (no Arc<BSR> exists there; the external runtime owns the lifecycle).

Fix

• foyer-common/src/spawn.rs: Redesigned BackgroundShutdownRuntime to hold Mutex<Option<Runtime>> + a stored Handle (spawning now goes through Deref<Target=Handle>; Handle exposes both spawn and spawn_blocking, which are the only methods used in-tree). Added an idempotent BackgroundShutdownRuntime::shutdown() that takes the Runtime out of the Mutex and calls shutdown_background() (or drop(runtime) under madsim); Drop calls it. Added Spawner::shutdown() (no-op for Spawner::Handle). shutdown_background is non-blocking (signals + returns; worker threads self-exit), so it is safe even when called from within the dedicated runtime itself.
• foyer-storage/src/engine/block/engine.rs: Added impl Drop for BlockEngineInner that calls self._spawner.shutdown(), forcing the dedicated runtime to tear down when the engine is dropped. The forced shutdown drops the spawned flusher/reclaimer task futures (releasing their captured Arc<…> clones), so the Arc<BackgroundShutdownRuntime> finally reaches zero.
• foyer-storage/Cargo.toml: Added rt-multi-thread,macros to the tokio dev-dependency so the dedicated-runtime regression tests (which build a new_multi_thread runtime) compile under single-package cargo nextest run -p foyer-storage.

Testing

Added 4 regression tests (gated #[cfg(not(madsim))]) that hold only a Weak<BackgroundShutdownRuntime> across drop(store)/drop(engine) and assert weak.upgrade().is_none() afterward (impossible before the fix — the bug report's repro showed weak.upgrade().is_some()=true):

• store::tests::test_dedicated_runtime_released_on_drop — drop after wait()
• store::tests::test_dedicated_runtime_released_on_close_then_drop — close() does not tear down (still alive after close()), but subsequent drop() releases the runtime
• store::tests::test_dedicated_runtime_no_accumulation — dropping N=4 caches releases all N runtimes (no linear growth)
• engine::block::engine::tests::test_dedicated_runtime_released_on_engine_drop — engine-level teardown works without the Store wrapper

Verification:

• Unit tests, typecheck, lint, and build all pass: foyer-common (36), foyer-storage (25 default / 28 with test_utils incl. fuzzy), foyer-memory (31), foyer (12 incl. hybrid fuzzy). No regressions on the default Spawner::Handle path.
• cargo fmt --all -- --check and nightly ffmt both clean; cargo clippy clean on the touched crates (only the pre-existing environmental chunks_exact_to_as_chunks unknown-lint note).
• madsim build + tests pass (103/103); the #[cfg(madsim)] drop(runtime) shutdown branch is exercised via BlockEngineInner::Drop in every test_store_* test.
• End-to-end: cargo run -p examples --example hybrid_full completes without hanging; cargo run -r -p foyer-bench -- --runtime dedicated ... completes with Close takes: 6.7ms (clean teardown instead of leaking until process exit).

---

Automatic Fixes PRs can be configured here.