Week 9 - Threading, Send, Sync, and the Memory Model¶

9.1 Conceptual Core¶

Send = "safe to transfer ownership to another thread."
Sync = "safe to share by &T across threads" (equivalently: &T: Send).
These are auto traits: structurally derived. A struct is Send if all its fields are Send, etc. Opt-out by including a PhantomData<*const ()> or by using !Send/!Sync negative impls (nightly).
Memory model: Rust's memory model is, today, the C++20 memory model (consume order excluded). You must internalize the happens-before relation, the meaning of Relaxed, Acquire, Release, AcqRel, SeqCst, and the difference between synchronization and visibility.

9.2 Mechanical Detail¶

std::thread::spawn vs thread::scope: scoped threads (stable since 1.63) allow borrowing from the parent stack frame because the join is guaranteed before the borrow ends. This is the right default for most parallelism.
Atomics (AtomicUsize, AtomicPtr<T>, AtomicBool, etc.): the API surface is small but every method takes an Ordering. The orderings are:
Relaxed-no synchronization; only atomicity. Used for counters where ordering doesn't matter (e.g., metrics).
Acquire / Release-pairwise synchronization. A Release store synchronizes with an Acquire load that observes the value, establishing happens-before from the store-side's prior writes to the load-side's subsequent reads.
AcqRel-both, on RMW (read-modify-write) operations.
SeqCst-total order across all SeqCst operations system-wide. Strongest, slowest, safest if you do not yet know what you're doing.
fence: standalone memory barriers. Rarely needed; usually a smell that the abstraction should be different.
Spin loops: std::hint::spin_loop() emits pause (x86) / yield (ARM). Critical for short-wait spinlocks; missing this hint hurts SMT performance and burns power.

9.3 Lab-"A Correct Spinlock"¶

Implement Spinlock<T> from scratch using AtomicBool: - lock() spins with Relaxed load, then Acquire CAS. - unlock() Release stores false. - Returns a SpinlockGuard<'_, T> whose Drop unlocks. - Verify with loom (run all interleavings-see week 10) that no two threads enter the critical section.

9.4 Idiomatic & Clippy Drill¶

clippy::mutex_atomic, clippy::mutex_integer, clippy::needless_collect, clippy::should_implement_trait. The first two flag the anti-pattern of Mutex<bool> where AtomicBool suffices.

9.5 Production Hardening Slice¶

Add a loom feature gate to your spinlock crate. Run cargo test --features loom in CI. Fail the build on any model-checker violation.