Skip to content

CHANGELOG.md

Latest commit

 

History

History
134 lines (108 loc) · 6.21 KB

CHANGELOG.md

File metadata and controls

134 lines (108 loc) · 6.21 KB
Raw

Changelog

All notable changes to this crate are documented here. The format is loosely based on Keep a Changelog, and this crate adheres to Semantic Versioning.

1.0.1

Fixed

  • wait_blocking() no longer breaks builds on wasm32 targets. EventListener::wait() (blocking) does not exist on target_family = "wasm" — the method is now gated with #[cfg(all(feature = "std", not(target_family = "wasm")))] so cargo check --all-features --target wasm32-unknown-unknown succeeds.
  • alloc + future is now a testable configuration. tests/future.rs previously imported std::time, std::sync::atomic, and runtime crates unconditionally, so cargo test --no-default-features --features alloc,future failed to compile. Std-only tests are now gated behind #[cfg(feature = "std")], and the manual-poll / remaining / clone / fmt / over-done / add-assign tests use core::* / alloc::* so they compile and run under alloc-only.
  • Counter overflow in add() is now caught in release builds. All three WaitGroup variants (sync, spin, future) used debug_assert! or unchecked += / fetch_add for the overflow check, which silently wrapped in release mode. A wrap from usize::MAX + 1 → 0 could reset the counter and let wait() return prematurely or hang. sync::WaitGroup now uses checked_add(...).expect(...), and spin / future use fetch_update with checked_add so the atomic is never corrupted.

Changed

  • CI now tests feature pairs and wasm with --all-features. Added explicit cargo build/test --no-default-features --features alloc,future and cargo build/test --all-features steps alongside the existing cargo hack --each-feature runs. The cross-build job now also runs cargo check --all-features per target, catching platform-specific conditional-compilation misses like the wasm wait_blocking issue above.

1.0.0

First stable release. Significant breaking changes from 0.9.x — see Migration from 0.9 below before upgrading.

The crate now commits to Semantic Versioning: breaking changes to the public API or a bump of the declared MSRV will require a major version bump.

Breaking changes

  • done() now returns the remaining count, not the previous value. In 0.9.x, future::WaitGroup::done() / no_std::WaitGroup::done() returned the counter value before the decrement, which contradicted the documented behaviour. It now consistently returns the count after decrementing. The blocking WaitGroup was already correct.

  • waitings() renamed to remaining() on all variants. The new name reads naturally (if wg.remaining() > 0 { … }) and is consistent across variants.

  • Old no_std module replaced by spin. The lock-free, atomic-counter variant now lives at wg::spin::WaitGroup. In no_std builds, wg::WaitGroup is a re-export of wg::spin::WaitGroup (backward-compatible for no_std users).

  • future::AsyncWaitGroup renamed to future::WaitGroup. Use the module-qualified name: wg::future::WaitGroup.

  • The alloc feature no longer pulls in crossbeam-utils; the spin backoff is now inlined into the crate.

Fixed

  • Busy-loop in future::WaitGroup::poll. The previous Pending branch called wake_by_ref() before returning Pending, causing the executor to re-poll the future continuously and burn 100% CPU until done() reached zero. The waker is now only registered once (via the listener) and the future properly yields.
  • Wrong return value from done() (see breaking changes above).
  • std::sync::Mutex poisoning no longer panics. The blocking WaitGroup now recovers the guard on poisoning via PoisonError::into_inner — a poisoned counter is not a memory-safety concern, and cascading panics across every thread touching the group were an over-reaction.
  • Redundant notify(usize::MAX) on every done(). The async variant now only notifies waiters when the counter actually reaches zero, instead of on every decrement.
  • crossbeam_utils::Backoff::new() was being reset inside the wait loop, so it never escalated past its first spin budget. Replaced with an inline adaptive backoff that spins and — on std — yields the OS thread once the spin budget is exhausted.
  • Silent over-done no longer triggers the misleading assert_eq!(x, 0). Calling done() on a zero counter remains a silent no-op returning 0.
  • The required-features = ["tokio"] on tests/future.rs pointed at a non-existent feature, so the integration test never ran in CI. Fixed to required-features = ["future"], and the CI job now exercises it.

Changed

  • Memory ordering tightened: add uses Release, done uses AcqRel/ Acquire, remaining / wait loads use Acquire. The previous SeqCst everywhere was unnecessarily strong — Release/Acquire provides the required happens-before edges for this structure.
  • #![forbid(unsafe_code)] is now enforced at the crate level.

Added

  • wg::spin::WaitGroup — a lock-free, atomic-counter WaitGroup available on both std and no_std. Uses an inline adaptive backoff that yields the OS thread on std and spins on pure no_std.
  • Dedicated integration tests for each variant.
  • Compile-time Send + Sync assertions for all three variants.
  • Declared MSRV (rust-version = "1.76.0", driven by parking_lot and triomphe floor requirements).

Migration from 0.9

  • Replace wg::AsyncWaitGroup with wg::future::WaitGroup.
  • Replace .waitings() with .remaining().
  • If you relied on done() returning the pre-decrement value on the async / no_std variants, adjust your code — it now returns the post-decrement value, matching the blocking variant and the documented contract.
  • required-features = ["tokio"] in consumer code should be required-features = ["future"].

0.9.2 and earlier

See the git history for pre-1.0 releases.