- Proper gate of
Pin(since 1.39 - we are not using onlyPin, but alsoPin::into_inner, #197).
- Some more careful orderings (#195).
- Support for Pin (#185, #183).
- Fix (hopefully) crash on ARM (#164).
- Fix Miri check (#186, #156).
- Fix support for Rust 1.31.0.
- Some minor clippy lints.
- Fix docs build (mutually exclusive features, #112).
- Support for no-std builds with the
experimental-thread-local. Needs nightly compiler. No stability guarantees with this feature (#93).
- Fix a data race reported by MIRI.
- Avoid violating stacked borrows (AFAIK these are still experimental and not normative, but better safe than sorry). (#80).
- The
AccessConvertwrapper is needed less often in practice (#77).
-
bug: Insufficient synchronization on weak platforms (#76).
Never observed in practice (it's suspected practical weak platforms like ARM are still stronger than the model), but still technically UB.
-
docs: Mention triomphe's
ThinArcaround the fat-pointer limitations.
- Support serde (by a feature).
- Allow const-initializing ArcSwapOption (
const_emptymethod).
- More helpful description of the
AsRawtrait (isn't implemented for ownedArc/Option<Arc>).
- Cache doc improvements.
- Allow mapping of DynAccess.
- Fix some lints.
- Don't leave threads running in tests/doctests. It's a bad form and annoys miri.
- Miri and 32 bit tests in CI.
- Making the writers lock-free. Soft-removing the IndependentStrategy, as it is no longer needed (hidden and the same as the DafultStrategy).
- Fix soundness bug around access::Map. Technically a breaking change, but unlikely to bite and breaking seems to be the least bad option. #45.
- Remove Clone implementation. People are often confused by it and it is easy to emulate by hand in the rare case it is actually needed.
- Get rid of the
load_signal_safe. It only complicates things and it is niche; signal-hook-registry has its own simplified version. - Avoid
from_ptr(as_ptr()). Slight change inRefCnt::incwhich technically is API breaking change, but this one should not matter in practice. - Extend documentation about clone behaviour.
- Few more traits for Guard (
From<T: RefCnt>,Default). - Get rid of
rcu_unwap, the whole concept is a trap. - Hide the whole gen lock thing.
- Introduce the
Strategy, as a high level way to choose how exactly the locking happens.- Not possible to implement by downstream users just yet, or call them.
- The CaS is its own trait for flexibility.
- Adding the SimpleGenLock experimental strategy.
- Not part of stability guarantees.
- Rename the
unstable-weaktoweakfeature. The support is now available on 1.45 (currently in beta).
- Adjust to
Weak::as_ptrfrom std (the weak pointer support, relying on unstable features). - Support running on miri (without some optimizations), so dependencies may run miri tests.
- Little optimization when waiting out the contention on write operations.
- Added
Guard::from_inner.
- Top-level docs rewrite (less rambling, hopefully more readable).
- Fix the
Displayimplementation onGuardto correctly delegate to the underlyingDisplayimplementation.
- The Access functionality ‒ ability to pass a handle to subpart of held data to somewhere with the ability to update itself.
- Mapped cache can take
FnMutas well asFn.
- Mapped caches ‒ to allow giving access to parts of config only.
- Support for Weak pointers.
- RefCnt implemented for Rc.
- Breaking: Big API cleanups.
- Peek is gone.
- Terminology of getting the data unified to
load. - There's only one kind of
Guardnow. - Guard derefs to the
Arc/Option<Arc>or similar. Cachegot moved to top level of the crate.- Several now unneeded semi-internal traits and trait methods got removed.
- Splitting benchmarks into a separate sub-crate.
- Minor documentation improvements.
- Prevention against UB due to dropping Guards and overflowing the guard counter (aborting instead, such problem is very degenerate anyway and wouldn't work in the first place).
- Tweak slot allocation to take smaller performance hit if some leases are held.
- Increase the number of lease slots per thread to 8.
- Added a cache for faster access by keeping an already loaded instance around.
- Fix Send/Sync for Guard and Lease (they were broken in the safe but uncomfortable direction ‒ not implementing them even if they could).
Lease<Option<_>>::unwrap(),expect()andinto_option()for convenient use.
- Use the correct
#[deprecated]syntax.
- Another locking store (
PrivateSharded) to complement the global and private unsharded ones. - Comparison to other crates/approaches in the docs.
- Updates to documentation, made it hopefully easier to digest.
- Added the ability to separate gen-locks of one ArcSwapAny from others.
- Some speed improvements by inlining.
- Simplified the
leasemethod internally, making it faster in optimistic cases.
- Another potentially weak ordering discovered (with even less practical effect than the previous).
- Increased potentially weak ordering (probably without any practical effect).
- Documentation link fix.
- Few convenience constructors.
- More tests (some randomized property testing).
compare_and_swapno longer takes&Guardas current as that is a sure way to create a deadlock.- Introduced
Leasefor temporary storage, which doesn't suffer from contention likeload, but doesn't block writes likeGuard. The downside is it slows down with number of held by the current thread. compare_and_swapandrcuuses leases.- Made the
ArcSwapas small as the pointer itself, by making the shards/counters and generation ID global. This comes at a theoretical cost of more contention when different threads use different instances.
- Added an
ArcSwapOption, which allows storing NULL values (as None) as well as a valid pointer. compare_and_swapaccepts borrowedArcascurrentand doesn't consume one ref count.- Sharding internal counters, to improve performance on read-mostly contented scenarios.
- Providing
peek_signal_safeas the only async signal safe method to use inside signal handlers. This removes the footgun with dropping theArcreturned fromloadinside a signal handler.
- The
peekmethod to use theArcinside without incrementing the reference count. - Some more (and hopefully better) benchmarks.
- Documentation fix (swap is not lock-free in current implementation).
- More freedom in the
rcuandrcu_unwrapreturn types.
rcusupport.compare_and_swapsupport.- Added some primitive benchmarks.
- Initial implementation.