This is an attempt to build useful high-level lock-free data structures, by designing simple, composable primitives and incrementally building complexity.
Most of the data structures built upon these primitives, are designed to perform zero allocation during their primary function. Allocation is only performed during setup, or when cloning handles to the data structures.
This allocation-light design is the primary differentiator from other crates providing lock-free algorithms, and it means that none of the containers provided here are unbounded.
This module contains simple, low-level building blocks, which can be used in isolation.
Currently, this contains:
AppendList
- an append-only list, which can be concurrently iterated.AtomicCell
- aCell
-like type, supporting only an atomicswap
operation.AtomicExt
- a set of extension methods to simplify working with atomics.IndexAllocator
- a type which can assign IDs from a contiguous slab.PrependList
- a prepend-only list, which also supports an atomicswap
operation.
This module creates an abstraction for shared ownership of data, where each owner is automatically assigned a unique ID.
There are multiple implementations:
BoundedIdHandle
is completely lock and allocation free, but has a predefined limit on the number of concurrent owners. Exceeding this limit will cause a panic.ResizingIdHandle
wraps the data structure in aparking_lot::RwLock
. Normal usage is still lock and allocation free, but exceeding the maximum number of concurrent owners will cause a write-lock to be taken, and the data structure to be automatically resized to accomodate the additional owners.
This module contains medium-level data structures, often based on the IdHandle
abstraction.
Currently, this contains:
Storage
- provides storage for values larger than ausize
, for use by other containers.Scratch
- provides a scratch space which each accessor can work in before its changes are atomically made visible to other accessors.AtomicCell
- an alternative to the primitiveAtomicCell
making slightly different trade-offs. It is slightly slower (approximately 15% slower in benchmarks), but can be composed into other data structures based around theIdHandle
abstraction.AtomicCellArray
- functionally equivalent to aVec<AtomicCell<T>>
, but much more memory efficient.MpscQueue
- a multiple-producer, single-consumer queue. This queue does not attempt to be fair, so it's possible for one producer to starve the others. The queue also does not provide a mechanism to "wake up" senders/receivers when it's possible to continue, and so it must be polled.MpmcQueue
- an experimental multiple-producer, multiple-consumer queue. No fairness guarantees, and no wake-up mechanism.
This module contains high-level data structures which are compatible with futures-rs.
Currently, this contains:
MpscQueue
- a multiple-producer, single-consumer queue. This queue is fair, so a single producer cannot starve other producers.MpmcQueue
- a multiple-producer, multiple-consumer queue. This queue is fair for producers, so a single producer cannot starve other producers, and prior to being closed, it is also fair for receivers. Once closed, any receiver can empty the queue.
- Fork it!
- Create your feature branch:
git checkout -b my-new-feature
- Commit your changes:
git commit -am 'Add some feature'
- Push to the branch:
git push origin my-new-feature
- Submit a pull request :D
MIT OR Apache-2.0