Scalable Delayed Dealloc

A scalable lock-free delayed memory reclaimer that emulates garbage collection by keeping track of memory reachability.

Its delayed deallocation algorithm is based on a variant of epoch-based reclamation where retired memory chunks are stored in thread-local storage until certain criteria are met. The crossbeam_epoch crate offers similar functionality, however users will find sdd easier to use as the lifetime of a memory chunk is safely managed. For instance, sdd::AtomicOwned and sdd::Owned retire the contained instance when they are dropped, and sdd::AtomicShared and sdd::Shared retire the instance when the last strong reference is dropped.

Features

• Lock-free epoch-based reclamation.
• Loom support: features = ["loom"].

Examples

This crate can be used without an unsafe block.

use sdd::{suspend, AtomicOwned, AtomicShared, Guard, Ptr, Shared, Tag};
use std::sync::atomic::Ordering::Relaxed;

// `atomic_shared` holds a strong reference to `17`.
let atomic_shared: AtomicShared<usize> = AtomicShared::new(17);

// `atomic_owned` owns `19`.
let atomic_owned: AtomicOwned<usize> = AtomicOwned::new(19);

// `guard` prevents the garbage collector from dropping reachable instances.
let guard = Guard::new();

// `ptr` cannot outlive `guard`.
let mut ptr: Ptr<usize> = atomic_shared.load(Relaxed, &guard);
assert_eq!(*ptr.as_ref().unwrap(), 17);

// `atomic_shared` can be tagged.
atomic_shared.update_tag_if(Tag::First, |p| p.tag() == Tag::None, Relaxed, Relaxed);

// `ptr` is not tagged, so CAS fails.
assert!(atomic_shared.compare_exchange(
    ptr,
    (Some(Shared::new(18)), Tag::First),
    Relaxed,
    Relaxed,
    &guard).is_err());

// `ptr` can be tagged.
ptr.set_tag(Tag::First);

// The ownership of the contained instance is transferred to the return value of CAS.
let prev: Shared<usize> = atomic_shared.compare_exchange(
    ptr,
    (Some(Shared::new(18)), Tag::Second),
    Relaxed,
    Relaxed,
    &guard).unwrap().0.unwrap();
assert_eq!(*prev, 17);

// `17` will be garbage-collected later.
drop(prev);

// `sdd::AtomicShared` can be converted into `sdd::Shared`.
let shared: Shared<usize> = atomic_shared.into_shared(Relaxed).unwrap();
assert_eq!(*shared, 18);

// `18` and `19` will be garbage-collected later.
drop(shared);
drop(atomic_owned);

// `17` is still valid as `guard` keeps the garbage collector from dropping it.
assert_eq!(*ptr.as_ref().unwrap(), 17);

// Execution of a closure can be deferred until all the current readers are gone.
guard.defer_execute(|| println!("deferred"));
drop(guard);

// If the thread is expected to lie dormant for a while, call `suspend()` to allow
// others to reclaim the memory.
suspend();

Memory Overhead

Retired instances are stored in intrusive queues in thread-local storage, and therefore additional space for Option<NonNull<dyn Collectible>> is allocated per instance.

Performance

The average time taken to enter and exit a protected region: 2.4 nanoseconds on Apple M2 Max.

Changelog