Fast, efficient, and robust memory reclamation for concurrent data structures.
Concurrent data structures are faced with the problem of deciding when it is safe to free memory. Although an object might have been logically removed, other threads that previously loaded it may still have access to it, and thus it is not safe to free immediately. Over the years, many algorithms have been devised to solve this problem. However, most traditional memory reclamation schemes make the tradeoff between performance, efficiency, and robustness. For example, [Epoch Based Reclamation] is fast and lightweight but lacks robustness in that a stalled thread can prevent the reclamation of all retired objects. [Hazard Pointers], another popular scheme, tracks individual pointers, making it efficient and robust but generally much slower.
Another problem that is often not considered is workload balancing. In most reclamation schemes, the thread that retires an object is the one that reclaims it. This leads to unbalanced reclamation in read-dominated workloads; parallelism is degraded when only a fraction of threads are writing. This is especially prevalent with the use of M:N threading models as provided by asynchronous runtimes like [Tokio].
Seize is based on the [Hyaline reclamation scheme]. It uses reference counting to determine when it is safe to free memory. However, counters are only used per batch of retired objects, allowing it to avoid the high overhead incurred by traditional reference counting schemes. Performance is typically on par or better than epoch based schemes, while memory efficiency is similar to hazard pointers. Reclamation is naturally balanced as the thread with the last reference to a batch is the one that frees memory. Epochs are also tracked to protect against stalled threads, making reclamation truly lock-free.
Seize is compatible with all modern hardware that supports common atomic operations such as FAA and CAS.
Seize tries to stay out of your way as much as possible. It works with raw pointers directly instead of creating safe wrapper types that end up being a hassle to work with in practice. Below is a step-by-step guide on how to get started.
Seize avoids the use of global state and encourages creating a designated collector per data structure. Collectors allow you to allocate, protect, and retire objects:
```rust,ignore use seize::Collector;
struct Stack
impl
Seize requires storing some metadata about the global epoch for each object that
is allocated. It also needs to reserve a couple words for retirement lists.
Because of this, values in a concurrent data structure must take the form of
AtomicPtr<seize::Linked<T>>, as opposed to just AtomicPtr<T>.
You can link a value to a collector with the link method. link_boxed is also
provided as a quick way to link an object to a collector, and allocate it:
```rust use std::sync::atomic::{AtomicPtr, Ordering}; use std::mem::ManuallyDrop; use seize::{Collector, Linked};
pub struct Stack
struct Node
impl
// ...
}
} ```
To begin a concurrent operation, you must mark the thread as active by calling
the enter method.
```rust,ignore impl Stack { pub fn push(&self, value: T) { // ...
let guard = self.collector.enter(); // <===
// ...
}
} ```
enter returns a guard that allows you to safely load an atomic pointer. Any
valid pointer loaded through a guard is guaranteed to stay valid until the
guard is dropped, or it is retired by the current thread. Importantly, if a
different thread retires an object while we you hold the guard, the collector
knows not to reclaim the object until the guard is dropped.
```rust,ignore impl Stack { pub fn push(&self, value: T) { let node = self.collector.linkboxed(Node { value: ManuallyDrop::new(value), next: AtomicPtr::new(ptr::nullmut()), });
let guard = self.collector.enter();
loop {
let head = guard.protect(&self.head); // <===
unsafe { (*node).next.store(head, Ordering::SeqCst) }
if self
.head
.compare_exchange(head, node, Ordering::SeqCst, Ordering::SeqCst)
.is_ok()
{
break;
}
}
// drop(guard);
}
} ```
Note that the lifetime of a guarded pointer is logically tied to that of the guard -- when the guard is dropped the pointer is invalidated -- but a raw pointer is returned for convenience.
Objects that have been removed from a data structure can be safely retired through the collector. When no thread holds a reference to it, it's memory will be reclaimed:
```rust,ignore
impl
loop {
let head = guard.protect(&self.head); // <===
if head.is_null() {
return None;
}
let next = guard.protect(&(*head).next); // <===
if self
.head
.compare_exchange(head, next, Ordering::SeqCst, Ordering::SeqCst)
.is_ok()
{
unsafe {
let data = ptr::read(&(*head).data);
self.collector.retire(head, |_| {}); // <===
return Some(ManuallyDrop::into_inner(data));
}
}
}
}
} ```
There are a couple important things to note about retiring an object:
An object can only be retired if it is no longer accessible to any thread that comes after. In the above code example this was ensured by swapping out the node before retiring it. Threads that loaded a value before it was retired are safe, but threads that come after are not.
Unlike in schemes like EBR, a guard does not protect objects retired by the current thread. If no other thread holds a reference to an object it may be reclaimed immediately. This makes the following code unsound:
rust,ignore
let ptr = guard.protect(&node);
collector.retire(ptr, |_| {});
println!("{}", (*ptr).value); // <===== unsound!
Retirement can be delayed until the guard is dropped by calling reclaim on
the guard, instead of on the collector directly:
rust,ignore
let ptr = guard.protect(&node);
guard.retire(ptr, |_| {});
println!("{}", (*ptr).value); // <===== ok!
drop(guard); // <===== ptr is invalidated!
You probably noticed that retire takes a function as a second parameter. This
function is known as a reclaimer, and is run when the collector decides it is
safe to free the retired object. Typically you will pass in a function from the
reclaim module. For example, values allocated with Box can use
reclaim::boxed:
```rust,ignore use seize::reclaim;
impl
The type annotation there is important. It is unsound to pass a reclaimer of a different type than the object being retired.
If you need to run custom reclamation code, you can write a custom reclaimer.
Functions passed to retire are called with a type-erased Link. This is
because retired values are connected to thread-local batches via linked lists,
losing any information. To extract the underlying value from a link, you can
call the cast method.
``rust,ignore
collector.reclaim(value, |link: Link| unsafe {
// SAFETY: the value passed to reclaim was of type
//*mut Linked
// SAFETY: the value was allocated with `link_boxed`
let value = Box::from_raw(ptr);
println!("dropping {}", value);
drop(value);
}); ```