Module arena 

Thread-safe bump arena allocator (native only — uses std::sync::Mutex). Thread-safe bump allocator implementing allocator-api2’s Allocator trait.

Why not bumpalo?

[bumpalo::Bump] is !Send + !Sync — bump allocation is inherently single-threaded because the bump pointer is a plain Cell<usize>. BEAM’s tokio multi-threaded runtime sends Arc<Put> between worker threads via channels, which requires Put: Send. If the BTreeMap inside Put uses a !Send allocator, the entire Put becomes !Send and cannot cross thread boundaries.

Design

SyncBumpArena wraps a Mutex<SyncBumpInner> behind an Arc. The mutex is only contended during allocation (construction of BTreeMap nodes). Read-only access to the BTreeMap (lookups, iteration, serialization) does not touch the allocator at all — the Allocator trait’s allocate and deallocate methods are only called during insertion and drop.

The inner state holds a chunk list (Vec<Chunk>) and a cursor pointing into the current chunk. When the current chunk is exhausted, a new chunk is allocated from the global allocator (doubling in size, starting at 4 KiB). deallocate is a no-op — all memory is freed at once when the last Arc reference drops.

Performance

• Allocation: Mutex::lock + pointer bump. The lock is held for nanoseconds (just advancing a cursor). In practice, contention is rare because BTreeMap construction happens in a single actor’s context.
• Drop: O(chunks) — free each chunk. O(1) relative to the number of entries in the BTreeMap. This is the primary win: std’s BTreeMap drop walks every node; ours just frees a few chunks.
• Clone: Arc::clone — one atomic increment.

Safety

The unsafe impl Allocator delegates to SyncBumpInner::allocate, which returns valid, aligned memory from a chunk. The memory remains valid until the Arc<SyncBumpInner> is dropped (i.e., until all clones of the arena are gone). deallocate is a no-op, which is sound for bump allocators.

Structs

SyncBumpArena

A thread-safe bump arena allocator implementing allocator-api2’s Allocator trait.