slabmalloc

Simple slab based malloc implementation in rust, in order to provide the necessary interface to rusts liballoc library. slabmalloc only relies on libcore and is designed to be used in kernel level code as the only interface a client needs to provide is the necessary mechanism to allocate and free 4KiB frames (or any other default page-size on non-x86 hardware).

Build

By default this library should compile using cargo build with nightly versions of the Rust compiler.

Add the following line to the Cargo.toml dependencies:

slabmalloc = ...

Due to the use of const_fn, if you use the library with a stable rustc the unstable feature needs to be disabled:

slabmalloc = { version = ..., default_features = false }

Documentation

• API Documentation
• Examples

API Usage

slabmalloc has two main components described here. However, if you just want to implement a GlobalAlloc trait have a look at the provided example.

It provides a ZoneAllocator to allocate arbitrary sized objects:

let object_size = 12;
let alignment = 4;
let layout = Layout::from_size_align(object_size, alignment).unwrap();

// We need something that can provide backing memory
// (4 KiB and 2 MiB pages) to our ZoneAllocator
// (see tests.rs for a dummy implementation).
let mut pager = Pager::new();
let page = pager.allocate_page().expect("Can't allocate a page");

let mut zone: ZoneAllocator = Default::default();
// Prematurely fill the ZoneAllocator with memory.
// Alternatively, the allocate call would return an
// error which we can capture to refill on-demand.
unsafe { zone.refill(layout, page)? };

let allocated = zone.allocate(layout)?;
zone.deallocate(allocated, layout)?;

And a SCAllocator to allocate fixed sized objects:

let object_size = 10;
let alignment = 8;
let layout = Layout::from_size_align(object_size, alignment).unwrap();

// We need something that can provide backing memory
// (4 KiB and 2 MiB pages) to our ZoneAllocator
// (see tests.rs for a dummy implementation).
let mut pager = Pager::new();
let page = pager.allocate_page().expect("Can't allocate a page");

let mut sa: SCAllocator<ObjectPage> = SCAllocator::new(object_size);
// Prematurely fill the SCAllocator with memory.
// Alternatively, the allocate call would return an
// error which we can capture to refill on-demand.
unsafe { sa.refill(page) };

sa.allocate(layout)?;

Performance

No real effort on optimizing or analyzing the performance as of yet. But if you insist, here are some silly, single-threaded benchmark numbers:

test tests::jemalloc_allocate_deallocate       ... bench:           6 ns/iter (+/- 0)
test tests::jemalloc_allocate_deallocate_big   ... bench:           7 ns/iter (+/- 0)
test tests::slabmalloc_allocate_deallocate     ... bench:          16 ns/iter (+/- 0)
test tests::slabmalloc_allocate_deallocate_big ... bench:          16 ns/iter (+/- 1)

For multiple threads it's beneficial to give every thread it's own instance of a ZoneAllocator. The dealloc can deal with deallocating any pointer on any any thread. The SMP design for that (atomic bitfield ops) is probably not that great of an idea compared to say, jemalloc's fully partitioned approach.

On Naming

We call our allocator slabmalloc; however the name can be confusing as slabmalloc differs a bit from the seminal paper by Jeff Bonwick describing the "slab allocator". slabmalloc really is just a malloc implementation with size classes and different allocators per class (a segregated-storage allocator), while incorporating some of the simple and effective ideas from slab allocation.

Some notable differences for folks familiar with the slab allocator:

• The slab allocator constructor asks for an object constructor and destructor function to initialize/deinitialize objects. slabmalloc is more malloc-like; it just deals with memory, not object caching.

• A slab in the slab allocator consists of one or more pages of virtually contiguous memory, carved up into equal-size chunks, with a reference count indicating how many of those chunks have been allocated. Instead, slabmalloc uses a (cache-line sized) bitmap to track objects within a slab. Similarly, the slab allocator builds a linked-list of free objects, whereas slabmalloc scans the bitmap in a slab to find a free slot.

• For large objects, the slab allocator does not embed meta-data within the slab page. Because, you could fit only one 2 KiB buffer on a 4 KiB page with the embedded slab data. Moreover, with large (multi-page) slabs it can not determine the slab data address from the buffer address. So a per-cache hash-table is used to map allocated objects to meta-data memory. In slabmalloc, the meta-data is always at the end of the page. It uses different slab sizes to ensure bigger objects are not allocated on small slab-pages. The problem of determining the slab base address is alleviated in rust as we also receive the object size on deallocations (we determine the size of the underlying slab by looking at the size of the object that is to be freed).