Talc

Talc is a performant and flexible memory allocator, with first class support for no_std and WebAssembly. It's suitable for projects such as operating system kernels, website backends, or arena allocation in single-threaded contexts.

Is your project targeting WASM? Check out usage and comparisons here.

Table of Contents

• Setup
• Benchmarks
• Algorithm
• Testing
• General Usage
• Advanced Usage
• Conditional Features
• Support Me
• Changelog

Setup

Use it as an arena allocator via the Allocator API as follows: ```rust

![feature(allocator_api)]

use talc::*; use core::alloc::{Allocator, Layout};

static mut ARENA: [u8; 10000] = [0; 10000];

fn main () { let talck = Talc::new(ErrOnOom).lock::

talck.allocator().allocate(Layout::new::<[u32; 16]>());

} ```

Or as a global allocator: ```rust

![feature(constmutrefs)]

use talc::*;

static mut ARENA: [u8; 10000] = [0; 10000];

[global_allocator]

static ALLOCATOR: Talck = Talc::new(unsafe { // if we're in a hosted environment, the Rust runtime may allocate before // main() is called, so we need to initialize the arena automatically ClaimOnOom::new(Span::from_array(&mut ARENA)) }).lock();

fn main() { let mut vec = Vec::with_capacity(100); vec.extend(0..300usize); } ```

See General Usage and Advanced Usage for more details.

Benchmarks

Macrobenchmarks (based on galloc's benchmarks)

The original benchmarks have been modified (e.g. replacing rand with fastrand) in order to alleviate the overhead. Additionally, alignment requirements are inversely exponentially frequent, ranging from 2^2 bytes to 2^18, with 2^2 and 2^3 being most common.

Random Actions Benchmark Results

The number of successful allocations, deallocations, and reallocations within the allotted time.

Note that these results are sensitive to the allocation sizes, ratio of allocations to deallocations, and other such factors.

Heap Efficiency Benchmark Results

The average occupied capacity upon first allocation failure when randomly allocating/deallocating/reallocating.

| Allocator | Average Random Actions Heap Efficiency | | --------------------- | -------------------------------------- | | dlmalloc | 99.07% | | talc | 98.87% | | linkedlistallocator | 98.28% | | galloc | 95.86% | | buddy_alloc | 58.75% |

Microbenchmarks (based on simplechunkallocator's benchmark)

Pre-fail allocations account for all allocations up until the first allocation failure, at which point heap pressure has become a major factor. Some allocators deal with heap pressure better than others, and many applications aren't concerned with such cases (where allocation failure results in a panic), hence they are separated out for separate consideration. Actual number of pre-fail allocations can be quite noisy due to random allocation sizes.

``` ignore RESULTS OF BENCHMARK: Talc 1980717 allocation attempts, 1397166 successful allocations, 27321 pre-fail allocations, 1386546 deallocations CATEGORY | OCTILE 0 1 2 3 4 5 6 7 8 | AVERAGE ---------------------|--------------------------------------------------------------------------|--------- All Allocations | 42 63 63 84 84 105 126 210 31752 | 137 ticks Pre-Fail Allocations | 42 63 84 84 84 105 105 126 3465 | 105 ticks Deallocations | 42 84 84 105 210 252 294 420 34062 | 239 ticks

RESULTS OF BENCHMARK: Buddy Allocator 2181289 allocation attempts, 1534468 successful allocations, 19225 pre-fail allocations, 1527694 deallocations CATEGORY | OCTILE 0 1 2 3 4 5 6 7 8 | AVERAGE ---------------------|--------------------------------------------------------------------------|--------- All Allocations | 21 42 42 63 63 63 63 63 288414 | 59 ticks Pre-Fail Allocations | 21 42 42 42 63 63 63 84 288414 | 79 ticks Deallocations | 42 63 63 63 63 84 84 126 21945 | 95 ticks

RESULTS OF BENCHMARK: Dlmalloc 1963524 allocation attempts, 1391789 successful allocations, 26241 pre-fail allocations, 1380568 deallocations CATEGORY | OCTILE 0 1 2 3 4 5 6 7 8 | AVERAGE ---------------------|--------------------------------------------------------------------------|--------- All Allocations | 42 63 84 147 168 189 210 315 25557 | 179 ticks Pre-Fail Allocations | 42 63 105 147 168 189 210 294 2289 | 173 ticks Deallocations | 42 105 126 210 252 294 378 441 62958 | 280 ticks

RESULTS OF BENCHMARK: Galloc 274406 allocation attempts, 200491 successful allocations, 24503 pre-fail allocations, 190673 deallocations CATEGORY | OCTILE 0 1 2 3 4 5 6 7 8 | AVERAGE ---------------------|--------------------------------------------------------------------------|--------- All Allocations | 42 63 84 273 12327 27489 42441 46263 110145 | 19458 ticks Pre-Fail Allocations | 42 42 42 63 63 63 63 861 22344 | 730 ticks Deallocations | 42 63 84 168 231 273 399 756 27153 | 322 ticks

RESULTS OF BENCHMARK: Linked List Allocator 133404 allocation attempts, 103843 successful allocations, 25115 pre-fail allocations, 93590 deallocations CATEGORY | OCTILE 0 1 2 3 4 5 6 7 8 | AVERAGE ---------------------|--------------------------------------------------------------------------|--------- All Allocations | 42 4263 9618 16212 24297 35028 47502 59724 729666 | 29867 ticks Pre-Fail Allocations | 42 819 2310 4095 6426 9198 12810 17955 836325 | 11266 ticks Deallocations | 42 3234 7056 11298 16338 22218 29442 39039 117957 | 19519 ticks ```

Q: Why does Buddy Allocator perform much better here than in the random actions benchmark?

A: The buddy allocator's performance is heavily dependant on the size of allocations in random actions, as it doesn't appear to reallocate efficiently. The microbenchmark results only measure allocation and deallocation, with no regard to reallocation. (The currently-used sizes of 1 to 20000 bytes leads to the results above in Random Actions.)

Algorithm

This is a dlmalloc-style linked list allocator with boundary tagging and bucketing, aimed at general-purpose use cases. Allocation is O(n) worst case, while in-place reallocations and deallocations are O(1).

The main differences compared to Galloc, using a similar algorithm, is that Talc doesn't bucket by alignment at all, assuming most allocations will require at most a machine-word size alignment. Instead, a much broader range of bucket sizes are used, which should often be more efficient.

Additionally, the layout of chunk metadata is rearranged to allow for smaller minimum-size chunks to reduce memory overhead of small allocations. The minimum chunk size is 3 * usize, with a single usize being reserved per allocation.

Testing

Tests on most of the helper types and Talc functions.

Other than that, lots of fuzzing of the allocator.

General Usage

Here is the list of Talc methods: * Constructors: * new * Information: * get_allocated_span - returns the minimum span containing all allocated memory * Management: * claim - claim memory to establishing a new heap * extend - extend the extent of a heap * truncate - reduce the extent of a heap * lock - wraps the Talc in a Talck, which supports the GlobalAlloc and Allocator APIs * Allocation: * malloc * free * grow * shrink

Read their documentation for more info.

Span is a handy little type for describing memory regions, because trying to manipulate Range<*mut u8> or *mut [u8] or base_ptr-size pairs tends to be inconvenient or annoying.

Advanced Usage

The most powerful feature of the allocator is that it has a modular OOM handling system, allowing you to fail out of or recover from allocation failure easily.

As an example, recovering by extending the heap is implemented below.

```rust use talc::*;

struct MyOomHandler { heap: Span, }

impl OomHandler for MyOomHandler { fn handle_oom(talc: &mut Talc, layout: core::alloc::Layout) -> Result<(), ()> { // alloc doesn't have enough memory, and we just got called! we must free up some memory // we'll go through an example of how to handle this situation

    // we can inspect `layout` to estimate how much we should free up for this allocation
    // or we can extend by any amount (increasing powers of two has good time complexity)
    // creating another heap would also work, but this isn't covered here

    // this function will be repeatedly called until we free up enough memory or 
    // we return Err(()) causing allocation failure. Be careful to avoid conditions where 
    // the heap isn't sufficiently extended indefinitely, causing an infinite loop

    // an arbitrary address limit for the sake of example
    const HEAP_TOP_LIMIT: *mut u8 = 0x80000000 as *mut u8;

    let old_heap: Span = talc.oom_handler.heap;

    // we're going to extend the heap upward, doubling its size
    // but we'll be sure not to extend past the limit
    let new_heap: Span = old_heap.extend(0, old_heap.size()).below(HEAP_TOP_LIMIT);

    if new_heap == old_heap {
        // we won't be extending the heap, so we should return Err
        return Err(());
    }

    unsafe {
        // we're assuming the new memory up to HEAP_TOP_LIMIT is allocatable
        talc.oom_handler.heap = talc.extend(old_heap, new_heap);
    }

    Ok(())
}

} ```

Conditional Features

• lock_api (default): Provides the Talck locking wrapper type that implements GlobalAlloc.
• allocator (default): Provides an Allocator trait implementation via Talck.

Support Me

If you find the project useful, please consider donating via Paypal. Thanks!

On the other hand, I'm looking for part-time programming work for which South Africans are eligible. If you know of any suitable vacancies, please get in touch. Here's my LinkedIn.

Changelog

v3.0.1

• Improved documentation
• Improved and updated benchmarks
  • Increased the range of allocation sizes on Random Actions. (sorry Buddy Allocator!)
  • Increased the number of iterations the Heap Efficiency benchmark does to produce more accurate and stable values.

v3.0.0

• Added support for multiple discontinuous heaps! This required some major API changes
  • new_arena no longer exists (use new and then claim)
  • init has been replaced with claim
  • claim, extend and truncate now return the new heap extent
  • InitOnOom is now ClaimOnOom.
  • All of the above now have different behavior and documentation.
• Each heap now has a fixed overhead of one usize at the bottom.

To migrate from v2 to v3, keep in mind that you must keep track of the heaps if you want to resize them, by storing the returned Spans. Read claim, extend and truncate's documentation for all the details.

v2.2.1

• Rewrote the allocator internals to place allocation metadata above the allocation.
  • This will have the largest impact on avoiding false sharing, where previously, the allocation metadata for one allocation would infringe on the cache-line of the allocation before it, even if a sufficiently high alignment was demanded. A marginal/negligible increase in single-threaded performance resulted, too.
• Removed heapexhaustion and replaced heapefficiency benchmarks.
• Improved documentation and other resources.
• Changed the WASM size measurement to include slightly less overhead.

v2.2.0

• Added dlmalloc to the benchmarks.
• WASM should now be fully supported via TalckWasm. Let me know what breaks ;)
  • Find more details here.

v2.1.0

• Tests are now passing on 32 bit targets.
• Documentation fixes and improvements for various items.
• Fixed using lock_api without allocator.
• Experimental WASM support has been added via TalckWasm on WASM targets.

v2.0.0

• Removed dependency on spin and switched to using lock_api (thanks Stefan Lankes)
  • You can specify the lock you want to use with talc.lock::<spin::Mutex<()>>() for example.
• Removed the requirement that the Talc struct must not be moved, and removed the mov function.
  • The arena is now used to store metadata, so extremely small arenas will result in allocation failure.
• Made the OOM handling system use generics and traits instead of a function pointer.
  • Use ErrOnOom to do what it says on the tin. InitOnOom is similar but inits to the given span if completely uninitialized. Implement OomHandler on any struct to implement your own behaviour (the OOM handler state can be accessed from handle_oom via talc.oom_handler).
• Changed the API and internals of Span and other changes to pass miri's Stacked Borrows checks.
  • Span now uses pointers exclusively and carries provenance.
• Updated the benchmarks in a number of ways, notably adding buddy_alloc and removing simple_chunk_allocator.