Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                
Skip to content
/ ferroc Public

A lock-free memory allocator

License

Apache-2.0, MIT licenses found

Licenses found

Apache-2.0
LICENSE-APACHE
MIT
LICENSE-MIT
Notifications You must be signed in to change notification settings

js2xxx/ferroc

Repository files navigation

Ferroc: A Multithread Lock-free Memory Allocator

Build Status Cargo Coverage Documentation License

Ferroc (combined from "ferrum" and "malloc") is a lock-free concurrent memory allocator written in Rust, primarily inspired by mimalloc.

This memory allocator is designed to work as fast as other mainstream memory allocators while providing flexible configurations such as embedded/bare-metal environment integrations.

Examples

If you simply want to utilize another memory allocator, you can use Ferroc as the global allocator with default features:

use ferroc::Ferroc;

#[global_allocator]
static FERROC: Ferroc = Ferroc;

fn main() {
    // Using the global allocator API.
    let _vec = vec![10; 100];

    // Manually allocate memory.
    let layout = std::alloc::Layout::new::<u8>();
    let ptr = Ferroc.allocate(layout).unwrap();
    unsafe { Ferroc.deallocate(ptr, layout) };

    // Immediately run some delayed clean-up operations.
    Ferroc.collect(/* force */false);
}

If you want more control over the allocator, you can disable the default features and enable the ones you need:

ferroc = {version = "*", default-features = false, features = ["base-mmap"]}
#![feature(allocator_api)]

use core::pin::pin;
use ferroc::{
    arena::Arenas,
    heap::{Heap, Context},
    base::Mmap,
};

fn main() {
    let arenas = Arenas::new(Mmap); // `Arenas` are `Send` & `Sync`...
    let cx = pin!(Context::new(&arenas));
    let heap = Heap::new(cx.as_ref()); // ...while `Context`s and `Heap`s are not.

    // Using the allocator API.
    let mut vec = Vec::new_in(&heap);
    vec.extend([1, 2, 3, 4]);
    assert_eq!(vec.iter().sum::<i32>(), 10);

    // Manually allocate memory.
    let layout = std::alloc::Layout::new::<u8>();
    let ptr = heap.allocate(layout).unwrap();
    unsafe { heap.deallocate(ptr.cast(), layout) }.unwrap();

    // Immediately run some delayed clean-up operations.
    heap.collect(/* force */false);
}

Cargo Features

  • Basic features: generic Arenas, Contexts and Heaps;
  • "base-static": Base allocator Static;
  • "base-mmap": Base allocator Mmap based on os-specific virtual memory managers (std and libc required);
  • "global": Global allocator instantiation macros config! and config_mod! (inner thread local statics are leaked by default);
  • "libc": libc dependency (currently required by pthread option in config*! if you want a pthread thread-local destructor);
  • "default": The default global allocator Ferroc provided by Mmap and pthread thread-local destructor (consisting of all the features above);
  • "c": fe_* C functions for C/C++ targets;
  • "c-bindgen": A generated C/C++ header "ferroc.h" in the root directory;
  • "c-override": Override default allocator functions such as malloc/free and operator new/delete, which can be useful for embedding Ferroc in a C/C++ project (see this section for more details);
  • "track-valgrind": Valgrind memory tracking support based on crabgrind, which requires Valgrind's version of at least 3.22;
  • "finer-grained": Add more object size types to small bins, decreasing fragmentation but also the minimal alignment from 16 to 8, potentially leading some programs that need SIMD to fail for misalignment.

Environment variables

There are several additional configurations that can be set via environment variables to control the behavior of Ferroc:

  • FE_SLAB_SHIFT: Set the slab size (in bits) for all the data structures, a.k.a. the minimal allocation unit of arenas. Default is 22.
  • FE_SHARD_SHIFT: Set the shard size (in bits) for all the data structures, a.k.a. the minimal allocation unit of slabs. Default is 16.

Those configurations are not exported as cargo features due to the requirement of additiveness.

Building process for CMake users

  1. Download and install the latest nightly Rust toolchain:
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs |\
    sh -- -y --toolchain nightly --profile minimal -c rust-src
  1. Just cmake and make it:
mkdir target && cd target # Required to be `target`. Don't change it to `build` or other names.
cmake .. && make
  1. If you want to install the library:
sudo make install

Common options like --install-prefix and --config are supported naturally by CMake.

There are also some custom options (via cmake -D) you can enable:

  • FE_TRACK_VALGRIND: See "track-valgrind" above;
  • FE_FINER_GRAINED: See "finer-grained" above.
  • FE_PGO_GATHER: Enable PGO (Profile-Guided Optimization) gathering.
  • FE_PGO_USE: Build with optimization from pre-gathered PGO data (which requires appending llvm-tools to the second line of step 1).

Additionally, environment variables (see the section above) can also be used to control the build.

Benchmarks

Using a subset of mimalloc-bench for benchmarking. Running on my laptop with 16GB of RAM and an Intel i7-10750H CPU @ 2.60GHz. The process is repeated 10 times.

Time consumed:

Time consumed #1 Time consumed #2

Memory consumed:

Memory consumed #1 Memory consumed #2

Caveats

This crate only supports the latest nightly Rust compiler currently and utilizes many unstable features. Use it with care.

License

Licensed under either of

at your option.