RFC 1974: global-allocators

libs (traits | types-libstd | attributes | impls-libstd | allocation)

Summary

Overhaul the global allocator APIs to put them on a path to stabilization, and switch the default allocator to the system allocator when the feature stabilizes.

This RFC is a refinement of the previous RFC 1183.

Motivation

The current API

The unstable allocator feature allows developers to select the global allocator which will be used in a program. A crate identifies itself as an allocator with the #![allocator] annotation, and declares a number of allocation functions with specific #[no_mangle] names and a C ABI. To override the default global allocator, a crate simply pulls an allocator in via an extern crate.

There are a couple of issues with the current approach:

A C-style ABI is error prone - nothing ensures that the signatures are correct, and if a function is omitted that error will be caught by the linker rather than compiler.

Allocators have some state, and with the current API, that state is forced to be truly global since bare functions can't carry state.

Since an allocator is automatically selected when it is pulled into the crate graph, it is painful to compose allocators. For example, one may want to create an allocator which records statistics about active allocations, or adds padding around allocations to attempt to detect buffer overflows in unsafe code. To do this currently, the underlying allocator would need to be split into two crates, one which contains all of the functionality and another which is tagged as an #![allocator].

jemalloc

Rust's default allocator has historically been jemalloc. While jemalloc does provide significant speedups over certain system allocators for some allocation heavy workflows, it has has been a source of problems. For example, it has deadlock issues on Windows, does not work with Valgrind, adds ~300KB to binaries, and has caused crashes on macOS 10.12. See this comment for more details. As a result, it is already disabled on many targets, including all of Windows. While there are certainly contexts in which jemalloc is a good choice, developers should be making that decision, not the compiler. The system allocator is a more reasonable and unsurprising default choice.

A third party crate allowing users to opt-into jemalloc would also open the door to provide access to some of the library's other features such as tracing, arena pinning, and diagnostic output dumps for code that depends on jemalloc directly.

Detailed design

Defining an allocator

Global allocators will use the Allocator trait defined in RFC 1398. However Allocator's methods take &mut self since it's designed to be used with individual collections. Since this allocator is global across threads, we can't take &mut self references to it. So, instead of implementing Allocator for the allocator type itself, it is implemented for shared references to the allocator. This is a bit strange, but similar to File's Read and Write implementations, for example.

pub struct Jemalloc;

impl<'a> Allocator for &'a Jemalloc {
    // ...
}

Using an allocator

The alloc::heap module will contain several items:

/// Defined in RFC 1398
pub struct Layout { ... }

/// Defined in RFC 1398
pub unsafe trait Allocator { ... }

/// An `Allocator` which uses the system allocator.
///
/// This uses `malloc`/`free` on Unix systems, and `HeapAlloc`/`HeapFree` on
/// Windows, for example.
pub struct System;

unsafe impl Allocator for System { ... }

unsafe impl<'a> Allocator for &'a System { ... }

/// An `Allocator` which uses the configured global allocator.
///
/// The global allocator is selected by defining a static instance of the
/// allocator and annotating it with `#[global_allocator]`. Only one global
/// allocator can be defined in a crate graph.
///
/// # Note
///
/// For techical reasons, only non-generic methods of the `Allocator` trait
/// will be forwarded to the selected global allocator in the current
/// implementation.
pub struct Heap;

unsafe impl Allocator for Heap { ... }

unsafe impl<'a> Allocator for &'a Heap { ... }

This module will be reexported as std::alloc, which will be the location at which it will be stabilized. The alloc crate is not proposed for stabilization at this time.

An example of setting the global allocator:

extern crate my_allocator;

use my_allocator::{MyAllocator, MY_ALLOCATOR_INIT};

#[global_allocator]
static ALLOCATOR: MyAllocator = MY_ALLOCATOR_INIT;

fn main() {
    ...
}

Note that ALLOCATOR is still a normal static value - it can be used like any other static would be.

The existing alloc_system and alloc_jemalloc crates will likely be deprecated and eventually removed. The alloc_system crate is replaced with the SystemAllocator structure in the standard library and the alloc_jemalloc crate will become available on crates.io. The alloc_jemalloc crate will likely look like:

pub struct Jemalloc;

unsafe impl Allocator for Jemalloc {
    // ...
}

unsafe impl<'a> Allocator for &'a Jemalloc {
    // ...
}

It is not proposed in this RFC to switch the per-platform default allocator just yet. Assuming everything goes smoothly, however, it will likely be defined as System as platforms transition away from jemalloc-by-default once the jemalloc-from-crates.io is stable and usable.

The compiler will also no longer forbid cyclic the cyclic dependency between a crate defining an implementation of an allocator and the alloc crate itself. As a vestige of the current implementation this is only to get around linkage errors where the liballoc rlib references symbols defined in the "allocator crate". With this RFC the compiler has far more control over the ABI and linkage here, so this restriction is no longer necessary.

How We Teach This

Global allocator selection would be a somewhat advanced topic - the system allocator is sufficient for most use cases. It is a new tool that developers can use to optimize for their program's specific workload when necessary.

It should be emphasized that in most cases, the "terminal" crate (i.e. the bin, cdylib or staticlib crate) should be the only thing selecting the global allocator. Libraries should be agnostic over the global allocator unless they are specifically designed to augment functionality of a specific allocator.

Defining an allocator is an even more advanced topic that should probably live in the Nomicon.

Drawbacks

Dropping the default of jemalloc will regress performance of some programs until they manually opt back into that allocator, which may produce confusion in the community as to why things suddenly became slower.

Depending on implementation of a trait for references to a type is unfortunate. It's pretty strange and unfamiliar to many Rust developers. Many global allocators are zero-sized as their state lives outside of the Rust structure, but a reference to the allocator will be 4 or 8 bytes. If developers wish to use global allocators as "normal" allocators in individual collections, allocator authors may have to implement Allocator twice - for the type and references to the type. One can forward to the other, but it's still work that would not need to be done ideally.

In theory, there could be a blanket implementation of impl<'a, T> Allocator for T where &'a T: Allocator, but the compiler is unfortunately not able to deal with this currently.

The Allocator trait defines some functions which have generic arguments. They're purely convenience functions, but if a global allocator overrides them, the custom implementations will not be used when going through the Heap type. This may be confusing.

Alternatives

We could define a separate GlobalAllocator trait with methods taking &self to avoid the strange implementation for references requirement. This does require the duplication of some or all of the API surface and documentation of Allocator to a second trait with only a difference in receiver type.

The GlobalAllocator trait could be responsible for simply returning a type which implements Allocator. This avoids the duplication or the strange implementation for references issues in the other possibilities, but can't be defined in a reasonable way without HKT, and is a somewhat strange layer of indirection.

Unresolved questions

Are System and Heap the right names for the two Allocator implementations in std::heap?

Should std::heap also have free functions which forward to the global allocator?