Alloy: opt-in tracing garbage collection for Rust

Alloy is a fork of the Rust language with support for opt-in tracing garbage collection (GC) using the Gc<T> type. It is a research project designed to help make writing cyclic data structures easier in Rust.

Alloy is not production-ready. However, it is sufficiently polished to be usable for real programs: it supports GC across multiple threads; has high-quality error messages; and reasonable performance.

:warning: Alloy won't be able trace objects for garbage collection unless you set the #[global_allocator] to use std::gc::GcAllocator.

Using Alloy to write a doubly-linked list

The following example program shows how we can use Alloy's Gc<T> smart pointer to write a doubly-linked list with three nodes:

#![feature(gc)]
use std::gc::{Gc, GcAllocator};
use std::cell::RefCell;

#[global_allocator]
static A: GcAllocator = GcAllocator;

struct Node {
    name: &'static str,
    prev: Option<Gc<RefCell<Node>>>,
    next: Option<Gc<RefCell<Node>>>,
}

fn main() {
    let c = Gc::new(RefCell::new(Node { name: "c", prev: None, next: None}));
    let b = Gc::new(RefCell::new(Node { name: "b", prev: None, next: Some(c)}));
    let a = Gc::new(RefCell::new(Node { name: "a", prev: None, next: Some(b)}));

    // Now patch in the previous nodes
    c.borrow_mut().next = Some(b);
    b.borrow_mut().next = Some(a);
}

This is similar to using Rust's Rc smart pointer, but instead, there is a garbage collector running in the background which will automatically free the Gc values when they're no longer used. There are two main ergonomic benefits to using Alloy:

1. The Gc type is Copy, so new pointers can be created easily without needing to clone them.
2. Alloy supports cyclic references by design, so there's no need to use Weak references.

Building Alloy

Dependencies

Make sure you have installed the dependencies:

• rustup
• python 3 or 2.7
• git
• A C compiler (when building for the host, cc is enough; cross-compiling may need additional compilers)
• curl
• pkg-config if you are compiling on Linux and targeting Linux
• libiconv (already included with glibc on Debian-based distros)
• g++, clang++, or MSVC with versions listed on LLVM's documentation
• ninja, or GNU make 3.81 or later (Ninja is recommended, especially on Windows)
• cmake 3.13.4 or later
• libstdc++-static may be required on some Linux distributions such as Fedora and Ubuntu

Build steps

Building Alloy from source is the same process as building the official Rust compiler from source. For a more detailed guide on how this is done, along with the different configuration options, follow the installation guide from the official Rust repository.

1. Clone the source with git:

   git clone https://github.com/softdevteam/alloy.git
   cd rust
   

2. Configure the build settings:

   ./configure
   

   If you plan to use x.py install to create an installation, it is recommended that you set the prefix value in the [install] section to a directory: ./configure --set install.prefix=<path>

3. Build and install:

   ./x.py build && ./x.py install
   

   When complete, ./x.py install will place several programs into $PREFIX/bin: rustc, the Rust compiler, and rustdoc, the API-documentation tool. By default, it will also include [Cargo], Rust's package manager. You can disable this behavior by passing --set build.extended=false to ./configure.

4. Add the Alloy toolchain to rustup:

   rustup toolchain link alloy /path/to/alloy/rustc
   

Rust programs which use cargo can now be built and run using Alloy instead of the official Rust compiler:

cargo +alloy build

How it works

The Gc<T> smart pointer

Alloy provides a new smart pointer type, Gc<T>, which allows for shared ownership of a value of type T allocated in the heap and managed by a garbage collector.

use std::gc::Gc;

fn main() {
    let a = Gc::new(123);
}

This creates a garbage collected object containing the u64 value 123.

Interior mutability

There is no way to mutate, or obtain a mutable reference (&mut T) to the contents of a Gc<T> once it has been allocated. This is because mutable references must not alias with any other references, and there is no way to know at compile-time whether there is only one Gc reference to the data.

As with other shared ownership types in Rust, interior mutability (e.g. RefCell, Mutex, etc) must be used when mutating the contents inside a Gc:

fn main() {
    let a = Gc::new(RefCell::new(123));
    *a.borrow_mut() = 456; // Mutate the value inside the GC
}

The collector

Alloy uses conservative garbage collection. This means that it does not have any specific knowledge about where references to objects are located. Instead, Alloy will assume that an object is still alive if it can be reached by a value on the stack (or in a register) which, if treated like a pointer, points to an object in the heap. The fields of those objects are then traced using the same approach, until all live objects in the program have been discovered.

This tends to work well in practice, however, it comes with an important caveat: you must not hide references from the GC. For example, data structures such as XOR lists are unsound because Alloy will never be able to reach their objects.

Behind the scenes, Alloy uses the Boehm Demers Weiser GC (BDWGC) for its garbage collection implementation. This supports incremental, generational, parallel (but not concurrent!)¹ collection.

Finalisation

Finalisers are a common component of most tracing GCs which are used to run code for cleanup once an object dies (e.g.~closing a file handle or a database connection).

Alloy takes a novel approach to finalisation compared to previous Rust GCs in that it uses existing drop methods as garbage collection finalisers, saving users the potentially error-prone task of manually writing both destructor and finaliser methods for GC managed objects.

When a Gc's underlying contents becomes unreachable, Alloy will call its finaliser, which means that drop is called on all the component types (in the same way that Rust automatically calls drop in an RAII context).

Finalisation order

To achieve Alloy's goal of making cyclic data structures easier to write, we made the decision to support the finalisation of objects with cycles. This means that no guarantees are made about the order in which finalisers are run. In order for this to be sound, you must not access other Gc objects from inside the drop method a Gc. Alloy ensures that this rule is followed by checking for potential misuses of drop at compile-time with Finaliser Safety Analysis (FSA). Consider the following example:

struct Node {
    next: Gc<usize>,
}

impl Drop for Node {
    fn drop(&mut self) {
        *self.next;
    }
}

fn main {
    let x = Gc::new(Node { Gc::new(123) });
}

If at any point you try to create Gc<Node>, such as in main above, Alloy will throw the following compiler error:

error: `Node { next: Gc::new(123) }` cannot be safely finalized.
  --> src/main.rs
   |
12 |     let x = Gc::new(Node { next: Gc::new(123) });
   |                     ^^^^^^^^^^^^^^^^^^^^^^^^^ has a drop method which cannot be safely finalized.
...
2  |         *self.next;
   |          ---------
   |          |
   |          caused by the expression here in `fn drop(&mut)` because
   |          it uses another `Gc` type.
   |
   = help: `Gc` finalizers are unordered, so this field may have already been dropped.
     It is not safe to dereference.

As this suggests, Finaliser Safety Analysis is only applied to a given drop method if its parent type is used in a Gc.

FSA is conservative and can rule out drop methods that a human can determine are in fact safe to be used as finalisers. For those situations you can unsafe impl the FinalizerSafe trait, which overrides FSA for a given type.

Concurrency-safe finalisation

Alloy runs finalisers on a dedicated finalisation thread. This is because finalisers can potentially run at any time during the execution of a Rust program, and if they try to acquire locks to shared data which is already being held, the program could crash or even deadlock. Finalising on a dedicated thread means that finalisation can wait safely until the user program no longer holds exclusive access. (See Destructors, Finalizers, and Synchronization for a more in-depth discussion of this issue).

However, in order to safely finalise objects on a separate thread, finalisers must access data in a thread-safe way. Alloy ensures that this happens at compile-time by using finaliser safety analysis to check that only values marked Send or Sync are used inside a drop method when used during finalisation.

Known limitations

• Alloy is limited to x86-64 architectures.
• Alloy uses the BDWGC's handlers for the SIGXCPU and SIGPWR signals to co-ordinate pausing threads so that GC can happen. It cannot be used with programs which also catch these signals.
• Alloy does not support semi-conservative collection (i.e. precise tracing through heap allocated struct / enum fields).
• Alloy has only been tested on Linux.

License

Rust is primarily distributed under the terms of both the MIT license and the Apache License (Version 2.0), with portions covered by various BSD-like licenses.

See LICENSE-APACHE, LICENSE-MIT, and COPYRIGHT for details.

Footnotes

1. A concurrent collector is one where threads doing GC work can run at the same time as normal program (i.e. mutator) threads. A parallel garbage collector simply means that the garbage collection workload can be parallelised across multiple worker threads. ↩