sendable

The sendable crate defines types that facilitate sending interconnected data between threads:

• SendRc, a single-threaded reference-counting pointer that can be sent between threads. You can think of it as a variant of Rc<T> that is Send if T is Send. This is unlike Rc<T> which is never Send, and also unlike Arc<T>, which requires T: Send + Sync to be Send.
• SendOption, which holds an Option<T> and is Send even if T is not Send. It is useful for store reference to a single-threaded arena which is sent separately.

When is SendRc useful?

You should consider SendRc if:

• your values form an acyclic graph or a hierarchy with cross-references;
• you build and use the hierarchy from a single thread;
• you need to occasionally move the whole thing to another thread.

Within the confines of a single thread, Rc and RefCell provide an ergonomic and safe representation of acyclic graphs. They are also efficient because single-threaded manipulation doesn't require atomics or locks, makes deref() trivial, and allows the compiler to inline borrow() and borrow_mut() and even optimize them away where they are not globally observable.

In programs that process many such graphs it comes very useful to be able to create them in one thread and ship them off to another for processing (and possibly to a third one for teardown). Given that types like RefCell and Cell are Send, the idea is not unthinkable. The trouble is with Rc, which is neither Send nor Sync, and for good reason. Even though it would be perfectly safe to move an entire hierarchy of Rc<RefCell<T>>s from one thread to another, the borrow checker doesn't allow it because it cannot statically prove that you have moved all of them. If some Rcs pointing to shared data remained in the original thread, unsynchronized access to the non-Sync cells and unsynchronized manipulation of the reference counts would be undefined behavior and wreak havoc.

If there were a way to demonstrate to Rust that you've sent all Rc<T> pointers to a particular shared value to a different thread, there would be no problem in doing so, as long as T itself is Send. Ability to carry out that demonstration is what SendRc brings to the table.

How does SendRc work?

When a SendRc is constructed, it stores the id of the current thread beside the value and the reference count. Before giving access to the value, and before manipulating the reference count through clone() and drop(), it checks that the current thread is still the expected thread, panicking otherwise.

Before SendRcs are moved to a different thread, each pointer must be explicitly "parked", i.e. registered for sending. Once parked, access to the value it points to is prohibited, even in the original thread. Once all SendRcs pointing to the shared value have been parked, they can be sent across the thread boundary, and re-enabled in the new thread at once. In a simple case of two SendRcs, the process looks like this:

// create two SendRcs pointing to a shared value
let mut r1 = SendRc::new(RefCell::new(1));
let mut r2 = SendRc::clone(&r1);

// prepare to send them to a different thread
let pre_send = SendRc::pre_send();
pre_send.park(&mut r1); // r1 and r2 cannot be dereferenced from this point
pre_send.park(&mut r2);
// ready() would panic if there were unparked SendRcs pointing to the value
let post_send = pre_send.ready();

// move everything to a different thread
std::thread::spawn(move || {
    // SendRcs are still unusable until unparked
    post_send.unpark();
    // they're again usable from this point, and only in this thread
    *r1.borrow_mut() += 1;
    assert_eq!(*r2.borrow(), 2);
});

Why not just use Arc?

Arc indeed allows moves between threads, but it fundamentally assumes that the underlying value will be shared among different threads. Arc requires T: Send + Sync in order for Arc<T> to be Send because if it only required T: Send, you could create an Arc<RefCell<u32>>, clone it, send the clone to a different thread, and call borrow_mut() from two threads on the same RefCell without synchronization. That is forbidden, and is why Arc<RefCell<T>> is not a thing in Rust.

SendRc can get away with it because it guards each access to the data with a thread check. When moving data across threads, it requires proof that all references to the allocated value in the previous thread were relinquished prior to the move. SendRc<RefCell<u32>> is sound because if you clone it and send the clone to a different thread, you won't be able to access the data, nor clone or even drop it - any of those would trigger a panic.

Using the standard library, one could fix the issue by switching to the full-blown Arc<Mutex<T>> or Arc<RwLock<T>>. However, that slows down access to data because it requires strongly-ordered atomics, poison checks, and calls into the pthread API. It also increases memory overhead due to the mandatory allocation of the system mutex. Even the most efficient mutex implementations like parking_lot don't come for free and bear the cost of synchronization. But even disregarding the cost, it's conceptually wrong to use Arc<Mutex<T>> if neither Arc nor Mutex are actually needed because the code doesn't access the value of T from multiple threads in parallel.

In summary, SendRc<T> is Send with certain guarantees enforced at run time, the same way an Arc<Mutex<T>> is Send + Sync with certain guarantees enforced at run time. They just serve different purposes.

Why not use an arena? Or unsafe?

To make an arena Send, the whole design must be devoted to that idea from the ground up. A simple approach of replacing every Rc with an arena id doesn't really work because in addition to the id, the object then needs a reference to the arena. It can't have a field of type Option<&Arena> or Option<Rc<Arena>> because such field would make the type non-Send if the arena contains RefCell.

There are arena-based designs that do work, but require more radical changes, such as decoupling storage of values from access and sharing. All data is then in the arena, and the accessors are created on-the-fly and have a lifetime connected to the lifetime of the arena. This requires dealing with the lifetime everywhere and is not easy to get right for non-experts.

Finally, one can avoid the arena by just using unsafe impl Send on a wrapper type that is used to send the whole world to the new thread, borrow checker be damned. That solution is hacky and gives up the guarantees afforded by Rust. If you make a mistake, say by leaving an Rc clone in the original thread, you're facing undefined behavior and core dumps much like in C++. In Rust we hope to do better, and SendRc is intended to provide a sound solution that addresses this scenario.

What about SendOption?

SendOption is a related proposition: a type that holds Option<T> and is always Send, regardless of whether T is Send. Surely that can't be safe?

What makes it work is that SendOption requires you to set the value to None before sending it to another thread. If the inner Option<T> is None, it doesn't matter if T is not Send because no T is actually getting sent anywhere. If you do send a non-None SendOption<T> into another thread, SendOption will use panic to prevent you from accessing it in any way (including by dropping it). Failure to abide by the rules results in a T that was effectively never "sent" to another thread, only its bits were shallow-copied and forgotten, and that's safe.

SendOption is designed for types which are composed of Send data, except for an optional field of a non-send type. The field is meant to be set and used only inside a particular thread, and will be None while being sent across threads, but since Rust can't prove that, a field of Option<NonSendType> makes the entire outer type non-Send. For example, a field with a SendOption<Rc<Arena>> could be used to create a Send type that refers to a single-threaded arena.

Is this really safe?

sendable adheres to Rust safety guarantees. In particular, it should never be possible to cause undefined behavior with safe code. Failure to abide by the rules that cannot be caught at compile time, e.g. by sending a SendRc to another thread and attempting to use it without the park/unpark procedure, is detected at run time and met with panic.

As with any crate that uses unsafe code, the compiler cannot statically verify that there is no soundness issue. However, the code is fairly straightforward in implementing the design outlined above, and every usage of unsafe is carefully documented. I went through several iterations of the design and the implementation before settling on the current approach and, while I did find the occasional issue, the underlying idea held up under scrutiny. MIRI finds no undefined behavior while running the tests.

You are invited to review the code - it is not large - and report any issues you encounter.

Are the run-time checks expensive?

While run-time checks performed by SendRc and SendOption are not free, they are reasonably cheap.

SendRc::deref() compares the integer id of the pinned-to thread fetched with a relaxed atomic load with the current thread, fetched from thread-local storage. It also checks that migration isn't in progress with a non-atomic integer comparison to zero. The relaxed atomic load compiles to an ordinary load on Intel, which is as cheap as it gets, and if you're worried, you can hold on to the reference to avoid repeating the checks. (The borrow checker will prevent you from sending the SendRc to another thread while there is an outstanding reference.) SendRc::clone() and SendRc::drop() do the same kind of check.

SendOption::deref() and SendOption::deref_mut() only check that the current thread is the expected one, using a similar relaxed load and compare as SendRc.

Regarding memory usage, SendRc's heap overhead is two u64s for the pinning and parking info, and a machine word for the reference count, i.e. on 64-bit architectures it's one u64 more than that of Rc. An individual SendRc is two machine words wide because it has to track the identity of each pointer. SendOption stores a u64 alongside the underlying option.

License

sendable is distributed under the terms of both the MIT license and the Apache License (Version 2.0). See LICENSE-APACHE and LICENSE-MIT for details. Contributing changes is assumed to signal agreement with these licensing terms.