You can move !Unpin

Although I am now mostly comfortable with Rust, some concepts still elude me. One of them is the exact meaning of Unpin. The documentation says:

The documentation of Unpin says:

Types that do not require any pinning guarantees.

Where pinning is described as:

The ability to rely on this guarantee that the value a pointer is pointing at (its pointee) will

Not be moved out of its memory location

More generally, remain valid at that same memory location

From this, you could naturally deduce that Unpin is the trait that allows Rust values to be moved.

Except it is not. You can move !Unpin values:

fn is_unpin<T: Unpin>(_: &T) {}
fn f<T>(_: T) {}

fn main() {
    let val = async { 1 }; // this is a Future, which is !Unpin
    //is_unpin(a); // error: the trait `Unpin` is not implemented
    let other_val = val; // move, ok
    f(other_val); // move, ok
}

I am definitely not the first one to get confused by Unpin. The thing is that, if you’re here, you are probably thinking of Unpin wrong. Unpin is actually a trait for very niche use cases, which you probably don’t need. And, to understand why, you first need to understand Pin and pinning.

Self-Referential Objects

Some objects contain references to part of themselves. Common examples are pointing to a slice of a buffer, or linked lists. However, in practice, the most common case in Rust is futures, from asynchronous code.

For example, consider the asynchronous function below.

async fn f() {
    let a = 3;
    let b = &a;
    tokio::time::sleep(Duration::from_secs(1)).await;
    println!("{b}");
}

When you write f(), it calls the asynchronous function. More specifically, it creates an impl Future object, its role is to store the data that needs to be kept while the function is not running. That is, across await points. The lifetime of the future will look like this:

Create the impl Future when calling f(); let’s call it future.
The user passes that future to a runtime, for instance with tokio::spawn(future).
At some point, the runtime calls future.poll().
The code from f is executed until the first await.
future stores the current state of the execution, which includes the position in the code (the await from line 4), as well as the local variables, a and b.
That first call to future.poll() returns Poll::Pending, so the runtimes knows it is not done yet.
At some point, the runtime calls future.poll() again.
The rest of the code from f is executed, resuming from the await from line 4.
Since we reached the end of the function, future.poll() returns Poll::Ready(()), where () is the value returned by the code above.
The runtime drops future.

Step 5 is where future becomes self-referential. Indeed, it must store both a and &a.

(Not) Moving

Self-referential objects must not be moved in memory. If you were to do so, it would change their address, and invalidate the references that point to them¹.

For example, that means that you cannot put such objects in a Vec. The Vec would need to move them in memory when reallocating. You would think that this means that Vec<T> requires T: Unpin. But that’s not how it works. Unpin does not mean objects can be moved.

In fact, Rust values can always be moved. As shown in the example in the introduction, you can even move objects which do not implement Unpin. You won’t be able to get Rust to complain that you cannot put some value into a vector.

Technically, safe Rust does let you build self-referential objects:

struct SelfReferential {
    self_ptr: *const Self,
}

fn main() {
    let mut s = SelfReferential {
        self_ptr: std::ptr::null(),
    };
    s.self_ptr = &s as *const SelfReferential
}

However, this only works with raw pointers, whose validity is not tracked by the compiler. If you move the object, that pointer becomes invalid, and using it would be undefined behavior (it would break the aliasing rules). This is why dereferencing pointers is unsafe: it is up to the user to make sure that the pointer remains valid.

So, safe Rust won’t let you use these self-references. And borrow rules will also prevent you from constructing such an object with references instead of raw pointers:

struct SelfReferential<'a> {
    self_ref: Option<&'a Self>,
}

fn main() {
    let mut s = SelfReferential {
        self_ref: None,
    };
    // error: cannot assign to `s.self_ref` because it is borrowed
    s.self_ref = Some(&s); 
}

In short, you can move any value in Rust, and that’s fine, because values cannot be self-referential.

Pinning

As discussed above, safe Rust won’t let you work directly with self-referential objects. So, all you can do is holding a pointer to that value.

Also, note that, to remain safe, such a pointer should not let you “take” that value to put it elsewhere (e.g. let x = *p, or std::mem::replace(&mut p, …)). That’s what Pin does. It is a safe interface for pointers to objects that should not be moved.

In other words, you can have a Pin<&T>, a Pin<&mut T>, a Pin<Rc<T>>, a Pin<Box<T>>. And these will let you interact with the T object, without ever letting you move it. The pin! macro lets you create a pinned pointer to a value on the stack², and the Box::pin method puts the value on the heap and returns a pinned pointer to it.

Note that, since Pin<…> is a pointer, it itself is totally fine to move. In fact, it is itself Unpin:

fn is_unpin<T: Unpin>(_: &T) {}

fn main() {
    let val = async { 1 };
    //is_unpin(a); // error: the trait `Unpin` is not implemented
    let ptr = std::pin::pin!(val);
    is_unpin(&ptr); // ok
    *ptr; // error: cannot move out of dereference of Pin<…>
}

Pinning Futures

As an aside, I would like to discuss something that sometimes surprised me when working with asynchronous code. Specifically, when you call an asynchronous function, you get an impl Future which is not pinned, but Future::poll takes Pin<&mut Self>.

The reason for that is actually simple: this allows the user to pin the future on the stack. If calling an asynchronous function returned a pinned future, that would require systematically using the heap, like you would do with Box::pin.

This is perfectly fine, because futures are not self-referential until you call Future::poll. When the user does, they have to do so on a pinned reference to the future, ensuring that they won’t be able to move it anymore. This is guaranteed by the safe interface of Pin.

Unpin

The title of this article is about Unpin, so I do need to talk about it.

Unpin is actually something very niche. To simplify a bit, the official documentation explains that it is automatically implemented for types that cannot be self-referential, and allows these types to be safely moved out of Pin. However, it does not explain why it would be useful. The best I could find is an article that uses it when implementing complex future wrapper; however it still requires unsafe.

Conclusion

There are three main takeaways from this article:

All Rust values can be moved
Pin is a pointer wrapper for objects that should not be moved
Unpin is a very niche trait that is only indirectly related to moving

Writing this article forced me to actually understand the relationship between Pin and Unpin, which is not what I expected. This will probably help my blazingly fast code in the future!

Technically, you could update references that point to these objects, but that involves quite a few other issues, and that’s not what Rust does. ↩︎
Strictly speaking, the value can still end-up on the heap, so the proper name is “local pinning”. ↩︎

Quentin Santos

Make it work. Make it right. Make it fast.