Arc and Mutex in Rust

When writing concurrent Rust you will encounter Arc and Mutex types sooner or later. And although Mutex might already sound familiar as it's a concept known in many languages, chances are you haven't heard about Arc before Rust. What's more, you can't fully understand these concepts without tying them to the ownership model in Rust. This post is my take on understanding Arc and Mutex in Rust.

Typically when you share data in a concurrent environment you either share memory or pass data as messages. You might often hear that passing messages (for example by using channels) is a preferred way to handle concurrency, but in Rust I don't think the safety or correctness differences are as big as in other languages due to the ownership model in Rust. Or more specifically: you can't have a data race in safe Rust. That's why when I choose between message passing or memory sharing in Rust, I do it mostly in relation to convinience, not safety.

If you choose to share data by sharing memory you will quickly encounter that you can't do much without Arc and Mutex. Arc is a smart pointer that let's you safely share a value between multiple threads. Mutex is a wrapper over another type, which allows safe mutability across threads. In order to fully understand these concepts, though, let's dive into the ownership model.

Ownership in Rust

If you tried to distill the ownership model in Rust, you would probably get the following points:

Let's see how it plays out. Given a User struct containing a String field named name we create a thread and print out a message for the user:

1use std::thread::spawn;
2
3#[derive(Debug)]
4struct User {
5 name: String
6}
7
8fn main() {
9 let user = User { name: "drogus".to_string() };
10
11 spawn(move || {
12 println!("Hello from the first thread {}", user.name);
13 }).join().unwrap();
14}

So far so good, the program compiles and prints the message. Now imagine we need to add a second thread that also has access to the user instance:

1fn main() {
2 let user = User { name: "drogus".to_string() };
3
4 let t1 = spawn(move || {
5 println!("Hello from the first thread {}", user.name);
6 });
7
8 let t2 = spawn(move || {
9 println!("Hello from the second thread {}", user.name);
10 });
11
12 t1.join().unwrap();
13 t2.join().unwrap();
14}

With this code we get the following error:

error[E0382]: use of moved value: `user.name`
  --> src/main.rs:15:20
   |
11 |     let t1 = spawn(move || {
   |                    ------- value moved into closure here
12 |         println!("Hello from the first thread {}", user.name);
   |                                                    --------- variable moved due to use in closure
...
15 |     let t2 = spawn(move || {
   |                    ^^^^^^^ value used here after move
16 |         println!("Hello from the second thread {}", user.name);
   |                                                    --------- use occurs due to use in closure
   |
   = note: move occurs because `user.name` has type `String`, which does not implement the `Copy` trait

What does the compiler want? The error reads "use of moved value user.name". The compiler is even nice enough to point us to specific places where the problem occurs. We first move the value to the first thread on line 11 and then we try to do the same thing with the second thread on line 15. If you look at the ownership rules, this shouldn't be surprising. A value can have only one owner. With the current version of the code we need to "move" the value to the first thread if we want to use it, and thus we can't move it to the other thread. It already changed ownership. But we don't mutate the data, right? Which means we can have multiple shared references. Let's try that.

1fn main() {
2 let user = User { name: "drogus".to_string() };
3
4 let t1 = spawn(|| {
5 println!("Hello from the first thread {}", &user.name);
6 });
7
8 let t2 = spawn(|| {
9 println!("Hello from the second thread {}", &user.name);
10 });
11
12 t1.join().unwrap();
13 t2.join().unwrap();
14}

I removed the move keyword in the thread closures and I made the threads borrow the user value immutably, or in other words get a shared reference, which is represented by the ampersand. With this code we get the following:

error[E0373]: closure may outlive the current function, but it borrows `user.name`, which is owned by the current function
  --> src/main.rs:15:20
   |
15 |     let t2 = spawn(|| {
   |                    ^^ may outlive borrowed value `user.name`
16 |         println!("Hello from the first thread {}", &user.name);
   |                                                     --------- `user.name` is borrowed here
   |
note: function requires argument type to outlive `'static`
  --> src/main.rs:15:14
   |
15 |       let t2 = spawn(|| {
   |  ______________^
16 | |         println!("Hello from the second thread {}", &user.name);
17 | |     });
   | |______^
help: to force the closure to take ownership of `user.name` (and any other referenced variables), use the `move` keyword
   |
15 |     let t2 = spawn(move || {
   |                    ++++

Now the error says that the closure can outlive the function. In other words the Rust compiler can't guarantee that the closure in the thread will finish before the main() function finishes. Threads are borrowing the user struct, but it's still owned by the main function. In this scenario if the main function finishes, the user struct goes out of scope and the memory is dropped. Thus if it was allowed to share the value with threads in that manner, there could be a scenario when a thread is trying to read freed memory. Which is an undefined behaviour and we certainly don't want that.

The note also says that it may help to move the variable user to the thread in order to avoid borrowing, but we're just coming from that scenario, so it's no good. Now, there are two easy solutions to fix this and one of them is to use Arc, but let's explore the other solution first: scoped threads.

Scoped threads

Scoped threads is a feature available either from an excellent crossbeam crate or as an experimental nightly feature in Rust. For the purpose of this article I will use crossbeam, but the API is very similar for both versions. After adding crossbeam = "0.8" to dependencies in Cargo.toml this code will work without problems

1use crossbeam::scope;
2
3#[derive(Debug)]
4struct User {
5 name: String,
6}
7
8fn main() {
9 let user = User {
10 name: "drogus".to_string(),
11 };
12
13 scope(|s| {
14 s.spawn(|_| {
15 println!("Hello from the first thread {}", &user.name);
16 });
17
18 s.spawn(|_| {
19 println!("Hello from the second thread {}", &user.name);
20 });
21 })
22 .unwrap();
23}

The way scoped threads work is all of the threads created in the scope are guaranteed to be finished before the scope closure finishes. Or in other words

One interesting thing to note here is that as a human reader we would have interpreted both of these programs as valid. In the version that Rust rejects we join both threads before the main() function finishes, so it would be actually safe to share the user value with the threads. This is, unfortunately, something that you may encounter when writing in Rust. Writing a compiler that would accept all of the valid programs is not possible, thus we're left with the next best thing: a compiler that will reject all invalid programs at a cost of being overly strict. Scoped threads is a future written specifically to allow us to write this code in a way that compiler can accept.

As useful as the scoped threads feature is, however, you can't always use it, for example when writing async code. Let's get to the Arc solution then.

Arc to the rescue

Arc is a smart pointer enabling sharing data between threads. Its name is a shortcut for "atomic reference counter". The way Arc works is essentially to wrap a value we're trying to share and act as a pointer to it. Arc keeps track of all of the copies of the pointer and as soon as the last pointer goes out of scope it can safely free the memory. The solution to our small problem with Arc would look something like this:

1use std::thread::spawn;
2use std::sync::Arc;
3
4#[derive(Debug)]
5struct User {
6 name: String
7}
8
9fn main() {
10 let user_original = Arc::new(User { name: "drogus".to_string() });
11
12 let user = user_original.clone();
13 let t1 = spawn(move || {
14 println!("Hello from the first thread {}", user.name);
15 });
16
17 let user = user_original.clone();
18 let t2 = spawn(move || {
19 println!("Hello from the first thread {}", user.name);
20 });
21
22 t1.join().unwrap();
23 t2.join().unwrap();
24}

Let's go through it step by step. First, on line 10, we create a user value, but we also wrap it with an Arc. Now the value is stored in memory and Arc acts only as a pointer. Whenever we clone the Arc we only clone the reference, not the user value itself. On lines 12 and 17 we clone the Arc and thus a copy of the pointer is moved to each of the threads. As you can see Arc allows us to share the data regardless of the lifetimes. In this example we will have three pointers to the user value. One created when an Arc is created, one created by cloning before starting the first thread and moved to the first thread and one created by cloning before starting the second thread and moved the first thread. As long as any of these pointers is alive, Rust will not free the memory. But when both the threads and the main function finish, all of the Arc pointers will get out of scope, dropped and as soon as the last one drops, it will also drop the user value.

Send and Sync

Let's go a bit deeper, though. If you look at the Arc documentation, you will see it implements Send and Sync traits, but only if the wrapped type also implements both Send and Sync. In order to understand what it means and why it's implemented this way let's start by defining Send and Sync.

The Rustonomicon defines Send and Sync as:

Feel free to read about these traits on Rustonomicon, but I'll also try to share my understanding here. Both Send and Sync are traits acting as markers - they don't have any implemented methods nor they require you to implement anything. What they allow is to notify the compiler about a type's ability to be shared or sent between threads. Let's start with Send, which is a bit more straightforward. What it means is that you can't send type which is !Send (read: not Send) to another thread. For example you can't send it through a channel nor can you move it to a thread. For example this code will not compile:

1#![feature(negative_impls)]
2
3#[derive(Debug)]
4struct Foo {}
5impl !Send for Foo {}
6
7fn main() {
8 let foo = Foo {};
9 spawn(move || {
10 dbg!(foo);
11 });
12}

Send and Sync are autoderived, meaning that for example if all of the attributes of a type are Send, the type will also be Send. This code uses an experimental feature called negative_impls, which lets us tell the compiler "I explicitly want to mark this type as !Send". Trying to compile this code will result in an error:

`Foo` cannot be sent between threads safely

The same would happen if you created a channel to send foo to a thread. So now what with Arc? As you might have guessed it will also not help, this will also error out in the same way (and the same would be true for a !Sync type as Arc needs both traits):

1#![feature(negative_impls)]
2
3#[derive(Debug)]
4struct Foo {}
5impl !Send for Foo {}
6
7fn main() {
8 let foo = Arc::new(Foo {});
9 spawn(move || {
10 dbg!(foo);
11 });
12}

Now, why is that the case? Isn't Arc supposed to be wrapping our type and give it more capabilities? While this is certainly true, Arc can't magically make our type threadsafe. I will give you a more in-depth example to show you why at the end of this article, but for now let's continue with learning on how to use these types.

What we learned so far is this: Arc enables us to share references to types that are Send + Sync between threads without us having to worry about lifetimes (because it's not a regular reference, but rather a smart pointer).

Modifying data with Mutex

Now let's talk about Mutex. Mutexes in many languages are treated like semaphores. You create a mutex object and you can guard a certain piece (or pieces) of the code with the mutex in a way that only one thread at a time can access the guarded place. In Rust Mutex behaves more like a wrapper. It consumes the underlying value and let's you access it only after locking the mutex. Typically Mutex is used with conjunction with Arc to make it easier to share it between threads. Let's look at the following example:

1use std::time::Duration;
2use std::{thread, thread::sleep};
3use std::sync::{Arc, Mutex};
4
5struct User {
6 name: String
7}
8
9fn main() {
10 let user_original = Arc::new(Mutex::new(User { name: String::from("drogus") }));
11
12 let user = user_original.clone();
13 let t1 = thread::spawn(move || {
14 let mut locked_user = user.lock().unwrap();
15 locked_user.name = String::from("piotr");
16 // after locked_user goes out of scope, mutex will be unlocked again,
17 // but you can also explicitly unlock it with:
18 // drop(locked_user);
19 });
20
21 let user = user_original.clone();
22 let t2 = thread::spawn(move || {
23 sleep(Duration::from_millis(10));
24
25 // it will print: Hello piotr
26 println!("Hello {}", user.lock().unwrap().name);
27 });
28
29 t1.join().unwrap();
30 t2.join().unwrap();
31}

Let's go over it. in the first line of the main() function we create an instance of the User struct and we wrap it with a Mutex and an Arc. With an Arc we can easily clone the pointer and thus share the mutex between threads. In the 13th line you can see the mutex is locked and since that moment the underlying value can be used exclusively by this thread. Then we modify the value in the next line. The mutex is unlocked once the locked guard goes out of scope or we manually drop it with drop(locked_user).

In the second thread we wait 10ms and print the name, which should be the name updated in the first thread. This time locking is done in one line, so the mutex will be dropped in the same statement.

One more thing that is worth mentioning is the unwrap() method we call after lock(). Mutex from the standard library has a notion of being poisoned. If a thread panics while the mutex is locked we can't be certain if the value inside Mutex is still valid and thus the default behaviour is to return an error instead of a guard. So Mutex can either return an Ok() variant with the wrapped value as an argument or an error. You can read more about it in the docs. In general leaving unrwap() methods in the production code is not recommended, but in case of Mutex it might be a valid strategy - if a mutex has been poisoned we might decide that the application state is invalid and crash the application.

Another interesting thing about Mutex is that as long as a type inside the Mutex is Send, Mutex will also be Sync. This is because Mutex ensures that only one thread can get access to the underlying value and thus it's safe to share Mutex between threads.

Mutex: add Sync to a Send type

As you may remember from the beginning of the article, Arc needs an underlying type to be Send + Sync in order for Arc to be Send + Sync too. Mutex only requires and underlying type to be Send in order for Mutex to be `Send

Mutex without Arc?

An interesting question that you may ask is if Mutex can be used without Arc. I encourage you to think about it a little before reading further: what does it mean that Mutex is Send + Sync for types that are Send?

If you get back to the first part of this post you can see what it means for the Arc type and in case of Mutex it means a very similar thing. If we can use something like scope threads it's entirely possible to use Mutex without Arc:

1use crossbeam::scope;
2use std::{sync::Mutex, thread::sleep, time::Duration};
3
4#[derive(Debug)]
5struct User {
6 name: String,
7}
8
9fn main() {
10 let user = Mutex::new(User {
11 name: "drogus".to_string(),
12 });
13
14 scope(|s| {
15 s.spawn(|_| {
16 user.lock().unwrap().name = String::from("piotr");
17 });
18
19 s.spawn(|_| {
20 sleep(Duration::from_millis(10));
21
22 // should print: Hello piotr
23 println!("Hello {}", user.lock().unwrap().name);
24 });
25 })
26 .unwrap();
27}

In this program we achieve the same goal. We are accessing a value behind a mutex in two separate threads, but we share mutexes by reference and not by using an Arc. But again, this is not always possible, for example in async code, so Mutex is very often used along with Arc.

Summary

I'm hoping that in this article I helped you understand what are Arc and Mutex types in Rust and how to use them. To sum it up I would say you would typically use Arc whenever you want to share data between threads and you can't do so using regular references. You would also use Mutex if you need to modify data you share between threads. And then you would use Arc<Mutex<...>> whenever you want to modify data you share between threads and you can't share a mutex using references.

Bonus: Why Arc needs type to be Sync

Now let's get back to the question of "why Arc needs the underlying type to be both Send and Sync to mark it as Send and Sync). Feel free to ignore this last section, though, it's not really needed for you to use Arc and Mutex in your code. It might help you understand markter traits a bit better.

Lets take Cell as an example. Cell wraps another type and enables "interior mutability" or in other words it allows us to modify a value inside an immutable struct. Cell is Send, but it's !Sync.

An example of using Cell would be:

1use std::cell::Cell;
2
3struct User {
4 age: Cell<usize>
5}
6
7fn main() {
8 let user = User { age: Cell::new(30) };
9
10 user.age.set(36);
11
12 // will print: Age: 36
13 println!("Age: {}", user.age.get());
14}

Cell is useful in some situations, but it isn't thread safe or in other words it's !Sync. If you somehow shared a value wrapped in a cell between multiple threads you could modify the same place in memory from two threads, for example:

1// this example will not compile, `Cell` is `!Sync` and thus
2// `Arc` will be `!Sync` and `!Send`
3use std::cell::Cell;
4
5struct User {
6 age: Cell<usize>
7}
8
9fn main() {
10 let user_original = Arc::new(User { age: Cell::new(30) });
11
12 let user = user_original.clone();
13 std::thread::spawn(move || {
14 user.age.set(2);
15 });
16
17 let user = user_original.clone();
18 std::thread::spawn(move || {
19 user.age.set(3);
20 });
21}

If that worked, it could result in an undefined behaviour. That's why Arc will not work with any type that is not Send nor Sync. At the same time Cell is Send, meaning that you can send it between threads. Why is that? Sending, or in other words moving, will not make a value accessible from more than one thread, it will have to always be only one thread. Once you move it to another, the previous thread doesn't own the value anymore. With that in mind, we can always mutate a Cell locally.

Bonus: why Arc needs type

At this point you might also wonder why Arc will not provide the Send trait for a !Send type, either. One of the types in Rust which is !Send is Rc. Rc is a cousin of Arc, but it's not "atomic", Rc expands to just "reference counter". Its role is pretty much the same as Arc, but it can only be used in a single thread. Not only it can't be shared between threads, but also it can't be moved between threads. Let's see why.

1// this code won't compile, Rc is !Send and !Sync
2use std::rc::Rc;
3
4fn main() {
5 let foo = Rc::new(1);
6
7 let foo_clone = foo.clone();
8 std::thread::spawn(move || {
9 dbg!(foo_clone);
10 });
11
12 let foo_clone = foo.clone();
13 std::thread::spawn(move || {
14 dbg!(foo_clone);
15 });
16}

This example won't compile, because Rc is !Sync + !Send. Its internal counter is not atomic and thus sharing it between threads could result in an inaccurate count of references. Now imagine that Arc would make !Send types Send:

1use std::rc::Rc;
2use std::sync::Arc;
3
4#[derive(Debug)]
5struct User {
6 name: Rc<String>,
7}
8unsafe impl Send for User {}
9unsafe impl Sync for User {}
10
11fn main() {
12 let foo = Arc::new(User {
13 name: Rc::new(String::from("drogus")),
14 });
15
16 let foo_clone = foo.clone();
17 std::thread::spawn(move || {
18 let name = foo_clone.name.clone();
19 });
20
21 let foo_clone = foo.clone();
22 std::thread::spawn(move || {
23 let name = foo_clone.name.clone();
24 });
25}

This example will compile, but it's wrong, please don't do it in your actual code! In here I define a User struct, which holds an Rc inside. Because Send and Sync are autoderived and Rc is !Send + !Sync, the User struct should also be !Send + !Sync, but we can explicitly tell the compiler to mark it differently, in this case Send + Sync, using unsafe impl syntax.

Now you can clearly see what would go wrong if Arc allowed !Send types to be moved between threads. In the example Arc clones are moved into separate threads and then nothing is stopping us from cloning the Rc type. And because Rc type is not thread safe, it could result in an inacurate count of references and thus could either free memory too soon or it could not free it at all even though it should.

I know that this article is a long one, so kudos too all of you that got it all the way here, thanks!