Concurrency

Concurrency and parallelism are incredibly important topics in computer science, and are also a hot topic in industry today. Computers are gaining more and more cores, yet many programmers aren't prepared to fully utilize them.

Rust's memory safety features also apply to its concurrency story too. Even concurrent Rust programs must be memory safe, having no data races. Rust's type system is up to the task, and gives you powerful ways to reason about concurrent code at compile time.

Before we talk about the concurrency features that come with Rust, it's important to understand something: Rust is low-level enough that the vast majority of this is provided by the standard library, not by the language. This means that if you don't like some aspect of the way Rust handles concurrency, you can implement an alternative way of doing things. mio is a real-world example of this principle in action.

Background: Send and Sync

Concurrency is difficult to reason about. In Rust, we have a strong, static type system to help us reason about our code. As such, Rust gives us two traits to help us make sense of code that can possibly be concurrent.

Send

The first trait we're going to talk about is Send. When a type T implements Send, it indicates that something of this type is able to have ownership transferred safely between threads.

This is important to enforce certain restrictions. For example, if we have a channel connecting two threads, we would want to be able to send some data down the channel and to the other thread. Therefore, we'd ensure that Send was implemented for that type.

In the opposite way, if we were wrapping a library with FFI that isn't threadsafe, we wouldn't want to implement Send, and so the compiler will help us enforce that it can't leave the current thread.

Sync

The second of these traits is called Sync. When a type T implements Sync, it indicates that something of this type has no possibility of introducing memory unsafety when used from multiple threads concurrently through shared references. This implies that types which don't have interior mutability are inherently Sync, which includes simple primitive types (like u8) and aggregate types containing them.

For sharing references across threads, Rust provides a wrapper type called Arc<T>. Arc<T> implements Send and Sync if and only if T implements both Send and Sync. For example, an object of type Arc<RefCell<U>> cannot be transferred across threads because RefCell does not implement Sync, consequently Arc<RefCell<U>> would not implement Send.

These two traits allow you to use the type system to make strong guarantees about the properties of your code under concurrency. Before we demonstrate why, we need to learn how to create a concurrent Rust program in the first place!

Threads

Rust's standard library provides a library for threads, which allow you to run Rust code in parallel. Here's a basic example of using std::thread:

use std::thread; fn main() { thread::spawn(|| { println!("Hello from a thread!"); }); }
use std::thread;

fn main() {
    thread::spawn(|| {
        println!("Hello from a thread!");
    });
}

The thread::spawn() method accepts a closure, which is executed in a new thread. It returns a handle to the thread, that can be used to wait for the child thread to finish and extract its result:

use std::thread; fn main() { let handle = thread::spawn(|| { "Hello from a thread!" }); println!("{}", handle.join().unwrap()); }
use std::thread;

fn main() {
    let handle = thread::spawn(|| {
        "Hello from a thread!"
    });

    println!("{}", handle.join().unwrap());
}

As closures can capture variables from their environment, we can also try to bring some data into the other thread:

use std::thread; fn main() { let x = 1; thread::spawn(|| { println!("x is {}", x); }); }
use std::thread;

fn main() {
    let x = 1;
    thread::spawn(|| {
        println!("x is {}", x);
    });
}

However, this gives us an error:

5:19: 7:6 error: closure may outlive the current function, but it
                 borrows `x`, which is owned by the current function
...
5:19: 7:6 help: to force the closure to take ownership of `x` (and any other referenced variables),
          use the `move` keyword, as shown:
      thread::spawn(move || {
          println!("x is {}", x);
      });

This is because by default closures capture variables by reference, and thus the closure only captures a reference to x. This is a problem, because the thread may outlive the scope of x, leading to a dangling pointer.

To fix this, we use a move closure as mentioned in the error message. move closures are explained in depth here; basically they move variables from their environment into themselves.

use std::thread; fn main() { let x = 1; thread::spawn(move || { println!("x is {}", x); }); }
use std::thread;

fn main() {
    let x = 1;
    thread::spawn(move || {
        println!("x is {}", x);
    });
}

Many languages have the ability to execute threads, but it's wildly unsafe. There are entire books about how to prevent errors that occur from shared mutable state. Rust helps out with its type system here as well, by preventing data races at compile time. Let's talk about how you actually share things between threads.

Safe Shared Mutable State

Due to Rust's type system, we have a concept that sounds like a lie: "safe shared mutable state." Many programmers agree that shared mutable state is very, very bad.

Someone once said this:

Shared mutable state is the root of all evil. Most languages attempt to deal with this problem through the 'mutable' part, but Rust deals with it by solving the 'shared' part.

The same ownership system that helps prevent using pointers incorrectly also helps rule out data races, one of the worst kinds of concurrency bugs.

As an example, here is a Rust program that could have a data race in many languages. It will not compile:

use std::thread; use std::time::Duration; fn main() { let mut data = vec![1, 2, 3]; for i in 0..3 { thread::spawn(move || { data[i] += 1; }); } thread::sleep(Duration::from_millis(50)); }
use std::thread;
use std::time::Duration;

fn main() {
    let mut data = vec![1, 2, 3];

    for i in 0..3 {
        thread::spawn(move || {
            data[i] += 1;
        });
    }

    thread::sleep(Duration::from_millis(50));
}

This gives us an error:

8:17 error: capture of moved value: `data`
        data[i] += 1;
        ^~~~

Rust knows this wouldn't be safe! If we had a reference to data in each thread, and the thread takes ownership of the reference, we'd have three owners! data gets moved out of main in the first call to spawn(), so subsequent calls in the loop cannot use this variable.

Note that this specific example will not cause a data race since different array indices are being accessed. But this can't be determined at compile time, and in a similar situation where i is a constant or is random, you would have a data race.

So, we need some type that lets us have more than one owning reference to a value. Usually, we'd use Rc<T> for this, which is a reference counted type that provides shared ownership. It has some runtime bookkeeping that keeps track of the number of references to it, hence the "reference count" part of its name.

Calling clone() on an Rc<T> will return a new owned reference and bump the internal reference count. We create one of these for each thread:

use std::thread; use std::time::Duration; use std::rc::Rc; fn main() { let mut data = Rc::new(vec![1, 2, 3]); for i in 0..3 { // create a new owned reference let data_ref = data.clone(); // use it in a thread thread::spawn(move || { data_ref[i] += 1; }); } thread::sleep(Duration::from_millis(50)); }
use std::thread;
use std::time::Duration;
use std::rc::Rc;

fn main() {
    let mut data = Rc::new(vec![1, 2, 3]);

    for i in 0..3 {
        // create a new owned reference
        let data_ref = data.clone();

        // use it in a thread
        thread::spawn(move || {
            data_ref[i] += 1;
        });
    }

    thread::sleep(Duration::from_millis(50));
}

This won't work, however, and will give us the error:

13:9: 13:22 error: the trait bound `alloc::rc::Rc<collections::vec::Vec<i32>> : core::marker::Send`
            is not satisfied
...
13:9: 13:22 note: `alloc::rc::Rc<collections::vec::Vec<i32>>`
            cannot be sent between threads safely

As the error message mentions, Rc cannot be sent between threads safely. This is because the internal reference count is not maintained in a thread safe matter and can have a data race.

To solve this, we'll use Arc<T>, Rust's standard atomic reference count type.

The Atomic part means Arc<T> can safely be accessed from multiple threads. To do this the compiler guarantees that mutations of the internal count use indivisible operations which can't have data races.

In essence, Arc<T> is a type that lets us share ownership of data across threads.

use std::thread; use std::sync::Arc; use std::time::Duration; fn main() { let mut data = Arc::new(vec![1, 2, 3]); for i in 0..3 { let data = data.clone(); thread::spawn(move || { data[i] += 1; }); } thread::sleep(Duration::from_millis(50)); }
use std::thread;
use std::sync::Arc;
use std::time::Duration;

fn main() {
    let mut data = Arc::new(vec![1, 2, 3]);

    for i in 0..3 {
        let data = data.clone();
        thread::spawn(move || {
            data[i] += 1;
        });
    }

    thread::sleep(Duration::from_millis(50));
}

Similarly to last time, we use clone() to create a new owned handle. This handle is then moved into the new thread.

And... still gives us an error.

<anon>:11:24 error: cannot borrow immutable borrowed content as mutable
<anon>:11                    data[i] += 1;
                             ^~~~

Arc<T> by default has immutable contents. It allows the sharing of data between threads, but shared mutable data is unsafe and when threads are involved can cause data races!

Usually when we wish to make something in an immutable position mutable, we use Cell<T> or RefCell<T> which allow safe mutation via runtime checks or otherwise (see also: Choosing Your Guarantees). However, similar to Rc, these are not thread safe. If we try using these, we will get an error about these types not being Sync, and the code will fail to compile.

It looks like we need some type that allows us to safely mutate a shared value across threads, for example a type that can ensure only one thread at a time is able to mutate the value inside it at any one time.

For that, we can use the Mutex<T> type!

Here's the working version:

use std::sync::{Arc, Mutex}; use std::thread; use std::time::Duration; fn main() { let data = Arc::new(Mutex::new(vec![1, 2, 3])); for i in 0..3 { let data = data.clone(); thread::spawn(move || { let mut data = data.lock().unwrap(); data[i] += 1; }); } thread::sleep(Duration::from_millis(50)); }
use std::sync::{Arc, Mutex};
use std::thread;
use std::time::Duration;

fn main() {
    let data = Arc::new(Mutex::new(vec![1, 2, 3]));

    for i in 0..3 {
        let data = data.clone();
        thread::spawn(move || {
            let mut data = data.lock().unwrap();
            data[i] += 1;
        });
    }

    thread::sleep(Duration::from_millis(50));
}

Note that the value of i is bound (copied) to the closure and not shared among the threads.

We're "locking" the mutex here. A mutex (short for "mutual exclusion"), as mentioned, only allows one thread at a time to access a value. When we wish to access the value, we use lock() on it. This will "lock" the mutex, and no other thread will be able to lock it (and hence, do anything with the value) until we're done with it. If a thread attempts to lock a mutex which is already locked, it will wait until the other thread releases the lock.

The lock "release" here is implicit; when the result of the lock (in this case, data) goes out of scope, the lock is automatically released.

Note that lock method of Mutex has this signature:

fn main() { fn lock(&self) -> LockResult<MutexGuard<T>> }
fn lock(&self) -> LockResult<MutexGuard<T>>

and because Send is not implemented for MutexGuard<T>, the guard cannot cross thread boundaries, ensuring thread-locality of lock acquire and release.

Let's examine the body of the thread more closely:

use std::sync::{Arc, Mutex}; use std::thread; use std::time::Duration; fn main() { let data = Arc::new(Mutex::new(vec![1, 2, 3])); for i in 0..3 { let data = data.clone(); thread::spawn(move || { let mut data = data.lock().unwrap(); data[i] += 1; }); } thread::sleep(Duration::from_millis(50)); }
thread::spawn(move || {
    let mut data = data.lock().unwrap();
    data[i] += 1;
});

First, we call lock(), which acquires the mutex's lock. Because this may fail, it returns a Result<T, E>, and because this is just an example, we unwrap() it to get a reference to the data. Real code would have more robust error handling here. We're then free to mutate it, since we have the lock.

Lastly, while the threads are running, we wait on a short timer. But this is not ideal: we may have picked a reasonable amount of time to wait but it's more likely we'll either be waiting longer than necessary or not long enough, depending on just how much time the threads actually take to finish computing when the program runs.

A more precise alternative to the timer would be to use one of the mechanisms provided by the Rust standard library for synchronizing threads with each other. Let's talk about one of them: channels.

Channels

Here's a version of our code that uses channels for synchronization, rather than waiting for a specific time:

use std::sync::{Arc, Mutex}; use std::thread; use std::sync::mpsc; fn main() { let data = Arc::new(Mutex::new(0)); // `tx` is the "transmitter" or "sender" // `rx` is the "receiver" let (tx, rx) = mpsc::channel(); for _ in 0..10 { let (data, tx) = (data.clone(), tx.clone()); thread::spawn(move || { let mut data = data.lock().unwrap(); *data += 1; tx.send(()).unwrap(); }); } for _ in 0..10 { rx.recv().unwrap(); } }
use std::sync::{Arc, Mutex};
use std::thread;
use std::sync::mpsc;

fn main() {
    let data = Arc::new(Mutex::new(0));

    // `tx` is the "transmitter" or "sender"
    // `rx` is the "receiver"
    let (tx, rx) = mpsc::channel();

    for _ in 0..10 {
        let (data, tx) = (data.clone(), tx.clone());

        thread::spawn(move || {
            let mut data = data.lock().unwrap();
            *data += 1;

            tx.send(()).unwrap();
        });
    }

    for _ in 0..10 {
        rx.recv().unwrap();
    }
}

We use the mpsc::channel() method to construct a new channel. We send a simple () down the channel, and then wait for ten of them to come back.

While this channel is sending a generic signal, we can send any data that is Send over the channel!

use std::thread; use std::sync::mpsc; fn main() { let (tx, rx) = mpsc::channel(); for i in 0..10 { let tx = tx.clone(); thread::spawn(move || { let answer = i * i; tx.send(answer).unwrap(); }); } for _ in 0..10 { println!("{}", rx.recv().unwrap()); } }
use std::thread;
use std::sync::mpsc;

fn main() {
    let (tx, rx) = mpsc::channel();

    for i in 0..10 {
        let tx = tx.clone();

        thread::spawn(move || {
            let answer = i * i;

            tx.send(answer).unwrap();
        });
    }

    for _ in 0..10 {
        println!("{}", rx.recv().unwrap());
    }
}

Here we create 10 threads, asking each to calculate the square of a number (i at the time of spawn()), and then send() back the answer over the channel.

Panics

A panic! will crash the currently executing thread. You can use Rust's threads as a simple isolation mechanism:

fn main() { use std::thread; let handle = thread::spawn(move || { panic!("oops!"); }); let result = handle.join(); assert!(result.is_err()); }
use std::thread;

let handle = thread::spawn(move || {
    panic!("oops!");
});

let result = handle.join();

assert!(result.is_err());

Thread.join() gives us a Result back, which allows us to check if the thread has panicked or not.