[TOC] # Fearless Concurrency Handling concurrent programming safely and efficiently is another of Rust’s major goals. *Concurrent programming*, where different parts of a program execute independently, and *parallel programming*, where different parts of a program execute at the same time, are becoming increasingly important as more computers take advantage of their multiple processors. Historically, programming in these contexts has been difficult and error prone: Rust hopes to change that. Initially, the Rust team thought that ensuring memory safety and preventing concurrency problems were two separate challenges to be solved with different methods. Over time, the team discovered that the ownership and type systems are a powerful set of tools to help manage memory safety *and* concurrency problems! By leveraging ownership and type checking, many concurrency errors are compile-time errors in Rust rather than runtime errors. Therefore, rather than making you spend lots of time trying to reproduce the exact circumstances under which a runtime concurrency bug occurs, incorrect code will refuse to compile and present an error explaining the problem. As a result, you can fix your code while you’re working on it rather than potentially after it has been shipped to production. We’ve nicknamed this aspect of Rust *fearless* *concurrency*. Fearless concurrency allows you to write code that is free of subtle bugs and is easy to refactor without introducing new bugs. > Note: For simplicity’s sake, we’ll refer to many of the problems as *concurrent* rather than being more precise by saying *concurrent and/or parallel*. If this book were about concurrency and/or parallelism, we’d be more specific. For this chapter, please mentally substitute *concurrent and/or parallel* whenever we use *concurrent*. Many languages are dogmatic about the solutions they offer for handling concurrent problems. For example, Erlang has elegant functionality for message-passing concurrency but has only obscure ways to share state between threads. Supporting only a subset of possible solutions is a reasonable strategy for higher-level languages because a higher-level language promises benefits from giving up some control to gain abstractions. However, lower-level languages are expected to provide the solution with the best performance in any given situation and have fewer abstractions over the hardware. Therefore, Rust offers a variety of tools for modeling problems in whatever way is appropriate for your situation and requirements. Here are the topics we’ll cover in this chapter: * How to create threads to run multiple pieces of code at the same time * *Message-passing* concurrency, where channels send messages between threads * *Shared-state* concurrency, where multiple threads have access to some piece of data * The `Sync` and `Send` traits, which extend Rust’s concurrency guarantees to user-defined types as well as types provided by the standard library ## Using Threads to Run Code Simultaneously In most current operating systems, an executed program’s code is run in a *process*, and the operating system will manage multiple processes at once. Within a program, you can also have independent parts that run simultaneously. The features that run these independent parts are called *threads*. For example, a web server could have multiple threads so that it could respond to more than one request at the same time. Splitting the computation in your program into multiple threads to run multiple tasks at the same time can improve performance, but it also adds complexity. Because threads can run simultaneously, there’s no inherent guarantee about the order in which parts of your code on different threads will run. This can lead to problems, such as: * Race conditions, where threads are accessing data or resources in an inconsistent order * Deadlocks, where two threads are waiting for each other, preventing both threads from continuing * Bugs that happen only in certain situations and are hard to reproduce and fix reliably Rust attempts to mitigate the negative effects of using threads, but programming in a multithreaded context still takes careful thought and requires a code structure that is different from that in programs running in a single thread. Programming languages implement threads in a few different ways, and many operating systems provide an API the language can call for creating new threads. The Rust standard library uses a *1:1* model of thread implementation, whereby a program uses one operating system thread per one language thread. There are crates that implement other models of threading that make different trade-offs to the 1:1 model. ### Creating a New Thread with spawn To create a new thread, we call the `thread::spawn` function and pass it a closure (we talked about closures in Chapter 13) containing the code we want to run in the new thread. The example in Listing 16-1 prints some text from a main thread and other text from a new thread. Filename: src/main.rs ``` use std::thread; use std::time::Duration; fn main() { thread::spawn(|| { for i in 1..10 { println!("hi number {i} from the spawned thread!"); thread::sleep(Duration::from_millis(1)); } }); for i in 1..5 { println!("hi number {i} from the main thread!"); thread::sleep(Duration::from_millis(1)); } } ``` Listing 16-1: Creating a new thread to print one thing while the main thread prints something else Note that when the main thread of a Rust program completes, all spawned threads are shut down, whether or not they have finished running. The output from this program might be a little different every time, but it will look similar to the following: ``` hi number 1 from the main thread! hi number 1 from the spawned thread! hi number 2 from the main thread! hi number 2 from the spawned thread! hi number 3 from the main thread! hi number 3 from the spawned thread! hi number 4 from the main thread! hi number 4 from the spawned thread! hi number 5 from the spawned thread! ``` The calls to `thread::sleep` force a thread to stop its execution for a short duration, allowing a different thread to run. The threads will probably take turns, but that isn’t guaranteed: it depends on how your operating system schedules the threads. In this run, the main thread printed first, even though the print statement from the spawned thread appears first in the code. And even though we told the spawned thread to print until `i` is 9, it only got to 5 before the main thread shut down. If you run this code and only see output from the main thread, or don’t see any overlap, try increasing the numbers in the ranges to create more opportunities for the operating system to switch between the threads. ### Waiting for All Threads to Finish Using join Handles The code in Listing 16-1 not only stops the spawned thread prematurely most of the time due to the main thread ending, but because there is no guarantee on the order in which threads run, we also can’t guarantee that the spawned thread will get to run at all! We can fix the problem of the spawned thread not running or of it ending prematurely by saving the return value of `thread::spawn` in a variable. The return type of `thread::spawn` is `JoinHandle`. A `JoinHandle` is an owned value that, when we call the `join` method on it, will wait for its thread to finish. Listing 16-2 shows how to use the `JoinHandle` of the thread we created in Listing 16-1 and call `join` to make sure the spawned thread finishes before `main` exits. Filename: src/main.rs ``` use std::thread; use std::time::Duration; fn main() { let handle = thread::spawn(|| { for i in 1..10 { println!("hi number {i} from the spawned thread!"); thread::sleep(Duration::from_millis(1)); } }); for i in 1..5 { println!("hi number {i} from the main thread!"); thread::sleep(Duration::from_millis(1)); } handle.join().unwrap(); } ``` Listing 16-2: Saving a `JoinHandle` from `thread::spawn` to guarantee the thread is run to completion Calling `join` on the handle blocks the thread currently running until the thread represented by the handle terminates. *Blocking* a thread means that thread is prevented from performing work or exiting. Because we’ve put the call to `join` after the main thread’s `for` loop, running Listing 16-2 should produce output similar to this: ``` hi number 1 from the main thread! hi number 2 from the main thread! hi number 1 from the spawned thread! hi number 3 from the main thread! hi number 2 from the spawned thread! hi number 4 from the main thread! hi number 3 from the spawned thread! hi number 4 from the spawned thread! hi number 5 from the spawned thread! hi number 6 from the spawned thread! hi number 7 from the spawned thread! hi number 8 from the spawned thread! hi number 9 from the spawned thread! ``` The two threads continue alternating, but the main thread waits because of the call to `handle.join()` and does not end until the spawned thread is finished. But let’s see what happens when we instead move `handle.join()` before the `for` loop in `main`, like this: Filename: src/main.rs ``` use std::thread; use std::time::Duration; fn main() { let handle = thread::spawn(|| { for i in 1..10 { println!("hi number {i} from the spawned thread!"); thread::sleep(Duration::from_millis(1)); } }); handle.join().unwrap(); for i in 1..5 { println!("hi number {i} from the main thread!"); thread::sleep(Duration::from_millis(1)); } } ``` The main thread will wait for the spawned thread to finish and then run its `for` loop, so the output won’t be interleaved anymore, as shown here: ``` hi number 1 from the spawned thread! hi number 2 from the spawned thread! hi number 3 from the spawned thread! hi number 4 from the spawned thread! hi number 5 from the spawned thread! hi number 6 from the spawned thread! hi number 7 from the spawned thread! hi number 8 from the spawned thread! hi number 9 from the spawned thread! hi number 1 from the main thread! hi number 2 from the main thread! hi number 3 from the main thread! hi number 4 from the main thread! ``` Small details, such as where `join` is called, can affect whether or not your threads run at the same time. ### Using move Closures with Threads We’ll often use the `move` keyword with closures passed to `thread::spawn` because the closure will then take ownership of the values it uses from the environment, thus transferring ownership of those values from one thread to another. In “Capturing the Environment with Closures” on page XX, we discussed `move` in the context of closures. Now we’ll concentrate more on the interaction between `move` and `thread::spawn`. Notice in Listing 16-1 that the closure we pass to `thread::spawn` takes no arguments: we’re not using any data from the main thread in the spawned thread’s code. To use data from the main thread in the spawned thread, the spawned thread’s closure must capture the values it needs. Listing 16-3 shows an attempt to create a vector in the main thread and use it in the spawned thread. However, this won’t work yet, as you’ll see in a moment. Filename: src/main.rs ``` use std::thread; fn main() { let v = vec![1, 2, 3]; let handle = thread::spawn(|| { println!("Here's a vector: {:?}", v); }); handle.join().unwrap(); } ``` Listing 16-3: Attempting to use a vector created by the main thread in another thread The closure uses `v`, so it will capture `v` and make it part of the closure’s environment. Because `thread::spawn` runs this closure in a new thread, we should be able to access `v` inside that new thread. But when we compile this example, we get the following error: ``` error[E0373]: closure may outlive the current function, but it borrows `v`, which is owned by the current function --> src/main.rs:6:32 | 6 | let handle = thread::spawn(|| { | ^^ may outlive borrowed value `v` 7 | println!("Here's a vector: {:?}", v); | - `v` is borrowed here | note: function requires argument type to outlive `'static` --> src/main.rs:6:18 | 6 | let handle = thread::spawn(|| { | __________________^ 7 | | println!("Here's a vector: {:?}", v); 8 | | }); | |______^ help: to force the closure to take ownership of `v` (and any other referenced variables), use the `move` keyword | 6 | let handle = thread::spawn(move || { | ++++ ``` Rust *infers* how to capture `v`, and because `println!` only needs a reference to `v`, the closure tries to borrow `v`. However, there’s a problem: Rust can’t tell how long the spawned thread will run, so it doesn’t know whether the reference to `v` will always be valid. Listing 16-4 provides a scenario that’s more likely to have a reference to `v` that won’t be valid. Filename: src/main.rs ``` use std::thread; fn main() { let v = vec![1, 2, 3]; let handle = thread::spawn(|| { println!("Here's a vector: {:?}", v); }); drop(v); // oh no! handle.join().unwrap(); } ``` Listing 16-4: A thread with a closure that attempts to capture a reference to `v` from a main thread that drops `v` If Rust allowed us to run this code, there’s a possibility that the spawned thread would be immediately put in the background without running at all. The spawned thread has a reference to `v` inside, but the main thread immediately drops `v`, using the `drop` function we discussed in Chapter 15. Then, when the spawned thread starts to execute, `v` is no longer valid, so a reference to it is also invalid. Oh no! To fix the compiler error in Listing 16-3, we can use the error message’s advice: ``` help: to force the closure to take ownership of `v` (and any other referenced variables), use the `move` keyword | 6 | let handle = thread::spawn(move || { | ++++ ``` By adding the `move` keyword before the closure, we force the closure to take ownership of the values it’s using rather than allowing Rust to infer that it should borrow the values. The modification to Listing 16-3 shown in Listing 16-5 will compile and run as we intend. Filename: src/main.rs ``` use std::thread; fn main() { let v = vec![1, 2, 3]; let handle = thread::spawn(move || { println!("Here's a vector: {:?}", v); }); handle.join().unwrap(); } ``` Listing 16-5: Using the `move` keyword to force a closure to take ownership of the values it uses We might be tempted to try the same thing to fix the code in Listing 16-4 where the main thread called `drop` by using a `move` closure. However, this fix will not work because what Listing 16-4 is trying to do is disallowed for a different reason. If we added `move` to the closure, we would move `v` into the closure’s environment, and we could no longer call `drop` on it in the main thread. We would get this compiler error instead: ``` error[E0382]: use of moved value: `v` --> src/main.rs:10:10 | 4 | let v = vec![1, 2, 3]; | - move occurs because `v` has type `Vec`, which does not implement the `Copy` trait 5 | 6 | let handle = thread::spawn(move || { | ------- value moved into closure here 7 | println!("Here's a vector: {:?}", v); | - variable moved due to use in closure ... 10 | drop(v); // oh no! | ^ value used here after move ``` Rust’s ownership rules have saved us again! We got an error from the code in Listing 16-3 because Rust was being conservative and only borrowing `v` for the thread, which meant the main thread could theoretically invalidate the spawned thread’s reference. By telling Rust to move ownership of `v` to the spawned thread, we’re guaranteeing Rust that the main thread won’t use `v` anymore. If we change Listing 16-4 in the same way, we’re then violating the ownership rules when we try to use `v` in the main thread. The `move` keyword overrides Rust’s conservative default of borrowing; it doesn’t let us violate the ownership rules. Now that we’ve covered what threads are and the methods supplied by the thread API, let’s look at some situations in which we can use threads. ## Using Message Passing to Transfer Data Between Threads One increasingly popular approach to ensuring safe concurrency is *message passing*, where threads or actors communicate by sending each other messages containing data. Here’s the idea in a slogan from the Go language documentation at *https://golang.org/doc/effective_go.html#concurrency*: “Do not communicate by sharing memory; instead, share memory by communicating.” To accomplish message-sending concurrency, Rust’s standard library provides an implementation of *channels*. A channel is a general programming concept by which data is sent from one thread to another. You can imagine a channel in programming as being like a directional channel of water, such as a stream or a river. If you put something like a rubber duck into a river, it will travel downstream to the end of the waterway. A channel has two halves: a transmitter and a receiver. The transmitter half is the upstream location where you put the rubber duck into the river, and the receiver half is where the rubber duck ends up downstream. One part of your code calls methods on the transmitter with the data you want to send, and another part checks the receiving end for arriving messages. A channel is said to be *closed* if either the transmitter or receiver half is dropped. Here, we’ll work up to a program that has one thread to generate values and send them down a channel, and another thread that will receive the values and print them out. We’ll be sending simple values between threads using a channel to illustrate the feature. Once you’re familiar with the technique, you could use channels for any threads that need to communicate with each other, such as a chat system or a system where many threads perform parts of a calculation and send the parts to one thread that aggregates the results. First, in Listing 16-6, we’ll create a channel but not do anything with it. Note that this won’t compile yet because Rust can’t tell what type of values we want to send over the channel. Filename: src/main.rs ``` use std::sync::mpsc; fn main() { let (tx, rx) = mpsc::channel(); } ``` Listing 16-6: Creating a channel and assigning the two halves to `tx` and `rx` We create a new channel using the `mpsc::channel` function; `mpsc` stands for *multiple producer, single consumer*. In short, the way Rust’s standard library implements channels means a channel can have multiple *sending* ends that produce values but only one *receiving* end that consumes those values. Imagine multiple streams flowing together into one big river: everything sent down any of the streams will end up in one river at the end. We’ll start with a single producer for now, but we’ll add multiple producers when we get this example working. The `mpsc::channel` function returns a tuple, the first element of which is the sending end—the transmitter—and the second element of which is the receiving end—the receiver. The abbreviations `tx` and `rx` are traditionally used in many fields for *transmitter* and *receiver*, respectively, so we name our variables as such to indicate each end. We’re using a `let` statement with a pattern that destructures the tuples; we’ll discuss the use of patterns in `let` statements and destructuring in Chapter 18. For now, know that using a `let` statement in this way is a convenient approach to extract the pieces of the tuple returned by `mpsc::channel`. Let’s move the transmitting end into a spawned thread and have it send one string so the spawned thread is communicating with the main thread, as shown in Listing 16-7. This is like putting a rubber duck in the river upstream or sending a chat message from one thread to another. Filename: src/main.rs ``` use std::sync::mpsc; use std::thread; fn main() { let (tx, rx) = mpsc::channel(); thread::spawn(move || { let val = String::from("hi"); tx.send(val).unwrap(); }); } ``` Listing 16-7: Moving `tx` to a spawned thread and sending `"hi"` Again, we’re using `thread::spawn` to create a new thread and then using `move` to move `tx` into the closure so the spawned thread owns `tx`. The spawned thread needs to own the transmitter to be able to send messages through the channel. The transmitter has a `send` method that takes the value we want to send. The `send` method returns a `Result` type, so if the receiver has already been dropped and there’s nowhere to send a value, the send operation will return an error. In this example, we’re calling `unwrap` to panic in case of an error. But in a real application, we would handle it properly: return to Chapter 9 to review strategies for proper error handling. In Listing 16-8, we’ll get the value from the receiver in the main thread. This is like retrieving the rubber duck from the water at the end of the river or receiving a chat message. Filename: src/main.rs ``` use std::sync::mpsc; use std::thread; fn main() { let (tx, rx) = mpsc::channel(); thread::spawn(move || { let val = String::from("hi"); tx.send(val).unwrap(); }); let received = rx.recv().unwrap(); println!("Got: {received}"); } ``` Listing 16-8: Receiving the value `"hi"` in the main thread and printing it The receiver has two useful methods: `recv` and `try_recv`. We’re using `recv`, short for *receive*, which will block the main thread’s execution and wait until a value is sent down the channel. Once a value is sent, `recv` will return it in a `Result`. When the transmitter closes, `recv` will return an error to signal that no more values will be coming. The `try_recv` method doesn’t block, but will instead return a `Result` immediately: an `Ok` value holding a message if one is available and an `Err` value if there aren’t any messages this time. Using `try_recv` is useful if this thread has other work to do while waiting for messages: we could write a loop that calls `try_recv` every so often, handles a message if one is available, and otherwise does other work for a little while until checking again. We’ve used `recv` in this example for simplicity; we don’t have any other work for the main thread to do other than wait for messages, so blocking the main thread is appropriate. When we run the code in Listing 16-8, we’ll see the value printed from the main thread: ``` Got: hi ``` Perfect! ### Channels and Ownership Transference The ownership rules play a vital role in message sending because they help you write safe, concurrent code. Preventing errors in concurrent programming is the advantage of thinking about ownership throughout your Rust programs. Let’s do an experiment to show how channels and ownership work together to prevent problems: we’ll try to use a `val` value in the spawned thread *after* we’ve sent it down the channel. Try compiling the code in Listing 16-9 to see why this code isn’t allowed. Filename: src/main.rs ``` use std::sync::mpsc; use std::thread; fn main() { let (tx, rx) = mpsc::channel(); thread::spawn(move || { let val = String::from("hi"); tx.send(val).unwrap(); println!("val is {val}"); }); let received = rx.recv().unwrap(); println!("Got: {received}"); } ``` Listing 16-9: Attempting to use `val` after we’ve sent it down the channel Here, we try to print `val` after we’ve sent it down the channel via `tx.send`. Allowing this would be a bad idea: once the value has been sent to another thread, that thread could modify or drop it before we try to use the value again. Potentially, the other thread’s modifications could cause errors or unexpected results due to inconsistent or nonexistent data. However, Rust gives us an error if we try to compile the code in Listing 16-9: ``` error[E0382]: borrow of moved value: `val` --> src/main.rs:10:31 | 8 | let val = String::from("hi"); | --- move occurs because `val` has type `String`, which does not implement the `Copy` trait 9 | tx.send(val).unwrap(); | --- value moved here 10 | println!("val is {val}"); | ^^^ value borrowed here after move ``` Our concurrency mistake has caused a compile-time error. The `send` function takes ownership of its parameter, and when the value is moved the receiver takes ownership of it. This stops us from accidentally using the value again after sending it; the ownership system checks that everything is okay. ### Sending Multiple Values and Seeing the Receiver Waiting The code in Listing 16-8 compiled and ran, but it didn’t clearly show us that two separate threads were talking to each other over the channel. In Listing 16-10 we’ve made some modifications that will prove the code in Listing 16-8 is running concurrently: the spawned thread will now send multiple messages and pause for a second between each message. Filename: src/main.rs ``` use std::sync::mpsc; use std::thread; use std::time::Duration; fn main() { let (tx, rx) = mpsc::channel(); thread::spawn(move || { let vals = vec![ String::from("hi"), String::from("from"), String::from("the"), String::from("thread"), ]; for val in vals { tx.send(val).unwrap(); thread::sleep(Duration::from_secs(1)); } }); for received in rx { println!("Got: {received}"); } } ``` Listing 16-10: Sending multiple messages and pausing between each one This time, the spawned thread has a vector of strings that we want to send to the main thread. We iterate over them, sending each individually, and pause between each by calling the `thread::sleep` function with a `Duration` value of one second. In the main thread, we’re not calling the `recv` function explicitly anymore: instead, we’re treating `rx` as an iterator. For each value received, we’re printing it. When the channel is closed, iteration will end. When running the code in Listing 16-10, you should see the following output with a one-second pause in between each line: ``` Got: hi Got: from Got: the Got: thread ``` Because we don’t have any code that pauses or delays in the `for` loop in the main thread, we can tell that the main thread is waiting to receive values from the spawned thread. ### Creating Multiple Producers by Cloning the Transmitter Earlier we mentioned that `mpsc` was an acronym for *multiple producer, single consumer*. Let’s put `mpsc` to use and expand the code in Listing 16-10 to create multiple threads that all send values to the same receiver. We can do so by cloning the transmitter, as shown in Listing 16-11. Filename: src/main.rs ``` --snip-- let (tx, rx) = mpsc::channel(); let tx1 = tx.clone(); thread::spawn(move || { let vals = vec![ String::from("hi"), String::from("from"), String::from("the"), String::from("thread"), ]; for val in vals { tx1.send(val).unwrap(); thread::sleep(Duration::from_secs(1)); } }); thread::spawn(move || { let vals = vec![ String::from("more"), String::from("messages"), String::from("for"), String::from("you"), ]; for val in vals { tx.send(val).unwrap(); thread::sleep(Duration::from_secs(1)); } }); for received in rx { println!("Got: {received}"); } --snip-- ``` Listing 16-11: Sending multiple messages from multiple producers This time, before we create the first spawned thread, we call `clone` on the transmitter. This will give us a new transmitter we can pass to the first spawned thread. We pass the original transmitter to a second spawned thread. This gives us two threads, each sending different messages to the one receiver. When you run the code, your output should look something like this: ``` Got: hi Got: more Got: from Got: messages Got: for Got: the Got: thread Got: you ``` You might see the values in another order, depending on your system. This is what makes concurrency interesting as well as difficult. If you experiment with `thread::sleep`, giving it various values in the different threads, each run will be more nondeterministic and create different output each time. Now that we’ve looked at how channels work, let’s look at a different method of concurrency. ## Shared-State Concurrency Message passing is a fine way to handle concurrency, but it’s not the only way. Another method would be for multiple threads to access the same shared data. Consider this part of the slogan from the Go language documentation again: “Do not communicate by sharing memory.” What would communicating by sharing memory look like? In addition, why would message-passing enthusiasts caution not to use memory sharing? In a way, channels in any programming language are similar to single ownership because once you transfer a value down a channel, you should no longer use that value. Shared-memory concurrency is like multiple ownership: multiple threads can access the same memory location at the same time. As you saw in Chapter 15, where smart pointers made multiple ownership possible, multiple ownership can add complexity because these different owners need managing. Rust’s type system and ownership rules greatly assist in getting this management correct. For an example, let’s look at mutexes, one of the more common concurrency primitives for shared memory. ### Using Mutexes to Allow Access to Data from One Thread at a Time *Mutex* is an abbreviation for *mutual exclusion*, as in a mutex allows only one thread to access some data at any given time. To access the data in a mutex, a thread must first signal that it wants access by asking to acquire the mutex’s *lock*. The lock is a data structure that is part of the mutex that keeps track of who currently has exclusive access to the data. Therefore, the mutex is described as *guarding* the data it holds via the locking system. Mutexes have a reputation for being difficult to use because you have to remember two rules: 1. You must attempt to acquire the lock before using the data. 1. When you’re done with the data that the mutex guards, you must unlock the data so other threads can acquire the lock. For a real-world metaphor for a mutex, imagine a panel discussion at a conference with only one microphone. Before a panelist can speak, they have to ask or signal that they want to use the microphone. When they get the microphone, they can talk for as long as they want to and then hand the microphone to the next panelist who requests to speak. If a panelist forgets to hand the microphone off when they’re finished with it, no one else is able to speak. If management of the shared microphone goes wrong, the panel won’t work as planned! Management of mutexes can be incredibly tricky to get right, which is why so many people are enthusiastic about channels. However, thanks to Rust’s type system and ownership rules, you can’t get locking and unlocking wrong. #### The API of Mutex As an example of how to use a mutex, let’s start by using a mutex in a single-threaded context, as shown in Listing 16-12. Filename: src/main.rs ``` use std::sync::Mutex; fn main() { 1 let m = Mutex::new(5); { 2 let mut num = m.lock().unwrap(); 3 *num = 6; 4 } 5 println!("m = {:?}", m); } ``` Listing 16-12: Exploring the API of `Mutex` in a single-threaded context for simplicity As with many types, we create a `Mutex` using the associated function `new` [1]. To access the data inside the mutex, we use the `lock` method to acquire the lock [2]. This call will block the current thread so it can’t do any work until it’s our turn to have the lock. The call to `lock` would fail if another thread holding the lock panicked. In that case, no one would ever be able to get the lock, so we’ve chosen to `unwrap` and have this thread panic if we’re in that situation. After we’ve acquired the lock, we can treat the return value, named `num` in this case, as a mutable reference to the data inside. The type system ensures that we acquire a lock before using the value in `m`. The type of `m` is `Mutex`, not `i32`, so we *must* call `lock` to be able to use the `i32` value. We can’t forget; the type system won’t let us access the inner `i32` otherwise. As you might suspect, `Mutex` is a smart pointer. More accurately, the call to `lock` *returns* a smart pointer called `MutexGuard`, wrapped in a `LockResult` that we handled with the call to `unwrap`. The `MutexGuard` smart pointer implements `Deref` to point at our inner data; the smart pointer also has a `Drop` implementation that releases the lock automatically when a `MutexGuard` goes out of scope, which happens at the end of the inner scope [4]. As a result, we don’t risk forgetting to release the lock and blocking the mutex from being used by other threads because the lock release happens automatically. After dropping the lock, we can print the mutex value and see that we were able to change the inner `i32` to `6` [5]. #### Sharing a Mutex Between Multiple Threads Now let’s try to share a value between multiple threads using `Mutex`. We’ll spin up 10 threads and have them each increment a counter value by 1, so the counter goes from 0 to 10. The example in Listing 16-13 will have a compiler error, and we’ll use that error to learn more about using `Mutex` and how Rust helps us use it correctly. Filename: src/main.rs ``` use std::sync::Mutex; use std::thread; fn main() { 1 let counter = Mutex::new(0); let mut handles = vec![]; 2 for _ in 0..10 { 3 let handle = thread::spawn(move || { 4 let mut num = counter.lock().unwrap(); 5 *num += 1; }); 6 handles.push(handle); } for handle in handles { 7 handle.join().unwrap(); } 8 println!("Result: {}", *counter.lock().unwrap()); } ``` Listing 16-13: Ten threads, each incrementing a counter guarded by a `Mutex` We create a `counter` variable to hold an `i32` inside a `Mutex` [1], as we did in Listing 16-12. Next, we create 10 threads by iterating over a range of numbers [2]. We use `thread::spawn` and give all the threads the same closure: one that moves the counter into the thread [3], acquires a lock on the `Mutex` by calling the `lock` method [4], and then adds 1 to the value in the mutex [5]. When a thread finishes running its closure, `num` will go out of scope and release the lock so another thread can acquire it. In the main thread, we collect all the join handles [6]. Then, as we did in Listing 16-2, we call `join` on each handle to make sure all the threads finish [7]. At that point, the main thread will acquire the lock and print the result of this program [8]. We hinted that this example wouldn’t compile. Now let’s find out why! ``` error[E0382]: use of moved value: `counter` --> src/main.rs:9:36 | 5 | let counter = Mutex::new(0); | ------- move occurs because `counter` has type `Mutex`, which does not implement the `Copy` trait ... 9 | let handle = thread::spawn(move || { | ^^^^^^^ value moved into closure here, in previous iteration of loop 10 | let mut num = counter.lock().unwrap(); | ------- use occurs due to use in closure ``` The error message states that the `counter` value was moved in the previous iteration of the loop. Rust is telling us that we can’t move the ownership of lock `counter` into multiple threads. Let’s fix the compiler error with the multiple-ownership method we discussed in Chapter 15. #### Multiple Ownership with Multiple Threads In Chapter 15, we gave a value to multiple owners by using the smart pointer `Rc` to create a reference counted value. Let’s do the same here and see what happens. We’ll wrap the `Mutex` in `Rc` in Listing 16-14 and clone the `Rc` before moving ownership to the thread. Filename: src/main.rs ``` use std::rc::Rc; use std::sync::Mutex; use std::thread; fn main() { let counter = Rc::new(Mutex::new(0)); let mut handles = vec![]; for _ in 0..10 { let counter = Rc::clone(&counter); let handle = thread::spawn(move || { let mut num = counter.lock().unwrap(); *num += 1; }); handles.push(handle); } for handle in handles { handle.join().unwrap(); } println!("Result: {}", *counter.lock().unwrap()); } ``` Listing 16-14: Attempting to use `Rc` to allow multiple threads to own the `Mutex` Once again, we compile and get… different errors! The compiler is teaching us a lot. ``` error[E0277]: `Rc>` cannot be sent between threads safely 1 --> src/main.rs:11:22 | 11 | let handle = thread::spawn(move || { | ______________________^^^^^^^^^^^^^_- | | | | | `Rc>` cannot be sent between threads safely 12 | | let mut num = counter.lock().unwrap(); 13 | | 14 | | *num += 1; 15 | | }); | |_________- within this `[closure@src/main.rs:11:36: 15:10]` | = help: within `[closure@src/main.rs:11:36: 15:10]`, the trait `Send` is not implemented for `Rc>` 2 = note: required because it appears within the type `[closure@src/main.rs:11:36: 15:10]` note: required by a bound in `spawn` ``` Wow, that error message is very wordy! Here’s the important part to focus on: ``Rc>` cannot be sent between threads safely` [1]. The compiler is also telling us the reason why: `the trait `Send` is not implemented for `Rc>`` [2]. We’ll talk about `Send` in the next section: it’s one of the traits that ensures the types we use with threads are meant for use in concurrent situations. Unfortunately, `Rc` is not safe to share across threads. When `Rc` manages the reference count, it adds to the count for each call to `clone` and subtracts from the count when each clone is dropped. But it doesn’t use any concurrency primitives to make sure that changes to the count can’t be interrupted by another thread. This could lead to wrong counts—subtle bugs that could in turn lead to memory leaks or a value being dropped before we’re done with it. What we need is a type exactly like `Rc` but one that makes changes to the reference count in a thread-safe way. #### Atomic Reference Counting with Arc Fortunately, `Arc` *is* a type like `Rc` that is safe to use in concurrent situations. The *a* stands for *atomic*, meaning it’s an *atomically reference counted* type. Atomics are an additional kind of concurrency primitive that we won’t cover in detail here: see the standard library documentation for `std::sync::atomic` for more details. At this point, you just need to know that atomics work like primitive types but are safe to share across threads. You might then wonder why all primitive types aren’t atomic and why standard library types aren’t implemented to use `Arc` by default. The reason is that thread safety comes with a performance penalty that you only want to pay when you really need to. If you’re just performing operations on values within a single thread, your code can run faster if it doesn’t have to enforce the guarantees atomics provide. Let’s return to our example: `Arc` and `Rc` have the same API, so we fix our program by changing the `use` line, the call to `new`, and the call to `clone`. The code in Listing 16-15 will finally compile and run. Filename: src/main.rs ``` use std::sync::{Arc, Mutex}; use std::thread; fn main() { let counter = Arc::new(Mutex::new(0)); let mut handles = vec![]; for _ in 0..10 { let counter = Arc::clone(&counter); let handle = thread::spawn(move || { let mut num = counter.lock().unwrap(); *num += 1; }); handles.push(handle); } for handle in handles { handle.join().unwrap(); } println!("Result: {}", *counter.lock().unwrap()); } ``` Listing 16-15: Using an `Arc` to wrap the `Mutex` to be able to share ownership across multiple threads This code will print the following: ``` Result: 10 ``` We did it! We counted from 0 to 10, which may not seem very impressive, but it did teach us a lot about `Mutex` and thread safety. You could also use this program’s structure to do more complicated operations than just incrementing a counter. Using this strategy, you can divide a calculation into independent parts, split those parts across threads, and then use a `Mutex` to have each thread update the final result with its part. Note that if you are doing simple numerical operations, there are types simpler than `Mutex` types provided by the `std::sync::atomic` module of the standard library. These types provide safe, concurrent, atomic access to primitive types. We chose to use `Mutex` with a primitive type for this example so we could concentrate on how `Mutex` works. ### Similarities Between RefCell/Rc and Mutex/Arc You might have noticed that `counter` is immutable but we could get a mutable reference to the value inside it; this means `Mutex` provides interior mutability, as the `Cell` family does. In the same way we used `RefCell` in Chapter 15 to allow us to mutate contents inside an `Rc`, we use `Mutex` to mutate contents inside an `Arc`. Another detail to note is that Rust can’t protect you from all kinds of logic errors when you use `Mutex`. Recall in Chapter 15 that using `Rc` came with the risk of creating reference cycles, where two `Rc` values refer to each other, causing memory leaks. Similarly, `Mutex` comes with the risk of creating *deadlocks*. These occur when an operation needs to lock two resources and two threads have each acquired one of the locks, causing them to wait for each other forever. If you’re interested in deadlocks, try creating a Rust program that has a deadlock; then research deadlock mitigation strategies for mutexes in any language and have a go at implementing them in Rust. The standard library API documentation for `Mutex` and `MutexGuard` offers useful information. We’ll round out this chapter by talking about the `Send` and `Sync` traits and how we can use them with custom types. ## Extensible Concurrency with the Send and Sync Traits Interestingly, the Rust language has *very* few concurrency features. Almost every concurrency feature we’ve talked about so far in this chapter has been part of the standard library, not the language. Your options for handling concurrency are not limited to the language or the standard library; you can write your own concurrency features or use those written by others. However, two concurrency concepts are embedded in the language: the `std::marker` traits `Send` and `Sync` . ### Allowing Transference of Ownership Between Threads with Send The `Send` marker trait indicates that ownership of values of the type implementing `Send` can be transferred between threads. Almost every Rust type is `Send`, but there are some exceptions, including `Rc`: this cannot be `Send` because if you cloned an `Rc` value and tried to transfer ownership of the clone to another thread, both threads might update the reference count at the same time. For this reason, `Rc` is implemented for use in single-threaded situations where you don’t want to pay the thread-safe performance penalty. Therefore, Rust’s type system and trait bounds ensure that you can never accidentally send an `Rc` value across threads unsafely. When we tried to do this in Listing 16-14, we got the error `the trait `Send` is not implemented for `Rc>``. When we switched to `Arc`, which is `Send`, the code compiled. Any type composed entirely of `Send` types is automatically marked as `Send` as well. Almost all primitive types are `Send`, aside from raw pointers, which we’ll discuss in Chapter 19. ### Allowing Access from Multiple Threads with Sync The `Sync` marker trait indicates that it is safe for the type implementing `Sync` to be referenced from multiple threads. In other words, any type `T` is `Sync` if `&T` (an immutable reference to `T`) is `Send`, meaning the reference can be sent safely to another thread. Similar to `Send`, primitive types are `Sync`, and types composed entirely of types that are `Sync` are also `Sync`. The smart pointer `Rc` is also not `Sync` for the same reasons that it’s not `Send`. The `RefCell` type (which we talked about in Chapter 15) and the family of related `Cell` types are not `Sync`. The implementation of borrow checking that `RefCell` does at runtime is not thread-safe. The smart pointer `Mutex` is `Sync` and can be used to share access with multiple threads, as you saw in “Sharing a Mutex Between Multiple Threads” on page XX. ### Implementing Send and Sync Manually Is Unsafe Because types that are made up of `Send` and `Sync` traits are automatically also `Send` and `Sync`, we don’t have to implement those traits manually. As marker traits, they don’t even have any methods to implement. They’re just useful for enforcing invariants related to concurrency. Manually implementing these traits involves implementing unsafe Rust code. We’ll talk about using unsafe Rust code in Chapter 19; for now, the important information is that building new concurrent types not made up of `Send` and `Sync` parts requires careful thought to uphold the safety guarantees. “The Rustonomicon” at *https://doc.rust-lang.org/stable/nomicon* has more information about these guarantees and how to uphold them. ## Summary This isn’t the last you’ll see of concurrency in this book: the project in Chapter 20 will use the concepts in this chapter in a more realistic situation than the smaller examples discussed here. As mentioned earlier, because very little of how Rust handles concurrency is part of the language, many concurrency solutions are implemented as crates. These evolve more quickly than the standard library, so be sure to search online for the current, state-of-the-art crates to use in multithreaded situations. The Rust standard library provides channels for message passing and smart pointer types, such as `Mutex` and `Arc`, that are safe to use in concurrent contexts. The type system and the borrow checker ensure that the code using these solutions won’t end up with data races or invalid references. Once you get your code to compile, you can rest assured that it will happily run on multiple threads without the kinds of hard-to-track-down bugs common in other languages. Concurrent programming is no longer a concept to be afraid of: go forth and make your programs concurrent, fearlessly! Next, we’ll talk about idiomatic ways to model problems and structure solutions as your Rust programs get bigger. In addition, we’ll discuss how Rust’s idioms relate to those you might be familiar with from object-oriented programming.