The Rust Programming Language

Turning Our Single-Threaded Server into a Multithreaded Server

Right now, the server will process each request in turn, meaning it won’t process a second connection until the first is finished processing. If the server received more and more requests, this serial execution would be less and less optimal. If the server receives a request that takes a long time to process, subsequent requests will have to wait until the long request is finished, even if the new requests can be processed quickly. We’ll need to fix this, but first, we’ll look at the problem in action.

Simulating a Slow Request in the Current Server Implementation

We’ll look at how a slow-processing request can affect other requests made to our current server implementation. Listing 21-10 implements handling a request to /sleep with a simulated slow response that will cause the server to sleep for 5 seconds before responding.

use std::{
    fs,
    io::{prelude::*, BufReader},
    net::{TcpListener, TcpStream},
    thread,
    time::Duration,
};
// --snip--

fn main() {
    let listener = TcpListener::bind("127.0.0.1:7878").unwrap();

    for stream in listener.incoming() {
        let stream = stream.unwrap();

        handle_connection(stream);
    }
}

fn handle_connection(mut stream: TcpStream) {
    // --snip--

    let buf_reader = BufReader::new(&stream);
    let request_line = buf_reader.lines().next().unwrap().unwrap();

    let (status_line, filename) = match &request_line[..] {
        "GET / HTTP/1.1" => ("HTTP/1.1 200 OK", "hello.html"),
        "GET /sleep HTTP/1.1" => {
            thread::sleep(Duration::from_secs(5));
            ("HTTP/1.1 200 OK", "hello.html")
        }
        _ => ("HTTP/1.1 404 NOT FOUND", "404.html"),
    };

    // --snip--

    let contents = fs::read_to_string(filename).unwrap();
    let length = contents.len();

    let response =
        format!("{status_line}\r\nContent-Length: {length}\r\n\r\n{contents}");

    stream.write_all(response.as_bytes()).unwrap();
}

We switched from if to match now that we have three cases. We need to explicitly match on a slice of request_line to pattern match against the string literal values; match doesn’t do automatic referencing and dereferencing like the equality method does.

The first arm is the same as the if block from Listing 21-9. The second arm matches a request to /sleep. When that request is received, the server will sleep for 5 seconds before rendering the successful HTML page. The third arm is the same as the else block from Listing 21-9.

You can see how primitive our server is: real libraries would handle the recognition of multiple requests in a much less verbose way!

Start the server using cargo run. Then open two browser windows: one for http://127.0.0.1:7878/ and the other for http://127.0.0.1:7878/sleep. If you enter the / URI a few times, as before, you’ll see it respond quickly. But if you enter /sleep and then load /, you’ll see that / waits until sleep has slept for its full 5 seconds before loading.

There are multiple techniques we could use to avoid requests backing up behind a slow request, including using async as we did Chapter 17; the one we’ll implement is a thread pool.

Improving Throughput with a Thread Pool

A thread pool is a group of spawned threads that are waiting and ready to handle a task. When the program receives a new task, it assigns one of the threads in the pool to the task, and that thread will process the task. The remaining threads in the pool are available to handle any other tasks that come in while the first thread is processing. When the first thread is done processing its task, it’s returned to the pool of idle threads, ready to handle a new task. A thread pool allows you to process connections concurrently, increasing the throughput of your server.

We’ll limit the number of threads in the pool to a small number to protect us from Denial of Service (DoS) attacks; if we had our program create a new thread for each request as it came in, someone making 10 million requests to our server could create havoc by using up all our server’s resources and grinding the processing of requests to a halt.

Rather than spawning unlimited threads, then, we’ll have a fixed number of threads waiting in the pool. Requests that come in are sent to the pool for processing. The pool will maintain a queue of incoming requests. Each of the threads in the pool will pop off a request from this queue, handle the request, and then ask the queue for another request. With this design, we can process up to N requests concurrently, where N is the number of threads. If each thread is responding to a long-running request, subsequent requests can still back up in the queue, but we’ve increased the number of long-running requests we can handle before reaching that point.

This technique is just one of many ways to improve the throughput of a web server. Other options you might explore are the fork/join model, the single-threaded async I/O model, or the multi-threaded async I/O model. If you’re interested in this topic, you can read more about other solutions and try to implement them; with a low-level language like Rust, all of these options are possible.

Before we begin implementing a thread pool, let’s talk about what using the pool should look like. When you’re trying to design code, writing the client interface first can help guide your design. Write the API of the code so it’s structured in the way you want to call it; then implement the functionality within that structure rather than implementing the functionality and then designing the public API.

Similar to how we used test-driven development in the project in Chapter 12, we’ll use compiler-driven development here. We’ll write the code that calls the functions we want, and then we’ll look at errors from the compiler to determine what we should change next to get the code to work. Before we do that, however, we’ll explore the technique we’re not going to use as a starting point.

Spawning a Thread for Each Request

First, let’s explore how our code might look if it did create a new thread for every connection. As mentioned earlier, this isn’t our final plan due to the problems with potentially spawning an unlimited number of threads, but it is a starting point to get a working multithreaded server first. Then we’ll add the thread pool as an improvement, and contrasting the two solutions will be easier. Listing 21-11 shows the changes to make to main to spawn a new thread to handle each stream within the for loop.

use std::{
    fs,
    io::{prelude::*, BufReader},
    net::{TcpListener, TcpStream},
    thread,
    time::Duration,
};

fn main() {
    let listener = TcpListener::bind("127.0.0.1:7878").unwrap();

    for stream in listener.incoming() {
        let stream = stream.unwrap();

        thread::spawn(|| {
            handle_connection(stream);
        });
    }
}

fn handle_connection(mut stream: TcpStream) {
    let buf_reader = BufReader::new(&stream);
    let request_line = buf_reader.lines().next().unwrap().unwrap();

    let (status_line, filename) = match &request_line[..] {
        "GET / HTTP/1.1" => ("HTTP/1.1 200 OK", "hello.html"),
        "GET /sleep HTTP/1.1" => {
            thread::sleep(Duration::from_secs(5));
            ("HTTP/1.1 200 OK", "hello.html")
        }
        _ => ("HTTP/1.1 404 NOT FOUND", "404.html"),
    };

    let contents = fs::read_to_string(filename).unwrap();
    let length = contents.len();

    let response =
        format!("{status_line}\r\nContent-Length: {length}\r\n\r\n{contents}");

    stream.write_all(response.as_bytes()).unwrap();
}

As you learned in Chapter 16, thread::spawn will create a new thread and then run the code in the closure in the new thread. If you run this code and load /sleep in your browser, then / in two more browser tabs, you’ll indeed see that the requests to / don’t have to wait for /sleep to finish. However, as we mentioned, this will eventually overwhelm the system because you’d be making new threads without any limit.

You may also recall from Chapter 17 that this is exactly the kind of situation where async and await really shine! Keep that in mind as we build the thread pool and think about how things would look different or the same with async.

Creating a Finite Number of Threads

We want our thread pool to work in a similar, familiar way so switching from threads to a thread pool doesn’t require large changes to the code that uses our API. Listing 21-12 shows the hypothetical interface for a ThreadPool struct we want to use instead of thread::spawn.

use std::{
    fs,
    io::{prelude::*, BufReader},
    net::{TcpListener, TcpStream},
    thread,
    time::Duration,
};

fn main() {
    let listener = TcpListener::bind("127.0.0.1:7878").unwrap();
    let pool = ThreadPool::new(4);

    for stream in listener.incoming() {
        let stream = stream.unwrap();

        pool.execute(|| {
            handle_connection(stream);
        });
    }
}

fn handle_connection(mut stream: TcpStream) {
    let buf_reader = BufReader::new(&stream);
    let request_line = buf_reader.lines().next().unwrap().unwrap();

    let (status_line, filename) = match &request_line[..] {
        "GET / HTTP/1.1" => ("HTTP/1.1 200 OK", "hello.html"),
        "GET /sleep HTTP/1.1" => {
            thread::sleep(Duration::from_secs(5));
            ("HTTP/1.1 200 OK", "hello.html")
        }
        _ => ("HTTP/1.1 404 NOT FOUND", "404.html"),
    };

    let contents = fs::read_to_string(filename).unwrap();
    let length = contents.len();

    let response =
        format!("{status_line}\r\nContent-Length: {length}\r\n\r\n{contents}");

    stream.write_all(response.as_bytes()).unwrap();
}

We use ThreadPool::new to create a new thread pool with a configurable number of threads, in this case four. Then, in the for loop, pool.execute has a similar interface as thread::spawn in that it takes a closure the pool should run for each stream. We need to implement pool.execute so it takes the closure and gives it to a thread in the pool to run. This code won’t yet compile, but we’ll try so the compiler can guide us in how to fix it.

Building ThreadPool Using Compiler Driven Development

Make the changes in Listing 21-12 to src/main.rs, and then let’s use the compiler errors from cargo check to drive our development. Here is the first error we get:

$ cargo check
    Checking hello v0.1.0 (file:///projects/hello)
error[E0433]: failed to resolve: use of undeclared type `ThreadPool`
  --> src/main.rs:11:16
   |
11 |     let pool = ThreadPool::new(4);
   |                ^^^^^^^^^^ use of undeclared type `ThreadPool`

For more information about this error, try `rustc --explain E0433`.
error: could not compile `hello` (bin "hello") due to 1 previous error

Great! This error tells us we need a ThreadPool type or module, so we’ll build one now. Our ThreadPool implementation will be independent of the kind of work our web server is doing. So, let’s switch the hello crate from a binary crate to a library crate to hold our ThreadPool implementation. After we change to a library crate, we could also use the separate thread pool library for any work we want to do using a thread pool, not just for serving web requests.

Create a src/lib.rs that contains the following, which is the simplest definition of a ThreadPool struct that we can have for now:

pub struct ThreadPool;

Then edit main.rs file to bring ThreadPool into scope from the library crate by adding the following code to the top of src/main.rs:

use hello::ThreadPool;
use std::{
    fs,
    io::{prelude::*, BufReader},
    net::{TcpListener, TcpStream},
    thread,
    time::Duration,
};

fn main() {
    let listener = TcpListener::bind("127.0.0.1:7878").unwrap();
    let pool = ThreadPool::new(4);

    for stream in listener.incoming() {
        let stream = stream.unwrap();

        pool.execute(|| {
            handle_connection(stream);
        });
    }
}

fn handle_connection(mut stream: TcpStream) {
    let buf_reader = BufReader::new(&stream);
    let request_line = buf_reader.lines().next().unwrap().unwrap();

    let (status_line, filename) = match &request_line[..] {
        "GET / HTTP/1.1" => ("HTTP/1.1 200 OK", "hello.html"),
        "GET /sleep HTTP/1.1" => {
            thread::sleep(Duration::from_secs(5));
            ("HTTP/1.1 200 OK", "hello.html")
        }
        _ => ("HTTP/1.1 404 NOT FOUND", "404.html"),
    };

    let contents = fs::read_to_string(filename).unwrap();
    let length = contents.len();

    let response =
        format!("{status_line}\r\nContent-Length: {length}\r\n\r\n{contents}");

    stream.write_all(response.as_bytes()).unwrap();
}

This code still won’t work, but let’s check it again to get the next error that we need to address:

$ cargo check
    Checking hello v0.1.0 (file:///projects/hello)
error[E0599]: no function or associated item named `new` found for struct `ThreadPool` in the current scope
  --> src/main.rs:12:28
   |
12 |     let pool = ThreadPool::new(4);
   |                            ^^^ function or associated item not found in `ThreadPool`

For more information about this error, try `rustc --explain E0599`.
error: could not compile `hello` (bin "hello") due to 1 previous error

This error indicates that next we need to create an associated function named new for ThreadPool. We also know that new needs to have one parameter that can accept 4 as an argument and should return a ThreadPool instance. Let’s implement the simplest new function that will have those characteristics:

pub struct ThreadPool;

impl ThreadPool {
    pub fn new(size: usize) -> ThreadPool {
        ThreadPool
    }
}

We chose usize as the type of the size parameter, because we know that a negative number of threads doesn’t make any sense. We also know we’ll use this 4 as the number of elements in a collection of threads, which is what the usize type is for, as discussed in the “Integer Types” section of Chapter 3.

Let’s check the code again:

$ cargo check
    Checking hello v0.1.0 (file:///projects/hello)
error[E0599]: no method named `execute` found for struct `ThreadPool` in the current scope
  --> src/main.rs:17:14
   |
17 |         pool.execute(|| {
   |         -----^^^^^^^ method not found in `ThreadPool`

For more information about this error, try `rustc --explain E0599`.
error: could not compile `hello` (bin "hello") due to 1 previous error

Now the error occurs because we don’t have an execute method on ThreadPool. Recall from the “Creating a Finite Number of Threads” section that we decided our thread pool should have an interface similar to thread::spawn. In addition, we’ll implement the execute function so it takes the closure it’s given and gives it to an idle thread in the pool to run.

We’ll define the execute method on ThreadPool to take a closure as a parameter. Recall from the “Moving Captured Values Out of the Closure and the Fn Traits” section in Chapter 13 that we can take closures as parameters with three different traits: Fn, FnMut, and FnOnce. We need to decide which kind of closure to use here. We know we’ll end up doing something similar to the standard library thread::spawn implementation, so we can look at what bounds the signature of thread::spawn has on its parameter. The documentation shows us the following:

pub fn spawn<F, T>(f: F) -> JoinHandle<T>
    where
        F: FnOnce() -> T,
        F: Send + 'static,
        T: Send + 'static,

The F type parameter is the one we’re concerned with here; the T type parameter is related to the return value, and we’re not concerned with that. We can see that spawn uses FnOnce as the trait bound on F. This is probably what we want as well, because we’ll eventually pass the argument we get in execute to spawn. We can be further confident that FnOnce is the trait we want to use because the thread for running a request will only execute that request’s closure one time, which matches the Once in FnOnce.

The F type parameter also has the trait bound Send and the lifetime bound 'static, which are useful in our situation: we need Send to transfer the closure from one thread to another and 'static because we don’t know how long the thread will take to execute. Let’s create an execute method on ThreadPool that will take a generic parameter of type F with these bounds:

pub struct ThreadPool;

impl ThreadPool {
    // --snip--
    pub fn new(size: usize) -> ThreadPool {
        ThreadPool
    }

    pub fn execute<F>(&self, f: F)
    where
        F: FnOnce() + Send + 'static,
    {
    }
}

We still use the () after FnOnce because this FnOnce represents a closure that takes no parameters and returns the unit type (). Just like function definitions, the return type can be omitted from the signature, but even if we have no parameters, we still need the parentheses.

Again, this is the simplest implementation of the execute method: it does nothing, but we’re trying only to make our code compile. Let’s check it again:

$ cargo check
    Checking hello v0.1.0 (file:///projects/hello)
    Finished `dev` profile [unoptimized + debuginfo] target(s) in 0.24s

It compiles! But note that if you try cargo run and make a request in the browser, you’ll see the errors in the browser that we saw at the beginning of the chapter. Our library isn’t actually calling the closure passed to execute yet!

Consider: what would be different here if we were going to execute a future instead of a closure?

Validating the Number of Threads in new

We aren’t doing anything with the parameters to new and execute. Let’s implement the bodies of these functions with the behavior we want. To start, let’s think about new. Earlier we chose an unsigned type for the size parameter, because a pool with a negative number of threads makes no sense. However, a pool with zero threads also makes no sense, yet zero is a perfectly valid usize. We’ll add code to check that size is greater than zero before we return a ThreadPool instance and have the program panic if it receives a zero by using the assert! macro, as shown in Listing 21-13.

pub struct ThreadPool;

impl ThreadPool {
    /// Create a new ThreadPool.
    ///
    /// The size is the number of threads in the pool.
    ///
    /// # Panics
    ///
    /// The `new` function will panic if the size is zero.
    pub fn new(size: usize) -> ThreadPool {
        assert!(size > 0);

        ThreadPool
    }

    // --snip--
    pub fn execute<F>(&self, f: F)
    where
        F: FnOnce() + Send + 'static,
    {
    }
}

We’ve also added some documentation for our ThreadPool with doc comments. Note that we followed good documentation practices by adding a section that calls out the situations in which our function can panic, as discussed in Chapter 14. Try running cargo doc --open and clicking the ThreadPool struct to see what the generated docs for new look like!

Instead of adding the assert! macro as we’ve done here, we could change new into build and return a Result like we did with Config::build in the I/O project in Listing 12-9. But we’ve decided in this case that trying to create a thread pool without any threads should be an unrecoverable error. If you’re feeling ambitious, try to write a function named build with the following signature to compare with the new function:

pub fn build(size: usize) -> Result<ThreadPool, PoolCreationError> {

Creating Space to Store the Threads

Now that we have a way to know we have a valid number of threads to store in the pool, we can create those threads and store them in the ThreadPool struct before returning the struct. But how do we “store” a thread? Let’s take another look at the thread::spawn signature:

pub fn spawn<F, T>(f: F) -> JoinHandle<T>
    where
        F: FnOnce() -> T,
        F: Send + 'static,
        T: Send + 'static,

The spawn function returns a JoinHandle<T>, where T is the type that the closure returns. Let’s try using JoinHandle too and see what happens. In our case, the closures we’re passing to the thread pool will handle the connection and not return anything, so T will be the unit type ().

The code in Listing 21-14 will compile but doesn’t create any threads yet. We’ve changed the definition of ThreadPool to hold a vector of thread::JoinHandle<()> instances, initialized the vector with a capacity of size, set up a for loop that will run some code to create the threads, and returned a ThreadPool instance containing them.

use std::thread;

pub struct ThreadPool {
    threads: Vec<thread::JoinHandle<()>>,
}

impl ThreadPool {
    // --snip--
    /// Create a new ThreadPool.
    ///
    /// The size is the number of threads in the pool.
    ///
    /// # Panics
    ///
    /// The `new` function will panic if the size is zero.
    pub fn new(size: usize) -> ThreadPool {
        assert!(size > 0);

        let mut threads = Vec::with_capacity(size);

        for _ in 0..size {
            // create some threads and store them in the vector
        }

        ThreadPool { threads }
    }
    // --snip--

    pub fn execute<F>(&self, f: F)
    where
        F: FnOnce() + Send + 'static,
    {
    }
}

We’ve brought std::thread into scope in the library crate, because we’re using thread::JoinHandle as the type of the items in the vector in ThreadPool.

Once a valid size is received, our ThreadPool creates a new vector that can hold size items. The with_capacity function performs the same task as Vec::new but with an important difference: it preallocates space in the vector. Because we know we need to store size elements in the vector, doing this allocation up front is slightly more efficient than using Vec::new, which resizes itself as elements are inserted.

When you run cargo check again, it should succeed.

A Worker Struct Responsible for Sending Code from the ThreadPool to a Thread

We left a comment in the for loop in Listing 21-14 regarding the creation of threads. Here, we’ll look at how we actually create threads. The standard library provides thread::spawn as a way to create threads, and thread::spawn expects to get some code the thread should run as soon as the thread is created. However, in our case, we want to create the threads and have them wait for code that we’ll send later. The standard library’s implementation of threads doesn’t include any way to do that; we have to implement it manually.

We’ll implement this behavior by introducing a new data structure between the ThreadPool and the threads that will manage this new behavior. We’ll call this data structure Worker, which is a common term in pooling implementations. The Worker picks up code that needs to be run and runs the code in the Worker’s thread. Think of people working in the kitchen at a restaurant: the workers wait until orders come in from customers, and then they’re responsible for taking those orders and fulfilling them.

Instead of storing a vector of JoinHandle<()> instances in the thread pool, we’ll store instances of the Worker struct. Each Worker will store a single JoinHandle<()> instance. Then we’ll implement a method on Worker that will take a closure of code to run and send it to the already running thread for execution. We’ll also give each worker an id so we can distinguish between the different workers in the pool when logging or debugging.

Here is the new process that will happen when we create a ThreadPool. We’ll implement the code that sends the closure to the thread after we have Worker set up in this way:

1. Define a Worker struct that holds an id and a JoinHandle<()>.
2. Change ThreadPool to hold a vector of Worker instances.
3. Define a Worker::new function that takes an id number and returns a Worker instance that holds the id and a thread spawned with an empty closure.
4. In ThreadPool::new, use the for loop counter to generate an id, create a new Worker with that id, and store the worker in the vector.

If you’re up for a challenge, try implementing these changes on your own before looking at the code in Listing 21-15.

Ready? Here is Listing 21-15 with one way to make the preceding modifications.

use std::thread;

pub struct ThreadPool {
    workers: Vec<Worker>,
}

impl ThreadPool {
    // --snip--
    /// Create a new ThreadPool.
    ///
    /// The size is the number of threads in the pool.
    ///
    /// # Panics
    ///
    /// The `new` function will panic if the size is zero.
    pub fn new(size: usize) -> ThreadPool {
        assert!(size > 0);

        let mut workers = Vec::with_capacity(size);

        for id in 0..size {
            workers.push(Worker::new(id));
        }

        ThreadPool { workers }
    }
    // --snip--

    pub fn execute<F>(&self, f: F)
    where
        F: FnOnce() + Send + 'static,
    {
    }
}

struct Worker {
    id: usize,
    thread: thread::JoinHandle<()>,
}

impl Worker {
    fn new(id: usize) -> Worker {
        let thread = thread::spawn(|| {});

        Worker { id, thread }
    }
}

We’ve changed the name of the field on ThreadPool from threads to workers because it’s now holding Worker instances instead of JoinHandle<()> instances. We use the counter in the for loop as an argument to Worker::new, and we store each new Worker in the vector named workers.

External code (like our server in src/main.rs) doesn’t need to know the implementation details regarding using a Worker struct within ThreadPool, so we make the Worker struct and its new function private. The Worker::new function uses the id we give it and stores a JoinHandle<()> instance that is created by spawning a new thread using an empty closure.

This code will compile and will store the number of Worker instances we specified as an argument to ThreadPool::new. But we’re still not processing the closure that we get in execute. Let’s look at how to do that next.

Sending Requests to Threads via Channels

The next problem we’ll tackle is that the closures given to thread::spawn do absolutely nothing. Currently, we get the closure we want to execute in the execute method. But we need to give thread::spawn a closure to run when we create each Worker during the creation of the ThreadPool.

We want the Worker structs that we just created to fetch the code to run from a queue held in the ThreadPool and send that code to its thread to run.

The channels we learned about in Chapter 16—a simple way to communicate between two threads—would be perfect for this use case. We’ll use a channel to function as the queue of jobs, and execute will send a job from the ThreadPool to the Worker instances, which will send the job to its thread. Here is the plan:

1. The ThreadPool will create a channel and hold on to the sender.
2. Each Worker will hold on to the receiver.
3. We’ll create a new Job struct that will hold the closures we want to send down the channel.
4. The execute method will send the job it wants to execute through the sender.
5. In its thread, the Worker will loop over its receiver and execute the closures of any jobs it receives.

Let’s start by creating a channel in ThreadPool::new and holding the sender in the ThreadPool instance, as shown in Listing 21-16. The Job struct doesn’t hold anything for now but will be the type of item we’re sending down the channel.

use std::{sync::mpsc, thread};

pub struct ThreadPool {
    workers: Vec<Worker>,
    sender: mpsc::Sender<Job>,
}

struct Job;

impl ThreadPool {
    // --snip--
    /// Create a new ThreadPool.
    ///
    /// The size is the number of threads in the pool.
    ///
    /// # Panics
    ///
    /// The `new` function will panic if the size is zero.
    pub fn new(size: usize) -> ThreadPool {
        assert!(size > 0);

        let (sender, receiver) = mpsc::channel();

        let mut workers = Vec::with_capacity(size);

        for id in 0..size {
            workers.push(Worker::new(id));
        }

        ThreadPool { workers, sender }
    }
    // --snip--

    pub fn execute<F>(&self, f: F)
    where
        F: FnOnce() + Send + 'static,
    {
    }
}

struct Worker {
    id: usize,
    thread: thread::JoinHandle<()>,
}

impl Worker {
    fn new(id: usize) -> Worker {
        let thread = thread::spawn(|| {});

        Worker { id, thread }
    }
}

In ThreadPool::new, we create our new channel and have the pool hold the sender. This will successfully compile.

Let’s try passing a receiver of the channel into each worker as the thread pool creates the channel. We know we want to use the receiver in the thread that the workers spawn, so we’ll reference the receiver parameter in the closure. The code in Listing 21-17 won’t quite compile yet.

use std::{sync::mpsc, thread};

pub struct ThreadPool {
    workers: Vec<Worker>,
    sender: mpsc::Sender<Job>,
}

struct Job;

impl ThreadPool {
    // --snip--
    /// Create a new ThreadPool.
    ///
    /// The size is the number of threads in the pool.
    ///
    /// # Panics
    ///
    /// The `new` function will panic if the size is zero.
    pub fn new(size: usize) -> ThreadPool {
        assert!(size > 0);

        let (sender, receiver) = mpsc::channel();

        let mut workers = Vec::with_capacity(size);

        for id in 0..size {
            workers.push(Worker::new(id, receiver));
        }

        ThreadPool { workers, sender }
    }
    // --snip--

    pub fn execute<F>(&self, f: F)
    where
        F: FnOnce() + Send + 'static,
    {
    }
}

// --snip--


struct Worker {
    id: usize,
    thread: thread::JoinHandle<()>,
}

impl Worker {
    fn new(id: usize, receiver: mpsc::Receiver<Job>) -> Worker {
        let thread = thread::spawn(|| {
            receiver;
        });

        Worker { id, thread }
    }
}

We’ve made some small and straightforward changes: we pass the receiver into Worker::new, and then we use it inside the closure.

When we try to check this code, we get this error:

$ cargo check
    Checking hello v0.1.0 (file:///projects/hello)
error[E0382]: use of moved value: `receiver`
  --> src/lib.rs:26:42
   |
21 |         let (sender, receiver) = mpsc::channel();
   |                      -------- move occurs because `receiver` has type `std::sync::mpsc::Receiver<Job>`, which does not implement the `Copy` trait
...
25 |         for id in 0..size {
   |         ----------------- inside of this loop
26 |             workers.push(Worker::new(id, receiver));
   |                                          ^^^^^^^^ value moved here, in previous iteration of loop
   |
note: consider changing this parameter type in method `new` to borrow instead if owning the value isn't necessary
  --> src/lib.rs:47:33
   |
47 |     fn new(id: usize, receiver: mpsc::Receiver<Job>) -> Worker {
   |        --- in this method       ^^^^^^^^^^^^^^^^^^^ this parameter takes ownership of the value
help: consider moving the expression out of the loop so it is only moved once
   |
25 ~         let mut value = Worker::new(id, receiver);
26 ~         for id in 0..size {
27 ~             workers.push(value);
   |

For more information about this error, try `rustc --explain E0382`.
error: could not compile `hello` (lib) due to 1 previous error

The code is trying to pass receiver to multiple Worker instances. This won’t work, as you’ll recall from Chapter 16: the channel implementation that Rust provides is multiple producer, single consumer. This means we can’t just clone the consuming end of the channel to fix this code. We also don’t want to send a message multiple times to multiple consumers; we want one list of messages with multiple workers such that each message gets processed once.

Additionally, taking a job off the channel queue involves mutating the receiver, so the threads need a safe way to share and modify receiver; otherwise, we might get race conditions (as covered in Chapter 16).

Recall the thread-safe smart pointers discussed in Chapter 16: to share ownership across multiple threads and allow the threads to mutate the value, we need to use Arc<Mutex<T>>. The Arc type will let multiple workers own the receiver, and Mutex will ensure that only one worker gets a job from the receiver at a time. Listing 21-18 shows the changes we need to make.

use std::{
    sync::{mpsc, Arc, Mutex},
    thread,
};
// --snip--

pub struct ThreadPool {
    workers: Vec<Worker>,
    sender: mpsc::Sender<Job>,
}

struct Job;

impl ThreadPool {
    // --snip--
    /// Create a new ThreadPool.
    ///
    /// The size is the number of threads in the pool.
    ///
    /// # Panics
    ///
    /// The `new` function will panic if the size is zero.
    pub fn new(size: usize) -> ThreadPool {
        assert!(size > 0);

        let (sender, receiver) = mpsc::channel();

        let receiver = Arc::new(Mutex::new(receiver));

        let mut workers = Vec::with_capacity(size);

        for id in 0..size {
            workers.push(Worker::new(id, Arc::clone(&receiver)));
        }

        ThreadPool { workers, sender }
    }

    // --snip--

    pub fn execute<F>(&self, f: F)
    where
        F: FnOnce() + Send + 'static,
    {
    }
}

// --snip--

struct Worker {
    id: usize,
    thread: thread::JoinHandle<()>,
}

impl Worker {
    fn new(id: usize, receiver: Arc<Mutex<mpsc::Receiver<Job>>>) -> Worker {
        // --snip--
        let thread = thread::spawn(|| {
            receiver;
        });

        Worker { id, thread }
    }
}

In ThreadPool::new, we put the receiver in an Arc and a Mutex. For each new worker, we clone the Arc to bump the reference count so the workers can share ownership of the receiver.

With these changes, the code compiles! We’re getting there!

Implementing the execute Method

Let’s finally implement the execute method on ThreadPool. We’ll also change Job from a struct to a type alias for a trait object that holds the type of closure that execute receives. As discussed in the “Creating Type Synonyms with Type Aliases” section of Chapter 20, type aliases allow us to make long types shorter for ease of use. Look at Listing 21-19.

use std::{
    sync::{mpsc, Arc, Mutex},
    thread,
};

pub struct ThreadPool {
    workers: Vec<Worker>,
    sender: mpsc::Sender<Job>,
}

// --snip--

type Job = Box<dyn FnOnce() + Send + 'static>;

impl ThreadPool {
    // --snip--
    /// Create a new ThreadPool.
    ///
    /// The size is the number of threads in the pool.
    ///
    /// # Panics
    ///
    /// The `new` function will panic if the size is zero.
    pub fn new(size: usize) -> ThreadPool {
        assert!(size > 0);

        let (sender, receiver) = mpsc::channel();

        let receiver = Arc::new(Mutex::new(receiver));

        let mut workers = Vec::with_capacity(size);

        for id in 0..size {
            workers.push(Worker::new(id, Arc::clone(&receiver)));
        }

        ThreadPool { workers, sender }
    }

    pub fn execute<F>(&self, f: F)
    where
        F: FnOnce() + Send + 'static,
    {
        let job = Box::new(f);

        self.sender.send(job).unwrap();
    }
}

// --snip--

struct Worker {
    id: usize,
    thread: thread::JoinHandle<()>,
}

impl Worker {
    fn new(id: usize, receiver: Arc<Mutex<mpsc::Receiver<Job>>>) -> Worker {
        let thread = thread::spawn(|| {
            receiver;
        });

        Worker { id, thread }
    }
}

After creating a new Job instance using the closure we get in execute, we send that job down the sending end of the channel. We’re calling unwrap on send for the case that sending fails. This might happen if, for example, we stop all our threads from executing, meaning the receiving end has stopped receiving new messages. At the moment, we can’t stop our threads from executing: our threads continue executing as long as the pool exists. The reason we use unwrap is that we know the failure case won’t happen, but the compiler doesn’t know that.

But we’re not quite done yet! In the worker, our closure being passed to thread::spawn still only references the receiving end of the channel. Instead, we need the closure to loop forever, asking the receiving end of the channel for a job and running the job when it gets one. Let’s make the change shown in Listing 21-20 to Worker::new.

use std::{
    sync::{mpsc, Arc, Mutex},
    thread,
};

pub struct ThreadPool {
    workers: Vec<Worker>,
    sender: mpsc::Sender<Job>,
}

type Job = Box<dyn FnOnce() + Send + 'static>;

impl ThreadPool {
    /// Create a new ThreadPool.
    ///
    /// The size is the number of threads in the pool.
    ///
    /// # Panics
    ///
    /// The `new` function will panic if the size is zero.
    pub fn new(size: usize) -> ThreadPool {
        assert!(size > 0);

        let (sender, receiver) = mpsc::channel();

        let receiver = Arc::new(Mutex::new(receiver));

        let mut workers = Vec::with_capacity(size);

        for id in 0..size {
            workers.push(Worker::new(id, Arc::clone(&receiver)));
        }

        ThreadPool { workers, sender }
    }

    pub fn execute<F>(&self, f: F)
    where
        F: FnOnce() + Send + 'static,
    {
        let job = Box::new(f);

        self.sender.send(job).unwrap();
    }
}

struct Worker {
    id: usize,
    thread: thread::JoinHandle<()>,
}

// --snip--

impl Worker {
    fn new(id: usize, receiver: Arc<Mutex<mpsc::Receiver<Job>>>) -> Worker {
        let thread = thread::spawn(move || loop {
            let job = receiver.lock().unwrap().recv().unwrap();

            println!("Worker {id} got a job; executing.");

            job();
        });

        Worker { id, thread }
    }
}

Here, we first call lock on the receiver to acquire the mutex, and then we call unwrap to panic on any errors. Acquiring a lock might fail if the mutex is in a poisoned state, which can happen if some other thread panicked while holding the lock rather than releasing the lock. In this situation, calling unwrap to have this thread panic is the correct action to take. Feel free to change this unwrap to an expect with an error message that is meaningful to you.

If we get the lock on the mutex, we call recv to receive a Job from the channel. A final unwrap moves past any errors here as well, which might occur if the thread holding the sender has shut down, similar to how the send method returns Err if the receiver shuts down.

The call to recv blocks, so if there is no job yet, the current thread will wait until a job becomes available. The Mutex<T> ensures that only one Worker thread at a time is trying to request a job.

Our thread pool is now in a working state! Give it a cargo run and make some requests:

$ cargo run
   Compiling hello v0.1.0 (file:///projects/hello)
warning: field is never read: `workers`
 --> src/lib.rs:7:5
  |
7 |     workers: Vec<Worker>,
  |     ^^^^^^^^^^^^^^^^^^^^
  |
  = note: `#[warn(dead_code)]` on by default

warning: field is never read: `id`
  --> src/lib.rs:48:5
   |
48 |     id: usize,
   |     ^^^^^^^^^

warning: field is never read: `thread`
  --> src/lib.rs:49:5
   |
49 |     thread: thread::JoinHandle<()>,
   |     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

warning: `hello` (lib) generated 3 warnings
    Finished dev [unoptimized + debuginfo] target(s) in 1.40s
     Running `target/debug/hello`
Worker 0 got a job; executing.
Worker 2 got a job; executing.
Worker 1 got a job; executing.
Worker 3 got a job; executing.
Worker 0 got a job; executing.
Worker 2 got a job; executing.
Worker 1 got a job; executing.
Worker 3 got a job; executing.
Worker 0 got a job; executing.
Worker 2 got a job; executing.

Success! We now have a thread pool that executes connections asynchronously. There are never more than four threads created, so our system won’t get overloaded if the server receives a lot of requests. If we make a request to /sleep, the server will be able to serve other requests by having another thread run them.

This is a good time to pause and consider how the code in Listings 21-18, 21-19, and 21-20 would be different if we were using futures instead of a closure for the work to be done. What types would change? How would the method signatures be different, if at all? What parts of the code would stay the same?

After learning about the while let loop in Chapters 17 and 18, you might be wondering why we didn’t write the worker thread code as shown in Listing 21-21.

use std::{
    sync::{mpsc, Arc, Mutex},
    thread,
};

pub struct ThreadPool {
    workers: Vec<Worker>,
    sender: mpsc::Sender<Job>,
}

type Job = Box<dyn FnOnce() + Send + 'static>;

impl ThreadPool {
    /// Create a new ThreadPool.
    ///
    /// The size is the number of threads in the pool.
    ///
    /// # Panics
    ///
    /// The `new` function will panic if the size is zero.
    pub fn new(size: usize) -> ThreadPool {
        assert!(size > 0);

        let (sender, receiver) = mpsc::channel();

        let receiver = Arc::new(Mutex::new(receiver));

        let mut workers = Vec::with_capacity(size);

        for id in 0..size {
            workers.push(Worker::new(id, Arc::clone(&receiver)));
        }

        ThreadPool { workers, sender }
    }

    pub fn execute<F>(&self, f: F)
    where
        F: FnOnce() + Send + 'static,
    {
        let job = Box::new(f);

        self.sender.send(job).unwrap();
    }
}

struct Worker {
    id: usize,
    thread: thread::JoinHandle<()>,
}
// --snip--

impl Worker {
    fn new(id: usize, receiver: Arc<Mutex<mpsc::Receiver<Job>>>) -> Worker {
        let thread = thread::spawn(move || {
            while let Ok(job) = receiver.lock().unwrap().recv() {
                println!("Worker {id} got a job; executing.");

                job();
            }
        });

        Worker { id, thread }
    }
}

This code compiles and runs but doesn’t result in the desired threading behavior: a slow request will still cause other requests to wait to be processed. The reason is somewhat subtle: the Mutex struct has no public unlock method because the ownership of the lock is based on the lifetime of the MutexGuard<T> within the LockResult<MutexGuard<T>> that the lock method returns. At compile time, the borrow checker can then enforce the rule that a resource guarded by a Mutex cannot be accessed unless we hold the lock. However, this implementation can also result in the lock being held longer than intended if we aren’t mindful of the lifetime of the MutexGuard<T>.

The code in Listing 21-20 that uses let job = receiver.lock().unwrap().recv().unwrap(); works because with let, any temporary values used in the expression on the right hand side of the equals sign are immediately dropped when the let statement ends. However, while let (and if let and match) does not drop temporary values until the end of the associated block. In Listing 21-21, the lock remains held for the duration of the call to job(), meaning other workers cannot receive jobs.