Reading and writing, part 5: concurrency and language design

This article is part of a five-part series on race conditions. The complete series is:

Over the course of this series we’ve explored the idea of a Location type, which characterises a resource that allows arbitrary reads and writes with no co-ordination or control mechanism:

Location<T> = {
    read() -> T;
    write(T);
}

We’ve seen many instances of this type, from a simple in-memory counter variable with asynchronous or multi-threaded writers in part 1, through a similar structure stored in a database in part 2, to complex objects being accessed by multiple users in a web application in part 3 and part 4. In this final instalment of the series, we’re going to loop back to looking at in-memory data in programs, and how some languages have rethought concurrent programming to make these access patterns either safe or non-existent.

In part 1 we looked at how object-oriented programming and mutexes are both ways to control access to mutable data, letting us limit the ways that external callers can change the object’s state and thus ensure it remains consistent. A closely related concept is that of the actor model, a framework for concurrent programming. An actor is a process, which has some local state, and can communicate with other actors exclusively by sending them messages. This resembles the OOP idea of sending an object commands by invoking its methods, rather than extracting state from inside it and manipulating that state directly.

We can build a toy version of the actor model in Ruby as follows. The Actor class is constructed with some other object that it retains a reference to. It has a mailbox, modelled as a Queue, a thread-safe structure that supports push and shift operations to add and remove values from a first-in first-out queue. The actor starts a thread, which continuously loops over the messages in the queue, sending each one to the underlying object.

class Actor
  def initialize(object)
    @object  = object
    @mailbox = Queue.new

    @thread = Thread.new do
      loop do
        message = @mailbox.shift
        @object.__send__(*message)
      end
    end
  end

  # ...

Queue#shift blocks if the queue is empty, returning the next pushed message as soon as one exists. __send__ is Ruby’s way of dynamically invoking a method by name; message here should be an array containing the name of the method to invoke, and any arguments to call it with.

Actor supports two further methods: send lets an external caller push messages into the mailbox – the only way actors are allowed to communicate. And, we’ll add a wait method which joins the actor’s thread, so we can make our program’s main thread block on one of the actors we create.

  # ...

  def send(*message)
    @mailbox.push(message)
  end

  def wait
    @thread.join
  end
end

In effect this Actor class places a mutex around an entire object. Its send method lets other actors send it commands, just like invoking the methods on a normal object, but rather than executing the command immediately in the caller’s thread, the command goes in a queue which is executed sequentially in a thread dedicated to the actor. It means all the messages sent to an actor are processed one after another, rather than overlapping in time as usually happens in multi-threaded programs.

To work with this class, we’ll take our original Counter class and remove the attr_reader :value declaration that let us read the @value field out of the object. Instead, we’ll add a get method, whose sender argument is a reference to an Actor, to which we send the message [:puts, @value] to make it print the value. In the actor model, we can’t just extract state from objects, we have to send the actor a message and have it send us one back with the information we want. The message :puts will invoke the sender’s puts method, which prints whatever it receives.

class Counter
  def initialize
    @value = 0
  end

  def inc
    @value += 1
  end

  def get(sender)
    sender.send(:puts, @value)
  end
end

Let’s try out this implementation with a short example program. We’ll create two actors, one wrapping a Counter instance and one wrapping the main object – the top-level environment in Ruby’s runtime. We’ll send the message :inc to the counter a thousand times, and then send it the message [:get, main] to make it send its final value to the main object which will print it.

counter = Actor.new(Counter.new)
main    = Actor.new(self)

1000.times { counter.send(:inc) }

counter.send(:get, main)
main.wait

This program prints 1000, as we’d expect as all our send(:inc) calls are sequential. Let’s now try the same thing using a set of threads:

counter = Actor.new(Counter.new)
main    = Actor.new(self)

threads = (1..4).map do
  Thread.new do
    1000.times { counter.send(:inc) }
  end
end

threads.each(&:join)
counter.send(:get, main)
main.wait

This works correctly, printing 4000. We didn’t have to use a mutex to make sure this concurrent access to the counter object is valid. The Actor class ensures the counter does not lose updates by enqueueing all the messages it receives and handing them off to the underlying Counter sequentially.

The actor model both extends and restricts object-oriented programming, making us think of objects as processes. An object encapsulates some state that must be kept consistent, and updates it according to a sequence of messages it receives. Effectively, we can only write setter methods that change an object’s state, rather than getter methods that extract that state. If we want to get a value out of an actor, we can ask it to send that value to us, but this model instead encourages flowing updates in one direction through a system by propagating events through a network of actors.

If there is a cycle in that network, there’s a clear thread of causality and messages to the same object cannot be interleaved in the way normal OOP allows. Consider this JavaScript program, where a call to Joe.hello() causes an invocation of Mike.hello(), which itself calls Joe.interrupt():

let Joe = {
  hello() {
    console.log('starting Joe.hello');
    Mike.hello();
    console.log('finishing Joe.hello');
  },

  interrupt() {
    console.log('INTERRUPTING JOE');
  }
};

let Mike = {
  hello() {
    Joe.interrupt();
  }
};

By calling Mike.hello(), Joe might indirectly be invoking its own methods, so it needs to make sure that whatever changes Joe.hello() makes are safe if Joe.interrupt() happens in their midst.

$ node joe_mike.js

starting Joe.hello
INTERRUPTING JOE
finishing Joe.hello

In the actor model, an object’s own methods can only be invoked immediately by that object. If it sends a message to another object that results in another message looping back to it, that new message will go in the mailbox and won’t be processed until the current message is done processing. Let’s write the above using the Actor class we developed above:

class Joe
  def hello(joe, mike)
    puts "starting Joe.hello"
    mike.send(:hello, joe)
    puts "finishing Joe.hello"
  end

  def interrupt
    puts "INTERRUPTING JOE"
  end
end

class Mike
  def hello(joe)
    joe.send(:interrupt)
  end
end

joe  = Actor.new(Joe.new)
mike = Actor.new(Mike.new)

joe.send(:hello, joe, mike)

joe.wait

Each object’s methods need to take arguments that let them refer to the actors in the system, otherwise we need to use global variables. When we run this program, the call mike.send(:hello, joe) causes an invocation of joe.send(:interrupt), but that message will be processed after Joe#hello has finished executing – the Actor wrapper makes sure of this.

$ ruby joe_mike.rb

starting Joe.hello
finishing Joe.hello
INTERRUPTING JOE

This is useful because it limits the extent to which an object’s own methods can be interleaved, reducing the number of possible code paths that can change the object’s state and making the system easier to reason about.

One language that puts the actor model at the heart of its mechanics is Erlang. In Erlang, actors are lightweight concurrent processes and a single runtime can span multiple machines, with actors communicating over the network. Rather than using objects with mutable state, Erlang processes wrap pure functions. State is handled by passing the starting value into a function, and then calling the function with a new value in response to a message the process receives.

For example, below is an example of how we’d write a counter in Erlang. It responds to two messages:

  • { inc }, which is a tuple containing the single word inc; this causes counter to call itself again with a new input: Value + 1. In Erlang, variables begin with uppercase letters.
  • { get, From }, a tuple of two words, the second of which is a variable identifying the process that sent the message; this sends Value to the process identified by From and then calls counter with the current Value.
counter(Value) ->
    receive
        { inc } -> counter(Value + 1);
        { get, From } -> From ! Value, counter(Value)
    end.

The recursive calls to counter serve to establish an infinite loop in which the process runs continuously, waiting for new messages and updating its state – the Value variable here. We create a running process from this function using the spawn function in the Erlang shell:

> Ctr = spawn(fun() -> counter(0) end).

We can then send it messages using the ! operator; here we send { inc } and then { get, self() }, where self() returns the current actor’s process ID or PID. This causes counter to send the current Value to the shell’s main process. Calling flush() makes the shell print any messages it’s been sent.

> Ctr ! { inc }.
> Ctr ! { get, self() }.
> flush().
Shell got 1
ok

We can send more of these messages and see that the counter updates as expected.

> Ctr ! { inc }.
> Ctr ! { inc }.
> Ctr ! { get, self() }.
> flush().
Shell got 3
ok

In the following snippet, we start a new counter process, and then spawn a thousand other processes that each increments the counter. The syntax is probably unfamiliar; this essentially tells Erlang to spawn a process that runs Ctr ! { inc } for every value from 1 to 1000.

> Ctr = spawn(fun() -> counter(0) end).
> [spawn(fun() -> Ctr ! { inc } end) || _ <- lists:seq(1, 1000)]

These processes all run concurrently and independently; they do not co-ordinate with each other over sending messages to the Ctr process. However, when all of them are complete we see that no updates have been lost and the counter has the correct value.

> Ctr ! { get, self() }.
> flush().
Shell got 1000
ok

Erlang solves the concurrency problem by removing mutation entirely. We do not update process state by reading copies of values, modifying our copy and then writing it back. Instead, processes listen for messages and restart themselves with new state in response to what they receive. A process’s state is a pure function of its inputs and the sequence of messages it’s been sent, and we cannot access process state from outside as is normal in procedural languages.

We also have the property that messages are not interleaved, so there is no ambiguity in what the result of some sequence of messages is. Let’s recreate our conversation between Joe and Mike in Erlang: joe() and mike() are two functions that loop infinitely by handling a message and then calling themselves.

joe() ->
    receive
        { hello, Mike } ->
            io:format("starting Joe.hello~n"),
            Mike ! { hello, self() },
            io:format("finishing Joe.hello~n");
        { interrupt } ->
            io:format("INTERRUPTING JOE~n")
    end,
    joe().

mike() ->
    receive
        { hello, Joe } ->
            Joe ! { interrupt }
    end,
    mike().

If we spawn two processes wrapping these functions, and send Joe ! { hello, Mike }, then we see the complete output from the { hello, Mike } message before we see the output from the { interrupt } one.

> Joe = spawn(fun() -> joe() end).
> Mike = spawn(fun() -> mike() end).

> Joe ! { hello, Mike }.
starting Joe.hello
finishing Joe.hello
INTERRUPTING JOE

This pattern of using pure functions that process a message stream to restart themselves with new values is how Erlang models servers, parsers and all sorts of other state machines. It’s also how the Haskell-inspired Elm user interface environment works: the UI is a pure function of the model state, user interaction with that UI generates messages, those messages are combined with the current model to generate a new one, and the loop continues.

This is also how people typically program with React, following the Flux architecture that emphasizes state flowing in a single direction through your code, rather than the two-way data binding that other frameworks put between the model and the UI. That cyclic triggering of events leads to the interleaving problem we saw above, and reproduces many of the bugs we’ve explored in this series. Programming with bare events is equivalent to performing uncontrolled reads and writes of all the components and model state in a UI, and is incredibly hard to do correctly without causing race conditions. I like how Manish Goregaokar puts it:

Aliasing with mutability in a sufficiently complex, single-threaded program is effectively the same thing as accessing data shared across multiple threads without a lock

The Problem With Single-threaded Shared Mutability

Aliasing here refers to having many variables or data structures holding references to the same value, and all of those references being allowed to perform mutations to the referenced object. Which brings us nicely to the other major example I want to look at in relation to language design: Rust’s ownership system.

In Rust, every value has a single owner: a variable, function parameter, location in a data structure, or other binding that contains the value. If we assign the value to another binding, the value is moved out of its previous owner and the old binding becomes unusable. For example, this program creates a Vec (short for vector, a dynamically-sized array) and binds it to the variable list – the list variable now owns the Vec value. We then perform the assignment let other = list, which transfers ownership of the Vec from list to other, making list unusable. The use of list in the final println! statement is thus illegal.

fn main() {
    let list = vec![1, 2, 3];
    let other = list;
    println!("list: {:?}", list);
}

This program will not compile; Rust spots the use of the invalidated binding list and gives us an error:

error[E0382]: borrow of moved value: `list`
 --> ownership.rs:6:28
  |
4 |     let list = vec![1, 2, 3];
  |         ---- move occurs because `list` has type `std::vec::Vec<i32>`,
  |              which does not implement the `Copy` trait
5 |     let other = list;
  |                 ---- value moved here
6 |     println!("list: {:?}", list);
  |                            ^^^^ value borrowed here after move

You don’t need to know what borrow means just yet. What this error means is that the binding list is used after the Vec value has been moved out of it, into other. The error shows us where the move happens, where list is used after its value is moved, and also why it’s moved: the type Vec<i32> (a vector of 32-bit integers) does not implement Copy, a trait that causes values to be copied in situations where the default behaviour is to move them. Only simple values like numbers and booleans are Copy, most complex structures like Vec are not.

Values are also moved when they’re passed to other functions. Here we pass list to the function do_nothing, which takes Vec<i32> is input. do_nothing does not use its input in any way – just passing list to it is sufficient to move the Vec from main into do_nothing, so it can no longer be used in main.

fn main() {
    let list = vec![1, 2, 3];
    do_nothing(list);
    println!("list: {:?}", list);
}

fn do_nothing(list: Vec<i32>) {}

We get a similar error from the compiler for this program, showing that the move happens where list is passed as an argument to do_nothing.

error[E0382]: borrow of moved value: `list`
 --> ownership.rs:6:28
  |
4 |     let list = vec![1, 2, 3];
  |         ---- move occurs because `list` has type `std::vec::Vec<i32>`,
  |              which does not implement the `Copy` trait
5 |     do_nothing(list);
  |                ---- value moved here
6 |     println!("list: {:?}", list);
  |                            ^^^^ value borrowed here after move

This suggests that all values are single-use: once you pass them to a function in order to do anything with them, they become unusable. This is true when we pass values to functions, but not when we pass references. A function taking Vec<i32> as input means it takes that type by value: the Vec is moved into the called function. By changing its input type to &Vec<i32>, that means the function takes a reference (&) to a Vec<i32>, which it can use to read the referenced value without taking ownership of it. For example, to print a value we only need to be able to read from it:

fn show(list: &Vec<i32>) {
    println!("list: {:?}", list);
}

A & reference is an immutable reference; it only allows the holder to read the referenced value, not to change it. A &mut is a mutable reference, and taking this kind of reference allows a function to mutate the referenced value, again without taking ownership of it. For example, this generic function takes &mut Vec<T>, a mutable reference to a Vec of any type T, and an item of type T, and adds the item to the list:

fn add<T>(list: &mut Vec<T>, item: T) {
    list.push(item);
}

The method call list.push(item) is syntactic sugar for Vec::push(list, item), and Vec<T>::push is just another function that’s defined to take a &mut Vec<T>. If it took a read-only reference &Vec<T>, the compiler would not let it mutate the Vec struct on which it operates – to change a structure you must have a mutable reference to it, or you must own it, a read-only reference is not enough. The show() function above would not be allowed to call list.push(), because it takes a read-only &Vec<i32>.

Use of & and &mut references is known as borrowing, creating a pointer to a value instead of passing a copy of it to the called function. We saw above the compiler telling us about a “borrow” of the list binding when we used println!, and that’s because println! is a macro for printing the given values, and it generates code that only needs to read the arguments, not change or take ownership of them.

Borrowing lets us use values without destroying them by moving them, but it comes with an important restriction. As well as having a single owner, a value may have either a single &mut reference or any number of & references to it in existence at any point in a program. That is, & and &mut references to the same value cannot coexist, and we cannot have two &mut references to the same value at the same time. These restrictions limit how values can be changed, and are designed to prevent data races of the kinds we’ve examined in this series.

For example, the following function is legal; it creates a & reference to list in order to call show(), and when that function call ends the & reference is immediately dropped. Therefore we can then create a &mut reference to call add(), which again the &mut reference is dropped immediately after this call, allowing us to call show() again.

fn main() {
    let mut list = vec![1, 2, 3];
    show(&list);
    add(&mut list, 4);
    show(&list);
}

This program compiles and prints the state of the Vec before and after adding the new element:

list: [1, 2, 3]
list: [1, 2, 3, 4]

What we’re not allowed to do is create a & reference, and then create a &mut to the same value while the & reference still exists. This program stores a &Vec<i32> reference to list in the variable list_ref, which is later used by being passed to show(). Between creating list_ref and using it, we try to call add(), which requires a &mut to the Vec.

fn main() {
    let mut list = vec![1, 2, 3];
    let list_ref = &list;
    add(&mut list, 4);
    show(list_ref);
}

This will not compile: the expression &mut list cannot exist inside a region of the program where &list exists. The compiler tells us as much by showing us where the conflicting references are created and used:

error[E0502]: cannot borrow `list` as mutable because it is also borrowed as
immutable
 --> ownership.rs:8:9
  |
6 |     let list_ref = &list;
  |                    ----- immutable borrow occurs here
7 |     add(&mut list, 4);
  |         ^^^^^^^^^ mutable borrow occurs here
8 |     show(list_ref);
  |          -------- immutable borrow later used here

The same thing happens if we try to crate a & reference at a point where a conflicting &mut reference exists:

fn main() {
    let mut list = vec![1, 2, 3];
    let list_mut = &mut list;
    show(&list);
    add(list_mut, 4);
}

This program fails to compile, with an error similar to the one above:

error[E0502]: cannot borrow `list` as immutable because it is also borrowed as
mutable
 --> ownership.rs:8:10
  |
6 |     let list_mut = &mut list;
  |                    --------- mutable borrow occurs here
7 |     show(&list);
  |          ^^^^^ immutable borrow occurs here
8 |     add(list_mut, 4);
  |         -------- mutable borrow later used here

The conflicting references do not need to point to exactly the same location; a reference covers the value it points at and any values contained within it, so & and &mut references conflict if the regions of memory they cover have any overlap. For example, here we say &mut list[1] to get a mutable reference to the second slot in the array, and then try to print the list before using list_mut to replace the second element of the array.

fn main() {
    let mut list = vec![1, 2, 3];
    let list_mut = &mut list[1];
    show(&list);
    *list_mut = 4;
}

This is not legal: we can’t take a reference to the whole list while holding a mutable reference to something inside it. Reversing these statements so that we print the list after we’ve finished using &mut list[1] is fine:

fn main() {
    let mut list = vec![1, 2, 3];
    let list_mut = &mut list[1];
    *list_mut = 4;
    show(&list);
}

This compiles and prints the Vec after its second element has been replaced.

[1, 4, 3]

Note that in higher-level languages like Ruby or JavaScript, the expression list[1] gives a reference to the object that is currently second in the list. That is, if we have the following list of JS objects:

let list = [{ id: 1 }, { id: 2 }, { id: 3 }];
let second = list[1];

Then second returns a reference to the object { id: 2 }, and it continues to point at this object if something is inserted at the front of the list:

// list = [{ id: 1 }, { id: 2 }, { id: 3 }]
//                    ---------
//                        ^
//                        |
//                     second

list.unshift({ id: 4 });

// list = [{ id: 4 }, { id: 1 }, { id: 2 }, { id: 3 }]
//                               ---------
//                                   ^
//                                   |
//                                second

So if we now said second.id = 5, the contents of list would be:

// list = [{ id: 4 }, { id: 1 }, { id: 5 }, { id: 3 }]

In Rust, references work more like pointers in C. The expressions &list[1] and &mut list[1] are pointers to the second slot in the array, as a fixed address in memory. So if we had second = &mut list[1] with the initial list state:

// list = vec![{ id: 1 }, { id: 2 }, { id: 3 }]
//                        ^
//                        |
//                     second

If another item was inserted at the front of the list, then second would continue to point at the same location, not at the same object it was pointing at before:

// list = vec![{ id: 4 }, { id: 1 }, { id: 2 }, { id: 3 }]
//                        ^
//                        |
//                     second

Adding an item to a Vec can also cause the entire structure to be moved in memory, as it may require more space to hold the new item. In that case, the second reference would end up pointing at memory that’s no longer in use. These are some of the many problems that are prevented by disallowing overlapping &mut references to co-exist: modifying a structure may invalidate any existing references into the structure, so you’re only allowed to create a &mut reference if there are no other references pointing into the same structure. As such, & references provide true immutability: the existence of a & reference to a structure prevents anything inside it, and sometimes the structure containing the referenced value, from being changed.

So unlike languages that embrace pure functions, like Erlang, Rust still allows mutation, it just restricts it with compile-time checks. You can let a function read from a structure by giving it a & reference, and be sure the referenced value will not change while that & reference is live. The borrow checker – the part of the compiler that enforces the “one &mut or many &s” rule – provides similar guarantees to a database letting you lock records or running transactions with serializable isolation. It lets many functions read a value at the same time but if a function wants to modify the value, it can only do that while no other function is reading or writing it. In fact when I was first hearing about Rust, it sounded a lot like people had applied database concurrency control theory to a programming language!

We’ve already seen how Rust prevents various read/write phenomena in single-threaded code, but it really shines when you want to write concurrent code. A driving force in the design of this system was to prevent data races in threaded code, by letting the compiler reject programs with concurrency bugs. For example, let’s try to rewrite one of our early examples of a counter that is incremented by a set of threads. We’ll start four threads, each of which tries to increment counter a thousand times.

use std::thread;

fn main() {
    let mut counter = 0;
    let mut threads = Vec::new();

    for _ in 0..4 {
        let th = thread::spawn(|| {
            for _ in 0..1000 {
                counter += 1;
            }
        });
        threads.push(th);
    }

    for th in threads {
        th.join().unwrap();
    }

    println!("counter: {}", counter);
}

As we saw in part 1, this program is not safe: allowing multiple threads to read and write a counter will result in lost updates. Rust knows this, and the compiler rejects this program. It gives us three distinct error messages, the first being:

error[E0499]: cannot borrow `counter` as mutable more than once at a time
  --> counter.rs:8:32
   |
8  |           let th = thread::spawn(|| {
   |                    -             ^^ mutable borrow starts here in previous
   |                    |                iteration of loop
   |  __________________|
   | |
9  | |             for _ in 0..1000 {
10 | |                 counter += 1;
   | |                 ------- borrows occur due to use of `counter` in closure
11 | |             }
12 | |         });
   | |__________- argument requires that `counter` is borrowed for `'static`

This says we’re trying to create multiple &mut references to counter at the same time, in the loop that starts the threads up. The &mut reference is needed in order to mutate counter in the statement counter += 1. The reason the &mut references end up coexisting is down to how closures work. The structure passed to thread::spawn(), that is the expression || { ... }, is a closure, which is a block of code that can be stored and executed at will, just like anonymous functions in JavaScript, blocks in Ruby, or lambdas in Java or Python. To do that, it needs to capture references to any variables it accesses that are defined outside the closure; here counter is declared outside the closure and the closure includes the statement counter += 1, so the closure object, which is just a struct in memory like any other, needs to hold a &mut reference to counter.

What makes the &mut references overlap in time is the lifetime of the closure: the compiler tells us “argument requires that counter is borrowed for 'static”. 'static is Rust jargon for “the duration of the whole program”, in other words it’s saying the closure will exist indefinitely with respect to the function that created it. Think about what thread::spawn() does: it doesn’t execute the closure immediately, waiting for it to complete. It starts a new thread of execution, so the closure will begin executing at some unknown time, concurrently with other threads, possibly after the enclosing function main() has returned. That means the closures we create inside the outer loop may exist at the same time, and therefore any references they capture must be ones that are allowed to coexist.

In effect, this code is equivalent to the following, where we try to push several &mut to the same value into a Vec:

fn main() {
    let mut counter = 0;
    let mut mut_refs = Vec::new();

    for _ in 0..4 {
        mut_refs.push(&mut counter);
    }
}

The thread code fails to compile for the same reason: it’s creating structures that contain &mut counter and those structures remain in existence indefinitely, which is what “borrowed for 'static” is getting at.

The second error we get is also related to the 'static lifetime of the closures passed to thread::spawn():

error[E0373]: closure may outlive the current function, but it borrows
`counter`, which is owned by the current function
  --> counter.rs:8:32
   |
8  |         let th = thread::spawn(|| {
   |                                ^^ may outlive borrowed value `counter`
9  |             for _ in 0..1000 {
10 |                 counter += 1;
   |                 ------- `counter` is borrowed here
   |
note: function requires argument type to outlive `'static`
  --> counter.rs:8:18
   |
8  |           let th = thread::spawn(|| {
   |  __________________^
9  | |             for _ in 0..1000 {
10 | |                 counter += 1;
11 | |             }
12 | |         });
   | |__________^
help: to force the closure to take ownership of `counter` (and any other
referenced variables), use the `move` keyword
   |
8  |         let th = thread::spawn(move || {
   |                                ^^^^^^^

The note here is telling us that thread::spawn() only accepts closures that “outlive 'static”, meaning they remain valid indefinitely rather than being bound to the lifetime of another value or function call. The closure contains a reference to counter, which is a local variable in the main() function, and which will therefore cease to exist when main() returns. But, the closure is required to be able to outlive the current function, so it’s illegal for it to hold references to values that will vanish when the function returns!

So, closures passed to thread::spawn() are not allowed to refer to local variables from the calling function. Instead, the closures must take ownership of any values they use, which can be achieved by putting the word move before the pipes.

The final error requires less explanation; it is simply a & vs &mut conflict that we’ve already seen in single-threaded code. We’re creating a set of &mut references to counter that will live for an indefinite amount of time (“borrowed for 'static”), and so it’s illegal to take a & reference to counter in order to print it at the end of main().

error[E0502]: cannot borrow `counter` as immutable because it is also borrowed
as mutable
  --> counter.rs:20:29
   |
8  |           let th = thread::spawn(|| {
   |                    -             -- mutable borrow occurs here
   |  __________________|
   | |
9  | |             for _ in 0..1000 {
10 | |                 counter += 1;
   | |                 ------- first borrow occurs due to use of `counter` in
   | |                         closure
11 | |             }
12 | |         });
   | |__________- argument requires that `counter` is borrowed for `'static`
...
20 |       println!("counter: {}", counter);
   |                               ^^^^^^^ immutable borrow occurs here

Let’s see if we can fix the errors relating to the closure by taking the compiler’s advice and adding move in front of the closure:

        let th = thread::spawn(move || {
            for _ in 0..1000 {
                counter += 1;
            }
        });

This does actually compile, but the result of running the resulting binary is surprising:

counter: 0

We get a hint about why it doesn’t do what we expect from this compiler warning:

warning: unused variable: `counter`
  --> counter.rs:10:17
   |
10 |                 counter += 1;
   |                 ^^^^^^^
   |
   = note: `#[warn(unused_variables)]` on by default
   = help: did you mean to capture by reference instead?

The counter variable we use inside each thread is unused, which means it’s never read. That’s because by adding move to the closure, we made it take ownership of all the variables it uses. Instead of storing a reference to counter, the closure instead contains a copy of it. Integers implement Copy, so any situation that would normally result in a value being moved results in integers being copied instead. So, each thread starts up with a copy of counter, it increments its local copy, and leaves counter itself untouched, and so the value of counter at the end of the program is zero.

It’s looking like we can’t write the program we want. We can’t have the threads holding &mut references to counter, so how can they all update it? The first part of the answer is that Rust provides a Mutex struct, that can be wrapped around any value to control access to it. Here, we’ll use Mutex::new(0) for our counter, and then any time we want to access it we have to call counter.lock().unwrap(), which gives us a mutable reference to the wrapped value, but only allowing one thread to have access at a time. This forces the threads to access the counter sequentially while incrementing it.

use std::sync::Mutex;
use std::thread;

fn main() {
    let counter = Mutex::new(0);
    let mut threads = Vec::new();

    for _ in 0..4 {
        let th = thread::spawn(move || {
            for _ in 0..1000 {
                let mut ctr = counter.lock().unwrap();
                *ctr += 1;
            }
        });
        threads.push(th);
    }

    for th in threads {
        th.join().unwrap();
    }

    let ctr = counter.lock().unwrap();
    println!("counter: {}", *ctr);
}

Note that we never explicitly unlock the Mutex. The ctr reference that we get by calling counter.lock() goes out of scope at the end of the loop body, and the compiler automatically destroys it at that point. For Mutex locks, the destruction procedure releases the lock so another thread can acquire it. All types in Rust can add custom behaviour when their values go out of scope by implementing the Drop trait.

Unfortunately this still is not enough to make the program compile:

error[E0382]: use of moved value: `counter`
  --> counter.rs:9:32
   |
5  |     let counter = Mutex::new(0);
   |         ------- move occurs because `counter` has type
   |                 `std::sync::Mutex<i32>`, which does not implement the
   |                 `Copy` trait
...
9  |         let th = thread::spawn(move || {
   |                                ^^^^^^^ value moved into closure here, in
   |                                        previous iteration of loop
10 |             for _ in 0..1000 {
11 |                 let mut ctr = counter.lock().unwrap();
   |                               ------- use occurs due to use in closure

The type Mutex<i32> is not Copy, so values of that type are moved in any situation where the ownership system determines movement is required. Here, the move closure includes a use of counter, and so the first thread we create takes ownership of counter and none of the subsequent threads can use it. We need a way for all the threads to share a reference to the Mutex and ultimately access the integer it wraps.

This is where the final ingredient in the solution comes in: Rust’s Arc type. ARC stands for atomic reference counter, and this type provides a way for many threads to share a reference to a value. We modify the first section of our program like so:

    let counter = Arc::new(Mutex::new(0));
    let mut threads = Vec::new();

    for _ in 0..4 {
        let ctr_ref = counter.clone();

        let th = thread::spawn(move || {
            for _ in 0..1000 {
                let mut ctr = ctr_ref.lock().unwrap();
                *ctr += 1;
            }
        });
        threads.push(th);
    }

There are two main changes here. First, counter is now initialised as Arc::new(Mutex::new(0)): an atomic reference counter to a mutex-wrapped 32-bit integer, or Arc<Mutex<i32>>. Second, instead of each thread taking ownership of the Mutex, it takes ownership of ctr_ref, a clone of the Arc. Cloning an Arc just copies a reference, not the wrapped value, in a way that’s safe to share the references across threads, so we can now give all the threads a way to access the shared counter and mutate it safely.

The program now compiles without warnings and prints the expected result:

counter: 4000

There’s a fair bit to learn in order to make this work, and this is part of why Rust’s learning curve is renowned for being steep. However, it does mean that if your program compiles it’s almost guaranteed not to contain any data races. It is possible to subvert these guarantees, for example I can break the above program by using separate Mutex locks for reading the counter value and then updating it:

        let th = thread::spawn(move || {
            for _ in 0..1000 {
                let tmp = *ctr_ref.lock().unwrap();
                *ctr_ref.lock().unwrap() = tmp + 1;
            }
        });

But this is more effort than writing the code that doesn’t have this bug, and still requires fixing all the errors we saw along the way in order to even be allowed to compile this program. Rust puts up a lot of barriers to you writing concurrency bugs, so much so that if you manage to get around them you’ve almost certainly done so deliberately.

This distinction reminds me of the difference in how PostgreSQL and MySQL handle failed INSERT queries that violate uniqueness constraints, as we saw in part 3. PostgreSQL makes the transaction unusable unless you use savepoints, whereas MySQL keeps on running your code in a way that might deadlock when you get to production. It’s an ergonomic change, marking the difference between you writing bugs deliberately or by accident. A system that raises errors on code that isn’t guaranteed to work is helping you avoid pushing bugs to production, while a system that is more lax might feel easier to use or to experiment with, but ultimate lets you ship subtle bugs that are harder to diagnose.

JavaScript has a similar problem. At first it feels easier to use than Rust – the learning curve is less steep, and it’s much less effort to get your programs to run at all. But writing production-quality concurrent JavaScript is extremely difficult, and the language itself gives you barely any help at all. In JS, you see weird race conditions happen in production and spend countless hours trying to find out why they’re happening. In Rust, the compiler won’t let you run a program that contains race conditions. Initially that feels incredibly restrictive and it can be a frustrating learning experience, but it will always tell you why your program doesn’t work, which is a substantial part of the work of debugging anything. Sometimes a system that feels easy at first is simply hiding complexity that you will need to figure out on your own later.

A similar distinction exists between the Puppet and Chef config management systems. Puppet has its own custom language for declaring what should be installed on your machines, and it’s highly restrictive – it has a limited syntax, and you have to explicitly declare all the dependencies between things you’re installing. Chef recipes are written in Ruby; they still use a declarative API but Chef executes them more like normal scripts, running a recipe from top-to-bottom without doing any static analysis first. Puppet feels restrictive but it can spot where your config makes no sense, for example because it contains conflicting or contradictory instructions, and it can do this because it has a custom language that only lets you write code that will have predictable results when executed. Chef can feel easier to get going with as it uses an already-established language, but it will also happily run configuration that makes no sense, for example you can declare that a package both should be and should not be installed, and the resulting machine state just depends on whatever order those instructions happen to run in.

All technical system designs seem to sit somewhere on this spectrum between feeling easy to get started with, and preventing you making mistakes. I would put JavaScript and Chef on the “easy to use” end, and Rust and Puppet on the “hard to break” end. Being hard to break usually means that you as a user have to explain more of your intent to the system, for example by using a static type system or by declaring all dependencies, and this effort pays off to the extent that the system catches bugs you wouldn’t have caught yourself. One of the first things I like to do when learning a new system is try to write programs that make no sense, to see if the system realises and tells me off, or executes them anyway. This sort of information is especially useful when maintaining systems, because I need to know how much the machine is going to spot bugs I’ve introduced and how much testing I need to do myself.

One of the interesting things about Rust is how much it’s pushed the ergonomics of language design, and tried to reconcile the needs of safety and ease-of-use, which often feel like they pull in opposing directions. Its designers are always looking for ways to make the language easier to use while keeping its guarantees about memory safety, augmenting the syntax to make certain tasks simpler and improving the error messages the compiler emits.

As programmers designing systems, UIs, APIs, we need to be asking how we can give users the functionality they need, in as simple an interface as possible, that also makes it easier to write correct code than to write bugs. Ultimately this comes down to the set of operations we expose and how we package and organise them, and there is a wealth of different perspectives we can draw on from programming languages, databases and other systems. Pay attention to how your system allows data to move around: everything else flows from there.