Rust 0.8
8a4f0fa6

1 Introduction

Rust provides safe concurrency through a combination of lightweight, memory-isolated tasks and message passing. This tutorial will describe the concurrency model in Rust, how it relates to the Rust type system, and introduce the fundamental library abstractions for constructing concurrent programs.

Rust tasks are not the same as traditional threads: rather, they are considered green threads, lightweight units of execution that the Rust runtime schedules cooperatively onto a small number of operating system threads. On a multi-core system Rust tasks will be scheduled in parallel by default. Because tasks are significantly cheaper to create than traditional threads, Rust can create hundreds of thousands of concurrent tasks on a typical 32-bit system. In general, all Rust code executes inside a task, including the main function.

In order to make efficient use of memory Rust tasks have dynamically sized stacks. A task begins its life with a small amount of stack space (currently in the low thousands of bytes, depending on platform), and acquires more stack as needed. Unlike in languages such as C, a Rust task cannot accidentally write to memory beyond the end of the stack, causing crashes or worse.

Tasks provide failure isolation and recovery. When a fatal error occurs in Rust code as a result of an explicit call to fail!(), an assertion failure, or another invalid operation, the runtime system destroys the entire task. Unlike in languages such as Java and C++, there is no way to catch an exception. Instead, tasks may monitor each other for failure.

Tasks use Rust's type system to provide strong memory safety guarantees. In particular, the type system guarantees that tasks cannot share mutable state with each other. Tasks communicate with each other by transferring owned data through the global exchange heap.

1.1 A note about the libraries

While Rust's type system provides the building blocks needed for safe and efficient tasks, all of the task functionality itself is implemented in the standard and extra libraries, which are still under development and do not always present a consistent or complete interface.

For your reference, these are the standard modules involved in Rust concurrency at this writing:

2 Basics

The programming interface for creating and managing tasks lives in the task module of the std library, and is thus available to all Rust code by default. At its simplest, creating a task is a matter of calling the spawn function with a closure argument. spawn executes the closure in the new task.

// Print something profound in a different task using a named function
fn print_message() { println("I am running in a different task!"); }
spawn(print_message);

// Print something more profound in a different task using a lambda expression
spawn( || println("I am also running in a different task!") );

// The canonical way to spawn is using `do` notation
do spawn {
    println("I too am running in a different task!");
}

In Rust, there is nothing special about creating tasks: a task is not a concept that appears in the language semantics. Instead, Rust's type system provides all the tools necessary to implement safe concurrency: particularly, owned types. The language leaves the implementation details to the standard library.

The spawn function has a very simple type signature: fn spawn(f: ~fn()). Because it accepts only owned closures, and owned closures contain only owned data, spawn can safely move the entire closure and all its associated state into an entirely different task for execution. Like any closure, the function passed to spawn may capture an environment that it carries across tasks.

// Generate some state locally
let child_task_number = generate_task_number();

do spawn {
   // Capture it in the remote task
   println(fmt!("I am child number %d", child_task_number));
}

2.1 Communication

Now that we have spawned a new task, it would be nice if we could communicate with it. Recall that Rust does not have shared mutable state, so one task may not manipulate variables owned by another task. Instead we use pipes.

A pipe is simply a pair of endpoints: one for sending messages and another for receiving messages. Pipes are low-level communication building-blocks and so come in a variety of forms, each one appropriate for a different use case. In what follows, we cover the most commonly used varieties.

The simplest way to create a pipe is to use the comm::stream function to create a (Port, Chan) pair. In Rust parlance, a channel is a sending endpoint of a pipe, and a port is the receiving endpoint. Consider the following example of calculating two results concurrently:

let (port, chan): (Port<int>, Chan<int>) = stream();

do spawn || {
    let result = some_expensive_computation();
    chan.send(result);
}

some_other_expensive_computation();
let result = port.recv();

Let's examine this example in detail. First, the let statement creates a stream for sending and receiving integers (the left-hand side of the let, (chan, port), is an example of a destructuring let: the pattern separates a tuple into its component parts).

let (port, chan): (Port<int>, Chan<int>) = stream();

The child task will use the channel to send data to the parent task, which will wait to receive the data on the port. The next statement spawns the child task.

do spawn || {
    let result = some_expensive_computation();
    chan.send(result);
}

Notice that the creation of the task closure transfers chan to the child task implicitly: the closure captures chan in its environment. Both Chan and Port are sendable types and may be captured into tasks or otherwise transferred between them. In the example, the child task runs an expensive computation, then sends the result over the captured channel.

Finally, the parent continues with some other expensive computation, then waits for the child's result to arrive on the port:

some_other_expensive_computation();
let result = port.recv();

The Port and Chan pair created by stream enables efficient communication between a single sender and a single receiver, but multiple senders cannot use a single Chan, and multiple receivers cannot use a single Port. What if our example needed to compute multiple results across a number of tasks? The following program is ill-typed:

let (port, chan) = stream();

do spawn {
    chan.send(some_expensive_computation());
}

// ERROR! The previous spawn statement already owns the channel,
// so the compiler will not allow it to be captured again
do spawn {
    chan.send(some_expensive_computation());
}

Instead we can use a SharedChan, a type that allows a single Chan to be shared by multiple senders.

let (port, chan) = stream();
let chan = SharedChan::new(chan);

for init_val in range(0u, 3) {
    // Create a new channel handle to distribute to the child task
    let child_chan = chan.clone();
    do spawn {
        child_chan.send(some_expensive_computation(init_val));
    }
}

let result = port.recv() + port.recv() + port.recv();

Here we transfer ownership of the channel into a new SharedChan value. Like Chan, SharedChan is a non-copyable, owned type (sometimes also referred to as an affine or linear type). Unlike with Chan, though, the programmer may duplicate a SharedChan, with the clone() method. A cloned SharedChan produces a new handle to the same channel, allowing multiple tasks to send data to a single port. Between spawn, stream and SharedChan, we have enough tools to implement many useful concurrency patterns.

Note that the above SharedChan example is somewhat contrived since you could also simply use three stream pairs, but it serves to illustrate the point. For reference, written with multiple streams, it might look like the example below.

// Create a vector of ports, one for each child task
let ports = do vec::from_fn(3) |init_val| {
    let (port, chan) = stream();
    do spawn {
        chan.send(some_expensive_computation(init_val));
    }
    port
};

// Wait on each port, accumulating the results
let result = ports.iter().fold(0, |accum, port| accum + port.recv() );

2.2 Backgrounding computations: Futures

With extra::future, rust has a mechanism for requesting a computation and getting the result later.

The basic example below illustrates this.
fn fib(n: uint) -> uint {
    // lengthy computation returning an uint
    12586269025
}

let mut delayed_fib = extra::future::Future::spawn (|| fib(50) );
make_a_sandwich();
println(fmt!("fib(50) = %?", delayed_fib.get()))

The call to future::spawn returns immediately a future object regardless of how long it takes to run fib(50). You can then make yourself a sandwich while the computation of fib is running. The result of the execution of the method is obtained by calling get on the future. This call will block until the value is available (i.e. the computation is complete). Note that the future needs to be mutable so that it can save the result for next time get is called.

Here is another example showing how futures allow you to background computations. The workload will be distributed on the available cores.
fn partial_sum(start: uint) -> f64 {
    let mut local_sum = 0f64;
    for num in range(start*100000, (start+1)*100000) {
        local_sum += (num as f64 + 1.0).pow(&-2.0);
    }
    local_sum
}

fn main() {
    let mut futures = vec::from_fn(1000, |ind| do extra::future::Future::spawn { partial_sum(ind) });

    let mut final_res = 0f64;
    for ft in futures.mut_iter()  {
        final_res += ft.get();
    }
    println(fmt!("π^2/6 is not far from : %?", final_res));
}

2.3 Sharing immutable data without copy: Arc

To share immutable data between tasks, a first approach would be to only use pipes as we have seen previously. A copy of the data to share would then be made for each task. In some cases, this would add up to a significant amount of wasted memory and would require copying the same data more than necessary.

To tackle this issue, one can use an Atomically Reference Counted wrapper (Arc) as implemented in the extra library of Rust. With an Arc, the data will no longer be copied for each task. The Arc acts as a reference to the shared data and only this reference is shared and cloned.

Here is a small example showing how to use Arcs. We wish to run concurrently several computations on a single large vector of floats. Each task needs the full vector to perform its duty.
use extra::arc::Arc;

fn pnorm(nums: &~[float], p: uint) -> float {
    nums.iter().fold(0.0, |a,b| a+(*b).pow(&(p as float)) ).pow(&(1f / (p as float)))
}

fn main() {
    let numbers = vec::from_fn(1000000, |_| rand::random::<float>());
    println(fmt!("Inf-norm = %?",  *numbers.iter().max().unwrap()));

    let numbers_arc = Arc::new(numbers);

    for num in range(1u, 10) {
        let (port, chan)  = stream();
        chan.send(numbers_arc.clone());

        do spawn {
            let local_arc : Arc<~[float]> = port.recv();
            let task_numbers = local_arc.get();
            println(fmt!("%u-norm = %?", num, pnorm(task_numbers, num)));
        }
    }
}
The function pnorm performs a simple computation on the vector (it computes the sum of its items at the power given as argument and takes the inverse power of this value). The Arc on the vector is created by the line
let numbers_arc=Arc::new(numbers);
and a clone of it is sent to each task
chan.send(numbers_arc.clone());

copying only the wrapper and not its contents.

Each task recovers the underlying data by
let task_numbers = local_arc.get();

and can use it as if it were local.

The arc module also implements Arcs around mutable data that are not covered here.

3 Handling task failure

Rust has a built-in mechanism for raising exceptions. The fail!() macro (which can also be written with an error string as an argument: fail!( ~reason)) and the assert! construct (which effectively calls fail!() if a boolean expression is false) are both ways to raise exceptions. When a task raises an exception the task unwinds its stack---running destructors and freeing memory along the way---and then exits. Unlike exceptions in C++, exceptions in Rust are unrecoverable within a single task: once a task fails, there is no way to "catch" the exception.

All tasks are, by default, linked to each other. That means that the fates of all tasks are intertwined: if one fails, so do all the others.

// Create a child task that fails
do spawn { fail!() }

// This will also fail because the task we spawned failed
do_some_work();

While it isn't possible for a task to recover from failure, tasks may notify each other of failure. The simplest way of handling task failure is with the try function, which is similar to spawn, but immediately blocks waiting for the child task to finish. try returns a value of type Result<T, ()>. Result is an enum type with two variants: Ok and Err. In this case, because the type arguments to Result are int and (), callers can pattern-match on a result to check whether it's an Ok result with an int field (representing a successful result) or an Err result (representing termination with an error).

let result: Result<int, ()> = do task::try {
    if some_condition() {
        calculate_result()
    } else {
        fail!("oops!");
    }
};
assert!(result.is_err());

Unlike spawn, the function spawned using try may return a value, which try will dutifully propagate back to the caller in a Result enum. If the child task terminates successfully, try will return an Ok result; if the child task fails, try will return an Error result.

Note: A failed task does not currently produce a useful error value (try always returns Err(())). In the future, it may be possible for tasks to intercept the value passed to fail!().

TODO: Need discussion of future_result in order to make failure modes useful.

But not all failures are created equal. In some cases you might need to abort the entire program (perhaps you're writing an assert which, if it trips, indicates an unrecoverable logic error); in other cases you might want to contain the failure at a certain boundary (perhaps a small piece of input from the outside world, which you happen to be processing in parallel, is malformed and its processing task can't proceed). Hence, you will need different linked failure modes.

3.1 Failure modes

By default, task failure is bidirectionally linked, which means that if either task fails, it kills the other one.

do spawn {
    do spawn {
        fail!();  // All three tasks will fail.
    }
    sleep_forever();  // Will get woken up by force, then fail
}
sleep_forever();  // Will get woken up by force, then fail

If you want parent tasks to be able to kill their children, but do not want a parent to fail automatically if one of its child task fails, you can call task::spawn_supervised for unidirectionally linked failure. The function task::try, which we saw previously, uses spawn_supervised internally, with additional logic to wait for the child task to finish before returning. Hence:

let (receiver, sender): (Port<int>, Chan<int>) = stream();
do spawn {  // Bidirectionally linked
    // Wait for the supervised child task to exist.
    let message = receiver.recv();
    // Kill both it and the parent task.
    assert!(message != 42);
}
do try {  // Unidirectionally linked
    sender.send(42);
    sleep_forever();  // Will get woken up by force
}
// Flow never reaches here -- parent task was killed too.

Supervised failure is useful in any situation where one task manages multiple fallible child tasks, and the parent task can recover if any child fails. On the other hand, if the parent (supervisor) fails, then there is nothing the children can do to recover, so they should also fail.

Supervised task failure propagates across multiple generations even if an intermediate generation has already exited:

do task::spawn_supervised {
    do task::spawn_supervised {
        sleep_forever();  // Will get woken up by force, then fail
    }
    // Intermediate task immediately exits
}
wait_for_a_while();
fail!();  // Will kill grandchild even if child has already exited

Finally, tasks can be configured to not propagate failure to each other at all, using task::spawn_unlinked for isolated failure.

let (time1, time2) = (random(), random());
do task::spawn_unlinked {
    sleep_for(time2);  // Won't get forced awake
    fail!();
}
sleep_for(time1);  // Won't get forced awake
fail!();
// It will take MAX(time1,time2) for the program to finish.

3.2 Creating a task with a bi-directional communication path

A very common thing to do is to spawn a child task where the parent and child both need to exchange messages with each other. The function extra::comm::DuplexStream() supports this pattern. We'll look briefly at how to use it.

To see how DuplexStream() works, we will create a child task that repeatedly receives a uint message, converts it to a string, and sends the string in response. The child terminates when it receives 0. Here is the function that implements the child task:

fn stringifier(channel: &DuplexStream<~str, uint>) {
    let mut value: uint;
    loop {
        value = channel.recv();
        channel.send(uint::to_str(value));
        if value == 0 { break; }
    }
}

The implementation of DuplexStream supports both sending and receiving. The stringifier function takes a DuplexStream that can send strings (the first type parameter) and receive uint messages (the second type parameter). The body itself simply loops, reading from the channel and then sending its response back. The actual response itself is simply the stringified version of the received value, uint::to_str(value).

Here is the code for the parent task:

let (from_child, to_child) = DuplexStream();

do spawn {
    stringifier(&to_child);
};

from_child.send(22);
assert!(from_child.recv() == ~"22");

from_child.send(23);
from_child.send(0);

assert!(from_child.recv() == ~"23");
assert!(from_child.recv() == ~"0");

The parent task first calls DuplexStream to create a pair of bidirectional endpoints. It then uses task::spawn to create the child task, which captures one end of the communication channel. As a result, both parent and child can send and receive data to and from the other.