Hierarchical state machines for designing event-driven systems.
Features
- Hierachical state machines
- State-local storage
- Compatible with
#![no_std], state machines are defined in ROM and no heap memory allocations. - (Optional) macro's for reducing boilerplate.
- Support for generics.
- Support for async actions and handlers (only on
std).
Overview
A simple blinky state machine:
┌─────────────────────────┐
│ Blinking │◀─────────┐
│ ┌───────────────┐ │ │
│ ┌─▶│ LedOn │──┐ │ ┌───────────────┐
│ │ └───────────────┘ │ │ │ NotBlinking │
│ │ ┌───────────────┐ │ │ └───────────────┘
│ └──│ LedOff │◀─┘ │ ▲
│ └───────────────┘ │──────────┘
└─────────────────────────┘
#[derive(Default)]
pub struct Blinky;
pub enum Event {
TimerElapsed,
ButtonPressed
}
#[state_machine(initial = "State::led_on()")]
impl Blinky {
#[state(superstate = "blinking")]
fn led_on(event: &Event) -> Response<State> {
match event {
Event::TimerElapsed => Transition(State::led_off()),
_ => Super
}
}
#[state(superstate = "blinking")]
fn led_off(event: &Event) -> Response<State> {
match event {
Event::TimerElapsed => Transition(State::led_on()),
_ => Super
}
}
#[superstate]
fn blinking(event: &Event) -> Response<State> {
match event {
Event::ButtonPressed => Transition(State::not_blinking()),
_ => Super
}
}
#[state]
fn not_blinking(event: &Event) -> Response<State> {
match event {
Event::ButtonPressed => Transition(State::led_on()),
_ => Super
}
}
}
fn main() {
let mut state_machine = Blinky::default().state_machine();
state_machine.handle(&Event::TimerElapsed);
state_machine.handle(&Event::ButtonPressed);
}(See the macro/blinky example for the full code with comments. Or see no_macro/blinky for a version without using macro's).
States are defined by writing methods inside the impl block and adding the #[state] attribute to them. When an event is submitted to the state machine, the method associated with the current state will be called to process it. By default this event is mapped to the event argument of the method.
#[state]
fn led_on(event: &Event) -> Response<State> {
Transition(State::led_off())
}Every state must return a Response. A Response can be one of three things:
Handled: The event has been handled.Transition: Transition to another state.Super: Defer the event to the parent superstate.
Superstates allow you to create a hierarchy of states. States can defer an event to their superstate by returning the Super response.
#[state(superstate = "blinking")]
fn led_on(event: &Event) -> Response<State> {
match event {
Event::TimerElapsed => Transition(State::led_off()),
Event::ButtonPressed => Super
}
}
#[superstate]
fn blinking(event: &Event) -> Response<State> {
match event {
Event::ButtonPressed => Transition(State::not_blinking()),
_ => Super
}
}Superstates can themselves also have superstates.
Actions run when entering or leaving states during a transition.
#[state(entry_action = "enter_led_on", exit_action = "exit_led_on")]
fn led_on(event: &Event) -> Response<State> {
Transition(State::led_off())
}
#[action]
fn enter_led_on() {
println!("Entered on");
}
#[action]
fn exit_led_on() {
println!("Exited on");
}If the type on which your state machine is implemented has any fields, you can access them inside all states, superstates or actions.
#[state]
fn led_on(&mut self, event: &Event) -> Response<State> {
match event {
Event::TimerElapsed => {
self.led = false;
Transition(State::led_off())
}
_ => Super
}
}Or alternatively, set led inside the entry action.
#[action]
fn enter_led_off(&mut self) {
self.led = false;
}Sometimes you have data that only exists in a certain state. Instead of adding this data to the shared storage and potentially having to unwrap an Option<T>, you can add it as an input to your state handler.
#[state]
fn led_on(counter: &mut u32, event: &Event) -> Response<State> {
match event {
Event::TimerElapsed => {
*counter -= 1;
if *counter == 0 {
Transition(State::led_off())
} else {
Handled
}
}
Event::ButtonPressed => Transition(State::led_on(10))
}
}counter is only available in the led_on state but can also be accessed in its superstates and actions.
When state machines are used in a larger systems it can sometimes be necessary to pass in an external mutable context. By default this context is mapped to the context argument of the method.
#[state]
fn led_on(context: &mut Context, event: &Event) -> Response<State> {
match event {
Event::TimerElapsed => {
context.do_something();
Handled
}
_ => Super
}
}You will then be required to use the handle_with_context method to submit events to the state machine.
state_machine.handle_with_context(&Event::TimerElapsed, &mut context);For logging purposes you can define two callbacks that will be called at specific points during state machine execution.
on_dispatchis called before an event is dispatched to a specific state or superstate.on_transitionis called after a transition has occured.
#[state_machine(
initial = "State::on()",
on_dispatch = "Self::on_dispatch",
on_transition = "Self::on_transition",
state(derive(Debug)),
superstate(derive(Debug))
)]
impl Blinky {
...
}
impl Blinky {
fn on_transition(&mut self, source: &State, target: &State) {
println!("transitioned from `{:?}` to `{:?}`", source, target);
}
fn on_dispatch(&mut self, state: StateOrSuperstate<Blinky>, event: &Event) {
println!("dispatched `{:?}` to `{:?}`", event, state);
}
}All handlers and actions can be made async. (This is only available on std for now and requires the async feature to be enabled).
#[state_machine(initial = "State::led_on()")]
impl Blinky {
#[state]
async fn led_on(event: &Event) -> Response<State> {
match event {
Event::TimerElapsed => Transition(State::led_off()),
_ => Super
}
}
}The #[state_machine] macro will then automatically detect that async functions are being used
and generate an async state machine.
async fn main() {
let mut state_machine = Blinky::default().state_machine();
state_machine.handle(&Event::TimerElapsed).await;
state_machine.handle(&Event::ButtonPressed).await;
}A lot of the implemenation details are dealt with by the #[state_machine] macro, but it's always valuable to understand what's happening behind the scenes. Furthermore, you'll see that the generated code is actually pretty straight-forward and could easily be written by hand, so if you prefer to avoid using macro's this is totally feasible.
The goal of statig is to represent a hierarchical state machine. Conceptually a hierarchical state machine can be tought of as tree.
┌ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ┐
Top
└ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ┘
│
┌────────────┴────────────┐
│ │
┌─────────────────────┐ ╔═════════════════════╗
│ Blinking │ ║ NotBlinking ║
│─────────────────────│ ╚═════════════════════╝
│ counter: &'a usize │
└─────────────────────┘
│
┌────────────┴────────────┐
│ │
╔═════════════════════╗ ╔═════════════════════╗
║ LedOn ║ ║ LedOff ║
║─────────────────────║ ║─────────────────────║
║ counter: usize ║ ║ counter: usize ║
╚═════════════════════╝ ╚═════════════════════╝
Nodes at the edge of the tree are called leaf-states and are represented by an enum in statig. If data only exists in a particular state we can give that state ownership of the data. This is referred to as 'state-local storage'. For example counter only exists in the LedOn and LedOff state.
enum State {
LedOn { counter: usize },
LedOff { counter: usize },
NotBlinking
}States such as Blinking are called superstates. They define shared behavior of their child states. Superstates are also represented by an enum, but instead of owning their data, they borrow it from the underlying state.
enum Superstate<'sub> {
Blinking { counter: &'sub usize }
}The association between states and their handlers is then expressed in the State and Superstate traits with the call_handler() method.
impl statig::State<Blinky> for State {
fn call_handler(&mut self, blinky: &mut Blinky, event: &Event) -> Response<Self> {
match self {
State::LedOn { counter } => blinky.led_on(counter, event),
State::LedOff { counter } => blinky.led_off(counter, event),
State::NotBlinking => blinky.not_blinking(event)
}
}
}
impl statig::Superstate<Blinky> for Superstate {
fn call_handler(&mut self, blinky: &mut Blinky, event: &Event) -> Response<Self> {
match self {
Superstate::Blinking { counter } => blinky.blinking(counter, event),
}
}
}The association between states and their actions is expressed in a similar fashion.
impl statig::State<Blinky> for State {
...
fn call_entry_action(&mut self, blinky: &mut Blinky) {
match self {
State::LedOn { counter } => blinky.enter_led_on(counter),
State::LedOff { counter } => blinky.enter_led_off(counter),
State::NotBlinking => blinky.enter_not_blinking()
}
}
fn call_exit_action(&mut self, blinky: &mut Blinky) {
match self {
State::LedOn { counter } => blinky.exit_led_on(counter),
State::LedOff { counter } => blinky.exit_led_off(counter),
State::NotBlinking => blinky.exit_not_blinking()
}
}
}
impl statig::Superstate<Blinky> for Superstate {
...
fn call_entry_action(&mut self, blinky: &mut Blinky) {
match self {
Superstate::Blinking { counter } => blinky.enter_blinking(counter),
}
}
fn call_exit_action(&mut self, blinky: &mut Blinky) {
match self {
Superstate::Blinking { counter } => blinky.exit_blinking(counter),
}
}
}The tree structure of states and their superstates is expressed in the superstate method of the State and Superstate trait.
impl statig::State<Blinky> for State {
...
fn superstate(&mut self) -> Option<Superstate<'_>> {
match self {
State::LedOn { counter } => Some(Superstate::Blinking { counter }),
State::LedOff { counter } => Some(Superstate::Blinking { counter }),
State::NotBlinking => None
}
}
}
impl<'sub> statig::Superstate<Blinky> for Superstate<'sub> {
...
fn superstate(&mut self) -> Option<Superstate<'_>> {
match self {
Superstate::Blinking { .. } => None
}
}
}When an event arrives, statig will first dispatch it to the current leaf state. If this state returns a Super response, it will then be dispatched to that state's superstate, which in turn returns its own response. Every time an event is defered to a superstate, statig will traverse upwards in the graph until it reaches the Top state. This is an implicit superstate that will consider every event as handled.
In case the returned response is a Transition, statig will perform a transition sequence by traversing the graph from the current source state to the target state by taking the shortest possible path. When this path is going upwards from the source state, every state that is passed will have its exit action executed. And then similarly when going downward, every state that is passed will have its entry action executed.
For example when transitioning from the LedOn state to the NotBlinking state the transition sequence looks like this:
- Exit the
LedOnstate - Exit the
Blinkingstate - Enter the
NotBlinkingstate
For comparison, the transition from the LedOn state to the LedOff state looks like this:
- Exit the
LedOnstate - Enter the
LedOffstate
We don't execute the exit or entry action of Blinking as this superstate is shared between the LedOn and LedOff state.
Entry and exit actions also have access to state-local storage, but note that exit actions operate on state-local storage of the source state and that entry actions operate on the state-local storage of the target state.
For example chaning the value of counter in the exit action of LedOn will have no effect on the value of counter in the LedOff state.
Finally, the StateMachine trait is implemented on the type that will be used for the shared storage.
impl IntoStateMachine for Blinky {
type State = State;
type Superstate<'sub> = Superstate<'sub>;
type Event<'evt> = Event;
type Context<'ctx> = Context;
const INITIAL: State = State::off(10);
}Short answer: nothing. #[state_machine] simply parses the underlying impl block and derives some code based on its content and adds it to your source file. Your code will still be there, unchanged. In fact #[state_machine] could have been a derive macro, but at the moment Rust only allows derive macros to be used on enums and structs. If you'd like to see what the generated code looks like take a look at the test with and without macros.
I would say they serve a different purpose. The typestate pattern is very useful for designing an API as it is able to enforce the validity of operations at compile time by making each state a unique type. But statig is designed to model a dynamic system where events originate externally and the order of operations is determined at run time. More concretely, this means that the state machine is going to sit in a loop where events are read from a queue and submitted to the state machine using the handle() method. If we want to do the same with a state machine that uses the typestate pattern we'd have to use an enum to wrap all our different states and match events to operations on these states. This means extra boilerplate code for little advantage as the order of operations is unknown so it can't be checked at compile time. On the other hand statig gives you the ability to create a hierarchy of states which I find to be invaluable as state machines grow in complexity.
The idea for this library came from reading the book Practical UML Statecharts in C/C++. I highly recommend it if you want to learn how to use state machines to design complex systems.