# rex-sm :crown:
### Design Features
- Make state machines modular units that can be reused
- State machines that use other state machines are aware of who they can use,
but state machines that are being used are not aware of state machines that use them
- No tasking/threading
- Handle multiple state machines concurrently
- Trigger other state machines, and wait for their completion, then resume
- Create/delete timers
- Send notifications outside the state machine group
- Send events to themselves
- Accept events from outside the state machine group
- Encourage state machine implementations to be represented in a `(<State>, <Event>)` format
### Inspiration
- [message-io](https://crates.io/crates/message-io)
- [async-hsm](https://crates.io/crates/async-hsm)
- [event-driven finite-state machine](https://en.wikipedia.org/wiki/Event-driven_finite-state_machine)
### Overview
`StateMachine`s are first registered with the `StateMachineManager`, which I will refer
to as simply the `Manager`. Every call to `Manager::cycle()` processes a single event.
A single event corresponds to running on a single state machine. The `Manager` accesses
the contents of the `Controller` and manipulates it. A single `Controller` is shared
amongst all state machines registered with the `Manager`.
There are two types of events `UserEvent`s and `SystemEvent`s. `UserEvent`s are passed to
`StateMachine::cycle()` while `SystemEvent`s are not. `StateMachine::cycle()` accepts a
`&mut Controller` and a `UserEvent`. The `StateMachine` uses the functions in the `Controller`
to add/remove events from the event queue; all functions do this except for timer related functions.
`SystemEvent`s are consumed by the manager and used to modify the `Controller` internals or send
data or notifications to outside the state machine group.
## Node based StateMachine Manager (in development)
## Goals
* decoupling state machine input processing from a given **state’s current enumerations**
* state signaling that all feeds into the same sink (the manager’s
`signal_queue`) ; this allows lifts and transits to be processed
**homogeneously** thus avoiding type opacity through `Box<dyn Signal>`
## In practice the design should give at most three message streams connected to a particular state machine down:
I/O
* One `Input` (handles both external and internal events)
```rust
Signal {
id: StateId<K>,
input: I,
}
```
Two outputs:
* `Signal` output (events meant to be processed as inputs for other state machines)
* `Notification` output (events meant to be processed by _anything_ that is not a state machine fed by a given `signal_queue`)
### The `StateMachineManager` owns:
* The state storage layer (using `StateStore`)
* the input event stream
* The state machine processors
### The `StateMachineManager` is responsible for:
* routing `Signal`s to the appropriate state machines
* Injecting `ProcessorContext`s into the state machines: this action is what allows state machines to cycle concurrently
### `StateStore` is responsible for:
* inserting & updating various state hierarchies
* operations are done concurrently by holding all node trees in `Arc<Mutex<_>>` containers.
### This allows `StateStore` storage to:
* Create multiple indices (Through fresh `DashMap` key insertions) pointing to the same tree by incrementing the `Arc` count and inserting a new entry per child node
* Allows independent interior mutability per state tree, decoupling unrelated states from resource contention
## Node based `TimeoutManager` (in development)
Considerations:
- only one timer is active per `StateId<K>`, State machines should not have
to keep track of `Operation::Set(Instant::now())`
emitted to notifications.
Thus, all timers should be indexable by `StateId<K>`.
- A newer `Operation::Set` for the same `StateId<K>` should override an old timer.
- A timeout should emit a `Signal` that is pertinent to the related state machine.
Approach:
* `TimeoutManager` implements a per tick polling approach to resolving
timeouts
* TimeoutManager accepts two types of inputs, set and cancel (timeout)
* Timeouts are stored within the `TimeoutLedger`
* `TimeoutLedger` contains a `BTreeMap` that indexes IDs by `Instant` and a
`HashMap` that indexes `Instant`s by ID This double indexing allows timeout
cancellations to go through without providing the `Instant` that they were
meant to remove