akka还有一个不常使用、但我觉得比较方便的一个模块，那就是FSM（有限状态机）。我们知道了akka中Actor模型的具体实现之后，就会发现，Akka的actor可以非常方便的实现FSM。其实在akka中实现FSM还是非常简单的，或者说我们在akka中有意无意的都在使用FSM。
在介绍分析akka中FSM源码之前，先简单介绍下FSM基本概念。一个有限状态机是一个设备，或者是一个设备模型，具有有限数量的状态，它可以在任何给定的时间根据输入进行操作，使得一个状态变换到另一个状态，或者是使一个输入或者一种行为的发生。一个有限状态机在任何瞬间只能处在一种状态。FSM有几个重要的概念：状态、行为、状态变换条件。状态就是当前FSM所处的状态，行为就是当前状态下FSM对输入的响应，状态变换就是指FSM在什么条件下需要更改当前的状态。
那么如何将FSM映射到akka中呢？其实非常简单。FSM就是一个actor；状态就是actor中的某个变量的值；行为其实可以定义为某个状态下的receive函数；状态转移，定义成某个函数就行了，只不过什么时候转移需要开发者自定义。当然这只是我们在没有分析源码之前的想当然，那具体是怎么样的呢？
/**
* This captures all of the managed state of the [[akka.actor.FSM]]: the state
* name, the state data, possibly custom timeout, stop reason and replies
* accumulated while processing the last message.
*/
case class State[S, D](stateName: S, stateData: D, timeout: Option[FiniteDuration] = None, stopReason: Option[Reason] = None, replies: List[Any] = Nil)
上面是状态的定义，有两个类型参数:S、D。分别是状态名称和状态数据，其中状态名称比较容易理解，状态数据可以认为是与当前状态关联的值。timeout标志当前状态持续的最长时间，这一点比较特殊，普通的状态机，如果没有主动更改状态会一直持续在当前状态的。stopReason当前状态停止的原因。replies暂时不做分析，可以理解为一种通知机制，退出当前状态时对listener的通知。
/**
* All messages sent to the [[akka.actor.FSM]] will be wrapped inside an
* `Event`, which allows pattern matching to extract both state and data.
*/
final case class Event[D](event: Any, stateData: D) extends NoSerializationVerificationNeeded
FSM还定义了一个Event，用来对用户数据的封装，用来做模式匹配。
type StateFunction = scala.PartialFunction[Event, State]
上面是FSM行为的定义，比较简单，就是一个PartialFunction，它输入一个Event，输出下一个状态。其实从这一点也可以看出，转义状态是需要开发手动指定的。
/*
* A fsm hakker is an awesome dude or dudette who either thinks about hacking or has to eat ;-)
*/
class FSMHakker(name: String, left: ActorRef, right: ActorRef) extends Actor with FSM[FSMHakkerState, TakenChopsticks] { //All hakkers start waiting
startWith(Waiting, TakenChopsticks(None, None)) when(Waiting) {
case Event(Think, _) =>
println("%s starts to think".format(name))
startThinking(5.seconds)
} //When a hakker is thinking it can become hungry
//and try to pick up its chopsticks and eat
when(Thinking) {
case Event(StateTimeout, _) =>
left ! Take
right ! Take
goto(Hungry)
} // When a hakker is hungry it tries to pick up its chopsticks and eat
// When it picks one up, it goes into wait for the other
// If the hakkers first attempt at grabbing a chopstick fails,
// it starts to wait for the response of the other grab
when(Hungry) {
case Event(Taken(`left`), _) =>
goto(WaitForOtherChopstick) using TakenChopsticks(Some(left), None)
case Event(Taken(`right`), _) =>
goto(WaitForOtherChopstick) using TakenChopsticks(None, Some(right))
case Event(Busy(_), _) =>
goto(FirstChopstickDenied)
} // When a hakker is waiting for the last chopstick it can either obtain it
// and start eating, or the other chopstick was busy, and the hakker goes
// back to think about how he should obtain his chopsticks :-)
when(WaitForOtherChopstick) {
case Event(Taken(`left`), TakenChopsticks(None, Some(right))) => startEating(left, right)
case Event(Taken(`right`), TakenChopsticks(Some(left), None)) => startEating(left, right)
case Event(Busy(chopstick), TakenChopsticks(leftOption, rightOption)) =>
leftOption.foreach(_ ! Put)
rightOption.foreach(_ ! Put)
startThinking(10.milliseconds)
} private def startEating(left: ActorRef, right: ActorRef): State = {
println("%s has picked up %s and %s and starts to eat".format(name, left.path.name, right.path.name))
goto(Eating) using TakenChopsticks(Some(left), Some(right)) forMax (5.seconds)
} // When the results of the other grab comes back,
// he needs to put it back if he got the other one.
// Then go back and think and try to grab the chopsticks again
when(FirstChopstickDenied) {
case Event(Taken(secondChopstick), _) =>
secondChopstick ! Put
startThinking(10.milliseconds)
case Event(Busy(chopstick), _) =>
startThinking(10.milliseconds)
} // When a hakker is eating, he can decide to start to think,
// then he puts down his chopsticks and starts to think
when(Eating) {
case Event(StateTimeout, _) =>
println("%s puts down his chopsticks and starts to think".format(name))
left ! Put
right ! Put
startThinking(5.seconds)
} // Initialize the hakker
initialize() private def startThinking(duration: FiniteDuration): State = {
goto(Thinking) using TakenChopsticks(None, None) forMax duration
}
}
上面是官方的一个demo，有几个重要的点需要注意。startWith用来定义状态机起始状态；when用来定义某个状态下对Event的处理行为；goto用来转移状态；using用来定义转移状态时，当前状态关联的值；initialize用来初始化。下面我们逐步分析FSM这些概念的实现源码。
/**
* Set initial state. Call this method from the constructor before the [[#initialize]] method.
* If different state is needed after a restart this method, followed by [[#initialize]], can
* be used in the actor life cycle hooks [[akka.actor.Actor#preStart]] and [[akka.actor.Actor#postRestart]].
*
* @param stateName initial state designator
* @param stateData initial state data
* @param timeout state timeout for the initial state, overriding the default timeout for that state
*/
final def startWith(stateName: S, stateData: D, timeout: Timeout = None): Unit =
currentState = FSM.State(stateName, stateData, timeout)
private var currentState: State = _
startWith实现非常简单，就是给currentState赋值，官网注释也说了，必须在initialize之前调用。挺不喜欢这种设计的，居然需要跟调用顺序绑定。
/**
* Insert a new StateFunction at the end of the processing chain for the
* given state. If the stateTimeout parameter is set, entering this state
* without a differing explicit timeout setting will trigger a StateTimeout
* event; the same is true when using #stay.
*
* @param stateName designator for the state
* @param stateTimeout default state timeout for this state
* @param stateFunction partial function describing response to input
*/
final def when(stateName: S, stateTimeout: FiniteDuration = null)(stateFunction: StateFunction): Unit =
register(stateName, stateFunction, Option(stateTimeout))
private def register(name: S, function: StateFunction, timeout: Timeout): Unit = {
if (stateFunctions contains name) {
stateFunctions(name) = stateFunctions(name) orElse function
stateTimeouts(name) = timeout orElse stateTimeouts(name)
} else {
stateFunctions(name) = function
stateTimeouts(name) = timeout
}
}
/*
* State definitions
*/
private val stateFunctions = mutable.Map[S, StateFunction]()
private val stateTimeouts = mutable.Map[S, Timeout]()
综合上面三段代码，其实when方法就是把状态名与对应的行为之间的对应关系保存到一个map中，但需要注意的是，map中保存的是状态行为的一个合并，如果对同一个状态多次调用when，则每次注册的函数会依次调用，如果已经命中某个Event则不会再往后传递执行。
/**
* Produce transition to other state.
* Return this from a state function in order to effect the transition.
*
* This method always triggers transition events, even for `A -> A` transitions.
* If you want to stay in the same state without triggering an state transition event use [[#stay]] instead.
*
* @param nextStateName state designator for the next state
* @return state transition descriptor
*/
final def goto(nextStateName: S): State = FSM.State(nextStateName, currentState.stateData)
goto看起来非常简单，它就是构造了下一个状态，那么为啥状态就会转移了呢？这个后面再分析。官网注释里面说，调用这个方法时会触发状态转移，而且会触发转移事件，即使是同状态的转义；如果不想触发状态转移事件，则可以使用stay方法，那么stay方法是什么呢？
/**
* Produce "empty" transition descriptor.
* Return this from a state function when no state change is to be effected.
*
* No transition event will be triggered by [[#stay]].
* If you want to trigger an event like `S -> S` for `onTransition` to handle use `goto` instead.
*
* @return descriptor for staying in current state
*/
final def stay(): State = goto(currentState.stateName).withNotification(false)
它居然还是调用了goto，不同的是调用了withNotification。
private[akka] def withNotification(notifies: Boolean): State[S, D] = {
if (notifies)
State(stateName, stateData, timeout, stopReason, replies)
else
new SilentState(stateName, stateData, timeout, stopReason, replies)
}
}
而当notifies是false时，withNotification返回了一个SilentState
private[akka] class SilentState[S, D](_stateName: S, _stateData: D, _timeout: Option[FiniteDuration], _stopReason: Option[Reason], _replies: List[Any])
extends State[S, D](_stateName, _stateData, _timeout, _stopReason, _replies) { /**
* INTERNAL API
*/
private[akka] override def notifies: Boolean = false override def copy(stateName: S = stateName, stateData: D = stateData, timeout: Option[FiniteDuration] = timeout, stopReason: Option[Reason] = stopReason, replies: List[Any] = replies): State[S, D] = {
new SilentState(stateName, stateData, timeout, stopReason, replies)
}
}
SilentState与State的唯一区别就是notifies的值是false。其实简单来说，goto和stay都是创建了下一个状态的值。
分析到这里，startWith/when/StateFunction/goto基本就可以定义一个状态机了。那么FSM是如何进行状态的变换，或者说是如何运行的呢？其实吧，如果你对actor比较了解，就一定知道，这是在receive里面实现的。
override def receive: Receive = {
case TimeoutMarker(gen) ⇒
if (generation == gen) {
processMsg(StateTimeout, "state timeout")
}
case t @ Timer(name, msg, repeat, gen, owner) ⇒
if ((owner eq this) && (timers contains name) && (timers(name).generation == gen)) {
if (timeoutFuture.isDefined) {
timeoutFuture.get.cancel()
timeoutFuture = None
}
generation += 1
if (!repeat) {
timers -= name
}
processMsg(msg, t)
}
case SubscribeTransitionCallBack(actorRef) ⇒
// TODO Use context.watch(actor) and receive Terminated(actor) to clean up list
listeners.add(actorRef)
// send current state back as reference point
actorRef ! CurrentState(self, currentState.stateName)
case Listen(actorRef) ⇒
// TODO Use context.watch(actor) and receive Terminated(actor) to clean up list
listeners.add(actorRef)
// send current state back as reference point
actorRef ! CurrentState(self, currentState.stateName)
case UnsubscribeTransitionCallBack(actorRef) ⇒
listeners.remove(actorRef)
case Deafen(actorRef) ⇒
listeners.remove(actorRef)
case value ⇒ {
if (timeoutFuture.isDefined) {
timeoutFuture.get.cancel()
timeoutFuture = None
}
generation += 1
processMsg(value, sender())
}
}
我们只关注最后一段代码processMsg(value, sender())。
private def processMsg(value: Any, source: AnyRef): Unit = {
val event = Event(value, currentState.stateData)
processEvent(event, source)
}
private[akka] def processEvent(event: Event, source: AnyRef): Unit = {
val stateFunc = stateFunctions(currentState.stateName)
val nextState = if (stateFunc isDefinedAt event) {
stateFunc(event)
} else {
// handleEventDefault ensures that this is always defined
handleEvent(event)
}
applyState(nextState)
}
private[akka] def applyState(nextState: State): Unit = {
nextState.stopReason match {
case None ⇒ makeTransition(nextState)
case _ ⇒
nextState.replies.reverse foreach { r ⇒ sender() ! r }
terminate(nextState)
context.stop(self)
}
}
private[akka] def makeTransition(nextState: State): Unit = {
if (!stateFunctions.contains(nextState.stateName)) {
terminate(stay withStopReason Failure("Next state %s does not exist".format(nextState.stateName)))
} else {
nextState.replies.reverse foreach { r ⇒ sender() ! r }
if (currentState.stateName != nextState.stateName || nextState.notifies) {
this.nextState = nextState
handleTransition(currentState.stateName, nextState.stateName)
gossip(Transition(self, currentState.stateName, nextState.stateName))
this.nextState = null
}
currentState = nextState
def scheduleTimeout(d: FiniteDuration): Some[Cancellable] = {
import context.dispatcher
Some(context.system.scheduler.scheduleOnce(d, self, TimeoutMarker(generation)))
}
currentState.timeout match {
case SomeMaxFiniteDuration ⇒ // effectively disable stateTimeout
case Some(d: FiniteDuration) if d.length >= 0 ⇒ timeoutFuture = scheduleTimeout(d)
case _ ⇒
val timeout = stateTimeouts(currentState.stateName)
if (timeout.isDefined) timeoutFuture = scheduleTimeout(timeout.get)
}
}
}
processMsg方法对用户输入的消息，进行了简单的包装，然后调用processEvent；processEvent根据当前状态名找到对应的状态函数，处理当前事件，状态函数返回的下一个状态（是用goto或stay返回的）传给applyState，进行状态转移。
下一个状态没有stopReason则进行正常的状态转移，否则就停止当前状态机。makeTransition是最终处理状态转移的方法，其实也比较简单，会首先判断是否需要通知，需要则通知相关的listener；然后更新当前状态为最新的值，当然了，还会处理超时时间。
其实分析道这里，FSM就没必要再深入研究了，非常简单。就是定义了一些DSL，用trait Actor的receive函数进行行为函数的调用和状态转移，如果你的状态不多，转移条件简单，完全没必要用FSM，自己用beconme/unbecome实现就行了。而且官方的FSM是用一个函数调用列表（或者说是一个函数指针，C++里面经常这样实现）实现的，并没有用beconme/unbecome，这一点倒让我挺意外的。当然了，如果你的场景确实是FSM，那最好还是用官方的实现，毕竟它已经把FSM可能遇到的问题都帮你处理好了。
但有一点需要说明那就是akka的FSM对超时的处理机制，完全需要自己处理，比如你通过forMax给某个状态设置了超时时间，那么在该状态对应的行为中你需要处理Event（FSM.StateTimeout）消息，否则akka是不知道在某个状态超时之后应该转移到哪个状态的。akka所做的就是在某个状态指定时间内没有收到消息时，发送一个超时事件给你，而指定的时间内收到消息，则会取消上一次的超时设置，设置新的超时时间。