Akka源码分析-local-DeathWatch
生命周期监控,也就是死亡监控,是akka编程中常用的机制。比如我们有了某个actor的ActorRef之后,希望在该actor死亡之后收到响应的消息,此时我们就可以使用watch函数达到这一目的。
class WatchActor extends Actor { val child = context.actorOf(Props.empty, "child") context.watch(child) // <-- this is the only call needed for registration var lastSender = context.system.deadLetters def receive = { case "kill" ⇒ context.stop(child); lastSender = sender() case Terminated(`child`) ⇒ lastSender ! "finished" } }
我们从官网的一个例子入手,其实DeathWatch用起来还是非常方便的,就是调用context.watch,在对应的actor由于某种原因stop之后,就会收到Terminated消息,该消息只有一个参数,那就是stop的ActorRef。看起来简单,那具体是怎么实现的呢?
/** * Registers this actor as a Monitor for the provided ActorRef. * This actor will receive a Terminated(subject) message when watched * actor is terminated. * * `watch` is idempotent if it is not mixed with `watchWith`. * * It will fail with an [[IllegalStateException]] if the same subject was watched before using `watchWith`. * To clear the termination message, unwatch first. * * *Warning*: This method is not thread-safe and must not be accessed from threads other * than the ordinary actor message processing thread, such as [[java.util.concurrent.CompletionStage]] and [[scala.concurrent.Future]] callbacks. * * @return the provided ActorRef */ def watch(subject: ActorRef): ActorRef
上面是ActorContex关于watch的官方注释,非常简单,就是watch一个actor,然后就会收到对应的Terminated消息,还说这个方法不是线程安全的。
如果读者看过我之前的源码分析文章的话,一定知道context就是ActorContext的实例,而ActorContext是ActorCell的一个功能截面,那么watch函数的具体实现应该就是在ActorCell里面了。由于ActorCell实现的接口比较多,就不再具体分析如何找到watch实现在哪个类了,直接告诉答案:dungeon.DeathWatch。
private[akka] trait DeathWatch { this: ActorCell ⇒
首先它是一个自我类型限定的trait,这种方式我之前吐槽过这里就不展开说了,来看看watch如何实现的。
override final def watch(subject: ActorRef): ActorRef = subject match { case a: InternalActorRef ⇒ if (a != self) { if (!watchingContains(a)) maintainAddressTerminatedSubscription(a) { a.sendSystemMessage(Watch(a, self)) // ➡➡➡ NEVER SEND THE SAME SYSTEM MESSAGE OBJECT TO TWO ACTORS ⬅⬅⬅ updateWatching(a, None) } else checkWatchingSame(a, None) } a }
从上面源码可以分析出几个简单的技术点:1、不能watch自身;2、如果已经被监控则调用checkWatchingSame;3、没有被监控过,就给被监控的actor发送Watch整个系统消息;4、没有监控过则更新监控信息。
/** * This map holds a [[None]] for actors for which we send a [[Terminated]] notification on termination, * ``Some(message)`` for actors for which we send a custom termination message. */ private var watching: Map[ActorRef, Option[Any]] = Map.empty
// when all actor references have uid, i.e. actorFor is removed private def watchingContains(subject: ActorRef): Boolean = watching.contains(subject) || (subject.path.uid != ActorCell.undefinedUid && watching.contains(new UndefinedUidActorRef(subject)))
判断是否已经监控过,这个具体实现比较有意思,watching是一个Map,首先判断Map中是否需包含该ActorRef;如果不包含该ActorRef,就去判断有没有UID,有UID则创建一个UndefinedUidActorRef,再去watching中判断是否包含。难道不奇怪么?既然都不包含了,创建一个UndefinedUidActorRef就有可能包含了?谁说不是呢,哈哈。其实也不是。我们来看看ActorRef是如何定义equals的。
/** * Equals takes path and the unique id of the actor cell into account. */ final override def equals(that: Any): Boolean = that match { case other: ActorRef ⇒ path.uid == other.path.uid && path == other.path case _ ⇒ false }
上面源码逻辑比较清晰,如果两个ActorRef相等,则一定是path相等,且对应的uid相等。ActorPath的判等就不再分析了,肯定是各个层次相同喽。
那么有没有可能path相同,而uid不同呢?当然可能了,如果一个actor被stop之后,再用相同的actorOf参数创建呢?此时uid是不同的,而path是相同的。
private[akka] class UndefinedUidActorRef(ref: ActorRef) extends MinimalActorRef { override val path = ref.path.withUid(ActorCell.undefinedUid) override def provider = throw new UnsupportedOperationException("UndefinedUidActorRef does not provide") }
UndefinedUidActorRef就是与原ActorRef路径相同,而uid是ActorCell.undefinedUid的一个新的ActorRef。
maintainAddressTerminatedSubscription,它会判断是不是本地actor,如果是本地actor则调用后面的block,对于远程actor会有一些特殊操作,这里不再分析。
private def updateWatching(ref: InternalActorRef, newMessage: Option[Any]): Unit = watching = watching.updated(ref, newMessage)
updateWatching比较简单,就是把要watch的actorRef插入到watching这个Map中去。你要问我这个ActorRef在Map中对应的value是啥,我也是拒绝回答的,你可以看看watchWith的用法,这里不再分析。下面我们来分析一下被监控的Actor收到Watching之后是如何做响应的。
case Watch(watchee, watcher) ⇒ addWatcher(watchee, watcher)
它命中了ActorCell.systemInvoke中的以上分支。
protected def addWatcher(watchee: ActorRef, watcher: ActorRef): Unit = { val watcheeSelf = watchee == self val watcherSelf = watcher == self if (watcheeSelf && !watcherSelf) { if (!watchedBy.contains(watcher)) maintainAddressTerminatedSubscription(watcher) { watchedBy += watcher if (system.settings.DebugLifecycle) publish(Debug(self.path.toString, clazz(actor), s"now watched by $watcher")) } } else if (!watcheeSelf && watcherSelf) { watch(watchee) } else { publish(Warning(self.path.toString, clazz(actor), "BUG: illegal Watch(%s,%s) for %s".format(watchee, watcher, self))) } }
正常情况下,会命中第一个if的第一个分支的代码,其实也比较简答,就是去watchedBy里面查找是否保存过watcher,如果没有就把它加到watchedBy里面。
private var watchedBy: Set[ActorRef] = ActorCell.emptyActorRefSet
watchedBy是一个set,也就是里面的ActorRef不重复。那如果这个actor被stop之后,啥时候通知对应的watchedBy呢?这个问题其实还是满复杂的。
如果想知道什么时候通知了watchedBy,就需要知道stop的逻辑,那么ActorCell的stop是如何实现的呢?
// ➡➡➡ NEVER SEND THE SAME SYSTEM MESSAGE OBJECT TO TWO ACTORS ⬅⬅⬅ final def stop(): Unit = try dispatcher.systemDispatch(this, Terminate()) catch handleException
stop在Dispatch这个trait里面实现,很简单,它又用当前dispatcher发送了一个Terminate消息给自己。
case Terminate() ⇒ terminate()
收到Terminate消息后,调用了terminate方法。
protected def terminate() { setReceiveTimeout(Duration.Undefined) cancelReceiveTimeout // prevent Deadletter(Terminated) messages unwatchWatchedActors(actor) // stop all children, which will turn childrenRefs into TerminatingChildrenContainer (if there are children) children foreach stop if (systemImpl.aborting) { // separate iteration because this is a very rare case that should not penalize normal operation children foreach { case ref: ActorRefScope if !ref.isLocal ⇒ self.sendSystemMessage(DeathWatchNotification(ref, true, false)) case _ ⇒ } } val wasTerminating = isTerminating if (setChildrenTerminationReason(ChildrenContainer.Termination)) { if (!wasTerminating) { // do not process normal messages while waiting for all children to terminate suspendNonRecursive() // do not propagate failures during shutdown to the supervisor setFailed(self) if (system.settings.DebugLifecycle) publish(Debug(self.path.toString, clazz(actor), "stopping")) } } else { setTerminated() finishTerminate() } }
terminate方法,逻辑清晰,它会通知子actor进行stop。那么子actor是如何stop的呢?
final def stop(actor: ActorRef): Unit = { if (childrenRefs.getByRef(actor).isDefined) { @tailrec def shallDie(ref: ActorRef): Boolean = { val c = childrenRefs swapChildrenRefs(c, c.shallDie(ref)) || shallDie(ref) } if (actor match { case r: RepointableRef ⇒ r.isStarted case _ ⇒ true }) shallDie(actor) } actor.asInstanceOf[InternalActorRef].stop() }
其实比较简单,就是判断当前actor是否存在,若存在且已经启动则调用swapChildrenRefs,最后调用这个子actor的stop()方法,进行递归stop。
override def shallDie(actor: ActorRef): ChildrenContainer = TerminatingChildrenContainer(c, Set(actor), UserRequest)
shallDie其实就是创建一个TerminatingChildrenContainer,然后去替换childrenRefs。
@tailrec final protected def setChildrenTerminationReason(reason: ChildrenContainer.SuspendReason): Boolean = { childrenRefs match { case c: ChildrenContainer.TerminatingChildrenContainer ⇒ swapChildrenRefs(c, c.copy(reason = reason)) || setChildrenTerminationReason(reason) case _ ⇒ false } }
最后一个if语句会调用setChildrenTerminationReason,此时childrenRefs已经是TerminatingChildrenContainer类型的了,所以会返回true。
private def finishTerminate() { val a = actor /* The following order is crucial for things to work properly. Only change this if you're very confident and lucky. * * Please note that if a parent is also a watcher then ChildTerminated and Terminated must be processed in this * specific order. */ try if (a ne null) a.aroundPostStop() catch handleNonFatalOrInterruptedException { e ⇒ publish(Error(e, self.path.toString, clazz(a), e.getMessage)) } finally try dispatcher.detach(this) finally try parent.sendSystemMessage(DeathWatchNotification(self, existenceConfirmed = true, addressTerminated = false)) finally try stopFunctionRefs() finally try tellWatchersWeDied() finally try unwatchWatchedActors(a) // stay here as we expect an emergency stop from handleInvokeFailure finally { if (system.settings.DebugLifecycle) publish(Debug(self.path.toString, clazz(a), "stopped")) clearActorFields(a, recreate = false) clearActorCellFields(this) actor = null } }
所以最终会调用finishTerminate,在finishTerminate代码中会去调用tellWatchersWeDied
protected def tellWatchersWeDied(): Unit = if (!watchedBy.isEmpty) { try { // Don't need to send to parent parent since it receives a DWN by default def sendTerminated(ifLocal: Boolean)(watcher: ActorRef): Unit = if (watcher.asInstanceOf[ActorRefScope].isLocal == ifLocal && watcher != parent) watcher.asInstanceOf[InternalActorRef].sendSystemMessage(DeathWatchNotification(self, existenceConfirmed = true, addressTerminated = false)) /* * It is important to notify the remote watchers first, otherwise RemoteDaemon might shut down, causing * the remoting to shut down as well. At this point Terminated messages to remote watchers are no longer * deliverable. * * The problematic case is: * 1. Terminated is sent to RemoteDaemon * 1a. RemoteDaemon is fast enough to notify the terminator actor in RemoteActorRefProvider * 1b. The terminator is fast enough to enqueue the shutdown command in the remoting * 2. Only at this point is the Terminated (to be sent remotely) enqueued in the mailbox of remoting * * If the remote watchers are notified first, then the mailbox of the Remoting will guarantee the correct order. */ watchedBy foreach sendTerminated(ifLocal = false) watchedBy foreach sendTerminated(ifLocal = true) } finally { maintainAddressTerminatedSubscription() { watchedBy = ActorCell.emptyActorRefSet } } }
tellWatchersWeDied做了什么呢?其实就是给watchedBy对应的actorRef发送DeathWatchNotification消息。请注意DeathWatchNotification的第一个参数是self,就是要stop的actor。
case DeathWatchNotification(a, ec, at) ⇒ watchedActorTerminated(a, ec, at)
而watcher收到DeathWatchNotification如何响应呢?
/** * When this actor is watching the subject of [[akka.actor.Terminated]] message * it will be propagated to user's receive. */ protected def watchedActorTerminated(actor: ActorRef, existenceConfirmed: Boolean, addressTerminated: Boolean): Unit = { watchingGet(actor) match { case None ⇒ // We're apparently no longer watching this actor. case Some(optionalMessage) ⇒ maintainAddressTerminatedSubscription(actor) { watching = removeFromMap(actor, watching) } if (!isTerminating) { self.tell(optionalMessage.getOrElse(Terminated(actor)(existenceConfirmed, addressTerminated)), actor) terminatedQueuedFor(actor) } } if (childrenRefs.getByRef(actor).isDefined) handleChildTerminated(actor) }
很明显watchedActorTerminated在当前actor处于正常状态,且已经监控了对应的actor时,会给自己发送一个Terminated(actor),或者Terminated(actor,msg)的消息。这样监控者就收到了被监控actor的Terminated消息了。
其实吧,抛开子actor状态的维护以及其他复杂的操作,简单来说就是,监控者保存自己监控了哪些actor,被监控者保存了自己被哪些actor监控了,在被监控者stop的最后一刻发送Terminated消息给监控者就好了。当然了,这还涉及到remote模式,此时就比较复杂,后面再分析。