Akka源码分析-Remote-发消息

　　上一篇博客我们介绍了remote模式下Actor的创建，其实与local的创建并没有太大区别，一般情况下还是使用LocalActorRef创建了Actor。那么发消息是否意味着也是相同的呢？

　　既然actorOf还是委托给了LocalActorRef，那么在本地创建的Actor发消息还是跟以前一样的，那么如果如何给远程的Actor发消息呢？我们一般是通过actorSelection或者给远程Actor发送一个Identify消息，来接收对应的ActorRef，然后再发消息。我们来分析一下这两者的区别。

　　首先来看actorSelection，不管是用ActorSystem或者ActorContext的actorSelection方法，最终都是调用了ActorRefFactory对应的方法。

/**
   * Construct an [[akka.actor.ActorSelection]] from the given path, which is
   * parsed for wildcards (these are replaced by regular expressions
   * internally). No attempt is made to verify the existence of any part of
   * the supplied path, it is recommended to send a message and gather the
   * replies in order to resolve the matching set of actors.
   */
  def actorSelection(path: String): ActorSelection = path match {
    case RelativeActorPath(elems) ⇒
      if (elems.isEmpty) ActorSelection(provider.deadLetters, "")
      else if (elems.head.isEmpty) ActorSelection(provider.rootGuardian, elems.tail)
      else ActorSelection(lookupRoot, elems)
    case ActorPathExtractor(address, elems) ⇒
      ActorSelection(provider.rootGuardianAt(address), elems)
    case _ ⇒
      ActorSelection(provider.deadLetters, "")
  }

　　我们发现它支持两种类型的path：RelativeActorPath、ActorPathExtractor。

/**
 * Extractor for so-called “relative actor paths” as in “relative URI”, not in
 * “relative to some actor”. Examples:
 *
 *  * "grand/child"
 *  * "/user/hello/world"
 */
object RelativeActorPath extends PathUtils {
  def unapply(addr: String): Option[immutable.Seq[String]] = {
    try {
      val uri = new URI(addr)
      if (uri.isAbsolute) None
      else Some(split(uri.getRawPath, uri.getRawFragment))
    } catch {
      case _: URISyntaxException ⇒ None
    }
  }
}

　　RelativeActorPath提取器比较简单，就是创建了一个URI对象，然后判断其是否为Absolute，如果是就返回None，如果不是就返回对应的elemes。对于远程Actor，我们一般会指定主机名、端口号，例如akka.tcp://actorSystemName@10.0.0.1:2552/user/actorName，根据URI的定义，这个URI的schema是akka.tcp，很显然是Absolute，那就会返回None。

/**
 * Given an ActorPath it returns the Address and the path elements if the path is well-formed
 */
object ActorPathExtractor extends PathUtils {
  def unapply(addr: String): Option[(Address, immutable.Iterable[String])] =
    try {
      val uri = new URI(addr)
      uri.getRawPath match {
        case null ⇒ None
        case path ⇒ AddressFromURIString.unapply(uri).map((_, split(path, uri.getRawFragment).drop(1)))
      }
    } catch {
      case _: URISyntaxException ⇒ None
    }
}

　　ActorPathExtractor这个提取器的名称定义的是有问题的，既然actorSelection只支持两种类型的路径选择：本地和远程。第一个解析器定义成相对路径，那么后面一个就直接是绝对路径好了啊，为啥用ActorPathExtractor这样蹩脚的命名？难道本地模式下，就不是ActorPath提取器了？我们来看看对于akka.tcp://actorSystemName@10.0.0.1:2552/user/actorName提取出了什么。经调试，address是akka.tcp://actorSystemName@10.0.0.1:2552，elems就是后面的user、actorName了。

　　也就是说remote模式下，如果有host、prot等信息就会返回ActorSelection(provider.rootGuardianAt(address), elems)这个类。不过好像无论哪种情况都返回这个类，好尴尬啊，但传入的第一个参数是不同的：provider.rootGuardianAt(address)。也就是说actorSelection这个函数是不区分当前的模式的，只要含有host/port就会传入provider.rootGuardianAt(address)，否则就传入provider.rootGuardian。如果在local模式下，也强制用actorSelection查找远程Actor会发生什么呢？我们来看看LocalActorRefProvider。

  override def rootGuardianAt(address: Address): ActorRef =
    if (address == rootPath.address) rootGuardian
    else deadLetters

　　local模式下，如果待查询actor的地址就是本地地址，则直接在本地返回查找；否则就返回deadLetters。其实是无法查找远程actor的。那么RemoteActorRefProvider呢？

def rootGuardianAt(address: Address): ActorRef = {
    if (hasAddress(address)) rootGuardian
    else try {
      new RemoteActorRef(transport, transport.localAddressForRemote(address),
        RootActorPath(address), Nobody, props = None, deploy = None)
    } catch {
      case NonFatal(e) ⇒
        log.error(e, "No root guardian at [{}]", address)
        new EmptyLocalActorRef(this, RootActorPath(address), eventStream)
    }
  }

　　当然了，它也会判断一下本地地址是否包含待查询地址（防止多网卡或其他特殊情况），如果包含，则意味着是本地Actor交给rootGuardian；否则就创建RemoteActorRef。

　　分析到这里我们知道了，其实在remote模式下，actorSelection返回了一个RemoteActorRef，还记得这个类的作用嘛？我们之前简单分析过，它其实是对远程Acotor的一个本地网络代理，也就是说所有通过actorSelection发送给远程actor的消息，都会经过他中转。

　　我们继续分析ActorSelection的源码

/**
   * Construct an ActorSelection from the given string representing a path
   * relative to the given target. This operation has to create all the
   * matching magic, so it is preferable to cache its result if the
   * intention is to send messages frequently.
   */
  def apply(anchorRef: ActorRef, elements: Iterable[String]): ActorSelection = {
    val compiled: immutable.IndexedSeq[SelectionPathElement] = elements.collect({
      case x if !x.isEmpty ⇒
        if ((x.indexOf('?') != -1) || (x.indexOf('*') != -1)) SelectChildPattern(x)
        else if (x == "..") SelectParent
        else SelectChildName(x)
    })(scala.collection.breakOut)
    new ActorSelection with ScalaActorSelection {
      override val anchor = anchorRef
      override val path = compiled
    }
  }

　　很显然这里的anchorRef是上面创建的RemoteActorRef实例，其中ActorSelection的anchor（锚定）是anchorRef。至此，一个ActorSelection创建完毕。那么如何发消息呢？这就需要分析tell或者！方法了。

  def tell(msg: Any, sender: ActorRef): Unit =
    ActorSelection.deliverSelection(anchor.asInstanceOf[InternalActorRef], sender,
      ActorSelectionMessage(msg, path, wildcardFanOut = false))

　　其实乍一看，我们应该明白，这就是在deliverSelection函数内部，把消息封装成ActorSelectionMessage发送给了anchor。

　　该函数首先判断sel的elements是否为空，很显然不为空，进入rec函数。该函数比较复杂而且还是一个尾递归函数，但我们知道此处的ref就是RemoteActorRef，那么RemoteActorRef是不是一个ActorRefWithCell呢？

private[akka] class RemoteActorRef private[akka] (
  remote:                RemoteTransport,
  val localAddressToUse: Address,
  val path:              ActorPath,
  val getParent:         InternalActorRef,
  props:                 Option[Props],
  deploy:                Option[Deploy])
  extends InternalActorRef with RemoteRef

　　那么rec就会走到case _的逻辑，也就是把消息转发给了前面创建的RemoteActorRef，我们来看看这个示例是如何实现tell的。

override def !(message: Any)(implicit sender: ActorRef = Actor.noSender): Unit = {
    if (message == null) throw InvalidMessageException("Message is null")
    try remote.send(message, OptionVal(sender), this) catch handleException(message, sender)
  }

　　RemoteActorRef这个类，通过remote把消息发送出去了，那么remote是什么呢？RemoteTransport是不是很熟悉？在ActorSystem启动的时候我们分析过这个对象，它是Remoting类的实例，Remoting里面send方法是怎样的呢？

override def send(message: Any, senderOption: OptionVal[ActorRef], recipient: RemoteActorRef): Unit = endpointManager match {
    case Some(manager) ⇒ manager.tell(Send(message, senderOption, recipient), sender = senderOption getOrElse Actor.noSender)
    case None          ⇒ throw new RemoteTransportExceptionNoStackTrace("Attempted to send remote message but Remoting is not running.", null)
  }

　　它又把消息转发给了manager，而manager就是endpointManager。endpointManager是不是也比较眼熟呢？前面文章中我们也见到过，这是一个EndpointManager实例，而EndpointManager是一个Actor。请注意这里用Send又对message进行了封装。EndpointManager是如何对Send消息进行反应的呢？

case s @ Send(message, senderOption, recipientRef, _) ⇒
      val recipientAddress = recipientRef.path.address

      def createAndRegisterWritingEndpoint(): ActorRef = {
        endpoints.registerWritableEndpoint(
          recipientAddress,
          uid = None,
          createEndpoint(
            recipientAddress,
            recipientRef.localAddressToUse,
            transportMapping(recipientRef.localAddressToUse),
            settings,
            handleOption = None,
            writing = true))
      }

      endpoints.writableEndpointWithPolicyFor(recipientAddress) match {
        case Some(Pass(endpoint, _)) ⇒
          endpoint ! s
        case Some(Gated(timeOfRelease)) ⇒
          if (timeOfRelease.isOverdue()) createAndRegisterWritingEndpoint() ! s
          else extendedSystem.deadLetters ! s
        case Some(Quarantined(uid, _)) ⇒
          // timeOfRelease is only used for garbage collection reasons, therefore it is ignored here. We still have
          // the Quarantined tombstone and we know what UID we don't want to accept, so use it.
          createAndRegisterWritingEndpoint() ! s
        case None ⇒
          createAndRegisterWritingEndpoint() ! s

      }

　　分析以上逻辑，简单来看，会先判断是不是存在一个endpoint，如果存在说明链接已经建立，可以直接发送，否则出于其他状态，就重新创建endpoint，然后把消息转发给该endpoint。

def registerWritableEndpoint(address: Address, uid: Option[Int], endpoint: ActorRef): ActorRef =
      addressToWritable.get(address) match {
        case Some(Pass(e, _)) ⇒
          throw new IllegalArgumentException(s"Attempting to overwrite existing endpoint [$e] with [$endpoint]")
        case _ ⇒
          // note that this overwrites Quarantine marker,
          // but that is ok since we keep the quarantined uid in addressToRefuseUid
          addressToWritable += address → Pass(endpoint, uid)
          writableToAddress += endpoint → address
          endpoint
      }

　　registerWritableEndpoint没有太复杂的逻辑，就是查询addressToWritable这个HashMap，如果不存在则把对应的endpoint加入缓存，并返回endpoint。而endpoint是通过createEndpoint创建的。

private def createEndpoint(
    remoteAddress:    Address,
    localAddress:     Address,
    transport:        AkkaProtocolTransport,
    endpointSettings: RemoteSettings,
    handleOption:     Option[AkkaProtocolHandle],
    writing:          Boolean): ActorRef = {
    require(transportMapping contains localAddress, "Transport mapping is not defined for the address")
    // refuseUid is ignored for read-only endpoints since the UID of the remote system is already known and has passed
    // quarantine checks
    val refuseUid = endpoints.refuseUid(remoteAddress)

    if (writing) context.watch(context.actorOf(
      RARP(extendedSystem).configureDispatcher(ReliableDeliverySupervisor.props(
        handleOption,
        localAddress,
        remoteAddress,
        refuseUid,
        transport,
        endpointSettings,
        AkkaPduProtobufCodec,
        receiveBuffers)).withDeploy(Deploy.local),
      "reliableEndpointWriter-" + AddressUrlEncoder(remoteAddress) + "-" + endpointId.next()))
    else context.watch(context.actorOf(
      RARP(extendedSystem).configureDispatcher(EndpointWriter.props(
        handleOption,
        localAddress,
        remoteAddress,
        refuseUid,
        transport,
        endpointSettings,
        AkkaPduProtobufCodec,
        receiveBuffers,
        reliableDeliverySupervisor = None)).withDeploy(Deploy.local),
      "endpointWriter-" + AddressUrlEncoder(remoteAddress) + "-" + endpointId.next()))
  }

　　createEndpoint最终创建了ReliableDeliverySupervisor这个Actor，也就是说RemoteActorRef最终又把消息发送给了ReliableDeliverySupervisor，ReliableDeliverySupervisor收到消息去调用handleSend方法。

  private def handleSend(send: Send): Unit =
    if (send.message.isInstanceOf[SystemMessage]) {
      val sequencedSend = send.copy(seqOpt = Some(nextSeq()))
      tryBuffer(sequencedSend)
      // If we have not confirmed the remote UID we cannot transfer the system message at this point just buffer it.
      // GotUid will kick resendAll() causing the messages to be properly written.
      // Flow control by not sending more when we already have many outstanding.
      if (uidConfirmed && resendBuffer.nonAcked.size <= settings.SysResendLimit)
        writer ! sequencedSend
    } else writer ! send

　　除去特殊情况，用户发的普通消息又发送给了writer，艾玛我去，真是绕啊。writer是什么呢？

var writer: ActorRef = createWriter()

private def createWriter(): ActorRef = {
    context.watch(context.actorOf(RARP(context.system).configureDispatcher(EndpointWriter.props(
      handleOrActive = currentHandle,
      localAddress = localAddress,
      remoteAddress = remoteAddress,
      refuseUid,
      transport = transport,
      settings = settings,
      AkkaPduProtobufCodec,
      receiveBuffers = receiveBuffers,
      reliableDeliverySupervisor = Some(self))).withDeploy(Deploy.local), "endpointWriter"))
  }

　　很显然这又是一个ACor！！！哎，继续查找EndpointWriter这个Actor喽

def receive = if (handle.isEmpty) initializing else writing

val writing: Receive = {
    case s: Send ⇒
      if (!writeSend(s)) {
        enqueueInBuffer(s)
        scheduleBackoffTimer()
        context.become(buffering)
      }

    // We are in Writing state, so buffer is empty, safe to stop here
    case FlushAndStop ⇒
      flushAndStop()

    case AckIdleCheckTimer if ackDeadline.isOverdue() ⇒
      trySendPureAck()
  }

　　这个Actor会先判断是否已经初始化，这里就假设初始化吧，初始化之后就会进入writing这个偏函数，对send类型的消息，又调用了writeSend函数。

　　这个函数简单来看，就是调用codec对消息进行序列化，然后创建了一个pdu，最终把pdu通过handle的write发送出去。handle又是什么呢？

var handle: Option[AkkaProtocolHandle] = handleOrActive

private[remote] class AkkaProtocolHandle(
  _localAddress:          Address,
  _remoteAddress:         Address,
  val readHandlerPromise: Promise[HandleEventListener],
  _wrappedHandle:         AssociationHandle,
  val handshakeInfo:      HandshakeInfo,
  private val stateActor: ActorRef,
  private val codec:      AkkaPduCodec)
  extends AbstractTransportAdapterHandle(_localAddress, _remoteAddress, _wrappedHandle, AkkaScheme) {

  override def write(payload: ByteString): Boolean = wrappedHandle.write(codec.constructPayload(payload))

  override def disassociate(): Unit = disassociate(Unknown)

  def disassociate(info: DisassociateInfo): Unit = stateActor ! DisassociateUnderlying(info)
}

　　handle最终是一个AkkaProtocolHandle，这个对象我们不再具体分析，我们可以认为这是一个本地与远程地址链接的通道，通过这个通道就可以与远程actor发送消息了。

　　分析到这个地方，actorSelection与远程通信的过程大概就梳理清楚了。为了方便理解，作者特意辛苦的画了一个流程图，以供参考。细心的读者一定会问，那我的消息通过handle发送出去了，对方怎么接收呢？接收之后怎么发送到指定actor的邮箱呢？这一点我们后面再分析。

　　actorSelection分析清楚了，剩下的就是通过ActorRef发送消息了。那么如何得到远程Actor的ActorRef呢？当然是“问”它了啊，怎么“问”呢？发消息啊。发什么消息呢？

/**
 * A message all Actors will understand, that when processed will reply with
 * [[akka.actor.ActorIdentity]] containing the `ActorRef`. The `messageId`
 * is returned in the `ActorIdentity` message as `correlationId`.
 */
@SerialVersionUID(1L)
final case class Identify(messageId: Any) extends AutoReceivedMessage with NotInfluenceReceiveTimeout

　　官网对Identify的注释非常清楚，这个消息继承了AutoReceivedMessage，所有的Actor都理解该消息，且受到该消息后会返回akka.actor.ActorIdentity消息，里面包含当前Actor的ActorRef。那么所有的Actor为啥都理解该消息呢？

//Memory consistency is handled by the Mailbox (reading mailbox status then processing messages, then writing mailbox status
  final def invoke(messageHandle: Envelope): Unit = {
    val influenceReceiveTimeout = !messageHandle.message.isInstanceOf[NotInfluenceReceiveTimeout]
    try {
      currentMessage = messageHandle
      if (influenceReceiveTimeout)
        cancelReceiveTimeout()
      messageHandle.message match {
        case msg: AutoReceivedMessage ⇒ autoReceiveMessage(messageHandle)
        case msg                      ⇒ receiveMessage(msg)
      }
      currentMessage = null // reset current message after successful invocation
    } catch handleNonFatalOrInterruptedException { e ⇒
      handleInvokeFailure(Nil, e)
    } finally {
      if (influenceReceiveTimeout)
        checkReceiveTimeout // Reschedule receive timeout
    }
  }

  def autoReceiveMessage(msg: Envelope): Unit = {
    if (system.settings.DebugAutoReceive)
      publish(Debug(self.path.toString, clazz(actor), "received AutoReceiveMessage " + msg))

    msg.message match {
      case t: Terminated              ⇒ receivedTerminated(t)
      case AddressTerminated(address) ⇒ addressTerminated(address)
      case Kill                       ⇒ throw ActorKilledException("Kill")
      case PoisonPill                 ⇒ self.stop()
      case sel: ActorSelectionMessage ⇒ receiveSelection(sel)
      case Identify(messageId)        ⇒ sender() ! ActorIdentity(messageId, Some(self))
    }
  }　　

　　如果读者看过我之前分析的文章对上面的代码一定还有印象，它是ActorCell里面处理消息的两个函数，invoke会先判断消息类型是不是AutoReceivedMessage，如果是就自己处理了，不会去调用开发者自定义的receive函数。而Identify属于AutoReceivedMessage，收到后给sender发送了ActorIdentity消息，该消息的第二个参数是当前Actor的ActorFef变量。这样本地的actor收到远程actor返回的ActorIdentity，就可以通过对方的ActorRef给它发送消息了。当然本地actor收到的ActorIdentity消息中，第二个参数应该是一个RemoteActorRef类型。如何通过RemoteActorRef发送消息，上文已经分析清楚了，其实actorSelection最终也是通过远程actor的ActorPath创建了对应的RemoteActorRef，来发送消息的。

　　至此给远程actor发消息的两种方法就讲解完毕了。其实还有第三种方式，就是在本地创建一个远程Actor，当然了最终还是需要通过RemoteActorRef发消息的，这个具体就不再详细介绍了。