Akka源码分析-Remote-网络链接

　　上一篇博客中，我们分析了Akka remote模式下消息发送的过程，但细心的读者一定发现没有介绍网络相关初始化、创建链接、释放链接的过程，本文就介绍一下相关的内容。

　　网络初始化就离不开ActorSystem的初始化，毕竟ActorSystem初始化之后就可以创建Actor并发送远程消息了。在ActorSystem初始化时，调用了RemoteActorRefProvider的init函数，init创建了Remoting这个RemoteTransport，并调用了start，而RemoteTransport作用是什么呢？

/**
 * INTERNAL API
 *
 * The remote transport is responsible for sending and receiving messages.
 * Each transport has an address, which it should provide in
 * Serialization.currentTransportInformation (thread-local) while serializing
 * actor references (which might also be part of messages). This address must
 * be available (i.e. fully initialized) by the time the first message is
 * received or when the start() method returns, whatever happens first.
 */
private[akka] abstract class RemoteTransport(val system: ExtendedActorSystem, val provider: RemoteActorRefProvider)

　　transport一般用来做网络传输层，负责收发消息。每个transport有一个本地address与之绑定，这个address一般就是我们配置的remote监听的本地地址（或者端口随机的本地地址）。

　　Remoting.start我们有分析过，简单来说就是创建了EndpointManager对象，并发送Listen消息等待返回，listen结束后发送StartupFinished消息。那么网络初始化（或绑定）的代码应该就是在Listen消息的处理过程中了，处理Listen消息时，调用了listen函数，这个之前也分析到过。

private def listens: Future[Seq[(AkkaProtocolTransport, Address, Promise[AssociationEventListener])]] = {
    /*
     * Constructs chains of adapters on top of each driver as given in configuration. The resulting structure looks
     * like the following:
     *   AkkaProtocolTransport <- Adapter <- ... <- Adapter <- Driver
     *
     * The transports variable contains only the heads of each chains (the AkkaProtocolTransport instances).
     */
    val transports: Seq[AkkaProtocolTransport] = for ((fqn, adapters, config) ← settings.Transports) yield {

      val args = Seq(classOf[ExtendedActorSystem] → context.system, classOf[Config] → config)

      // Loads the driver -- the bottom element of the chain.
      // The chain at this point:
      //   Driver
      val driver = extendedSystem.dynamicAccess
        .createInstanceFor[Transport](fqn, args).recover({

          case exception ⇒ throw new IllegalArgumentException(
            s"Cannot instantiate transport [$fqn]. " +
              "Make sure it extends [akka.remote.transport.Transport] and has constructor with " +
              "[akka.actor.ExtendedActorSystem] and [com.typesafe.config.Config] parameters", exception)

        }).get

      // Iteratively decorates the bottom level driver with a list of adapters.
      // The chain at this point:
      //   Adapter <- ... <- Adapter <- Driver
      val wrappedTransport =
        adapters.map { TransportAdaptersExtension.get(context.system).getAdapterProvider }.foldLeft(driver) {
          (t: Transport, provider: TransportAdapterProvider) ⇒
            // The TransportAdapterProvider will wrap the given Transport and returns with a wrapped one
            provider.create(t, context.system.asInstanceOf[ExtendedActorSystem])
        }

      // Apply AkkaProtocolTransport wrapper to the end of the chain
      // The chain at this point:
      //   AkkaProtocolTransport <- Adapter <- ... <- Adapter <- Driver
      new AkkaProtocolTransport(wrappedTransport, context.system, new AkkaProtocolSettings(conf), AkkaPduProtobufCodec)
    }

    // Collect all transports, listen addresses and listener promises in one future
    Future.sequence(transports.map { transport ⇒
      transport.listen map { case (address, listenerPromise) ⇒ (transport, address, listenerPromise) }
    })
  }

　　我们来着重分析下listen函数，它读取了settings.Transports并为之创建了对应的AkkaProtocolTransport对象，settings.Transports是什么呢？

val Transports: immutable.Seq[(String, immutable.Seq[String], Config)] = transportNames.map { name ⇒
    val transportConfig = transportConfigFor(name)
    (
      transportConfig.getString("transport-class"),
      immutableSeq(transportConfig.getStringList("applied-adapters")).reverse,
      transportConfig)
  }

  val Adapters: Map[String, String] = configToMap(getConfig("akka.remote.adapters"))

  private def transportNames: immutable.Seq[String] = immutableSeq(getStringList("akka.remote.enabled-transports"))

　　其实就是transport-class，applied-adapters，transportConfig三元组的seq。通过源码中的reference.conf，知道transport-class的值默认是akka.remote.transport.netty.NettyTransport，applied-adapters为空。

      # The class given here must implement the akka.remote.transport.Transport
      # interface and offer a public constructor which takes two arguments:
      #  1) akka.actor.ExtendedActorSystem
      #  2) com.typesafe.config.Config
      transport-class = "akka.remote.transport.netty.NettyTransport"

      # Transport drivers can be augmented with adapters by adding their
      # name to the applied-adapters list. The last adapter in the
      # list is the adapter immediately above the driver, while
      # the first one is the top of the stack below the standard
      # Akka protocol
      applied-adapters = []

　　NettyTransport我们应该知道具体是做啥的了吧，就是用来具体创建链接、监听链接状态，收发消息的。另外adapters默认为空，所以这段for代码，就是加载了transport-class对应类的实例，然后把它作为参数传给了AkkaProtocolTransport。由于可以配置多个transport并监听不同的地址，所以这里是一个seq，不过默认只有一个。

　　listen函数的最后一行调用了所有transport的listen函数，并返回transport/address/listenerPromise三元组seq。

/**
 * Implementation of the Akka protocol as a Transport that wraps an underlying Transport instance.
 *
 * Features provided by this transport are:
 *  - Soft-state associations via the use of heartbeats and failure detectors
 *  - Secure-cookie handling
 *  - Transparent origin address handling
 *  - pluggable codecs to encode and decode Akka PDUs
 */
private[remote] class AkkaProtocolTransport(
  wrappedTransport:     Transport,
  private val system:   ActorSystem,
  private val settings: AkkaProtocolSettings,
  private val codec:    AkkaPduCodec) extends ActorTransportAdapter(wrappedTransport, system)

　　AkkaProtocolTransport其实是对底层协议的代理、封装，并提供其他一些特性，例如心跳和失败检测，编解码插件等。那AkkaProtocolTransport的listen在哪里实现呢？

override def listen: Future[(Address, Promise[AssociationEventListener])] = {
    val upstreamListenerPromise: Promise[AssociationEventListener] = Promise()

    for {
      (listenAddress, listenerPromise) ← wrappedTransport.listen
      // Enforce ordering between the signalling of "listen ready" to upstream
      // and initialization happening in interceptListen
      _ ← listenerPromise.tryCompleteWith(interceptListen(listenAddress, upstreamListenerPromise.future)).future
    } yield (augmentScheme(listenAddress), upstreamListenerPromise)
  }

　　通过它的继承关系我们找到了AbstractTransportAdapter，其实也就是调用wrappedTransport的listen，那不就是在调用NettyTransport的listen么？

override def listen: Future[(Address, Promise[AssociationEventListener])] = {
    for {
      address ← addressToSocketAddress(Address("", "", settings.BindHostname, settings.BindPortSelector))
    } yield {
      try {
        val newServerChannel = inboundBootstrap match {
          case b: ServerBootstrap         ⇒ b.bind(address)
          case b: ConnectionlessBootstrap ⇒ b.bind(address)
        }

        // Block reads until a handler actor is registered
        newServerChannel.setReadable(false)
        channelGroup.add(newServerChannel)

        serverChannel = newServerChannel

        addressFromSocketAddress(newServerChannel.getLocalAddress, schemeIdentifier, system.name, Some(settings.Hostname),
          if (settings.PortSelector == 0) None else Some(settings.PortSelector)) match {
            case Some(address) ⇒
              addressFromSocketAddress(newServerChannel.getLocalAddress, schemeIdentifier, system.name, None, None) match {
                case Some(address) ⇒ boundTo = address
                case None          ⇒ throw new NettyTransportException(s"Unknown local address type [${newServerChannel.getLocalAddress.getClass.getName}]")
              }
              localAddress = address
              associationListenerPromise.future.foreach { _ ⇒ newServerChannel.setReadable(true) }
              (address, associationListenerPromise)
            case None ⇒ throw new NettyTransportException(s"Unknown local address type [${newServerChannel.getLocalAddress.getClass.getName}]")
          }
      } catch {
        case NonFatal(e) ⇒ {
          log.error("failed to bind to {}, shutting down Netty transport", address)
          try { shutdown() } catch { case NonFatal(e) ⇒ } // ignore possible exception during shutdown
          throw e
        }
      }
    }
  }

def addressToSocketAddress(addr: Address): Future[InetSocketAddress] = addr match {
    case Address(_, _, Some(host), Some(port)) ⇒ Future { blocking { new InetSocketAddress(InetAddress.getByName(host), port) } }
    case _                                     ⇒ Future.failed(new IllegalArgumentException(s"Address [$addr] does not contain host or port information."))
  }

　　简单来说就是根据address配置去创建一个InetSocketAddress并进行绑定、创建chennel等其他网络初始化。

　　至此EndpointManager对Listen消息处理完毕，就是初始化网络状态，进行监听，由于最后还收到了StartupFinished，所以EndpointManager进入了accepting状态。

    case StartupFinished ⇒
      context.become(accepting)

　　EndpointManager.accepting之前只分析了对Send消息的处理，其实还有一个非常重要的消息处理过程：接收连接请求、并创建链接。也就是对InboundAssociation消息的处理。那InboundAssociation是如何产生的呢？这还需要回过头分析NettyTransport。其中有一个字段，在listen中也用到了：inboundBootstrap

private val inboundBootstrap: Bootstrap = settings.TransportMode match {
    case Tcp ⇒ setupBootstrap(new ServerBootstrap(serverChannelFactory), serverPipelineFactory)
    case Udp ⇒ setupBootstrap(new ConnectionlessBootstrap(serverChannelFactory), serverPipelineFactory)
  }

　　根据mode判断创建哪种类型的Bootstrap。

val TransportMode: Mode = getString("transport-protocol") match {
    case "tcp"   ⇒ Tcp
    case "udp"   ⇒ Udp
    case unknown ⇒ throw new ConfigurationException(s"Unknown transport: [$unknown]")
  }

　　很显然是创建ServerBootstrap。此过程中还传入了一个非常关键的对象：serverPipelineFactory。

private val serverPipelineFactory: ChannelPipelineFactory = new ChannelPipelineFactory {
    override def getPipeline: ChannelPipeline = {
      val pipeline = newPipeline
      if (EnableSsl) pipeline.addFirst("SslHandler", sslHandler(isClient = false))
      val handler = if (isDatagram) new UdpServerHandler(NettyTransport.this, associationListenerPromise.future)
      else new TcpServerHandler(NettyTransport.this, associationListenerPromise.future, log)
      pipeline.addLast("ServerHandler", handler)
      pipeline
    }
  }

　　根据上下文以及netty基础概念得知，ChannelPipelineFactory，而getPipeline中addLast函数添加了一个非常重要的handler：TcpServerHandler。

private[remote] class TcpServerHandler(_transport: NettyTransport, _associationListenerFuture: Future[AssociationEventListener], val log: LoggingAdapter)
  extends ServerHandler(_transport, _associationListenerFuture) with TcpHandlers {

  override def onConnect(ctx: ChannelHandlerContext, e: ChannelStateEvent): Unit =
    initInbound(e.getChannel, e.getChannel.getRemoteAddress, null)

}

　　TcpServerHandler在OnConnect时会调用initInbound。

final protected def initInbound(channel: Channel, remoteSocketAddress: SocketAddress, msg: ChannelBuffer): Unit = {
    channel.setReadable(false)
    associationListenerFuture.foreach {
      listener ⇒
        val remoteAddress = NettyTransport.addressFromSocketAddress(remoteSocketAddress, transport.schemeIdentifier,
          transport.system.name, hostName = None, port = None).getOrElse(
            throw new NettyTransportException(s"Unknown inbound remote address type [${remoteSocketAddress.getClass.getName}]"))
        init(channel, remoteSocketAddress, remoteAddress, msg) { listener notify InboundAssociation(_) }
    }
  }

　　initInbound会调用init函数，init源码如下：

final protected def init(channel: Channel, remoteSocketAddress: SocketAddress, remoteAddress: Address, msg: ChannelBuffer)(
    op: (AssociationHandle ⇒ Any)): Unit = {
    import transport._
    NettyTransport.addressFromSocketAddress(channel.getLocalAddress, schemeIdentifier, system.name, Some(settings.Hostname), None) match {
      case Some(localAddress) ⇒
        val handle = createHandle(channel, localAddress, remoteAddress)
        handle.readHandlerPromise.future.foreach {
          listener ⇒
            registerListener(channel, listener, msg, remoteSocketAddress.asInstanceOf[InetSocketAddress])
            channel.setReadable(true)
        }
        op(handle)

      case _ ⇒ NettyTransport.gracefulClose(channel)
    }
  }

　　init简单来说就是创建一个handle，并调用op(handle)，联系上下文我们知道op应该就是“{ listener notify InboundAssociation(_) }”这段代码。就是用InboundAssociation封装handle，并调用listener的notify！！！我们好像找到了InboundAssociation消息的来源！！！那listener究竟是在哪里赋值的呢？

这个就比较曲折了，具体过程不再展示，只说结果。就是在listens成功之后，给self发送了一个ListensResult消息，收到消息后，有一段代码是在赋值：promise.success(ActorAssociationEventListener(self))。这里的promise就是上面代码中的listener。

    case Listen(addressesPromise) ⇒
      listens map { ListensResult(addressesPromise, _) } recover {
        case NonFatal(e) ⇒ ListensFailure(addressesPromise, e)
      } pipeTo self
    case ListensResult(addressesPromise, results) ⇒
      transportMapping = results.groupBy {
        case (_, transportAddress, _) ⇒ transportAddress
      } map {
        case (a, t) if t.size > 1 ⇒
          throw new RemoteTransportException(s"There are more than one transports listening on local address [$a]", null)
        case (a, t) ⇒ a → t.head._1
      }
      // Register to each transport as listener and collect mapping to addresses
      val transportsAndAddresses = results map {
        case (transport, address, promise) ⇒
          promise.success(ActorAssociationEventListener(self))
          transport → address
      }
      addressesPromise.success(transportsAndAddresses)

　　ActorAssociationEventListener源码非常简单，就是把收到的消息发送给actor，而这里的actor就是上面代码中的self，其实就是EndpointManager。

  /**
   * Class to convert ordinary [[akka.actor.ActorRef]] instances to an AssociationEventListener. The adapter will
   * forward event objects as messages to the provided ActorRef.
   * @param actor
   */
  final case class ActorAssociationEventListener(actor: ActorRef) extends AssociationEventListener {
    override def notify(ev: AssociationEvent): Unit = actor ! ev
  }

　　至此我们就回到了EndpointManager中对InboundAssociation消息的处理，handleInboundAssociation就不再详细分析，这应该就是建立连接的过程。至此我们就可以正常的发送消息了，毕竟本地服务的socket已经与远程ActorSystem对应的socket建立了链接。但远程ActorSystem网络对象收到消息之后如何分发给指定的Actor呢？我们下篇博客继续分析。